diff --git a/go.mod b/go.mod index 545ff148..5c87a72c 100644 --- a/go.mod +++ b/go.mod @@ -15,13 +15,17 @@ require ( github.com/hashicorp/golang-lru/v2 v2.0.7 github.com/holiman/uint256 v1.2.4 github.com/ipfs/go-merkledag v0.11.0 + github.com/pkg/errors v0.9.1 github.com/rs/zerolog v1.31.0 github.com/spf13/cobra v1.7.0 github.com/stretchr/testify v1.8.4 github.com/theckman/yacspin v0.13.12 + gorgonia.org/cu v0.9.6 k8s.io/apimachinery v0.28.3 ) +replace gorgonia.org/cu => github.com/hunjixin/cu v0.0.0-20240618140529-d11ba74b75b6 + require ( dario.cat/mergo v1.0.0 // indirect github.com/Microsoft/go-winio v0.6.1 // indirect diff --git a/go.sum b/go.sum index 3b083a1d..bd31b45e 100644 --- a/go.sum +++ b/go.sum @@ -1,7 +1,9 @@ +cloud.google.com/go v0.26.0/go.mod h1:aQUYkXzVsufM+DwF1aE+0xfcU+56JwCaLick0ClmMTw= dario.cat/mergo v1.0.0 h1:AGCNq9Evsj31mOgNPcLyXc+4PNABt905YmuqPYYpBWk= dario.cat/mergo v1.0.0/go.mod h1:uNxQE+84aUszobStD9th8a29P2fMDhsBdgRYvZOxGmk= github.com/BurntSushi/toml v0.3.1 h1:WXkYYl6Yr3qBf1K79EBnL4mak0OimBfB0XUf9Vl28OQ= github.com/BurntSushi/toml v0.3.1/go.mod h1:xHWCNGjB5oqiDr8zfno3MHue2Ht5sIBksp03qcyfWMU= +github.com/BurntSushi/xgb v0.0.0-20160522181843-27f122750802/go.mod h1:IVnqGOEym/WlBOVXweHU+Q+/VP0lqqI8lqeDx9IjBqo= github.com/DataDog/zstd v1.4.5 h1:EndNeuB0l9syBZhut0wns3gV1hL8zX8LIu6ZiVHWLIQ= github.com/Microsoft/go-winio v0.5.2/go.mod h1:WpS1mjBmmwHBEWmogvA2mj8546UReBk4v8QkMxJ6pZY= github.com/Microsoft/go-winio v0.6.1 h1:9/kr64B9VUZrLm5YYwbGtUJnMgqWVOdUAXu6Migciow= @@ -13,8 +15,12 @@ github.com/StackExchange/wmi v1.2.1/go.mod h1:rcmrprowKIVzvc+NUiLncP2uuArMWLCbu9 github.com/VictoriaMetrics/fastcache v1.12.1 h1:i0mICQuojGDL3KblA7wUNlY5lOK6a4bwt3uRKnkZU40= github.com/acomagu/bufpipe v1.0.4 h1:e3H4WUzM3npvo5uv95QuJM3cQspFNtFBzvJ2oNjKIDQ= github.com/acomagu/bufpipe v1.0.4/go.mod h1:mxdxdup/WdsKVreO5GpW4+M/1CE2sMG4jeGJ2sYmHc4= +github.com/ajstarks/svgo v0.0.0-20180226025133-644b8db467af/go.mod h1:K08gAheRH3/J6wwsYMMT4xOr94bZjxIelGM0+d/wbFw= github.com/anmitsu/go-shlex v0.0.0-20200514113438-38f4b401e2be h1:9AeTilPcZAjCFIImctFaOjnTIavg87rW78vTPkQqLI8= +github.com/apache/arrow/go/arrow v0.0.0-20201229220542-30ce2eb5d4dc/go.mod h1:c9sxoIT3YgLxH4UhLOCKaBlEojuMhVYpk4Ntv3opUTQ= +github.com/apache/arrow/go/arrow v0.0.0-20210105145422-88aaea5262db/go.mod h1:c9sxoIT3YgLxH4UhLOCKaBlEojuMhVYpk4Ntv3opUTQ= github.com/armon/go-socks5 v0.0.0-20160902184237-e75332964ef5 h1:0CwZNZbxp69SHPdPJAN/hZIm0C4OItdklCFmMRWYpio= +github.com/awalterschulze/gographviz v0.0.0-20190221210632-1e9ccb565bca/go.mod h1:GEV5wmg4YquNw7v1kkyoX9etIk8yVmXj+AkDHuuETHs= github.com/benbjohnson/clock v1.1.0/go.mod h1:J11/hYXuz8f4ySSvYwY0FKfm+ezbsZBKZxNJlLklBHA= github.com/benbjohnson/clock v1.3.0 h1:ip6w0uFQkncKQ979AypyG0ER7mqUSBdKLOgAle/AT8A= github.com/beorn7/perks v1.0.1 h1:VlbKKnNfV8bJzeqoa4cOKqO6bYr3WgKZxO8Z16+hsOM= @@ -24,10 +30,18 @@ github.com/btcsuite/btcd/btcec/v2 v2.2.0 h1:fzn1qaOt32TuLjFlkzYSsBC35Q3KUjT1SwPx github.com/btcsuite/btcd/btcec/v2 v2.2.0/go.mod h1:U7MHm051Al6XmscBQ0BoNydpOTsFAn707034b5nY8zU= github.com/btcsuite/btcd/chaincfg/chainhash v1.0.1 h1:q0rUy8C/TYNBQS1+CGKw68tLOFYSNEs0TFnxxnS9+4U= github.com/bwesterb/go-ristretto v1.2.3/go.mod h1:fUIoIZaG73pV5biE2Blr2xEzDoMj7NFEuV9ekS419A0= +github.com/census-instrumentation/opencensus-proto v0.2.1/go.mod h1:f6KPmirojxKA12rnyqOA5BBL4O983OfeGPqjHWSTneU= github.com/cespare/cp v0.1.0 h1:SE+dxFebS7Iik5LK0tsi1k9ZCxEaFX4AjQmoyA+1dJk= github.com/cespare/xxhash/v2 v2.2.0 h1:DC2CZ1Ep5Y4k3ZQ899DldepgrayRUGE6BBZ/cd9Cj44= +github.com/chewxy/hm v1.0.0/go.mod h1:qg9YI4q6Fkj/whwHR1D+bOGeF7SniIP40VweVepLjg0= +github.com/chewxy/math32 v1.0.0/go.mod h1:Miac6hA1ohdDUTagnvJy/q+aNnEk16qWUdb8ZVhvCN0= +github.com/chewxy/math32 v1.0.6/go.mod h1:dOB2rcuFrCn6UHrze36WSLVPKtzPMRAQvBvUwkSsLqs= +github.com/chewxy/math32 v1.0.7-0.20210223031236-a3549c8cb6a9/go.mod h1:dOB2rcuFrCn6UHrze36WSLVPKtzPMRAQvBvUwkSsLqs= +github.com/client9/misspell v0.3.4/go.mod h1:qj6jICC3Q7zFZvVWo7KLAzC3yx5G7kyvSDkc90ppPyw= +github.com/cloudflare/cfssl v0.0.0-20190808011637-b1ec8c586c2a/go.mod h1:yMWuSON2oQp+43nFtAV/uvKQIFpSPerB57DCt9t8sSA= github.com/cloudflare/circl v1.3.3 h1:fE/Qz0QdIGqeWfnwq0RE0R7MI51s0M2E4Ga9kq5AEMs= github.com/cloudflare/circl v1.3.3/go.mod h1:5XYMA4rFBvNIrhs50XuiBJ15vF2pZn4nnUKZrLbUZFA= +github.com/cncf/udpa/go v0.0.0-20191209042840-269d4d468f6f/go.mod h1:M8M6+tZqaGXZJjfX53e64911xZQV5JYwmTeXPW+k8Sc= github.com/cockroachdb/errors v1.8.1 h1:A5+txlVZfOqFBDa4mGz2bUWSp0aHElvHX2bKkdbQu+Y= github.com/cockroachdb/logtags v0.0.0-20190617123548-eb05cc24525f h1:o/kfcElHqOiXqcou5a3rIlMc7oJbMQkeLk0VQJ7zgqY= github.com/cockroachdb/pebble v0.0.0-20230928194634-aa077af62593 h1:aPEJyR4rPBvDmeyi+l/FS/VtA00IWvjeFvjen1m1l1A= @@ -46,6 +60,7 @@ github.com/crate-crypto/go-kzg-4844 v0.3.0/go.mod h1:SBP7ikXEgDnUPONgm33HtuDZEDt github.com/cskr/pubsub v1.0.2 h1:vlOzMhl6PFn60gRlTQQsIfVwaPB/B/8MziK8FhEPt/0= github.com/cyphar/filepath-securejoin v0.2.4 h1:Ugdm7cg7i6ZK6x3xDF1oEu1nfkyfH53EtKeQYTC3kyg= github.com/cyphar/filepath-securejoin v0.2.4/go.mod h1:aPGpWjXOXUn2NCNjFvBE6aRxGGx79pTxQpKOJNYHHl4= +github.com/cznic/cc v0.0.0-20181122101902-d673e9b70d4d/go.mod h1:m3fD/V+XTB35Kh9zw6dzjMY+We0Q7PMf6LLIC4vuG9k= github.com/davecgh/go-spew v1.1.0/go.mod h1:J7Y8YcW2NihsgmVo/mv3lAwl/skON4iLHjSsI+c5H38= github.com/davecgh/go-spew v1.1.1 h1:vj9j/u1bqnvCEfJOwUhtlOARqs3+rkHYY13jYWTU97c= github.com/davecgh/go-spew v1.1.1/go.mod h1:J7Y8YcW2NihsgmVo/mv3lAwl/skON4iLHjSsI+c5H38= @@ -54,16 +69,23 @@ github.com/deckarep/golang-set/v2 v2.1.0/go.mod h1:VAky9rY/yGXJOLEDv3OMci+7wtDpO github.com/decred/dcrd/crypto/blake256 v1.0.0 h1:/8DMNYp9SGi5f0w7uCm6d6M4OU2rGFK09Y2A4Xv7EE0= github.com/decred/dcrd/dcrec/secp256k1/v4 v4.1.0 h1:HbphB4TFFXpv7MNrT52FGrrgVXF1owhMVTHFZIlnvd4= github.com/decred/dcrd/dcrec/secp256k1/v4 v4.1.0/go.mod h1:DZGJHZMqrU4JJqFAWUS2UO1+lbSKsdiOoYi9Zzey7Fc= +github.com/dustin/go-humanize v1.0.0/go.mod h1:HtrtbFcZ19U5GC7JDqmcUSB87Iq5E25KnS6fMYU6eOk= github.com/elazarl/goproxy v0.0.0-20230808193330-2592e75ae04a h1:mATvB/9r/3gvcejNsXKSkQ6lcIaNec2nyfOdlTBR2lU= github.com/emirpasic/gods v1.18.1 h1:FXtiHYKDGKCW2KzwZKx0iC0PQmdlorYgdFG9jPXJ1Bc= github.com/emirpasic/gods v1.18.1/go.mod h1:8tpGGwCnJ5H4r6BWwaV6OrWmMoPhUl5jm/FMNAnJvWQ= +github.com/envoyproxy/go-control-plane v0.9.0/go.mod h1:YTl/9mNaCwkRvm6d1a2C3ymFceY/DCBVvsKhRF0iEA4= +github.com/envoyproxy/go-control-plane v0.9.1-0.20191026205805-5f8ba28d4473/go.mod h1:YTl/9mNaCwkRvm6d1a2C3ymFceY/DCBVvsKhRF0iEA4= +github.com/envoyproxy/go-control-plane v0.9.4/go.mod h1:6rpuAdCZL397s3pYoYcLgu1mIlRU8Am5FuJP05cCM98= +github.com/envoyproxy/protoc-gen-validate v0.1.0/go.mod h1:iSmxcyjqTsJpI2R4NaDN7+kN2VEUnK/pcBlmesArF7c= github.com/ethereum/c-kzg-4844 v0.3.1 h1:sR65+68+WdnMKxseNWxSJuAv2tsUrihTpVBTfM/U5Zg= github.com/ethereum/c-kzg-4844 v0.3.1/go.mod h1:VewdlzQmpT5QSrVhbBuGoCdFJkpaJlO1aQputP83wc0= github.com/ethereum/go-ethereum v1.13.4 h1:25HJnaWVg3q1O7Z62LaaI6S9wVq8QCw3K88g8wEzrcM= github.com/ethereum/go-ethereum v1.13.4/go.mod h1:I0U5VewuuTzvBtVzKo7b3hJzDhXOUtn9mJW7SsIPB0Q= +github.com/fatih/color v1.10.0/go.mod h1:ELkj/draVOlAH/xkhN6mQ50Qd0MPOk5AAr3maGEBuJM= github.com/fatih/color v1.15.0 h1:kOqh6YHBtK8aywxGerMG2Eq3H6Qgoqeo13Bk2Mv/nBs= github.com/fatih/color v1.15.0/go.mod h1:0h5ZqXfHYED7Bhv2ZJamyIOUej9KtShiJESRwBDUSsw= github.com/fjl/memsize v0.0.0-20190710130421-bcb5799ab5e5 h1:FtmdgXiUlNeRsoNMFlKLDt+S+6hbjVMEW6RGQ7aUf7c= +github.com/fogleman/gg v1.2.1-0.20190220221249-0403632d5b90/go.mod h1:R/bRT+9gY/C5z7JzPU0zXsXHKM4/ayA+zqcVNZzPa1k= github.com/frankban/quicktest v1.14.4 h1:g2rn0vABPOOXmZUj+vbmUp0lPoXEMuhTpIluN0XL9UY= github.com/fsnotify/fsnotify v1.6.0 h1:n+5WquG0fcWoWp6xPWfHdbskMCQaFnG6PfBrh1Ky4HY= github.com/fsnotify/fsnotify v1.6.0/go.mod h1:sl3t1tCWJFWoRz9R8WJCbQihKKwmorjAbSClcnxKAGw= @@ -89,14 +111,38 @@ github.com/go-stack/stack v1.8.1/go.mod h1:dcoOX6HbPZSZptuspn9bctJ+N/CnF5gGygcUP github.com/godbus/dbus/v5 v5.0.4/go.mod h1:xhWf0FNVPg57R7Z0UbKHbJfkEywrmjJnf7w5xrFpKfA= github.com/gofrs/flock v0.8.1 h1:+gYjHKf32LDeiEEFhQaotPbLuUXjY5ZqxKgXy7n59aw= github.com/gogo/protobuf v1.2.1/go.mod h1:hp+jE20tsWTFYpLwKvXlhS1hjn+gTNwPg2I6zVXpSg4= +github.com/gogo/protobuf v1.3.1/go.mod h1:SlYgWuQ5SjCEi6WLHjHCa1yvBfUnHcTbrrZtXPKa29o= github.com/gogo/protobuf v1.3.2 h1:Ov1cvc58UF3b5XjBnZv7+opcTcQFZebYjWzi34vdm4Q= github.com/gogo/protobuf v1.3.2/go.mod h1:P1XiOD3dCwIKUDQYPy72D8LYyHL2YPYrpS2s69NZV8Q= github.com/golang-jwt/jwt/v4 v4.5.0 h1:7cYmW1XlMY7h7ii7UhUyChSgS5wUJEnm9uZVTGqOWzg= +github.com/golang/freetype v0.0.0-20170609003504-e2365dfdc4a0/go.mod h1:E/TSTwGwJL78qG/PmXZO1EjYhfJinVAhrmmHX6Z8B9k= +github.com/golang/glog v0.0.0-20160126235308-23def4e6c14b/go.mod h1:SBH7ygxi8pfUlaOkMMuAQtPIUF8ecWP5IEl/CR7VP2Q= github.com/golang/groupcache v0.0.0-20210331224755-41bb18bfe9da h1:oI5xCqsCo564l8iNU+DwB5epxmsaqB+rhGL0m5jtYqE= github.com/golang/groupcache v0.0.0-20210331224755-41bb18bfe9da/go.mod h1:cIg4eruTrX1D+g88fzRXU5OdNfaM+9IcxsU14FzY7Hc= +github.com/golang/mock v1.1.1/go.mod h1:oTYuIxOrZwtPieC+H1uAHpcLFnEyAGVDL/k47Jfbm0A= +github.com/golang/protobuf v1.2.0/go.mod h1:6lQm79b+lXiMfvg/cZm0SGofjICqVBUtrP5yJMmIC1U= +github.com/golang/protobuf v1.3.2/go.mod h1:6lQm79b+lXiMfvg/cZm0SGofjICqVBUtrP5yJMmIC1U= +github.com/golang/protobuf v1.3.3/go.mod h1:vzj43D7+SQXF/4pzW/hwtAqwc6iTitCiVSaWz5lYuqw= +github.com/golang/protobuf v1.4.0-rc.1/go.mod h1:ceaxUfeHdC40wWswd/P6IGgMaK3YpKi5j83Wpe3EHw8= +github.com/golang/protobuf v1.4.0-rc.1.0.20200221234624-67d41d38c208/go.mod h1:xKAWHe0F5eneWXFV3EuXVDTCmh+JuBKY0li0aMyXATA= +github.com/golang/protobuf v1.4.0-rc.2/go.mod h1:LlEzMj4AhA7rCAGe4KMBDvJI+AwstrUpVNzEA03Pprs= +github.com/golang/protobuf v1.4.0-rc.4.0.20200313231945-b860323f09d0/go.mod h1:WU3c8KckQ9AFe+yFwt9sWVRKCVIyN9cPHBJSNnbL67w= +github.com/golang/protobuf v1.4.0/go.mod h1:jodUvKwWbYaEsadDk5Fwe5c77LiNKVO9IDvqG2KuDX0= +github.com/golang/protobuf v1.4.1/go.mod h1:U8fpvMrcmy5pZrNK1lt4xCsGvpyWQ/VVv6QDs8UjoX8= +github.com/golang/protobuf v1.4.2/go.mod h1:oDoupMAO8OvCJWAcko0GGGIgR6R6ocIYbsSw735rRwI= +github.com/golang/protobuf v1.4.3/go.mod h1:oDoupMAO8OvCJWAcko0GGGIgR6R6ocIYbsSw735rRwI= github.com/golang/protobuf v1.5.0/go.mod h1:FsONVRAS9T7sI+LIUmWTfcYkHO4aIWwzhcaSAoJOfIk= github.com/golang/protobuf v1.5.3 h1:KhyjKVUg7Usr/dYsdSqoFveMYd5ko72D+zANwlG1mmg= github.com/golang/snappy v0.0.5-0.20220116011046-fa5810519dcb h1:PBC98N2aIaM3XXiurYmW7fx4GZkL8feAMVq7nEjURHk= +github.com/gonum/blas v0.0.0-20181208220705-f22b278b28ac/go.mod h1:P32wAyui1PQ58Oce/KYkOqQv8cVw1zAapXOl+dRFGbc= +github.com/google/flatbuffers v1.11.0/go.mod h1:1AeVuKshWv4vARoZatz6mlQ0JxURH0Kv5+zNeJKJCa8= +github.com/google/flatbuffers v1.12.0/go.mod h1:1AeVuKshWv4vARoZatz6mlQ0JxURH0Kv5+zNeJKJCa8= +github.com/google/go-cmp v0.2.0/go.mod h1:oXzfMopK8JAjlY9xF4vHSVASa0yLyX7SntLO5aqRK0M= +github.com/google/go-cmp v0.3.0/go.mod h1:8QqcDgzrUqlUb/G2PQTWiueGozuR1884gddMywk6iLU= +github.com/google/go-cmp v0.3.1/go.mod h1:8QqcDgzrUqlUb/G2PQTWiueGozuR1884gddMywk6iLU= +github.com/google/go-cmp v0.4.0/go.mod h1:v8dTdLbMG2kIc/vJvl+f65V22dbkXbowE6jgT/gNBxE= +github.com/google/go-cmp v0.5.0/go.mod h1:v8dTdLbMG2kIc/vJvl+f65V22dbkXbowE6jgT/gNBxE= +github.com/google/go-cmp v0.5.3/go.mod h1:v8dTdLbMG2kIc/vJvl+f65V22dbkXbowE6jgT/gNBxE= github.com/google/go-cmp v0.5.5/go.mod h1:v8dTdLbMG2kIc/vJvl+f65V22dbkXbowE6jgT/gNBxE= github.com/google/go-cmp v0.5.7/go.mod h1:n+brtR0CgQNWTVd5ZUFpTBC8YFBDLK/h/bpaJ8/DtOE= github.com/google/go-cmp v0.5.9 h1:O2Tfq5qg4qc4AmwVlvv0oLiVAGB7enBSJ2x2DqQFi38= @@ -107,6 +153,7 @@ github.com/google/uuid v1.1.1/go.mod h1:TIyPZe4MgqvfeYDBFedMoGGpEw/LqOeaOT+nhxU+ github.com/google/uuid v1.3.0 h1:t6JiXgmwXMjEs8VusXIJk2BXHsn+wx8BZdTaoZ5fu7I= github.com/google/uuid v1.3.0/go.mod h1:TIyPZe4MgqvfeYDBFedMoGGpEw/LqOeaOT+nhxU+yHo= github.com/gopherjs/gopherjs v0.0.0-20181017120253-0766667cb4d1 h1:EGx4pi6eqNxGaHF6qqu48+N2wcFQ5qg5FXgOdqsJ5d8= +github.com/gorgonia/bindgen v0.0.0-20210223094355-432cd89e7765/go.mod h1:BLHSe436vhQKRfm6wxJgebeK4fDY+ER/8jV3vVH9yYU= github.com/gorilla/mux v1.8.0 h1:i40aqfkR1h2SlN9hojwV5ZA91wcXFOvkdNIeFDP5koI= github.com/gorilla/mux v1.8.0/go.mod h1:DVbg23sWSpFRCP0SfiEN6jmj59UnW/n46BH5rLB71So= github.com/gorilla/websocket v1.5.0 h1:PPwGk2jz7EePpoHN/+ClbZu8SPxiqlu12wZP/3sWmnc= @@ -129,6 +176,8 @@ github.com/holiman/bloomfilter/v2 v2.0.3 h1:73e0e/V0tCydx14a0SCYS/EWCxgwLZ18CZcZ github.com/holiman/uint256 v1.2.4 h1:jUc4Nk8fm9jZabQuqr2JzednajVmBpC+oiTiXZJEApU= github.com/holiman/uint256 v1.2.4/go.mod h1:EOMSn4q6Nyt9P6efbI3bueV4e1b3dGlUCXeiRV4ng7E= github.com/huin/goupnp v1.3.0 h1:UvLUlWDNpoUdYzb2TCn+MuTWtcjXKSza2n6CBdQ0xXc= +github.com/hunjixin/cu v0.0.0-20240618140529-d11ba74b75b6 h1:cBUBkngmqWtdvTTy1YK39UPneUtIWm/gmbsiCkqzNd8= +github.com/hunjixin/cu v0.0.0-20240618140529-d11ba74b75b6/go.mod h1:nR6RAm64n9htu6Orv1NVbsMJXHjnsC3SHPfgcxI08e4= github.com/inconshreveable/mousetrap v1.1.0 h1:wN+x4NVGpMsO7ErUn/mUI3vEoE6Jt13X2s0bqwp9tc8= github.com/inconshreveable/mousetrap v1.1.0/go.mod h1:vpF70FUmC8bwa3OWnCshd2FqLfsEA9PFc4w1p2J65bw= github.com/ipfs/bbloom v0.0.4 h1:Gi+8EGJ2y5qiD5FbsbpX/TMNcJw8gSqr7eyjHa4Fhvs= @@ -194,9 +243,11 @@ github.com/jbenet/go-context v0.0.0-20150711004518-d14ea06fba99/go.mod h1:1lJo3i github.com/jbenet/goprocess v0.1.4 h1:DRGOFReOMqqDNXwW70QkacFW0YN9QnwLV0Vqk+3oU0o= github.com/jbenet/goprocess v0.1.4/go.mod h1:5yspPrukOVuOLORacaBi858NqyClJPQxYZlqdZVfqY4= github.com/jtolds/gls v4.2.1+incompatible h1:fSuqC+Gmlu6l/ZYAoZzx2pyucC8Xza35fpRVWLVmUEE= +github.com/jung-kurt/gofpdf v1.0.3-0.20190309125859-24315acbbda5/go.mod h1:7Id9E/uU8ce6rXgefFLlgrJj/GYY22cpxn+r32jIOes= github.com/kevinburke/ssh_config v1.2.0 h1:x584FjTGwHzMwvHx18PXxbBVzfnxogHaAReU4gf13a4= github.com/kevinburke/ssh_config v1.2.0/go.mod h1:CT57kijsi8u/K/BOFA39wgDQJ9CxiF4nAY/ojJ6r6mM= github.com/kisielk/errcheck v1.1.0/go.mod h1:EZBBE59ingxPouuu3KfxchcWSUPOHkagtvWXihfKN4Q= +github.com/kisielk/errcheck v1.2.0/go.mod h1:/BMXB+zMLi60iA8Vv6Ksmxu/1UDYcXs4uQLJ+jE2L00= github.com/kisielk/errcheck v1.5.0/go.mod h1:pFxgyoBC7bSaBwPgfKdkLd5X25qrDl4LWUI2bnpBCr8= github.com/kisielk/gotool v1.0.0/go.mod h1:XhKaO+MFFWcvkIS/tQcRk01m1F5IRFswLeQ+oQHNcck= github.com/klauspost/compress v1.15.15 h1:EF27CXIuDsYJ6mmvtBRlEuB2UVOqHG1tAXgZ7yIO+lw= @@ -213,6 +264,7 @@ github.com/kr/text v0.1.0/go.mod h1:4Jbv+DJW3UT/LiOwJeYQe1efqtUx/iVham/4vfdArNI= github.com/kr/text v0.2.0 h1:5Nx0Ya0ZqY2ygV366QzturHI13Jq95ApcVaJBhpS+AY= github.com/kylelemons/godebug v1.1.0 h1:RPNrshWIDI6G2gRW9EHilWtl7Z6Sb1BR0xunSBf0SNc= github.com/leanovate/gopter v0.2.9 h1:fQjYxZaynp97ozCzfOyOuAGOU4aU/z37zf/tOujFk7c= +github.com/leesper/go_rng v0.0.0-20171009123644-5344a9259b21/go.mod h1:N0SVk0uhy+E1PZ3C9ctsPRlvOPAFPkCNlcPBDkt0N3U= github.com/libp2p/go-buffer-pool v0.1.0 h1:oK4mSFcQz7cTQIfqbe4MIj9gLW+mnanjyFtc6cdF0Y8= github.com/libp2p/go-cidranger v1.1.0 h1:ewPN8EZ0dd1LSnrtuwd4709PXVcITVeuwbag38yPW7c= github.com/libp2p/go-libp2p v0.22.0 h1:2Tce0kHOp5zASFKJbNzRElvh0iZwdtG5uZheNW8chIw= @@ -226,9 +278,11 @@ github.com/libp2p/go-openssl v0.1.0 h1:LBkKEcUv6vtZIQLVTegAil8jbNpJErQ9AnT+bWV+O github.com/matryer/is v1.2.0 h1:92UTHpy8CDwaJ08GqLDzhhuixiBUUD1p3AU6PHddz4A= github.com/matryer/is v1.2.0/go.mod h1:2fLPjFQM9rhQ15aVEtbuwhJinnOqrmgXPNdZsdwlWXA= github.com/mattn/go-colorable v0.1.1/go.mod h1:FuOcm+DKB9mbwrcAfNl7/TZVBZ6rcnceauSikq3lYCQ= +github.com/mattn/go-colorable v0.1.8/go.mod h1:u6P/XSegPjTcexA+o6vUJrdnUu04hMope9wVRipJSqc= github.com/mattn/go-colorable v0.1.13 h1:fFA4WZxdEF4tXPZVKMLwD8oUnCTTo08duU7wxecdEvA= github.com/mattn/go-colorable v0.1.13/go.mod h1:7S9/ev0klgBDR4GtXTXX8a3vIGJpMovkB8vQcUbaXHg= github.com/mattn/go-isatty v0.0.5/go.mod h1:Iq45c/XA43vh69/j3iqttzPXn0bhXyGjM0Hdxcsrc5s= +github.com/mattn/go-isatty v0.0.12/go.mod h1:cbi8OIDigv2wuxKPP5vlRcQ1OAZbq2CE4Kysco4FUpU= github.com/mattn/go-isatty v0.0.14/go.mod h1:7GGIvUiUoEMVVmxf/4nioHXj79iQHKdU27kJ6hsGG94= github.com/mattn/go-isatty v0.0.16/go.mod h1:kYGgaQfpe5nmfYZH+SKPsOc2e4SrIfOl2e/yFXSvRLM= github.com/mattn/go-isatty v0.0.19 h1:JITubQf0MOLdlGRuRq+jtsDlekdYPia9ZFsB8h/APPA= @@ -289,9 +343,13 @@ github.com/pmezard/go-difflib v1.0.0/go.mod h1:iKH77koFhYxTK1pcRnkKkqfTogsbg7gZN github.com/polydawn/refmt v0.0.0-20201211092308-30ac6d18308e h1:ZOcivgkkFRnjfoTcGsDq3UQYiBmekwLA+qg0OjyB/ls= github.com/polydawn/refmt v0.0.0-20201211092308-30ac6d18308e/go.mod h1:uIp+gprXxxrWSjjklXD+mN4wed/tMfjMMmN/9+JsA9o= github.com/prometheus/client_golang v1.12.0 h1:C+UIj/QWtmqY13Arb8kwMt5j34/0Z2iKamrJ+ryC0Gg= +github.com/prometheus/client_model v0.0.0-20190812154241-14fe0d1b01d4/go.mod h1:xMI15A0UPsDsEKsMN9yxemIoYk6Tm2C1GtYGdfGttqA= github.com/prometheus/client_model v0.2.1-0.20210607210712-147c58e9608a h1:CmF68hwI0XsOQ5UwlBopMi2Ow4Pbg32akc4KIVCOm+Y= github.com/prometheus/common v0.32.1 h1:hWIdL3N2HoUx3B8j3YN9mWor0qhY/NlEKZEaXxuIRh4= github.com/prometheus/procfs v0.7.3 h1:4jVXhlkAyzOScmCkXBTOLRLTz8EeU+eyjrwB/EPq0VU= +github.com/remyoudompheng/bigfft v0.0.0-20170806203942-52369c62f446/go.mod h1:uYEyJGbgTkfkS4+E/PavXkNJcbFIpEtjt2B0KDQ5+9M= +github.com/remyoudompheng/bigfft v0.0.0-20190728182440-6a916e37a237/go.mod h1:qqbHyh8v60DhA7CoWK5oRCqLrMHRGoxYCSS9EjAz6Eo= +github.com/remyoudompheng/bigfft v0.0.0-20200410134404-eec4a21b6bb0/go.mod h1:qqbHyh8v60DhA7CoWK5oRCqLrMHRGoxYCSS9EjAz6Eo= github.com/rivo/uniseg v0.2.0 h1:S1pD9weZBuJdFmowNwbpi7BJ8TNftyUImj/0WQi72jY= github.com/rivo/uniseg v0.2.0/go.mod h1:J6wj4VEh+S6ZtnVlnTBMWIodfgj8LQOQFoIToxlJtxc= github.com/rogpeppe/go-internal v1.3.0/go.mod h1:M8bDsm7K2OlrFYOpmOWEs/qY81heoFRclV5y23lUDJ4= @@ -320,9 +378,12 @@ github.com/spf13/pflag v1.0.5 h1:iy+VFUOCP1a+8yFto/drg2CJ5u0yRoB7fZw3DKv/JXA= github.com/spf13/pflag v1.0.5/go.mod h1:McXfInJRrz4CZXVZOBLb0bTZqETkiAhM9Iw0y3An2Bg= github.com/status-im/keycard-go v0.2.0 h1:QDLFswOQu1r5jsycloeQh3bVU8n/NatHHaZobtDnDzA= github.com/stretchr/objx v0.1.0/go.mod h1:HFkY916IF+rwdDfMAkV7OtwuqBVzrE8GR6GFx+wExME= +github.com/stretchr/testify v1.1.4/go.mod h1:a8OnRcib4nhh0OaRAV+Yts87kKdq0PP7pXfy6kDkUVs= +github.com/stretchr/testify v1.2.0/go.mod h1:a8OnRcib4nhh0OaRAV+Yts87kKdq0PP7pXfy6kDkUVs= github.com/stretchr/testify v1.2.2/go.mod h1:a8OnRcib4nhh0OaRAV+Yts87kKdq0PP7pXfy6kDkUVs= github.com/stretchr/testify v1.3.0/go.mod h1:M5WIy9Dh21IEIfnGCwXGc5bZfKNJtfHm1UVUgZn+9EI= github.com/stretchr/testify v1.4.0/go.mod h1:j7eGeouHqKxXV5pUuKE4zz7dFj8WfuZ+81PSLYec5m4= +github.com/stretchr/testify v1.6.1/go.mod h1:6Fq8oRcR53rry900zMqJjRRixrwX3KX962/h/Wwjteg= github.com/stretchr/testify v1.7.0/go.mod h1:6Fq8oRcR53rry900zMqJjRRixrwX3KX962/h/Wwjteg= github.com/stretchr/testify v1.7.1/go.mod h1:6Fq8oRcR53rry900zMqJjRRixrwX3KX962/h/Wwjteg= github.com/stretchr/testify v1.8.4 h1:CcVxjf3Q8PM0mHUKJCdn+eZZtm5yQwehR5yeSVQQcUk= @@ -344,6 +405,7 @@ github.com/whyrusleeping/go-logging v0.0.0-20170515211332-0457bb6b88fc/go.mod h1 github.com/xanzy/ssh-agent v0.3.3 h1:+/15pJfg/RsTxqYcX6fHqOXZwwMP+2VyYWJeWM2qQFM= github.com/xanzy/ssh-agent v0.3.3/go.mod h1:6dzNDKs0J9rVPHPhaGCukekBHKqfl+L3KghI1Bc68Uw= github.com/xrash/smetrics v0.0.0-20201216005158-039620a65673 h1:bAn7/zixMGCfxrRTfdpNzjtPYqr8smhKouy9mxVdGPU= +github.com/xtgo/set v1.0.0/go.mod h1:d3NHzGzSa0NmB2NhFyECA+QdRp29oEn2xbT+TpeFoM8= github.com/yuin/goldmark v1.1.27/go.mod h1:3hX8gzYuyVAZsxl0MRgGTJEmQBFcNTphYh9decYSb74= github.com/yuin/goldmark v1.2.1/go.mod h1:3hX8gzYuyVAZsxl0MRgGTJEmQBFcNTphYh9decYSb74= github.com/yuin/goldmark v1.3.5/go.mod h1:mwnBkeHKe2W/ZEtQ+71ViKU8L12m81fl3OWwC1Zlc8k= @@ -367,6 +429,7 @@ go.uber.org/zap v1.16.0/go.mod h1:MA8QOfq0BHJwdXa996Y4dYkAqRKB8/1K1QMMZVaNZjQ= go.uber.org/zap v1.19.1/go.mod h1:j3DNczoxDZroyBnOT1L/Q79cfUMGZxlv/9dzN7SM1rI= go.uber.org/zap v1.22.0 h1:Zcye5DUgBloQ9BaT4qc9BnjOFog5TvBSAGkJ3Nf70c0= go.uber.org/zap v1.22.0/go.mod h1:H4siCOZOrAolnUPJEkfaSjDqyP+BDS0DdDWzwcgt3+U= +go4.org/unsafe/assume-no-moving-gc v0.0.0-20201222180813-1025295fd063/go.mod h1:FftLjUGFEDu5k8lt0ddY+HcrH/qU/0qk+H8j9/nTl3E= golang.org/x/crypto v0.0.0-20190211182817-74369b46fc67/go.mod h1:6SG95UA2DQfeDnfUPMdvaQW0Q7yPrPDi9nlGo2tz2b4= golang.org/x/crypto v0.0.0-20190308221718-c2843e01d9a2/go.mod h1:djNgcEr1/C05ACkg1iLfiJU5Ep61QUkGW8qpdssI0+w= golang.org/x/crypto v0.0.0-20190510104115-cbcb75029529/go.mod h1:yigFU9vqHzYiE8UmvKecakEJjdnWj3jj499lnFckfCI= @@ -379,9 +442,21 @@ golang.org/x/crypto v0.3.1-0.20221117191849-2c476679df9a/go.mod h1:hebNnKkNXi2Uz golang.org/x/crypto v0.7.0/go.mod h1:pYwdfH91IfpZVANVyUOhSIPZaFoJGxTFbZhFTx+dXZU= golang.org/x/crypto v0.14.0 h1:wBqGXzWJW6m1XrIKlAH0Hs1JJ7+9KBwnIO8v66Q9cHc= golang.org/x/crypto v0.14.0/go.mod h1:MVFd36DqK4CsrnJYDkBA3VC4m2GkXAM0PvzMCn4JQf4= +golang.org/x/exp v0.0.0-20180321215751-8460e604b9de/go.mod h1:CJ0aWSM057203Lf6IL+f9T1iT9GByDxfZKAQTCR3kQA= +golang.org/x/exp v0.0.0-20180807140117-3d87b88a115f/go.mod h1:CJ0aWSM057203Lf6IL+f9T1iT9GByDxfZKAQTCR3kQA= +golang.org/x/exp v0.0.0-20181106170214-d68db9428509/go.mod h1:CJ0aWSM057203Lf6IL+f9T1iT9GByDxfZKAQTCR3kQA= +golang.org/x/exp v0.0.0-20190121172915-509febef88a4/go.mod h1:CJ0aWSM057203Lf6IL+f9T1iT9GByDxfZKAQTCR3kQA= +golang.org/x/exp v0.0.0-20190125153040-c74c464bbbf2/go.mod h1:CJ0aWSM057203Lf6IL+f9T1iT9GByDxfZKAQTCR3kQA= +golang.org/x/exp v0.0.0-20190312203227-4b39c73a6495/go.mod h1:ZjyILWgesfNpC6sMxTJOJm9Kp84zZh5NQWvqDGG3Qr8= golang.org/x/exp v0.0.0-20230905200255-921286631fa9 h1:GoHiUyI/Tp2nVkLI2mCxVkOjsbSXD66ic0XW0js0R9g= golang.org/x/exp v0.0.0-20230905200255-921286631fa9/go.mod h1:S2oDrQGGwySpoQPVqRShND87VCbxmc6bL1Yd2oYrm6k= +golang.org/x/image v0.0.0-20180708004352-c73c2afc3b81/go.mod h1:ux5Hcp/YLpHSI86hEcLt0YII63i6oz57MZXIpbrjZUs= +golang.org/x/image v0.0.0-20190227222117-0694c2d4d067/go.mod h1:kZ7UVZpmo3dzQBMxlp+ypCbDeSB+sBbTgSJuh5dn5js= +golang.org/x/lint v0.0.0-20181026193005-c67002cb31c3/go.mod h1:UVdnD1Gm6xHRNCYTkRU2/jEulfH38KcIWyp/GAMgvoE= +golang.org/x/lint v0.0.0-20190227174305-5b3e6a55c961/go.mod h1:wehouNa3lNwaWXcvxsM5YxQ5yQlVC4a0KAMCusXpPoU= +golang.org/x/lint v0.0.0-20190313153728-d0100b6bd8b3/go.mod h1:6SW0HCj/g11FgYtHlgUYUwCkIfeOF89ocIRzGO/8vkc= golang.org/x/lint v0.0.0-20190930215403-16217165b5de/go.mod h1:6SW0HCj/g11FgYtHlgUYUwCkIfeOF89ocIRzGO/8vkc= +golang.org/x/mobile v0.0.0-20190312151609-d3739f865fa6/go.mod h1:z+o9i4GpDbdi3rU15maQ/Ox0txvL9dWGYEHz965HBQE= golang.org/x/mod v0.0.0-20190513183733-4bf6d317e70e/go.mod h1:mXi4GBBbnImb6dmsKGUJ2LatrhH/nqhxcFungHvyanc= golang.org/x/mod v0.2.0/go.mod h1:s0Qsj1ACt9ePp/hMypM3fl4fZqREWJwdYDEqhRiZZUA= golang.org/x/mod v0.3.0/go.mod h1:s0Qsj1ACt9ePp/hMypM3fl4fZqREWJwdYDEqhRiZZUA= @@ -390,11 +465,15 @@ golang.org/x/mod v0.6.0-dev.0.20220419223038-86c51ed26bb4/go.mod h1:jJ57K6gSWd91 golang.org/x/mod v0.8.0/go.mod h1:iBbtSCu2XBx23ZKBPSOrRkjjQPZFPuis4dIYUhu/chs= golang.org/x/mod v0.12.0 h1:rmsUpXtvNzj340zd98LZ4KntptpfRHwpFOHG188oHXc= golang.org/x/mod v0.12.0/go.mod h1:iBbtSCu2XBx23ZKBPSOrRkjjQPZFPuis4dIYUhu/chs= +golang.org/x/net v0.0.0-20180724234803-3673e40ba225/go.mod h1:mL1N/T3taQHkDXs73rZJwtUhF3w3ftmwwsq0BUmARs4= +golang.org/x/net v0.0.0-20180826012351-8a410e7b638d/go.mod h1:mL1N/T3taQHkDXs73rZJwtUhF3w3ftmwwsq0BUmARs4= +golang.org/x/net v0.0.0-20190213061140-3a22650c66bd/go.mod h1:mL1N/T3taQHkDXs73rZJwtUhF3w3ftmwwsq0BUmARs4= golang.org/x/net v0.0.0-20190227160552-c95aed5357e7/go.mod h1:mL1N/T3taQHkDXs73rZJwtUhF3w3ftmwwsq0BUmARs4= golang.org/x/net v0.0.0-20190311183353-d8887717615a/go.mod h1:t9HGtf8HONx5eT2rtn7q6eTqICYqUVnKs3thJo3Qplg= golang.org/x/net v0.0.0-20190404232315-eb5bcb51f2a3/go.mod h1:t9HGtf8HONx5eT2rtn7q6eTqICYqUVnKs3thJo3Qplg= golang.org/x/net v0.0.0-20190620200207-3b0461eec859/go.mod h1:z5CRVTTTmAJ677TzLLGU+0bjPO0LkuOLi4/5GtJWs/s= golang.org/x/net v0.0.0-20200226121028-0de0cce0169b/go.mod h1:z5CRVTTTmAJ677TzLLGU+0bjPO0LkuOLi4/5GtJWs/s= +golang.org/x/net v0.0.0-20200904194848-62affa334b73/go.mod h1:/O7V0waA8r7cgGh81Ro3o1hOxt32SMVPicZroKQ2sZA= golang.org/x/net v0.0.0-20201021035429-f5854403a974/go.mod h1:sp8m0HH+o8qH0wwXwYZr8TS3Oi6o0r6Gce1SSxlDquU= golang.org/x/net v0.0.0-20210226172049-e18ecbb05110/go.mod h1:m0MpNAwzfU5UDzcl9v0D8zg8gWTRqZa9RBIspLL5mdg= golang.org/x/net v0.0.0-20210405180319-a5a99cb37ef4/go.mod h1:p54w0d4576C0XHj96bSt6lcn1PtDYWL6XObtHCRCNQM= @@ -405,6 +484,9 @@ golang.org/x/net v0.6.0/go.mod h1:2Tu9+aMcznHK/AK1HMvgo6xiTLG5rD5rZLDS+rp2Bjs= golang.org/x/net v0.8.0/go.mod h1:QVkue5JL9kW//ek3r6jTKnTFis1tRmNAW2P1shuFdJc= golang.org/x/net v0.17.0 h1:pVaXccu2ozPjCXewfr1S7xza/zcXTity9cCdXQYSjIM= golang.org/x/net v0.17.0/go.mod h1:NxSsAGuq816PNPmqtQdLE42eU2Fs7NoRIZrHJAlaCOE= +golang.org/x/oauth2 v0.0.0-20180821212333-d2e6202438be/go.mod h1:N/0e6XlmueqKjAGxoOufVs8QHGRruUQn6yWY3a++T0U= +golang.org/x/sync v0.0.0-20180314180146-1d60e4601c6f/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM= +golang.org/x/sync v0.0.0-20181108010431-42b317875d0f/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM= golang.org/x/sync v0.0.0-20190423024810-112230192c58/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM= golang.org/x/sync v0.0.0-20190911185100-cd5d95a43a6e/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM= golang.org/x/sync v0.0.0-20201020160332-67f06af15bc9/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM= @@ -413,12 +495,18 @@ golang.org/x/sync v0.0.0-20220722155255-886fb9371eb4/go.mod h1:RxMgew5VJxzue5/jJ golang.org/x/sync v0.1.0/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM= golang.org/x/sync v0.3.0 h1:ftCYgMx6zT/asHUrPw8BLLscYtGznsLAnjq5RH9P66E= golang.org/x/sync v0.3.0/go.mod h1:FU7BRWz2tNW+3quACPkgCx/L+uEAv1htQ0V83Z9Rj+Y= +golang.org/x/sys v0.0.0-20180830151530-49385e6e1522/go.mod h1:STP8DvDyc/dI5b8T5hshtkjS+E42TnysNCUPdjciGhY= golang.org/x/sys v0.0.0-20190215142949-d0b11bdaac8a/go.mod h1:STP8DvDyc/dI5b8T5hshtkjS+E42TnysNCUPdjciGhY= golang.org/x/sys v0.0.0-20190219092855-153ac476189d/go.mod h1:STP8DvDyc/dI5b8T5hshtkjS+E42TnysNCUPdjciGhY= golang.org/x/sys v0.0.0-20190222072716-a9d3bda3a223/go.mod h1:STP8DvDyc/dI5b8T5hshtkjS+E42TnysNCUPdjciGhY= +golang.org/x/sys v0.0.0-20190312061237-fead79001313/go.mod h1:h1NjWce9XRLGQEsW7wpKNCjG9DtNlClVuFLEZdDNbEs= golang.org/x/sys v0.0.0-20190412213103-97732733099d/go.mod h1:h1NjWce9XRLGQEsW7wpKNCjG9DtNlClVuFLEZdDNbEs= golang.org/x/sys v0.0.0-20190916202348-b4ddaad3f8a3/go.mod h1:h1NjWce9XRLGQEsW7wpKNCjG9DtNlClVuFLEZdDNbEs= golang.org/x/sys v0.0.0-20191026070338-33540a1f6037/go.mod h1:h1NjWce9XRLGQEsW7wpKNCjG9DtNlClVuFLEZdDNbEs= +golang.org/x/sys v0.0.0-20200116001909-b77594299b42/go.mod h1:h1NjWce9XRLGQEsW7wpKNCjG9DtNlClVuFLEZdDNbEs= +golang.org/x/sys v0.0.0-20200223170610-d5e6a3e2c0ae/go.mod h1:h1NjWce9XRLGQEsW7wpKNCjG9DtNlClVuFLEZdDNbEs= +golang.org/x/sys v0.0.0-20200323222414-85ca7c5b95cd/go.mod h1:h1NjWce9XRLGQEsW7wpKNCjG9DtNlClVuFLEZdDNbEs= +golang.org/x/sys v0.0.0-20200909081042-eff7692f9009/go.mod h1:h1NjWce9XRLGQEsW7wpKNCjG9DtNlClVuFLEZdDNbEs= golang.org/x/sys v0.0.0-20200930185726-fdedc70b468f/go.mod h1:h1NjWce9XRLGQEsW7wpKNCjG9DtNlClVuFLEZdDNbEs= golang.org/x/sys v0.0.0-20201119102817-f84b799fce68/go.mod h1:h1NjWce9XRLGQEsW7wpKNCjG9DtNlClVuFLEZdDNbEs= golang.org/x/sys v0.0.0-20210124154548-22da62e12c0c/go.mod h1:h1NjWce9XRLGQEsW7wpKNCjG9DtNlClVuFLEZdDNbEs= @@ -458,8 +546,15 @@ golang.org/x/text v0.8.0/go.mod h1:e1OnstbJyHTd6l/uOt8jFFHp6TRDWZR/bV3emEE/zU8= golang.org/x/text v0.13.0 h1:ablQoSUd0tRdKxZewP80B+BaqeKJuVhuRxj/dkrun3k= golang.org/x/time v0.3.0 h1:rg5rLMjNzMS1RkNLzCG38eapWhnYLFYXDXj2gOlr8j4= golang.org/x/tools v0.0.0-20180221164845-07fd8470d635/go.mod h1:n7NCudcB/nEzxVGmLbDWY5pfWTLqBcC2KZ6jyYvM4mQ= +golang.org/x/tools v0.0.0-20180525024113-a5b4c53f6e8b/go.mod h1:n7NCudcB/nEzxVGmLbDWY5pfWTLqBcC2KZ6jyYvM4mQ= golang.org/x/tools v0.0.0-20180917221912-90fa682c2a6e/go.mod h1:n7NCudcB/nEzxVGmLbDWY5pfWTLqBcC2KZ6jyYvM4mQ= +golang.org/x/tools v0.0.0-20181030221726-6c7e314b6563/go.mod h1:n7NCudcB/nEzxVGmLbDWY5pfWTLqBcC2KZ6jyYvM4mQ= +golang.org/x/tools v0.0.0-20190114222345-bf090417da8b/go.mod h1:n7NCudcB/nEzxVGmLbDWY5pfWTLqBcC2KZ6jyYvM4mQ= +golang.org/x/tools v0.0.0-20190206041539-40960b6deb8e/go.mod h1:n7NCudcB/nEzxVGmLbDWY5pfWTLqBcC2KZ6jyYvM4mQ= +golang.org/x/tools v0.0.0-20190226205152-f727befe758c/go.mod h1:9Yl7xja0Znq3iFh3HoIrodX9oNMXvdceNzlUR8zjMvY= golang.org/x/tools v0.0.0-20190311212946-11955173bddd/go.mod h1:LCzVGOaR6xXOjkQ3onu1FJEFr0SW1gC7cKk1uF8kGRs= +golang.org/x/tools v0.0.0-20190312151545-0bb0c0a6e846/go.mod h1:LCzVGOaR6xXOjkQ3onu1FJEFr0SW1gC7cKk1uF8kGRs= +golang.org/x/tools v0.0.0-20190524140312-2c0ae7006135/go.mod h1:RgjU9mgBXZiqYHBnxXauZ1Gv1EHHAz9KjViQ78xBX0Q= golang.org/x/tools v0.0.0-20190621195816-6e04913cbbac/go.mod h1:/rFqwRUd4F7ZHNgwSSTFct+R/Kf4OFW1sUzUTQQTgfc= golang.org/x/tools v0.0.0-20191029041327-9cc4af7d6b2c/go.mod h1:b+2E5dAYhXwXZwtnZ6UAqBI28+e2cm9otk0dWdXHAEo= golang.org/x/tools v0.0.0-20191029190741-b9c20aec41a5/go.mod h1:b+2E5dAYhXwXZwtnZ6UAqBI28+e2cm9otk0dWdXHAEo= @@ -475,6 +570,34 @@ golang.org/x/xerrors v0.0.0-20190717185122-a985d3407aa7/go.mod h1:I/5z698sn9Ka8T golang.org/x/xerrors v0.0.0-20191011141410-1b5146add898/go.mod h1:I/5z698sn9Ka8TeJc9MKroUUfqBBauWjQqLJ2OPfmY0= golang.org/x/xerrors v0.0.0-20191204190536-9bdfabe68543/go.mod h1:I/5z698sn9Ka8TeJc9MKroUUfqBBauWjQqLJ2OPfmY0= golang.org/x/xerrors v0.0.0-20200804184101-5ec99f83aff1/go.mod h1:I/5z698sn9Ka8TeJc9MKroUUfqBBauWjQqLJ2OPfmY0= +gonum.org/v1/gonum v0.0.0-20180816165407-929014505bf4/go.mod h1:Y+Yx5eoAFn32cQvJDxZx5Dpnq+c3wtXuadVZAcxbbBo= +gonum.org/v1/gonum v0.8.1-0.20200930085651-eea0b5cb5cc9/go.mod h1:oe/vMfY3deqTw+1EZJhuvEW2iwGF1bW9wwu7XCu0+v0= +gonum.org/v1/gonum v0.8.2/go.mod h1:oe/vMfY3deqTw+1EZJhuvEW2iwGF1bW9wwu7XCu0+v0= +gonum.org/v1/netlib v0.0.0-20190313105609-8cb42192e0e0/go.mod h1:wa6Ws7BG/ESfp6dHfk7C6KdzKA7wR7u/rKwOGE66zvw= +gonum.org/v1/netlib v0.0.0-20201012070519-2390d26c3658/go.mod h1:zQa7n16lh3Z6FbSTYgjG+KNhz1bA/b9t3plFEaGMp+A= +gonum.org/v1/plot v0.0.0-20190515093506-e2840ee46a6b/go.mod h1:Wt8AAjI+ypCyYX3nZBvf6cAIx93T+c/OS2HFAYskSZc= +google.golang.org/appengine v1.1.0/go.mod h1:EbEs0AVv82hx2wNQdGPgUI5lhzA/G0D9YwlJXL52JkM= +google.golang.org/appengine v1.4.0/go.mod h1:xpcJRLb0r/rnEns0DIKYYv+WjYCduHsrkT7/EB5XEv4= +google.golang.org/genproto v0.0.0-20180817151627-c66870c02cf8/go.mod h1:JiN7NxoALGmiZfu7CAH4rXhgtRTLTxftemlI0sWmxmc= +google.golang.org/genproto v0.0.0-20190819201941-24fa4b261c55/go.mod h1:DMBHOl98Agz4BDEuKkezgsaosCRResVns1a3J2ZsMNc= +google.golang.org/genproto v0.0.0-20200526211855-cb27e3aa2013/go.mod h1:NbSheEEYHJ7i3ixzK3sjbqSGDJWnxyFXZblF3eUsNvo= +google.golang.org/genproto v0.0.0-20200911024640-645f7a48b24f/go.mod h1:FWY/as6DDZQgahTzZj3fqbO1CbirC29ZNUFHwi0/+no= +google.golang.org/grpc v1.19.0/go.mod h1:mqu4LbDTu4XGKhr4mRzUsmM4RtVoemTSY81AxZiDr8c= +google.golang.org/grpc v1.23.0/go.mod h1:Y5yQAOtifL1yxbo5wqy6BxZv8vAUGQwXBOALyacEbxg= +google.golang.org/grpc v1.25.1/go.mod h1:c3i+UQWmh7LiEpx4sFZnkU36qjEYZ0imhYfXVyQciAY= +google.golang.org/grpc v1.27.0/go.mod h1:qbnxyOmOxrQa7FizSgH+ReBfzJrCY1pSN7KXBS8abTk= +google.golang.org/grpc v1.32.0/go.mod h1:N36X2cJ7JwdamYAgDz+s+rVMFjt3numwzf/HckM8pak= +google.golang.org/grpc/cmd/protoc-gen-go-grpc v0.0.0-20200910201057-6591123024b3/go.mod h1:6Kw0yEErY5E/yWrBtf03jp27GLLJujG4z/JK95pnjjw= +google.golang.org/protobuf v0.0.0-20200109180630-ec00e32a8dfd/go.mod h1:DFci5gLYBciE7Vtevhsrf46CRTquxDuWsQurQQe4oz8= +google.golang.org/protobuf v0.0.0-20200221191635-4d8936d0db64/go.mod h1:kwYJMbMJ01Woi6D6+Kah6886xMZcty6N08ah7+eCXa0= +google.golang.org/protobuf v0.0.0-20200228230310-ab0ca4ff8a60/go.mod h1:cfTl7dwQJ+fmap5saPgwCLgHXTUD7jkjRqWcaiX5VyM= +google.golang.org/protobuf v1.20.1-0.20200309200217-e05f789c0967/go.mod h1:A+miEFZTKqfCUM6K7xSMQL9OKL/b6hQv+e19PK+JZNE= +google.golang.org/protobuf v1.21.0/go.mod h1:47Nbq4nVaFHyn7ilMalzfO3qCViNmqZ2kzikPIcrTAo= +google.golang.org/protobuf v1.22.0/go.mod h1:EGpADcykh3NcUnDUJcl1+ZksZNG86OlYog2l/sGQquU= +google.golang.org/protobuf v1.23.0/go.mod h1:EGpADcykh3NcUnDUJcl1+ZksZNG86OlYog2l/sGQquU= +google.golang.org/protobuf v1.23.1-0.20200526195155-81db48ad09cc/go.mod h1:EGpADcykh3NcUnDUJcl1+ZksZNG86OlYog2l/sGQquU= +google.golang.org/protobuf v1.24.0/go.mod h1:r/3tXBNzIEhYS9I1OUVjXDlt8tc493IdKGjtUeSXeh4= +google.golang.org/protobuf v1.25.0/go.mod h1:9JNX74DMeImyA3h4bdi1ymwjUzf21/xIlbajtzgsN7c= google.golang.org/protobuf v1.26.0-rc.1/go.mod h1:jlhhOSvTdKEhbULTjvd4ARK9grFBp09yW+WbY/TyQbw= google.golang.org/protobuf v1.30.0 h1:kPPoIgf3TsEvrm0PFe15JQ+570QVxYzEvvHqChK+cng= google.golang.org/protobuf v1.30.0/go.mod h1:HV8QOd/L58Z+nl8r43ehVNZIU/HEI6OcFqwMG9pJV4I= @@ -482,6 +605,7 @@ gopkg.in/check.v1 v0.0.0-20161208181325-20d25e280405/go.mod h1:Co6ibVJAznAaIkqp8 gopkg.in/check.v1 v1.0.0-20180628173108-788fd7840127/go.mod h1:Co6ibVJAznAaIkqp8huTwlJQCZ016jof/cbN4VW5Yz0= gopkg.in/check.v1 v1.0.0-20190902080502-41f04d3bba15/go.mod h1:Co6ibVJAznAaIkqp8huTwlJQCZ016jof/cbN4VW5Yz0= gopkg.in/check.v1 v1.0.0-20201130134442-10cb98267c6c h1:Hei/4ADfdWqJk1ZMxUNpqntNwaWcugrBjAiHlqqRiVk= +gopkg.in/cheggaaa/pb.v1 v1.0.27/go.mod h1:V/YB90LKu/1FcN3WVnfiiE5oMCibMjukxqG/qStrOgw= gopkg.in/errgo.v2 v2.1.0/go.mod h1:hNsd1EY+bozCKY1Ytp96fpM3vjJbqLJn88ws8XvfDNI= gopkg.in/natefinch/lumberjack.v2 v2.0.0 h1:1Lc07Kr7qY4U2YPouBjpCLxpiyxIVoxqXgkXLknAOE8= gopkg.in/warnings.v0 v0.1.2 h1:wFXVbFY8DY5/xOe1ECiWdKCzZlxgshcYVNkBHstARME= @@ -494,10 +618,33 @@ gopkg.in/yaml.v3 v3.0.0-20200313102051-9f266ea9e77c/go.mod h1:K4uyk7z7BCEPqu6E+C gopkg.in/yaml.v3 v3.0.0-20210107192922-496545a6307b/go.mod h1:K4uyk7z7BCEPqu6E+C64Yfv1cQ7kz7rIZviUmN+EgEM= gopkg.in/yaml.v3 v3.0.1 h1:fxVm/GzAzEWqLHuvctI91KS9hhNmmWOoWu0XTYJS7CA= gopkg.in/yaml.v3 v3.0.1/go.mod h1:K4uyk7z7BCEPqu6E+C64Yfv1cQ7kz7rIZviUmN+EgEM= +gorgonia.org/dawson v1.2.0/go.mod h1:Px1mcziba8YUBIDsbzGwbKJ11uIblv/zkln4jNrZ9Ws= +gorgonia.org/gorgonia v0.9.17/go.mod h1:g66b5Z6ATUdhVqYl2ZAAwblv5hnGW08vNinGLcnrceI= +gorgonia.org/tensor v0.9.17/go.mod h1:75SMdLLhZ+2oB0/EE8lFEIt1Caoykdd4bz1mAe59deg= +gorgonia.org/tensor v0.9.20/go.mod h1:75SMdLLhZ+2oB0/EE8lFEIt1Caoykdd4bz1mAe59deg= +gorgonia.org/vecf32 v0.9.0/go.mod h1:NCc+5D2oxddRL11hd+pCB1PEyXWOyiQxfZ/1wwhOXCA= +gorgonia.org/vecf64 v0.9.0/go.mod h1:hp7IOWCnRiVQKON73kkC/AUMtEXyf9kGlVrtPQ9ccVA= +honnef.co/go/tools v0.0.0-20190102054323-c2f93a96b099/go.mod h1:rf3lG4BRIbNafJWhAfAdb/ePZxsR/4RtNHQocxwk9r4= +honnef.co/go/tools v0.0.0-20190523083050-ea95bdfd59fc/go.mod h1:rf3lG4BRIbNafJWhAfAdb/ePZxsR/4RtNHQocxwk9r4= honnef.co/go/tools v0.0.1-2019.2.3/go.mod h1:a3bituU0lyd329TUQxRnasdCoJDkEUEAqEt0JzvZhAg= k8s.io/apimachinery v0.28.3 h1:B1wYx8txOaCQG0HmYF6nbpU8dg6HvA06x5tEffvOe7A= k8s.io/apimachinery v0.28.3/go.mod h1:uQTKmIqs+rAYaq+DFaoD2X7pcjLOqbQX2AOiO0nIpb8= lukechampine.com/blake3 v1.1.7 h1:GgRMhmdsuK8+ii6UZFDL8Nb+VyMwadAgcJyfYHxG6n0= lukechampine.com/blake3 v1.1.7/go.mod h1:tkKEOtDkNtklkXtLNEOGNq5tcV90tJiA1vAA12R78LA= +modernc.org/cc v1.0.0/go.mod h1:1Sk4//wdnYJiUIxnW8ddKpaOJCF37yAdqYnkxUpaYxw= +modernc.org/cc v1.0.1/go.mod h1:uj1/YV+GYVdtSfGOgOtY62Jz8YIiEC0EzZNq481HIQs= +modernc.org/fileutil v1.0.0/go.mod h1:JHsWpkrk/CnVV1H/eGlFf85BEpfkrp56ro8nojIq9Q8= +modernc.org/golex v1.0.0/go.mod h1:b/QX9oBD/LhixY6NDh+IdGv17hgB+51fET1i2kPSmvk= +modernc.org/golex v1.0.1/go.mod h1:QCA53QtsT1NdGkaZZkF5ezFwk4IXh4BGNafAARTC254= +modernc.org/internal v1.0.0/go.mod h1:VUD/+JAkhCpvkUitlEOnhpVxCgsBI90oTzSCRcqQVSM= +modernc.org/ir v1.0.0/go.mod h1:wxK1nK3PS04CASoUY+HJr+FQywv4+D38y2sRrd71y7s= +modernc.org/lex v1.0.0/go.mod h1:G6rxMTy3cH2iA0iXL/HRRv4Znu8MK4higxph/lE7ypk= +modernc.org/lexer v1.0.0/go.mod h1:F/Dld0YKYdZCLQ7bD0USbWL4YKCyTDRDHiDTOs0q0vk= +modernc.org/mathutil v1.0.0/go.mod h1:wU0vUrJsVWBZ4P6e7xtFJEhFSNsfRLJ8H458uRjg03k= +modernc.org/mathutil v1.1.1/go.mod h1:mZW8CKdRPY1v87qxC/wUdX5O1qDzXMP5TH3wjfpga6E= +modernc.org/strutil v1.1.0/go.mod h1:lstksw84oURvj9y3tn8lGvRxyRC1S2+g5uuIzNfIOBs= +modernc.org/token v1.0.0/go.mod h1:UGzOrNV1mAFSEB63lOFHIpNRUVMvYTc6yu1SMY/XTDM= +modernc.org/xc v1.0.0/go.mod h1:mRNCo0bvLjGhHO9WsyuKVU4q0ceiDDDoEeWDJHrNx8I= +rsc.io/pdf v0.1.1/go.mod h1:n8OzWcQ6Sp37PL01nO98y4iUCRdTGarVfzxY20ICaU4= rsc.io/tmplfunc v0.0.3 h1:53XFQh69AfOa8Tw0Jm7t+GV7KZhOi6jzsCzTtKbMvzU= rsc.io/tmplfunc v0.0.3/go.mod h1:AG3sTPzElb1Io3Yg4voV9AGZJuleGAwaVRxL9M49PhA= diff --git a/pkg/options/resource-provider.go b/pkg/options/resource-provider.go index eac8d270..09016682 100644 --- a/pkg/options/resource-provider.go +++ b/pkg/options/resource-provider.go @@ -22,7 +22,11 @@ func NewResourceProviderOptions() resourceprovider.ResourceProviderOptions { func GetDefaultResourceProviderPowOptions() resourceprovider.ResourceProviderPowOptions { return resourceprovider.ResourceProviderPowOptions{ - EnablePow: GetDefaultServeOptionBool("ENABLE_POW", false), + EnablePow: GetDefaultServeOptionBool("ENABLE_POW", false), + NumWorkers: GetDefaultServeOptionInt("NUM_WORKER", 0), + + CudaGridSize: GetDefaultServeOptionInt("CUDA_GRID_SIZE", 256), + CudaBlockSize: GetDefaultServeOptionInt("CUDA_BLOCK_SIZE", 512), } } @@ -84,6 +88,19 @@ func AddResourceProviderPowCliFlags(cmd *cobra.Command, options *resourceprovide &options.EnablePow, "enable-pow", options.EnablePow, `Start pow mining (ENABLE_POW)`, ) + cmd.PersistentFlags().IntVar( + &options.NumWorkers, "num-worker", options.NumWorkers, + `Start pow mining (NUM_WORKER)`, + ) + + cmd.PersistentFlags().IntVar( + &options.CudaGridSize, "cuda-grid-size", options.CudaGridSize, + `Cuda grid size (CUDA_GRID_SIZE)`, + ) + cmd.PersistentFlags().IntVar( + &options.CudaBlockSize, "cuda-block-size", options.CudaBlockSize, + `Cuda block size (CUDA_BLOCK_SIZE)`, + ) } func AddResourceProviderCliFlags(cmd *cobra.Command, options *resourceprovider.ResourceProviderOptions) { diff --git a/pkg/resourceprovider/cpuworker.go b/pkg/resourceprovider/cpuworker.go new file mode 100644 index 00000000..33919a18 --- /dev/null +++ b/pkg/resourceprovider/cpuworker.go @@ -0,0 +1,109 @@ +package resourceprovider + +import ( + "context" + "math/big" + "sync/atomic" + "time" + + "github.com/ethereum/go-ethereum/crypto" + "github.com/holiman/uint256" + "github.com/rs/zerolog/log" +) + +var ( + bigOne = uint256.NewInt(1) +) + +type CpuWorker struct { + cfg *WorkerConfig + + state atomic.Int32 + quit chan chan struct{} +} + +func NewCpuWorker(cfg *WorkerConfig) (Worker, error) { + return &CpuWorker{ + cfg: cfg, + quit: make(chan chan struct{}, 1), + }, nil +} + +func (w *CpuWorker) Stop() { + if w.state.Load() == 0 { + return + } + + //cancel previous task + waitChan := make(chan struct{}) + select { + case w.quit <- waitChan: // may already exit + <-waitChan + default: + } +} + +func (w *CpuWorker) FindSolution(ctx context.Context, task *Task) { + w.state.Store(1) + defer w.state.Store(0) + + nonce := task.From.Clone() + startTime := time.Now() + + hashesCompleted := uint64(0) + ticker := time.NewTicker(time.Second * hashUpdateSecs) + defer ticker.Stop() + +OUT: + for { + select { + case <-ctx.Done(): + break OUT + case respCh := <-w.quit: + respCh <- struct{}{} + return + case <-ticker.C: + w.cfg.updateHashes <- hashesCompleted + hashesCompleted = 0 + default: + // Non-blocking select to fall through + } + + if nonce.Cmp(task.End) >= 0 { + return + } + hashNumber, err := calculateHashNumber(task.Challenge, nonce.ToBig()) + if err != nil { + log.Err(err).Msg("Calculate hash number") + return + } + hashesCompleted += 1 + + // Check if the hash is below the target difficulty + if hashNumber.Cmp(task.Difficulty) == -1 { + log.Info().Str("Elapsed Time", time.Since(startTime).String()). + Str("challenge", new(big.Int).SetBytes(task.Challenge[:]).String()). + Str("Nonce", nonce.String()). + Str("HashNumber", hashNumber.String()). + Msg("Success!") + w.cfg.resultCh <- TaskResult{ + Id: task.Id, + Nonce: nonce.Clone(), + } + } + + nonce.Add(nonce, bigOne) + } +} + +func calculateHashNumber(challenge [32]byte, nonce *big.Int) (*uint256.Int, error) { + data, err := formatMinerArgs(challenge, nonce) + if err != nil { + return nil, err + } + + // Calculate Keccak-256 hash + hashResult := crypto.Keccak256(data) + + return new(uint256.Int).SetBytes(hashResult), nil +} diff --git a/pkg/resourceprovider/worker_test.go b/pkg/resourceprovider/cpuworker_test.go similarity index 100% rename from pkg/resourceprovider/worker_test.go rename to pkg/resourceprovider/cpuworker_test.go diff --git a/pkg/resourceprovider/cudaminer/config.h b/pkg/resourceprovider/cudaminer/config.h new file mode 100644 index 00000000..a6a45cd9 --- /dev/null +++ b/pkg/resourceprovider/cudaminer/config.h @@ -0,0 +1,26 @@ +/* + * Type Definitions for CUDA Hashing Algos + * + * Date: 12 June 2019 + * Revision: 1 + * + * This file is released into the Public Domain. + */ + +#pragma once +#define USE_MD2 1 +#define USE_MD5 1 +#define USE_SHA1 1 +#define USE_SHA256 1 + +#define CUDA_HASH 1 +#define OCL_HASH 0 + +typedef unsigned char BYTE; +typedef unsigned int WORD; +typedef unsigned long long LONG; + +#include +#include +#include +#include \ No newline at end of file diff --git a/pkg/resourceprovider/cudaminer/keccak.cu b/pkg/resourceprovider/cudaminer/keccak.cu new file mode 100644 index 00000000..2ff0708c --- /dev/null +++ b/pkg/resourceprovider/cudaminer/keccak.cu @@ -0,0 +1,329 @@ +/* + * keccak.cu Implementation of Keccak/SHA3 digest + * + * Date: 12 June 2019 + * Revision: 1 + * + * This file is released into the Public Domain. + */ + + +extern "C" +{ + #include "keccak.cuh" +} + +#define KECCAK_ROUND 24 +#define KECCAK_STATE_SIZE 25 +#define KECCAK_Q_SIZE 192 + +__constant__ uint64_t CUDA_KECCAK_CONSTS[24] = { 0x0000000000000001, 0x0000000000008082, + 0x800000000000808a, 0x8000000080008000, 0x000000000000808b, 0x0000000080000001, 0x8000000080008081, + 0x8000000000008009, 0x000000000000008a, 0x0000000000000088, 0x0000000080008009, 0x000000008000000a, + 0x000000008000808b, 0x800000000000008b, 0x8000000000008089, 0x8000000000008003, 0x8000000000008002, + 0x8000000000000080, 0x000000000000800a, 0x800000008000000a, 0x8000000080008081, 0x8000000000008080, + 0x0000000080000001, 0x8000000080008008 }; + + +__constant__ uint64_t digestbitlen = 256; +__constant__ uint64_t rate_bits = 1088; +__constant__ uint64_t rate_BYTEs = 136; +__constant__ uint64_t absorb_round = 17; +typedef struct { + int64_t state[KECCAK_STATE_SIZE]; + uint8_t q[KECCAK_Q_SIZE]; + + uint64_t bits_in_queue; + +} cuda_keccak_ctx_t; +typedef cuda_keccak_ctx_t CUDA_KECCAK_CTX; + +__device__ uint64_t cuda_keccak_leuint64(void *in) +{ + uint64_t a; + memcpy(&a, in, 8); + return a; +} + +__device__ int64_t cuda_keccak_MIN(int64_t a, int64_t b) +{ + if (a > b) return b; + return a; +} + +__device__ uint64_t cuda_keccak_UMIN(uint64_t a, uint64_t b) +{ + if (a > b) return b; + return a; +} + +__device__ void cuda_keccak_extract(cuda_keccak_ctx_t *ctx) +{ + uint64_t len = rate_bits >> 6; + int64_t a; + int s = sizeof(uint64_t); + + for (int i = 0;i < len;i++) { + a = cuda_keccak_leuint64((int64_t*)&ctx->state[i]); + memcpy(ctx->q + (i * s), &a, s); + } +} +__device__ __forceinline__ uint64_t cuda_keccak_ROTL64(uint64_t a, uint64_t b) { + return (a << b) | (a >> (64 - b)); +} + +__device__ void cuda_keccak_permutations(cuda_keccak_ctx_t *ctx) { + int64_t* A = ctx->state; + + #pragma unroll 24 + for (int i = 0; i < KECCAK_ROUND; i++) { + int64_t C[5], D[5]; + + // Theta + C[0] = A[0] ^ A[5] ^ A[10] ^ A[15] ^ A[20]; + C[1] = A[1] ^ A[6] ^ A[11] ^ A[16] ^ A[21]; + C[2] = A[2] ^ A[7] ^ A[12] ^ A[17] ^ A[22]; + C[3] = A[3] ^ A[8] ^ A[13] ^ A[18] ^ A[23]; + C[4] = A[4] ^ A[9] ^ A[14] ^ A[19] ^ A[24]; + + D[0] = cuda_keccak_ROTL64(C[1], 1) ^ C[4]; + D[1] = cuda_keccak_ROTL64(C[2], 1) ^ C[0]; + D[2] = cuda_keccak_ROTL64(C[3], 1) ^ C[1]; + D[3] = cuda_keccak_ROTL64(C[4], 1) ^ C[2]; + D[4] = cuda_keccak_ROTL64(C[0], 1) ^ C[3]; + + #pragma unroll 25 + for (int j = 0; j < 25; j += 5) { + A[j] ^= D[0]; + A[j + 1] ^= D[1]; + A[j + 2] ^= D[2]; + A[j + 3] ^= D[3]; + A[j + 4] ^= D[4]; + } + + // Rho Pi + int64_t B[25]; + B[0] = A[0]; + B[1] = cuda_keccak_ROTL64(A[6], 44); + B[2] = cuda_keccak_ROTL64(A[12], 43); + B[3] = cuda_keccak_ROTL64(A[18], 21); + B[4] = cuda_keccak_ROTL64(A[24], 14); + B[5] = cuda_keccak_ROTL64(A[3], 28); + B[6] = cuda_keccak_ROTL64(A[9], 20); + B[7] = cuda_keccak_ROTL64(A[10], 3); + B[8] = cuda_keccak_ROTL64(A[16], 45); + B[9] = cuda_keccak_ROTL64(A[22], 61); + B[10] = cuda_keccak_ROTL64(A[1], 1); + B[11] = cuda_keccak_ROTL64(A[7], 6); + B[12] = cuda_keccak_ROTL64(A[13], 25); + B[13] = cuda_keccak_ROTL64(A[19], 8); + B[14] = cuda_keccak_ROTL64(A[20], 18); + B[15] = cuda_keccak_ROTL64(A[4], 27); + B[16] = cuda_keccak_ROTL64(A[5], 36); + B[17] = cuda_keccak_ROTL64(A[11], 10); + B[18] = cuda_keccak_ROTL64(A[17], 15); + B[19] = cuda_keccak_ROTL64(A[23], 56); + B[20] = cuda_keccak_ROTL64(A[2], 62); + B[21] = cuda_keccak_ROTL64(A[8], 55); + B[22] = cuda_keccak_ROTL64(A[14], 39); + B[23] = cuda_keccak_ROTL64(A[15], 41); + B[24] = cuda_keccak_ROTL64(A[21], 2); + + // Chi + #pragma unroll 24 + for (int j = 0; j < 25; j += 5) { + #pragma unroll 5 + for (int k = 0; k < 5; ++k) { + A[j + k] = B[j + k] ^ (~B[j + (k + 1) % 5] & B[j + (k + 2) % 5]); + } + } + + // Iota + A[0] ^= CUDA_KECCAK_CONSTS[i]; + } +} + + +__device__ void cuda_keccak_absorb(cuda_keccak_ctx_t *ctx, uint8_t* in) +{ + + uint64_t offset = 0; + for (uint64_t i = 0; i < absorb_round; ++i) {//10 + ctx->state[i] ^= cuda_keccak_leuint64(in + offset);//18 + offset += 8;//9 + } + + cuda_keccak_permutations(ctx);//8 +} + +__device__ void cuda_keccak_pad(cuda_keccak_ctx_t *ctx) +{ + ctx->q[ctx->bits_in_queue >> 3] |= (1L << (ctx->bits_in_queue & 7)); //6 + + if (++(ctx->bits_in_queue) == rate_bits) {//9 + cuda_keccak_absorb(ctx, ctx->q);//8 + ctx->bits_in_queue = 0;//53 + } + + uint64_t full = ctx->bits_in_queue >> 6; //7 + uint64_t partial = ctx->bits_in_queue & 63; //8 + + uint64_t offset = 0; + for (int i = 0; i < full; ++i) {//52 + ctx->state[i] ^= cuda_keccak_leuint64(ctx->q + offset);//52 + offset += 8;//52 + } + + if (partial > 0) {//8 + uint64_t mask = (1L << partial) - 1;//17 + ctx->state[full] ^= cuda_keccak_leuint64(ctx->q + offset) & mask;//16 + } + + ctx->state[(rate_bits - 1) >> 6] ^= 9223372036854775808ULL;/* 1 << 63 */ //9 + + cuda_keccak_permutations(ctx);//8 + cuda_keccak_extract(ctx);//58 + + ctx->bits_in_queue = rate_bits;//37 +} + + +/* + * Digestbitlen must be 128 224 256 288 384 512 + */ +__device__ void cuda_keccak_init(cuda_keccak_ctx_t *ctx) +{ + memset(ctx, 0, sizeof(cuda_keccak_ctx_t)); + ctx->bits_in_queue = 0;//11 +} + +__device__ void cuda_keccak_update(cuda_keccak_ctx_t *ctx, uint8_t *in, uint64_t inlen) +{ + int64_t BYTEs = ctx->bits_in_queue >> 3; + int64_t count = 0; + while (count < inlen) {//46 + if (BYTEs == 0 && count <= ((int64_t)(inlen - rate_BYTEs))) {//12 + do { + cuda_keccak_absorb(ctx, in + count);//8 + count += rate_BYTEs;//56 + } while (count <= ((int64_t)(inlen - rate_BYTEs)));//46 + } else { + int64_t partial = cuda_keccak_MIN(rate_BYTEs - BYTEs, inlen - count);//12 + memcpy(ctx->q + BYTEs, in + count, partial);//12 + + BYTEs += partial;//10 + count += partial;//8 + + if (BYTEs == rate_BYTEs) {//10 + cuda_keccak_absorb(ctx, ctx->q);//8 + BYTEs = 0; + } + } + } + ctx->bits_in_queue = BYTEs << 3;//8 +} + +__device__ void cuda_keccak_final_rev(cuda_keccak_ctx_t *ctx, uint8_t *out) +{ + cuda_keccak_pad(ctx); + uint64_t i = 0;//6 + + while (i < digestbitlen) {//46 + if (ctx->bits_in_queue == 0) {//9 + cuda_keccak_permutations(ctx);//8 + cuda_keccak_extract(ctx);//56 + ctx->bits_in_queue = rate_bits;//7 + } + + uint64_t partial_block = cuda_keccak_UMIN(ctx->bits_in_queue, digestbitlen - i);//9 + + //directly reverse? + int start = 31- (i >> 3); + uint8_t* pos = ctx->q + (rate_BYTEs - (ctx->bits_in_queue >> 3)); + for (int j = 0; j< (partial_block >> 3); j++) { + out[start-j] = pos[j]; + } + + ctx->bits_in_queue -= partial_block;//11 + i += partial_block;//11 + } +} + + + +__noinline__ __device__ static bool hashbelowtarget(const uint64_t *const __restrict__ hash, const uint64_t *const __restrict__ target) +{ + if (hash[3] > target[3])//46 + return false; + if (hash[3] < target[3])//46 + return true; + if (hash[2] > target[2])//45 + return false; + if (hash[2] < target[2])//45 + return true; + + if (hash[1] > target[1])//43 + return false; + if (hash[1] < target[1])//43 + return true; + if (hash[0] > target[0])//39 + return false; + + return true; +} + +__device__ uint64_t *addUint256(const uint64_t *a, const uint64_t b) +{ + uint64_t *result = new uint64_t[4];//47 + uint64_t sum = a[0] + b;//10 + result[0] = sum;//10 + + uint64_t carry = (sum < a[0]) ? 1 : 0;//12 + for (int i = 1; i < 4; i++)//13 + { + sum = a[i] + carry;//16 + result[i] = sum;//14 + carry = (sum < a[i]) ? 1 : 0;//14 + } + + return result; +} +__device__ void reverse32BytesInPlace(uint8_t *data, uint8_t *out) +{ + for (int i = 0; i < 32; i++)//13 + { + out[i] = data[31-i]; + } +} + +extern "C" __global__ __launch_bounds__(1024) + + void kernel_lilypad_pow(uint8_t* chanllenge, uint64_t* startNonce, uint64_t* target, uint64_t n_batch, uint8_t* resNonce) +{ + uint64_t thread = blockIdx.x * blockDim.x + threadIdx.x; //4 + if (thread >= n_batch) {//36 + return; + } + + //pack input + uint8_t in[64]; + memcpy(in, chanllenge, 32); + //increase nonce + uint8_t* nonce = (uint8_t*)addUint256(startNonce, thread);//35 + uint8_t nonce_rev[32]; + reverse32BytesInPlace(nonce, nonce_rev);//18 + memcpy(in+32, nonce_rev, 32); + + + uint8_t out[32]; + CUDA_KECCAK_CTX ctx; + cuda_keccak_init(&ctx); //6 + cuda_keccak_update(&ctx, in,64); //12 + cuda_keccak_final_rev(&ctx, out); //6 + + if (hashbelowtarget((uint64_t*)out, target)) {//49 + memcpy(resNonce, nonce_rev, 32); + } + + delete nonce;//45 +} diff --git a/pkg/resourceprovider/cudaminer/keccak.cuh b/pkg/resourceprovider/cudaminer/keccak.cuh new file mode 100644 index 00000000..d1819df6 --- /dev/null +++ b/pkg/resourceprovider/cudaminer/keccak.cuh @@ -0,0 +1,12 @@ +/* + * keccak.cuh CUDA Implementation of BLAKE2B Hashing + * + * Date: 12 June 2019 + * Revision: 1 + * + * This file is released into the Public Domain. + */ + + +#pragma once +#include "config.h" diff --git a/pkg/resourceprovider/cudaminer/keccak.ptx b/pkg/resourceprovider/cudaminer/keccak.ptx new file mode 100644 index 00000000..a782188b --- /dev/null +++ b/pkg/resourceprovider/cudaminer/keccak.ptx @@ -0,0 +1,55324 @@ +// +// Generated by NVIDIA NVVM Compiler +// +// Compiler Build ID: CL-32267302 +// Cuda compilation tools, release 12.0, V12.0.140 +// Based on NVVM 7.0.1 +// + +.version 8.0 +.target sm_52 +.address_size 64 + +.extern .func (.param .b64 func_retval0) malloc +( + .param .b64 malloc_param_0 +) +; +.extern .func free +( + .param .b64 free_param_0 +) +; +.const .align 8 .b8 CUDA_KECCAK_CONSTS[192] = {1, 0, 0, 0, 0, 0, 0, 0, 130, 128, 0, 0, 0, 0, 0, 0, 138, 128, 0, 0, 0, 0, 0, 128, 0, 128, 0, 128, 0, 0, 0, 128, 139, 128, 0, 0, 0, 0, 0, 0, 1, 0, 0, 128, 0, 0, 0, 0, 129, 128, 0, 128, 0, 0, 0, 128, 9, 128, 0, 0, 0, 0, 0, 128, 138, 0, 0, 0, 0, 0, 0, 0, 136, 0, 0, 0, 0, 0, 0, 0, 9, 128, 0, 128, 0, 0, 0, 0, 10, 0, 0, 128, 0, 0, 0, 0, 139, 128, 0, 128, 0, 0, 0, 0, 139, 0, 0, 0, 0, 0, 0, 128, 137, 128, 0, 0, 0, 0, 0, 128, 3, 128, 0, 0, 0, 0, 0, 128, 2, 128, 0, 0, 0, 0, 0, 128, 128, 0, 0, 0, 0, 0, 0, 128, 10, 128, 0, 0, 0, 0, 0, 0, 10, 0, 0, 128, 0, 0, 0, 128, 129, 128, 0, 128, 0, 0, 0, 128, 128, 128, 0, 0, 0, 0, 0, 128, 1, 0, 0, 128, 0, 0, 0, 0, 8, 128, 0, 128, 0, 0, 0, 128}; +.const .align 8 .u64 digestbitlen = 256; +.const .align 8 .u64 rate_bits = 1088; +.const .align 8 .u64 rate_BYTEs = 136; +.const .align 8 .u64 absorb_round = 17; + +.func (.param .b32 func_retval0) _ZN39_INTERNAL_467e079b_9_keccak_cu_bbb2fa6e15hashbelowtargetEPKyS1_( + .param .b64 _ZN39_INTERNAL_467e079b_9_keccak_cu_bbb2fa6e15hashbelowtargetEPKyS1__param_0, + .param .b64 _ZN39_INTERNAL_467e079b_9_keccak_cu_bbb2fa6e15hashbelowtargetEPKyS1__param_1 +) +{ + .reg .pred %p<8>; + .reg .b16 %rs<10>; + .reg .b32 %r<2>; + .reg .b64 %rd<13>; + + + ld.param.u64 %rd9, [_ZN39_INTERNAL_467e079b_9_keccak_cu_bbb2fa6e15hashbelowtargetEPKyS1__param_0]; + ld.param.u64 %rd10, [_ZN39_INTERNAL_467e079b_9_keccak_cu_bbb2fa6e15hashbelowtargetEPKyS1__param_1]; + cvta.to.global.u64 %rd2, %rd10; + cvta.to.local.u64 %rd1, %rd9; + ld.global.u64 %rd3, [%rd2+24]; + ld.local.u64 %rd4, [%rd1+24]; + setp.gt.u64 %p1, %rd4, %rd3; + mov.u16 %rs3, 0; + mov.u16 %rs9, %rs3; + @%p1 bra $L__BB0_7; + + setp.lt.u64 %p2, %rd4, %rd3; + mov.u16 %rs4, 1; + mov.u16 %rs9, %rs4; + @%p2 bra $L__BB0_7; + + ld.global.u64 %rd5, [%rd2+16]; + ld.local.u64 %rd6, [%rd1+16]; + setp.gt.u64 %p3, %rd6, %rd5; + mov.u16 %rs9, %rs3; + @%p3 bra $L__BB0_7; + + setp.lt.u64 %p4, %rd6, %rd5; + mov.u16 %rs9, %rs4; + @%p4 bra $L__BB0_7; + + ld.global.u64 %rd7, [%rd2+8]; + ld.local.u64 %rd8, [%rd1+8]; + setp.gt.u64 %p5, %rd8, %rd7; + mov.u16 %rs9, %rs3; + @%p5 bra $L__BB0_7; + + setp.lt.u64 %p6, %rd8, %rd7; + mov.u16 %rs9, %rs4; + @%p6 bra $L__BB0_7; + + ld.local.u64 %rd11, [%rd1]; + ld.global.u64 %rd12, [%rd2]; + setp.le.u64 %p7, %rd11, %rd12; + selp.u16 %rs9, 1, 0, %p7; + +$L__BB0_7: + cvt.u32.u16 %r1, %rs9; + st.param.b32 [func_retval0+0], %r1; + ret; + +} + // .globl kernel_lilypad_pow +.visible .entry kernel_lilypad_pow( + .param .u64 kernel_lilypad_pow_param_0, + .param .u64 kernel_lilypad_pow_param_1, + .param .u64 kernel_lilypad_pow_param_2, + .param .u64 kernel_lilypad_pow_param_3, + .param .u64 kernel_lilypad_pow_param_4 +) +.maxntid 1024, 1, 1 +{ + .local .align 16 .b8 __local_depot1[496]; + .reg .b64 %SP; + .reg .b64 %SPL; + .reg .pred %p<60>; + .reg .b16 %rs<56>; + .reg .b32 %r<14043>; + .reg .b64 %rd<19868>; + + + mov.u64 %SPL, __local_depot1; + cvta.local.u64 %SP, %SPL; + ld.param.u64 %rd161, [kernel_lilypad_pow_param_0]; + ld.param.u64 %rd162, [kernel_lilypad_pow_param_1]; + ld.param.u64 %rd165, [kernel_lilypad_pow_param_3]; + add.u64 %rd1, %SPL, 0; + add.u64 %rd2, %SPL, 96; + mov.u32 %r20, %ntid.x; + mov.u32 %r21, %ctaid.x; + mov.u32 %r22, %tid.x; + mad.lo.s32 %r23, %r21, %r20, %r22; + cvt.u64.u32 %rd3, %r23; + setp.ge.u64 %p1, %rd3, %rd165; + @%p1 bra $L__BB1_76; + + cvta.to.global.u64 %rd4, %rd162; + cvta.to.global.u64 %rd5, %rd161; + mov.u64 %rd19826, 0; + mov.u32 %r14032, 0; + +$L__BB1_2: + add.s64 %rd169, %rd5, %rd19826; + ld.global.u8 %rs25, [%rd169]; + add.s64 %rd170, %rd1, %rd19826; + st.local.u8 [%rd170], %rs25; + add.s64 %rd19826, %rd19826, 1; + add.s32 %r14032, %r14032, 1; + setp.lt.u32 %p2, %r14032, 32; + @%p2 bra $L__BB1_2; + + mov.u64 %rd172, 32; + { // callseq 0, 0 + .reg .b32 temp_param_reg; + .param .b64 param0; + st.param.b64 [param0+0], %rd172; + .param .b64 retval0; + call.uni (retval0), + malloc, + ( + param0 + ); + ld.param.b64 %rd8, [retval0+0]; + } // callseq 0 + ld.global.u64 %rd173, [%rd4]; + mov.u64 %rd19828, 0; + add.s64 %rd9, %rd173, %rd3; + st.u64 [%rd8], %rd9; + ld.global.u64 %rd174, [%rd4]; + setp.lt.u64 %p3, %rd9, %rd174; + selp.u64 %rd175, 1, 0, %p3; + ld.global.u64 %rd176, [%rd4+8]; + add.s64 %rd177, %rd176, %rd175; + st.u64 [%rd8+8], %rd177; + ld.global.u64 %rd178, [%rd4+8]; + setp.lt.u64 %p4, %rd177, %rd178; + selp.u64 %rd179, 1, 0, %p4; + ld.global.u64 %rd180, [%rd4+16]; + add.s64 %rd181, %rd180, %rd179; + st.u64 [%rd8+16], %rd181; + ld.global.u64 %rd182, [%rd4+16]; + setp.lt.u64 %p5, %rd181, %rd182; + selp.u64 %rd183, 1, 0, %p5; + ld.global.u64 %rd184, [%rd4+24]; + add.s64 %rd185, %rd184, %rd183; + st.u64 [%rd8+24], %rd185; + shr.u64 %rd186, %rd185, 56; + cvt.u16.u64 %rs1, %rd186; + shr.u64 %rd187, %rd185, 48; + cvt.u16.u64 %rs2, %rd187; + shr.u64 %rd188, %rd185, 40; + cvt.u32.u64 %r25, %rd188; + cvt.u16.u64 %rs3, %rd188; + shr.u64 %rd189, %rd185, 32; + cvt.u16.u64 %rs4, %rd189; + shr.u64 %rd190, %rd185, 24; + cvt.u32.u64 %r26, %rd190; + cvt.u16.u64 %rs5, %rd190; + shr.u64 %rd191, %rd185, 16; + cvt.u16.u64 %rs6, %rd191; + shr.u64 %rd192, %rd185, 8; + cvt.u32.u64 %r27, %rd192; + cvt.u16.u64 %rs7, %rd192; + cvt.u16.u64 %rs8, %rd185; + shr.u64 %rd193, %rd181, 56; + cvt.u16.u64 %rs9, %rd193; + shr.u64 %rd194, %rd181, 48; + cvt.u16.u64 %rs10, %rd194; + shr.u64 %rd195, %rd181, 40; + cvt.u32.u64 %r28, %rd195; + cvt.u16.u64 %rs11, %rd195; + shr.u64 %rd196, %rd181, 32; + cvt.u16.u64 %rs12, %rd196; + shr.u64 %rd197, %rd181, 24; + cvt.u32.u64 %r29, %rd197; + cvt.u16.u64 %rs13, %rd197; + shr.u64 %rd198, %rd181, 16; + cvt.u16.u64 %rs14, %rd198; + shr.u64 %rd199, %rd181, 8; + cvt.u32.u64 %r30, %rd199; + cvt.u16.u64 %rs15, %rd199; + cvt.u16.u64 %rs16, %rd181; + shr.u64 %rd200, %rd177, 56; + cvt.u16.u64 %rs17, %rd200; + shr.u64 %rd201, %rd177, 48; + cvt.u16.u64 %rs18, %rd201; + shr.u64 %rd202, %rd177, 40; + cvt.u32.u64 %r31, %rd202; + cvt.u16.u64 %rs19, %rd202; + shr.u64 %rd203, %rd177, 32; + cvt.u16.u64 %rs20, %rd203; + shr.u64 %rd204, %rd177, 24; + cvt.u32.u64 %r32, %rd204; + cvt.u16.u64 %rs21, %rd204; + shr.u64 %rd205, %rd177, 16; + cvt.u16.u64 %rs22, %rd205; + shr.u64 %rd206, %rd177, 8; + cvt.u32.u64 %r33, %rd206; + cvt.u16.u64 %rs23, %rd206; + cvt.u16.u64 %rs24, %rd177; + shr.u64 %rd207, %rd9, 56; + shr.u64 %rd208, %rd9, 48; + shr.u64 %rd209, %rd9, 40; + cvt.u32.u64 %r34, %rd209; + shr.u64 %rd210, %rd9, 32; + cvt.u32.u64 %r35, %rd9; + shr.u64 %rd211, %rd9, 16; + shl.b16 %rs26, %rs2, 8; + or.b16 %rs27, %rs1, %rs26; + cvt.u32.u64 %r36, %rd189; + and.b32 %r37, %r25, 255; + prmt.b32 %r38, %r36, %r37, 30212; + cvt.u16.u32 %rs28, %r38; + cvt.u32.u64 %r39, %rd191; + and.b32 %r40, %r26, 255; + prmt.b32 %r41, %r39, %r40, 30212; + cvt.u32.u64 %r42, %rd185; + and.b32 %r43, %r27, 255; + prmt.b32 %r44, %r42, %r43, 30212; + shl.b16 %rs29, %rs10, 8; + or.b16 %rs30, %rs9, %rs29; + cvt.u32.u64 %r45, %rd196; + and.b32 %r46, %r28, 255; + prmt.b32 %r47, %r45, %r46, 30212; + cvt.u16.u32 %rs31, %r47; + cvt.u32.u64 %r48, %rd198; + and.b32 %r49, %r29, 255; + prmt.b32 %r50, %r48, %r49, 30212; + cvt.u32.u64 %r51, %rd181; + and.b32 %r52, %r30, 255; + prmt.b32 %r53, %r51, %r52, 30212; + prmt.b32 %r54, %r53, %r50, 4180; + mov.b32 %r55, {%rs30, %rs31}; + prmt.b32 %r56, %r44, %r41, 4180; + mov.b32 %r57, {%rs27, %rs28}; + st.local.v4.u32 [%rd1+32], {%r57, %r56, %r55, %r54}; + shl.b16 %rs32, %rs18, 8; + or.b16 %rs33, %rs17, %rs32; + cvt.u32.u64 %r58, %rd203; + and.b32 %r59, %r31, 255; + prmt.b32 %r60, %r58, %r59, 30212; + cvt.u16.u32 %rs34, %r60; + cvt.u32.u64 %r61, %rd205; + and.b32 %r62, %r32, 255; + prmt.b32 %r63, %r61, %r62, 30212; + cvt.u32.u64 %r64, %rd177; + and.b32 %r65, %r33, 255; + prmt.b32 %r66, %r64, %r65, 30212; + cvt.u16.u64 %rs35, %rd207; + cvt.u16.u64 %rs36, %rd208; + shl.b16 %rs37, %rs36, 8; + or.b16 %rs38, %rs35, %rs37; + cvt.u32.u64 %r67, %rd210; + and.b32 %r68, %r34, 255; + prmt.b32 %r69, %r67, %r68, 30212; + cvt.u16.u32 %rs39, %r69; + cvt.u16.u64 %rs40, %rd9; + shl.b16 %rs41, %rs40, 8; + shr.u16 %rs42, %rs40, 8; + or.b16 %rs43, %rs42, %rs41; + shr.u32 %r70, %r35, 24; + cvt.u32.u64 %r71, %rd211; + prmt.b32 %r72, %r71, %r70, 30212; + cvt.u16.u32 %rs44, %r72; + mov.b32 %r73, {%rs44, %rs43}; + mov.b32 %r74, {%rs38, %rs39}; + prmt.b32 %r75, %r66, %r63, 4180; + mov.b32 %r76, {%rs33, %rs34}; + st.local.v4.u32 [%rd1+48], {%r76, %r75, %r74, %r73}; + mov.u64 %rd19827, %rd2; + +$L__BB1_4: + mov.u32 %r77, 0; + st.local.v2.u32 [%rd19827], {%r77, %r77}; + add.s64 %rd19827, %rd19827, 8; + add.s64 %rd19828, %rd19828, 1; + setp.lt.u64 %p6, %rd19828, 49; + @%p6 bra $L__BB1_4; + + mov.u64 %rd19844, 0; + st.local.u64 [%rd2+392], %rd19844; + ld.const.u64 %rd14, [rate_BYTEs]; + ld.const.u64 %rd15, [absorb_round]; + ld.const.u64 %rd16, [CUDA_KECCAK_CONSTS]; + ld.const.u64 %rd17, [CUDA_KECCAK_CONSTS+8]; + ld.const.u64 %rd18, [CUDA_KECCAK_CONSTS+16]; + ld.const.u64 %rd19, [CUDA_KECCAK_CONSTS+24]; + ld.const.u64 %rd20, [CUDA_KECCAK_CONSTS+32]; + ld.const.u64 %rd21, [CUDA_KECCAK_CONSTS+40]; + ld.const.u64 %rd22, [CUDA_KECCAK_CONSTS+48]; + ld.const.u64 %rd23, [CUDA_KECCAK_CONSTS+56]; + ld.const.u64 %rd24, [CUDA_KECCAK_CONSTS+64]; + ld.const.u64 %rd25, [CUDA_KECCAK_CONSTS+72]; + ld.const.u64 %rd26, [CUDA_KECCAK_CONSTS+80]; + ld.const.u64 %rd27, [CUDA_KECCAK_CONSTS+88]; + ld.const.u64 %rd28, [CUDA_KECCAK_CONSTS+96]; + ld.const.u64 %rd29, [CUDA_KECCAK_CONSTS+104]; + ld.const.u64 %rd30, [CUDA_KECCAK_CONSTS+112]; + ld.const.u64 %rd31, [CUDA_KECCAK_CONSTS+120]; + ld.const.u64 %rd32, [CUDA_KECCAK_CONSTS+128]; + ld.const.u64 %rd33, [CUDA_KECCAK_CONSTS+136]; + ld.const.u64 %rd34, [CUDA_KECCAK_CONSTS+144]; + ld.const.u64 %rd35, [CUDA_KECCAK_CONSTS+152]; + ld.const.u64 %rd36, [CUDA_KECCAK_CONSTS+160]; + ld.const.u64 %rd37, [CUDA_KECCAK_CONSTS+168]; + ld.const.u64 %rd38, [CUDA_KECCAK_CONSTS+176]; + ld.const.u64 %rd39, [CUDA_KECCAK_CONSTS+184]; + mov.u64 %rd19843, %rd19844; + +$L__BB1_6: + mov.u64 %rd214, 64; + sub.s64 %rd215, %rd214, %rd14; + setp.gt.s64 %p7, %rd19844, %rd215; + setp.ne.s64 %p8, %rd19843, 0; + or.pred %p9, %p8, %p7; + @%p9 bra $L__BB1_16; + bra.uni $L__BB1_7; + +$L__BB1_16: + sub.s64 %rd4103, %rd14, %rd19843; + sub.s64 %rd4105, %rd214, %rd19844; + min.s64 %rd56, %rd4103, %rd4105; + setp.eq.s64 %p17, %rd56, 0; + @%p17 bra $L__BB1_19; + + mov.u64 %rd19837, 0; + +$L__BB1_18: + add.s64 %rd4107, %rd19844, %rd19837; + add.s64 %rd4108, %rd1, %rd4107; + ld.local.u8 %rs45, [%rd4108]; + add.s64 %rd4109, %rd19843, %rd19837; + add.s64 %rd4110, %rd2, %rd4109; + st.local.u8 [%rd4110+200], %rs45; + add.s64 %rd19837, %rd19837, 1; + setp.lt.u64 %p18, %rd19837, %rd56; + @%p18 bra $L__BB1_18; + +$L__BB1_19: + add.s64 %rd19843, %rd56, %rd19843; + add.s64 %rd19844, %rd56, %rd19844; + setp.ne.s64 %p19, %rd19843, %rd14; + @%p19 bra $L__BB1_29; + + setp.eq.s64 %p20, %rd15, 0; + @%p20 bra $L__BB1_28; + + max.u64 %rd4113, %rd15, 1; + add.s64 %rd4114, %rd4113, -1; + setp.lt.u64 %p21, %rd4114, 3; + mov.u64 %rd19841, 0; + mov.u64 %rd19842, %rd19841; + @%p21 bra $L__BB1_24; + + and.b64 %rd4118, %rd4113, 3; + sub.s64 %rd19840, %rd4113, %rd4118; + +$L__BB1_23: + add.s64 %rd4119, %rd2, %rd19842; + ld.local.u64 %rd4120, [%rd4119]; + ld.local.u8 %rd4121, [%rd4119+200]; + ld.local.u8 %rd4122, [%rd4119+201]; + bfi.b64 %rd4123, %rd4122, %rd4121, 8, 8; + ld.local.u8 %rd4124, [%rd4119+202]; + ld.local.u8 %rd4125, [%rd4119+203]; + bfi.b64 %rd4126, %rd4125, %rd4124, 8, 8; + bfi.b64 %rd4127, %rd4126, %rd4123, 16, 16; + ld.local.u8 %rd4128, [%rd4119+204]; + ld.local.u8 %rd4129, [%rd4119+205]; + bfi.b64 %rd4130, %rd4129, %rd4128, 8, 8; + ld.local.u8 %rd4131, [%rd4119+206]; + ld.local.u8 %rd4132, [%rd4119+207]; + bfi.b64 %rd4133, %rd4132, %rd4131, 8, 8; + bfi.b64 %rd4134, %rd4133, %rd4130, 16, 16; + bfi.b64 %rd4135, %rd4134, %rd4127, 32, 32; + xor.b64 %rd4136, %rd4120, %rd4135; + st.local.u64 [%rd4119], %rd4136; + ld.local.u64 %rd4137, [%rd4119+8]; + ld.local.u8 %rd4138, [%rd4119+208]; + ld.local.u8 %rd4139, [%rd4119+209]; + bfi.b64 %rd4140, %rd4139, %rd4138, 8, 8; + ld.local.u8 %rd4141, [%rd4119+210]; + ld.local.u8 %rd4142, [%rd4119+211]; + bfi.b64 %rd4143, %rd4142, %rd4141, 8, 8; + bfi.b64 %rd4144, %rd4143, %rd4140, 16, 16; + ld.local.u8 %rd4145, [%rd4119+212]; + ld.local.u8 %rd4146, [%rd4119+213]; + bfi.b64 %rd4147, %rd4146, %rd4145, 8, 8; + ld.local.u8 %rd4148, [%rd4119+214]; + ld.local.u8 %rd4149, [%rd4119+215]; + bfi.b64 %rd4150, %rd4149, %rd4148, 8, 8; + bfi.b64 %rd4151, %rd4150, %rd4147, 16, 16; + bfi.b64 %rd4152, %rd4151, %rd4144, 32, 32; + xor.b64 %rd4153, %rd4137, %rd4152; + st.local.u64 [%rd4119+8], %rd4153; + ld.local.u64 %rd4154, [%rd4119+16]; + ld.local.u8 %rd4155, [%rd4119+216]; + ld.local.u8 %rd4156, [%rd4119+217]; + bfi.b64 %rd4157, %rd4156, %rd4155, 8, 8; + ld.local.u8 %rd4158, [%rd4119+218]; + ld.local.u8 %rd4159, [%rd4119+219]; + bfi.b64 %rd4160, %rd4159, %rd4158, 8, 8; + bfi.b64 %rd4161, %rd4160, %rd4157, 16, 16; + ld.local.u8 %rd4162, [%rd4119+220]; + ld.local.u8 %rd4163, [%rd4119+221]; + bfi.b64 %rd4164, %rd4163, %rd4162, 8, 8; + ld.local.u8 %rd4165, [%rd4119+222]; + ld.local.u8 %rd4166, [%rd4119+223]; + bfi.b64 %rd4167, %rd4166, %rd4165, 8, 8; + bfi.b64 %rd4168, %rd4167, %rd4164, 16, 16; + bfi.b64 %rd4169, %rd4168, %rd4161, 32, 32; + xor.b64 %rd4170, %rd4154, %rd4169; + st.local.u64 [%rd4119+16], %rd4170; + ld.local.u64 %rd4171, [%rd4119+24]; + ld.local.u8 %rd4172, [%rd4119+224]; + ld.local.u8 %rd4173, [%rd4119+225]; + bfi.b64 %rd4174, %rd4173, %rd4172, 8, 8; + ld.local.u8 %rd4175, [%rd4119+226]; + ld.local.u8 %rd4176, [%rd4119+227]; + bfi.b64 %rd4177, %rd4176, %rd4175, 8, 8; + bfi.b64 %rd4178, %rd4177, %rd4174, 16, 16; + ld.local.u8 %rd4179, [%rd4119+228]; + ld.local.u8 %rd4180, [%rd4119+229]; + bfi.b64 %rd4181, %rd4180, %rd4179, 8, 8; + ld.local.u8 %rd4182, [%rd4119+230]; + ld.local.u8 %rd4183, [%rd4119+231]; + bfi.b64 %rd4184, %rd4183, %rd4182, 8, 8; + bfi.b64 %rd4185, %rd4184, %rd4181, 16, 16; + bfi.b64 %rd4186, %rd4185, %rd4178, 32, 32; + xor.b64 %rd4187, %rd4171, %rd4186; + st.local.u64 [%rd4119+24], %rd4187; + add.s64 %rd19842, %rd19842, 32; + add.s64 %rd19841, %rd19841, 4; + add.s64 %rd19840, %rd19840, -4; + setp.ne.s64 %p22, %rd19840, 0; + @%p22 bra $L__BB1_23; + +$L__BB1_24: + and.b64 %rd4189, %rd4113, 3; + setp.eq.s64 %p23, %rd4189, 0; + @%p23 bra $L__BB1_28; + + setp.eq.s64 %p24, %rd4189, 1; + add.s64 %rd4192, %rd2, %rd19842; + add.s64 %rd70, %rd4192, 200; + shl.b64 %rd4193, %rd19841, 3; + add.s64 %rd71, %rd2, %rd4193; + ld.local.u64 %rd4194, [%rd71]; + ld.local.u8 %rd4195, [%rd4192+200]; + ld.local.u8 %rd4196, [%rd4192+201]; + bfi.b64 %rd4197, %rd4196, %rd4195, 8, 8; + ld.local.u8 %rd4198, [%rd4192+202]; + ld.local.u8 %rd4199, [%rd4192+203]; + bfi.b64 %rd4200, %rd4199, %rd4198, 8, 8; + bfi.b64 %rd4201, %rd4200, %rd4197, 16, 16; + ld.local.u8 %rd4202, [%rd4192+204]; + ld.local.u8 %rd4203, [%rd4192+205]; + bfi.b64 %rd4204, %rd4203, %rd4202, 8, 8; + ld.local.u8 %rd4205, [%rd4192+206]; + ld.local.u8 %rd4206, [%rd4192+207]; + bfi.b64 %rd4207, %rd4206, %rd4205, 8, 8; + bfi.b64 %rd4208, %rd4207, %rd4204, 16, 16; + bfi.b64 %rd4209, %rd4208, %rd4201, 32, 32; + xor.b64 %rd4210, %rd4194, %rd4209; + st.local.u64 [%rd71], %rd4210; + @%p24 bra $L__BB1_28; + + setp.eq.s64 %p25, %rd4189, 2; + ld.local.u64 %rd4213, [%rd71+8]; + ld.local.u8 %rd4214, [%rd70+8]; + ld.local.u8 %rd4215, [%rd70+9]; + bfi.b64 %rd4216, %rd4215, %rd4214, 8, 8; + ld.local.u8 %rd4217, [%rd70+10]; + ld.local.u8 %rd4218, [%rd70+11]; + bfi.b64 %rd4219, %rd4218, %rd4217, 8, 8; + bfi.b64 %rd4220, %rd4219, %rd4216, 16, 16; + ld.local.u8 %rd4221, [%rd70+12]; + ld.local.u8 %rd4222, [%rd70+13]; + bfi.b64 %rd4223, %rd4222, %rd4221, 8, 8; + ld.local.u8 %rd4224, [%rd70+14]; + ld.local.u8 %rd4225, [%rd70+15]; + bfi.b64 %rd4226, %rd4225, %rd4224, 8, 8; + bfi.b64 %rd4227, %rd4226, %rd4223, 16, 16; + bfi.b64 %rd4228, %rd4227, %rd4220, 32, 32; + xor.b64 %rd4229, %rd4213, %rd4228; + st.local.u64 [%rd71+8], %rd4229; + @%p25 bra $L__BB1_28; + + ld.local.u64 %rd4230, [%rd71+16]; + ld.local.u8 %rd4231, [%rd70+16]; + ld.local.u8 %rd4232, [%rd70+17]; + bfi.b64 %rd4233, %rd4232, %rd4231, 8, 8; + ld.local.u8 %rd4234, [%rd70+18]; + ld.local.u8 %rd4235, [%rd70+19]; + bfi.b64 %rd4236, %rd4235, %rd4234, 8, 8; + bfi.b64 %rd4237, %rd4236, %rd4233, 16, 16; + ld.local.u8 %rd4238, [%rd70+20]; + ld.local.u8 %rd4239, [%rd70+21]; + bfi.b64 %rd4240, %rd4239, %rd4238, 8, 8; + ld.local.u8 %rd4241, [%rd70+22]; + ld.local.u8 %rd4242, [%rd70+23]; + bfi.b64 %rd4243, %rd4242, %rd4241, 8, 8; + bfi.b64 %rd4244, %rd4243, %rd4240, 16, 16; + bfi.b64 %rd4245, %rd4244, %rd4237, 32, 32; + xor.b64 %rd4246, %rd4230, %rd4245; + st.local.u64 [%rd71+16], %rd4246; + +$L__BB1_28: + ld.local.u64 %rd4248, [%rd2+40]; + mov.u64 %rd19843, 0; + ld.local.u64 %rd4249, [%rd2]; + xor.b64 %rd4250, %rd4248, %rd4249; + ld.local.u64 %rd4251, [%rd2+80]; + xor.b64 %rd4252, %rd4250, %rd4251; + ld.local.u64 %rd4253, [%rd2+120]; + xor.b64 %rd4254, %rd4252, %rd4253; + ld.local.u64 %rd4255, [%rd2+160]; + xor.b64 %rd4256, %rd4254, %rd4255; + ld.local.u64 %rd4257, [%rd2+48]; + ld.local.u64 %rd4258, [%rd2+8]; + xor.b64 %rd4259, %rd4257, %rd4258; + ld.local.u64 %rd4260, [%rd2+88]; + xor.b64 %rd4261, %rd4259, %rd4260; + ld.local.u64 %rd4262, [%rd2+128]; + xor.b64 %rd4263, %rd4261, %rd4262; + ld.local.u64 %rd4264, [%rd2+168]; + xor.b64 %rd4265, %rd4263, %rd4264; + ld.local.u64 %rd4266, [%rd2+56]; + ld.local.u64 %rd4267, [%rd2+16]; + xor.b64 %rd4268, %rd4266, %rd4267; + ld.local.u64 %rd4269, [%rd2+96]; + xor.b64 %rd4270, %rd4268, %rd4269; + ld.local.u64 %rd4271, [%rd2+136]; + xor.b64 %rd4272, %rd4270, %rd4271; + ld.local.u64 %rd4273, [%rd2+176]; + xor.b64 %rd4274, %rd4272, %rd4273; + ld.local.u64 %rd4275, [%rd2+64]; + ld.local.u64 %rd4276, [%rd2+24]; + xor.b64 %rd4277, %rd4275, %rd4276; + ld.local.u64 %rd4278, [%rd2+104]; + xor.b64 %rd4279, %rd4277, %rd4278; + ld.local.u64 %rd4280, [%rd2+144]; + xor.b64 %rd4281, %rd4279, %rd4280; + ld.local.u64 %rd4282, [%rd2+184]; + xor.b64 %rd4283, %rd4281, %rd4282; + ld.local.u64 %rd4284, [%rd2+72]; + ld.local.u64 %rd4285, [%rd2+32]; + xor.b64 %rd4286, %rd4284, %rd4285; + ld.local.u64 %rd4287, [%rd2+112]; + xor.b64 %rd4288, %rd4286, %rd4287; + ld.local.u64 %rd4289, [%rd2+152]; + xor.b64 %rd4290, %rd4288, %rd4289; + ld.local.u64 %rd4291, [%rd2+192]; + xor.b64 %rd4292, %rd4290, %rd4291; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2862}, %rd4265; + } + { + .reg .b32 %dummy; + mov.b64 {%r2863,%dummy}, %rd4265; + } + shf.l.wrap.b32 %r2864, %r2863, %r2862, 1; + shf.l.wrap.b32 %r2865, %r2862, %r2863, 1; + mov.b64 %rd4293, {%r2865, %r2864}; + xor.b64 %rd4294, %rd4292, %rd4293; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2866}, %rd4274; + } + { + .reg .b32 %dummy; + mov.b64 {%r2867,%dummy}, %rd4274; + } + shf.l.wrap.b32 %r2868, %r2867, %r2866, 1; + shf.l.wrap.b32 %r2869, %r2866, %r2867, 1; + mov.b64 %rd4295, {%r2869, %r2868}; + xor.b64 %rd4296, %rd4295, %rd4256; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2870}, %rd4283; + } + { + .reg .b32 %dummy; + mov.b64 {%r2871,%dummy}, %rd4283; + } + shf.l.wrap.b32 %r2872, %r2871, %r2870, 1; + shf.l.wrap.b32 %r2873, %r2870, %r2871, 1; + mov.b64 %rd4297, {%r2873, %r2872}; + xor.b64 %rd4298, %rd4297, %rd4265; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2874}, %rd4292; + } + { + .reg .b32 %dummy; + mov.b64 {%r2875,%dummy}, %rd4292; + } + shf.l.wrap.b32 %r2876, %r2875, %r2874, 1; + shf.l.wrap.b32 %r2877, %r2874, %r2875, 1; + mov.b64 %rd4299, {%r2877, %r2876}; + xor.b64 %rd4300, %rd4299, %rd4274; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2878}, %rd4256; + } + { + .reg .b32 %dummy; + mov.b64 {%r2879,%dummy}, %rd4256; + } + shf.l.wrap.b32 %r2880, %r2879, %r2878, 1; + shf.l.wrap.b32 %r2881, %r2878, %r2879, 1; + mov.b64 %rd4301, {%r2881, %r2880}; + xor.b64 %rd4302, %rd4283, %rd4301; + xor.b64 %rd4303, %rd4294, %rd4249; + xor.b64 %rd4304, %rd4296, %rd4258; + xor.b64 %rd4305, %rd4298, %rd4267; + xor.b64 %rd4306, %rd4300, %rd4276; + xor.b64 %rd4307, %rd4302, %rd4285; + xor.b64 %rd4308, %rd4294, %rd4248; + xor.b64 %rd4309, %rd4296, %rd4257; + xor.b64 %rd4310, %rd4298, %rd4266; + xor.b64 %rd4311, %rd4300, %rd4275; + xor.b64 %rd4312, %rd4302, %rd4284; + xor.b64 %rd4313, %rd4294, %rd4251; + xor.b64 %rd4314, %rd4296, %rd4260; + xor.b64 %rd4315, %rd4298, %rd4269; + xor.b64 %rd4316, %rd4300, %rd4278; + xor.b64 %rd4317, %rd4302, %rd4287; + xor.b64 %rd4318, %rd4294, %rd4253; + xor.b64 %rd4319, %rd4296, %rd4262; + xor.b64 %rd4320, %rd4298, %rd4271; + xor.b64 %rd4321, %rd4300, %rd4280; + xor.b64 %rd4322, %rd4302, %rd4289; + xor.b64 %rd4323, %rd4294, %rd4255; + xor.b64 %rd4324, %rd4296, %rd4264; + xor.b64 %rd4325, %rd4298, %rd4273; + xor.b64 %rd4326, %rd4300, %rd4282; + xor.b64 %rd4327, %rd4302, %rd4291; + { + .reg .b32 %dummy; + mov.b64 {%r2882,%dummy}, %rd4309; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2883}, %rd4309; + } + shf.r.wrap.b32 %r2884, %r2883, %r2882, 20; + shf.r.wrap.b32 %r2885, %r2882, %r2883, 20; + mov.b64 %rd4328, {%r2885, %r2884}; + { + .reg .b32 %dummy; + mov.b64 {%r2886,%dummy}, %rd4315; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2887}, %rd4315; + } + shf.r.wrap.b32 %r2888, %r2887, %r2886, 21; + shf.r.wrap.b32 %r2889, %r2886, %r2887, 21; + mov.b64 %rd4329, {%r2889, %r2888}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2890}, %rd4321; + } + { + .reg .b32 %dummy; + mov.b64 {%r2891,%dummy}, %rd4321; + } + shf.l.wrap.b32 %r2892, %r2891, %r2890, 21; + shf.l.wrap.b32 %r2893, %r2890, %r2891, 21; + mov.b64 %rd4330, {%r2893, %r2892}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2894}, %rd4327; + } + { + .reg .b32 %dummy; + mov.b64 {%r2895,%dummy}, %rd4327; + } + shf.l.wrap.b32 %r2896, %r2895, %r2894, 14; + shf.l.wrap.b32 %r2897, %r2894, %r2895, 14; + mov.b64 %rd4331, {%r2897, %r2896}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2898}, %rd4306; + } + { + .reg .b32 %dummy; + mov.b64 {%r2899,%dummy}, %rd4306; + } + shf.l.wrap.b32 %r2900, %r2899, %r2898, 28; + shf.l.wrap.b32 %r2901, %r2898, %r2899, 28; + mov.b64 %rd4332, {%r2901, %r2900}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2902}, %rd4312; + } + { + .reg .b32 %dummy; + mov.b64 {%r2903,%dummy}, %rd4312; + } + shf.l.wrap.b32 %r2904, %r2903, %r2902, 20; + shf.l.wrap.b32 %r2905, %r2902, %r2903, 20; + mov.b64 %rd4333, {%r2905, %r2904}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2906}, %rd4313; + } + { + .reg .b32 %dummy; + mov.b64 {%r2907,%dummy}, %rd4313; + } + shf.l.wrap.b32 %r2908, %r2907, %r2906, 3; + shf.l.wrap.b32 %r2909, %r2906, %r2907, 3; + mov.b64 %rd4334, {%r2909, %r2908}; + { + .reg .b32 %dummy; + mov.b64 {%r2910,%dummy}, %rd4319; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2911}, %rd4319; + } + shf.r.wrap.b32 %r2912, %r2911, %r2910, 19; + shf.r.wrap.b32 %r2913, %r2910, %r2911, 19; + mov.b64 %rd4335, {%r2913, %r2912}; + { + .reg .b32 %dummy; + mov.b64 {%r2914,%dummy}, %rd4325; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2915}, %rd4325; + } + shf.r.wrap.b32 %r2916, %r2915, %r2914, 3; + shf.r.wrap.b32 %r2917, %r2914, %r2915, 3; + mov.b64 %rd4336, {%r2917, %r2916}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2918}, %rd4304; + } + { + .reg .b32 %dummy; + mov.b64 {%r2919,%dummy}, %rd4304; + } + shf.l.wrap.b32 %r2920, %r2919, %r2918, 1; + shf.l.wrap.b32 %r2921, %r2918, %r2919, 1; + mov.b64 %rd4337, {%r2921, %r2920}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2922}, %rd4310; + } + { + .reg .b32 %dummy; + mov.b64 {%r2923,%dummy}, %rd4310; + } + shf.l.wrap.b32 %r2924, %r2923, %r2922, 6; + shf.l.wrap.b32 %r2925, %r2922, %r2923, 6; + mov.b64 %rd4338, {%r2925, %r2924}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2926}, %rd4316; + } + { + .reg .b32 %dummy; + mov.b64 {%r2927,%dummy}, %rd4316; + } + shf.l.wrap.b32 %r2928, %r2927, %r2926, 25; + shf.l.wrap.b32 %r2929, %r2926, %r2927, 25; + mov.b64 %rd4339, {%r2929, %r2928}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2930}, %rd4322; + } + { + .reg .b32 %dummy; + mov.b64 {%r2931,%dummy}, %rd4322; + } + shf.l.wrap.b32 %r2932, %r2931, %r2930, 8; + shf.l.wrap.b32 %r2933, %r2930, %r2931, 8; + mov.b64 %rd4340, {%r2933, %r2932}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2934}, %rd4323; + } + { + .reg .b32 %dummy; + mov.b64 {%r2935,%dummy}, %rd4323; + } + shf.l.wrap.b32 %r2936, %r2935, %r2934, 18; + shf.l.wrap.b32 %r2937, %r2934, %r2935, 18; + mov.b64 %rd4341, {%r2937, %r2936}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2938}, %rd4307; + } + { + .reg .b32 %dummy; + mov.b64 {%r2939,%dummy}, %rd4307; + } + shf.l.wrap.b32 %r2940, %r2939, %r2938, 27; + shf.l.wrap.b32 %r2941, %r2938, %r2939, 27; + mov.b64 %rd4342, {%r2941, %r2940}; + { + .reg .b32 %dummy; + mov.b64 {%r2942,%dummy}, %rd4308; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2943}, %rd4308; + } + shf.r.wrap.b32 %r2944, %r2943, %r2942, 28; + shf.r.wrap.b32 %r2945, %r2942, %r2943, 28; + mov.b64 %rd4343, {%r2945, %r2944}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2946}, %rd4314; + } + { + .reg .b32 %dummy; + mov.b64 {%r2947,%dummy}, %rd4314; + } + shf.l.wrap.b32 %r2948, %r2947, %r2946, 10; + shf.l.wrap.b32 %r2949, %r2946, %r2947, 10; + mov.b64 %rd4344, {%r2949, %r2948}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2950}, %rd4320; + } + { + .reg .b32 %dummy; + mov.b64 {%r2951,%dummy}, %rd4320; + } + shf.l.wrap.b32 %r2952, %r2951, %r2950, 15; + shf.l.wrap.b32 %r2953, %r2950, %r2951, 15; + mov.b64 %rd4345, {%r2953, %r2952}; + { + .reg .b32 %dummy; + mov.b64 {%r2954,%dummy}, %rd4326; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2955}, %rd4326; + } + shf.r.wrap.b32 %r2956, %r2955, %r2954, 8; + shf.r.wrap.b32 %r2957, %r2954, %r2955, 8; + mov.b64 %rd4346, {%r2957, %r2956}; + { + .reg .b32 %dummy; + mov.b64 {%r2958,%dummy}, %rd4305; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2959}, %rd4305; + } + shf.r.wrap.b32 %r2960, %r2959, %r2958, 2; + shf.r.wrap.b32 %r2961, %r2958, %r2959, 2; + mov.b64 %rd4347, {%r2961, %r2960}; + { + .reg .b32 %dummy; + mov.b64 {%r2962,%dummy}, %rd4311; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2963}, %rd4311; + } + shf.r.wrap.b32 %r2964, %r2963, %r2962, 9; + shf.r.wrap.b32 %r2965, %r2962, %r2963, 9; + mov.b64 %rd4348, {%r2965, %r2964}; + { + .reg .b32 %dummy; + mov.b64 {%r2966,%dummy}, %rd4317; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2967}, %rd4317; + } + shf.r.wrap.b32 %r2968, %r2967, %r2966, 25; + shf.r.wrap.b32 %r2969, %r2966, %r2967, 25; + mov.b64 %rd4349, {%r2969, %r2968}; + { + .reg .b32 %dummy; + mov.b64 {%r2970,%dummy}, %rd4318; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2971}, %rd4318; + } + shf.r.wrap.b32 %r2972, %r2971, %r2970, 23; + shf.r.wrap.b32 %r2973, %r2970, %r2971, 23; + mov.b64 %rd4350, {%r2973, %r2972}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2974}, %rd4324; + } + { + .reg .b32 %dummy; + mov.b64 {%r2975,%dummy}, %rd4324; + } + shf.l.wrap.b32 %r2976, %r2975, %r2974, 2; + shf.l.wrap.b32 %r2977, %r2974, %r2975, 2; + mov.b64 %rd4351, {%r2977, %r2976}; + not.b64 %rd4352, %rd4328; + and.b64 %rd4353, %rd4329, %rd4352; + xor.b64 %rd4354, %rd4353, %rd4303; + not.b64 %rd4355, %rd4329; + and.b64 %rd4356, %rd4330, %rd4355; + xor.b64 %rd4357, %rd4356, %rd4328; + not.b64 %rd4358, %rd4330; + and.b64 %rd4359, %rd4331, %rd4358; + xor.b64 %rd4360, %rd4359, %rd4329; + not.b64 %rd4361, %rd4331; + and.b64 %rd4362, %rd4303, %rd4361; + xor.b64 %rd4363, %rd4362, %rd4330; + not.b64 %rd4364, %rd4303; + and.b64 %rd4365, %rd4328, %rd4364; + xor.b64 %rd4366, %rd4365, %rd4331; + not.b64 %rd4367, %rd4333; + and.b64 %rd4368, %rd4334, %rd4367; + xor.b64 %rd4369, %rd4368, %rd4332; + not.b64 %rd4370, %rd4334; + and.b64 %rd4371, %rd4335, %rd4370; + xor.b64 %rd4372, %rd4371, %rd4333; + not.b64 %rd4373, %rd4335; + and.b64 %rd4374, %rd4336, %rd4373; + xor.b64 %rd4375, %rd4374, %rd4334; + not.b64 %rd4376, %rd4336; + and.b64 %rd4377, %rd4332, %rd4376; + xor.b64 %rd4378, %rd4377, %rd4335; + not.b64 %rd4379, %rd4332; + and.b64 %rd4380, %rd4333, %rd4379; + xor.b64 %rd4381, %rd4380, %rd4336; + not.b64 %rd4382, %rd4338; + and.b64 %rd4383, %rd4339, %rd4382; + xor.b64 %rd4384, %rd4383, %rd4337; + not.b64 %rd4385, %rd4339; + and.b64 %rd4386, %rd4340, %rd4385; + xor.b64 %rd4387, %rd4386, %rd4338; + not.b64 %rd4388, %rd4340; + and.b64 %rd4389, %rd4341, %rd4388; + xor.b64 %rd4390, %rd4389, %rd4339; + not.b64 %rd4391, %rd4341; + and.b64 %rd4392, %rd4337, %rd4391; + xor.b64 %rd4393, %rd4392, %rd4340; + not.b64 %rd4394, %rd4337; + and.b64 %rd4395, %rd4338, %rd4394; + xor.b64 %rd4396, %rd4395, %rd4341; + not.b64 %rd4397, %rd4343; + and.b64 %rd4398, %rd4344, %rd4397; + xor.b64 %rd4399, %rd4398, %rd4342; + not.b64 %rd4400, %rd4344; + and.b64 %rd4401, %rd4345, %rd4400; + xor.b64 %rd4402, %rd4401, %rd4343; + not.b64 %rd4403, %rd4345; + and.b64 %rd4404, %rd4346, %rd4403; + xor.b64 %rd4405, %rd4404, %rd4344; + not.b64 %rd4406, %rd4346; + and.b64 %rd4407, %rd4342, %rd4406; + xor.b64 %rd4408, %rd4407, %rd4345; + not.b64 %rd4409, %rd4342; + and.b64 %rd4410, %rd4343, %rd4409; + xor.b64 %rd4411, %rd4410, %rd4346; + not.b64 %rd4412, %rd4348; + and.b64 %rd4413, %rd4349, %rd4412; + xor.b64 %rd4414, %rd4413, %rd4347; + not.b64 %rd4415, %rd4349; + and.b64 %rd4416, %rd4350, %rd4415; + xor.b64 %rd4417, %rd4416, %rd4348; + not.b64 %rd4418, %rd4350; + and.b64 %rd4419, %rd4351, %rd4418; + xor.b64 %rd4420, %rd4419, %rd4349; + not.b64 %rd4421, %rd4351; + and.b64 %rd4422, %rd4347, %rd4421; + xor.b64 %rd4423, %rd4422, %rd4350; + not.b64 %rd4424, %rd4347; + and.b64 %rd4425, %rd4348, %rd4424; + xor.b64 %rd4426, %rd4425, %rd4351; + xor.b64 %rd4427, %rd4354, %rd16; + xor.b64 %rd4428, %rd4369, %rd4427; + xor.b64 %rd4429, %rd4428, %rd4384; + xor.b64 %rd4430, %rd4429, %rd4399; + xor.b64 %rd4431, %rd4430, %rd4414; + xor.b64 %rd4432, %rd4372, %rd4357; + xor.b64 %rd4433, %rd4432, %rd4387; + xor.b64 %rd4434, %rd4433, %rd4402; + xor.b64 %rd4435, %rd4434, %rd4417; + xor.b64 %rd4436, %rd4375, %rd4360; + xor.b64 %rd4437, %rd4436, %rd4390; + xor.b64 %rd4438, %rd4437, %rd4405; + xor.b64 %rd4439, %rd4438, %rd4420; + xor.b64 %rd4440, %rd4378, %rd4363; + xor.b64 %rd4441, %rd4440, %rd4393; + xor.b64 %rd4442, %rd4441, %rd4408; + xor.b64 %rd4443, %rd4442, %rd4423; + xor.b64 %rd4444, %rd4381, %rd4366; + xor.b64 %rd4445, %rd4444, %rd4396; + xor.b64 %rd4446, %rd4445, %rd4411; + xor.b64 %rd4447, %rd4446, %rd4426; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2978}, %rd4435; + } + { + .reg .b32 %dummy; + mov.b64 {%r2979,%dummy}, %rd4435; + } + shf.l.wrap.b32 %r2980, %r2979, %r2978, 1; + shf.l.wrap.b32 %r2981, %r2978, %r2979, 1; + mov.b64 %rd4448, {%r2981, %r2980}; + xor.b64 %rd4449, %rd4447, %rd4448; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2982}, %rd4439; + } + { + .reg .b32 %dummy; + mov.b64 {%r2983,%dummy}, %rd4439; + } + shf.l.wrap.b32 %r2984, %r2983, %r2982, 1; + shf.l.wrap.b32 %r2985, %r2982, %r2983, 1; + mov.b64 %rd4450, {%r2985, %r2984}; + xor.b64 %rd4451, %rd4450, %rd4431; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2986}, %rd4443; + } + { + .reg .b32 %dummy; + mov.b64 {%r2987,%dummy}, %rd4443; + } + shf.l.wrap.b32 %r2988, %r2987, %r2986, 1; + shf.l.wrap.b32 %r2989, %r2986, %r2987, 1; + mov.b64 %rd4452, {%r2989, %r2988}; + xor.b64 %rd4453, %rd4452, %rd4435; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2990}, %rd4447; + } + { + .reg .b32 %dummy; + mov.b64 {%r2991,%dummy}, %rd4447; + } + shf.l.wrap.b32 %r2992, %r2991, %r2990, 1; + shf.l.wrap.b32 %r2993, %r2990, %r2991, 1; + mov.b64 %rd4454, {%r2993, %r2992}; + xor.b64 %rd4455, %rd4454, %rd4439; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2994}, %rd4431; + } + { + .reg .b32 %dummy; + mov.b64 {%r2995,%dummy}, %rd4431; + } + shf.l.wrap.b32 %r2996, %r2995, %r2994, 1; + shf.l.wrap.b32 %r2997, %r2994, %r2995, 1; + mov.b64 %rd4456, {%r2997, %r2996}; + xor.b64 %rd4457, %rd4443, %rd4456; + xor.b64 %rd4458, %rd4449, %rd4427; + xor.b64 %rd4459, %rd4451, %rd4357; + xor.b64 %rd4460, %rd4453, %rd4360; + xor.b64 %rd4461, %rd4455, %rd4363; + xor.b64 %rd4462, %rd4457, %rd4366; + xor.b64 %rd4463, %rd4449, %rd4369; + xor.b64 %rd4464, %rd4451, %rd4372; + xor.b64 %rd4465, %rd4453, %rd4375; + xor.b64 %rd4466, %rd4455, %rd4378; + xor.b64 %rd4467, %rd4457, %rd4381; + xor.b64 %rd4468, %rd4449, %rd4384; + xor.b64 %rd4469, %rd4451, %rd4387; + xor.b64 %rd4470, %rd4453, %rd4390; + xor.b64 %rd4471, %rd4455, %rd4393; + xor.b64 %rd4472, %rd4457, %rd4396; + xor.b64 %rd4473, %rd4449, %rd4399; + xor.b64 %rd4474, %rd4451, %rd4402; + xor.b64 %rd4475, %rd4453, %rd4405; + xor.b64 %rd4476, %rd4455, %rd4408; + xor.b64 %rd4477, %rd4457, %rd4411; + xor.b64 %rd4478, %rd4449, %rd4414; + xor.b64 %rd4479, %rd4451, %rd4417; + xor.b64 %rd4480, %rd4453, %rd4420; + xor.b64 %rd4481, %rd4455, %rd4423; + xor.b64 %rd4482, %rd4457, %rd4426; + { + .reg .b32 %dummy; + mov.b64 {%r2998,%dummy}, %rd4464; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2999}, %rd4464; + } + shf.r.wrap.b32 %r3000, %r2999, %r2998, 20; + shf.r.wrap.b32 %r3001, %r2998, %r2999, 20; + mov.b64 %rd4483, {%r3001, %r3000}; + { + .reg .b32 %dummy; + mov.b64 {%r3002,%dummy}, %rd4470; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3003}, %rd4470; + } + shf.r.wrap.b32 %r3004, %r3003, %r3002, 21; + shf.r.wrap.b32 %r3005, %r3002, %r3003, 21; + mov.b64 %rd4484, {%r3005, %r3004}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3006}, %rd4476; + } + { + .reg .b32 %dummy; + mov.b64 {%r3007,%dummy}, %rd4476; + } + shf.l.wrap.b32 %r3008, %r3007, %r3006, 21; + shf.l.wrap.b32 %r3009, %r3006, %r3007, 21; + mov.b64 %rd4485, {%r3009, %r3008}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3010}, %rd4482; + } + { + .reg .b32 %dummy; + mov.b64 {%r3011,%dummy}, %rd4482; + } + shf.l.wrap.b32 %r3012, %r3011, %r3010, 14; + shf.l.wrap.b32 %r3013, %r3010, %r3011, 14; + mov.b64 %rd4486, {%r3013, %r3012}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3014}, %rd4461; + } + { + .reg .b32 %dummy; + mov.b64 {%r3015,%dummy}, %rd4461; + } + shf.l.wrap.b32 %r3016, %r3015, %r3014, 28; + shf.l.wrap.b32 %r3017, %r3014, %r3015, 28; + mov.b64 %rd4487, {%r3017, %r3016}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3018}, %rd4467; + } + { + .reg .b32 %dummy; + mov.b64 {%r3019,%dummy}, %rd4467; + } + shf.l.wrap.b32 %r3020, %r3019, %r3018, 20; + shf.l.wrap.b32 %r3021, %r3018, %r3019, 20; + mov.b64 %rd4488, {%r3021, %r3020}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3022}, %rd4468; + } + { + .reg .b32 %dummy; + mov.b64 {%r3023,%dummy}, %rd4468; + } + shf.l.wrap.b32 %r3024, %r3023, %r3022, 3; + shf.l.wrap.b32 %r3025, %r3022, %r3023, 3; + mov.b64 %rd4489, {%r3025, %r3024}; + { + .reg .b32 %dummy; + mov.b64 {%r3026,%dummy}, %rd4474; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3027}, %rd4474; + } + shf.r.wrap.b32 %r3028, %r3027, %r3026, 19; + shf.r.wrap.b32 %r3029, %r3026, %r3027, 19; + mov.b64 %rd4490, {%r3029, %r3028}; + { + .reg .b32 %dummy; + mov.b64 {%r3030,%dummy}, %rd4480; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3031}, %rd4480; + } + shf.r.wrap.b32 %r3032, %r3031, %r3030, 3; + shf.r.wrap.b32 %r3033, %r3030, %r3031, 3; + mov.b64 %rd4491, {%r3033, %r3032}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3034}, %rd4459; + } + { + .reg .b32 %dummy; + mov.b64 {%r3035,%dummy}, %rd4459; + } + shf.l.wrap.b32 %r3036, %r3035, %r3034, 1; + shf.l.wrap.b32 %r3037, %r3034, %r3035, 1; + mov.b64 %rd4492, {%r3037, %r3036}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3038}, %rd4465; + } + { + .reg .b32 %dummy; + mov.b64 {%r3039,%dummy}, %rd4465; + } + shf.l.wrap.b32 %r3040, %r3039, %r3038, 6; + shf.l.wrap.b32 %r3041, %r3038, %r3039, 6; + mov.b64 %rd4493, {%r3041, %r3040}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3042}, %rd4471; + } + { + .reg .b32 %dummy; + mov.b64 {%r3043,%dummy}, %rd4471; + } + shf.l.wrap.b32 %r3044, %r3043, %r3042, 25; + shf.l.wrap.b32 %r3045, %r3042, %r3043, 25; + mov.b64 %rd4494, {%r3045, %r3044}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3046}, %rd4477; + } + { + .reg .b32 %dummy; + mov.b64 {%r3047,%dummy}, %rd4477; + } + shf.l.wrap.b32 %r3048, %r3047, %r3046, 8; + shf.l.wrap.b32 %r3049, %r3046, %r3047, 8; + mov.b64 %rd4495, {%r3049, %r3048}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3050}, %rd4478; + } + { + .reg .b32 %dummy; + mov.b64 {%r3051,%dummy}, %rd4478; + } + shf.l.wrap.b32 %r3052, %r3051, %r3050, 18; + shf.l.wrap.b32 %r3053, %r3050, %r3051, 18; + mov.b64 %rd4496, {%r3053, %r3052}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3054}, %rd4462; + } + { + .reg .b32 %dummy; + mov.b64 {%r3055,%dummy}, %rd4462; + } + shf.l.wrap.b32 %r3056, %r3055, %r3054, 27; + shf.l.wrap.b32 %r3057, %r3054, %r3055, 27; + mov.b64 %rd4497, {%r3057, %r3056}; + { + .reg .b32 %dummy; + mov.b64 {%r3058,%dummy}, %rd4463; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3059}, %rd4463; + } + shf.r.wrap.b32 %r3060, %r3059, %r3058, 28; + shf.r.wrap.b32 %r3061, %r3058, %r3059, 28; + mov.b64 %rd4498, {%r3061, %r3060}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3062}, %rd4469; + } + { + .reg .b32 %dummy; + mov.b64 {%r3063,%dummy}, %rd4469; + } + shf.l.wrap.b32 %r3064, %r3063, %r3062, 10; + shf.l.wrap.b32 %r3065, %r3062, %r3063, 10; + mov.b64 %rd4499, {%r3065, %r3064}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3066}, %rd4475; + } + { + .reg .b32 %dummy; + mov.b64 {%r3067,%dummy}, %rd4475; + } + shf.l.wrap.b32 %r3068, %r3067, %r3066, 15; + shf.l.wrap.b32 %r3069, %r3066, %r3067, 15; + mov.b64 %rd4500, {%r3069, %r3068}; + { + .reg .b32 %dummy; + mov.b64 {%r3070,%dummy}, %rd4481; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3071}, %rd4481; + } + shf.r.wrap.b32 %r3072, %r3071, %r3070, 8; + shf.r.wrap.b32 %r3073, %r3070, %r3071, 8; + mov.b64 %rd4501, {%r3073, %r3072}; + { + .reg .b32 %dummy; + mov.b64 {%r3074,%dummy}, %rd4460; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3075}, %rd4460; + } + shf.r.wrap.b32 %r3076, %r3075, %r3074, 2; + shf.r.wrap.b32 %r3077, %r3074, %r3075, 2; + mov.b64 %rd4502, {%r3077, %r3076}; + { + .reg .b32 %dummy; + mov.b64 {%r3078,%dummy}, %rd4466; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3079}, %rd4466; + } + shf.r.wrap.b32 %r3080, %r3079, %r3078, 9; + shf.r.wrap.b32 %r3081, %r3078, %r3079, 9; + mov.b64 %rd4503, {%r3081, %r3080}; + { + .reg .b32 %dummy; + mov.b64 {%r3082,%dummy}, %rd4472; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3083}, %rd4472; + } + shf.r.wrap.b32 %r3084, %r3083, %r3082, 25; + shf.r.wrap.b32 %r3085, %r3082, %r3083, 25; + mov.b64 %rd4504, {%r3085, %r3084}; + { + .reg .b32 %dummy; + mov.b64 {%r3086,%dummy}, %rd4473; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3087}, %rd4473; + } + shf.r.wrap.b32 %r3088, %r3087, %r3086, 23; + shf.r.wrap.b32 %r3089, %r3086, %r3087, 23; + mov.b64 %rd4505, {%r3089, %r3088}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3090}, %rd4479; + } + { + .reg .b32 %dummy; + mov.b64 {%r3091,%dummy}, %rd4479; + } + shf.l.wrap.b32 %r3092, %r3091, %r3090, 2; + shf.l.wrap.b32 %r3093, %r3090, %r3091, 2; + mov.b64 %rd4506, {%r3093, %r3092}; + not.b64 %rd4507, %rd4483; + and.b64 %rd4508, %rd4484, %rd4507; + xor.b64 %rd4509, %rd4508, %rd4458; + not.b64 %rd4510, %rd4484; + and.b64 %rd4511, %rd4485, %rd4510; + xor.b64 %rd4512, %rd4511, %rd4483; + not.b64 %rd4513, %rd4485; + and.b64 %rd4514, %rd4486, %rd4513; + xor.b64 %rd4515, %rd4514, %rd4484; + not.b64 %rd4516, %rd4486; + and.b64 %rd4517, %rd4458, %rd4516; + xor.b64 %rd4518, %rd4517, %rd4485; + not.b64 %rd4519, %rd4458; + and.b64 %rd4520, %rd4483, %rd4519; + xor.b64 %rd4521, %rd4520, %rd4486; + not.b64 %rd4522, %rd4488; + and.b64 %rd4523, %rd4489, %rd4522; + xor.b64 %rd4524, %rd4523, %rd4487; + not.b64 %rd4525, %rd4489; + and.b64 %rd4526, %rd4490, %rd4525; + xor.b64 %rd4527, %rd4526, %rd4488; + not.b64 %rd4528, %rd4490; + and.b64 %rd4529, %rd4491, %rd4528; + xor.b64 %rd4530, %rd4529, %rd4489; + not.b64 %rd4531, %rd4491; + and.b64 %rd4532, %rd4487, %rd4531; + xor.b64 %rd4533, %rd4532, %rd4490; + not.b64 %rd4534, %rd4487; + and.b64 %rd4535, %rd4488, %rd4534; + xor.b64 %rd4536, %rd4535, %rd4491; + not.b64 %rd4537, %rd4493; + and.b64 %rd4538, %rd4494, %rd4537; + xor.b64 %rd4539, %rd4538, %rd4492; + not.b64 %rd4540, %rd4494; + and.b64 %rd4541, %rd4495, %rd4540; + xor.b64 %rd4542, %rd4541, %rd4493; + not.b64 %rd4543, %rd4495; + and.b64 %rd4544, %rd4496, %rd4543; + xor.b64 %rd4545, %rd4544, %rd4494; + not.b64 %rd4546, %rd4496; + and.b64 %rd4547, %rd4492, %rd4546; + xor.b64 %rd4548, %rd4547, %rd4495; + not.b64 %rd4549, %rd4492; + and.b64 %rd4550, %rd4493, %rd4549; + xor.b64 %rd4551, %rd4550, %rd4496; + not.b64 %rd4552, %rd4498; + and.b64 %rd4553, %rd4499, %rd4552; + xor.b64 %rd4554, %rd4553, %rd4497; + not.b64 %rd4555, %rd4499; + and.b64 %rd4556, %rd4500, %rd4555; + xor.b64 %rd4557, %rd4556, %rd4498; + not.b64 %rd4558, %rd4500; + and.b64 %rd4559, %rd4501, %rd4558; + xor.b64 %rd4560, %rd4559, %rd4499; + not.b64 %rd4561, %rd4501; + and.b64 %rd4562, %rd4497, %rd4561; + xor.b64 %rd4563, %rd4562, %rd4500; + not.b64 %rd4564, %rd4497; + and.b64 %rd4565, %rd4498, %rd4564; + xor.b64 %rd4566, %rd4565, %rd4501; + not.b64 %rd4567, %rd4503; + and.b64 %rd4568, %rd4504, %rd4567; + xor.b64 %rd4569, %rd4568, %rd4502; + not.b64 %rd4570, %rd4504; + and.b64 %rd4571, %rd4505, %rd4570; + xor.b64 %rd4572, %rd4571, %rd4503; + not.b64 %rd4573, %rd4505; + and.b64 %rd4574, %rd4506, %rd4573; + xor.b64 %rd4575, %rd4574, %rd4504; + not.b64 %rd4576, %rd4506; + and.b64 %rd4577, %rd4502, %rd4576; + xor.b64 %rd4578, %rd4577, %rd4505; + not.b64 %rd4579, %rd4502; + and.b64 %rd4580, %rd4503, %rd4579; + xor.b64 %rd4581, %rd4580, %rd4506; + xor.b64 %rd4582, %rd4509, %rd17; + xor.b64 %rd4583, %rd4524, %rd4582; + xor.b64 %rd4584, %rd4583, %rd4539; + xor.b64 %rd4585, %rd4584, %rd4554; + xor.b64 %rd4586, %rd4585, %rd4569; + xor.b64 %rd4587, %rd4527, %rd4512; + xor.b64 %rd4588, %rd4587, %rd4542; + xor.b64 %rd4589, %rd4588, %rd4557; + xor.b64 %rd4590, %rd4589, %rd4572; + xor.b64 %rd4591, %rd4530, %rd4515; + xor.b64 %rd4592, %rd4591, %rd4545; + xor.b64 %rd4593, %rd4592, %rd4560; + xor.b64 %rd4594, %rd4593, %rd4575; + xor.b64 %rd4595, %rd4533, %rd4518; + xor.b64 %rd4596, %rd4595, %rd4548; + xor.b64 %rd4597, %rd4596, %rd4563; + xor.b64 %rd4598, %rd4597, %rd4578; + xor.b64 %rd4599, %rd4536, %rd4521; + xor.b64 %rd4600, %rd4599, %rd4551; + xor.b64 %rd4601, %rd4600, %rd4566; + xor.b64 %rd4602, %rd4601, %rd4581; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3094}, %rd4590; + } + { + .reg .b32 %dummy; + mov.b64 {%r3095,%dummy}, %rd4590; + } + shf.l.wrap.b32 %r3096, %r3095, %r3094, 1; + shf.l.wrap.b32 %r3097, %r3094, %r3095, 1; + mov.b64 %rd4603, {%r3097, %r3096}; + xor.b64 %rd4604, %rd4602, %rd4603; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3098}, %rd4594; + } + { + .reg .b32 %dummy; + mov.b64 {%r3099,%dummy}, %rd4594; + } + shf.l.wrap.b32 %r3100, %r3099, %r3098, 1; + shf.l.wrap.b32 %r3101, %r3098, %r3099, 1; + mov.b64 %rd4605, {%r3101, %r3100}; + xor.b64 %rd4606, %rd4605, %rd4586; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3102}, %rd4598; + } + { + .reg .b32 %dummy; + mov.b64 {%r3103,%dummy}, %rd4598; + } + shf.l.wrap.b32 %r3104, %r3103, %r3102, 1; + shf.l.wrap.b32 %r3105, %r3102, %r3103, 1; + mov.b64 %rd4607, {%r3105, %r3104}; + xor.b64 %rd4608, %rd4607, %rd4590; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3106}, %rd4602; + } + { + .reg .b32 %dummy; + mov.b64 {%r3107,%dummy}, %rd4602; + } + shf.l.wrap.b32 %r3108, %r3107, %r3106, 1; + shf.l.wrap.b32 %r3109, %r3106, %r3107, 1; + mov.b64 %rd4609, {%r3109, %r3108}; + xor.b64 %rd4610, %rd4609, %rd4594; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3110}, %rd4586; + } + { + .reg .b32 %dummy; + mov.b64 {%r3111,%dummy}, %rd4586; + } + shf.l.wrap.b32 %r3112, %r3111, %r3110, 1; + shf.l.wrap.b32 %r3113, %r3110, %r3111, 1; + mov.b64 %rd4611, {%r3113, %r3112}; + xor.b64 %rd4612, %rd4598, %rd4611; + xor.b64 %rd4613, %rd4604, %rd4582; + xor.b64 %rd4614, %rd4606, %rd4512; + xor.b64 %rd4615, %rd4608, %rd4515; + xor.b64 %rd4616, %rd4610, %rd4518; + xor.b64 %rd4617, %rd4612, %rd4521; + xor.b64 %rd4618, %rd4604, %rd4524; + xor.b64 %rd4619, %rd4606, %rd4527; + xor.b64 %rd4620, %rd4608, %rd4530; + xor.b64 %rd4621, %rd4610, %rd4533; + xor.b64 %rd4622, %rd4612, %rd4536; + xor.b64 %rd4623, %rd4604, %rd4539; + xor.b64 %rd4624, %rd4606, %rd4542; + xor.b64 %rd4625, %rd4608, %rd4545; + xor.b64 %rd4626, %rd4610, %rd4548; + xor.b64 %rd4627, %rd4612, %rd4551; + xor.b64 %rd4628, %rd4604, %rd4554; + xor.b64 %rd4629, %rd4606, %rd4557; + xor.b64 %rd4630, %rd4608, %rd4560; + xor.b64 %rd4631, %rd4610, %rd4563; + xor.b64 %rd4632, %rd4612, %rd4566; + xor.b64 %rd4633, %rd4604, %rd4569; + xor.b64 %rd4634, %rd4606, %rd4572; + xor.b64 %rd4635, %rd4608, %rd4575; + xor.b64 %rd4636, %rd4610, %rd4578; + xor.b64 %rd4637, %rd4612, %rd4581; + { + .reg .b32 %dummy; + mov.b64 {%r3114,%dummy}, %rd4619; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3115}, %rd4619; + } + shf.r.wrap.b32 %r3116, %r3115, %r3114, 20; + shf.r.wrap.b32 %r3117, %r3114, %r3115, 20; + mov.b64 %rd4638, {%r3117, %r3116}; + { + .reg .b32 %dummy; + mov.b64 {%r3118,%dummy}, %rd4625; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3119}, %rd4625; + } + shf.r.wrap.b32 %r3120, %r3119, %r3118, 21; + shf.r.wrap.b32 %r3121, %r3118, %r3119, 21; + mov.b64 %rd4639, {%r3121, %r3120}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3122}, %rd4631; + } + { + .reg .b32 %dummy; + mov.b64 {%r3123,%dummy}, %rd4631; + } + shf.l.wrap.b32 %r3124, %r3123, %r3122, 21; + shf.l.wrap.b32 %r3125, %r3122, %r3123, 21; + mov.b64 %rd4640, {%r3125, %r3124}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3126}, %rd4637; + } + { + .reg .b32 %dummy; + mov.b64 {%r3127,%dummy}, %rd4637; + } + shf.l.wrap.b32 %r3128, %r3127, %r3126, 14; + shf.l.wrap.b32 %r3129, %r3126, %r3127, 14; + mov.b64 %rd4641, {%r3129, %r3128}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3130}, %rd4616; + } + { + .reg .b32 %dummy; + mov.b64 {%r3131,%dummy}, %rd4616; + } + shf.l.wrap.b32 %r3132, %r3131, %r3130, 28; + shf.l.wrap.b32 %r3133, %r3130, %r3131, 28; + mov.b64 %rd4642, {%r3133, %r3132}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3134}, %rd4622; + } + { + .reg .b32 %dummy; + mov.b64 {%r3135,%dummy}, %rd4622; + } + shf.l.wrap.b32 %r3136, %r3135, %r3134, 20; + shf.l.wrap.b32 %r3137, %r3134, %r3135, 20; + mov.b64 %rd4643, {%r3137, %r3136}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3138}, %rd4623; + } + { + .reg .b32 %dummy; + mov.b64 {%r3139,%dummy}, %rd4623; + } + shf.l.wrap.b32 %r3140, %r3139, %r3138, 3; + shf.l.wrap.b32 %r3141, %r3138, %r3139, 3; + mov.b64 %rd4644, {%r3141, %r3140}; + { + .reg .b32 %dummy; + mov.b64 {%r3142,%dummy}, %rd4629; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3143}, %rd4629; + } + shf.r.wrap.b32 %r3144, %r3143, %r3142, 19; + shf.r.wrap.b32 %r3145, %r3142, %r3143, 19; + mov.b64 %rd4645, {%r3145, %r3144}; + { + .reg .b32 %dummy; + mov.b64 {%r3146,%dummy}, %rd4635; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3147}, %rd4635; + } + shf.r.wrap.b32 %r3148, %r3147, %r3146, 3; + shf.r.wrap.b32 %r3149, %r3146, %r3147, 3; + mov.b64 %rd4646, {%r3149, %r3148}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3150}, %rd4614; + } + { + .reg .b32 %dummy; + mov.b64 {%r3151,%dummy}, %rd4614; + } + shf.l.wrap.b32 %r3152, %r3151, %r3150, 1; + shf.l.wrap.b32 %r3153, %r3150, %r3151, 1; + mov.b64 %rd4647, {%r3153, %r3152}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3154}, %rd4620; + } + { + .reg .b32 %dummy; + mov.b64 {%r3155,%dummy}, %rd4620; + } + shf.l.wrap.b32 %r3156, %r3155, %r3154, 6; + shf.l.wrap.b32 %r3157, %r3154, %r3155, 6; + mov.b64 %rd4648, {%r3157, %r3156}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3158}, %rd4626; + } + { + .reg .b32 %dummy; + mov.b64 {%r3159,%dummy}, %rd4626; + } + shf.l.wrap.b32 %r3160, %r3159, %r3158, 25; + shf.l.wrap.b32 %r3161, %r3158, %r3159, 25; + mov.b64 %rd4649, {%r3161, %r3160}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3162}, %rd4632; + } + { + .reg .b32 %dummy; + mov.b64 {%r3163,%dummy}, %rd4632; + } + shf.l.wrap.b32 %r3164, %r3163, %r3162, 8; + shf.l.wrap.b32 %r3165, %r3162, %r3163, 8; + mov.b64 %rd4650, {%r3165, %r3164}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3166}, %rd4633; + } + { + .reg .b32 %dummy; + mov.b64 {%r3167,%dummy}, %rd4633; + } + shf.l.wrap.b32 %r3168, %r3167, %r3166, 18; + shf.l.wrap.b32 %r3169, %r3166, %r3167, 18; + mov.b64 %rd4651, {%r3169, %r3168}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3170}, %rd4617; + } + { + .reg .b32 %dummy; + mov.b64 {%r3171,%dummy}, %rd4617; + } + shf.l.wrap.b32 %r3172, %r3171, %r3170, 27; + shf.l.wrap.b32 %r3173, %r3170, %r3171, 27; + mov.b64 %rd4652, {%r3173, %r3172}; + { + .reg .b32 %dummy; + mov.b64 {%r3174,%dummy}, %rd4618; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3175}, %rd4618; + } + shf.r.wrap.b32 %r3176, %r3175, %r3174, 28; + shf.r.wrap.b32 %r3177, %r3174, %r3175, 28; + mov.b64 %rd4653, {%r3177, %r3176}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3178}, %rd4624; + } + { + .reg .b32 %dummy; + mov.b64 {%r3179,%dummy}, %rd4624; + } + shf.l.wrap.b32 %r3180, %r3179, %r3178, 10; + shf.l.wrap.b32 %r3181, %r3178, %r3179, 10; + mov.b64 %rd4654, {%r3181, %r3180}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3182}, %rd4630; + } + { + .reg .b32 %dummy; + mov.b64 {%r3183,%dummy}, %rd4630; + } + shf.l.wrap.b32 %r3184, %r3183, %r3182, 15; + shf.l.wrap.b32 %r3185, %r3182, %r3183, 15; + mov.b64 %rd4655, {%r3185, %r3184}; + { + .reg .b32 %dummy; + mov.b64 {%r3186,%dummy}, %rd4636; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3187}, %rd4636; + } + shf.r.wrap.b32 %r3188, %r3187, %r3186, 8; + shf.r.wrap.b32 %r3189, %r3186, %r3187, 8; + mov.b64 %rd4656, {%r3189, %r3188}; + { + .reg .b32 %dummy; + mov.b64 {%r3190,%dummy}, %rd4615; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3191}, %rd4615; + } + shf.r.wrap.b32 %r3192, %r3191, %r3190, 2; + shf.r.wrap.b32 %r3193, %r3190, %r3191, 2; + mov.b64 %rd4657, {%r3193, %r3192}; + { + .reg .b32 %dummy; + mov.b64 {%r3194,%dummy}, %rd4621; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3195}, %rd4621; + } + shf.r.wrap.b32 %r3196, %r3195, %r3194, 9; + shf.r.wrap.b32 %r3197, %r3194, %r3195, 9; + mov.b64 %rd4658, {%r3197, %r3196}; + { + .reg .b32 %dummy; + mov.b64 {%r3198,%dummy}, %rd4627; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3199}, %rd4627; + } + shf.r.wrap.b32 %r3200, %r3199, %r3198, 25; + shf.r.wrap.b32 %r3201, %r3198, %r3199, 25; + mov.b64 %rd4659, {%r3201, %r3200}; + { + .reg .b32 %dummy; + mov.b64 {%r3202,%dummy}, %rd4628; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3203}, %rd4628; + } + shf.r.wrap.b32 %r3204, %r3203, %r3202, 23; + shf.r.wrap.b32 %r3205, %r3202, %r3203, 23; + mov.b64 %rd4660, {%r3205, %r3204}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3206}, %rd4634; + } + { + .reg .b32 %dummy; + mov.b64 {%r3207,%dummy}, %rd4634; + } + shf.l.wrap.b32 %r3208, %r3207, %r3206, 2; + shf.l.wrap.b32 %r3209, %r3206, %r3207, 2; + mov.b64 %rd4661, {%r3209, %r3208}; + not.b64 %rd4662, %rd4638; + and.b64 %rd4663, %rd4639, %rd4662; + xor.b64 %rd4664, %rd4663, %rd4613; + not.b64 %rd4665, %rd4639; + and.b64 %rd4666, %rd4640, %rd4665; + xor.b64 %rd4667, %rd4666, %rd4638; + not.b64 %rd4668, %rd4640; + and.b64 %rd4669, %rd4641, %rd4668; + xor.b64 %rd4670, %rd4669, %rd4639; + not.b64 %rd4671, %rd4641; + and.b64 %rd4672, %rd4613, %rd4671; + xor.b64 %rd4673, %rd4672, %rd4640; + not.b64 %rd4674, %rd4613; + and.b64 %rd4675, %rd4638, %rd4674; + xor.b64 %rd4676, %rd4675, %rd4641; + not.b64 %rd4677, %rd4643; + and.b64 %rd4678, %rd4644, %rd4677; + xor.b64 %rd4679, %rd4678, %rd4642; + not.b64 %rd4680, %rd4644; + and.b64 %rd4681, %rd4645, %rd4680; + xor.b64 %rd4682, %rd4681, %rd4643; + not.b64 %rd4683, %rd4645; + and.b64 %rd4684, %rd4646, %rd4683; + xor.b64 %rd4685, %rd4684, %rd4644; + not.b64 %rd4686, %rd4646; + and.b64 %rd4687, %rd4642, %rd4686; + xor.b64 %rd4688, %rd4687, %rd4645; + not.b64 %rd4689, %rd4642; + and.b64 %rd4690, %rd4643, %rd4689; + xor.b64 %rd4691, %rd4690, %rd4646; + not.b64 %rd4692, %rd4648; + and.b64 %rd4693, %rd4649, %rd4692; + xor.b64 %rd4694, %rd4693, %rd4647; + not.b64 %rd4695, %rd4649; + and.b64 %rd4696, %rd4650, %rd4695; + xor.b64 %rd4697, %rd4696, %rd4648; + not.b64 %rd4698, %rd4650; + and.b64 %rd4699, %rd4651, %rd4698; + xor.b64 %rd4700, %rd4699, %rd4649; + not.b64 %rd4701, %rd4651; + and.b64 %rd4702, %rd4647, %rd4701; + xor.b64 %rd4703, %rd4702, %rd4650; + not.b64 %rd4704, %rd4647; + and.b64 %rd4705, %rd4648, %rd4704; + xor.b64 %rd4706, %rd4705, %rd4651; + not.b64 %rd4707, %rd4653; + and.b64 %rd4708, %rd4654, %rd4707; + xor.b64 %rd4709, %rd4708, %rd4652; + not.b64 %rd4710, %rd4654; + and.b64 %rd4711, %rd4655, %rd4710; + xor.b64 %rd4712, %rd4711, %rd4653; + not.b64 %rd4713, %rd4655; + and.b64 %rd4714, %rd4656, %rd4713; + xor.b64 %rd4715, %rd4714, %rd4654; + not.b64 %rd4716, %rd4656; + and.b64 %rd4717, %rd4652, %rd4716; + xor.b64 %rd4718, %rd4717, %rd4655; + not.b64 %rd4719, %rd4652; + and.b64 %rd4720, %rd4653, %rd4719; + xor.b64 %rd4721, %rd4720, %rd4656; + not.b64 %rd4722, %rd4658; + and.b64 %rd4723, %rd4659, %rd4722; + xor.b64 %rd4724, %rd4723, %rd4657; + not.b64 %rd4725, %rd4659; + and.b64 %rd4726, %rd4660, %rd4725; + xor.b64 %rd4727, %rd4726, %rd4658; + not.b64 %rd4728, %rd4660; + and.b64 %rd4729, %rd4661, %rd4728; + xor.b64 %rd4730, %rd4729, %rd4659; + not.b64 %rd4731, %rd4661; + and.b64 %rd4732, %rd4657, %rd4731; + xor.b64 %rd4733, %rd4732, %rd4660; + not.b64 %rd4734, %rd4657; + and.b64 %rd4735, %rd4658, %rd4734; + xor.b64 %rd4736, %rd4735, %rd4661; + xor.b64 %rd4737, %rd4664, %rd18; + xor.b64 %rd4738, %rd4679, %rd4737; + xor.b64 %rd4739, %rd4738, %rd4694; + xor.b64 %rd4740, %rd4739, %rd4709; + xor.b64 %rd4741, %rd4740, %rd4724; + xor.b64 %rd4742, %rd4682, %rd4667; + xor.b64 %rd4743, %rd4742, %rd4697; + xor.b64 %rd4744, %rd4743, %rd4712; + xor.b64 %rd4745, %rd4744, %rd4727; + xor.b64 %rd4746, %rd4685, %rd4670; + xor.b64 %rd4747, %rd4746, %rd4700; + xor.b64 %rd4748, %rd4747, %rd4715; + xor.b64 %rd4749, %rd4748, %rd4730; + xor.b64 %rd4750, %rd4688, %rd4673; + xor.b64 %rd4751, %rd4750, %rd4703; + xor.b64 %rd4752, %rd4751, %rd4718; + xor.b64 %rd4753, %rd4752, %rd4733; + xor.b64 %rd4754, %rd4691, %rd4676; + xor.b64 %rd4755, %rd4754, %rd4706; + xor.b64 %rd4756, %rd4755, %rd4721; + xor.b64 %rd4757, %rd4756, %rd4736; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3210}, %rd4745; + } + { + .reg .b32 %dummy; + mov.b64 {%r3211,%dummy}, %rd4745; + } + shf.l.wrap.b32 %r3212, %r3211, %r3210, 1; + shf.l.wrap.b32 %r3213, %r3210, %r3211, 1; + mov.b64 %rd4758, {%r3213, %r3212}; + xor.b64 %rd4759, %rd4757, %rd4758; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3214}, %rd4749; + } + { + .reg .b32 %dummy; + mov.b64 {%r3215,%dummy}, %rd4749; + } + shf.l.wrap.b32 %r3216, %r3215, %r3214, 1; + shf.l.wrap.b32 %r3217, %r3214, %r3215, 1; + mov.b64 %rd4760, {%r3217, %r3216}; + xor.b64 %rd4761, %rd4760, %rd4741; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3218}, %rd4753; + } + { + .reg .b32 %dummy; + mov.b64 {%r3219,%dummy}, %rd4753; + } + shf.l.wrap.b32 %r3220, %r3219, %r3218, 1; + shf.l.wrap.b32 %r3221, %r3218, %r3219, 1; + mov.b64 %rd4762, {%r3221, %r3220}; + xor.b64 %rd4763, %rd4762, %rd4745; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3222}, %rd4757; + } + { + .reg .b32 %dummy; + mov.b64 {%r3223,%dummy}, %rd4757; + } + shf.l.wrap.b32 %r3224, %r3223, %r3222, 1; + shf.l.wrap.b32 %r3225, %r3222, %r3223, 1; + mov.b64 %rd4764, {%r3225, %r3224}; + xor.b64 %rd4765, %rd4764, %rd4749; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3226}, %rd4741; + } + { + .reg .b32 %dummy; + mov.b64 {%r3227,%dummy}, %rd4741; + } + shf.l.wrap.b32 %r3228, %r3227, %r3226, 1; + shf.l.wrap.b32 %r3229, %r3226, %r3227, 1; + mov.b64 %rd4766, {%r3229, %r3228}; + xor.b64 %rd4767, %rd4753, %rd4766; + xor.b64 %rd4768, %rd4759, %rd4737; + xor.b64 %rd4769, %rd4761, %rd4667; + xor.b64 %rd4770, %rd4763, %rd4670; + xor.b64 %rd4771, %rd4765, %rd4673; + xor.b64 %rd4772, %rd4767, %rd4676; + xor.b64 %rd4773, %rd4759, %rd4679; + xor.b64 %rd4774, %rd4761, %rd4682; + xor.b64 %rd4775, %rd4763, %rd4685; + xor.b64 %rd4776, %rd4765, %rd4688; + xor.b64 %rd4777, %rd4767, %rd4691; + xor.b64 %rd4778, %rd4759, %rd4694; + xor.b64 %rd4779, %rd4761, %rd4697; + xor.b64 %rd4780, %rd4763, %rd4700; + xor.b64 %rd4781, %rd4765, %rd4703; + xor.b64 %rd4782, %rd4767, %rd4706; + xor.b64 %rd4783, %rd4759, %rd4709; + xor.b64 %rd4784, %rd4761, %rd4712; + xor.b64 %rd4785, %rd4763, %rd4715; + xor.b64 %rd4786, %rd4765, %rd4718; + xor.b64 %rd4787, %rd4767, %rd4721; + xor.b64 %rd4788, %rd4759, %rd4724; + xor.b64 %rd4789, %rd4761, %rd4727; + xor.b64 %rd4790, %rd4763, %rd4730; + xor.b64 %rd4791, %rd4765, %rd4733; + xor.b64 %rd4792, %rd4767, %rd4736; + { + .reg .b32 %dummy; + mov.b64 {%r3230,%dummy}, %rd4774; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3231}, %rd4774; + } + shf.r.wrap.b32 %r3232, %r3231, %r3230, 20; + shf.r.wrap.b32 %r3233, %r3230, %r3231, 20; + mov.b64 %rd4793, {%r3233, %r3232}; + { + .reg .b32 %dummy; + mov.b64 {%r3234,%dummy}, %rd4780; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3235}, %rd4780; + } + shf.r.wrap.b32 %r3236, %r3235, %r3234, 21; + shf.r.wrap.b32 %r3237, %r3234, %r3235, 21; + mov.b64 %rd4794, {%r3237, %r3236}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3238}, %rd4786; + } + { + .reg .b32 %dummy; + mov.b64 {%r3239,%dummy}, %rd4786; + } + shf.l.wrap.b32 %r3240, %r3239, %r3238, 21; + shf.l.wrap.b32 %r3241, %r3238, %r3239, 21; + mov.b64 %rd4795, {%r3241, %r3240}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3242}, %rd4792; + } + { + .reg .b32 %dummy; + mov.b64 {%r3243,%dummy}, %rd4792; + } + shf.l.wrap.b32 %r3244, %r3243, %r3242, 14; + shf.l.wrap.b32 %r3245, %r3242, %r3243, 14; + mov.b64 %rd4796, {%r3245, %r3244}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3246}, %rd4771; + } + { + .reg .b32 %dummy; + mov.b64 {%r3247,%dummy}, %rd4771; + } + shf.l.wrap.b32 %r3248, %r3247, %r3246, 28; + shf.l.wrap.b32 %r3249, %r3246, %r3247, 28; + mov.b64 %rd4797, {%r3249, %r3248}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3250}, %rd4777; + } + { + .reg .b32 %dummy; + mov.b64 {%r3251,%dummy}, %rd4777; + } + shf.l.wrap.b32 %r3252, %r3251, %r3250, 20; + shf.l.wrap.b32 %r3253, %r3250, %r3251, 20; + mov.b64 %rd4798, {%r3253, %r3252}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3254}, %rd4778; + } + { + .reg .b32 %dummy; + mov.b64 {%r3255,%dummy}, %rd4778; + } + shf.l.wrap.b32 %r3256, %r3255, %r3254, 3; + shf.l.wrap.b32 %r3257, %r3254, %r3255, 3; + mov.b64 %rd4799, {%r3257, %r3256}; + { + .reg .b32 %dummy; + mov.b64 {%r3258,%dummy}, %rd4784; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3259}, %rd4784; + } + shf.r.wrap.b32 %r3260, %r3259, %r3258, 19; + shf.r.wrap.b32 %r3261, %r3258, %r3259, 19; + mov.b64 %rd4800, {%r3261, %r3260}; + { + .reg .b32 %dummy; + mov.b64 {%r3262,%dummy}, %rd4790; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3263}, %rd4790; + } + shf.r.wrap.b32 %r3264, %r3263, %r3262, 3; + shf.r.wrap.b32 %r3265, %r3262, %r3263, 3; + mov.b64 %rd4801, {%r3265, %r3264}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3266}, %rd4769; + } + { + .reg .b32 %dummy; + mov.b64 {%r3267,%dummy}, %rd4769; + } + shf.l.wrap.b32 %r3268, %r3267, %r3266, 1; + shf.l.wrap.b32 %r3269, %r3266, %r3267, 1; + mov.b64 %rd4802, {%r3269, %r3268}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3270}, %rd4775; + } + { + .reg .b32 %dummy; + mov.b64 {%r3271,%dummy}, %rd4775; + } + shf.l.wrap.b32 %r3272, %r3271, %r3270, 6; + shf.l.wrap.b32 %r3273, %r3270, %r3271, 6; + mov.b64 %rd4803, {%r3273, %r3272}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3274}, %rd4781; + } + { + .reg .b32 %dummy; + mov.b64 {%r3275,%dummy}, %rd4781; + } + shf.l.wrap.b32 %r3276, %r3275, %r3274, 25; + shf.l.wrap.b32 %r3277, %r3274, %r3275, 25; + mov.b64 %rd4804, {%r3277, %r3276}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3278}, %rd4787; + } + { + .reg .b32 %dummy; + mov.b64 {%r3279,%dummy}, %rd4787; + } + shf.l.wrap.b32 %r3280, %r3279, %r3278, 8; + shf.l.wrap.b32 %r3281, %r3278, %r3279, 8; + mov.b64 %rd4805, {%r3281, %r3280}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3282}, %rd4788; + } + { + .reg .b32 %dummy; + mov.b64 {%r3283,%dummy}, %rd4788; + } + shf.l.wrap.b32 %r3284, %r3283, %r3282, 18; + shf.l.wrap.b32 %r3285, %r3282, %r3283, 18; + mov.b64 %rd4806, {%r3285, %r3284}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3286}, %rd4772; + } + { + .reg .b32 %dummy; + mov.b64 {%r3287,%dummy}, %rd4772; + } + shf.l.wrap.b32 %r3288, %r3287, %r3286, 27; + shf.l.wrap.b32 %r3289, %r3286, %r3287, 27; + mov.b64 %rd4807, {%r3289, %r3288}; + { + .reg .b32 %dummy; + mov.b64 {%r3290,%dummy}, %rd4773; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3291}, %rd4773; + } + shf.r.wrap.b32 %r3292, %r3291, %r3290, 28; + shf.r.wrap.b32 %r3293, %r3290, %r3291, 28; + mov.b64 %rd4808, {%r3293, %r3292}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3294}, %rd4779; + } + { + .reg .b32 %dummy; + mov.b64 {%r3295,%dummy}, %rd4779; + } + shf.l.wrap.b32 %r3296, %r3295, %r3294, 10; + shf.l.wrap.b32 %r3297, %r3294, %r3295, 10; + mov.b64 %rd4809, {%r3297, %r3296}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3298}, %rd4785; + } + { + .reg .b32 %dummy; + mov.b64 {%r3299,%dummy}, %rd4785; + } + shf.l.wrap.b32 %r3300, %r3299, %r3298, 15; + shf.l.wrap.b32 %r3301, %r3298, %r3299, 15; + mov.b64 %rd4810, {%r3301, %r3300}; + { + .reg .b32 %dummy; + mov.b64 {%r3302,%dummy}, %rd4791; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3303}, %rd4791; + } + shf.r.wrap.b32 %r3304, %r3303, %r3302, 8; + shf.r.wrap.b32 %r3305, %r3302, %r3303, 8; + mov.b64 %rd4811, {%r3305, %r3304}; + { + .reg .b32 %dummy; + mov.b64 {%r3306,%dummy}, %rd4770; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3307}, %rd4770; + } + shf.r.wrap.b32 %r3308, %r3307, %r3306, 2; + shf.r.wrap.b32 %r3309, %r3306, %r3307, 2; + mov.b64 %rd4812, {%r3309, %r3308}; + { + .reg .b32 %dummy; + mov.b64 {%r3310,%dummy}, %rd4776; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3311}, %rd4776; + } + shf.r.wrap.b32 %r3312, %r3311, %r3310, 9; + shf.r.wrap.b32 %r3313, %r3310, %r3311, 9; + mov.b64 %rd4813, {%r3313, %r3312}; + { + .reg .b32 %dummy; + mov.b64 {%r3314,%dummy}, %rd4782; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3315}, %rd4782; + } + shf.r.wrap.b32 %r3316, %r3315, %r3314, 25; + shf.r.wrap.b32 %r3317, %r3314, %r3315, 25; + mov.b64 %rd4814, {%r3317, %r3316}; + { + .reg .b32 %dummy; + mov.b64 {%r3318,%dummy}, %rd4783; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3319}, %rd4783; + } + shf.r.wrap.b32 %r3320, %r3319, %r3318, 23; + shf.r.wrap.b32 %r3321, %r3318, %r3319, 23; + mov.b64 %rd4815, {%r3321, %r3320}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3322}, %rd4789; + } + { + .reg .b32 %dummy; + mov.b64 {%r3323,%dummy}, %rd4789; + } + shf.l.wrap.b32 %r3324, %r3323, %r3322, 2; + shf.l.wrap.b32 %r3325, %r3322, %r3323, 2; + mov.b64 %rd4816, {%r3325, %r3324}; + not.b64 %rd4817, %rd4793; + and.b64 %rd4818, %rd4794, %rd4817; + xor.b64 %rd4819, %rd4818, %rd4768; + not.b64 %rd4820, %rd4794; + and.b64 %rd4821, %rd4795, %rd4820; + xor.b64 %rd4822, %rd4821, %rd4793; + not.b64 %rd4823, %rd4795; + and.b64 %rd4824, %rd4796, %rd4823; + xor.b64 %rd4825, %rd4824, %rd4794; + not.b64 %rd4826, %rd4796; + and.b64 %rd4827, %rd4768, %rd4826; + xor.b64 %rd4828, %rd4827, %rd4795; + not.b64 %rd4829, %rd4768; + and.b64 %rd4830, %rd4793, %rd4829; + xor.b64 %rd4831, %rd4830, %rd4796; + not.b64 %rd4832, %rd4798; + and.b64 %rd4833, %rd4799, %rd4832; + xor.b64 %rd4834, %rd4833, %rd4797; + not.b64 %rd4835, %rd4799; + and.b64 %rd4836, %rd4800, %rd4835; + xor.b64 %rd4837, %rd4836, %rd4798; + not.b64 %rd4838, %rd4800; + and.b64 %rd4839, %rd4801, %rd4838; + xor.b64 %rd4840, %rd4839, %rd4799; + not.b64 %rd4841, %rd4801; + and.b64 %rd4842, %rd4797, %rd4841; + xor.b64 %rd4843, %rd4842, %rd4800; + not.b64 %rd4844, %rd4797; + and.b64 %rd4845, %rd4798, %rd4844; + xor.b64 %rd4846, %rd4845, %rd4801; + not.b64 %rd4847, %rd4803; + and.b64 %rd4848, %rd4804, %rd4847; + xor.b64 %rd4849, %rd4848, %rd4802; + not.b64 %rd4850, %rd4804; + and.b64 %rd4851, %rd4805, %rd4850; + xor.b64 %rd4852, %rd4851, %rd4803; + not.b64 %rd4853, %rd4805; + and.b64 %rd4854, %rd4806, %rd4853; + xor.b64 %rd4855, %rd4854, %rd4804; + not.b64 %rd4856, %rd4806; + and.b64 %rd4857, %rd4802, %rd4856; + xor.b64 %rd4858, %rd4857, %rd4805; + not.b64 %rd4859, %rd4802; + and.b64 %rd4860, %rd4803, %rd4859; + xor.b64 %rd4861, %rd4860, %rd4806; + not.b64 %rd4862, %rd4808; + and.b64 %rd4863, %rd4809, %rd4862; + xor.b64 %rd4864, %rd4863, %rd4807; + not.b64 %rd4865, %rd4809; + and.b64 %rd4866, %rd4810, %rd4865; + xor.b64 %rd4867, %rd4866, %rd4808; + not.b64 %rd4868, %rd4810; + and.b64 %rd4869, %rd4811, %rd4868; + xor.b64 %rd4870, %rd4869, %rd4809; + not.b64 %rd4871, %rd4811; + and.b64 %rd4872, %rd4807, %rd4871; + xor.b64 %rd4873, %rd4872, %rd4810; + not.b64 %rd4874, %rd4807; + and.b64 %rd4875, %rd4808, %rd4874; + xor.b64 %rd4876, %rd4875, %rd4811; + not.b64 %rd4877, %rd4813; + and.b64 %rd4878, %rd4814, %rd4877; + xor.b64 %rd4879, %rd4878, %rd4812; + not.b64 %rd4880, %rd4814; + and.b64 %rd4881, %rd4815, %rd4880; + xor.b64 %rd4882, %rd4881, %rd4813; + not.b64 %rd4883, %rd4815; + and.b64 %rd4884, %rd4816, %rd4883; + xor.b64 %rd4885, %rd4884, %rd4814; + not.b64 %rd4886, %rd4816; + and.b64 %rd4887, %rd4812, %rd4886; + xor.b64 %rd4888, %rd4887, %rd4815; + not.b64 %rd4889, %rd4812; + and.b64 %rd4890, %rd4813, %rd4889; + xor.b64 %rd4891, %rd4890, %rd4816; + xor.b64 %rd4892, %rd4819, %rd19; + xor.b64 %rd4893, %rd4834, %rd4892; + xor.b64 %rd4894, %rd4893, %rd4849; + xor.b64 %rd4895, %rd4894, %rd4864; + xor.b64 %rd4896, %rd4895, %rd4879; + xor.b64 %rd4897, %rd4837, %rd4822; + xor.b64 %rd4898, %rd4897, %rd4852; + xor.b64 %rd4899, %rd4898, %rd4867; + xor.b64 %rd4900, %rd4899, %rd4882; + xor.b64 %rd4901, %rd4840, %rd4825; + xor.b64 %rd4902, %rd4901, %rd4855; + xor.b64 %rd4903, %rd4902, %rd4870; + xor.b64 %rd4904, %rd4903, %rd4885; + xor.b64 %rd4905, %rd4843, %rd4828; + xor.b64 %rd4906, %rd4905, %rd4858; + xor.b64 %rd4907, %rd4906, %rd4873; + xor.b64 %rd4908, %rd4907, %rd4888; + xor.b64 %rd4909, %rd4846, %rd4831; + xor.b64 %rd4910, %rd4909, %rd4861; + xor.b64 %rd4911, %rd4910, %rd4876; + xor.b64 %rd4912, %rd4911, %rd4891; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3326}, %rd4900; + } + { + .reg .b32 %dummy; + mov.b64 {%r3327,%dummy}, %rd4900; + } + shf.l.wrap.b32 %r3328, %r3327, %r3326, 1; + shf.l.wrap.b32 %r3329, %r3326, %r3327, 1; + mov.b64 %rd4913, {%r3329, %r3328}; + xor.b64 %rd4914, %rd4912, %rd4913; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3330}, %rd4904; + } + { + .reg .b32 %dummy; + mov.b64 {%r3331,%dummy}, %rd4904; + } + shf.l.wrap.b32 %r3332, %r3331, %r3330, 1; + shf.l.wrap.b32 %r3333, %r3330, %r3331, 1; + mov.b64 %rd4915, {%r3333, %r3332}; + xor.b64 %rd4916, %rd4915, %rd4896; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3334}, %rd4908; + } + { + .reg .b32 %dummy; + mov.b64 {%r3335,%dummy}, %rd4908; + } + shf.l.wrap.b32 %r3336, %r3335, %r3334, 1; + shf.l.wrap.b32 %r3337, %r3334, %r3335, 1; + mov.b64 %rd4917, {%r3337, %r3336}; + xor.b64 %rd4918, %rd4917, %rd4900; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3338}, %rd4912; + } + { + .reg .b32 %dummy; + mov.b64 {%r3339,%dummy}, %rd4912; + } + shf.l.wrap.b32 %r3340, %r3339, %r3338, 1; + shf.l.wrap.b32 %r3341, %r3338, %r3339, 1; + mov.b64 %rd4919, {%r3341, %r3340}; + xor.b64 %rd4920, %rd4919, %rd4904; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3342}, %rd4896; + } + { + .reg .b32 %dummy; + mov.b64 {%r3343,%dummy}, %rd4896; + } + shf.l.wrap.b32 %r3344, %r3343, %r3342, 1; + shf.l.wrap.b32 %r3345, %r3342, %r3343, 1; + mov.b64 %rd4921, {%r3345, %r3344}; + xor.b64 %rd4922, %rd4908, %rd4921; + xor.b64 %rd4923, %rd4914, %rd4892; + xor.b64 %rd4924, %rd4916, %rd4822; + xor.b64 %rd4925, %rd4918, %rd4825; + xor.b64 %rd4926, %rd4920, %rd4828; + xor.b64 %rd4927, %rd4922, %rd4831; + xor.b64 %rd4928, %rd4914, %rd4834; + xor.b64 %rd4929, %rd4916, %rd4837; + xor.b64 %rd4930, %rd4918, %rd4840; + xor.b64 %rd4931, %rd4920, %rd4843; + xor.b64 %rd4932, %rd4922, %rd4846; + xor.b64 %rd4933, %rd4914, %rd4849; + xor.b64 %rd4934, %rd4916, %rd4852; + xor.b64 %rd4935, %rd4918, %rd4855; + xor.b64 %rd4936, %rd4920, %rd4858; + xor.b64 %rd4937, %rd4922, %rd4861; + xor.b64 %rd4938, %rd4914, %rd4864; + xor.b64 %rd4939, %rd4916, %rd4867; + xor.b64 %rd4940, %rd4918, %rd4870; + xor.b64 %rd4941, %rd4920, %rd4873; + xor.b64 %rd4942, %rd4922, %rd4876; + xor.b64 %rd4943, %rd4914, %rd4879; + xor.b64 %rd4944, %rd4916, %rd4882; + xor.b64 %rd4945, %rd4918, %rd4885; + xor.b64 %rd4946, %rd4920, %rd4888; + xor.b64 %rd4947, %rd4922, %rd4891; + { + .reg .b32 %dummy; + mov.b64 {%r3346,%dummy}, %rd4929; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3347}, %rd4929; + } + shf.r.wrap.b32 %r3348, %r3347, %r3346, 20; + shf.r.wrap.b32 %r3349, %r3346, %r3347, 20; + mov.b64 %rd4948, {%r3349, %r3348}; + { + .reg .b32 %dummy; + mov.b64 {%r3350,%dummy}, %rd4935; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3351}, %rd4935; + } + shf.r.wrap.b32 %r3352, %r3351, %r3350, 21; + shf.r.wrap.b32 %r3353, %r3350, %r3351, 21; + mov.b64 %rd4949, {%r3353, %r3352}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3354}, %rd4941; + } + { + .reg .b32 %dummy; + mov.b64 {%r3355,%dummy}, %rd4941; + } + shf.l.wrap.b32 %r3356, %r3355, %r3354, 21; + shf.l.wrap.b32 %r3357, %r3354, %r3355, 21; + mov.b64 %rd4950, {%r3357, %r3356}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3358}, %rd4947; + } + { + .reg .b32 %dummy; + mov.b64 {%r3359,%dummy}, %rd4947; + } + shf.l.wrap.b32 %r3360, %r3359, %r3358, 14; + shf.l.wrap.b32 %r3361, %r3358, %r3359, 14; + mov.b64 %rd4951, {%r3361, %r3360}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3362}, %rd4926; + } + { + .reg .b32 %dummy; + mov.b64 {%r3363,%dummy}, %rd4926; + } + shf.l.wrap.b32 %r3364, %r3363, %r3362, 28; + shf.l.wrap.b32 %r3365, %r3362, %r3363, 28; + mov.b64 %rd4952, {%r3365, %r3364}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3366}, %rd4932; + } + { + .reg .b32 %dummy; + mov.b64 {%r3367,%dummy}, %rd4932; + } + shf.l.wrap.b32 %r3368, %r3367, %r3366, 20; + shf.l.wrap.b32 %r3369, %r3366, %r3367, 20; + mov.b64 %rd4953, {%r3369, %r3368}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3370}, %rd4933; + } + { + .reg .b32 %dummy; + mov.b64 {%r3371,%dummy}, %rd4933; + } + shf.l.wrap.b32 %r3372, %r3371, %r3370, 3; + shf.l.wrap.b32 %r3373, %r3370, %r3371, 3; + mov.b64 %rd4954, {%r3373, %r3372}; + { + .reg .b32 %dummy; + mov.b64 {%r3374,%dummy}, %rd4939; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3375}, %rd4939; + } + shf.r.wrap.b32 %r3376, %r3375, %r3374, 19; + shf.r.wrap.b32 %r3377, %r3374, %r3375, 19; + mov.b64 %rd4955, {%r3377, %r3376}; + { + .reg .b32 %dummy; + mov.b64 {%r3378,%dummy}, %rd4945; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3379}, %rd4945; + } + shf.r.wrap.b32 %r3380, %r3379, %r3378, 3; + shf.r.wrap.b32 %r3381, %r3378, %r3379, 3; + mov.b64 %rd4956, {%r3381, %r3380}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3382}, %rd4924; + } + { + .reg .b32 %dummy; + mov.b64 {%r3383,%dummy}, %rd4924; + } + shf.l.wrap.b32 %r3384, %r3383, %r3382, 1; + shf.l.wrap.b32 %r3385, %r3382, %r3383, 1; + mov.b64 %rd4957, {%r3385, %r3384}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3386}, %rd4930; + } + { + .reg .b32 %dummy; + mov.b64 {%r3387,%dummy}, %rd4930; + } + shf.l.wrap.b32 %r3388, %r3387, %r3386, 6; + shf.l.wrap.b32 %r3389, %r3386, %r3387, 6; + mov.b64 %rd4958, {%r3389, %r3388}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3390}, %rd4936; + } + { + .reg .b32 %dummy; + mov.b64 {%r3391,%dummy}, %rd4936; + } + shf.l.wrap.b32 %r3392, %r3391, %r3390, 25; + shf.l.wrap.b32 %r3393, %r3390, %r3391, 25; + mov.b64 %rd4959, {%r3393, %r3392}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3394}, %rd4942; + } + { + .reg .b32 %dummy; + mov.b64 {%r3395,%dummy}, %rd4942; + } + shf.l.wrap.b32 %r3396, %r3395, %r3394, 8; + shf.l.wrap.b32 %r3397, %r3394, %r3395, 8; + mov.b64 %rd4960, {%r3397, %r3396}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3398}, %rd4943; + } + { + .reg .b32 %dummy; + mov.b64 {%r3399,%dummy}, %rd4943; + } + shf.l.wrap.b32 %r3400, %r3399, %r3398, 18; + shf.l.wrap.b32 %r3401, %r3398, %r3399, 18; + mov.b64 %rd4961, {%r3401, %r3400}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3402}, %rd4927; + } + { + .reg .b32 %dummy; + mov.b64 {%r3403,%dummy}, %rd4927; + } + shf.l.wrap.b32 %r3404, %r3403, %r3402, 27; + shf.l.wrap.b32 %r3405, %r3402, %r3403, 27; + mov.b64 %rd4962, {%r3405, %r3404}; + { + .reg .b32 %dummy; + mov.b64 {%r3406,%dummy}, %rd4928; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3407}, %rd4928; + } + shf.r.wrap.b32 %r3408, %r3407, %r3406, 28; + shf.r.wrap.b32 %r3409, %r3406, %r3407, 28; + mov.b64 %rd4963, {%r3409, %r3408}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3410}, %rd4934; + } + { + .reg .b32 %dummy; + mov.b64 {%r3411,%dummy}, %rd4934; + } + shf.l.wrap.b32 %r3412, %r3411, %r3410, 10; + shf.l.wrap.b32 %r3413, %r3410, %r3411, 10; + mov.b64 %rd4964, {%r3413, %r3412}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3414}, %rd4940; + } + { + .reg .b32 %dummy; + mov.b64 {%r3415,%dummy}, %rd4940; + } + shf.l.wrap.b32 %r3416, %r3415, %r3414, 15; + shf.l.wrap.b32 %r3417, %r3414, %r3415, 15; + mov.b64 %rd4965, {%r3417, %r3416}; + { + .reg .b32 %dummy; + mov.b64 {%r3418,%dummy}, %rd4946; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3419}, %rd4946; + } + shf.r.wrap.b32 %r3420, %r3419, %r3418, 8; + shf.r.wrap.b32 %r3421, %r3418, %r3419, 8; + mov.b64 %rd4966, {%r3421, %r3420}; + { + .reg .b32 %dummy; + mov.b64 {%r3422,%dummy}, %rd4925; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3423}, %rd4925; + } + shf.r.wrap.b32 %r3424, %r3423, %r3422, 2; + shf.r.wrap.b32 %r3425, %r3422, %r3423, 2; + mov.b64 %rd4967, {%r3425, %r3424}; + { + .reg .b32 %dummy; + mov.b64 {%r3426,%dummy}, %rd4931; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3427}, %rd4931; + } + shf.r.wrap.b32 %r3428, %r3427, %r3426, 9; + shf.r.wrap.b32 %r3429, %r3426, %r3427, 9; + mov.b64 %rd4968, {%r3429, %r3428}; + { + .reg .b32 %dummy; + mov.b64 {%r3430,%dummy}, %rd4937; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3431}, %rd4937; + } + shf.r.wrap.b32 %r3432, %r3431, %r3430, 25; + shf.r.wrap.b32 %r3433, %r3430, %r3431, 25; + mov.b64 %rd4969, {%r3433, %r3432}; + { + .reg .b32 %dummy; + mov.b64 {%r3434,%dummy}, %rd4938; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3435}, %rd4938; + } + shf.r.wrap.b32 %r3436, %r3435, %r3434, 23; + shf.r.wrap.b32 %r3437, %r3434, %r3435, 23; + mov.b64 %rd4970, {%r3437, %r3436}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3438}, %rd4944; + } + { + .reg .b32 %dummy; + mov.b64 {%r3439,%dummy}, %rd4944; + } + shf.l.wrap.b32 %r3440, %r3439, %r3438, 2; + shf.l.wrap.b32 %r3441, %r3438, %r3439, 2; + mov.b64 %rd4971, {%r3441, %r3440}; + not.b64 %rd4972, %rd4948; + and.b64 %rd4973, %rd4949, %rd4972; + xor.b64 %rd4974, %rd4973, %rd4923; + not.b64 %rd4975, %rd4949; + and.b64 %rd4976, %rd4950, %rd4975; + xor.b64 %rd4977, %rd4976, %rd4948; + not.b64 %rd4978, %rd4950; + and.b64 %rd4979, %rd4951, %rd4978; + xor.b64 %rd4980, %rd4979, %rd4949; + not.b64 %rd4981, %rd4951; + and.b64 %rd4982, %rd4923, %rd4981; + xor.b64 %rd4983, %rd4982, %rd4950; + not.b64 %rd4984, %rd4923; + and.b64 %rd4985, %rd4948, %rd4984; + xor.b64 %rd4986, %rd4985, %rd4951; + not.b64 %rd4987, %rd4953; + and.b64 %rd4988, %rd4954, %rd4987; + xor.b64 %rd4989, %rd4988, %rd4952; + not.b64 %rd4990, %rd4954; + and.b64 %rd4991, %rd4955, %rd4990; + xor.b64 %rd4992, %rd4991, %rd4953; + not.b64 %rd4993, %rd4955; + and.b64 %rd4994, %rd4956, %rd4993; + xor.b64 %rd4995, %rd4994, %rd4954; + not.b64 %rd4996, %rd4956; + and.b64 %rd4997, %rd4952, %rd4996; + xor.b64 %rd4998, %rd4997, %rd4955; + not.b64 %rd4999, %rd4952; + and.b64 %rd5000, %rd4953, %rd4999; + xor.b64 %rd5001, %rd5000, %rd4956; + not.b64 %rd5002, %rd4958; + and.b64 %rd5003, %rd4959, %rd5002; + xor.b64 %rd5004, %rd5003, %rd4957; + not.b64 %rd5005, %rd4959; + and.b64 %rd5006, %rd4960, %rd5005; + xor.b64 %rd5007, %rd5006, %rd4958; + not.b64 %rd5008, %rd4960; + and.b64 %rd5009, %rd4961, %rd5008; + xor.b64 %rd5010, %rd5009, %rd4959; + not.b64 %rd5011, %rd4961; + and.b64 %rd5012, %rd4957, %rd5011; + xor.b64 %rd5013, %rd5012, %rd4960; + not.b64 %rd5014, %rd4957; + and.b64 %rd5015, %rd4958, %rd5014; + xor.b64 %rd5016, %rd5015, %rd4961; + not.b64 %rd5017, %rd4963; + and.b64 %rd5018, %rd4964, %rd5017; + xor.b64 %rd5019, %rd5018, %rd4962; + not.b64 %rd5020, %rd4964; + and.b64 %rd5021, %rd4965, %rd5020; + xor.b64 %rd5022, %rd5021, %rd4963; + not.b64 %rd5023, %rd4965; + and.b64 %rd5024, %rd4966, %rd5023; + xor.b64 %rd5025, %rd5024, %rd4964; + not.b64 %rd5026, %rd4966; + and.b64 %rd5027, %rd4962, %rd5026; + xor.b64 %rd5028, %rd5027, %rd4965; + not.b64 %rd5029, %rd4962; + and.b64 %rd5030, %rd4963, %rd5029; + xor.b64 %rd5031, %rd5030, %rd4966; + not.b64 %rd5032, %rd4968; + and.b64 %rd5033, %rd4969, %rd5032; + xor.b64 %rd5034, %rd5033, %rd4967; + not.b64 %rd5035, %rd4969; + and.b64 %rd5036, %rd4970, %rd5035; + xor.b64 %rd5037, %rd5036, %rd4968; + not.b64 %rd5038, %rd4970; + and.b64 %rd5039, %rd4971, %rd5038; + xor.b64 %rd5040, %rd5039, %rd4969; + not.b64 %rd5041, %rd4971; + and.b64 %rd5042, %rd4967, %rd5041; + xor.b64 %rd5043, %rd5042, %rd4970; + not.b64 %rd5044, %rd4967; + and.b64 %rd5045, %rd4968, %rd5044; + xor.b64 %rd5046, %rd5045, %rd4971; + xor.b64 %rd5047, %rd4974, %rd20; + xor.b64 %rd5048, %rd4989, %rd5047; + xor.b64 %rd5049, %rd5048, %rd5004; + xor.b64 %rd5050, %rd5049, %rd5019; + xor.b64 %rd5051, %rd5050, %rd5034; + xor.b64 %rd5052, %rd4992, %rd4977; + xor.b64 %rd5053, %rd5052, %rd5007; + xor.b64 %rd5054, %rd5053, %rd5022; + xor.b64 %rd5055, %rd5054, %rd5037; + xor.b64 %rd5056, %rd4995, %rd4980; + xor.b64 %rd5057, %rd5056, %rd5010; + xor.b64 %rd5058, %rd5057, %rd5025; + xor.b64 %rd5059, %rd5058, %rd5040; + xor.b64 %rd5060, %rd4998, %rd4983; + xor.b64 %rd5061, %rd5060, %rd5013; + xor.b64 %rd5062, %rd5061, %rd5028; + xor.b64 %rd5063, %rd5062, %rd5043; + xor.b64 %rd5064, %rd5001, %rd4986; + xor.b64 %rd5065, %rd5064, %rd5016; + xor.b64 %rd5066, %rd5065, %rd5031; + xor.b64 %rd5067, %rd5066, %rd5046; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3442}, %rd5055; + } + { + .reg .b32 %dummy; + mov.b64 {%r3443,%dummy}, %rd5055; + } + shf.l.wrap.b32 %r3444, %r3443, %r3442, 1; + shf.l.wrap.b32 %r3445, %r3442, %r3443, 1; + mov.b64 %rd5068, {%r3445, %r3444}; + xor.b64 %rd5069, %rd5067, %rd5068; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3446}, %rd5059; + } + { + .reg .b32 %dummy; + mov.b64 {%r3447,%dummy}, %rd5059; + } + shf.l.wrap.b32 %r3448, %r3447, %r3446, 1; + shf.l.wrap.b32 %r3449, %r3446, %r3447, 1; + mov.b64 %rd5070, {%r3449, %r3448}; + xor.b64 %rd5071, %rd5070, %rd5051; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3450}, %rd5063; + } + { + .reg .b32 %dummy; + mov.b64 {%r3451,%dummy}, %rd5063; + } + shf.l.wrap.b32 %r3452, %r3451, %r3450, 1; + shf.l.wrap.b32 %r3453, %r3450, %r3451, 1; + mov.b64 %rd5072, {%r3453, %r3452}; + xor.b64 %rd5073, %rd5072, %rd5055; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3454}, %rd5067; + } + { + .reg .b32 %dummy; + mov.b64 {%r3455,%dummy}, %rd5067; + } + shf.l.wrap.b32 %r3456, %r3455, %r3454, 1; + shf.l.wrap.b32 %r3457, %r3454, %r3455, 1; + mov.b64 %rd5074, {%r3457, %r3456}; + xor.b64 %rd5075, %rd5074, %rd5059; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3458}, %rd5051; + } + { + .reg .b32 %dummy; + mov.b64 {%r3459,%dummy}, %rd5051; + } + shf.l.wrap.b32 %r3460, %r3459, %r3458, 1; + shf.l.wrap.b32 %r3461, %r3458, %r3459, 1; + mov.b64 %rd5076, {%r3461, %r3460}; + xor.b64 %rd5077, %rd5063, %rd5076; + xor.b64 %rd5078, %rd5069, %rd5047; + xor.b64 %rd5079, %rd5071, %rd4977; + xor.b64 %rd5080, %rd5073, %rd4980; + xor.b64 %rd5081, %rd5075, %rd4983; + xor.b64 %rd5082, %rd5077, %rd4986; + xor.b64 %rd5083, %rd5069, %rd4989; + xor.b64 %rd5084, %rd5071, %rd4992; + xor.b64 %rd5085, %rd5073, %rd4995; + xor.b64 %rd5086, %rd5075, %rd4998; + xor.b64 %rd5087, %rd5077, %rd5001; + xor.b64 %rd5088, %rd5069, %rd5004; + xor.b64 %rd5089, %rd5071, %rd5007; + xor.b64 %rd5090, %rd5073, %rd5010; + xor.b64 %rd5091, %rd5075, %rd5013; + xor.b64 %rd5092, %rd5077, %rd5016; + xor.b64 %rd5093, %rd5069, %rd5019; + xor.b64 %rd5094, %rd5071, %rd5022; + xor.b64 %rd5095, %rd5073, %rd5025; + xor.b64 %rd5096, %rd5075, %rd5028; + xor.b64 %rd5097, %rd5077, %rd5031; + xor.b64 %rd5098, %rd5069, %rd5034; + xor.b64 %rd5099, %rd5071, %rd5037; + xor.b64 %rd5100, %rd5073, %rd5040; + xor.b64 %rd5101, %rd5075, %rd5043; + xor.b64 %rd5102, %rd5077, %rd5046; + { + .reg .b32 %dummy; + mov.b64 {%r3462,%dummy}, %rd5084; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3463}, %rd5084; + } + shf.r.wrap.b32 %r3464, %r3463, %r3462, 20; + shf.r.wrap.b32 %r3465, %r3462, %r3463, 20; + mov.b64 %rd5103, {%r3465, %r3464}; + { + .reg .b32 %dummy; + mov.b64 {%r3466,%dummy}, %rd5090; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3467}, %rd5090; + } + shf.r.wrap.b32 %r3468, %r3467, %r3466, 21; + shf.r.wrap.b32 %r3469, %r3466, %r3467, 21; + mov.b64 %rd5104, {%r3469, %r3468}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3470}, %rd5096; + } + { + .reg .b32 %dummy; + mov.b64 {%r3471,%dummy}, %rd5096; + } + shf.l.wrap.b32 %r3472, %r3471, %r3470, 21; + shf.l.wrap.b32 %r3473, %r3470, %r3471, 21; + mov.b64 %rd5105, {%r3473, %r3472}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3474}, %rd5102; + } + { + .reg .b32 %dummy; + mov.b64 {%r3475,%dummy}, %rd5102; + } + shf.l.wrap.b32 %r3476, %r3475, %r3474, 14; + shf.l.wrap.b32 %r3477, %r3474, %r3475, 14; + mov.b64 %rd5106, {%r3477, %r3476}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3478}, %rd5081; + } + { + .reg .b32 %dummy; + mov.b64 {%r3479,%dummy}, %rd5081; + } + shf.l.wrap.b32 %r3480, %r3479, %r3478, 28; + shf.l.wrap.b32 %r3481, %r3478, %r3479, 28; + mov.b64 %rd5107, {%r3481, %r3480}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3482}, %rd5087; + } + { + .reg .b32 %dummy; + mov.b64 {%r3483,%dummy}, %rd5087; + } + shf.l.wrap.b32 %r3484, %r3483, %r3482, 20; + shf.l.wrap.b32 %r3485, %r3482, %r3483, 20; + mov.b64 %rd5108, {%r3485, %r3484}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3486}, %rd5088; + } + { + .reg .b32 %dummy; + mov.b64 {%r3487,%dummy}, %rd5088; + } + shf.l.wrap.b32 %r3488, %r3487, %r3486, 3; + shf.l.wrap.b32 %r3489, %r3486, %r3487, 3; + mov.b64 %rd5109, {%r3489, %r3488}; + { + .reg .b32 %dummy; + mov.b64 {%r3490,%dummy}, %rd5094; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3491}, %rd5094; + } + shf.r.wrap.b32 %r3492, %r3491, %r3490, 19; + shf.r.wrap.b32 %r3493, %r3490, %r3491, 19; + mov.b64 %rd5110, {%r3493, %r3492}; + { + .reg .b32 %dummy; + mov.b64 {%r3494,%dummy}, %rd5100; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3495}, %rd5100; + } + shf.r.wrap.b32 %r3496, %r3495, %r3494, 3; + shf.r.wrap.b32 %r3497, %r3494, %r3495, 3; + mov.b64 %rd5111, {%r3497, %r3496}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3498}, %rd5079; + } + { + .reg .b32 %dummy; + mov.b64 {%r3499,%dummy}, %rd5079; + } + shf.l.wrap.b32 %r3500, %r3499, %r3498, 1; + shf.l.wrap.b32 %r3501, %r3498, %r3499, 1; + mov.b64 %rd5112, {%r3501, %r3500}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3502}, %rd5085; + } + { + .reg .b32 %dummy; + mov.b64 {%r3503,%dummy}, %rd5085; + } + shf.l.wrap.b32 %r3504, %r3503, %r3502, 6; + shf.l.wrap.b32 %r3505, %r3502, %r3503, 6; + mov.b64 %rd5113, {%r3505, %r3504}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3506}, %rd5091; + } + { + .reg .b32 %dummy; + mov.b64 {%r3507,%dummy}, %rd5091; + } + shf.l.wrap.b32 %r3508, %r3507, %r3506, 25; + shf.l.wrap.b32 %r3509, %r3506, %r3507, 25; + mov.b64 %rd5114, {%r3509, %r3508}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3510}, %rd5097; + } + { + .reg .b32 %dummy; + mov.b64 {%r3511,%dummy}, %rd5097; + } + shf.l.wrap.b32 %r3512, %r3511, %r3510, 8; + shf.l.wrap.b32 %r3513, %r3510, %r3511, 8; + mov.b64 %rd5115, {%r3513, %r3512}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3514}, %rd5098; + } + { + .reg .b32 %dummy; + mov.b64 {%r3515,%dummy}, %rd5098; + } + shf.l.wrap.b32 %r3516, %r3515, %r3514, 18; + shf.l.wrap.b32 %r3517, %r3514, %r3515, 18; + mov.b64 %rd5116, {%r3517, %r3516}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3518}, %rd5082; + } + { + .reg .b32 %dummy; + mov.b64 {%r3519,%dummy}, %rd5082; + } + shf.l.wrap.b32 %r3520, %r3519, %r3518, 27; + shf.l.wrap.b32 %r3521, %r3518, %r3519, 27; + mov.b64 %rd5117, {%r3521, %r3520}; + { + .reg .b32 %dummy; + mov.b64 {%r3522,%dummy}, %rd5083; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3523}, %rd5083; + } + shf.r.wrap.b32 %r3524, %r3523, %r3522, 28; + shf.r.wrap.b32 %r3525, %r3522, %r3523, 28; + mov.b64 %rd5118, {%r3525, %r3524}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3526}, %rd5089; + } + { + .reg .b32 %dummy; + mov.b64 {%r3527,%dummy}, %rd5089; + } + shf.l.wrap.b32 %r3528, %r3527, %r3526, 10; + shf.l.wrap.b32 %r3529, %r3526, %r3527, 10; + mov.b64 %rd5119, {%r3529, %r3528}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3530}, %rd5095; + } + { + .reg .b32 %dummy; + mov.b64 {%r3531,%dummy}, %rd5095; + } + shf.l.wrap.b32 %r3532, %r3531, %r3530, 15; + shf.l.wrap.b32 %r3533, %r3530, %r3531, 15; + mov.b64 %rd5120, {%r3533, %r3532}; + { + .reg .b32 %dummy; + mov.b64 {%r3534,%dummy}, %rd5101; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3535}, %rd5101; + } + shf.r.wrap.b32 %r3536, %r3535, %r3534, 8; + shf.r.wrap.b32 %r3537, %r3534, %r3535, 8; + mov.b64 %rd5121, {%r3537, %r3536}; + { + .reg .b32 %dummy; + mov.b64 {%r3538,%dummy}, %rd5080; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3539}, %rd5080; + } + shf.r.wrap.b32 %r3540, %r3539, %r3538, 2; + shf.r.wrap.b32 %r3541, %r3538, %r3539, 2; + mov.b64 %rd5122, {%r3541, %r3540}; + { + .reg .b32 %dummy; + mov.b64 {%r3542,%dummy}, %rd5086; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3543}, %rd5086; + } + shf.r.wrap.b32 %r3544, %r3543, %r3542, 9; + shf.r.wrap.b32 %r3545, %r3542, %r3543, 9; + mov.b64 %rd5123, {%r3545, %r3544}; + { + .reg .b32 %dummy; + mov.b64 {%r3546,%dummy}, %rd5092; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3547}, %rd5092; + } + shf.r.wrap.b32 %r3548, %r3547, %r3546, 25; + shf.r.wrap.b32 %r3549, %r3546, %r3547, 25; + mov.b64 %rd5124, {%r3549, %r3548}; + { + .reg .b32 %dummy; + mov.b64 {%r3550,%dummy}, %rd5093; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3551}, %rd5093; + } + shf.r.wrap.b32 %r3552, %r3551, %r3550, 23; + shf.r.wrap.b32 %r3553, %r3550, %r3551, 23; + mov.b64 %rd5125, {%r3553, %r3552}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3554}, %rd5099; + } + { + .reg .b32 %dummy; + mov.b64 {%r3555,%dummy}, %rd5099; + } + shf.l.wrap.b32 %r3556, %r3555, %r3554, 2; + shf.l.wrap.b32 %r3557, %r3554, %r3555, 2; + mov.b64 %rd5126, {%r3557, %r3556}; + not.b64 %rd5127, %rd5103; + and.b64 %rd5128, %rd5104, %rd5127; + xor.b64 %rd5129, %rd5128, %rd5078; + not.b64 %rd5130, %rd5104; + and.b64 %rd5131, %rd5105, %rd5130; + xor.b64 %rd5132, %rd5131, %rd5103; + not.b64 %rd5133, %rd5105; + and.b64 %rd5134, %rd5106, %rd5133; + xor.b64 %rd5135, %rd5134, %rd5104; + not.b64 %rd5136, %rd5106; + and.b64 %rd5137, %rd5078, %rd5136; + xor.b64 %rd5138, %rd5137, %rd5105; + not.b64 %rd5139, %rd5078; + and.b64 %rd5140, %rd5103, %rd5139; + xor.b64 %rd5141, %rd5140, %rd5106; + not.b64 %rd5142, %rd5108; + and.b64 %rd5143, %rd5109, %rd5142; + xor.b64 %rd5144, %rd5143, %rd5107; + not.b64 %rd5145, %rd5109; + and.b64 %rd5146, %rd5110, %rd5145; + xor.b64 %rd5147, %rd5146, %rd5108; + not.b64 %rd5148, %rd5110; + and.b64 %rd5149, %rd5111, %rd5148; + xor.b64 %rd5150, %rd5149, %rd5109; + not.b64 %rd5151, %rd5111; + and.b64 %rd5152, %rd5107, %rd5151; + xor.b64 %rd5153, %rd5152, %rd5110; + not.b64 %rd5154, %rd5107; + and.b64 %rd5155, %rd5108, %rd5154; + xor.b64 %rd5156, %rd5155, %rd5111; + not.b64 %rd5157, %rd5113; + and.b64 %rd5158, %rd5114, %rd5157; + xor.b64 %rd5159, %rd5158, %rd5112; + not.b64 %rd5160, %rd5114; + and.b64 %rd5161, %rd5115, %rd5160; + xor.b64 %rd5162, %rd5161, %rd5113; + not.b64 %rd5163, %rd5115; + and.b64 %rd5164, %rd5116, %rd5163; + xor.b64 %rd5165, %rd5164, %rd5114; + not.b64 %rd5166, %rd5116; + and.b64 %rd5167, %rd5112, %rd5166; + xor.b64 %rd5168, %rd5167, %rd5115; + not.b64 %rd5169, %rd5112; + and.b64 %rd5170, %rd5113, %rd5169; + xor.b64 %rd5171, %rd5170, %rd5116; + not.b64 %rd5172, %rd5118; + and.b64 %rd5173, %rd5119, %rd5172; + xor.b64 %rd5174, %rd5173, %rd5117; + not.b64 %rd5175, %rd5119; + and.b64 %rd5176, %rd5120, %rd5175; + xor.b64 %rd5177, %rd5176, %rd5118; + not.b64 %rd5178, %rd5120; + and.b64 %rd5179, %rd5121, %rd5178; + xor.b64 %rd5180, %rd5179, %rd5119; + not.b64 %rd5181, %rd5121; + and.b64 %rd5182, %rd5117, %rd5181; + xor.b64 %rd5183, %rd5182, %rd5120; + not.b64 %rd5184, %rd5117; + and.b64 %rd5185, %rd5118, %rd5184; + xor.b64 %rd5186, %rd5185, %rd5121; + not.b64 %rd5187, %rd5123; + and.b64 %rd5188, %rd5124, %rd5187; + xor.b64 %rd5189, %rd5188, %rd5122; + not.b64 %rd5190, %rd5124; + and.b64 %rd5191, %rd5125, %rd5190; + xor.b64 %rd5192, %rd5191, %rd5123; + not.b64 %rd5193, %rd5125; + and.b64 %rd5194, %rd5126, %rd5193; + xor.b64 %rd5195, %rd5194, %rd5124; + not.b64 %rd5196, %rd5126; + and.b64 %rd5197, %rd5122, %rd5196; + xor.b64 %rd5198, %rd5197, %rd5125; + not.b64 %rd5199, %rd5122; + and.b64 %rd5200, %rd5123, %rd5199; + xor.b64 %rd5201, %rd5200, %rd5126; + xor.b64 %rd5202, %rd5129, %rd21; + xor.b64 %rd5203, %rd5144, %rd5202; + xor.b64 %rd5204, %rd5203, %rd5159; + xor.b64 %rd5205, %rd5204, %rd5174; + xor.b64 %rd5206, %rd5205, %rd5189; + xor.b64 %rd5207, %rd5147, %rd5132; + xor.b64 %rd5208, %rd5207, %rd5162; + xor.b64 %rd5209, %rd5208, %rd5177; + xor.b64 %rd5210, %rd5209, %rd5192; + xor.b64 %rd5211, %rd5150, %rd5135; + xor.b64 %rd5212, %rd5211, %rd5165; + xor.b64 %rd5213, %rd5212, %rd5180; + xor.b64 %rd5214, %rd5213, %rd5195; + xor.b64 %rd5215, %rd5153, %rd5138; + xor.b64 %rd5216, %rd5215, %rd5168; + xor.b64 %rd5217, %rd5216, %rd5183; + xor.b64 %rd5218, %rd5217, %rd5198; + xor.b64 %rd5219, %rd5156, %rd5141; + xor.b64 %rd5220, %rd5219, %rd5171; + xor.b64 %rd5221, %rd5220, %rd5186; + xor.b64 %rd5222, %rd5221, %rd5201; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3558}, %rd5210; + } + { + .reg .b32 %dummy; + mov.b64 {%r3559,%dummy}, %rd5210; + } + shf.l.wrap.b32 %r3560, %r3559, %r3558, 1; + shf.l.wrap.b32 %r3561, %r3558, %r3559, 1; + mov.b64 %rd5223, {%r3561, %r3560}; + xor.b64 %rd5224, %rd5222, %rd5223; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3562}, %rd5214; + } + { + .reg .b32 %dummy; + mov.b64 {%r3563,%dummy}, %rd5214; + } + shf.l.wrap.b32 %r3564, %r3563, %r3562, 1; + shf.l.wrap.b32 %r3565, %r3562, %r3563, 1; + mov.b64 %rd5225, {%r3565, %r3564}; + xor.b64 %rd5226, %rd5225, %rd5206; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3566}, %rd5218; + } + { + .reg .b32 %dummy; + mov.b64 {%r3567,%dummy}, %rd5218; + } + shf.l.wrap.b32 %r3568, %r3567, %r3566, 1; + shf.l.wrap.b32 %r3569, %r3566, %r3567, 1; + mov.b64 %rd5227, {%r3569, %r3568}; + xor.b64 %rd5228, %rd5227, %rd5210; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3570}, %rd5222; + } + { + .reg .b32 %dummy; + mov.b64 {%r3571,%dummy}, %rd5222; + } + shf.l.wrap.b32 %r3572, %r3571, %r3570, 1; + shf.l.wrap.b32 %r3573, %r3570, %r3571, 1; + mov.b64 %rd5229, {%r3573, %r3572}; + xor.b64 %rd5230, %rd5229, %rd5214; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3574}, %rd5206; + } + { + .reg .b32 %dummy; + mov.b64 {%r3575,%dummy}, %rd5206; + } + shf.l.wrap.b32 %r3576, %r3575, %r3574, 1; + shf.l.wrap.b32 %r3577, %r3574, %r3575, 1; + mov.b64 %rd5231, {%r3577, %r3576}; + xor.b64 %rd5232, %rd5218, %rd5231; + xor.b64 %rd5233, %rd5224, %rd5202; + xor.b64 %rd5234, %rd5226, %rd5132; + xor.b64 %rd5235, %rd5228, %rd5135; + xor.b64 %rd5236, %rd5230, %rd5138; + xor.b64 %rd5237, %rd5232, %rd5141; + xor.b64 %rd5238, %rd5224, %rd5144; + xor.b64 %rd5239, %rd5226, %rd5147; + xor.b64 %rd5240, %rd5228, %rd5150; + xor.b64 %rd5241, %rd5230, %rd5153; + xor.b64 %rd5242, %rd5232, %rd5156; + xor.b64 %rd5243, %rd5224, %rd5159; + xor.b64 %rd5244, %rd5226, %rd5162; + xor.b64 %rd5245, %rd5228, %rd5165; + xor.b64 %rd5246, %rd5230, %rd5168; + xor.b64 %rd5247, %rd5232, %rd5171; + xor.b64 %rd5248, %rd5224, %rd5174; + xor.b64 %rd5249, %rd5226, %rd5177; + xor.b64 %rd5250, %rd5228, %rd5180; + xor.b64 %rd5251, %rd5230, %rd5183; + xor.b64 %rd5252, %rd5232, %rd5186; + xor.b64 %rd5253, %rd5224, %rd5189; + xor.b64 %rd5254, %rd5226, %rd5192; + xor.b64 %rd5255, %rd5228, %rd5195; + xor.b64 %rd5256, %rd5230, %rd5198; + xor.b64 %rd5257, %rd5232, %rd5201; + { + .reg .b32 %dummy; + mov.b64 {%r3578,%dummy}, %rd5239; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3579}, %rd5239; + } + shf.r.wrap.b32 %r3580, %r3579, %r3578, 20; + shf.r.wrap.b32 %r3581, %r3578, %r3579, 20; + mov.b64 %rd5258, {%r3581, %r3580}; + { + .reg .b32 %dummy; + mov.b64 {%r3582,%dummy}, %rd5245; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3583}, %rd5245; + } + shf.r.wrap.b32 %r3584, %r3583, %r3582, 21; + shf.r.wrap.b32 %r3585, %r3582, %r3583, 21; + mov.b64 %rd5259, {%r3585, %r3584}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3586}, %rd5251; + } + { + .reg .b32 %dummy; + mov.b64 {%r3587,%dummy}, %rd5251; + } + shf.l.wrap.b32 %r3588, %r3587, %r3586, 21; + shf.l.wrap.b32 %r3589, %r3586, %r3587, 21; + mov.b64 %rd5260, {%r3589, %r3588}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3590}, %rd5257; + } + { + .reg .b32 %dummy; + mov.b64 {%r3591,%dummy}, %rd5257; + } + shf.l.wrap.b32 %r3592, %r3591, %r3590, 14; + shf.l.wrap.b32 %r3593, %r3590, %r3591, 14; + mov.b64 %rd5261, {%r3593, %r3592}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3594}, %rd5236; + } + { + .reg .b32 %dummy; + mov.b64 {%r3595,%dummy}, %rd5236; + } + shf.l.wrap.b32 %r3596, %r3595, %r3594, 28; + shf.l.wrap.b32 %r3597, %r3594, %r3595, 28; + mov.b64 %rd5262, {%r3597, %r3596}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3598}, %rd5242; + } + { + .reg .b32 %dummy; + mov.b64 {%r3599,%dummy}, %rd5242; + } + shf.l.wrap.b32 %r3600, %r3599, %r3598, 20; + shf.l.wrap.b32 %r3601, %r3598, %r3599, 20; + mov.b64 %rd5263, {%r3601, %r3600}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3602}, %rd5243; + } + { + .reg .b32 %dummy; + mov.b64 {%r3603,%dummy}, %rd5243; + } + shf.l.wrap.b32 %r3604, %r3603, %r3602, 3; + shf.l.wrap.b32 %r3605, %r3602, %r3603, 3; + mov.b64 %rd5264, {%r3605, %r3604}; + { + .reg .b32 %dummy; + mov.b64 {%r3606,%dummy}, %rd5249; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3607}, %rd5249; + } + shf.r.wrap.b32 %r3608, %r3607, %r3606, 19; + shf.r.wrap.b32 %r3609, %r3606, %r3607, 19; + mov.b64 %rd5265, {%r3609, %r3608}; + { + .reg .b32 %dummy; + mov.b64 {%r3610,%dummy}, %rd5255; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3611}, %rd5255; + } + shf.r.wrap.b32 %r3612, %r3611, %r3610, 3; + shf.r.wrap.b32 %r3613, %r3610, %r3611, 3; + mov.b64 %rd5266, {%r3613, %r3612}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3614}, %rd5234; + } + { + .reg .b32 %dummy; + mov.b64 {%r3615,%dummy}, %rd5234; + } + shf.l.wrap.b32 %r3616, %r3615, %r3614, 1; + shf.l.wrap.b32 %r3617, %r3614, %r3615, 1; + mov.b64 %rd5267, {%r3617, %r3616}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3618}, %rd5240; + } + { + .reg .b32 %dummy; + mov.b64 {%r3619,%dummy}, %rd5240; + } + shf.l.wrap.b32 %r3620, %r3619, %r3618, 6; + shf.l.wrap.b32 %r3621, %r3618, %r3619, 6; + mov.b64 %rd5268, {%r3621, %r3620}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3622}, %rd5246; + } + { + .reg .b32 %dummy; + mov.b64 {%r3623,%dummy}, %rd5246; + } + shf.l.wrap.b32 %r3624, %r3623, %r3622, 25; + shf.l.wrap.b32 %r3625, %r3622, %r3623, 25; + mov.b64 %rd5269, {%r3625, %r3624}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3626}, %rd5252; + } + { + .reg .b32 %dummy; + mov.b64 {%r3627,%dummy}, %rd5252; + } + shf.l.wrap.b32 %r3628, %r3627, %r3626, 8; + shf.l.wrap.b32 %r3629, %r3626, %r3627, 8; + mov.b64 %rd5270, {%r3629, %r3628}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3630}, %rd5253; + } + { + .reg .b32 %dummy; + mov.b64 {%r3631,%dummy}, %rd5253; + } + shf.l.wrap.b32 %r3632, %r3631, %r3630, 18; + shf.l.wrap.b32 %r3633, %r3630, %r3631, 18; + mov.b64 %rd5271, {%r3633, %r3632}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3634}, %rd5237; + } + { + .reg .b32 %dummy; + mov.b64 {%r3635,%dummy}, %rd5237; + } + shf.l.wrap.b32 %r3636, %r3635, %r3634, 27; + shf.l.wrap.b32 %r3637, %r3634, %r3635, 27; + mov.b64 %rd5272, {%r3637, %r3636}; + { + .reg .b32 %dummy; + mov.b64 {%r3638,%dummy}, %rd5238; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3639}, %rd5238; + } + shf.r.wrap.b32 %r3640, %r3639, %r3638, 28; + shf.r.wrap.b32 %r3641, %r3638, %r3639, 28; + mov.b64 %rd5273, {%r3641, %r3640}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3642}, %rd5244; + } + { + .reg .b32 %dummy; + mov.b64 {%r3643,%dummy}, %rd5244; + } + shf.l.wrap.b32 %r3644, %r3643, %r3642, 10; + shf.l.wrap.b32 %r3645, %r3642, %r3643, 10; + mov.b64 %rd5274, {%r3645, %r3644}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3646}, %rd5250; + } + { + .reg .b32 %dummy; + mov.b64 {%r3647,%dummy}, %rd5250; + } + shf.l.wrap.b32 %r3648, %r3647, %r3646, 15; + shf.l.wrap.b32 %r3649, %r3646, %r3647, 15; + mov.b64 %rd5275, {%r3649, %r3648}; + { + .reg .b32 %dummy; + mov.b64 {%r3650,%dummy}, %rd5256; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3651}, %rd5256; + } + shf.r.wrap.b32 %r3652, %r3651, %r3650, 8; + shf.r.wrap.b32 %r3653, %r3650, %r3651, 8; + mov.b64 %rd5276, {%r3653, %r3652}; + { + .reg .b32 %dummy; + mov.b64 {%r3654,%dummy}, %rd5235; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3655}, %rd5235; + } + shf.r.wrap.b32 %r3656, %r3655, %r3654, 2; + shf.r.wrap.b32 %r3657, %r3654, %r3655, 2; + mov.b64 %rd5277, {%r3657, %r3656}; + { + .reg .b32 %dummy; + mov.b64 {%r3658,%dummy}, %rd5241; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3659}, %rd5241; + } + shf.r.wrap.b32 %r3660, %r3659, %r3658, 9; + shf.r.wrap.b32 %r3661, %r3658, %r3659, 9; + mov.b64 %rd5278, {%r3661, %r3660}; + { + .reg .b32 %dummy; + mov.b64 {%r3662,%dummy}, %rd5247; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3663}, %rd5247; + } + shf.r.wrap.b32 %r3664, %r3663, %r3662, 25; + shf.r.wrap.b32 %r3665, %r3662, %r3663, 25; + mov.b64 %rd5279, {%r3665, %r3664}; + { + .reg .b32 %dummy; + mov.b64 {%r3666,%dummy}, %rd5248; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3667}, %rd5248; + } + shf.r.wrap.b32 %r3668, %r3667, %r3666, 23; + shf.r.wrap.b32 %r3669, %r3666, %r3667, 23; + mov.b64 %rd5280, {%r3669, %r3668}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3670}, %rd5254; + } + { + .reg .b32 %dummy; + mov.b64 {%r3671,%dummy}, %rd5254; + } + shf.l.wrap.b32 %r3672, %r3671, %r3670, 2; + shf.l.wrap.b32 %r3673, %r3670, %r3671, 2; + mov.b64 %rd5281, {%r3673, %r3672}; + not.b64 %rd5282, %rd5258; + and.b64 %rd5283, %rd5259, %rd5282; + xor.b64 %rd5284, %rd5283, %rd5233; + not.b64 %rd5285, %rd5259; + and.b64 %rd5286, %rd5260, %rd5285; + xor.b64 %rd5287, %rd5286, %rd5258; + not.b64 %rd5288, %rd5260; + and.b64 %rd5289, %rd5261, %rd5288; + xor.b64 %rd5290, %rd5289, %rd5259; + not.b64 %rd5291, %rd5261; + and.b64 %rd5292, %rd5233, %rd5291; + xor.b64 %rd5293, %rd5292, %rd5260; + not.b64 %rd5294, %rd5233; + and.b64 %rd5295, %rd5258, %rd5294; + xor.b64 %rd5296, %rd5295, %rd5261; + not.b64 %rd5297, %rd5263; + and.b64 %rd5298, %rd5264, %rd5297; + xor.b64 %rd5299, %rd5298, %rd5262; + not.b64 %rd5300, %rd5264; + and.b64 %rd5301, %rd5265, %rd5300; + xor.b64 %rd5302, %rd5301, %rd5263; + not.b64 %rd5303, %rd5265; + and.b64 %rd5304, %rd5266, %rd5303; + xor.b64 %rd5305, %rd5304, %rd5264; + not.b64 %rd5306, %rd5266; + and.b64 %rd5307, %rd5262, %rd5306; + xor.b64 %rd5308, %rd5307, %rd5265; + not.b64 %rd5309, %rd5262; + and.b64 %rd5310, %rd5263, %rd5309; + xor.b64 %rd5311, %rd5310, %rd5266; + not.b64 %rd5312, %rd5268; + and.b64 %rd5313, %rd5269, %rd5312; + xor.b64 %rd5314, %rd5313, %rd5267; + not.b64 %rd5315, %rd5269; + and.b64 %rd5316, %rd5270, %rd5315; + xor.b64 %rd5317, %rd5316, %rd5268; + not.b64 %rd5318, %rd5270; + and.b64 %rd5319, %rd5271, %rd5318; + xor.b64 %rd5320, %rd5319, %rd5269; + not.b64 %rd5321, %rd5271; + and.b64 %rd5322, %rd5267, %rd5321; + xor.b64 %rd5323, %rd5322, %rd5270; + not.b64 %rd5324, %rd5267; + and.b64 %rd5325, %rd5268, %rd5324; + xor.b64 %rd5326, %rd5325, %rd5271; + not.b64 %rd5327, %rd5273; + and.b64 %rd5328, %rd5274, %rd5327; + xor.b64 %rd5329, %rd5328, %rd5272; + not.b64 %rd5330, %rd5274; + and.b64 %rd5331, %rd5275, %rd5330; + xor.b64 %rd5332, %rd5331, %rd5273; + not.b64 %rd5333, %rd5275; + and.b64 %rd5334, %rd5276, %rd5333; + xor.b64 %rd5335, %rd5334, %rd5274; + not.b64 %rd5336, %rd5276; + and.b64 %rd5337, %rd5272, %rd5336; + xor.b64 %rd5338, %rd5337, %rd5275; + not.b64 %rd5339, %rd5272; + and.b64 %rd5340, %rd5273, %rd5339; + xor.b64 %rd5341, %rd5340, %rd5276; + not.b64 %rd5342, %rd5278; + and.b64 %rd5343, %rd5279, %rd5342; + xor.b64 %rd5344, %rd5343, %rd5277; + not.b64 %rd5345, %rd5279; + and.b64 %rd5346, %rd5280, %rd5345; + xor.b64 %rd5347, %rd5346, %rd5278; + not.b64 %rd5348, %rd5280; + and.b64 %rd5349, %rd5281, %rd5348; + xor.b64 %rd5350, %rd5349, %rd5279; + not.b64 %rd5351, %rd5281; + and.b64 %rd5352, %rd5277, %rd5351; + xor.b64 %rd5353, %rd5352, %rd5280; + not.b64 %rd5354, %rd5277; + and.b64 %rd5355, %rd5278, %rd5354; + xor.b64 %rd5356, %rd5355, %rd5281; + xor.b64 %rd5357, %rd5284, %rd22; + xor.b64 %rd5358, %rd5299, %rd5357; + xor.b64 %rd5359, %rd5358, %rd5314; + xor.b64 %rd5360, %rd5359, %rd5329; + xor.b64 %rd5361, %rd5360, %rd5344; + xor.b64 %rd5362, %rd5302, %rd5287; + xor.b64 %rd5363, %rd5362, %rd5317; + xor.b64 %rd5364, %rd5363, %rd5332; + xor.b64 %rd5365, %rd5364, %rd5347; + xor.b64 %rd5366, %rd5305, %rd5290; + xor.b64 %rd5367, %rd5366, %rd5320; + xor.b64 %rd5368, %rd5367, %rd5335; + xor.b64 %rd5369, %rd5368, %rd5350; + xor.b64 %rd5370, %rd5308, %rd5293; + xor.b64 %rd5371, %rd5370, %rd5323; + xor.b64 %rd5372, %rd5371, %rd5338; + xor.b64 %rd5373, %rd5372, %rd5353; + xor.b64 %rd5374, %rd5311, %rd5296; + xor.b64 %rd5375, %rd5374, %rd5326; + xor.b64 %rd5376, %rd5375, %rd5341; + xor.b64 %rd5377, %rd5376, %rd5356; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3674}, %rd5365; + } + { + .reg .b32 %dummy; + mov.b64 {%r3675,%dummy}, %rd5365; + } + shf.l.wrap.b32 %r3676, %r3675, %r3674, 1; + shf.l.wrap.b32 %r3677, %r3674, %r3675, 1; + mov.b64 %rd5378, {%r3677, %r3676}; + xor.b64 %rd5379, %rd5377, %rd5378; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3678}, %rd5369; + } + { + .reg .b32 %dummy; + mov.b64 {%r3679,%dummy}, %rd5369; + } + shf.l.wrap.b32 %r3680, %r3679, %r3678, 1; + shf.l.wrap.b32 %r3681, %r3678, %r3679, 1; + mov.b64 %rd5380, {%r3681, %r3680}; + xor.b64 %rd5381, %rd5380, %rd5361; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3682}, %rd5373; + } + { + .reg .b32 %dummy; + mov.b64 {%r3683,%dummy}, %rd5373; + } + shf.l.wrap.b32 %r3684, %r3683, %r3682, 1; + shf.l.wrap.b32 %r3685, %r3682, %r3683, 1; + mov.b64 %rd5382, {%r3685, %r3684}; + xor.b64 %rd5383, %rd5382, %rd5365; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3686}, %rd5377; + } + { + .reg .b32 %dummy; + mov.b64 {%r3687,%dummy}, %rd5377; + } + shf.l.wrap.b32 %r3688, %r3687, %r3686, 1; + shf.l.wrap.b32 %r3689, %r3686, %r3687, 1; + mov.b64 %rd5384, {%r3689, %r3688}; + xor.b64 %rd5385, %rd5384, %rd5369; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3690}, %rd5361; + } + { + .reg .b32 %dummy; + mov.b64 {%r3691,%dummy}, %rd5361; + } + shf.l.wrap.b32 %r3692, %r3691, %r3690, 1; + shf.l.wrap.b32 %r3693, %r3690, %r3691, 1; + mov.b64 %rd5386, {%r3693, %r3692}; + xor.b64 %rd5387, %rd5373, %rd5386; + xor.b64 %rd5388, %rd5379, %rd5357; + xor.b64 %rd5389, %rd5381, %rd5287; + xor.b64 %rd5390, %rd5383, %rd5290; + xor.b64 %rd5391, %rd5385, %rd5293; + xor.b64 %rd5392, %rd5387, %rd5296; + xor.b64 %rd5393, %rd5379, %rd5299; + xor.b64 %rd5394, %rd5381, %rd5302; + xor.b64 %rd5395, %rd5383, %rd5305; + xor.b64 %rd5396, %rd5385, %rd5308; + xor.b64 %rd5397, %rd5387, %rd5311; + xor.b64 %rd5398, %rd5379, %rd5314; + xor.b64 %rd5399, %rd5381, %rd5317; + xor.b64 %rd5400, %rd5383, %rd5320; + xor.b64 %rd5401, %rd5385, %rd5323; + xor.b64 %rd5402, %rd5387, %rd5326; + xor.b64 %rd5403, %rd5379, %rd5329; + xor.b64 %rd5404, %rd5381, %rd5332; + xor.b64 %rd5405, %rd5383, %rd5335; + xor.b64 %rd5406, %rd5385, %rd5338; + xor.b64 %rd5407, %rd5387, %rd5341; + xor.b64 %rd5408, %rd5379, %rd5344; + xor.b64 %rd5409, %rd5381, %rd5347; + xor.b64 %rd5410, %rd5383, %rd5350; + xor.b64 %rd5411, %rd5385, %rd5353; + xor.b64 %rd5412, %rd5387, %rd5356; + { + .reg .b32 %dummy; + mov.b64 {%r3694,%dummy}, %rd5394; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3695}, %rd5394; + } + shf.r.wrap.b32 %r3696, %r3695, %r3694, 20; + shf.r.wrap.b32 %r3697, %r3694, %r3695, 20; + mov.b64 %rd5413, {%r3697, %r3696}; + { + .reg .b32 %dummy; + mov.b64 {%r3698,%dummy}, %rd5400; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3699}, %rd5400; + } + shf.r.wrap.b32 %r3700, %r3699, %r3698, 21; + shf.r.wrap.b32 %r3701, %r3698, %r3699, 21; + mov.b64 %rd5414, {%r3701, %r3700}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3702}, %rd5406; + } + { + .reg .b32 %dummy; + mov.b64 {%r3703,%dummy}, %rd5406; + } + shf.l.wrap.b32 %r3704, %r3703, %r3702, 21; + shf.l.wrap.b32 %r3705, %r3702, %r3703, 21; + mov.b64 %rd5415, {%r3705, %r3704}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3706}, %rd5412; + } + { + .reg .b32 %dummy; + mov.b64 {%r3707,%dummy}, %rd5412; + } + shf.l.wrap.b32 %r3708, %r3707, %r3706, 14; + shf.l.wrap.b32 %r3709, %r3706, %r3707, 14; + mov.b64 %rd5416, {%r3709, %r3708}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3710}, %rd5391; + } + { + .reg .b32 %dummy; + mov.b64 {%r3711,%dummy}, %rd5391; + } + shf.l.wrap.b32 %r3712, %r3711, %r3710, 28; + shf.l.wrap.b32 %r3713, %r3710, %r3711, 28; + mov.b64 %rd5417, {%r3713, %r3712}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3714}, %rd5397; + } + { + .reg .b32 %dummy; + mov.b64 {%r3715,%dummy}, %rd5397; + } + shf.l.wrap.b32 %r3716, %r3715, %r3714, 20; + shf.l.wrap.b32 %r3717, %r3714, %r3715, 20; + mov.b64 %rd5418, {%r3717, %r3716}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3718}, %rd5398; + } + { + .reg .b32 %dummy; + mov.b64 {%r3719,%dummy}, %rd5398; + } + shf.l.wrap.b32 %r3720, %r3719, %r3718, 3; + shf.l.wrap.b32 %r3721, %r3718, %r3719, 3; + mov.b64 %rd5419, {%r3721, %r3720}; + { + .reg .b32 %dummy; + mov.b64 {%r3722,%dummy}, %rd5404; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3723}, %rd5404; + } + shf.r.wrap.b32 %r3724, %r3723, %r3722, 19; + shf.r.wrap.b32 %r3725, %r3722, %r3723, 19; + mov.b64 %rd5420, {%r3725, %r3724}; + { + .reg .b32 %dummy; + mov.b64 {%r3726,%dummy}, %rd5410; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3727}, %rd5410; + } + shf.r.wrap.b32 %r3728, %r3727, %r3726, 3; + shf.r.wrap.b32 %r3729, %r3726, %r3727, 3; + mov.b64 %rd5421, {%r3729, %r3728}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3730}, %rd5389; + } + { + .reg .b32 %dummy; + mov.b64 {%r3731,%dummy}, %rd5389; + } + shf.l.wrap.b32 %r3732, %r3731, %r3730, 1; + shf.l.wrap.b32 %r3733, %r3730, %r3731, 1; + mov.b64 %rd5422, {%r3733, %r3732}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3734}, %rd5395; + } + { + .reg .b32 %dummy; + mov.b64 {%r3735,%dummy}, %rd5395; + } + shf.l.wrap.b32 %r3736, %r3735, %r3734, 6; + shf.l.wrap.b32 %r3737, %r3734, %r3735, 6; + mov.b64 %rd5423, {%r3737, %r3736}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3738}, %rd5401; + } + { + .reg .b32 %dummy; + mov.b64 {%r3739,%dummy}, %rd5401; + } + shf.l.wrap.b32 %r3740, %r3739, %r3738, 25; + shf.l.wrap.b32 %r3741, %r3738, %r3739, 25; + mov.b64 %rd5424, {%r3741, %r3740}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3742}, %rd5407; + } + { + .reg .b32 %dummy; + mov.b64 {%r3743,%dummy}, %rd5407; + } + shf.l.wrap.b32 %r3744, %r3743, %r3742, 8; + shf.l.wrap.b32 %r3745, %r3742, %r3743, 8; + mov.b64 %rd5425, {%r3745, %r3744}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3746}, %rd5408; + } + { + .reg .b32 %dummy; + mov.b64 {%r3747,%dummy}, %rd5408; + } + shf.l.wrap.b32 %r3748, %r3747, %r3746, 18; + shf.l.wrap.b32 %r3749, %r3746, %r3747, 18; + mov.b64 %rd5426, {%r3749, %r3748}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3750}, %rd5392; + } + { + .reg .b32 %dummy; + mov.b64 {%r3751,%dummy}, %rd5392; + } + shf.l.wrap.b32 %r3752, %r3751, %r3750, 27; + shf.l.wrap.b32 %r3753, %r3750, %r3751, 27; + mov.b64 %rd5427, {%r3753, %r3752}; + { + .reg .b32 %dummy; + mov.b64 {%r3754,%dummy}, %rd5393; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3755}, %rd5393; + } + shf.r.wrap.b32 %r3756, %r3755, %r3754, 28; + shf.r.wrap.b32 %r3757, %r3754, %r3755, 28; + mov.b64 %rd5428, {%r3757, %r3756}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3758}, %rd5399; + } + { + .reg .b32 %dummy; + mov.b64 {%r3759,%dummy}, %rd5399; + } + shf.l.wrap.b32 %r3760, %r3759, %r3758, 10; + shf.l.wrap.b32 %r3761, %r3758, %r3759, 10; + mov.b64 %rd5429, {%r3761, %r3760}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3762}, %rd5405; + } + { + .reg .b32 %dummy; + mov.b64 {%r3763,%dummy}, %rd5405; + } + shf.l.wrap.b32 %r3764, %r3763, %r3762, 15; + shf.l.wrap.b32 %r3765, %r3762, %r3763, 15; + mov.b64 %rd5430, {%r3765, %r3764}; + { + .reg .b32 %dummy; + mov.b64 {%r3766,%dummy}, %rd5411; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3767}, %rd5411; + } + shf.r.wrap.b32 %r3768, %r3767, %r3766, 8; + shf.r.wrap.b32 %r3769, %r3766, %r3767, 8; + mov.b64 %rd5431, {%r3769, %r3768}; + { + .reg .b32 %dummy; + mov.b64 {%r3770,%dummy}, %rd5390; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3771}, %rd5390; + } + shf.r.wrap.b32 %r3772, %r3771, %r3770, 2; + shf.r.wrap.b32 %r3773, %r3770, %r3771, 2; + mov.b64 %rd5432, {%r3773, %r3772}; + { + .reg .b32 %dummy; + mov.b64 {%r3774,%dummy}, %rd5396; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3775}, %rd5396; + } + shf.r.wrap.b32 %r3776, %r3775, %r3774, 9; + shf.r.wrap.b32 %r3777, %r3774, %r3775, 9; + mov.b64 %rd5433, {%r3777, %r3776}; + { + .reg .b32 %dummy; + mov.b64 {%r3778,%dummy}, %rd5402; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3779}, %rd5402; + } + shf.r.wrap.b32 %r3780, %r3779, %r3778, 25; + shf.r.wrap.b32 %r3781, %r3778, %r3779, 25; + mov.b64 %rd5434, {%r3781, %r3780}; + { + .reg .b32 %dummy; + mov.b64 {%r3782,%dummy}, %rd5403; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3783}, %rd5403; + } + shf.r.wrap.b32 %r3784, %r3783, %r3782, 23; + shf.r.wrap.b32 %r3785, %r3782, %r3783, 23; + mov.b64 %rd5435, {%r3785, %r3784}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3786}, %rd5409; + } + { + .reg .b32 %dummy; + mov.b64 {%r3787,%dummy}, %rd5409; + } + shf.l.wrap.b32 %r3788, %r3787, %r3786, 2; + shf.l.wrap.b32 %r3789, %r3786, %r3787, 2; + mov.b64 %rd5436, {%r3789, %r3788}; + not.b64 %rd5437, %rd5413; + and.b64 %rd5438, %rd5414, %rd5437; + xor.b64 %rd5439, %rd5438, %rd5388; + not.b64 %rd5440, %rd5414; + and.b64 %rd5441, %rd5415, %rd5440; + xor.b64 %rd5442, %rd5441, %rd5413; + not.b64 %rd5443, %rd5415; + and.b64 %rd5444, %rd5416, %rd5443; + xor.b64 %rd5445, %rd5444, %rd5414; + not.b64 %rd5446, %rd5416; + and.b64 %rd5447, %rd5388, %rd5446; + xor.b64 %rd5448, %rd5447, %rd5415; + not.b64 %rd5449, %rd5388; + and.b64 %rd5450, %rd5413, %rd5449; + xor.b64 %rd5451, %rd5450, %rd5416; + not.b64 %rd5452, %rd5418; + and.b64 %rd5453, %rd5419, %rd5452; + xor.b64 %rd5454, %rd5453, %rd5417; + not.b64 %rd5455, %rd5419; + and.b64 %rd5456, %rd5420, %rd5455; + xor.b64 %rd5457, %rd5456, %rd5418; + not.b64 %rd5458, %rd5420; + and.b64 %rd5459, %rd5421, %rd5458; + xor.b64 %rd5460, %rd5459, %rd5419; + not.b64 %rd5461, %rd5421; + and.b64 %rd5462, %rd5417, %rd5461; + xor.b64 %rd5463, %rd5462, %rd5420; + not.b64 %rd5464, %rd5417; + and.b64 %rd5465, %rd5418, %rd5464; + xor.b64 %rd5466, %rd5465, %rd5421; + not.b64 %rd5467, %rd5423; + and.b64 %rd5468, %rd5424, %rd5467; + xor.b64 %rd5469, %rd5468, %rd5422; + not.b64 %rd5470, %rd5424; + and.b64 %rd5471, %rd5425, %rd5470; + xor.b64 %rd5472, %rd5471, %rd5423; + not.b64 %rd5473, %rd5425; + and.b64 %rd5474, %rd5426, %rd5473; + xor.b64 %rd5475, %rd5474, %rd5424; + not.b64 %rd5476, %rd5426; + and.b64 %rd5477, %rd5422, %rd5476; + xor.b64 %rd5478, %rd5477, %rd5425; + not.b64 %rd5479, %rd5422; + and.b64 %rd5480, %rd5423, %rd5479; + xor.b64 %rd5481, %rd5480, %rd5426; + not.b64 %rd5482, %rd5428; + and.b64 %rd5483, %rd5429, %rd5482; + xor.b64 %rd5484, %rd5483, %rd5427; + not.b64 %rd5485, %rd5429; + and.b64 %rd5486, %rd5430, %rd5485; + xor.b64 %rd5487, %rd5486, %rd5428; + not.b64 %rd5488, %rd5430; + and.b64 %rd5489, %rd5431, %rd5488; + xor.b64 %rd5490, %rd5489, %rd5429; + not.b64 %rd5491, %rd5431; + and.b64 %rd5492, %rd5427, %rd5491; + xor.b64 %rd5493, %rd5492, %rd5430; + not.b64 %rd5494, %rd5427; + and.b64 %rd5495, %rd5428, %rd5494; + xor.b64 %rd5496, %rd5495, %rd5431; + not.b64 %rd5497, %rd5433; + and.b64 %rd5498, %rd5434, %rd5497; + xor.b64 %rd5499, %rd5498, %rd5432; + not.b64 %rd5500, %rd5434; + and.b64 %rd5501, %rd5435, %rd5500; + xor.b64 %rd5502, %rd5501, %rd5433; + not.b64 %rd5503, %rd5435; + and.b64 %rd5504, %rd5436, %rd5503; + xor.b64 %rd5505, %rd5504, %rd5434; + not.b64 %rd5506, %rd5436; + and.b64 %rd5507, %rd5432, %rd5506; + xor.b64 %rd5508, %rd5507, %rd5435; + not.b64 %rd5509, %rd5432; + and.b64 %rd5510, %rd5433, %rd5509; + xor.b64 %rd5511, %rd5510, %rd5436; + xor.b64 %rd5512, %rd5439, %rd23; + xor.b64 %rd5513, %rd5454, %rd5512; + xor.b64 %rd5514, %rd5513, %rd5469; + xor.b64 %rd5515, %rd5514, %rd5484; + xor.b64 %rd5516, %rd5515, %rd5499; + xor.b64 %rd5517, %rd5457, %rd5442; + xor.b64 %rd5518, %rd5517, %rd5472; + xor.b64 %rd5519, %rd5518, %rd5487; + xor.b64 %rd5520, %rd5519, %rd5502; + xor.b64 %rd5521, %rd5460, %rd5445; + xor.b64 %rd5522, %rd5521, %rd5475; + xor.b64 %rd5523, %rd5522, %rd5490; + xor.b64 %rd5524, %rd5523, %rd5505; + xor.b64 %rd5525, %rd5463, %rd5448; + xor.b64 %rd5526, %rd5525, %rd5478; + xor.b64 %rd5527, %rd5526, %rd5493; + xor.b64 %rd5528, %rd5527, %rd5508; + xor.b64 %rd5529, %rd5466, %rd5451; + xor.b64 %rd5530, %rd5529, %rd5481; + xor.b64 %rd5531, %rd5530, %rd5496; + xor.b64 %rd5532, %rd5531, %rd5511; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3790}, %rd5520; + } + { + .reg .b32 %dummy; + mov.b64 {%r3791,%dummy}, %rd5520; + } + shf.l.wrap.b32 %r3792, %r3791, %r3790, 1; + shf.l.wrap.b32 %r3793, %r3790, %r3791, 1; + mov.b64 %rd5533, {%r3793, %r3792}; + xor.b64 %rd5534, %rd5532, %rd5533; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3794}, %rd5524; + } + { + .reg .b32 %dummy; + mov.b64 {%r3795,%dummy}, %rd5524; + } + shf.l.wrap.b32 %r3796, %r3795, %r3794, 1; + shf.l.wrap.b32 %r3797, %r3794, %r3795, 1; + mov.b64 %rd5535, {%r3797, %r3796}; + xor.b64 %rd5536, %rd5535, %rd5516; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3798}, %rd5528; + } + { + .reg .b32 %dummy; + mov.b64 {%r3799,%dummy}, %rd5528; + } + shf.l.wrap.b32 %r3800, %r3799, %r3798, 1; + shf.l.wrap.b32 %r3801, %r3798, %r3799, 1; + mov.b64 %rd5537, {%r3801, %r3800}; + xor.b64 %rd5538, %rd5537, %rd5520; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3802}, %rd5532; + } + { + .reg .b32 %dummy; + mov.b64 {%r3803,%dummy}, %rd5532; + } + shf.l.wrap.b32 %r3804, %r3803, %r3802, 1; + shf.l.wrap.b32 %r3805, %r3802, %r3803, 1; + mov.b64 %rd5539, {%r3805, %r3804}; + xor.b64 %rd5540, %rd5539, %rd5524; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3806}, %rd5516; + } + { + .reg .b32 %dummy; + mov.b64 {%r3807,%dummy}, %rd5516; + } + shf.l.wrap.b32 %r3808, %r3807, %r3806, 1; + shf.l.wrap.b32 %r3809, %r3806, %r3807, 1; + mov.b64 %rd5541, {%r3809, %r3808}; + xor.b64 %rd5542, %rd5528, %rd5541; + xor.b64 %rd5543, %rd5534, %rd5512; + xor.b64 %rd5544, %rd5536, %rd5442; + xor.b64 %rd5545, %rd5538, %rd5445; + xor.b64 %rd5546, %rd5540, %rd5448; + xor.b64 %rd5547, %rd5542, %rd5451; + xor.b64 %rd5548, %rd5534, %rd5454; + xor.b64 %rd5549, %rd5536, %rd5457; + xor.b64 %rd5550, %rd5538, %rd5460; + xor.b64 %rd5551, %rd5540, %rd5463; + xor.b64 %rd5552, %rd5542, %rd5466; + xor.b64 %rd5553, %rd5534, %rd5469; + xor.b64 %rd5554, %rd5536, %rd5472; + xor.b64 %rd5555, %rd5538, %rd5475; + xor.b64 %rd5556, %rd5540, %rd5478; + xor.b64 %rd5557, %rd5542, %rd5481; + xor.b64 %rd5558, %rd5534, %rd5484; + xor.b64 %rd5559, %rd5536, %rd5487; + xor.b64 %rd5560, %rd5538, %rd5490; + xor.b64 %rd5561, %rd5540, %rd5493; + xor.b64 %rd5562, %rd5542, %rd5496; + xor.b64 %rd5563, %rd5534, %rd5499; + xor.b64 %rd5564, %rd5536, %rd5502; + xor.b64 %rd5565, %rd5538, %rd5505; + xor.b64 %rd5566, %rd5540, %rd5508; + xor.b64 %rd5567, %rd5542, %rd5511; + { + .reg .b32 %dummy; + mov.b64 {%r3810,%dummy}, %rd5549; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3811}, %rd5549; + } + shf.r.wrap.b32 %r3812, %r3811, %r3810, 20; + shf.r.wrap.b32 %r3813, %r3810, %r3811, 20; + mov.b64 %rd5568, {%r3813, %r3812}; + { + .reg .b32 %dummy; + mov.b64 {%r3814,%dummy}, %rd5555; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3815}, %rd5555; + } + shf.r.wrap.b32 %r3816, %r3815, %r3814, 21; + shf.r.wrap.b32 %r3817, %r3814, %r3815, 21; + mov.b64 %rd5569, {%r3817, %r3816}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3818}, %rd5561; + } + { + .reg .b32 %dummy; + mov.b64 {%r3819,%dummy}, %rd5561; + } + shf.l.wrap.b32 %r3820, %r3819, %r3818, 21; + shf.l.wrap.b32 %r3821, %r3818, %r3819, 21; + mov.b64 %rd5570, {%r3821, %r3820}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3822}, %rd5567; + } + { + .reg .b32 %dummy; + mov.b64 {%r3823,%dummy}, %rd5567; + } + shf.l.wrap.b32 %r3824, %r3823, %r3822, 14; + shf.l.wrap.b32 %r3825, %r3822, %r3823, 14; + mov.b64 %rd5571, {%r3825, %r3824}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3826}, %rd5546; + } + { + .reg .b32 %dummy; + mov.b64 {%r3827,%dummy}, %rd5546; + } + shf.l.wrap.b32 %r3828, %r3827, %r3826, 28; + shf.l.wrap.b32 %r3829, %r3826, %r3827, 28; + mov.b64 %rd5572, {%r3829, %r3828}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3830}, %rd5552; + } + { + .reg .b32 %dummy; + mov.b64 {%r3831,%dummy}, %rd5552; + } + shf.l.wrap.b32 %r3832, %r3831, %r3830, 20; + shf.l.wrap.b32 %r3833, %r3830, %r3831, 20; + mov.b64 %rd5573, {%r3833, %r3832}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3834}, %rd5553; + } + { + .reg .b32 %dummy; + mov.b64 {%r3835,%dummy}, %rd5553; + } + shf.l.wrap.b32 %r3836, %r3835, %r3834, 3; + shf.l.wrap.b32 %r3837, %r3834, %r3835, 3; + mov.b64 %rd5574, {%r3837, %r3836}; + { + .reg .b32 %dummy; + mov.b64 {%r3838,%dummy}, %rd5559; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3839}, %rd5559; + } + shf.r.wrap.b32 %r3840, %r3839, %r3838, 19; + shf.r.wrap.b32 %r3841, %r3838, %r3839, 19; + mov.b64 %rd5575, {%r3841, %r3840}; + { + .reg .b32 %dummy; + mov.b64 {%r3842,%dummy}, %rd5565; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3843}, %rd5565; + } + shf.r.wrap.b32 %r3844, %r3843, %r3842, 3; + shf.r.wrap.b32 %r3845, %r3842, %r3843, 3; + mov.b64 %rd5576, {%r3845, %r3844}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3846}, %rd5544; + } + { + .reg .b32 %dummy; + mov.b64 {%r3847,%dummy}, %rd5544; + } + shf.l.wrap.b32 %r3848, %r3847, %r3846, 1; + shf.l.wrap.b32 %r3849, %r3846, %r3847, 1; + mov.b64 %rd5577, {%r3849, %r3848}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3850}, %rd5550; + } + { + .reg .b32 %dummy; + mov.b64 {%r3851,%dummy}, %rd5550; + } + shf.l.wrap.b32 %r3852, %r3851, %r3850, 6; + shf.l.wrap.b32 %r3853, %r3850, %r3851, 6; + mov.b64 %rd5578, {%r3853, %r3852}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3854}, %rd5556; + } + { + .reg .b32 %dummy; + mov.b64 {%r3855,%dummy}, %rd5556; + } + shf.l.wrap.b32 %r3856, %r3855, %r3854, 25; + shf.l.wrap.b32 %r3857, %r3854, %r3855, 25; + mov.b64 %rd5579, {%r3857, %r3856}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3858}, %rd5562; + } + { + .reg .b32 %dummy; + mov.b64 {%r3859,%dummy}, %rd5562; + } + shf.l.wrap.b32 %r3860, %r3859, %r3858, 8; + shf.l.wrap.b32 %r3861, %r3858, %r3859, 8; + mov.b64 %rd5580, {%r3861, %r3860}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3862}, %rd5563; + } + { + .reg .b32 %dummy; + mov.b64 {%r3863,%dummy}, %rd5563; + } + shf.l.wrap.b32 %r3864, %r3863, %r3862, 18; + shf.l.wrap.b32 %r3865, %r3862, %r3863, 18; + mov.b64 %rd5581, {%r3865, %r3864}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3866}, %rd5547; + } + { + .reg .b32 %dummy; + mov.b64 {%r3867,%dummy}, %rd5547; + } + shf.l.wrap.b32 %r3868, %r3867, %r3866, 27; + shf.l.wrap.b32 %r3869, %r3866, %r3867, 27; + mov.b64 %rd5582, {%r3869, %r3868}; + { + .reg .b32 %dummy; + mov.b64 {%r3870,%dummy}, %rd5548; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3871}, %rd5548; + } + shf.r.wrap.b32 %r3872, %r3871, %r3870, 28; + shf.r.wrap.b32 %r3873, %r3870, %r3871, 28; + mov.b64 %rd5583, {%r3873, %r3872}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3874}, %rd5554; + } + { + .reg .b32 %dummy; + mov.b64 {%r3875,%dummy}, %rd5554; + } + shf.l.wrap.b32 %r3876, %r3875, %r3874, 10; + shf.l.wrap.b32 %r3877, %r3874, %r3875, 10; + mov.b64 %rd5584, {%r3877, %r3876}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3878}, %rd5560; + } + { + .reg .b32 %dummy; + mov.b64 {%r3879,%dummy}, %rd5560; + } + shf.l.wrap.b32 %r3880, %r3879, %r3878, 15; + shf.l.wrap.b32 %r3881, %r3878, %r3879, 15; + mov.b64 %rd5585, {%r3881, %r3880}; + { + .reg .b32 %dummy; + mov.b64 {%r3882,%dummy}, %rd5566; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3883}, %rd5566; + } + shf.r.wrap.b32 %r3884, %r3883, %r3882, 8; + shf.r.wrap.b32 %r3885, %r3882, %r3883, 8; + mov.b64 %rd5586, {%r3885, %r3884}; + { + .reg .b32 %dummy; + mov.b64 {%r3886,%dummy}, %rd5545; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3887}, %rd5545; + } + shf.r.wrap.b32 %r3888, %r3887, %r3886, 2; + shf.r.wrap.b32 %r3889, %r3886, %r3887, 2; + mov.b64 %rd5587, {%r3889, %r3888}; + { + .reg .b32 %dummy; + mov.b64 {%r3890,%dummy}, %rd5551; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3891}, %rd5551; + } + shf.r.wrap.b32 %r3892, %r3891, %r3890, 9; + shf.r.wrap.b32 %r3893, %r3890, %r3891, 9; + mov.b64 %rd5588, {%r3893, %r3892}; + { + .reg .b32 %dummy; + mov.b64 {%r3894,%dummy}, %rd5557; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3895}, %rd5557; + } + shf.r.wrap.b32 %r3896, %r3895, %r3894, 25; + shf.r.wrap.b32 %r3897, %r3894, %r3895, 25; + mov.b64 %rd5589, {%r3897, %r3896}; + { + .reg .b32 %dummy; + mov.b64 {%r3898,%dummy}, %rd5558; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3899}, %rd5558; + } + shf.r.wrap.b32 %r3900, %r3899, %r3898, 23; + shf.r.wrap.b32 %r3901, %r3898, %r3899, 23; + mov.b64 %rd5590, {%r3901, %r3900}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3902}, %rd5564; + } + { + .reg .b32 %dummy; + mov.b64 {%r3903,%dummy}, %rd5564; + } + shf.l.wrap.b32 %r3904, %r3903, %r3902, 2; + shf.l.wrap.b32 %r3905, %r3902, %r3903, 2; + mov.b64 %rd5591, {%r3905, %r3904}; + not.b64 %rd5592, %rd5568; + and.b64 %rd5593, %rd5569, %rd5592; + xor.b64 %rd5594, %rd5593, %rd5543; + not.b64 %rd5595, %rd5569; + and.b64 %rd5596, %rd5570, %rd5595; + xor.b64 %rd5597, %rd5596, %rd5568; + not.b64 %rd5598, %rd5570; + and.b64 %rd5599, %rd5571, %rd5598; + xor.b64 %rd5600, %rd5599, %rd5569; + not.b64 %rd5601, %rd5571; + and.b64 %rd5602, %rd5543, %rd5601; + xor.b64 %rd5603, %rd5602, %rd5570; + not.b64 %rd5604, %rd5543; + and.b64 %rd5605, %rd5568, %rd5604; + xor.b64 %rd5606, %rd5605, %rd5571; + not.b64 %rd5607, %rd5573; + and.b64 %rd5608, %rd5574, %rd5607; + xor.b64 %rd5609, %rd5608, %rd5572; + not.b64 %rd5610, %rd5574; + and.b64 %rd5611, %rd5575, %rd5610; + xor.b64 %rd5612, %rd5611, %rd5573; + not.b64 %rd5613, %rd5575; + and.b64 %rd5614, %rd5576, %rd5613; + xor.b64 %rd5615, %rd5614, %rd5574; + not.b64 %rd5616, %rd5576; + and.b64 %rd5617, %rd5572, %rd5616; + xor.b64 %rd5618, %rd5617, %rd5575; + not.b64 %rd5619, %rd5572; + and.b64 %rd5620, %rd5573, %rd5619; + xor.b64 %rd5621, %rd5620, %rd5576; + not.b64 %rd5622, %rd5578; + and.b64 %rd5623, %rd5579, %rd5622; + xor.b64 %rd5624, %rd5623, %rd5577; + not.b64 %rd5625, %rd5579; + and.b64 %rd5626, %rd5580, %rd5625; + xor.b64 %rd5627, %rd5626, %rd5578; + not.b64 %rd5628, %rd5580; + and.b64 %rd5629, %rd5581, %rd5628; + xor.b64 %rd5630, %rd5629, %rd5579; + not.b64 %rd5631, %rd5581; + and.b64 %rd5632, %rd5577, %rd5631; + xor.b64 %rd5633, %rd5632, %rd5580; + not.b64 %rd5634, %rd5577; + and.b64 %rd5635, %rd5578, %rd5634; + xor.b64 %rd5636, %rd5635, %rd5581; + not.b64 %rd5637, %rd5583; + and.b64 %rd5638, %rd5584, %rd5637; + xor.b64 %rd5639, %rd5638, %rd5582; + not.b64 %rd5640, %rd5584; + and.b64 %rd5641, %rd5585, %rd5640; + xor.b64 %rd5642, %rd5641, %rd5583; + not.b64 %rd5643, %rd5585; + and.b64 %rd5644, %rd5586, %rd5643; + xor.b64 %rd5645, %rd5644, %rd5584; + not.b64 %rd5646, %rd5586; + and.b64 %rd5647, %rd5582, %rd5646; + xor.b64 %rd5648, %rd5647, %rd5585; + not.b64 %rd5649, %rd5582; + and.b64 %rd5650, %rd5583, %rd5649; + xor.b64 %rd5651, %rd5650, %rd5586; + not.b64 %rd5652, %rd5588; + and.b64 %rd5653, %rd5589, %rd5652; + xor.b64 %rd5654, %rd5653, %rd5587; + not.b64 %rd5655, %rd5589; + and.b64 %rd5656, %rd5590, %rd5655; + xor.b64 %rd5657, %rd5656, %rd5588; + not.b64 %rd5658, %rd5590; + and.b64 %rd5659, %rd5591, %rd5658; + xor.b64 %rd5660, %rd5659, %rd5589; + not.b64 %rd5661, %rd5591; + and.b64 %rd5662, %rd5587, %rd5661; + xor.b64 %rd5663, %rd5662, %rd5590; + not.b64 %rd5664, %rd5587; + and.b64 %rd5665, %rd5588, %rd5664; + xor.b64 %rd5666, %rd5665, %rd5591; + xor.b64 %rd5667, %rd5594, %rd24; + xor.b64 %rd5668, %rd5609, %rd5667; + xor.b64 %rd5669, %rd5668, %rd5624; + xor.b64 %rd5670, %rd5669, %rd5639; + xor.b64 %rd5671, %rd5670, %rd5654; + xor.b64 %rd5672, %rd5612, %rd5597; + xor.b64 %rd5673, %rd5672, %rd5627; + xor.b64 %rd5674, %rd5673, %rd5642; + xor.b64 %rd5675, %rd5674, %rd5657; + xor.b64 %rd5676, %rd5615, %rd5600; + xor.b64 %rd5677, %rd5676, %rd5630; + xor.b64 %rd5678, %rd5677, %rd5645; + xor.b64 %rd5679, %rd5678, %rd5660; + xor.b64 %rd5680, %rd5618, %rd5603; + xor.b64 %rd5681, %rd5680, %rd5633; + xor.b64 %rd5682, %rd5681, %rd5648; + xor.b64 %rd5683, %rd5682, %rd5663; + xor.b64 %rd5684, %rd5621, %rd5606; + xor.b64 %rd5685, %rd5684, %rd5636; + xor.b64 %rd5686, %rd5685, %rd5651; + xor.b64 %rd5687, %rd5686, %rd5666; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3906}, %rd5675; + } + { + .reg .b32 %dummy; + mov.b64 {%r3907,%dummy}, %rd5675; + } + shf.l.wrap.b32 %r3908, %r3907, %r3906, 1; + shf.l.wrap.b32 %r3909, %r3906, %r3907, 1; + mov.b64 %rd5688, {%r3909, %r3908}; + xor.b64 %rd5689, %rd5687, %rd5688; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3910}, %rd5679; + } + { + .reg .b32 %dummy; + mov.b64 {%r3911,%dummy}, %rd5679; + } + shf.l.wrap.b32 %r3912, %r3911, %r3910, 1; + shf.l.wrap.b32 %r3913, %r3910, %r3911, 1; + mov.b64 %rd5690, {%r3913, %r3912}; + xor.b64 %rd5691, %rd5690, %rd5671; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3914}, %rd5683; + } + { + .reg .b32 %dummy; + mov.b64 {%r3915,%dummy}, %rd5683; + } + shf.l.wrap.b32 %r3916, %r3915, %r3914, 1; + shf.l.wrap.b32 %r3917, %r3914, %r3915, 1; + mov.b64 %rd5692, {%r3917, %r3916}; + xor.b64 %rd5693, %rd5692, %rd5675; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3918}, %rd5687; + } + { + .reg .b32 %dummy; + mov.b64 {%r3919,%dummy}, %rd5687; + } + shf.l.wrap.b32 %r3920, %r3919, %r3918, 1; + shf.l.wrap.b32 %r3921, %r3918, %r3919, 1; + mov.b64 %rd5694, {%r3921, %r3920}; + xor.b64 %rd5695, %rd5694, %rd5679; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3922}, %rd5671; + } + { + .reg .b32 %dummy; + mov.b64 {%r3923,%dummy}, %rd5671; + } + shf.l.wrap.b32 %r3924, %r3923, %r3922, 1; + shf.l.wrap.b32 %r3925, %r3922, %r3923, 1; + mov.b64 %rd5696, {%r3925, %r3924}; + xor.b64 %rd5697, %rd5683, %rd5696; + xor.b64 %rd5698, %rd5689, %rd5667; + xor.b64 %rd5699, %rd5691, %rd5597; + xor.b64 %rd5700, %rd5693, %rd5600; + xor.b64 %rd5701, %rd5695, %rd5603; + xor.b64 %rd5702, %rd5697, %rd5606; + xor.b64 %rd5703, %rd5689, %rd5609; + xor.b64 %rd5704, %rd5691, %rd5612; + xor.b64 %rd5705, %rd5693, %rd5615; + xor.b64 %rd5706, %rd5695, %rd5618; + xor.b64 %rd5707, %rd5697, %rd5621; + xor.b64 %rd5708, %rd5689, %rd5624; + xor.b64 %rd5709, %rd5691, %rd5627; + xor.b64 %rd5710, %rd5693, %rd5630; + xor.b64 %rd5711, %rd5695, %rd5633; + xor.b64 %rd5712, %rd5697, %rd5636; + xor.b64 %rd5713, %rd5689, %rd5639; + xor.b64 %rd5714, %rd5691, %rd5642; + xor.b64 %rd5715, %rd5693, %rd5645; + xor.b64 %rd5716, %rd5695, %rd5648; + xor.b64 %rd5717, %rd5697, %rd5651; + xor.b64 %rd5718, %rd5689, %rd5654; + xor.b64 %rd5719, %rd5691, %rd5657; + xor.b64 %rd5720, %rd5693, %rd5660; + xor.b64 %rd5721, %rd5695, %rd5663; + xor.b64 %rd5722, %rd5697, %rd5666; + { + .reg .b32 %dummy; + mov.b64 {%r3926,%dummy}, %rd5704; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3927}, %rd5704; + } + shf.r.wrap.b32 %r3928, %r3927, %r3926, 20; + shf.r.wrap.b32 %r3929, %r3926, %r3927, 20; + mov.b64 %rd5723, {%r3929, %r3928}; + { + .reg .b32 %dummy; + mov.b64 {%r3930,%dummy}, %rd5710; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3931}, %rd5710; + } + shf.r.wrap.b32 %r3932, %r3931, %r3930, 21; + shf.r.wrap.b32 %r3933, %r3930, %r3931, 21; + mov.b64 %rd5724, {%r3933, %r3932}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3934}, %rd5716; + } + { + .reg .b32 %dummy; + mov.b64 {%r3935,%dummy}, %rd5716; + } + shf.l.wrap.b32 %r3936, %r3935, %r3934, 21; + shf.l.wrap.b32 %r3937, %r3934, %r3935, 21; + mov.b64 %rd5725, {%r3937, %r3936}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3938}, %rd5722; + } + { + .reg .b32 %dummy; + mov.b64 {%r3939,%dummy}, %rd5722; + } + shf.l.wrap.b32 %r3940, %r3939, %r3938, 14; + shf.l.wrap.b32 %r3941, %r3938, %r3939, 14; + mov.b64 %rd5726, {%r3941, %r3940}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3942}, %rd5701; + } + { + .reg .b32 %dummy; + mov.b64 {%r3943,%dummy}, %rd5701; + } + shf.l.wrap.b32 %r3944, %r3943, %r3942, 28; + shf.l.wrap.b32 %r3945, %r3942, %r3943, 28; + mov.b64 %rd5727, {%r3945, %r3944}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3946}, %rd5707; + } + { + .reg .b32 %dummy; + mov.b64 {%r3947,%dummy}, %rd5707; + } + shf.l.wrap.b32 %r3948, %r3947, %r3946, 20; + shf.l.wrap.b32 %r3949, %r3946, %r3947, 20; + mov.b64 %rd5728, {%r3949, %r3948}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3950}, %rd5708; + } + { + .reg .b32 %dummy; + mov.b64 {%r3951,%dummy}, %rd5708; + } + shf.l.wrap.b32 %r3952, %r3951, %r3950, 3; + shf.l.wrap.b32 %r3953, %r3950, %r3951, 3; + mov.b64 %rd5729, {%r3953, %r3952}; + { + .reg .b32 %dummy; + mov.b64 {%r3954,%dummy}, %rd5714; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3955}, %rd5714; + } + shf.r.wrap.b32 %r3956, %r3955, %r3954, 19; + shf.r.wrap.b32 %r3957, %r3954, %r3955, 19; + mov.b64 %rd5730, {%r3957, %r3956}; + { + .reg .b32 %dummy; + mov.b64 {%r3958,%dummy}, %rd5720; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3959}, %rd5720; + } + shf.r.wrap.b32 %r3960, %r3959, %r3958, 3; + shf.r.wrap.b32 %r3961, %r3958, %r3959, 3; + mov.b64 %rd5731, {%r3961, %r3960}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3962}, %rd5699; + } + { + .reg .b32 %dummy; + mov.b64 {%r3963,%dummy}, %rd5699; + } + shf.l.wrap.b32 %r3964, %r3963, %r3962, 1; + shf.l.wrap.b32 %r3965, %r3962, %r3963, 1; + mov.b64 %rd5732, {%r3965, %r3964}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3966}, %rd5705; + } + { + .reg .b32 %dummy; + mov.b64 {%r3967,%dummy}, %rd5705; + } + shf.l.wrap.b32 %r3968, %r3967, %r3966, 6; + shf.l.wrap.b32 %r3969, %r3966, %r3967, 6; + mov.b64 %rd5733, {%r3969, %r3968}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3970}, %rd5711; + } + { + .reg .b32 %dummy; + mov.b64 {%r3971,%dummy}, %rd5711; + } + shf.l.wrap.b32 %r3972, %r3971, %r3970, 25; + shf.l.wrap.b32 %r3973, %r3970, %r3971, 25; + mov.b64 %rd5734, {%r3973, %r3972}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3974}, %rd5717; + } + { + .reg .b32 %dummy; + mov.b64 {%r3975,%dummy}, %rd5717; + } + shf.l.wrap.b32 %r3976, %r3975, %r3974, 8; + shf.l.wrap.b32 %r3977, %r3974, %r3975, 8; + mov.b64 %rd5735, {%r3977, %r3976}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3978}, %rd5718; + } + { + .reg .b32 %dummy; + mov.b64 {%r3979,%dummy}, %rd5718; + } + shf.l.wrap.b32 %r3980, %r3979, %r3978, 18; + shf.l.wrap.b32 %r3981, %r3978, %r3979, 18; + mov.b64 %rd5736, {%r3981, %r3980}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3982}, %rd5702; + } + { + .reg .b32 %dummy; + mov.b64 {%r3983,%dummy}, %rd5702; + } + shf.l.wrap.b32 %r3984, %r3983, %r3982, 27; + shf.l.wrap.b32 %r3985, %r3982, %r3983, 27; + mov.b64 %rd5737, {%r3985, %r3984}; + { + .reg .b32 %dummy; + mov.b64 {%r3986,%dummy}, %rd5703; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3987}, %rd5703; + } + shf.r.wrap.b32 %r3988, %r3987, %r3986, 28; + shf.r.wrap.b32 %r3989, %r3986, %r3987, 28; + mov.b64 %rd5738, {%r3989, %r3988}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3990}, %rd5709; + } + { + .reg .b32 %dummy; + mov.b64 {%r3991,%dummy}, %rd5709; + } + shf.l.wrap.b32 %r3992, %r3991, %r3990, 10; + shf.l.wrap.b32 %r3993, %r3990, %r3991, 10; + mov.b64 %rd5739, {%r3993, %r3992}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3994}, %rd5715; + } + { + .reg .b32 %dummy; + mov.b64 {%r3995,%dummy}, %rd5715; + } + shf.l.wrap.b32 %r3996, %r3995, %r3994, 15; + shf.l.wrap.b32 %r3997, %r3994, %r3995, 15; + mov.b64 %rd5740, {%r3997, %r3996}; + { + .reg .b32 %dummy; + mov.b64 {%r3998,%dummy}, %rd5721; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r3999}, %rd5721; + } + shf.r.wrap.b32 %r4000, %r3999, %r3998, 8; + shf.r.wrap.b32 %r4001, %r3998, %r3999, 8; + mov.b64 %rd5741, {%r4001, %r4000}; + { + .reg .b32 %dummy; + mov.b64 {%r4002,%dummy}, %rd5700; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4003}, %rd5700; + } + shf.r.wrap.b32 %r4004, %r4003, %r4002, 2; + shf.r.wrap.b32 %r4005, %r4002, %r4003, 2; + mov.b64 %rd5742, {%r4005, %r4004}; + { + .reg .b32 %dummy; + mov.b64 {%r4006,%dummy}, %rd5706; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4007}, %rd5706; + } + shf.r.wrap.b32 %r4008, %r4007, %r4006, 9; + shf.r.wrap.b32 %r4009, %r4006, %r4007, 9; + mov.b64 %rd5743, {%r4009, %r4008}; + { + .reg .b32 %dummy; + mov.b64 {%r4010,%dummy}, %rd5712; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4011}, %rd5712; + } + shf.r.wrap.b32 %r4012, %r4011, %r4010, 25; + shf.r.wrap.b32 %r4013, %r4010, %r4011, 25; + mov.b64 %rd5744, {%r4013, %r4012}; + { + .reg .b32 %dummy; + mov.b64 {%r4014,%dummy}, %rd5713; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4015}, %rd5713; + } + shf.r.wrap.b32 %r4016, %r4015, %r4014, 23; + shf.r.wrap.b32 %r4017, %r4014, %r4015, 23; + mov.b64 %rd5745, {%r4017, %r4016}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4018}, %rd5719; + } + { + .reg .b32 %dummy; + mov.b64 {%r4019,%dummy}, %rd5719; + } + shf.l.wrap.b32 %r4020, %r4019, %r4018, 2; + shf.l.wrap.b32 %r4021, %r4018, %r4019, 2; + mov.b64 %rd5746, {%r4021, %r4020}; + not.b64 %rd5747, %rd5723; + and.b64 %rd5748, %rd5724, %rd5747; + xor.b64 %rd5749, %rd5748, %rd5698; + not.b64 %rd5750, %rd5724; + and.b64 %rd5751, %rd5725, %rd5750; + xor.b64 %rd5752, %rd5751, %rd5723; + not.b64 %rd5753, %rd5725; + and.b64 %rd5754, %rd5726, %rd5753; + xor.b64 %rd5755, %rd5754, %rd5724; + not.b64 %rd5756, %rd5726; + and.b64 %rd5757, %rd5698, %rd5756; + xor.b64 %rd5758, %rd5757, %rd5725; + not.b64 %rd5759, %rd5698; + and.b64 %rd5760, %rd5723, %rd5759; + xor.b64 %rd5761, %rd5760, %rd5726; + not.b64 %rd5762, %rd5728; + and.b64 %rd5763, %rd5729, %rd5762; + xor.b64 %rd5764, %rd5763, %rd5727; + not.b64 %rd5765, %rd5729; + and.b64 %rd5766, %rd5730, %rd5765; + xor.b64 %rd5767, %rd5766, %rd5728; + not.b64 %rd5768, %rd5730; + and.b64 %rd5769, %rd5731, %rd5768; + xor.b64 %rd5770, %rd5769, %rd5729; + not.b64 %rd5771, %rd5731; + and.b64 %rd5772, %rd5727, %rd5771; + xor.b64 %rd5773, %rd5772, %rd5730; + not.b64 %rd5774, %rd5727; + and.b64 %rd5775, %rd5728, %rd5774; + xor.b64 %rd5776, %rd5775, %rd5731; + not.b64 %rd5777, %rd5733; + and.b64 %rd5778, %rd5734, %rd5777; + xor.b64 %rd5779, %rd5778, %rd5732; + not.b64 %rd5780, %rd5734; + and.b64 %rd5781, %rd5735, %rd5780; + xor.b64 %rd5782, %rd5781, %rd5733; + not.b64 %rd5783, %rd5735; + and.b64 %rd5784, %rd5736, %rd5783; + xor.b64 %rd5785, %rd5784, %rd5734; + not.b64 %rd5786, %rd5736; + and.b64 %rd5787, %rd5732, %rd5786; + xor.b64 %rd5788, %rd5787, %rd5735; + not.b64 %rd5789, %rd5732; + and.b64 %rd5790, %rd5733, %rd5789; + xor.b64 %rd5791, %rd5790, %rd5736; + not.b64 %rd5792, %rd5738; + and.b64 %rd5793, %rd5739, %rd5792; + xor.b64 %rd5794, %rd5793, %rd5737; + not.b64 %rd5795, %rd5739; + and.b64 %rd5796, %rd5740, %rd5795; + xor.b64 %rd5797, %rd5796, %rd5738; + not.b64 %rd5798, %rd5740; + and.b64 %rd5799, %rd5741, %rd5798; + xor.b64 %rd5800, %rd5799, %rd5739; + not.b64 %rd5801, %rd5741; + and.b64 %rd5802, %rd5737, %rd5801; + xor.b64 %rd5803, %rd5802, %rd5740; + not.b64 %rd5804, %rd5737; + and.b64 %rd5805, %rd5738, %rd5804; + xor.b64 %rd5806, %rd5805, %rd5741; + not.b64 %rd5807, %rd5743; + and.b64 %rd5808, %rd5744, %rd5807; + xor.b64 %rd5809, %rd5808, %rd5742; + not.b64 %rd5810, %rd5744; + and.b64 %rd5811, %rd5745, %rd5810; + xor.b64 %rd5812, %rd5811, %rd5743; + not.b64 %rd5813, %rd5745; + and.b64 %rd5814, %rd5746, %rd5813; + xor.b64 %rd5815, %rd5814, %rd5744; + not.b64 %rd5816, %rd5746; + and.b64 %rd5817, %rd5742, %rd5816; + xor.b64 %rd5818, %rd5817, %rd5745; + not.b64 %rd5819, %rd5742; + and.b64 %rd5820, %rd5743, %rd5819; + xor.b64 %rd5821, %rd5820, %rd5746; + xor.b64 %rd5822, %rd5749, %rd25; + xor.b64 %rd5823, %rd5764, %rd5822; + xor.b64 %rd5824, %rd5823, %rd5779; + xor.b64 %rd5825, %rd5824, %rd5794; + xor.b64 %rd5826, %rd5825, %rd5809; + xor.b64 %rd5827, %rd5767, %rd5752; + xor.b64 %rd5828, %rd5827, %rd5782; + xor.b64 %rd5829, %rd5828, %rd5797; + xor.b64 %rd5830, %rd5829, %rd5812; + xor.b64 %rd5831, %rd5770, %rd5755; + xor.b64 %rd5832, %rd5831, %rd5785; + xor.b64 %rd5833, %rd5832, %rd5800; + xor.b64 %rd5834, %rd5833, %rd5815; + xor.b64 %rd5835, %rd5773, %rd5758; + xor.b64 %rd5836, %rd5835, %rd5788; + xor.b64 %rd5837, %rd5836, %rd5803; + xor.b64 %rd5838, %rd5837, %rd5818; + xor.b64 %rd5839, %rd5776, %rd5761; + xor.b64 %rd5840, %rd5839, %rd5791; + xor.b64 %rd5841, %rd5840, %rd5806; + xor.b64 %rd5842, %rd5841, %rd5821; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4022}, %rd5830; + } + { + .reg .b32 %dummy; + mov.b64 {%r4023,%dummy}, %rd5830; + } + shf.l.wrap.b32 %r4024, %r4023, %r4022, 1; + shf.l.wrap.b32 %r4025, %r4022, %r4023, 1; + mov.b64 %rd5843, {%r4025, %r4024}; + xor.b64 %rd5844, %rd5842, %rd5843; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4026}, %rd5834; + } + { + .reg .b32 %dummy; + mov.b64 {%r4027,%dummy}, %rd5834; + } + shf.l.wrap.b32 %r4028, %r4027, %r4026, 1; + shf.l.wrap.b32 %r4029, %r4026, %r4027, 1; + mov.b64 %rd5845, {%r4029, %r4028}; + xor.b64 %rd5846, %rd5845, %rd5826; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4030}, %rd5838; + } + { + .reg .b32 %dummy; + mov.b64 {%r4031,%dummy}, %rd5838; + } + shf.l.wrap.b32 %r4032, %r4031, %r4030, 1; + shf.l.wrap.b32 %r4033, %r4030, %r4031, 1; + mov.b64 %rd5847, {%r4033, %r4032}; + xor.b64 %rd5848, %rd5847, %rd5830; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4034}, %rd5842; + } + { + .reg .b32 %dummy; + mov.b64 {%r4035,%dummy}, %rd5842; + } + shf.l.wrap.b32 %r4036, %r4035, %r4034, 1; + shf.l.wrap.b32 %r4037, %r4034, %r4035, 1; + mov.b64 %rd5849, {%r4037, %r4036}; + xor.b64 %rd5850, %rd5849, %rd5834; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4038}, %rd5826; + } + { + .reg .b32 %dummy; + mov.b64 {%r4039,%dummy}, %rd5826; + } + shf.l.wrap.b32 %r4040, %r4039, %r4038, 1; + shf.l.wrap.b32 %r4041, %r4038, %r4039, 1; + mov.b64 %rd5851, {%r4041, %r4040}; + xor.b64 %rd5852, %rd5838, %rd5851; + xor.b64 %rd5853, %rd5844, %rd5822; + xor.b64 %rd5854, %rd5846, %rd5752; + xor.b64 %rd5855, %rd5848, %rd5755; + xor.b64 %rd5856, %rd5850, %rd5758; + xor.b64 %rd5857, %rd5852, %rd5761; + xor.b64 %rd5858, %rd5844, %rd5764; + xor.b64 %rd5859, %rd5846, %rd5767; + xor.b64 %rd5860, %rd5848, %rd5770; + xor.b64 %rd5861, %rd5850, %rd5773; + xor.b64 %rd5862, %rd5852, %rd5776; + xor.b64 %rd5863, %rd5844, %rd5779; + xor.b64 %rd5864, %rd5846, %rd5782; + xor.b64 %rd5865, %rd5848, %rd5785; + xor.b64 %rd5866, %rd5850, %rd5788; + xor.b64 %rd5867, %rd5852, %rd5791; + xor.b64 %rd5868, %rd5844, %rd5794; + xor.b64 %rd5869, %rd5846, %rd5797; + xor.b64 %rd5870, %rd5848, %rd5800; + xor.b64 %rd5871, %rd5850, %rd5803; + xor.b64 %rd5872, %rd5852, %rd5806; + xor.b64 %rd5873, %rd5844, %rd5809; + xor.b64 %rd5874, %rd5846, %rd5812; + xor.b64 %rd5875, %rd5848, %rd5815; + xor.b64 %rd5876, %rd5850, %rd5818; + xor.b64 %rd5877, %rd5852, %rd5821; + { + .reg .b32 %dummy; + mov.b64 {%r4042,%dummy}, %rd5859; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4043}, %rd5859; + } + shf.r.wrap.b32 %r4044, %r4043, %r4042, 20; + shf.r.wrap.b32 %r4045, %r4042, %r4043, 20; + mov.b64 %rd5878, {%r4045, %r4044}; + { + .reg .b32 %dummy; + mov.b64 {%r4046,%dummy}, %rd5865; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4047}, %rd5865; + } + shf.r.wrap.b32 %r4048, %r4047, %r4046, 21; + shf.r.wrap.b32 %r4049, %r4046, %r4047, 21; + mov.b64 %rd5879, {%r4049, %r4048}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4050}, %rd5871; + } + { + .reg .b32 %dummy; + mov.b64 {%r4051,%dummy}, %rd5871; + } + shf.l.wrap.b32 %r4052, %r4051, %r4050, 21; + shf.l.wrap.b32 %r4053, %r4050, %r4051, 21; + mov.b64 %rd5880, {%r4053, %r4052}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4054}, %rd5877; + } + { + .reg .b32 %dummy; + mov.b64 {%r4055,%dummy}, %rd5877; + } + shf.l.wrap.b32 %r4056, %r4055, %r4054, 14; + shf.l.wrap.b32 %r4057, %r4054, %r4055, 14; + mov.b64 %rd5881, {%r4057, %r4056}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4058}, %rd5856; + } + { + .reg .b32 %dummy; + mov.b64 {%r4059,%dummy}, %rd5856; + } + shf.l.wrap.b32 %r4060, %r4059, %r4058, 28; + shf.l.wrap.b32 %r4061, %r4058, %r4059, 28; + mov.b64 %rd5882, {%r4061, %r4060}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4062}, %rd5862; + } + { + .reg .b32 %dummy; + mov.b64 {%r4063,%dummy}, %rd5862; + } + shf.l.wrap.b32 %r4064, %r4063, %r4062, 20; + shf.l.wrap.b32 %r4065, %r4062, %r4063, 20; + mov.b64 %rd5883, {%r4065, %r4064}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4066}, %rd5863; + } + { + .reg .b32 %dummy; + mov.b64 {%r4067,%dummy}, %rd5863; + } + shf.l.wrap.b32 %r4068, %r4067, %r4066, 3; + shf.l.wrap.b32 %r4069, %r4066, %r4067, 3; + mov.b64 %rd5884, {%r4069, %r4068}; + { + .reg .b32 %dummy; + mov.b64 {%r4070,%dummy}, %rd5869; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4071}, %rd5869; + } + shf.r.wrap.b32 %r4072, %r4071, %r4070, 19; + shf.r.wrap.b32 %r4073, %r4070, %r4071, 19; + mov.b64 %rd5885, {%r4073, %r4072}; + { + .reg .b32 %dummy; + mov.b64 {%r4074,%dummy}, %rd5875; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4075}, %rd5875; + } + shf.r.wrap.b32 %r4076, %r4075, %r4074, 3; + shf.r.wrap.b32 %r4077, %r4074, %r4075, 3; + mov.b64 %rd5886, {%r4077, %r4076}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4078}, %rd5854; + } + { + .reg .b32 %dummy; + mov.b64 {%r4079,%dummy}, %rd5854; + } + shf.l.wrap.b32 %r4080, %r4079, %r4078, 1; + shf.l.wrap.b32 %r4081, %r4078, %r4079, 1; + mov.b64 %rd5887, {%r4081, %r4080}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4082}, %rd5860; + } + { + .reg .b32 %dummy; + mov.b64 {%r4083,%dummy}, %rd5860; + } + shf.l.wrap.b32 %r4084, %r4083, %r4082, 6; + shf.l.wrap.b32 %r4085, %r4082, %r4083, 6; + mov.b64 %rd5888, {%r4085, %r4084}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4086}, %rd5866; + } + { + .reg .b32 %dummy; + mov.b64 {%r4087,%dummy}, %rd5866; + } + shf.l.wrap.b32 %r4088, %r4087, %r4086, 25; + shf.l.wrap.b32 %r4089, %r4086, %r4087, 25; + mov.b64 %rd5889, {%r4089, %r4088}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4090}, %rd5872; + } + { + .reg .b32 %dummy; + mov.b64 {%r4091,%dummy}, %rd5872; + } + shf.l.wrap.b32 %r4092, %r4091, %r4090, 8; + shf.l.wrap.b32 %r4093, %r4090, %r4091, 8; + mov.b64 %rd5890, {%r4093, %r4092}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4094}, %rd5873; + } + { + .reg .b32 %dummy; + mov.b64 {%r4095,%dummy}, %rd5873; + } + shf.l.wrap.b32 %r4096, %r4095, %r4094, 18; + shf.l.wrap.b32 %r4097, %r4094, %r4095, 18; + mov.b64 %rd5891, {%r4097, %r4096}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4098}, %rd5857; + } + { + .reg .b32 %dummy; + mov.b64 {%r4099,%dummy}, %rd5857; + } + shf.l.wrap.b32 %r4100, %r4099, %r4098, 27; + shf.l.wrap.b32 %r4101, %r4098, %r4099, 27; + mov.b64 %rd5892, {%r4101, %r4100}; + { + .reg .b32 %dummy; + mov.b64 {%r4102,%dummy}, %rd5858; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4103}, %rd5858; + } + shf.r.wrap.b32 %r4104, %r4103, %r4102, 28; + shf.r.wrap.b32 %r4105, %r4102, %r4103, 28; + mov.b64 %rd5893, {%r4105, %r4104}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4106}, %rd5864; + } + { + .reg .b32 %dummy; + mov.b64 {%r4107,%dummy}, %rd5864; + } + shf.l.wrap.b32 %r4108, %r4107, %r4106, 10; + shf.l.wrap.b32 %r4109, %r4106, %r4107, 10; + mov.b64 %rd5894, {%r4109, %r4108}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4110}, %rd5870; + } + { + .reg .b32 %dummy; + mov.b64 {%r4111,%dummy}, %rd5870; + } + shf.l.wrap.b32 %r4112, %r4111, %r4110, 15; + shf.l.wrap.b32 %r4113, %r4110, %r4111, 15; + mov.b64 %rd5895, {%r4113, %r4112}; + { + .reg .b32 %dummy; + mov.b64 {%r4114,%dummy}, %rd5876; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4115}, %rd5876; + } + shf.r.wrap.b32 %r4116, %r4115, %r4114, 8; + shf.r.wrap.b32 %r4117, %r4114, %r4115, 8; + mov.b64 %rd5896, {%r4117, %r4116}; + { + .reg .b32 %dummy; + mov.b64 {%r4118,%dummy}, %rd5855; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4119}, %rd5855; + } + shf.r.wrap.b32 %r4120, %r4119, %r4118, 2; + shf.r.wrap.b32 %r4121, %r4118, %r4119, 2; + mov.b64 %rd5897, {%r4121, %r4120}; + { + .reg .b32 %dummy; + mov.b64 {%r4122,%dummy}, %rd5861; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4123}, %rd5861; + } + shf.r.wrap.b32 %r4124, %r4123, %r4122, 9; + shf.r.wrap.b32 %r4125, %r4122, %r4123, 9; + mov.b64 %rd5898, {%r4125, %r4124}; + { + .reg .b32 %dummy; + mov.b64 {%r4126,%dummy}, %rd5867; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4127}, %rd5867; + } + shf.r.wrap.b32 %r4128, %r4127, %r4126, 25; + shf.r.wrap.b32 %r4129, %r4126, %r4127, 25; + mov.b64 %rd5899, {%r4129, %r4128}; + { + .reg .b32 %dummy; + mov.b64 {%r4130,%dummy}, %rd5868; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4131}, %rd5868; + } + shf.r.wrap.b32 %r4132, %r4131, %r4130, 23; + shf.r.wrap.b32 %r4133, %r4130, %r4131, 23; + mov.b64 %rd5900, {%r4133, %r4132}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4134}, %rd5874; + } + { + .reg .b32 %dummy; + mov.b64 {%r4135,%dummy}, %rd5874; + } + shf.l.wrap.b32 %r4136, %r4135, %r4134, 2; + shf.l.wrap.b32 %r4137, %r4134, %r4135, 2; + mov.b64 %rd5901, {%r4137, %r4136}; + not.b64 %rd5902, %rd5878; + and.b64 %rd5903, %rd5879, %rd5902; + xor.b64 %rd5904, %rd5903, %rd5853; + not.b64 %rd5905, %rd5879; + and.b64 %rd5906, %rd5880, %rd5905; + xor.b64 %rd5907, %rd5906, %rd5878; + not.b64 %rd5908, %rd5880; + and.b64 %rd5909, %rd5881, %rd5908; + xor.b64 %rd5910, %rd5909, %rd5879; + not.b64 %rd5911, %rd5881; + and.b64 %rd5912, %rd5853, %rd5911; + xor.b64 %rd5913, %rd5912, %rd5880; + not.b64 %rd5914, %rd5853; + and.b64 %rd5915, %rd5878, %rd5914; + xor.b64 %rd5916, %rd5915, %rd5881; + not.b64 %rd5917, %rd5883; + and.b64 %rd5918, %rd5884, %rd5917; + xor.b64 %rd5919, %rd5918, %rd5882; + not.b64 %rd5920, %rd5884; + and.b64 %rd5921, %rd5885, %rd5920; + xor.b64 %rd5922, %rd5921, %rd5883; + not.b64 %rd5923, %rd5885; + and.b64 %rd5924, %rd5886, %rd5923; + xor.b64 %rd5925, %rd5924, %rd5884; + not.b64 %rd5926, %rd5886; + and.b64 %rd5927, %rd5882, %rd5926; + xor.b64 %rd5928, %rd5927, %rd5885; + not.b64 %rd5929, %rd5882; + and.b64 %rd5930, %rd5883, %rd5929; + xor.b64 %rd5931, %rd5930, %rd5886; + not.b64 %rd5932, %rd5888; + and.b64 %rd5933, %rd5889, %rd5932; + xor.b64 %rd5934, %rd5933, %rd5887; + not.b64 %rd5935, %rd5889; + and.b64 %rd5936, %rd5890, %rd5935; + xor.b64 %rd5937, %rd5936, %rd5888; + not.b64 %rd5938, %rd5890; + and.b64 %rd5939, %rd5891, %rd5938; + xor.b64 %rd5940, %rd5939, %rd5889; + not.b64 %rd5941, %rd5891; + and.b64 %rd5942, %rd5887, %rd5941; + xor.b64 %rd5943, %rd5942, %rd5890; + not.b64 %rd5944, %rd5887; + and.b64 %rd5945, %rd5888, %rd5944; + xor.b64 %rd5946, %rd5945, %rd5891; + not.b64 %rd5947, %rd5893; + and.b64 %rd5948, %rd5894, %rd5947; + xor.b64 %rd5949, %rd5948, %rd5892; + not.b64 %rd5950, %rd5894; + and.b64 %rd5951, %rd5895, %rd5950; + xor.b64 %rd5952, %rd5951, %rd5893; + not.b64 %rd5953, %rd5895; + and.b64 %rd5954, %rd5896, %rd5953; + xor.b64 %rd5955, %rd5954, %rd5894; + not.b64 %rd5956, %rd5896; + and.b64 %rd5957, %rd5892, %rd5956; + xor.b64 %rd5958, %rd5957, %rd5895; + not.b64 %rd5959, %rd5892; + and.b64 %rd5960, %rd5893, %rd5959; + xor.b64 %rd5961, %rd5960, %rd5896; + not.b64 %rd5962, %rd5898; + and.b64 %rd5963, %rd5899, %rd5962; + xor.b64 %rd5964, %rd5963, %rd5897; + not.b64 %rd5965, %rd5899; + and.b64 %rd5966, %rd5900, %rd5965; + xor.b64 %rd5967, %rd5966, %rd5898; + not.b64 %rd5968, %rd5900; + and.b64 %rd5969, %rd5901, %rd5968; + xor.b64 %rd5970, %rd5969, %rd5899; + not.b64 %rd5971, %rd5901; + and.b64 %rd5972, %rd5897, %rd5971; + xor.b64 %rd5973, %rd5972, %rd5900; + not.b64 %rd5974, %rd5897; + and.b64 %rd5975, %rd5898, %rd5974; + xor.b64 %rd5976, %rd5975, %rd5901; + xor.b64 %rd5977, %rd5904, %rd26; + xor.b64 %rd5978, %rd5919, %rd5977; + xor.b64 %rd5979, %rd5978, %rd5934; + xor.b64 %rd5980, %rd5979, %rd5949; + xor.b64 %rd5981, %rd5980, %rd5964; + xor.b64 %rd5982, %rd5922, %rd5907; + xor.b64 %rd5983, %rd5982, %rd5937; + xor.b64 %rd5984, %rd5983, %rd5952; + xor.b64 %rd5985, %rd5984, %rd5967; + xor.b64 %rd5986, %rd5925, %rd5910; + xor.b64 %rd5987, %rd5986, %rd5940; + xor.b64 %rd5988, %rd5987, %rd5955; + xor.b64 %rd5989, %rd5988, %rd5970; + xor.b64 %rd5990, %rd5928, %rd5913; + xor.b64 %rd5991, %rd5990, %rd5943; + xor.b64 %rd5992, %rd5991, %rd5958; + xor.b64 %rd5993, %rd5992, %rd5973; + xor.b64 %rd5994, %rd5931, %rd5916; + xor.b64 %rd5995, %rd5994, %rd5946; + xor.b64 %rd5996, %rd5995, %rd5961; + xor.b64 %rd5997, %rd5996, %rd5976; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4138}, %rd5985; + } + { + .reg .b32 %dummy; + mov.b64 {%r4139,%dummy}, %rd5985; + } + shf.l.wrap.b32 %r4140, %r4139, %r4138, 1; + shf.l.wrap.b32 %r4141, %r4138, %r4139, 1; + mov.b64 %rd5998, {%r4141, %r4140}; + xor.b64 %rd5999, %rd5997, %rd5998; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4142}, %rd5989; + } + { + .reg .b32 %dummy; + mov.b64 {%r4143,%dummy}, %rd5989; + } + shf.l.wrap.b32 %r4144, %r4143, %r4142, 1; + shf.l.wrap.b32 %r4145, %r4142, %r4143, 1; + mov.b64 %rd6000, {%r4145, %r4144}; + xor.b64 %rd6001, %rd6000, %rd5981; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4146}, %rd5993; + } + { + .reg .b32 %dummy; + mov.b64 {%r4147,%dummy}, %rd5993; + } + shf.l.wrap.b32 %r4148, %r4147, %r4146, 1; + shf.l.wrap.b32 %r4149, %r4146, %r4147, 1; + mov.b64 %rd6002, {%r4149, %r4148}; + xor.b64 %rd6003, %rd6002, %rd5985; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4150}, %rd5997; + } + { + .reg .b32 %dummy; + mov.b64 {%r4151,%dummy}, %rd5997; + } + shf.l.wrap.b32 %r4152, %r4151, %r4150, 1; + shf.l.wrap.b32 %r4153, %r4150, %r4151, 1; + mov.b64 %rd6004, {%r4153, %r4152}; + xor.b64 %rd6005, %rd6004, %rd5989; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4154}, %rd5981; + } + { + .reg .b32 %dummy; + mov.b64 {%r4155,%dummy}, %rd5981; + } + shf.l.wrap.b32 %r4156, %r4155, %r4154, 1; + shf.l.wrap.b32 %r4157, %r4154, %r4155, 1; + mov.b64 %rd6006, {%r4157, %r4156}; + xor.b64 %rd6007, %rd5993, %rd6006; + xor.b64 %rd6008, %rd5999, %rd5977; + xor.b64 %rd6009, %rd6001, %rd5907; + xor.b64 %rd6010, %rd6003, %rd5910; + xor.b64 %rd6011, %rd6005, %rd5913; + xor.b64 %rd6012, %rd6007, %rd5916; + xor.b64 %rd6013, %rd5999, %rd5919; + xor.b64 %rd6014, %rd6001, %rd5922; + xor.b64 %rd6015, %rd6003, %rd5925; + xor.b64 %rd6016, %rd6005, %rd5928; + xor.b64 %rd6017, %rd6007, %rd5931; + xor.b64 %rd6018, %rd5999, %rd5934; + xor.b64 %rd6019, %rd6001, %rd5937; + xor.b64 %rd6020, %rd6003, %rd5940; + xor.b64 %rd6021, %rd6005, %rd5943; + xor.b64 %rd6022, %rd6007, %rd5946; + xor.b64 %rd6023, %rd5999, %rd5949; + xor.b64 %rd6024, %rd6001, %rd5952; + xor.b64 %rd6025, %rd6003, %rd5955; + xor.b64 %rd6026, %rd6005, %rd5958; + xor.b64 %rd6027, %rd6007, %rd5961; + xor.b64 %rd6028, %rd5999, %rd5964; + xor.b64 %rd6029, %rd6001, %rd5967; + xor.b64 %rd6030, %rd6003, %rd5970; + xor.b64 %rd6031, %rd6005, %rd5973; + xor.b64 %rd6032, %rd6007, %rd5976; + { + .reg .b32 %dummy; + mov.b64 {%r4158,%dummy}, %rd6014; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4159}, %rd6014; + } + shf.r.wrap.b32 %r4160, %r4159, %r4158, 20; + shf.r.wrap.b32 %r4161, %r4158, %r4159, 20; + mov.b64 %rd6033, {%r4161, %r4160}; + { + .reg .b32 %dummy; + mov.b64 {%r4162,%dummy}, %rd6020; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4163}, %rd6020; + } + shf.r.wrap.b32 %r4164, %r4163, %r4162, 21; + shf.r.wrap.b32 %r4165, %r4162, %r4163, 21; + mov.b64 %rd6034, {%r4165, %r4164}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4166}, %rd6026; + } + { + .reg .b32 %dummy; + mov.b64 {%r4167,%dummy}, %rd6026; + } + shf.l.wrap.b32 %r4168, %r4167, %r4166, 21; + shf.l.wrap.b32 %r4169, %r4166, %r4167, 21; + mov.b64 %rd6035, {%r4169, %r4168}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4170}, %rd6032; + } + { + .reg .b32 %dummy; + mov.b64 {%r4171,%dummy}, %rd6032; + } + shf.l.wrap.b32 %r4172, %r4171, %r4170, 14; + shf.l.wrap.b32 %r4173, %r4170, %r4171, 14; + mov.b64 %rd6036, {%r4173, %r4172}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4174}, %rd6011; + } + { + .reg .b32 %dummy; + mov.b64 {%r4175,%dummy}, %rd6011; + } + shf.l.wrap.b32 %r4176, %r4175, %r4174, 28; + shf.l.wrap.b32 %r4177, %r4174, %r4175, 28; + mov.b64 %rd6037, {%r4177, %r4176}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4178}, %rd6017; + } + { + .reg .b32 %dummy; + mov.b64 {%r4179,%dummy}, %rd6017; + } + shf.l.wrap.b32 %r4180, %r4179, %r4178, 20; + shf.l.wrap.b32 %r4181, %r4178, %r4179, 20; + mov.b64 %rd6038, {%r4181, %r4180}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4182}, %rd6018; + } + { + .reg .b32 %dummy; + mov.b64 {%r4183,%dummy}, %rd6018; + } + shf.l.wrap.b32 %r4184, %r4183, %r4182, 3; + shf.l.wrap.b32 %r4185, %r4182, %r4183, 3; + mov.b64 %rd6039, {%r4185, %r4184}; + { + .reg .b32 %dummy; + mov.b64 {%r4186,%dummy}, %rd6024; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4187}, %rd6024; + } + shf.r.wrap.b32 %r4188, %r4187, %r4186, 19; + shf.r.wrap.b32 %r4189, %r4186, %r4187, 19; + mov.b64 %rd6040, {%r4189, %r4188}; + { + .reg .b32 %dummy; + mov.b64 {%r4190,%dummy}, %rd6030; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4191}, %rd6030; + } + shf.r.wrap.b32 %r4192, %r4191, %r4190, 3; + shf.r.wrap.b32 %r4193, %r4190, %r4191, 3; + mov.b64 %rd6041, {%r4193, %r4192}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4194}, %rd6009; + } + { + .reg .b32 %dummy; + mov.b64 {%r4195,%dummy}, %rd6009; + } + shf.l.wrap.b32 %r4196, %r4195, %r4194, 1; + shf.l.wrap.b32 %r4197, %r4194, %r4195, 1; + mov.b64 %rd6042, {%r4197, %r4196}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4198}, %rd6015; + } + { + .reg .b32 %dummy; + mov.b64 {%r4199,%dummy}, %rd6015; + } + shf.l.wrap.b32 %r4200, %r4199, %r4198, 6; + shf.l.wrap.b32 %r4201, %r4198, %r4199, 6; + mov.b64 %rd6043, {%r4201, %r4200}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4202}, %rd6021; + } + { + .reg .b32 %dummy; + mov.b64 {%r4203,%dummy}, %rd6021; + } + shf.l.wrap.b32 %r4204, %r4203, %r4202, 25; + shf.l.wrap.b32 %r4205, %r4202, %r4203, 25; + mov.b64 %rd6044, {%r4205, %r4204}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4206}, %rd6027; + } + { + .reg .b32 %dummy; + mov.b64 {%r4207,%dummy}, %rd6027; + } + shf.l.wrap.b32 %r4208, %r4207, %r4206, 8; + shf.l.wrap.b32 %r4209, %r4206, %r4207, 8; + mov.b64 %rd6045, {%r4209, %r4208}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4210}, %rd6028; + } + { + .reg .b32 %dummy; + mov.b64 {%r4211,%dummy}, %rd6028; + } + shf.l.wrap.b32 %r4212, %r4211, %r4210, 18; + shf.l.wrap.b32 %r4213, %r4210, %r4211, 18; + mov.b64 %rd6046, {%r4213, %r4212}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4214}, %rd6012; + } + { + .reg .b32 %dummy; + mov.b64 {%r4215,%dummy}, %rd6012; + } + shf.l.wrap.b32 %r4216, %r4215, %r4214, 27; + shf.l.wrap.b32 %r4217, %r4214, %r4215, 27; + mov.b64 %rd6047, {%r4217, %r4216}; + { + .reg .b32 %dummy; + mov.b64 {%r4218,%dummy}, %rd6013; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4219}, %rd6013; + } + shf.r.wrap.b32 %r4220, %r4219, %r4218, 28; + shf.r.wrap.b32 %r4221, %r4218, %r4219, 28; + mov.b64 %rd6048, {%r4221, %r4220}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4222}, %rd6019; + } + { + .reg .b32 %dummy; + mov.b64 {%r4223,%dummy}, %rd6019; + } + shf.l.wrap.b32 %r4224, %r4223, %r4222, 10; + shf.l.wrap.b32 %r4225, %r4222, %r4223, 10; + mov.b64 %rd6049, {%r4225, %r4224}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4226}, %rd6025; + } + { + .reg .b32 %dummy; + mov.b64 {%r4227,%dummy}, %rd6025; + } + shf.l.wrap.b32 %r4228, %r4227, %r4226, 15; + shf.l.wrap.b32 %r4229, %r4226, %r4227, 15; + mov.b64 %rd6050, {%r4229, %r4228}; + { + .reg .b32 %dummy; + mov.b64 {%r4230,%dummy}, %rd6031; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4231}, %rd6031; + } + shf.r.wrap.b32 %r4232, %r4231, %r4230, 8; + shf.r.wrap.b32 %r4233, %r4230, %r4231, 8; + mov.b64 %rd6051, {%r4233, %r4232}; + { + .reg .b32 %dummy; + mov.b64 {%r4234,%dummy}, %rd6010; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4235}, %rd6010; + } + shf.r.wrap.b32 %r4236, %r4235, %r4234, 2; + shf.r.wrap.b32 %r4237, %r4234, %r4235, 2; + mov.b64 %rd6052, {%r4237, %r4236}; + { + .reg .b32 %dummy; + mov.b64 {%r4238,%dummy}, %rd6016; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4239}, %rd6016; + } + shf.r.wrap.b32 %r4240, %r4239, %r4238, 9; + shf.r.wrap.b32 %r4241, %r4238, %r4239, 9; + mov.b64 %rd6053, {%r4241, %r4240}; + { + .reg .b32 %dummy; + mov.b64 {%r4242,%dummy}, %rd6022; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4243}, %rd6022; + } + shf.r.wrap.b32 %r4244, %r4243, %r4242, 25; + shf.r.wrap.b32 %r4245, %r4242, %r4243, 25; + mov.b64 %rd6054, {%r4245, %r4244}; + { + .reg .b32 %dummy; + mov.b64 {%r4246,%dummy}, %rd6023; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4247}, %rd6023; + } + shf.r.wrap.b32 %r4248, %r4247, %r4246, 23; + shf.r.wrap.b32 %r4249, %r4246, %r4247, 23; + mov.b64 %rd6055, {%r4249, %r4248}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4250}, %rd6029; + } + { + .reg .b32 %dummy; + mov.b64 {%r4251,%dummy}, %rd6029; + } + shf.l.wrap.b32 %r4252, %r4251, %r4250, 2; + shf.l.wrap.b32 %r4253, %r4250, %r4251, 2; + mov.b64 %rd6056, {%r4253, %r4252}; + not.b64 %rd6057, %rd6033; + and.b64 %rd6058, %rd6034, %rd6057; + xor.b64 %rd6059, %rd6058, %rd6008; + not.b64 %rd6060, %rd6034; + and.b64 %rd6061, %rd6035, %rd6060; + xor.b64 %rd6062, %rd6061, %rd6033; + not.b64 %rd6063, %rd6035; + and.b64 %rd6064, %rd6036, %rd6063; + xor.b64 %rd6065, %rd6064, %rd6034; + not.b64 %rd6066, %rd6036; + and.b64 %rd6067, %rd6008, %rd6066; + xor.b64 %rd6068, %rd6067, %rd6035; + not.b64 %rd6069, %rd6008; + and.b64 %rd6070, %rd6033, %rd6069; + xor.b64 %rd6071, %rd6070, %rd6036; + not.b64 %rd6072, %rd6038; + and.b64 %rd6073, %rd6039, %rd6072; + xor.b64 %rd6074, %rd6073, %rd6037; + not.b64 %rd6075, %rd6039; + and.b64 %rd6076, %rd6040, %rd6075; + xor.b64 %rd6077, %rd6076, %rd6038; + not.b64 %rd6078, %rd6040; + and.b64 %rd6079, %rd6041, %rd6078; + xor.b64 %rd6080, %rd6079, %rd6039; + not.b64 %rd6081, %rd6041; + and.b64 %rd6082, %rd6037, %rd6081; + xor.b64 %rd6083, %rd6082, %rd6040; + not.b64 %rd6084, %rd6037; + and.b64 %rd6085, %rd6038, %rd6084; + xor.b64 %rd6086, %rd6085, %rd6041; + not.b64 %rd6087, %rd6043; + and.b64 %rd6088, %rd6044, %rd6087; + xor.b64 %rd6089, %rd6088, %rd6042; + not.b64 %rd6090, %rd6044; + and.b64 %rd6091, %rd6045, %rd6090; + xor.b64 %rd6092, %rd6091, %rd6043; + not.b64 %rd6093, %rd6045; + and.b64 %rd6094, %rd6046, %rd6093; + xor.b64 %rd6095, %rd6094, %rd6044; + not.b64 %rd6096, %rd6046; + and.b64 %rd6097, %rd6042, %rd6096; + xor.b64 %rd6098, %rd6097, %rd6045; + not.b64 %rd6099, %rd6042; + and.b64 %rd6100, %rd6043, %rd6099; + xor.b64 %rd6101, %rd6100, %rd6046; + not.b64 %rd6102, %rd6048; + and.b64 %rd6103, %rd6049, %rd6102; + xor.b64 %rd6104, %rd6103, %rd6047; + not.b64 %rd6105, %rd6049; + and.b64 %rd6106, %rd6050, %rd6105; + xor.b64 %rd6107, %rd6106, %rd6048; + not.b64 %rd6108, %rd6050; + and.b64 %rd6109, %rd6051, %rd6108; + xor.b64 %rd6110, %rd6109, %rd6049; + not.b64 %rd6111, %rd6051; + and.b64 %rd6112, %rd6047, %rd6111; + xor.b64 %rd6113, %rd6112, %rd6050; + not.b64 %rd6114, %rd6047; + and.b64 %rd6115, %rd6048, %rd6114; + xor.b64 %rd6116, %rd6115, %rd6051; + not.b64 %rd6117, %rd6053; + and.b64 %rd6118, %rd6054, %rd6117; + xor.b64 %rd6119, %rd6118, %rd6052; + not.b64 %rd6120, %rd6054; + and.b64 %rd6121, %rd6055, %rd6120; + xor.b64 %rd6122, %rd6121, %rd6053; + not.b64 %rd6123, %rd6055; + and.b64 %rd6124, %rd6056, %rd6123; + xor.b64 %rd6125, %rd6124, %rd6054; + not.b64 %rd6126, %rd6056; + and.b64 %rd6127, %rd6052, %rd6126; + xor.b64 %rd6128, %rd6127, %rd6055; + not.b64 %rd6129, %rd6052; + and.b64 %rd6130, %rd6053, %rd6129; + xor.b64 %rd6131, %rd6130, %rd6056; + xor.b64 %rd6132, %rd6059, %rd27; + xor.b64 %rd6133, %rd6074, %rd6132; + xor.b64 %rd6134, %rd6133, %rd6089; + xor.b64 %rd6135, %rd6134, %rd6104; + xor.b64 %rd6136, %rd6135, %rd6119; + xor.b64 %rd6137, %rd6077, %rd6062; + xor.b64 %rd6138, %rd6137, %rd6092; + xor.b64 %rd6139, %rd6138, %rd6107; + xor.b64 %rd6140, %rd6139, %rd6122; + xor.b64 %rd6141, %rd6080, %rd6065; + xor.b64 %rd6142, %rd6141, %rd6095; + xor.b64 %rd6143, %rd6142, %rd6110; + xor.b64 %rd6144, %rd6143, %rd6125; + xor.b64 %rd6145, %rd6083, %rd6068; + xor.b64 %rd6146, %rd6145, %rd6098; + xor.b64 %rd6147, %rd6146, %rd6113; + xor.b64 %rd6148, %rd6147, %rd6128; + xor.b64 %rd6149, %rd6086, %rd6071; + xor.b64 %rd6150, %rd6149, %rd6101; + xor.b64 %rd6151, %rd6150, %rd6116; + xor.b64 %rd6152, %rd6151, %rd6131; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4254}, %rd6140; + } + { + .reg .b32 %dummy; + mov.b64 {%r4255,%dummy}, %rd6140; + } + shf.l.wrap.b32 %r4256, %r4255, %r4254, 1; + shf.l.wrap.b32 %r4257, %r4254, %r4255, 1; + mov.b64 %rd6153, {%r4257, %r4256}; + xor.b64 %rd6154, %rd6152, %rd6153; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4258}, %rd6144; + } + { + .reg .b32 %dummy; + mov.b64 {%r4259,%dummy}, %rd6144; + } + shf.l.wrap.b32 %r4260, %r4259, %r4258, 1; + shf.l.wrap.b32 %r4261, %r4258, %r4259, 1; + mov.b64 %rd6155, {%r4261, %r4260}; + xor.b64 %rd6156, %rd6155, %rd6136; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4262}, %rd6148; + } + { + .reg .b32 %dummy; + mov.b64 {%r4263,%dummy}, %rd6148; + } + shf.l.wrap.b32 %r4264, %r4263, %r4262, 1; + shf.l.wrap.b32 %r4265, %r4262, %r4263, 1; + mov.b64 %rd6157, {%r4265, %r4264}; + xor.b64 %rd6158, %rd6157, %rd6140; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4266}, %rd6152; + } + { + .reg .b32 %dummy; + mov.b64 {%r4267,%dummy}, %rd6152; + } + shf.l.wrap.b32 %r4268, %r4267, %r4266, 1; + shf.l.wrap.b32 %r4269, %r4266, %r4267, 1; + mov.b64 %rd6159, {%r4269, %r4268}; + xor.b64 %rd6160, %rd6159, %rd6144; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4270}, %rd6136; + } + { + .reg .b32 %dummy; + mov.b64 {%r4271,%dummy}, %rd6136; + } + shf.l.wrap.b32 %r4272, %r4271, %r4270, 1; + shf.l.wrap.b32 %r4273, %r4270, %r4271, 1; + mov.b64 %rd6161, {%r4273, %r4272}; + xor.b64 %rd6162, %rd6148, %rd6161; + xor.b64 %rd6163, %rd6154, %rd6132; + xor.b64 %rd6164, %rd6156, %rd6062; + xor.b64 %rd6165, %rd6158, %rd6065; + xor.b64 %rd6166, %rd6160, %rd6068; + xor.b64 %rd6167, %rd6162, %rd6071; + xor.b64 %rd6168, %rd6154, %rd6074; + xor.b64 %rd6169, %rd6156, %rd6077; + xor.b64 %rd6170, %rd6158, %rd6080; + xor.b64 %rd6171, %rd6160, %rd6083; + xor.b64 %rd6172, %rd6162, %rd6086; + xor.b64 %rd6173, %rd6154, %rd6089; + xor.b64 %rd6174, %rd6156, %rd6092; + xor.b64 %rd6175, %rd6158, %rd6095; + xor.b64 %rd6176, %rd6160, %rd6098; + xor.b64 %rd6177, %rd6162, %rd6101; + xor.b64 %rd6178, %rd6154, %rd6104; + xor.b64 %rd6179, %rd6156, %rd6107; + xor.b64 %rd6180, %rd6158, %rd6110; + xor.b64 %rd6181, %rd6160, %rd6113; + xor.b64 %rd6182, %rd6162, %rd6116; + xor.b64 %rd6183, %rd6154, %rd6119; + xor.b64 %rd6184, %rd6156, %rd6122; + xor.b64 %rd6185, %rd6158, %rd6125; + xor.b64 %rd6186, %rd6160, %rd6128; + xor.b64 %rd6187, %rd6162, %rd6131; + { + .reg .b32 %dummy; + mov.b64 {%r4274,%dummy}, %rd6169; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4275}, %rd6169; + } + shf.r.wrap.b32 %r4276, %r4275, %r4274, 20; + shf.r.wrap.b32 %r4277, %r4274, %r4275, 20; + mov.b64 %rd6188, {%r4277, %r4276}; + { + .reg .b32 %dummy; + mov.b64 {%r4278,%dummy}, %rd6175; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4279}, %rd6175; + } + shf.r.wrap.b32 %r4280, %r4279, %r4278, 21; + shf.r.wrap.b32 %r4281, %r4278, %r4279, 21; + mov.b64 %rd6189, {%r4281, %r4280}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4282}, %rd6181; + } + { + .reg .b32 %dummy; + mov.b64 {%r4283,%dummy}, %rd6181; + } + shf.l.wrap.b32 %r4284, %r4283, %r4282, 21; + shf.l.wrap.b32 %r4285, %r4282, %r4283, 21; + mov.b64 %rd6190, {%r4285, %r4284}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4286}, %rd6187; + } + { + .reg .b32 %dummy; + mov.b64 {%r4287,%dummy}, %rd6187; + } + shf.l.wrap.b32 %r4288, %r4287, %r4286, 14; + shf.l.wrap.b32 %r4289, %r4286, %r4287, 14; + mov.b64 %rd6191, {%r4289, %r4288}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4290}, %rd6166; + } + { + .reg .b32 %dummy; + mov.b64 {%r4291,%dummy}, %rd6166; + } + shf.l.wrap.b32 %r4292, %r4291, %r4290, 28; + shf.l.wrap.b32 %r4293, %r4290, %r4291, 28; + mov.b64 %rd6192, {%r4293, %r4292}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4294}, %rd6172; + } + { + .reg .b32 %dummy; + mov.b64 {%r4295,%dummy}, %rd6172; + } + shf.l.wrap.b32 %r4296, %r4295, %r4294, 20; + shf.l.wrap.b32 %r4297, %r4294, %r4295, 20; + mov.b64 %rd6193, {%r4297, %r4296}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4298}, %rd6173; + } + { + .reg .b32 %dummy; + mov.b64 {%r4299,%dummy}, %rd6173; + } + shf.l.wrap.b32 %r4300, %r4299, %r4298, 3; + shf.l.wrap.b32 %r4301, %r4298, %r4299, 3; + mov.b64 %rd6194, {%r4301, %r4300}; + { + .reg .b32 %dummy; + mov.b64 {%r4302,%dummy}, %rd6179; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4303}, %rd6179; + } + shf.r.wrap.b32 %r4304, %r4303, %r4302, 19; + shf.r.wrap.b32 %r4305, %r4302, %r4303, 19; + mov.b64 %rd6195, {%r4305, %r4304}; + { + .reg .b32 %dummy; + mov.b64 {%r4306,%dummy}, %rd6185; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4307}, %rd6185; + } + shf.r.wrap.b32 %r4308, %r4307, %r4306, 3; + shf.r.wrap.b32 %r4309, %r4306, %r4307, 3; + mov.b64 %rd6196, {%r4309, %r4308}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4310}, %rd6164; + } + { + .reg .b32 %dummy; + mov.b64 {%r4311,%dummy}, %rd6164; + } + shf.l.wrap.b32 %r4312, %r4311, %r4310, 1; + shf.l.wrap.b32 %r4313, %r4310, %r4311, 1; + mov.b64 %rd6197, {%r4313, %r4312}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4314}, %rd6170; + } + { + .reg .b32 %dummy; + mov.b64 {%r4315,%dummy}, %rd6170; + } + shf.l.wrap.b32 %r4316, %r4315, %r4314, 6; + shf.l.wrap.b32 %r4317, %r4314, %r4315, 6; + mov.b64 %rd6198, {%r4317, %r4316}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4318}, %rd6176; + } + { + .reg .b32 %dummy; + mov.b64 {%r4319,%dummy}, %rd6176; + } + shf.l.wrap.b32 %r4320, %r4319, %r4318, 25; + shf.l.wrap.b32 %r4321, %r4318, %r4319, 25; + mov.b64 %rd6199, {%r4321, %r4320}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4322}, %rd6182; + } + { + .reg .b32 %dummy; + mov.b64 {%r4323,%dummy}, %rd6182; + } + shf.l.wrap.b32 %r4324, %r4323, %r4322, 8; + shf.l.wrap.b32 %r4325, %r4322, %r4323, 8; + mov.b64 %rd6200, {%r4325, %r4324}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4326}, %rd6183; + } + { + .reg .b32 %dummy; + mov.b64 {%r4327,%dummy}, %rd6183; + } + shf.l.wrap.b32 %r4328, %r4327, %r4326, 18; + shf.l.wrap.b32 %r4329, %r4326, %r4327, 18; + mov.b64 %rd6201, {%r4329, %r4328}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4330}, %rd6167; + } + { + .reg .b32 %dummy; + mov.b64 {%r4331,%dummy}, %rd6167; + } + shf.l.wrap.b32 %r4332, %r4331, %r4330, 27; + shf.l.wrap.b32 %r4333, %r4330, %r4331, 27; + mov.b64 %rd6202, {%r4333, %r4332}; + { + .reg .b32 %dummy; + mov.b64 {%r4334,%dummy}, %rd6168; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4335}, %rd6168; + } + shf.r.wrap.b32 %r4336, %r4335, %r4334, 28; + shf.r.wrap.b32 %r4337, %r4334, %r4335, 28; + mov.b64 %rd6203, {%r4337, %r4336}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4338}, %rd6174; + } + { + .reg .b32 %dummy; + mov.b64 {%r4339,%dummy}, %rd6174; + } + shf.l.wrap.b32 %r4340, %r4339, %r4338, 10; + shf.l.wrap.b32 %r4341, %r4338, %r4339, 10; + mov.b64 %rd6204, {%r4341, %r4340}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4342}, %rd6180; + } + { + .reg .b32 %dummy; + mov.b64 {%r4343,%dummy}, %rd6180; + } + shf.l.wrap.b32 %r4344, %r4343, %r4342, 15; + shf.l.wrap.b32 %r4345, %r4342, %r4343, 15; + mov.b64 %rd6205, {%r4345, %r4344}; + { + .reg .b32 %dummy; + mov.b64 {%r4346,%dummy}, %rd6186; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4347}, %rd6186; + } + shf.r.wrap.b32 %r4348, %r4347, %r4346, 8; + shf.r.wrap.b32 %r4349, %r4346, %r4347, 8; + mov.b64 %rd6206, {%r4349, %r4348}; + { + .reg .b32 %dummy; + mov.b64 {%r4350,%dummy}, %rd6165; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4351}, %rd6165; + } + shf.r.wrap.b32 %r4352, %r4351, %r4350, 2; + shf.r.wrap.b32 %r4353, %r4350, %r4351, 2; + mov.b64 %rd6207, {%r4353, %r4352}; + { + .reg .b32 %dummy; + mov.b64 {%r4354,%dummy}, %rd6171; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4355}, %rd6171; + } + shf.r.wrap.b32 %r4356, %r4355, %r4354, 9; + shf.r.wrap.b32 %r4357, %r4354, %r4355, 9; + mov.b64 %rd6208, {%r4357, %r4356}; + { + .reg .b32 %dummy; + mov.b64 {%r4358,%dummy}, %rd6177; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4359}, %rd6177; + } + shf.r.wrap.b32 %r4360, %r4359, %r4358, 25; + shf.r.wrap.b32 %r4361, %r4358, %r4359, 25; + mov.b64 %rd6209, {%r4361, %r4360}; + { + .reg .b32 %dummy; + mov.b64 {%r4362,%dummy}, %rd6178; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4363}, %rd6178; + } + shf.r.wrap.b32 %r4364, %r4363, %r4362, 23; + shf.r.wrap.b32 %r4365, %r4362, %r4363, 23; + mov.b64 %rd6210, {%r4365, %r4364}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4366}, %rd6184; + } + { + .reg .b32 %dummy; + mov.b64 {%r4367,%dummy}, %rd6184; + } + shf.l.wrap.b32 %r4368, %r4367, %r4366, 2; + shf.l.wrap.b32 %r4369, %r4366, %r4367, 2; + mov.b64 %rd6211, {%r4369, %r4368}; + not.b64 %rd6212, %rd6188; + and.b64 %rd6213, %rd6189, %rd6212; + xor.b64 %rd6214, %rd6213, %rd6163; + not.b64 %rd6215, %rd6189; + and.b64 %rd6216, %rd6190, %rd6215; + xor.b64 %rd6217, %rd6216, %rd6188; + not.b64 %rd6218, %rd6190; + and.b64 %rd6219, %rd6191, %rd6218; + xor.b64 %rd6220, %rd6219, %rd6189; + not.b64 %rd6221, %rd6191; + and.b64 %rd6222, %rd6163, %rd6221; + xor.b64 %rd6223, %rd6222, %rd6190; + not.b64 %rd6224, %rd6163; + and.b64 %rd6225, %rd6188, %rd6224; + xor.b64 %rd6226, %rd6225, %rd6191; + not.b64 %rd6227, %rd6193; + and.b64 %rd6228, %rd6194, %rd6227; + xor.b64 %rd6229, %rd6228, %rd6192; + not.b64 %rd6230, %rd6194; + and.b64 %rd6231, %rd6195, %rd6230; + xor.b64 %rd6232, %rd6231, %rd6193; + not.b64 %rd6233, %rd6195; + and.b64 %rd6234, %rd6196, %rd6233; + xor.b64 %rd6235, %rd6234, %rd6194; + not.b64 %rd6236, %rd6196; + and.b64 %rd6237, %rd6192, %rd6236; + xor.b64 %rd6238, %rd6237, %rd6195; + not.b64 %rd6239, %rd6192; + and.b64 %rd6240, %rd6193, %rd6239; + xor.b64 %rd6241, %rd6240, %rd6196; + not.b64 %rd6242, %rd6198; + and.b64 %rd6243, %rd6199, %rd6242; + xor.b64 %rd6244, %rd6243, %rd6197; + not.b64 %rd6245, %rd6199; + and.b64 %rd6246, %rd6200, %rd6245; + xor.b64 %rd6247, %rd6246, %rd6198; + not.b64 %rd6248, %rd6200; + and.b64 %rd6249, %rd6201, %rd6248; + xor.b64 %rd6250, %rd6249, %rd6199; + not.b64 %rd6251, %rd6201; + and.b64 %rd6252, %rd6197, %rd6251; + xor.b64 %rd6253, %rd6252, %rd6200; + not.b64 %rd6254, %rd6197; + and.b64 %rd6255, %rd6198, %rd6254; + xor.b64 %rd6256, %rd6255, %rd6201; + not.b64 %rd6257, %rd6203; + and.b64 %rd6258, %rd6204, %rd6257; + xor.b64 %rd6259, %rd6258, %rd6202; + not.b64 %rd6260, %rd6204; + and.b64 %rd6261, %rd6205, %rd6260; + xor.b64 %rd6262, %rd6261, %rd6203; + not.b64 %rd6263, %rd6205; + and.b64 %rd6264, %rd6206, %rd6263; + xor.b64 %rd6265, %rd6264, %rd6204; + not.b64 %rd6266, %rd6206; + and.b64 %rd6267, %rd6202, %rd6266; + xor.b64 %rd6268, %rd6267, %rd6205; + not.b64 %rd6269, %rd6202; + and.b64 %rd6270, %rd6203, %rd6269; + xor.b64 %rd6271, %rd6270, %rd6206; + not.b64 %rd6272, %rd6208; + and.b64 %rd6273, %rd6209, %rd6272; + xor.b64 %rd6274, %rd6273, %rd6207; + not.b64 %rd6275, %rd6209; + and.b64 %rd6276, %rd6210, %rd6275; + xor.b64 %rd6277, %rd6276, %rd6208; + not.b64 %rd6278, %rd6210; + and.b64 %rd6279, %rd6211, %rd6278; + xor.b64 %rd6280, %rd6279, %rd6209; + not.b64 %rd6281, %rd6211; + and.b64 %rd6282, %rd6207, %rd6281; + xor.b64 %rd6283, %rd6282, %rd6210; + not.b64 %rd6284, %rd6207; + and.b64 %rd6285, %rd6208, %rd6284; + xor.b64 %rd6286, %rd6285, %rd6211; + xor.b64 %rd6287, %rd6214, %rd28; + xor.b64 %rd6288, %rd6229, %rd6287; + xor.b64 %rd6289, %rd6288, %rd6244; + xor.b64 %rd6290, %rd6289, %rd6259; + xor.b64 %rd6291, %rd6290, %rd6274; + xor.b64 %rd6292, %rd6232, %rd6217; + xor.b64 %rd6293, %rd6292, %rd6247; + xor.b64 %rd6294, %rd6293, %rd6262; + xor.b64 %rd6295, %rd6294, %rd6277; + xor.b64 %rd6296, %rd6235, %rd6220; + xor.b64 %rd6297, %rd6296, %rd6250; + xor.b64 %rd6298, %rd6297, %rd6265; + xor.b64 %rd6299, %rd6298, %rd6280; + xor.b64 %rd6300, %rd6238, %rd6223; + xor.b64 %rd6301, %rd6300, %rd6253; + xor.b64 %rd6302, %rd6301, %rd6268; + xor.b64 %rd6303, %rd6302, %rd6283; + xor.b64 %rd6304, %rd6241, %rd6226; + xor.b64 %rd6305, %rd6304, %rd6256; + xor.b64 %rd6306, %rd6305, %rd6271; + xor.b64 %rd6307, %rd6306, %rd6286; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4370}, %rd6295; + } + { + .reg .b32 %dummy; + mov.b64 {%r4371,%dummy}, %rd6295; + } + shf.l.wrap.b32 %r4372, %r4371, %r4370, 1; + shf.l.wrap.b32 %r4373, %r4370, %r4371, 1; + mov.b64 %rd6308, {%r4373, %r4372}; + xor.b64 %rd6309, %rd6307, %rd6308; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4374}, %rd6299; + } + { + .reg .b32 %dummy; + mov.b64 {%r4375,%dummy}, %rd6299; + } + shf.l.wrap.b32 %r4376, %r4375, %r4374, 1; + shf.l.wrap.b32 %r4377, %r4374, %r4375, 1; + mov.b64 %rd6310, {%r4377, %r4376}; + xor.b64 %rd6311, %rd6310, %rd6291; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4378}, %rd6303; + } + { + .reg .b32 %dummy; + mov.b64 {%r4379,%dummy}, %rd6303; + } + shf.l.wrap.b32 %r4380, %r4379, %r4378, 1; + shf.l.wrap.b32 %r4381, %r4378, %r4379, 1; + mov.b64 %rd6312, {%r4381, %r4380}; + xor.b64 %rd6313, %rd6312, %rd6295; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4382}, %rd6307; + } + { + .reg .b32 %dummy; + mov.b64 {%r4383,%dummy}, %rd6307; + } + shf.l.wrap.b32 %r4384, %r4383, %r4382, 1; + shf.l.wrap.b32 %r4385, %r4382, %r4383, 1; + mov.b64 %rd6314, {%r4385, %r4384}; + xor.b64 %rd6315, %rd6314, %rd6299; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4386}, %rd6291; + } + { + .reg .b32 %dummy; + mov.b64 {%r4387,%dummy}, %rd6291; + } + shf.l.wrap.b32 %r4388, %r4387, %r4386, 1; + shf.l.wrap.b32 %r4389, %r4386, %r4387, 1; + mov.b64 %rd6316, {%r4389, %r4388}; + xor.b64 %rd6317, %rd6303, %rd6316; + xor.b64 %rd6318, %rd6309, %rd6287; + xor.b64 %rd6319, %rd6311, %rd6217; + xor.b64 %rd6320, %rd6313, %rd6220; + xor.b64 %rd6321, %rd6315, %rd6223; + xor.b64 %rd6322, %rd6317, %rd6226; + xor.b64 %rd6323, %rd6309, %rd6229; + xor.b64 %rd6324, %rd6311, %rd6232; + xor.b64 %rd6325, %rd6313, %rd6235; + xor.b64 %rd6326, %rd6315, %rd6238; + xor.b64 %rd6327, %rd6317, %rd6241; + xor.b64 %rd6328, %rd6309, %rd6244; + xor.b64 %rd6329, %rd6311, %rd6247; + xor.b64 %rd6330, %rd6313, %rd6250; + xor.b64 %rd6331, %rd6315, %rd6253; + xor.b64 %rd6332, %rd6317, %rd6256; + xor.b64 %rd6333, %rd6309, %rd6259; + xor.b64 %rd6334, %rd6311, %rd6262; + xor.b64 %rd6335, %rd6313, %rd6265; + xor.b64 %rd6336, %rd6315, %rd6268; + xor.b64 %rd6337, %rd6317, %rd6271; + xor.b64 %rd6338, %rd6309, %rd6274; + xor.b64 %rd6339, %rd6311, %rd6277; + xor.b64 %rd6340, %rd6313, %rd6280; + xor.b64 %rd6341, %rd6315, %rd6283; + xor.b64 %rd6342, %rd6317, %rd6286; + { + .reg .b32 %dummy; + mov.b64 {%r4390,%dummy}, %rd6324; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4391}, %rd6324; + } + shf.r.wrap.b32 %r4392, %r4391, %r4390, 20; + shf.r.wrap.b32 %r4393, %r4390, %r4391, 20; + mov.b64 %rd6343, {%r4393, %r4392}; + { + .reg .b32 %dummy; + mov.b64 {%r4394,%dummy}, %rd6330; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4395}, %rd6330; + } + shf.r.wrap.b32 %r4396, %r4395, %r4394, 21; + shf.r.wrap.b32 %r4397, %r4394, %r4395, 21; + mov.b64 %rd6344, {%r4397, %r4396}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4398}, %rd6336; + } + { + .reg .b32 %dummy; + mov.b64 {%r4399,%dummy}, %rd6336; + } + shf.l.wrap.b32 %r4400, %r4399, %r4398, 21; + shf.l.wrap.b32 %r4401, %r4398, %r4399, 21; + mov.b64 %rd6345, {%r4401, %r4400}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4402}, %rd6342; + } + { + .reg .b32 %dummy; + mov.b64 {%r4403,%dummy}, %rd6342; + } + shf.l.wrap.b32 %r4404, %r4403, %r4402, 14; + shf.l.wrap.b32 %r4405, %r4402, %r4403, 14; + mov.b64 %rd6346, {%r4405, %r4404}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4406}, %rd6321; + } + { + .reg .b32 %dummy; + mov.b64 {%r4407,%dummy}, %rd6321; + } + shf.l.wrap.b32 %r4408, %r4407, %r4406, 28; + shf.l.wrap.b32 %r4409, %r4406, %r4407, 28; + mov.b64 %rd6347, {%r4409, %r4408}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4410}, %rd6327; + } + { + .reg .b32 %dummy; + mov.b64 {%r4411,%dummy}, %rd6327; + } + shf.l.wrap.b32 %r4412, %r4411, %r4410, 20; + shf.l.wrap.b32 %r4413, %r4410, %r4411, 20; + mov.b64 %rd6348, {%r4413, %r4412}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4414}, %rd6328; + } + { + .reg .b32 %dummy; + mov.b64 {%r4415,%dummy}, %rd6328; + } + shf.l.wrap.b32 %r4416, %r4415, %r4414, 3; + shf.l.wrap.b32 %r4417, %r4414, %r4415, 3; + mov.b64 %rd6349, {%r4417, %r4416}; + { + .reg .b32 %dummy; + mov.b64 {%r4418,%dummy}, %rd6334; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4419}, %rd6334; + } + shf.r.wrap.b32 %r4420, %r4419, %r4418, 19; + shf.r.wrap.b32 %r4421, %r4418, %r4419, 19; + mov.b64 %rd6350, {%r4421, %r4420}; + { + .reg .b32 %dummy; + mov.b64 {%r4422,%dummy}, %rd6340; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4423}, %rd6340; + } + shf.r.wrap.b32 %r4424, %r4423, %r4422, 3; + shf.r.wrap.b32 %r4425, %r4422, %r4423, 3; + mov.b64 %rd6351, {%r4425, %r4424}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4426}, %rd6319; + } + { + .reg .b32 %dummy; + mov.b64 {%r4427,%dummy}, %rd6319; + } + shf.l.wrap.b32 %r4428, %r4427, %r4426, 1; + shf.l.wrap.b32 %r4429, %r4426, %r4427, 1; + mov.b64 %rd6352, {%r4429, %r4428}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4430}, %rd6325; + } + { + .reg .b32 %dummy; + mov.b64 {%r4431,%dummy}, %rd6325; + } + shf.l.wrap.b32 %r4432, %r4431, %r4430, 6; + shf.l.wrap.b32 %r4433, %r4430, %r4431, 6; + mov.b64 %rd6353, {%r4433, %r4432}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4434}, %rd6331; + } + { + .reg .b32 %dummy; + mov.b64 {%r4435,%dummy}, %rd6331; + } + shf.l.wrap.b32 %r4436, %r4435, %r4434, 25; + shf.l.wrap.b32 %r4437, %r4434, %r4435, 25; + mov.b64 %rd6354, {%r4437, %r4436}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4438}, %rd6337; + } + { + .reg .b32 %dummy; + mov.b64 {%r4439,%dummy}, %rd6337; + } + shf.l.wrap.b32 %r4440, %r4439, %r4438, 8; + shf.l.wrap.b32 %r4441, %r4438, %r4439, 8; + mov.b64 %rd6355, {%r4441, %r4440}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4442}, %rd6338; + } + { + .reg .b32 %dummy; + mov.b64 {%r4443,%dummy}, %rd6338; + } + shf.l.wrap.b32 %r4444, %r4443, %r4442, 18; + shf.l.wrap.b32 %r4445, %r4442, %r4443, 18; + mov.b64 %rd6356, {%r4445, %r4444}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4446}, %rd6322; + } + { + .reg .b32 %dummy; + mov.b64 {%r4447,%dummy}, %rd6322; + } + shf.l.wrap.b32 %r4448, %r4447, %r4446, 27; + shf.l.wrap.b32 %r4449, %r4446, %r4447, 27; + mov.b64 %rd6357, {%r4449, %r4448}; + { + .reg .b32 %dummy; + mov.b64 {%r4450,%dummy}, %rd6323; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4451}, %rd6323; + } + shf.r.wrap.b32 %r4452, %r4451, %r4450, 28; + shf.r.wrap.b32 %r4453, %r4450, %r4451, 28; + mov.b64 %rd6358, {%r4453, %r4452}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4454}, %rd6329; + } + { + .reg .b32 %dummy; + mov.b64 {%r4455,%dummy}, %rd6329; + } + shf.l.wrap.b32 %r4456, %r4455, %r4454, 10; + shf.l.wrap.b32 %r4457, %r4454, %r4455, 10; + mov.b64 %rd6359, {%r4457, %r4456}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4458}, %rd6335; + } + { + .reg .b32 %dummy; + mov.b64 {%r4459,%dummy}, %rd6335; + } + shf.l.wrap.b32 %r4460, %r4459, %r4458, 15; + shf.l.wrap.b32 %r4461, %r4458, %r4459, 15; + mov.b64 %rd6360, {%r4461, %r4460}; + { + .reg .b32 %dummy; + mov.b64 {%r4462,%dummy}, %rd6341; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4463}, %rd6341; + } + shf.r.wrap.b32 %r4464, %r4463, %r4462, 8; + shf.r.wrap.b32 %r4465, %r4462, %r4463, 8; + mov.b64 %rd6361, {%r4465, %r4464}; + { + .reg .b32 %dummy; + mov.b64 {%r4466,%dummy}, %rd6320; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4467}, %rd6320; + } + shf.r.wrap.b32 %r4468, %r4467, %r4466, 2; + shf.r.wrap.b32 %r4469, %r4466, %r4467, 2; + mov.b64 %rd6362, {%r4469, %r4468}; + { + .reg .b32 %dummy; + mov.b64 {%r4470,%dummy}, %rd6326; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4471}, %rd6326; + } + shf.r.wrap.b32 %r4472, %r4471, %r4470, 9; + shf.r.wrap.b32 %r4473, %r4470, %r4471, 9; + mov.b64 %rd6363, {%r4473, %r4472}; + { + .reg .b32 %dummy; + mov.b64 {%r4474,%dummy}, %rd6332; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4475}, %rd6332; + } + shf.r.wrap.b32 %r4476, %r4475, %r4474, 25; + shf.r.wrap.b32 %r4477, %r4474, %r4475, 25; + mov.b64 %rd6364, {%r4477, %r4476}; + { + .reg .b32 %dummy; + mov.b64 {%r4478,%dummy}, %rd6333; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4479}, %rd6333; + } + shf.r.wrap.b32 %r4480, %r4479, %r4478, 23; + shf.r.wrap.b32 %r4481, %r4478, %r4479, 23; + mov.b64 %rd6365, {%r4481, %r4480}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4482}, %rd6339; + } + { + .reg .b32 %dummy; + mov.b64 {%r4483,%dummy}, %rd6339; + } + shf.l.wrap.b32 %r4484, %r4483, %r4482, 2; + shf.l.wrap.b32 %r4485, %r4482, %r4483, 2; + mov.b64 %rd6366, {%r4485, %r4484}; + not.b64 %rd6367, %rd6343; + and.b64 %rd6368, %rd6344, %rd6367; + xor.b64 %rd6369, %rd6368, %rd6318; + not.b64 %rd6370, %rd6344; + and.b64 %rd6371, %rd6345, %rd6370; + xor.b64 %rd6372, %rd6371, %rd6343; + not.b64 %rd6373, %rd6345; + and.b64 %rd6374, %rd6346, %rd6373; + xor.b64 %rd6375, %rd6374, %rd6344; + not.b64 %rd6376, %rd6346; + and.b64 %rd6377, %rd6318, %rd6376; + xor.b64 %rd6378, %rd6377, %rd6345; + not.b64 %rd6379, %rd6318; + and.b64 %rd6380, %rd6343, %rd6379; + xor.b64 %rd6381, %rd6380, %rd6346; + not.b64 %rd6382, %rd6348; + and.b64 %rd6383, %rd6349, %rd6382; + xor.b64 %rd6384, %rd6383, %rd6347; + not.b64 %rd6385, %rd6349; + and.b64 %rd6386, %rd6350, %rd6385; + xor.b64 %rd6387, %rd6386, %rd6348; + not.b64 %rd6388, %rd6350; + and.b64 %rd6389, %rd6351, %rd6388; + xor.b64 %rd6390, %rd6389, %rd6349; + not.b64 %rd6391, %rd6351; + and.b64 %rd6392, %rd6347, %rd6391; + xor.b64 %rd6393, %rd6392, %rd6350; + not.b64 %rd6394, %rd6347; + and.b64 %rd6395, %rd6348, %rd6394; + xor.b64 %rd6396, %rd6395, %rd6351; + not.b64 %rd6397, %rd6353; + and.b64 %rd6398, %rd6354, %rd6397; + xor.b64 %rd6399, %rd6398, %rd6352; + not.b64 %rd6400, %rd6354; + and.b64 %rd6401, %rd6355, %rd6400; + xor.b64 %rd6402, %rd6401, %rd6353; + not.b64 %rd6403, %rd6355; + and.b64 %rd6404, %rd6356, %rd6403; + xor.b64 %rd6405, %rd6404, %rd6354; + not.b64 %rd6406, %rd6356; + and.b64 %rd6407, %rd6352, %rd6406; + xor.b64 %rd6408, %rd6407, %rd6355; + not.b64 %rd6409, %rd6352; + and.b64 %rd6410, %rd6353, %rd6409; + xor.b64 %rd6411, %rd6410, %rd6356; + not.b64 %rd6412, %rd6358; + and.b64 %rd6413, %rd6359, %rd6412; + xor.b64 %rd6414, %rd6413, %rd6357; + not.b64 %rd6415, %rd6359; + and.b64 %rd6416, %rd6360, %rd6415; + xor.b64 %rd6417, %rd6416, %rd6358; + not.b64 %rd6418, %rd6360; + and.b64 %rd6419, %rd6361, %rd6418; + xor.b64 %rd6420, %rd6419, %rd6359; + not.b64 %rd6421, %rd6361; + and.b64 %rd6422, %rd6357, %rd6421; + xor.b64 %rd6423, %rd6422, %rd6360; + not.b64 %rd6424, %rd6357; + and.b64 %rd6425, %rd6358, %rd6424; + xor.b64 %rd6426, %rd6425, %rd6361; + not.b64 %rd6427, %rd6363; + and.b64 %rd6428, %rd6364, %rd6427; + xor.b64 %rd6429, %rd6428, %rd6362; + not.b64 %rd6430, %rd6364; + and.b64 %rd6431, %rd6365, %rd6430; + xor.b64 %rd6432, %rd6431, %rd6363; + not.b64 %rd6433, %rd6365; + and.b64 %rd6434, %rd6366, %rd6433; + xor.b64 %rd6435, %rd6434, %rd6364; + not.b64 %rd6436, %rd6366; + and.b64 %rd6437, %rd6362, %rd6436; + xor.b64 %rd6438, %rd6437, %rd6365; + not.b64 %rd6439, %rd6362; + and.b64 %rd6440, %rd6363, %rd6439; + xor.b64 %rd6441, %rd6440, %rd6366; + xor.b64 %rd6442, %rd6369, %rd29; + xor.b64 %rd6443, %rd6384, %rd6442; + xor.b64 %rd6444, %rd6443, %rd6399; + xor.b64 %rd6445, %rd6444, %rd6414; + xor.b64 %rd6446, %rd6445, %rd6429; + xor.b64 %rd6447, %rd6387, %rd6372; + xor.b64 %rd6448, %rd6447, %rd6402; + xor.b64 %rd6449, %rd6448, %rd6417; + xor.b64 %rd6450, %rd6449, %rd6432; + xor.b64 %rd6451, %rd6390, %rd6375; + xor.b64 %rd6452, %rd6451, %rd6405; + xor.b64 %rd6453, %rd6452, %rd6420; + xor.b64 %rd6454, %rd6453, %rd6435; + xor.b64 %rd6455, %rd6393, %rd6378; + xor.b64 %rd6456, %rd6455, %rd6408; + xor.b64 %rd6457, %rd6456, %rd6423; + xor.b64 %rd6458, %rd6457, %rd6438; + xor.b64 %rd6459, %rd6396, %rd6381; + xor.b64 %rd6460, %rd6459, %rd6411; + xor.b64 %rd6461, %rd6460, %rd6426; + xor.b64 %rd6462, %rd6461, %rd6441; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4486}, %rd6450; + } + { + .reg .b32 %dummy; + mov.b64 {%r4487,%dummy}, %rd6450; + } + shf.l.wrap.b32 %r4488, %r4487, %r4486, 1; + shf.l.wrap.b32 %r4489, %r4486, %r4487, 1; + mov.b64 %rd6463, {%r4489, %r4488}; + xor.b64 %rd6464, %rd6462, %rd6463; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4490}, %rd6454; + } + { + .reg .b32 %dummy; + mov.b64 {%r4491,%dummy}, %rd6454; + } + shf.l.wrap.b32 %r4492, %r4491, %r4490, 1; + shf.l.wrap.b32 %r4493, %r4490, %r4491, 1; + mov.b64 %rd6465, {%r4493, %r4492}; + xor.b64 %rd6466, %rd6465, %rd6446; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4494}, %rd6458; + } + { + .reg .b32 %dummy; + mov.b64 {%r4495,%dummy}, %rd6458; + } + shf.l.wrap.b32 %r4496, %r4495, %r4494, 1; + shf.l.wrap.b32 %r4497, %r4494, %r4495, 1; + mov.b64 %rd6467, {%r4497, %r4496}; + xor.b64 %rd6468, %rd6467, %rd6450; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4498}, %rd6462; + } + { + .reg .b32 %dummy; + mov.b64 {%r4499,%dummy}, %rd6462; + } + shf.l.wrap.b32 %r4500, %r4499, %r4498, 1; + shf.l.wrap.b32 %r4501, %r4498, %r4499, 1; + mov.b64 %rd6469, {%r4501, %r4500}; + xor.b64 %rd6470, %rd6469, %rd6454; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4502}, %rd6446; + } + { + .reg .b32 %dummy; + mov.b64 {%r4503,%dummy}, %rd6446; + } + shf.l.wrap.b32 %r4504, %r4503, %r4502, 1; + shf.l.wrap.b32 %r4505, %r4502, %r4503, 1; + mov.b64 %rd6471, {%r4505, %r4504}; + xor.b64 %rd6472, %rd6458, %rd6471; + xor.b64 %rd6473, %rd6464, %rd6442; + xor.b64 %rd6474, %rd6466, %rd6372; + xor.b64 %rd6475, %rd6468, %rd6375; + xor.b64 %rd6476, %rd6470, %rd6378; + xor.b64 %rd6477, %rd6472, %rd6381; + xor.b64 %rd6478, %rd6464, %rd6384; + xor.b64 %rd6479, %rd6466, %rd6387; + xor.b64 %rd6480, %rd6468, %rd6390; + xor.b64 %rd6481, %rd6470, %rd6393; + xor.b64 %rd6482, %rd6472, %rd6396; + xor.b64 %rd6483, %rd6464, %rd6399; + xor.b64 %rd6484, %rd6466, %rd6402; + xor.b64 %rd6485, %rd6468, %rd6405; + xor.b64 %rd6486, %rd6470, %rd6408; + xor.b64 %rd6487, %rd6472, %rd6411; + xor.b64 %rd6488, %rd6464, %rd6414; + xor.b64 %rd6489, %rd6466, %rd6417; + xor.b64 %rd6490, %rd6468, %rd6420; + xor.b64 %rd6491, %rd6470, %rd6423; + xor.b64 %rd6492, %rd6472, %rd6426; + xor.b64 %rd6493, %rd6464, %rd6429; + xor.b64 %rd6494, %rd6466, %rd6432; + xor.b64 %rd6495, %rd6468, %rd6435; + xor.b64 %rd6496, %rd6470, %rd6438; + xor.b64 %rd6497, %rd6472, %rd6441; + { + .reg .b32 %dummy; + mov.b64 {%r4506,%dummy}, %rd6479; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4507}, %rd6479; + } + shf.r.wrap.b32 %r4508, %r4507, %r4506, 20; + shf.r.wrap.b32 %r4509, %r4506, %r4507, 20; + mov.b64 %rd6498, {%r4509, %r4508}; + { + .reg .b32 %dummy; + mov.b64 {%r4510,%dummy}, %rd6485; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4511}, %rd6485; + } + shf.r.wrap.b32 %r4512, %r4511, %r4510, 21; + shf.r.wrap.b32 %r4513, %r4510, %r4511, 21; + mov.b64 %rd6499, {%r4513, %r4512}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4514}, %rd6491; + } + { + .reg .b32 %dummy; + mov.b64 {%r4515,%dummy}, %rd6491; + } + shf.l.wrap.b32 %r4516, %r4515, %r4514, 21; + shf.l.wrap.b32 %r4517, %r4514, %r4515, 21; + mov.b64 %rd6500, {%r4517, %r4516}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4518}, %rd6497; + } + { + .reg .b32 %dummy; + mov.b64 {%r4519,%dummy}, %rd6497; + } + shf.l.wrap.b32 %r4520, %r4519, %r4518, 14; + shf.l.wrap.b32 %r4521, %r4518, %r4519, 14; + mov.b64 %rd6501, {%r4521, %r4520}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4522}, %rd6476; + } + { + .reg .b32 %dummy; + mov.b64 {%r4523,%dummy}, %rd6476; + } + shf.l.wrap.b32 %r4524, %r4523, %r4522, 28; + shf.l.wrap.b32 %r4525, %r4522, %r4523, 28; + mov.b64 %rd6502, {%r4525, %r4524}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4526}, %rd6482; + } + { + .reg .b32 %dummy; + mov.b64 {%r4527,%dummy}, %rd6482; + } + shf.l.wrap.b32 %r4528, %r4527, %r4526, 20; + shf.l.wrap.b32 %r4529, %r4526, %r4527, 20; + mov.b64 %rd6503, {%r4529, %r4528}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4530}, %rd6483; + } + { + .reg .b32 %dummy; + mov.b64 {%r4531,%dummy}, %rd6483; + } + shf.l.wrap.b32 %r4532, %r4531, %r4530, 3; + shf.l.wrap.b32 %r4533, %r4530, %r4531, 3; + mov.b64 %rd6504, {%r4533, %r4532}; + { + .reg .b32 %dummy; + mov.b64 {%r4534,%dummy}, %rd6489; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4535}, %rd6489; + } + shf.r.wrap.b32 %r4536, %r4535, %r4534, 19; + shf.r.wrap.b32 %r4537, %r4534, %r4535, 19; + mov.b64 %rd6505, {%r4537, %r4536}; + { + .reg .b32 %dummy; + mov.b64 {%r4538,%dummy}, %rd6495; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4539}, %rd6495; + } + shf.r.wrap.b32 %r4540, %r4539, %r4538, 3; + shf.r.wrap.b32 %r4541, %r4538, %r4539, 3; + mov.b64 %rd6506, {%r4541, %r4540}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4542}, %rd6474; + } + { + .reg .b32 %dummy; + mov.b64 {%r4543,%dummy}, %rd6474; + } + shf.l.wrap.b32 %r4544, %r4543, %r4542, 1; + shf.l.wrap.b32 %r4545, %r4542, %r4543, 1; + mov.b64 %rd6507, {%r4545, %r4544}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4546}, %rd6480; + } + { + .reg .b32 %dummy; + mov.b64 {%r4547,%dummy}, %rd6480; + } + shf.l.wrap.b32 %r4548, %r4547, %r4546, 6; + shf.l.wrap.b32 %r4549, %r4546, %r4547, 6; + mov.b64 %rd6508, {%r4549, %r4548}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4550}, %rd6486; + } + { + .reg .b32 %dummy; + mov.b64 {%r4551,%dummy}, %rd6486; + } + shf.l.wrap.b32 %r4552, %r4551, %r4550, 25; + shf.l.wrap.b32 %r4553, %r4550, %r4551, 25; + mov.b64 %rd6509, {%r4553, %r4552}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4554}, %rd6492; + } + { + .reg .b32 %dummy; + mov.b64 {%r4555,%dummy}, %rd6492; + } + shf.l.wrap.b32 %r4556, %r4555, %r4554, 8; + shf.l.wrap.b32 %r4557, %r4554, %r4555, 8; + mov.b64 %rd6510, {%r4557, %r4556}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4558}, %rd6493; + } + { + .reg .b32 %dummy; + mov.b64 {%r4559,%dummy}, %rd6493; + } + shf.l.wrap.b32 %r4560, %r4559, %r4558, 18; + shf.l.wrap.b32 %r4561, %r4558, %r4559, 18; + mov.b64 %rd6511, {%r4561, %r4560}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4562}, %rd6477; + } + { + .reg .b32 %dummy; + mov.b64 {%r4563,%dummy}, %rd6477; + } + shf.l.wrap.b32 %r4564, %r4563, %r4562, 27; + shf.l.wrap.b32 %r4565, %r4562, %r4563, 27; + mov.b64 %rd6512, {%r4565, %r4564}; + { + .reg .b32 %dummy; + mov.b64 {%r4566,%dummy}, %rd6478; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4567}, %rd6478; + } + shf.r.wrap.b32 %r4568, %r4567, %r4566, 28; + shf.r.wrap.b32 %r4569, %r4566, %r4567, 28; + mov.b64 %rd6513, {%r4569, %r4568}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4570}, %rd6484; + } + { + .reg .b32 %dummy; + mov.b64 {%r4571,%dummy}, %rd6484; + } + shf.l.wrap.b32 %r4572, %r4571, %r4570, 10; + shf.l.wrap.b32 %r4573, %r4570, %r4571, 10; + mov.b64 %rd6514, {%r4573, %r4572}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4574}, %rd6490; + } + { + .reg .b32 %dummy; + mov.b64 {%r4575,%dummy}, %rd6490; + } + shf.l.wrap.b32 %r4576, %r4575, %r4574, 15; + shf.l.wrap.b32 %r4577, %r4574, %r4575, 15; + mov.b64 %rd6515, {%r4577, %r4576}; + { + .reg .b32 %dummy; + mov.b64 {%r4578,%dummy}, %rd6496; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4579}, %rd6496; + } + shf.r.wrap.b32 %r4580, %r4579, %r4578, 8; + shf.r.wrap.b32 %r4581, %r4578, %r4579, 8; + mov.b64 %rd6516, {%r4581, %r4580}; + { + .reg .b32 %dummy; + mov.b64 {%r4582,%dummy}, %rd6475; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4583}, %rd6475; + } + shf.r.wrap.b32 %r4584, %r4583, %r4582, 2; + shf.r.wrap.b32 %r4585, %r4582, %r4583, 2; + mov.b64 %rd6517, {%r4585, %r4584}; + { + .reg .b32 %dummy; + mov.b64 {%r4586,%dummy}, %rd6481; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4587}, %rd6481; + } + shf.r.wrap.b32 %r4588, %r4587, %r4586, 9; + shf.r.wrap.b32 %r4589, %r4586, %r4587, 9; + mov.b64 %rd6518, {%r4589, %r4588}; + { + .reg .b32 %dummy; + mov.b64 {%r4590,%dummy}, %rd6487; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4591}, %rd6487; + } + shf.r.wrap.b32 %r4592, %r4591, %r4590, 25; + shf.r.wrap.b32 %r4593, %r4590, %r4591, 25; + mov.b64 %rd6519, {%r4593, %r4592}; + { + .reg .b32 %dummy; + mov.b64 {%r4594,%dummy}, %rd6488; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4595}, %rd6488; + } + shf.r.wrap.b32 %r4596, %r4595, %r4594, 23; + shf.r.wrap.b32 %r4597, %r4594, %r4595, 23; + mov.b64 %rd6520, {%r4597, %r4596}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4598}, %rd6494; + } + { + .reg .b32 %dummy; + mov.b64 {%r4599,%dummy}, %rd6494; + } + shf.l.wrap.b32 %r4600, %r4599, %r4598, 2; + shf.l.wrap.b32 %r4601, %r4598, %r4599, 2; + mov.b64 %rd6521, {%r4601, %r4600}; + not.b64 %rd6522, %rd6498; + and.b64 %rd6523, %rd6499, %rd6522; + xor.b64 %rd6524, %rd6523, %rd6473; + not.b64 %rd6525, %rd6499; + and.b64 %rd6526, %rd6500, %rd6525; + xor.b64 %rd6527, %rd6526, %rd6498; + not.b64 %rd6528, %rd6500; + and.b64 %rd6529, %rd6501, %rd6528; + xor.b64 %rd6530, %rd6529, %rd6499; + not.b64 %rd6531, %rd6501; + and.b64 %rd6532, %rd6473, %rd6531; + xor.b64 %rd6533, %rd6532, %rd6500; + not.b64 %rd6534, %rd6473; + and.b64 %rd6535, %rd6498, %rd6534; + xor.b64 %rd6536, %rd6535, %rd6501; + not.b64 %rd6537, %rd6503; + and.b64 %rd6538, %rd6504, %rd6537; + xor.b64 %rd6539, %rd6538, %rd6502; + not.b64 %rd6540, %rd6504; + and.b64 %rd6541, %rd6505, %rd6540; + xor.b64 %rd6542, %rd6541, %rd6503; + not.b64 %rd6543, %rd6505; + and.b64 %rd6544, %rd6506, %rd6543; + xor.b64 %rd6545, %rd6544, %rd6504; + not.b64 %rd6546, %rd6506; + and.b64 %rd6547, %rd6502, %rd6546; + xor.b64 %rd6548, %rd6547, %rd6505; + not.b64 %rd6549, %rd6502; + and.b64 %rd6550, %rd6503, %rd6549; + xor.b64 %rd6551, %rd6550, %rd6506; + not.b64 %rd6552, %rd6508; + and.b64 %rd6553, %rd6509, %rd6552; + xor.b64 %rd6554, %rd6553, %rd6507; + not.b64 %rd6555, %rd6509; + and.b64 %rd6556, %rd6510, %rd6555; + xor.b64 %rd6557, %rd6556, %rd6508; + not.b64 %rd6558, %rd6510; + and.b64 %rd6559, %rd6511, %rd6558; + xor.b64 %rd6560, %rd6559, %rd6509; + not.b64 %rd6561, %rd6511; + and.b64 %rd6562, %rd6507, %rd6561; + xor.b64 %rd6563, %rd6562, %rd6510; + not.b64 %rd6564, %rd6507; + and.b64 %rd6565, %rd6508, %rd6564; + xor.b64 %rd6566, %rd6565, %rd6511; + not.b64 %rd6567, %rd6513; + and.b64 %rd6568, %rd6514, %rd6567; + xor.b64 %rd6569, %rd6568, %rd6512; + not.b64 %rd6570, %rd6514; + and.b64 %rd6571, %rd6515, %rd6570; + xor.b64 %rd6572, %rd6571, %rd6513; + not.b64 %rd6573, %rd6515; + and.b64 %rd6574, %rd6516, %rd6573; + xor.b64 %rd6575, %rd6574, %rd6514; + not.b64 %rd6576, %rd6516; + and.b64 %rd6577, %rd6512, %rd6576; + xor.b64 %rd6578, %rd6577, %rd6515; + not.b64 %rd6579, %rd6512; + and.b64 %rd6580, %rd6513, %rd6579; + xor.b64 %rd6581, %rd6580, %rd6516; + not.b64 %rd6582, %rd6518; + and.b64 %rd6583, %rd6519, %rd6582; + xor.b64 %rd6584, %rd6583, %rd6517; + not.b64 %rd6585, %rd6519; + and.b64 %rd6586, %rd6520, %rd6585; + xor.b64 %rd6587, %rd6586, %rd6518; + not.b64 %rd6588, %rd6520; + and.b64 %rd6589, %rd6521, %rd6588; + xor.b64 %rd6590, %rd6589, %rd6519; + not.b64 %rd6591, %rd6521; + and.b64 %rd6592, %rd6517, %rd6591; + xor.b64 %rd6593, %rd6592, %rd6520; + not.b64 %rd6594, %rd6517; + and.b64 %rd6595, %rd6518, %rd6594; + xor.b64 %rd6596, %rd6595, %rd6521; + xor.b64 %rd6597, %rd6524, %rd30; + xor.b64 %rd6598, %rd6539, %rd6597; + xor.b64 %rd6599, %rd6598, %rd6554; + xor.b64 %rd6600, %rd6599, %rd6569; + xor.b64 %rd6601, %rd6600, %rd6584; + xor.b64 %rd6602, %rd6542, %rd6527; + xor.b64 %rd6603, %rd6602, %rd6557; + xor.b64 %rd6604, %rd6603, %rd6572; + xor.b64 %rd6605, %rd6604, %rd6587; + xor.b64 %rd6606, %rd6545, %rd6530; + xor.b64 %rd6607, %rd6606, %rd6560; + xor.b64 %rd6608, %rd6607, %rd6575; + xor.b64 %rd6609, %rd6608, %rd6590; + xor.b64 %rd6610, %rd6548, %rd6533; + xor.b64 %rd6611, %rd6610, %rd6563; + xor.b64 %rd6612, %rd6611, %rd6578; + xor.b64 %rd6613, %rd6612, %rd6593; + xor.b64 %rd6614, %rd6551, %rd6536; + xor.b64 %rd6615, %rd6614, %rd6566; + xor.b64 %rd6616, %rd6615, %rd6581; + xor.b64 %rd6617, %rd6616, %rd6596; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4602}, %rd6605; + } + { + .reg .b32 %dummy; + mov.b64 {%r4603,%dummy}, %rd6605; + } + shf.l.wrap.b32 %r4604, %r4603, %r4602, 1; + shf.l.wrap.b32 %r4605, %r4602, %r4603, 1; + mov.b64 %rd6618, {%r4605, %r4604}; + xor.b64 %rd6619, %rd6617, %rd6618; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4606}, %rd6609; + } + { + .reg .b32 %dummy; + mov.b64 {%r4607,%dummy}, %rd6609; + } + shf.l.wrap.b32 %r4608, %r4607, %r4606, 1; + shf.l.wrap.b32 %r4609, %r4606, %r4607, 1; + mov.b64 %rd6620, {%r4609, %r4608}; + xor.b64 %rd6621, %rd6620, %rd6601; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4610}, %rd6613; + } + { + .reg .b32 %dummy; + mov.b64 {%r4611,%dummy}, %rd6613; + } + shf.l.wrap.b32 %r4612, %r4611, %r4610, 1; + shf.l.wrap.b32 %r4613, %r4610, %r4611, 1; + mov.b64 %rd6622, {%r4613, %r4612}; + xor.b64 %rd6623, %rd6622, %rd6605; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4614}, %rd6617; + } + { + .reg .b32 %dummy; + mov.b64 {%r4615,%dummy}, %rd6617; + } + shf.l.wrap.b32 %r4616, %r4615, %r4614, 1; + shf.l.wrap.b32 %r4617, %r4614, %r4615, 1; + mov.b64 %rd6624, {%r4617, %r4616}; + xor.b64 %rd6625, %rd6624, %rd6609; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4618}, %rd6601; + } + { + .reg .b32 %dummy; + mov.b64 {%r4619,%dummy}, %rd6601; + } + shf.l.wrap.b32 %r4620, %r4619, %r4618, 1; + shf.l.wrap.b32 %r4621, %r4618, %r4619, 1; + mov.b64 %rd6626, {%r4621, %r4620}; + xor.b64 %rd6627, %rd6613, %rd6626; + xor.b64 %rd6628, %rd6619, %rd6597; + xor.b64 %rd6629, %rd6621, %rd6527; + xor.b64 %rd6630, %rd6623, %rd6530; + xor.b64 %rd6631, %rd6625, %rd6533; + xor.b64 %rd6632, %rd6627, %rd6536; + xor.b64 %rd6633, %rd6619, %rd6539; + xor.b64 %rd6634, %rd6621, %rd6542; + xor.b64 %rd6635, %rd6623, %rd6545; + xor.b64 %rd6636, %rd6625, %rd6548; + xor.b64 %rd6637, %rd6627, %rd6551; + xor.b64 %rd6638, %rd6619, %rd6554; + xor.b64 %rd6639, %rd6621, %rd6557; + xor.b64 %rd6640, %rd6623, %rd6560; + xor.b64 %rd6641, %rd6625, %rd6563; + xor.b64 %rd6642, %rd6627, %rd6566; + xor.b64 %rd6643, %rd6619, %rd6569; + xor.b64 %rd6644, %rd6621, %rd6572; + xor.b64 %rd6645, %rd6623, %rd6575; + xor.b64 %rd6646, %rd6625, %rd6578; + xor.b64 %rd6647, %rd6627, %rd6581; + xor.b64 %rd6648, %rd6619, %rd6584; + xor.b64 %rd6649, %rd6621, %rd6587; + xor.b64 %rd6650, %rd6623, %rd6590; + xor.b64 %rd6651, %rd6625, %rd6593; + xor.b64 %rd6652, %rd6627, %rd6596; + { + .reg .b32 %dummy; + mov.b64 {%r4622,%dummy}, %rd6634; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4623}, %rd6634; + } + shf.r.wrap.b32 %r4624, %r4623, %r4622, 20; + shf.r.wrap.b32 %r4625, %r4622, %r4623, 20; + mov.b64 %rd6653, {%r4625, %r4624}; + { + .reg .b32 %dummy; + mov.b64 {%r4626,%dummy}, %rd6640; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4627}, %rd6640; + } + shf.r.wrap.b32 %r4628, %r4627, %r4626, 21; + shf.r.wrap.b32 %r4629, %r4626, %r4627, 21; + mov.b64 %rd6654, {%r4629, %r4628}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4630}, %rd6646; + } + { + .reg .b32 %dummy; + mov.b64 {%r4631,%dummy}, %rd6646; + } + shf.l.wrap.b32 %r4632, %r4631, %r4630, 21; + shf.l.wrap.b32 %r4633, %r4630, %r4631, 21; + mov.b64 %rd6655, {%r4633, %r4632}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4634}, %rd6652; + } + { + .reg .b32 %dummy; + mov.b64 {%r4635,%dummy}, %rd6652; + } + shf.l.wrap.b32 %r4636, %r4635, %r4634, 14; + shf.l.wrap.b32 %r4637, %r4634, %r4635, 14; + mov.b64 %rd6656, {%r4637, %r4636}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4638}, %rd6631; + } + { + .reg .b32 %dummy; + mov.b64 {%r4639,%dummy}, %rd6631; + } + shf.l.wrap.b32 %r4640, %r4639, %r4638, 28; + shf.l.wrap.b32 %r4641, %r4638, %r4639, 28; + mov.b64 %rd6657, {%r4641, %r4640}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4642}, %rd6637; + } + { + .reg .b32 %dummy; + mov.b64 {%r4643,%dummy}, %rd6637; + } + shf.l.wrap.b32 %r4644, %r4643, %r4642, 20; + shf.l.wrap.b32 %r4645, %r4642, %r4643, 20; + mov.b64 %rd6658, {%r4645, %r4644}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4646}, %rd6638; + } + { + .reg .b32 %dummy; + mov.b64 {%r4647,%dummy}, %rd6638; + } + shf.l.wrap.b32 %r4648, %r4647, %r4646, 3; + shf.l.wrap.b32 %r4649, %r4646, %r4647, 3; + mov.b64 %rd6659, {%r4649, %r4648}; + { + .reg .b32 %dummy; + mov.b64 {%r4650,%dummy}, %rd6644; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4651}, %rd6644; + } + shf.r.wrap.b32 %r4652, %r4651, %r4650, 19; + shf.r.wrap.b32 %r4653, %r4650, %r4651, 19; + mov.b64 %rd6660, {%r4653, %r4652}; + { + .reg .b32 %dummy; + mov.b64 {%r4654,%dummy}, %rd6650; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4655}, %rd6650; + } + shf.r.wrap.b32 %r4656, %r4655, %r4654, 3; + shf.r.wrap.b32 %r4657, %r4654, %r4655, 3; + mov.b64 %rd6661, {%r4657, %r4656}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4658}, %rd6629; + } + { + .reg .b32 %dummy; + mov.b64 {%r4659,%dummy}, %rd6629; + } + shf.l.wrap.b32 %r4660, %r4659, %r4658, 1; + shf.l.wrap.b32 %r4661, %r4658, %r4659, 1; + mov.b64 %rd6662, {%r4661, %r4660}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4662}, %rd6635; + } + { + .reg .b32 %dummy; + mov.b64 {%r4663,%dummy}, %rd6635; + } + shf.l.wrap.b32 %r4664, %r4663, %r4662, 6; + shf.l.wrap.b32 %r4665, %r4662, %r4663, 6; + mov.b64 %rd6663, {%r4665, %r4664}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4666}, %rd6641; + } + { + .reg .b32 %dummy; + mov.b64 {%r4667,%dummy}, %rd6641; + } + shf.l.wrap.b32 %r4668, %r4667, %r4666, 25; + shf.l.wrap.b32 %r4669, %r4666, %r4667, 25; + mov.b64 %rd6664, {%r4669, %r4668}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4670}, %rd6647; + } + { + .reg .b32 %dummy; + mov.b64 {%r4671,%dummy}, %rd6647; + } + shf.l.wrap.b32 %r4672, %r4671, %r4670, 8; + shf.l.wrap.b32 %r4673, %r4670, %r4671, 8; + mov.b64 %rd6665, {%r4673, %r4672}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4674}, %rd6648; + } + { + .reg .b32 %dummy; + mov.b64 {%r4675,%dummy}, %rd6648; + } + shf.l.wrap.b32 %r4676, %r4675, %r4674, 18; + shf.l.wrap.b32 %r4677, %r4674, %r4675, 18; + mov.b64 %rd6666, {%r4677, %r4676}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4678}, %rd6632; + } + { + .reg .b32 %dummy; + mov.b64 {%r4679,%dummy}, %rd6632; + } + shf.l.wrap.b32 %r4680, %r4679, %r4678, 27; + shf.l.wrap.b32 %r4681, %r4678, %r4679, 27; + mov.b64 %rd6667, {%r4681, %r4680}; + { + .reg .b32 %dummy; + mov.b64 {%r4682,%dummy}, %rd6633; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4683}, %rd6633; + } + shf.r.wrap.b32 %r4684, %r4683, %r4682, 28; + shf.r.wrap.b32 %r4685, %r4682, %r4683, 28; + mov.b64 %rd6668, {%r4685, %r4684}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4686}, %rd6639; + } + { + .reg .b32 %dummy; + mov.b64 {%r4687,%dummy}, %rd6639; + } + shf.l.wrap.b32 %r4688, %r4687, %r4686, 10; + shf.l.wrap.b32 %r4689, %r4686, %r4687, 10; + mov.b64 %rd6669, {%r4689, %r4688}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4690}, %rd6645; + } + { + .reg .b32 %dummy; + mov.b64 {%r4691,%dummy}, %rd6645; + } + shf.l.wrap.b32 %r4692, %r4691, %r4690, 15; + shf.l.wrap.b32 %r4693, %r4690, %r4691, 15; + mov.b64 %rd6670, {%r4693, %r4692}; + { + .reg .b32 %dummy; + mov.b64 {%r4694,%dummy}, %rd6651; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4695}, %rd6651; + } + shf.r.wrap.b32 %r4696, %r4695, %r4694, 8; + shf.r.wrap.b32 %r4697, %r4694, %r4695, 8; + mov.b64 %rd6671, {%r4697, %r4696}; + { + .reg .b32 %dummy; + mov.b64 {%r4698,%dummy}, %rd6630; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4699}, %rd6630; + } + shf.r.wrap.b32 %r4700, %r4699, %r4698, 2; + shf.r.wrap.b32 %r4701, %r4698, %r4699, 2; + mov.b64 %rd6672, {%r4701, %r4700}; + { + .reg .b32 %dummy; + mov.b64 {%r4702,%dummy}, %rd6636; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4703}, %rd6636; + } + shf.r.wrap.b32 %r4704, %r4703, %r4702, 9; + shf.r.wrap.b32 %r4705, %r4702, %r4703, 9; + mov.b64 %rd6673, {%r4705, %r4704}; + { + .reg .b32 %dummy; + mov.b64 {%r4706,%dummy}, %rd6642; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4707}, %rd6642; + } + shf.r.wrap.b32 %r4708, %r4707, %r4706, 25; + shf.r.wrap.b32 %r4709, %r4706, %r4707, 25; + mov.b64 %rd6674, {%r4709, %r4708}; + { + .reg .b32 %dummy; + mov.b64 {%r4710,%dummy}, %rd6643; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4711}, %rd6643; + } + shf.r.wrap.b32 %r4712, %r4711, %r4710, 23; + shf.r.wrap.b32 %r4713, %r4710, %r4711, 23; + mov.b64 %rd6675, {%r4713, %r4712}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4714}, %rd6649; + } + { + .reg .b32 %dummy; + mov.b64 {%r4715,%dummy}, %rd6649; + } + shf.l.wrap.b32 %r4716, %r4715, %r4714, 2; + shf.l.wrap.b32 %r4717, %r4714, %r4715, 2; + mov.b64 %rd6676, {%r4717, %r4716}; + not.b64 %rd6677, %rd6653; + and.b64 %rd6678, %rd6654, %rd6677; + xor.b64 %rd6679, %rd6678, %rd6628; + not.b64 %rd6680, %rd6654; + and.b64 %rd6681, %rd6655, %rd6680; + xor.b64 %rd6682, %rd6681, %rd6653; + not.b64 %rd6683, %rd6655; + and.b64 %rd6684, %rd6656, %rd6683; + xor.b64 %rd6685, %rd6684, %rd6654; + not.b64 %rd6686, %rd6656; + and.b64 %rd6687, %rd6628, %rd6686; + xor.b64 %rd6688, %rd6687, %rd6655; + not.b64 %rd6689, %rd6628; + and.b64 %rd6690, %rd6653, %rd6689; + xor.b64 %rd6691, %rd6690, %rd6656; + not.b64 %rd6692, %rd6658; + and.b64 %rd6693, %rd6659, %rd6692; + xor.b64 %rd6694, %rd6693, %rd6657; + not.b64 %rd6695, %rd6659; + and.b64 %rd6696, %rd6660, %rd6695; + xor.b64 %rd6697, %rd6696, %rd6658; + not.b64 %rd6698, %rd6660; + and.b64 %rd6699, %rd6661, %rd6698; + xor.b64 %rd6700, %rd6699, %rd6659; + not.b64 %rd6701, %rd6661; + and.b64 %rd6702, %rd6657, %rd6701; + xor.b64 %rd6703, %rd6702, %rd6660; + not.b64 %rd6704, %rd6657; + and.b64 %rd6705, %rd6658, %rd6704; + xor.b64 %rd6706, %rd6705, %rd6661; + not.b64 %rd6707, %rd6663; + and.b64 %rd6708, %rd6664, %rd6707; + xor.b64 %rd6709, %rd6708, %rd6662; + not.b64 %rd6710, %rd6664; + and.b64 %rd6711, %rd6665, %rd6710; + xor.b64 %rd6712, %rd6711, %rd6663; + not.b64 %rd6713, %rd6665; + and.b64 %rd6714, %rd6666, %rd6713; + xor.b64 %rd6715, %rd6714, %rd6664; + not.b64 %rd6716, %rd6666; + and.b64 %rd6717, %rd6662, %rd6716; + xor.b64 %rd6718, %rd6717, %rd6665; + not.b64 %rd6719, %rd6662; + and.b64 %rd6720, %rd6663, %rd6719; + xor.b64 %rd6721, %rd6720, %rd6666; + not.b64 %rd6722, %rd6668; + and.b64 %rd6723, %rd6669, %rd6722; + xor.b64 %rd6724, %rd6723, %rd6667; + not.b64 %rd6725, %rd6669; + and.b64 %rd6726, %rd6670, %rd6725; + xor.b64 %rd6727, %rd6726, %rd6668; + not.b64 %rd6728, %rd6670; + and.b64 %rd6729, %rd6671, %rd6728; + xor.b64 %rd6730, %rd6729, %rd6669; + not.b64 %rd6731, %rd6671; + and.b64 %rd6732, %rd6667, %rd6731; + xor.b64 %rd6733, %rd6732, %rd6670; + not.b64 %rd6734, %rd6667; + and.b64 %rd6735, %rd6668, %rd6734; + xor.b64 %rd6736, %rd6735, %rd6671; + not.b64 %rd6737, %rd6673; + and.b64 %rd6738, %rd6674, %rd6737; + xor.b64 %rd6739, %rd6738, %rd6672; + not.b64 %rd6740, %rd6674; + and.b64 %rd6741, %rd6675, %rd6740; + xor.b64 %rd6742, %rd6741, %rd6673; + not.b64 %rd6743, %rd6675; + and.b64 %rd6744, %rd6676, %rd6743; + xor.b64 %rd6745, %rd6744, %rd6674; + not.b64 %rd6746, %rd6676; + and.b64 %rd6747, %rd6672, %rd6746; + xor.b64 %rd6748, %rd6747, %rd6675; + not.b64 %rd6749, %rd6672; + and.b64 %rd6750, %rd6673, %rd6749; + xor.b64 %rd6751, %rd6750, %rd6676; + xor.b64 %rd6752, %rd6679, %rd31; + xor.b64 %rd6753, %rd6694, %rd6752; + xor.b64 %rd6754, %rd6753, %rd6709; + xor.b64 %rd6755, %rd6754, %rd6724; + xor.b64 %rd6756, %rd6755, %rd6739; + xor.b64 %rd6757, %rd6697, %rd6682; + xor.b64 %rd6758, %rd6757, %rd6712; + xor.b64 %rd6759, %rd6758, %rd6727; + xor.b64 %rd6760, %rd6759, %rd6742; + xor.b64 %rd6761, %rd6700, %rd6685; + xor.b64 %rd6762, %rd6761, %rd6715; + xor.b64 %rd6763, %rd6762, %rd6730; + xor.b64 %rd6764, %rd6763, %rd6745; + xor.b64 %rd6765, %rd6703, %rd6688; + xor.b64 %rd6766, %rd6765, %rd6718; + xor.b64 %rd6767, %rd6766, %rd6733; + xor.b64 %rd6768, %rd6767, %rd6748; + xor.b64 %rd6769, %rd6706, %rd6691; + xor.b64 %rd6770, %rd6769, %rd6721; + xor.b64 %rd6771, %rd6770, %rd6736; + xor.b64 %rd6772, %rd6771, %rd6751; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4718}, %rd6760; + } + { + .reg .b32 %dummy; + mov.b64 {%r4719,%dummy}, %rd6760; + } + shf.l.wrap.b32 %r4720, %r4719, %r4718, 1; + shf.l.wrap.b32 %r4721, %r4718, %r4719, 1; + mov.b64 %rd6773, {%r4721, %r4720}; + xor.b64 %rd6774, %rd6772, %rd6773; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4722}, %rd6764; + } + { + .reg .b32 %dummy; + mov.b64 {%r4723,%dummy}, %rd6764; + } + shf.l.wrap.b32 %r4724, %r4723, %r4722, 1; + shf.l.wrap.b32 %r4725, %r4722, %r4723, 1; + mov.b64 %rd6775, {%r4725, %r4724}; + xor.b64 %rd6776, %rd6775, %rd6756; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4726}, %rd6768; + } + { + .reg .b32 %dummy; + mov.b64 {%r4727,%dummy}, %rd6768; + } + shf.l.wrap.b32 %r4728, %r4727, %r4726, 1; + shf.l.wrap.b32 %r4729, %r4726, %r4727, 1; + mov.b64 %rd6777, {%r4729, %r4728}; + xor.b64 %rd6778, %rd6777, %rd6760; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4730}, %rd6772; + } + { + .reg .b32 %dummy; + mov.b64 {%r4731,%dummy}, %rd6772; + } + shf.l.wrap.b32 %r4732, %r4731, %r4730, 1; + shf.l.wrap.b32 %r4733, %r4730, %r4731, 1; + mov.b64 %rd6779, {%r4733, %r4732}; + xor.b64 %rd6780, %rd6779, %rd6764; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4734}, %rd6756; + } + { + .reg .b32 %dummy; + mov.b64 {%r4735,%dummy}, %rd6756; + } + shf.l.wrap.b32 %r4736, %r4735, %r4734, 1; + shf.l.wrap.b32 %r4737, %r4734, %r4735, 1; + mov.b64 %rd6781, {%r4737, %r4736}; + xor.b64 %rd6782, %rd6768, %rd6781; + xor.b64 %rd6783, %rd6774, %rd6752; + xor.b64 %rd6784, %rd6776, %rd6682; + xor.b64 %rd6785, %rd6778, %rd6685; + xor.b64 %rd6786, %rd6780, %rd6688; + xor.b64 %rd6787, %rd6782, %rd6691; + xor.b64 %rd6788, %rd6774, %rd6694; + xor.b64 %rd6789, %rd6776, %rd6697; + xor.b64 %rd6790, %rd6778, %rd6700; + xor.b64 %rd6791, %rd6780, %rd6703; + xor.b64 %rd6792, %rd6782, %rd6706; + xor.b64 %rd6793, %rd6774, %rd6709; + xor.b64 %rd6794, %rd6776, %rd6712; + xor.b64 %rd6795, %rd6778, %rd6715; + xor.b64 %rd6796, %rd6780, %rd6718; + xor.b64 %rd6797, %rd6782, %rd6721; + xor.b64 %rd6798, %rd6774, %rd6724; + xor.b64 %rd6799, %rd6776, %rd6727; + xor.b64 %rd6800, %rd6778, %rd6730; + xor.b64 %rd6801, %rd6780, %rd6733; + xor.b64 %rd6802, %rd6782, %rd6736; + xor.b64 %rd6803, %rd6774, %rd6739; + xor.b64 %rd6804, %rd6776, %rd6742; + xor.b64 %rd6805, %rd6778, %rd6745; + xor.b64 %rd6806, %rd6780, %rd6748; + xor.b64 %rd6807, %rd6782, %rd6751; + { + .reg .b32 %dummy; + mov.b64 {%r4738,%dummy}, %rd6789; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4739}, %rd6789; + } + shf.r.wrap.b32 %r4740, %r4739, %r4738, 20; + shf.r.wrap.b32 %r4741, %r4738, %r4739, 20; + mov.b64 %rd6808, {%r4741, %r4740}; + { + .reg .b32 %dummy; + mov.b64 {%r4742,%dummy}, %rd6795; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4743}, %rd6795; + } + shf.r.wrap.b32 %r4744, %r4743, %r4742, 21; + shf.r.wrap.b32 %r4745, %r4742, %r4743, 21; + mov.b64 %rd6809, {%r4745, %r4744}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4746}, %rd6801; + } + { + .reg .b32 %dummy; + mov.b64 {%r4747,%dummy}, %rd6801; + } + shf.l.wrap.b32 %r4748, %r4747, %r4746, 21; + shf.l.wrap.b32 %r4749, %r4746, %r4747, 21; + mov.b64 %rd6810, {%r4749, %r4748}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4750}, %rd6807; + } + { + .reg .b32 %dummy; + mov.b64 {%r4751,%dummy}, %rd6807; + } + shf.l.wrap.b32 %r4752, %r4751, %r4750, 14; + shf.l.wrap.b32 %r4753, %r4750, %r4751, 14; + mov.b64 %rd6811, {%r4753, %r4752}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4754}, %rd6786; + } + { + .reg .b32 %dummy; + mov.b64 {%r4755,%dummy}, %rd6786; + } + shf.l.wrap.b32 %r4756, %r4755, %r4754, 28; + shf.l.wrap.b32 %r4757, %r4754, %r4755, 28; + mov.b64 %rd6812, {%r4757, %r4756}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4758}, %rd6792; + } + { + .reg .b32 %dummy; + mov.b64 {%r4759,%dummy}, %rd6792; + } + shf.l.wrap.b32 %r4760, %r4759, %r4758, 20; + shf.l.wrap.b32 %r4761, %r4758, %r4759, 20; + mov.b64 %rd6813, {%r4761, %r4760}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4762}, %rd6793; + } + { + .reg .b32 %dummy; + mov.b64 {%r4763,%dummy}, %rd6793; + } + shf.l.wrap.b32 %r4764, %r4763, %r4762, 3; + shf.l.wrap.b32 %r4765, %r4762, %r4763, 3; + mov.b64 %rd6814, {%r4765, %r4764}; + { + .reg .b32 %dummy; + mov.b64 {%r4766,%dummy}, %rd6799; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4767}, %rd6799; + } + shf.r.wrap.b32 %r4768, %r4767, %r4766, 19; + shf.r.wrap.b32 %r4769, %r4766, %r4767, 19; + mov.b64 %rd6815, {%r4769, %r4768}; + { + .reg .b32 %dummy; + mov.b64 {%r4770,%dummy}, %rd6805; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4771}, %rd6805; + } + shf.r.wrap.b32 %r4772, %r4771, %r4770, 3; + shf.r.wrap.b32 %r4773, %r4770, %r4771, 3; + mov.b64 %rd6816, {%r4773, %r4772}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4774}, %rd6784; + } + { + .reg .b32 %dummy; + mov.b64 {%r4775,%dummy}, %rd6784; + } + shf.l.wrap.b32 %r4776, %r4775, %r4774, 1; + shf.l.wrap.b32 %r4777, %r4774, %r4775, 1; + mov.b64 %rd6817, {%r4777, %r4776}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4778}, %rd6790; + } + { + .reg .b32 %dummy; + mov.b64 {%r4779,%dummy}, %rd6790; + } + shf.l.wrap.b32 %r4780, %r4779, %r4778, 6; + shf.l.wrap.b32 %r4781, %r4778, %r4779, 6; + mov.b64 %rd6818, {%r4781, %r4780}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4782}, %rd6796; + } + { + .reg .b32 %dummy; + mov.b64 {%r4783,%dummy}, %rd6796; + } + shf.l.wrap.b32 %r4784, %r4783, %r4782, 25; + shf.l.wrap.b32 %r4785, %r4782, %r4783, 25; + mov.b64 %rd6819, {%r4785, %r4784}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4786}, %rd6802; + } + { + .reg .b32 %dummy; + mov.b64 {%r4787,%dummy}, %rd6802; + } + shf.l.wrap.b32 %r4788, %r4787, %r4786, 8; + shf.l.wrap.b32 %r4789, %r4786, %r4787, 8; + mov.b64 %rd6820, {%r4789, %r4788}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4790}, %rd6803; + } + { + .reg .b32 %dummy; + mov.b64 {%r4791,%dummy}, %rd6803; + } + shf.l.wrap.b32 %r4792, %r4791, %r4790, 18; + shf.l.wrap.b32 %r4793, %r4790, %r4791, 18; + mov.b64 %rd6821, {%r4793, %r4792}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4794}, %rd6787; + } + { + .reg .b32 %dummy; + mov.b64 {%r4795,%dummy}, %rd6787; + } + shf.l.wrap.b32 %r4796, %r4795, %r4794, 27; + shf.l.wrap.b32 %r4797, %r4794, %r4795, 27; + mov.b64 %rd6822, {%r4797, %r4796}; + { + .reg .b32 %dummy; + mov.b64 {%r4798,%dummy}, %rd6788; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4799}, %rd6788; + } + shf.r.wrap.b32 %r4800, %r4799, %r4798, 28; + shf.r.wrap.b32 %r4801, %r4798, %r4799, 28; + mov.b64 %rd6823, {%r4801, %r4800}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4802}, %rd6794; + } + { + .reg .b32 %dummy; + mov.b64 {%r4803,%dummy}, %rd6794; + } + shf.l.wrap.b32 %r4804, %r4803, %r4802, 10; + shf.l.wrap.b32 %r4805, %r4802, %r4803, 10; + mov.b64 %rd6824, {%r4805, %r4804}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4806}, %rd6800; + } + { + .reg .b32 %dummy; + mov.b64 {%r4807,%dummy}, %rd6800; + } + shf.l.wrap.b32 %r4808, %r4807, %r4806, 15; + shf.l.wrap.b32 %r4809, %r4806, %r4807, 15; + mov.b64 %rd6825, {%r4809, %r4808}; + { + .reg .b32 %dummy; + mov.b64 {%r4810,%dummy}, %rd6806; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4811}, %rd6806; + } + shf.r.wrap.b32 %r4812, %r4811, %r4810, 8; + shf.r.wrap.b32 %r4813, %r4810, %r4811, 8; + mov.b64 %rd6826, {%r4813, %r4812}; + { + .reg .b32 %dummy; + mov.b64 {%r4814,%dummy}, %rd6785; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4815}, %rd6785; + } + shf.r.wrap.b32 %r4816, %r4815, %r4814, 2; + shf.r.wrap.b32 %r4817, %r4814, %r4815, 2; + mov.b64 %rd6827, {%r4817, %r4816}; + { + .reg .b32 %dummy; + mov.b64 {%r4818,%dummy}, %rd6791; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4819}, %rd6791; + } + shf.r.wrap.b32 %r4820, %r4819, %r4818, 9; + shf.r.wrap.b32 %r4821, %r4818, %r4819, 9; + mov.b64 %rd6828, {%r4821, %r4820}; + { + .reg .b32 %dummy; + mov.b64 {%r4822,%dummy}, %rd6797; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4823}, %rd6797; + } + shf.r.wrap.b32 %r4824, %r4823, %r4822, 25; + shf.r.wrap.b32 %r4825, %r4822, %r4823, 25; + mov.b64 %rd6829, {%r4825, %r4824}; + { + .reg .b32 %dummy; + mov.b64 {%r4826,%dummy}, %rd6798; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4827}, %rd6798; + } + shf.r.wrap.b32 %r4828, %r4827, %r4826, 23; + shf.r.wrap.b32 %r4829, %r4826, %r4827, 23; + mov.b64 %rd6830, {%r4829, %r4828}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4830}, %rd6804; + } + { + .reg .b32 %dummy; + mov.b64 {%r4831,%dummy}, %rd6804; + } + shf.l.wrap.b32 %r4832, %r4831, %r4830, 2; + shf.l.wrap.b32 %r4833, %r4830, %r4831, 2; + mov.b64 %rd6831, {%r4833, %r4832}; + not.b64 %rd6832, %rd6808; + and.b64 %rd6833, %rd6809, %rd6832; + xor.b64 %rd6834, %rd6833, %rd6783; + not.b64 %rd6835, %rd6809; + and.b64 %rd6836, %rd6810, %rd6835; + xor.b64 %rd6837, %rd6836, %rd6808; + not.b64 %rd6838, %rd6810; + and.b64 %rd6839, %rd6811, %rd6838; + xor.b64 %rd6840, %rd6839, %rd6809; + not.b64 %rd6841, %rd6811; + and.b64 %rd6842, %rd6783, %rd6841; + xor.b64 %rd6843, %rd6842, %rd6810; + not.b64 %rd6844, %rd6783; + and.b64 %rd6845, %rd6808, %rd6844; + xor.b64 %rd6846, %rd6845, %rd6811; + not.b64 %rd6847, %rd6813; + and.b64 %rd6848, %rd6814, %rd6847; + xor.b64 %rd6849, %rd6848, %rd6812; + not.b64 %rd6850, %rd6814; + and.b64 %rd6851, %rd6815, %rd6850; + xor.b64 %rd6852, %rd6851, %rd6813; + not.b64 %rd6853, %rd6815; + and.b64 %rd6854, %rd6816, %rd6853; + xor.b64 %rd6855, %rd6854, %rd6814; + not.b64 %rd6856, %rd6816; + and.b64 %rd6857, %rd6812, %rd6856; + xor.b64 %rd6858, %rd6857, %rd6815; + not.b64 %rd6859, %rd6812; + and.b64 %rd6860, %rd6813, %rd6859; + xor.b64 %rd6861, %rd6860, %rd6816; + not.b64 %rd6862, %rd6818; + and.b64 %rd6863, %rd6819, %rd6862; + xor.b64 %rd6864, %rd6863, %rd6817; + not.b64 %rd6865, %rd6819; + and.b64 %rd6866, %rd6820, %rd6865; + xor.b64 %rd6867, %rd6866, %rd6818; + not.b64 %rd6868, %rd6820; + and.b64 %rd6869, %rd6821, %rd6868; + xor.b64 %rd6870, %rd6869, %rd6819; + not.b64 %rd6871, %rd6821; + and.b64 %rd6872, %rd6817, %rd6871; + xor.b64 %rd6873, %rd6872, %rd6820; + not.b64 %rd6874, %rd6817; + and.b64 %rd6875, %rd6818, %rd6874; + xor.b64 %rd6876, %rd6875, %rd6821; + not.b64 %rd6877, %rd6823; + and.b64 %rd6878, %rd6824, %rd6877; + xor.b64 %rd6879, %rd6878, %rd6822; + not.b64 %rd6880, %rd6824; + and.b64 %rd6881, %rd6825, %rd6880; + xor.b64 %rd6882, %rd6881, %rd6823; + not.b64 %rd6883, %rd6825; + and.b64 %rd6884, %rd6826, %rd6883; + xor.b64 %rd6885, %rd6884, %rd6824; + not.b64 %rd6886, %rd6826; + and.b64 %rd6887, %rd6822, %rd6886; + xor.b64 %rd6888, %rd6887, %rd6825; + not.b64 %rd6889, %rd6822; + and.b64 %rd6890, %rd6823, %rd6889; + xor.b64 %rd6891, %rd6890, %rd6826; + not.b64 %rd6892, %rd6828; + and.b64 %rd6893, %rd6829, %rd6892; + xor.b64 %rd6894, %rd6893, %rd6827; + not.b64 %rd6895, %rd6829; + and.b64 %rd6896, %rd6830, %rd6895; + xor.b64 %rd6897, %rd6896, %rd6828; + not.b64 %rd6898, %rd6830; + and.b64 %rd6899, %rd6831, %rd6898; + xor.b64 %rd6900, %rd6899, %rd6829; + not.b64 %rd6901, %rd6831; + and.b64 %rd6902, %rd6827, %rd6901; + xor.b64 %rd6903, %rd6902, %rd6830; + not.b64 %rd6904, %rd6827; + and.b64 %rd6905, %rd6828, %rd6904; + xor.b64 %rd6906, %rd6905, %rd6831; + xor.b64 %rd6907, %rd6834, %rd32; + xor.b64 %rd6908, %rd6849, %rd6907; + xor.b64 %rd6909, %rd6908, %rd6864; + xor.b64 %rd6910, %rd6909, %rd6879; + xor.b64 %rd6911, %rd6910, %rd6894; + xor.b64 %rd6912, %rd6852, %rd6837; + xor.b64 %rd6913, %rd6912, %rd6867; + xor.b64 %rd6914, %rd6913, %rd6882; + xor.b64 %rd6915, %rd6914, %rd6897; + xor.b64 %rd6916, %rd6855, %rd6840; + xor.b64 %rd6917, %rd6916, %rd6870; + xor.b64 %rd6918, %rd6917, %rd6885; + xor.b64 %rd6919, %rd6918, %rd6900; + xor.b64 %rd6920, %rd6858, %rd6843; + xor.b64 %rd6921, %rd6920, %rd6873; + xor.b64 %rd6922, %rd6921, %rd6888; + xor.b64 %rd6923, %rd6922, %rd6903; + xor.b64 %rd6924, %rd6861, %rd6846; + xor.b64 %rd6925, %rd6924, %rd6876; + xor.b64 %rd6926, %rd6925, %rd6891; + xor.b64 %rd6927, %rd6926, %rd6906; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4834}, %rd6915; + } + { + .reg .b32 %dummy; + mov.b64 {%r4835,%dummy}, %rd6915; + } + shf.l.wrap.b32 %r4836, %r4835, %r4834, 1; + shf.l.wrap.b32 %r4837, %r4834, %r4835, 1; + mov.b64 %rd6928, {%r4837, %r4836}; + xor.b64 %rd6929, %rd6927, %rd6928; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4838}, %rd6919; + } + { + .reg .b32 %dummy; + mov.b64 {%r4839,%dummy}, %rd6919; + } + shf.l.wrap.b32 %r4840, %r4839, %r4838, 1; + shf.l.wrap.b32 %r4841, %r4838, %r4839, 1; + mov.b64 %rd6930, {%r4841, %r4840}; + xor.b64 %rd6931, %rd6930, %rd6911; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4842}, %rd6923; + } + { + .reg .b32 %dummy; + mov.b64 {%r4843,%dummy}, %rd6923; + } + shf.l.wrap.b32 %r4844, %r4843, %r4842, 1; + shf.l.wrap.b32 %r4845, %r4842, %r4843, 1; + mov.b64 %rd6932, {%r4845, %r4844}; + xor.b64 %rd6933, %rd6932, %rd6915; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4846}, %rd6927; + } + { + .reg .b32 %dummy; + mov.b64 {%r4847,%dummy}, %rd6927; + } + shf.l.wrap.b32 %r4848, %r4847, %r4846, 1; + shf.l.wrap.b32 %r4849, %r4846, %r4847, 1; + mov.b64 %rd6934, {%r4849, %r4848}; + xor.b64 %rd6935, %rd6934, %rd6919; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4850}, %rd6911; + } + { + .reg .b32 %dummy; + mov.b64 {%r4851,%dummy}, %rd6911; + } + shf.l.wrap.b32 %r4852, %r4851, %r4850, 1; + shf.l.wrap.b32 %r4853, %r4850, %r4851, 1; + mov.b64 %rd6936, {%r4853, %r4852}; + xor.b64 %rd6937, %rd6923, %rd6936; + xor.b64 %rd6938, %rd6929, %rd6907; + xor.b64 %rd6939, %rd6931, %rd6837; + xor.b64 %rd6940, %rd6933, %rd6840; + xor.b64 %rd6941, %rd6935, %rd6843; + xor.b64 %rd6942, %rd6937, %rd6846; + xor.b64 %rd6943, %rd6929, %rd6849; + xor.b64 %rd6944, %rd6931, %rd6852; + xor.b64 %rd6945, %rd6933, %rd6855; + xor.b64 %rd6946, %rd6935, %rd6858; + xor.b64 %rd6947, %rd6937, %rd6861; + xor.b64 %rd6948, %rd6929, %rd6864; + xor.b64 %rd6949, %rd6931, %rd6867; + xor.b64 %rd6950, %rd6933, %rd6870; + xor.b64 %rd6951, %rd6935, %rd6873; + xor.b64 %rd6952, %rd6937, %rd6876; + xor.b64 %rd6953, %rd6929, %rd6879; + xor.b64 %rd6954, %rd6931, %rd6882; + xor.b64 %rd6955, %rd6933, %rd6885; + xor.b64 %rd6956, %rd6935, %rd6888; + xor.b64 %rd6957, %rd6937, %rd6891; + xor.b64 %rd6958, %rd6929, %rd6894; + xor.b64 %rd6959, %rd6931, %rd6897; + xor.b64 %rd6960, %rd6933, %rd6900; + xor.b64 %rd6961, %rd6935, %rd6903; + xor.b64 %rd6962, %rd6937, %rd6906; + { + .reg .b32 %dummy; + mov.b64 {%r4854,%dummy}, %rd6944; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4855}, %rd6944; + } + shf.r.wrap.b32 %r4856, %r4855, %r4854, 20; + shf.r.wrap.b32 %r4857, %r4854, %r4855, 20; + mov.b64 %rd6963, {%r4857, %r4856}; + { + .reg .b32 %dummy; + mov.b64 {%r4858,%dummy}, %rd6950; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4859}, %rd6950; + } + shf.r.wrap.b32 %r4860, %r4859, %r4858, 21; + shf.r.wrap.b32 %r4861, %r4858, %r4859, 21; + mov.b64 %rd6964, {%r4861, %r4860}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4862}, %rd6956; + } + { + .reg .b32 %dummy; + mov.b64 {%r4863,%dummy}, %rd6956; + } + shf.l.wrap.b32 %r4864, %r4863, %r4862, 21; + shf.l.wrap.b32 %r4865, %r4862, %r4863, 21; + mov.b64 %rd6965, {%r4865, %r4864}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4866}, %rd6962; + } + { + .reg .b32 %dummy; + mov.b64 {%r4867,%dummy}, %rd6962; + } + shf.l.wrap.b32 %r4868, %r4867, %r4866, 14; + shf.l.wrap.b32 %r4869, %r4866, %r4867, 14; + mov.b64 %rd6966, {%r4869, %r4868}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4870}, %rd6941; + } + { + .reg .b32 %dummy; + mov.b64 {%r4871,%dummy}, %rd6941; + } + shf.l.wrap.b32 %r4872, %r4871, %r4870, 28; + shf.l.wrap.b32 %r4873, %r4870, %r4871, 28; + mov.b64 %rd6967, {%r4873, %r4872}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4874}, %rd6947; + } + { + .reg .b32 %dummy; + mov.b64 {%r4875,%dummy}, %rd6947; + } + shf.l.wrap.b32 %r4876, %r4875, %r4874, 20; + shf.l.wrap.b32 %r4877, %r4874, %r4875, 20; + mov.b64 %rd6968, {%r4877, %r4876}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4878}, %rd6948; + } + { + .reg .b32 %dummy; + mov.b64 {%r4879,%dummy}, %rd6948; + } + shf.l.wrap.b32 %r4880, %r4879, %r4878, 3; + shf.l.wrap.b32 %r4881, %r4878, %r4879, 3; + mov.b64 %rd6969, {%r4881, %r4880}; + { + .reg .b32 %dummy; + mov.b64 {%r4882,%dummy}, %rd6954; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4883}, %rd6954; + } + shf.r.wrap.b32 %r4884, %r4883, %r4882, 19; + shf.r.wrap.b32 %r4885, %r4882, %r4883, 19; + mov.b64 %rd6970, {%r4885, %r4884}; + { + .reg .b32 %dummy; + mov.b64 {%r4886,%dummy}, %rd6960; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4887}, %rd6960; + } + shf.r.wrap.b32 %r4888, %r4887, %r4886, 3; + shf.r.wrap.b32 %r4889, %r4886, %r4887, 3; + mov.b64 %rd6971, {%r4889, %r4888}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4890}, %rd6939; + } + { + .reg .b32 %dummy; + mov.b64 {%r4891,%dummy}, %rd6939; + } + shf.l.wrap.b32 %r4892, %r4891, %r4890, 1; + shf.l.wrap.b32 %r4893, %r4890, %r4891, 1; + mov.b64 %rd6972, {%r4893, %r4892}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4894}, %rd6945; + } + { + .reg .b32 %dummy; + mov.b64 {%r4895,%dummy}, %rd6945; + } + shf.l.wrap.b32 %r4896, %r4895, %r4894, 6; + shf.l.wrap.b32 %r4897, %r4894, %r4895, 6; + mov.b64 %rd6973, {%r4897, %r4896}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4898}, %rd6951; + } + { + .reg .b32 %dummy; + mov.b64 {%r4899,%dummy}, %rd6951; + } + shf.l.wrap.b32 %r4900, %r4899, %r4898, 25; + shf.l.wrap.b32 %r4901, %r4898, %r4899, 25; + mov.b64 %rd6974, {%r4901, %r4900}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4902}, %rd6957; + } + { + .reg .b32 %dummy; + mov.b64 {%r4903,%dummy}, %rd6957; + } + shf.l.wrap.b32 %r4904, %r4903, %r4902, 8; + shf.l.wrap.b32 %r4905, %r4902, %r4903, 8; + mov.b64 %rd6975, {%r4905, %r4904}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4906}, %rd6958; + } + { + .reg .b32 %dummy; + mov.b64 {%r4907,%dummy}, %rd6958; + } + shf.l.wrap.b32 %r4908, %r4907, %r4906, 18; + shf.l.wrap.b32 %r4909, %r4906, %r4907, 18; + mov.b64 %rd6976, {%r4909, %r4908}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4910}, %rd6942; + } + { + .reg .b32 %dummy; + mov.b64 {%r4911,%dummy}, %rd6942; + } + shf.l.wrap.b32 %r4912, %r4911, %r4910, 27; + shf.l.wrap.b32 %r4913, %r4910, %r4911, 27; + mov.b64 %rd6977, {%r4913, %r4912}; + { + .reg .b32 %dummy; + mov.b64 {%r4914,%dummy}, %rd6943; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4915}, %rd6943; + } + shf.r.wrap.b32 %r4916, %r4915, %r4914, 28; + shf.r.wrap.b32 %r4917, %r4914, %r4915, 28; + mov.b64 %rd6978, {%r4917, %r4916}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4918}, %rd6949; + } + { + .reg .b32 %dummy; + mov.b64 {%r4919,%dummy}, %rd6949; + } + shf.l.wrap.b32 %r4920, %r4919, %r4918, 10; + shf.l.wrap.b32 %r4921, %r4918, %r4919, 10; + mov.b64 %rd6979, {%r4921, %r4920}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4922}, %rd6955; + } + { + .reg .b32 %dummy; + mov.b64 {%r4923,%dummy}, %rd6955; + } + shf.l.wrap.b32 %r4924, %r4923, %r4922, 15; + shf.l.wrap.b32 %r4925, %r4922, %r4923, 15; + mov.b64 %rd6980, {%r4925, %r4924}; + { + .reg .b32 %dummy; + mov.b64 {%r4926,%dummy}, %rd6961; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4927}, %rd6961; + } + shf.r.wrap.b32 %r4928, %r4927, %r4926, 8; + shf.r.wrap.b32 %r4929, %r4926, %r4927, 8; + mov.b64 %rd6981, {%r4929, %r4928}; + { + .reg .b32 %dummy; + mov.b64 {%r4930,%dummy}, %rd6940; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4931}, %rd6940; + } + shf.r.wrap.b32 %r4932, %r4931, %r4930, 2; + shf.r.wrap.b32 %r4933, %r4930, %r4931, 2; + mov.b64 %rd6982, {%r4933, %r4932}; + { + .reg .b32 %dummy; + mov.b64 {%r4934,%dummy}, %rd6946; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4935}, %rd6946; + } + shf.r.wrap.b32 %r4936, %r4935, %r4934, 9; + shf.r.wrap.b32 %r4937, %r4934, %r4935, 9; + mov.b64 %rd6983, {%r4937, %r4936}; + { + .reg .b32 %dummy; + mov.b64 {%r4938,%dummy}, %rd6952; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4939}, %rd6952; + } + shf.r.wrap.b32 %r4940, %r4939, %r4938, 25; + shf.r.wrap.b32 %r4941, %r4938, %r4939, 25; + mov.b64 %rd6984, {%r4941, %r4940}; + { + .reg .b32 %dummy; + mov.b64 {%r4942,%dummy}, %rd6953; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4943}, %rd6953; + } + shf.r.wrap.b32 %r4944, %r4943, %r4942, 23; + shf.r.wrap.b32 %r4945, %r4942, %r4943, 23; + mov.b64 %rd6985, {%r4945, %r4944}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4946}, %rd6959; + } + { + .reg .b32 %dummy; + mov.b64 {%r4947,%dummy}, %rd6959; + } + shf.l.wrap.b32 %r4948, %r4947, %r4946, 2; + shf.l.wrap.b32 %r4949, %r4946, %r4947, 2; + mov.b64 %rd6986, {%r4949, %r4948}; + not.b64 %rd6987, %rd6963; + and.b64 %rd6988, %rd6964, %rd6987; + xor.b64 %rd6989, %rd6988, %rd6938; + not.b64 %rd6990, %rd6964; + and.b64 %rd6991, %rd6965, %rd6990; + xor.b64 %rd6992, %rd6991, %rd6963; + not.b64 %rd6993, %rd6965; + and.b64 %rd6994, %rd6966, %rd6993; + xor.b64 %rd6995, %rd6994, %rd6964; + not.b64 %rd6996, %rd6966; + and.b64 %rd6997, %rd6938, %rd6996; + xor.b64 %rd6998, %rd6997, %rd6965; + not.b64 %rd6999, %rd6938; + and.b64 %rd7000, %rd6963, %rd6999; + xor.b64 %rd7001, %rd7000, %rd6966; + not.b64 %rd7002, %rd6968; + and.b64 %rd7003, %rd6969, %rd7002; + xor.b64 %rd7004, %rd7003, %rd6967; + not.b64 %rd7005, %rd6969; + and.b64 %rd7006, %rd6970, %rd7005; + xor.b64 %rd7007, %rd7006, %rd6968; + not.b64 %rd7008, %rd6970; + and.b64 %rd7009, %rd6971, %rd7008; + xor.b64 %rd7010, %rd7009, %rd6969; + not.b64 %rd7011, %rd6971; + and.b64 %rd7012, %rd6967, %rd7011; + xor.b64 %rd7013, %rd7012, %rd6970; + not.b64 %rd7014, %rd6967; + and.b64 %rd7015, %rd6968, %rd7014; + xor.b64 %rd7016, %rd7015, %rd6971; + not.b64 %rd7017, %rd6973; + and.b64 %rd7018, %rd6974, %rd7017; + xor.b64 %rd7019, %rd7018, %rd6972; + not.b64 %rd7020, %rd6974; + and.b64 %rd7021, %rd6975, %rd7020; + xor.b64 %rd7022, %rd7021, %rd6973; + not.b64 %rd7023, %rd6975; + and.b64 %rd7024, %rd6976, %rd7023; + xor.b64 %rd7025, %rd7024, %rd6974; + not.b64 %rd7026, %rd6976; + and.b64 %rd7027, %rd6972, %rd7026; + xor.b64 %rd7028, %rd7027, %rd6975; + not.b64 %rd7029, %rd6972; + and.b64 %rd7030, %rd6973, %rd7029; + xor.b64 %rd7031, %rd7030, %rd6976; + not.b64 %rd7032, %rd6978; + and.b64 %rd7033, %rd6979, %rd7032; + xor.b64 %rd7034, %rd7033, %rd6977; + not.b64 %rd7035, %rd6979; + and.b64 %rd7036, %rd6980, %rd7035; + xor.b64 %rd7037, %rd7036, %rd6978; + not.b64 %rd7038, %rd6980; + and.b64 %rd7039, %rd6981, %rd7038; + xor.b64 %rd7040, %rd7039, %rd6979; + not.b64 %rd7041, %rd6981; + and.b64 %rd7042, %rd6977, %rd7041; + xor.b64 %rd7043, %rd7042, %rd6980; + not.b64 %rd7044, %rd6977; + and.b64 %rd7045, %rd6978, %rd7044; + xor.b64 %rd7046, %rd7045, %rd6981; + not.b64 %rd7047, %rd6983; + and.b64 %rd7048, %rd6984, %rd7047; + xor.b64 %rd7049, %rd7048, %rd6982; + not.b64 %rd7050, %rd6984; + and.b64 %rd7051, %rd6985, %rd7050; + xor.b64 %rd7052, %rd7051, %rd6983; + not.b64 %rd7053, %rd6985; + and.b64 %rd7054, %rd6986, %rd7053; + xor.b64 %rd7055, %rd7054, %rd6984; + not.b64 %rd7056, %rd6986; + and.b64 %rd7057, %rd6982, %rd7056; + xor.b64 %rd7058, %rd7057, %rd6985; + not.b64 %rd7059, %rd6982; + and.b64 %rd7060, %rd6983, %rd7059; + xor.b64 %rd7061, %rd7060, %rd6986; + xor.b64 %rd7062, %rd6989, %rd33; + xor.b64 %rd7063, %rd7004, %rd7062; + xor.b64 %rd7064, %rd7063, %rd7019; + xor.b64 %rd7065, %rd7064, %rd7034; + xor.b64 %rd7066, %rd7065, %rd7049; + xor.b64 %rd7067, %rd7007, %rd6992; + xor.b64 %rd7068, %rd7067, %rd7022; + xor.b64 %rd7069, %rd7068, %rd7037; + xor.b64 %rd7070, %rd7069, %rd7052; + xor.b64 %rd7071, %rd7010, %rd6995; + xor.b64 %rd7072, %rd7071, %rd7025; + xor.b64 %rd7073, %rd7072, %rd7040; + xor.b64 %rd7074, %rd7073, %rd7055; + xor.b64 %rd7075, %rd7013, %rd6998; + xor.b64 %rd7076, %rd7075, %rd7028; + xor.b64 %rd7077, %rd7076, %rd7043; + xor.b64 %rd7078, %rd7077, %rd7058; + xor.b64 %rd7079, %rd7016, %rd7001; + xor.b64 %rd7080, %rd7079, %rd7031; + xor.b64 %rd7081, %rd7080, %rd7046; + xor.b64 %rd7082, %rd7081, %rd7061; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4950}, %rd7070; + } + { + .reg .b32 %dummy; + mov.b64 {%r4951,%dummy}, %rd7070; + } + shf.l.wrap.b32 %r4952, %r4951, %r4950, 1; + shf.l.wrap.b32 %r4953, %r4950, %r4951, 1; + mov.b64 %rd7083, {%r4953, %r4952}; + xor.b64 %rd7084, %rd7082, %rd7083; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4954}, %rd7074; + } + { + .reg .b32 %dummy; + mov.b64 {%r4955,%dummy}, %rd7074; + } + shf.l.wrap.b32 %r4956, %r4955, %r4954, 1; + shf.l.wrap.b32 %r4957, %r4954, %r4955, 1; + mov.b64 %rd7085, {%r4957, %r4956}; + xor.b64 %rd7086, %rd7085, %rd7066; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4958}, %rd7078; + } + { + .reg .b32 %dummy; + mov.b64 {%r4959,%dummy}, %rd7078; + } + shf.l.wrap.b32 %r4960, %r4959, %r4958, 1; + shf.l.wrap.b32 %r4961, %r4958, %r4959, 1; + mov.b64 %rd7087, {%r4961, %r4960}; + xor.b64 %rd7088, %rd7087, %rd7070; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4962}, %rd7082; + } + { + .reg .b32 %dummy; + mov.b64 {%r4963,%dummy}, %rd7082; + } + shf.l.wrap.b32 %r4964, %r4963, %r4962, 1; + shf.l.wrap.b32 %r4965, %r4962, %r4963, 1; + mov.b64 %rd7089, {%r4965, %r4964}; + xor.b64 %rd7090, %rd7089, %rd7074; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4966}, %rd7066; + } + { + .reg .b32 %dummy; + mov.b64 {%r4967,%dummy}, %rd7066; + } + shf.l.wrap.b32 %r4968, %r4967, %r4966, 1; + shf.l.wrap.b32 %r4969, %r4966, %r4967, 1; + mov.b64 %rd7091, {%r4969, %r4968}; + xor.b64 %rd7092, %rd7078, %rd7091; + xor.b64 %rd7093, %rd7084, %rd7062; + xor.b64 %rd7094, %rd7086, %rd6992; + xor.b64 %rd7095, %rd7088, %rd6995; + xor.b64 %rd7096, %rd7090, %rd6998; + xor.b64 %rd7097, %rd7092, %rd7001; + xor.b64 %rd7098, %rd7084, %rd7004; + xor.b64 %rd7099, %rd7086, %rd7007; + xor.b64 %rd7100, %rd7088, %rd7010; + xor.b64 %rd7101, %rd7090, %rd7013; + xor.b64 %rd7102, %rd7092, %rd7016; + xor.b64 %rd7103, %rd7084, %rd7019; + xor.b64 %rd7104, %rd7086, %rd7022; + xor.b64 %rd7105, %rd7088, %rd7025; + xor.b64 %rd7106, %rd7090, %rd7028; + xor.b64 %rd7107, %rd7092, %rd7031; + xor.b64 %rd7108, %rd7084, %rd7034; + xor.b64 %rd7109, %rd7086, %rd7037; + xor.b64 %rd7110, %rd7088, %rd7040; + xor.b64 %rd7111, %rd7090, %rd7043; + xor.b64 %rd7112, %rd7092, %rd7046; + xor.b64 %rd7113, %rd7084, %rd7049; + xor.b64 %rd7114, %rd7086, %rd7052; + xor.b64 %rd7115, %rd7088, %rd7055; + xor.b64 %rd7116, %rd7090, %rd7058; + xor.b64 %rd7117, %rd7092, %rd7061; + { + .reg .b32 %dummy; + mov.b64 {%r4970,%dummy}, %rd7099; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4971}, %rd7099; + } + shf.r.wrap.b32 %r4972, %r4971, %r4970, 20; + shf.r.wrap.b32 %r4973, %r4970, %r4971, 20; + mov.b64 %rd7118, {%r4973, %r4972}; + { + .reg .b32 %dummy; + mov.b64 {%r4974,%dummy}, %rd7105; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4975}, %rd7105; + } + shf.r.wrap.b32 %r4976, %r4975, %r4974, 21; + shf.r.wrap.b32 %r4977, %r4974, %r4975, 21; + mov.b64 %rd7119, {%r4977, %r4976}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4978}, %rd7111; + } + { + .reg .b32 %dummy; + mov.b64 {%r4979,%dummy}, %rd7111; + } + shf.l.wrap.b32 %r4980, %r4979, %r4978, 21; + shf.l.wrap.b32 %r4981, %r4978, %r4979, 21; + mov.b64 %rd7120, {%r4981, %r4980}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4982}, %rd7117; + } + { + .reg .b32 %dummy; + mov.b64 {%r4983,%dummy}, %rd7117; + } + shf.l.wrap.b32 %r4984, %r4983, %r4982, 14; + shf.l.wrap.b32 %r4985, %r4982, %r4983, 14; + mov.b64 %rd7121, {%r4985, %r4984}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4986}, %rd7096; + } + { + .reg .b32 %dummy; + mov.b64 {%r4987,%dummy}, %rd7096; + } + shf.l.wrap.b32 %r4988, %r4987, %r4986, 28; + shf.l.wrap.b32 %r4989, %r4986, %r4987, 28; + mov.b64 %rd7122, {%r4989, %r4988}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4990}, %rd7102; + } + { + .reg .b32 %dummy; + mov.b64 {%r4991,%dummy}, %rd7102; + } + shf.l.wrap.b32 %r4992, %r4991, %r4990, 20; + shf.l.wrap.b32 %r4993, %r4990, %r4991, 20; + mov.b64 %rd7123, {%r4993, %r4992}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4994}, %rd7103; + } + { + .reg .b32 %dummy; + mov.b64 {%r4995,%dummy}, %rd7103; + } + shf.l.wrap.b32 %r4996, %r4995, %r4994, 3; + shf.l.wrap.b32 %r4997, %r4994, %r4995, 3; + mov.b64 %rd7124, {%r4997, %r4996}; + { + .reg .b32 %dummy; + mov.b64 {%r4998,%dummy}, %rd7109; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r4999}, %rd7109; + } + shf.r.wrap.b32 %r5000, %r4999, %r4998, 19; + shf.r.wrap.b32 %r5001, %r4998, %r4999, 19; + mov.b64 %rd7125, {%r5001, %r5000}; + { + .reg .b32 %dummy; + mov.b64 {%r5002,%dummy}, %rd7115; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5003}, %rd7115; + } + shf.r.wrap.b32 %r5004, %r5003, %r5002, 3; + shf.r.wrap.b32 %r5005, %r5002, %r5003, 3; + mov.b64 %rd7126, {%r5005, %r5004}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5006}, %rd7094; + } + { + .reg .b32 %dummy; + mov.b64 {%r5007,%dummy}, %rd7094; + } + shf.l.wrap.b32 %r5008, %r5007, %r5006, 1; + shf.l.wrap.b32 %r5009, %r5006, %r5007, 1; + mov.b64 %rd7127, {%r5009, %r5008}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5010}, %rd7100; + } + { + .reg .b32 %dummy; + mov.b64 {%r5011,%dummy}, %rd7100; + } + shf.l.wrap.b32 %r5012, %r5011, %r5010, 6; + shf.l.wrap.b32 %r5013, %r5010, %r5011, 6; + mov.b64 %rd7128, {%r5013, %r5012}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5014}, %rd7106; + } + { + .reg .b32 %dummy; + mov.b64 {%r5015,%dummy}, %rd7106; + } + shf.l.wrap.b32 %r5016, %r5015, %r5014, 25; + shf.l.wrap.b32 %r5017, %r5014, %r5015, 25; + mov.b64 %rd7129, {%r5017, %r5016}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5018}, %rd7112; + } + { + .reg .b32 %dummy; + mov.b64 {%r5019,%dummy}, %rd7112; + } + shf.l.wrap.b32 %r5020, %r5019, %r5018, 8; + shf.l.wrap.b32 %r5021, %r5018, %r5019, 8; + mov.b64 %rd7130, {%r5021, %r5020}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5022}, %rd7113; + } + { + .reg .b32 %dummy; + mov.b64 {%r5023,%dummy}, %rd7113; + } + shf.l.wrap.b32 %r5024, %r5023, %r5022, 18; + shf.l.wrap.b32 %r5025, %r5022, %r5023, 18; + mov.b64 %rd7131, {%r5025, %r5024}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5026}, %rd7097; + } + { + .reg .b32 %dummy; + mov.b64 {%r5027,%dummy}, %rd7097; + } + shf.l.wrap.b32 %r5028, %r5027, %r5026, 27; + shf.l.wrap.b32 %r5029, %r5026, %r5027, 27; + mov.b64 %rd7132, {%r5029, %r5028}; + { + .reg .b32 %dummy; + mov.b64 {%r5030,%dummy}, %rd7098; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5031}, %rd7098; + } + shf.r.wrap.b32 %r5032, %r5031, %r5030, 28; + shf.r.wrap.b32 %r5033, %r5030, %r5031, 28; + mov.b64 %rd7133, {%r5033, %r5032}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5034}, %rd7104; + } + { + .reg .b32 %dummy; + mov.b64 {%r5035,%dummy}, %rd7104; + } + shf.l.wrap.b32 %r5036, %r5035, %r5034, 10; + shf.l.wrap.b32 %r5037, %r5034, %r5035, 10; + mov.b64 %rd7134, {%r5037, %r5036}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5038}, %rd7110; + } + { + .reg .b32 %dummy; + mov.b64 {%r5039,%dummy}, %rd7110; + } + shf.l.wrap.b32 %r5040, %r5039, %r5038, 15; + shf.l.wrap.b32 %r5041, %r5038, %r5039, 15; + mov.b64 %rd7135, {%r5041, %r5040}; + { + .reg .b32 %dummy; + mov.b64 {%r5042,%dummy}, %rd7116; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5043}, %rd7116; + } + shf.r.wrap.b32 %r5044, %r5043, %r5042, 8; + shf.r.wrap.b32 %r5045, %r5042, %r5043, 8; + mov.b64 %rd7136, {%r5045, %r5044}; + { + .reg .b32 %dummy; + mov.b64 {%r5046,%dummy}, %rd7095; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5047}, %rd7095; + } + shf.r.wrap.b32 %r5048, %r5047, %r5046, 2; + shf.r.wrap.b32 %r5049, %r5046, %r5047, 2; + mov.b64 %rd7137, {%r5049, %r5048}; + { + .reg .b32 %dummy; + mov.b64 {%r5050,%dummy}, %rd7101; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5051}, %rd7101; + } + shf.r.wrap.b32 %r5052, %r5051, %r5050, 9; + shf.r.wrap.b32 %r5053, %r5050, %r5051, 9; + mov.b64 %rd7138, {%r5053, %r5052}; + { + .reg .b32 %dummy; + mov.b64 {%r5054,%dummy}, %rd7107; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5055}, %rd7107; + } + shf.r.wrap.b32 %r5056, %r5055, %r5054, 25; + shf.r.wrap.b32 %r5057, %r5054, %r5055, 25; + mov.b64 %rd7139, {%r5057, %r5056}; + { + .reg .b32 %dummy; + mov.b64 {%r5058,%dummy}, %rd7108; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5059}, %rd7108; + } + shf.r.wrap.b32 %r5060, %r5059, %r5058, 23; + shf.r.wrap.b32 %r5061, %r5058, %r5059, 23; + mov.b64 %rd7140, {%r5061, %r5060}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5062}, %rd7114; + } + { + .reg .b32 %dummy; + mov.b64 {%r5063,%dummy}, %rd7114; + } + shf.l.wrap.b32 %r5064, %r5063, %r5062, 2; + shf.l.wrap.b32 %r5065, %r5062, %r5063, 2; + mov.b64 %rd7141, {%r5065, %r5064}; + not.b64 %rd7142, %rd7118; + and.b64 %rd7143, %rd7119, %rd7142; + xor.b64 %rd7144, %rd7143, %rd7093; + not.b64 %rd7145, %rd7119; + and.b64 %rd7146, %rd7120, %rd7145; + xor.b64 %rd7147, %rd7146, %rd7118; + not.b64 %rd7148, %rd7120; + and.b64 %rd7149, %rd7121, %rd7148; + xor.b64 %rd7150, %rd7149, %rd7119; + not.b64 %rd7151, %rd7121; + and.b64 %rd7152, %rd7093, %rd7151; + xor.b64 %rd7153, %rd7152, %rd7120; + not.b64 %rd7154, %rd7093; + and.b64 %rd7155, %rd7118, %rd7154; + xor.b64 %rd7156, %rd7155, %rd7121; + not.b64 %rd7157, %rd7123; + and.b64 %rd7158, %rd7124, %rd7157; + xor.b64 %rd7159, %rd7158, %rd7122; + not.b64 %rd7160, %rd7124; + and.b64 %rd7161, %rd7125, %rd7160; + xor.b64 %rd7162, %rd7161, %rd7123; + not.b64 %rd7163, %rd7125; + and.b64 %rd7164, %rd7126, %rd7163; + xor.b64 %rd7165, %rd7164, %rd7124; + not.b64 %rd7166, %rd7126; + and.b64 %rd7167, %rd7122, %rd7166; + xor.b64 %rd7168, %rd7167, %rd7125; + not.b64 %rd7169, %rd7122; + and.b64 %rd7170, %rd7123, %rd7169; + xor.b64 %rd7171, %rd7170, %rd7126; + not.b64 %rd7172, %rd7128; + and.b64 %rd7173, %rd7129, %rd7172; + xor.b64 %rd7174, %rd7173, %rd7127; + not.b64 %rd7175, %rd7129; + and.b64 %rd7176, %rd7130, %rd7175; + xor.b64 %rd7177, %rd7176, %rd7128; + not.b64 %rd7178, %rd7130; + and.b64 %rd7179, %rd7131, %rd7178; + xor.b64 %rd7180, %rd7179, %rd7129; + not.b64 %rd7181, %rd7131; + and.b64 %rd7182, %rd7127, %rd7181; + xor.b64 %rd7183, %rd7182, %rd7130; + not.b64 %rd7184, %rd7127; + and.b64 %rd7185, %rd7128, %rd7184; + xor.b64 %rd7186, %rd7185, %rd7131; + not.b64 %rd7187, %rd7133; + and.b64 %rd7188, %rd7134, %rd7187; + xor.b64 %rd7189, %rd7188, %rd7132; + not.b64 %rd7190, %rd7134; + and.b64 %rd7191, %rd7135, %rd7190; + xor.b64 %rd7192, %rd7191, %rd7133; + not.b64 %rd7193, %rd7135; + and.b64 %rd7194, %rd7136, %rd7193; + xor.b64 %rd7195, %rd7194, %rd7134; + not.b64 %rd7196, %rd7136; + and.b64 %rd7197, %rd7132, %rd7196; + xor.b64 %rd7198, %rd7197, %rd7135; + not.b64 %rd7199, %rd7132; + and.b64 %rd7200, %rd7133, %rd7199; + xor.b64 %rd7201, %rd7200, %rd7136; + not.b64 %rd7202, %rd7138; + and.b64 %rd7203, %rd7139, %rd7202; + xor.b64 %rd7204, %rd7203, %rd7137; + not.b64 %rd7205, %rd7139; + and.b64 %rd7206, %rd7140, %rd7205; + xor.b64 %rd7207, %rd7206, %rd7138; + not.b64 %rd7208, %rd7140; + and.b64 %rd7209, %rd7141, %rd7208; + xor.b64 %rd7210, %rd7209, %rd7139; + not.b64 %rd7211, %rd7141; + and.b64 %rd7212, %rd7137, %rd7211; + xor.b64 %rd7213, %rd7212, %rd7140; + not.b64 %rd7214, %rd7137; + and.b64 %rd7215, %rd7138, %rd7214; + xor.b64 %rd7216, %rd7215, %rd7141; + xor.b64 %rd7217, %rd7144, %rd34; + xor.b64 %rd7218, %rd7159, %rd7217; + xor.b64 %rd7219, %rd7218, %rd7174; + xor.b64 %rd7220, %rd7219, %rd7189; + xor.b64 %rd7221, %rd7220, %rd7204; + xor.b64 %rd7222, %rd7162, %rd7147; + xor.b64 %rd7223, %rd7222, %rd7177; + xor.b64 %rd7224, %rd7223, %rd7192; + xor.b64 %rd7225, %rd7224, %rd7207; + xor.b64 %rd7226, %rd7165, %rd7150; + xor.b64 %rd7227, %rd7226, %rd7180; + xor.b64 %rd7228, %rd7227, %rd7195; + xor.b64 %rd7229, %rd7228, %rd7210; + xor.b64 %rd7230, %rd7168, %rd7153; + xor.b64 %rd7231, %rd7230, %rd7183; + xor.b64 %rd7232, %rd7231, %rd7198; + xor.b64 %rd7233, %rd7232, %rd7213; + xor.b64 %rd7234, %rd7171, %rd7156; + xor.b64 %rd7235, %rd7234, %rd7186; + xor.b64 %rd7236, %rd7235, %rd7201; + xor.b64 %rd7237, %rd7236, %rd7216; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5066}, %rd7225; + } + { + .reg .b32 %dummy; + mov.b64 {%r5067,%dummy}, %rd7225; + } + shf.l.wrap.b32 %r5068, %r5067, %r5066, 1; + shf.l.wrap.b32 %r5069, %r5066, %r5067, 1; + mov.b64 %rd7238, {%r5069, %r5068}; + xor.b64 %rd7239, %rd7237, %rd7238; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5070}, %rd7229; + } + { + .reg .b32 %dummy; + mov.b64 {%r5071,%dummy}, %rd7229; + } + shf.l.wrap.b32 %r5072, %r5071, %r5070, 1; + shf.l.wrap.b32 %r5073, %r5070, %r5071, 1; + mov.b64 %rd7240, {%r5073, %r5072}; + xor.b64 %rd7241, %rd7240, %rd7221; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5074}, %rd7233; + } + { + .reg .b32 %dummy; + mov.b64 {%r5075,%dummy}, %rd7233; + } + shf.l.wrap.b32 %r5076, %r5075, %r5074, 1; + shf.l.wrap.b32 %r5077, %r5074, %r5075, 1; + mov.b64 %rd7242, {%r5077, %r5076}; + xor.b64 %rd7243, %rd7242, %rd7225; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5078}, %rd7237; + } + { + .reg .b32 %dummy; + mov.b64 {%r5079,%dummy}, %rd7237; + } + shf.l.wrap.b32 %r5080, %r5079, %r5078, 1; + shf.l.wrap.b32 %r5081, %r5078, %r5079, 1; + mov.b64 %rd7244, {%r5081, %r5080}; + xor.b64 %rd7245, %rd7244, %rd7229; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5082}, %rd7221; + } + { + .reg .b32 %dummy; + mov.b64 {%r5083,%dummy}, %rd7221; + } + shf.l.wrap.b32 %r5084, %r5083, %r5082, 1; + shf.l.wrap.b32 %r5085, %r5082, %r5083, 1; + mov.b64 %rd7246, {%r5085, %r5084}; + xor.b64 %rd7247, %rd7233, %rd7246; + xor.b64 %rd7248, %rd7239, %rd7217; + xor.b64 %rd7249, %rd7241, %rd7147; + xor.b64 %rd7250, %rd7243, %rd7150; + xor.b64 %rd7251, %rd7245, %rd7153; + xor.b64 %rd7252, %rd7247, %rd7156; + xor.b64 %rd7253, %rd7239, %rd7159; + xor.b64 %rd7254, %rd7241, %rd7162; + xor.b64 %rd7255, %rd7243, %rd7165; + xor.b64 %rd7256, %rd7245, %rd7168; + xor.b64 %rd7257, %rd7247, %rd7171; + xor.b64 %rd7258, %rd7239, %rd7174; + xor.b64 %rd7259, %rd7241, %rd7177; + xor.b64 %rd7260, %rd7243, %rd7180; + xor.b64 %rd7261, %rd7245, %rd7183; + xor.b64 %rd7262, %rd7247, %rd7186; + xor.b64 %rd7263, %rd7239, %rd7189; + xor.b64 %rd7264, %rd7241, %rd7192; + xor.b64 %rd7265, %rd7243, %rd7195; + xor.b64 %rd7266, %rd7245, %rd7198; + xor.b64 %rd7267, %rd7247, %rd7201; + xor.b64 %rd7268, %rd7239, %rd7204; + xor.b64 %rd7269, %rd7241, %rd7207; + xor.b64 %rd7270, %rd7243, %rd7210; + xor.b64 %rd7271, %rd7245, %rd7213; + xor.b64 %rd7272, %rd7247, %rd7216; + { + .reg .b32 %dummy; + mov.b64 {%r5086,%dummy}, %rd7254; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5087}, %rd7254; + } + shf.r.wrap.b32 %r5088, %r5087, %r5086, 20; + shf.r.wrap.b32 %r5089, %r5086, %r5087, 20; + mov.b64 %rd7273, {%r5089, %r5088}; + { + .reg .b32 %dummy; + mov.b64 {%r5090,%dummy}, %rd7260; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5091}, %rd7260; + } + shf.r.wrap.b32 %r5092, %r5091, %r5090, 21; + shf.r.wrap.b32 %r5093, %r5090, %r5091, 21; + mov.b64 %rd7274, {%r5093, %r5092}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5094}, %rd7266; + } + { + .reg .b32 %dummy; + mov.b64 {%r5095,%dummy}, %rd7266; + } + shf.l.wrap.b32 %r5096, %r5095, %r5094, 21; + shf.l.wrap.b32 %r5097, %r5094, %r5095, 21; + mov.b64 %rd7275, {%r5097, %r5096}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5098}, %rd7272; + } + { + .reg .b32 %dummy; + mov.b64 {%r5099,%dummy}, %rd7272; + } + shf.l.wrap.b32 %r5100, %r5099, %r5098, 14; + shf.l.wrap.b32 %r5101, %r5098, %r5099, 14; + mov.b64 %rd7276, {%r5101, %r5100}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5102}, %rd7251; + } + { + .reg .b32 %dummy; + mov.b64 {%r5103,%dummy}, %rd7251; + } + shf.l.wrap.b32 %r5104, %r5103, %r5102, 28; + shf.l.wrap.b32 %r5105, %r5102, %r5103, 28; + mov.b64 %rd7277, {%r5105, %r5104}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5106}, %rd7257; + } + { + .reg .b32 %dummy; + mov.b64 {%r5107,%dummy}, %rd7257; + } + shf.l.wrap.b32 %r5108, %r5107, %r5106, 20; + shf.l.wrap.b32 %r5109, %r5106, %r5107, 20; + mov.b64 %rd7278, {%r5109, %r5108}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5110}, %rd7258; + } + { + .reg .b32 %dummy; + mov.b64 {%r5111,%dummy}, %rd7258; + } + shf.l.wrap.b32 %r5112, %r5111, %r5110, 3; + shf.l.wrap.b32 %r5113, %r5110, %r5111, 3; + mov.b64 %rd7279, {%r5113, %r5112}; + { + .reg .b32 %dummy; + mov.b64 {%r5114,%dummy}, %rd7264; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5115}, %rd7264; + } + shf.r.wrap.b32 %r5116, %r5115, %r5114, 19; + shf.r.wrap.b32 %r5117, %r5114, %r5115, 19; + mov.b64 %rd7280, {%r5117, %r5116}; + { + .reg .b32 %dummy; + mov.b64 {%r5118,%dummy}, %rd7270; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5119}, %rd7270; + } + shf.r.wrap.b32 %r5120, %r5119, %r5118, 3; + shf.r.wrap.b32 %r5121, %r5118, %r5119, 3; + mov.b64 %rd7281, {%r5121, %r5120}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5122}, %rd7249; + } + { + .reg .b32 %dummy; + mov.b64 {%r5123,%dummy}, %rd7249; + } + shf.l.wrap.b32 %r5124, %r5123, %r5122, 1; + shf.l.wrap.b32 %r5125, %r5122, %r5123, 1; + mov.b64 %rd7282, {%r5125, %r5124}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5126}, %rd7255; + } + { + .reg .b32 %dummy; + mov.b64 {%r5127,%dummy}, %rd7255; + } + shf.l.wrap.b32 %r5128, %r5127, %r5126, 6; + shf.l.wrap.b32 %r5129, %r5126, %r5127, 6; + mov.b64 %rd7283, {%r5129, %r5128}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5130}, %rd7261; + } + { + .reg .b32 %dummy; + mov.b64 {%r5131,%dummy}, %rd7261; + } + shf.l.wrap.b32 %r5132, %r5131, %r5130, 25; + shf.l.wrap.b32 %r5133, %r5130, %r5131, 25; + mov.b64 %rd7284, {%r5133, %r5132}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5134}, %rd7267; + } + { + .reg .b32 %dummy; + mov.b64 {%r5135,%dummy}, %rd7267; + } + shf.l.wrap.b32 %r5136, %r5135, %r5134, 8; + shf.l.wrap.b32 %r5137, %r5134, %r5135, 8; + mov.b64 %rd7285, {%r5137, %r5136}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5138}, %rd7268; + } + { + .reg .b32 %dummy; + mov.b64 {%r5139,%dummy}, %rd7268; + } + shf.l.wrap.b32 %r5140, %r5139, %r5138, 18; + shf.l.wrap.b32 %r5141, %r5138, %r5139, 18; + mov.b64 %rd7286, {%r5141, %r5140}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5142}, %rd7252; + } + { + .reg .b32 %dummy; + mov.b64 {%r5143,%dummy}, %rd7252; + } + shf.l.wrap.b32 %r5144, %r5143, %r5142, 27; + shf.l.wrap.b32 %r5145, %r5142, %r5143, 27; + mov.b64 %rd7287, {%r5145, %r5144}; + { + .reg .b32 %dummy; + mov.b64 {%r5146,%dummy}, %rd7253; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5147}, %rd7253; + } + shf.r.wrap.b32 %r5148, %r5147, %r5146, 28; + shf.r.wrap.b32 %r5149, %r5146, %r5147, 28; + mov.b64 %rd7288, {%r5149, %r5148}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5150}, %rd7259; + } + { + .reg .b32 %dummy; + mov.b64 {%r5151,%dummy}, %rd7259; + } + shf.l.wrap.b32 %r5152, %r5151, %r5150, 10; + shf.l.wrap.b32 %r5153, %r5150, %r5151, 10; + mov.b64 %rd7289, {%r5153, %r5152}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5154}, %rd7265; + } + { + .reg .b32 %dummy; + mov.b64 {%r5155,%dummy}, %rd7265; + } + shf.l.wrap.b32 %r5156, %r5155, %r5154, 15; + shf.l.wrap.b32 %r5157, %r5154, %r5155, 15; + mov.b64 %rd7290, {%r5157, %r5156}; + { + .reg .b32 %dummy; + mov.b64 {%r5158,%dummy}, %rd7271; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5159}, %rd7271; + } + shf.r.wrap.b32 %r5160, %r5159, %r5158, 8; + shf.r.wrap.b32 %r5161, %r5158, %r5159, 8; + mov.b64 %rd7291, {%r5161, %r5160}; + { + .reg .b32 %dummy; + mov.b64 {%r5162,%dummy}, %rd7250; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5163}, %rd7250; + } + shf.r.wrap.b32 %r5164, %r5163, %r5162, 2; + shf.r.wrap.b32 %r5165, %r5162, %r5163, 2; + mov.b64 %rd7292, {%r5165, %r5164}; + { + .reg .b32 %dummy; + mov.b64 {%r5166,%dummy}, %rd7256; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5167}, %rd7256; + } + shf.r.wrap.b32 %r5168, %r5167, %r5166, 9; + shf.r.wrap.b32 %r5169, %r5166, %r5167, 9; + mov.b64 %rd7293, {%r5169, %r5168}; + { + .reg .b32 %dummy; + mov.b64 {%r5170,%dummy}, %rd7262; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5171}, %rd7262; + } + shf.r.wrap.b32 %r5172, %r5171, %r5170, 25; + shf.r.wrap.b32 %r5173, %r5170, %r5171, 25; + mov.b64 %rd7294, {%r5173, %r5172}; + { + .reg .b32 %dummy; + mov.b64 {%r5174,%dummy}, %rd7263; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5175}, %rd7263; + } + shf.r.wrap.b32 %r5176, %r5175, %r5174, 23; + shf.r.wrap.b32 %r5177, %r5174, %r5175, 23; + mov.b64 %rd7295, {%r5177, %r5176}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5178}, %rd7269; + } + { + .reg .b32 %dummy; + mov.b64 {%r5179,%dummy}, %rd7269; + } + shf.l.wrap.b32 %r5180, %r5179, %r5178, 2; + shf.l.wrap.b32 %r5181, %r5178, %r5179, 2; + mov.b64 %rd7296, {%r5181, %r5180}; + not.b64 %rd7297, %rd7273; + and.b64 %rd7298, %rd7274, %rd7297; + xor.b64 %rd7299, %rd7298, %rd7248; + not.b64 %rd7300, %rd7274; + and.b64 %rd7301, %rd7275, %rd7300; + xor.b64 %rd7302, %rd7301, %rd7273; + not.b64 %rd7303, %rd7275; + and.b64 %rd7304, %rd7276, %rd7303; + xor.b64 %rd7305, %rd7304, %rd7274; + not.b64 %rd7306, %rd7276; + and.b64 %rd7307, %rd7248, %rd7306; + xor.b64 %rd7308, %rd7307, %rd7275; + not.b64 %rd7309, %rd7248; + and.b64 %rd7310, %rd7273, %rd7309; + xor.b64 %rd7311, %rd7310, %rd7276; + not.b64 %rd7312, %rd7278; + and.b64 %rd7313, %rd7279, %rd7312; + xor.b64 %rd7314, %rd7313, %rd7277; + not.b64 %rd7315, %rd7279; + and.b64 %rd7316, %rd7280, %rd7315; + xor.b64 %rd7317, %rd7316, %rd7278; + not.b64 %rd7318, %rd7280; + and.b64 %rd7319, %rd7281, %rd7318; + xor.b64 %rd7320, %rd7319, %rd7279; + not.b64 %rd7321, %rd7281; + and.b64 %rd7322, %rd7277, %rd7321; + xor.b64 %rd7323, %rd7322, %rd7280; + not.b64 %rd7324, %rd7277; + and.b64 %rd7325, %rd7278, %rd7324; + xor.b64 %rd7326, %rd7325, %rd7281; + not.b64 %rd7327, %rd7283; + and.b64 %rd7328, %rd7284, %rd7327; + xor.b64 %rd7329, %rd7328, %rd7282; + not.b64 %rd7330, %rd7284; + and.b64 %rd7331, %rd7285, %rd7330; + xor.b64 %rd7332, %rd7331, %rd7283; + not.b64 %rd7333, %rd7285; + and.b64 %rd7334, %rd7286, %rd7333; + xor.b64 %rd7335, %rd7334, %rd7284; + not.b64 %rd7336, %rd7286; + and.b64 %rd7337, %rd7282, %rd7336; + xor.b64 %rd7338, %rd7337, %rd7285; + not.b64 %rd7339, %rd7282; + and.b64 %rd7340, %rd7283, %rd7339; + xor.b64 %rd7341, %rd7340, %rd7286; + not.b64 %rd7342, %rd7288; + and.b64 %rd7343, %rd7289, %rd7342; + xor.b64 %rd7344, %rd7343, %rd7287; + not.b64 %rd7345, %rd7289; + and.b64 %rd7346, %rd7290, %rd7345; + xor.b64 %rd7347, %rd7346, %rd7288; + not.b64 %rd7348, %rd7290; + and.b64 %rd7349, %rd7291, %rd7348; + xor.b64 %rd7350, %rd7349, %rd7289; + not.b64 %rd7351, %rd7291; + and.b64 %rd7352, %rd7287, %rd7351; + xor.b64 %rd7353, %rd7352, %rd7290; + not.b64 %rd7354, %rd7287; + and.b64 %rd7355, %rd7288, %rd7354; + xor.b64 %rd7356, %rd7355, %rd7291; + not.b64 %rd7357, %rd7293; + and.b64 %rd7358, %rd7294, %rd7357; + xor.b64 %rd7359, %rd7358, %rd7292; + not.b64 %rd7360, %rd7294; + and.b64 %rd7361, %rd7295, %rd7360; + xor.b64 %rd7362, %rd7361, %rd7293; + not.b64 %rd7363, %rd7295; + and.b64 %rd7364, %rd7296, %rd7363; + xor.b64 %rd7365, %rd7364, %rd7294; + not.b64 %rd7366, %rd7296; + and.b64 %rd7367, %rd7292, %rd7366; + xor.b64 %rd7368, %rd7367, %rd7295; + not.b64 %rd7369, %rd7292; + and.b64 %rd7370, %rd7293, %rd7369; + xor.b64 %rd7371, %rd7370, %rd7296; + xor.b64 %rd7372, %rd7299, %rd35; + xor.b64 %rd7373, %rd7314, %rd7372; + xor.b64 %rd7374, %rd7373, %rd7329; + xor.b64 %rd7375, %rd7374, %rd7344; + xor.b64 %rd7376, %rd7375, %rd7359; + xor.b64 %rd7377, %rd7317, %rd7302; + xor.b64 %rd7378, %rd7377, %rd7332; + xor.b64 %rd7379, %rd7378, %rd7347; + xor.b64 %rd7380, %rd7379, %rd7362; + xor.b64 %rd7381, %rd7320, %rd7305; + xor.b64 %rd7382, %rd7381, %rd7335; + xor.b64 %rd7383, %rd7382, %rd7350; + xor.b64 %rd7384, %rd7383, %rd7365; + xor.b64 %rd7385, %rd7323, %rd7308; + xor.b64 %rd7386, %rd7385, %rd7338; + xor.b64 %rd7387, %rd7386, %rd7353; + xor.b64 %rd7388, %rd7387, %rd7368; + xor.b64 %rd7389, %rd7326, %rd7311; + xor.b64 %rd7390, %rd7389, %rd7341; + xor.b64 %rd7391, %rd7390, %rd7356; + xor.b64 %rd7392, %rd7391, %rd7371; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5182}, %rd7380; + } + { + .reg .b32 %dummy; + mov.b64 {%r5183,%dummy}, %rd7380; + } + shf.l.wrap.b32 %r5184, %r5183, %r5182, 1; + shf.l.wrap.b32 %r5185, %r5182, %r5183, 1; + mov.b64 %rd7393, {%r5185, %r5184}; + xor.b64 %rd7394, %rd7392, %rd7393; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5186}, %rd7384; + } + { + .reg .b32 %dummy; + mov.b64 {%r5187,%dummy}, %rd7384; + } + shf.l.wrap.b32 %r5188, %r5187, %r5186, 1; + shf.l.wrap.b32 %r5189, %r5186, %r5187, 1; + mov.b64 %rd7395, {%r5189, %r5188}; + xor.b64 %rd7396, %rd7395, %rd7376; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5190}, %rd7388; + } + { + .reg .b32 %dummy; + mov.b64 {%r5191,%dummy}, %rd7388; + } + shf.l.wrap.b32 %r5192, %r5191, %r5190, 1; + shf.l.wrap.b32 %r5193, %r5190, %r5191, 1; + mov.b64 %rd7397, {%r5193, %r5192}; + xor.b64 %rd7398, %rd7397, %rd7380; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5194}, %rd7392; + } + { + .reg .b32 %dummy; + mov.b64 {%r5195,%dummy}, %rd7392; + } + shf.l.wrap.b32 %r5196, %r5195, %r5194, 1; + shf.l.wrap.b32 %r5197, %r5194, %r5195, 1; + mov.b64 %rd7399, {%r5197, %r5196}; + xor.b64 %rd7400, %rd7399, %rd7384; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5198}, %rd7376; + } + { + .reg .b32 %dummy; + mov.b64 {%r5199,%dummy}, %rd7376; + } + shf.l.wrap.b32 %r5200, %r5199, %r5198, 1; + shf.l.wrap.b32 %r5201, %r5198, %r5199, 1; + mov.b64 %rd7401, {%r5201, %r5200}; + xor.b64 %rd7402, %rd7388, %rd7401; + xor.b64 %rd7403, %rd7394, %rd7372; + xor.b64 %rd7404, %rd7396, %rd7302; + xor.b64 %rd7405, %rd7398, %rd7305; + xor.b64 %rd7406, %rd7400, %rd7308; + xor.b64 %rd7407, %rd7402, %rd7311; + xor.b64 %rd7408, %rd7394, %rd7314; + xor.b64 %rd7409, %rd7396, %rd7317; + xor.b64 %rd7410, %rd7398, %rd7320; + xor.b64 %rd7411, %rd7400, %rd7323; + xor.b64 %rd7412, %rd7402, %rd7326; + xor.b64 %rd7413, %rd7394, %rd7329; + xor.b64 %rd7414, %rd7396, %rd7332; + xor.b64 %rd7415, %rd7398, %rd7335; + xor.b64 %rd7416, %rd7400, %rd7338; + xor.b64 %rd7417, %rd7402, %rd7341; + xor.b64 %rd7418, %rd7394, %rd7344; + xor.b64 %rd7419, %rd7396, %rd7347; + xor.b64 %rd7420, %rd7398, %rd7350; + xor.b64 %rd7421, %rd7400, %rd7353; + xor.b64 %rd7422, %rd7402, %rd7356; + xor.b64 %rd7423, %rd7394, %rd7359; + xor.b64 %rd7424, %rd7396, %rd7362; + xor.b64 %rd7425, %rd7398, %rd7365; + xor.b64 %rd7426, %rd7400, %rd7368; + xor.b64 %rd7427, %rd7402, %rd7371; + { + .reg .b32 %dummy; + mov.b64 {%r5202,%dummy}, %rd7409; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5203}, %rd7409; + } + shf.r.wrap.b32 %r5204, %r5203, %r5202, 20; + shf.r.wrap.b32 %r5205, %r5202, %r5203, 20; + mov.b64 %rd7428, {%r5205, %r5204}; + { + .reg .b32 %dummy; + mov.b64 {%r5206,%dummy}, %rd7415; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5207}, %rd7415; + } + shf.r.wrap.b32 %r5208, %r5207, %r5206, 21; + shf.r.wrap.b32 %r5209, %r5206, %r5207, 21; + mov.b64 %rd7429, {%r5209, %r5208}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5210}, %rd7421; + } + { + .reg .b32 %dummy; + mov.b64 {%r5211,%dummy}, %rd7421; + } + shf.l.wrap.b32 %r5212, %r5211, %r5210, 21; + shf.l.wrap.b32 %r5213, %r5210, %r5211, 21; + mov.b64 %rd7430, {%r5213, %r5212}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5214}, %rd7427; + } + { + .reg .b32 %dummy; + mov.b64 {%r5215,%dummy}, %rd7427; + } + shf.l.wrap.b32 %r5216, %r5215, %r5214, 14; + shf.l.wrap.b32 %r5217, %r5214, %r5215, 14; + mov.b64 %rd7431, {%r5217, %r5216}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5218}, %rd7406; + } + { + .reg .b32 %dummy; + mov.b64 {%r5219,%dummy}, %rd7406; + } + shf.l.wrap.b32 %r5220, %r5219, %r5218, 28; + shf.l.wrap.b32 %r5221, %r5218, %r5219, 28; + mov.b64 %rd7432, {%r5221, %r5220}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5222}, %rd7412; + } + { + .reg .b32 %dummy; + mov.b64 {%r5223,%dummy}, %rd7412; + } + shf.l.wrap.b32 %r5224, %r5223, %r5222, 20; + shf.l.wrap.b32 %r5225, %r5222, %r5223, 20; + mov.b64 %rd7433, {%r5225, %r5224}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5226}, %rd7413; + } + { + .reg .b32 %dummy; + mov.b64 {%r5227,%dummy}, %rd7413; + } + shf.l.wrap.b32 %r5228, %r5227, %r5226, 3; + shf.l.wrap.b32 %r5229, %r5226, %r5227, 3; + mov.b64 %rd7434, {%r5229, %r5228}; + { + .reg .b32 %dummy; + mov.b64 {%r5230,%dummy}, %rd7419; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5231}, %rd7419; + } + shf.r.wrap.b32 %r5232, %r5231, %r5230, 19; + shf.r.wrap.b32 %r5233, %r5230, %r5231, 19; + mov.b64 %rd7435, {%r5233, %r5232}; + { + .reg .b32 %dummy; + mov.b64 {%r5234,%dummy}, %rd7425; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5235}, %rd7425; + } + shf.r.wrap.b32 %r5236, %r5235, %r5234, 3; + shf.r.wrap.b32 %r5237, %r5234, %r5235, 3; + mov.b64 %rd7436, {%r5237, %r5236}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5238}, %rd7404; + } + { + .reg .b32 %dummy; + mov.b64 {%r5239,%dummy}, %rd7404; + } + shf.l.wrap.b32 %r5240, %r5239, %r5238, 1; + shf.l.wrap.b32 %r5241, %r5238, %r5239, 1; + mov.b64 %rd7437, {%r5241, %r5240}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5242}, %rd7410; + } + { + .reg .b32 %dummy; + mov.b64 {%r5243,%dummy}, %rd7410; + } + shf.l.wrap.b32 %r5244, %r5243, %r5242, 6; + shf.l.wrap.b32 %r5245, %r5242, %r5243, 6; + mov.b64 %rd7438, {%r5245, %r5244}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5246}, %rd7416; + } + { + .reg .b32 %dummy; + mov.b64 {%r5247,%dummy}, %rd7416; + } + shf.l.wrap.b32 %r5248, %r5247, %r5246, 25; + shf.l.wrap.b32 %r5249, %r5246, %r5247, 25; + mov.b64 %rd7439, {%r5249, %r5248}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5250}, %rd7422; + } + { + .reg .b32 %dummy; + mov.b64 {%r5251,%dummy}, %rd7422; + } + shf.l.wrap.b32 %r5252, %r5251, %r5250, 8; + shf.l.wrap.b32 %r5253, %r5250, %r5251, 8; + mov.b64 %rd7440, {%r5253, %r5252}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5254}, %rd7423; + } + { + .reg .b32 %dummy; + mov.b64 {%r5255,%dummy}, %rd7423; + } + shf.l.wrap.b32 %r5256, %r5255, %r5254, 18; + shf.l.wrap.b32 %r5257, %r5254, %r5255, 18; + mov.b64 %rd7441, {%r5257, %r5256}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5258}, %rd7407; + } + { + .reg .b32 %dummy; + mov.b64 {%r5259,%dummy}, %rd7407; + } + shf.l.wrap.b32 %r5260, %r5259, %r5258, 27; + shf.l.wrap.b32 %r5261, %r5258, %r5259, 27; + mov.b64 %rd7442, {%r5261, %r5260}; + { + .reg .b32 %dummy; + mov.b64 {%r5262,%dummy}, %rd7408; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5263}, %rd7408; + } + shf.r.wrap.b32 %r5264, %r5263, %r5262, 28; + shf.r.wrap.b32 %r5265, %r5262, %r5263, 28; + mov.b64 %rd7443, {%r5265, %r5264}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5266}, %rd7414; + } + { + .reg .b32 %dummy; + mov.b64 {%r5267,%dummy}, %rd7414; + } + shf.l.wrap.b32 %r5268, %r5267, %r5266, 10; + shf.l.wrap.b32 %r5269, %r5266, %r5267, 10; + mov.b64 %rd7444, {%r5269, %r5268}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5270}, %rd7420; + } + { + .reg .b32 %dummy; + mov.b64 {%r5271,%dummy}, %rd7420; + } + shf.l.wrap.b32 %r5272, %r5271, %r5270, 15; + shf.l.wrap.b32 %r5273, %r5270, %r5271, 15; + mov.b64 %rd7445, {%r5273, %r5272}; + { + .reg .b32 %dummy; + mov.b64 {%r5274,%dummy}, %rd7426; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5275}, %rd7426; + } + shf.r.wrap.b32 %r5276, %r5275, %r5274, 8; + shf.r.wrap.b32 %r5277, %r5274, %r5275, 8; + mov.b64 %rd7446, {%r5277, %r5276}; + { + .reg .b32 %dummy; + mov.b64 {%r5278,%dummy}, %rd7405; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5279}, %rd7405; + } + shf.r.wrap.b32 %r5280, %r5279, %r5278, 2; + shf.r.wrap.b32 %r5281, %r5278, %r5279, 2; + mov.b64 %rd7447, {%r5281, %r5280}; + { + .reg .b32 %dummy; + mov.b64 {%r5282,%dummy}, %rd7411; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5283}, %rd7411; + } + shf.r.wrap.b32 %r5284, %r5283, %r5282, 9; + shf.r.wrap.b32 %r5285, %r5282, %r5283, 9; + mov.b64 %rd7448, {%r5285, %r5284}; + { + .reg .b32 %dummy; + mov.b64 {%r5286,%dummy}, %rd7417; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5287}, %rd7417; + } + shf.r.wrap.b32 %r5288, %r5287, %r5286, 25; + shf.r.wrap.b32 %r5289, %r5286, %r5287, 25; + mov.b64 %rd7449, {%r5289, %r5288}; + { + .reg .b32 %dummy; + mov.b64 {%r5290,%dummy}, %rd7418; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5291}, %rd7418; + } + shf.r.wrap.b32 %r5292, %r5291, %r5290, 23; + shf.r.wrap.b32 %r5293, %r5290, %r5291, 23; + mov.b64 %rd7450, {%r5293, %r5292}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5294}, %rd7424; + } + { + .reg .b32 %dummy; + mov.b64 {%r5295,%dummy}, %rd7424; + } + shf.l.wrap.b32 %r5296, %r5295, %r5294, 2; + shf.l.wrap.b32 %r5297, %r5294, %r5295, 2; + mov.b64 %rd7451, {%r5297, %r5296}; + not.b64 %rd7452, %rd7428; + and.b64 %rd7453, %rd7429, %rd7452; + xor.b64 %rd7454, %rd7453, %rd7403; + not.b64 %rd7455, %rd7429; + and.b64 %rd7456, %rd7430, %rd7455; + xor.b64 %rd7457, %rd7456, %rd7428; + not.b64 %rd7458, %rd7430; + and.b64 %rd7459, %rd7431, %rd7458; + xor.b64 %rd7460, %rd7459, %rd7429; + not.b64 %rd7461, %rd7431; + and.b64 %rd7462, %rd7403, %rd7461; + xor.b64 %rd7463, %rd7462, %rd7430; + not.b64 %rd7464, %rd7403; + and.b64 %rd7465, %rd7428, %rd7464; + xor.b64 %rd7466, %rd7465, %rd7431; + not.b64 %rd7467, %rd7433; + and.b64 %rd7468, %rd7434, %rd7467; + xor.b64 %rd7469, %rd7468, %rd7432; + not.b64 %rd7470, %rd7434; + and.b64 %rd7471, %rd7435, %rd7470; + xor.b64 %rd7472, %rd7471, %rd7433; + not.b64 %rd7473, %rd7435; + and.b64 %rd7474, %rd7436, %rd7473; + xor.b64 %rd7475, %rd7474, %rd7434; + not.b64 %rd7476, %rd7436; + and.b64 %rd7477, %rd7432, %rd7476; + xor.b64 %rd7478, %rd7477, %rd7435; + not.b64 %rd7479, %rd7432; + and.b64 %rd7480, %rd7433, %rd7479; + xor.b64 %rd7481, %rd7480, %rd7436; + not.b64 %rd7482, %rd7438; + and.b64 %rd7483, %rd7439, %rd7482; + xor.b64 %rd7484, %rd7483, %rd7437; + not.b64 %rd7485, %rd7439; + and.b64 %rd7486, %rd7440, %rd7485; + xor.b64 %rd7487, %rd7486, %rd7438; + not.b64 %rd7488, %rd7440; + and.b64 %rd7489, %rd7441, %rd7488; + xor.b64 %rd7490, %rd7489, %rd7439; + not.b64 %rd7491, %rd7441; + and.b64 %rd7492, %rd7437, %rd7491; + xor.b64 %rd7493, %rd7492, %rd7440; + not.b64 %rd7494, %rd7437; + and.b64 %rd7495, %rd7438, %rd7494; + xor.b64 %rd7496, %rd7495, %rd7441; + not.b64 %rd7497, %rd7443; + and.b64 %rd7498, %rd7444, %rd7497; + xor.b64 %rd7499, %rd7498, %rd7442; + not.b64 %rd7500, %rd7444; + and.b64 %rd7501, %rd7445, %rd7500; + xor.b64 %rd7502, %rd7501, %rd7443; + not.b64 %rd7503, %rd7445; + and.b64 %rd7504, %rd7446, %rd7503; + xor.b64 %rd7505, %rd7504, %rd7444; + not.b64 %rd7506, %rd7446; + and.b64 %rd7507, %rd7442, %rd7506; + xor.b64 %rd7508, %rd7507, %rd7445; + not.b64 %rd7509, %rd7442; + and.b64 %rd7510, %rd7443, %rd7509; + xor.b64 %rd7511, %rd7510, %rd7446; + not.b64 %rd7512, %rd7448; + and.b64 %rd7513, %rd7449, %rd7512; + xor.b64 %rd7514, %rd7513, %rd7447; + not.b64 %rd7515, %rd7449; + and.b64 %rd7516, %rd7450, %rd7515; + xor.b64 %rd7517, %rd7516, %rd7448; + not.b64 %rd7518, %rd7450; + and.b64 %rd7519, %rd7451, %rd7518; + xor.b64 %rd7520, %rd7519, %rd7449; + not.b64 %rd7521, %rd7451; + and.b64 %rd7522, %rd7447, %rd7521; + xor.b64 %rd7523, %rd7522, %rd7450; + not.b64 %rd7524, %rd7447; + and.b64 %rd7525, %rd7448, %rd7524; + xor.b64 %rd7526, %rd7525, %rd7451; + xor.b64 %rd7527, %rd7454, %rd36; + xor.b64 %rd7528, %rd7469, %rd7527; + xor.b64 %rd7529, %rd7528, %rd7484; + xor.b64 %rd7530, %rd7529, %rd7499; + xor.b64 %rd7531, %rd7530, %rd7514; + xor.b64 %rd7532, %rd7472, %rd7457; + xor.b64 %rd7533, %rd7532, %rd7487; + xor.b64 %rd7534, %rd7533, %rd7502; + xor.b64 %rd7535, %rd7534, %rd7517; + xor.b64 %rd7536, %rd7475, %rd7460; + xor.b64 %rd7537, %rd7536, %rd7490; + xor.b64 %rd7538, %rd7537, %rd7505; + xor.b64 %rd7539, %rd7538, %rd7520; + xor.b64 %rd7540, %rd7478, %rd7463; + xor.b64 %rd7541, %rd7540, %rd7493; + xor.b64 %rd7542, %rd7541, %rd7508; + xor.b64 %rd7543, %rd7542, %rd7523; + xor.b64 %rd7544, %rd7481, %rd7466; + xor.b64 %rd7545, %rd7544, %rd7496; + xor.b64 %rd7546, %rd7545, %rd7511; + xor.b64 %rd7547, %rd7546, %rd7526; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5298}, %rd7535; + } + { + .reg .b32 %dummy; + mov.b64 {%r5299,%dummy}, %rd7535; + } + shf.l.wrap.b32 %r5300, %r5299, %r5298, 1; + shf.l.wrap.b32 %r5301, %r5298, %r5299, 1; + mov.b64 %rd7548, {%r5301, %r5300}; + xor.b64 %rd7549, %rd7547, %rd7548; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5302}, %rd7539; + } + { + .reg .b32 %dummy; + mov.b64 {%r5303,%dummy}, %rd7539; + } + shf.l.wrap.b32 %r5304, %r5303, %r5302, 1; + shf.l.wrap.b32 %r5305, %r5302, %r5303, 1; + mov.b64 %rd7550, {%r5305, %r5304}; + xor.b64 %rd7551, %rd7550, %rd7531; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5306}, %rd7543; + } + { + .reg .b32 %dummy; + mov.b64 {%r5307,%dummy}, %rd7543; + } + shf.l.wrap.b32 %r5308, %r5307, %r5306, 1; + shf.l.wrap.b32 %r5309, %r5306, %r5307, 1; + mov.b64 %rd7552, {%r5309, %r5308}; + xor.b64 %rd7553, %rd7552, %rd7535; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5310}, %rd7547; + } + { + .reg .b32 %dummy; + mov.b64 {%r5311,%dummy}, %rd7547; + } + shf.l.wrap.b32 %r5312, %r5311, %r5310, 1; + shf.l.wrap.b32 %r5313, %r5310, %r5311, 1; + mov.b64 %rd7554, {%r5313, %r5312}; + xor.b64 %rd7555, %rd7554, %rd7539; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5314}, %rd7531; + } + { + .reg .b32 %dummy; + mov.b64 {%r5315,%dummy}, %rd7531; + } + shf.l.wrap.b32 %r5316, %r5315, %r5314, 1; + shf.l.wrap.b32 %r5317, %r5314, %r5315, 1; + mov.b64 %rd7556, {%r5317, %r5316}; + xor.b64 %rd7557, %rd7543, %rd7556; + xor.b64 %rd7558, %rd7549, %rd7527; + xor.b64 %rd7559, %rd7551, %rd7457; + xor.b64 %rd7560, %rd7553, %rd7460; + xor.b64 %rd7561, %rd7555, %rd7463; + xor.b64 %rd7562, %rd7557, %rd7466; + xor.b64 %rd7563, %rd7549, %rd7469; + xor.b64 %rd7564, %rd7551, %rd7472; + xor.b64 %rd7565, %rd7553, %rd7475; + xor.b64 %rd7566, %rd7555, %rd7478; + xor.b64 %rd7567, %rd7557, %rd7481; + xor.b64 %rd7568, %rd7549, %rd7484; + xor.b64 %rd7569, %rd7551, %rd7487; + xor.b64 %rd7570, %rd7553, %rd7490; + xor.b64 %rd7571, %rd7555, %rd7493; + xor.b64 %rd7572, %rd7557, %rd7496; + xor.b64 %rd7573, %rd7549, %rd7499; + xor.b64 %rd7574, %rd7551, %rd7502; + xor.b64 %rd7575, %rd7553, %rd7505; + xor.b64 %rd7576, %rd7555, %rd7508; + xor.b64 %rd7577, %rd7557, %rd7511; + xor.b64 %rd7578, %rd7549, %rd7514; + xor.b64 %rd7579, %rd7551, %rd7517; + xor.b64 %rd7580, %rd7553, %rd7520; + xor.b64 %rd7581, %rd7555, %rd7523; + xor.b64 %rd7582, %rd7557, %rd7526; + { + .reg .b32 %dummy; + mov.b64 {%r5318,%dummy}, %rd7564; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5319}, %rd7564; + } + shf.r.wrap.b32 %r5320, %r5319, %r5318, 20; + shf.r.wrap.b32 %r5321, %r5318, %r5319, 20; + mov.b64 %rd7583, {%r5321, %r5320}; + { + .reg .b32 %dummy; + mov.b64 {%r5322,%dummy}, %rd7570; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5323}, %rd7570; + } + shf.r.wrap.b32 %r5324, %r5323, %r5322, 21; + shf.r.wrap.b32 %r5325, %r5322, %r5323, 21; + mov.b64 %rd7584, {%r5325, %r5324}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5326}, %rd7576; + } + { + .reg .b32 %dummy; + mov.b64 {%r5327,%dummy}, %rd7576; + } + shf.l.wrap.b32 %r5328, %r5327, %r5326, 21; + shf.l.wrap.b32 %r5329, %r5326, %r5327, 21; + mov.b64 %rd7585, {%r5329, %r5328}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5330}, %rd7582; + } + { + .reg .b32 %dummy; + mov.b64 {%r5331,%dummy}, %rd7582; + } + shf.l.wrap.b32 %r5332, %r5331, %r5330, 14; + shf.l.wrap.b32 %r5333, %r5330, %r5331, 14; + mov.b64 %rd7586, {%r5333, %r5332}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5334}, %rd7561; + } + { + .reg .b32 %dummy; + mov.b64 {%r5335,%dummy}, %rd7561; + } + shf.l.wrap.b32 %r5336, %r5335, %r5334, 28; + shf.l.wrap.b32 %r5337, %r5334, %r5335, 28; + mov.b64 %rd7587, {%r5337, %r5336}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5338}, %rd7567; + } + { + .reg .b32 %dummy; + mov.b64 {%r5339,%dummy}, %rd7567; + } + shf.l.wrap.b32 %r5340, %r5339, %r5338, 20; + shf.l.wrap.b32 %r5341, %r5338, %r5339, 20; + mov.b64 %rd7588, {%r5341, %r5340}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5342}, %rd7568; + } + { + .reg .b32 %dummy; + mov.b64 {%r5343,%dummy}, %rd7568; + } + shf.l.wrap.b32 %r5344, %r5343, %r5342, 3; + shf.l.wrap.b32 %r5345, %r5342, %r5343, 3; + mov.b64 %rd7589, {%r5345, %r5344}; + { + .reg .b32 %dummy; + mov.b64 {%r5346,%dummy}, %rd7574; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5347}, %rd7574; + } + shf.r.wrap.b32 %r5348, %r5347, %r5346, 19; + shf.r.wrap.b32 %r5349, %r5346, %r5347, 19; + mov.b64 %rd7590, {%r5349, %r5348}; + { + .reg .b32 %dummy; + mov.b64 {%r5350,%dummy}, %rd7580; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5351}, %rd7580; + } + shf.r.wrap.b32 %r5352, %r5351, %r5350, 3; + shf.r.wrap.b32 %r5353, %r5350, %r5351, 3; + mov.b64 %rd7591, {%r5353, %r5352}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5354}, %rd7559; + } + { + .reg .b32 %dummy; + mov.b64 {%r5355,%dummy}, %rd7559; + } + shf.l.wrap.b32 %r5356, %r5355, %r5354, 1; + shf.l.wrap.b32 %r5357, %r5354, %r5355, 1; + mov.b64 %rd7592, {%r5357, %r5356}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5358}, %rd7565; + } + { + .reg .b32 %dummy; + mov.b64 {%r5359,%dummy}, %rd7565; + } + shf.l.wrap.b32 %r5360, %r5359, %r5358, 6; + shf.l.wrap.b32 %r5361, %r5358, %r5359, 6; + mov.b64 %rd7593, {%r5361, %r5360}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5362}, %rd7571; + } + { + .reg .b32 %dummy; + mov.b64 {%r5363,%dummy}, %rd7571; + } + shf.l.wrap.b32 %r5364, %r5363, %r5362, 25; + shf.l.wrap.b32 %r5365, %r5362, %r5363, 25; + mov.b64 %rd7594, {%r5365, %r5364}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5366}, %rd7577; + } + { + .reg .b32 %dummy; + mov.b64 {%r5367,%dummy}, %rd7577; + } + shf.l.wrap.b32 %r5368, %r5367, %r5366, 8; + shf.l.wrap.b32 %r5369, %r5366, %r5367, 8; + mov.b64 %rd7595, {%r5369, %r5368}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5370}, %rd7578; + } + { + .reg .b32 %dummy; + mov.b64 {%r5371,%dummy}, %rd7578; + } + shf.l.wrap.b32 %r5372, %r5371, %r5370, 18; + shf.l.wrap.b32 %r5373, %r5370, %r5371, 18; + mov.b64 %rd7596, {%r5373, %r5372}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5374}, %rd7562; + } + { + .reg .b32 %dummy; + mov.b64 {%r5375,%dummy}, %rd7562; + } + shf.l.wrap.b32 %r5376, %r5375, %r5374, 27; + shf.l.wrap.b32 %r5377, %r5374, %r5375, 27; + mov.b64 %rd7597, {%r5377, %r5376}; + { + .reg .b32 %dummy; + mov.b64 {%r5378,%dummy}, %rd7563; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5379}, %rd7563; + } + shf.r.wrap.b32 %r5380, %r5379, %r5378, 28; + shf.r.wrap.b32 %r5381, %r5378, %r5379, 28; + mov.b64 %rd7598, {%r5381, %r5380}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5382}, %rd7569; + } + { + .reg .b32 %dummy; + mov.b64 {%r5383,%dummy}, %rd7569; + } + shf.l.wrap.b32 %r5384, %r5383, %r5382, 10; + shf.l.wrap.b32 %r5385, %r5382, %r5383, 10; + mov.b64 %rd7599, {%r5385, %r5384}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5386}, %rd7575; + } + { + .reg .b32 %dummy; + mov.b64 {%r5387,%dummy}, %rd7575; + } + shf.l.wrap.b32 %r5388, %r5387, %r5386, 15; + shf.l.wrap.b32 %r5389, %r5386, %r5387, 15; + mov.b64 %rd7600, {%r5389, %r5388}; + { + .reg .b32 %dummy; + mov.b64 {%r5390,%dummy}, %rd7581; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5391}, %rd7581; + } + shf.r.wrap.b32 %r5392, %r5391, %r5390, 8; + shf.r.wrap.b32 %r5393, %r5390, %r5391, 8; + mov.b64 %rd7601, {%r5393, %r5392}; + { + .reg .b32 %dummy; + mov.b64 {%r5394,%dummy}, %rd7560; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5395}, %rd7560; + } + shf.r.wrap.b32 %r5396, %r5395, %r5394, 2; + shf.r.wrap.b32 %r5397, %r5394, %r5395, 2; + mov.b64 %rd7602, {%r5397, %r5396}; + { + .reg .b32 %dummy; + mov.b64 {%r5398,%dummy}, %rd7566; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5399}, %rd7566; + } + shf.r.wrap.b32 %r5400, %r5399, %r5398, 9; + shf.r.wrap.b32 %r5401, %r5398, %r5399, 9; + mov.b64 %rd7603, {%r5401, %r5400}; + { + .reg .b32 %dummy; + mov.b64 {%r5402,%dummy}, %rd7572; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5403}, %rd7572; + } + shf.r.wrap.b32 %r5404, %r5403, %r5402, 25; + shf.r.wrap.b32 %r5405, %r5402, %r5403, 25; + mov.b64 %rd7604, {%r5405, %r5404}; + { + .reg .b32 %dummy; + mov.b64 {%r5406,%dummy}, %rd7573; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5407}, %rd7573; + } + shf.r.wrap.b32 %r5408, %r5407, %r5406, 23; + shf.r.wrap.b32 %r5409, %r5406, %r5407, 23; + mov.b64 %rd7605, {%r5409, %r5408}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5410}, %rd7579; + } + { + .reg .b32 %dummy; + mov.b64 {%r5411,%dummy}, %rd7579; + } + shf.l.wrap.b32 %r5412, %r5411, %r5410, 2; + shf.l.wrap.b32 %r5413, %r5410, %r5411, 2; + mov.b64 %rd7606, {%r5413, %r5412}; + not.b64 %rd7607, %rd7583; + and.b64 %rd7608, %rd7584, %rd7607; + xor.b64 %rd7609, %rd7608, %rd7558; + not.b64 %rd7610, %rd7584; + and.b64 %rd7611, %rd7585, %rd7610; + xor.b64 %rd7612, %rd7611, %rd7583; + not.b64 %rd7613, %rd7585; + and.b64 %rd7614, %rd7586, %rd7613; + xor.b64 %rd7615, %rd7614, %rd7584; + not.b64 %rd7616, %rd7586; + and.b64 %rd7617, %rd7558, %rd7616; + xor.b64 %rd7618, %rd7617, %rd7585; + not.b64 %rd7619, %rd7558; + and.b64 %rd7620, %rd7583, %rd7619; + xor.b64 %rd7621, %rd7620, %rd7586; + not.b64 %rd7622, %rd7588; + and.b64 %rd7623, %rd7589, %rd7622; + xor.b64 %rd7624, %rd7623, %rd7587; + not.b64 %rd7625, %rd7589; + and.b64 %rd7626, %rd7590, %rd7625; + xor.b64 %rd7627, %rd7626, %rd7588; + not.b64 %rd7628, %rd7590; + and.b64 %rd7629, %rd7591, %rd7628; + xor.b64 %rd7630, %rd7629, %rd7589; + not.b64 %rd7631, %rd7591; + and.b64 %rd7632, %rd7587, %rd7631; + xor.b64 %rd7633, %rd7632, %rd7590; + not.b64 %rd7634, %rd7587; + and.b64 %rd7635, %rd7588, %rd7634; + xor.b64 %rd7636, %rd7635, %rd7591; + not.b64 %rd7637, %rd7593; + and.b64 %rd7638, %rd7594, %rd7637; + xor.b64 %rd7639, %rd7638, %rd7592; + not.b64 %rd7640, %rd7594; + and.b64 %rd7641, %rd7595, %rd7640; + xor.b64 %rd7642, %rd7641, %rd7593; + not.b64 %rd7643, %rd7595; + and.b64 %rd7644, %rd7596, %rd7643; + xor.b64 %rd7645, %rd7644, %rd7594; + not.b64 %rd7646, %rd7596; + and.b64 %rd7647, %rd7592, %rd7646; + xor.b64 %rd7648, %rd7647, %rd7595; + not.b64 %rd7649, %rd7592; + and.b64 %rd7650, %rd7593, %rd7649; + xor.b64 %rd7651, %rd7650, %rd7596; + not.b64 %rd7652, %rd7598; + and.b64 %rd7653, %rd7599, %rd7652; + xor.b64 %rd7654, %rd7653, %rd7597; + not.b64 %rd7655, %rd7599; + and.b64 %rd7656, %rd7600, %rd7655; + xor.b64 %rd7657, %rd7656, %rd7598; + not.b64 %rd7658, %rd7600; + and.b64 %rd7659, %rd7601, %rd7658; + xor.b64 %rd7660, %rd7659, %rd7599; + not.b64 %rd7661, %rd7601; + and.b64 %rd7662, %rd7597, %rd7661; + xor.b64 %rd7663, %rd7662, %rd7600; + not.b64 %rd7664, %rd7597; + and.b64 %rd7665, %rd7598, %rd7664; + xor.b64 %rd7666, %rd7665, %rd7601; + not.b64 %rd7667, %rd7603; + and.b64 %rd7668, %rd7604, %rd7667; + xor.b64 %rd7669, %rd7668, %rd7602; + not.b64 %rd7670, %rd7604; + and.b64 %rd7671, %rd7605, %rd7670; + xor.b64 %rd7672, %rd7671, %rd7603; + not.b64 %rd7673, %rd7605; + and.b64 %rd7674, %rd7606, %rd7673; + xor.b64 %rd7675, %rd7674, %rd7604; + not.b64 %rd7676, %rd7606; + and.b64 %rd7677, %rd7602, %rd7676; + xor.b64 %rd7678, %rd7677, %rd7605; + not.b64 %rd7679, %rd7602; + and.b64 %rd7680, %rd7603, %rd7679; + xor.b64 %rd7681, %rd7680, %rd7606; + xor.b64 %rd7682, %rd7609, %rd37; + xor.b64 %rd7683, %rd7624, %rd7682; + xor.b64 %rd7684, %rd7683, %rd7639; + xor.b64 %rd7685, %rd7684, %rd7654; + xor.b64 %rd7686, %rd7685, %rd7669; + xor.b64 %rd7687, %rd7627, %rd7612; + xor.b64 %rd7688, %rd7687, %rd7642; + xor.b64 %rd7689, %rd7688, %rd7657; + xor.b64 %rd7690, %rd7689, %rd7672; + xor.b64 %rd7691, %rd7630, %rd7615; + xor.b64 %rd7692, %rd7691, %rd7645; + xor.b64 %rd7693, %rd7692, %rd7660; + xor.b64 %rd7694, %rd7693, %rd7675; + xor.b64 %rd7695, %rd7633, %rd7618; + xor.b64 %rd7696, %rd7695, %rd7648; + xor.b64 %rd7697, %rd7696, %rd7663; + xor.b64 %rd7698, %rd7697, %rd7678; + xor.b64 %rd7699, %rd7636, %rd7621; + xor.b64 %rd7700, %rd7699, %rd7651; + xor.b64 %rd7701, %rd7700, %rd7666; + xor.b64 %rd7702, %rd7701, %rd7681; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5414}, %rd7690; + } + { + .reg .b32 %dummy; + mov.b64 {%r5415,%dummy}, %rd7690; + } + shf.l.wrap.b32 %r5416, %r5415, %r5414, 1; + shf.l.wrap.b32 %r5417, %r5414, %r5415, 1; + mov.b64 %rd7703, {%r5417, %r5416}; + xor.b64 %rd7704, %rd7702, %rd7703; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5418}, %rd7694; + } + { + .reg .b32 %dummy; + mov.b64 {%r5419,%dummy}, %rd7694; + } + shf.l.wrap.b32 %r5420, %r5419, %r5418, 1; + shf.l.wrap.b32 %r5421, %r5418, %r5419, 1; + mov.b64 %rd7705, {%r5421, %r5420}; + xor.b64 %rd7706, %rd7705, %rd7686; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5422}, %rd7698; + } + { + .reg .b32 %dummy; + mov.b64 {%r5423,%dummy}, %rd7698; + } + shf.l.wrap.b32 %r5424, %r5423, %r5422, 1; + shf.l.wrap.b32 %r5425, %r5422, %r5423, 1; + mov.b64 %rd7707, {%r5425, %r5424}; + xor.b64 %rd7708, %rd7707, %rd7690; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5426}, %rd7702; + } + { + .reg .b32 %dummy; + mov.b64 {%r5427,%dummy}, %rd7702; + } + shf.l.wrap.b32 %r5428, %r5427, %r5426, 1; + shf.l.wrap.b32 %r5429, %r5426, %r5427, 1; + mov.b64 %rd7709, {%r5429, %r5428}; + xor.b64 %rd7710, %rd7709, %rd7694; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5430}, %rd7686; + } + { + .reg .b32 %dummy; + mov.b64 {%r5431,%dummy}, %rd7686; + } + shf.l.wrap.b32 %r5432, %r5431, %r5430, 1; + shf.l.wrap.b32 %r5433, %r5430, %r5431, 1; + mov.b64 %rd7711, {%r5433, %r5432}; + xor.b64 %rd7712, %rd7698, %rd7711; + xor.b64 %rd7713, %rd7704, %rd7682; + xor.b64 %rd7714, %rd7706, %rd7612; + xor.b64 %rd7715, %rd7708, %rd7615; + xor.b64 %rd7716, %rd7710, %rd7618; + xor.b64 %rd7717, %rd7712, %rd7621; + xor.b64 %rd7718, %rd7704, %rd7624; + xor.b64 %rd7719, %rd7706, %rd7627; + xor.b64 %rd7720, %rd7708, %rd7630; + xor.b64 %rd7721, %rd7710, %rd7633; + xor.b64 %rd7722, %rd7712, %rd7636; + xor.b64 %rd7723, %rd7704, %rd7639; + xor.b64 %rd7724, %rd7706, %rd7642; + xor.b64 %rd7725, %rd7708, %rd7645; + xor.b64 %rd7726, %rd7710, %rd7648; + xor.b64 %rd7727, %rd7712, %rd7651; + xor.b64 %rd7728, %rd7704, %rd7654; + xor.b64 %rd7729, %rd7706, %rd7657; + xor.b64 %rd7730, %rd7708, %rd7660; + xor.b64 %rd7731, %rd7710, %rd7663; + xor.b64 %rd7732, %rd7712, %rd7666; + xor.b64 %rd7733, %rd7704, %rd7669; + xor.b64 %rd7734, %rd7706, %rd7672; + xor.b64 %rd7735, %rd7708, %rd7675; + xor.b64 %rd7736, %rd7710, %rd7678; + xor.b64 %rd7737, %rd7712, %rd7681; + { + .reg .b32 %dummy; + mov.b64 {%r5434,%dummy}, %rd7719; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5435}, %rd7719; + } + shf.r.wrap.b32 %r5436, %r5435, %r5434, 20; + shf.r.wrap.b32 %r5437, %r5434, %r5435, 20; + mov.b64 %rd7738, {%r5437, %r5436}; + { + .reg .b32 %dummy; + mov.b64 {%r5438,%dummy}, %rd7725; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5439}, %rd7725; + } + shf.r.wrap.b32 %r5440, %r5439, %r5438, 21; + shf.r.wrap.b32 %r5441, %r5438, %r5439, 21; + mov.b64 %rd7739, {%r5441, %r5440}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5442}, %rd7731; + } + { + .reg .b32 %dummy; + mov.b64 {%r5443,%dummy}, %rd7731; + } + shf.l.wrap.b32 %r5444, %r5443, %r5442, 21; + shf.l.wrap.b32 %r5445, %r5442, %r5443, 21; + mov.b64 %rd7740, {%r5445, %r5444}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5446}, %rd7737; + } + { + .reg .b32 %dummy; + mov.b64 {%r5447,%dummy}, %rd7737; + } + shf.l.wrap.b32 %r5448, %r5447, %r5446, 14; + shf.l.wrap.b32 %r5449, %r5446, %r5447, 14; + mov.b64 %rd7741, {%r5449, %r5448}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5450}, %rd7716; + } + { + .reg .b32 %dummy; + mov.b64 {%r5451,%dummy}, %rd7716; + } + shf.l.wrap.b32 %r5452, %r5451, %r5450, 28; + shf.l.wrap.b32 %r5453, %r5450, %r5451, 28; + mov.b64 %rd7742, {%r5453, %r5452}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5454}, %rd7722; + } + { + .reg .b32 %dummy; + mov.b64 {%r5455,%dummy}, %rd7722; + } + shf.l.wrap.b32 %r5456, %r5455, %r5454, 20; + shf.l.wrap.b32 %r5457, %r5454, %r5455, 20; + mov.b64 %rd7743, {%r5457, %r5456}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5458}, %rd7723; + } + { + .reg .b32 %dummy; + mov.b64 {%r5459,%dummy}, %rd7723; + } + shf.l.wrap.b32 %r5460, %r5459, %r5458, 3; + shf.l.wrap.b32 %r5461, %r5458, %r5459, 3; + mov.b64 %rd7744, {%r5461, %r5460}; + { + .reg .b32 %dummy; + mov.b64 {%r5462,%dummy}, %rd7729; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5463}, %rd7729; + } + shf.r.wrap.b32 %r5464, %r5463, %r5462, 19; + shf.r.wrap.b32 %r5465, %r5462, %r5463, 19; + mov.b64 %rd7745, {%r5465, %r5464}; + { + .reg .b32 %dummy; + mov.b64 {%r5466,%dummy}, %rd7735; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5467}, %rd7735; + } + shf.r.wrap.b32 %r5468, %r5467, %r5466, 3; + shf.r.wrap.b32 %r5469, %r5466, %r5467, 3; + mov.b64 %rd7746, {%r5469, %r5468}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5470}, %rd7714; + } + { + .reg .b32 %dummy; + mov.b64 {%r5471,%dummy}, %rd7714; + } + shf.l.wrap.b32 %r5472, %r5471, %r5470, 1; + shf.l.wrap.b32 %r5473, %r5470, %r5471, 1; + mov.b64 %rd7747, {%r5473, %r5472}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5474}, %rd7720; + } + { + .reg .b32 %dummy; + mov.b64 {%r5475,%dummy}, %rd7720; + } + shf.l.wrap.b32 %r5476, %r5475, %r5474, 6; + shf.l.wrap.b32 %r5477, %r5474, %r5475, 6; + mov.b64 %rd7748, {%r5477, %r5476}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5478}, %rd7726; + } + { + .reg .b32 %dummy; + mov.b64 {%r5479,%dummy}, %rd7726; + } + shf.l.wrap.b32 %r5480, %r5479, %r5478, 25; + shf.l.wrap.b32 %r5481, %r5478, %r5479, 25; + mov.b64 %rd7749, {%r5481, %r5480}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5482}, %rd7732; + } + { + .reg .b32 %dummy; + mov.b64 {%r5483,%dummy}, %rd7732; + } + shf.l.wrap.b32 %r5484, %r5483, %r5482, 8; + shf.l.wrap.b32 %r5485, %r5482, %r5483, 8; + mov.b64 %rd7750, {%r5485, %r5484}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5486}, %rd7733; + } + { + .reg .b32 %dummy; + mov.b64 {%r5487,%dummy}, %rd7733; + } + shf.l.wrap.b32 %r5488, %r5487, %r5486, 18; + shf.l.wrap.b32 %r5489, %r5486, %r5487, 18; + mov.b64 %rd7751, {%r5489, %r5488}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5490}, %rd7717; + } + { + .reg .b32 %dummy; + mov.b64 {%r5491,%dummy}, %rd7717; + } + shf.l.wrap.b32 %r5492, %r5491, %r5490, 27; + shf.l.wrap.b32 %r5493, %r5490, %r5491, 27; + mov.b64 %rd7752, {%r5493, %r5492}; + { + .reg .b32 %dummy; + mov.b64 {%r5494,%dummy}, %rd7718; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5495}, %rd7718; + } + shf.r.wrap.b32 %r5496, %r5495, %r5494, 28; + shf.r.wrap.b32 %r5497, %r5494, %r5495, 28; + mov.b64 %rd7753, {%r5497, %r5496}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5498}, %rd7724; + } + { + .reg .b32 %dummy; + mov.b64 {%r5499,%dummy}, %rd7724; + } + shf.l.wrap.b32 %r5500, %r5499, %r5498, 10; + shf.l.wrap.b32 %r5501, %r5498, %r5499, 10; + mov.b64 %rd7754, {%r5501, %r5500}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5502}, %rd7730; + } + { + .reg .b32 %dummy; + mov.b64 {%r5503,%dummy}, %rd7730; + } + shf.l.wrap.b32 %r5504, %r5503, %r5502, 15; + shf.l.wrap.b32 %r5505, %r5502, %r5503, 15; + mov.b64 %rd7755, {%r5505, %r5504}; + { + .reg .b32 %dummy; + mov.b64 {%r5506,%dummy}, %rd7736; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5507}, %rd7736; + } + shf.r.wrap.b32 %r5508, %r5507, %r5506, 8; + shf.r.wrap.b32 %r5509, %r5506, %r5507, 8; + mov.b64 %rd7756, {%r5509, %r5508}; + { + .reg .b32 %dummy; + mov.b64 {%r5510,%dummy}, %rd7715; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5511}, %rd7715; + } + shf.r.wrap.b32 %r5512, %r5511, %r5510, 2; + shf.r.wrap.b32 %r5513, %r5510, %r5511, 2; + mov.b64 %rd7757, {%r5513, %r5512}; + { + .reg .b32 %dummy; + mov.b64 {%r5514,%dummy}, %rd7721; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5515}, %rd7721; + } + shf.r.wrap.b32 %r5516, %r5515, %r5514, 9; + shf.r.wrap.b32 %r5517, %r5514, %r5515, 9; + mov.b64 %rd7758, {%r5517, %r5516}; + { + .reg .b32 %dummy; + mov.b64 {%r5518,%dummy}, %rd7727; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5519}, %rd7727; + } + shf.r.wrap.b32 %r5520, %r5519, %r5518, 25; + shf.r.wrap.b32 %r5521, %r5518, %r5519, 25; + mov.b64 %rd7759, {%r5521, %r5520}; + { + .reg .b32 %dummy; + mov.b64 {%r5522,%dummy}, %rd7728; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5523}, %rd7728; + } + shf.r.wrap.b32 %r5524, %r5523, %r5522, 23; + shf.r.wrap.b32 %r5525, %r5522, %r5523, 23; + mov.b64 %rd7760, {%r5525, %r5524}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5526}, %rd7734; + } + { + .reg .b32 %dummy; + mov.b64 {%r5527,%dummy}, %rd7734; + } + shf.l.wrap.b32 %r5528, %r5527, %r5526, 2; + shf.l.wrap.b32 %r5529, %r5526, %r5527, 2; + mov.b64 %rd7761, {%r5529, %r5528}; + not.b64 %rd7762, %rd7738; + and.b64 %rd7763, %rd7739, %rd7762; + xor.b64 %rd7764, %rd7763, %rd7713; + not.b64 %rd7765, %rd7739; + and.b64 %rd7766, %rd7740, %rd7765; + xor.b64 %rd7767, %rd7766, %rd7738; + not.b64 %rd7768, %rd7740; + and.b64 %rd7769, %rd7741, %rd7768; + xor.b64 %rd7770, %rd7769, %rd7739; + not.b64 %rd7771, %rd7741; + and.b64 %rd7772, %rd7713, %rd7771; + xor.b64 %rd7773, %rd7772, %rd7740; + not.b64 %rd7774, %rd7713; + and.b64 %rd7775, %rd7738, %rd7774; + xor.b64 %rd7776, %rd7775, %rd7741; + not.b64 %rd7777, %rd7743; + and.b64 %rd7778, %rd7744, %rd7777; + xor.b64 %rd7779, %rd7778, %rd7742; + not.b64 %rd7780, %rd7744; + and.b64 %rd7781, %rd7745, %rd7780; + xor.b64 %rd7782, %rd7781, %rd7743; + not.b64 %rd7783, %rd7745; + and.b64 %rd7784, %rd7746, %rd7783; + xor.b64 %rd7785, %rd7784, %rd7744; + not.b64 %rd7786, %rd7746; + and.b64 %rd7787, %rd7742, %rd7786; + xor.b64 %rd7788, %rd7787, %rd7745; + not.b64 %rd7789, %rd7742; + and.b64 %rd7790, %rd7743, %rd7789; + xor.b64 %rd7791, %rd7790, %rd7746; + not.b64 %rd7792, %rd7748; + and.b64 %rd7793, %rd7749, %rd7792; + xor.b64 %rd7794, %rd7793, %rd7747; + not.b64 %rd7795, %rd7749; + and.b64 %rd7796, %rd7750, %rd7795; + xor.b64 %rd7797, %rd7796, %rd7748; + not.b64 %rd7798, %rd7750; + and.b64 %rd7799, %rd7751, %rd7798; + xor.b64 %rd7800, %rd7799, %rd7749; + not.b64 %rd7801, %rd7751; + and.b64 %rd7802, %rd7747, %rd7801; + xor.b64 %rd7803, %rd7802, %rd7750; + not.b64 %rd7804, %rd7747; + and.b64 %rd7805, %rd7748, %rd7804; + xor.b64 %rd7806, %rd7805, %rd7751; + not.b64 %rd7807, %rd7753; + and.b64 %rd7808, %rd7754, %rd7807; + xor.b64 %rd7809, %rd7808, %rd7752; + not.b64 %rd7810, %rd7754; + and.b64 %rd7811, %rd7755, %rd7810; + xor.b64 %rd7812, %rd7811, %rd7753; + not.b64 %rd7813, %rd7755; + and.b64 %rd7814, %rd7756, %rd7813; + xor.b64 %rd7815, %rd7814, %rd7754; + not.b64 %rd7816, %rd7756; + and.b64 %rd7817, %rd7752, %rd7816; + xor.b64 %rd7818, %rd7817, %rd7755; + not.b64 %rd7819, %rd7752; + and.b64 %rd7820, %rd7753, %rd7819; + xor.b64 %rd7821, %rd7820, %rd7756; + not.b64 %rd7822, %rd7758; + and.b64 %rd7823, %rd7759, %rd7822; + xor.b64 %rd7824, %rd7823, %rd7757; + not.b64 %rd7825, %rd7759; + and.b64 %rd7826, %rd7760, %rd7825; + xor.b64 %rd7827, %rd7826, %rd7758; + not.b64 %rd7828, %rd7760; + and.b64 %rd7829, %rd7761, %rd7828; + xor.b64 %rd7830, %rd7829, %rd7759; + not.b64 %rd7831, %rd7761; + and.b64 %rd7832, %rd7757, %rd7831; + xor.b64 %rd7833, %rd7832, %rd7760; + not.b64 %rd7834, %rd7757; + and.b64 %rd7835, %rd7758, %rd7834; + xor.b64 %rd7836, %rd7835, %rd7761; + xor.b64 %rd7837, %rd7764, %rd38; + xor.b64 %rd7838, %rd7779, %rd7837; + xor.b64 %rd7839, %rd7838, %rd7794; + xor.b64 %rd7840, %rd7839, %rd7809; + xor.b64 %rd7841, %rd7840, %rd7824; + xor.b64 %rd7842, %rd7782, %rd7767; + xor.b64 %rd7843, %rd7842, %rd7797; + xor.b64 %rd7844, %rd7843, %rd7812; + xor.b64 %rd7845, %rd7844, %rd7827; + xor.b64 %rd7846, %rd7785, %rd7770; + xor.b64 %rd7847, %rd7846, %rd7800; + xor.b64 %rd7848, %rd7847, %rd7815; + xor.b64 %rd7849, %rd7848, %rd7830; + xor.b64 %rd7850, %rd7788, %rd7773; + xor.b64 %rd7851, %rd7850, %rd7803; + xor.b64 %rd7852, %rd7851, %rd7818; + xor.b64 %rd7853, %rd7852, %rd7833; + xor.b64 %rd7854, %rd7791, %rd7776; + xor.b64 %rd7855, %rd7854, %rd7806; + xor.b64 %rd7856, %rd7855, %rd7821; + xor.b64 %rd7857, %rd7856, %rd7836; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5530}, %rd7845; + } + { + .reg .b32 %dummy; + mov.b64 {%r5531,%dummy}, %rd7845; + } + shf.l.wrap.b32 %r5532, %r5531, %r5530, 1; + shf.l.wrap.b32 %r5533, %r5530, %r5531, 1; + mov.b64 %rd7858, {%r5533, %r5532}; + xor.b64 %rd7859, %rd7857, %rd7858; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5534}, %rd7849; + } + { + .reg .b32 %dummy; + mov.b64 {%r5535,%dummy}, %rd7849; + } + shf.l.wrap.b32 %r5536, %r5535, %r5534, 1; + shf.l.wrap.b32 %r5537, %r5534, %r5535, 1; + mov.b64 %rd7860, {%r5537, %r5536}; + xor.b64 %rd7861, %rd7860, %rd7841; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5538}, %rd7853; + } + { + .reg .b32 %dummy; + mov.b64 {%r5539,%dummy}, %rd7853; + } + shf.l.wrap.b32 %r5540, %r5539, %r5538, 1; + shf.l.wrap.b32 %r5541, %r5538, %r5539, 1; + mov.b64 %rd7862, {%r5541, %r5540}; + xor.b64 %rd7863, %rd7862, %rd7845; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5542}, %rd7857; + } + { + .reg .b32 %dummy; + mov.b64 {%r5543,%dummy}, %rd7857; + } + shf.l.wrap.b32 %r5544, %r5543, %r5542, 1; + shf.l.wrap.b32 %r5545, %r5542, %r5543, 1; + mov.b64 %rd7864, {%r5545, %r5544}; + xor.b64 %rd7865, %rd7864, %rd7849; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5546}, %rd7841; + } + { + .reg .b32 %dummy; + mov.b64 {%r5547,%dummy}, %rd7841; + } + shf.l.wrap.b32 %r5548, %r5547, %r5546, 1; + shf.l.wrap.b32 %r5549, %r5546, %r5547, 1; + mov.b64 %rd7866, {%r5549, %r5548}; + xor.b64 %rd7867, %rd7853, %rd7866; + xor.b64 %rd7868, %rd7859, %rd7837; + xor.b64 %rd7869, %rd7861, %rd7767; + xor.b64 %rd7870, %rd7863, %rd7770; + xor.b64 %rd7871, %rd7865, %rd7773; + xor.b64 %rd7872, %rd7867, %rd7776; + xor.b64 %rd7873, %rd7859, %rd7779; + xor.b64 %rd7874, %rd7861, %rd7782; + xor.b64 %rd7875, %rd7863, %rd7785; + xor.b64 %rd7876, %rd7865, %rd7788; + xor.b64 %rd7877, %rd7867, %rd7791; + xor.b64 %rd7878, %rd7859, %rd7794; + xor.b64 %rd7879, %rd7861, %rd7797; + xor.b64 %rd7880, %rd7863, %rd7800; + xor.b64 %rd7881, %rd7865, %rd7803; + xor.b64 %rd7882, %rd7867, %rd7806; + xor.b64 %rd7883, %rd7859, %rd7809; + xor.b64 %rd7884, %rd7861, %rd7812; + xor.b64 %rd7885, %rd7863, %rd7815; + xor.b64 %rd7886, %rd7865, %rd7818; + xor.b64 %rd7887, %rd7867, %rd7821; + xor.b64 %rd7888, %rd7859, %rd7824; + xor.b64 %rd7889, %rd7861, %rd7827; + xor.b64 %rd7890, %rd7863, %rd7830; + xor.b64 %rd7891, %rd7865, %rd7833; + xor.b64 %rd7892, %rd7867, %rd7836; + { + .reg .b32 %dummy; + mov.b64 {%r5550,%dummy}, %rd7874; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5551}, %rd7874; + } + shf.r.wrap.b32 %r5552, %r5551, %r5550, 20; + shf.r.wrap.b32 %r5553, %r5550, %r5551, 20; + mov.b64 %rd7893, {%r5553, %r5552}; + { + .reg .b32 %dummy; + mov.b64 {%r5554,%dummy}, %rd7880; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5555}, %rd7880; + } + shf.r.wrap.b32 %r5556, %r5555, %r5554, 21; + shf.r.wrap.b32 %r5557, %r5554, %r5555, 21; + mov.b64 %rd7894, {%r5557, %r5556}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5558}, %rd7886; + } + { + .reg .b32 %dummy; + mov.b64 {%r5559,%dummy}, %rd7886; + } + shf.l.wrap.b32 %r5560, %r5559, %r5558, 21; + shf.l.wrap.b32 %r5561, %r5558, %r5559, 21; + mov.b64 %rd7895, {%r5561, %r5560}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5562}, %rd7892; + } + { + .reg .b32 %dummy; + mov.b64 {%r5563,%dummy}, %rd7892; + } + shf.l.wrap.b32 %r5564, %r5563, %r5562, 14; + shf.l.wrap.b32 %r5565, %r5562, %r5563, 14; + mov.b64 %rd7896, {%r5565, %r5564}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5566}, %rd7871; + } + { + .reg .b32 %dummy; + mov.b64 {%r5567,%dummy}, %rd7871; + } + shf.l.wrap.b32 %r5568, %r5567, %r5566, 28; + shf.l.wrap.b32 %r5569, %r5566, %r5567, 28; + mov.b64 %rd7897, {%r5569, %r5568}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5570}, %rd7877; + } + { + .reg .b32 %dummy; + mov.b64 {%r5571,%dummy}, %rd7877; + } + shf.l.wrap.b32 %r5572, %r5571, %r5570, 20; + shf.l.wrap.b32 %r5573, %r5570, %r5571, 20; + mov.b64 %rd7898, {%r5573, %r5572}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5574}, %rd7878; + } + { + .reg .b32 %dummy; + mov.b64 {%r5575,%dummy}, %rd7878; + } + shf.l.wrap.b32 %r5576, %r5575, %r5574, 3; + shf.l.wrap.b32 %r5577, %r5574, %r5575, 3; + mov.b64 %rd7899, {%r5577, %r5576}; + { + .reg .b32 %dummy; + mov.b64 {%r5578,%dummy}, %rd7884; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5579}, %rd7884; + } + shf.r.wrap.b32 %r5580, %r5579, %r5578, 19; + shf.r.wrap.b32 %r5581, %r5578, %r5579, 19; + mov.b64 %rd7900, {%r5581, %r5580}; + { + .reg .b32 %dummy; + mov.b64 {%r5582,%dummy}, %rd7890; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5583}, %rd7890; + } + shf.r.wrap.b32 %r5584, %r5583, %r5582, 3; + shf.r.wrap.b32 %r5585, %r5582, %r5583, 3; + mov.b64 %rd7901, {%r5585, %r5584}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5586}, %rd7869; + } + { + .reg .b32 %dummy; + mov.b64 {%r5587,%dummy}, %rd7869; + } + shf.l.wrap.b32 %r5588, %r5587, %r5586, 1; + shf.l.wrap.b32 %r5589, %r5586, %r5587, 1; + mov.b64 %rd7902, {%r5589, %r5588}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5590}, %rd7875; + } + { + .reg .b32 %dummy; + mov.b64 {%r5591,%dummy}, %rd7875; + } + shf.l.wrap.b32 %r5592, %r5591, %r5590, 6; + shf.l.wrap.b32 %r5593, %r5590, %r5591, 6; + mov.b64 %rd7903, {%r5593, %r5592}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5594}, %rd7881; + } + { + .reg .b32 %dummy; + mov.b64 {%r5595,%dummy}, %rd7881; + } + shf.l.wrap.b32 %r5596, %r5595, %r5594, 25; + shf.l.wrap.b32 %r5597, %r5594, %r5595, 25; + mov.b64 %rd7904, {%r5597, %r5596}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5598}, %rd7887; + } + { + .reg .b32 %dummy; + mov.b64 {%r5599,%dummy}, %rd7887; + } + shf.l.wrap.b32 %r5600, %r5599, %r5598, 8; + shf.l.wrap.b32 %r5601, %r5598, %r5599, 8; + mov.b64 %rd7905, {%r5601, %r5600}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5602}, %rd7888; + } + { + .reg .b32 %dummy; + mov.b64 {%r5603,%dummy}, %rd7888; + } + shf.l.wrap.b32 %r5604, %r5603, %r5602, 18; + shf.l.wrap.b32 %r5605, %r5602, %r5603, 18; + mov.b64 %rd7906, {%r5605, %r5604}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5606}, %rd7872; + } + { + .reg .b32 %dummy; + mov.b64 {%r5607,%dummy}, %rd7872; + } + shf.l.wrap.b32 %r5608, %r5607, %r5606, 27; + shf.l.wrap.b32 %r5609, %r5606, %r5607, 27; + mov.b64 %rd7907, {%r5609, %r5608}; + { + .reg .b32 %dummy; + mov.b64 {%r5610,%dummy}, %rd7873; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5611}, %rd7873; + } + shf.r.wrap.b32 %r5612, %r5611, %r5610, 28; + shf.r.wrap.b32 %r5613, %r5610, %r5611, 28; + mov.b64 %rd7908, {%r5613, %r5612}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5614}, %rd7879; + } + { + .reg .b32 %dummy; + mov.b64 {%r5615,%dummy}, %rd7879; + } + shf.l.wrap.b32 %r5616, %r5615, %r5614, 10; + shf.l.wrap.b32 %r5617, %r5614, %r5615, 10; + mov.b64 %rd7909, {%r5617, %r5616}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5618}, %rd7885; + } + { + .reg .b32 %dummy; + mov.b64 {%r5619,%dummy}, %rd7885; + } + shf.l.wrap.b32 %r5620, %r5619, %r5618, 15; + shf.l.wrap.b32 %r5621, %r5618, %r5619, 15; + mov.b64 %rd7910, {%r5621, %r5620}; + { + .reg .b32 %dummy; + mov.b64 {%r5622,%dummy}, %rd7891; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5623}, %rd7891; + } + shf.r.wrap.b32 %r5624, %r5623, %r5622, 8; + shf.r.wrap.b32 %r5625, %r5622, %r5623, 8; + mov.b64 %rd7911, {%r5625, %r5624}; + { + .reg .b32 %dummy; + mov.b64 {%r5626,%dummy}, %rd7870; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5627}, %rd7870; + } + shf.r.wrap.b32 %r5628, %r5627, %r5626, 2; + shf.r.wrap.b32 %r5629, %r5626, %r5627, 2; + mov.b64 %rd7912, {%r5629, %r5628}; + { + .reg .b32 %dummy; + mov.b64 {%r5630,%dummy}, %rd7876; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5631}, %rd7876; + } + shf.r.wrap.b32 %r5632, %r5631, %r5630, 9; + shf.r.wrap.b32 %r5633, %r5630, %r5631, 9; + mov.b64 %rd7913, {%r5633, %r5632}; + { + .reg .b32 %dummy; + mov.b64 {%r5634,%dummy}, %rd7882; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5635}, %rd7882; + } + shf.r.wrap.b32 %r5636, %r5635, %r5634, 25; + shf.r.wrap.b32 %r5637, %r5634, %r5635, 25; + mov.b64 %rd7914, {%r5637, %r5636}; + { + .reg .b32 %dummy; + mov.b64 {%r5638,%dummy}, %rd7883; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5639}, %rd7883; + } + shf.r.wrap.b32 %r5640, %r5639, %r5638, 23; + shf.r.wrap.b32 %r5641, %r5638, %r5639, 23; + mov.b64 %rd7915, {%r5641, %r5640}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5642}, %rd7889; + } + { + .reg .b32 %dummy; + mov.b64 {%r5643,%dummy}, %rd7889; + } + shf.l.wrap.b32 %r5644, %r5643, %r5642, 2; + shf.l.wrap.b32 %r5645, %r5642, %r5643, 2; + mov.b64 %rd7916, {%r5645, %r5644}; + not.b64 %rd7917, %rd7893; + and.b64 %rd7918, %rd7894, %rd7917; + xor.b64 %rd7919, %rd7918, %rd7868; + not.b64 %rd7920, %rd7894; + and.b64 %rd7921, %rd7895, %rd7920; + xor.b64 %rd7922, %rd7921, %rd7893; + st.local.u64 [%rd2+8], %rd7922; + not.b64 %rd7923, %rd7895; + and.b64 %rd7924, %rd7896, %rd7923; + xor.b64 %rd7925, %rd7924, %rd7894; + st.local.u64 [%rd2+16], %rd7925; + not.b64 %rd7926, %rd7896; + and.b64 %rd7927, %rd7868, %rd7926; + xor.b64 %rd7928, %rd7927, %rd7895; + st.local.u64 [%rd2+24], %rd7928; + not.b64 %rd7929, %rd7868; + and.b64 %rd7930, %rd7893, %rd7929; + xor.b64 %rd7931, %rd7930, %rd7896; + st.local.u64 [%rd2+32], %rd7931; + not.b64 %rd7932, %rd7898; + and.b64 %rd7933, %rd7899, %rd7932; + xor.b64 %rd7934, %rd7933, %rd7897; + st.local.u64 [%rd2+40], %rd7934; + not.b64 %rd7935, %rd7899; + and.b64 %rd7936, %rd7900, %rd7935; + xor.b64 %rd7937, %rd7936, %rd7898; + st.local.u64 [%rd2+48], %rd7937; + not.b64 %rd7938, %rd7900; + and.b64 %rd7939, %rd7901, %rd7938; + xor.b64 %rd7940, %rd7939, %rd7899; + st.local.u64 [%rd2+56], %rd7940; + not.b64 %rd7941, %rd7901; + and.b64 %rd7942, %rd7897, %rd7941; + xor.b64 %rd7943, %rd7942, %rd7900; + st.local.u64 [%rd2+64], %rd7943; + not.b64 %rd7944, %rd7897; + and.b64 %rd7945, %rd7898, %rd7944; + xor.b64 %rd7946, %rd7945, %rd7901; + st.local.u64 [%rd2+72], %rd7946; + not.b64 %rd7947, %rd7903; + and.b64 %rd7948, %rd7904, %rd7947; + xor.b64 %rd7949, %rd7948, %rd7902; + st.local.u64 [%rd2+80], %rd7949; + not.b64 %rd7950, %rd7904; + and.b64 %rd7951, %rd7905, %rd7950; + xor.b64 %rd7952, %rd7951, %rd7903; + st.local.u64 [%rd2+88], %rd7952; + not.b64 %rd7953, %rd7905; + and.b64 %rd7954, %rd7906, %rd7953; + xor.b64 %rd7955, %rd7954, %rd7904; + st.local.u64 [%rd2+96], %rd7955; + not.b64 %rd7956, %rd7906; + and.b64 %rd7957, %rd7902, %rd7956; + xor.b64 %rd7958, %rd7957, %rd7905; + st.local.u64 [%rd2+104], %rd7958; + not.b64 %rd7959, %rd7902; + and.b64 %rd7960, %rd7903, %rd7959; + xor.b64 %rd7961, %rd7960, %rd7906; + st.local.u64 [%rd2+112], %rd7961; + not.b64 %rd7962, %rd7908; + and.b64 %rd7963, %rd7909, %rd7962; + xor.b64 %rd7964, %rd7963, %rd7907; + st.local.u64 [%rd2+120], %rd7964; + not.b64 %rd7965, %rd7909; + and.b64 %rd7966, %rd7910, %rd7965; + xor.b64 %rd7967, %rd7966, %rd7908; + st.local.u64 [%rd2+128], %rd7967; + not.b64 %rd7968, %rd7910; + and.b64 %rd7969, %rd7911, %rd7968; + xor.b64 %rd7970, %rd7969, %rd7909; + st.local.u64 [%rd2+136], %rd7970; + not.b64 %rd7971, %rd7911; + and.b64 %rd7972, %rd7907, %rd7971; + xor.b64 %rd7973, %rd7972, %rd7910; + st.local.u64 [%rd2+144], %rd7973; + not.b64 %rd7974, %rd7907; + and.b64 %rd7975, %rd7908, %rd7974; + xor.b64 %rd7976, %rd7975, %rd7911; + st.local.u64 [%rd2+152], %rd7976; + not.b64 %rd7977, %rd7913; + and.b64 %rd7978, %rd7914, %rd7977; + xor.b64 %rd7979, %rd7978, %rd7912; + st.local.u64 [%rd2+160], %rd7979; + not.b64 %rd7980, %rd7914; + and.b64 %rd7981, %rd7915, %rd7980; + xor.b64 %rd7982, %rd7981, %rd7913; + st.local.u64 [%rd2+168], %rd7982; + not.b64 %rd7983, %rd7915; + and.b64 %rd7984, %rd7916, %rd7983; + xor.b64 %rd7985, %rd7984, %rd7914; + st.local.u64 [%rd2+176], %rd7985; + not.b64 %rd7986, %rd7916; + and.b64 %rd7987, %rd7912, %rd7986; + xor.b64 %rd7988, %rd7987, %rd7915; + st.local.u64 [%rd2+184], %rd7988; + not.b64 %rd7989, %rd7912; + and.b64 %rd7990, %rd7913, %rd7989; + xor.b64 %rd7991, %rd7990, %rd7916; + st.local.u64 [%rd2+192], %rd7991; + xor.b64 %rd7992, %rd7919, %rd39; + st.local.u64 [%rd2], %rd7992; + bra.uni $L__BB1_29; + +$L__BB1_7: + setp.eq.s64 %p10, %rd15, 0; + @%p10 bra $L__BB1_15; + + max.u64 %rd218, %rd15, 1; + add.s64 %rd219, %rd218, -1; + setp.lt.u64 %p11, %rd219, 3; + mov.u64 %rd19835, 0; + mov.u64 %rd19836, %rd19835; + @%p11 bra $L__BB1_11; + + and.b64 %rd223, %rd218, 3; + sub.s64 %rd19834, %rd218, %rd223; + +$L__BB1_10: + add.s64 %rd224, %rd19836, %rd19844; + add.s64 %rd225, %rd1, %rd224; + shl.b64 %rd226, %rd19835, 3; + add.s64 %rd227, %rd2, %rd226; + ld.local.u64 %rd228, [%rd227]; + ld.local.u8 %rd229, [%rd225]; + ld.local.u8 %rd230, [%rd225+1]; + bfi.b64 %rd231, %rd230, %rd229, 8, 8; + ld.local.u8 %rd232, [%rd225+2]; + ld.local.u8 %rd233, [%rd225+3]; + bfi.b64 %rd234, %rd233, %rd232, 8, 8; + bfi.b64 %rd235, %rd234, %rd231, 16, 16; + ld.local.u8 %rd236, [%rd225+4]; + ld.local.u8 %rd237, [%rd225+5]; + bfi.b64 %rd238, %rd237, %rd236, 8, 8; + ld.local.u8 %rd239, [%rd225+6]; + ld.local.u8 %rd240, [%rd225+7]; + bfi.b64 %rd241, %rd240, %rd239, 8, 8; + bfi.b64 %rd242, %rd241, %rd238, 16, 16; + bfi.b64 %rd243, %rd242, %rd235, 32, 32; + xor.b64 %rd244, %rd228, %rd243; + st.local.u64 [%rd227], %rd244; + ld.local.u64 %rd245, [%rd227+8]; + ld.local.u8 %rd246, [%rd225+8]; + ld.local.u8 %rd247, [%rd225+9]; + bfi.b64 %rd248, %rd247, %rd246, 8, 8; + ld.local.u8 %rd249, [%rd225+10]; + ld.local.u8 %rd250, [%rd225+11]; + bfi.b64 %rd251, %rd250, %rd249, 8, 8; + bfi.b64 %rd252, %rd251, %rd248, 16, 16; + ld.local.u8 %rd253, [%rd225+12]; + ld.local.u8 %rd254, [%rd225+13]; + bfi.b64 %rd255, %rd254, %rd253, 8, 8; + ld.local.u8 %rd256, [%rd225+14]; + ld.local.u8 %rd257, [%rd225+15]; + bfi.b64 %rd258, %rd257, %rd256, 8, 8; + bfi.b64 %rd259, %rd258, %rd255, 16, 16; + bfi.b64 %rd260, %rd259, %rd252, 32, 32; + xor.b64 %rd261, %rd245, %rd260; + st.local.u64 [%rd227+8], %rd261; + ld.local.u64 %rd262, [%rd227+16]; + ld.local.u8 %rd263, [%rd225+16]; + ld.local.u8 %rd264, [%rd225+17]; + bfi.b64 %rd265, %rd264, %rd263, 8, 8; + ld.local.u8 %rd266, [%rd225+18]; + ld.local.u8 %rd267, [%rd225+19]; + bfi.b64 %rd268, %rd267, %rd266, 8, 8; + bfi.b64 %rd269, %rd268, %rd265, 16, 16; + ld.local.u8 %rd270, [%rd225+20]; + ld.local.u8 %rd271, [%rd225+21]; + bfi.b64 %rd272, %rd271, %rd270, 8, 8; + ld.local.u8 %rd273, [%rd225+22]; + ld.local.u8 %rd274, [%rd225+23]; + bfi.b64 %rd275, %rd274, %rd273, 8, 8; + bfi.b64 %rd276, %rd275, %rd272, 16, 16; + bfi.b64 %rd277, %rd276, %rd269, 32, 32; + xor.b64 %rd278, %rd262, %rd277; + st.local.u64 [%rd227+16], %rd278; + ld.local.u64 %rd279, [%rd227+24]; + ld.local.u8 %rd280, [%rd225+24]; + ld.local.u8 %rd281, [%rd225+25]; + bfi.b64 %rd282, %rd281, %rd280, 8, 8; + ld.local.u8 %rd283, [%rd225+26]; + ld.local.u8 %rd284, [%rd225+27]; + bfi.b64 %rd285, %rd284, %rd283, 8, 8; + bfi.b64 %rd286, %rd285, %rd282, 16, 16; + ld.local.u8 %rd287, [%rd225+28]; + ld.local.u8 %rd288, [%rd225+29]; + bfi.b64 %rd289, %rd288, %rd287, 8, 8; + ld.local.u8 %rd290, [%rd225+30]; + ld.local.u8 %rd291, [%rd225+31]; + bfi.b64 %rd292, %rd291, %rd290, 8, 8; + bfi.b64 %rd293, %rd292, %rd289, 16, 16; + bfi.b64 %rd294, %rd293, %rd286, 32, 32; + xor.b64 %rd295, %rd279, %rd294; + st.local.u64 [%rd227+24], %rd295; + add.s64 %rd19836, %rd19836, 32; + add.s64 %rd19835, %rd19835, 4; + add.s64 %rd19834, %rd19834, -4; + setp.ne.s64 %p12, %rd19834, 0; + @%p12 bra $L__BB1_10; + +$L__BB1_11: + and.b64 %rd297, %rd218, 3; + setp.eq.s64 %p13, %rd297, 0; + @%p13 bra $L__BB1_15; + + setp.eq.s64 %p14, %rd297, 1; + add.s64 %rd300, %rd19836, %rd19844; + add.s64 %rd52, %rd1, %rd300; + shl.b64 %rd301, %rd19835, 3; + add.s64 %rd53, %rd2, %rd301; + ld.local.u64 %rd302, [%rd53]; + ld.local.u8 %rd303, [%rd52]; + ld.local.u8 %rd304, [%rd52+1]; + bfi.b64 %rd305, %rd304, %rd303, 8, 8; + ld.local.u8 %rd306, [%rd52+2]; + ld.local.u8 %rd307, [%rd52+3]; + bfi.b64 %rd308, %rd307, %rd306, 8, 8; + bfi.b64 %rd309, %rd308, %rd305, 16, 16; + ld.local.u8 %rd310, [%rd52+4]; + ld.local.u8 %rd311, [%rd52+5]; + bfi.b64 %rd312, %rd311, %rd310, 8, 8; + ld.local.u8 %rd313, [%rd52+6]; + ld.local.u8 %rd314, [%rd52+7]; + bfi.b64 %rd315, %rd314, %rd313, 8, 8; + bfi.b64 %rd316, %rd315, %rd312, 16, 16; + bfi.b64 %rd317, %rd316, %rd309, 32, 32; + xor.b64 %rd318, %rd302, %rd317; + st.local.u64 [%rd53], %rd318; + @%p14 bra $L__BB1_15; + + setp.eq.s64 %p15, %rd297, 2; + ld.local.u64 %rd321, [%rd53+8]; + ld.local.u8 %rd322, [%rd52+8]; + ld.local.u8 %rd323, [%rd52+9]; + bfi.b64 %rd324, %rd323, %rd322, 8, 8; + ld.local.u8 %rd325, [%rd52+10]; + ld.local.u8 %rd326, [%rd52+11]; + bfi.b64 %rd327, %rd326, %rd325, 8, 8; + bfi.b64 %rd328, %rd327, %rd324, 16, 16; + ld.local.u8 %rd329, [%rd52+12]; + ld.local.u8 %rd330, [%rd52+13]; + bfi.b64 %rd331, %rd330, %rd329, 8, 8; + ld.local.u8 %rd332, [%rd52+14]; + ld.local.u8 %rd333, [%rd52+15]; + bfi.b64 %rd334, %rd333, %rd332, 8, 8; + bfi.b64 %rd335, %rd334, %rd331, 16, 16; + bfi.b64 %rd336, %rd335, %rd328, 32, 32; + xor.b64 %rd337, %rd321, %rd336; + st.local.u64 [%rd53+8], %rd337; + @%p15 bra $L__BB1_15; + + ld.local.u64 %rd338, [%rd53+16]; + ld.local.u8 %rd339, [%rd52+16]; + ld.local.u8 %rd340, [%rd52+17]; + bfi.b64 %rd341, %rd340, %rd339, 8, 8; + ld.local.u8 %rd342, [%rd52+18]; + ld.local.u8 %rd343, [%rd52+19]; + bfi.b64 %rd344, %rd343, %rd342, 8, 8; + bfi.b64 %rd345, %rd344, %rd341, 16, 16; + ld.local.u8 %rd346, [%rd52+20]; + ld.local.u8 %rd347, [%rd52+21]; + bfi.b64 %rd348, %rd347, %rd346, 8, 8; + ld.local.u8 %rd349, [%rd52+22]; + ld.local.u8 %rd350, [%rd52+23]; + bfi.b64 %rd351, %rd350, %rd349, 8, 8; + bfi.b64 %rd352, %rd351, %rd348, 16, 16; + bfi.b64 %rd353, %rd352, %rd345, 32, 32; + xor.b64 %rd354, %rd338, %rd353; + st.local.u64 [%rd53+16], %rd354; + +$L__BB1_15: + ld.local.u64 %rd356, [%rd2+40]; + mov.u64 %rd19843, 0; + ld.local.u64 %rd357, [%rd2]; + xor.b64 %rd358, %rd356, %rd357; + ld.local.u64 %rd359, [%rd2+80]; + xor.b64 %rd360, %rd358, %rd359; + ld.local.u64 %rd361, [%rd2+120]; + xor.b64 %rd362, %rd360, %rd361; + ld.local.u64 %rd363, [%rd2+160]; + xor.b64 %rd364, %rd362, %rd363; + ld.local.u64 %rd365, [%rd2+48]; + ld.local.u64 %rd366, [%rd2+8]; + xor.b64 %rd367, %rd365, %rd366; + ld.local.u64 %rd368, [%rd2+88]; + xor.b64 %rd369, %rd367, %rd368; + ld.local.u64 %rd370, [%rd2+128]; + xor.b64 %rd371, %rd369, %rd370; + ld.local.u64 %rd372, [%rd2+168]; + xor.b64 %rd373, %rd371, %rd372; + ld.local.u64 %rd374, [%rd2+56]; + ld.local.u64 %rd375, [%rd2+16]; + xor.b64 %rd376, %rd374, %rd375; + ld.local.u64 %rd377, [%rd2+96]; + xor.b64 %rd378, %rd376, %rd377; + ld.local.u64 %rd379, [%rd2+136]; + xor.b64 %rd380, %rd378, %rd379; + ld.local.u64 %rd381, [%rd2+176]; + xor.b64 %rd382, %rd380, %rd381; + mov.u64 %rd383, 64; + ld.local.u64 %rd384, [%rd2+64]; + ld.local.u64 %rd385, [%rd2+24]; + xor.b64 %rd386, %rd384, %rd385; + ld.local.u64 %rd387, [%rd2+104]; + xor.b64 %rd388, %rd386, %rd387; + ld.local.u64 %rd389, [%rd2+144]; + xor.b64 %rd390, %rd388, %rd389; + ld.local.u64 %rd391, [%rd2+184]; + xor.b64 %rd392, %rd390, %rd391; + ld.local.u64 %rd393, [%rd2+72]; + ld.local.u64 %rd394, [%rd2+32]; + xor.b64 %rd395, %rd393, %rd394; + ld.local.u64 %rd396, [%rd2+112]; + xor.b64 %rd397, %rd395, %rd396; + ld.local.u64 %rd398, [%rd2+152]; + xor.b64 %rd399, %rd397, %rd398; + ld.local.u64 %rd400, [%rd2+192]; + xor.b64 %rd401, %rd399, %rd400; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r78}, %rd373; + } + { + .reg .b32 %dummy; + mov.b64 {%r79,%dummy}, %rd373; + } + shf.l.wrap.b32 %r80, %r79, %r78, 1; + shf.l.wrap.b32 %r81, %r78, %r79, 1; + mov.b64 %rd402, {%r81, %r80}; + xor.b64 %rd403, %rd401, %rd402; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r82}, %rd382; + } + { + .reg .b32 %dummy; + mov.b64 {%r83,%dummy}, %rd382; + } + shf.l.wrap.b32 %r84, %r83, %r82, 1; + shf.l.wrap.b32 %r85, %r82, %r83, 1; + mov.b64 %rd404, {%r85, %r84}; + xor.b64 %rd405, %rd404, %rd364; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r86}, %rd392; + } + { + .reg .b32 %dummy; + mov.b64 {%r87,%dummy}, %rd392; + } + shf.l.wrap.b32 %r88, %r87, %r86, 1; + shf.l.wrap.b32 %r89, %r86, %r87, 1; + mov.b64 %rd406, {%r89, %r88}; + xor.b64 %rd407, %rd406, %rd373; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r90}, %rd401; + } + { + .reg .b32 %dummy; + mov.b64 {%r91,%dummy}, %rd401; + } + shf.l.wrap.b32 %r92, %r91, %r90, 1; + shf.l.wrap.b32 %r93, %r90, %r91, 1; + mov.b64 %rd408, {%r93, %r92}; + xor.b64 %rd409, %rd408, %rd382; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r94}, %rd364; + } + { + .reg .b32 %dummy; + mov.b64 {%r95,%dummy}, %rd364; + } + shf.l.wrap.b32 %r96, %r95, %r94, 1; + shf.l.wrap.b32 %r97, %r94, %r95, 1; + mov.b64 %rd410, {%r97, %r96}; + xor.b64 %rd411, %rd392, %rd410; + xor.b64 %rd412, %rd403, %rd357; + xor.b64 %rd413, %rd405, %rd366; + xor.b64 %rd414, %rd407, %rd375; + xor.b64 %rd415, %rd409, %rd385; + xor.b64 %rd416, %rd411, %rd394; + xor.b64 %rd417, %rd403, %rd356; + xor.b64 %rd418, %rd405, %rd365; + xor.b64 %rd419, %rd407, %rd374; + xor.b64 %rd420, %rd409, %rd384; + xor.b64 %rd421, %rd411, %rd393; + xor.b64 %rd422, %rd403, %rd359; + xor.b64 %rd423, %rd405, %rd368; + xor.b64 %rd424, %rd407, %rd377; + xor.b64 %rd425, %rd409, %rd387; + xor.b64 %rd426, %rd411, %rd396; + xor.b64 %rd427, %rd403, %rd361; + xor.b64 %rd428, %rd405, %rd370; + xor.b64 %rd429, %rd407, %rd379; + xor.b64 %rd430, %rd409, %rd389; + xor.b64 %rd431, %rd411, %rd398; + xor.b64 %rd432, %rd403, %rd363; + xor.b64 %rd433, %rd405, %rd372; + xor.b64 %rd434, %rd407, %rd381; + xor.b64 %rd435, %rd409, %rd391; + xor.b64 %rd436, %rd411, %rd400; + { + .reg .b32 %dummy; + mov.b64 {%r98,%dummy}, %rd418; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r99}, %rd418; + } + shf.r.wrap.b32 %r100, %r99, %r98, 20; + shf.r.wrap.b32 %r101, %r98, %r99, 20; + mov.b64 %rd437, {%r101, %r100}; + { + .reg .b32 %dummy; + mov.b64 {%r102,%dummy}, %rd424; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r103}, %rd424; + } + shf.r.wrap.b32 %r104, %r103, %r102, 21; + shf.r.wrap.b32 %r105, %r102, %r103, 21; + mov.b64 %rd438, {%r105, %r104}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r106}, %rd430; + } + { + .reg .b32 %dummy; + mov.b64 {%r107,%dummy}, %rd430; + } + shf.l.wrap.b32 %r108, %r107, %r106, 21; + shf.l.wrap.b32 %r109, %r106, %r107, 21; + mov.b64 %rd439, {%r109, %r108}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r110}, %rd436; + } + { + .reg .b32 %dummy; + mov.b64 {%r111,%dummy}, %rd436; + } + shf.l.wrap.b32 %r112, %r111, %r110, 14; + shf.l.wrap.b32 %r113, %r110, %r111, 14; + mov.b64 %rd440, {%r113, %r112}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r114}, %rd415; + } + { + .reg .b32 %dummy; + mov.b64 {%r115,%dummy}, %rd415; + } + shf.l.wrap.b32 %r116, %r115, %r114, 28; + shf.l.wrap.b32 %r117, %r114, %r115, 28; + mov.b64 %rd441, {%r117, %r116}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r118}, %rd421; + } + { + .reg .b32 %dummy; + mov.b64 {%r119,%dummy}, %rd421; + } + shf.l.wrap.b32 %r120, %r119, %r118, 20; + shf.l.wrap.b32 %r121, %r118, %r119, 20; + mov.b64 %rd442, {%r121, %r120}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r122}, %rd422; + } + { + .reg .b32 %dummy; + mov.b64 {%r123,%dummy}, %rd422; + } + shf.l.wrap.b32 %r124, %r123, %r122, 3; + shf.l.wrap.b32 %r125, %r122, %r123, 3; + mov.b64 %rd443, {%r125, %r124}; + { + .reg .b32 %dummy; + mov.b64 {%r126,%dummy}, %rd428; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r127}, %rd428; + } + shf.r.wrap.b32 %r128, %r127, %r126, 19; + shf.r.wrap.b32 %r129, %r126, %r127, 19; + mov.b64 %rd444, {%r129, %r128}; + { + .reg .b32 %dummy; + mov.b64 {%r130,%dummy}, %rd434; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r131}, %rd434; + } + shf.r.wrap.b32 %r132, %r131, %r130, 3; + shf.r.wrap.b32 %r133, %r130, %r131, 3; + mov.b64 %rd445, {%r133, %r132}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r134}, %rd413; + } + { + .reg .b32 %dummy; + mov.b64 {%r135,%dummy}, %rd413; + } + shf.l.wrap.b32 %r136, %r135, %r134, 1; + shf.l.wrap.b32 %r137, %r134, %r135, 1; + mov.b64 %rd446, {%r137, %r136}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r138}, %rd419; + } + { + .reg .b32 %dummy; + mov.b64 {%r139,%dummy}, %rd419; + } + shf.l.wrap.b32 %r140, %r139, %r138, 6; + shf.l.wrap.b32 %r141, %r138, %r139, 6; + mov.b64 %rd447, {%r141, %r140}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r142}, %rd425; + } + { + .reg .b32 %dummy; + mov.b64 {%r143,%dummy}, %rd425; + } + shf.l.wrap.b32 %r144, %r143, %r142, 25; + shf.l.wrap.b32 %r145, %r142, %r143, 25; + mov.b64 %rd448, {%r145, %r144}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r146}, %rd431; + } + { + .reg .b32 %dummy; + mov.b64 {%r147,%dummy}, %rd431; + } + shf.l.wrap.b32 %r148, %r147, %r146, 8; + shf.l.wrap.b32 %r149, %r146, %r147, 8; + mov.b64 %rd449, {%r149, %r148}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r150}, %rd432; + } + { + .reg .b32 %dummy; + mov.b64 {%r151,%dummy}, %rd432; + } + shf.l.wrap.b32 %r152, %r151, %r150, 18; + shf.l.wrap.b32 %r153, %r150, %r151, 18; + mov.b64 %rd450, {%r153, %r152}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r154}, %rd416; + } + { + .reg .b32 %dummy; + mov.b64 {%r155,%dummy}, %rd416; + } + shf.l.wrap.b32 %r156, %r155, %r154, 27; + shf.l.wrap.b32 %r157, %r154, %r155, 27; + mov.b64 %rd451, {%r157, %r156}; + { + .reg .b32 %dummy; + mov.b64 {%r158,%dummy}, %rd417; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r159}, %rd417; + } + shf.r.wrap.b32 %r160, %r159, %r158, 28; + shf.r.wrap.b32 %r161, %r158, %r159, 28; + mov.b64 %rd452, {%r161, %r160}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r162}, %rd423; + } + { + .reg .b32 %dummy; + mov.b64 {%r163,%dummy}, %rd423; + } + shf.l.wrap.b32 %r164, %r163, %r162, 10; + shf.l.wrap.b32 %r165, %r162, %r163, 10; + mov.b64 %rd453, {%r165, %r164}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r166}, %rd429; + } + { + .reg .b32 %dummy; + mov.b64 {%r167,%dummy}, %rd429; + } + shf.l.wrap.b32 %r168, %r167, %r166, 15; + shf.l.wrap.b32 %r169, %r166, %r167, 15; + mov.b64 %rd454, {%r169, %r168}; + { + .reg .b32 %dummy; + mov.b64 {%r170,%dummy}, %rd435; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r171}, %rd435; + } + shf.r.wrap.b32 %r172, %r171, %r170, 8; + shf.r.wrap.b32 %r173, %r170, %r171, 8; + mov.b64 %rd455, {%r173, %r172}; + { + .reg .b32 %dummy; + mov.b64 {%r174,%dummy}, %rd414; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r175}, %rd414; + } + shf.r.wrap.b32 %r176, %r175, %r174, 2; + shf.r.wrap.b32 %r177, %r174, %r175, 2; + mov.b64 %rd456, {%r177, %r176}; + { + .reg .b32 %dummy; + mov.b64 {%r178,%dummy}, %rd420; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r179}, %rd420; + } + shf.r.wrap.b32 %r180, %r179, %r178, 9; + shf.r.wrap.b32 %r181, %r178, %r179, 9; + mov.b64 %rd457, {%r181, %r180}; + { + .reg .b32 %dummy; + mov.b64 {%r182,%dummy}, %rd426; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r183}, %rd426; + } + shf.r.wrap.b32 %r184, %r183, %r182, 25; + shf.r.wrap.b32 %r185, %r182, %r183, 25; + mov.b64 %rd458, {%r185, %r184}; + { + .reg .b32 %dummy; + mov.b64 {%r186,%dummy}, %rd427; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r187}, %rd427; + } + shf.r.wrap.b32 %r188, %r187, %r186, 23; + shf.r.wrap.b32 %r189, %r186, %r187, 23; + mov.b64 %rd459, {%r189, %r188}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r190}, %rd433; + } + { + .reg .b32 %dummy; + mov.b64 {%r191,%dummy}, %rd433; + } + shf.l.wrap.b32 %r192, %r191, %r190, 2; + shf.l.wrap.b32 %r193, %r190, %r191, 2; + mov.b64 %rd460, {%r193, %r192}; + not.b64 %rd461, %rd437; + and.b64 %rd462, %rd438, %rd461; + xor.b64 %rd463, %rd462, %rd412; + not.b64 %rd464, %rd438; + and.b64 %rd465, %rd439, %rd464; + xor.b64 %rd466, %rd465, %rd437; + not.b64 %rd467, %rd439; + and.b64 %rd468, %rd440, %rd467; + xor.b64 %rd469, %rd468, %rd438; + not.b64 %rd470, %rd440; + and.b64 %rd471, %rd412, %rd470; + xor.b64 %rd472, %rd471, %rd439; + not.b64 %rd473, %rd412; + and.b64 %rd474, %rd437, %rd473; + xor.b64 %rd475, %rd474, %rd440; + not.b64 %rd476, %rd442; + and.b64 %rd477, %rd443, %rd476; + xor.b64 %rd478, %rd477, %rd441; + not.b64 %rd479, %rd443; + and.b64 %rd480, %rd444, %rd479; + xor.b64 %rd481, %rd480, %rd442; + not.b64 %rd482, %rd444; + and.b64 %rd483, %rd445, %rd482; + xor.b64 %rd484, %rd483, %rd443; + not.b64 %rd485, %rd445; + and.b64 %rd486, %rd441, %rd485; + xor.b64 %rd487, %rd486, %rd444; + not.b64 %rd488, %rd441; + and.b64 %rd489, %rd442, %rd488; + xor.b64 %rd490, %rd489, %rd445; + not.b64 %rd491, %rd447; + and.b64 %rd492, %rd448, %rd491; + xor.b64 %rd493, %rd492, %rd446; + not.b64 %rd494, %rd448; + and.b64 %rd495, %rd449, %rd494; + xor.b64 %rd496, %rd495, %rd447; + not.b64 %rd497, %rd449; + and.b64 %rd498, %rd450, %rd497; + xor.b64 %rd499, %rd498, %rd448; + not.b64 %rd500, %rd450; + and.b64 %rd501, %rd446, %rd500; + xor.b64 %rd502, %rd501, %rd449; + not.b64 %rd503, %rd446; + and.b64 %rd504, %rd447, %rd503; + xor.b64 %rd505, %rd504, %rd450; + not.b64 %rd506, %rd452; + and.b64 %rd507, %rd453, %rd506; + xor.b64 %rd508, %rd507, %rd451; + not.b64 %rd509, %rd453; + and.b64 %rd510, %rd454, %rd509; + xor.b64 %rd511, %rd510, %rd452; + not.b64 %rd512, %rd454; + and.b64 %rd513, %rd455, %rd512; + xor.b64 %rd514, %rd513, %rd453; + not.b64 %rd515, %rd455; + and.b64 %rd516, %rd451, %rd515; + xor.b64 %rd517, %rd516, %rd454; + not.b64 %rd518, %rd451; + and.b64 %rd519, %rd452, %rd518; + xor.b64 %rd520, %rd519, %rd455; + not.b64 %rd521, %rd457; + and.b64 %rd522, %rd458, %rd521; + xor.b64 %rd523, %rd522, %rd456; + not.b64 %rd524, %rd458; + and.b64 %rd525, %rd459, %rd524; + xor.b64 %rd526, %rd525, %rd457; + not.b64 %rd527, %rd459; + and.b64 %rd528, %rd460, %rd527; + xor.b64 %rd529, %rd528, %rd458; + not.b64 %rd530, %rd460; + and.b64 %rd531, %rd456, %rd530; + xor.b64 %rd532, %rd531, %rd459; + not.b64 %rd533, %rd456; + and.b64 %rd534, %rd457, %rd533; + xor.b64 %rd535, %rd534, %rd460; + xor.b64 %rd536, %rd463, %rd16; + xor.b64 %rd537, %rd478, %rd536; + xor.b64 %rd538, %rd537, %rd493; + xor.b64 %rd539, %rd538, %rd508; + xor.b64 %rd540, %rd539, %rd523; + xor.b64 %rd541, %rd481, %rd466; + xor.b64 %rd542, %rd541, %rd496; + xor.b64 %rd543, %rd542, %rd511; + xor.b64 %rd544, %rd543, %rd526; + xor.b64 %rd545, %rd484, %rd469; + xor.b64 %rd546, %rd545, %rd499; + xor.b64 %rd547, %rd546, %rd514; + xor.b64 %rd548, %rd547, %rd529; + xor.b64 %rd549, %rd487, %rd472; + xor.b64 %rd550, %rd549, %rd502; + xor.b64 %rd551, %rd550, %rd517; + xor.b64 %rd552, %rd551, %rd532; + xor.b64 %rd553, %rd490, %rd475; + xor.b64 %rd554, %rd553, %rd505; + xor.b64 %rd555, %rd554, %rd520; + xor.b64 %rd556, %rd555, %rd535; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r194}, %rd544; + } + { + .reg .b32 %dummy; + mov.b64 {%r195,%dummy}, %rd544; + } + shf.l.wrap.b32 %r196, %r195, %r194, 1; + shf.l.wrap.b32 %r197, %r194, %r195, 1; + mov.b64 %rd557, {%r197, %r196}; + xor.b64 %rd558, %rd556, %rd557; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r198}, %rd548; + } + { + .reg .b32 %dummy; + mov.b64 {%r199,%dummy}, %rd548; + } + shf.l.wrap.b32 %r200, %r199, %r198, 1; + shf.l.wrap.b32 %r201, %r198, %r199, 1; + mov.b64 %rd559, {%r201, %r200}; + xor.b64 %rd560, %rd559, %rd540; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r202}, %rd552; + } + { + .reg .b32 %dummy; + mov.b64 {%r203,%dummy}, %rd552; + } + shf.l.wrap.b32 %r204, %r203, %r202, 1; + shf.l.wrap.b32 %r205, %r202, %r203, 1; + mov.b64 %rd561, {%r205, %r204}; + xor.b64 %rd562, %rd561, %rd544; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r206}, %rd556; + } + { + .reg .b32 %dummy; + mov.b64 {%r207,%dummy}, %rd556; + } + shf.l.wrap.b32 %r208, %r207, %r206, 1; + shf.l.wrap.b32 %r209, %r206, %r207, 1; + mov.b64 %rd563, {%r209, %r208}; + xor.b64 %rd564, %rd563, %rd548; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r210}, %rd540; + } + { + .reg .b32 %dummy; + mov.b64 {%r211,%dummy}, %rd540; + } + shf.l.wrap.b32 %r212, %r211, %r210, 1; + shf.l.wrap.b32 %r213, %r210, %r211, 1; + mov.b64 %rd565, {%r213, %r212}; + xor.b64 %rd566, %rd552, %rd565; + xor.b64 %rd567, %rd558, %rd536; + xor.b64 %rd568, %rd560, %rd466; + xor.b64 %rd569, %rd562, %rd469; + xor.b64 %rd570, %rd564, %rd472; + xor.b64 %rd571, %rd566, %rd475; + xor.b64 %rd572, %rd558, %rd478; + xor.b64 %rd573, %rd560, %rd481; + xor.b64 %rd574, %rd562, %rd484; + xor.b64 %rd575, %rd564, %rd487; + xor.b64 %rd576, %rd566, %rd490; + xor.b64 %rd577, %rd558, %rd493; + xor.b64 %rd578, %rd560, %rd496; + xor.b64 %rd579, %rd562, %rd499; + xor.b64 %rd580, %rd564, %rd502; + xor.b64 %rd581, %rd566, %rd505; + xor.b64 %rd582, %rd558, %rd508; + xor.b64 %rd583, %rd560, %rd511; + xor.b64 %rd584, %rd562, %rd514; + xor.b64 %rd585, %rd564, %rd517; + xor.b64 %rd586, %rd566, %rd520; + xor.b64 %rd587, %rd558, %rd523; + xor.b64 %rd588, %rd560, %rd526; + xor.b64 %rd589, %rd562, %rd529; + xor.b64 %rd590, %rd564, %rd532; + xor.b64 %rd591, %rd566, %rd535; + { + .reg .b32 %dummy; + mov.b64 {%r214,%dummy}, %rd573; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r215}, %rd573; + } + shf.r.wrap.b32 %r216, %r215, %r214, 20; + shf.r.wrap.b32 %r217, %r214, %r215, 20; + mov.b64 %rd592, {%r217, %r216}; + { + .reg .b32 %dummy; + mov.b64 {%r218,%dummy}, %rd579; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r219}, %rd579; + } + shf.r.wrap.b32 %r220, %r219, %r218, 21; + shf.r.wrap.b32 %r221, %r218, %r219, 21; + mov.b64 %rd593, {%r221, %r220}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r222}, %rd585; + } + { + .reg .b32 %dummy; + mov.b64 {%r223,%dummy}, %rd585; + } + shf.l.wrap.b32 %r224, %r223, %r222, 21; + shf.l.wrap.b32 %r225, %r222, %r223, 21; + mov.b64 %rd594, {%r225, %r224}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r226}, %rd591; + } + { + .reg .b32 %dummy; + mov.b64 {%r227,%dummy}, %rd591; + } + shf.l.wrap.b32 %r228, %r227, %r226, 14; + shf.l.wrap.b32 %r229, %r226, %r227, 14; + mov.b64 %rd595, {%r229, %r228}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r230}, %rd570; + } + { + .reg .b32 %dummy; + mov.b64 {%r231,%dummy}, %rd570; + } + shf.l.wrap.b32 %r232, %r231, %r230, 28; + shf.l.wrap.b32 %r233, %r230, %r231, 28; + mov.b64 %rd596, {%r233, %r232}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r234}, %rd576; + } + { + .reg .b32 %dummy; + mov.b64 {%r235,%dummy}, %rd576; + } + shf.l.wrap.b32 %r236, %r235, %r234, 20; + shf.l.wrap.b32 %r237, %r234, %r235, 20; + mov.b64 %rd597, {%r237, %r236}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r238}, %rd577; + } + { + .reg .b32 %dummy; + mov.b64 {%r239,%dummy}, %rd577; + } + shf.l.wrap.b32 %r240, %r239, %r238, 3; + shf.l.wrap.b32 %r241, %r238, %r239, 3; + mov.b64 %rd598, {%r241, %r240}; + { + .reg .b32 %dummy; + mov.b64 {%r242,%dummy}, %rd583; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r243}, %rd583; + } + shf.r.wrap.b32 %r244, %r243, %r242, 19; + shf.r.wrap.b32 %r245, %r242, %r243, 19; + mov.b64 %rd599, {%r245, %r244}; + { + .reg .b32 %dummy; + mov.b64 {%r246,%dummy}, %rd589; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r247}, %rd589; + } + shf.r.wrap.b32 %r248, %r247, %r246, 3; + shf.r.wrap.b32 %r249, %r246, %r247, 3; + mov.b64 %rd600, {%r249, %r248}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r250}, %rd568; + } + { + .reg .b32 %dummy; + mov.b64 {%r251,%dummy}, %rd568; + } + shf.l.wrap.b32 %r252, %r251, %r250, 1; + shf.l.wrap.b32 %r253, %r250, %r251, 1; + mov.b64 %rd601, {%r253, %r252}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r254}, %rd574; + } + { + .reg .b32 %dummy; + mov.b64 {%r255,%dummy}, %rd574; + } + shf.l.wrap.b32 %r256, %r255, %r254, 6; + shf.l.wrap.b32 %r257, %r254, %r255, 6; + mov.b64 %rd602, {%r257, %r256}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r258}, %rd580; + } + { + .reg .b32 %dummy; + mov.b64 {%r259,%dummy}, %rd580; + } + shf.l.wrap.b32 %r260, %r259, %r258, 25; + shf.l.wrap.b32 %r261, %r258, %r259, 25; + mov.b64 %rd603, {%r261, %r260}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r262}, %rd586; + } + { + .reg .b32 %dummy; + mov.b64 {%r263,%dummy}, %rd586; + } + shf.l.wrap.b32 %r264, %r263, %r262, 8; + shf.l.wrap.b32 %r265, %r262, %r263, 8; + mov.b64 %rd604, {%r265, %r264}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r266}, %rd587; + } + { + .reg .b32 %dummy; + mov.b64 {%r267,%dummy}, %rd587; + } + shf.l.wrap.b32 %r268, %r267, %r266, 18; + shf.l.wrap.b32 %r269, %r266, %r267, 18; + mov.b64 %rd605, {%r269, %r268}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r270}, %rd571; + } + { + .reg .b32 %dummy; + mov.b64 {%r271,%dummy}, %rd571; + } + shf.l.wrap.b32 %r272, %r271, %r270, 27; + shf.l.wrap.b32 %r273, %r270, %r271, 27; + mov.b64 %rd606, {%r273, %r272}; + { + .reg .b32 %dummy; + mov.b64 {%r274,%dummy}, %rd572; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r275}, %rd572; + } + shf.r.wrap.b32 %r276, %r275, %r274, 28; + shf.r.wrap.b32 %r277, %r274, %r275, 28; + mov.b64 %rd607, {%r277, %r276}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r278}, %rd578; + } + { + .reg .b32 %dummy; + mov.b64 {%r279,%dummy}, %rd578; + } + shf.l.wrap.b32 %r280, %r279, %r278, 10; + shf.l.wrap.b32 %r281, %r278, %r279, 10; + mov.b64 %rd608, {%r281, %r280}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r282}, %rd584; + } + { + .reg .b32 %dummy; + mov.b64 {%r283,%dummy}, %rd584; + } + shf.l.wrap.b32 %r284, %r283, %r282, 15; + shf.l.wrap.b32 %r285, %r282, %r283, 15; + mov.b64 %rd609, {%r285, %r284}; + { + .reg .b32 %dummy; + mov.b64 {%r286,%dummy}, %rd590; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r287}, %rd590; + } + shf.r.wrap.b32 %r288, %r287, %r286, 8; + shf.r.wrap.b32 %r289, %r286, %r287, 8; + mov.b64 %rd610, {%r289, %r288}; + { + .reg .b32 %dummy; + mov.b64 {%r290,%dummy}, %rd569; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r291}, %rd569; + } + shf.r.wrap.b32 %r292, %r291, %r290, 2; + shf.r.wrap.b32 %r293, %r290, %r291, 2; + mov.b64 %rd611, {%r293, %r292}; + { + .reg .b32 %dummy; + mov.b64 {%r294,%dummy}, %rd575; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r295}, %rd575; + } + shf.r.wrap.b32 %r296, %r295, %r294, 9; + shf.r.wrap.b32 %r297, %r294, %r295, 9; + mov.b64 %rd612, {%r297, %r296}; + { + .reg .b32 %dummy; + mov.b64 {%r298,%dummy}, %rd581; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r299}, %rd581; + } + shf.r.wrap.b32 %r300, %r299, %r298, 25; + shf.r.wrap.b32 %r301, %r298, %r299, 25; + mov.b64 %rd613, {%r301, %r300}; + { + .reg .b32 %dummy; + mov.b64 {%r302,%dummy}, %rd582; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r303}, %rd582; + } + shf.r.wrap.b32 %r304, %r303, %r302, 23; + shf.r.wrap.b32 %r305, %r302, %r303, 23; + mov.b64 %rd614, {%r305, %r304}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r306}, %rd588; + } + { + .reg .b32 %dummy; + mov.b64 {%r307,%dummy}, %rd588; + } + shf.l.wrap.b32 %r308, %r307, %r306, 2; + shf.l.wrap.b32 %r309, %r306, %r307, 2; + mov.b64 %rd615, {%r309, %r308}; + not.b64 %rd616, %rd592; + and.b64 %rd617, %rd593, %rd616; + xor.b64 %rd618, %rd617, %rd567; + not.b64 %rd619, %rd593; + and.b64 %rd620, %rd594, %rd619; + xor.b64 %rd621, %rd620, %rd592; + not.b64 %rd622, %rd594; + and.b64 %rd623, %rd595, %rd622; + xor.b64 %rd624, %rd623, %rd593; + not.b64 %rd625, %rd595; + and.b64 %rd626, %rd567, %rd625; + xor.b64 %rd627, %rd626, %rd594; + not.b64 %rd628, %rd567; + and.b64 %rd629, %rd592, %rd628; + xor.b64 %rd630, %rd629, %rd595; + not.b64 %rd631, %rd597; + and.b64 %rd632, %rd598, %rd631; + xor.b64 %rd633, %rd632, %rd596; + not.b64 %rd634, %rd598; + and.b64 %rd635, %rd599, %rd634; + xor.b64 %rd636, %rd635, %rd597; + not.b64 %rd637, %rd599; + and.b64 %rd638, %rd600, %rd637; + xor.b64 %rd639, %rd638, %rd598; + not.b64 %rd640, %rd600; + and.b64 %rd641, %rd596, %rd640; + xor.b64 %rd642, %rd641, %rd599; + not.b64 %rd643, %rd596; + and.b64 %rd644, %rd597, %rd643; + xor.b64 %rd645, %rd644, %rd600; + not.b64 %rd646, %rd602; + and.b64 %rd647, %rd603, %rd646; + xor.b64 %rd648, %rd647, %rd601; + not.b64 %rd649, %rd603; + and.b64 %rd650, %rd604, %rd649; + xor.b64 %rd651, %rd650, %rd602; + not.b64 %rd652, %rd604; + and.b64 %rd653, %rd605, %rd652; + xor.b64 %rd654, %rd653, %rd603; + not.b64 %rd655, %rd605; + and.b64 %rd656, %rd601, %rd655; + xor.b64 %rd657, %rd656, %rd604; + not.b64 %rd658, %rd601; + and.b64 %rd659, %rd602, %rd658; + xor.b64 %rd660, %rd659, %rd605; + not.b64 %rd661, %rd607; + and.b64 %rd662, %rd608, %rd661; + xor.b64 %rd663, %rd662, %rd606; + not.b64 %rd664, %rd608; + and.b64 %rd665, %rd609, %rd664; + xor.b64 %rd666, %rd665, %rd607; + not.b64 %rd667, %rd609; + and.b64 %rd668, %rd610, %rd667; + xor.b64 %rd669, %rd668, %rd608; + not.b64 %rd670, %rd610; + and.b64 %rd671, %rd606, %rd670; + xor.b64 %rd672, %rd671, %rd609; + not.b64 %rd673, %rd606; + and.b64 %rd674, %rd607, %rd673; + xor.b64 %rd675, %rd674, %rd610; + not.b64 %rd676, %rd612; + and.b64 %rd677, %rd613, %rd676; + xor.b64 %rd678, %rd677, %rd611; + not.b64 %rd679, %rd613; + and.b64 %rd680, %rd614, %rd679; + xor.b64 %rd681, %rd680, %rd612; + not.b64 %rd682, %rd614; + and.b64 %rd683, %rd615, %rd682; + xor.b64 %rd684, %rd683, %rd613; + not.b64 %rd685, %rd615; + and.b64 %rd686, %rd611, %rd685; + xor.b64 %rd687, %rd686, %rd614; + not.b64 %rd688, %rd611; + and.b64 %rd689, %rd612, %rd688; + xor.b64 %rd690, %rd689, %rd615; + xor.b64 %rd691, %rd618, %rd17; + xor.b64 %rd692, %rd633, %rd691; + xor.b64 %rd693, %rd692, %rd648; + xor.b64 %rd694, %rd693, %rd663; + xor.b64 %rd695, %rd694, %rd678; + xor.b64 %rd696, %rd636, %rd621; + xor.b64 %rd697, %rd696, %rd651; + xor.b64 %rd698, %rd697, %rd666; + xor.b64 %rd699, %rd698, %rd681; + xor.b64 %rd700, %rd639, %rd624; + xor.b64 %rd701, %rd700, %rd654; + xor.b64 %rd702, %rd701, %rd669; + xor.b64 %rd703, %rd702, %rd684; + xor.b64 %rd704, %rd642, %rd627; + xor.b64 %rd705, %rd704, %rd657; + xor.b64 %rd706, %rd705, %rd672; + xor.b64 %rd707, %rd706, %rd687; + xor.b64 %rd708, %rd645, %rd630; + xor.b64 %rd709, %rd708, %rd660; + xor.b64 %rd710, %rd709, %rd675; + xor.b64 %rd711, %rd710, %rd690; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r310}, %rd699; + } + { + .reg .b32 %dummy; + mov.b64 {%r311,%dummy}, %rd699; + } + shf.l.wrap.b32 %r312, %r311, %r310, 1; + shf.l.wrap.b32 %r313, %r310, %r311, 1; + mov.b64 %rd712, {%r313, %r312}; + xor.b64 %rd713, %rd711, %rd712; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r314}, %rd703; + } + { + .reg .b32 %dummy; + mov.b64 {%r315,%dummy}, %rd703; + } + shf.l.wrap.b32 %r316, %r315, %r314, 1; + shf.l.wrap.b32 %r317, %r314, %r315, 1; + mov.b64 %rd714, {%r317, %r316}; + xor.b64 %rd715, %rd714, %rd695; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r318}, %rd707; + } + { + .reg .b32 %dummy; + mov.b64 {%r319,%dummy}, %rd707; + } + shf.l.wrap.b32 %r320, %r319, %r318, 1; + shf.l.wrap.b32 %r321, %r318, %r319, 1; + mov.b64 %rd716, {%r321, %r320}; + xor.b64 %rd717, %rd716, %rd699; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r322}, %rd711; + } + { + .reg .b32 %dummy; + mov.b64 {%r323,%dummy}, %rd711; + } + shf.l.wrap.b32 %r324, %r323, %r322, 1; + shf.l.wrap.b32 %r325, %r322, %r323, 1; + mov.b64 %rd718, {%r325, %r324}; + xor.b64 %rd719, %rd718, %rd703; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r326}, %rd695; + } + { + .reg .b32 %dummy; + mov.b64 {%r327,%dummy}, %rd695; + } + shf.l.wrap.b32 %r328, %r327, %r326, 1; + shf.l.wrap.b32 %r329, %r326, %r327, 1; + mov.b64 %rd720, {%r329, %r328}; + xor.b64 %rd721, %rd707, %rd720; + xor.b64 %rd722, %rd713, %rd691; + xor.b64 %rd723, %rd715, %rd621; + xor.b64 %rd724, %rd717, %rd624; + xor.b64 %rd725, %rd719, %rd627; + xor.b64 %rd726, %rd721, %rd630; + xor.b64 %rd727, %rd713, %rd633; + xor.b64 %rd728, %rd715, %rd636; + xor.b64 %rd729, %rd717, %rd639; + xor.b64 %rd730, %rd719, %rd642; + xor.b64 %rd731, %rd721, %rd645; + xor.b64 %rd732, %rd713, %rd648; + xor.b64 %rd733, %rd715, %rd651; + xor.b64 %rd734, %rd717, %rd654; + xor.b64 %rd735, %rd719, %rd657; + xor.b64 %rd736, %rd721, %rd660; + xor.b64 %rd737, %rd713, %rd663; + xor.b64 %rd738, %rd715, %rd666; + xor.b64 %rd739, %rd717, %rd669; + xor.b64 %rd740, %rd719, %rd672; + xor.b64 %rd741, %rd721, %rd675; + xor.b64 %rd742, %rd713, %rd678; + xor.b64 %rd743, %rd715, %rd681; + xor.b64 %rd744, %rd717, %rd684; + xor.b64 %rd745, %rd719, %rd687; + xor.b64 %rd746, %rd721, %rd690; + { + .reg .b32 %dummy; + mov.b64 {%r330,%dummy}, %rd728; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r331}, %rd728; + } + shf.r.wrap.b32 %r332, %r331, %r330, 20; + shf.r.wrap.b32 %r333, %r330, %r331, 20; + mov.b64 %rd747, {%r333, %r332}; + { + .reg .b32 %dummy; + mov.b64 {%r334,%dummy}, %rd734; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r335}, %rd734; + } + shf.r.wrap.b32 %r336, %r335, %r334, 21; + shf.r.wrap.b32 %r337, %r334, %r335, 21; + mov.b64 %rd748, {%r337, %r336}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r338}, %rd740; + } + { + .reg .b32 %dummy; + mov.b64 {%r339,%dummy}, %rd740; + } + shf.l.wrap.b32 %r340, %r339, %r338, 21; + shf.l.wrap.b32 %r341, %r338, %r339, 21; + mov.b64 %rd749, {%r341, %r340}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r342}, %rd746; + } + { + .reg .b32 %dummy; + mov.b64 {%r343,%dummy}, %rd746; + } + shf.l.wrap.b32 %r344, %r343, %r342, 14; + shf.l.wrap.b32 %r345, %r342, %r343, 14; + mov.b64 %rd750, {%r345, %r344}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r346}, %rd725; + } + { + .reg .b32 %dummy; + mov.b64 {%r347,%dummy}, %rd725; + } + shf.l.wrap.b32 %r348, %r347, %r346, 28; + shf.l.wrap.b32 %r349, %r346, %r347, 28; + mov.b64 %rd751, {%r349, %r348}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r350}, %rd731; + } + { + .reg .b32 %dummy; + mov.b64 {%r351,%dummy}, %rd731; + } + shf.l.wrap.b32 %r352, %r351, %r350, 20; + shf.l.wrap.b32 %r353, %r350, %r351, 20; + mov.b64 %rd752, {%r353, %r352}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r354}, %rd732; + } + { + .reg .b32 %dummy; + mov.b64 {%r355,%dummy}, %rd732; + } + shf.l.wrap.b32 %r356, %r355, %r354, 3; + shf.l.wrap.b32 %r357, %r354, %r355, 3; + mov.b64 %rd753, {%r357, %r356}; + { + .reg .b32 %dummy; + mov.b64 {%r358,%dummy}, %rd738; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r359}, %rd738; + } + shf.r.wrap.b32 %r360, %r359, %r358, 19; + shf.r.wrap.b32 %r361, %r358, %r359, 19; + mov.b64 %rd754, {%r361, %r360}; + { + .reg .b32 %dummy; + mov.b64 {%r362,%dummy}, %rd744; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r363}, %rd744; + } + shf.r.wrap.b32 %r364, %r363, %r362, 3; + shf.r.wrap.b32 %r365, %r362, %r363, 3; + mov.b64 %rd755, {%r365, %r364}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r366}, %rd723; + } + { + .reg .b32 %dummy; + mov.b64 {%r367,%dummy}, %rd723; + } + shf.l.wrap.b32 %r368, %r367, %r366, 1; + shf.l.wrap.b32 %r369, %r366, %r367, 1; + mov.b64 %rd756, {%r369, %r368}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r370}, %rd729; + } + { + .reg .b32 %dummy; + mov.b64 {%r371,%dummy}, %rd729; + } + shf.l.wrap.b32 %r372, %r371, %r370, 6; + shf.l.wrap.b32 %r373, %r370, %r371, 6; + mov.b64 %rd757, {%r373, %r372}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r374}, %rd735; + } + { + .reg .b32 %dummy; + mov.b64 {%r375,%dummy}, %rd735; + } + shf.l.wrap.b32 %r376, %r375, %r374, 25; + shf.l.wrap.b32 %r377, %r374, %r375, 25; + mov.b64 %rd758, {%r377, %r376}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r378}, %rd741; + } + { + .reg .b32 %dummy; + mov.b64 {%r379,%dummy}, %rd741; + } + shf.l.wrap.b32 %r380, %r379, %r378, 8; + shf.l.wrap.b32 %r381, %r378, %r379, 8; + mov.b64 %rd759, {%r381, %r380}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r382}, %rd742; + } + { + .reg .b32 %dummy; + mov.b64 {%r383,%dummy}, %rd742; + } + shf.l.wrap.b32 %r384, %r383, %r382, 18; + shf.l.wrap.b32 %r385, %r382, %r383, 18; + mov.b64 %rd760, {%r385, %r384}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r386}, %rd726; + } + { + .reg .b32 %dummy; + mov.b64 {%r387,%dummy}, %rd726; + } + shf.l.wrap.b32 %r388, %r387, %r386, 27; + shf.l.wrap.b32 %r389, %r386, %r387, 27; + mov.b64 %rd761, {%r389, %r388}; + { + .reg .b32 %dummy; + mov.b64 {%r390,%dummy}, %rd727; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r391}, %rd727; + } + shf.r.wrap.b32 %r392, %r391, %r390, 28; + shf.r.wrap.b32 %r393, %r390, %r391, 28; + mov.b64 %rd762, {%r393, %r392}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r394}, %rd733; + } + { + .reg .b32 %dummy; + mov.b64 {%r395,%dummy}, %rd733; + } + shf.l.wrap.b32 %r396, %r395, %r394, 10; + shf.l.wrap.b32 %r397, %r394, %r395, 10; + mov.b64 %rd763, {%r397, %r396}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r398}, %rd739; + } + { + .reg .b32 %dummy; + mov.b64 {%r399,%dummy}, %rd739; + } + shf.l.wrap.b32 %r400, %r399, %r398, 15; + shf.l.wrap.b32 %r401, %r398, %r399, 15; + mov.b64 %rd764, {%r401, %r400}; + { + .reg .b32 %dummy; + mov.b64 {%r402,%dummy}, %rd745; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r403}, %rd745; + } + shf.r.wrap.b32 %r404, %r403, %r402, 8; + shf.r.wrap.b32 %r405, %r402, %r403, 8; + mov.b64 %rd765, {%r405, %r404}; + { + .reg .b32 %dummy; + mov.b64 {%r406,%dummy}, %rd724; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r407}, %rd724; + } + shf.r.wrap.b32 %r408, %r407, %r406, 2; + shf.r.wrap.b32 %r409, %r406, %r407, 2; + mov.b64 %rd766, {%r409, %r408}; + { + .reg .b32 %dummy; + mov.b64 {%r410,%dummy}, %rd730; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r411}, %rd730; + } + shf.r.wrap.b32 %r412, %r411, %r410, 9; + shf.r.wrap.b32 %r413, %r410, %r411, 9; + mov.b64 %rd767, {%r413, %r412}; + { + .reg .b32 %dummy; + mov.b64 {%r414,%dummy}, %rd736; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r415}, %rd736; + } + shf.r.wrap.b32 %r416, %r415, %r414, 25; + shf.r.wrap.b32 %r417, %r414, %r415, 25; + mov.b64 %rd768, {%r417, %r416}; + { + .reg .b32 %dummy; + mov.b64 {%r418,%dummy}, %rd737; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r419}, %rd737; + } + shf.r.wrap.b32 %r420, %r419, %r418, 23; + shf.r.wrap.b32 %r421, %r418, %r419, 23; + mov.b64 %rd769, {%r421, %r420}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r422}, %rd743; + } + { + .reg .b32 %dummy; + mov.b64 {%r423,%dummy}, %rd743; + } + shf.l.wrap.b32 %r424, %r423, %r422, 2; + shf.l.wrap.b32 %r425, %r422, %r423, 2; + mov.b64 %rd770, {%r425, %r424}; + not.b64 %rd771, %rd747; + and.b64 %rd772, %rd748, %rd771; + xor.b64 %rd773, %rd772, %rd722; + not.b64 %rd774, %rd748; + and.b64 %rd775, %rd749, %rd774; + xor.b64 %rd776, %rd775, %rd747; + not.b64 %rd777, %rd749; + and.b64 %rd778, %rd750, %rd777; + xor.b64 %rd779, %rd778, %rd748; + not.b64 %rd780, %rd750; + and.b64 %rd781, %rd722, %rd780; + xor.b64 %rd782, %rd781, %rd749; + not.b64 %rd783, %rd722; + and.b64 %rd784, %rd747, %rd783; + xor.b64 %rd785, %rd784, %rd750; + not.b64 %rd786, %rd752; + and.b64 %rd787, %rd753, %rd786; + xor.b64 %rd788, %rd787, %rd751; + not.b64 %rd789, %rd753; + and.b64 %rd790, %rd754, %rd789; + xor.b64 %rd791, %rd790, %rd752; + not.b64 %rd792, %rd754; + and.b64 %rd793, %rd755, %rd792; + xor.b64 %rd794, %rd793, %rd753; + not.b64 %rd795, %rd755; + and.b64 %rd796, %rd751, %rd795; + xor.b64 %rd797, %rd796, %rd754; + not.b64 %rd798, %rd751; + and.b64 %rd799, %rd752, %rd798; + xor.b64 %rd800, %rd799, %rd755; + not.b64 %rd801, %rd757; + and.b64 %rd802, %rd758, %rd801; + xor.b64 %rd803, %rd802, %rd756; + not.b64 %rd804, %rd758; + and.b64 %rd805, %rd759, %rd804; + xor.b64 %rd806, %rd805, %rd757; + not.b64 %rd807, %rd759; + and.b64 %rd808, %rd760, %rd807; + xor.b64 %rd809, %rd808, %rd758; + not.b64 %rd810, %rd760; + and.b64 %rd811, %rd756, %rd810; + xor.b64 %rd812, %rd811, %rd759; + not.b64 %rd813, %rd756; + and.b64 %rd814, %rd757, %rd813; + xor.b64 %rd815, %rd814, %rd760; + not.b64 %rd816, %rd762; + and.b64 %rd817, %rd763, %rd816; + xor.b64 %rd818, %rd817, %rd761; + not.b64 %rd819, %rd763; + and.b64 %rd820, %rd764, %rd819; + xor.b64 %rd821, %rd820, %rd762; + not.b64 %rd822, %rd764; + and.b64 %rd823, %rd765, %rd822; + xor.b64 %rd824, %rd823, %rd763; + not.b64 %rd825, %rd765; + and.b64 %rd826, %rd761, %rd825; + xor.b64 %rd827, %rd826, %rd764; + not.b64 %rd828, %rd761; + and.b64 %rd829, %rd762, %rd828; + xor.b64 %rd830, %rd829, %rd765; + not.b64 %rd831, %rd767; + and.b64 %rd832, %rd768, %rd831; + xor.b64 %rd833, %rd832, %rd766; + not.b64 %rd834, %rd768; + and.b64 %rd835, %rd769, %rd834; + xor.b64 %rd836, %rd835, %rd767; + not.b64 %rd837, %rd769; + and.b64 %rd838, %rd770, %rd837; + xor.b64 %rd839, %rd838, %rd768; + not.b64 %rd840, %rd770; + and.b64 %rd841, %rd766, %rd840; + xor.b64 %rd842, %rd841, %rd769; + not.b64 %rd843, %rd766; + and.b64 %rd844, %rd767, %rd843; + xor.b64 %rd845, %rd844, %rd770; + xor.b64 %rd846, %rd773, %rd18; + xor.b64 %rd847, %rd788, %rd846; + xor.b64 %rd848, %rd847, %rd803; + xor.b64 %rd849, %rd848, %rd818; + xor.b64 %rd850, %rd849, %rd833; + xor.b64 %rd851, %rd791, %rd776; + xor.b64 %rd852, %rd851, %rd806; + xor.b64 %rd853, %rd852, %rd821; + xor.b64 %rd854, %rd853, %rd836; + xor.b64 %rd855, %rd794, %rd779; + xor.b64 %rd856, %rd855, %rd809; + xor.b64 %rd857, %rd856, %rd824; + xor.b64 %rd858, %rd857, %rd839; + xor.b64 %rd859, %rd797, %rd782; + xor.b64 %rd860, %rd859, %rd812; + xor.b64 %rd861, %rd860, %rd827; + xor.b64 %rd862, %rd861, %rd842; + xor.b64 %rd863, %rd800, %rd785; + xor.b64 %rd864, %rd863, %rd815; + xor.b64 %rd865, %rd864, %rd830; + xor.b64 %rd866, %rd865, %rd845; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r426}, %rd854; + } + { + .reg .b32 %dummy; + mov.b64 {%r427,%dummy}, %rd854; + } + shf.l.wrap.b32 %r428, %r427, %r426, 1; + shf.l.wrap.b32 %r429, %r426, %r427, 1; + mov.b64 %rd867, {%r429, %r428}; + xor.b64 %rd868, %rd866, %rd867; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r430}, %rd858; + } + { + .reg .b32 %dummy; + mov.b64 {%r431,%dummy}, %rd858; + } + shf.l.wrap.b32 %r432, %r431, %r430, 1; + shf.l.wrap.b32 %r433, %r430, %r431, 1; + mov.b64 %rd869, {%r433, %r432}; + xor.b64 %rd870, %rd869, %rd850; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r434}, %rd862; + } + { + .reg .b32 %dummy; + mov.b64 {%r435,%dummy}, %rd862; + } + shf.l.wrap.b32 %r436, %r435, %r434, 1; + shf.l.wrap.b32 %r437, %r434, %r435, 1; + mov.b64 %rd871, {%r437, %r436}; + xor.b64 %rd872, %rd871, %rd854; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r438}, %rd866; + } + { + .reg .b32 %dummy; + mov.b64 {%r439,%dummy}, %rd866; + } + shf.l.wrap.b32 %r440, %r439, %r438, 1; + shf.l.wrap.b32 %r441, %r438, %r439, 1; + mov.b64 %rd873, {%r441, %r440}; + xor.b64 %rd874, %rd873, %rd858; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r442}, %rd850; + } + { + .reg .b32 %dummy; + mov.b64 {%r443,%dummy}, %rd850; + } + shf.l.wrap.b32 %r444, %r443, %r442, 1; + shf.l.wrap.b32 %r445, %r442, %r443, 1; + mov.b64 %rd875, {%r445, %r444}; + xor.b64 %rd876, %rd862, %rd875; + xor.b64 %rd877, %rd868, %rd846; + xor.b64 %rd878, %rd870, %rd776; + xor.b64 %rd879, %rd872, %rd779; + xor.b64 %rd880, %rd874, %rd782; + xor.b64 %rd881, %rd876, %rd785; + xor.b64 %rd882, %rd868, %rd788; + xor.b64 %rd883, %rd870, %rd791; + xor.b64 %rd884, %rd872, %rd794; + xor.b64 %rd885, %rd874, %rd797; + xor.b64 %rd886, %rd876, %rd800; + xor.b64 %rd887, %rd868, %rd803; + xor.b64 %rd888, %rd870, %rd806; + xor.b64 %rd889, %rd872, %rd809; + xor.b64 %rd890, %rd874, %rd812; + xor.b64 %rd891, %rd876, %rd815; + xor.b64 %rd892, %rd868, %rd818; + xor.b64 %rd893, %rd870, %rd821; + xor.b64 %rd894, %rd872, %rd824; + xor.b64 %rd895, %rd874, %rd827; + xor.b64 %rd896, %rd876, %rd830; + xor.b64 %rd897, %rd868, %rd833; + xor.b64 %rd898, %rd870, %rd836; + xor.b64 %rd899, %rd872, %rd839; + xor.b64 %rd900, %rd874, %rd842; + xor.b64 %rd901, %rd876, %rd845; + { + .reg .b32 %dummy; + mov.b64 {%r446,%dummy}, %rd883; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r447}, %rd883; + } + shf.r.wrap.b32 %r448, %r447, %r446, 20; + shf.r.wrap.b32 %r449, %r446, %r447, 20; + mov.b64 %rd902, {%r449, %r448}; + { + .reg .b32 %dummy; + mov.b64 {%r450,%dummy}, %rd889; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r451}, %rd889; + } + shf.r.wrap.b32 %r452, %r451, %r450, 21; + shf.r.wrap.b32 %r453, %r450, %r451, 21; + mov.b64 %rd903, {%r453, %r452}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r454}, %rd895; + } + { + .reg .b32 %dummy; + mov.b64 {%r455,%dummy}, %rd895; + } + shf.l.wrap.b32 %r456, %r455, %r454, 21; + shf.l.wrap.b32 %r457, %r454, %r455, 21; + mov.b64 %rd904, {%r457, %r456}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r458}, %rd901; + } + { + .reg .b32 %dummy; + mov.b64 {%r459,%dummy}, %rd901; + } + shf.l.wrap.b32 %r460, %r459, %r458, 14; + shf.l.wrap.b32 %r461, %r458, %r459, 14; + mov.b64 %rd905, {%r461, %r460}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r462}, %rd880; + } + { + .reg .b32 %dummy; + mov.b64 {%r463,%dummy}, %rd880; + } + shf.l.wrap.b32 %r464, %r463, %r462, 28; + shf.l.wrap.b32 %r465, %r462, %r463, 28; + mov.b64 %rd906, {%r465, %r464}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r466}, %rd886; + } + { + .reg .b32 %dummy; + mov.b64 {%r467,%dummy}, %rd886; + } + shf.l.wrap.b32 %r468, %r467, %r466, 20; + shf.l.wrap.b32 %r469, %r466, %r467, 20; + mov.b64 %rd907, {%r469, %r468}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r470}, %rd887; + } + { + .reg .b32 %dummy; + mov.b64 {%r471,%dummy}, %rd887; + } + shf.l.wrap.b32 %r472, %r471, %r470, 3; + shf.l.wrap.b32 %r473, %r470, %r471, 3; + mov.b64 %rd908, {%r473, %r472}; + { + .reg .b32 %dummy; + mov.b64 {%r474,%dummy}, %rd893; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r475}, %rd893; + } + shf.r.wrap.b32 %r476, %r475, %r474, 19; + shf.r.wrap.b32 %r477, %r474, %r475, 19; + mov.b64 %rd909, {%r477, %r476}; + { + .reg .b32 %dummy; + mov.b64 {%r478,%dummy}, %rd899; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r479}, %rd899; + } + shf.r.wrap.b32 %r480, %r479, %r478, 3; + shf.r.wrap.b32 %r481, %r478, %r479, 3; + mov.b64 %rd910, {%r481, %r480}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r482}, %rd878; + } + { + .reg .b32 %dummy; + mov.b64 {%r483,%dummy}, %rd878; + } + shf.l.wrap.b32 %r484, %r483, %r482, 1; + shf.l.wrap.b32 %r485, %r482, %r483, 1; + mov.b64 %rd911, {%r485, %r484}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r486}, %rd884; + } + { + .reg .b32 %dummy; + mov.b64 {%r487,%dummy}, %rd884; + } + shf.l.wrap.b32 %r488, %r487, %r486, 6; + shf.l.wrap.b32 %r489, %r486, %r487, 6; + mov.b64 %rd912, {%r489, %r488}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r490}, %rd890; + } + { + .reg .b32 %dummy; + mov.b64 {%r491,%dummy}, %rd890; + } + shf.l.wrap.b32 %r492, %r491, %r490, 25; + shf.l.wrap.b32 %r493, %r490, %r491, 25; + mov.b64 %rd913, {%r493, %r492}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r494}, %rd896; + } + { + .reg .b32 %dummy; + mov.b64 {%r495,%dummy}, %rd896; + } + shf.l.wrap.b32 %r496, %r495, %r494, 8; + shf.l.wrap.b32 %r497, %r494, %r495, 8; + mov.b64 %rd914, {%r497, %r496}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r498}, %rd897; + } + { + .reg .b32 %dummy; + mov.b64 {%r499,%dummy}, %rd897; + } + shf.l.wrap.b32 %r500, %r499, %r498, 18; + shf.l.wrap.b32 %r501, %r498, %r499, 18; + mov.b64 %rd915, {%r501, %r500}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r502}, %rd881; + } + { + .reg .b32 %dummy; + mov.b64 {%r503,%dummy}, %rd881; + } + shf.l.wrap.b32 %r504, %r503, %r502, 27; + shf.l.wrap.b32 %r505, %r502, %r503, 27; + mov.b64 %rd916, {%r505, %r504}; + { + .reg .b32 %dummy; + mov.b64 {%r506,%dummy}, %rd882; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r507}, %rd882; + } + shf.r.wrap.b32 %r508, %r507, %r506, 28; + shf.r.wrap.b32 %r509, %r506, %r507, 28; + mov.b64 %rd917, {%r509, %r508}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r510}, %rd888; + } + { + .reg .b32 %dummy; + mov.b64 {%r511,%dummy}, %rd888; + } + shf.l.wrap.b32 %r512, %r511, %r510, 10; + shf.l.wrap.b32 %r513, %r510, %r511, 10; + mov.b64 %rd918, {%r513, %r512}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r514}, %rd894; + } + { + .reg .b32 %dummy; + mov.b64 {%r515,%dummy}, %rd894; + } + shf.l.wrap.b32 %r516, %r515, %r514, 15; + shf.l.wrap.b32 %r517, %r514, %r515, 15; + mov.b64 %rd919, {%r517, %r516}; + { + .reg .b32 %dummy; + mov.b64 {%r518,%dummy}, %rd900; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r519}, %rd900; + } + shf.r.wrap.b32 %r520, %r519, %r518, 8; + shf.r.wrap.b32 %r521, %r518, %r519, 8; + mov.b64 %rd920, {%r521, %r520}; + { + .reg .b32 %dummy; + mov.b64 {%r522,%dummy}, %rd879; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r523}, %rd879; + } + shf.r.wrap.b32 %r524, %r523, %r522, 2; + shf.r.wrap.b32 %r525, %r522, %r523, 2; + mov.b64 %rd921, {%r525, %r524}; + { + .reg .b32 %dummy; + mov.b64 {%r526,%dummy}, %rd885; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r527}, %rd885; + } + shf.r.wrap.b32 %r528, %r527, %r526, 9; + shf.r.wrap.b32 %r529, %r526, %r527, 9; + mov.b64 %rd922, {%r529, %r528}; + { + .reg .b32 %dummy; + mov.b64 {%r530,%dummy}, %rd891; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r531}, %rd891; + } + shf.r.wrap.b32 %r532, %r531, %r530, 25; + shf.r.wrap.b32 %r533, %r530, %r531, 25; + mov.b64 %rd923, {%r533, %r532}; + { + .reg .b32 %dummy; + mov.b64 {%r534,%dummy}, %rd892; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r535}, %rd892; + } + shf.r.wrap.b32 %r536, %r535, %r534, 23; + shf.r.wrap.b32 %r537, %r534, %r535, 23; + mov.b64 %rd924, {%r537, %r536}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r538}, %rd898; + } + { + .reg .b32 %dummy; + mov.b64 {%r539,%dummy}, %rd898; + } + shf.l.wrap.b32 %r540, %r539, %r538, 2; + shf.l.wrap.b32 %r541, %r538, %r539, 2; + mov.b64 %rd925, {%r541, %r540}; + not.b64 %rd926, %rd902; + and.b64 %rd927, %rd903, %rd926; + xor.b64 %rd928, %rd927, %rd877; + not.b64 %rd929, %rd903; + and.b64 %rd930, %rd904, %rd929; + xor.b64 %rd931, %rd930, %rd902; + not.b64 %rd932, %rd904; + and.b64 %rd933, %rd905, %rd932; + xor.b64 %rd934, %rd933, %rd903; + not.b64 %rd935, %rd905; + and.b64 %rd936, %rd877, %rd935; + xor.b64 %rd937, %rd936, %rd904; + not.b64 %rd938, %rd877; + and.b64 %rd939, %rd902, %rd938; + xor.b64 %rd940, %rd939, %rd905; + not.b64 %rd941, %rd907; + and.b64 %rd942, %rd908, %rd941; + xor.b64 %rd943, %rd942, %rd906; + not.b64 %rd944, %rd908; + and.b64 %rd945, %rd909, %rd944; + xor.b64 %rd946, %rd945, %rd907; + not.b64 %rd947, %rd909; + and.b64 %rd948, %rd910, %rd947; + xor.b64 %rd949, %rd948, %rd908; + not.b64 %rd950, %rd910; + and.b64 %rd951, %rd906, %rd950; + xor.b64 %rd952, %rd951, %rd909; + not.b64 %rd953, %rd906; + and.b64 %rd954, %rd907, %rd953; + xor.b64 %rd955, %rd954, %rd910; + not.b64 %rd956, %rd912; + and.b64 %rd957, %rd913, %rd956; + xor.b64 %rd958, %rd957, %rd911; + not.b64 %rd959, %rd913; + and.b64 %rd960, %rd914, %rd959; + xor.b64 %rd961, %rd960, %rd912; + not.b64 %rd962, %rd914; + and.b64 %rd963, %rd915, %rd962; + xor.b64 %rd964, %rd963, %rd913; + not.b64 %rd965, %rd915; + and.b64 %rd966, %rd911, %rd965; + xor.b64 %rd967, %rd966, %rd914; + not.b64 %rd968, %rd911; + and.b64 %rd969, %rd912, %rd968; + xor.b64 %rd970, %rd969, %rd915; + not.b64 %rd971, %rd917; + and.b64 %rd972, %rd918, %rd971; + xor.b64 %rd973, %rd972, %rd916; + not.b64 %rd974, %rd918; + and.b64 %rd975, %rd919, %rd974; + xor.b64 %rd976, %rd975, %rd917; + not.b64 %rd977, %rd919; + and.b64 %rd978, %rd920, %rd977; + xor.b64 %rd979, %rd978, %rd918; + not.b64 %rd980, %rd920; + and.b64 %rd981, %rd916, %rd980; + xor.b64 %rd982, %rd981, %rd919; + not.b64 %rd983, %rd916; + and.b64 %rd984, %rd917, %rd983; + xor.b64 %rd985, %rd984, %rd920; + not.b64 %rd986, %rd922; + and.b64 %rd987, %rd923, %rd986; + xor.b64 %rd988, %rd987, %rd921; + not.b64 %rd989, %rd923; + and.b64 %rd990, %rd924, %rd989; + xor.b64 %rd991, %rd990, %rd922; + not.b64 %rd992, %rd924; + and.b64 %rd993, %rd925, %rd992; + xor.b64 %rd994, %rd993, %rd923; + not.b64 %rd995, %rd925; + and.b64 %rd996, %rd921, %rd995; + xor.b64 %rd997, %rd996, %rd924; + not.b64 %rd998, %rd921; + and.b64 %rd999, %rd922, %rd998; + xor.b64 %rd1000, %rd999, %rd925; + xor.b64 %rd1001, %rd928, %rd19; + xor.b64 %rd1002, %rd943, %rd1001; + xor.b64 %rd1003, %rd1002, %rd958; + xor.b64 %rd1004, %rd1003, %rd973; + xor.b64 %rd1005, %rd1004, %rd988; + xor.b64 %rd1006, %rd946, %rd931; + xor.b64 %rd1007, %rd1006, %rd961; + xor.b64 %rd1008, %rd1007, %rd976; + xor.b64 %rd1009, %rd1008, %rd991; + xor.b64 %rd1010, %rd949, %rd934; + xor.b64 %rd1011, %rd1010, %rd964; + xor.b64 %rd1012, %rd1011, %rd979; + xor.b64 %rd1013, %rd1012, %rd994; + xor.b64 %rd1014, %rd952, %rd937; + xor.b64 %rd1015, %rd1014, %rd967; + xor.b64 %rd1016, %rd1015, %rd982; + xor.b64 %rd1017, %rd1016, %rd997; + xor.b64 %rd1018, %rd955, %rd940; + xor.b64 %rd1019, %rd1018, %rd970; + xor.b64 %rd1020, %rd1019, %rd985; + xor.b64 %rd1021, %rd1020, %rd1000; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r542}, %rd1009; + } + { + .reg .b32 %dummy; + mov.b64 {%r543,%dummy}, %rd1009; + } + shf.l.wrap.b32 %r544, %r543, %r542, 1; + shf.l.wrap.b32 %r545, %r542, %r543, 1; + mov.b64 %rd1022, {%r545, %r544}; + xor.b64 %rd1023, %rd1021, %rd1022; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r546}, %rd1013; + } + { + .reg .b32 %dummy; + mov.b64 {%r547,%dummy}, %rd1013; + } + shf.l.wrap.b32 %r548, %r547, %r546, 1; + shf.l.wrap.b32 %r549, %r546, %r547, 1; + mov.b64 %rd1024, {%r549, %r548}; + xor.b64 %rd1025, %rd1024, %rd1005; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r550}, %rd1017; + } + { + .reg .b32 %dummy; + mov.b64 {%r551,%dummy}, %rd1017; + } + shf.l.wrap.b32 %r552, %r551, %r550, 1; + shf.l.wrap.b32 %r553, %r550, %r551, 1; + mov.b64 %rd1026, {%r553, %r552}; + xor.b64 %rd1027, %rd1026, %rd1009; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r554}, %rd1021; + } + { + .reg .b32 %dummy; + mov.b64 {%r555,%dummy}, %rd1021; + } + shf.l.wrap.b32 %r556, %r555, %r554, 1; + shf.l.wrap.b32 %r557, %r554, %r555, 1; + mov.b64 %rd1028, {%r557, %r556}; + xor.b64 %rd1029, %rd1028, %rd1013; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r558}, %rd1005; + } + { + .reg .b32 %dummy; + mov.b64 {%r559,%dummy}, %rd1005; + } + shf.l.wrap.b32 %r560, %r559, %r558, 1; + shf.l.wrap.b32 %r561, %r558, %r559, 1; + mov.b64 %rd1030, {%r561, %r560}; + xor.b64 %rd1031, %rd1017, %rd1030; + xor.b64 %rd1032, %rd1023, %rd1001; + xor.b64 %rd1033, %rd1025, %rd931; + xor.b64 %rd1034, %rd1027, %rd934; + xor.b64 %rd1035, %rd1029, %rd937; + xor.b64 %rd1036, %rd1031, %rd940; + xor.b64 %rd1037, %rd1023, %rd943; + xor.b64 %rd1038, %rd1025, %rd946; + xor.b64 %rd1039, %rd1027, %rd949; + xor.b64 %rd1040, %rd1029, %rd952; + xor.b64 %rd1041, %rd1031, %rd955; + xor.b64 %rd1042, %rd1023, %rd958; + xor.b64 %rd1043, %rd1025, %rd961; + xor.b64 %rd1044, %rd1027, %rd964; + xor.b64 %rd1045, %rd1029, %rd967; + xor.b64 %rd1046, %rd1031, %rd970; + xor.b64 %rd1047, %rd1023, %rd973; + xor.b64 %rd1048, %rd1025, %rd976; + xor.b64 %rd1049, %rd1027, %rd979; + xor.b64 %rd1050, %rd1029, %rd982; + xor.b64 %rd1051, %rd1031, %rd985; + xor.b64 %rd1052, %rd1023, %rd988; + xor.b64 %rd1053, %rd1025, %rd991; + xor.b64 %rd1054, %rd1027, %rd994; + xor.b64 %rd1055, %rd1029, %rd997; + xor.b64 %rd1056, %rd1031, %rd1000; + { + .reg .b32 %dummy; + mov.b64 {%r562,%dummy}, %rd1038; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r563}, %rd1038; + } + shf.r.wrap.b32 %r564, %r563, %r562, 20; + shf.r.wrap.b32 %r565, %r562, %r563, 20; + mov.b64 %rd1057, {%r565, %r564}; + { + .reg .b32 %dummy; + mov.b64 {%r566,%dummy}, %rd1044; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r567}, %rd1044; + } + shf.r.wrap.b32 %r568, %r567, %r566, 21; + shf.r.wrap.b32 %r569, %r566, %r567, 21; + mov.b64 %rd1058, {%r569, %r568}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r570}, %rd1050; + } + { + .reg .b32 %dummy; + mov.b64 {%r571,%dummy}, %rd1050; + } + shf.l.wrap.b32 %r572, %r571, %r570, 21; + shf.l.wrap.b32 %r573, %r570, %r571, 21; + mov.b64 %rd1059, {%r573, %r572}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r574}, %rd1056; + } + { + .reg .b32 %dummy; + mov.b64 {%r575,%dummy}, %rd1056; + } + shf.l.wrap.b32 %r576, %r575, %r574, 14; + shf.l.wrap.b32 %r577, %r574, %r575, 14; + mov.b64 %rd1060, {%r577, %r576}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r578}, %rd1035; + } + { + .reg .b32 %dummy; + mov.b64 {%r579,%dummy}, %rd1035; + } + shf.l.wrap.b32 %r580, %r579, %r578, 28; + shf.l.wrap.b32 %r581, %r578, %r579, 28; + mov.b64 %rd1061, {%r581, %r580}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r582}, %rd1041; + } + { + .reg .b32 %dummy; + mov.b64 {%r583,%dummy}, %rd1041; + } + shf.l.wrap.b32 %r584, %r583, %r582, 20; + shf.l.wrap.b32 %r585, %r582, %r583, 20; + mov.b64 %rd1062, {%r585, %r584}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r586}, %rd1042; + } + { + .reg .b32 %dummy; + mov.b64 {%r587,%dummy}, %rd1042; + } + shf.l.wrap.b32 %r588, %r587, %r586, 3; + shf.l.wrap.b32 %r589, %r586, %r587, 3; + mov.b64 %rd1063, {%r589, %r588}; + { + .reg .b32 %dummy; + mov.b64 {%r590,%dummy}, %rd1048; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r591}, %rd1048; + } + shf.r.wrap.b32 %r592, %r591, %r590, 19; + shf.r.wrap.b32 %r593, %r590, %r591, 19; + mov.b64 %rd1064, {%r593, %r592}; + { + .reg .b32 %dummy; + mov.b64 {%r594,%dummy}, %rd1054; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r595}, %rd1054; + } + shf.r.wrap.b32 %r596, %r595, %r594, 3; + shf.r.wrap.b32 %r597, %r594, %r595, 3; + mov.b64 %rd1065, {%r597, %r596}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r598}, %rd1033; + } + { + .reg .b32 %dummy; + mov.b64 {%r599,%dummy}, %rd1033; + } + shf.l.wrap.b32 %r600, %r599, %r598, 1; + shf.l.wrap.b32 %r601, %r598, %r599, 1; + mov.b64 %rd1066, {%r601, %r600}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r602}, %rd1039; + } + { + .reg .b32 %dummy; + mov.b64 {%r603,%dummy}, %rd1039; + } + shf.l.wrap.b32 %r604, %r603, %r602, 6; + shf.l.wrap.b32 %r605, %r602, %r603, 6; + mov.b64 %rd1067, {%r605, %r604}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r606}, %rd1045; + } + { + .reg .b32 %dummy; + mov.b64 {%r607,%dummy}, %rd1045; + } + shf.l.wrap.b32 %r608, %r607, %r606, 25; + shf.l.wrap.b32 %r609, %r606, %r607, 25; + mov.b64 %rd1068, {%r609, %r608}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r610}, %rd1051; + } + { + .reg .b32 %dummy; + mov.b64 {%r611,%dummy}, %rd1051; + } + shf.l.wrap.b32 %r612, %r611, %r610, 8; + shf.l.wrap.b32 %r613, %r610, %r611, 8; + mov.b64 %rd1069, {%r613, %r612}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r614}, %rd1052; + } + { + .reg .b32 %dummy; + mov.b64 {%r615,%dummy}, %rd1052; + } + shf.l.wrap.b32 %r616, %r615, %r614, 18; + shf.l.wrap.b32 %r617, %r614, %r615, 18; + mov.b64 %rd1070, {%r617, %r616}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r618}, %rd1036; + } + { + .reg .b32 %dummy; + mov.b64 {%r619,%dummy}, %rd1036; + } + shf.l.wrap.b32 %r620, %r619, %r618, 27; + shf.l.wrap.b32 %r621, %r618, %r619, 27; + mov.b64 %rd1071, {%r621, %r620}; + { + .reg .b32 %dummy; + mov.b64 {%r622,%dummy}, %rd1037; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r623}, %rd1037; + } + shf.r.wrap.b32 %r624, %r623, %r622, 28; + shf.r.wrap.b32 %r625, %r622, %r623, 28; + mov.b64 %rd1072, {%r625, %r624}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r626}, %rd1043; + } + { + .reg .b32 %dummy; + mov.b64 {%r627,%dummy}, %rd1043; + } + shf.l.wrap.b32 %r628, %r627, %r626, 10; + shf.l.wrap.b32 %r629, %r626, %r627, 10; + mov.b64 %rd1073, {%r629, %r628}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r630}, %rd1049; + } + { + .reg .b32 %dummy; + mov.b64 {%r631,%dummy}, %rd1049; + } + shf.l.wrap.b32 %r632, %r631, %r630, 15; + shf.l.wrap.b32 %r633, %r630, %r631, 15; + mov.b64 %rd1074, {%r633, %r632}; + { + .reg .b32 %dummy; + mov.b64 {%r634,%dummy}, %rd1055; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r635}, %rd1055; + } + shf.r.wrap.b32 %r636, %r635, %r634, 8; + shf.r.wrap.b32 %r637, %r634, %r635, 8; + mov.b64 %rd1075, {%r637, %r636}; + { + .reg .b32 %dummy; + mov.b64 {%r638,%dummy}, %rd1034; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r639}, %rd1034; + } + shf.r.wrap.b32 %r640, %r639, %r638, 2; + shf.r.wrap.b32 %r641, %r638, %r639, 2; + mov.b64 %rd1076, {%r641, %r640}; + { + .reg .b32 %dummy; + mov.b64 {%r642,%dummy}, %rd1040; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r643}, %rd1040; + } + shf.r.wrap.b32 %r644, %r643, %r642, 9; + shf.r.wrap.b32 %r645, %r642, %r643, 9; + mov.b64 %rd1077, {%r645, %r644}; + { + .reg .b32 %dummy; + mov.b64 {%r646,%dummy}, %rd1046; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r647}, %rd1046; + } + shf.r.wrap.b32 %r648, %r647, %r646, 25; + shf.r.wrap.b32 %r649, %r646, %r647, 25; + mov.b64 %rd1078, {%r649, %r648}; + { + .reg .b32 %dummy; + mov.b64 {%r650,%dummy}, %rd1047; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r651}, %rd1047; + } + shf.r.wrap.b32 %r652, %r651, %r650, 23; + shf.r.wrap.b32 %r653, %r650, %r651, 23; + mov.b64 %rd1079, {%r653, %r652}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r654}, %rd1053; + } + { + .reg .b32 %dummy; + mov.b64 {%r655,%dummy}, %rd1053; + } + shf.l.wrap.b32 %r656, %r655, %r654, 2; + shf.l.wrap.b32 %r657, %r654, %r655, 2; + mov.b64 %rd1080, {%r657, %r656}; + not.b64 %rd1081, %rd1057; + and.b64 %rd1082, %rd1058, %rd1081; + xor.b64 %rd1083, %rd1082, %rd1032; + not.b64 %rd1084, %rd1058; + and.b64 %rd1085, %rd1059, %rd1084; + xor.b64 %rd1086, %rd1085, %rd1057; + not.b64 %rd1087, %rd1059; + and.b64 %rd1088, %rd1060, %rd1087; + xor.b64 %rd1089, %rd1088, %rd1058; + not.b64 %rd1090, %rd1060; + and.b64 %rd1091, %rd1032, %rd1090; + xor.b64 %rd1092, %rd1091, %rd1059; + not.b64 %rd1093, %rd1032; + and.b64 %rd1094, %rd1057, %rd1093; + xor.b64 %rd1095, %rd1094, %rd1060; + not.b64 %rd1096, %rd1062; + and.b64 %rd1097, %rd1063, %rd1096; + xor.b64 %rd1098, %rd1097, %rd1061; + not.b64 %rd1099, %rd1063; + and.b64 %rd1100, %rd1064, %rd1099; + xor.b64 %rd1101, %rd1100, %rd1062; + not.b64 %rd1102, %rd1064; + and.b64 %rd1103, %rd1065, %rd1102; + xor.b64 %rd1104, %rd1103, %rd1063; + not.b64 %rd1105, %rd1065; + and.b64 %rd1106, %rd1061, %rd1105; + xor.b64 %rd1107, %rd1106, %rd1064; + not.b64 %rd1108, %rd1061; + and.b64 %rd1109, %rd1062, %rd1108; + xor.b64 %rd1110, %rd1109, %rd1065; + not.b64 %rd1111, %rd1067; + and.b64 %rd1112, %rd1068, %rd1111; + xor.b64 %rd1113, %rd1112, %rd1066; + not.b64 %rd1114, %rd1068; + and.b64 %rd1115, %rd1069, %rd1114; + xor.b64 %rd1116, %rd1115, %rd1067; + not.b64 %rd1117, %rd1069; + and.b64 %rd1118, %rd1070, %rd1117; + xor.b64 %rd1119, %rd1118, %rd1068; + not.b64 %rd1120, %rd1070; + and.b64 %rd1121, %rd1066, %rd1120; + xor.b64 %rd1122, %rd1121, %rd1069; + not.b64 %rd1123, %rd1066; + and.b64 %rd1124, %rd1067, %rd1123; + xor.b64 %rd1125, %rd1124, %rd1070; + not.b64 %rd1126, %rd1072; + and.b64 %rd1127, %rd1073, %rd1126; + xor.b64 %rd1128, %rd1127, %rd1071; + not.b64 %rd1129, %rd1073; + and.b64 %rd1130, %rd1074, %rd1129; + xor.b64 %rd1131, %rd1130, %rd1072; + not.b64 %rd1132, %rd1074; + and.b64 %rd1133, %rd1075, %rd1132; + xor.b64 %rd1134, %rd1133, %rd1073; + not.b64 %rd1135, %rd1075; + and.b64 %rd1136, %rd1071, %rd1135; + xor.b64 %rd1137, %rd1136, %rd1074; + not.b64 %rd1138, %rd1071; + and.b64 %rd1139, %rd1072, %rd1138; + xor.b64 %rd1140, %rd1139, %rd1075; + not.b64 %rd1141, %rd1077; + and.b64 %rd1142, %rd1078, %rd1141; + xor.b64 %rd1143, %rd1142, %rd1076; + not.b64 %rd1144, %rd1078; + and.b64 %rd1145, %rd1079, %rd1144; + xor.b64 %rd1146, %rd1145, %rd1077; + not.b64 %rd1147, %rd1079; + and.b64 %rd1148, %rd1080, %rd1147; + xor.b64 %rd1149, %rd1148, %rd1078; + not.b64 %rd1150, %rd1080; + and.b64 %rd1151, %rd1076, %rd1150; + xor.b64 %rd1152, %rd1151, %rd1079; + not.b64 %rd1153, %rd1076; + and.b64 %rd1154, %rd1077, %rd1153; + xor.b64 %rd1155, %rd1154, %rd1080; + xor.b64 %rd1156, %rd1083, %rd20; + xor.b64 %rd1157, %rd1098, %rd1156; + xor.b64 %rd1158, %rd1157, %rd1113; + xor.b64 %rd1159, %rd1158, %rd1128; + xor.b64 %rd1160, %rd1159, %rd1143; + xor.b64 %rd1161, %rd1101, %rd1086; + xor.b64 %rd1162, %rd1161, %rd1116; + xor.b64 %rd1163, %rd1162, %rd1131; + xor.b64 %rd1164, %rd1163, %rd1146; + xor.b64 %rd1165, %rd1104, %rd1089; + xor.b64 %rd1166, %rd1165, %rd1119; + xor.b64 %rd1167, %rd1166, %rd1134; + xor.b64 %rd1168, %rd1167, %rd1149; + xor.b64 %rd1169, %rd1107, %rd1092; + xor.b64 %rd1170, %rd1169, %rd1122; + xor.b64 %rd1171, %rd1170, %rd1137; + xor.b64 %rd1172, %rd1171, %rd1152; + xor.b64 %rd1173, %rd1110, %rd1095; + xor.b64 %rd1174, %rd1173, %rd1125; + xor.b64 %rd1175, %rd1174, %rd1140; + xor.b64 %rd1176, %rd1175, %rd1155; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r658}, %rd1164; + } + { + .reg .b32 %dummy; + mov.b64 {%r659,%dummy}, %rd1164; + } + shf.l.wrap.b32 %r660, %r659, %r658, 1; + shf.l.wrap.b32 %r661, %r658, %r659, 1; + mov.b64 %rd1177, {%r661, %r660}; + xor.b64 %rd1178, %rd1176, %rd1177; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r662}, %rd1168; + } + { + .reg .b32 %dummy; + mov.b64 {%r663,%dummy}, %rd1168; + } + shf.l.wrap.b32 %r664, %r663, %r662, 1; + shf.l.wrap.b32 %r665, %r662, %r663, 1; + mov.b64 %rd1179, {%r665, %r664}; + xor.b64 %rd1180, %rd1179, %rd1160; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r666}, %rd1172; + } + { + .reg .b32 %dummy; + mov.b64 {%r667,%dummy}, %rd1172; + } + shf.l.wrap.b32 %r668, %r667, %r666, 1; + shf.l.wrap.b32 %r669, %r666, %r667, 1; + mov.b64 %rd1181, {%r669, %r668}; + xor.b64 %rd1182, %rd1181, %rd1164; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r670}, %rd1176; + } + { + .reg .b32 %dummy; + mov.b64 {%r671,%dummy}, %rd1176; + } + shf.l.wrap.b32 %r672, %r671, %r670, 1; + shf.l.wrap.b32 %r673, %r670, %r671, 1; + mov.b64 %rd1183, {%r673, %r672}; + xor.b64 %rd1184, %rd1183, %rd1168; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r674}, %rd1160; + } + { + .reg .b32 %dummy; + mov.b64 {%r675,%dummy}, %rd1160; + } + shf.l.wrap.b32 %r676, %r675, %r674, 1; + shf.l.wrap.b32 %r677, %r674, %r675, 1; + mov.b64 %rd1185, {%r677, %r676}; + xor.b64 %rd1186, %rd1172, %rd1185; + xor.b64 %rd1187, %rd1178, %rd1156; + xor.b64 %rd1188, %rd1180, %rd1086; + xor.b64 %rd1189, %rd1182, %rd1089; + xor.b64 %rd1190, %rd1184, %rd1092; + xor.b64 %rd1191, %rd1186, %rd1095; + xor.b64 %rd1192, %rd1178, %rd1098; + xor.b64 %rd1193, %rd1180, %rd1101; + xor.b64 %rd1194, %rd1182, %rd1104; + xor.b64 %rd1195, %rd1184, %rd1107; + xor.b64 %rd1196, %rd1186, %rd1110; + xor.b64 %rd1197, %rd1178, %rd1113; + xor.b64 %rd1198, %rd1180, %rd1116; + xor.b64 %rd1199, %rd1182, %rd1119; + xor.b64 %rd1200, %rd1184, %rd1122; + xor.b64 %rd1201, %rd1186, %rd1125; + xor.b64 %rd1202, %rd1178, %rd1128; + xor.b64 %rd1203, %rd1180, %rd1131; + xor.b64 %rd1204, %rd1182, %rd1134; + xor.b64 %rd1205, %rd1184, %rd1137; + xor.b64 %rd1206, %rd1186, %rd1140; + xor.b64 %rd1207, %rd1178, %rd1143; + xor.b64 %rd1208, %rd1180, %rd1146; + xor.b64 %rd1209, %rd1182, %rd1149; + xor.b64 %rd1210, %rd1184, %rd1152; + xor.b64 %rd1211, %rd1186, %rd1155; + { + .reg .b32 %dummy; + mov.b64 {%r678,%dummy}, %rd1193; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r679}, %rd1193; + } + shf.r.wrap.b32 %r680, %r679, %r678, 20; + shf.r.wrap.b32 %r681, %r678, %r679, 20; + mov.b64 %rd1212, {%r681, %r680}; + { + .reg .b32 %dummy; + mov.b64 {%r682,%dummy}, %rd1199; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r683}, %rd1199; + } + shf.r.wrap.b32 %r684, %r683, %r682, 21; + shf.r.wrap.b32 %r685, %r682, %r683, 21; + mov.b64 %rd1213, {%r685, %r684}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r686}, %rd1205; + } + { + .reg .b32 %dummy; + mov.b64 {%r687,%dummy}, %rd1205; + } + shf.l.wrap.b32 %r688, %r687, %r686, 21; + shf.l.wrap.b32 %r689, %r686, %r687, 21; + mov.b64 %rd1214, {%r689, %r688}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r690}, %rd1211; + } + { + .reg .b32 %dummy; + mov.b64 {%r691,%dummy}, %rd1211; + } + shf.l.wrap.b32 %r692, %r691, %r690, 14; + shf.l.wrap.b32 %r693, %r690, %r691, 14; + mov.b64 %rd1215, {%r693, %r692}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r694}, %rd1190; + } + { + .reg .b32 %dummy; + mov.b64 {%r695,%dummy}, %rd1190; + } + shf.l.wrap.b32 %r696, %r695, %r694, 28; + shf.l.wrap.b32 %r697, %r694, %r695, 28; + mov.b64 %rd1216, {%r697, %r696}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r698}, %rd1196; + } + { + .reg .b32 %dummy; + mov.b64 {%r699,%dummy}, %rd1196; + } + shf.l.wrap.b32 %r700, %r699, %r698, 20; + shf.l.wrap.b32 %r701, %r698, %r699, 20; + mov.b64 %rd1217, {%r701, %r700}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r702}, %rd1197; + } + { + .reg .b32 %dummy; + mov.b64 {%r703,%dummy}, %rd1197; + } + shf.l.wrap.b32 %r704, %r703, %r702, 3; + shf.l.wrap.b32 %r705, %r702, %r703, 3; + mov.b64 %rd1218, {%r705, %r704}; + { + .reg .b32 %dummy; + mov.b64 {%r706,%dummy}, %rd1203; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r707}, %rd1203; + } + shf.r.wrap.b32 %r708, %r707, %r706, 19; + shf.r.wrap.b32 %r709, %r706, %r707, 19; + mov.b64 %rd1219, {%r709, %r708}; + { + .reg .b32 %dummy; + mov.b64 {%r710,%dummy}, %rd1209; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r711}, %rd1209; + } + shf.r.wrap.b32 %r712, %r711, %r710, 3; + shf.r.wrap.b32 %r713, %r710, %r711, 3; + mov.b64 %rd1220, {%r713, %r712}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r714}, %rd1188; + } + { + .reg .b32 %dummy; + mov.b64 {%r715,%dummy}, %rd1188; + } + shf.l.wrap.b32 %r716, %r715, %r714, 1; + shf.l.wrap.b32 %r717, %r714, %r715, 1; + mov.b64 %rd1221, {%r717, %r716}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r718}, %rd1194; + } + { + .reg .b32 %dummy; + mov.b64 {%r719,%dummy}, %rd1194; + } + shf.l.wrap.b32 %r720, %r719, %r718, 6; + shf.l.wrap.b32 %r721, %r718, %r719, 6; + mov.b64 %rd1222, {%r721, %r720}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r722}, %rd1200; + } + { + .reg .b32 %dummy; + mov.b64 {%r723,%dummy}, %rd1200; + } + shf.l.wrap.b32 %r724, %r723, %r722, 25; + shf.l.wrap.b32 %r725, %r722, %r723, 25; + mov.b64 %rd1223, {%r725, %r724}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r726}, %rd1206; + } + { + .reg .b32 %dummy; + mov.b64 {%r727,%dummy}, %rd1206; + } + shf.l.wrap.b32 %r728, %r727, %r726, 8; + shf.l.wrap.b32 %r729, %r726, %r727, 8; + mov.b64 %rd1224, {%r729, %r728}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r730}, %rd1207; + } + { + .reg .b32 %dummy; + mov.b64 {%r731,%dummy}, %rd1207; + } + shf.l.wrap.b32 %r732, %r731, %r730, 18; + shf.l.wrap.b32 %r733, %r730, %r731, 18; + mov.b64 %rd1225, {%r733, %r732}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r734}, %rd1191; + } + { + .reg .b32 %dummy; + mov.b64 {%r735,%dummy}, %rd1191; + } + shf.l.wrap.b32 %r736, %r735, %r734, 27; + shf.l.wrap.b32 %r737, %r734, %r735, 27; + mov.b64 %rd1226, {%r737, %r736}; + { + .reg .b32 %dummy; + mov.b64 {%r738,%dummy}, %rd1192; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r739}, %rd1192; + } + shf.r.wrap.b32 %r740, %r739, %r738, 28; + shf.r.wrap.b32 %r741, %r738, %r739, 28; + mov.b64 %rd1227, {%r741, %r740}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r742}, %rd1198; + } + { + .reg .b32 %dummy; + mov.b64 {%r743,%dummy}, %rd1198; + } + shf.l.wrap.b32 %r744, %r743, %r742, 10; + shf.l.wrap.b32 %r745, %r742, %r743, 10; + mov.b64 %rd1228, {%r745, %r744}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r746}, %rd1204; + } + { + .reg .b32 %dummy; + mov.b64 {%r747,%dummy}, %rd1204; + } + shf.l.wrap.b32 %r748, %r747, %r746, 15; + shf.l.wrap.b32 %r749, %r746, %r747, 15; + mov.b64 %rd1229, {%r749, %r748}; + { + .reg .b32 %dummy; + mov.b64 {%r750,%dummy}, %rd1210; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r751}, %rd1210; + } + shf.r.wrap.b32 %r752, %r751, %r750, 8; + shf.r.wrap.b32 %r753, %r750, %r751, 8; + mov.b64 %rd1230, {%r753, %r752}; + { + .reg .b32 %dummy; + mov.b64 {%r754,%dummy}, %rd1189; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r755}, %rd1189; + } + shf.r.wrap.b32 %r756, %r755, %r754, 2; + shf.r.wrap.b32 %r757, %r754, %r755, 2; + mov.b64 %rd1231, {%r757, %r756}; + { + .reg .b32 %dummy; + mov.b64 {%r758,%dummy}, %rd1195; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r759}, %rd1195; + } + shf.r.wrap.b32 %r760, %r759, %r758, 9; + shf.r.wrap.b32 %r761, %r758, %r759, 9; + mov.b64 %rd1232, {%r761, %r760}; + { + .reg .b32 %dummy; + mov.b64 {%r762,%dummy}, %rd1201; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r763}, %rd1201; + } + shf.r.wrap.b32 %r764, %r763, %r762, 25; + shf.r.wrap.b32 %r765, %r762, %r763, 25; + mov.b64 %rd1233, {%r765, %r764}; + { + .reg .b32 %dummy; + mov.b64 {%r766,%dummy}, %rd1202; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r767}, %rd1202; + } + shf.r.wrap.b32 %r768, %r767, %r766, 23; + shf.r.wrap.b32 %r769, %r766, %r767, 23; + mov.b64 %rd1234, {%r769, %r768}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r770}, %rd1208; + } + { + .reg .b32 %dummy; + mov.b64 {%r771,%dummy}, %rd1208; + } + shf.l.wrap.b32 %r772, %r771, %r770, 2; + shf.l.wrap.b32 %r773, %r770, %r771, 2; + mov.b64 %rd1235, {%r773, %r772}; + not.b64 %rd1236, %rd1212; + and.b64 %rd1237, %rd1213, %rd1236; + xor.b64 %rd1238, %rd1237, %rd1187; + not.b64 %rd1239, %rd1213; + and.b64 %rd1240, %rd1214, %rd1239; + xor.b64 %rd1241, %rd1240, %rd1212; + not.b64 %rd1242, %rd1214; + and.b64 %rd1243, %rd1215, %rd1242; + xor.b64 %rd1244, %rd1243, %rd1213; + not.b64 %rd1245, %rd1215; + and.b64 %rd1246, %rd1187, %rd1245; + xor.b64 %rd1247, %rd1246, %rd1214; + not.b64 %rd1248, %rd1187; + and.b64 %rd1249, %rd1212, %rd1248; + xor.b64 %rd1250, %rd1249, %rd1215; + not.b64 %rd1251, %rd1217; + and.b64 %rd1252, %rd1218, %rd1251; + xor.b64 %rd1253, %rd1252, %rd1216; + not.b64 %rd1254, %rd1218; + and.b64 %rd1255, %rd1219, %rd1254; + xor.b64 %rd1256, %rd1255, %rd1217; + not.b64 %rd1257, %rd1219; + and.b64 %rd1258, %rd1220, %rd1257; + xor.b64 %rd1259, %rd1258, %rd1218; + not.b64 %rd1260, %rd1220; + and.b64 %rd1261, %rd1216, %rd1260; + xor.b64 %rd1262, %rd1261, %rd1219; + not.b64 %rd1263, %rd1216; + and.b64 %rd1264, %rd1217, %rd1263; + xor.b64 %rd1265, %rd1264, %rd1220; + not.b64 %rd1266, %rd1222; + and.b64 %rd1267, %rd1223, %rd1266; + xor.b64 %rd1268, %rd1267, %rd1221; + not.b64 %rd1269, %rd1223; + and.b64 %rd1270, %rd1224, %rd1269; + xor.b64 %rd1271, %rd1270, %rd1222; + not.b64 %rd1272, %rd1224; + and.b64 %rd1273, %rd1225, %rd1272; + xor.b64 %rd1274, %rd1273, %rd1223; + not.b64 %rd1275, %rd1225; + and.b64 %rd1276, %rd1221, %rd1275; + xor.b64 %rd1277, %rd1276, %rd1224; + not.b64 %rd1278, %rd1221; + and.b64 %rd1279, %rd1222, %rd1278; + xor.b64 %rd1280, %rd1279, %rd1225; + not.b64 %rd1281, %rd1227; + and.b64 %rd1282, %rd1228, %rd1281; + xor.b64 %rd1283, %rd1282, %rd1226; + not.b64 %rd1284, %rd1228; + and.b64 %rd1285, %rd1229, %rd1284; + xor.b64 %rd1286, %rd1285, %rd1227; + not.b64 %rd1287, %rd1229; + and.b64 %rd1288, %rd1230, %rd1287; + xor.b64 %rd1289, %rd1288, %rd1228; + not.b64 %rd1290, %rd1230; + and.b64 %rd1291, %rd1226, %rd1290; + xor.b64 %rd1292, %rd1291, %rd1229; + not.b64 %rd1293, %rd1226; + and.b64 %rd1294, %rd1227, %rd1293; + xor.b64 %rd1295, %rd1294, %rd1230; + not.b64 %rd1296, %rd1232; + and.b64 %rd1297, %rd1233, %rd1296; + xor.b64 %rd1298, %rd1297, %rd1231; + not.b64 %rd1299, %rd1233; + and.b64 %rd1300, %rd1234, %rd1299; + xor.b64 %rd1301, %rd1300, %rd1232; + not.b64 %rd1302, %rd1234; + and.b64 %rd1303, %rd1235, %rd1302; + xor.b64 %rd1304, %rd1303, %rd1233; + not.b64 %rd1305, %rd1235; + and.b64 %rd1306, %rd1231, %rd1305; + xor.b64 %rd1307, %rd1306, %rd1234; + not.b64 %rd1308, %rd1231; + and.b64 %rd1309, %rd1232, %rd1308; + xor.b64 %rd1310, %rd1309, %rd1235; + xor.b64 %rd1311, %rd1238, %rd21; + xor.b64 %rd1312, %rd1253, %rd1311; + xor.b64 %rd1313, %rd1312, %rd1268; + xor.b64 %rd1314, %rd1313, %rd1283; + xor.b64 %rd1315, %rd1314, %rd1298; + xor.b64 %rd1316, %rd1256, %rd1241; + xor.b64 %rd1317, %rd1316, %rd1271; + xor.b64 %rd1318, %rd1317, %rd1286; + xor.b64 %rd1319, %rd1318, %rd1301; + xor.b64 %rd1320, %rd1259, %rd1244; + xor.b64 %rd1321, %rd1320, %rd1274; + xor.b64 %rd1322, %rd1321, %rd1289; + xor.b64 %rd1323, %rd1322, %rd1304; + xor.b64 %rd1324, %rd1262, %rd1247; + xor.b64 %rd1325, %rd1324, %rd1277; + xor.b64 %rd1326, %rd1325, %rd1292; + xor.b64 %rd1327, %rd1326, %rd1307; + xor.b64 %rd1328, %rd1265, %rd1250; + xor.b64 %rd1329, %rd1328, %rd1280; + xor.b64 %rd1330, %rd1329, %rd1295; + xor.b64 %rd1331, %rd1330, %rd1310; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r774}, %rd1319; + } + { + .reg .b32 %dummy; + mov.b64 {%r775,%dummy}, %rd1319; + } + shf.l.wrap.b32 %r776, %r775, %r774, 1; + shf.l.wrap.b32 %r777, %r774, %r775, 1; + mov.b64 %rd1332, {%r777, %r776}; + xor.b64 %rd1333, %rd1331, %rd1332; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r778}, %rd1323; + } + { + .reg .b32 %dummy; + mov.b64 {%r779,%dummy}, %rd1323; + } + shf.l.wrap.b32 %r780, %r779, %r778, 1; + shf.l.wrap.b32 %r781, %r778, %r779, 1; + mov.b64 %rd1334, {%r781, %r780}; + xor.b64 %rd1335, %rd1334, %rd1315; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r782}, %rd1327; + } + { + .reg .b32 %dummy; + mov.b64 {%r783,%dummy}, %rd1327; + } + shf.l.wrap.b32 %r784, %r783, %r782, 1; + shf.l.wrap.b32 %r785, %r782, %r783, 1; + mov.b64 %rd1336, {%r785, %r784}; + xor.b64 %rd1337, %rd1336, %rd1319; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r786}, %rd1331; + } + { + .reg .b32 %dummy; + mov.b64 {%r787,%dummy}, %rd1331; + } + shf.l.wrap.b32 %r788, %r787, %r786, 1; + shf.l.wrap.b32 %r789, %r786, %r787, 1; + mov.b64 %rd1338, {%r789, %r788}; + xor.b64 %rd1339, %rd1338, %rd1323; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r790}, %rd1315; + } + { + .reg .b32 %dummy; + mov.b64 {%r791,%dummy}, %rd1315; + } + shf.l.wrap.b32 %r792, %r791, %r790, 1; + shf.l.wrap.b32 %r793, %r790, %r791, 1; + mov.b64 %rd1340, {%r793, %r792}; + xor.b64 %rd1341, %rd1327, %rd1340; + xor.b64 %rd1342, %rd1333, %rd1311; + xor.b64 %rd1343, %rd1335, %rd1241; + xor.b64 %rd1344, %rd1337, %rd1244; + xor.b64 %rd1345, %rd1339, %rd1247; + xor.b64 %rd1346, %rd1341, %rd1250; + xor.b64 %rd1347, %rd1333, %rd1253; + xor.b64 %rd1348, %rd1335, %rd1256; + xor.b64 %rd1349, %rd1337, %rd1259; + xor.b64 %rd1350, %rd1339, %rd1262; + xor.b64 %rd1351, %rd1341, %rd1265; + xor.b64 %rd1352, %rd1333, %rd1268; + xor.b64 %rd1353, %rd1335, %rd1271; + xor.b64 %rd1354, %rd1337, %rd1274; + xor.b64 %rd1355, %rd1339, %rd1277; + xor.b64 %rd1356, %rd1341, %rd1280; + xor.b64 %rd1357, %rd1333, %rd1283; + xor.b64 %rd1358, %rd1335, %rd1286; + xor.b64 %rd1359, %rd1337, %rd1289; + xor.b64 %rd1360, %rd1339, %rd1292; + xor.b64 %rd1361, %rd1341, %rd1295; + xor.b64 %rd1362, %rd1333, %rd1298; + xor.b64 %rd1363, %rd1335, %rd1301; + xor.b64 %rd1364, %rd1337, %rd1304; + xor.b64 %rd1365, %rd1339, %rd1307; + xor.b64 %rd1366, %rd1341, %rd1310; + { + .reg .b32 %dummy; + mov.b64 {%r794,%dummy}, %rd1348; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r795}, %rd1348; + } + shf.r.wrap.b32 %r796, %r795, %r794, 20; + shf.r.wrap.b32 %r797, %r794, %r795, 20; + mov.b64 %rd1367, {%r797, %r796}; + { + .reg .b32 %dummy; + mov.b64 {%r798,%dummy}, %rd1354; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r799}, %rd1354; + } + shf.r.wrap.b32 %r800, %r799, %r798, 21; + shf.r.wrap.b32 %r801, %r798, %r799, 21; + mov.b64 %rd1368, {%r801, %r800}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r802}, %rd1360; + } + { + .reg .b32 %dummy; + mov.b64 {%r803,%dummy}, %rd1360; + } + shf.l.wrap.b32 %r804, %r803, %r802, 21; + shf.l.wrap.b32 %r805, %r802, %r803, 21; + mov.b64 %rd1369, {%r805, %r804}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r806}, %rd1366; + } + { + .reg .b32 %dummy; + mov.b64 {%r807,%dummy}, %rd1366; + } + shf.l.wrap.b32 %r808, %r807, %r806, 14; + shf.l.wrap.b32 %r809, %r806, %r807, 14; + mov.b64 %rd1370, {%r809, %r808}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r810}, %rd1345; + } + { + .reg .b32 %dummy; + mov.b64 {%r811,%dummy}, %rd1345; + } + shf.l.wrap.b32 %r812, %r811, %r810, 28; + shf.l.wrap.b32 %r813, %r810, %r811, 28; + mov.b64 %rd1371, {%r813, %r812}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r814}, %rd1351; + } + { + .reg .b32 %dummy; + mov.b64 {%r815,%dummy}, %rd1351; + } + shf.l.wrap.b32 %r816, %r815, %r814, 20; + shf.l.wrap.b32 %r817, %r814, %r815, 20; + mov.b64 %rd1372, {%r817, %r816}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r818}, %rd1352; + } + { + .reg .b32 %dummy; + mov.b64 {%r819,%dummy}, %rd1352; + } + shf.l.wrap.b32 %r820, %r819, %r818, 3; + shf.l.wrap.b32 %r821, %r818, %r819, 3; + mov.b64 %rd1373, {%r821, %r820}; + { + .reg .b32 %dummy; + mov.b64 {%r822,%dummy}, %rd1358; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r823}, %rd1358; + } + shf.r.wrap.b32 %r824, %r823, %r822, 19; + shf.r.wrap.b32 %r825, %r822, %r823, 19; + mov.b64 %rd1374, {%r825, %r824}; + { + .reg .b32 %dummy; + mov.b64 {%r826,%dummy}, %rd1364; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r827}, %rd1364; + } + shf.r.wrap.b32 %r828, %r827, %r826, 3; + shf.r.wrap.b32 %r829, %r826, %r827, 3; + mov.b64 %rd1375, {%r829, %r828}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r830}, %rd1343; + } + { + .reg .b32 %dummy; + mov.b64 {%r831,%dummy}, %rd1343; + } + shf.l.wrap.b32 %r832, %r831, %r830, 1; + shf.l.wrap.b32 %r833, %r830, %r831, 1; + mov.b64 %rd1376, {%r833, %r832}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r834}, %rd1349; + } + { + .reg .b32 %dummy; + mov.b64 {%r835,%dummy}, %rd1349; + } + shf.l.wrap.b32 %r836, %r835, %r834, 6; + shf.l.wrap.b32 %r837, %r834, %r835, 6; + mov.b64 %rd1377, {%r837, %r836}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r838}, %rd1355; + } + { + .reg .b32 %dummy; + mov.b64 {%r839,%dummy}, %rd1355; + } + shf.l.wrap.b32 %r840, %r839, %r838, 25; + shf.l.wrap.b32 %r841, %r838, %r839, 25; + mov.b64 %rd1378, {%r841, %r840}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r842}, %rd1361; + } + { + .reg .b32 %dummy; + mov.b64 {%r843,%dummy}, %rd1361; + } + shf.l.wrap.b32 %r844, %r843, %r842, 8; + shf.l.wrap.b32 %r845, %r842, %r843, 8; + mov.b64 %rd1379, {%r845, %r844}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r846}, %rd1362; + } + { + .reg .b32 %dummy; + mov.b64 {%r847,%dummy}, %rd1362; + } + shf.l.wrap.b32 %r848, %r847, %r846, 18; + shf.l.wrap.b32 %r849, %r846, %r847, 18; + mov.b64 %rd1380, {%r849, %r848}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r850}, %rd1346; + } + { + .reg .b32 %dummy; + mov.b64 {%r851,%dummy}, %rd1346; + } + shf.l.wrap.b32 %r852, %r851, %r850, 27; + shf.l.wrap.b32 %r853, %r850, %r851, 27; + mov.b64 %rd1381, {%r853, %r852}; + { + .reg .b32 %dummy; + mov.b64 {%r854,%dummy}, %rd1347; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r855}, %rd1347; + } + shf.r.wrap.b32 %r856, %r855, %r854, 28; + shf.r.wrap.b32 %r857, %r854, %r855, 28; + mov.b64 %rd1382, {%r857, %r856}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r858}, %rd1353; + } + { + .reg .b32 %dummy; + mov.b64 {%r859,%dummy}, %rd1353; + } + shf.l.wrap.b32 %r860, %r859, %r858, 10; + shf.l.wrap.b32 %r861, %r858, %r859, 10; + mov.b64 %rd1383, {%r861, %r860}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r862}, %rd1359; + } + { + .reg .b32 %dummy; + mov.b64 {%r863,%dummy}, %rd1359; + } + shf.l.wrap.b32 %r864, %r863, %r862, 15; + shf.l.wrap.b32 %r865, %r862, %r863, 15; + mov.b64 %rd1384, {%r865, %r864}; + { + .reg .b32 %dummy; + mov.b64 {%r866,%dummy}, %rd1365; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r867}, %rd1365; + } + shf.r.wrap.b32 %r868, %r867, %r866, 8; + shf.r.wrap.b32 %r869, %r866, %r867, 8; + mov.b64 %rd1385, {%r869, %r868}; + { + .reg .b32 %dummy; + mov.b64 {%r870,%dummy}, %rd1344; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r871}, %rd1344; + } + shf.r.wrap.b32 %r872, %r871, %r870, 2; + shf.r.wrap.b32 %r873, %r870, %r871, 2; + mov.b64 %rd1386, {%r873, %r872}; + { + .reg .b32 %dummy; + mov.b64 {%r874,%dummy}, %rd1350; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r875}, %rd1350; + } + shf.r.wrap.b32 %r876, %r875, %r874, 9; + shf.r.wrap.b32 %r877, %r874, %r875, 9; + mov.b64 %rd1387, {%r877, %r876}; + { + .reg .b32 %dummy; + mov.b64 {%r878,%dummy}, %rd1356; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r879}, %rd1356; + } + shf.r.wrap.b32 %r880, %r879, %r878, 25; + shf.r.wrap.b32 %r881, %r878, %r879, 25; + mov.b64 %rd1388, {%r881, %r880}; + { + .reg .b32 %dummy; + mov.b64 {%r882,%dummy}, %rd1357; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r883}, %rd1357; + } + shf.r.wrap.b32 %r884, %r883, %r882, 23; + shf.r.wrap.b32 %r885, %r882, %r883, 23; + mov.b64 %rd1389, {%r885, %r884}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r886}, %rd1363; + } + { + .reg .b32 %dummy; + mov.b64 {%r887,%dummy}, %rd1363; + } + shf.l.wrap.b32 %r888, %r887, %r886, 2; + shf.l.wrap.b32 %r889, %r886, %r887, 2; + mov.b64 %rd1390, {%r889, %r888}; + not.b64 %rd1391, %rd1367; + and.b64 %rd1392, %rd1368, %rd1391; + xor.b64 %rd1393, %rd1392, %rd1342; + not.b64 %rd1394, %rd1368; + and.b64 %rd1395, %rd1369, %rd1394; + xor.b64 %rd1396, %rd1395, %rd1367; + not.b64 %rd1397, %rd1369; + and.b64 %rd1398, %rd1370, %rd1397; + xor.b64 %rd1399, %rd1398, %rd1368; + not.b64 %rd1400, %rd1370; + and.b64 %rd1401, %rd1342, %rd1400; + xor.b64 %rd1402, %rd1401, %rd1369; + not.b64 %rd1403, %rd1342; + and.b64 %rd1404, %rd1367, %rd1403; + xor.b64 %rd1405, %rd1404, %rd1370; + not.b64 %rd1406, %rd1372; + and.b64 %rd1407, %rd1373, %rd1406; + xor.b64 %rd1408, %rd1407, %rd1371; + not.b64 %rd1409, %rd1373; + and.b64 %rd1410, %rd1374, %rd1409; + xor.b64 %rd1411, %rd1410, %rd1372; + not.b64 %rd1412, %rd1374; + and.b64 %rd1413, %rd1375, %rd1412; + xor.b64 %rd1414, %rd1413, %rd1373; + not.b64 %rd1415, %rd1375; + and.b64 %rd1416, %rd1371, %rd1415; + xor.b64 %rd1417, %rd1416, %rd1374; + not.b64 %rd1418, %rd1371; + and.b64 %rd1419, %rd1372, %rd1418; + xor.b64 %rd1420, %rd1419, %rd1375; + not.b64 %rd1421, %rd1377; + and.b64 %rd1422, %rd1378, %rd1421; + xor.b64 %rd1423, %rd1422, %rd1376; + not.b64 %rd1424, %rd1378; + and.b64 %rd1425, %rd1379, %rd1424; + xor.b64 %rd1426, %rd1425, %rd1377; + not.b64 %rd1427, %rd1379; + and.b64 %rd1428, %rd1380, %rd1427; + xor.b64 %rd1429, %rd1428, %rd1378; + not.b64 %rd1430, %rd1380; + and.b64 %rd1431, %rd1376, %rd1430; + xor.b64 %rd1432, %rd1431, %rd1379; + not.b64 %rd1433, %rd1376; + and.b64 %rd1434, %rd1377, %rd1433; + xor.b64 %rd1435, %rd1434, %rd1380; + not.b64 %rd1436, %rd1382; + and.b64 %rd1437, %rd1383, %rd1436; + xor.b64 %rd1438, %rd1437, %rd1381; + not.b64 %rd1439, %rd1383; + and.b64 %rd1440, %rd1384, %rd1439; + xor.b64 %rd1441, %rd1440, %rd1382; + not.b64 %rd1442, %rd1384; + and.b64 %rd1443, %rd1385, %rd1442; + xor.b64 %rd1444, %rd1443, %rd1383; + not.b64 %rd1445, %rd1385; + and.b64 %rd1446, %rd1381, %rd1445; + xor.b64 %rd1447, %rd1446, %rd1384; + not.b64 %rd1448, %rd1381; + and.b64 %rd1449, %rd1382, %rd1448; + xor.b64 %rd1450, %rd1449, %rd1385; + not.b64 %rd1451, %rd1387; + and.b64 %rd1452, %rd1388, %rd1451; + xor.b64 %rd1453, %rd1452, %rd1386; + not.b64 %rd1454, %rd1388; + and.b64 %rd1455, %rd1389, %rd1454; + xor.b64 %rd1456, %rd1455, %rd1387; + not.b64 %rd1457, %rd1389; + and.b64 %rd1458, %rd1390, %rd1457; + xor.b64 %rd1459, %rd1458, %rd1388; + not.b64 %rd1460, %rd1390; + and.b64 %rd1461, %rd1386, %rd1460; + xor.b64 %rd1462, %rd1461, %rd1389; + not.b64 %rd1463, %rd1386; + and.b64 %rd1464, %rd1387, %rd1463; + xor.b64 %rd1465, %rd1464, %rd1390; + xor.b64 %rd1466, %rd1393, %rd22; + xor.b64 %rd1467, %rd1408, %rd1466; + xor.b64 %rd1468, %rd1467, %rd1423; + xor.b64 %rd1469, %rd1468, %rd1438; + xor.b64 %rd1470, %rd1469, %rd1453; + xor.b64 %rd1471, %rd1411, %rd1396; + xor.b64 %rd1472, %rd1471, %rd1426; + xor.b64 %rd1473, %rd1472, %rd1441; + xor.b64 %rd1474, %rd1473, %rd1456; + xor.b64 %rd1475, %rd1414, %rd1399; + xor.b64 %rd1476, %rd1475, %rd1429; + xor.b64 %rd1477, %rd1476, %rd1444; + xor.b64 %rd1478, %rd1477, %rd1459; + xor.b64 %rd1479, %rd1417, %rd1402; + xor.b64 %rd1480, %rd1479, %rd1432; + xor.b64 %rd1481, %rd1480, %rd1447; + xor.b64 %rd1482, %rd1481, %rd1462; + xor.b64 %rd1483, %rd1420, %rd1405; + xor.b64 %rd1484, %rd1483, %rd1435; + xor.b64 %rd1485, %rd1484, %rd1450; + xor.b64 %rd1486, %rd1485, %rd1465; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r890}, %rd1474; + } + { + .reg .b32 %dummy; + mov.b64 {%r891,%dummy}, %rd1474; + } + shf.l.wrap.b32 %r892, %r891, %r890, 1; + shf.l.wrap.b32 %r893, %r890, %r891, 1; + mov.b64 %rd1487, {%r893, %r892}; + xor.b64 %rd1488, %rd1486, %rd1487; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r894}, %rd1478; + } + { + .reg .b32 %dummy; + mov.b64 {%r895,%dummy}, %rd1478; + } + shf.l.wrap.b32 %r896, %r895, %r894, 1; + shf.l.wrap.b32 %r897, %r894, %r895, 1; + mov.b64 %rd1489, {%r897, %r896}; + xor.b64 %rd1490, %rd1489, %rd1470; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r898}, %rd1482; + } + { + .reg .b32 %dummy; + mov.b64 {%r899,%dummy}, %rd1482; + } + shf.l.wrap.b32 %r900, %r899, %r898, 1; + shf.l.wrap.b32 %r901, %r898, %r899, 1; + mov.b64 %rd1491, {%r901, %r900}; + xor.b64 %rd1492, %rd1491, %rd1474; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r902}, %rd1486; + } + { + .reg .b32 %dummy; + mov.b64 {%r903,%dummy}, %rd1486; + } + shf.l.wrap.b32 %r904, %r903, %r902, 1; + shf.l.wrap.b32 %r905, %r902, %r903, 1; + mov.b64 %rd1493, {%r905, %r904}; + xor.b64 %rd1494, %rd1493, %rd1478; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r906}, %rd1470; + } + { + .reg .b32 %dummy; + mov.b64 {%r907,%dummy}, %rd1470; + } + shf.l.wrap.b32 %r908, %r907, %r906, 1; + shf.l.wrap.b32 %r909, %r906, %r907, 1; + mov.b64 %rd1495, {%r909, %r908}; + xor.b64 %rd1496, %rd1482, %rd1495; + xor.b64 %rd1497, %rd1488, %rd1466; + xor.b64 %rd1498, %rd1490, %rd1396; + xor.b64 %rd1499, %rd1492, %rd1399; + xor.b64 %rd1500, %rd1494, %rd1402; + xor.b64 %rd1501, %rd1496, %rd1405; + xor.b64 %rd1502, %rd1488, %rd1408; + xor.b64 %rd1503, %rd1490, %rd1411; + xor.b64 %rd1504, %rd1492, %rd1414; + xor.b64 %rd1505, %rd1494, %rd1417; + xor.b64 %rd1506, %rd1496, %rd1420; + xor.b64 %rd1507, %rd1488, %rd1423; + xor.b64 %rd1508, %rd1490, %rd1426; + xor.b64 %rd1509, %rd1492, %rd1429; + xor.b64 %rd1510, %rd1494, %rd1432; + xor.b64 %rd1511, %rd1496, %rd1435; + xor.b64 %rd1512, %rd1488, %rd1438; + xor.b64 %rd1513, %rd1490, %rd1441; + xor.b64 %rd1514, %rd1492, %rd1444; + xor.b64 %rd1515, %rd1494, %rd1447; + xor.b64 %rd1516, %rd1496, %rd1450; + xor.b64 %rd1517, %rd1488, %rd1453; + xor.b64 %rd1518, %rd1490, %rd1456; + xor.b64 %rd1519, %rd1492, %rd1459; + xor.b64 %rd1520, %rd1494, %rd1462; + xor.b64 %rd1521, %rd1496, %rd1465; + { + .reg .b32 %dummy; + mov.b64 {%r910,%dummy}, %rd1503; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r911}, %rd1503; + } + shf.r.wrap.b32 %r912, %r911, %r910, 20; + shf.r.wrap.b32 %r913, %r910, %r911, 20; + mov.b64 %rd1522, {%r913, %r912}; + { + .reg .b32 %dummy; + mov.b64 {%r914,%dummy}, %rd1509; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r915}, %rd1509; + } + shf.r.wrap.b32 %r916, %r915, %r914, 21; + shf.r.wrap.b32 %r917, %r914, %r915, 21; + mov.b64 %rd1523, {%r917, %r916}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r918}, %rd1515; + } + { + .reg .b32 %dummy; + mov.b64 {%r919,%dummy}, %rd1515; + } + shf.l.wrap.b32 %r920, %r919, %r918, 21; + shf.l.wrap.b32 %r921, %r918, %r919, 21; + mov.b64 %rd1524, {%r921, %r920}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r922}, %rd1521; + } + { + .reg .b32 %dummy; + mov.b64 {%r923,%dummy}, %rd1521; + } + shf.l.wrap.b32 %r924, %r923, %r922, 14; + shf.l.wrap.b32 %r925, %r922, %r923, 14; + mov.b64 %rd1525, {%r925, %r924}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r926}, %rd1500; + } + { + .reg .b32 %dummy; + mov.b64 {%r927,%dummy}, %rd1500; + } + shf.l.wrap.b32 %r928, %r927, %r926, 28; + shf.l.wrap.b32 %r929, %r926, %r927, 28; + mov.b64 %rd1526, {%r929, %r928}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r930}, %rd1506; + } + { + .reg .b32 %dummy; + mov.b64 {%r931,%dummy}, %rd1506; + } + shf.l.wrap.b32 %r932, %r931, %r930, 20; + shf.l.wrap.b32 %r933, %r930, %r931, 20; + mov.b64 %rd1527, {%r933, %r932}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r934}, %rd1507; + } + { + .reg .b32 %dummy; + mov.b64 {%r935,%dummy}, %rd1507; + } + shf.l.wrap.b32 %r936, %r935, %r934, 3; + shf.l.wrap.b32 %r937, %r934, %r935, 3; + mov.b64 %rd1528, {%r937, %r936}; + { + .reg .b32 %dummy; + mov.b64 {%r938,%dummy}, %rd1513; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r939}, %rd1513; + } + shf.r.wrap.b32 %r940, %r939, %r938, 19; + shf.r.wrap.b32 %r941, %r938, %r939, 19; + mov.b64 %rd1529, {%r941, %r940}; + { + .reg .b32 %dummy; + mov.b64 {%r942,%dummy}, %rd1519; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r943}, %rd1519; + } + shf.r.wrap.b32 %r944, %r943, %r942, 3; + shf.r.wrap.b32 %r945, %r942, %r943, 3; + mov.b64 %rd1530, {%r945, %r944}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r946}, %rd1498; + } + { + .reg .b32 %dummy; + mov.b64 {%r947,%dummy}, %rd1498; + } + shf.l.wrap.b32 %r948, %r947, %r946, 1; + shf.l.wrap.b32 %r949, %r946, %r947, 1; + mov.b64 %rd1531, {%r949, %r948}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r950}, %rd1504; + } + { + .reg .b32 %dummy; + mov.b64 {%r951,%dummy}, %rd1504; + } + shf.l.wrap.b32 %r952, %r951, %r950, 6; + shf.l.wrap.b32 %r953, %r950, %r951, 6; + mov.b64 %rd1532, {%r953, %r952}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r954}, %rd1510; + } + { + .reg .b32 %dummy; + mov.b64 {%r955,%dummy}, %rd1510; + } + shf.l.wrap.b32 %r956, %r955, %r954, 25; + shf.l.wrap.b32 %r957, %r954, %r955, 25; + mov.b64 %rd1533, {%r957, %r956}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r958}, %rd1516; + } + { + .reg .b32 %dummy; + mov.b64 {%r959,%dummy}, %rd1516; + } + shf.l.wrap.b32 %r960, %r959, %r958, 8; + shf.l.wrap.b32 %r961, %r958, %r959, 8; + mov.b64 %rd1534, {%r961, %r960}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r962}, %rd1517; + } + { + .reg .b32 %dummy; + mov.b64 {%r963,%dummy}, %rd1517; + } + shf.l.wrap.b32 %r964, %r963, %r962, 18; + shf.l.wrap.b32 %r965, %r962, %r963, 18; + mov.b64 %rd1535, {%r965, %r964}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r966}, %rd1501; + } + { + .reg .b32 %dummy; + mov.b64 {%r967,%dummy}, %rd1501; + } + shf.l.wrap.b32 %r968, %r967, %r966, 27; + shf.l.wrap.b32 %r969, %r966, %r967, 27; + mov.b64 %rd1536, {%r969, %r968}; + { + .reg .b32 %dummy; + mov.b64 {%r970,%dummy}, %rd1502; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r971}, %rd1502; + } + shf.r.wrap.b32 %r972, %r971, %r970, 28; + shf.r.wrap.b32 %r973, %r970, %r971, 28; + mov.b64 %rd1537, {%r973, %r972}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r974}, %rd1508; + } + { + .reg .b32 %dummy; + mov.b64 {%r975,%dummy}, %rd1508; + } + shf.l.wrap.b32 %r976, %r975, %r974, 10; + shf.l.wrap.b32 %r977, %r974, %r975, 10; + mov.b64 %rd1538, {%r977, %r976}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r978}, %rd1514; + } + { + .reg .b32 %dummy; + mov.b64 {%r979,%dummy}, %rd1514; + } + shf.l.wrap.b32 %r980, %r979, %r978, 15; + shf.l.wrap.b32 %r981, %r978, %r979, 15; + mov.b64 %rd1539, {%r981, %r980}; + { + .reg .b32 %dummy; + mov.b64 {%r982,%dummy}, %rd1520; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r983}, %rd1520; + } + shf.r.wrap.b32 %r984, %r983, %r982, 8; + shf.r.wrap.b32 %r985, %r982, %r983, 8; + mov.b64 %rd1540, {%r985, %r984}; + { + .reg .b32 %dummy; + mov.b64 {%r986,%dummy}, %rd1499; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r987}, %rd1499; + } + shf.r.wrap.b32 %r988, %r987, %r986, 2; + shf.r.wrap.b32 %r989, %r986, %r987, 2; + mov.b64 %rd1541, {%r989, %r988}; + { + .reg .b32 %dummy; + mov.b64 {%r990,%dummy}, %rd1505; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r991}, %rd1505; + } + shf.r.wrap.b32 %r992, %r991, %r990, 9; + shf.r.wrap.b32 %r993, %r990, %r991, 9; + mov.b64 %rd1542, {%r993, %r992}; + { + .reg .b32 %dummy; + mov.b64 {%r994,%dummy}, %rd1511; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r995}, %rd1511; + } + shf.r.wrap.b32 %r996, %r995, %r994, 25; + shf.r.wrap.b32 %r997, %r994, %r995, 25; + mov.b64 %rd1543, {%r997, %r996}; + { + .reg .b32 %dummy; + mov.b64 {%r998,%dummy}, %rd1512; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r999}, %rd1512; + } + shf.r.wrap.b32 %r1000, %r999, %r998, 23; + shf.r.wrap.b32 %r1001, %r998, %r999, 23; + mov.b64 %rd1544, {%r1001, %r1000}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1002}, %rd1518; + } + { + .reg .b32 %dummy; + mov.b64 {%r1003,%dummy}, %rd1518; + } + shf.l.wrap.b32 %r1004, %r1003, %r1002, 2; + shf.l.wrap.b32 %r1005, %r1002, %r1003, 2; + mov.b64 %rd1545, {%r1005, %r1004}; + not.b64 %rd1546, %rd1522; + and.b64 %rd1547, %rd1523, %rd1546; + xor.b64 %rd1548, %rd1547, %rd1497; + not.b64 %rd1549, %rd1523; + and.b64 %rd1550, %rd1524, %rd1549; + xor.b64 %rd1551, %rd1550, %rd1522; + not.b64 %rd1552, %rd1524; + and.b64 %rd1553, %rd1525, %rd1552; + xor.b64 %rd1554, %rd1553, %rd1523; + not.b64 %rd1555, %rd1525; + and.b64 %rd1556, %rd1497, %rd1555; + xor.b64 %rd1557, %rd1556, %rd1524; + not.b64 %rd1558, %rd1497; + and.b64 %rd1559, %rd1522, %rd1558; + xor.b64 %rd1560, %rd1559, %rd1525; + not.b64 %rd1561, %rd1527; + and.b64 %rd1562, %rd1528, %rd1561; + xor.b64 %rd1563, %rd1562, %rd1526; + not.b64 %rd1564, %rd1528; + and.b64 %rd1565, %rd1529, %rd1564; + xor.b64 %rd1566, %rd1565, %rd1527; + not.b64 %rd1567, %rd1529; + and.b64 %rd1568, %rd1530, %rd1567; + xor.b64 %rd1569, %rd1568, %rd1528; + not.b64 %rd1570, %rd1530; + and.b64 %rd1571, %rd1526, %rd1570; + xor.b64 %rd1572, %rd1571, %rd1529; + not.b64 %rd1573, %rd1526; + and.b64 %rd1574, %rd1527, %rd1573; + xor.b64 %rd1575, %rd1574, %rd1530; + not.b64 %rd1576, %rd1532; + and.b64 %rd1577, %rd1533, %rd1576; + xor.b64 %rd1578, %rd1577, %rd1531; + not.b64 %rd1579, %rd1533; + and.b64 %rd1580, %rd1534, %rd1579; + xor.b64 %rd1581, %rd1580, %rd1532; + not.b64 %rd1582, %rd1534; + and.b64 %rd1583, %rd1535, %rd1582; + xor.b64 %rd1584, %rd1583, %rd1533; + not.b64 %rd1585, %rd1535; + and.b64 %rd1586, %rd1531, %rd1585; + xor.b64 %rd1587, %rd1586, %rd1534; + not.b64 %rd1588, %rd1531; + and.b64 %rd1589, %rd1532, %rd1588; + xor.b64 %rd1590, %rd1589, %rd1535; + not.b64 %rd1591, %rd1537; + and.b64 %rd1592, %rd1538, %rd1591; + xor.b64 %rd1593, %rd1592, %rd1536; + not.b64 %rd1594, %rd1538; + and.b64 %rd1595, %rd1539, %rd1594; + xor.b64 %rd1596, %rd1595, %rd1537; + not.b64 %rd1597, %rd1539; + and.b64 %rd1598, %rd1540, %rd1597; + xor.b64 %rd1599, %rd1598, %rd1538; + not.b64 %rd1600, %rd1540; + and.b64 %rd1601, %rd1536, %rd1600; + xor.b64 %rd1602, %rd1601, %rd1539; + not.b64 %rd1603, %rd1536; + and.b64 %rd1604, %rd1537, %rd1603; + xor.b64 %rd1605, %rd1604, %rd1540; + not.b64 %rd1606, %rd1542; + and.b64 %rd1607, %rd1543, %rd1606; + xor.b64 %rd1608, %rd1607, %rd1541; + not.b64 %rd1609, %rd1543; + and.b64 %rd1610, %rd1544, %rd1609; + xor.b64 %rd1611, %rd1610, %rd1542; + not.b64 %rd1612, %rd1544; + and.b64 %rd1613, %rd1545, %rd1612; + xor.b64 %rd1614, %rd1613, %rd1543; + not.b64 %rd1615, %rd1545; + and.b64 %rd1616, %rd1541, %rd1615; + xor.b64 %rd1617, %rd1616, %rd1544; + not.b64 %rd1618, %rd1541; + and.b64 %rd1619, %rd1542, %rd1618; + xor.b64 %rd1620, %rd1619, %rd1545; + xor.b64 %rd1621, %rd1548, %rd23; + xor.b64 %rd1622, %rd1563, %rd1621; + xor.b64 %rd1623, %rd1622, %rd1578; + xor.b64 %rd1624, %rd1623, %rd1593; + xor.b64 %rd1625, %rd1624, %rd1608; + xor.b64 %rd1626, %rd1566, %rd1551; + xor.b64 %rd1627, %rd1626, %rd1581; + xor.b64 %rd1628, %rd1627, %rd1596; + xor.b64 %rd1629, %rd1628, %rd1611; + xor.b64 %rd1630, %rd1569, %rd1554; + xor.b64 %rd1631, %rd1630, %rd1584; + xor.b64 %rd1632, %rd1631, %rd1599; + xor.b64 %rd1633, %rd1632, %rd1614; + xor.b64 %rd1634, %rd1572, %rd1557; + xor.b64 %rd1635, %rd1634, %rd1587; + xor.b64 %rd1636, %rd1635, %rd1602; + xor.b64 %rd1637, %rd1636, %rd1617; + xor.b64 %rd1638, %rd1575, %rd1560; + xor.b64 %rd1639, %rd1638, %rd1590; + xor.b64 %rd1640, %rd1639, %rd1605; + xor.b64 %rd1641, %rd1640, %rd1620; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1006}, %rd1629; + } + { + .reg .b32 %dummy; + mov.b64 {%r1007,%dummy}, %rd1629; + } + shf.l.wrap.b32 %r1008, %r1007, %r1006, 1; + shf.l.wrap.b32 %r1009, %r1006, %r1007, 1; + mov.b64 %rd1642, {%r1009, %r1008}; + xor.b64 %rd1643, %rd1641, %rd1642; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1010}, %rd1633; + } + { + .reg .b32 %dummy; + mov.b64 {%r1011,%dummy}, %rd1633; + } + shf.l.wrap.b32 %r1012, %r1011, %r1010, 1; + shf.l.wrap.b32 %r1013, %r1010, %r1011, 1; + mov.b64 %rd1644, {%r1013, %r1012}; + xor.b64 %rd1645, %rd1644, %rd1625; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1014}, %rd1637; + } + { + .reg .b32 %dummy; + mov.b64 {%r1015,%dummy}, %rd1637; + } + shf.l.wrap.b32 %r1016, %r1015, %r1014, 1; + shf.l.wrap.b32 %r1017, %r1014, %r1015, 1; + mov.b64 %rd1646, {%r1017, %r1016}; + xor.b64 %rd1647, %rd1646, %rd1629; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1018}, %rd1641; + } + { + .reg .b32 %dummy; + mov.b64 {%r1019,%dummy}, %rd1641; + } + shf.l.wrap.b32 %r1020, %r1019, %r1018, 1; + shf.l.wrap.b32 %r1021, %r1018, %r1019, 1; + mov.b64 %rd1648, {%r1021, %r1020}; + xor.b64 %rd1649, %rd1648, %rd1633; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1022}, %rd1625; + } + { + .reg .b32 %dummy; + mov.b64 {%r1023,%dummy}, %rd1625; + } + shf.l.wrap.b32 %r1024, %r1023, %r1022, 1; + shf.l.wrap.b32 %r1025, %r1022, %r1023, 1; + mov.b64 %rd1650, {%r1025, %r1024}; + xor.b64 %rd1651, %rd1637, %rd1650; + xor.b64 %rd1652, %rd1643, %rd1621; + xor.b64 %rd1653, %rd1645, %rd1551; + xor.b64 %rd1654, %rd1647, %rd1554; + xor.b64 %rd1655, %rd1649, %rd1557; + xor.b64 %rd1656, %rd1651, %rd1560; + xor.b64 %rd1657, %rd1643, %rd1563; + xor.b64 %rd1658, %rd1645, %rd1566; + xor.b64 %rd1659, %rd1647, %rd1569; + xor.b64 %rd1660, %rd1649, %rd1572; + xor.b64 %rd1661, %rd1651, %rd1575; + xor.b64 %rd1662, %rd1643, %rd1578; + xor.b64 %rd1663, %rd1645, %rd1581; + xor.b64 %rd1664, %rd1647, %rd1584; + xor.b64 %rd1665, %rd1649, %rd1587; + xor.b64 %rd1666, %rd1651, %rd1590; + xor.b64 %rd1667, %rd1643, %rd1593; + xor.b64 %rd1668, %rd1645, %rd1596; + xor.b64 %rd1669, %rd1647, %rd1599; + xor.b64 %rd1670, %rd1649, %rd1602; + xor.b64 %rd1671, %rd1651, %rd1605; + xor.b64 %rd1672, %rd1643, %rd1608; + xor.b64 %rd1673, %rd1645, %rd1611; + xor.b64 %rd1674, %rd1647, %rd1614; + xor.b64 %rd1675, %rd1649, %rd1617; + xor.b64 %rd1676, %rd1651, %rd1620; + { + .reg .b32 %dummy; + mov.b64 {%r1026,%dummy}, %rd1658; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1027}, %rd1658; + } + shf.r.wrap.b32 %r1028, %r1027, %r1026, 20; + shf.r.wrap.b32 %r1029, %r1026, %r1027, 20; + mov.b64 %rd1677, {%r1029, %r1028}; + { + .reg .b32 %dummy; + mov.b64 {%r1030,%dummy}, %rd1664; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1031}, %rd1664; + } + shf.r.wrap.b32 %r1032, %r1031, %r1030, 21; + shf.r.wrap.b32 %r1033, %r1030, %r1031, 21; + mov.b64 %rd1678, {%r1033, %r1032}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1034}, %rd1670; + } + { + .reg .b32 %dummy; + mov.b64 {%r1035,%dummy}, %rd1670; + } + shf.l.wrap.b32 %r1036, %r1035, %r1034, 21; + shf.l.wrap.b32 %r1037, %r1034, %r1035, 21; + mov.b64 %rd1679, {%r1037, %r1036}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1038}, %rd1676; + } + { + .reg .b32 %dummy; + mov.b64 {%r1039,%dummy}, %rd1676; + } + shf.l.wrap.b32 %r1040, %r1039, %r1038, 14; + shf.l.wrap.b32 %r1041, %r1038, %r1039, 14; + mov.b64 %rd1680, {%r1041, %r1040}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1042}, %rd1655; + } + { + .reg .b32 %dummy; + mov.b64 {%r1043,%dummy}, %rd1655; + } + shf.l.wrap.b32 %r1044, %r1043, %r1042, 28; + shf.l.wrap.b32 %r1045, %r1042, %r1043, 28; + mov.b64 %rd1681, {%r1045, %r1044}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1046}, %rd1661; + } + { + .reg .b32 %dummy; + mov.b64 {%r1047,%dummy}, %rd1661; + } + shf.l.wrap.b32 %r1048, %r1047, %r1046, 20; + shf.l.wrap.b32 %r1049, %r1046, %r1047, 20; + mov.b64 %rd1682, {%r1049, %r1048}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1050}, %rd1662; + } + { + .reg .b32 %dummy; + mov.b64 {%r1051,%dummy}, %rd1662; + } + shf.l.wrap.b32 %r1052, %r1051, %r1050, 3; + shf.l.wrap.b32 %r1053, %r1050, %r1051, 3; + mov.b64 %rd1683, {%r1053, %r1052}; + { + .reg .b32 %dummy; + mov.b64 {%r1054,%dummy}, %rd1668; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1055}, %rd1668; + } + shf.r.wrap.b32 %r1056, %r1055, %r1054, 19; + shf.r.wrap.b32 %r1057, %r1054, %r1055, 19; + mov.b64 %rd1684, {%r1057, %r1056}; + { + .reg .b32 %dummy; + mov.b64 {%r1058,%dummy}, %rd1674; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1059}, %rd1674; + } + shf.r.wrap.b32 %r1060, %r1059, %r1058, 3; + shf.r.wrap.b32 %r1061, %r1058, %r1059, 3; + mov.b64 %rd1685, {%r1061, %r1060}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1062}, %rd1653; + } + { + .reg .b32 %dummy; + mov.b64 {%r1063,%dummy}, %rd1653; + } + shf.l.wrap.b32 %r1064, %r1063, %r1062, 1; + shf.l.wrap.b32 %r1065, %r1062, %r1063, 1; + mov.b64 %rd1686, {%r1065, %r1064}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1066}, %rd1659; + } + { + .reg .b32 %dummy; + mov.b64 {%r1067,%dummy}, %rd1659; + } + shf.l.wrap.b32 %r1068, %r1067, %r1066, 6; + shf.l.wrap.b32 %r1069, %r1066, %r1067, 6; + mov.b64 %rd1687, {%r1069, %r1068}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1070}, %rd1665; + } + { + .reg .b32 %dummy; + mov.b64 {%r1071,%dummy}, %rd1665; + } + shf.l.wrap.b32 %r1072, %r1071, %r1070, 25; + shf.l.wrap.b32 %r1073, %r1070, %r1071, 25; + mov.b64 %rd1688, {%r1073, %r1072}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1074}, %rd1671; + } + { + .reg .b32 %dummy; + mov.b64 {%r1075,%dummy}, %rd1671; + } + shf.l.wrap.b32 %r1076, %r1075, %r1074, 8; + shf.l.wrap.b32 %r1077, %r1074, %r1075, 8; + mov.b64 %rd1689, {%r1077, %r1076}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1078}, %rd1672; + } + { + .reg .b32 %dummy; + mov.b64 {%r1079,%dummy}, %rd1672; + } + shf.l.wrap.b32 %r1080, %r1079, %r1078, 18; + shf.l.wrap.b32 %r1081, %r1078, %r1079, 18; + mov.b64 %rd1690, {%r1081, %r1080}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1082}, %rd1656; + } + { + .reg .b32 %dummy; + mov.b64 {%r1083,%dummy}, %rd1656; + } + shf.l.wrap.b32 %r1084, %r1083, %r1082, 27; + shf.l.wrap.b32 %r1085, %r1082, %r1083, 27; + mov.b64 %rd1691, {%r1085, %r1084}; + { + .reg .b32 %dummy; + mov.b64 {%r1086,%dummy}, %rd1657; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1087}, %rd1657; + } + shf.r.wrap.b32 %r1088, %r1087, %r1086, 28; + shf.r.wrap.b32 %r1089, %r1086, %r1087, 28; + mov.b64 %rd1692, {%r1089, %r1088}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1090}, %rd1663; + } + { + .reg .b32 %dummy; + mov.b64 {%r1091,%dummy}, %rd1663; + } + shf.l.wrap.b32 %r1092, %r1091, %r1090, 10; + shf.l.wrap.b32 %r1093, %r1090, %r1091, 10; + mov.b64 %rd1693, {%r1093, %r1092}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1094}, %rd1669; + } + { + .reg .b32 %dummy; + mov.b64 {%r1095,%dummy}, %rd1669; + } + shf.l.wrap.b32 %r1096, %r1095, %r1094, 15; + shf.l.wrap.b32 %r1097, %r1094, %r1095, 15; + mov.b64 %rd1694, {%r1097, %r1096}; + { + .reg .b32 %dummy; + mov.b64 {%r1098,%dummy}, %rd1675; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1099}, %rd1675; + } + shf.r.wrap.b32 %r1100, %r1099, %r1098, 8; + shf.r.wrap.b32 %r1101, %r1098, %r1099, 8; + mov.b64 %rd1695, {%r1101, %r1100}; + { + .reg .b32 %dummy; + mov.b64 {%r1102,%dummy}, %rd1654; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1103}, %rd1654; + } + shf.r.wrap.b32 %r1104, %r1103, %r1102, 2; + shf.r.wrap.b32 %r1105, %r1102, %r1103, 2; + mov.b64 %rd1696, {%r1105, %r1104}; + { + .reg .b32 %dummy; + mov.b64 {%r1106,%dummy}, %rd1660; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1107}, %rd1660; + } + shf.r.wrap.b32 %r1108, %r1107, %r1106, 9; + shf.r.wrap.b32 %r1109, %r1106, %r1107, 9; + mov.b64 %rd1697, {%r1109, %r1108}; + { + .reg .b32 %dummy; + mov.b64 {%r1110,%dummy}, %rd1666; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1111}, %rd1666; + } + shf.r.wrap.b32 %r1112, %r1111, %r1110, 25; + shf.r.wrap.b32 %r1113, %r1110, %r1111, 25; + mov.b64 %rd1698, {%r1113, %r1112}; + { + .reg .b32 %dummy; + mov.b64 {%r1114,%dummy}, %rd1667; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1115}, %rd1667; + } + shf.r.wrap.b32 %r1116, %r1115, %r1114, 23; + shf.r.wrap.b32 %r1117, %r1114, %r1115, 23; + mov.b64 %rd1699, {%r1117, %r1116}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1118}, %rd1673; + } + { + .reg .b32 %dummy; + mov.b64 {%r1119,%dummy}, %rd1673; + } + shf.l.wrap.b32 %r1120, %r1119, %r1118, 2; + shf.l.wrap.b32 %r1121, %r1118, %r1119, 2; + mov.b64 %rd1700, {%r1121, %r1120}; + not.b64 %rd1701, %rd1677; + and.b64 %rd1702, %rd1678, %rd1701; + xor.b64 %rd1703, %rd1702, %rd1652; + not.b64 %rd1704, %rd1678; + and.b64 %rd1705, %rd1679, %rd1704; + xor.b64 %rd1706, %rd1705, %rd1677; + not.b64 %rd1707, %rd1679; + and.b64 %rd1708, %rd1680, %rd1707; + xor.b64 %rd1709, %rd1708, %rd1678; + not.b64 %rd1710, %rd1680; + and.b64 %rd1711, %rd1652, %rd1710; + xor.b64 %rd1712, %rd1711, %rd1679; + not.b64 %rd1713, %rd1652; + and.b64 %rd1714, %rd1677, %rd1713; + xor.b64 %rd1715, %rd1714, %rd1680; + not.b64 %rd1716, %rd1682; + and.b64 %rd1717, %rd1683, %rd1716; + xor.b64 %rd1718, %rd1717, %rd1681; + not.b64 %rd1719, %rd1683; + and.b64 %rd1720, %rd1684, %rd1719; + xor.b64 %rd1721, %rd1720, %rd1682; + not.b64 %rd1722, %rd1684; + and.b64 %rd1723, %rd1685, %rd1722; + xor.b64 %rd1724, %rd1723, %rd1683; + not.b64 %rd1725, %rd1685; + and.b64 %rd1726, %rd1681, %rd1725; + xor.b64 %rd1727, %rd1726, %rd1684; + not.b64 %rd1728, %rd1681; + and.b64 %rd1729, %rd1682, %rd1728; + xor.b64 %rd1730, %rd1729, %rd1685; + not.b64 %rd1731, %rd1687; + and.b64 %rd1732, %rd1688, %rd1731; + xor.b64 %rd1733, %rd1732, %rd1686; + not.b64 %rd1734, %rd1688; + and.b64 %rd1735, %rd1689, %rd1734; + xor.b64 %rd1736, %rd1735, %rd1687; + not.b64 %rd1737, %rd1689; + and.b64 %rd1738, %rd1690, %rd1737; + xor.b64 %rd1739, %rd1738, %rd1688; + not.b64 %rd1740, %rd1690; + and.b64 %rd1741, %rd1686, %rd1740; + xor.b64 %rd1742, %rd1741, %rd1689; + not.b64 %rd1743, %rd1686; + and.b64 %rd1744, %rd1687, %rd1743; + xor.b64 %rd1745, %rd1744, %rd1690; + not.b64 %rd1746, %rd1692; + and.b64 %rd1747, %rd1693, %rd1746; + xor.b64 %rd1748, %rd1747, %rd1691; + not.b64 %rd1749, %rd1693; + and.b64 %rd1750, %rd1694, %rd1749; + xor.b64 %rd1751, %rd1750, %rd1692; + not.b64 %rd1752, %rd1694; + and.b64 %rd1753, %rd1695, %rd1752; + xor.b64 %rd1754, %rd1753, %rd1693; + not.b64 %rd1755, %rd1695; + and.b64 %rd1756, %rd1691, %rd1755; + xor.b64 %rd1757, %rd1756, %rd1694; + not.b64 %rd1758, %rd1691; + and.b64 %rd1759, %rd1692, %rd1758; + xor.b64 %rd1760, %rd1759, %rd1695; + not.b64 %rd1761, %rd1697; + and.b64 %rd1762, %rd1698, %rd1761; + xor.b64 %rd1763, %rd1762, %rd1696; + not.b64 %rd1764, %rd1698; + and.b64 %rd1765, %rd1699, %rd1764; + xor.b64 %rd1766, %rd1765, %rd1697; + not.b64 %rd1767, %rd1699; + and.b64 %rd1768, %rd1700, %rd1767; + xor.b64 %rd1769, %rd1768, %rd1698; + not.b64 %rd1770, %rd1700; + and.b64 %rd1771, %rd1696, %rd1770; + xor.b64 %rd1772, %rd1771, %rd1699; + not.b64 %rd1773, %rd1696; + and.b64 %rd1774, %rd1697, %rd1773; + xor.b64 %rd1775, %rd1774, %rd1700; + xor.b64 %rd1776, %rd1703, %rd24; + xor.b64 %rd1777, %rd1718, %rd1776; + xor.b64 %rd1778, %rd1777, %rd1733; + xor.b64 %rd1779, %rd1778, %rd1748; + xor.b64 %rd1780, %rd1779, %rd1763; + xor.b64 %rd1781, %rd1721, %rd1706; + xor.b64 %rd1782, %rd1781, %rd1736; + xor.b64 %rd1783, %rd1782, %rd1751; + xor.b64 %rd1784, %rd1783, %rd1766; + xor.b64 %rd1785, %rd1724, %rd1709; + xor.b64 %rd1786, %rd1785, %rd1739; + xor.b64 %rd1787, %rd1786, %rd1754; + xor.b64 %rd1788, %rd1787, %rd1769; + xor.b64 %rd1789, %rd1727, %rd1712; + xor.b64 %rd1790, %rd1789, %rd1742; + xor.b64 %rd1791, %rd1790, %rd1757; + xor.b64 %rd1792, %rd1791, %rd1772; + xor.b64 %rd1793, %rd1730, %rd1715; + xor.b64 %rd1794, %rd1793, %rd1745; + xor.b64 %rd1795, %rd1794, %rd1760; + xor.b64 %rd1796, %rd1795, %rd1775; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1122}, %rd1784; + } + { + .reg .b32 %dummy; + mov.b64 {%r1123,%dummy}, %rd1784; + } + shf.l.wrap.b32 %r1124, %r1123, %r1122, 1; + shf.l.wrap.b32 %r1125, %r1122, %r1123, 1; + mov.b64 %rd1797, {%r1125, %r1124}; + xor.b64 %rd1798, %rd1796, %rd1797; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1126}, %rd1788; + } + { + .reg .b32 %dummy; + mov.b64 {%r1127,%dummy}, %rd1788; + } + shf.l.wrap.b32 %r1128, %r1127, %r1126, 1; + shf.l.wrap.b32 %r1129, %r1126, %r1127, 1; + mov.b64 %rd1799, {%r1129, %r1128}; + xor.b64 %rd1800, %rd1799, %rd1780; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1130}, %rd1792; + } + { + .reg .b32 %dummy; + mov.b64 {%r1131,%dummy}, %rd1792; + } + shf.l.wrap.b32 %r1132, %r1131, %r1130, 1; + shf.l.wrap.b32 %r1133, %r1130, %r1131, 1; + mov.b64 %rd1801, {%r1133, %r1132}; + xor.b64 %rd1802, %rd1801, %rd1784; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1134}, %rd1796; + } + { + .reg .b32 %dummy; + mov.b64 {%r1135,%dummy}, %rd1796; + } + shf.l.wrap.b32 %r1136, %r1135, %r1134, 1; + shf.l.wrap.b32 %r1137, %r1134, %r1135, 1; + mov.b64 %rd1803, {%r1137, %r1136}; + xor.b64 %rd1804, %rd1803, %rd1788; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1138}, %rd1780; + } + { + .reg .b32 %dummy; + mov.b64 {%r1139,%dummy}, %rd1780; + } + shf.l.wrap.b32 %r1140, %r1139, %r1138, 1; + shf.l.wrap.b32 %r1141, %r1138, %r1139, 1; + mov.b64 %rd1805, {%r1141, %r1140}; + xor.b64 %rd1806, %rd1792, %rd1805; + xor.b64 %rd1807, %rd1798, %rd1776; + xor.b64 %rd1808, %rd1800, %rd1706; + xor.b64 %rd1809, %rd1802, %rd1709; + xor.b64 %rd1810, %rd1804, %rd1712; + xor.b64 %rd1811, %rd1806, %rd1715; + xor.b64 %rd1812, %rd1798, %rd1718; + xor.b64 %rd1813, %rd1800, %rd1721; + xor.b64 %rd1814, %rd1802, %rd1724; + xor.b64 %rd1815, %rd1804, %rd1727; + xor.b64 %rd1816, %rd1806, %rd1730; + xor.b64 %rd1817, %rd1798, %rd1733; + xor.b64 %rd1818, %rd1800, %rd1736; + xor.b64 %rd1819, %rd1802, %rd1739; + xor.b64 %rd1820, %rd1804, %rd1742; + xor.b64 %rd1821, %rd1806, %rd1745; + xor.b64 %rd1822, %rd1798, %rd1748; + xor.b64 %rd1823, %rd1800, %rd1751; + xor.b64 %rd1824, %rd1802, %rd1754; + xor.b64 %rd1825, %rd1804, %rd1757; + xor.b64 %rd1826, %rd1806, %rd1760; + xor.b64 %rd1827, %rd1798, %rd1763; + xor.b64 %rd1828, %rd1800, %rd1766; + xor.b64 %rd1829, %rd1802, %rd1769; + xor.b64 %rd1830, %rd1804, %rd1772; + xor.b64 %rd1831, %rd1806, %rd1775; + { + .reg .b32 %dummy; + mov.b64 {%r1142,%dummy}, %rd1813; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1143}, %rd1813; + } + shf.r.wrap.b32 %r1144, %r1143, %r1142, 20; + shf.r.wrap.b32 %r1145, %r1142, %r1143, 20; + mov.b64 %rd1832, {%r1145, %r1144}; + { + .reg .b32 %dummy; + mov.b64 {%r1146,%dummy}, %rd1819; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1147}, %rd1819; + } + shf.r.wrap.b32 %r1148, %r1147, %r1146, 21; + shf.r.wrap.b32 %r1149, %r1146, %r1147, 21; + mov.b64 %rd1833, {%r1149, %r1148}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1150}, %rd1825; + } + { + .reg .b32 %dummy; + mov.b64 {%r1151,%dummy}, %rd1825; + } + shf.l.wrap.b32 %r1152, %r1151, %r1150, 21; + shf.l.wrap.b32 %r1153, %r1150, %r1151, 21; + mov.b64 %rd1834, {%r1153, %r1152}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1154}, %rd1831; + } + { + .reg .b32 %dummy; + mov.b64 {%r1155,%dummy}, %rd1831; + } + shf.l.wrap.b32 %r1156, %r1155, %r1154, 14; + shf.l.wrap.b32 %r1157, %r1154, %r1155, 14; + mov.b64 %rd1835, {%r1157, %r1156}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1158}, %rd1810; + } + { + .reg .b32 %dummy; + mov.b64 {%r1159,%dummy}, %rd1810; + } + shf.l.wrap.b32 %r1160, %r1159, %r1158, 28; + shf.l.wrap.b32 %r1161, %r1158, %r1159, 28; + mov.b64 %rd1836, {%r1161, %r1160}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1162}, %rd1816; + } + { + .reg .b32 %dummy; + mov.b64 {%r1163,%dummy}, %rd1816; + } + shf.l.wrap.b32 %r1164, %r1163, %r1162, 20; + shf.l.wrap.b32 %r1165, %r1162, %r1163, 20; + mov.b64 %rd1837, {%r1165, %r1164}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1166}, %rd1817; + } + { + .reg .b32 %dummy; + mov.b64 {%r1167,%dummy}, %rd1817; + } + shf.l.wrap.b32 %r1168, %r1167, %r1166, 3; + shf.l.wrap.b32 %r1169, %r1166, %r1167, 3; + mov.b64 %rd1838, {%r1169, %r1168}; + { + .reg .b32 %dummy; + mov.b64 {%r1170,%dummy}, %rd1823; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1171}, %rd1823; + } + shf.r.wrap.b32 %r1172, %r1171, %r1170, 19; + shf.r.wrap.b32 %r1173, %r1170, %r1171, 19; + mov.b64 %rd1839, {%r1173, %r1172}; + { + .reg .b32 %dummy; + mov.b64 {%r1174,%dummy}, %rd1829; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1175}, %rd1829; + } + shf.r.wrap.b32 %r1176, %r1175, %r1174, 3; + shf.r.wrap.b32 %r1177, %r1174, %r1175, 3; + mov.b64 %rd1840, {%r1177, %r1176}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1178}, %rd1808; + } + { + .reg .b32 %dummy; + mov.b64 {%r1179,%dummy}, %rd1808; + } + shf.l.wrap.b32 %r1180, %r1179, %r1178, 1; + shf.l.wrap.b32 %r1181, %r1178, %r1179, 1; + mov.b64 %rd1841, {%r1181, %r1180}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1182}, %rd1814; + } + { + .reg .b32 %dummy; + mov.b64 {%r1183,%dummy}, %rd1814; + } + shf.l.wrap.b32 %r1184, %r1183, %r1182, 6; + shf.l.wrap.b32 %r1185, %r1182, %r1183, 6; + mov.b64 %rd1842, {%r1185, %r1184}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1186}, %rd1820; + } + { + .reg .b32 %dummy; + mov.b64 {%r1187,%dummy}, %rd1820; + } + shf.l.wrap.b32 %r1188, %r1187, %r1186, 25; + shf.l.wrap.b32 %r1189, %r1186, %r1187, 25; + mov.b64 %rd1843, {%r1189, %r1188}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1190}, %rd1826; + } + { + .reg .b32 %dummy; + mov.b64 {%r1191,%dummy}, %rd1826; + } + shf.l.wrap.b32 %r1192, %r1191, %r1190, 8; + shf.l.wrap.b32 %r1193, %r1190, %r1191, 8; + mov.b64 %rd1844, {%r1193, %r1192}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1194}, %rd1827; + } + { + .reg .b32 %dummy; + mov.b64 {%r1195,%dummy}, %rd1827; + } + shf.l.wrap.b32 %r1196, %r1195, %r1194, 18; + shf.l.wrap.b32 %r1197, %r1194, %r1195, 18; + mov.b64 %rd1845, {%r1197, %r1196}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1198}, %rd1811; + } + { + .reg .b32 %dummy; + mov.b64 {%r1199,%dummy}, %rd1811; + } + shf.l.wrap.b32 %r1200, %r1199, %r1198, 27; + shf.l.wrap.b32 %r1201, %r1198, %r1199, 27; + mov.b64 %rd1846, {%r1201, %r1200}; + { + .reg .b32 %dummy; + mov.b64 {%r1202,%dummy}, %rd1812; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1203}, %rd1812; + } + shf.r.wrap.b32 %r1204, %r1203, %r1202, 28; + shf.r.wrap.b32 %r1205, %r1202, %r1203, 28; + mov.b64 %rd1847, {%r1205, %r1204}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1206}, %rd1818; + } + { + .reg .b32 %dummy; + mov.b64 {%r1207,%dummy}, %rd1818; + } + shf.l.wrap.b32 %r1208, %r1207, %r1206, 10; + shf.l.wrap.b32 %r1209, %r1206, %r1207, 10; + mov.b64 %rd1848, {%r1209, %r1208}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1210}, %rd1824; + } + { + .reg .b32 %dummy; + mov.b64 {%r1211,%dummy}, %rd1824; + } + shf.l.wrap.b32 %r1212, %r1211, %r1210, 15; + shf.l.wrap.b32 %r1213, %r1210, %r1211, 15; + mov.b64 %rd1849, {%r1213, %r1212}; + { + .reg .b32 %dummy; + mov.b64 {%r1214,%dummy}, %rd1830; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1215}, %rd1830; + } + shf.r.wrap.b32 %r1216, %r1215, %r1214, 8; + shf.r.wrap.b32 %r1217, %r1214, %r1215, 8; + mov.b64 %rd1850, {%r1217, %r1216}; + { + .reg .b32 %dummy; + mov.b64 {%r1218,%dummy}, %rd1809; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1219}, %rd1809; + } + shf.r.wrap.b32 %r1220, %r1219, %r1218, 2; + shf.r.wrap.b32 %r1221, %r1218, %r1219, 2; + mov.b64 %rd1851, {%r1221, %r1220}; + { + .reg .b32 %dummy; + mov.b64 {%r1222,%dummy}, %rd1815; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1223}, %rd1815; + } + shf.r.wrap.b32 %r1224, %r1223, %r1222, 9; + shf.r.wrap.b32 %r1225, %r1222, %r1223, 9; + mov.b64 %rd1852, {%r1225, %r1224}; + { + .reg .b32 %dummy; + mov.b64 {%r1226,%dummy}, %rd1821; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1227}, %rd1821; + } + shf.r.wrap.b32 %r1228, %r1227, %r1226, 25; + shf.r.wrap.b32 %r1229, %r1226, %r1227, 25; + mov.b64 %rd1853, {%r1229, %r1228}; + { + .reg .b32 %dummy; + mov.b64 {%r1230,%dummy}, %rd1822; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1231}, %rd1822; + } + shf.r.wrap.b32 %r1232, %r1231, %r1230, 23; + shf.r.wrap.b32 %r1233, %r1230, %r1231, 23; + mov.b64 %rd1854, {%r1233, %r1232}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1234}, %rd1828; + } + { + .reg .b32 %dummy; + mov.b64 {%r1235,%dummy}, %rd1828; + } + shf.l.wrap.b32 %r1236, %r1235, %r1234, 2; + shf.l.wrap.b32 %r1237, %r1234, %r1235, 2; + mov.b64 %rd1855, {%r1237, %r1236}; + not.b64 %rd1856, %rd1832; + and.b64 %rd1857, %rd1833, %rd1856; + xor.b64 %rd1858, %rd1857, %rd1807; + not.b64 %rd1859, %rd1833; + and.b64 %rd1860, %rd1834, %rd1859; + xor.b64 %rd1861, %rd1860, %rd1832; + not.b64 %rd1862, %rd1834; + and.b64 %rd1863, %rd1835, %rd1862; + xor.b64 %rd1864, %rd1863, %rd1833; + not.b64 %rd1865, %rd1835; + and.b64 %rd1866, %rd1807, %rd1865; + xor.b64 %rd1867, %rd1866, %rd1834; + not.b64 %rd1868, %rd1807; + and.b64 %rd1869, %rd1832, %rd1868; + xor.b64 %rd1870, %rd1869, %rd1835; + not.b64 %rd1871, %rd1837; + and.b64 %rd1872, %rd1838, %rd1871; + xor.b64 %rd1873, %rd1872, %rd1836; + not.b64 %rd1874, %rd1838; + and.b64 %rd1875, %rd1839, %rd1874; + xor.b64 %rd1876, %rd1875, %rd1837; + not.b64 %rd1877, %rd1839; + and.b64 %rd1878, %rd1840, %rd1877; + xor.b64 %rd1879, %rd1878, %rd1838; + not.b64 %rd1880, %rd1840; + and.b64 %rd1881, %rd1836, %rd1880; + xor.b64 %rd1882, %rd1881, %rd1839; + not.b64 %rd1883, %rd1836; + and.b64 %rd1884, %rd1837, %rd1883; + xor.b64 %rd1885, %rd1884, %rd1840; + not.b64 %rd1886, %rd1842; + and.b64 %rd1887, %rd1843, %rd1886; + xor.b64 %rd1888, %rd1887, %rd1841; + not.b64 %rd1889, %rd1843; + and.b64 %rd1890, %rd1844, %rd1889; + xor.b64 %rd1891, %rd1890, %rd1842; + not.b64 %rd1892, %rd1844; + and.b64 %rd1893, %rd1845, %rd1892; + xor.b64 %rd1894, %rd1893, %rd1843; + not.b64 %rd1895, %rd1845; + and.b64 %rd1896, %rd1841, %rd1895; + xor.b64 %rd1897, %rd1896, %rd1844; + not.b64 %rd1898, %rd1841; + and.b64 %rd1899, %rd1842, %rd1898; + xor.b64 %rd1900, %rd1899, %rd1845; + not.b64 %rd1901, %rd1847; + and.b64 %rd1902, %rd1848, %rd1901; + xor.b64 %rd1903, %rd1902, %rd1846; + not.b64 %rd1904, %rd1848; + and.b64 %rd1905, %rd1849, %rd1904; + xor.b64 %rd1906, %rd1905, %rd1847; + not.b64 %rd1907, %rd1849; + and.b64 %rd1908, %rd1850, %rd1907; + xor.b64 %rd1909, %rd1908, %rd1848; + not.b64 %rd1910, %rd1850; + and.b64 %rd1911, %rd1846, %rd1910; + xor.b64 %rd1912, %rd1911, %rd1849; + not.b64 %rd1913, %rd1846; + and.b64 %rd1914, %rd1847, %rd1913; + xor.b64 %rd1915, %rd1914, %rd1850; + not.b64 %rd1916, %rd1852; + and.b64 %rd1917, %rd1853, %rd1916; + xor.b64 %rd1918, %rd1917, %rd1851; + not.b64 %rd1919, %rd1853; + and.b64 %rd1920, %rd1854, %rd1919; + xor.b64 %rd1921, %rd1920, %rd1852; + not.b64 %rd1922, %rd1854; + and.b64 %rd1923, %rd1855, %rd1922; + xor.b64 %rd1924, %rd1923, %rd1853; + not.b64 %rd1925, %rd1855; + and.b64 %rd1926, %rd1851, %rd1925; + xor.b64 %rd1927, %rd1926, %rd1854; + not.b64 %rd1928, %rd1851; + and.b64 %rd1929, %rd1852, %rd1928; + xor.b64 %rd1930, %rd1929, %rd1855; + xor.b64 %rd1931, %rd1858, %rd25; + xor.b64 %rd1932, %rd1873, %rd1931; + xor.b64 %rd1933, %rd1932, %rd1888; + xor.b64 %rd1934, %rd1933, %rd1903; + xor.b64 %rd1935, %rd1934, %rd1918; + xor.b64 %rd1936, %rd1876, %rd1861; + xor.b64 %rd1937, %rd1936, %rd1891; + xor.b64 %rd1938, %rd1937, %rd1906; + xor.b64 %rd1939, %rd1938, %rd1921; + xor.b64 %rd1940, %rd1879, %rd1864; + xor.b64 %rd1941, %rd1940, %rd1894; + xor.b64 %rd1942, %rd1941, %rd1909; + xor.b64 %rd1943, %rd1942, %rd1924; + xor.b64 %rd1944, %rd1882, %rd1867; + xor.b64 %rd1945, %rd1944, %rd1897; + xor.b64 %rd1946, %rd1945, %rd1912; + xor.b64 %rd1947, %rd1946, %rd1927; + xor.b64 %rd1948, %rd1885, %rd1870; + xor.b64 %rd1949, %rd1948, %rd1900; + xor.b64 %rd1950, %rd1949, %rd1915; + xor.b64 %rd1951, %rd1950, %rd1930; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1238}, %rd1939; + } + { + .reg .b32 %dummy; + mov.b64 {%r1239,%dummy}, %rd1939; + } + shf.l.wrap.b32 %r1240, %r1239, %r1238, 1; + shf.l.wrap.b32 %r1241, %r1238, %r1239, 1; + mov.b64 %rd1952, {%r1241, %r1240}; + xor.b64 %rd1953, %rd1951, %rd1952; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1242}, %rd1943; + } + { + .reg .b32 %dummy; + mov.b64 {%r1243,%dummy}, %rd1943; + } + shf.l.wrap.b32 %r1244, %r1243, %r1242, 1; + shf.l.wrap.b32 %r1245, %r1242, %r1243, 1; + mov.b64 %rd1954, {%r1245, %r1244}; + xor.b64 %rd1955, %rd1954, %rd1935; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1246}, %rd1947; + } + { + .reg .b32 %dummy; + mov.b64 {%r1247,%dummy}, %rd1947; + } + shf.l.wrap.b32 %r1248, %r1247, %r1246, 1; + shf.l.wrap.b32 %r1249, %r1246, %r1247, 1; + mov.b64 %rd1956, {%r1249, %r1248}; + xor.b64 %rd1957, %rd1956, %rd1939; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1250}, %rd1951; + } + { + .reg .b32 %dummy; + mov.b64 {%r1251,%dummy}, %rd1951; + } + shf.l.wrap.b32 %r1252, %r1251, %r1250, 1; + shf.l.wrap.b32 %r1253, %r1250, %r1251, 1; + mov.b64 %rd1958, {%r1253, %r1252}; + xor.b64 %rd1959, %rd1958, %rd1943; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1254}, %rd1935; + } + { + .reg .b32 %dummy; + mov.b64 {%r1255,%dummy}, %rd1935; + } + shf.l.wrap.b32 %r1256, %r1255, %r1254, 1; + shf.l.wrap.b32 %r1257, %r1254, %r1255, 1; + mov.b64 %rd1960, {%r1257, %r1256}; + xor.b64 %rd1961, %rd1947, %rd1960; + xor.b64 %rd1962, %rd1953, %rd1931; + xor.b64 %rd1963, %rd1955, %rd1861; + xor.b64 %rd1964, %rd1957, %rd1864; + xor.b64 %rd1965, %rd1959, %rd1867; + xor.b64 %rd1966, %rd1961, %rd1870; + xor.b64 %rd1967, %rd1953, %rd1873; + xor.b64 %rd1968, %rd1955, %rd1876; + xor.b64 %rd1969, %rd1957, %rd1879; + xor.b64 %rd1970, %rd1959, %rd1882; + xor.b64 %rd1971, %rd1961, %rd1885; + xor.b64 %rd1972, %rd1953, %rd1888; + xor.b64 %rd1973, %rd1955, %rd1891; + xor.b64 %rd1974, %rd1957, %rd1894; + xor.b64 %rd1975, %rd1959, %rd1897; + xor.b64 %rd1976, %rd1961, %rd1900; + xor.b64 %rd1977, %rd1953, %rd1903; + xor.b64 %rd1978, %rd1955, %rd1906; + xor.b64 %rd1979, %rd1957, %rd1909; + xor.b64 %rd1980, %rd1959, %rd1912; + xor.b64 %rd1981, %rd1961, %rd1915; + xor.b64 %rd1982, %rd1953, %rd1918; + xor.b64 %rd1983, %rd1955, %rd1921; + xor.b64 %rd1984, %rd1957, %rd1924; + xor.b64 %rd1985, %rd1959, %rd1927; + xor.b64 %rd1986, %rd1961, %rd1930; + { + .reg .b32 %dummy; + mov.b64 {%r1258,%dummy}, %rd1968; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1259}, %rd1968; + } + shf.r.wrap.b32 %r1260, %r1259, %r1258, 20; + shf.r.wrap.b32 %r1261, %r1258, %r1259, 20; + mov.b64 %rd1987, {%r1261, %r1260}; + { + .reg .b32 %dummy; + mov.b64 {%r1262,%dummy}, %rd1974; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1263}, %rd1974; + } + shf.r.wrap.b32 %r1264, %r1263, %r1262, 21; + shf.r.wrap.b32 %r1265, %r1262, %r1263, 21; + mov.b64 %rd1988, {%r1265, %r1264}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1266}, %rd1980; + } + { + .reg .b32 %dummy; + mov.b64 {%r1267,%dummy}, %rd1980; + } + shf.l.wrap.b32 %r1268, %r1267, %r1266, 21; + shf.l.wrap.b32 %r1269, %r1266, %r1267, 21; + mov.b64 %rd1989, {%r1269, %r1268}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1270}, %rd1986; + } + { + .reg .b32 %dummy; + mov.b64 {%r1271,%dummy}, %rd1986; + } + shf.l.wrap.b32 %r1272, %r1271, %r1270, 14; + shf.l.wrap.b32 %r1273, %r1270, %r1271, 14; + mov.b64 %rd1990, {%r1273, %r1272}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1274}, %rd1965; + } + { + .reg .b32 %dummy; + mov.b64 {%r1275,%dummy}, %rd1965; + } + shf.l.wrap.b32 %r1276, %r1275, %r1274, 28; + shf.l.wrap.b32 %r1277, %r1274, %r1275, 28; + mov.b64 %rd1991, {%r1277, %r1276}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1278}, %rd1971; + } + { + .reg .b32 %dummy; + mov.b64 {%r1279,%dummy}, %rd1971; + } + shf.l.wrap.b32 %r1280, %r1279, %r1278, 20; + shf.l.wrap.b32 %r1281, %r1278, %r1279, 20; + mov.b64 %rd1992, {%r1281, %r1280}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1282}, %rd1972; + } + { + .reg .b32 %dummy; + mov.b64 {%r1283,%dummy}, %rd1972; + } + shf.l.wrap.b32 %r1284, %r1283, %r1282, 3; + shf.l.wrap.b32 %r1285, %r1282, %r1283, 3; + mov.b64 %rd1993, {%r1285, %r1284}; + { + .reg .b32 %dummy; + mov.b64 {%r1286,%dummy}, %rd1978; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1287}, %rd1978; + } + shf.r.wrap.b32 %r1288, %r1287, %r1286, 19; + shf.r.wrap.b32 %r1289, %r1286, %r1287, 19; + mov.b64 %rd1994, {%r1289, %r1288}; + { + .reg .b32 %dummy; + mov.b64 {%r1290,%dummy}, %rd1984; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1291}, %rd1984; + } + shf.r.wrap.b32 %r1292, %r1291, %r1290, 3; + shf.r.wrap.b32 %r1293, %r1290, %r1291, 3; + mov.b64 %rd1995, {%r1293, %r1292}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1294}, %rd1963; + } + { + .reg .b32 %dummy; + mov.b64 {%r1295,%dummy}, %rd1963; + } + shf.l.wrap.b32 %r1296, %r1295, %r1294, 1; + shf.l.wrap.b32 %r1297, %r1294, %r1295, 1; + mov.b64 %rd1996, {%r1297, %r1296}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1298}, %rd1969; + } + { + .reg .b32 %dummy; + mov.b64 {%r1299,%dummy}, %rd1969; + } + shf.l.wrap.b32 %r1300, %r1299, %r1298, 6; + shf.l.wrap.b32 %r1301, %r1298, %r1299, 6; + mov.b64 %rd1997, {%r1301, %r1300}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1302}, %rd1975; + } + { + .reg .b32 %dummy; + mov.b64 {%r1303,%dummy}, %rd1975; + } + shf.l.wrap.b32 %r1304, %r1303, %r1302, 25; + shf.l.wrap.b32 %r1305, %r1302, %r1303, 25; + mov.b64 %rd1998, {%r1305, %r1304}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1306}, %rd1981; + } + { + .reg .b32 %dummy; + mov.b64 {%r1307,%dummy}, %rd1981; + } + shf.l.wrap.b32 %r1308, %r1307, %r1306, 8; + shf.l.wrap.b32 %r1309, %r1306, %r1307, 8; + mov.b64 %rd1999, {%r1309, %r1308}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1310}, %rd1982; + } + { + .reg .b32 %dummy; + mov.b64 {%r1311,%dummy}, %rd1982; + } + shf.l.wrap.b32 %r1312, %r1311, %r1310, 18; + shf.l.wrap.b32 %r1313, %r1310, %r1311, 18; + mov.b64 %rd2000, {%r1313, %r1312}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1314}, %rd1966; + } + { + .reg .b32 %dummy; + mov.b64 {%r1315,%dummy}, %rd1966; + } + shf.l.wrap.b32 %r1316, %r1315, %r1314, 27; + shf.l.wrap.b32 %r1317, %r1314, %r1315, 27; + mov.b64 %rd2001, {%r1317, %r1316}; + { + .reg .b32 %dummy; + mov.b64 {%r1318,%dummy}, %rd1967; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1319}, %rd1967; + } + shf.r.wrap.b32 %r1320, %r1319, %r1318, 28; + shf.r.wrap.b32 %r1321, %r1318, %r1319, 28; + mov.b64 %rd2002, {%r1321, %r1320}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1322}, %rd1973; + } + { + .reg .b32 %dummy; + mov.b64 {%r1323,%dummy}, %rd1973; + } + shf.l.wrap.b32 %r1324, %r1323, %r1322, 10; + shf.l.wrap.b32 %r1325, %r1322, %r1323, 10; + mov.b64 %rd2003, {%r1325, %r1324}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1326}, %rd1979; + } + { + .reg .b32 %dummy; + mov.b64 {%r1327,%dummy}, %rd1979; + } + shf.l.wrap.b32 %r1328, %r1327, %r1326, 15; + shf.l.wrap.b32 %r1329, %r1326, %r1327, 15; + mov.b64 %rd2004, {%r1329, %r1328}; + { + .reg .b32 %dummy; + mov.b64 {%r1330,%dummy}, %rd1985; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1331}, %rd1985; + } + shf.r.wrap.b32 %r1332, %r1331, %r1330, 8; + shf.r.wrap.b32 %r1333, %r1330, %r1331, 8; + mov.b64 %rd2005, {%r1333, %r1332}; + { + .reg .b32 %dummy; + mov.b64 {%r1334,%dummy}, %rd1964; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1335}, %rd1964; + } + shf.r.wrap.b32 %r1336, %r1335, %r1334, 2; + shf.r.wrap.b32 %r1337, %r1334, %r1335, 2; + mov.b64 %rd2006, {%r1337, %r1336}; + { + .reg .b32 %dummy; + mov.b64 {%r1338,%dummy}, %rd1970; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1339}, %rd1970; + } + shf.r.wrap.b32 %r1340, %r1339, %r1338, 9; + shf.r.wrap.b32 %r1341, %r1338, %r1339, 9; + mov.b64 %rd2007, {%r1341, %r1340}; + { + .reg .b32 %dummy; + mov.b64 {%r1342,%dummy}, %rd1976; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1343}, %rd1976; + } + shf.r.wrap.b32 %r1344, %r1343, %r1342, 25; + shf.r.wrap.b32 %r1345, %r1342, %r1343, 25; + mov.b64 %rd2008, {%r1345, %r1344}; + { + .reg .b32 %dummy; + mov.b64 {%r1346,%dummy}, %rd1977; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1347}, %rd1977; + } + shf.r.wrap.b32 %r1348, %r1347, %r1346, 23; + shf.r.wrap.b32 %r1349, %r1346, %r1347, 23; + mov.b64 %rd2009, {%r1349, %r1348}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1350}, %rd1983; + } + { + .reg .b32 %dummy; + mov.b64 {%r1351,%dummy}, %rd1983; + } + shf.l.wrap.b32 %r1352, %r1351, %r1350, 2; + shf.l.wrap.b32 %r1353, %r1350, %r1351, 2; + mov.b64 %rd2010, {%r1353, %r1352}; + not.b64 %rd2011, %rd1987; + and.b64 %rd2012, %rd1988, %rd2011; + xor.b64 %rd2013, %rd2012, %rd1962; + not.b64 %rd2014, %rd1988; + and.b64 %rd2015, %rd1989, %rd2014; + xor.b64 %rd2016, %rd2015, %rd1987; + not.b64 %rd2017, %rd1989; + and.b64 %rd2018, %rd1990, %rd2017; + xor.b64 %rd2019, %rd2018, %rd1988; + not.b64 %rd2020, %rd1990; + and.b64 %rd2021, %rd1962, %rd2020; + xor.b64 %rd2022, %rd2021, %rd1989; + not.b64 %rd2023, %rd1962; + and.b64 %rd2024, %rd1987, %rd2023; + xor.b64 %rd2025, %rd2024, %rd1990; + not.b64 %rd2026, %rd1992; + and.b64 %rd2027, %rd1993, %rd2026; + xor.b64 %rd2028, %rd2027, %rd1991; + not.b64 %rd2029, %rd1993; + and.b64 %rd2030, %rd1994, %rd2029; + xor.b64 %rd2031, %rd2030, %rd1992; + not.b64 %rd2032, %rd1994; + and.b64 %rd2033, %rd1995, %rd2032; + xor.b64 %rd2034, %rd2033, %rd1993; + not.b64 %rd2035, %rd1995; + and.b64 %rd2036, %rd1991, %rd2035; + xor.b64 %rd2037, %rd2036, %rd1994; + not.b64 %rd2038, %rd1991; + and.b64 %rd2039, %rd1992, %rd2038; + xor.b64 %rd2040, %rd2039, %rd1995; + not.b64 %rd2041, %rd1997; + and.b64 %rd2042, %rd1998, %rd2041; + xor.b64 %rd2043, %rd2042, %rd1996; + not.b64 %rd2044, %rd1998; + and.b64 %rd2045, %rd1999, %rd2044; + xor.b64 %rd2046, %rd2045, %rd1997; + not.b64 %rd2047, %rd1999; + and.b64 %rd2048, %rd2000, %rd2047; + xor.b64 %rd2049, %rd2048, %rd1998; + not.b64 %rd2050, %rd2000; + and.b64 %rd2051, %rd1996, %rd2050; + xor.b64 %rd2052, %rd2051, %rd1999; + not.b64 %rd2053, %rd1996; + and.b64 %rd2054, %rd1997, %rd2053; + xor.b64 %rd2055, %rd2054, %rd2000; + not.b64 %rd2056, %rd2002; + and.b64 %rd2057, %rd2003, %rd2056; + xor.b64 %rd2058, %rd2057, %rd2001; + not.b64 %rd2059, %rd2003; + and.b64 %rd2060, %rd2004, %rd2059; + xor.b64 %rd2061, %rd2060, %rd2002; + not.b64 %rd2062, %rd2004; + and.b64 %rd2063, %rd2005, %rd2062; + xor.b64 %rd2064, %rd2063, %rd2003; + not.b64 %rd2065, %rd2005; + and.b64 %rd2066, %rd2001, %rd2065; + xor.b64 %rd2067, %rd2066, %rd2004; + not.b64 %rd2068, %rd2001; + and.b64 %rd2069, %rd2002, %rd2068; + xor.b64 %rd2070, %rd2069, %rd2005; + not.b64 %rd2071, %rd2007; + and.b64 %rd2072, %rd2008, %rd2071; + xor.b64 %rd2073, %rd2072, %rd2006; + not.b64 %rd2074, %rd2008; + and.b64 %rd2075, %rd2009, %rd2074; + xor.b64 %rd2076, %rd2075, %rd2007; + not.b64 %rd2077, %rd2009; + and.b64 %rd2078, %rd2010, %rd2077; + xor.b64 %rd2079, %rd2078, %rd2008; + not.b64 %rd2080, %rd2010; + and.b64 %rd2081, %rd2006, %rd2080; + xor.b64 %rd2082, %rd2081, %rd2009; + not.b64 %rd2083, %rd2006; + and.b64 %rd2084, %rd2007, %rd2083; + xor.b64 %rd2085, %rd2084, %rd2010; + xor.b64 %rd2086, %rd2013, %rd26; + xor.b64 %rd2087, %rd2028, %rd2086; + xor.b64 %rd2088, %rd2087, %rd2043; + xor.b64 %rd2089, %rd2088, %rd2058; + xor.b64 %rd2090, %rd2089, %rd2073; + xor.b64 %rd2091, %rd2031, %rd2016; + xor.b64 %rd2092, %rd2091, %rd2046; + xor.b64 %rd2093, %rd2092, %rd2061; + xor.b64 %rd2094, %rd2093, %rd2076; + xor.b64 %rd2095, %rd2034, %rd2019; + xor.b64 %rd2096, %rd2095, %rd2049; + xor.b64 %rd2097, %rd2096, %rd2064; + xor.b64 %rd2098, %rd2097, %rd2079; + xor.b64 %rd2099, %rd2037, %rd2022; + xor.b64 %rd2100, %rd2099, %rd2052; + xor.b64 %rd2101, %rd2100, %rd2067; + xor.b64 %rd2102, %rd2101, %rd2082; + xor.b64 %rd2103, %rd2040, %rd2025; + xor.b64 %rd2104, %rd2103, %rd2055; + xor.b64 %rd2105, %rd2104, %rd2070; + xor.b64 %rd2106, %rd2105, %rd2085; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1354}, %rd2094; + } + { + .reg .b32 %dummy; + mov.b64 {%r1355,%dummy}, %rd2094; + } + shf.l.wrap.b32 %r1356, %r1355, %r1354, 1; + shf.l.wrap.b32 %r1357, %r1354, %r1355, 1; + mov.b64 %rd2107, {%r1357, %r1356}; + xor.b64 %rd2108, %rd2106, %rd2107; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1358}, %rd2098; + } + { + .reg .b32 %dummy; + mov.b64 {%r1359,%dummy}, %rd2098; + } + shf.l.wrap.b32 %r1360, %r1359, %r1358, 1; + shf.l.wrap.b32 %r1361, %r1358, %r1359, 1; + mov.b64 %rd2109, {%r1361, %r1360}; + xor.b64 %rd2110, %rd2109, %rd2090; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1362}, %rd2102; + } + { + .reg .b32 %dummy; + mov.b64 {%r1363,%dummy}, %rd2102; + } + shf.l.wrap.b32 %r1364, %r1363, %r1362, 1; + shf.l.wrap.b32 %r1365, %r1362, %r1363, 1; + mov.b64 %rd2111, {%r1365, %r1364}; + xor.b64 %rd2112, %rd2111, %rd2094; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1366}, %rd2106; + } + { + .reg .b32 %dummy; + mov.b64 {%r1367,%dummy}, %rd2106; + } + shf.l.wrap.b32 %r1368, %r1367, %r1366, 1; + shf.l.wrap.b32 %r1369, %r1366, %r1367, 1; + mov.b64 %rd2113, {%r1369, %r1368}; + xor.b64 %rd2114, %rd2113, %rd2098; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1370}, %rd2090; + } + { + .reg .b32 %dummy; + mov.b64 {%r1371,%dummy}, %rd2090; + } + shf.l.wrap.b32 %r1372, %r1371, %r1370, 1; + shf.l.wrap.b32 %r1373, %r1370, %r1371, 1; + mov.b64 %rd2115, {%r1373, %r1372}; + xor.b64 %rd2116, %rd2102, %rd2115; + xor.b64 %rd2117, %rd2108, %rd2086; + xor.b64 %rd2118, %rd2110, %rd2016; + xor.b64 %rd2119, %rd2112, %rd2019; + xor.b64 %rd2120, %rd2114, %rd2022; + xor.b64 %rd2121, %rd2116, %rd2025; + xor.b64 %rd2122, %rd2108, %rd2028; + xor.b64 %rd2123, %rd2110, %rd2031; + xor.b64 %rd2124, %rd2112, %rd2034; + xor.b64 %rd2125, %rd2114, %rd2037; + xor.b64 %rd2126, %rd2116, %rd2040; + xor.b64 %rd2127, %rd2108, %rd2043; + xor.b64 %rd2128, %rd2110, %rd2046; + xor.b64 %rd2129, %rd2112, %rd2049; + xor.b64 %rd2130, %rd2114, %rd2052; + xor.b64 %rd2131, %rd2116, %rd2055; + xor.b64 %rd2132, %rd2108, %rd2058; + xor.b64 %rd2133, %rd2110, %rd2061; + xor.b64 %rd2134, %rd2112, %rd2064; + xor.b64 %rd2135, %rd2114, %rd2067; + xor.b64 %rd2136, %rd2116, %rd2070; + xor.b64 %rd2137, %rd2108, %rd2073; + xor.b64 %rd2138, %rd2110, %rd2076; + xor.b64 %rd2139, %rd2112, %rd2079; + xor.b64 %rd2140, %rd2114, %rd2082; + xor.b64 %rd2141, %rd2116, %rd2085; + { + .reg .b32 %dummy; + mov.b64 {%r1374,%dummy}, %rd2123; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1375}, %rd2123; + } + shf.r.wrap.b32 %r1376, %r1375, %r1374, 20; + shf.r.wrap.b32 %r1377, %r1374, %r1375, 20; + mov.b64 %rd2142, {%r1377, %r1376}; + { + .reg .b32 %dummy; + mov.b64 {%r1378,%dummy}, %rd2129; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1379}, %rd2129; + } + shf.r.wrap.b32 %r1380, %r1379, %r1378, 21; + shf.r.wrap.b32 %r1381, %r1378, %r1379, 21; + mov.b64 %rd2143, {%r1381, %r1380}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1382}, %rd2135; + } + { + .reg .b32 %dummy; + mov.b64 {%r1383,%dummy}, %rd2135; + } + shf.l.wrap.b32 %r1384, %r1383, %r1382, 21; + shf.l.wrap.b32 %r1385, %r1382, %r1383, 21; + mov.b64 %rd2144, {%r1385, %r1384}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1386}, %rd2141; + } + { + .reg .b32 %dummy; + mov.b64 {%r1387,%dummy}, %rd2141; + } + shf.l.wrap.b32 %r1388, %r1387, %r1386, 14; + shf.l.wrap.b32 %r1389, %r1386, %r1387, 14; + mov.b64 %rd2145, {%r1389, %r1388}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1390}, %rd2120; + } + { + .reg .b32 %dummy; + mov.b64 {%r1391,%dummy}, %rd2120; + } + shf.l.wrap.b32 %r1392, %r1391, %r1390, 28; + shf.l.wrap.b32 %r1393, %r1390, %r1391, 28; + mov.b64 %rd2146, {%r1393, %r1392}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1394}, %rd2126; + } + { + .reg .b32 %dummy; + mov.b64 {%r1395,%dummy}, %rd2126; + } + shf.l.wrap.b32 %r1396, %r1395, %r1394, 20; + shf.l.wrap.b32 %r1397, %r1394, %r1395, 20; + mov.b64 %rd2147, {%r1397, %r1396}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1398}, %rd2127; + } + { + .reg .b32 %dummy; + mov.b64 {%r1399,%dummy}, %rd2127; + } + shf.l.wrap.b32 %r1400, %r1399, %r1398, 3; + shf.l.wrap.b32 %r1401, %r1398, %r1399, 3; + mov.b64 %rd2148, {%r1401, %r1400}; + { + .reg .b32 %dummy; + mov.b64 {%r1402,%dummy}, %rd2133; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1403}, %rd2133; + } + shf.r.wrap.b32 %r1404, %r1403, %r1402, 19; + shf.r.wrap.b32 %r1405, %r1402, %r1403, 19; + mov.b64 %rd2149, {%r1405, %r1404}; + { + .reg .b32 %dummy; + mov.b64 {%r1406,%dummy}, %rd2139; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1407}, %rd2139; + } + shf.r.wrap.b32 %r1408, %r1407, %r1406, 3; + shf.r.wrap.b32 %r1409, %r1406, %r1407, 3; + mov.b64 %rd2150, {%r1409, %r1408}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1410}, %rd2118; + } + { + .reg .b32 %dummy; + mov.b64 {%r1411,%dummy}, %rd2118; + } + shf.l.wrap.b32 %r1412, %r1411, %r1410, 1; + shf.l.wrap.b32 %r1413, %r1410, %r1411, 1; + mov.b64 %rd2151, {%r1413, %r1412}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1414}, %rd2124; + } + { + .reg .b32 %dummy; + mov.b64 {%r1415,%dummy}, %rd2124; + } + shf.l.wrap.b32 %r1416, %r1415, %r1414, 6; + shf.l.wrap.b32 %r1417, %r1414, %r1415, 6; + mov.b64 %rd2152, {%r1417, %r1416}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1418}, %rd2130; + } + { + .reg .b32 %dummy; + mov.b64 {%r1419,%dummy}, %rd2130; + } + shf.l.wrap.b32 %r1420, %r1419, %r1418, 25; + shf.l.wrap.b32 %r1421, %r1418, %r1419, 25; + mov.b64 %rd2153, {%r1421, %r1420}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1422}, %rd2136; + } + { + .reg .b32 %dummy; + mov.b64 {%r1423,%dummy}, %rd2136; + } + shf.l.wrap.b32 %r1424, %r1423, %r1422, 8; + shf.l.wrap.b32 %r1425, %r1422, %r1423, 8; + mov.b64 %rd2154, {%r1425, %r1424}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1426}, %rd2137; + } + { + .reg .b32 %dummy; + mov.b64 {%r1427,%dummy}, %rd2137; + } + shf.l.wrap.b32 %r1428, %r1427, %r1426, 18; + shf.l.wrap.b32 %r1429, %r1426, %r1427, 18; + mov.b64 %rd2155, {%r1429, %r1428}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1430}, %rd2121; + } + { + .reg .b32 %dummy; + mov.b64 {%r1431,%dummy}, %rd2121; + } + shf.l.wrap.b32 %r1432, %r1431, %r1430, 27; + shf.l.wrap.b32 %r1433, %r1430, %r1431, 27; + mov.b64 %rd2156, {%r1433, %r1432}; + { + .reg .b32 %dummy; + mov.b64 {%r1434,%dummy}, %rd2122; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1435}, %rd2122; + } + shf.r.wrap.b32 %r1436, %r1435, %r1434, 28; + shf.r.wrap.b32 %r1437, %r1434, %r1435, 28; + mov.b64 %rd2157, {%r1437, %r1436}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1438}, %rd2128; + } + { + .reg .b32 %dummy; + mov.b64 {%r1439,%dummy}, %rd2128; + } + shf.l.wrap.b32 %r1440, %r1439, %r1438, 10; + shf.l.wrap.b32 %r1441, %r1438, %r1439, 10; + mov.b64 %rd2158, {%r1441, %r1440}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1442}, %rd2134; + } + { + .reg .b32 %dummy; + mov.b64 {%r1443,%dummy}, %rd2134; + } + shf.l.wrap.b32 %r1444, %r1443, %r1442, 15; + shf.l.wrap.b32 %r1445, %r1442, %r1443, 15; + mov.b64 %rd2159, {%r1445, %r1444}; + { + .reg .b32 %dummy; + mov.b64 {%r1446,%dummy}, %rd2140; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1447}, %rd2140; + } + shf.r.wrap.b32 %r1448, %r1447, %r1446, 8; + shf.r.wrap.b32 %r1449, %r1446, %r1447, 8; + mov.b64 %rd2160, {%r1449, %r1448}; + { + .reg .b32 %dummy; + mov.b64 {%r1450,%dummy}, %rd2119; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1451}, %rd2119; + } + shf.r.wrap.b32 %r1452, %r1451, %r1450, 2; + shf.r.wrap.b32 %r1453, %r1450, %r1451, 2; + mov.b64 %rd2161, {%r1453, %r1452}; + { + .reg .b32 %dummy; + mov.b64 {%r1454,%dummy}, %rd2125; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1455}, %rd2125; + } + shf.r.wrap.b32 %r1456, %r1455, %r1454, 9; + shf.r.wrap.b32 %r1457, %r1454, %r1455, 9; + mov.b64 %rd2162, {%r1457, %r1456}; + { + .reg .b32 %dummy; + mov.b64 {%r1458,%dummy}, %rd2131; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1459}, %rd2131; + } + shf.r.wrap.b32 %r1460, %r1459, %r1458, 25; + shf.r.wrap.b32 %r1461, %r1458, %r1459, 25; + mov.b64 %rd2163, {%r1461, %r1460}; + { + .reg .b32 %dummy; + mov.b64 {%r1462,%dummy}, %rd2132; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1463}, %rd2132; + } + shf.r.wrap.b32 %r1464, %r1463, %r1462, 23; + shf.r.wrap.b32 %r1465, %r1462, %r1463, 23; + mov.b64 %rd2164, {%r1465, %r1464}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1466}, %rd2138; + } + { + .reg .b32 %dummy; + mov.b64 {%r1467,%dummy}, %rd2138; + } + shf.l.wrap.b32 %r1468, %r1467, %r1466, 2; + shf.l.wrap.b32 %r1469, %r1466, %r1467, 2; + mov.b64 %rd2165, {%r1469, %r1468}; + not.b64 %rd2166, %rd2142; + and.b64 %rd2167, %rd2143, %rd2166; + xor.b64 %rd2168, %rd2167, %rd2117; + not.b64 %rd2169, %rd2143; + and.b64 %rd2170, %rd2144, %rd2169; + xor.b64 %rd2171, %rd2170, %rd2142; + not.b64 %rd2172, %rd2144; + and.b64 %rd2173, %rd2145, %rd2172; + xor.b64 %rd2174, %rd2173, %rd2143; + not.b64 %rd2175, %rd2145; + and.b64 %rd2176, %rd2117, %rd2175; + xor.b64 %rd2177, %rd2176, %rd2144; + not.b64 %rd2178, %rd2117; + and.b64 %rd2179, %rd2142, %rd2178; + xor.b64 %rd2180, %rd2179, %rd2145; + not.b64 %rd2181, %rd2147; + and.b64 %rd2182, %rd2148, %rd2181; + xor.b64 %rd2183, %rd2182, %rd2146; + not.b64 %rd2184, %rd2148; + and.b64 %rd2185, %rd2149, %rd2184; + xor.b64 %rd2186, %rd2185, %rd2147; + not.b64 %rd2187, %rd2149; + and.b64 %rd2188, %rd2150, %rd2187; + xor.b64 %rd2189, %rd2188, %rd2148; + not.b64 %rd2190, %rd2150; + and.b64 %rd2191, %rd2146, %rd2190; + xor.b64 %rd2192, %rd2191, %rd2149; + not.b64 %rd2193, %rd2146; + and.b64 %rd2194, %rd2147, %rd2193; + xor.b64 %rd2195, %rd2194, %rd2150; + not.b64 %rd2196, %rd2152; + and.b64 %rd2197, %rd2153, %rd2196; + xor.b64 %rd2198, %rd2197, %rd2151; + not.b64 %rd2199, %rd2153; + and.b64 %rd2200, %rd2154, %rd2199; + xor.b64 %rd2201, %rd2200, %rd2152; + not.b64 %rd2202, %rd2154; + and.b64 %rd2203, %rd2155, %rd2202; + xor.b64 %rd2204, %rd2203, %rd2153; + not.b64 %rd2205, %rd2155; + and.b64 %rd2206, %rd2151, %rd2205; + xor.b64 %rd2207, %rd2206, %rd2154; + not.b64 %rd2208, %rd2151; + and.b64 %rd2209, %rd2152, %rd2208; + xor.b64 %rd2210, %rd2209, %rd2155; + not.b64 %rd2211, %rd2157; + and.b64 %rd2212, %rd2158, %rd2211; + xor.b64 %rd2213, %rd2212, %rd2156; + not.b64 %rd2214, %rd2158; + and.b64 %rd2215, %rd2159, %rd2214; + xor.b64 %rd2216, %rd2215, %rd2157; + not.b64 %rd2217, %rd2159; + and.b64 %rd2218, %rd2160, %rd2217; + xor.b64 %rd2219, %rd2218, %rd2158; + not.b64 %rd2220, %rd2160; + and.b64 %rd2221, %rd2156, %rd2220; + xor.b64 %rd2222, %rd2221, %rd2159; + not.b64 %rd2223, %rd2156; + and.b64 %rd2224, %rd2157, %rd2223; + xor.b64 %rd2225, %rd2224, %rd2160; + not.b64 %rd2226, %rd2162; + and.b64 %rd2227, %rd2163, %rd2226; + xor.b64 %rd2228, %rd2227, %rd2161; + not.b64 %rd2229, %rd2163; + and.b64 %rd2230, %rd2164, %rd2229; + xor.b64 %rd2231, %rd2230, %rd2162; + not.b64 %rd2232, %rd2164; + and.b64 %rd2233, %rd2165, %rd2232; + xor.b64 %rd2234, %rd2233, %rd2163; + not.b64 %rd2235, %rd2165; + and.b64 %rd2236, %rd2161, %rd2235; + xor.b64 %rd2237, %rd2236, %rd2164; + not.b64 %rd2238, %rd2161; + and.b64 %rd2239, %rd2162, %rd2238; + xor.b64 %rd2240, %rd2239, %rd2165; + xor.b64 %rd2241, %rd2168, %rd27; + xor.b64 %rd2242, %rd2183, %rd2241; + xor.b64 %rd2243, %rd2242, %rd2198; + xor.b64 %rd2244, %rd2243, %rd2213; + xor.b64 %rd2245, %rd2244, %rd2228; + xor.b64 %rd2246, %rd2186, %rd2171; + xor.b64 %rd2247, %rd2246, %rd2201; + xor.b64 %rd2248, %rd2247, %rd2216; + xor.b64 %rd2249, %rd2248, %rd2231; + xor.b64 %rd2250, %rd2189, %rd2174; + xor.b64 %rd2251, %rd2250, %rd2204; + xor.b64 %rd2252, %rd2251, %rd2219; + xor.b64 %rd2253, %rd2252, %rd2234; + xor.b64 %rd2254, %rd2192, %rd2177; + xor.b64 %rd2255, %rd2254, %rd2207; + xor.b64 %rd2256, %rd2255, %rd2222; + xor.b64 %rd2257, %rd2256, %rd2237; + xor.b64 %rd2258, %rd2195, %rd2180; + xor.b64 %rd2259, %rd2258, %rd2210; + xor.b64 %rd2260, %rd2259, %rd2225; + xor.b64 %rd2261, %rd2260, %rd2240; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1470}, %rd2249; + } + { + .reg .b32 %dummy; + mov.b64 {%r1471,%dummy}, %rd2249; + } + shf.l.wrap.b32 %r1472, %r1471, %r1470, 1; + shf.l.wrap.b32 %r1473, %r1470, %r1471, 1; + mov.b64 %rd2262, {%r1473, %r1472}; + xor.b64 %rd2263, %rd2261, %rd2262; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1474}, %rd2253; + } + { + .reg .b32 %dummy; + mov.b64 {%r1475,%dummy}, %rd2253; + } + shf.l.wrap.b32 %r1476, %r1475, %r1474, 1; + shf.l.wrap.b32 %r1477, %r1474, %r1475, 1; + mov.b64 %rd2264, {%r1477, %r1476}; + xor.b64 %rd2265, %rd2264, %rd2245; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1478}, %rd2257; + } + { + .reg .b32 %dummy; + mov.b64 {%r1479,%dummy}, %rd2257; + } + shf.l.wrap.b32 %r1480, %r1479, %r1478, 1; + shf.l.wrap.b32 %r1481, %r1478, %r1479, 1; + mov.b64 %rd2266, {%r1481, %r1480}; + xor.b64 %rd2267, %rd2266, %rd2249; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1482}, %rd2261; + } + { + .reg .b32 %dummy; + mov.b64 {%r1483,%dummy}, %rd2261; + } + shf.l.wrap.b32 %r1484, %r1483, %r1482, 1; + shf.l.wrap.b32 %r1485, %r1482, %r1483, 1; + mov.b64 %rd2268, {%r1485, %r1484}; + xor.b64 %rd2269, %rd2268, %rd2253; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1486}, %rd2245; + } + { + .reg .b32 %dummy; + mov.b64 {%r1487,%dummy}, %rd2245; + } + shf.l.wrap.b32 %r1488, %r1487, %r1486, 1; + shf.l.wrap.b32 %r1489, %r1486, %r1487, 1; + mov.b64 %rd2270, {%r1489, %r1488}; + xor.b64 %rd2271, %rd2257, %rd2270; + xor.b64 %rd2272, %rd2263, %rd2241; + xor.b64 %rd2273, %rd2265, %rd2171; + xor.b64 %rd2274, %rd2267, %rd2174; + xor.b64 %rd2275, %rd2269, %rd2177; + xor.b64 %rd2276, %rd2271, %rd2180; + xor.b64 %rd2277, %rd2263, %rd2183; + xor.b64 %rd2278, %rd2265, %rd2186; + xor.b64 %rd2279, %rd2267, %rd2189; + xor.b64 %rd2280, %rd2269, %rd2192; + xor.b64 %rd2281, %rd2271, %rd2195; + xor.b64 %rd2282, %rd2263, %rd2198; + xor.b64 %rd2283, %rd2265, %rd2201; + xor.b64 %rd2284, %rd2267, %rd2204; + xor.b64 %rd2285, %rd2269, %rd2207; + xor.b64 %rd2286, %rd2271, %rd2210; + xor.b64 %rd2287, %rd2263, %rd2213; + xor.b64 %rd2288, %rd2265, %rd2216; + xor.b64 %rd2289, %rd2267, %rd2219; + xor.b64 %rd2290, %rd2269, %rd2222; + xor.b64 %rd2291, %rd2271, %rd2225; + xor.b64 %rd2292, %rd2263, %rd2228; + xor.b64 %rd2293, %rd2265, %rd2231; + xor.b64 %rd2294, %rd2267, %rd2234; + xor.b64 %rd2295, %rd2269, %rd2237; + xor.b64 %rd2296, %rd2271, %rd2240; + { + .reg .b32 %dummy; + mov.b64 {%r1490,%dummy}, %rd2278; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1491}, %rd2278; + } + shf.r.wrap.b32 %r1492, %r1491, %r1490, 20; + shf.r.wrap.b32 %r1493, %r1490, %r1491, 20; + mov.b64 %rd2297, {%r1493, %r1492}; + { + .reg .b32 %dummy; + mov.b64 {%r1494,%dummy}, %rd2284; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1495}, %rd2284; + } + shf.r.wrap.b32 %r1496, %r1495, %r1494, 21; + shf.r.wrap.b32 %r1497, %r1494, %r1495, 21; + mov.b64 %rd2298, {%r1497, %r1496}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1498}, %rd2290; + } + { + .reg .b32 %dummy; + mov.b64 {%r1499,%dummy}, %rd2290; + } + shf.l.wrap.b32 %r1500, %r1499, %r1498, 21; + shf.l.wrap.b32 %r1501, %r1498, %r1499, 21; + mov.b64 %rd2299, {%r1501, %r1500}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1502}, %rd2296; + } + { + .reg .b32 %dummy; + mov.b64 {%r1503,%dummy}, %rd2296; + } + shf.l.wrap.b32 %r1504, %r1503, %r1502, 14; + shf.l.wrap.b32 %r1505, %r1502, %r1503, 14; + mov.b64 %rd2300, {%r1505, %r1504}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1506}, %rd2275; + } + { + .reg .b32 %dummy; + mov.b64 {%r1507,%dummy}, %rd2275; + } + shf.l.wrap.b32 %r1508, %r1507, %r1506, 28; + shf.l.wrap.b32 %r1509, %r1506, %r1507, 28; + mov.b64 %rd2301, {%r1509, %r1508}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1510}, %rd2281; + } + { + .reg .b32 %dummy; + mov.b64 {%r1511,%dummy}, %rd2281; + } + shf.l.wrap.b32 %r1512, %r1511, %r1510, 20; + shf.l.wrap.b32 %r1513, %r1510, %r1511, 20; + mov.b64 %rd2302, {%r1513, %r1512}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1514}, %rd2282; + } + { + .reg .b32 %dummy; + mov.b64 {%r1515,%dummy}, %rd2282; + } + shf.l.wrap.b32 %r1516, %r1515, %r1514, 3; + shf.l.wrap.b32 %r1517, %r1514, %r1515, 3; + mov.b64 %rd2303, {%r1517, %r1516}; + { + .reg .b32 %dummy; + mov.b64 {%r1518,%dummy}, %rd2288; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1519}, %rd2288; + } + shf.r.wrap.b32 %r1520, %r1519, %r1518, 19; + shf.r.wrap.b32 %r1521, %r1518, %r1519, 19; + mov.b64 %rd2304, {%r1521, %r1520}; + { + .reg .b32 %dummy; + mov.b64 {%r1522,%dummy}, %rd2294; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1523}, %rd2294; + } + shf.r.wrap.b32 %r1524, %r1523, %r1522, 3; + shf.r.wrap.b32 %r1525, %r1522, %r1523, 3; + mov.b64 %rd2305, {%r1525, %r1524}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1526}, %rd2273; + } + { + .reg .b32 %dummy; + mov.b64 {%r1527,%dummy}, %rd2273; + } + shf.l.wrap.b32 %r1528, %r1527, %r1526, 1; + shf.l.wrap.b32 %r1529, %r1526, %r1527, 1; + mov.b64 %rd2306, {%r1529, %r1528}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1530}, %rd2279; + } + { + .reg .b32 %dummy; + mov.b64 {%r1531,%dummy}, %rd2279; + } + shf.l.wrap.b32 %r1532, %r1531, %r1530, 6; + shf.l.wrap.b32 %r1533, %r1530, %r1531, 6; + mov.b64 %rd2307, {%r1533, %r1532}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1534}, %rd2285; + } + { + .reg .b32 %dummy; + mov.b64 {%r1535,%dummy}, %rd2285; + } + shf.l.wrap.b32 %r1536, %r1535, %r1534, 25; + shf.l.wrap.b32 %r1537, %r1534, %r1535, 25; + mov.b64 %rd2308, {%r1537, %r1536}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1538}, %rd2291; + } + { + .reg .b32 %dummy; + mov.b64 {%r1539,%dummy}, %rd2291; + } + shf.l.wrap.b32 %r1540, %r1539, %r1538, 8; + shf.l.wrap.b32 %r1541, %r1538, %r1539, 8; + mov.b64 %rd2309, {%r1541, %r1540}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1542}, %rd2292; + } + { + .reg .b32 %dummy; + mov.b64 {%r1543,%dummy}, %rd2292; + } + shf.l.wrap.b32 %r1544, %r1543, %r1542, 18; + shf.l.wrap.b32 %r1545, %r1542, %r1543, 18; + mov.b64 %rd2310, {%r1545, %r1544}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1546}, %rd2276; + } + { + .reg .b32 %dummy; + mov.b64 {%r1547,%dummy}, %rd2276; + } + shf.l.wrap.b32 %r1548, %r1547, %r1546, 27; + shf.l.wrap.b32 %r1549, %r1546, %r1547, 27; + mov.b64 %rd2311, {%r1549, %r1548}; + { + .reg .b32 %dummy; + mov.b64 {%r1550,%dummy}, %rd2277; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1551}, %rd2277; + } + shf.r.wrap.b32 %r1552, %r1551, %r1550, 28; + shf.r.wrap.b32 %r1553, %r1550, %r1551, 28; + mov.b64 %rd2312, {%r1553, %r1552}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1554}, %rd2283; + } + { + .reg .b32 %dummy; + mov.b64 {%r1555,%dummy}, %rd2283; + } + shf.l.wrap.b32 %r1556, %r1555, %r1554, 10; + shf.l.wrap.b32 %r1557, %r1554, %r1555, 10; + mov.b64 %rd2313, {%r1557, %r1556}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1558}, %rd2289; + } + { + .reg .b32 %dummy; + mov.b64 {%r1559,%dummy}, %rd2289; + } + shf.l.wrap.b32 %r1560, %r1559, %r1558, 15; + shf.l.wrap.b32 %r1561, %r1558, %r1559, 15; + mov.b64 %rd2314, {%r1561, %r1560}; + { + .reg .b32 %dummy; + mov.b64 {%r1562,%dummy}, %rd2295; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1563}, %rd2295; + } + shf.r.wrap.b32 %r1564, %r1563, %r1562, 8; + shf.r.wrap.b32 %r1565, %r1562, %r1563, 8; + mov.b64 %rd2315, {%r1565, %r1564}; + { + .reg .b32 %dummy; + mov.b64 {%r1566,%dummy}, %rd2274; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1567}, %rd2274; + } + shf.r.wrap.b32 %r1568, %r1567, %r1566, 2; + shf.r.wrap.b32 %r1569, %r1566, %r1567, 2; + mov.b64 %rd2316, {%r1569, %r1568}; + { + .reg .b32 %dummy; + mov.b64 {%r1570,%dummy}, %rd2280; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1571}, %rd2280; + } + shf.r.wrap.b32 %r1572, %r1571, %r1570, 9; + shf.r.wrap.b32 %r1573, %r1570, %r1571, 9; + mov.b64 %rd2317, {%r1573, %r1572}; + { + .reg .b32 %dummy; + mov.b64 {%r1574,%dummy}, %rd2286; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1575}, %rd2286; + } + shf.r.wrap.b32 %r1576, %r1575, %r1574, 25; + shf.r.wrap.b32 %r1577, %r1574, %r1575, 25; + mov.b64 %rd2318, {%r1577, %r1576}; + { + .reg .b32 %dummy; + mov.b64 {%r1578,%dummy}, %rd2287; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1579}, %rd2287; + } + shf.r.wrap.b32 %r1580, %r1579, %r1578, 23; + shf.r.wrap.b32 %r1581, %r1578, %r1579, 23; + mov.b64 %rd2319, {%r1581, %r1580}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1582}, %rd2293; + } + { + .reg .b32 %dummy; + mov.b64 {%r1583,%dummy}, %rd2293; + } + shf.l.wrap.b32 %r1584, %r1583, %r1582, 2; + shf.l.wrap.b32 %r1585, %r1582, %r1583, 2; + mov.b64 %rd2320, {%r1585, %r1584}; + not.b64 %rd2321, %rd2297; + and.b64 %rd2322, %rd2298, %rd2321; + xor.b64 %rd2323, %rd2322, %rd2272; + not.b64 %rd2324, %rd2298; + and.b64 %rd2325, %rd2299, %rd2324; + xor.b64 %rd2326, %rd2325, %rd2297; + not.b64 %rd2327, %rd2299; + and.b64 %rd2328, %rd2300, %rd2327; + xor.b64 %rd2329, %rd2328, %rd2298; + not.b64 %rd2330, %rd2300; + and.b64 %rd2331, %rd2272, %rd2330; + xor.b64 %rd2332, %rd2331, %rd2299; + not.b64 %rd2333, %rd2272; + and.b64 %rd2334, %rd2297, %rd2333; + xor.b64 %rd2335, %rd2334, %rd2300; + not.b64 %rd2336, %rd2302; + and.b64 %rd2337, %rd2303, %rd2336; + xor.b64 %rd2338, %rd2337, %rd2301; + not.b64 %rd2339, %rd2303; + and.b64 %rd2340, %rd2304, %rd2339; + xor.b64 %rd2341, %rd2340, %rd2302; + not.b64 %rd2342, %rd2304; + and.b64 %rd2343, %rd2305, %rd2342; + xor.b64 %rd2344, %rd2343, %rd2303; + not.b64 %rd2345, %rd2305; + and.b64 %rd2346, %rd2301, %rd2345; + xor.b64 %rd2347, %rd2346, %rd2304; + not.b64 %rd2348, %rd2301; + and.b64 %rd2349, %rd2302, %rd2348; + xor.b64 %rd2350, %rd2349, %rd2305; + not.b64 %rd2351, %rd2307; + and.b64 %rd2352, %rd2308, %rd2351; + xor.b64 %rd2353, %rd2352, %rd2306; + not.b64 %rd2354, %rd2308; + and.b64 %rd2355, %rd2309, %rd2354; + xor.b64 %rd2356, %rd2355, %rd2307; + not.b64 %rd2357, %rd2309; + and.b64 %rd2358, %rd2310, %rd2357; + xor.b64 %rd2359, %rd2358, %rd2308; + not.b64 %rd2360, %rd2310; + and.b64 %rd2361, %rd2306, %rd2360; + xor.b64 %rd2362, %rd2361, %rd2309; + not.b64 %rd2363, %rd2306; + and.b64 %rd2364, %rd2307, %rd2363; + xor.b64 %rd2365, %rd2364, %rd2310; + not.b64 %rd2366, %rd2312; + and.b64 %rd2367, %rd2313, %rd2366; + xor.b64 %rd2368, %rd2367, %rd2311; + not.b64 %rd2369, %rd2313; + and.b64 %rd2370, %rd2314, %rd2369; + xor.b64 %rd2371, %rd2370, %rd2312; + not.b64 %rd2372, %rd2314; + and.b64 %rd2373, %rd2315, %rd2372; + xor.b64 %rd2374, %rd2373, %rd2313; + not.b64 %rd2375, %rd2315; + and.b64 %rd2376, %rd2311, %rd2375; + xor.b64 %rd2377, %rd2376, %rd2314; + not.b64 %rd2378, %rd2311; + and.b64 %rd2379, %rd2312, %rd2378; + xor.b64 %rd2380, %rd2379, %rd2315; + not.b64 %rd2381, %rd2317; + and.b64 %rd2382, %rd2318, %rd2381; + xor.b64 %rd2383, %rd2382, %rd2316; + not.b64 %rd2384, %rd2318; + and.b64 %rd2385, %rd2319, %rd2384; + xor.b64 %rd2386, %rd2385, %rd2317; + not.b64 %rd2387, %rd2319; + and.b64 %rd2388, %rd2320, %rd2387; + xor.b64 %rd2389, %rd2388, %rd2318; + not.b64 %rd2390, %rd2320; + and.b64 %rd2391, %rd2316, %rd2390; + xor.b64 %rd2392, %rd2391, %rd2319; + not.b64 %rd2393, %rd2316; + and.b64 %rd2394, %rd2317, %rd2393; + xor.b64 %rd2395, %rd2394, %rd2320; + xor.b64 %rd2396, %rd2323, %rd28; + xor.b64 %rd2397, %rd2338, %rd2396; + xor.b64 %rd2398, %rd2397, %rd2353; + xor.b64 %rd2399, %rd2398, %rd2368; + xor.b64 %rd2400, %rd2399, %rd2383; + xor.b64 %rd2401, %rd2341, %rd2326; + xor.b64 %rd2402, %rd2401, %rd2356; + xor.b64 %rd2403, %rd2402, %rd2371; + xor.b64 %rd2404, %rd2403, %rd2386; + xor.b64 %rd2405, %rd2344, %rd2329; + xor.b64 %rd2406, %rd2405, %rd2359; + xor.b64 %rd2407, %rd2406, %rd2374; + xor.b64 %rd2408, %rd2407, %rd2389; + xor.b64 %rd2409, %rd2347, %rd2332; + xor.b64 %rd2410, %rd2409, %rd2362; + xor.b64 %rd2411, %rd2410, %rd2377; + xor.b64 %rd2412, %rd2411, %rd2392; + xor.b64 %rd2413, %rd2350, %rd2335; + xor.b64 %rd2414, %rd2413, %rd2365; + xor.b64 %rd2415, %rd2414, %rd2380; + xor.b64 %rd2416, %rd2415, %rd2395; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1586}, %rd2404; + } + { + .reg .b32 %dummy; + mov.b64 {%r1587,%dummy}, %rd2404; + } + shf.l.wrap.b32 %r1588, %r1587, %r1586, 1; + shf.l.wrap.b32 %r1589, %r1586, %r1587, 1; + mov.b64 %rd2417, {%r1589, %r1588}; + xor.b64 %rd2418, %rd2416, %rd2417; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1590}, %rd2408; + } + { + .reg .b32 %dummy; + mov.b64 {%r1591,%dummy}, %rd2408; + } + shf.l.wrap.b32 %r1592, %r1591, %r1590, 1; + shf.l.wrap.b32 %r1593, %r1590, %r1591, 1; + mov.b64 %rd2419, {%r1593, %r1592}; + xor.b64 %rd2420, %rd2419, %rd2400; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1594}, %rd2412; + } + { + .reg .b32 %dummy; + mov.b64 {%r1595,%dummy}, %rd2412; + } + shf.l.wrap.b32 %r1596, %r1595, %r1594, 1; + shf.l.wrap.b32 %r1597, %r1594, %r1595, 1; + mov.b64 %rd2421, {%r1597, %r1596}; + xor.b64 %rd2422, %rd2421, %rd2404; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1598}, %rd2416; + } + { + .reg .b32 %dummy; + mov.b64 {%r1599,%dummy}, %rd2416; + } + shf.l.wrap.b32 %r1600, %r1599, %r1598, 1; + shf.l.wrap.b32 %r1601, %r1598, %r1599, 1; + mov.b64 %rd2423, {%r1601, %r1600}; + xor.b64 %rd2424, %rd2423, %rd2408; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1602}, %rd2400; + } + { + .reg .b32 %dummy; + mov.b64 {%r1603,%dummy}, %rd2400; + } + shf.l.wrap.b32 %r1604, %r1603, %r1602, 1; + shf.l.wrap.b32 %r1605, %r1602, %r1603, 1; + mov.b64 %rd2425, {%r1605, %r1604}; + xor.b64 %rd2426, %rd2412, %rd2425; + xor.b64 %rd2427, %rd2418, %rd2396; + xor.b64 %rd2428, %rd2420, %rd2326; + xor.b64 %rd2429, %rd2422, %rd2329; + xor.b64 %rd2430, %rd2424, %rd2332; + xor.b64 %rd2431, %rd2426, %rd2335; + xor.b64 %rd2432, %rd2418, %rd2338; + xor.b64 %rd2433, %rd2420, %rd2341; + xor.b64 %rd2434, %rd2422, %rd2344; + xor.b64 %rd2435, %rd2424, %rd2347; + xor.b64 %rd2436, %rd2426, %rd2350; + xor.b64 %rd2437, %rd2418, %rd2353; + xor.b64 %rd2438, %rd2420, %rd2356; + xor.b64 %rd2439, %rd2422, %rd2359; + xor.b64 %rd2440, %rd2424, %rd2362; + xor.b64 %rd2441, %rd2426, %rd2365; + xor.b64 %rd2442, %rd2418, %rd2368; + xor.b64 %rd2443, %rd2420, %rd2371; + xor.b64 %rd2444, %rd2422, %rd2374; + xor.b64 %rd2445, %rd2424, %rd2377; + xor.b64 %rd2446, %rd2426, %rd2380; + xor.b64 %rd2447, %rd2418, %rd2383; + xor.b64 %rd2448, %rd2420, %rd2386; + xor.b64 %rd2449, %rd2422, %rd2389; + xor.b64 %rd2450, %rd2424, %rd2392; + xor.b64 %rd2451, %rd2426, %rd2395; + { + .reg .b32 %dummy; + mov.b64 {%r1606,%dummy}, %rd2433; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1607}, %rd2433; + } + shf.r.wrap.b32 %r1608, %r1607, %r1606, 20; + shf.r.wrap.b32 %r1609, %r1606, %r1607, 20; + mov.b64 %rd2452, {%r1609, %r1608}; + { + .reg .b32 %dummy; + mov.b64 {%r1610,%dummy}, %rd2439; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1611}, %rd2439; + } + shf.r.wrap.b32 %r1612, %r1611, %r1610, 21; + shf.r.wrap.b32 %r1613, %r1610, %r1611, 21; + mov.b64 %rd2453, {%r1613, %r1612}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1614}, %rd2445; + } + { + .reg .b32 %dummy; + mov.b64 {%r1615,%dummy}, %rd2445; + } + shf.l.wrap.b32 %r1616, %r1615, %r1614, 21; + shf.l.wrap.b32 %r1617, %r1614, %r1615, 21; + mov.b64 %rd2454, {%r1617, %r1616}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1618}, %rd2451; + } + { + .reg .b32 %dummy; + mov.b64 {%r1619,%dummy}, %rd2451; + } + shf.l.wrap.b32 %r1620, %r1619, %r1618, 14; + shf.l.wrap.b32 %r1621, %r1618, %r1619, 14; + mov.b64 %rd2455, {%r1621, %r1620}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1622}, %rd2430; + } + { + .reg .b32 %dummy; + mov.b64 {%r1623,%dummy}, %rd2430; + } + shf.l.wrap.b32 %r1624, %r1623, %r1622, 28; + shf.l.wrap.b32 %r1625, %r1622, %r1623, 28; + mov.b64 %rd2456, {%r1625, %r1624}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1626}, %rd2436; + } + { + .reg .b32 %dummy; + mov.b64 {%r1627,%dummy}, %rd2436; + } + shf.l.wrap.b32 %r1628, %r1627, %r1626, 20; + shf.l.wrap.b32 %r1629, %r1626, %r1627, 20; + mov.b64 %rd2457, {%r1629, %r1628}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1630}, %rd2437; + } + { + .reg .b32 %dummy; + mov.b64 {%r1631,%dummy}, %rd2437; + } + shf.l.wrap.b32 %r1632, %r1631, %r1630, 3; + shf.l.wrap.b32 %r1633, %r1630, %r1631, 3; + mov.b64 %rd2458, {%r1633, %r1632}; + { + .reg .b32 %dummy; + mov.b64 {%r1634,%dummy}, %rd2443; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1635}, %rd2443; + } + shf.r.wrap.b32 %r1636, %r1635, %r1634, 19; + shf.r.wrap.b32 %r1637, %r1634, %r1635, 19; + mov.b64 %rd2459, {%r1637, %r1636}; + { + .reg .b32 %dummy; + mov.b64 {%r1638,%dummy}, %rd2449; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1639}, %rd2449; + } + shf.r.wrap.b32 %r1640, %r1639, %r1638, 3; + shf.r.wrap.b32 %r1641, %r1638, %r1639, 3; + mov.b64 %rd2460, {%r1641, %r1640}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1642}, %rd2428; + } + { + .reg .b32 %dummy; + mov.b64 {%r1643,%dummy}, %rd2428; + } + shf.l.wrap.b32 %r1644, %r1643, %r1642, 1; + shf.l.wrap.b32 %r1645, %r1642, %r1643, 1; + mov.b64 %rd2461, {%r1645, %r1644}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1646}, %rd2434; + } + { + .reg .b32 %dummy; + mov.b64 {%r1647,%dummy}, %rd2434; + } + shf.l.wrap.b32 %r1648, %r1647, %r1646, 6; + shf.l.wrap.b32 %r1649, %r1646, %r1647, 6; + mov.b64 %rd2462, {%r1649, %r1648}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1650}, %rd2440; + } + { + .reg .b32 %dummy; + mov.b64 {%r1651,%dummy}, %rd2440; + } + shf.l.wrap.b32 %r1652, %r1651, %r1650, 25; + shf.l.wrap.b32 %r1653, %r1650, %r1651, 25; + mov.b64 %rd2463, {%r1653, %r1652}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1654}, %rd2446; + } + { + .reg .b32 %dummy; + mov.b64 {%r1655,%dummy}, %rd2446; + } + shf.l.wrap.b32 %r1656, %r1655, %r1654, 8; + shf.l.wrap.b32 %r1657, %r1654, %r1655, 8; + mov.b64 %rd2464, {%r1657, %r1656}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1658}, %rd2447; + } + { + .reg .b32 %dummy; + mov.b64 {%r1659,%dummy}, %rd2447; + } + shf.l.wrap.b32 %r1660, %r1659, %r1658, 18; + shf.l.wrap.b32 %r1661, %r1658, %r1659, 18; + mov.b64 %rd2465, {%r1661, %r1660}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1662}, %rd2431; + } + { + .reg .b32 %dummy; + mov.b64 {%r1663,%dummy}, %rd2431; + } + shf.l.wrap.b32 %r1664, %r1663, %r1662, 27; + shf.l.wrap.b32 %r1665, %r1662, %r1663, 27; + mov.b64 %rd2466, {%r1665, %r1664}; + { + .reg .b32 %dummy; + mov.b64 {%r1666,%dummy}, %rd2432; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1667}, %rd2432; + } + shf.r.wrap.b32 %r1668, %r1667, %r1666, 28; + shf.r.wrap.b32 %r1669, %r1666, %r1667, 28; + mov.b64 %rd2467, {%r1669, %r1668}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1670}, %rd2438; + } + { + .reg .b32 %dummy; + mov.b64 {%r1671,%dummy}, %rd2438; + } + shf.l.wrap.b32 %r1672, %r1671, %r1670, 10; + shf.l.wrap.b32 %r1673, %r1670, %r1671, 10; + mov.b64 %rd2468, {%r1673, %r1672}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1674}, %rd2444; + } + { + .reg .b32 %dummy; + mov.b64 {%r1675,%dummy}, %rd2444; + } + shf.l.wrap.b32 %r1676, %r1675, %r1674, 15; + shf.l.wrap.b32 %r1677, %r1674, %r1675, 15; + mov.b64 %rd2469, {%r1677, %r1676}; + { + .reg .b32 %dummy; + mov.b64 {%r1678,%dummy}, %rd2450; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1679}, %rd2450; + } + shf.r.wrap.b32 %r1680, %r1679, %r1678, 8; + shf.r.wrap.b32 %r1681, %r1678, %r1679, 8; + mov.b64 %rd2470, {%r1681, %r1680}; + { + .reg .b32 %dummy; + mov.b64 {%r1682,%dummy}, %rd2429; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1683}, %rd2429; + } + shf.r.wrap.b32 %r1684, %r1683, %r1682, 2; + shf.r.wrap.b32 %r1685, %r1682, %r1683, 2; + mov.b64 %rd2471, {%r1685, %r1684}; + { + .reg .b32 %dummy; + mov.b64 {%r1686,%dummy}, %rd2435; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1687}, %rd2435; + } + shf.r.wrap.b32 %r1688, %r1687, %r1686, 9; + shf.r.wrap.b32 %r1689, %r1686, %r1687, 9; + mov.b64 %rd2472, {%r1689, %r1688}; + { + .reg .b32 %dummy; + mov.b64 {%r1690,%dummy}, %rd2441; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1691}, %rd2441; + } + shf.r.wrap.b32 %r1692, %r1691, %r1690, 25; + shf.r.wrap.b32 %r1693, %r1690, %r1691, 25; + mov.b64 %rd2473, {%r1693, %r1692}; + { + .reg .b32 %dummy; + mov.b64 {%r1694,%dummy}, %rd2442; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1695}, %rd2442; + } + shf.r.wrap.b32 %r1696, %r1695, %r1694, 23; + shf.r.wrap.b32 %r1697, %r1694, %r1695, 23; + mov.b64 %rd2474, {%r1697, %r1696}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1698}, %rd2448; + } + { + .reg .b32 %dummy; + mov.b64 {%r1699,%dummy}, %rd2448; + } + shf.l.wrap.b32 %r1700, %r1699, %r1698, 2; + shf.l.wrap.b32 %r1701, %r1698, %r1699, 2; + mov.b64 %rd2475, {%r1701, %r1700}; + not.b64 %rd2476, %rd2452; + and.b64 %rd2477, %rd2453, %rd2476; + xor.b64 %rd2478, %rd2477, %rd2427; + not.b64 %rd2479, %rd2453; + and.b64 %rd2480, %rd2454, %rd2479; + xor.b64 %rd2481, %rd2480, %rd2452; + not.b64 %rd2482, %rd2454; + and.b64 %rd2483, %rd2455, %rd2482; + xor.b64 %rd2484, %rd2483, %rd2453; + not.b64 %rd2485, %rd2455; + and.b64 %rd2486, %rd2427, %rd2485; + xor.b64 %rd2487, %rd2486, %rd2454; + not.b64 %rd2488, %rd2427; + and.b64 %rd2489, %rd2452, %rd2488; + xor.b64 %rd2490, %rd2489, %rd2455; + not.b64 %rd2491, %rd2457; + and.b64 %rd2492, %rd2458, %rd2491; + xor.b64 %rd2493, %rd2492, %rd2456; + not.b64 %rd2494, %rd2458; + and.b64 %rd2495, %rd2459, %rd2494; + xor.b64 %rd2496, %rd2495, %rd2457; + not.b64 %rd2497, %rd2459; + and.b64 %rd2498, %rd2460, %rd2497; + xor.b64 %rd2499, %rd2498, %rd2458; + not.b64 %rd2500, %rd2460; + and.b64 %rd2501, %rd2456, %rd2500; + xor.b64 %rd2502, %rd2501, %rd2459; + not.b64 %rd2503, %rd2456; + and.b64 %rd2504, %rd2457, %rd2503; + xor.b64 %rd2505, %rd2504, %rd2460; + not.b64 %rd2506, %rd2462; + and.b64 %rd2507, %rd2463, %rd2506; + xor.b64 %rd2508, %rd2507, %rd2461; + not.b64 %rd2509, %rd2463; + and.b64 %rd2510, %rd2464, %rd2509; + xor.b64 %rd2511, %rd2510, %rd2462; + not.b64 %rd2512, %rd2464; + and.b64 %rd2513, %rd2465, %rd2512; + xor.b64 %rd2514, %rd2513, %rd2463; + not.b64 %rd2515, %rd2465; + and.b64 %rd2516, %rd2461, %rd2515; + xor.b64 %rd2517, %rd2516, %rd2464; + not.b64 %rd2518, %rd2461; + and.b64 %rd2519, %rd2462, %rd2518; + xor.b64 %rd2520, %rd2519, %rd2465; + not.b64 %rd2521, %rd2467; + and.b64 %rd2522, %rd2468, %rd2521; + xor.b64 %rd2523, %rd2522, %rd2466; + not.b64 %rd2524, %rd2468; + and.b64 %rd2525, %rd2469, %rd2524; + xor.b64 %rd2526, %rd2525, %rd2467; + not.b64 %rd2527, %rd2469; + and.b64 %rd2528, %rd2470, %rd2527; + xor.b64 %rd2529, %rd2528, %rd2468; + not.b64 %rd2530, %rd2470; + and.b64 %rd2531, %rd2466, %rd2530; + xor.b64 %rd2532, %rd2531, %rd2469; + not.b64 %rd2533, %rd2466; + and.b64 %rd2534, %rd2467, %rd2533; + xor.b64 %rd2535, %rd2534, %rd2470; + not.b64 %rd2536, %rd2472; + and.b64 %rd2537, %rd2473, %rd2536; + xor.b64 %rd2538, %rd2537, %rd2471; + not.b64 %rd2539, %rd2473; + and.b64 %rd2540, %rd2474, %rd2539; + xor.b64 %rd2541, %rd2540, %rd2472; + not.b64 %rd2542, %rd2474; + and.b64 %rd2543, %rd2475, %rd2542; + xor.b64 %rd2544, %rd2543, %rd2473; + not.b64 %rd2545, %rd2475; + and.b64 %rd2546, %rd2471, %rd2545; + xor.b64 %rd2547, %rd2546, %rd2474; + not.b64 %rd2548, %rd2471; + and.b64 %rd2549, %rd2472, %rd2548; + xor.b64 %rd2550, %rd2549, %rd2475; + xor.b64 %rd2551, %rd2478, %rd29; + xor.b64 %rd2552, %rd2493, %rd2551; + xor.b64 %rd2553, %rd2552, %rd2508; + xor.b64 %rd2554, %rd2553, %rd2523; + xor.b64 %rd2555, %rd2554, %rd2538; + xor.b64 %rd2556, %rd2496, %rd2481; + xor.b64 %rd2557, %rd2556, %rd2511; + xor.b64 %rd2558, %rd2557, %rd2526; + xor.b64 %rd2559, %rd2558, %rd2541; + xor.b64 %rd2560, %rd2499, %rd2484; + xor.b64 %rd2561, %rd2560, %rd2514; + xor.b64 %rd2562, %rd2561, %rd2529; + xor.b64 %rd2563, %rd2562, %rd2544; + xor.b64 %rd2564, %rd2502, %rd2487; + xor.b64 %rd2565, %rd2564, %rd2517; + xor.b64 %rd2566, %rd2565, %rd2532; + xor.b64 %rd2567, %rd2566, %rd2547; + xor.b64 %rd2568, %rd2505, %rd2490; + xor.b64 %rd2569, %rd2568, %rd2520; + xor.b64 %rd2570, %rd2569, %rd2535; + xor.b64 %rd2571, %rd2570, %rd2550; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1702}, %rd2559; + } + { + .reg .b32 %dummy; + mov.b64 {%r1703,%dummy}, %rd2559; + } + shf.l.wrap.b32 %r1704, %r1703, %r1702, 1; + shf.l.wrap.b32 %r1705, %r1702, %r1703, 1; + mov.b64 %rd2572, {%r1705, %r1704}; + xor.b64 %rd2573, %rd2571, %rd2572; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1706}, %rd2563; + } + { + .reg .b32 %dummy; + mov.b64 {%r1707,%dummy}, %rd2563; + } + shf.l.wrap.b32 %r1708, %r1707, %r1706, 1; + shf.l.wrap.b32 %r1709, %r1706, %r1707, 1; + mov.b64 %rd2574, {%r1709, %r1708}; + xor.b64 %rd2575, %rd2574, %rd2555; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1710}, %rd2567; + } + { + .reg .b32 %dummy; + mov.b64 {%r1711,%dummy}, %rd2567; + } + shf.l.wrap.b32 %r1712, %r1711, %r1710, 1; + shf.l.wrap.b32 %r1713, %r1710, %r1711, 1; + mov.b64 %rd2576, {%r1713, %r1712}; + xor.b64 %rd2577, %rd2576, %rd2559; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1714}, %rd2571; + } + { + .reg .b32 %dummy; + mov.b64 {%r1715,%dummy}, %rd2571; + } + shf.l.wrap.b32 %r1716, %r1715, %r1714, 1; + shf.l.wrap.b32 %r1717, %r1714, %r1715, 1; + mov.b64 %rd2578, {%r1717, %r1716}; + xor.b64 %rd2579, %rd2578, %rd2563; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1718}, %rd2555; + } + { + .reg .b32 %dummy; + mov.b64 {%r1719,%dummy}, %rd2555; + } + shf.l.wrap.b32 %r1720, %r1719, %r1718, 1; + shf.l.wrap.b32 %r1721, %r1718, %r1719, 1; + mov.b64 %rd2580, {%r1721, %r1720}; + xor.b64 %rd2581, %rd2567, %rd2580; + xor.b64 %rd2582, %rd2573, %rd2551; + xor.b64 %rd2583, %rd2575, %rd2481; + xor.b64 %rd2584, %rd2577, %rd2484; + xor.b64 %rd2585, %rd2579, %rd2487; + xor.b64 %rd2586, %rd2581, %rd2490; + xor.b64 %rd2587, %rd2573, %rd2493; + xor.b64 %rd2588, %rd2575, %rd2496; + xor.b64 %rd2589, %rd2577, %rd2499; + xor.b64 %rd2590, %rd2579, %rd2502; + xor.b64 %rd2591, %rd2581, %rd2505; + xor.b64 %rd2592, %rd2573, %rd2508; + xor.b64 %rd2593, %rd2575, %rd2511; + xor.b64 %rd2594, %rd2577, %rd2514; + xor.b64 %rd2595, %rd2579, %rd2517; + xor.b64 %rd2596, %rd2581, %rd2520; + xor.b64 %rd2597, %rd2573, %rd2523; + xor.b64 %rd2598, %rd2575, %rd2526; + xor.b64 %rd2599, %rd2577, %rd2529; + xor.b64 %rd2600, %rd2579, %rd2532; + xor.b64 %rd2601, %rd2581, %rd2535; + xor.b64 %rd2602, %rd2573, %rd2538; + xor.b64 %rd2603, %rd2575, %rd2541; + xor.b64 %rd2604, %rd2577, %rd2544; + xor.b64 %rd2605, %rd2579, %rd2547; + xor.b64 %rd2606, %rd2581, %rd2550; + { + .reg .b32 %dummy; + mov.b64 {%r1722,%dummy}, %rd2588; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1723}, %rd2588; + } + shf.r.wrap.b32 %r1724, %r1723, %r1722, 20; + shf.r.wrap.b32 %r1725, %r1722, %r1723, 20; + mov.b64 %rd2607, {%r1725, %r1724}; + { + .reg .b32 %dummy; + mov.b64 {%r1726,%dummy}, %rd2594; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1727}, %rd2594; + } + shf.r.wrap.b32 %r1728, %r1727, %r1726, 21; + shf.r.wrap.b32 %r1729, %r1726, %r1727, 21; + mov.b64 %rd2608, {%r1729, %r1728}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1730}, %rd2600; + } + { + .reg .b32 %dummy; + mov.b64 {%r1731,%dummy}, %rd2600; + } + shf.l.wrap.b32 %r1732, %r1731, %r1730, 21; + shf.l.wrap.b32 %r1733, %r1730, %r1731, 21; + mov.b64 %rd2609, {%r1733, %r1732}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1734}, %rd2606; + } + { + .reg .b32 %dummy; + mov.b64 {%r1735,%dummy}, %rd2606; + } + shf.l.wrap.b32 %r1736, %r1735, %r1734, 14; + shf.l.wrap.b32 %r1737, %r1734, %r1735, 14; + mov.b64 %rd2610, {%r1737, %r1736}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1738}, %rd2585; + } + { + .reg .b32 %dummy; + mov.b64 {%r1739,%dummy}, %rd2585; + } + shf.l.wrap.b32 %r1740, %r1739, %r1738, 28; + shf.l.wrap.b32 %r1741, %r1738, %r1739, 28; + mov.b64 %rd2611, {%r1741, %r1740}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1742}, %rd2591; + } + { + .reg .b32 %dummy; + mov.b64 {%r1743,%dummy}, %rd2591; + } + shf.l.wrap.b32 %r1744, %r1743, %r1742, 20; + shf.l.wrap.b32 %r1745, %r1742, %r1743, 20; + mov.b64 %rd2612, {%r1745, %r1744}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1746}, %rd2592; + } + { + .reg .b32 %dummy; + mov.b64 {%r1747,%dummy}, %rd2592; + } + shf.l.wrap.b32 %r1748, %r1747, %r1746, 3; + shf.l.wrap.b32 %r1749, %r1746, %r1747, 3; + mov.b64 %rd2613, {%r1749, %r1748}; + { + .reg .b32 %dummy; + mov.b64 {%r1750,%dummy}, %rd2598; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1751}, %rd2598; + } + shf.r.wrap.b32 %r1752, %r1751, %r1750, 19; + shf.r.wrap.b32 %r1753, %r1750, %r1751, 19; + mov.b64 %rd2614, {%r1753, %r1752}; + { + .reg .b32 %dummy; + mov.b64 {%r1754,%dummy}, %rd2604; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1755}, %rd2604; + } + shf.r.wrap.b32 %r1756, %r1755, %r1754, 3; + shf.r.wrap.b32 %r1757, %r1754, %r1755, 3; + mov.b64 %rd2615, {%r1757, %r1756}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1758}, %rd2583; + } + { + .reg .b32 %dummy; + mov.b64 {%r1759,%dummy}, %rd2583; + } + shf.l.wrap.b32 %r1760, %r1759, %r1758, 1; + shf.l.wrap.b32 %r1761, %r1758, %r1759, 1; + mov.b64 %rd2616, {%r1761, %r1760}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1762}, %rd2589; + } + { + .reg .b32 %dummy; + mov.b64 {%r1763,%dummy}, %rd2589; + } + shf.l.wrap.b32 %r1764, %r1763, %r1762, 6; + shf.l.wrap.b32 %r1765, %r1762, %r1763, 6; + mov.b64 %rd2617, {%r1765, %r1764}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1766}, %rd2595; + } + { + .reg .b32 %dummy; + mov.b64 {%r1767,%dummy}, %rd2595; + } + shf.l.wrap.b32 %r1768, %r1767, %r1766, 25; + shf.l.wrap.b32 %r1769, %r1766, %r1767, 25; + mov.b64 %rd2618, {%r1769, %r1768}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1770}, %rd2601; + } + { + .reg .b32 %dummy; + mov.b64 {%r1771,%dummy}, %rd2601; + } + shf.l.wrap.b32 %r1772, %r1771, %r1770, 8; + shf.l.wrap.b32 %r1773, %r1770, %r1771, 8; + mov.b64 %rd2619, {%r1773, %r1772}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1774}, %rd2602; + } + { + .reg .b32 %dummy; + mov.b64 {%r1775,%dummy}, %rd2602; + } + shf.l.wrap.b32 %r1776, %r1775, %r1774, 18; + shf.l.wrap.b32 %r1777, %r1774, %r1775, 18; + mov.b64 %rd2620, {%r1777, %r1776}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1778}, %rd2586; + } + { + .reg .b32 %dummy; + mov.b64 {%r1779,%dummy}, %rd2586; + } + shf.l.wrap.b32 %r1780, %r1779, %r1778, 27; + shf.l.wrap.b32 %r1781, %r1778, %r1779, 27; + mov.b64 %rd2621, {%r1781, %r1780}; + { + .reg .b32 %dummy; + mov.b64 {%r1782,%dummy}, %rd2587; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1783}, %rd2587; + } + shf.r.wrap.b32 %r1784, %r1783, %r1782, 28; + shf.r.wrap.b32 %r1785, %r1782, %r1783, 28; + mov.b64 %rd2622, {%r1785, %r1784}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1786}, %rd2593; + } + { + .reg .b32 %dummy; + mov.b64 {%r1787,%dummy}, %rd2593; + } + shf.l.wrap.b32 %r1788, %r1787, %r1786, 10; + shf.l.wrap.b32 %r1789, %r1786, %r1787, 10; + mov.b64 %rd2623, {%r1789, %r1788}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1790}, %rd2599; + } + { + .reg .b32 %dummy; + mov.b64 {%r1791,%dummy}, %rd2599; + } + shf.l.wrap.b32 %r1792, %r1791, %r1790, 15; + shf.l.wrap.b32 %r1793, %r1790, %r1791, 15; + mov.b64 %rd2624, {%r1793, %r1792}; + { + .reg .b32 %dummy; + mov.b64 {%r1794,%dummy}, %rd2605; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1795}, %rd2605; + } + shf.r.wrap.b32 %r1796, %r1795, %r1794, 8; + shf.r.wrap.b32 %r1797, %r1794, %r1795, 8; + mov.b64 %rd2625, {%r1797, %r1796}; + { + .reg .b32 %dummy; + mov.b64 {%r1798,%dummy}, %rd2584; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1799}, %rd2584; + } + shf.r.wrap.b32 %r1800, %r1799, %r1798, 2; + shf.r.wrap.b32 %r1801, %r1798, %r1799, 2; + mov.b64 %rd2626, {%r1801, %r1800}; + { + .reg .b32 %dummy; + mov.b64 {%r1802,%dummy}, %rd2590; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1803}, %rd2590; + } + shf.r.wrap.b32 %r1804, %r1803, %r1802, 9; + shf.r.wrap.b32 %r1805, %r1802, %r1803, 9; + mov.b64 %rd2627, {%r1805, %r1804}; + { + .reg .b32 %dummy; + mov.b64 {%r1806,%dummy}, %rd2596; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1807}, %rd2596; + } + shf.r.wrap.b32 %r1808, %r1807, %r1806, 25; + shf.r.wrap.b32 %r1809, %r1806, %r1807, 25; + mov.b64 %rd2628, {%r1809, %r1808}; + { + .reg .b32 %dummy; + mov.b64 {%r1810,%dummy}, %rd2597; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1811}, %rd2597; + } + shf.r.wrap.b32 %r1812, %r1811, %r1810, 23; + shf.r.wrap.b32 %r1813, %r1810, %r1811, 23; + mov.b64 %rd2629, {%r1813, %r1812}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1814}, %rd2603; + } + { + .reg .b32 %dummy; + mov.b64 {%r1815,%dummy}, %rd2603; + } + shf.l.wrap.b32 %r1816, %r1815, %r1814, 2; + shf.l.wrap.b32 %r1817, %r1814, %r1815, 2; + mov.b64 %rd2630, {%r1817, %r1816}; + not.b64 %rd2631, %rd2607; + and.b64 %rd2632, %rd2608, %rd2631; + xor.b64 %rd2633, %rd2632, %rd2582; + not.b64 %rd2634, %rd2608; + and.b64 %rd2635, %rd2609, %rd2634; + xor.b64 %rd2636, %rd2635, %rd2607; + not.b64 %rd2637, %rd2609; + and.b64 %rd2638, %rd2610, %rd2637; + xor.b64 %rd2639, %rd2638, %rd2608; + not.b64 %rd2640, %rd2610; + and.b64 %rd2641, %rd2582, %rd2640; + xor.b64 %rd2642, %rd2641, %rd2609; + not.b64 %rd2643, %rd2582; + and.b64 %rd2644, %rd2607, %rd2643; + xor.b64 %rd2645, %rd2644, %rd2610; + not.b64 %rd2646, %rd2612; + and.b64 %rd2647, %rd2613, %rd2646; + xor.b64 %rd2648, %rd2647, %rd2611; + not.b64 %rd2649, %rd2613; + and.b64 %rd2650, %rd2614, %rd2649; + xor.b64 %rd2651, %rd2650, %rd2612; + not.b64 %rd2652, %rd2614; + and.b64 %rd2653, %rd2615, %rd2652; + xor.b64 %rd2654, %rd2653, %rd2613; + not.b64 %rd2655, %rd2615; + and.b64 %rd2656, %rd2611, %rd2655; + xor.b64 %rd2657, %rd2656, %rd2614; + not.b64 %rd2658, %rd2611; + and.b64 %rd2659, %rd2612, %rd2658; + xor.b64 %rd2660, %rd2659, %rd2615; + not.b64 %rd2661, %rd2617; + and.b64 %rd2662, %rd2618, %rd2661; + xor.b64 %rd2663, %rd2662, %rd2616; + not.b64 %rd2664, %rd2618; + and.b64 %rd2665, %rd2619, %rd2664; + xor.b64 %rd2666, %rd2665, %rd2617; + not.b64 %rd2667, %rd2619; + and.b64 %rd2668, %rd2620, %rd2667; + xor.b64 %rd2669, %rd2668, %rd2618; + not.b64 %rd2670, %rd2620; + and.b64 %rd2671, %rd2616, %rd2670; + xor.b64 %rd2672, %rd2671, %rd2619; + not.b64 %rd2673, %rd2616; + and.b64 %rd2674, %rd2617, %rd2673; + xor.b64 %rd2675, %rd2674, %rd2620; + not.b64 %rd2676, %rd2622; + and.b64 %rd2677, %rd2623, %rd2676; + xor.b64 %rd2678, %rd2677, %rd2621; + not.b64 %rd2679, %rd2623; + and.b64 %rd2680, %rd2624, %rd2679; + xor.b64 %rd2681, %rd2680, %rd2622; + not.b64 %rd2682, %rd2624; + and.b64 %rd2683, %rd2625, %rd2682; + xor.b64 %rd2684, %rd2683, %rd2623; + not.b64 %rd2685, %rd2625; + and.b64 %rd2686, %rd2621, %rd2685; + xor.b64 %rd2687, %rd2686, %rd2624; + not.b64 %rd2688, %rd2621; + and.b64 %rd2689, %rd2622, %rd2688; + xor.b64 %rd2690, %rd2689, %rd2625; + not.b64 %rd2691, %rd2627; + and.b64 %rd2692, %rd2628, %rd2691; + xor.b64 %rd2693, %rd2692, %rd2626; + not.b64 %rd2694, %rd2628; + and.b64 %rd2695, %rd2629, %rd2694; + xor.b64 %rd2696, %rd2695, %rd2627; + not.b64 %rd2697, %rd2629; + and.b64 %rd2698, %rd2630, %rd2697; + xor.b64 %rd2699, %rd2698, %rd2628; + not.b64 %rd2700, %rd2630; + and.b64 %rd2701, %rd2626, %rd2700; + xor.b64 %rd2702, %rd2701, %rd2629; + not.b64 %rd2703, %rd2626; + and.b64 %rd2704, %rd2627, %rd2703; + xor.b64 %rd2705, %rd2704, %rd2630; + xor.b64 %rd2706, %rd2633, %rd30; + xor.b64 %rd2707, %rd2648, %rd2706; + xor.b64 %rd2708, %rd2707, %rd2663; + xor.b64 %rd2709, %rd2708, %rd2678; + xor.b64 %rd2710, %rd2709, %rd2693; + xor.b64 %rd2711, %rd2651, %rd2636; + xor.b64 %rd2712, %rd2711, %rd2666; + xor.b64 %rd2713, %rd2712, %rd2681; + xor.b64 %rd2714, %rd2713, %rd2696; + xor.b64 %rd2715, %rd2654, %rd2639; + xor.b64 %rd2716, %rd2715, %rd2669; + xor.b64 %rd2717, %rd2716, %rd2684; + xor.b64 %rd2718, %rd2717, %rd2699; + xor.b64 %rd2719, %rd2657, %rd2642; + xor.b64 %rd2720, %rd2719, %rd2672; + xor.b64 %rd2721, %rd2720, %rd2687; + xor.b64 %rd2722, %rd2721, %rd2702; + xor.b64 %rd2723, %rd2660, %rd2645; + xor.b64 %rd2724, %rd2723, %rd2675; + xor.b64 %rd2725, %rd2724, %rd2690; + xor.b64 %rd2726, %rd2725, %rd2705; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1818}, %rd2714; + } + { + .reg .b32 %dummy; + mov.b64 {%r1819,%dummy}, %rd2714; + } + shf.l.wrap.b32 %r1820, %r1819, %r1818, 1; + shf.l.wrap.b32 %r1821, %r1818, %r1819, 1; + mov.b64 %rd2727, {%r1821, %r1820}; + xor.b64 %rd2728, %rd2726, %rd2727; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1822}, %rd2718; + } + { + .reg .b32 %dummy; + mov.b64 {%r1823,%dummy}, %rd2718; + } + shf.l.wrap.b32 %r1824, %r1823, %r1822, 1; + shf.l.wrap.b32 %r1825, %r1822, %r1823, 1; + mov.b64 %rd2729, {%r1825, %r1824}; + xor.b64 %rd2730, %rd2729, %rd2710; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1826}, %rd2722; + } + { + .reg .b32 %dummy; + mov.b64 {%r1827,%dummy}, %rd2722; + } + shf.l.wrap.b32 %r1828, %r1827, %r1826, 1; + shf.l.wrap.b32 %r1829, %r1826, %r1827, 1; + mov.b64 %rd2731, {%r1829, %r1828}; + xor.b64 %rd2732, %rd2731, %rd2714; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1830}, %rd2726; + } + { + .reg .b32 %dummy; + mov.b64 {%r1831,%dummy}, %rd2726; + } + shf.l.wrap.b32 %r1832, %r1831, %r1830, 1; + shf.l.wrap.b32 %r1833, %r1830, %r1831, 1; + mov.b64 %rd2733, {%r1833, %r1832}; + xor.b64 %rd2734, %rd2733, %rd2718; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1834}, %rd2710; + } + { + .reg .b32 %dummy; + mov.b64 {%r1835,%dummy}, %rd2710; + } + shf.l.wrap.b32 %r1836, %r1835, %r1834, 1; + shf.l.wrap.b32 %r1837, %r1834, %r1835, 1; + mov.b64 %rd2735, {%r1837, %r1836}; + xor.b64 %rd2736, %rd2722, %rd2735; + xor.b64 %rd2737, %rd2728, %rd2706; + xor.b64 %rd2738, %rd2730, %rd2636; + xor.b64 %rd2739, %rd2732, %rd2639; + xor.b64 %rd2740, %rd2734, %rd2642; + xor.b64 %rd2741, %rd2736, %rd2645; + xor.b64 %rd2742, %rd2728, %rd2648; + xor.b64 %rd2743, %rd2730, %rd2651; + xor.b64 %rd2744, %rd2732, %rd2654; + xor.b64 %rd2745, %rd2734, %rd2657; + xor.b64 %rd2746, %rd2736, %rd2660; + xor.b64 %rd2747, %rd2728, %rd2663; + xor.b64 %rd2748, %rd2730, %rd2666; + xor.b64 %rd2749, %rd2732, %rd2669; + xor.b64 %rd2750, %rd2734, %rd2672; + xor.b64 %rd2751, %rd2736, %rd2675; + xor.b64 %rd2752, %rd2728, %rd2678; + xor.b64 %rd2753, %rd2730, %rd2681; + xor.b64 %rd2754, %rd2732, %rd2684; + xor.b64 %rd2755, %rd2734, %rd2687; + xor.b64 %rd2756, %rd2736, %rd2690; + xor.b64 %rd2757, %rd2728, %rd2693; + xor.b64 %rd2758, %rd2730, %rd2696; + xor.b64 %rd2759, %rd2732, %rd2699; + xor.b64 %rd2760, %rd2734, %rd2702; + xor.b64 %rd2761, %rd2736, %rd2705; + { + .reg .b32 %dummy; + mov.b64 {%r1838,%dummy}, %rd2743; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1839}, %rd2743; + } + shf.r.wrap.b32 %r1840, %r1839, %r1838, 20; + shf.r.wrap.b32 %r1841, %r1838, %r1839, 20; + mov.b64 %rd2762, {%r1841, %r1840}; + { + .reg .b32 %dummy; + mov.b64 {%r1842,%dummy}, %rd2749; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1843}, %rd2749; + } + shf.r.wrap.b32 %r1844, %r1843, %r1842, 21; + shf.r.wrap.b32 %r1845, %r1842, %r1843, 21; + mov.b64 %rd2763, {%r1845, %r1844}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1846}, %rd2755; + } + { + .reg .b32 %dummy; + mov.b64 {%r1847,%dummy}, %rd2755; + } + shf.l.wrap.b32 %r1848, %r1847, %r1846, 21; + shf.l.wrap.b32 %r1849, %r1846, %r1847, 21; + mov.b64 %rd2764, {%r1849, %r1848}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1850}, %rd2761; + } + { + .reg .b32 %dummy; + mov.b64 {%r1851,%dummy}, %rd2761; + } + shf.l.wrap.b32 %r1852, %r1851, %r1850, 14; + shf.l.wrap.b32 %r1853, %r1850, %r1851, 14; + mov.b64 %rd2765, {%r1853, %r1852}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1854}, %rd2740; + } + { + .reg .b32 %dummy; + mov.b64 {%r1855,%dummy}, %rd2740; + } + shf.l.wrap.b32 %r1856, %r1855, %r1854, 28; + shf.l.wrap.b32 %r1857, %r1854, %r1855, 28; + mov.b64 %rd2766, {%r1857, %r1856}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1858}, %rd2746; + } + { + .reg .b32 %dummy; + mov.b64 {%r1859,%dummy}, %rd2746; + } + shf.l.wrap.b32 %r1860, %r1859, %r1858, 20; + shf.l.wrap.b32 %r1861, %r1858, %r1859, 20; + mov.b64 %rd2767, {%r1861, %r1860}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1862}, %rd2747; + } + { + .reg .b32 %dummy; + mov.b64 {%r1863,%dummy}, %rd2747; + } + shf.l.wrap.b32 %r1864, %r1863, %r1862, 3; + shf.l.wrap.b32 %r1865, %r1862, %r1863, 3; + mov.b64 %rd2768, {%r1865, %r1864}; + { + .reg .b32 %dummy; + mov.b64 {%r1866,%dummy}, %rd2753; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1867}, %rd2753; + } + shf.r.wrap.b32 %r1868, %r1867, %r1866, 19; + shf.r.wrap.b32 %r1869, %r1866, %r1867, 19; + mov.b64 %rd2769, {%r1869, %r1868}; + { + .reg .b32 %dummy; + mov.b64 {%r1870,%dummy}, %rd2759; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1871}, %rd2759; + } + shf.r.wrap.b32 %r1872, %r1871, %r1870, 3; + shf.r.wrap.b32 %r1873, %r1870, %r1871, 3; + mov.b64 %rd2770, {%r1873, %r1872}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1874}, %rd2738; + } + { + .reg .b32 %dummy; + mov.b64 {%r1875,%dummy}, %rd2738; + } + shf.l.wrap.b32 %r1876, %r1875, %r1874, 1; + shf.l.wrap.b32 %r1877, %r1874, %r1875, 1; + mov.b64 %rd2771, {%r1877, %r1876}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1878}, %rd2744; + } + { + .reg .b32 %dummy; + mov.b64 {%r1879,%dummy}, %rd2744; + } + shf.l.wrap.b32 %r1880, %r1879, %r1878, 6; + shf.l.wrap.b32 %r1881, %r1878, %r1879, 6; + mov.b64 %rd2772, {%r1881, %r1880}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1882}, %rd2750; + } + { + .reg .b32 %dummy; + mov.b64 {%r1883,%dummy}, %rd2750; + } + shf.l.wrap.b32 %r1884, %r1883, %r1882, 25; + shf.l.wrap.b32 %r1885, %r1882, %r1883, 25; + mov.b64 %rd2773, {%r1885, %r1884}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1886}, %rd2756; + } + { + .reg .b32 %dummy; + mov.b64 {%r1887,%dummy}, %rd2756; + } + shf.l.wrap.b32 %r1888, %r1887, %r1886, 8; + shf.l.wrap.b32 %r1889, %r1886, %r1887, 8; + mov.b64 %rd2774, {%r1889, %r1888}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1890}, %rd2757; + } + { + .reg .b32 %dummy; + mov.b64 {%r1891,%dummy}, %rd2757; + } + shf.l.wrap.b32 %r1892, %r1891, %r1890, 18; + shf.l.wrap.b32 %r1893, %r1890, %r1891, 18; + mov.b64 %rd2775, {%r1893, %r1892}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1894}, %rd2741; + } + { + .reg .b32 %dummy; + mov.b64 {%r1895,%dummy}, %rd2741; + } + shf.l.wrap.b32 %r1896, %r1895, %r1894, 27; + shf.l.wrap.b32 %r1897, %r1894, %r1895, 27; + mov.b64 %rd2776, {%r1897, %r1896}; + { + .reg .b32 %dummy; + mov.b64 {%r1898,%dummy}, %rd2742; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1899}, %rd2742; + } + shf.r.wrap.b32 %r1900, %r1899, %r1898, 28; + shf.r.wrap.b32 %r1901, %r1898, %r1899, 28; + mov.b64 %rd2777, {%r1901, %r1900}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1902}, %rd2748; + } + { + .reg .b32 %dummy; + mov.b64 {%r1903,%dummy}, %rd2748; + } + shf.l.wrap.b32 %r1904, %r1903, %r1902, 10; + shf.l.wrap.b32 %r1905, %r1902, %r1903, 10; + mov.b64 %rd2778, {%r1905, %r1904}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1906}, %rd2754; + } + { + .reg .b32 %dummy; + mov.b64 {%r1907,%dummy}, %rd2754; + } + shf.l.wrap.b32 %r1908, %r1907, %r1906, 15; + shf.l.wrap.b32 %r1909, %r1906, %r1907, 15; + mov.b64 %rd2779, {%r1909, %r1908}; + { + .reg .b32 %dummy; + mov.b64 {%r1910,%dummy}, %rd2760; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1911}, %rd2760; + } + shf.r.wrap.b32 %r1912, %r1911, %r1910, 8; + shf.r.wrap.b32 %r1913, %r1910, %r1911, 8; + mov.b64 %rd2780, {%r1913, %r1912}; + { + .reg .b32 %dummy; + mov.b64 {%r1914,%dummy}, %rd2739; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1915}, %rd2739; + } + shf.r.wrap.b32 %r1916, %r1915, %r1914, 2; + shf.r.wrap.b32 %r1917, %r1914, %r1915, 2; + mov.b64 %rd2781, {%r1917, %r1916}; + { + .reg .b32 %dummy; + mov.b64 {%r1918,%dummy}, %rd2745; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1919}, %rd2745; + } + shf.r.wrap.b32 %r1920, %r1919, %r1918, 9; + shf.r.wrap.b32 %r1921, %r1918, %r1919, 9; + mov.b64 %rd2782, {%r1921, %r1920}; + { + .reg .b32 %dummy; + mov.b64 {%r1922,%dummy}, %rd2751; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1923}, %rd2751; + } + shf.r.wrap.b32 %r1924, %r1923, %r1922, 25; + shf.r.wrap.b32 %r1925, %r1922, %r1923, 25; + mov.b64 %rd2783, {%r1925, %r1924}; + { + .reg .b32 %dummy; + mov.b64 {%r1926,%dummy}, %rd2752; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1927}, %rd2752; + } + shf.r.wrap.b32 %r1928, %r1927, %r1926, 23; + shf.r.wrap.b32 %r1929, %r1926, %r1927, 23; + mov.b64 %rd2784, {%r1929, %r1928}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1930}, %rd2758; + } + { + .reg .b32 %dummy; + mov.b64 {%r1931,%dummy}, %rd2758; + } + shf.l.wrap.b32 %r1932, %r1931, %r1930, 2; + shf.l.wrap.b32 %r1933, %r1930, %r1931, 2; + mov.b64 %rd2785, {%r1933, %r1932}; + not.b64 %rd2786, %rd2762; + and.b64 %rd2787, %rd2763, %rd2786; + xor.b64 %rd2788, %rd2787, %rd2737; + not.b64 %rd2789, %rd2763; + and.b64 %rd2790, %rd2764, %rd2789; + xor.b64 %rd2791, %rd2790, %rd2762; + not.b64 %rd2792, %rd2764; + and.b64 %rd2793, %rd2765, %rd2792; + xor.b64 %rd2794, %rd2793, %rd2763; + not.b64 %rd2795, %rd2765; + and.b64 %rd2796, %rd2737, %rd2795; + xor.b64 %rd2797, %rd2796, %rd2764; + not.b64 %rd2798, %rd2737; + and.b64 %rd2799, %rd2762, %rd2798; + xor.b64 %rd2800, %rd2799, %rd2765; + not.b64 %rd2801, %rd2767; + and.b64 %rd2802, %rd2768, %rd2801; + xor.b64 %rd2803, %rd2802, %rd2766; + not.b64 %rd2804, %rd2768; + and.b64 %rd2805, %rd2769, %rd2804; + xor.b64 %rd2806, %rd2805, %rd2767; + not.b64 %rd2807, %rd2769; + and.b64 %rd2808, %rd2770, %rd2807; + xor.b64 %rd2809, %rd2808, %rd2768; + not.b64 %rd2810, %rd2770; + and.b64 %rd2811, %rd2766, %rd2810; + xor.b64 %rd2812, %rd2811, %rd2769; + not.b64 %rd2813, %rd2766; + and.b64 %rd2814, %rd2767, %rd2813; + xor.b64 %rd2815, %rd2814, %rd2770; + not.b64 %rd2816, %rd2772; + and.b64 %rd2817, %rd2773, %rd2816; + xor.b64 %rd2818, %rd2817, %rd2771; + not.b64 %rd2819, %rd2773; + and.b64 %rd2820, %rd2774, %rd2819; + xor.b64 %rd2821, %rd2820, %rd2772; + not.b64 %rd2822, %rd2774; + and.b64 %rd2823, %rd2775, %rd2822; + xor.b64 %rd2824, %rd2823, %rd2773; + not.b64 %rd2825, %rd2775; + and.b64 %rd2826, %rd2771, %rd2825; + xor.b64 %rd2827, %rd2826, %rd2774; + not.b64 %rd2828, %rd2771; + and.b64 %rd2829, %rd2772, %rd2828; + xor.b64 %rd2830, %rd2829, %rd2775; + not.b64 %rd2831, %rd2777; + and.b64 %rd2832, %rd2778, %rd2831; + xor.b64 %rd2833, %rd2832, %rd2776; + not.b64 %rd2834, %rd2778; + and.b64 %rd2835, %rd2779, %rd2834; + xor.b64 %rd2836, %rd2835, %rd2777; + not.b64 %rd2837, %rd2779; + and.b64 %rd2838, %rd2780, %rd2837; + xor.b64 %rd2839, %rd2838, %rd2778; + not.b64 %rd2840, %rd2780; + and.b64 %rd2841, %rd2776, %rd2840; + xor.b64 %rd2842, %rd2841, %rd2779; + not.b64 %rd2843, %rd2776; + and.b64 %rd2844, %rd2777, %rd2843; + xor.b64 %rd2845, %rd2844, %rd2780; + not.b64 %rd2846, %rd2782; + and.b64 %rd2847, %rd2783, %rd2846; + xor.b64 %rd2848, %rd2847, %rd2781; + not.b64 %rd2849, %rd2783; + and.b64 %rd2850, %rd2784, %rd2849; + xor.b64 %rd2851, %rd2850, %rd2782; + not.b64 %rd2852, %rd2784; + and.b64 %rd2853, %rd2785, %rd2852; + xor.b64 %rd2854, %rd2853, %rd2783; + not.b64 %rd2855, %rd2785; + and.b64 %rd2856, %rd2781, %rd2855; + xor.b64 %rd2857, %rd2856, %rd2784; + not.b64 %rd2858, %rd2781; + and.b64 %rd2859, %rd2782, %rd2858; + xor.b64 %rd2860, %rd2859, %rd2785; + xor.b64 %rd2861, %rd2788, %rd31; + xor.b64 %rd2862, %rd2803, %rd2861; + xor.b64 %rd2863, %rd2862, %rd2818; + xor.b64 %rd2864, %rd2863, %rd2833; + xor.b64 %rd2865, %rd2864, %rd2848; + xor.b64 %rd2866, %rd2806, %rd2791; + xor.b64 %rd2867, %rd2866, %rd2821; + xor.b64 %rd2868, %rd2867, %rd2836; + xor.b64 %rd2869, %rd2868, %rd2851; + xor.b64 %rd2870, %rd2809, %rd2794; + xor.b64 %rd2871, %rd2870, %rd2824; + xor.b64 %rd2872, %rd2871, %rd2839; + xor.b64 %rd2873, %rd2872, %rd2854; + xor.b64 %rd2874, %rd2812, %rd2797; + xor.b64 %rd2875, %rd2874, %rd2827; + xor.b64 %rd2876, %rd2875, %rd2842; + xor.b64 %rd2877, %rd2876, %rd2857; + xor.b64 %rd2878, %rd2815, %rd2800; + xor.b64 %rd2879, %rd2878, %rd2830; + xor.b64 %rd2880, %rd2879, %rd2845; + xor.b64 %rd2881, %rd2880, %rd2860; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1934}, %rd2869; + } + { + .reg .b32 %dummy; + mov.b64 {%r1935,%dummy}, %rd2869; + } + shf.l.wrap.b32 %r1936, %r1935, %r1934, 1; + shf.l.wrap.b32 %r1937, %r1934, %r1935, 1; + mov.b64 %rd2882, {%r1937, %r1936}; + xor.b64 %rd2883, %rd2881, %rd2882; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1938}, %rd2873; + } + { + .reg .b32 %dummy; + mov.b64 {%r1939,%dummy}, %rd2873; + } + shf.l.wrap.b32 %r1940, %r1939, %r1938, 1; + shf.l.wrap.b32 %r1941, %r1938, %r1939, 1; + mov.b64 %rd2884, {%r1941, %r1940}; + xor.b64 %rd2885, %rd2884, %rd2865; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1942}, %rd2877; + } + { + .reg .b32 %dummy; + mov.b64 {%r1943,%dummy}, %rd2877; + } + shf.l.wrap.b32 %r1944, %r1943, %r1942, 1; + shf.l.wrap.b32 %r1945, %r1942, %r1943, 1; + mov.b64 %rd2886, {%r1945, %r1944}; + xor.b64 %rd2887, %rd2886, %rd2869; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1946}, %rd2881; + } + { + .reg .b32 %dummy; + mov.b64 {%r1947,%dummy}, %rd2881; + } + shf.l.wrap.b32 %r1948, %r1947, %r1946, 1; + shf.l.wrap.b32 %r1949, %r1946, %r1947, 1; + mov.b64 %rd2888, {%r1949, %r1948}; + xor.b64 %rd2889, %rd2888, %rd2873; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1950}, %rd2865; + } + { + .reg .b32 %dummy; + mov.b64 {%r1951,%dummy}, %rd2865; + } + shf.l.wrap.b32 %r1952, %r1951, %r1950, 1; + shf.l.wrap.b32 %r1953, %r1950, %r1951, 1; + mov.b64 %rd2890, {%r1953, %r1952}; + xor.b64 %rd2891, %rd2877, %rd2890; + xor.b64 %rd2892, %rd2883, %rd2861; + xor.b64 %rd2893, %rd2885, %rd2791; + xor.b64 %rd2894, %rd2887, %rd2794; + xor.b64 %rd2895, %rd2889, %rd2797; + xor.b64 %rd2896, %rd2891, %rd2800; + xor.b64 %rd2897, %rd2883, %rd2803; + xor.b64 %rd2898, %rd2885, %rd2806; + xor.b64 %rd2899, %rd2887, %rd2809; + xor.b64 %rd2900, %rd2889, %rd2812; + xor.b64 %rd2901, %rd2891, %rd2815; + xor.b64 %rd2902, %rd2883, %rd2818; + xor.b64 %rd2903, %rd2885, %rd2821; + xor.b64 %rd2904, %rd2887, %rd2824; + xor.b64 %rd2905, %rd2889, %rd2827; + xor.b64 %rd2906, %rd2891, %rd2830; + xor.b64 %rd2907, %rd2883, %rd2833; + xor.b64 %rd2908, %rd2885, %rd2836; + xor.b64 %rd2909, %rd2887, %rd2839; + xor.b64 %rd2910, %rd2889, %rd2842; + xor.b64 %rd2911, %rd2891, %rd2845; + xor.b64 %rd2912, %rd2883, %rd2848; + xor.b64 %rd2913, %rd2885, %rd2851; + xor.b64 %rd2914, %rd2887, %rd2854; + xor.b64 %rd2915, %rd2889, %rd2857; + xor.b64 %rd2916, %rd2891, %rd2860; + { + .reg .b32 %dummy; + mov.b64 {%r1954,%dummy}, %rd2898; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1955}, %rd2898; + } + shf.r.wrap.b32 %r1956, %r1955, %r1954, 20; + shf.r.wrap.b32 %r1957, %r1954, %r1955, 20; + mov.b64 %rd2917, {%r1957, %r1956}; + { + .reg .b32 %dummy; + mov.b64 {%r1958,%dummy}, %rd2904; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1959}, %rd2904; + } + shf.r.wrap.b32 %r1960, %r1959, %r1958, 21; + shf.r.wrap.b32 %r1961, %r1958, %r1959, 21; + mov.b64 %rd2918, {%r1961, %r1960}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1962}, %rd2910; + } + { + .reg .b32 %dummy; + mov.b64 {%r1963,%dummy}, %rd2910; + } + shf.l.wrap.b32 %r1964, %r1963, %r1962, 21; + shf.l.wrap.b32 %r1965, %r1962, %r1963, 21; + mov.b64 %rd2919, {%r1965, %r1964}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1966}, %rd2916; + } + { + .reg .b32 %dummy; + mov.b64 {%r1967,%dummy}, %rd2916; + } + shf.l.wrap.b32 %r1968, %r1967, %r1966, 14; + shf.l.wrap.b32 %r1969, %r1966, %r1967, 14; + mov.b64 %rd2920, {%r1969, %r1968}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1970}, %rd2895; + } + { + .reg .b32 %dummy; + mov.b64 {%r1971,%dummy}, %rd2895; + } + shf.l.wrap.b32 %r1972, %r1971, %r1970, 28; + shf.l.wrap.b32 %r1973, %r1970, %r1971, 28; + mov.b64 %rd2921, {%r1973, %r1972}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1974}, %rd2901; + } + { + .reg .b32 %dummy; + mov.b64 {%r1975,%dummy}, %rd2901; + } + shf.l.wrap.b32 %r1976, %r1975, %r1974, 20; + shf.l.wrap.b32 %r1977, %r1974, %r1975, 20; + mov.b64 %rd2922, {%r1977, %r1976}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1978}, %rd2902; + } + { + .reg .b32 %dummy; + mov.b64 {%r1979,%dummy}, %rd2902; + } + shf.l.wrap.b32 %r1980, %r1979, %r1978, 3; + shf.l.wrap.b32 %r1981, %r1978, %r1979, 3; + mov.b64 %rd2923, {%r1981, %r1980}; + { + .reg .b32 %dummy; + mov.b64 {%r1982,%dummy}, %rd2908; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1983}, %rd2908; + } + shf.r.wrap.b32 %r1984, %r1983, %r1982, 19; + shf.r.wrap.b32 %r1985, %r1982, %r1983, 19; + mov.b64 %rd2924, {%r1985, %r1984}; + { + .reg .b32 %dummy; + mov.b64 {%r1986,%dummy}, %rd2914; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1987}, %rd2914; + } + shf.r.wrap.b32 %r1988, %r1987, %r1986, 3; + shf.r.wrap.b32 %r1989, %r1986, %r1987, 3; + mov.b64 %rd2925, {%r1989, %r1988}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1990}, %rd2893; + } + { + .reg .b32 %dummy; + mov.b64 {%r1991,%dummy}, %rd2893; + } + shf.l.wrap.b32 %r1992, %r1991, %r1990, 1; + shf.l.wrap.b32 %r1993, %r1990, %r1991, 1; + mov.b64 %rd2926, {%r1993, %r1992}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1994}, %rd2899; + } + { + .reg .b32 %dummy; + mov.b64 {%r1995,%dummy}, %rd2899; + } + shf.l.wrap.b32 %r1996, %r1995, %r1994, 6; + shf.l.wrap.b32 %r1997, %r1994, %r1995, 6; + mov.b64 %rd2927, {%r1997, %r1996}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r1998}, %rd2905; + } + { + .reg .b32 %dummy; + mov.b64 {%r1999,%dummy}, %rd2905; + } + shf.l.wrap.b32 %r2000, %r1999, %r1998, 25; + shf.l.wrap.b32 %r2001, %r1998, %r1999, 25; + mov.b64 %rd2928, {%r2001, %r2000}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2002}, %rd2911; + } + { + .reg .b32 %dummy; + mov.b64 {%r2003,%dummy}, %rd2911; + } + shf.l.wrap.b32 %r2004, %r2003, %r2002, 8; + shf.l.wrap.b32 %r2005, %r2002, %r2003, 8; + mov.b64 %rd2929, {%r2005, %r2004}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2006}, %rd2912; + } + { + .reg .b32 %dummy; + mov.b64 {%r2007,%dummy}, %rd2912; + } + shf.l.wrap.b32 %r2008, %r2007, %r2006, 18; + shf.l.wrap.b32 %r2009, %r2006, %r2007, 18; + mov.b64 %rd2930, {%r2009, %r2008}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2010}, %rd2896; + } + { + .reg .b32 %dummy; + mov.b64 {%r2011,%dummy}, %rd2896; + } + shf.l.wrap.b32 %r2012, %r2011, %r2010, 27; + shf.l.wrap.b32 %r2013, %r2010, %r2011, 27; + mov.b64 %rd2931, {%r2013, %r2012}; + { + .reg .b32 %dummy; + mov.b64 {%r2014,%dummy}, %rd2897; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2015}, %rd2897; + } + shf.r.wrap.b32 %r2016, %r2015, %r2014, 28; + shf.r.wrap.b32 %r2017, %r2014, %r2015, 28; + mov.b64 %rd2932, {%r2017, %r2016}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2018}, %rd2903; + } + { + .reg .b32 %dummy; + mov.b64 {%r2019,%dummy}, %rd2903; + } + shf.l.wrap.b32 %r2020, %r2019, %r2018, 10; + shf.l.wrap.b32 %r2021, %r2018, %r2019, 10; + mov.b64 %rd2933, {%r2021, %r2020}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2022}, %rd2909; + } + { + .reg .b32 %dummy; + mov.b64 {%r2023,%dummy}, %rd2909; + } + shf.l.wrap.b32 %r2024, %r2023, %r2022, 15; + shf.l.wrap.b32 %r2025, %r2022, %r2023, 15; + mov.b64 %rd2934, {%r2025, %r2024}; + { + .reg .b32 %dummy; + mov.b64 {%r2026,%dummy}, %rd2915; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2027}, %rd2915; + } + shf.r.wrap.b32 %r2028, %r2027, %r2026, 8; + shf.r.wrap.b32 %r2029, %r2026, %r2027, 8; + mov.b64 %rd2935, {%r2029, %r2028}; + { + .reg .b32 %dummy; + mov.b64 {%r2030,%dummy}, %rd2894; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2031}, %rd2894; + } + shf.r.wrap.b32 %r2032, %r2031, %r2030, 2; + shf.r.wrap.b32 %r2033, %r2030, %r2031, 2; + mov.b64 %rd2936, {%r2033, %r2032}; + { + .reg .b32 %dummy; + mov.b64 {%r2034,%dummy}, %rd2900; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2035}, %rd2900; + } + shf.r.wrap.b32 %r2036, %r2035, %r2034, 9; + shf.r.wrap.b32 %r2037, %r2034, %r2035, 9; + mov.b64 %rd2937, {%r2037, %r2036}; + { + .reg .b32 %dummy; + mov.b64 {%r2038,%dummy}, %rd2906; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2039}, %rd2906; + } + shf.r.wrap.b32 %r2040, %r2039, %r2038, 25; + shf.r.wrap.b32 %r2041, %r2038, %r2039, 25; + mov.b64 %rd2938, {%r2041, %r2040}; + { + .reg .b32 %dummy; + mov.b64 {%r2042,%dummy}, %rd2907; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2043}, %rd2907; + } + shf.r.wrap.b32 %r2044, %r2043, %r2042, 23; + shf.r.wrap.b32 %r2045, %r2042, %r2043, 23; + mov.b64 %rd2939, {%r2045, %r2044}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2046}, %rd2913; + } + { + .reg .b32 %dummy; + mov.b64 {%r2047,%dummy}, %rd2913; + } + shf.l.wrap.b32 %r2048, %r2047, %r2046, 2; + shf.l.wrap.b32 %r2049, %r2046, %r2047, 2; + mov.b64 %rd2940, {%r2049, %r2048}; + not.b64 %rd2941, %rd2917; + and.b64 %rd2942, %rd2918, %rd2941; + xor.b64 %rd2943, %rd2942, %rd2892; + not.b64 %rd2944, %rd2918; + and.b64 %rd2945, %rd2919, %rd2944; + xor.b64 %rd2946, %rd2945, %rd2917; + not.b64 %rd2947, %rd2919; + and.b64 %rd2948, %rd2920, %rd2947; + xor.b64 %rd2949, %rd2948, %rd2918; + not.b64 %rd2950, %rd2920; + and.b64 %rd2951, %rd2892, %rd2950; + xor.b64 %rd2952, %rd2951, %rd2919; + not.b64 %rd2953, %rd2892; + and.b64 %rd2954, %rd2917, %rd2953; + xor.b64 %rd2955, %rd2954, %rd2920; + not.b64 %rd2956, %rd2922; + and.b64 %rd2957, %rd2923, %rd2956; + xor.b64 %rd2958, %rd2957, %rd2921; + not.b64 %rd2959, %rd2923; + and.b64 %rd2960, %rd2924, %rd2959; + xor.b64 %rd2961, %rd2960, %rd2922; + not.b64 %rd2962, %rd2924; + and.b64 %rd2963, %rd2925, %rd2962; + xor.b64 %rd2964, %rd2963, %rd2923; + not.b64 %rd2965, %rd2925; + and.b64 %rd2966, %rd2921, %rd2965; + xor.b64 %rd2967, %rd2966, %rd2924; + not.b64 %rd2968, %rd2921; + and.b64 %rd2969, %rd2922, %rd2968; + xor.b64 %rd2970, %rd2969, %rd2925; + not.b64 %rd2971, %rd2927; + and.b64 %rd2972, %rd2928, %rd2971; + xor.b64 %rd2973, %rd2972, %rd2926; + not.b64 %rd2974, %rd2928; + and.b64 %rd2975, %rd2929, %rd2974; + xor.b64 %rd2976, %rd2975, %rd2927; + not.b64 %rd2977, %rd2929; + and.b64 %rd2978, %rd2930, %rd2977; + xor.b64 %rd2979, %rd2978, %rd2928; + not.b64 %rd2980, %rd2930; + and.b64 %rd2981, %rd2926, %rd2980; + xor.b64 %rd2982, %rd2981, %rd2929; + not.b64 %rd2983, %rd2926; + and.b64 %rd2984, %rd2927, %rd2983; + xor.b64 %rd2985, %rd2984, %rd2930; + not.b64 %rd2986, %rd2932; + and.b64 %rd2987, %rd2933, %rd2986; + xor.b64 %rd2988, %rd2987, %rd2931; + not.b64 %rd2989, %rd2933; + and.b64 %rd2990, %rd2934, %rd2989; + xor.b64 %rd2991, %rd2990, %rd2932; + not.b64 %rd2992, %rd2934; + and.b64 %rd2993, %rd2935, %rd2992; + xor.b64 %rd2994, %rd2993, %rd2933; + not.b64 %rd2995, %rd2935; + and.b64 %rd2996, %rd2931, %rd2995; + xor.b64 %rd2997, %rd2996, %rd2934; + not.b64 %rd2998, %rd2931; + and.b64 %rd2999, %rd2932, %rd2998; + xor.b64 %rd3000, %rd2999, %rd2935; + not.b64 %rd3001, %rd2937; + and.b64 %rd3002, %rd2938, %rd3001; + xor.b64 %rd3003, %rd3002, %rd2936; + not.b64 %rd3004, %rd2938; + and.b64 %rd3005, %rd2939, %rd3004; + xor.b64 %rd3006, %rd3005, %rd2937; + not.b64 %rd3007, %rd2939; + and.b64 %rd3008, %rd2940, %rd3007; + xor.b64 %rd3009, %rd3008, %rd2938; + not.b64 %rd3010, %rd2940; + and.b64 %rd3011, %rd2936, %rd3010; + xor.b64 %rd3012, %rd3011, %rd2939; + not.b64 %rd3013, %rd2936; + and.b64 %rd3014, %rd2937, %rd3013; + xor.b64 %rd3015, %rd3014, %rd2940; + xor.b64 %rd3016, %rd2943, %rd32; + xor.b64 %rd3017, %rd2958, %rd3016; + xor.b64 %rd3018, %rd3017, %rd2973; + xor.b64 %rd3019, %rd3018, %rd2988; + xor.b64 %rd3020, %rd3019, %rd3003; + xor.b64 %rd3021, %rd2961, %rd2946; + xor.b64 %rd3022, %rd3021, %rd2976; + xor.b64 %rd3023, %rd3022, %rd2991; + xor.b64 %rd3024, %rd3023, %rd3006; + xor.b64 %rd3025, %rd2964, %rd2949; + xor.b64 %rd3026, %rd3025, %rd2979; + xor.b64 %rd3027, %rd3026, %rd2994; + xor.b64 %rd3028, %rd3027, %rd3009; + xor.b64 %rd3029, %rd2967, %rd2952; + xor.b64 %rd3030, %rd3029, %rd2982; + xor.b64 %rd3031, %rd3030, %rd2997; + xor.b64 %rd3032, %rd3031, %rd3012; + xor.b64 %rd3033, %rd2970, %rd2955; + xor.b64 %rd3034, %rd3033, %rd2985; + xor.b64 %rd3035, %rd3034, %rd3000; + xor.b64 %rd3036, %rd3035, %rd3015; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2050}, %rd3024; + } + { + .reg .b32 %dummy; + mov.b64 {%r2051,%dummy}, %rd3024; + } + shf.l.wrap.b32 %r2052, %r2051, %r2050, 1; + shf.l.wrap.b32 %r2053, %r2050, %r2051, 1; + mov.b64 %rd3037, {%r2053, %r2052}; + xor.b64 %rd3038, %rd3036, %rd3037; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2054}, %rd3028; + } + { + .reg .b32 %dummy; + mov.b64 {%r2055,%dummy}, %rd3028; + } + shf.l.wrap.b32 %r2056, %r2055, %r2054, 1; + shf.l.wrap.b32 %r2057, %r2054, %r2055, 1; + mov.b64 %rd3039, {%r2057, %r2056}; + xor.b64 %rd3040, %rd3039, %rd3020; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2058}, %rd3032; + } + { + .reg .b32 %dummy; + mov.b64 {%r2059,%dummy}, %rd3032; + } + shf.l.wrap.b32 %r2060, %r2059, %r2058, 1; + shf.l.wrap.b32 %r2061, %r2058, %r2059, 1; + mov.b64 %rd3041, {%r2061, %r2060}; + xor.b64 %rd3042, %rd3041, %rd3024; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2062}, %rd3036; + } + { + .reg .b32 %dummy; + mov.b64 {%r2063,%dummy}, %rd3036; + } + shf.l.wrap.b32 %r2064, %r2063, %r2062, 1; + shf.l.wrap.b32 %r2065, %r2062, %r2063, 1; + mov.b64 %rd3043, {%r2065, %r2064}; + xor.b64 %rd3044, %rd3043, %rd3028; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2066}, %rd3020; + } + { + .reg .b32 %dummy; + mov.b64 {%r2067,%dummy}, %rd3020; + } + shf.l.wrap.b32 %r2068, %r2067, %r2066, 1; + shf.l.wrap.b32 %r2069, %r2066, %r2067, 1; + mov.b64 %rd3045, {%r2069, %r2068}; + xor.b64 %rd3046, %rd3032, %rd3045; + xor.b64 %rd3047, %rd3038, %rd3016; + xor.b64 %rd3048, %rd3040, %rd2946; + xor.b64 %rd3049, %rd3042, %rd2949; + xor.b64 %rd3050, %rd3044, %rd2952; + xor.b64 %rd3051, %rd3046, %rd2955; + xor.b64 %rd3052, %rd3038, %rd2958; + xor.b64 %rd3053, %rd3040, %rd2961; + xor.b64 %rd3054, %rd3042, %rd2964; + xor.b64 %rd3055, %rd3044, %rd2967; + xor.b64 %rd3056, %rd3046, %rd2970; + xor.b64 %rd3057, %rd3038, %rd2973; + xor.b64 %rd3058, %rd3040, %rd2976; + xor.b64 %rd3059, %rd3042, %rd2979; + xor.b64 %rd3060, %rd3044, %rd2982; + xor.b64 %rd3061, %rd3046, %rd2985; + xor.b64 %rd3062, %rd3038, %rd2988; + xor.b64 %rd3063, %rd3040, %rd2991; + xor.b64 %rd3064, %rd3042, %rd2994; + xor.b64 %rd3065, %rd3044, %rd2997; + xor.b64 %rd3066, %rd3046, %rd3000; + xor.b64 %rd3067, %rd3038, %rd3003; + xor.b64 %rd3068, %rd3040, %rd3006; + xor.b64 %rd3069, %rd3042, %rd3009; + xor.b64 %rd3070, %rd3044, %rd3012; + xor.b64 %rd3071, %rd3046, %rd3015; + { + .reg .b32 %dummy; + mov.b64 {%r2070,%dummy}, %rd3053; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2071}, %rd3053; + } + shf.r.wrap.b32 %r2072, %r2071, %r2070, 20; + shf.r.wrap.b32 %r2073, %r2070, %r2071, 20; + mov.b64 %rd3072, {%r2073, %r2072}; + { + .reg .b32 %dummy; + mov.b64 {%r2074,%dummy}, %rd3059; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2075}, %rd3059; + } + shf.r.wrap.b32 %r2076, %r2075, %r2074, 21; + shf.r.wrap.b32 %r2077, %r2074, %r2075, 21; + mov.b64 %rd3073, {%r2077, %r2076}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2078}, %rd3065; + } + { + .reg .b32 %dummy; + mov.b64 {%r2079,%dummy}, %rd3065; + } + shf.l.wrap.b32 %r2080, %r2079, %r2078, 21; + shf.l.wrap.b32 %r2081, %r2078, %r2079, 21; + mov.b64 %rd3074, {%r2081, %r2080}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2082}, %rd3071; + } + { + .reg .b32 %dummy; + mov.b64 {%r2083,%dummy}, %rd3071; + } + shf.l.wrap.b32 %r2084, %r2083, %r2082, 14; + shf.l.wrap.b32 %r2085, %r2082, %r2083, 14; + mov.b64 %rd3075, {%r2085, %r2084}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2086}, %rd3050; + } + { + .reg .b32 %dummy; + mov.b64 {%r2087,%dummy}, %rd3050; + } + shf.l.wrap.b32 %r2088, %r2087, %r2086, 28; + shf.l.wrap.b32 %r2089, %r2086, %r2087, 28; + mov.b64 %rd3076, {%r2089, %r2088}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2090}, %rd3056; + } + { + .reg .b32 %dummy; + mov.b64 {%r2091,%dummy}, %rd3056; + } + shf.l.wrap.b32 %r2092, %r2091, %r2090, 20; + shf.l.wrap.b32 %r2093, %r2090, %r2091, 20; + mov.b64 %rd3077, {%r2093, %r2092}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2094}, %rd3057; + } + { + .reg .b32 %dummy; + mov.b64 {%r2095,%dummy}, %rd3057; + } + shf.l.wrap.b32 %r2096, %r2095, %r2094, 3; + shf.l.wrap.b32 %r2097, %r2094, %r2095, 3; + mov.b64 %rd3078, {%r2097, %r2096}; + { + .reg .b32 %dummy; + mov.b64 {%r2098,%dummy}, %rd3063; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2099}, %rd3063; + } + shf.r.wrap.b32 %r2100, %r2099, %r2098, 19; + shf.r.wrap.b32 %r2101, %r2098, %r2099, 19; + mov.b64 %rd3079, {%r2101, %r2100}; + { + .reg .b32 %dummy; + mov.b64 {%r2102,%dummy}, %rd3069; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2103}, %rd3069; + } + shf.r.wrap.b32 %r2104, %r2103, %r2102, 3; + shf.r.wrap.b32 %r2105, %r2102, %r2103, 3; + mov.b64 %rd3080, {%r2105, %r2104}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2106}, %rd3048; + } + { + .reg .b32 %dummy; + mov.b64 {%r2107,%dummy}, %rd3048; + } + shf.l.wrap.b32 %r2108, %r2107, %r2106, 1; + shf.l.wrap.b32 %r2109, %r2106, %r2107, 1; + mov.b64 %rd3081, {%r2109, %r2108}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2110}, %rd3054; + } + { + .reg .b32 %dummy; + mov.b64 {%r2111,%dummy}, %rd3054; + } + shf.l.wrap.b32 %r2112, %r2111, %r2110, 6; + shf.l.wrap.b32 %r2113, %r2110, %r2111, 6; + mov.b64 %rd3082, {%r2113, %r2112}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2114}, %rd3060; + } + { + .reg .b32 %dummy; + mov.b64 {%r2115,%dummy}, %rd3060; + } + shf.l.wrap.b32 %r2116, %r2115, %r2114, 25; + shf.l.wrap.b32 %r2117, %r2114, %r2115, 25; + mov.b64 %rd3083, {%r2117, %r2116}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2118}, %rd3066; + } + { + .reg .b32 %dummy; + mov.b64 {%r2119,%dummy}, %rd3066; + } + shf.l.wrap.b32 %r2120, %r2119, %r2118, 8; + shf.l.wrap.b32 %r2121, %r2118, %r2119, 8; + mov.b64 %rd3084, {%r2121, %r2120}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2122}, %rd3067; + } + { + .reg .b32 %dummy; + mov.b64 {%r2123,%dummy}, %rd3067; + } + shf.l.wrap.b32 %r2124, %r2123, %r2122, 18; + shf.l.wrap.b32 %r2125, %r2122, %r2123, 18; + mov.b64 %rd3085, {%r2125, %r2124}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2126}, %rd3051; + } + { + .reg .b32 %dummy; + mov.b64 {%r2127,%dummy}, %rd3051; + } + shf.l.wrap.b32 %r2128, %r2127, %r2126, 27; + shf.l.wrap.b32 %r2129, %r2126, %r2127, 27; + mov.b64 %rd3086, {%r2129, %r2128}; + { + .reg .b32 %dummy; + mov.b64 {%r2130,%dummy}, %rd3052; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2131}, %rd3052; + } + shf.r.wrap.b32 %r2132, %r2131, %r2130, 28; + shf.r.wrap.b32 %r2133, %r2130, %r2131, 28; + mov.b64 %rd3087, {%r2133, %r2132}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2134}, %rd3058; + } + { + .reg .b32 %dummy; + mov.b64 {%r2135,%dummy}, %rd3058; + } + shf.l.wrap.b32 %r2136, %r2135, %r2134, 10; + shf.l.wrap.b32 %r2137, %r2134, %r2135, 10; + mov.b64 %rd3088, {%r2137, %r2136}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2138}, %rd3064; + } + { + .reg .b32 %dummy; + mov.b64 {%r2139,%dummy}, %rd3064; + } + shf.l.wrap.b32 %r2140, %r2139, %r2138, 15; + shf.l.wrap.b32 %r2141, %r2138, %r2139, 15; + mov.b64 %rd3089, {%r2141, %r2140}; + { + .reg .b32 %dummy; + mov.b64 {%r2142,%dummy}, %rd3070; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2143}, %rd3070; + } + shf.r.wrap.b32 %r2144, %r2143, %r2142, 8; + shf.r.wrap.b32 %r2145, %r2142, %r2143, 8; + mov.b64 %rd3090, {%r2145, %r2144}; + { + .reg .b32 %dummy; + mov.b64 {%r2146,%dummy}, %rd3049; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2147}, %rd3049; + } + shf.r.wrap.b32 %r2148, %r2147, %r2146, 2; + shf.r.wrap.b32 %r2149, %r2146, %r2147, 2; + mov.b64 %rd3091, {%r2149, %r2148}; + { + .reg .b32 %dummy; + mov.b64 {%r2150,%dummy}, %rd3055; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2151}, %rd3055; + } + shf.r.wrap.b32 %r2152, %r2151, %r2150, 9; + shf.r.wrap.b32 %r2153, %r2150, %r2151, 9; + mov.b64 %rd3092, {%r2153, %r2152}; + { + .reg .b32 %dummy; + mov.b64 {%r2154,%dummy}, %rd3061; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2155}, %rd3061; + } + shf.r.wrap.b32 %r2156, %r2155, %r2154, 25; + shf.r.wrap.b32 %r2157, %r2154, %r2155, 25; + mov.b64 %rd3093, {%r2157, %r2156}; + { + .reg .b32 %dummy; + mov.b64 {%r2158,%dummy}, %rd3062; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2159}, %rd3062; + } + shf.r.wrap.b32 %r2160, %r2159, %r2158, 23; + shf.r.wrap.b32 %r2161, %r2158, %r2159, 23; + mov.b64 %rd3094, {%r2161, %r2160}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2162}, %rd3068; + } + { + .reg .b32 %dummy; + mov.b64 {%r2163,%dummy}, %rd3068; + } + shf.l.wrap.b32 %r2164, %r2163, %r2162, 2; + shf.l.wrap.b32 %r2165, %r2162, %r2163, 2; + mov.b64 %rd3095, {%r2165, %r2164}; + not.b64 %rd3096, %rd3072; + and.b64 %rd3097, %rd3073, %rd3096; + xor.b64 %rd3098, %rd3097, %rd3047; + not.b64 %rd3099, %rd3073; + and.b64 %rd3100, %rd3074, %rd3099; + xor.b64 %rd3101, %rd3100, %rd3072; + not.b64 %rd3102, %rd3074; + and.b64 %rd3103, %rd3075, %rd3102; + xor.b64 %rd3104, %rd3103, %rd3073; + not.b64 %rd3105, %rd3075; + and.b64 %rd3106, %rd3047, %rd3105; + xor.b64 %rd3107, %rd3106, %rd3074; + not.b64 %rd3108, %rd3047; + and.b64 %rd3109, %rd3072, %rd3108; + xor.b64 %rd3110, %rd3109, %rd3075; + not.b64 %rd3111, %rd3077; + and.b64 %rd3112, %rd3078, %rd3111; + xor.b64 %rd3113, %rd3112, %rd3076; + not.b64 %rd3114, %rd3078; + and.b64 %rd3115, %rd3079, %rd3114; + xor.b64 %rd3116, %rd3115, %rd3077; + not.b64 %rd3117, %rd3079; + and.b64 %rd3118, %rd3080, %rd3117; + xor.b64 %rd3119, %rd3118, %rd3078; + not.b64 %rd3120, %rd3080; + and.b64 %rd3121, %rd3076, %rd3120; + xor.b64 %rd3122, %rd3121, %rd3079; + not.b64 %rd3123, %rd3076; + and.b64 %rd3124, %rd3077, %rd3123; + xor.b64 %rd3125, %rd3124, %rd3080; + not.b64 %rd3126, %rd3082; + and.b64 %rd3127, %rd3083, %rd3126; + xor.b64 %rd3128, %rd3127, %rd3081; + not.b64 %rd3129, %rd3083; + and.b64 %rd3130, %rd3084, %rd3129; + xor.b64 %rd3131, %rd3130, %rd3082; + not.b64 %rd3132, %rd3084; + and.b64 %rd3133, %rd3085, %rd3132; + xor.b64 %rd3134, %rd3133, %rd3083; + not.b64 %rd3135, %rd3085; + and.b64 %rd3136, %rd3081, %rd3135; + xor.b64 %rd3137, %rd3136, %rd3084; + not.b64 %rd3138, %rd3081; + and.b64 %rd3139, %rd3082, %rd3138; + xor.b64 %rd3140, %rd3139, %rd3085; + not.b64 %rd3141, %rd3087; + and.b64 %rd3142, %rd3088, %rd3141; + xor.b64 %rd3143, %rd3142, %rd3086; + not.b64 %rd3144, %rd3088; + and.b64 %rd3145, %rd3089, %rd3144; + xor.b64 %rd3146, %rd3145, %rd3087; + not.b64 %rd3147, %rd3089; + and.b64 %rd3148, %rd3090, %rd3147; + xor.b64 %rd3149, %rd3148, %rd3088; + not.b64 %rd3150, %rd3090; + and.b64 %rd3151, %rd3086, %rd3150; + xor.b64 %rd3152, %rd3151, %rd3089; + not.b64 %rd3153, %rd3086; + and.b64 %rd3154, %rd3087, %rd3153; + xor.b64 %rd3155, %rd3154, %rd3090; + not.b64 %rd3156, %rd3092; + and.b64 %rd3157, %rd3093, %rd3156; + xor.b64 %rd3158, %rd3157, %rd3091; + not.b64 %rd3159, %rd3093; + and.b64 %rd3160, %rd3094, %rd3159; + xor.b64 %rd3161, %rd3160, %rd3092; + not.b64 %rd3162, %rd3094; + and.b64 %rd3163, %rd3095, %rd3162; + xor.b64 %rd3164, %rd3163, %rd3093; + not.b64 %rd3165, %rd3095; + and.b64 %rd3166, %rd3091, %rd3165; + xor.b64 %rd3167, %rd3166, %rd3094; + not.b64 %rd3168, %rd3091; + and.b64 %rd3169, %rd3092, %rd3168; + xor.b64 %rd3170, %rd3169, %rd3095; + xor.b64 %rd3171, %rd3098, %rd33; + xor.b64 %rd3172, %rd3113, %rd3171; + xor.b64 %rd3173, %rd3172, %rd3128; + xor.b64 %rd3174, %rd3173, %rd3143; + xor.b64 %rd3175, %rd3174, %rd3158; + xor.b64 %rd3176, %rd3116, %rd3101; + xor.b64 %rd3177, %rd3176, %rd3131; + xor.b64 %rd3178, %rd3177, %rd3146; + xor.b64 %rd3179, %rd3178, %rd3161; + xor.b64 %rd3180, %rd3119, %rd3104; + xor.b64 %rd3181, %rd3180, %rd3134; + xor.b64 %rd3182, %rd3181, %rd3149; + xor.b64 %rd3183, %rd3182, %rd3164; + xor.b64 %rd3184, %rd3122, %rd3107; + xor.b64 %rd3185, %rd3184, %rd3137; + xor.b64 %rd3186, %rd3185, %rd3152; + xor.b64 %rd3187, %rd3186, %rd3167; + xor.b64 %rd3188, %rd3125, %rd3110; + xor.b64 %rd3189, %rd3188, %rd3140; + xor.b64 %rd3190, %rd3189, %rd3155; + xor.b64 %rd3191, %rd3190, %rd3170; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2166}, %rd3179; + } + { + .reg .b32 %dummy; + mov.b64 {%r2167,%dummy}, %rd3179; + } + shf.l.wrap.b32 %r2168, %r2167, %r2166, 1; + shf.l.wrap.b32 %r2169, %r2166, %r2167, 1; + mov.b64 %rd3192, {%r2169, %r2168}; + xor.b64 %rd3193, %rd3191, %rd3192; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2170}, %rd3183; + } + { + .reg .b32 %dummy; + mov.b64 {%r2171,%dummy}, %rd3183; + } + shf.l.wrap.b32 %r2172, %r2171, %r2170, 1; + shf.l.wrap.b32 %r2173, %r2170, %r2171, 1; + mov.b64 %rd3194, {%r2173, %r2172}; + xor.b64 %rd3195, %rd3194, %rd3175; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2174}, %rd3187; + } + { + .reg .b32 %dummy; + mov.b64 {%r2175,%dummy}, %rd3187; + } + shf.l.wrap.b32 %r2176, %r2175, %r2174, 1; + shf.l.wrap.b32 %r2177, %r2174, %r2175, 1; + mov.b64 %rd3196, {%r2177, %r2176}; + xor.b64 %rd3197, %rd3196, %rd3179; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2178}, %rd3191; + } + { + .reg .b32 %dummy; + mov.b64 {%r2179,%dummy}, %rd3191; + } + shf.l.wrap.b32 %r2180, %r2179, %r2178, 1; + shf.l.wrap.b32 %r2181, %r2178, %r2179, 1; + mov.b64 %rd3198, {%r2181, %r2180}; + xor.b64 %rd3199, %rd3198, %rd3183; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2182}, %rd3175; + } + { + .reg .b32 %dummy; + mov.b64 {%r2183,%dummy}, %rd3175; + } + shf.l.wrap.b32 %r2184, %r2183, %r2182, 1; + shf.l.wrap.b32 %r2185, %r2182, %r2183, 1; + mov.b64 %rd3200, {%r2185, %r2184}; + xor.b64 %rd3201, %rd3187, %rd3200; + xor.b64 %rd3202, %rd3193, %rd3171; + xor.b64 %rd3203, %rd3195, %rd3101; + xor.b64 %rd3204, %rd3197, %rd3104; + xor.b64 %rd3205, %rd3199, %rd3107; + xor.b64 %rd3206, %rd3201, %rd3110; + xor.b64 %rd3207, %rd3193, %rd3113; + xor.b64 %rd3208, %rd3195, %rd3116; + xor.b64 %rd3209, %rd3197, %rd3119; + xor.b64 %rd3210, %rd3199, %rd3122; + xor.b64 %rd3211, %rd3201, %rd3125; + xor.b64 %rd3212, %rd3193, %rd3128; + xor.b64 %rd3213, %rd3195, %rd3131; + xor.b64 %rd3214, %rd3197, %rd3134; + xor.b64 %rd3215, %rd3199, %rd3137; + xor.b64 %rd3216, %rd3201, %rd3140; + xor.b64 %rd3217, %rd3193, %rd3143; + xor.b64 %rd3218, %rd3195, %rd3146; + xor.b64 %rd3219, %rd3197, %rd3149; + xor.b64 %rd3220, %rd3199, %rd3152; + xor.b64 %rd3221, %rd3201, %rd3155; + xor.b64 %rd3222, %rd3193, %rd3158; + xor.b64 %rd3223, %rd3195, %rd3161; + xor.b64 %rd3224, %rd3197, %rd3164; + xor.b64 %rd3225, %rd3199, %rd3167; + xor.b64 %rd3226, %rd3201, %rd3170; + { + .reg .b32 %dummy; + mov.b64 {%r2186,%dummy}, %rd3208; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2187}, %rd3208; + } + shf.r.wrap.b32 %r2188, %r2187, %r2186, 20; + shf.r.wrap.b32 %r2189, %r2186, %r2187, 20; + mov.b64 %rd3227, {%r2189, %r2188}; + { + .reg .b32 %dummy; + mov.b64 {%r2190,%dummy}, %rd3214; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2191}, %rd3214; + } + shf.r.wrap.b32 %r2192, %r2191, %r2190, 21; + shf.r.wrap.b32 %r2193, %r2190, %r2191, 21; + mov.b64 %rd3228, {%r2193, %r2192}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2194}, %rd3220; + } + { + .reg .b32 %dummy; + mov.b64 {%r2195,%dummy}, %rd3220; + } + shf.l.wrap.b32 %r2196, %r2195, %r2194, 21; + shf.l.wrap.b32 %r2197, %r2194, %r2195, 21; + mov.b64 %rd3229, {%r2197, %r2196}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2198}, %rd3226; + } + { + .reg .b32 %dummy; + mov.b64 {%r2199,%dummy}, %rd3226; + } + shf.l.wrap.b32 %r2200, %r2199, %r2198, 14; + shf.l.wrap.b32 %r2201, %r2198, %r2199, 14; + mov.b64 %rd3230, {%r2201, %r2200}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2202}, %rd3205; + } + { + .reg .b32 %dummy; + mov.b64 {%r2203,%dummy}, %rd3205; + } + shf.l.wrap.b32 %r2204, %r2203, %r2202, 28; + shf.l.wrap.b32 %r2205, %r2202, %r2203, 28; + mov.b64 %rd3231, {%r2205, %r2204}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2206}, %rd3211; + } + { + .reg .b32 %dummy; + mov.b64 {%r2207,%dummy}, %rd3211; + } + shf.l.wrap.b32 %r2208, %r2207, %r2206, 20; + shf.l.wrap.b32 %r2209, %r2206, %r2207, 20; + mov.b64 %rd3232, {%r2209, %r2208}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2210}, %rd3212; + } + { + .reg .b32 %dummy; + mov.b64 {%r2211,%dummy}, %rd3212; + } + shf.l.wrap.b32 %r2212, %r2211, %r2210, 3; + shf.l.wrap.b32 %r2213, %r2210, %r2211, 3; + mov.b64 %rd3233, {%r2213, %r2212}; + { + .reg .b32 %dummy; + mov.b64 {%r2214,%dummy}, %rd3218; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2215}, %rd3218; + } + shf.r.wrap.b32 %r2216, %r2215, %r2214, 19; + shf.r.wrap.b32 %r2217, %r2214, %r2215, 19; + mov.b64 %rd3234, {%r2217, %r2216}; + { + .reg .b32 %dummy; + mov.b64 {%r2218,%dummy}, %rd3224; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2219}, %rd3224; + } + shf.r.wrap.b32 %r2220, %r2219, %r2218, 3; + shf.r.wrap.b32 %r2221, %r2218, %r2219, 3; + mov.b64 %rd3235, {%r2221, %r2220}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2222}, %rd3203; + } + { + .reg .b32 %dummy; + mov.b64 {%r2223,%dummy}, %rd3203; + } + shf.l.wrap.b32 %r2224, %r2223, %r2222, 1; + shf.l.wrap.b32 %r2225, %r2222, %r2223, 1; + mov.b64 %rd3236, {%r2225, %r2224}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2226}, %rd3209; + } + { + .reg .b32 %dummy; + mov.b64 {%r2227,%dummy}, %rd3209; + } + shf.l.wrap.b32 %r2228, %r2227, %r2226, 6; + shf.l.wrap.b32 %r2229, %r2226, %r2227, 6; + mov.b64 %rd3237, {%r2229, %r2228}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2230}, %rd3215; + } + { + .reg .b32 %dummy; + mov.b64 {%r2231,%dummy}, %rd3215; + } + shf.l.wrap.b32 %r2232, %r2231, %r2230, 25; + shf.l.wrap.b32 %r2233, %r2230, %r2231, 25; + mov.b64 %rd3238, {%r2233, %r2232}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2234}, %rd3221; + } + { + .reg .b32 %dummy; + mov.b64 {%r2235,%dummy}, %rd3221; + } + shf.l.wrap.b32 %r2236, %r2235, %r2234, 8; + shf.l.wrap.b32 %r2237, %r2234, %r2235, 8; + mov.b64 %rd3239, {%r2237, %r2236}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2238}, %rd3222; + } + { + .reg .b32 %dummy; + mov.b64 {%r2239,%dummy}, %rd3222; + } + shf.l.wrap.b32 %r2240, %r2239, %r2238, 18; + shf.l.wrap.b32 %r2241, %r2238, %r2239, 18; + mov.b64 %rd3240, {%r2241, %r2240}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2242}, %rd3206; + } + { + .reg .b32 %dummy; + mov.b64 {%r2243,%dummy}, %rd3206; + } + shf.l.wrap.b32 %r2244, %r2243, %r2242, 27; + shf.l.wrap.b32 %r2245, %r2242, %r2243, 27; + mov.b64 %rd3241, {%r2245, %r2244}; + { + .reg .b32 %dummy; + mov.b64 {%r2246,%dummy}, %rd3207; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2247}, %rd3207; + } + shf.r.wrap.b32 %r2248, %r2247, %r2246, 28; + shf.r.wrap.b32 %r2249, %r2246, %r2247, 28; + mov.b64 %rd3242, {%r2249, %r2248}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2250}, %rd3213; + } + { + .reg .b32 %dummy; + mov.b64 {%r2251,%dummy}, %rd3213; + } + shf.l.wrap.b32 %r2252, %r2251, %r2250, 10; + shf.l.wrap.b32 %r2253, %r2250, %r2251, 10; + mov.b64 %rd3243, {%r2253, %r2252}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2254}, %rd3219; + } + { + .reg .b32 %dummy; + mov.b64 {%r2255,%dummy}, %rd3219; + } + shf.l.wrap.b32 %r2256, %r2255, %r2254, 15; + shf.l.wrap.b32 %r2257, %r2254, %r2255, 15; + mov.b64 %rd3244, {%r2257, %r2256}; + { + .reg .b32 %dummy; + mov.b64 {%r2258,%dummy}, %rd3225; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2259}, %rd3225; + } + shf.r.wrap.b32 %r2260, %r2259, %r2258, 8; + shf.r.wrap.b32 %r2261, %r2258, %r2259, 8; + mov.b64 %rd3245, {%r2261, %r2260}; + { + .reg .b32 %dummy; + mov.b64 {%r2262,%dummy}, %rd3204; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2263}, %rd3204; + } + shf.r.wrap.b32 %r2264, %r2263, %r2262, 2; + shf.r.wrap.b32 %r2265, %r2262, %r2263, 2; + mov.b64 %rd3246, {%r2265, %r2264}; + { + .reg .b32 %dummy; + mov.b64 {%r2266,%dummy}, %rd3210; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2267}, %rd3210; + } + shf.r.wrap.b32 %r2268, %r2267, %r2266, 9; + shf.r.wrap.b32 %r2269, %r2266, %r2267, 9; + mov.b64 %rd3247, {%r2269, %r2268}; + { + .reg .b32 %dummy; + mov.b64 {%r2270,%dummy}, %rd3216; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2271}, %rd3216; + } + shf.r.wrap.b32 %r2272, %r2271, %r2270, 25; + shf.r.wrap.b32 %r2273, %r2270, %r2271, 25; + mov.b64 %rd3248, {%r2273, %r2272}; + { + .reg .b32 %dummy; + mov.b64 {%r2274,%dummy}, %rd3217; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2275}, %rd3217; + } + shf.r.wrap.b32 %r2276, %r2275, %r2274, 23; + shf.r.wrap.b32 %r2277, %r2274, %r2275, 23; + mov.b64 %rd3249, {%r2277, %r2276}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2278}, %rd3223; + } + { + .reg .b32 %dummy; + mov.b64 {%r2279,%dummy}, %rd3223; + } + shf.l.wrap.b32 %r2280, %r2279, %r2278, 2; + shf.l.wrap.b32 %r2281, %r2278, %r2279, 2; + mov.b64 %rd3250, {%r2281, %r2280}; + not.b64 %rd3251, %rd3227; + and.b64 %rd3252, %rd3228, %rd3251; + xor.b64 %rd3253, %rd3252, %rd3202; + not.b64 %rd3254, %rd3228; + and.b64 %rd3255, %rd3229, %rd3254; + xor.b64 %rd3256, %rd3255, %rd3227; + not.b64 %rd3257, %rd3229; + and.b64 %rd3258, %rd3230, %rd3257; + xor.b64 %rd3259, %rd3258, %rd3228; + not.b64 %rd3260, %rd3230; + and.b64 %rd3261, %rd3202, %rd3260; + xor.b64 %rd3262, %rd3261, %rd3229; + not.b64 %rd3263, %rd3202; + and.b64 %rd3264, %rd3227, %rd3263; + xor.b64 %rd3265, %rd3264, %rd3230; + not.b64 %rd3266, %rd3232; + and.b64 %rd3267, %rd3233, %rd3266; + xor.b64 %rd3268, %rd3267, %rd3231; + not.b64 %rd3269, %rd3233; + and.b64 %rd3270, %rd3234, %rd3269; + xor.b64 %rd3271, %rd3270, %rd3232; + not.b64 %rd3272, %rd3234; + and.b64 %rd3273, %rd3235, %rd3272; + xor.b64 %rd3274, %rd3273, %rd3233; + not.b64 %rd3275, %rd3235; + and.b64 %rd3276, %rd3231, %rd3275; + xor.b64 %rd3277, %rd3276, %rd3234; + not.b64 %rd3278, %rd3231; + and.b64 %rd3279, %rd3232, %rd3278; + xor.b64 %rd3280, %rd3279, %rd3235; + not.b64 %rd3281, %rd3237; + and.b64 %rd3282, %rd3238, %rd3281; + xor.b64 %rd3283, %rd3282, %rd3236; + not.b64 %rd3284, %rd3238; + and.b64 %rd3285, %rd3239, %rd3284; + xor.b64 %rd3286, %rd3285, %rd3237; + not.b64 %rd3287, %rd3239; + and.b64 %rd3288, %rd3240, %rd3287; + xor.b64 %rd3289, %rd3288, %rd3238; + not.b64 %rd3290, %rd3240; + and.b64 %rd3291, %rd3236, %rd3290; + xor.b64 %rd3292, %rd3291, %rd3239; + not.b64 %rd3293, %rd3236; + and.b64 %rd3294, %rd3237, %rd3293; + xor.b64 %rd3295, %rd3294, %rd3240; + not.b64 %rd3296, %rd3242; + and.b64 %rd3297, %rd3243, %rd3296; + xor.b64 %rd3298, %rd3297, %rd3241; + not.b64 %rd3299, %rd3243; + and.b64 %rd3300, %rd3244, %rd3299; + xor.b64 %rd3301, %rd3300, %rd3242; + not.b64 %rd3302, %rd3244; + and.b64 %rd3303, %rd3245, %rd3302; + xor.b64 %rd3304, %rd3303, %rd3243; + not.b64 %rd3305, %rd3245; + and.b64 %rd3306, %rd3241, %rd3305; + xor.b64 %rd3307, %rd3306, %rd3244; + not.b64 %rd3308, %rd3241; + and.b64 %rd3309, %rd3242, %rd3308; + xor.b64 %rd3310, %rd3309, %rd3245; + not.b64 %rd3311, %rd3247; + and.b64 %rd3312, %rd3248, %rd3311; + xor.b64 %rd3313, %rd3312, %rd3246; + not.b64 %rd3314, %rd3248; + and.b64 %rd3315, %rd3249, %rd3314; + xor.b64 %rd3316, %rd3315, %rd3247; + not.b64 %rd3317, %rd3249; + and.b64 %rd3318, %rd3250, %rd3317; + xor.b64 %rd3319, %rd3318, %rd3248; + not.b64 %rd3320, %rd3250; + and.b64 %rd3321, %rd3246, %rd3320; + xor.b64 %rd3322, %rd3321, %rd3249; + not.b64 %rd3323, %rd3246; + and.b64 %rd3324, %rd3247, %rd3323; + xor.b64 %rd3325, %rd3324, %rd3250; + xor.b64 %rd3326, %rd3253, %rd34; + xor.b64 %rd3327, %rd3268, %rd3326; + xor.b64 %rd3328, %rd3327, %rd3283; + xor.b64 %rd3329, %rd3328, %rd3298; + xor.b64 %rd3330, %rd3329, %rd3313; + xor.b64 %rd3331, %rd3271, %rd3256; + xor.b64 %rd3332, %rd3331, %rd3286; + xor.b64 %rd3333, %rd3332, %rd3301; + xor.b64 %rd3334, %rd3333, %rd3316; + xor.b64 %rd3335, %rd3274, %rd3259; + xor.b64 %rd3336, %rd3335, %rd3289; + xor.b64 %rd3337, %rd3336, %rd3304; + xor.b64 %rd3338, %rd3337, %rd3319; + xor.b64 %rd3339, %rd3277, %rd3262; + xor.b64 %rd3340, %rd3339, %rd3292; + xor.b64 %rd3341, %rd3340, %rd3307; + xor.b64 %rd3342, %rd3341, %rd3322; + xor.b64 %rd3343, %rd3280, %rd3265; + xor.b64 %rd3344, %rd3343, %rd3295; + xor.b64 %rd3345, %rd3344, %rd3310; + xor.b64 %rd3346, %rd3345, %rd3325; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2282}, %rd3334; + } + { + .reg .b32 %dummy; + mov.b64 {%r2283,%dummy}, %rd3334; + } + shf.l.wrap.b32 %r2284, %r2283, %r2282, 1; + shf.l.wrap.b32 %r2285, %r2282, %r2283, 1; + mov.b64 %rd3347, {%r2285, %r2284}; + xor.b64 %rd3348, %rd3346, %rd3347; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2286}, %rd3338; + } + { + .reg .b32 %dummy; + mov.b64 {%r2287,%dummy}, %rd3338; + } + shf.l.wrap.b32 %r2288, %r2287, %r2286, 1; + shf.l.wrap.b32 %r2289, %r2286, %r2287, 1; + mov.b64 %rd3349, {%r2289, %r2288}; + xor.b64 %rd3350, %rd3349, %rd3330; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2290}, %rd3342; + } + { + .reg .b32 %dummy; + mov.b64 {%r2291,%dummy}, %rd3342; + } + shf.l.wrap.b32 %r2292, %r2291, %r2290, 1; + shf.l.wrap.b32 %r2293, %r2290, %r2291, 1; + mov.b64 %rd3351, {%r2293, %r2292}; + xor.b64 %rd3352, %rd3351, %rd3334; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2294}, %rd3346; + } + { + .reg .b32 %dummy; + mov.b64 {%r2295,%dummy}, %rd3346; + } + shf.l.wrap.b32 %r2296, %r2295, %r2294, 1; + shf.l.wrap.b32 %r2297, %r2294, %r2295, 1; + mov.b64 %rd3353, {%r2297, %r2296}; + xor.b64 %rd3354, %rd3353, %rd3338; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2298}, %rd3330; + } + { + .reg .b32 %dummy; + mov.b64 {%r2299,%dummy}, %rd3330; + } + shf.l.wrap.b32 %r2300, %r2299, %r2298, 1; + shf.l.wrap.b32 %r2301, %r2298, %r2299, 1; + mov.b64 %rd3355, {%r2301, %r2300}; + xor.b64 %rd3356, %rd3342, %rd3355; + xor.b64 %rd3357, %rd3348, %rd3326; + xor.b64 %rd3358, %rd3350, %rd3256; + xor.b64 %rd3359, %rd3352, %rd3259; + xor.b64 %rd3360, %rd3354, %rd3262; + xor.b64 %rd3361, %rd3356, %rd3265; + xor.b64 %rd3362, %rd3348, %rd3268; + xor.b64 %rd3363, %rd3350, %rd3271; + xor.b64 %rd3364, %rd3352, %rd3274; + xor.b64 %rd3365, %rd3354, %rd3277; + xor.b64 %rd3366, %rd3356, %rd3280; + xor.b64 %rd3367, %rd3348, %rd3283; + xor.b64 %rd3368, %rd3350, %rd3286; + xor.b64 %rd3369, %rd3352, %rd3289; + xor.b64 %rd3370, %rd3354, %rd3292; + xor.b64 %rd3371, %rd3356, %rd3295; + xor.b64 %rd3372, %rd3348, %rd3298; + xor.b64 %rd3373, %rd3350, %rd3301; + xor.b64 %rd3374, %rd3352, %rd3304; + xor.b64 %rd3375, %rd3354, %rd3307; + xor.b64 %rd3376, %rd3356, %rd3310; + xor.b64 %rd3377, %rd3348, %rd3313; + xor.b64 %rd3378, %rd3350, %rd3316; + xor.b64 %rd3379, %rd3352, %rd3319; + xor.b64 %rd3380, %rd3354, %rd3322; + xor.b64 %rd3381, %rd3356, %rd3325; + { + .reg .b32 %dummy; + mov.b64 {%r2302,%dummy}, %rd3363; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2303}, %rd3363; + } + shf.r.wrap.b32 %r2304, %r2303, %r2302, 20; + shf.r.wrap.b32 %r2305, %r2302, %r2303, 20; + mov.b64 %rd3382, {%r2305, %r2304}; + { + .reg .b32 %dummy; + mov.b64 {%r2306,%dummy}, %rd3369; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2307}, %rd3369; + } + shf.r.wrap.b32 %r2308, %r2307, %r2306, 21; + shf.r.wrap.b32 %r2309, %r2306, %r2307, 21; + mov.b64 %rd3383, {%r2309, %r2308}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2310}, %rd3375; + } + { + .reg .b32 %dummy; + mov.b64 {%r2311,%dummy}, %rd3375; + } + shf.l.wrap.b32 %r2312, %r2311, %r2310, 21; + shf.l.wrap.b32 %r2313, %r2310, %r2311, 21; + mov.b64 %rd3384, {%r2313, %r2312}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2314}, %rd3381; + } + { + .reg .b32 %dummy; + mov.b64 {%r2315,%dummy}, %rd3381; + } + shf.l.wrap.b32 %r2316, %r2315, %r2314, 14; + shf.l.wrap.b32 %r2317, %r2314, %r2315, 14; + mov.b64 %rd3385, {%r2317, %r2316}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2318}, %rd3360; + } + { + .reg .b32 %dummy; + mov.b64 {%r2319,%dummy}, %rd3360; + } + shf.l.wrap.b32 %r2320, %r2319, %r2318, 28; + shf.l.wrap.b32 %r2321, %r2318, %r2319, 28; + mov.b64 %rd3386, {%r2321, %r2320}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2322}, %rd3366; + } + { + .reg .b32 %dummy; + mov.b64 {%r2323,%dummy}, %rd3366; + } + shf.l.wrap.b32 %r2324, %r2323, %r2322, 20; + shf.l.wrap.b32 %r2325, %r2322, %r2323, 20; + mov.b64 %rd3387, {%r2325, %r2324}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2326}, %rd3367; + } + { + .reg .b32 %dummy; + mov.b64 {%r2327,%dummy}, %rd3367; + } + shf.l.wrap.b32 %r2328, %r2327, %r2326, 3; + shf.l.wrap.b32 %r2329, %r2326, %r2327, 3; + mov.b64 %rd3388, {%r2329, %r2328}; + { + .reg .b32 %dummy; + mov.b64 {%r2330,%dummy}, %rd3373; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2331}, %rd3373; + } + shf.r.wrap.b32 %r2332, %r2331, %r2330, 19; + shf.r.wrap.b32 %r2333, %r2330, %r2331, 19; + mov.b64 %rd3389, {%r2333, %r2332}; + { + .reg .b32 %dummy; + mov.b64 {%r2334,%dummy}, %rd3379; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2335}, %rd3379; + } + shf.r.wrap.b32 %r2336, %r2335, %r2334, 3; + shf.r.wrap.b32 %r2337, %r2334, %r2335, 3; + mov.b64 %rd3390, {%r2337, %r2336}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2338}, %rd3358; + } + { + .reg .b32 %dummy; + mov.b64 {%r2339,%dummy}, %rd3358; + } + shf.l.wrap.b32 %r2340, %r2339, %r2338, 1; + shf.l.wrap.b32 %r2341, %r2338, %r2339, 1; + mov.b64 %rd3391, {%r2341, %r2340}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2342}, %rd3364; + } + { + .reg .b32 %dummy; + mov.b64 {%r2343,%dummy}, %rd3364; + } + shf.l.wrap.b32 %r2344, %r2343, %r2342, 6; + shf.l.wrap.b32 %r2345, %r2342, %r2343, 6; + mov.b64 %rd3392, {%r2345, %r2344}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2346}, %rd3370; + } + { + .reg .b32 %dummy; + mov.b64 {%r2347,%dummy}, %rd3370; + } + shf.l.wrap.b32 %r2348, %r2347, %r2346, 25; + shf.l.wrap.b32 %r2349, %r2346, %r2347, 25; + mov.b64 %rd3393, {%r2349, %r2348}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2350}, %rd3376; + } + { + .reg .b32 %dummy; + mov.b64 {%r2351,%dummy}, %rd3376; + } + shf.l.wrap.b32 %r2352, %r2351, %r2350, 8; + shf.l.wrap.b32 %r2353, %r2350, %r2351, 8; + mov.b64 %rd3394, {%r2353, %r2352}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2354}, %rd3377; + } + { + .reg .b32 %dummy; + mov.b64 {%r2355,%dummy}, %rd3377; + } + shf.l.wrap.b32 %r2356, %r2355, %r2354, 18; + shf.l.wrap.b32 %r2357, %r2354, %r2355, 18; + mov.b64 %rd3395, {%r2357, %r2356}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2358}, %rd3361; + } + { + .reg .b32 %dummy; + mov.b64 {%r2359,%dummy}, %rd3361; + } + shf.l.wrap.b32 %r2360, %r2359, %r2358, 27; + shf.l.wrap.b32 %r2361, %r2358, %r2359, 27; + mov.b64 %rd3396, {%r2361, %r2360}; + { + .reg .b32 %dummy; + mov.b64 {%r2362,%dummy}, %rd3362; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2363}, %rd3362; + } + shf.r.wrap.b32 %r2364, %r2363, %r2362, 28; + shf.r.wrap.b32 %r2365, %r2362, %r2363, 28; + mov.b64 %rd3397, {%r2365, %r2364}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2366}, %rd3368; + } + { + .reg .b32 %dummy; + mov.b64 {%r2367,%dummy}, %rd3368; + } + shf.l.wrap.b32 %r2368, %r2367, %r2366, 10; + shf.l.wrap.b32 %r2369, %r2366, %r2367, 10; + mov.b64 %rd3398, {%r2369, %r2368}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2370}, %rd3374; + } + { + .reg .b32 %dummy; + mov.b64 {%r2371,%dummy}, %rd3374; + } + shf.l.wrap.b32 %r2372, %r2371, %r2370, 15; + shf.l.wrap.b32 %r2373, %r2370, %r2371, 15; + mov.b64 %rd3399, {%r2373, %r2372}; + { + .reg .b32 %dummy; + mov.b64 {%r2374,%dummy}, %rd3380; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2375}, %rd3380; + } + shf.r.wrap.b32 %r2376, %r2375, %r2374, 8; + shf.r.wrap.b32 %r2377, %r2374, %r2375, 8; + mov.b64 %rd3400, {%r2377, %r2376}; + { + .reg .b32 %dummy; + mov.b64 {%r2378,%dummy}, %rd3359; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2379}, %rd3359; + } + shf.r.wrap.b32 %r2380, %r2379, %r2378, 2; + shf.r.wrap.b32 %r2381, %r2378, %r2379, 2; + mov.b64 %rd3401, {%r2381, %r2380}; + { + .reg .b32 %dummy; + mov.b64 {%r2382,%dummy}, %rd3365; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2383}, %rd3365; + } + shf.r.wrap.b32 %r2384, %r2383, %r2382, 9; + shf.r.wrap.b32 %r2385, %r2382, %r2383, 9; + mov.b64 %rd3402, {%r2385, %r2384}; + { + .reg .b32 %dummy; + mov.b64 {%r2386,%dummy}, %rd3371; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2387}, %rd3371; + } + shf.r.wrap.b32 %r2388, %r2387, %r2386, 25; + shf.r.wrap.b32 %r2389, %r2386, %r2387, 25; + mov.b64 %rd3403, {%r2389, %r2388}; + { + .reg .b32 %dummy; + mov.b64 {%r2390,%dummy}, %rd3372; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2391}, %rd3372; + } + shf.r.wrap.b32 %r2392, %r2391, %r2390, 23; + shf.r.wrap.b32 %r2393, %r2390, %r2391, 23; + mov.b64 %rd3404, {%r2393, %r2392}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2394}, %rd3378; + } + { + .reg .b32 %dummy; + mov.b64 {%r2395,%dummy}, %rd3378; + } + shf.l.wrap.b32 %r2396, %r2395, %r2394, 2; + shf.l.wrap.b32 %r2397, %r2394, %r2395, 2; + mov.b64 %rd3405, {%r2397, %r2396}; + not.b64 %rd3406, %rd3382; + and.b64 %rd3407, %rd3383, %rd3406; + xor.b64 %rd3408, %rd3407, %rd3357; + not.b64 %rd3409, %rd3383; + and.b64 %rd3410, %rd3384, %rd3409; + xor.b64 %rd3411, %rd3410, %rd3382; + not.b64 %rd3412, %rd3384; + and.b64 %rd3413, %rd3385, %rd3412; + xor.b64 %rd3414, %rd3413, %rd3383; + not.b64 %rd3415, %rd3385; + and.b64 %rd3416, %rd3357, %rd3415; + xor.b64 %rd3417, %rd3416, %rd3384; + not.b64 %rd3418, %rd3357; + and.b64 %rd3419, %rd3382, %rd3418; + xor.b64 %rd3420, %rd3419, %rd3385; + not.b64 %rd3421, %rd3387; + and.b64 %rd3422, %rd3388, %rd3421; + xor.b64 %rd3423, %rd3422, %rd3386; + not.b64 %rd3424, %rd3388; + and.b64 %rd3425, %rd3389, %rd3424; + xor.b64 %rd3426, %rd3425, %rd3387; + not.b64 %rd3427, %rd3389; + and.b64 %rd3428, %rd3390, %rd3427; + xor.b64 %rd3429, %rd3428, %rd3388; + not.b64 %rd3430, %rd3390; + and.b64 %rd3431, %rd3386, %rd3430; + xor.b64 %rd3432, %rd3431, %rd3389; + not.b64 %rd3433, %rd3386; + and.b64 %rd3434, %rd3387, %rd3433; + xor.b64 %rd3435, %rd3434, %rd3390; + not.b64 %rd3436, %rd3392; + and.b64 %rd3437, %rd3393, %rd3436; + xor.b64 %rd3438, %rd3437, %rd3391; + not.b64 %rd3439, %rd3393; + and.b64 %rd3440, %rd3394, %rd3439; + xor.b64 %rd3441, %rd3440, %rd3392; + not.b64 %rd3442, %rd3394; + and.b64 %rd3443, %rd3395, %rd3442; + xor.b64 %rd3444, %rd3443, %rd3393; + not.b64 %rd3445, %rd3395; + and.b64 %rd3446, %rd3391, %rd3445; + xor.b64 %rd3447, %rd3446, %rd3394; + not.b64 %rd3448, %rd3391; + and.b64 %rd3449, %rd3392, %rd3448; + xor.b64 %rd3450, %rd3449, %rd3395; + not.b64 %rd3451, %rd3397; + and.b64 %rd3452, %rd3398, %rd3451; + xor.b64 %rd3453, %rd3452, %rd3396; + not.b64 %rd3454, %rd3398; + and.b64 %rd3455, %rd3399, %rd3454; + xor.b64 %rd3456, %rd3455, %rd3397; + not.b64 %rd3457, %rd3399; + and.b64 %rd3458, %rd3400, %rd3457; + xor.b64 %rd3459, %rd3458, %rd3398; + not.b64 %rd3460, %rd3400; + and.b64 %rd3461, %rd3396, %rd3460; + xor.b64 %rd3462, %rd3461, %rd3399; + not.b64 %rd3463, %rd3396; + and.b64 %rd3464, %rd3397, %rd3463; + xor.b64 %rd3465, %rd3464, %rd3400; + not.b64 %rd3466, %rd3402; + and.b64 %rd3467, %rd3403, %rd3466; + xor.b64 %rd3468, %rd3467, %rd3401; + not.b64 %rd3469, %rd3403; + and.b64 %rd3470, %rd3404, %rd3469; + xor.b64 %rd3471, %rd3470, %rd3402; + not.b64 %rd3472, %rd3404; + and.b64 %rd3473, %rd3405, %rd3472; + xor.b64 %rd3474, %rd3473, %rd3403; + not.b64 %rd3475, %rd3405; + and.b64 %rd3476, %rd3401, %rd3475; + xor.b64 %rd3477, %rd3476, %rd3404; + not.b64 %rd3478, %rd3401; + and.b64 %rd3479, %rd3402, %rd3478; + xor.b64 %rd3480, %rd3479, %rd3405; + xor.b64 %rd3481, %rd3408, %rd35; + xor.b64 %rd3482, %rd3423, %rd3481; + xor.b64 %rd3483, %rd3482, %rd3438; + xor.b64 %rd3484, %rd3483, %rd3453; + xor.b64 %rd3485, %rd3484, %rd3468; + xor.b64 %rd3486, %rd3426, %rd3411; + xor.b64 %rd3487, %rd3486, %rd3441; + xor.b64 %rd3488, %rd3487, %rd3456; + xor.b64 %rd3489, %rd3488, %rd3471; + xor.b64 %rd3490, %rd3429, %rd3414; + xor.b64 %rd3491, %rd3490, %rd3444; + xor.b64 %rd3492, %rd3491, %rd3459; + xor.b64 %rd3493, %rd3492, %rd3474; + xor.b64 %rd3494, %rd3432, %rd3417; + xor.b64 %rd3495, %rd3494, %rd3447; + xor.b64 %rd3496, %rd3495, %rd3462; + xor.b64 %rd3497, %rd3496, %rd3477; + xor.b64 %rd3498, %rd3435, %rd3420; + xor.b64 %rd3499, %rd3498, %rd3450; + xor.b64 %rd3500, %rd3499, %rd3465; + xor.b64 %rd3501, %rd3500, %rd3480; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2398}, %rd3489; + } + { + .reg .b32 %dummy; + mov.b64 {%r2399,%dummy}, %rd3489; + } + shf.l.wrap.b32 %r2400, %r2399, %r2398, 1; + shf.l.wrap.b32 %r2401, %r2398, %r2399, 1; + mov.b64 %rd3502, {%r2401, %r2400}; + xor.b64 %rd3503, %rd3501, %rd3502; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2402}, %rd3493; + } + { + .reg .b32 %dummy; + mov.b64 {%r2403,%dummy}, %rd3493; + } + shf.l.wrap.b32 %r2404, %r2403, %r2402, 1; + shf.l.wrap.b32 %r2405, %r2402, %r2403, 1; + mov.b64 %rd3504, {%r2405, %r2404}; + xor.b64 %rd3505, %rd3504, %rd3485; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2406}, %rd3497; + } + { + .reg .b32 %dummy; + mov.b64 {%r2407,%dummy}, %rd3497; + } + shf.l.wrap.b32 %r2408, %r2407, %r2406, 1; + shf.l.wrap.b32 %r2409, %r2406, %r2407, 1; + mov.b64 %rd3506, {%r2409, %r2408}; + xor.b64 %rd3507, %rd3506, %rd3489; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2410}, %rd3501; + } + { + .reg .b32 %dummy; + mov.b64 {%r2411,%dummy}, %rd3501; + } + shf.l.wrap.b32 %r2412, %r2411, %r2410, 1; + shf.l.wrap.b32 %r2413, %r2410, %r2411, 1; + mov.b64 %rd3508, {%r2413, %r2412}; + xor.b64 %rd3509, %rd3508, %rd3493; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2414}, %rd3485; + } + { + .reg .b32 %dummy; + mov.b64 {%r2415,%dummy}, %rd3485; + } + shf.l.wrap.b32 %r2416, %r2415, %r2414, 1; + shf.l.wrap.b32 %r2417, %r2414, %r2415, 1; + mov.b64 %rd3510, {%r2417, %r2416}; + xor.b64 %rd3511, %rd3497, %rd3510; + xor.b64 %rd3512, %rd3503, %rd3481; + xor.b64 %rd3513, %rd3505, %rd3411; + xor.b64 %rd3514, %rd3507, %rd3414; + xor.b64 %rd3515, %rd3509, %rd3417; + xor.b64 %rd3516, %rd3511, %rd3420; + xor.b64 %rd3517, %rd3503, %rd3423; + xor.b64 %rd3518, %rd3505, %rd3426; + xor.b64 %rd3519, %rd3507, %rd3429; + xor.b64 %rd3520, %rd3509, %rd3432; + xor.b64 %rd3521, %rd3511, %rd3435; + xor.b64 %rd3522, %rd3503, %rd3438; + xor.b64 %rd3523, %rd3505, %rd3441; + xor.b64 %rd3524, %rd3507, %rd3444; + xor.b64 %rd3525, %rd3509, %rd3447; + xor.b64 %rd3526, %rd3511, %rd3450; + xor.b64 %rd3527, %rd3503, %rd3453; + xor.b64 %rd3528, %rd3505, %rd3456; + xor.b64 %rd3529, %rd3507, %rd3459; + xor.b64 %rd3530, %rd3509, %rd3462; + xor.b64 %rd3531, %rd3511, %rd3465; + xor.b64 %rd3532, %rd3503, %rd3468; + xor.b64 %rd3533, %rd3505, %rd3471; + xor.b64 %rd3534, %rd3507, %rd3474; + xor.b64 %rd3535, %rd3509, %rd3477; + xor.b64 %rd3536, %rd3511, %rd3480; + { + .reg .b32 %dummy; + mov.b64 {%r2418,%dummy}, %rd3518; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2419}, %rd3518; + } + shf.r.wrap.b32 %r2420, %r2419, %r2418, 20; + shf.r.wrap.b32 %r2421, %r2418, %r2419, 20; + mov.b64 %rd3537, {%r2421, %r2420}; + { + .reg .b32 %dummy; + mov.b64 {%r2422,%dummy}, %rd3524; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2423}, %rd3524; + } + shf.r.wrap.b32 %r2424, %r2423, %r2422, 21; + shf.r.wrap.b32 %r2425, %r2422, %r2423, 21; + mov.b64 %rd3538, {%r2425, %r2424}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2426}, %rd3530; + } + { + .reg .b32 %dummy; + mov.b64 {%r2427,%dummy}, %rd3530; + } + shf.l.wrap.b32 %r2428, %r2427, %r2426, 21; + shf.l.wrap.b32 %r2429, %r2426, %r2427, 21; + mov.b64 %rd3539, {%r2429, %r2428}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2430}, %rd3536; + } + { + .reg .b32 %dummy; + mov.b64 {%r2431,%dummy}, %rd3536; + } + shf.l.wrap.b32 %r2432, %r2431, %r2430, 14; + shf.l.wrap.b32 %r2433, %r2430, %r2431, 14; + mov.b64 %rd3540, {%r2433, %r2432}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2434}, %rd3515; + } + { + .reg .b32 %dummy; + mov.b64 {%r2435,%dummy}, %rd3515; + } + shf.l.wrap.b32 %r2436, %r2435, %r2434, 28; + shf.l.wrap.b32 %r2437, %r2434, %r2435, 28; + mov.b64 %rd3541, {%r2437, %r2436}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2438}, %rd3521; + } + { + .reg .b32 %dummy; + mov.b64 {%r2439,%dummy}, %rd3521; + } + shf.l.wrap.b32 %r2440, %r2439, %r2438, 20; + shf.l.wrap.b32 %r2441, %r2438, %r2439, 20; + mov.b64 %rd3542, {%r2441, %r2440}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2442}, %rd3522; + } + { + .reg .b32 %dummy; + mov.b64 {%r2443,%dummy}, %rd3522; + } + shf.l.wrap.b32 %r2444, %r2443, %r2442, 3; + shf.l.wrap.b32 %r2445, %r2442, %r2443, 3; + mov.b64 %rd3543, {%r2445, %r2444}; + { + .reg .b32 %dummy; + mov.b64 {%r2446,%dummy}, %rd3528; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2447}, %rd3528; + } + shf.r.wrap.b32 %r2448, %r2447, %r2446, 19; + shf.r.wrap.b32 %r2449, %r2446, %r2447, 19; + mov.b64 %rd3544, {%r2449, %r2448}; + { + .reg .b32 %dummy; + mov.b64 {%r2450,%dummy}, %rd3534; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2451}, %rd3534; + } + shf.r.wrap.b32 %r2452, %r2451, %r2450, 3; + shf.r.wrap.b32 %r2453, %r2450, %r2451, 3; + mov.b64 %rd3545, {%r2453, %r2452}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2454}, %rd3513; + } + { + .reg .b32 %dummy; + mov.b64 {%r2455,%dummy}, %rd3513; + } + shf.l.wrap.b32 %r2456, %r2455, %r2454, 1; + shf.l.wrap.b32 %r2457, %r2454, %r2455, 1; + mov.b64 %rd3546, {%r2457, %r2456}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2458}, %rd3519; + } + { + .reg .b32 %dummy; + mov.b64 {%r2459,%dummy}, %rd3519; + } + shf.l.wrap.b32 %r2460, %r2459, %r2458, 6; + shf.l.wrap.b32 %r2461, %r2458, %r2459, 6; + mov.b64 %rd3547, {%r2461, %r2460}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2462}, %rd3525; + } + { + .reg .b32 %dummy; + mov.b64 {%r2463,%dummy}, %rd3525; + } + shf.l.wrap.b32 %r2464, %r2463, %r2462, 25; + shf.l.wrap.b32 %r2465, %r2462, %r2463, 25; + mov.b64 %rd3548, {%r2465, %r2464}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2466}, %rd3531; + } + { + .reg .b32 %dummy; + mov.b64 {%r2467,%dummy}, %rd3531; + } + shf.l.wrap.b32 %r2468, %r2467, %r2466, 8; + shf.l.wrap.b32 %r2469, %r2466, %r2467, 8; + mov.b64 %rd3549, {%r2469, %r2468}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2470}, %rd3532; + } + { + .reg .b32 %dummy; + mov.b64 {%r2471,%dummy}, %rd3532; + } + shf.l.wrap.b32 %r2472, %r2471, %r2470, 18; + shf.l.wrap.b32 %r2473, %r2470, %r2471, 18; + mov.b64 %rd3550, {%r2473, %r2472}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2474}, %rd3516; + } + { + .reg .b32 %dummy; + mov.b64 {%r2475,%dummy}, %rd3516; + } + shf.l.wrap.b32 %r2476, %r2475, %r2474, 27; + shf.l.wrap.b32 %r2477, %r2474, %r2475, 27; + mov.b64 %rd3551, {%r2477, %r2476}; + { + .reg .b32 %dummy; + mov.b64 {%r2478,%dummy}, %rd3517; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2479}, %rd3517; + } + shf.r.wrap.b32 %r2480, %r2479, %r2478, 28; + shf.r.wrap.b32 %r2481, %r2478, %r2479, 28; + mov.b64 %rd3552, {%r2481, %r2480}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2482}, %rd3523; + } + { + .reg .b32 %dummy; + mov.b64 {%r2483,%dummy}, %rd3523; + } + shf.l.wrap.b32 %r2484, %r2483, %r2482, 10; + shf.l.wrap.b32 %r2485, %r2482, %r2483, 10; + mov.b64 %rd3553, {%r2485, %r2484}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2486}, %rd3529; + } + { + .reg .b32 %dummy; + mov.b64 {%r2487,%dummy}, %rd3529; + } + shf.l.wrap.b32 %r2488, %r2487, %r2486, 15; + shf.l.wrap.b32 %r2489, %r2486, %r2487, 15; + mov.b64 %rd3554, {%r2489, %r2488}; + { + .reg .b32 %dummy; + mov.b64 {%r2490,%dummy}, %rd3535; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2491}, %rd3535; + } + shf.r.wrap.b32 %r2492, %r2491, %r2490, 8; + shf.r.wrap.b32 %r2493, %r2490, %r2491, 8; + mov.b64 %rd3555, {%r2493, %r2492}; + { + .reg .b32 %dummy; + mov.b64 {%r2494,%dummy}, %rd3514; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2495}, %rd3514; + } + shf.r.wrap.b32 %r2496, %r2495, %r2494, 2; + shf.r.wrap.b32 %r2497, %r2494, %r2495, 2; + mov.b64 %rd3556, {%r2497, %r2496}; + { + .reg .b32 %dummy; + mov.b64 {%r2498,%dummy}, %rd3520; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2499}, %rd3520; + } + shf.r.wrap.b32 %r2500, %r2499, %r2498, 9; + shf.r.wrap.b32 %r2501, %r2498, %r2499, 9; + mov.b64 %rd3557, {%r2501, %r2500}; + { + .reg .b32 %dummy; + mov.b64 {%r2502,%dummy}, %rd3526; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2503}, %rd3526; + } + shf.r.wrap.b32 %r2504, %r2503, %r2502, 25; + shf.r.wrap.b32 %r2505, %r2502, %r2503, 25; + mov.b64 %rd3558, {%r2505, %r2504}; + { + .reg .b32 %dummy; + mov.b64 {%r2506,%dummy}, %rd3527; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2507}, %rd3527; + } + shf.r.wrap.b32 %r2508, %r2507, %r2506, 23; + shf.r.wrap.b32 %r2509, %r2506, %r2507, 23; + mov.b64 %rd3559, {%r2509, %r2508}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2510}, %rd3533; + } + { + .reg .b32 %dummy; + mov.b64 {%r2511,%dummy}, %rd3533; + } + shf.l.wrap.b32 %r2512, %r2511, %r2510, 2; + shf.l.wrap.b32 %r2513, %r2510, %r2511, 2; + mov.b64 %rd3560, {%r2513, %r2512}; + not.b64 %rd3561, %rd3537; + and.b64 %rd3562, %rd3538, %rd3561; + xor.b64 %rd3563, %rd3562, %rd3512; + not.b64 %rd3564, %rd3538; + and.b64 %rd3565, %rd3539, %rd3564; + xor.b64 %rd3566, %rd3565, %rd3537; + not.b64 %rd3567, %rd3539; + and.b64 %rd3568, %rd3540, %rd3567; + xor.b64 %rd3569, %rd3568, %rd3538; + not.b64 %rd3570, %rd3540; + and.b64 %rd3571, %rd3512, %rd3570; + xor.b64 %rd3572, %rd3571, %rd3539; + not.b64 %rd3573, %rd3512; + and.b64 %rd3574, %rd3537, %rd3573; + xor.b64 %rd3575, %rd3574, %rd3540; + not.b64 %rd3576, %rd3542; + and.b64 %rd3577, %rd3543, %rd3576; + xor.b64 %rd3578, %rd3577, %rd3541; + not.b64 %rd3579, %rd3543; + and.b64 %rd3580, %rd3544, %rd3579; + xor.b64 %rd3581, %rd3580, %rd3542; + not.b64 %rd3582, %rd3544; + and.b64 %rd3583, %rd3545, %rd3582; + xor.b64 %rd3584, %rd3583, %rd3543; + not.b64 %rd3585, %rd3545; + and.b64 %rd3586, %rd3541, %rd3585; + xor.b64 %rd3587, %rd3586, %rd3544; + not.b64 %rd3588, %rd3541; + and.b64 %rd3589, %rd3542, %rd3588; + xor.b64 %rd3590, %rd3589, %rd3545; + not.b64 %rd3591, %rd3547; + and.b64 %rd3592, %rd3548, %rd3591; + xor.b64 %rd3593, %rd3592, %rd3546; + not.b64 %rd3594, %rd3548; + and.b64 %rd3595, %rd3549, %rd3594; + xor.b64 %rd3596, %rd3595, %rd3547; + not.b64 %rd3597, %rd3549; + and.b64 %rd3598, %rd3550, %rd3597; + xor.b64 %rd3599, %rd3598, %rd3548; + not.b64 %rd3600, %rd3550; + and.b64 %rd3601, %rd3546, %rd3600; + xor.b64 %rd3602, %rd3601, %rd3549; + not.b64 %rd3603, %rd3546; + and.b64 %rd3604, %rd3547, %rd3603; + xor.b64 %rd3605, %rd3604, %rd3550; + not.b64 %rd3606, %rd3552; + and.b64 %rd3607, %rd3553, %rd3606; + xor.b64 %rd3608, %rd3607, %rd3551; + not.b64 %rd3609, %rd3553; + and.b64 %rd3610, %rd3554, %rd3609; + xor.b64 %rd3611, %rd3610, %rd3552; + not.b64 %rd3612, %rd3554; + and.b64 %rd3613, %rd3555, %rd3612; + xor.b64 %rd3614, %rd3613, %rd3553; + not.b64 %rd3615, %rd3555; + and.b64 %rd3616, %rd3551, %rd3615; + xor.b64 %rd3617, %rd3616, %rd3554; + not.b64 %rd3618, %rd3551; + and.b64 %rd3619, %rd3552, %rd3618; + xor.b64 %rd3620, %rd3619, %rd3555; + not.b64 %rd3621, %rd3557; + and.b64 %rd3622, %rd3558, %rd3621; + xor.b64 %rd3623, %rd3622, %rd3556; + not.b64 %rd3624, %rd3558; + and.b64 %rd3625, %rd3559, %rd3624; + xor.b64 %rd3626, %rd3625, %rd3557; + not.b64 %rd3627, %rd3559; + and.b64 %rd3628, %rd3560, %rd3627; + xor.b64 %rd3629, %rd3628, %rd3558; + not.b64 %rd3630, %rd3560; + and.b64 %rd3631, %rd3556, %rd3630; + xor.b64 %rd3632, %rd3631, %rd3559; + not.b64 %rd3633, %rd3556; + and.b64 %rd3634, %rd3557, %rd3633; + xor.b64 %rd3635, %rd3634, %rd3560; + xor.b64 %rd3636, %rd3563, %rd36; + xor.b64 %rd3637, %rd3578, %rd3636; + xor.b64 %rd3638, %rd3637, %rd3593; + xor.b64 %rd3639, %rd3638, %rd3608; + xor.b64 %rd3640, %rd3639, %rd3623; + xor.b64 %rd3641, %rd3581, %rd3566; + xor.b64 %rd3642, %rd3641, %rd3596; + xor.b64 %rd3643, %rd3642, %rd3611; + xor.b64 %rd3644, %rd3643, %rd3626; + xor.b64 %rd3645, %rd3584, %rd3569; + xor.b64 %rd3646, %rd3645, %rd3599; + xor.b64 %rd3647, %rd3646, %rd3614; + xor.b64 %rd3648, %rd3647, %rd3629; + xor.b64 %rd3649, %rd3587, %rd3572; + xor.b64 %rd3650, %rd3649, %rd3602; + xor.b64 %rd3651, %rd3650, %rd3617; + xor.b64 %rd3652, %rd3651, %rd3632; + xor.b64 %rd3653, %rd3590, %rd3575; + xor.b64 %rd3654, %rd3653, %rd3605; + xor.b64 %rd3655, %rd3654, %rd3620; + xor.b64 %rd3656, %rd3655, %rd3635; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2514}, %rd3644; + } + { + .reg .b32 %dummy; + mov.b64 {%r2515,%dummy}, %rd3644; + } + shf.l.wrap.b32 %r2516, %r2515, %r2514, 1; + shf.l.wrap.b32 %r2517, %r2514, %r2515, 1; + mov.b64 %rd3657, {%r2517, %r2516}; + xor.b64 %rd3658, %rd3656, %rd3657; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2518}, %rd3648; + } + { + .reg .b32 %dummy; + mov.b64 {%r2519,%dummy}, %rd3648; + } + shf.l.wrap.b32 %r2520, %r2519, %r2518, 1; + shf.l.wrap.b32 %r2521, %r2518, %r2519, 1; + mov.b64 %rd3659, {%r2521, %r2520}; + xor.b64 %rd3660, %rd3659, %rd3640; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2522}, %rd3652; + } + { + .reg .b32 %dummy; + mov.b64 {%r2523,%dummy}, %rd3652; + } + shf.l.wrap.b32 %r2524, %r2523, %r2522, 1; + shf.l.wrap.b32 %r2525, %r2522, %r2523, 1; + mov.b64 %rd3661, {%r2525, %r2524}; + xor.b64 %rd3662, %rd3661, %rd3644; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2526}, %rd3656; + } + { + .reg .b32 %dummy; + mov.b64 {%r2527,%dummy}, %rd3656; + } + shf.l.wrap.b32 %r2528, %r2527, %r2526, 1; + shf.l.wrap.b32 %r2529, %r2526, %r2527, 1; + mov.b64 %rd3663, {%r2529, %r2528}; + xor.b64 %rd3664, %rd3663, %rd3648; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2530}, %rd3640; + } + { + .reg .b32 %dummy; + mov.b64 {%r2531,%dummy}, %rd3640; + } + shf.l.wrap.b32 %r2532, %r2531, %r2530, 1; + shf.l.wrap.b32 %r2533, %r2530, %r2531, 1; + mov.b64 %rd3665, {%r2533, %r2532}; + xor.b64 %rd3666, %rd3652, %rd3665; + xor.b64 %rd3667, %rd3658, %rd3636; + xor.b64 %rd3668, %rd3660, %rd3566; + xor.b64 %rd3669, %rd3662, %rd3569; + xor.b64 %rd3670, %rd3664, %rd3572; + xor.b64 %rd3671, %rd3666, %rd3575; + xor.b64 %rd3672, %rd3658, %rd3578; + xor.b64 %rd3673, %rd3660, %rd3581; + xor.b64 %rd3674, %rd3662, %rd3584; + xor.b64 %rd3675, %rd3664, %rd3587; + xor.b64 %rd3676, %rd3666, %rd3590; + xor.b64 %rd3677, %rd3658, %rd3593; + xor.b64 %rd3678, %rd3660, %rd3596; + xor.b64 %rd3679, %rd3662, %rd3599; + xor.b64 %rd3680, %rd3664, %rd3602; + xor.b64 %rd3681, %rd3666, %rd3605; + xor.b64 %rd3682, %rd3658, %rd3608; + xor.b64 %rd3683, %rd3660, %rd3611; + xor.b64 %rd3684, %rd3662, %rd3614; + xor.b64 %rd3685, %rd3664, %rd3617; + xor.b64 %rd3686, %rd3666, %rd3620; + xor.b64 %rd3687, %rd3658, %rd3623; + xor.b64 %rd3688, %rd3660, %rd3626; + xor.b64 %rd3689, %rd3662, %rd3629; + xor.b64 %rd3690, %rd3664, %rd3632; + xor.b64 %rd3691, %rd3666, %rd3635; + { + .reg .b32 %dummy; + mov.b64 {%r2534,%dummy}, %rd3673; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2535}, %rd3673; + } + shf.r.wrap.b32 %r2536, %r2535, %r2534, 20; + shf.r.wrap.b32 %r2537, %r2534, %r2535, 20; + mov.b64 %rd3692, {%r2537, %r2536}; + { + .reg .b32 %dummy; + mov.b64 {%r2538,%dummy}, %rd3679; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2539}, %rd3679; + } + shf.r.wrap.b32 %r2540, %r2539, %r2538, 21; + shf.r.wrap.b32 %r2541, %r2538, %r2539, 21; + mov.b64 %rd3693, {%r2541, %r2540}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2542}, %rd3685; + } + { + .reg .b32 %dummy; + mov.b64 {%r2543,%dummy}, %rd3685; + } + shf.l.wrap.b32 %r2544, %r2543, %r2542, 21; + shf.l.wrap.b32 %r2545, %r2542, %r2543, 21; + mov.b64 %rd3694, {%r2545, %r2544}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2546}, %rd3691; + } + { + .reg .b32 %dummy; + mov.b64 {%r2547,%dummy}, %rd3691; + } + shf.l.wrap.b32 %r2548, %r2547, %r2546, 14; + shf.l.wrap.b32 %r2549, %r2546, %r2547, 14; + mov.b64 %rd3695, {%r2549, %r2548}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2550}, %rd3670; + } + { + .reg .b32 %dummy; + mov.b64 {%r2551,%dummy}, %rd3670; + } + shf.l.wrap.b32 %r2552, %r2551, %r2550, 28; + shf.l.wrap.b32 %r2553, %r2550, %r2551, 28; + mov.b64 %rd3696, {%r2553, %r2552}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2554}, %rd3676; + } + { + .reg .b32 %dummy; + mov.b64 {%r2555,%dummy}, %rd3676; + } + shf.l.wrap.b32 %r2556, %r2555, %r2554, 20; + shf.l.wrap.b32 %r2557, %r2554, %r2555, 20; + mov.b64 %rd3697, {%r2557, %r2556}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2558}, %rd3677; + } + { + .reg .b32 %dummy; + mov.b64 {%r2559,%dummy}, %rd3677; + } + shf.l.wrap.b32 %r2560, %r2559, %r2558, 3; + shf.l.wrap.b32 %r2561, %r2558, %r2559, 3; + mov.b64 %rd3698, {%r2561, %r2560}; + { + .reg .b32 %dummy; + mov.b64 {%r2562,%dummy}, %rd3683; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2563}, %rd3683; + } + shf.r.wrap.b32 %r2564, %r2563, %r2562, 19; + shf.r.wrap.b32 %r2565, %r2562, %r2563, 19; + mov.b64 %rd3699, {%r2565, %r2564}; + { + .reg .b32 %dummy; + mov.b64 {%r2566,%dummy}, %rd3689; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2567}, %rd3689; + } + shf.r.wrap.b32 %r2568, %r2567, %r2566, 3; + shf.r.wrap.b32 %r2569, %r2566, %r2567, 3; + mov.b64 %rd3700, {%r2569, %r2568}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2570}, %rd3668; + } + { + .reg .b32 %dummy; + mov.b64 {%r2571,%dummy}, %rd3668; + } + shf.l.wrap.b32 %r2572, %r2571, %r2570, 1; + shf.l.wrap.b32 %r2573, %r2570, %r2571, 1; + mov.b64 %rd3701, {%r2573, %r2572}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2574}, %rd3674; + } + { + .reg .b32 %dummy; + mov.b64 {%r2575,%dummy}, %rd3674; + } + shf.l.wrap.b32 %r2576, %r2575, %r2574, 6; + shf.l.wrap.b32 %r2577, %r2574, %r2575, 6; + mov.b64 %rd3702, {%r2577, %r2576}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2578}, %rd3680; + } + { + .reg .b32 %dummy; + mov.b64 {%r2579,%dummy}, %rd3680; + } + shf.l.wrap.b32 %r2580, %r2579, %r2578, 25; + shf.l.wrap.b32 %r2581, %r2578, %r2579, 25; + mov.b64 %rd3703, {%r2581, %r2580}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2582}, %rd3686; + } + { + .reg .b32 %dummy; + mov.b64 {%r2583,%dummy}, %rd3686; + } + shf.l.wrap.b32 %r2584, %r2583, %r2582, 8; + shf.l.wrap.b32 %r2585, %r2582, %r2583, 8; + mov.b64 %rd3704, {%r2585, %r2584}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2586}, %rd3687; + } + { + .reg .b32 %dummy; + mov.b64 {%r2587,%dummy}, %rd3687; + } + shf.l.wrap.b32 %r2588, %r2587, %r2586, 18; + shf.l.wrap.b32 %r2589, %r2586, %r2587, 18; + mov.b64 %rd3705, {%r2589, %r2588}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2590}, %rd3671; + } + { + .reg .b32 %dummy; + mov.b64 {%r2591,%dummy}, %rd3671; + } + shf.l.wrap.b32 %r2592, %r2591, %r2590, 27; + shf.l.wrap.b32 %r2593, %r2590, %r2591, 27; + mov.b64 %rd3706, {%r2593, %r2592}; + { + .reg .b32 %dummy; + mov.b64 {%r2594,%dummy}, %rd3672; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2595}, %rd3672; + } + shf.r.wrap.b32 %r2596, %r2595, %r2594, 28; + shf.r.wrap.b32 %r2597, %r2594, %r2595, 28; + mov.b64 %rd3707, {%r2597, %r2596}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2598}, %rd3678; + } + { + .reg .b32 %dummy; + mov.b64 {%r2599,%dummy}, %rd3678; + } + shf.l.wrap.b32 %r2600, %r2599, %r2598, 10; + shf.l.wrap.b32 %r2601, %r2598, %r2599, 10; + mov.b64 %rd3708, {%r2601, %r2600}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2602}, %rd3684; + } + { + .reg .b32 %dummy; + mov.b64 {%r2603,%dummy}, %rd3684; + } + shf.l.wrap.b32 %r2604, %r2603, %r2602, 15; + shf.l.wrap.b32 %r2605, %r2602, %r2603, 15; + mov.b64 %rd3709, {%r2605, %r2604}; + { + .reg .b32 %dummy; + mov.b64 {%r2606,%dummy}, %rd3690; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2607}, %rd3690; + } + shf.r.wrap.b32 %r2608, %r2607, %r2606, 8; + shf.r.wrap.b32 %r2609, %r2606, %r2607, 8; + mov.b64 %rd3710, {%r2609, %r2608}; + { + .reg .b32 %dummy; + mov.b64 {%r2610,%dummy}, %rd3669; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2611}, %rd3669; + } + shf.r.wrap.b32 %r2612, %r2611, %r2610, 2; + shf.r.wrap.b32 %r2613, %r2610, %r2611, 2; + mov.b64 %rd3711, {%r2613, %r2612}; + { + .reg .b32 %dummy; + mov.b64 {%r2614,%dummy}, %rd3675; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2615}, %rd3675; + } + shf.r.wrap.b32 %r2616, %r2615, %r2614, 9; + shf.r.wrap.b32 %r2617, %r2614, %r2615, 9; + mov.b64 %rd3712, {%r2617, %r2616}; + { + .reg .b32 %dummy; + mov.b64 {%r2618,%dummy}, %rd3681; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2619}, %rd3681; + } + shf.r.wrap.b32 %r2620, %r2619, %r2618, 25; + shf.r.wrap.b32 %r2621, %r2618, %r2619, 25; + mov.b64 %rd3713, {%r2621, %r2620}; + { + .reg .b32 %dummy; + mov.b64 {%r2622,%dummy}, %rd3682; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2623}, %rd3682; + } + shf.r.wrap.b32 %r2624, %r2623, %r2622, 23; + shf.r.wrap.b32 %r2625, %r2622, %r2623, 23; + mov.b64 %rd3714, {%r2625, %r2624}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2626}, %rd3688; + } + { + .reg .b32 %dummy; + mov.b64 {%r2627,%dummy}, %rd3688; + } + shf.l.wrap.b32 %r2628, %r2627, %r2626, 2; + shf.l.wrap.b32 %r2629, %r2626, %r2627, 2; + mov.b64 %rd3715, {%r2629, %r2628}; + not.b64 %rd3716, %rd3692; + and.b64 %rd3717, %rd3693, %rd3716; + xor.b64 %rd3718, %rd3717, %rd3667; + not.b64 %rd3719, %rd3693; + and.b64 %rd3720, %rd3694, %rd3719; + xor.b64 %rd3721, %rd3720, %rd3692; + not.b64 %rd3722, %rd3694; + and.b64 %rd3723, %rd3695, %rd3722; + xor.b64 %rd3724, %rd3723, %rd3693; + not.b64 %rd3725, %rd3695; + and.b64 %rd3726, %rd3667, %rd3725; + xor.b64 %rd3727, %rd3726, %rd3694; + not.b64 %rd3728, %rd3667; + and.b64 %rd3729, %rd3692, %rd3728; + xor.b64 %rd3730, %rd3729, %rd3695; + not.b64 %rd3731, %rd3697; + and.b64 %rd3732, %rd3698, %rd3731; + xor.b64 %rd3733, %rd3732, %rd3696; + not.b64 %rd3734, %rd3698; + and.b64 %rd3735, %rd3699, %rd3734; + xor.b64 %rd3736, %rd3735, %rd3697; + not.b64 %rd3737, %rd3699; + and.b64 %rd3738, %rd3700, %rd3737; + xor.b64 %rd3739, %rd3738, %rd3698; + not.b64 %rd3740, %rd3700; + and.b64 %rd3741, %rd3696, %rd3740; + xor.b64 %rd3742, %rd3741, %rd3699; + not.b64 %rd3743, %rd3696; + and.b64 %rd3744, %rd3697, %rd3743; + xor.b64 %rd3745, %rd3744, %rd3700; + not.b64 %rd3746, %rd3702; + and.b64 %rd3747, %rd3703, %rd3746; + xor.b64 %rd3748, %rd3747, %rd3701; + not.b64 %rd3749, %rd3703; + and.b64 %rd3750, %rd3704, %rd3749; + xor.b64 %rd3751, %rd3750, %rd3702; + not.b64 %rd3752, %rd3704; + and.b64 %rd3753, %rd3705, %rd3752; + xor.b64 %rd3754, %rd3753, %rd3703; + not.b64 %rd3755, %rd3705; + and.b64 %rd3756, %rd3701, %rd3755; + xor.b64 %rd3757, %rd3756, %rd3704; + not.b64 %rd3758, %rd3701; + and.b64 %rd3759, %rd3702, %rd3758; + xor.b64 %rd3760, %rd3759, %rd3705; + not.b64 %rd3761, %rd3707; + and.b64 %rd3762, %rd3708, %rd3761; + xor.b64 %rd3763, %rd3762, %rd3706; + not.b64 %rd3764, %rd3708; + and.b64 %rd3765, %rd3709, %rd3764; + xor.b64 %rd3766, %rd3765, %rd3707; + not.b64 %rd3767, %rd3709; + and.b64 %rd3768, %rd3710, %rd3767; + xor.b64 %rd3769, %rd3768, %rd3708; + not.b64 %rd3770, %rd3710; + and.b64 %rd3771, %rd3706, %rd3770; + xor.b64 %rd3772, %rd3771, %rd3709; + not.b64 %rd3773, %rd3706; + and.b64 %rd3774, %rd3707, %rd3773; + xor.b64 %rd3775, %rd3774, %rd3710; + not.b64 %rd3776, %rd3712; + and.b64 %rd3777, %rd3713, %rd3776; + xor.b64 %rd3778, %rd3777, %rd3711; + not.b64 %rd3779, %rd3713; + and.b64 %rd3780, %rd3714, %rd3779; + xor.b64 %rd3781, %rd3780, %rd3712; + not.b64 %rd3782, %rd3714; + and.b64 %rd3783, %rd3715, %rd3782; + xor.b64 %rd3784, %rd3783, %rd3713; + not.b64 %rd3785, %rd3715; + and.b64 %rd3786, %rd3711, %rd3785; + xor.b64 %rd3787, %rd3786, %rd3714; + not.b64 %rd3788, %rd3711; + and.b64 %rd3789, %rd3712, %rd3788; + xor.b64 %rd3790, %rd3789, %rd3715; + xor.b64 %rd3791, %rd3718, %rd37; + xor.b64 %rd3792, %rd3733, %rd3791; + xor.b64 %rd3793, %rd3792, %rd3748; + xor.b64 %rd3794, %rd3793, %rd3763; + xor.b64 %rd3795, %rd3794, %rd3778; + xor.b64 %rd3796, %rd3736, %rd3721; + xor.b64 %rd3797, %rd3796, %rd3751; + xor.b64 %rd3798, %rd3797, %rd3766; + xor.b64 %rd3799, %rd3798, %rd3781; + xor.b64 %rd3800, %rd3739, %rd3724; + xor.b64 %rd3801, %rd3800, %rd3754; + xor.b64 %rd3802, %rd3801, %rd3769; + xor.b64 %rd3803, %rd3802, %rd3784; + xor.b64 %rd3804, %rd3742, %rd3727; + xor.b64 %rd3805, %rd3804, %rd3757; + xor.b64 %rd3806, %rd3805, %rd3772; + xor.b64 %rd3807, %rd3806, %rd3787; + xor.b64 %rd3808, %rd3745, %rd3730; + xor.b64 %rd3809, %rd3808, %rd3760; + xor.b64 %rd3810, %rd3809, %rd3775; + xor.b64 %rd3811, %rd3810, %rd3790; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2630}, %rd3799; + } + { + .reg .b32 %dummy; + mov.b64 {%r2631,%dummy}, %rd3799; + } + shf.l.wrap.b32 %r2632, %r2631, %r2630, 1; + shf.l.wrap.b32 %r2633, %r2630, %r2631, 1; + mov.b64 %rd3812, {%r2633, %r2632}; + xor.b64 %rd3813, %rd3811, %rd3812; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2634}, %rd3803; + } + { + .reg .b32 %dummy; + mov.b64 {%r2635,%dummy}, %rd3803; + } + shf.l.wrap.b32 %r2636, %r2635, %r2634, 1; + shf.l.wrap.b32 %r2637, %r2634, %r2635, 1; + mov.b64 %rd3814, {%r2637, %r2636}; + xor.b64 %rd3815, %rd3814, %rd3795; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2638}, %rd3807; + } + { + .reg .b32 %dummy; + mov.b64 {%r2639,%dummy}, %rd3807; + } + shf.l.wrap.b32 %r2640, %r2639, %r2638, 1; + shf.l.wrap.b32 %r2641, %r2638, %r2639, 1; + mov.b64 %rd3816, {%r2641, %r2640}; + xor.b64 %rd3817, %rd3816, %rd3799; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2642}, %rd3811; + } + { + .reg .b32 %dummy; + mov.b64 {%r2643,%dummy}, %rd3811; + } + shf.l.wrap.b32 %r2644, %r2643, %r2642, 1; + shf.l.wrap.b32 %r2645, %r2642, %r2643, 1; + mov.b64 %rd3818, {%r2645, %r2644}; + xor.b64 %rd3819, %rd3818, %rd3803; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2646}, %rd3795; + } + { + .reg .b32 %dummy; + mov.b64 {%r2647,%dummy}, %rd3795; + } + shf.l.wrap.b32 %r2648, %r2647, %r2646, 1; + shf.l.wrap.b32 %r2649, %r2646, %r2647, 1; + mov.b64 %rd3820, {%r2649, %r2648}; + xor.b64 %rd3821, %rd3807, %rd3820; + xor.b64 %rd3822, %rd3813, %rd3791; + xor.b64 %rd3823, %rd3815, %rd3721; + xor.b64 %rd3824, %rd3817, %rd3724; + xor.b64 %rd3825, %rd3819, %rd3727; + xor.b64 %rd3826, %rd3821, %rd3730; + xor.b64 %rd3827, %rd3813, %rd3733; + xor.b64 %rd3828, %rd3815, %rd3736; + xor.b64 %rd3829, %rd3817, %rd3739; + xor.b64 %rd3830, %rd3819, %rd3742; + xor.b64 %rd3831, %rd3821, %rd3745; + xor.b64 %rd3832, %rd3813, %rd3748; + xor.b64 %rd3833, %rd3815, %rd3751; + xor.b64 %rd3834, %rd3817, %rd3754; + xor.b64 %rd3835, %rd3819, %rd3757; + xor.b64 %rd3836, %rd3821, %rd3760; + xor.b64 %rd3837, %rd3813, %rd3763; + xor.b64 %rd3838, %rd3815, %rd3766; + xor.b64 %rd3839, %rd3817, %rd3769; + xor.b64 %rd3840, %rd3819, %rd3772; + xor.b64 %rd3841, %rd3821, %rd3775; + xor.b64 %rd3842, %rd3813, %rd3778; + xor.b64 %rd3843, %rd3815, %rd3781; + xor.b64 %rd3844, %rd3817, %rd3784; + xor.b64 %rd3845, %rd3819, %rd3787; + xor.b64 %rd3846, %rd3821, %rd3790; + { + .reg .b32 %dummy; + mov.b64 {%r2650,%dummy}, %rd3828; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2651}, %rd3828; + } + shf.r.wrap.b32 %r2652, %r2651, %r2650, 20; + shf.r.wrap.b32 %r2653, %r2650, %r2651, 20; + mov.b64 %rd3847, {%r2653, %r2652}; + { + .reg .b32 %dummy; + mov.b64 {%r2654,%dummy}, %rd3834; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2655}, %rd3834; + } + shf.r.wrap.b32 %r2656, %r2655, %r2654, 21; + shf.r.wrap.b32 %r2657, %r2654, %r2655, 21; + mov.b64 %rd3848, {%r2657, %r2656}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2658}, %rd3840; + } + { + .reg .b32 %dummy; + mov.b64 {%r2659,%dummy}, %rd3840; + } + shf.l.wrap.b32 %r2660, %r2659, %r2658, 21; + shf.l.wrap.b32 %r2661, %r2658, %r2659, 21; + mov.b64 %rd3849, {%r2661, %r2660}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2662}, %rd3846; + } + { + .reg .b32 %dummy; + mov.b64 {%r2663,%dummy}, %rd3846; + } + shf.l.wrap.b32 %r2664, %r2663, %r2662, 14; + shf.l.wrap.b32 %r2665, %r2662, %r2663, 14; + mov.b64 %rd3850, {%r2665, %r2664}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2666}, %rd3825; + } + { + .reg .b32 %dummy; + mov.b64 {%r2667,%dummy}, %rd3825; + } + shf.l.wrap.b32 %r2668, %r2667, %r2666, 28; + shf.l.wrap.b32 %r2669, %r2666, %r2667, 28; + mov.b64 %rd3851, {%r2669, %r2668}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2670}, %rd3831; + } + { + .reg .b32 %dummy; + mov.b64 {%r2671,%dummy}, %rd3831; + } + shf.l.wrap.b32 %r2672, %r2671, %r2670, 20; + shf.l.wrap.b32 %r2673, %r2670, %r2671, 20; + mov.b64 %rd3852, {%r2673, %r2672}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2674}, %rd3832; + } + { + .reg .b32 %dummy; + mov.b64 {%r2675,%dummy}, %rd3832; + } + shf.l.wrap.b32 %r2676, %r2675, %r2674, 3; + shf.l.wrap.b32 %r2677, %r2674, %r2675, 3; + mov.b64 %rd3853, {%r2677, %r2676}; + { + .reg .b32 %dummy; + mov.b64 {%r2678,%dummy}, %rd3838; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2679}, %rd3838; + } + shf.r.wrap.b32 %r2680, %r2679, %r2678, 19; + shf.r.wrap.b32 %r2681, %r2678, %r2679, 19; + mov.b64 %rd3854, {%r2681, %r2680}; + { + .reg .b32 %dummy; + mov.b64 {%r2682,%dummy}, %rd3844; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2683}, %rd3844; + } + shf.r.wrap.b32 %r2684, %r2683, %r2682, 3; + shf.r.wrap.b32 %r2685, %r2682, %r2683, 3; + mov.b64 %rd3855, {%r2685, %r2684}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2686}, %rd3823; + } + { + .reg .b32 %dummy; + mov.b64 {%r2687,%dummy}, %rd3823; + } + shf.l.wrap.b32 %r2688, %r2687, %r2686, 1; + shf.l.wrap.b32 %r2689, %r2686, %r2687, 1; + mov.b64 %rd3856, {%r2689, %r2688}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2690}, %rd3829; + } + { + .reg .b32 %dummy; + mov.b64 {%r2691,%dummy}, %rd3829; + } + shf.l.wrap.b32 %r2692, %r2691, %r2690, 6; + shf.l.wrap.b32 %r2693, %r2690, %r2691, 6; + mov.b64 %rd3857, {%r2693, %r2692}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2694}, %rd3835; + } + { + .reg .b32 %dummy; + mov.b64 {%r2695,%dummy}, %rd3835; + } + shf.l.wrap.b32 %r2696, %r2695, %r2694, 25; + shf.l.wrap.b32 %r2697, %r2694, %r2695, 25; + mov.b64 %rd3858, {%r2697, %r2696}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2698}, %rd3841; + } + { + .reg .b32 %dummy; + mov.b64 {%r2699,%dummy}, %rd3841; + } + shf.l.wrap.b32 %r2700, %r2699, %r2698, 8; + shf.l.wrap.b32 %r2701, %r2698, %r2699, 8; + mov.b64 %rd3859, {%r2701, %r2700}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2702}, %rd3842; + } + { + .reg .b32 %dummy; + mov.b64 {%r2703,%dummy}, %rd3842; + } + shf.l.wrap.b32 %r2704, %r2703, %r2702, 18; + shf.l.wrap.b32 %r2705, %r2702, %r2703, 18; + mov.b64 %rd3860, {%r2705, %r2704}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2706}, %rd3826; + } + { + .reg .b32 %dummy; + mov.b64 {%r2707,%dummy}, %rd3826; + } + shf.l.wrap.b32 %r2708, %r2707, %r2706, 27; + shf.l.wrap.b32 %r2709, %r2706, %r2707, 27; + mov.b64 %rd3861, {%r2709, %r2708}; + { + .reg .b32 %dummy; + mov.b64 {%r2710,%dummy}, %rd3827; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2711}, %rd3827; + } + shf.r.wrap.b32 %r2712, %r2711, %r2710, 28; + shf.r.wrap.b32 %r2713, %r2710, %r2711, 28; + mov.b64 %rd3862, {%r2713, %r2712}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2714}, %rd3833; + } + { + .reg .b32 %dummy; + mov.b64 {%r2715,%dummy}, %rd3833; + } + shf.l.wrap.b32 %r2716, %r2715, %r2714, 10; + shf.l.wrap.b32 %r2717, %r2714, %r2715, 10; + mov.b64 %rd3863, {%r2717, %r2716}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2718}, %rd3839; + } + { + .reg .b32 %dummy; + mov.b64 {%r2719,%dummy}, %rd3839; + } + shf.l.wrap.b32 %r2720, %r2719, %r2718, 15; + shf.l.wrap.b32 %r2721, %r2718, %r2719, 15; + mov.b64 %rd3864, {%r2721, %r2720}; + { + .reg .b32 %dummy; + mov.b64 {%r2722,%dummy}, %rd3845; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2723}, %rd3845; + } + shf.r.wrap.b32 %r2724, %r2723, %r2722, 8; + shf.r.wrap.b32 %r2725, %r2722, %r2723, 8; + mov.b64 %rd3865, {%r2725, %r2724}; + { + .reg .b32 %dummy; + mov.b64 {%r2726,%dummy}, %rd3824; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2727}, %rd3824; + } + shf.r.wrap.b32 %r2728, %r2727, %r2726, 2; + shf.r.wrap.b32 %r2729, %r2726, %r2727, 2; + mov.b64 %rd3866, {%r2729, %r2728}; + { + .reg .b32 %dummy; + mov.b64 {%r2730,%dummy}, %rd3830; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2731}, %rd3830; + } + shf.r.wrap.b32 %r2732, %r2731, %r2730, 9; + shf.r.wrap.b32 %r2733, %r2730, %r2731, 9; + mov.b64 %rd3867, {%r2733, %r2732}; + { + .reg .b32 %dummy; + mov.b64 {%r2734,%dummy}, %rd3836; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2735}, %rd3836; + } + shf.r.wrap.b32 %r2736, %r2735, %r2734, 25; + shf.r.wrap.b32 %r2737, %r2734, %r2735, 25; + mov.b64 %rd3868, {%r2737, %r2736}; + { + .reg .b32 %dummy; + mov.b64 {%r2738,%dummy}, %rd3837; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2739}, %rd3837; + } + shf.r.wrap.b32 %r2740, %r2739, %r2738, 23; + shf.r.wrap.b32 %r2741, %r2738, %r2739, 23; + mov.b64 %rd3869, {%r2741, %r2740}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2742}, %rd3843; + } + { + .reg .b32 %dummy; + mov.b64 {%r2743,%dummy}, %rd3843; + } + shf.l.wrap.b32 %r2744, %r2743, %r2742, 2; + shf.l.wrap.b32 %r2745, %r2742, %r2743, 2; + mov.b64 %rd3870, {%r2745, %r2744}; + not.b64 %rd3871, %rd3847; + and.b64 %rd3872, %rd3848, %rd3871; + xor.b64 %rd3873, %rd3872, %rd3822; + not.b64 %rd3874, %rd3848; + and.b64 %rd3875, %rd3849, %rd3874; + xor.b64 %rd3876, %rd3875, %rd3847; + not.b64 %rd3877, %rd3849; + and.b64 %rd3878, %rd3850, %rd3877; + xor.b64 %rd3879, %rd3878, %rd3848; + not.b64 %rd3880, %rd3850; + and.b64 %rd3881, %rd3822, %rd3880; + xor.b64 %rd3882, %rd3881, %rd3849; + not.b64 %rd3883, %rd3822; + and.b64 %rd3884, %rd3847, %rd3883; + xor.b64 %rd3885, %rd3884, %rd3850; + not.b64 %rd3886, %rd3852; + and.b64 %rd3887, %rd3853, %rd3886; + xor.b64 %rd3888, %rd3887, %rd3851; + not.b64 %rd3889, %rd3853; + and.b64 %rd3890, %rd3854, %rd3889; + xor.b64 %rd3891, %rd3890, %rd3852; + not.b64 %rd3892, %rd3854; + and.b64 %rd3893, %rd3855, %rd3892; + xor.b64 %rd3894, %rd3893, %rd3853; + not.b64 %rd3895, %rd3855; + and.b64 %rd3896, %rd3851, %rd3895; + xor.b64 %rd3897, %rd3896, %rd3854; + not.b64 %rd3898, %rd3851; + and.b64 %rd3899, %rd3852, %rd3898; + xor.b64 %rd3900, %rd3899, %rd3855; + not.b64 %rd3901, %rd3857; + and.b64 %rd3902, %rd3858, %rd3901; + xor.b64 %rd3903, %rd3902, %rd3856; + not.b64 %rd3904, %rd3858; + and.b64 %rd3905, %rd3859, %rd3904; + xor.b64 %rd3906, %rd3905, %rd3857; + not.b64 %rd3907, %rd3859; + and.b64 %rd3908, %rd3860, %rd3907; + xor.b64 %rd3909, %rd3908, %rd3858; + not.b64 %rd3910, %rd3860; + and.b64 %rd3911, %rd3856, %rd3910; + xor.b64 %rd3912, %rd3911, %rd3859; + not.b64 %rd3913, %rd3856; + and.b64 %rd3914, %rd3857, %rd3913; + xor.b64 %rd3915, %rd3914, %rd3860; + not.b64 %rd3916, %rd3862; + and.b64 %rd3917, %rd3863, %rd3916; + xor.b64 %rd3918, %rd3917, %rd3861; + not.b64 %rd3919, %rd3863; + and.b64 %rd3920, %rd3864, %rd3919; + xor.b64 %rd3921, %rd3920, %rd3862; + not.b64 %rd3922, %rd3864; + and.b64 %rd3923, %rd3865, %rd3922; + xor.b64 %rd3924, %rd3923, %rd3863; + not.b64 %rd3925, %rd3865; + and.b64 %rd3926, %rd3861, %rd3925; + xor.b64 %rd3927, %rd3926, %rd3864; + not.b64 %rd3928, %rd3861; + and.b64 %rd3929, %rd3862, %rd3928; + xor.b64 %rd3930, %rd3929, %rd3865; + not.b64 %rd3931, %rd3867; + and.b64 %rd3932, %rd3868, %rd3931; + xor.b64 %rd3933, %rd3932, %rd3866; + not.b64 %rd3934, %rd3868; + and.b64 %rd3935, %rd3869, %rd3934; + xor.b64 %rd3936, %rd3935, %rd3867; + not.b64 %rd3937, %rd3869; + and.b64 %rd3938, %rd3870, %rd3937; + xor.b64 %rd3939, %rd3938, %rd3868; + not.b64 %rd3940, %rd3870; + and.b64 %rd3941, %rd3866, %rd3940; + xor.b64 %rd3942, %rd3941, %rd3869; + not.b64 %rd3943, %rd3866; + and.b64 %rd3944, %rd3867, %rd3943; + xor.b64 %rd3945, %rd3944, %rd3870; + xor.b64 %rd3946, %rd3873, %rd38; + xor.b64 %rd3947, %rd3888, %rd3946; + xor.b64 %rd3948, %rd3947, %rd3903; + xor.b64 %rd3949, %rd3948, %rd3918; + xor.b64 %rd3950, %rd3949, %rd3933; + xor.b64 %rd3951, %rd3891, %rd3876; + xor.b64 %rd3952, %rd3951, %rd3906; + xor.b64 %rd3953, %rd3952, %rd3921; + xor.b64 %rd3954, %rd3953, %rd3936; + xor.b64 %rd3955, %rd3894, %rd3879; + xor.b64 %rd3956, %rd3955, %rd3909; + xor.b64 %rd3957, %rd3956, %rd3924; + xor.b64 %rd3958, %rd3957, %rd3939; + xor.b64 %rd3959, %rd3897, %rd3882; + xor.b64 %rd3960, %rd3959, %rd3912; + xor.b64 %rd3961, %rd3960, %rd3927; + xor.b64 %rd3962, %rd3961, %rd3942; + xor.b64 %rd3963, %rd3900, %rd3885; + xor.b64 %rd3964, %rd3963, %rd3915; + xor.b64 %rd3965, %rd3964, %rd3930; + xor.b64 %rd3966, %rd3965, %rd3945; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2746}, %rd3954; + } + { + .reg .b32 %dummy; + mov.b64 {%r2747,%dummy}, %rd3954; + } + shf.l.wrap.b32 %r2748, %r2747, %r2746, 1; + shf.l.wrap.b32 %r2749, %r2746, %r2747, 1; + mov.b64 %rd3967, {%r2749, %r2748}; + xor.b64 %rd3968, %rd3966, %rd3967; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2750}, %rd3958; + } + { + .reg .b32 %dummy; + mov.b64 {%r2751,%dummy}, %rd3958; + } + shf.l.wrap.b32 %r2752, %r2751, %r2750, 1; + shf.l.wrap.b32 %r2753, %r2750, %r2751, 1; + mov.b64 %rd3969, {%r2753, %r2752}; + xor.b64 %rd3970, %rd3969, %rd3950; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2754}, %rd3962; + } + { + .reg .b32 %dummy; + mov.b64 {%r2755,%dummy}, %rd3962; + } + shf.l.wrap.b32 %r2756, %r2755, %r2754, 1; + shf.l.wrap.b32 %r2757, %r2754, %r2755, 1; + mov.b64 %rd3971, {%r2757, %r2756}; + xor.b64 %rd3972, %rd3971, %rd3954; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2758}, %rd3966; + } + { + .reg .b32 %dummy; + mov.b64 {%r2759,%dummy}, %rd3966; + } + shf.l.wrap.b32 %r2760, %r2759, %r2758, 1; + shf.l.wrap.b32 %r2761, %r2758, %r2759, 1; + mov.b64 %rd3973, {%r2761, %r2760}; + xor.b64 %rd3974, %rd3973, %rd3958; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2762}, %rd3950; + } + { + .reg .b32 %dummy; + mov.b64 {%r2763,%dummy}, %rd3950; + } + shf.l.wrap.b32 %r2764, %r2763, %r2762, 1; + shf.l.wrap.b32 %r2765, %r2762, %r2763, 1; + mov.b64 %rd3975, {%r2765, %r2764}; + xor.b64 %rd3976, %rd3962, %rd3975; + xor.b64 %rd3977, %rd3968, %rd3946; + xor.b64 %rd3978, %rd3970, %rd3876; + xor.b64 %rd3979, %rd3972, %rd3879; + xor.b64 %rd3980, %rd3974, %rd3882; + xor.b64 %rd3981, %rd3976, %rd3885; + xor.b64 %rd3982, %rd3968, %rd3888; + xor.b64 %rd3983, %rd3970, %rd3891; + xor.b64 %rd3984, %rd3972, %rd3894; + xor.b64 %rd3985, %rd3974, %rd3897; + xor.b64 %rd3986, %rd3976, %rd3900; + xor.b64 %rd3987, %rd3968, %rd3903; + xor.b64 %rd3988, %rd3970, %rd3906; + xor.b64 %rd3989, %rd3972, %rd3909; + xor.b64 %rd3990, %rd3974, %rd3912; + xor.b64 %rd3991, %rd3976, %rd3915; + xor.b64 %rd3992, %rd3968, %rd3918; + xor.b64 %rd3993, %rd3970, %rd3921; + xor.b64 %rd3994, %rd3972, %rd3924; + xor.b64 %rd3995, %rd3974, %rd3927; + xor.b64 %rd3996, %rd3976, %rd3930; + xor.b64 %rd3997, %rd3968, %rd3933; + xor.b64 %rd3998, %rd3970, %rd3936; + xor.b64 %rd3999, %rd3972, %rd3939; + xor.b64 %rd4000, %rd3974, %rd3942; + xor.b64 %rd4001, %rd3976, %rd3945; + { + .reg .b32 %dummy; + mov.b64 {%r2766,%dummy}, %rd3983; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2767}, %rd3983; + } + shf.r.wrap.b32 %r2768, %r2767, %r2766, 20; + shf.r.wrap.b32 %r2769, %r2766, %r2767, 20; + mov.b64 %rd4002, {%r2769, %r2768}; + { + .reg .b32 %dummy; + mov.b64 {%r2770,%dummy}, %rd3989; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2771}, %rd3989; + } + shf.r.wrap.b32 %r2772, %r2771, %r2770, 21; + shf.r.wrap.b32 %r2773, %r2770, %r2771, 21; + mov.b64 %rd4003, {%r2773, %r2772}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2774}, %rd3995; + } + { + .reg .b32 %dummy; + mov.b64 {%r2775,%dummy}, %rd3995; + } + shf.l.wrap.b32 %r2776, %r2775, %r2774, 21; + shf.l.wrap.b32 %r2777, %r2774, %r2775, 21; + mov.b64 %rd4004, {%r2777, %r2776}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2778}, %rd4001; + } + { + .reg .b32 %dummy; + mov.b64 {%r2779,%dummy}, %rd4001; + } + shf.l.wrap.b32 %r2780, %r2779, %r2778, 14; + shf.l.wrap.b32 %r2781, %r2778, %r2779, 14; + mov.b64 %rd4005, {%r2781, %r2780}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2782}, %rd3980; + } + { + .reg .b32 %dummy; + mov.b64 {%r2783,%dummy}, %rd3980; + } + shf.l.wrap.b32 %r2784, %r2783, %r2782, 28; + shf.l.wrap.b32 %r2785, %r2782, %r2783, 28; + mov.b64 %rd4006, {%r2785, %r2784}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2786}, %rd3986; + } + { + .reg .b32 %dummy; + mov.b64 {%r2787,%dummy}, %rd3986; + } + shf.l.wrap.b32 %r2788, %r2787, %r2786, 20; + shf.l.wrap.b32 %r2789, %r2786, %r2787, 20; + mov.b64 %rd4007, {%r2789, %r2788}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2790}, %rd3987; + } + { + .reg .b32 %dummy; + mov.b64 {%r2791,%dummy}, %rd3987; + } + shf.l.wrap.b32 %r2792, %r2791, %r2790, 3; + shf.l.wrap.b32 %r2793, %r2790, %r2791, 3; + mov.b64 %rd4008, {%r2793, %r2792}; + { + .reg .b32 %dummy; + mov.b64 {%r2794,%dummy}, %rd3993; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2795}, %rd3993; + } + shf.r.wrap.b32 %r2796, %r2795, %r2794, 19; + shf.r.wrap.b32 %r2797, %r2794, %r2795, 19; + mov.b64 %rd4009, {%r2797, %r2796}; + { + .reg .b32 %dummy; + mov.b64 {%r2798,%dummy}, %rd3999; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2799}, %rd3999; + } + shf.r.wrap.b32 %r2800, %r2799, %r2798, 3; + shf.r.wrap.b32 %r2801, %r2798, %r2799, 3; + mov.b64 %rd4010, {%r2801, %r2800}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2802}, %rd3978; + } + { + .reg .b32 %dummy; + mov.b64 {%r2803,%dummy}, %rd3978; + } + shf.l.wrap.b32 %r2804, %r2803, %r2802, 1; + shf.l.wrap.b32 %r2805, %r2802, %r2803, 1; + mov.b64 %rd4011, {%r2805, %r2804}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2806}, %rd3984; + } + { + .reg .b32 %dummy; + mov.b64 {%r2807,%dummy}, %rd3984; + } + shf.l.wrap.b32 %r2808, %r2807, %r2806, 6; + shf.l.wrap.b32 %r2809, %r2806, %r2807, 6; + mov.b64 %rd4012, {%r2809, %r2808}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2810}, %rd3990; + } + { + .reg .b32 %dummy; + mov.b64 {%r2811,%dummy}, %rd3990; + } + shf.l.wrap.b32 %r2812, %r2811, %r2810, 25; + shf.l.wrap.b32 %r2813, %r2810, %r2811, 25; + mov.b64 %rd4013, {%r2813, %r2812}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2814}, %rd3996; + } + { + .reg .b32 %dummy; + mov.b64 {%r2815,%dummy}, %rd3996; + } + shf.l.wrap.b32 %r2816, %r2815, %r2814, 8; + shf.l.wrap.b32 %r2817, %r2814, %r2815, 8; + mov.b64 %rd4014, {%r2817, %r2816}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2818}, %rd3997; + } + { + .reg .b32 %dummy; + mov.b64 {%r2819,%dummy}, %rd3997; + } + shf.l.wrap.b32 %r2820, %r2819, %r2818, 18; + shf.l.wrap.b32 %r2821, %r2818, %r2819, 18; + mov.b64 %rd4015, {%r2821, %r2820}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2822}, %rd3981; + } + { + .reg .b32 %dummy; + mov.b64 {%r2823,%dummy}, %rd3981; + } + shf.l.wrap.b32 %r2824, %r2823, %r2822, 27; + shf.l.wrap.b32 %r2825, %r2822, %r2823, 27; + mov.b64 %rd4016, {%r2825, %r2824}; + { + .reg .b32 %dummy; + mov.b64 {%r2826,%dummy}, %rd3982; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2827}, %rd3982; + } + shf.r.wrap.b32 %r2828, %r2827, %r2826, 28; + shf.r.wrap.b32 %r2829, %r2826, %r2827, 28; + mov.b64 %rd4017, {%r2829, %r2828}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2830}, %rd3988; + } + { + .reg .b32 %dummy; + mov.b64 {%r2831,%dummy}, %rd3988; + } + shf.l.wrap.b32 %r2832, %r2831, %r2830, 10; + shf.l.wrap.b32 %r2833, %r2830, %r2831, 10; + mov.b64 %rd4018, {%r2833, %r2832}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2834}, %rd3994; + } + { + .reg .b32 %dummy; + mov.b64 {%r2835,%dummy}, %rd3994; + } + shf.l.wrap.b32 %r2836, %r2835, %r2834, 15; + shf.l.wrap.b32 %r2837, %r2834, %r2835, 15; + mov.b64 %rd4019, {%r2837, %r2836}; + { + .reg .b32 %dummy; + mov.b64 {%r2838,%dummy}, %rd4000; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2839}, %rd4000; + } + shf.r.wrap.b32 %r2840, %r2839, %r2838, 8; + shf.r.wrap.b32 %r2841, %r2838, %r2839, 8; + mov.b64 %rd4020, {%r2841, %r2840}; + { + .reg .b32 %dummy; + mov.b64 {%r2842,%dummy}, %rd3979; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2843}, %rd3979; + } + shf.r.wrap.b32 %r2844, %r2843, %r2842, 2; + shf.r.wrap.b32 %r2845, %r2842, %r2843, 2; + mov.b64 %rd4021, {%r2845, %r2844}; + { + .reg .b32 %dummy; + mov.b64 {%r2846,%dummy}, %rd3985; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2847}, %rd3985; + } + shf.r.wrap.b32 %r2848, %r2847, %r2846, 9; + shf.r.wrap.b32 %r2849, %r2846, %r2847, 9; + mov.b64 %rd4022, {%r2849, %r2848}; + { + .reg .b32 %dummy; + mov.b64 {%r2850,%dummy}, %rd3991; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2851}, %rd3991; + } + shf.r.wrap.b32 %r2852, %r2851, %r2850, 25; + shf.r.wrap.b32 %r2853, %r2850, %r2851, 25; + mov.b64 %rd4023, {%r2853, %r2852}; + { + .reg .b32 %dummy; + mov.b64 {%r2854,%dummy}, %rd3992; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2855}, %rd3992; + } + shf.r.wrap.b32 %r2856, %r2855, %r2854, 23; + shf.r.wrap.b32 %r2857, %r2854, %r2855, 23; + mov.b64 %rd4024, {%r2857, %r2856}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r2858}, %rd3998; + } + { + .reg .b32 %dummy; + mov.b64 {%r2859,%dummy}, %rd3998; + } + shf.l.wrap.b32 %r2860, %r2859, %r2858, 2; + shf.l.wrap.b32 %r2861, %r2858, %r2859, 2; + mov.b64 %rd4025, {%r2861, %r2860}; + not.b64 %rd4026, %rd4002; + and.b64 %rd4027, %rd4003, %rd4026; + xor.b64 %rd4028, %rd4027, %rd3977; + not.b64 %rd4029, %rd4003; + and.b64 %rd4030, %rd4004, %rd4029; + xor.b64 %rd4031, %rd4030, %rd4002; + st.local.u64 [%rd2+8], %rd4031; + not.b64 %rd4032, %rd4004; + and.b64 %rd4033, %rd4005, %rd4032; + xor.b64 %rd4034, %rd4033, %rd4003; + st.local.u64 [%rd2+16], %rd4034; + not.b64 %rd4035, %rd4005; + and.b64 %rd4036, %rd3977, %rd4035; + xor.b64 %rd4037, %rd4036, %rd4004; + st.local.u64 [%rd2+24], %rd4037; + not.b64 %rd4038, %rd3977; + and.b64 %rd4039, %rd4002, %rd4038; + xor.b64 %rd4040, %rd4039, %rd4005; + st.local.u64 [%rd2+32], %rd4040; + not.b64 %rd4041, %rd4007; + and.b64 %rd4042, %rd4008, %rd4041; + xor.b64 %rd4043, %rd4042, %rd4006; + st.local.u64 [%rd2+40], %rd4043; + not.b64 %rd4044, %rd4008; + and.b64 %rd4045, %rd4009, %rd4044; + xor.b64 %rd4046, %rd4045, %rd4007; + st.local.u64 [%rd2+48], %rd4046; + not.b64 %rd4047, %rd4009; + and.b64 %rd4048, %rd4010, %rd4047; + xor.b64 %rd4049, %rd4048, %rd4008; + st.local.u64 [%rd2+56], %rd4049; + not.b64 %rd4050, %rd4010; + and.b64 %rd4051, %rd4006, %rd4050; + xor.b64 %rd4052, %rd4051, %rd4009; + st.local.u64 [%rd2+64], %rd4052; + not.b64 %rd4053, %rd4006; + and.b64 %rd4054, %rd4007, %rd4053; + xor.b64 %rd4055, %rd4054, %rd4010; + st.local.u64 [%rd2+72], %rd4055; + not.b64 %rd4056, %rd4012; + and.b64 %rd4057, %rd4013, %rd4056; + xor.b64 %rd4058, %rd4057, %rd4011; + st.local.u64 [%rd2+80], %rd4058; + not.b64 %rd4059, %rd4013; + and.b64 %rd4060, %rd4014, %rd4059; + xor.b64 %rd4061, %rd4060, %rd4012; + st.local.u64 [%rd2+88], %rd4061; + not.b64 %rd4062, %rd4014; + and.b64 %rd4063, %rd4015, %rd4062; + xor.b64 %rd4064, %rd4063, %rd4013; + st.local.u64 [%rd2+96], %rd4064; + not.b64 %rd4065, %rd4015; + and.b64 %rd4066, %rd4011, %rd4065; + xor.b64 %rd4067, %rd4066, %rd4014; + st.local.u64 [%rd2+104], %rd4067; + not.b64 %rd4068, %rd4011; + and.b64 %rd4069, %rd4012, %rd4068; + xor.b64 %rd4070, %rd4069, %rd4015; + st.local.u64 [%rd2+112], %rd4070; + not.b64 %rd4071, %rd4017; + and.b64 %rd4072, %rd4018, %rd4071; + xor.b64 %rd4073, %rd4072, %rd4016; + st.local.u64 [%rd2+120], %rd4073; + not.b64 %rd4074, %rd4018; + and.b64 %rd4075, %rd4019, %rd4074; + xor.b64 %rd4076, %rd4075, %rd4017; + st.local.u64 [%rd2+128], %rd4076; + not.b64 %rd4077, %rd4019; + and.b64 %rd4078, %rd4020, %rd4077; + xor.b64 %rd4079, %rd4078, %rd4018; + st.local.u64 [%rd2+136], %rd4079; + not.b64 %rd4080, %rd4020; + and.b64 %rd4081, %rd4016, %rd4080; + xor.b64 %rd4082, %rd4081, %rd4019; + st.local.u64 [%rd2+144], %rd4082; + not.b64 %rd4083, %rd4016; + and.b64 %rd4084, %rd4017, %rd4083; + xor.b64 %rd4085, %rd4084, %rd4020; + st.local.u64 [%rd2+152], %rd4085; + not.b64 %rd4086, %rd4022; + and.b64 %rd4087, %rd4023, %rd4086; + xor.b64 %rd4088, %rd4087, %rd4021; + st.local.u64 [%rd2+160], %rd4088; + not.b64 %rd4089, %rd4023; + and.b64 %rd4090, %rd4024, %rd4089; + xor.b64 %rd4091, %rd4090, %rd4022; + st.local.u64 [%rd2+168], %rd4091; + not.b64 %rd4092, %rd4024; + and.b64 %rd4093, %rd4025, %rd4092; + xor.b64 %rd4094, %rd4093, %rd4023; + st.local.u64 [%rd2+176], %rd4094; + not.b64 %rd4095, %rd4025; + and.b64 %rd4096, %rd4021, %rd4095; + xor.b64 %rd4097, %rd4096, %rd4024; + st.local.u64 [%rd2+184], %rd4097; + not.b64 %rd4098, %rd4021; + and.b64 %rd4099, %rd4022, %rd4098; + xor.b64 %rd4100, %rd4099, %rd4025; + st.local.u64 [%rd2+192], %rd4100; + xor.b64 %rd4101, %rd4028, %rd39; + st.local.u64 [%rd2], %rd4101; + sub.s64 %rd4102, %rd383, %rd14; + add.s64 %rd19844, %rd14, %rd19844; + setp.gt.s64 %p16, %rd19844, %rd4102; + @%p16 bra $L__BB1_29; + bra.uni $L__BB1_7; + +$L__BB1_29: + setp.lt.u64 %p26, %rd19844, 64; + @%p26 bra $L__BB1_6; + + shl.b64 %rd7993, %rd19843, 3; + st.local.u64 [%rd2+392], %rd7993; + and.b64 %rd7994, %rd19843, 2305843009213693951; + add.s64 %rd7995, %rd2, %rd7994; + ld.local.u8 %rs46, [%rd7995+200]; + or.b16 %rs47, %rs46, 1; + st.local.u8 [%rd7995+200], %rs47; + ld.local.u64 %rd7996, [%rd2+392]; + add.s64 %rd19853, %rd7996, 1; + st.local.u64 [%rd2+392], %rd19853; + ld.const.u64 %rd7997, [rate_bits]; + setp.ne.s64 %p27, %rd19853, %rd7997; + @%p27 bra $L__BB1_38; + + setp.eq.s64 %p28, %rd15, 0; + @%p28 bra $L__BB1_37; + + add.s64 %rd8000, %rd15, -1; + and.b64 %rd19852, %rd15, 3; + setp.lt.u64 %p29, %rd8000, 3; + mov.u64 %rd19848, 0; + mov.u64 %rd19849, %rd19848; + @%p29 bra $L__BB1_35; + + sub.s64 %rd19847, %rd15, %rd19852; + +$L__BB1_34: + add.s64 %rd8003, %rd2, %rd19849; + ld.local.u64 %rd8004, [%rd8003]; + ld.local.u8 %rd8005, [%rd8003+200]; + ld.local.u8 %rd8006, [%rd8003+201]; + bfi.b64 %rd8007, %rd8006, %rd8005, 8, 8; + ld.local.u8 %rd8008, [%rd8003+202]; + ld.local.u8 %rd8009, [%rd8003+203]; + bfi.b64 %rd8010, %rd8009, %rd8008, 8, 8; + bfi.b64 %rd8011, %rd8010, %rd8007, 16, 16; + ld.local.u8 %rd8012, [%rd8003+204]; + ld.local.u8 %rd8013, [%rd8003+205]; + bfi.b64 %rd8014, %rd8013, %rd8012, 8, 8; + ld.local.u8 %rd8015, [%rd8003+206]; + ld.local.u8 %rd8016, [%rd8003+207]; + bfi.b64 %rd8017, %rd8016, %rd8015, 8, 8; + bfi.b64 %rd8018, %rd8017, %rd8014, 16, 16; + bfi.b64 %rd8019, %rd8018, %rd8011, 32, 32; + xor.b64 %rd8020, %rd8004, %rd8019; + st.local.u64 [%rd8003], %rd8020; + ld.local.u64 %rd8021, [%rd8003+8]; + ld.local.u8 %rd8022, [%rd8003+208]; + ld.local.u8 %rd8023, [%rd8003+209]; + bfi.b64 %rd8024, %rd8023, %rd8022, 8, 8; + ld.local.u8 %rd8025, [%rd8003+210]; + ld.local.u8 %rd8026, [%rd8003+211]; + bfi.b64 %rd8027, %rd8026, %rd8025, 8, 8; + bfi.b64 %rd8028, %rd8027, %rd8024, 16, 16; + ld.local.u8 %rd8029, [%rd8003+212]; + ld.local.u8 %rd8030, [%rd8003+213]; + bfi.b64 %rd8031, %rd8030, %rd8029, 8, 8; + ld.local.u8 %rd8032, [%rd8003+214]; + ld.local.u8 %rd8033, [%rd8003+215]; + bfi.b64 %rd8034, %rd8033, %rd8032, 8, 8; + bfi.b64 %rd8035, %rd8034, %rd8031, 16, 16; + bfi.b64 %rd8036, %rd8035, %rd8028, 32, 32; + xor.b64 %rd8037, %rd8021, %rd8036; + st.local.u64 [%rd8003+8], %rd8037; + ld.local.u64 %rd8038, [%rd8003+16]; + ld.local.u8 %rd8039, [%rd8003+216]; + ld.local.u8 %rd8040, [%rd8003+217]; + bfi.b64 %rd8041, %rd8040, %rd8039, 8, 8; + ld.local.u8 %rd8042, [%rd8003+218]; + ld.local.u8 %rd8043, [%rd8003+219]; + bfi.b64 %rd8044, %rd8043, %rd8042, 8, 8; + bfi.b64 %rd8045, %rd8044, %rd8041, 16, 16; + ld.local.u8 %rd8046, [%rd8003+220]; + ld.local.u8 %rd8047, [%rd8003+221]; + bfi.b64 %rd8048, %rd8047, %rd8046, 8, 8; + ld.local.u8 %rd8049, [%rd8003+222]; + ld.local.u8 %rd8050, [%rd8003+223]; + bfi.b64 %rd8051, %rd8050, %rd8049, 8, 8; + bfi.b64 %rd8052, %rd8051, %rd8048, 16, 16; + bfi.b64 %rd8053, %rd8052, %rd8045, 32, 32; + xor.b64 %rd8054, %rd8038, %rd8053; + st.local.u64 [%rd8003+16], %rd8054; + ld.local.u64 %rd8055, [%rd8003+24]; + ld.local.u8 %rd8056, [%rd8003+224]; + ld.local.u8 %rd8057, [%rd8003+225]; + bfi.b64 %rd8058, %rd8057, %rd8056, 8, 8; + ld.local.u8 %rd8059, [%rd8003+226]; + ld.local.u8 %rd8060, [%rd8003+227]; + bfi.b64 %rd8061, %rd8060, %rd8059, 8, 8; + bfi.b64 %rd8062, %rd8061, %rd8058, 16, 16; + ld.local.u8 %rd8063, [%rd8003+228]; + ld.local.u8 %rd8064, [%rd8003+229]; + bfi.b64 %rd8065, %rd8064, %rd8063, 8, 8; + ld.local.u8 %rd8066, [%rd8003+230]; + ld.local.u8 %rd8067, [%rd8003+231]; + bfi.b64 %rd8068, %rd8067, %rd8066, 8, 8; + bfi.b64 %rd8069, %rd8068, %rd8065, 16, 16; + bfi.b64 %rd8070, %rd8069, %rd8062, 32, 32; + xor.b64 %rd8071, %rd8055, %rd8070; + st.local.u64 [%rd8003+24], %rd8071; + add.s64 %rd19849, %rd19849, 32; + add.s64 %rd19848, %rd19848, 4; + add.s64 %rd19847, %rd19847, -4; + setp.ne.s64 %p30, %rd19847, 0; + @%p30 bra $L__BB1_34; + +$L__BB1_35: + setp.eq.s64 %p31, %rd19852, 0; + @%p31 bra $L__BB1_37; + +$L__BB1_36: + .pragma "nounroll"; + add.s64 %rd8072, %rd2, %rd19849; + shl.b64 %rd8073, %rd19848, 3; + add.s64 %rd8074, %rd2, %rd8073; + ld.local.u64 %rd8075, [%rd8074]; + ld.local.u8 %rd8076, [%rd8072+200]; + ld.local.u8 %rd8077, [%rd8072+201]; + bfi.b64 %rd8078, %rd8077, %rd8076, 8, 8; + ld.local.u8 %rd8079, [%rd8072+202]; + ld.local.u8 %rd8080, [%rd8072+203]; + bfi.b64 %rd8081, %rd8080, %rd8079, 8, 8; + bfi.b64 %rd8082, %rd8081, %rd8078, 16, 16; + ld.local.u8 %rd8083, [%rd8072+204]; + ld.local.u8 %rd8084, [%rd8072+205]; + bfi.b64 %rd8085, %rd8084, %rd8083, 8, 8; + ld.local.u8 %rd8086, [%rd8072+206]; + ld.local.u8 %rd8087, [%rd8072+207]; + bfi.b64 %rd8088, %rd8087, %rd8086, 8, 8; + bfi.b64 %rd8089, %rd8088, %rd8085, 16, 16; + bfi.b64 %rd8090, %rd8089, %rd8082, 32, 32; + xor.b64 %rd8091, %rd8075, %rd8090; + st.local.u64 [%rd8074], %rd8091; + add.s64 %rd19849, %rd19849, 8; + add.s64 %rd19848, %rd19848, 1; + add.s64 %rd19852, %rd19852, -1; + setp.ne.s64 %p32, %rd19852, 0; + @%p32 bra $L__BB1_36; + +$L__BB1_37: + ld.local.u64 %rd8093, [%rd2+40]; + mov.u64 %rd19853, 0; + ld.local.u64 %rd8094, [%rd2]; + xor.b64 %rd8095, %rd8093, %rd8094; + ld.local.u64 %rd8096, [%rd2+80]; + xor.b64 %rd8097, %rd8095, %rd8096; + ld.local.u64 %rd8098, [%rd2+120]; + xor.b64 %rd8099, %rd8097, %rd8098; + ld.local.u64 %rd8100, [%rd2+160]; + xor.b64 %rd8101, %rd8099, %rd8100; + ld.local.u64 %rd8102, [%rd2+48]; + ld.local.u64 %rd8103, [%rd2+8]; + xor.b64 %rd8104, %rd8102, %rd8103; + ld.local.u64 %rd8105, [%rd2+88]; + xor.b64 %rd8106, %rd8104, %rd8105; + ld.local.u64 %rd8107, [%rd2+128]; + xor.b64 %rd8108, %rd8106, %rd8107; + ld.local.u64 %rd8109, [%rd2+168]; + xor.b64 %rd8110, %rd8108, %rd8109; + ld.local.u64 %rd8111, [%rd2+56]; + ld.local.u64 %rd8112, [%rd2+16]; + xor.b64 %rd8113, %rd8111, %rd8112; + ld.local.u64 %rd8114, [%rd2+96]; + xor.b64 %rd8115, %rd8113, %rd8114; + ld.local.u64 %rd8116, [%rd2+136]; + xor.b64 %rd8117, %rd8115, %rd8116; + ld.local.u64 %rd8118, [%rd2+176]; + xor.b64 %rd8119, %rd8117, %rd8118; + ld.local.u64 %rd8120, [%rd2+64]; + ld.local.u64 %rd8121, [%rd2+24]; + xor.b64 %rd8122, %rd8120, %rd8121; + ld.local.u64 %rd8123, [%rd2+104]; + xor.b64 %rd8124, %rd8122, %rd8123; + ld.local.u64 %rd8125, [%rd2+144]; + xor.b64 %rd8126, %rd8124, %rd8125; + ld.local.u64 %rd8127, [%rd2+184]; + xor.b64 %rd8128, %rd8126, %rd8127; + ld.local.u64 %rd8129, [%rd2+72]; + ld.local.u64 %rd8130, [%rd2+32]; + xor.b64 %rd8131, %rd8129, %rd8130; + ld.local.u64 %rd8132, [%rd2+112]; + xor.b64 %rd8133, %rd8131, %rd8132; + ld.local.u64 %rd8134, [%rd2+152]; + xor.b64 %rd8135, %rd8133, %rd8134; + ld.local.u64 %rd8136, [%rd2+192]; + xor.b64 %rd8137, %rd8135, %rd8136; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5646}, %rd8110; + } + { + .reg .b32 %dummy; + mov.b64 {%r5647,%dummy}, %rd8110; + } + shf.l.wrap.b32 %r5648, %r5647, %r5646, 1; + shf.l.wrap.b32 %r5649, %r5646, %r5647, 1; + mov.b64 %rd8138, {%r5649, %r5648}; + xor.b64 %rd8139, %rd8137, %rd8138; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5650}, %rd8119; + } + { + .reg .b32 %dummy; + mov.b64 {%r5651,%dummy}, %rd8119; + } + shf.l.wrap.b32 %r5652, %r5651, %r5650, 1; + shf.l.wrap.b32 %r5653, %r5650, %r5651, 1; + mov.b64 %rd8140, {%r5653, %r5652}; + xor.b64 %rd8141, %rd8140, %rd8101; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5654}, %rd8128; + } + { + .reg .b32 %dummy; + mov.b64 {%r5655,%dummy}, %rd8128; + } + shf.l.wrap.b32 %r5656, %r5655, %r5654, 1; + shf.l.wrap.b32 %r5657, %r5654, %r5655, 1; + mov.b64 %rd8142, {%r5657, %r5656}; + xor.b64 %rd8143, %rd8142, %rd8110; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5658}, %rd8137; + } + { + .reg .b32 %dummy; + mov.b64 {%r5659,%dummy}, %rd8137; + } + shf.l.wrap.b32 %r5660, %r5659, %r5658, 1; + shf.l.wrap.b32 %r5661, %r5658, %r5659, 1; + mov.b64 %rd8144, {%r5661, %r5660}; + xor.b64 %rd8145, %rd8144, %rd8119; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5662}, %rd8101; + } + { + .reg .b32 %dummy; + mov.b64 {%r5663,%dummy}, %rd8101; + } + shf.l.wrap.b32 %r5664, %r5663, %r5662, 1; + shf.l.wrap.b32 %r5665, %r5662, %r5663, 1; + mov.b64 %rd8146, {%r5665, %r5664}; + xor.b64 %rd8147, %rd8128, %rd8146; + xor.b64 %rd8148, %rd8139, %rd8094; + xor.b64 %rd8149, %rd8141, %rd8103; + xor.b64 %rd8150, %rd8143, %rd8112; + xor.b64 %rd8151, %rd8145, %rd8121; + xor.b64 %rd8152, %rd8147, %rd8130; + xor.b64 %rd8153, %rd8139, %rd8093; + xor.b64 %rd8154, %rd8141, %rd8102; + xor.b64 %rd8155, %rd8143, %rd8111; + xor.b64 %rd8156, %rd8145, %rd8120; + xor.b64 %rd8157, %rd8147, %rd8129; + xor.b64 %rd8158, %rd8139, %rd8096; + xor.b64 %rd8159, %rd8141, %rd8105; + xor.b64 %rd8160, %rd8143, %rd8114; + xor.b64 %rd8161, %rd8145, %rd8123; + xor.b64 %rd8162, %rd8147, %rd8132; + xor.b64 %rd8163, %rd8139, %rd8098; + xor.b64 %rd8164, %rd8141, %rd8107; + xor.b64 %rd8165, %rd8143, %rd8116; + xor.b64 %rd8166, %rd8145, %rd8125; + xor.b64 %rd8167, %rd8147, %rd8134; + xor.b64 %rd8168, %rd8139, %rd8100; + xor.b64 %rd8169, %rd8141, %rd8109; + xor.b64 %rd8170, %rd8143, %rd8118; + xor.b64 %rd8171, %rd8145, %rd8127; + xor.b64 %rd8172, %rd8147, %rd8136; + { + .reg .b32 %dummy; + mov.b64 {%r5666,%dummy}, %rd8154; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5667}, %rd8154; + } + shf.r.wrap.b32 %r5668, %r5667, %r5666, 20; + shf.r.wrap.b32 %r5669, %r5666, %r5667, 20; + mov.b64 %rd8173, {%r5669, %r5668}; + { + .reg .b32 %dummy; + mov.b64 {%r5670,%dummy}, %rd8160; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5671}, %rd8160; + } + shf.r.wrap.b32 %r5672, %r5671, %r5670, 21; + shf.r.wrap.b32 %r5673, %r5670, %r5671, 21; + mov.b64 %rd8174, {%r5673, %r5672}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5674}, %rd8166; + } + { + .reg .b32 %dummy; + mov.b64 {%r5675,%dummy}, %rd8166; + } + shf.l.wrap.b32 %r5676, %r5675, %r5674, 21; + shf.l.wrap.b32 %r5677, %r5674, %r5675, 21; + mov.b64 %rd8175, {%r5677, %r5676}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5678}, %rd8172; + } + { + .reg .b32 %dummy; + mov.b64 {%r5679,%dummy}, %rd8172; + } + shf.l.wrap.b32 %r5680, %r5679, %r5678, 14; + shf.l.wrap.b32 %r5681, %r5678, %r5679, 14; + mov.b64 %rd8176, {%r5681, %r5680}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5682}, %rd8151; + } + { + .reg .b32 %dummy; + mov.b64 {%r5683,%dummy}, %rd8151; + } + shf.l.wrap.b32 %r5684, %r5683, %r5682, 28; + shf.l.wrap.b32 %r5685, %r5682, %r5683, 28; + mov.b64 %rd8177, {%r5685, %r5684}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5686}, %rd8157; + } + { + .reg .b32 %dummy; + mov.b64 {%r5687,%dummy}, %rd8157; + } + shf.l.wrap.b32 %r5688, %r5687, %r5686, 20; + shf.l.wrap.b32 %r5689, %r5686, %r5687, 20; + mov.b64 %rd8178, {%r5689, %r5688}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5690}, %rd8158; + } + { + .reg .b32 %dummy; + mov.b64 {%r5691,%dummy}, %rd8158; + } + shf.l.wrap.b32 %r5692, %r5691, %r5690, 3; + shf.l.wrap.b32 %r5693, %r5690, %r5691, 3; + mov.b64 %rd8179, {%r5693, %r5692}; + { + .reg .b32 %dummy; + mov.b64 {%r5694,%dummy}, %rd8164; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5695}, %rd8164; + } + shf.r.wrap.b32 %r5696, %r5695, %r5694, 19; + shf.r.wrap.b32 %r5697, %r5694, %r5695, 19; + mov.b64 %rd8180, {%r5697, %r5696}; + { + .reg .b32 %dummy; + mov.b64 {%r5698,%dummy}, %rd8170; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5699}, %rd8170; + } + shf.r.wrap.b32 %r5700, %r5699, %r5698, 3; + shf.r.wrap.b32 %r5701, %r5698, %r5699, 3; + mov.b64 %rd8181, {%r5701, %r5700}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5702}, %rd8149; + } + { + .reg .b32 %dummy; + mov.b64 {%r5703,%dummy}, %rd8149; + } + shf.l.wrap.b32 %r5704, %r5703, %r5702, 1; + shf.l.wrap.b32 %r5705, %r5702, %r5703, 1; + mov.b64 %rd8182, {%r5705, %r5704}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5706}, %rd8155; + } + { + .reg .b32 %dummy; + mov.b64 {%r5707,%dummy}, %rd8155; + } + shf.l.wrap.b32 %r5708, %r5707, %r5706, 6; + shf.l.wrap.b32 %r5709, %r5706, %r5707, 6; + mov.b64 %rd8183, {%r5709, %r5708}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5710}, %rd8161; + } + { + .reg .b32 %dummy; + mov.b64 {%r5711,%dummy}, %rd8161; + } + shf.l.wrap.b32 %r5712, %r5711, %r5710, 25; + shf.l.wrap.b32 %r5713, %r5710, %r5711, 25; + mov.b64 %rd8184, {%r5713, %r5712}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5714}, %rd8167; + } + { + .reg .b32 %dummy; + mov.b64 {%r5715,%dummy}, %rd8167; + } + shf.l.wrap.b32 %r5716, %r5715, %r5714, 8; + shf.l.wrap.b32 %r5717, %r5714, %r5715, 8; + mov.b64 %rd8185, {%r5717, %r5716}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5718}, %rd8168; + } + { + .reg .b32 %dummy; + mov.b64 {%r5719,%dummy}, %rd8168; + } + shf.l.wrap.b32 %r5720, %r5719, %r5718, 18; + shf.l.wrap.b32 %r5721, %r5718, %r5719, 18; + mov.b64 %rd8186, {%r5721, %r5720}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5722}, %rd8152; + } + { + .reg .b32 %dummy; + mov.b64 {%r5723,%dummy}, %rd8152; + } + shf.l.wrap.b32 %r5724, %r5723, %r5722, 27; + shf.l.wrap.b32 %r5725, %r5722, %r5723, 27; + mov.b64 %rd8187, {%r5725, %r5724}; + { + .reg .b32 %dummy; + mov.b64 {%r5726,%dummy}, %rd8153; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5727}, %rd8153; + } + shf.r.wrap.b32 %r5728, %r5727, %r5726, 28; + shf.r.wrap.b32 %r5729, %r5726, %r5727, 28; + mov.b64 %rd8188, {%r5729, %r5728}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5730}, %rd8159; + } + { + .reg .b32 %dummy; + mov.b64 {%r5731,%dummy}, %rd8159; + } + shf.l.wrap.b32 %r5732, %r5731, %r5730, 10; + shf.l.wrap.b32 %r5733, %r5730, %r5731, 10; + mov.b64 %rd8189, {%r5733, %r5732}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5734}, %rd8165; + } + { + .reg .b32 %dummy; + mov.b64 {%r5735,%dummy}, %rd8165; + } + shf.l.wrap.b32 %r5736, %r5735, %r5734, 15; + shf.l.wrap.b32 %r5737, %r5734, %r5735, 15; + mov.b64 %rd8190, {%r5737, %r5736}; + { + .reg .b32 %dummy; + mov.b64 {%r5738,%dummy}, %rd8171; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5739}, %rd8171; + } + shf.r.wrap.b32 %r5740, %r5739, %r5738, 8; + shf.r.wrap.b32 %r5741, %r5738, %r5739, 8; + mov.b64 %rd8191, {%r5741, %r5740}; + { + .reg .b32 %dummy; + mov.b64 {%r5742,%dummy}, %rd8150; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5743}, %rd8150; + } + shf.r.wrap.b32 %r5744, %r5743, %r5742, 2; + shf.r.wrap.b32 %r5745, %r5742, %r5743, 2; + mov.b64 %rd8192, {%r5745, %r5744}; + { + .reg .b32 %dummy; + mov.b64 {%r5746,%dummy}, %rd8156; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5747}, %rd8156; + } + shf.r.wrap.b32 %r5748, %r5747, %r5746, 9; + shf.r.wrap.b32 %r5749, %r5746, %r5747, 9; + mov.b64 %rd8193, {%r5749, %r5748}; + { + .reg .b32 %dummy; + mov.b64 {%r5750,%dummy}, %rd8162; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5751}, %rd8162; + } + shf.r.wrap.b32 %r5752, %r5751, %r5750, 25; + shf.r.wrap.b32 %r5753, %r5750, %r5751, 25; + mov.b64 %rd8194, {%r5753, %r5752}; + { + .reg .b32 %dummy; + mov.b64 {%r5754,%dummy}, %rd8163; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5755}, %rd8163; + } + shf.r.wrap.b32 %r5756, %r5755, %r5754, 23; + shf.r.wrap.b32 %r5757, %r5754, %r5755, 23; + mov.b64 %rd8195, {%r5757, %r5756}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5758}, %rd8169; + } + { + .reg .b32 %dummy; + mov.b64 {%r5759,%dummy}, %rd8169; + } + shf.l.wrap.b32 %r5760, %r5759, %r5758, 2; + shf.l.wrap.b32 %r5761, %r5758, %r5759, 2; + mov.b64 %rd8196, {%r5761, %r5760}; + not.b64 %rd8197, %rd8173; + and.b64 %rd8198, %rd8174, %rd8197; + xor.b64 %rd8199, %rd8198, %rd8148; + not.b64 %rd8200, %rd8174; + and.b64 %rd8201, %rd8175, %rd8200; + xor.b64 %rd8202, %rd8201, %rd8173; + not.b64 %rd8203, %rd8175; + and.b64 %rd8204, %rd8176, %rd8203; + xor.b64 %rd8205, %rd8204, %rd8174; + not.b64 %rd8206, %rd8176; + and.b64 %rd8207, %rd8148, %rd8206; + xor.b64 %rd8208, %rd8207, %rd8175; + not.b64 %rd8209, %rd8148; + and.b64 %rd8210, %rd8173, %rd8209; + xor.b64 %rd8211, %rd8210, %rd8176; + not.b64 %rd8212, %rd8178; + and.b64 %rd8213, %rd8179, %rd8212; + xor.b64 %rd8214, %rd8213, %rd8177; + not.b64 %rd8215, %rd8179; + and.b64 %rd8216, %rd8180, %rd8215; + xor.b64 %rd8217, %rd8216, %rd8178; + not.b64 %rd8218, %rd8180; + and.b64 %rd8219, %rd8181, %rd8218; + xor.b64 %rd8220, %rd8219, %rd8179; + not.b64 %rd8221, %rd8181; + and.b64 %rd8222, %rd8177, %rd8221; + xor.b64 %rd8223, %rd8222, %rd8180; + not.b64 %rd8224, %rd8177; + and.b64 %rd8225, %rd8178, %rd8224; + xor.b64 %rd8226, %rd8225, %rd8181; + not.b64 %rd8227, %rd8183; + and.b64 %rd8228, %rd8184, %rd8227; + xor.b64 %rd8229, %rd8228, %rd8182; + not.b64 %rd8230, %rd8184; + and.b64 %rd8231, %rd8185, %rd8230; + xor.b64 %rd8232, %rd8231, %rd8183; + not.b64 %rd8233, %rd8185; + and.b64 %rd8234, %rd8186, %rd8233; + xor.b64 %rd8235, %rd8234, %rd8184; + not.b64 %rd8236, %rd8186; + and.b64 %rd8237, %rd8182, %rd8236; + xor.b64 %rd8238, %rd8237, %rd8185; + not.b64 %rd8239, %rd8182; + and.b64 %rd8240, %rd8183, %rd8239; + xor.b64 %rd8241, %rd8240, %rd8186; + not.b64 %rd8242, %rd8188; + and.b64 %rd8243, %rd8189, %rd8242; + xor.b64 %rd8244, %rd8243, %rd8187; + not.b64 %rd8245, %rd8189; + and.b64 %rd8246, %rd8190, %rd8245; + xor.b64 %rd8247, %rd8246, %rd8188; + not.b64 %rd8248, %rd8190; + and.b64 %rd8249, %rd8191, %rd8248; + xor.b64 %rd8250, %rd8249, %rd8189; + not.b64 %rd8251, %rd8191; + and.b64 %rd8252, %rd8187, %rd8251; + xor.b64 %rd8253, %rd8252, %rd8190; + not.b64 %rd8254, %rd8187; + and.b64 %rd8255, %rd8188, %rd8254; + xor.b64 %rd8256, %rd8255, %rd8191; + not.b64 %rd8257, %rd8193; + and.b64 %rd8258, %rd8194, %rd8257; + xor.b64 %rd8259, %rd8258, %rd8192; + not.b64 %rd8260, %rd8194; + and.b64 %rd8261, %rd8195, %rd8260; + xor.b64 %rd8262, %rd8261, %rd8193; + not.b64 %rd8263, %rd8195; + and.b64 %rd8264, %rd8196, %rd8263; + xor.b64 %rd8265, %rd8264, %rd8194; + not.b64 %rd8266, %rd8196; + and.b64 %rd8267, %rd8192, %rd8266; + xor.b64 %rd8268, %rd8267, %rd8195; + not.b64 %rd8269, %rd8192; + and.b64 %rd8270, %rd8193, %rd8269; + xor.b64 %rd8271, %rd8270, %rd8196; + xor.b64 %rd8273, %rd8199, %rd16; + xor.b64 %rd8274, %rd8214, %rd8273; + xor.b64 %rd8275, %rd8274, %rd8229; + xor.b64 %rd8276, %rd8275, %rd8244; + xor.b64 %rd8277, %rd8276, %rd8259; + xor.b64 %rd8278, %rd8217, %rd8202; + xor.b64 %rd8279, %rd8278, %rd8232; + xor.b64 %rd8280, %rd8279, %rd8247; + xor.b64 %rd8281, %rd8280, %rd8262; + xor.b64 %rd8282, %rd8220, %rd8205; + xor.b64 %rd8283, %rd8282, %rd8235; + xor.b64 %rd8284, %rd8283, %rd8250; + xor.b64 %rd8285, %rd8284, %rd8265; + xor.b64 %rd8286, %rd8223, %rd8208; + xor.b64 %rd8287, %rd8286, %rd8238; + xor.b64 %rd8288, %rd8287, %rd8253; + xor.b64 %rd8289, %rd8288, %rd8268; + xor.b64 %rd8290, %rd8226, %rd8211; + xor.b64 %rd8291, %rd8290, %rd8241; + xor.b64 %rd8292, %rd8291, %rd8256; + xor.b64 %rd8293, %rd8292, %rd8271; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5762}, %rd8281; + } + { + .reg .b32 %dummy; + mov.b64 {%r5763,%dummy}, %rd8281; + } + shf.l.wrap.b32 %r5764, %r5763, %r5762, 1; + shf.l.wrap.b32 %r5765, %r5762, %r5763, 1; + mov.b64 %rd8294, {%r5765, %r5764}; + xor.b64 %rd8295, %rd8293, %rd8294; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5766}, %rd8285; + } + { + .reg .b32 %dummy; + mov.b64 {%r5767,%dummy}, %rd8285; + } + shf.l.wrap.b32 %r5768, %r5767, %r5766, 1; + shf.l.wrap.b32 %r5769, %r5766, %r5767, 1; + mov.b64 %rd8296, {%r5769, %r5768}; + xor.b64 %rd8297, %rd8296, %rd8277; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5770}, %rd8289; + } + { + .reg .b32 %dummy; + mov.b64 {%r5771,%dummy}, %rd8289; + } + shf.l.wrap.b32 %r5772, %r5771, %r5770, 1; + shf.l.wrap.b32 %r5773, %r5770, %r5771, 1; + mov.b64 %rd8298, {%r5773, %r5772}; + xor.b64 %rd8299, %rd8298, %rd8281; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5774}, %rd8293; + } + { + .reg .b32 %dummy; + mov.b64 {%r5775,%dummy}, %rd8293; + } + shf.l.wrap.b32 %r5776, %r5775, %r5774, 1; + shf.l.wrap.b32 %r5777, %r5774, %r5775, 1; + mov.b64 %rd8300, {%r5777, %r5776}; + xor.b64 %rd8301, %rd8300, %rd8285; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5778}, %rd8277; + } + { + .reg .b32 %dummy; + mov.b64 {%r5779,%dummy}, %rd8277; + } + shf.l.wrap.b32 %r5780, %r5779, %r5778, 1; + shf.l.wrap.b32 %r5781, %r5778, %r5779, 1; + mov.b64 %rd8302, {%r5781, %r5780}; + xor.b64 %rd8303, %rd8289, %rd8302; + xor.b64 %rd8304, %rd8295, %rd8273; + xor.b64 %rd8305, %rd8297, %rd8202; + xor.b64 %rd8306, %rd8299, %rd8205; + xor.b64 %rd8307, %rd8301, %rd8208; + xor.b64 %rd8308, %rd8303, %rd8211; + xor.b64 %rd8309, %rd8295, %rd8214; + xor.b64 %rd8310, %rd8297, %rd8217; + xor.b64 %rd8311, %rd8299, %rd8220; + xor.b64 %rd8312, %rd8301, %rd8223; + xor.b64 %rd8313, %rd8303, %rd8226; + xor.b64 %rd8314, %rd8295, %rd8229; + xor.b64 %rd8315, %rd8297, %rd8232; + xor.b64 %rd8316, %rd8299, %rd8235; + xor.b64 %rd8317, %rd8301, %rd8238; + xor.b64 %rd8318, %rd8303, %rd8241; + xor.b64 %rd8319, %rd8295, %rd8244; + xor.b64 %rd8320, %rd8297, %rd8247; + xor.b64 %rd8321, %rd8299, %rd8250; + xor.b64 %rd8322, %rd8301, %rd8253; + xor.b64 %rd8323, %rd8303, %rd8256; + xor.b64 %rd8324, %rd8295, %rd8259; + xor.b64 %rd8325, %rd8297, %rd8262; + xor.b64 %rd8326, %rd8299, %rd8265; + xor.b64 %rd8327, %rd8301, %rd8268; + xor.b64 %rd8328, %rd8303, %rd8271; + { + .reg .b32 %dummy; + mov.b64 {%r5782,%dummy}, %rd8310; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5783}, %rd8310; + } + shf.r.wrap.b32 %r5784, %r5783, %r5782, 20; + shf.r.wrap.b32 %r5785, %r5782, %r5783, 20; + mov.b64 %rd8329, {%r5785, %r5784}; + { + .reg .b32 %dummy; + mov.b64 {%r5786,%dummy}, %rd8316; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5787}, %rd8316; + } + shf.r.wrap.b32 %r5788, %r5787, %r5786, 21; + shf.r.wrap.b32 %r5789, %r5786, %r5787, 21; + mov.b64 %rd8330, {%r5789, %r5788}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5790}, %rd8322; + } + { + .reg .b32 %dummy; + mov.b64 {%r5791,%dummy}, %rd8322; + } + shf.l.wrap.b32 %r5792, %r5791, %r5790, 21; + shf.l.wrap.b32 %r5793, %r5790, %r5791, 21; + mov.b64 %rd8331, {%r5793, %r5792}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5794}, %rd8328; + } + { + .reg .b32 %dummy; + mov.b64 {%r5795,%dummy}, %rd8328; + } + shf.l.wrap.b32 %r5796, %r5795, %r5794, 14; + shf.l.wrap.b32 %r5797, %r5794, %r5795, 14; + mov.b64 %rd8332, {%r5797, %r5796}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5798}, %rd8307; + } + { + .reg .b32 %dummy; + mov.b64 {%r5799,%dummy}, %rd8307; + } + shf.l.wrap.b32 %r5800, %r5799, %r5798, 28; + shf.l.wrap.b32 %r5801, %r5798, %r5799, 28; + mov.b64 %rd8333, {%r5801, %r5800}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5802}, %rd8313; + } + { + .reg .b32 %dummy; + mov.b64 {%r5803,%dummy}, %rd8313; + } + shf.l.wrap.b32 %r5804, %r5803, %r5802, 20; + shf.l.wrap.b32 %r5805, %r5802, %r5803, 20; + mov.b64 %rd8334, {%r5805, %r5804}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5806}, %rd8314; + } + { + .reg .b32 %dummy; + mov.b64 {%r5807,%dummy}, %rd8314; + } + shf.l.wrap.b32 %r5808, %r5807, %r5806, 3; + shf.l.wrap.b32 %r5809, %r5806, %r5807, 3; + mov.b64 %rd8335, {%r5809, %r5808}; + { + .reg .b32 %dummy; + mov.b64 {%r5810,%dummy}, %rd8320; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5811}, %rd8320; + } + shf.r.wrap.b32 %r5812, %r5811, %r5810, 19; + shf.r.wrap.b32 %r5813, %r5810, %r5811, 19; + mov.b64 %rd8336, {%r5813, %r5812}; + { + .reg .b32 %dummy; + mov.b64 {%r5814,%dummy}, %rd8326; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5815}, %rd8326; + } + shf.r.wrap.b32 %r5816, %r5815, %r5814, 3; + shf.r.wrap.b32 %r5817, %r5814, %r5815, 3; + mov.b64 %rd8337, {%r5817, %r5816}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5818}, %rd8305; + } + { + .reg .b32 %dummy; + mov.b64 {%r5819,%dummy}, %rd8305; + } + shf.l.wrap.b32 %r5820, %r5819, %r5818, 1; + shf.l.wrap.b32 %r5821, %r5818, %r5819, 1; + mov.b64 %rd8338, {%r5821, %r5820}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5822}, %rd8311; + } + { + .reg .b32 %dummy; + mov.b64 {%r5823,%dummy}, %rd8311; + } + shf.l.wrap.b32 %r5824, %r5823, %r5822, 6; + shf.l.wrap.b32 %r5825, %r5822, %r5823, 6; + mov.b64 %rd8339, {%r5825, %r5824}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5826}, %rd8317; + } + { + .reg .b32 %dummy; + mov.b64 {%r5827,%dummy}, %rd8317; + } + shf.l.wrap.b32 %r5828, %r5827, %r5826, 25; + shf.l.wrap.b32 %r5829, %r5826, %r5827, 25; + mov.b64 %rd8340, {%r5829, %r5828}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5830}, %rd8323; + } + { + .reg .b32 %dummy; + mov.b64 {%r5831,%dummy}, %rd8323; + } + shf.l.wrap.b32 %r5832, %r5831, %r5830, 8; + shf.l.wrap.b32 %r5833, %r5830, %r5831, 8; + mov.b64 %rd8341, {%r5833, %r5832}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5834}, %rd8324; + } + { + .reg .b32 %dummy; + mov.b64 {%r5835,%dummy}, %rd8324; + } + shf.l.wrap.b32 %r5836, %r5835, %r5834, 18; + shf.l.wrap.b32 %r5837, %r5834, %r5835, 18; + mov.b64 %rd8342, {%r5837, %r5836}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5838}, %rd8308; + } + { + .reg .b32 %dummy; + mov.b64 {%r5839,%dummy}, %rd8308; + } + shf.l.wrap.b32 %r5840, %r5839, %r5838, 27; + shf.l.wrap.b32 %r5841, %r5838, %r5839, 27; + mov.b64 %rd8343, {%r5841, %r5840}; + { + .reg .b32 %dummy; + mov.b64 {%r5842,%dummy}, %rd8309; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5843}, %rd8309; + } + shf.r.wrap.b32 %r5844, %r5843, %r5842, 28; + shf.r.wrap.b32 %r5845, %r5842, %r5843, 28; + mov.b64 %rd8344, {%r5845, %r5844}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5846}, %rd8315; + } + { + .reg .b32 %dummy; + mov.b64 {%r5847,%dummy}, %rd8315; + } + shf.l.wrap.b32 %r5848, %r5847, %r5846, 10; + shf.l.wrap.b32 %r5849, %r5846, %r5847, 10; + mov.b64 %rd8345, {%r5849, %r5848}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5850}, %rd8321; + } + { + .reg .b32 %dummy; + mov.b64 {%r5851,%dummy}, %rd8321; + } + shf.l.wrap.b32 %r5852, %r5851, %r5850, 15; + shf.l.wrap.b32 %r5853, %r5850, %r5851, 15; + mov.b64 %rd8346, {%r5853, %r5852}; + { + .reg .b32 %dummy; + mov.b64 {%r5854,%dummy}, %rd8327; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5855}, %rd8327; + } + shf.r.wrap.b32 %r5856, %r5855, %r5854, 8; + shf.r.wrap.b32 %r5857, %r5854, %r5855, 8; + mov.b64 %rd8347, {%r5857, %r5856}; + { + .reg .b32 %dummy; + mov.b64 {%r5858,%dummy}, %rd8306; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5859}, %rd8306; + } + shf.r.wrap.b32 %r5860, %r5859, %r5858, 2; + shf.r.wrap.b32 %r5861, %r5858, %r5859, 2; + mov.b64 %rd8348, {%r5861, %r5860}; + { + .reg .b32 %dummy; + mov.b64 {%r5862,%dummy}, %rd8312; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5863}, %rd8312; + } + shf.r.wrap.b32 %r5864, %r5863, %r5862, 9; + shf.r.wrap.b32 %r5865, %r5862, %r5863, 9; + mov.b64 %rd8349, {%r5865, %r5864}; + { + .reg .b32 %dummy; + mov.b64 {%r5866,%dummy}, %rd8318; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5867}, %rd8318; + } + shf.r.wrap.b32 %r5868, %r5867, %r5866, 25; + shf.r.wrap.b32 %r5869, %r5866, %r5867, 25; + mov.b64 %rd8350, {%r5869, %r5868}; + { + .reg .b32 %dummy; + mov.b64 {%r5870,%dummy}, %rd8319; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5871}, %rd8319; + } + shf.r.wrap.b32 %r5872, %r5871, %r5870, 23; + shf.r.wrap.b32 %r5873, %r5870, %r5871, 23; + mov.b64 %rd8351, {%r5873, %r5872}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5874}, %rd8325; + } + { + .reg .b32 %dummy; + mov.b64 {%r5875,%dummy}, %rd8325; + } + shf.l.wrap.b32 %r5876, %r5875, %r5874, 2; + shf.l.wrap.b32 %r5877, %r5874, %r5875, 2; + mov.b64 %rd8352, {%r5877, %r5876}; + not.b64 %rd8353, %rd8329; + and.b64 %rd8354, %rd8330, %rd8353; + xor.b64 %rd8355, %rd8354, %rd8304; + not.b64 %rd8356, %rd8330; + and.b64 %rd8357, %rd8331, %rd8356; + xor.b64 %rd8358, %rd8357, %rd8329; + not.b64 %rd8359, %rd8331; + and.b64 %rd8360, %rd8332, %rd8359; + xor.b64 %rd8361, %rd8360, %rd8330; + not.b64 %rd8362, %rd8332; + and.b64 %rd8363, %rd8304, %rd8362; + xor.b64 %rd8364, %rd8363, %rd8331; + not.b64 %rd8365, %rd8304; + and.b64 %rd8366, %rd8329, %rd8365; + xor.b64 %rd8367, %rd8366, %rd8332; + not.b64 %rd8368, %rd8334; + and.b64 %rd8369, %rd8335, %rd8368; + xor.b64 %rd8370, %rd8369, %rd8333; + not.b64 %rd8371, %rd8335; + and.b64 %rd8372, %rd8336, %rd8371; + xor.b64 %rd8373, %rd8372, %rd8334; + not.b64 %rd8374, %rd8336; + and.b64 %rd8375, %rd8337, %rd8374; + xor.b64 %rd8376, %rd8375, %rd8335; + not.b64 %rd8377, %rd8337; + and.b64 %rd8378, %rd8333, %rd8377; + xor.b64 %rd8379, %rd8378, %rd8336; + not.b64 %rd8380, %rd8333; + and.b64 %rd8381, %rd8334, %rd8380; + xor.b64 %rd8382, %rd8381, %rd8337; + not.b64 %rd8383, %rd8339; + and.b64 %rd8384, %rd8340, %rd8383; + xor.b64 %rd8385, %rd8384, %rd8338; + not.b64 %rd8386, %rd8340; + and.b64 %rd8387, %rd8341, %rd8386; + xor.b64 %rd8388, %rd8387, %rd8339; + not.b64 %rd8389, %rd8341; + and.b64 %rd8390, %rd8342, %rd8389; + xor.b64 %rd8391, %rd8390, %rd8340; + not.b64 %rd8392, %rd8342; + and.b64 %rd8393, %rd8338, %rd8392; + xor.b64 %rd8394, %rd8393, %rd8341; + not.b64 %rd8395, %rd8338; + and.b64 %rd8396, %rd8339, %rd8395; + xor.b64 %rd8397, %rd8396, %rd8342; + not.b64 %rd8398, %rd8344; + and.b64 %rd8399, %rd8345, %rd8398; + xor.b64 %rd8400, %rd8399, %rd8343; + not.b64 %rd8401, %rd8345; + and.b64 %rd8402, %rd8346, %rd8401; + xor.b64 %rd8403, %rd8402, %rd8344; + not.b64 %rd8404, %rd8346; + and.b64 %rd8405, %rd8347, %rd8404; + xor.b64 %rd8406, %rd8405, %rd8345; + not.b64 %rd8407, %rd8347; + and.b64 %rd8408, %rd8343, %rd8407; + xor.b64 %rd8409, %rd8408, %rd8346; + not.b64 %rd8410, %rd8343; + and.b64 %rd8411, %rd8344, %rd8410; + xor.b64 %rd8412, %rd8411, %rd8347; + not.b64 %rd8413, %rd8349; + and.b64 %rd8414, %rd8350, %rd8413; + xor.b64 %rd8415, %rd8414, %rd8348; + not.b64 %rd8416, %rd8350; + and.b64 %rd8417, %rd8351, %rd8416; + xor.b64 %rd8418, %rd8417, %rd8349; + not.b64 %rd8419, %rd8351; + and.b64 %rd8420, %rd8352, %rd8419; + xor.b64 %rd8421, %rd8420, %rd8350; + not.b64 %rd8422, %rd8352; + and.b64 %rd8423, %rd8348, %rd8422; + xor.b64 %rd8424, %rd8423, %rd8351; + not.b64 %rd8425, %rd8348; + and.b64 %rd8426, %rd8349, %rd8425; + xor.b64 %rd8427, %rd8426, %rd8352; + xor.b64 %rd8429, %rd8355, %rd17; + xor.b64 %rd8430, %rd8370, %rd8429; + xor.b64 %rd8431, %rd8430, %rd8385; + xor.b64 %rd8432, %rd8431, %rd8400; + xor.b64 %rd8433, %rd8432, %rd8415; + xor.b64 %rd8434, %rd8373, %rd8358; + xor.b64 %rd8435, %rd8434, %rd8388; + xor.b64 %rd8436, %rd8435, %rd8403; + xor.b64 %rd8437, %rd8436, %rd8418; + xor.b64 %rd8438, %rd8376, %rd8361; + xor.b64 %rd8439, %rd8438, %rd8391; + xor.b64 %rd8440, %rd8439, %rd8406; + xor.b64 %rd8441, %rd8440, %rd8421; + xor.b64 %rd8442, %rd8379, %rd8364; + xor.b64 %rd8443, %rd8442, %rd8394; + xor.b64 %rd8444, %rd8443, %rd8409; + xor.b64 %rd8445, %rd8444, %rd8424; + xor.b64 %rd8446, %rd8382, %rd8367; + xor.b64 %rd8447, %rd8446, %rd8397; + xor.b64 %rd8448, %rd8447, %rd8412; + xor.b64 %rd8449, %rd8448, %rd8427; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5878}, %rd8437; + } + { + .reg .b32 %dummy; + mov.b64 {%r5879,%dummy}, %rd8437; + } + shf.l.wrap.b32 %r5880, %r5879, %r5878, 1; + shf.l.wrap.b32 %r5881, %r5878, %r5879, 1; + mov.b64 %rd8450, {%r5881, %r5880}; + xor.b64 %rd8451, %rd8449, %rd8450; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5882}, %rd8441; + } + { + .reg .b32 %dummy; + mov.b64 {%r5883,%dummy}, %rd8441; + } + shf.l.wrap.b32 %r5884, %r5883, %r5882, 1; + shf.l.wrap.b32 %r5885, %r5882, %r5883, 1; + mov.b64 %rd8452, {%r5885, %r5884}; + xor.b64 %rd8453, %rd8452, %rd8433; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5886}, %rd8445; + } + { + .reg .b32 %dummy; + mov.b64 {%r5887,%dummy}, %rd8445; + } + shf.l.wrap.b32 %r5888, %r5887, %r5886, 1; + shf.l.wrap.b32 %r5889, %r5886, %r5887, 1; + mov.b64 %rd8454, {%r5889, %r5888}; + xor.b64 %rd8455, %rd8454, %rd8437; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5890}, %rd8449; + } + { + .reg .b32 %dummy; + mov.b64 {%r5891,%dummy}, %rd8449; + } + shf.l.wrap.b32 %r5892, %r5891, %r5890, 1; + shf.l.wrap.b32 %r5893, %r5890, %r5891, 1; + mov.b64 %rd8456, {%r5893, %r5892}; + xor.b64 %rd8457, %rd8456, %rd8441; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5894}, %rd8433; + } + { + .reg .b32 %dummy; + mov.b64 {%r5895,%dummy}, %rd8433; + } + shf.l.wrap.b32 %r5896, %r5895, %r5894, 1; + shf.l.wrap.b32 %r5897, %r5894, %r5895, 1; + mov.b64 %rd8458, {%r5897, %r5896}; + xor.b64 %rd8459, %rd8445, %rd8458; + xor.b64 %rd8460, %rd8451, %rd8429; + xor.b64 %rd8461, %rd8453, %rd8358; + xor.b64 %rd8462, %rd8455, %rd8361; + xor.b64 %rd8463, %rd8457, %rd8364; + xor.b64 %rd8464, %rd8459, %rd8367; + xor.b64 %rd8465, %rd8451, %rd8370; + xor.b64 %rd8466, %rd8453, %rd8373; + xor.b64 %rd8467, %rd8455, %rd8376; + xor.b64 %rd8468, %rd8457, %rd8379; + xor.b64 %rd8469, %rd8459, %rd8382; + xor.b64 %rd8470, %rd8451, %rd8385; + xor.b64 %rd8471, %rd8453, %rd8388; + xor.b64 %rd8472, %rd8455, %rd8391; + xor.b64 %rd8473, %rd8457, %rd8394; + xor.b64 %rd8474, %rd8459, %rd8397; + xor.b64 %rd8475, %rd8451, %rd8400; + xor.b64 %rd8476, %rd8453, %rd8403; + xor.b64 %rd8477, %rd8455, %rd8406; + xor.b64 %rd8478, %rd8457, %rd8409; + xor.b64 %rd8479, %rd8459, %rd8412; + xor.b64 %rd8480, %rd8451, %rd8415; + xor.b64 %rd8481, %rd8453, %rd8418; + xor.b64 %rd8482, %rd8455, %rd8421; + xor.b64 %rd8483, %rd8457, %rd8424; + xor.b64 %rd8484, %rd8459, %rd8427; + { + .reg .b32 %dummy; + mov.b64 {%r5898,%dummy}, %rd8466; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5899}, %rd8466; + } + shf.r.wrap.b32 %r5900, %r5899, %r5898, 20; + shf.r.wrap.b32 %r5901, %r5898, %r5899, 20; + mov.b64 %rd8485, {%r5901, %r5900}; + { + .reg .b32 %dummy; + mov.b64 {%r5902,%dummy}, %rd8472; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5903}, %rd8472; + } + shf.r.wrap.b32 %r5904, %r5903, %r5902, 21; + shf.r.wrap.b32 %r5905, %r5902, %r5903, 21; + mov.b64 %rd8486, {%r5905, %r5904}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5906}, %rd8478; + } + { + .reg .b32 %dummy; + mov.b64 {%r5907,%dummy}, %rd8478; + } + shf.l.wrap.b32 %r5908, %r5907, %r5906, 21; + shf.l.wrap.b32 %r5909, %r5906, %r5907, 21; + mov.b64 %rd8487, {%r5909, %r5908}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5910}, %rd8484; + } + { + .reg .b32 %dummy; + mov.b64 {%r5911,%dummy}, %rd8484; + } + shf.l.wrap.b32 %r5912, %r5911, %r5910, 14; + shf.l.wrap.b32 %r5913, %r5910, %r5911, 14; + mov.b64 %rd8488, {%r5913, %r5912}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5914}, %rd8463; + } + { + .reg .b32 %dummy; + mov.b64 {%r5915,%dummy}, %rd8463; + } + shf.l.wrap.b32 %r5916, %r5915, %r5914, 28; + shf.l.wrap.b32 %r5917, %r5914, %r5915, 28; + mov.b64 %rd8489, {%r5917, %r5916}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5918}, %rd8469; + } + { + .reg .b32 %dummy; + mov.b64 {%r5919,%dummy}, %rd8469; + } + shf.l.wrap.b32 %r5920, %r5919, %r5918, 20; + shf.l.wrap.b32 %r5921, %r5918, %r5919, 20; + mov.b64 %rd8490, {%r5921, %r5920}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5922}, %rd8470; + } + { + .reg .b32 %dummy; + mov.b64 {%r5923,%dummy}, %rd8470; + } + shf.l.wrap.b32 %r5924, %r5923, %r5922, 3; + shf.l.wrap.b32 %r5925, %r5922, %r5923, 3; + mov.b64 %rd8491, {%r5925, %r5924}; + { + .reg .b32 %dummy; + mov.b64 {%r5926,%dummy}, %rd8476; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5927}, %rd8476; + } + shf.r.wrap.b32 %r5928, %r5927, %r5926, 19; + shf.r.wrap.b32 %r5929, %r5926, %r5927, 19; + mov.b64 %rd8492, {%r5929, %r5928}; + { + .reg .b32 %dummy; + mov.b64 {%r5930,%dummy}, %rd8482; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5931}, %rd8482; + } + shf.r.wrap.b32 %r5932, %r5931, %r5930, 3; + shf.r.wrap.b32 %r5933, %r5930, %r5931, 3; + mov.b64 %rd8493, {%r5933, %r5932}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5934}, %rd8461; + } + { + .reg .b32 %dummy; + mov.b64 {%r5935,%dummy}, %rd8461; + } + shf.l.wrap.b32 %r5936, %r5935, %r5934, 1; + shf.l.wrap.b32 %r5937, %r5934, %r5935, 1; + mov.b64 %rd8494, {%r5937, %r5936}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5938}, %rd8467; + } + { + .reg .b32 %dummy; + mov.b64 {%r5939,%dummy}, %rd8467; + } + shf.l.wrap.b32 %r5940, %r5939, %r5938, 6; + shf.l.wrap.b32 %r5941, %r5938, %r5939, 6; + mov.b64 %rd8495, {%r5941, %r5940}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5942}, %rd8473; + } + { + .reg .b32 %dummy; + mov.b64 {%r5943,%dummy}, %rd8473; + } + shf.l.wrap.b32 %r5944, %r5943, %r5942, 25; + shf.l.wrap.b32 %r5945, %r5942, %r5943, 25; + mov.b64 %rd8496, {%r5945, %r5944}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5946}, %rd8479; + } + { + .reg .b32 %dummy; + mov.b64 {%r5947,%dummy}, %rd8479; + } + shf.l.wrap.b32 %r5948, %r5947, %r5946, 8; + shf.l.wrap.b32 %r5949, %r5946, %r5947, 8; + mov.b64 %rd8497, {%r5949, %r5948}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5950}, %rd8480; + } + { + .reg .b32 %dummy; + mov.b64 {%r5951,%dummy}, %rd8480; + } + shf.l.wrap.b32 %r5952, %r5951, %r5950, 18; + shf.l.wrap.b32 %r5953, %r5950, %r5951, 18; + mov.b64 %rd8498, {%r5953, %r5952}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5954}, %rd8464; + } + { + .reg .b32 %dummy; + mov.b64 {%r5955,%dummy}, %rd8464; + } + shf.l.wrap.b32 %r5956, %r5955, %r5954, 27; + shf.l.wrap.b32 %r5957, %r5954, %r5955, 27; + mov.b64 %rd8499, {%r5957, %r5956}; + { + .reg .b32 %dummy; + mov.b64 {%r5958,%dummy}, %rd8465; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5959}, %rd8465; + } + shf.r.wrap.b32 %r5960, %r5959, %r5958, 28; + shf.r.wrap.b32 %r5961, %r5958, %r5959, 28; + mov.b64 %rd8500, {%r5961, %r5960}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5962}, %rd8471; + } + { + .reg .b32 %dummy; + mov.b64 {%r5963,%dummy}, %rd8471; + } + shf.l.wrap.b32 %r5964, %r5963, %r5962, 10; + shf.l.wrap.b32 %r5965, %r5962, %r5963, 10; + mov.b64 %rd8501, {%r5965, %r5964}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5966}, %rd8477; + } + { + .reg .b32 %dummy; + mov.b64 {%r5967,%dummy}, %rd8477; + } + shf.l.wrap.b32 %r5968, %r5967, %r5966, 15; + shf.l.wrap.b32 %r5969, %r5966, %r5967, 15; + mov.b64 %rd8502, {%r5969, %r5968}; + { + .reg .b32 %dummy; + mov.b64 {%r5970,%dummy}, %rd8483; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5971}, %rd8483; + } + shf.r.wrap.b32 %r5972, %r5971, %r5970, 8; + shf.r.wrap.b32 %r5973, %r5970, %r5971, 8; + mov.b64 %rd8503, {%r5973, %r5972}; + { + .reg .b32 %dummy; + mov.b64 {%r5974,%dummy}, %rd8462; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5975}, %rd8462; + } + shf.r.wrap.b32 %r5976, %r5975, %r5974, 2; + shf.r.wrap.b32 %r5977, %r5974, %r5975, 2; + mov.b64 %rd8504, {%r5977, %r5976}; + { + .reg .b32 %dummy; + mov.b64 {%r5978,%dummy}, %rd8468; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5979}, %rd8468; + } + shf.r.wrap.b32 %r5980, %r5979, %r5978, 9; + shf.r.wrap.b32 %r5981, %r5978, %r5979, 9; + mov.b64 %rd8505, {%r5981, %r5980}; + { + .reg .b32 %dummy; + mov.b64 {%r5982,%dummy}, %rd8474; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5983}, %rd8474; + } + shf.r.wrap.b32 %r5984, %r5983, %r5982, 25; + shf.r.wrap.b32 %r5985, %r5982, %r5983, 25; + mov.b64 %rd8506, {%r5985, %r5984}; + { + .reg .b32 %dummy; + mov.b64 {%r5986,%dummy}, %rd8475; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5987}, %rd8475; + } + shf.r.wrap.b32 %r5988, %r5987, %r5986, 23; + shf.r.wrap.b32 %r5989, %r5986, %r5987, 23; + mov.b64 %rd8507, {%r5989, %r5988}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5990}, %rd8481; + } + { + .reg .b32 %dummy; + mov.b64 {%r5991,%dummy}, %rd8481; + } + shf.l.wrap.b32 %r5992, %r5991, %r5990, 2; + shf.l.wrap.b32 %r5993, %r5990, %r5991, 2; + mov.b64 %rd8508, {%r5993, %r5992}; + not.b64 %rd8509, %rd8485; + and.b64 %rd8510, %rd8486, %rd8509; + xor.b64 %rd8511, %rd8510, %rd8460; + not.b64 %rd8512, %rd8486; + and.b64 %rd8513, %rd8487, %rd8512; + xor.b64 %rd8514, %rd8513, %rd8485; + not.b64 %rd8515, %rd8487; + and.b64 %rd8516, %rd8488, %rd8515; + xor.b64 %rd8517, %rd8516, %rd8486; + not.b64 %rd8518, %rd8488; + and.b64 %rd8519, %rd8460, %rd8518; + xor.b64 %rd8520, %rd8519, %rd8487; + not.b64 %rd8521, %rd8460; + and.b64 %rd8522, %rd8485, %rd8521; + xor.b64 %rd8523, %rd8522, %rd8488; + not.b64 %rd8524, %rd8490; + and.b64 %rd8525, %rd8491, %rd8524; + xor.b64 %rd8526, %rd8525, %rd8489; + not.b64 %rd8527, %rd8491; + and.b64 %rd8528, %rd8492, %rd8527; + xor.b64 %rd8529, %rd8528, %rd8490; + not.b64 %rd8530, %rd8492; + and.b64 %rd8531, %rd8493, %rd8530; + xor.b64 %rd8532, %rd8531, %rd8491; + not.b64 %rd8533, %rd8493; + and.b64 %rd8534, %rd8489, %rd8533; + xor.b64 %rd8535, %rd8534, %rd8492; + not.b64 %rd8536, %rd8489; + and.b64 %rd8537, %rd8490, %rd8536; + xor.b64 %rd8538, %rd8537, %rd8493; + not.b64 %rd8539, %rd8495; + and.b64 %rd8540, %rd8496, %rd8539; + xor.b64 %rd8541, %rd8540, %rd8494; + not.b64 %rd8542, %rd8496; + and.b64 %rd8543, %rd8497, %rd8542; + xor.b64 %rd8544, %rd8543, %rd8495; + not.b64 %rd8545, %rd8497; + and.b64 %rd8546, %rd8498, %rd8545; + xor.b64 %rd8547, %rd8546, %rd8496; + not.b64 %rd8548, %rd8498; + and.b64 %rd8549, %rd8494, %rd8548; + xor.b64 %rd8550, %rd8549, %rd8497; + not.b64 %rd8551, %rd8494; + and.b64 %rd8552, %rd8495, %rd8551; + xor.b64 %rd8553, %rd8552, %rd8498; + not.b64 %rd8554, %rd8500; + and.b64 %rd8555, %rd8501, %rd8554; + xor.b64 %rd8556, %rd8555, %rd8499; + not.b64 %rd8557, %rd8501; + and.b64 %rd8558, %rd8502, %rd8557; + xor.b64 %rd8559, %rd8558, %rd8500; + not.b64 %rd8560, %rd8502; + and.b64 %rd8561, %rd8503, %rd8560; + xor.b64 %rd8562, %rd8561, %rd8501; + not.b64 %rd8563, %rd8503; + and.b64 %rd8564, %rd8499, %rd8563; + xor.b64 %rd8565, %rd8564, %rd8502; + not.b64 %rd8566, %rd8499; + and.b64 %rd8567, %rd8500, %rd8566; + xor.b64 %rd8568, %rd8567, %rd8503; + not.b64 %rd8569, %rd8505; + and.b64 %rd8570, %rd8506, %rd8569; + xor.b64 %rd8571, %rd8570, %rd8504; + not.b64 %rd8572, %rd8506; + and.b64 %rd8573, %rd8507, %rd8572; + xor.b64 %rd8574, %rd8573, %rd8505; + not.b64 %rd8575, %rd8507; + and.b64 %rd8576, %rd8508, %rd8575; + xor.b64 %rd8577, %rd8576, %rd8506; + not.b64 %rd8578, %rd8508; + and.b64 %rd8579, %rd8504, %rd8578; + xor.b64 %rd8580, %rd8579, %rd8507; + not.b64 %rd8581, %rd8504; + and.b64 %rd8582, %rd8505, %rd8581; + xor.b64 %rd8583, %rd8582, %rd8508; + xor.b64 %rd8585, %rd8511, %rd18; + xor.b64 %rd8586, %rd8526, %rd8585; + xor.b64 %rd8587, %rd8586, %rd8541; + xor.b64 %rd8588, %rd8587, %rd8556; + xor.b64 %rd8589, %rd8588, %rd8571; + xor.b64 %rd8590, %rd8529, %rd8514; + xor.b64 %rd8591, %rd8590, %rd8544; + xor.b64 %rd8592, %rd8591, %rd8559; + xor.b64 %rd8593, %rd8592, %rd8574; + xor.b64 %rd8594, %rd8532, %rd8517; + xor.b64 %rd8595, %rd8594, %rd8547; + xor.b64 %rd8596, %rd8595, %rd8562; + xor.b64 %rd8597, %rd8596, %rd8577; + xor.b64 %rd8598, %rd8535, %rd8520; + xor.b64 %rd8599, %rd8598, %rd8550; + xor.b64 %rd8600, %rd8599, %rd8565; + xor.b64 %rd8601, %rd8600, %rd8580; + xor.b64 %rd8602, %rd8538, %rd8523; + xor.b64 %rd8603, %rd8602, %rd8553; + xor.b64 %rd8604, %rd8603, %rd8568; + xor.b64 %rd8605, %rd8604, %rd8583; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5994}, %rd8593; + } + { + .reg .b32 %dummy; + mov.b64 {%r5995,%dummy}, %rd8593; + } + shf.l.wrap.b32 %r5996, %r5995, %r5994, 1; + shf.l.wrap.b32 %r5997, %r5994, %r5995, 1; + mov.b64 %rd8606, {%r5997, %r5996}; + xor.b64 %rd8607, %rd8605, %rd8606; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r5998}, %rd8597; + } + { + .reg .b32 %dummy; + mov.b64 {%r5999,%dummy}, %rd8597; + } + shf.l.wrap.b32 %r6000, %r5999, %r5998, 1; + shf.l.wrap.b32 %r6001, %r5998, %r5999, 1; + mov.b64 %rd8608, {%r6001, %r6000}; + xor.b64 %rd8609, %rd8608, %rd8589; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6002}, %rd8601; + } + { + .reg .b32 %dummy; + mov.b64 {%r6003,%dummy}, %rd8601; + } + shf.l.wrap.b32 %r6004, %r6003, %r6002, 1; + shf.l.wrap.b32 %r6005, %r6002, %r6003, 1; + mov.b64 %rd8610, {%r6005, %r6004}; + xor.b64 %rd8611, %rd8610, %rd8593; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6006}, %rd8605; + } + { + .reg .b32 %dummy; + mov.b64 {%r6007,%dummy}, %rd8605; + } + shf.l.wrap.b32 %r6008, %r6007, %r6006, 1; + shf.l.wrap.b32 %r6009, %r6006, %r6007, 1; + mov.b64 %rd8612, {%r6009, %r6008}; + xor.b64 %rd8613, %rd8612, %rd8597; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6010}, %rd8589; + } + { + .reg .b32 %dummy; + mov.b64 {%r6011,%dummy}, %rd8589; + } + shf.l.wrap.b32 %r6012, %r6011, %r6010, 1; + shf.l.wrap.b32 %r6013, %r6010, %r6011, 1; + mov.b64 %rd8614, {%r6013, %r6012}; + xor.b64 %rd8615, %rd8601, %rd8614; + xor.b64 %rd8616, %rd8607, %rd8585; + xor.b64 %rd8617, %rd8609, %rd8514; + xor.b64 %rd8618, %rd8611, %rd8517; + xor.b64 %rd8619, %rd8613, %rd8520; + xor.b64 %rd8620, %rd8615, %rd8523; + xor.b64 %rd8621, %rd8607, %rd8526; + xor.b64 %rd8622, %rd8609, %rd8529; + xor.b64 %rd8623, %rd8611, %rd8532; + xor.b64 %rd8624, %rd8613, %rd8535; + xor.b64 %rd8625, %rd8615, %rd8538; + xor.b64 %rd8626, %rd8607, %rd8541; + xor.b64 %rd8627, %rd8609, %rd8544; + xor.b64 %rd8628, %rd8611, %rd8547; + xor.b64 %rd8629, %rd8613, %rd8550; + xor.b64 %rd8630, %rd8615, %rd8553; + xor.b64 %rd8631, %rd8607, %rd8556; + xor.b64 %rd8632, %rd8609, %rd8559; + xor.b64 %rd8633, %rd8611, %rd8562; + xor.b64 %rd8634, %rd8613, %rd8565; + xor.b64 %rd8635, %rd8615, %rd8568; + xor.b64 %rd8636, %rd8607, %rd8571; + xor.b64 %rd8637, %rd8609, %rd8574; + xor.b64 %rd8638, %rd8611, %rd8577; + xor.b64 %rd8639, %rd8613, %rd8580; + xor.b64 %rd8640, %rd8615, %rd8583; + { + .reg .b32 %dummy; + mov.b64 {%r6014,%dummy}, %rd8622; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6015}, %rd8622; + } + shf.r.wrap.b32 %r6016, %r6015, %r6014, 20; + shf.r.wrap.b32 %r6017, %r6014, %r6015, 20; + mov.b64 %rd8641, {%r6017, %r6016}; + { + .reg .b32 %dummy; + mov.b64 {%r6018,%dummy}, %rd8628; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6019}, %rd8628; + } + shf.r.wrap.b32 %r6020, %r6019, %r6018, 21; + shf.r.wrap.b32 %r6021, %r6018, %r6019, 21; + mov.b64 %rd8642, {%r6021, %r6020}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6022}, %rd8634; + } + { + .reg .b32 %dummy; + mov.b64 {%r6023,%dummy}, %rd8634; + } + shf.l.wrap.b32 %r6024, %r6023, %r6022, 21; + shf.l.wrap.b32 %r6025, %r6022, %r6023, 21; + mov.b64 %rd8643, {%r6025, %r6024}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6026}, %rd8640; + } + { + .reg .b32 %dummy; + mov.b64 {%r6027,%dummy}, %rd8640; + } + shf.l.wrap.b32 %r6028, %r6027, %r6026, 14; + shf.l.wrap.b32 %r6029, %r6026, %r6027, 14; + mov.b64 %rd8644, {%r6029, %r6028}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6030}, %rd8619; + } + { + .reg .b32 %dummy; + mov.b64 {%r6031,%dummy}, %rd8619; + } + shf.l.wrap.b32 %r6032, %r6031, %r6030, 28; + shf.l.wrap.b32 %r6033, %r6030, %r6031, 28; + mov.b64 %rd8645, {%r6033, %r6032}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6034}, %rd8625; + } + { + .reg .b32 %dummy; + mov.b64 {%r6035,%dummy}, %rd8625; + } + shf.l.wrap.b32 %r6036, %r6035, %r6034, 20; + shf.l.wrap.b32 %r6037, %r6034, %r6035, 20; + mov.b64 %rd8646, {%r6037, %r6036}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6038}, %rd8626; + } + { + .reg .b32 %dummy; + mov.b64 {%r6039,%dummy}, %rd8626; + } + shf.l.wrap.b32 %r6040, %r6039, %r6038, 3; + shf.l.wrap.b32 %r6041, %r6038, %r6039, 3; + mov.b64 %rd8647, {%r6041, %r6040}; + { + .reg .b32 %dummy; + mov.b64 {%r6042,%dummy}, %rd8632; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6043}, %rd8632; + } + shf.r.wrap.b32 %r6044, %r6043, %r6042, 19; + shf.r.wrap.b32 %r6045, %r6042, %r6043, 19; + mov.b64 %rd8648, {%r6045, %r6044}; + { + .reg .b32 %dummy; + mov.b64 {%r6046,%dummy}, %rd8638; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6047}, %rd8638; + } + shf.r.wrap.b32 %r6048, %r6047, %r6046, 3; + shf.r.wrap.b32 %r6049, %r6046, %r6047, 3; + mov.b64 %rd8649, {%r6049, %r6048}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6050}, %rd8617; + } + { + .reg .b32 %dummy; + mov.b64 {%r6051,%dummy}, %rd8617; + } + shf.l.wrap.b32 %r6052, %r6051, %r6050, 1; + shf.l.wrap.b32 %r6053, %r6050, %r6051, 1; + mov.b64 %rd8650, {%r6053, %r6052}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6054}, %rd8623; + } + { + .reg .b32 %dummy; + mov.b64 {%r6055,%dummy}, %rd8623; + } + shf.l.wrap.b32 %r6056, %r6055, %r6054, 6; + shf.l.wrap.b32 %r6057, %r6054, %r6055, 6; + mov.b64 %rd8651, {%r6057, %r6056}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6058}, %rd8629; + } + { + .reg .b32 %dummy; + mov.b64 {%r6059,%dummy}, %rd8629; + } + shf.l.wrap.b32 %r6060, %r6059, %r6058, 25; + shf.l.wrap.b32 %r6061, %r6058, %r6059, 25; + mov.b64 %rd8652, {%r6061, %r6060}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6062}, %rd8635; + } + { + .reg .b32 %dummy; + mov.b64 {%r6063,%dummy}, %rd8635; + } + shf.l.wrap.b32 %r6064, %r6063, %r6062, 8; + shf.l.wrap.b32 %r6065, %r6062, %r6063, 8; + mov.b64 %rd8653, {%r6065, %r6064}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6066}, %rd8636; + } + { + .reg .b32 %dummy; + mov.b64 {%r6067,%dummy}, %rd8636; + } + shf.l.wrap.b32 %r6068, %r6067, %r6066, 18; + shf.l.wrap.b32 %r6069, %r6066, %r6067, 18; + mov.b64 %rd8654, {%r6069, %r6068}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6070}, %rd8620; + } + { + .reg .b32 %dummy; + mov.b64 {%r6071,%dummy}, %rd8620; + } + shf.l.wrap.b32 %r6072, %r6071, %r6070, 27; + shf.l.wrap.b32 %r6073, %r6070, %r6071, 27; + mov.b64 %rd8655, {%r6073, %r6072}; + { + .reg .b32 %dummy; + mov.b64 {%r6074,%dummy}, %rd8621; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6075}, %rd8621; + } + shf.r.wrap.b32 %r6076, %r6075, %r6074, 28; + shf.r.wrap.b32 %r6077, %r6074, %r6075, 28; + mov.b64 %rd8656, {%r6077, %r6076}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6078}, %rd8627; + } + { + .reg .b32 %dummy; + mov.b64 {%r6079,%dummy}, %rd8627; + } + shf.l.wrap.b32 %r6080, %r6079, %r6078, 10; + shf.l.wrap.b32 %r6081, %r6078, %r6079, 10; + mov.b64 %rd8657, {%r6081, %r6080}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6082}, %rd8633; + } + { + .reg .b32 %dummy; + mov.b64 {%r6083,%dummy}, %rd8633; + } + shf.l.wrap.b32 %r6084, %r6083, %r6082, 15; + shf.l.wrap.b32 %r6085, %r6082, %r6083, 15; + mov.b64 %rd8658, {%r6085, %r6084}; + { + .reg .b32 %dummy; + mov.b64 {%r6086,%dummy}, %rd8639; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6087}, %rd8639; + } + shf.r.wrap.b32 %r6088, %r6087, %r6086, 8; + shf.r.wrap.b32 %r6089, %r6086, %r6087, 8; + mov.b64 %rd8659, {%r6089, %r6088}; + { + .reg .b32 %dummy; + mov.b64 {%r6090,%dummy}, %rd8618; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6091}, %rd8618; + } + shf.r.wrap.b32 %r6092, %r6091, %r6090, 2; + shf.r.wrap.b32 %r6093, %r6090, %r6091, 2; + mov.b64 %rd8660, {%r6093, %r6092}; + { + .reg .b32 %dummy; + mov.b64 {%r6094,%dummy}, %rd8624; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6095}, %rd8624; + } + shf.r.wrap.b32 %r6096, %r6095, %r6094, 9; + shf.r.wrap.b32 %r6097, %r6094, %r6095, 9; + mov.b64 %rd8661, {%r6097, %r6096}; + { + .reg .b32 %dummy; + mov.b64 {%r6098,%dummy}, %rd8630; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6099}, %rd8630; + } + shf.r.wrap.b32 %r6100, %r6099, %r6098, 25; + shf.r.wrap.b32 %r6101, %r6098, %r6099, 25; + mov.b64 %rd8662, {%r6101, %r6100}; + { + .reg .b32 %dummy; + mov.b64 {%r6102,%dummy}, %rd8631; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6103}, %rd8631; + } + shf.r.wrap.b32 %r6104, %r6103, %r6102, 23; + shf.r.wrap.b32 %r6105, %r6102, %r6103, 23; + mov.b64 %rd8663, {%r6105, %r6104}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6106}, %rd8637; + } + { + .reg .b32 %dummy; + mov.b64 {%r6107,%dummy}, %rd8637; + } + shf.l.wrap.b32 %r6108, %r6107, %r6106, 2; + shf.l.wrap.b32 %r6109, %r6106, %r6107, 2; + mov.b64 %rd8664, {%r6109, %r6108}; + not.b64 %rd8665, %rd8641; + and.b64 %rd8666, %rd8642, %rd8665; + xor.b64 %rd8667, %rd8666, %rd8616; + not.b64 %rd8668, %rd8642; + and.b64 %rd8669, %rd8643, %rd8668; + xor.b64 %rd8670, %rd8669, %rd8641; + not.b64 %rd8671, %rd8643; + and.b64 %rd8672, %rd8644, %rd8671; + xor.b64 %rd8673, %rd8672, %rd8642; + not.b64 %rd8674, %rd8644; + and.b64 %rd8675, %rd8616, %rd8674; + xor.b64 %rd8676, %rd8675, %rd8643; + not.b64 %rd8677, %rd8616; + and.b64 %rd8678, %rd8641, %rd8677; + xor.b64 %rd8679, %rd8678, %rd8644; + not.b64 %rd8680, %rd8646; + and.b64 %rd8681, %rd8647, %rd8680; + xor.b64 %rd8682, %rd8681, %rd8645; + not.b64 %rd8683, %rd8647; + and.b64 %rd8684, %rd8648, %rd8683; + xor.b64 %rd8685, %rd8684, %rd8646; + not.b64 %rd8686, %rd8648; + and.b64 %rd8687, %rd8649, %rd8686; + xor.b64 %rd8688, %rd8687, %rd8647; + not.b64 %rd8689, %rd8649; + and.b64 %rd8690, %rd8645, %rd8689; + xor.b64 %rd8691, %rd8690, %rd8648; + not.b64 %rd8692, %rd8645; + and.b64 %rd8693, %rd8646, %rd8692; + xor.b64 %rd8694, %rd8693, %rd8649; + not.b64 %rd8695, %rd8651; + and.b64 %rd8696, %rd8652, %rd8695; + xor.b64 %rd8697, %rd8696, %rd8650; + not.b64 %rd8698, %rd8652; + and.b64 %rd8699, %rd8653, %rd8698; + xor.b64 %rd8700, %rd8699, %rd8651; + not.b64 %rd8701, %rd8653; + and.b64 %rd8702, %rd8654, %rd8701; + xor.b64 %rd8703, %rd8702, %rd8652; + not.b64 %rd8704, %rd8654; + and.b64 %rd8705, %rd8650, %rd8704; + xor.b64 %rd8706, %rd8705, %rd8653; + not.b64 %rd8707, %rd8650; + and.b64 %rd8708, %rd8651, %rd8707; + xor.b64 %rd8709, %rd8708, %rd8654; + not.b64 %rd8710, %rd8656; + and.b64 %rd8711, %rd8657, %rd8710; + xor.b64 %rd8712, %rd8711, %rd8655; + not.b64 %rd8713, %rd8657; + and.b64 %rd8714, %rd8658, %rd8713; + xor.b64 %rd8715, %rd8714, %rd8656; + not.b64 %rd8716, %rd8658; + and.b64 %rd8717, %rd8659, %rd8716; + xor.b64 %rd8718, %rd8717, %rd8657; + not.b64 %rd8719, %rd8659; + and.b64 %rd8720, %rd8655, %rd8719; + xor.b64 %rd8721, %rd8720, %rd8658; + not.b64 %rd8722, %rd8655; + and.b64 %rd8723, %rd8656, %rd8722; + xor.b64 %rd8724, %rd8723, %rd8659; + not.b64 %rd8725, %rd8661; + and.b64 %rd8726, %rd8662, %rd8725; + xor.b64 %rd8727, %rd8726, %rd8660; + not.b64 %rd8728, %rd8662; + and.b64 %rd8729, %rd8663, %rd8728; + xor.b64 %rd8730, %rd8729, %rd8661; + not.b64 %rd8731, %rd8663; + and.b64 %rd8732, %rd8664, %rd8731; + xor.b64 %rd8733, %rd8732, %rd8662; + not.b64 %rd8734, %rd8664; + and.b64 %rd8735, %rd8660, %rd8734; + xor.b64 %rd8736, %rd8735, %rd8663; + not.b64 %rd8737, %rd8660; + and.b64 %rd8738, %rd8661, %rd8737; + xor.b64 %rd8739, %rd8738, %rd8664; + xor.b64 %rd8741, %rd8667, %rd19; + xor.b64 %rd8742, %rd8682, %rd8741; + xor.b64 %rd8743, %rd8742, %rd8697; + xor.b64 %rd8744, %rd8743, %rd8712; + xor.b64 %rd8745, %rd8744, %rd8727; + xor.b64 %rd8746, %rd8685, %rd8670; + xor.b64 %rd8747, %rd8746, %rd8700; + xor.b64 %rd8748, %rd8747, %rd8715; + xor.b64 %rd8749, %rd8748, %rd8730; + xor.b64 %rd8750, %rd8688, %rd8673; + xor.b64 %rd8751, %rd8750, %rd8703; + xor.b64 %rd8752, %rd8751, %rd8718; + xor.b64 %rd8753, %rd8752, %rd8733; + xor.b64 %rd8754, %rd8691, %rd8676; + xor.b64 %rd8755, %rd8754, %rd8706; + xor.b64 %rd8756, %rd8755, %rd8721; + xor.b64 %rd8757, %rd8756, %rd8736; + xor.b64 %rd8758, %rd8694, %rd8679; + xor.b64 %rd8759, %rd8758, %rd8709; + xor.b64 %rd8760, %rd8759, %rd8724; + xor.b64 %rd8761, %rd8760, %rd8739; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6110}, %rd8749; + } + { + .reg .b32 %dummy; + mov.b64 {%r6111,%dummy}, %rd8749; + } + shf.l.wrap.b32 %r6112, %r6111, %r6110, 1; + shf.l.wrap.b32 %r6113, %r6110, %r6111, 1; + mov.b64 %rd8762, {%r6113, %r6112}; + xor.b64 %rd8763, %rd8761, %rd8762; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6114}, %rd8753; + } + { + .reg .b32 %dummy; + mov.b64 {%r6115,%dummy}, %rd8753; + } + shf.l.wrap.b32 %r6116, %r6115, %r6114, 1; + shf.l.wrap.b32 %r6117, %r6114, %r6115, 1; + mov.b64 %rd8764, {%r6117, %r6116}; + xor.b64 %rd8765, %rd8764, %rd8745; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6118}, %rd8757; + } + { + .reg .b32 %dummy; + mov.b64 {%r6119,%dummy}, %rd8757; + } + shf.l.wrap.b32 %r6120, %r6119, %r6118, 1; + shf.l.wrap.b32 %r6121, %r6118, %r6119, 1; + mov.b64 %rd8766, {%r6121, %r6120}; + xor.b64 %rd8767, %rd8766, %rd8749; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6122}, %rd8761; + } + { + .reg .b32 %dummy; + mov.b64 {%r6123,%dummy}, %rd8761; + } + shf.l.wrap.b32 %r6124, %r6123, %r6122, 1; + shf.l.wrap.b32 %r6125, %r6122, %r6123, 1; + mov.b64 %rd8768, {%r6125, %r6124}; + xor.b64 %rd8769, %rd8768, %rd8753; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6126}, %rd8745; + } + { + .reg .b32 %dummy; + mov.b64 {%r6127,%dummy}, %rd8745; + } + shf.l.wrap.b32 %r6128, %r6127, %r6126, 1; + shf.l.wrap.b32 %r6129, %r6126, %r6127, 1; + mov.b64 %rd8770, {%r6129, %r6128}; + xor.b64 %rd8771, %rd8757, %rd8770; + xor.b64 %rd8772, %rd8763, %rd8741; + xor.b64 %rd8773, %rd8765, %rd8670; + xor.b64 %rd8774, %rd8767, %rd8673; + xor.b64 %rd8775, %rd8769, %rd8676; + xor.b64 %rd8776, %rd8771, %rd8679; + xor.b64 %rd8777, %rd8763, %rd8682; + xor.b64 %rd8778, %rd8765, %rd8685; + xor.b64 %rd8779, %rd8767, %rd8688; + xor.b64 %rd8780, %rd8769, %rd8691; + xor.b64 %rd8781, %rd8771, %rd8694; + xor.b64 %rd8782, %rd8763, %rd8697; + xor.b64 %rd8783, %rd8765, %rd8700; + xor.b64 %rd8784, %rd8767, %rd8703; + xor.b64 %rd8785, %rd8769, %rd8706; + xor.b64 %rd8786, %rd8771, %rd8709; + xor.b64 %rd8787, %rd8763, %rd8712; + xor.b64 %rd8788, %rd8765, %rd8715; + xor.b64 %rd8789, %rd8767, %rd8718; + xor.b64 %rd8790, %rd8769, %rd8721; + xor.b64 %rd8791, %rd8771, %rd8724; + xor.b64 %rd8792, %rd8763, %rd8727; + xor.b64 %rd8793, %rd8765, %rd8730; + xor.b64 %rd8794, %rd8767, %rd8733; + xor.b64 %rd8795, %rd8769, %rd8736; + xor.b64 %rd8796, %rd8771, %rd8739; + { + .reg .b32 %dummy; + mov.b64 {%r6130,%dummy}, %rd8778; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6131}, %rd8778; + } + shf.r.wrap.b32 %r6132, %r6131, %r6130, 20; + shf.r.wrap.b32 %r6133, %r6130, %r6131, 20; + mov.b64 %rd8797, {%r6133, %r6132}; + { + .reg .b32 %dummy; + mov.b64 {%r6134,%dummy}, %rd8784; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6135}, %rd8784; + } + shf.r.wrap.b32 %r6136, %r6135, %r6134, 21; + shf.r.wrap.b32 %r6137, %r6134, %r6135, 21; + mov.b64 %rd8798, {%r6137, %r6136}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6138}, %rd8790; + } + { + .reg .b32 %dummy; + mov.b64 {%r6139,%dummy}, %rd8790; + } + shf.l.wrap.b32 %r6140, %r6139, %r6138, 21; + shf.l.wrap.b32 %r6141, %r6138, %r6139, 21; + mov.b64 %rd8799, {%r6141, %r6140}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6142}, %rd8796; + } + { + .reg .b32 %dummy; + mov.b64 {%r6143,%dummy}, %rd8796; + } + shf.l.wrap.b32 %r6144, %r6143, %r6142, 14; + shf.l.wrap.b32 %r6145, %r6142, %r6143, 14; + mov.b64 %rd8800, {%r6145, %r6144}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6146}, %rd8775; + } + { + .reg .b32 %dummy; + mov.b64 {%r6147,%dummy}, %rd8775; + } + shf.l.wrap.b32 %r6148, %r6147, %r6146, 28; + shf.l.wrap.b32 %r6149, %r6146, %r6147, 28; + mov.b64 %rd8801, {%r6149, %r6148}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6150}, %rd8781; + } + { + .reg .b32 %dummy; + mov.b64 {%r6151,%dummy}, %rd8781; + } + shf.l.wrap.b32 %r6152, %r6151, %r6150, 20; + shf.l.wrap.b32 %r6153, %r6150, %r6151, 20; + mov.b64 %rd8802, {%r6153, %r6152}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6154}, %rd8782; + } + { + .reg .b32 %dummy; + mov.b64 {%r6155,%dummy}, %rd8782; + } + shf.l.wrap.b32 %r6156, %r6155, %r6154, 3; + shf.l.wrap.b32 %r6157, %r6154, %r6155, 3; + mov.b64 %rd8803, {%r6157, %r6156}; + { + .reg .b32 %dummy; + mov.b64 {%r6158,%dummy}, %rd8788; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6159}, %rd8788; + } + shf.r.wrap.b32 %r6160, %r6159, %r6158, 19; + shf.r.wrap.b32 %r6161, %r6158, %r6159, 19; + mov.b64 %rd8804, {%r6161, %r6160}; + { + .reg .b32 %dummy; + mov.b64 {%r6162,%dummy}, %rd8794; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6163}, %rd8794; + } + shf.r.wrap.b32 %r6164, %r6163, %r6162, 3; + shf.r.wrap.b32 %r6165, %r6162, %r6163, 3; + mov.b64 %rd8805, {%r6165, %r6164}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6166}, %rd8773; + } + { + .reg .b32 %dummy; + mov.b64 {%r6167,%dummy}, %rd8773; + } + shf.l.wrap.b32 %r6168, %r6167, %r6166, 1; + shf.l.wrap.b32 %r6169, %r6166, %r6167, 1; + mov.b64 %rd8806, {%r6169, %r6168}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6170}, %rd8779; + } + { + .reg .b32 %dummy; + mov.b64 {%r6171,%dummy}, %rd8779; + } + shf.l.wrap.b32 %r6172, %r6171, %r6170, 6; + shf.l.wrap.b32 %r6173, %r6170, %r6171, 6; + mov.b64 %rd8807, {%r6173, %r6172}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6174}, %rd8785; + } + { + .reg .b32 %dummy; + mov.b64 {%r6175,%dummy}, %rd8785; + } + shf.l.wrap.b32 %r6176, %r6175, %r6174, 25; + shf.l.wrap.b32 %r6177, %r6174, %r6175, 25; + mov.b64 %rd8808, {%r6177, %r6176}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6178}, %rd8791; + } + { + .reg .b32 %dummy; + mov.b64 {%r6179,%dummy}, %rd8791; + } + shf.l.wrap.b32 %r6180, %r6179, %r6178, 8; + shf.l.wrap.b32 %r6181, %r6178, %r6179, 8; + mov.b64 %rd8809, {%r6181, %r6180}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6182}, %rd8792; + } + { + .reg .b32 %dummy; + mov.b64 {%r6183,%dummy}, %rd8792; + } + shf.l.wrap.b32 %r6184, %r6183, %r6182, 18; + shf.l.wrap.b32 %r6185, %r6182, %r6183, 18; + mov.b64 %rd8810, {%r6185, %r6184}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6186}, %rd8776; + } + { + .reg .b32 %dummy; + mov.b64 {%r6187,%dummy}, %rd8776; + } + shf.l.wrap.b32 %r6188, %r6187, %r6186, 27; + shf.l.wrap.b32 %r6189, %r6186, %r6187, 27; + mov.b64 %rd8811, {%r6189, %r6188}; + { + .reg .b32 %dummy; + mov.b64 {%r6190,%dummy}, %rd8777; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6191}, %rd8777; + } + shf.r.wrap.b32 %r6192, %r6191, %r6190, 28; + shf.r.wrap.b32 %r6193, %r6190, %r6191, 28; + mov.b64 %rd8812, {%r6193, %r6192}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6194}, %rd8783; + } + { + .reg .b32 %dummy; + mov.b64 {%r6195,%dummy}, %rd8783; + } + shf.l.wrap.b32 %r6196, %r6195, %r6194, 10; + shf.l.wrap.b32 %r6197, %r6194, %r6195, 10; + mov.b64 %rd8813, {%r6197, %r6196}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6198}, %rd8789; + } + { + .reg .b32 %dummy; + mov.b64 {%r6199,%dummy}, %rd8789; + } + shf.l.wrap.b32 %r6200, %r6199, %r6198, 15; + shf.l.wrap.b32 %r6201, %r6198, %r6199, 15; + mov.b64 %rd8814, {%r6201, %r6200}; + { + .reg .b32 %dummy; + mov.b64 {%r6202,%dummy}, %rd8795; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6203}, %rd8795; + } + shf.r.wrap.b32 %r6204, %r6203, %r6202, 8; + shf.r.wrap.b32 %r6205, %r6202, %r6203, 8; + mov.b64 %rd8815, {%r6205, %r6204}; + { + .reg .b32 %dummy; + mov.b64 {%r6206,%dummy}, %rd8774; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6207}, %rd8774; + } + shf.r.wrap.b32 %r6208, %r6207, %r6206, 2; + shf.r.wrap.b32 %r6209, %r6206, %r6207, 2; + mov.b64 %rd8816, {%r6209, %r6208}; + { + .reg .b32 %dummy; + mov.b64 {%r6210,%dummy}, %rd8780; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6211}, %rd8780; + } + shf.r.wrap.b32 %r6212, %r6211, %r6210, 9; + shf.r.wrap.b32 %r6213, %r6210, %r6211, 9; + mov.b64 %rd8817, {%r6213, %r6212}; + { + .reg .b32 %dummy; + mov.b64 {%r6214,%dummy}, %rd8786; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6215}, %rd8786; + } + shf.r.wrap.b32 %r6216, %r6215, %r6214, 25; + shf.r.wrap.b32 %r6217, %r6214, %r6215, 25; + mov.b64 %rd8818, {%r6217, %r6216}; + { + .reg .b32 %dummy; + mov.b64 {%r6218,%dummy}, %rd8787; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6219}, %rd8787; + } + shf.r.wrap.b32 %r6220, %r6219, %r6218, 23; + shf.r.wrap.b32 %r6221, %r6218, %r6219, 23; + mov.b64 %rd8819, {%r6221, %r6220}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6222}, %rd8793; + } + { + .reg .b32 %dummy; + mov.b64 {%r6223,%dummy}, %rd8793; + } + shf.l.wrap.b32 %r6224, %r6223, %r6222, 2; + shf.l.wrap.b32 %r6225, %r6222, %r6223, 2; + mov.b64 %rd8820, {%r6225, %r6224}; + not.b64 %rd8821, %rd8797; + and.b64 %rd8822, %rd8798, %rd8821; + xor.b64 %rd8823, %rd8822, %rd8772; + not.b64 %rd8824, %rd8798; + and.b64 %rd8825, %rd8799, %rd8824; + xor.b64 %rd8826, %rd8825, %rd8797; + not.b64 %rd8827, %rd8799; + and.b64 %rd8828, %rd8800, %rd8827; + xor.b64 %rd8829, %rd8828, %rd8798; + not.b64 %rd8830, %rd8800; + and.b64 %rd8831, %rd8772, %rd8830; + xor.b64 %rd8832, %rd8831, %rd8799; + not.b64 %rd8833, %rd8772; + and.b64 %rd8834, %rd8797, %rd8833; + xor.b64 %rd8835, %rd8834, %rd8800; + not.b64 %rd8836, %rd8802; + and.b64 %rd8837, %rd8803, %rd8836; + xor.b64 %rd8838, %rd8837, %rd8801; + not.b64 %rd8839, %rd8803; + and.b64 %rd8840, %rd8804, %rd8839; + xor.b64 %rd8841, %rd8840, %rd8802; + not.b64 %rd8842, %rd8804; + and.b64 %rd8843, %rd8805, %rd8842; + xor.b64 %rd8844, %rd8843, %rd8803; + not.b64 %rd8845, %rd8805; + and.b64 %rd8846, %rd8801, %rd8845; + xor.b64 %rd8847, %rd8846, %rd8804; + not.b64 %rd8848, %rd8801; + and.b64 %rd8849, %rd8802, %rd8848; + xor.b64 %rd8850, %rd8849, %rd8805; + not.b64 %rd8851, %rd8807; + and.b64 %rd8852, %rd8808, %rd8851; + xor.b64 %rd8853, %rd8852, %rd8806; + not.b64 %rd8854, %rd8808; + and.b64 %rd8855, %rd8809, %rd8854; + xor.b64 %rd8856, %rd8855, %rd8807; + not.b64 %rd8857, %rd8809; + and.b64 %rd8858, %rd8810, %rd8857; + xor.b64 %rd8859, %rd8858, %rd8808; + not.b64 %rd8860, %rd8810; + and.b64 %rd8861, %rd8806, %rd8860; + xor.b64 %rd8862, %rd8861, %rd8809; + not.b64 %rd8863, %rd8806; + and.b64 %rd8864, %rd8807, %rd8863; + xor.b64 %rd8865, %rd8864, %rd8810; + not.b64 %rd8866, %rd8812; + and.b64 %rd8867, %rd8813, %rd8866; + xor.b64 %rd8868, %rd8867, %rd8811; + not.b64 %rd8869, %rd8813; + and.b64 %rd8870, %rd8814, %rd8869; + xor.b64 %rd8871, %rd8870, %rd8812; + not.b64 %rd8872, %rd8814; + and.b64 %rd8873, %rd8815, %rd8872; + xor.b64 %rd8874, %rd8873, %rd8813; + not.b64 %rd8875, %rd8815; + and.b64 %rd8876, %rd8811, %rd8875; + xor.b64 %rd8877, %rd8876, %rd8814; + not.b64 %rd8878, %rd8811; + and.b64 %rd8879, %rd8812, %rd8878; + xor.b64 %rd8880, %rd8879, %rd8815; + not.b64 %rd8881, %rd8817; + and.b64 %rd8882, %rd8818, %rd8881; + xor.b64 %rd8883, %rd8882, %rd8816; + not.b64 %rd8884, %rd8818; + and.b64 %rd8885, %rd8819, %rd8884; + xor.b64 %rd8886, %rd8885, %rd8817; + not.b64 %rd8887, %rd8819; + and.b64 %rd8888, %rd8820, %rd8887; + xor.b64 %rd8889, %rd8888, %rd8818; + not.b64 %rd8890, %rd8820; + and.b64 %rd8891, %rd8816, %rd8890; + xor.b64 %rd8892, %rd8891, %rd8819; + not.b64 %rd8893, %rd8816; + and.b64 %rd8894, %rd8817, %rd8893; + xor.b64 %rd8895, %rd8894, %rd8820; + xor.b64 %rd8897, %rd8823, %rd20; + xor.b64 %rd8898, %rd8838, %rd8897; + xor.b64 %rd8899, %rd8898, %rd8853; + xor.b64 %rd8900, %rd8899, %rd8868; + xor.b64 %rd8901, %rd8900, %rd8883; + xor.b64 %rd8902, %rd8841, %rd8826; + xor.b64 %rd8903, %rd8902, %rd8856; + xor.b64 %rd8904, %rd8903, %rd8871; + xor.b64 %rd8905, %rd8904, %rd8886; + xor.b64 %rd8906, %rd8844, %rd8829; + xor.b64 %rd8907, %rd8906, %rd8859; + xor.b64 %rd8908, %rd8907, %rd8874; + xor.b64 %rd8909, %rd8908, %rd8889; + xor.b64 %rd8910, %rd8847, %rd8832; + xor.b64 %rd8911, %rd8910, %rd8862; + xor.b64 %rd8912, %rd8911, %rd8877; + xor.b64 %rd8913, %rd8912, %rd8892; + xor.b64 %rd8914, %rd8850, %rd8835; + xor.b64 %rd8915, %rd8914, %rd8865; + xor.b64 %rd8916, %rd8915, %rd8880; + xor.b64 %rd8917, %rd8916, %rd8895; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6226}, %rd8905; + } + { + .reg .b32 %dummy; + mov.b64 {%r6227,%dummy}, %rd8905; + } + shf.l.wrap.b32 %r6228, %r6227, %r6226, 1; + shf.l.wrap.b32 %r6229, %r6226, %r6227, 1; + mov.b64 %rd8918, {%r6229, %r6228}; + xor.b64 %rd8919, %rd8917, %rd8918; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6230}, %rd8909; + } + { + .reg .b32 %dummy; + mov.b64 {%r6231,%dummy}, %rd8909; + } + shf.l.wrap.b32 %r6232, %r6231, %r6230, 1; + shf.l.wrap.b32 %r6233, %r6230, %r6231, 1; + mov.b64 %rd8920, {%r6233, %r6232}; + xor.b64 %rd8921, %rd8920, %rd8901; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6234}, %rd8913; + } + { + .reg .b32 %dummy; + mov.b64 {%r6235,%dummy}, %rd8913; + } + shf.l.wrap.b32 %r6236, %r6235, %r6234, 1; + shf.l.wrap.b32 %r6237, %r6234, %r6235, 1; + mov.b64 %rd8922, {%r6237, %r6236}; + xor.b64 %rd8923, %rd8922, %rd8905; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6238}, %rd8917; + } + { + .reg .b32 %dummy; + mov.b64 {%r6239,%dummy}, %rd8917; + } + shf.l.wrap.b32 %r6240, %r6239, %r6238, 1; + shf.l.wrap.b32 %r6241, %r6238, %r6239, 1; + mov.b64 %rd8924, {%r6241, %r6240}; + xor.b64 %rd8925, %rd8924, %rd8909; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6242}, %rd8901; + } + { + .reg .b32 %dummy; + mov.b64 {%r6243,%dummy}, %rd8901; + } + shf.l.wrap.b32 %r6244, %r6243, %r6242, 1; + shf.l.wrap.b32 %r6245, %r6242, %r6243, 1; + mov.b64 %rd8926, {%r6245, %r6244}; + xor.b64 %rd8927, %rd8913, %rd8926; + xor.b64 %rd8928, %rd8919, %rd8897; + xor.b64 %rd8929, %rd8921, %rd8826; + xor.b64 %rd8930, %rd8923, %rd8829; + xor.b64 %rd8931, %rd8925, %rd8832; + xor.b64 %rd8932, %rd8927, %rd8835; + xor.b64 %rd8933, %rd8919, %rd8838; + xor.b64 %rd8934, %rd8921, %rd8841; + xor.b64 %rd8935, %rd8923, %rd8844; + xor.b64 %rd8936, %rd8925, %rd8847; + xor.b64 %rd8937, %rd8927, %rd8850; + xor.b64 %rd8938, %rd8919, %rd8853; + xor.b64 %rd8939, %rd8921, %rd8856; + xor.b64 %rd8940, %rd8923, %rd8859; + xor.b64 %rd8941, %rd8925, %rd8862; + xor.b64 %rd8942, %rd8927, %rd8865; + xor.b64 %rd8943, %rd8919, %rd8868; + xor.b64 %rd8944, %rd8921, %rd8871; + xor.b64 %rd8945, %rd8923, %rd8874; + xor.b64 %rd8946, %rd8925, %rd8877; + xor.b64 %rd8947, %rd8927, %rd8880; + xor.b64 %rd8948, %rd8919, %rd8883; + xor.b64 %rd8949, %rd8921, %rd8886; + xor.b64 %rd8950, %rd8923, %rd8889; + xor.b64 %rd8951, %rd8925, %rd8892; + xor.b64 %rd8952, %rd8927, %rd8895; + { + .reg .b32 %dummy; + mov.b64 {%r6246,%dummy}, %rd8934; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6247}, %rd8934; + } + shf.r.wrap.b32 %r6248, %r6247, %r6246, 20; + shf.r.wrap.b32 %r6249, %r6246, %r6247, 20; + mov.b64 %rd8953, {%r6249, %r6248}; + { + .reg .b32 %dummy; + mov.b64 {%r6250,%dummy}, %rd8940; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6251}, %rd8940; + } + shf.r.wrap.b32 %r6252, %r6251, %r6250, 21; + shf.r.wrap.b32 %r6253, %r6250, %r6251, 21; + mov.b64 %rd8954, {%r6253, %r6252}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6254}, %rd8946; + } + { + .reg .b32 %dummy; + mov.b64 {%r6255,%dummy}, %rd8946; + } + shf.l.wrap.b32 %r6256, %r6255, %r6254, 21; + shf.l.wrap.b32 %r6257, %r6254, %r6255, 21; + mov.b64 %rd8955, {%r6257, %r6256}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6258}, %rd8952; + } + { + .reg .b32 %dummy; + mov.b64 {%r6259,%dummy}, %rd8952; + } + shf.l.wrap.b32 %r6260, %r6259, %r6258, 14; + shf.l.wrap.b32 %r6261, %r6258, %r6259, 14; + mov.b64 %rd8956, {%r6261, %r6260}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6262}, %rd8931; + } + { + .reg .b32 %dummy; + mov.b64 {%r6263,%dummy}, %rd8931; + } + shf.l.wrap.b32 %r6264, %r6263, %r6262, 28; + shf.l.wrap.b32 %r6265, %r6262, %r6263, 28; + mov.b64 %rd8957, {%r6265, %r6264}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6266}, %rd8937; + } + { + .reg .b32 %dummy; + mov.b64 {%r6267,%dummy}, %rd8937; + } + shf.l.wrap.b32 %r6268, %r6267, %r6266, 20; + shf.l.wrap.b32 %r6269, %r6266, %r6267, 20; + mov.b64 %rd8958, {%r6269, %r6268}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6270}, %rd8938; + } + { + .reg .b32 %dummy; + mov.b64 {%r6271,%dummy}, %rd8938; + } + shf.l.wrap.b32 %r6272, %r6271, %r6270, 3; + shf.l.wrap.b32 %r6273, %r6270, %r6271, 3; + mov.b64 %rd8959, {%r6273, %r6272}; + { + .reg .b32 %dummy; + mov.b64 {%r6274,%dummy}, %rd8944; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6275}, %rd8944; + } + shf.r.wrap.b32 %r6276, %r6275, %r6274, 19; + shf.r.wrap.b32 %r6277, %r6274, %r6275, 19; + mov.b64 %rd8960, {%r6277, %r6276}; + { + .reg .b32 %dummy; + mov.b64 {%r6278,%dummy}, %rd8950; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6279}, %rd8950; + } + shf.r.wrap.b32 %r6280, %r6279, %r6278, 3; + shf.r.wrap.b32 %r6281, %r6278, %r6279, 3; + mov.b64 %rd8961, {%r6281, %r6280}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6282}, %rd8929; + } + { + .reg .b32 %dummy; + mov.b64 {%r6283,%dummy}, %rd8929; + } + shf.l.wrap.b32 %r6284, %r6283, %r6282, 1; + shf.l.wrap.b32 %r6285, %r6282, %r6283, 1; + mov.b64 %rd8962, {%r6285, %r6284}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6286}, %rd8935; + } + { + .reg .b32 %dummy; + mov.b64 {%r6287,%dummy}, %rd8935; + } + shf.l.wrap.b32 %r6288, %r6287, %r6286, 6; + shf.l.wrap.b32 %r6289, %r6286, %r6287, 6; + mov.b64 %rd8963, {%r6289, %r6288}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6290}, %rd8941; + } + { + .reg .b32 %dummy; + mov.b64 {%r6291,%dummy}, %rd8941; + } + shf.l.wrap.b32 %r6292, %r6291, %r6290, 25; + shf.l.wrap.b32 %r6293, %r6290, %r6291, 25; + mov.b64 %rd8964, {%r6293, %r6292}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6294}, %rd8947; + } + { + .reg .b32 %dummy; + mov.b64 {%r6295,%dummy}, %rd8947; + } + shf.l.wrap.b32 %r6296, %r6295, %r6294, 8; + shf.l.wrap.b32 %r6297, %r6294, %r6295, 8; + mov.b64 %rd8965, {%r6297, %r6296}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6298}, %rd8948; + } + { + .reg .b32 %dummy; + mov.b64 {%r6299,%dummy}, %rd8948; + } + shf.l.wrap.b32 %r6300, %r6299, %r6298, 18; + shf.l.wrap.b32 %r6301, %r6298, %r6299, 18; + mov.b64 %rd8966, {%r6301, %r6300}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6302}, %rd8932; + } + { + .reg .b32 %dummy; + mov.b64 {%r6303,%dummy}, %rd8932; + } + shf.l.wrap.b32 %r6304, %r6303, %r6302, 27; + shf.l.wrap.b32 %r6305, %r6302, %r6303, 27; + mov.b64 %rd8967, {%r6305, %r6304}; + { + .reg .b32 %dummy; + mov.b64 {%r6306,%dummy}, %rd8933; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6307}, %rd8933; + } + shf.r.wrap.b32 %r6308, %r6307, %r6306, 28; + shf.r.wrap.b32 %r6309, %r6306, %r6307, 28; + mov.b64 %rd8968, {%r6309, %r6308}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6310}, %rd8939; + } + { + .reg .b32 %dummy; + mov.b64 {%r6311,%dummy}, %rd8939; + } + shf.l.wrap.b32 %r6312, %r6311, %r6310, 10; + shf.l.wrap.b32 %r6313, %r6310, %r6311, 10; + mov.b64 %rd8969, {%r6313, %r6312}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6314}, %rd8945; + } + { + .reg .b32 %dummy; + mov.b64 {%r6315,%dummy}, %rd8945; + } + shf.l.wrap.b32 %r6316, %r6315, %r6314, 15; + shf.l.wrap.b32 %r6317, %r6314, %r6315, 15; + mov.b64 %rd8970, {%r6317, %r6316}; + { + .reg .b32 %dummy; + mov.b64 {%r6318,%dummy}, %rd8951; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6319}, %rd8951; + } + shf.r.wrap.b32 %r6320, %r6319, %r6318, 8; + shf.r.wrap.b32 %r6321, %r6318, %r6319, 8; + mov.b64 %rd8971, {%r6321, %r6320}; + { + .reg .b32 %dummy; + mov.b64 {%r6322,%dummy}, %rd8930; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6323}, %rd8930; + } + shf.r.wrap.b32 %r6324, %r6323, %r6322, 2; + shf.r.wrap.b32 %r6325, %r6322, %r6323, 2; + mov.b64 %rd8972, {%r6325, %r6324}; + { + .reg .b32 %dummy; + mov.b64 {%r6326,%dummy}, %rd8936; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6327}, %rd8936; + } + shf.r.wrap.b32 %r6328, %r6327, %r6326, 9; + shf.r.wrap.b32 %r6329, %r6326, %r6327, 9; + mov.b64 %rd8973, {%r6329, %r6328}; + { + .reg .b32 %dummy; + mov.b64 {%r6330,%dummy}, %rd8942; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6331}, %rd8942; + } + shf.r.wrap.b32 %r6332, %r6331, %r6330, 25; + shf.r.wrap.b32 %r6333, %r6330, %r6331, 25; + mov.b64 %rd8974, {%r6333, %r6332}; + { + .reg .b32 %dummy; + mov.b64 {%r6334,%dummy}, %rd8943; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6335}, %rd8943; + } + shf.r.wrap.b32 %r6336, %r6335, %r6334, 23; + shf.r.wrap.b32 %r6337, %r6334, %r6335, 23; + mov.b64 %rd8975, {%r6337, %r6336}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6338}, %rd8949; + } + { + .reg .b32 %dummy; + mov.b64 {%r6339,%dummy}, %rd8949; + } + shf.l.wrap.b32 %r6340, %r6339, %r6338, 2; + shf.l.wrap.b32 %r6341, %r6338, %r6339, 2; + mov.b64 %rd8976, {%r6341, %r6340}; + not.b64 %rd8977, %rd8953; + and.b64 %rd8978, %rd8954, %rd8977; + xor.b64 %rd8979, %rd8978, %rd8928; + not.b64 %rd8980, %rd8954; + and.b64 %rd8981, %rd8955, %rd8980; + xor.b64 %rd8982, %rd8981, %rd8953; + not.b64 %rd8983, %rd8955; + and.b64 %rd8984, %rd8956, %rd8983; + xor.b64 %rd8985, %rd8984, %rd8954; + not.b64 %rd8986, %rd8956; + and.b64 %rd8987, %rd8928, %rd8986; + xor.b64 %rd8988, %rd8987, %rd8955; + not.b64 %rd8989, %rd8928; + and.b64 %rd8990, %rd8953, %rd8989; + xor.b64 %rd8991, %rd8990, %rd8956; + not.b64 %rd8992, %rd8958; + and.b64 %rd8993, %rd8959, %rd8992; + xor.b64 %rd8994, %rd8993, %rd8957; + not.b64 %rd8995, %rd8959; + and.b64 %rd8996, %rd8960, %rd8995; + xor.b64 %rd8997, %rd8996, %rd8958; + not.b64 %rd8998, %rd8960; + and.b64 %rd8999, %rd8961, %rd8998; + xor.b64 %rd9000, %rd8999, %rd8959; + not.b64 %rd9001, %rd8961; + and.b64 %rd9002, %rd8957, %rd9001; + xor.b64 %rd9003, %rd9002, %rd8960; + not.b64 %rd9004, %rd8957; + and.b64 %rd9005, %rd8958, %rd9004; + xor.b64 %rd9006, %rd9005, %rd8961; + not.b64 %rd9007, %rd8963; + and.b64 %rd9008, %rd8964, %rd9007; + xor.b64 %rd9009, %rd9008, %rd8962; + not.b64 %rd9010, %rd8964; + and.b64 %rd9011, %rd8965, %rd9010; + xor.b64 %rd9012, %rd9011, %rd8963; + not.b64 %rd9013, %rd8965; + and.b64 %rd9014, %rd8966, %rd9013; + xor.b64 %rd9015, %rd9014, %rd8964; + not.b64 %rd9016, %rd8966; + and.b64 %rd9017, %rd8962, %rd9016; + xor.b64 %rd9018, %rd9017, %rd8965; + not.b64 %rd9019, %rd8962; + and.b64 %rd9020, %rd8963, %rd9019; + xor.b64 %rd9021, %rd9020, %rd8966; + not.b64 %rd9022, %rd8968; + and.b64 %rd9023, %rd8969, %rd9022; + xor.b64 %rd9024, %rd9023, %rd8967; + not.b64 %rd9025, %rd8969; + and.b64 %rd9026, %rd8970, %rd9025; + xor.b64 %rd9027, %rd9026, %rd8968; + not.b64 %rd9028, %rd8970; + and.b64 %rd9029, %rd8971, %rd9028; + xor.b64 %rd9030, %rd9029, %rd8969; + not.b64 %rd9031, %rd8971; + and.b64 %rd9032, %rd8967, %rd9031; + xor.b64 %rd9033, %rd9032, %rd8970; + not.b64 %rd9034, %rd8967; + and.b64 %rd9035, %rd8968, %rd9034; + xor.b64 %rd9036, %rd9035, %rd8971; + not.b64 %rd9037, %rd8973; + and.b64 %rd9038, %rd8974, %rd9037; + xor.b64 %rd9039, %rd9038, %rd8972; + not.b64 %rd9040, %rd8974; + and.b64 %rd9041, %rd8975, %rd9040; + xor.b64 %rd9042, %rd9041, %rd8973; + not.b64 %rd9043, %rd8975; + and.b64 %rd9044, %rd8976, %rd9043; + xor.b64 %rd9045, %rd9044, %rd8974; + not.b64 %rd9046, %rd8976; + and.b64 %rd9047, %rd8972, %rd9046; + xor.b64 %rd9048, %rd9047, %rd8975; + not.b64 %rd9049, %rd8972; + and.b64 %rd9050, %rd8973, %rd9049; + xor.b64 %rd9051, %rd9050, %rd8976; + xor.b64 %rd9053, %rd8979, %rd21; + xor.b64 %rd9054, %rd8994, %rd9053; + xor.b64 %rd9055, %rd9054, %rd9009; + xor.b64 %rd9056, %rd9055, %rd9024; + xor.b64 %rd9057, %rd9056, %rd9039; + xor.b64 %rd9058, %rd8997, %rd8982; + xor.b64 %rd9059, %rd9058, %rd9012; + xor.b64 %rd9060, %rd9059, %rd9027; + xor.b64 %rd9061, %rd9060, %rd9042; + xor.b64 %rd9062, %rd9000, %rd8985; + xor.b64 %rd9063, %rd9062, %rd9015; + xor.b64 %rd9064, %rd9063, %rd9030; + xor.b64 %rd9065, %rd9064, %rd9045; + xor.b64 %rd9066, %rd9003, %rd8988; + xor.b64 %rd9067, %rd9066, %rd9018; + xor.b64 %rd9068, %rd9067, %rd9033; + xor.b64 %rd9069, %rd9068, %rd9048; + xor.b64 %rd9070, %rd9006, %rd8991; + xor.b64 %rd9071, %rd9070, %rd9021; + xor.b64 %rd9072, %rd9071, %rd9036; + xor.b64 %rd9073, %rd9072, %rd9051; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6342}, %rd9061; + } + { + .reg .b32 %dummy; + mov.b64 {%r6343,%dummy}, %rd9061; + } + shf.l.wrap.b32 %r6344, %r6343, %r6342, 1; + shf.l.wrap.b32 %r6345, %r6342, %r6343, 1; + mov.b64 %rd9074, {%r6345, %r6344}; + xor.b64 %rd9075, %rd9073, %rd9074; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6346}, %rd9065; + } + { + .reg .b32 %dummy; + mov.b64 {%r6347,%dummy}, %rd9065; + } + shf.l.wrap.b32 %r6348, %r6347, %r6346, 1; + shf.l.wrap.b32 %r6349, %r6346, %r6347, 1; + mov.b64 %rd9076, {%r6349, %r6348}; + xor.b64 %rd9077, %rd9076, %rd9057; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6350}, %rd9069; + } + { + .reg .b32 %dummy; + mov.b64 {%r6351,%dummy}, %rd9069; + } + shf.l.wrap.b32 %r6352, %r6351, %r6350, 1; + shf.l.wrap.b32 %r6353, %r6350, %r6351, 1; + mov.b64 %rd9078, {%r6353, %r6352}; + xor.b64 %rd9079, %rd9078, %rd9061; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6354}, %rd9073; + } + { + .reg .b32 %dummy; + mov.b64 {%r6355,%dummy}, %rd9073; + } + shf.l.wrap.b32 %r6356, %r6355, %r6354, 1; + shf.l.wrap.b32 %r6357, %r6354, %r6355, 1; + mov.b64 %rd9080, {%r6357, %r6356}; + xor.b64 %rd9081, %rd9080, %rd9065; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6358}, %rd9057; + } + { + .reg .b32 %dummy; + mov.b64 {%r6359,%dummy}, %rd9057; + } + shf.l.wrap.b32 %r6360, %r6359, %r6358, 1; + shf.l.wrap.b32 %r6361, %r6358, %r6359, 1; + mov.b64 %rd9082, {%r6361, %r6360}; + xor.b64 %rd9083, %rd9069, %rd9082; + xor.b64 %rd9084, %rd9075, %rd9053; + xor.b64 %rd9085, %rd9077, %rd8982; + xor.b64 %rd9086, %rd9079, %rd8985; + xor.b64 %rd9087, %rd9081, %rd8988; + xor.b64 %rd9088, %rd9083, %rd8991; + xor.b64 %rd9089, %rd9075, %rd8994; + xor.b64 %rd9090, %rd9077, %rd8997; + xor.b64 %rd9091, %rd9079, %rd9000; + xor.b64 %rd9092, %rd9081, %rd9003; + xor.b64 %rd9093, %rd9083, %rd9006; + xor.b64 %rd9094, %rd9075, %rd9009; + xor.b64 %rd9095, %rd9077, %rd9012; + xor.b64 %rd9096, %rd9079, %rd9015; + xor.b64 %rd9097, %rd9081, %rd9018; + xor.b64 %rd9098, %rd9083, %rd9021; + xor.b64 %rd9099, %rd9075, %rd9024; + xor.b64 %rd9100, %rd9077, %rd9027; + xor.b64 %rd9101, %rd9079, %rd9030; + xor.b64 %rd9102, %rd9081, %rd9033; + xor.b64 %rd9103, %rd9083, %rd9036; + xor.b64 %rd9104, %rd9075, %rd9039; + xor.b64 %rd9105, %rd9077, %rd9042; + xor.b64 %rd9106, %rd9079, %rd9045; + xor.b64 %rd9107, %rd9081, %rd9048; + xor.b64 %rd9108, %rd9083, %rd9051; + { + .reg .b32 %dummy; + mov.b64 {%r6362,%dummy}, %rd9090; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6363}, %rd9090; + } + shf.r.wrap.b32 %r6364, %r6363, %r6362, 20; + shf.r.wrap.b32 %r6365, %r6362, %r6363, 20; + mov.b64 %rd9109, {%r6365, %r6364}; + { + .reg .b32 %dummy; + mov.b64 {%r6366,%dummy}, %rd9096; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6367}, %rd9096; + } + shf.r.wrap.b32 %r6368, %r6367, %r6366, 21; + shf.r.wrap.b32 %r6369, %r6366, %r6367, 21; + mov.b64 %rd9110, {%r6369, %r6368}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6370}, %rd9102; + } + { + .reg .b32 %dummy; + mov.b64 {%r6371,%dummy}, %rd9102; + } + shf.l.wrap.b32 %r6372, %r6371, %r6370, 21; + shf.l.wrap.b32 %r6373, %r6370, %r6371, 21; + mov.b64 %rd9111, {%r6373, %r6372}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6374}, %rd9108; + } + { + .reg .b32 %dummy; + mov.b64 {%r6375,%dummy}, %rd9108; + } + shf.l.wrap.b32 %r6376, %r6375, %r6374, 14; + shf.l.wrap.b32 %r6377, %r6374, %r6375, 14; + mov.b64 %rd9112, {%r6377, %r6376}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6378}, %rd9087; + } + { + .reg .b32 %dummy; + mov.b64 {%r6379,%dummy}, %rd9087; + } + shf.l.wrap.b32 %r6380, %r6379, %r6378, 28; + shf.l.wrap.b32 %r6381, %r6378, %r6379, 28; + mov.b64 %rd9113, {%r6381, %r6380}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6382}, %rd9093; + } + { + .reg .b32 %dummy; + mov.b64 {%r6383,%dummy}, %rd9093; + } + shf.l.wrap.b32 %r6384, %r6383, %r6382, 20; + shf.l.wrap.b32 %r6385, %r6382, %r6383, 20; + mov.b64 %rd9114, {%r6385, %r6384}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6386}, %rd9094; + } + { + .reg .b32 %dummy; + mov.b64 {%r6387,%dummy}, %rd9094; + } + shf.l.wrap.b32 %r6388, %r6387, %r6386, 3; + shf.l.wrap.b32 %r6389, %r6386, %r6387, 3; + mov.b64 %rd9115, {%r6389, %r6388}; + { + .reg .b32 %dummy; + mov.b64 {%r6390,%dummy}, %rd9100; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6391}, %rd9100; + } + shf.r.wrap.b32 %r6392, %r6391, %r6390, 19; + shf.r.wrap.b32 %r6393, %r6390, %r6391, 19; + mov.b64 %rd9116, {%r6393, %r6392}; + { + .reg .b32 %dummy; + mov.b64 {%r6394,%dummy}, %rd9106; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6395}, %rd9106; + } + shf.r.wrap.b32 %r6396, %r6395, %r6394, 3; + shf.r.wrap.b32 %r6397, %r6394, %r6395, 3; + mov.b64 %rd9117, {%r6397, %r6396}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6398}, %rd9085; + } + { + .reg .b32 %dummy; + mov.b64 {%r6399,%dummy}, %rd9085; + } + shf.l.wrap.b32 %r6400, %r6399, %r6398, 1; + shf.l.wrap.b32 %r6401, %r6398, %r6399, 1; + mov.b64 %rd9118, {%r6401, %r6400}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6402}, %rd9091; + } + { + .reg .b32 %dummy; + mov.b64 {%r6403,%dummy}, %rd9091; + } + shf.l.wrap.b32 %r6404, %r6403, %r6402, 6; + shf.l.wrap.b32 %r6405, %r6402, %r6403, 6; + mov.b64 %rd9119, {%r6405, %r6404}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6406}, %rd9097; + } + { + .reg .b32 %dummy; + mov.b64 {%r6407,%dummy}, %rd9097; + } + shf.l.wrap.b32 %r6408, %r6407, %r6406, 25; + shf.l.wrap.b32 %r6409, %r6406, %r6407, 25; + mov.b64 %rd9120, {%r6409, %r6408}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6410}, %rd9103; + } + { + .reg .b32 %dummy; + mov.b64 {%r6411,%dummy}, %rd9103; + } + shf.l.wrap.b32 %r6412, %r6411, %r6410, 8; + shf.l.wrap.b32 %r6413, %r6410, %r6411, 8; + mov.b64 %rd9121, {%r6413, %r6412}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6414}, %rd9104; + } + { + .reg .b32 %dummy; + mov.b64 {%r6415,%dummy}, %rd9104; + } + shf.l.wrap.b32 %r6416, %r6415, %r6414, 18; + shf.l.wrap.b32 %r6417, %r6414, %r6415, 18; + mov.b64 %rd9122, {%r6417, %r6416}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6418}, %rd9088; + } + { + .reg .b32 %dummy; + mov.b64 {%r6419,%dummy}, %rd9088; + } + shf.l.wrap.b32 %r6420, %r6419, %r6418, 27; + shf.l.wrap.b32 %r6421, %r6418, %r6419, 27; + mov.b64 %rd9123, {%r6421, %r6420}; + { + .reg .b32 %dummy; + mov.b64 {%r6422,%dummy}, %rd9089; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6423}, %rd9089; + } + shf.r.wrap.b32 %r6424, %r6423, %r6422, 28; + shf.r.wrap.b32 %r6425, %r6422, %r6423, 28; + mov.b64 %rd9124, {%r6425, %r6424}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6426}, %rd9095; + } + { + .reg .b32 %dummy; + mov.b64 {%r6427,%dummy}, %rd9095; + } + shf.l.wrap.b32 %r6428, %r6427, %r6426, 10; + shf.l.wrap.b32 %r6429, %r6426, %r6427, 10; + mov.b64 %rd9125, {%r6429, %r6428}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6430}, %rd9101; + } + { + .reg .b32 %dummy; + mov.b64 {%r6431,%dummy}, %rd9101; + } + shf.l.wrap.b32 %r6432, %r6431, %r6430, 15; + shf.l.wrap.b32 %r6433, %r6430, %r6431, 15; + mov.b64 %rd9126, {%r6433, %r6432}; + { + .reg .b32 %dummy; + mov.b64 {%r6434,%dummy}, %rd9107; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6435}, %rd9107; + } + shf.r.wrap.b32 %r6436, %r6435, %r6434, 8; + shf.r.wrap.b32 %r6437, %r6434, %r6435, 8; + mov.b64 %rd9127, {%r6437, %r6436}; + { + .reg .b32 %dummy; + mov.b64 {%r6438,%dummy}, %rd9086; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6439}, %rd9086; + } + shf.r.wrap.b32 %r6440, %r6439, %r6438, 2; + shf.r.wrap.b32 %r6441, %r6438, %r6439, 2; + mov.b64 %rd9128, {%r6441, %r6440}; + { + .reg .b32 %dummy; + mov.b64 {%r6442,%dummy}, %rd9092; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6443}, %rd9092; + } + shf.r.wrap.b32 %r6444, %r6443, %r6442, 9; + shf.r.wrap.b32 %r6445, %r6442, %r6443, 9; + mov.b64 %rd9129, {%r6445, %r6444}; + { + .reg .b32 %dummy; + mov.b64 {%r6446,%dummy}, %rd9098; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6447}, %rd9098; + } + shf.r.wrap.b32 %r6448, %r6447, %r6446, 25; + shf.r.wrap.b32 %r6449, %r6446, %r6447, 25; + mov.b64 %rd9130, {%r6449, %r6448}; + { + .reg .b32 %dummy; + mov.b64 {%r6450,%dummy}, %rd9099; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6451}, %rd9099; + } + shf.r.wrap.b32 %r6452, %r6451, %r6450, 23; + shf.r.wrap.b32 %r6453, %r6450, %r6451, 23; + mov.b64 %rd9131, {%r6453, %r6452}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6454}, %rd9105; + } + { + .reg .b32 %dummy; + mov.b64 {%r6455,%dummy}, %rd9105; + } + shf.l.wrap.b32 %r6456, %r6455, %r6454, 2; + shf.l.wrap.b32 %r6457, %r6454, %r6455, 2; + mov.b64 %rd9132, {%r6457, %r6456}; + not.b64 %rd9133, %rd9109; + and.b64 %rd9134, %rd9110, %rd9133; + xor.b64 %rd9135, %rd9134, %rd9084; + not.b64 %rd9136, %rd9110; + and.b64 %rd9137, %rd9111, %rd9136; + xor.b64 %rd9138, %rd9137, %rd9109; + not.b64 %rd9139, %rd9111; + and.b64 %rd9140, %rd9112, %rd9139; + xor.b64 %rd9141, %rd9140, %rd9110; + not.b64 %rd9142, %rd9112; + and.b64 %rd9143, %rd9084, %rd9142; + xor.b64 %rd9144, %rd9143, %rd9111; + not.b64 %rd9145, %rd9084; + and.b64 %rd9146, %rd9109, %rd9145; + xor.b64 %rd9147, %rd9146, %rd9112; + not.b64 %rd9148, %rd9114; + and.b64 %rd9149, %rd9115, %rd9148; + xor.b64 %rd9150, %rd9149, %rd9113; + not.b64 %rd9151, %rd9115; + and.b64 %rd9152, %rd9116, %rd9151; + xor.b64 %rd9153, %rd9152, %rd9114; + not.b64 %rd9154, %rd9116; + and.b64 %rd9155, %rd9117, %rd9154; + xor.b64 %rd9156, %rd9155, %rd9115; + not.b64 %rd9157, %rd9117; + and.b64 %rd9158, %rd9113, %rd9157; + xor.b64 %rd9159, %rd9158, %rd9116; + not.b64 %rd9160, %rd9113; + and.b64 %rd9161, %rd9114, %rd9160; + xor.b64 %rd9162, %rd9161, %rd9117; + not.b64 %rd9163, %rd9119; + and.b64 %rd9164, %rd9120, %rd9163; + xor.b64 %rd9165, %rd9164, %rd9118; + not.b64 %rd9166, %rd9120; + and.b64 %rd9167, %rd9121, %rd9166; + xor.b64 %rd9168, %rd9167, %rd9119; + not.b64 %rd9169, %rd9121; + and.b64 %rd9170, %rd9122, %rd9169; + xor.b64 %rd9171, %rd9170, %rd9120; + not.b64 %rd9172, %rd9122; + and.b64 %rd9173, %rd9118, %rd9172; + xor.b64 %rd9174, %rd9173, %rd9121; + not.b64 %rd9175, %rd9118; + and.b64 %rd9176, %rd9119, %rd9175; + xor.b64 %rd9177, %rd9176, %rd9122; + not.b64 %rd9178, %rd9124; + and.b64 %rd9179, %rd9125, %rd9178; + xor.b64 %rd9180, %rd9179, %rd9123; + not.b64 %rd9181, %rd9125; + and.b64 %rd9182, %rd9126, %rd9181; + xor.b64 %rd9183, %rd9182, %rd9124; + not.b64 %rd9184, %rd9126; + and.b64 %rd9185, %rd9127, %rd9184; + xor.b64 %rd9186, %rd9185, %rd9125; + not.b64 %rd9187, %rd9127; + and.b64 %rd9188, %rd9123, %rd9187; + xor.b64 %rd9189, %rd9188, %rd9126; + not.b64 %rd9190, %rd9123; + and.b64 %rd9191, %rd9124, %rd9190; + xor.b64 %rd9192, %rd9191, %rd9127; + not.b64 %rd9193, %rd9129; + and.b64 %rd9194, %rd9130, %rd9193; + xor.b64 %rd9195, %rd9194, %rd9128; + not.b64 %rd9196, %rd9130; + and.b64 %rd9197, %rd9131, %rd9196; + xor.b64 %rd9198, %rd9197, %rd9129; + not.b64 %rd9199, %rd9131; + and.b64 %rd9200, %rd9132, %rd9199; + xor.b64 %rd9201, %rd9200, %rd9130; + not.b64 %rd9202, %rd9132; + and.b64 %rd9203, %rd9128, %rd9202; + xor.b64 %rd9204, %rd9203, %rd9131; + not.b64 %rd9205, %rd9128; + and.b64 %rd9206, %rd9129, %rd9205; + xor.b64 %rd9207, %rd9206, %rd9132; + xor.b64 %rd9209, %rd9135, %rd22; + xor.b64 %rd9210, %rd9150, %rd9209; + xor.b64 %rd9211, %rd9210, %rd9165; + xor.b64 %rd9212, %rd9211, %rd9180; + xor.b64 %rd9213, %rd9212, %rd9195; + xor.b64 %rd9214, %rd9153, %rd9138; + xor.b64 %rd9215, %rd9214, %rd9168; + xor.b64 %rd9216, %rd9215, %rd9183; + xor.b64 %rd9217, %rd9216, %rd9198; + xor.b64 %rd9218, %rd9156, %rd9141; + xor.b64 %rd9219, %rd9218, %rd9171; + xor.b64 %rd9220, %rd9219, %rd9186; + xor.b64 %rd9221, %rd9220, %rd9201; + xor.b64 %rd9222, %rd9159, %rd9144; + xor.b64 %rd9223, %rd9222, %rd9174; + xor.b64 %rd9224, %rd9223, %rd9189; + xor.b64 %rd9225, %rd9224, %rd9204; + xor.b64 %rd9226, %rd9162, %rd9147; + xor.b64 %rd9227, %rd9226, %rd9177; + xor.b64 %rd9228, %rd9227, %rd9192; + xor.b64 %rd9229, %rd9228, %rd9207; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6458}, %rd9217; + } + { + .reg .b32 %dummy; + mov.b64 {%r6459,%dummy}, %rd9217; + } + shf.l.wrap.b32 %r6460, %r6459, %r6458, 1; + shf.l.wrap.b32 %r6461, %r6458, %r6459, 1; + mov.b64 %rd9230, {%r6461, %r6460}; + xor.b64 %rd9231, %rd9229, %rd9230; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6462}, %rd9221; + } + { + .reg .b32 %dummy; + mov.b64 {%r6463,%dummy}, %rd9221; + } + shf.l.wrap.b32 %r6464, %r6463, %r6462, 1; + shf.l.wrap.b32 %r6465, %r6462, %r6463, 1; + mov.b64 %rd9232, {%r6465, %r6464}; + xor.b64 %rd9233, %rd9232, %rd9213; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6466}, %rd9225; + } + { + .reg .b32 %dummy; + mov.b64 {%r6467,%dummy}, %rd9225; + } + shf.l.wrap.b32 %r6468, %r6467, %r6466, 1; + shf.l.wrap.b32 %r6469, %r6466, %r6467, 1; + mov.b64 %rd9234, {%r6469, %r6468}; + xor.b64 %rd9235, %rd9234, %rd9217; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6470}, %rd9229; + } + { + .reg .b32 %dummy; + mov.b64 {%r6471,%dummy}, %rd9229; + } + shf.l.wrap.b32 %r6472, %r6471, %r6470, 1; + shf.l.wrap.b32 %r6473, %r6470, %r6471, 1; + mov.b64 %rd9236, {%r6473, %r6472}; + xor.b64 %rd9237, %rd9236, %rd9221; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6474}, %rd9213; + } + { + .reg .b32 %dummy; + mov.b64 {%r6475,%dummy}, %rd9213; + } + shf.l.wrap.b32 %r6476, %r6475, %r6474, 1; + shf.l.wrap.b32 %r6477, %r6474, %r6475, 1; + mov.b64 %rd9238, {%r6477, %r6476}; + xor.b64 %rd9239, %rd9225, %rd9238; + xor.b64 %rd9240, %rd9231, %rd9209; + xor.b64 %rd9241, %rd9233, %rd9138; + xor.b64 %rd9242, %rd9235, %rd9141; + xor.b64 %rd9243, %rd9237, %rd9144; + xor.b64 %rd9244, %rd9239, %rd9147; + xor.b64 %rd9245, %rd9231, %rd9150; + xor.b64 %rd9246, %rd9233, %rd9153; + xor.b64 %rd9247, %rd9235, %rd9156; + xor.b64 %rd9248, %rd9237, %rd9159; + xor.b64 %rd9249, %rd9239, %rd9162; + xor.b64 %rd9250, %rd9231, %rd9165; + xor.b64 %rd9251, %rd9233, %rd9168; + xor.b64 %rd9252, %rd9235, %rd9171; + xor.b64 %rd9253, %rd9237, %rd9174; + xor.b64 %rd9254, %rd9239, %rd9177; + xor.b64 %rd9255, %rd9231, %rd9180; + xor.b64 %rd9256, %rd9233, %rd9183; + xor.b64 %rd9257, %rd9235, %rd9186; + xor.b64 %rd9258, %rd9237, %rd9189; + xor.b64 %rd9259, %rd9239, %rd9192; + xor.b64 %rd9260, %rd9231, %rd9195; + xor.b64 %rd9261, %rd9233, %rd9198; + xor.b64 %rd9262, %rd9235, %rd9201; + xor.b64 %rd9263, %rd9237, %rd9204; + xor.b64 %rd9264, %rd9239, %rd9207; + { + .reg .b32 %dummy; + mov.b64 {%r6478,%dummy}, %rd9246; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6479}, %rd9246; + } + shf.r.wrap.b32 %r6480, %r6479, %r6478, 20; + shf.r.wrap.b32 %r6481, %r6478, %r6479, 20; + mov.b64 %rd9265, {%r6481, %r6480}; + { + .reg .b32 %dummy; + mov.b64 {%r6482,%dummy}, %rd9252; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6483}, %rd9252; + } + shf.r.wrap.b32 %r6484, %r6483, %r6482, 21; + shf.r.wrap.b32 %r6485, %r6482, %r6483, 21; + mov.b64 %rd9266, {%r6485, %r6484}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6486}, %rd9258; + } + { + .reg .b32 %dummy; + mov.b64 {%r6487,%dummy}, %rd9258; + } + shf.l.wrap.b32 %r6488, %r6487, %r6486, 21; + shf.l.wrap.b32 %r6489, %r6486, %r6487, 21; + mov.b64 %rd9267, {%r6489, %r6488}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6490}, %rd9264; + } + { + .reg .b32 %dummy; + mov.b64 {%r6491,%dummy}, %rd9264; + } + shf.l.wrap.b32 %r6492, %r6491, %r6490, 14; + shf.l.wrap.b32 %r6493, %r6490, %r6491, 14; + mov.b64 %rd9268, {%r6493, %r6492}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6494}, %rd9243; + } + { + .reg .b32 %dummy; + mov.b64 {%r6495,%dummy}, %rd9243; + } + shf.l.wrap.b32 %r6496, %r6495, %r6494, 28; + shf.l.wrap.b32 %r6497, %r6494, %r6495, 28; + mov.b64 %rd9269, {%r6497, %r6496}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6498}, %rd9249; + } + { + .reg .b32 %dummy; + mov.b64 {%r6499,%dummy}, %rd9249; + } + shf.l.wrap.b32 %r6500, %r6499, %r6498, 20; + shf.l.wrap.b32 %r6501, %r6498, %r6499, 20; + mov.b64 %rd9270, {%r6501, %r6500}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6502}, %rd9250; + } + { + .reg .b32 %dummy; + mov.b64 {%r6503,%dummy}, %rd9250; + } + shf.l.wrap.b32 %r6504, %r6503, %r6502, 3; + shf.l.wrap.b32 %r6505, %r6502, %r6503, 3; + mov.b64 %rd9271, {%r6505, %r6504}; + { + .reg .b32 %dummy; + mov.b64 {%r6506,%dummy}, %rd9256; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6507}, %rd9256; + } + shf.r.wrap.b32 %r6508, %r6507, %r6506, 19; + shf.r.wrap.b32 %r6509, %r6506, %r6507, 19; + mov.b64 %rd9272, {%r6509, %r6508}; + { + .reg .b32 %dummy; + mov.b64 {%r6510,%dummy}, %rd9262; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6511}, %rd9262; + } + shf.r.wrap.b32 %r6512, %r6511, %r6510, 3; + shf.r.wrap.b32 %r6513, %r6510, %r6511, 3; + mov.b64 %rd9273, {%r6513, %r6512}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6514}, %rd9241; + } + { + .reg .b32 %dummy; + mov.b64 {%r6515,%dummy}, %rd9241; + } + shf.l.wrap.b32 %r6516, %r6515, %r6514, 1; + shf.l.wrap.b32 %r6517, %r6514, %r6515, 1; + mov.b64 %rd9274, {%r6517, %r6516}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6518}, %rd9247; + } + { + .reg .b32 %dummy; + mov.b64 {%r6519,%dummy}, %rd9247; + } + shf.l.wrap.b32 %r6520, %r6519, %r6518, 6; + shf.l.wrap.b32 %r6521, %r6518, %r6519, 6; + mov.b64 %rd9275, {%r6521, %r6520}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6522}, %rd9253; + } + { + .reg .b32 %dummy; + mov.b64 {%r6523,%dummy}, %rd9253; + } + shf.l.wrap.b32 %r6524, %r6523, %r6522, 25; + shf.l.wrap.b32 %r6525, %r6522, %r6523, 25; + mov.b64 %rd9276, {%r6525, %r6524}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6526}, %rd9259; + } + { + .reg .b32 %dummy; + mov.b64 {%r6527,%dummy}, %rd9259; + } + shf.l.wrap.b32 %r6528, %r6527, %r6526, 8; + shf.l.wrap.b32 %r6529, %r6526, %r6527, 8; + mov.b64 %rd9277, {%r6529, %r6528}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6530}, %rd9260; + } + { + .reg .b32 %dummy; + mov.b64 {%r6531,%dummy}, %rd9260; + } + shf.l.wrap.b32 %r6532, %r6531, %r6530, 18; + shf.l.wrap.b32 %r6533, %r6530, %r6531, 18; + mov.b64 %rd9278, {%r6533, %r6532}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6534}, %rd9244; + } + { + .reg .b32 %dummy; + mov.b64 {%r6535,%dummy}, %rd9244; + } + shf.l.wrap.b32 %r6536, %r6535, %r6534, 27; + shf.l.wrap.b32 %r6537, %r6534, %r6535, 27; + mov.b64 %rd9279, {%r6537, %r6536}; + { + .reg .b32 %dummy; + mov.b64 {%r6538,%dummy}, %rd9245; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6539}, %rd9245; + } + shf.r.wrap.b32 %r6540, %r6539, %r6538, 28; + shf.r.wrap.b32 %r6541, %r6538, %r6539, 28; + mov.b64 %rd9280, {%r6541, %r6540}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6542}, %rd9251; + } + { + .reg .b32 %dummy; + mov.b64 {%r6543,%dummy}, %rd9251; + } + shf.l.wrap.b32 %r6544, %r6543, %r6542, 10; + shf.l.wrap.b32 %r6545, %r6542, %r6543, 10; + mov.b64 %rd9281, {%r6545, %r6544}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6546}, %rd9257; + } + { + .reg .b32 %dummy; + mov.b64 {%r6547,%dummy}, %rd9257; + } + shf.l.wrap.b32 %r6548, %r6547, %r6546, 15; + shf.l.wrap.b32 %r6549, %r6546, %r6547, 15; + mov.b64 %rd9282, {%r6549, %r6548}; + { + .reg .b32 %dummy; + mov.b64 {%r6550,%dummy}, %rd9263; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6551}, %rd9263; + } + shf.r.wrap.b32 %r6552, %r6551, %r6550, 8; + shf.r.wrap.b32 %r6553, %r6550, %r6551, 8; + mov.b64 %rd9283, {%r6553, %r6552}; + { + .reg .b32 %dummy; + mov.b64 {%r6554,%dummy}, %rd9242; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6555}, %rd9242; + } + shf.r.wrap.b32 %r6556, %r6555, %r6554, 2; + shf.r.wrap.b32 %r6557, %r6554, %r6555, 2; + mov.b64 %rd9284, {%r6557, %r6556}; + { + .reg .b32 %dummy; + mov.b64 {%r6558,%dummy}, %rd9248; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6559}, %rd9248; + } + shf.r.wrap.b32 %r6560, %r6559, %r6558, 9; + shf.r.wrap.b32 %r6561, %r6558, %r6559, 9; + mov.b64 %rd9285, {%r6561, %r6560}; + { + .reg .b32 %dummy; + mov.b64 {%r6562,%dummy}, %rd9254; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6563}, %rd9254; + } + shf.r.wrap.b32 %r6564, %r6563, %r6562, 25; + shf.r.wrap.b32 %r6565, %r6562, %r6563, 25; + mov.b64 %rd9286, {%r6565, %r6564}; + { + .reg .b32 %dummy; + mov.b64 {%r6566,%dummy}, %rd9255; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6567}, %rd9255; + } + shf.r.wrap.b32 %r6568, %r6567, %r6566, 23; + shf.r.wrap.b32 %r6569, %r6566, %r6567, 23; + mov.b64 %rd9287, {%r6569, %r6568}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6570}, %rd9261; + } + { + .reg .b32 %dummy; + mov.b64 {%r6571,%dummy}, %rd9261; + } + shf.l.wrap.b32 %r6572, %r6571, %r6570, 2; + shf.l.wrap.b32 %r6573, %r6570, %r6571, 2; + mov.b64 %rd9288, {%r6573, %r6572}; + not.b64 %rd9289, %rd9265; + and.b64 %rd9290, %rd9266, %rd9289; + xor.b64 %rd9291, %rd9290, %rd9240; + not.b64 %rd9292, %rd9266; + and.b64 %rd9293, %rd9267, %rd9292; + xor.b64 %rd9294, %rd9293, %rd9265; + not.b64 %rd9295, %rd9267; + and.b64 %rd9296, %rd9268, %rd9295; + xor.b64 %rd9297, %rd9296, %rd9266; + not.b64 %rd9298, %rd9268; + and.b64 %rd9299, %rd9240, %rd9298; + xor.b64 %rd9300, %rd9299, %rd9267; + not.b64 %rd9301, %rd9240; + and.b64 %rd9302, %rd9265, %rd9301; + xor.b64 %rd9303, %rd9302, %rd9268; + not.b64 %rd9304, %rd9270; + and.b64 %rd9305, %rd9271, %rd9304; + xor.b64 %rd9306, %rd9305, %rd9269; + not.b64 %rd9307, %rd9271; + and.b64 %rd9308, %rd9272, %rd9307; + xor.b64 %rd9309, %rd9308, %rd9270; + not.b64 %rd9310, %rd9272; + and.b64 %rd9311, %rd9273, %rd9310; + xor.b64 %rd9312, %rd9311, %rd9271; + not.b64 %rd9313, %rd9273; + and.b64 %rd9314, %rd9269, %rd9313; + xor.b64 %rd9315, %rd9314, %rd9272; + not.b64 %rd9316, %rd9269; + and.b64 %rd9317, %rd9270, %rd9316; + xor.b64 %rd9318, %rd9317, %rd9273; + not.b64 %rd9319, %rd9275; + and.b64 %rd9320, %rd9276, %rd9319; + xor.b64 %rd9321, %rd9320, %rd9274; + not.b64 %rd9322, %rd9276; + and.b64 %rd9323, %rd9277, %rd9322; + xor.b64 %rd9324, %rd9323, %rd9275; + not.b64 %rd9325, %rd9277; + and.b64 %rd9326, %rd9278, %rd9325; + xor.b64 %rd9327, %rd9326, %rd9276; + not.b64 %rd9328, %rd9278; + and.b64 %rd9329, %rd9274, %rd9328; + xor.b64 %rd9330, %rd9329, %rd9277; + not.b64 %rd9331, %rd9274; + and.b64 %rd9332, %rd9275, %rd9331; + xor.b64 %rd9333, %rd9332, %rd9278; + not.b64 %rd9334, %rd9280; + and.b64 %rd9335, %rd9281, %rd9334; + xor.b64 %rd9336, %rd9335, %rd9279; + not.b64 %rd9337, %rd9281; + and.b64 %rd9338, %rd9282, %rd9337; + xor.b64 %rd9339, %rd9338, %rd9280; + not.b64 %rd9340, %rd9282; + and.b64 %rd9341, %rd9283, %rd9340; + xor.b64 %rd9342, %rd9341, %rd9281; + not.b64 %rd9343, %rd9283; + and.b64 %rd9344, %rd9279, %rd9343; + xor.b64 %rd9345, %rd9344, %rd9282; + not.b64 %rd9346, %rd9279; + and.b64 %rd9347, %rd9280, %rd9346; + xor.b64 %rd9348, %rd9347, %rd9283; + not.b64 %rd9349, %rd9285; + and.b64 %rd9350, %rd9286, %rd9349; + xor.b64 %rd9351, %rd9350, %rd9284; + not.b64 %rd9352, %rd9286; + and.b64 %rd9353, %rd9287, %rd9352; + xor.b64 %rd9354, %rd9353, %rd9285; + not.b64 %rd9355, %rd9287; + and.b64 %rd9356, %rd9288, %rd9355; + xor.b64 %rd9357, %rd9356, %rd9286; + not.b64 %rd9358, %rd9288; + and.b64 %rd9359, %rd9284, %rd9358; + xor.b64 %rd9360, %rd9359, %rd9287; + not.b64 %rd9361, %rd9284; + and.b64 %rd9362, %rd9285, %rd9361; + xor.b64 %rd9363, %rd9362, %rd9288; + xor.b64 %rd9365, %rd9291, %rd23; + xor.b64 %rd9366, %rd9306, %rd9365; + xor.b64 %rd9367, %rd9366, %rd9321; + xor.b64 %rd9368, %rd9367, %rd9336; + xor.b64 %rd9369, %rd9368, %rd9351; + xor.b64 %rd9370, %rd9309, %rd9294; + xor.b64 %rd9371, %rd9370, %rd9324; + xor.b64 %rd9372, %rd9371, %rd9339; + xor.b64 %rd9373, %rd9372, %rd9354; + xor.b64 %rd9374, %rd9312, %rd9297; + xor.b64 %rd9375, %rd9374, %rd9327; + xor.b64 %rd9376, %rd9375, %rd9342; + xor.b64 %rd9377, %rd9376, %rd9357; + xor.b64 %rd9378, %rd9315, %rd9300; + xor.b64 %rd9379, %rd9378, %rd9330; + xor.b64 %rd9380, %rd9379, %rd9345; + xor.b64 %rd9381, %rd9380, %rd9360; + xor.b64 %rd9382, %rd9318, %rd9303; + xor.b64 %rd9383, %rd9382, %rd9333; + xor.b64 %rd9384, %rd9383, %rd9348; + xor.b64 %rd9385, %rd9384, %rd9363; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6574}, %rd9373; + } + { + .reg .b32 %dummy; + mov.b64 {%r6575,%dummy}, %rd9373; + } + shf.l.wrap.b32 %r6576, %r6575, %r6574, 1; + shf.l.wrap.b32 %r6577, %r6574, %r6575, 1; + mov.b64 %rd9386, {%r6577, %r6576}; + xor.b64 %rd9387, %rd9385, %rd9386; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6578}, %rd9377; + } + { + .reg .b32 %dummy; + mov.b64 {%r6579,%dummy}, %rd9377; + } + shf.l.wrap.b32 %r6580, %r6579, %r6578, 1; + shf.l.wrap.b32 %r6581, %r6578, %r6579, 1; + mov.b64 %rd9388, {%r6581, %r6580}; + xor.b64 %rd9389, %rd9388, %rd9369; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6582}, %rd9381; + } + { + .reg .b32 %dummy; + mov.b64 {%r6583,%dummy}, %rd9381; + } + shf.l.wrap.b32 %r6584, %r6583, %r6582, 1; + shf.l.wrap.b32 %r6585, %r6582, %r6583, 1; + mov.b64 %rd9390, {%r6585, %r6584}; + xor.b64 %rd9391, %rd9390, %rd9373; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6586}, %rd9385; + } + { + .reg .b32 %dummy; + mov.b64 {%r6587,%dummy}, %rd9385; + } + shf.l.wrap.b32 %r6588, %r6587, %r6586, 1; + shf.l.wrap.b32 %r6589, %r6586, %r6587, 1; + mov.b64 %rd9392, {%r6589, %r6588}; + xor.b64 %rd9393, %rd9392, %rd9377; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6590}, %rd9369; + } + { + .reg .b32 %dummy; + mov.b64 {%r6591,%dummy}, %rd9369; + } + shf.l.wrap.b32 %r6592, %r6591, %r6590, 1; + shf.l.wrap.b32 %r6593, %r6590, %r6591, 1; + mov.b64 %rd9394, {%r6593, %r6592}; + xor.b64 %rd9395, %rd9381, %rd9394; + xor.b64 %rd9396, %rd9387, %rd9365; + xor.b64 %rd9397, %rd9389, %rd9294; + xor.b64 %rd9398, %rd9391, %rd9297; + xor.b64 %rd9399, %rd9393, %rd9300; + xor.b64 %rd9400, %rd9395, %rd9303; + xor.b64 %rd9401, %rd9387, %rd9306; + xor.b64 %rd9402, %rd9389, %rd9309; + xor.b64 %rd9403, %rd9391, %rd9312; + xor.b64 %rd9404, %rd9393, %rd9315; + xor.b64 %rd9405, %rd9395, %rd9318; + xor.b64 %rd9406, %rd9387, %rd9321; + xor.b64 %rd9407, %rd9389, %rd9324; + xor.b64 %rd9408, %rd9391, %rd9327; + xor.b64 %rd9409, %rd9393, %rd9330; + xor.b64 %rd9410, %rd9395, %rd9333; + xor.b64 %rd9411, %rd9387, %rd9336; + xor.b64 %rd9412, %rd9389, %rd9339; + xor.b64 %rd9413, %rd9391, %rd9342; + xor.b64 %rd9414, %rd9393, %rd9345; + xor.b64 %rd9415, %rd9395, %rd9348; + xor.b64 %rd9416, %rd9387, %rd9351; + xor.b64 %rd9417, %rd9389, %rd9354; + xor.b64 %rd9418, %rd9391, %rd9357; + xor.b64 %rd9419, %rd9393, %rd9360; + xor.b64 %rd9420, %rd9395, %rd9363; + { + .reg .b32 %dummy; + mov.b64 {%r6594,%dummy}, %rd9402; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6595}, %rd9402; + } + shf.r.wrap.b32 %r6596, %r6595, %r6594, 20; + shf.r.wrap.b32 %r6597, %r6594, %r6595, 20; + mov.b64 %rd9421, {%r6597, %r6596}; + { + .reg .b32 %dummy; + mov.b64 {%r6598,%dummy}, %rd9408; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6599}, %rd9408; + } + shf.r.wrap.b32 %r6600, %r6599, %r6598, 21; + shf.r.wrap.b32 %r6601, %r6598, %r6599, 21; + mov.b64 %rd9422, {%r6601, %r6600}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6602}, %rd9414; + } + { + .reg .b32 %dummy; + mov.b64 {%r6603,%dummy}, %rd9414; + } + shf.l.wrap.b32 %r6604, %r6603, %r6602, 21; + shf.l.wrap.b32 %r6605, %r6602, %r6603, 21; + mov.b64 %rd9423, {%r6605, %r6604}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6606}, %rd9420; + } + { + .reg .b32 %dummy; + mov.b64 {%r6607,%dummy}, %rd9420; + } + shf.l.wrap.b32 %r6608, %r6607, %r6606, 14; + shf.l.wrap.b32 %r6609, %r6606, %r6607, 14; + mov.b64 %rd9424, {%r6609, %r6608}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6610}, %rd9399; + } + { + .reg .b32 %dummy; + mov.b64 {%r6611,%dummy}, %rd9399; + } + shf.l.wrap.b32 %r6612, %r6611, %r6610, 28; + shf.l.wrap.b32 %r6613, %r6610, %r6611, 28; + mov.b64 %rd9425, {%r6613, %r6612}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6614}, %rd9405; + } + { + .reg .b32 %dummy; + mov.b64 {%r6615,%dummy}, %rd9405; + } + shf.l.wrap.b32 %r6616, %r6615, %r6614, 20; + shf.l.wrap.b32 %r6617, %r6614, %r6615, 20; + mov.b64 %rd9426, {%r6617, %r6616}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6618}, %rd9406; + } + { + .reg .b32 %dummy; + mov.b64 {%r6619,%dummy}, %rd9406; + } + shf.l.wrap.b32 %r6620, %r6619, %r6618, 3; + shf.l.wrap.b32 %r6621, %r6618, %r6619, 3; + mov.b64 %rd9427, {%r6621, %r6620}; + { + .reg .b32 %dummy; + mov.b64 {%r6622,%dummy}, %rd9412; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6623}, %rd9412; + } + shf.r.wrap.b32 %r6624, %r6623, %r6622, 19; + shf.r.wrap.b32 %r6625, %r6622, %r6623, 19; + mov.b64 %rd9428, {%r6625, %r6624}; + { + .reg .b32 %dummy; + mov.b64 {%r6626,%dummy}, %rd9418; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6627}, %rd9418; + } + shf.r.wrap.b32 %r6628, %r6627, %r6626, 3; + shf.r.wrap.b32 %r6629, %r6626, %r6627, 3; + mov.b64 %rd9429, {%r6629, %r6628}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6630}, %rd9397; + } + { + .reg .b32 %dummy; + mov.b64 {%r6631,%dummy}, %rd9397; + } + shf.l.wrap.b32 %r6632, %r6631, %r6630, 1; + shf.l.wrap.b32 %r6633, %r6630, %r6631, 1; + mov.b64 %rd9430, {%r6633, %r6632}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6634}, %rd9403; + } + { + .reg .b32 %dummy; + mov.b64 {%r6635,%dummy}, %rd9403; + } + shf.l.wrap.b32 %r6636, %r6635, %r6634, 6; + shf.l.wrap.b32 %r6637, %r6634, %r6635, 6; + mov.b64 %rd9431, {%r6637, %r6636}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6638}, %rd9409; + } + { + .reg .b32 %dummy; + mov.b64 {%r6639,%dummy}, %rd9409; + } + shf.l.wrap.b32 %r6640, %r6639, %r6638, 25; + shf.l.wrap.b32 %r6641, %r6638, %r6639, 25; + mov.b64 %rd9432, {%r6641, %r6640}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6642}, %rd9415; + } + { + .reg .b32 %dummy; + mov.b64 {%r6643,%dummy}, %rd9415; + } + shf.l.wrap.b32 %r6644, %r6643, %r6642, 8; + shf.l.wrap.b32 %r6645, %r6642, %r6643, 8; + mov.b64 %rd9433, {%r6645, %r6644}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6646}, %rd9416; + } + { + .reg .b32 %dummy; + mov.b64 {%r6647,%dummy}, %rd9416; + } + shf.l.wrap.b32 %r6648, %r6647, %r6646, 18; + shf.l.wrap.b32 %r6649, %r6646, %r6647, 18; + mov.b64 %rd9434, {%r6649, %r6648}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6650}, %rd9400; + } + { + .reg .b32 %dummy; + mov.b64 {%r6651,%dummy}, %rd9400; + } + shf.l.wrap.b32 %r6652, %r6651, %r6650, 27; + shf.l.wrap.b32 %r6653, %r6650, %r6651, 27; + mov.b64 %rd9435, {%r6653, %r6652}; + { + .reg .b32 %dummy; + mov.b64 {%r6654,%dummy}, %rd9401; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6655}, %rd9401; + } + shf.r.wrap.b32 %r6656, %r6655, %r6654, 28; + shf.r.wrap.b32 %r6657, %r6654, %r6655, 28; + mov.b64 %rd9436, {%r6657, %r6656}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6658}, %rd9407; + } + { + .reg .b32 %dummy; + mov.b64 {%r6659,%dummy}, %rd9407; + } + shf.l.wrap.b32 %r6660, %r6659, %r6658, 10; + shf.l.wrap.b32 %r6661, %r6658, %r6659, 10; + mov.b64 %rd9437, {%r6661, %r6660}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6662}, %rd9413; + } + { + .reg .b32 %dummy; + mov.b64 {%r6663,%dummy}, %rd9413; + } + shf.l.wrap.b32 %r6664, %r6663, %r6662, 15; + shf.l.wrap.b32 %r6665, %r6662, %r6663, 15; + mov.b64 %rd9438, {%r6665, %r6664}; + { + .reg .b32 %dummy; + mov.b64 {%r6666,%dummy}, %rd9419; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6667}, %rd9419; + } + shf.r.wrap.b32 %r6668, %r6667, %r6666, 8; + shf.r.wrap.b32 %r6669, %r6666, %r6667, 8; + mov.b64 %rd9439, {%r6669, %r6668}; + { + .reg .b32 %dummy; + mov.b64 {%r6670,%dummy}, %rd9398; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6671}, %rd9398; + } + shf.r.wrap.b32 %r6672, %r6671, %r6670, 2; + shf.r.wrap.b32 %r6673, %r6670, %r6671, 2; + mov.b64 %rd9440, {%r6673, %r6672}; + { + .reg .b32 %dummy; + mov.b64 {%r6674,%dummy}, %rd9404; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6675}, %rd9404; + } + shf.r.wrap.b32 %r6676, %r6675, %r6674, 9; + shf.r.wrap.b32 %r6677, %r6674, %r6675, 9; + mov.b64 %rd9441, {%r6677, %r6676}; + { + .reg .b32 %dummy; + mov.b64 {%r6678,%dummy}, %rd9410; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6679}, %rd9410; + } + shf.r.wrap.b32 %r6680, %r6679, %r6678, 25; + shf.r.wrap.b32 %r6681, %r6678, %r6679, 25; + mov.b64 %rd9442, {%r6681, %r6680}; + { + .reg .b32 %dummy; + mov.b64 {%r6682,%dummy}, %rd9411; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6683}, %rd9411; + } + shf.r.wrap.b32 %r6684, %r6683, %r6682, 23; + shf.r.wrap.b32 %r6685, %r6682, %r6683, 23; + mov.b64 %rd9443, {%r6685, %r6684}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6686}, %rd9417; + } + { + .reg .b32 %dummy; + mov.b64 {%r6687,%dummy}, %rd9417; + } + shf.l.wrap.b32 %r6688, %r6687, %r6686, 2; + shf.l.wrap.b32 %r6689, %r6686, %r6687, 2; + mov.b64 %rd9444, {%r6689, %r6688}; + not.b64 %rd9445, %rd9421; + and.b64 %rd9446, %rd9422, %rd9445; + xor.b64 %rd9447, %rd9446, %rd9396; + not.b64 %rd9448, %rd9422; + and.b64 %rd9449, %rd9423, %rd9448; + xor.b64 %rd9450, %rd9449, %rd9421; + not.b64 %rd9451, %rd9423; + and.b64 %rd9452, %rd9424, %rd9451; + xor.b64 %rd9453, %rd9452, %rd9422; + not.b64 %rd9454, %rd9424; + and.b64 %rd9455, %rd9396, %rd9454; + xor.b64 %rd9456, %rd9455, %rd9423; + not.b64 %rd9457, %rd9396; + and.b64 %rd9458, %rd9421, %rd9457; + xor.b64 %rd9459, %rd9458, %rd9424; + not.b64 %rd9460, %rd9426; + and.b64 %rd9461, %rd9427, %rd9460; + xor.b64 %rd9462, %rd9461, %rd9425; + not.b64 %rd9463, %rd9427; + and.b64 %rd9464, %rd9428, %rd9463; + xor.b64 %rd9465, %rd9464, %rd9426; + not.b64 %rd9466, %rd9428; + and.b64 %rd9467, %rd9429, %rd9466; + xor.b64 %rd9468, %rd9467, %rd9427; + not.b64 %rd9469, %rd9429; + and.b64 %rd9470, %rd9425, %rd9469; + xor.b64 %rd9471, %rd9470, %rd9428; + not.b64 %rd9472, %rd9425; + and.b64 %rd9473, %rd9426, %rd9472; + xor.b64 %rd9474, %rd9473, %rd9429; + not.b64 %rd9475, %rd9431; + and.b64 %rd9476, %rd9432, %rd9475; + xor.b64 %rd9477, %rd9476, %rd9430; + not.b64 %rd9478, %rd9432; + and.b64 %rd9479, %rd9433, %rd9478; + xor.b64 %rd9480, %rd9479, %rd9431; + not.b64 %rd9481, %rd9433; + and.b64 %rd9482, %rd9434, %rd9481; + xor.b64 %rd9483, %rd9482, %rd9432; + not.b64 %rd9484, %rd9434; + and.b64 %rd9485, %rd9430, %rd9484; + xor.b64 %rd9486, %rd9485, %rd9433; + not.b64 %rd9487, %rd9430; + and.b64 %rd9488, %rd9431, %rd9487; + xor.b64 %rd9489, %rd9488, %rd9434; + not.b64 %rd9490, %rd9436; + and.b64 %rd9491, %rd9437, %rd9490; + xor.b64 %rd9492, %rd9491, %rd9435; + not.b64 %rd9493, %rd9437; + and.b64 %rd9494, %rd9438, %rd9493; + xor.b64 %rd9495, %rd9494, %rd9436; + not.b64 %rd9496, %rd9438; + and.b64 %rd9497, %rd9439, %rd9496; + xor.b64 %rd9498, %rd9497, %rd9437; + not.b64 %rd9499, %rd9439; + and.b64 %rd9500, %rd9435, %rd9499; + xor.b64 %rd9501, %rd9500, %rd9438; + not.b64 %rd9502, %rd9435; + and.b64 %rd9503, %rd9436, %rd9502; + xor.b64 %rd9504, %rd9503, %rd9439; + not.b64 %rd9505, %rd9441; + and.b64 %rd9506, %rd9442, %rd9505; + xor.b64 %rd9507, %rd9506, %rd9440; + not.b64 %rd9508, %rd9442; + and.b64 %rd9509, %rd9443, %rd9508; + xor.b64 %rd9510, %rd9509, %rd9441; + not.b64 %rd9511, %rd9443; + and.b64 %rd9512, %rd9444, %rd9511; + xor.b64 %rd9513, %rd9512, %rd9442; + not.b64 %rd9514, %rd9444; + and.b64 %rd9515, %rd9440, %rd9514; + xor.b64 %rd9516, %rd9515, %rd9443; + not.b64 %rd9517, %rd9440; + and.b64 %rd9518, %rd9441, %rd9517; + xor.b64 %rd9519, %rd9518, %rd9444; + xor.b64 %rd9521, %rd9447, %rd24; + xor.b64 %rd9522, %rd9462, %rd9521; + xor.b64 %rd9523, %rd9522, %rd9477; + xor.b64 %rd9524, %rd9523, %rd9492; + xor.b64 %rd9525, %rd9524, %rd9507; + xor.b64 %rd9526, %rd9465, %rd9450; + xor.b64 %rd9527, %rd9526, %rd9480; + xor.b64 %rd9528, %rd9527, %rd9495; + xor.b64 %rd9529, %rd9528, %rd9510; + xor.b64 %rd9530, %rd9468, %rd9453; + xor.b64 %rd9531, %rd9530, %rd9483; + xor.b64 %rd9532, %rd9531, %rd9498; + xor.b64 %rd9533, %rd9532, %rd9513; + xor.b64 %rd9534, %rd9471, %rd9456; + xor.b64 %rd9535, %rd9534, %rd9486; + xor.b64 %rd9536, %rd9535, %rd9501; + xor.b64 %rd9537, %rd9536, %rd9516; + xor.b64 %rd9538, %rd9474, %rd9459; + xor.b64 %rd9539, %rd9538, %rd9489; + xor.b64 %rd9540, %rd9539, %rd9504; + xor.b64 %rd9541, %rd9540, %rd9519; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6690}, %rd9529; + } + { + .reg .b32 %dummy; + mov.b64 {%r6691,%dummy}, %rd9529; + } + shf.l.wrap.b32 %r6692, %r6691, %r6690, 1; + shf.l.wrap.b32 %r6693, %r6690, %r6691, 1; + mov.b64 %rd9542, {%r6693, %r6692}; + xor.b64 %rd9543, %rd9541, %rd9542; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6694}, %rd9533; + } + { + .reg .b32 %dummy; + mov.b64 {%r6695,%dummy}, %rd9533; + } + shf.l.wrap.b32 %r6696, %r6695, %r6694, 1; + shf.l.wrap.b32 %r6697, %r6694, %r6695, 1; + mov.b64 %rd9544, {%r6697, %r6696}; + xor.b64 %rd9545, %rd9544, %rd9525; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6698}, %rd9537; + } + { + .reg .b32 %dummy; + mov.b64 {%r6699,%dummy}, %rd9537; + } + shf.l.wrap.b32 %r6700, %r6699, %r6698, 1; + shf.l.wrap.b32 %r6701, %r6698, %r6699, 1; + mov.b64 %rd9546, {%r6701, %r6700}; + xor.b64 %rd9547, %rd9546, %rd9529; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6702}, %rd9541; + } + { + .reg .b32 %dummy; + mov.b64 {%r6703,%dummy}, %rd9541; + } + shf.l.wrap.b32 %r6704, %r6703, %r6702, 1; + shf.l.wrap.b32 %r6705, %r6702, %r6703, 1; + mov.b64 %rd9548, {%r6705, %r6704}; + xor.b64 %rd9549, %rd9548, %rd9533; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6706}, %rd9525; + } + { + .reg .b32 %dummy; + mov.b64 {%r6707,%dummy}, %rd9525; + } + shf.l.wrap.b32 %r6708, %r6707, %r6706, 1; + shf.l.wrap.b32 %r6709, %r6706, %r6707, 1; + mov.b64 %rd9550, {%r6709, %r6708}; + xor.b64 %rd9551, %rd9537, %rd9550; + xor.b64 %rd9552, %rd9543, %rd9521; + xor.b64 %rd9553, %rd9545, %rd9450; + xor.b64 %rd9554, %rd9547, %rd9453; + xor.b64 %rd9555, %rd9549, %rd9456; + xor.b64 %rd9556, %rd9551, %rd9459; + xor.b64 %rd9557, %rd9543, %rd9462; + xor.b64 %rd9558, %rd9545, %rd9465; + xor.b64 %rd9559, %rd9547, %rd9468; + xor.b64 %rd9560, %rd9549, %rd9471; + xor.b64 %rd9561, %rd9551, %rd9474; + xor.b64 %rd9562, %rd9543, %rd9477; + xor.b64 %rd9563, %rd9545, %rd9480; + xor.b64 %rd9564, %rd9547, %rd9483; + xor.b64 %rd9565, %rd9549, %rd9486; + xor.b64 %rd9566, %rd9551, %rd9489; + xor.b64 %rd9567, %rd9543, %rd9492; + xor.b64 %rd9568, %rd9545, %rd9495; + xor.b64 %rd9569, %rd9547, %rd9498; + xor.b64 %rd9570, %rd9549, %rd9501; + xor.b64 %rd9571, %rd9551, %rd9504; + xor.b64 %rd9572, %rd9543, %rd9507; + xor.b64 %rd9573, %rd9545, %rd9510; + xor.b64 %rd9574, %rd9547, %rd9513; + xor.b64 %rd9575, %rd9549, %rd9516; + xor.b64 %rd9576, %rd9551, %rd9519; + { + .reg .b32 %dummy; + mov.b64 {%r6710,%dummy}, %rd9558; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6711}, %rd9558; + } + shf.r.wrap.b32 %r6712, %r6711, %r6710, 20; + shf.r.wrap.b32 %r6713, %r6710, %r6711, 20; + mov.b64 %rd9577, {%r6713, %r6712}; + { + .reg .b32 %dummy; + mov.b64 {%r6714,%dummy}, %rd9564; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6715}, %rd9564; + } + shf.r.wrap.b32 %r6716, %r6715, %r6714, 21; + shf.r.wrap.b32 %r6717, %r6714, %r6715, 21; + mov.b64 %rd9578, {%r6717, %r6716}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6718}, %rd9570; + } + { + .reg .b32 %dummy; + mov.b64 {%r6719,%dummy}, %rd9570; + } + shf.l.wrap.b32 %r6720, %r6719, %r6718, 21; + shf.l.wrap.b32 %r6721, %r6718, %r6719, 21; + mov.b64 %rd9579, {%r6721, %r6720}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6722}, %rd9576; + } + { + .reg .b32 %dummy; + mov.b64 {%r6723,%dummy}, %rd9576; + } + shf.l.wrap.b32 %r6724, %r6723, %r6722, 14; + shf.l.wrap.b32 %r6725, %r6722, %r6723, 14; + mov.b64 %rd9580, {%r6725, %r6724}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6726}, %rd9555; + } + { + .reg .b32 %dummy; + mov.b64 {%r6727,%dummy}, %rd9555; + } + shf.l.wrap.b32 %r6728, %r6727, %r6726, 28; + shf.l.wrap.b32 %r6729, %r6726, %r6727, 28; + mov.b64 %rd9581, {%r6729, %r6728}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6730}, %rd9561; + } + { + .reg .b32 %dummy; + mov.b64 {%r6731,%dummy}, %rd9561; + } + shf.l.wrap.b32 %r6732, %r6731, %r6730, 20; + shf.l.wrap.b32 %r6733, %r6730, %r6731, 20; + mov.b64 %rd9582, {%r6733, %r6732}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6734}, %rd9562; + } + { + .reg .b32 %dummy; + mov.b64 {%r6735,%dummy}, %rd9562; + } + shf.l.wrap.b32 %r6736, %r6735, %r6734, 3; + shf.l.wrap.b32 %r6737, %r6734, %r6735, 3; + mov.b64 %rd9583, {%r6737, %r6736}; + { + .reg .b32 %dummy; + mov.b64 {%r6738,%dummy}, %rd9568; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6739}, %rd9568; + } + shf.r.wrap.b32 %r6740, %r6739, %r6738, 19; + shf.r.wrap.b32 %r6741, %r6738, %r6739, 19; + mov.b64 %rd9584, {%r6741, %r6740}; + { + .reg .b32 %dummy; + mov.b64 {%r6742,%dummy}, %rd9574; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6743}, %rd9574; + } + shf.r.wrap.b32 %r6744, %r6743, %r6742, 3; + shf.r.wrap.b32 %r6745, %r6742, %r6743, 3; + mov.b64 %rd9585, {%r6745, %r6744}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6746}, %rd9553; + } + { + .reg .b32 %dummy; + mov.b64 {%r6747,%dummy}, %rd9553; + } + shf.l.wrap.b32 %r6748, %r6747, %r6746, 1; + shf.l.wrap.b32 %r6749, %r6746, %r6747, 1; + mov.b64 %rd9586, {%r6749, %r6748}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6750}, %rd9559; + } + { + .reg .b32 %dummy; + mov.b64 {%r6751,%dummy}, %rd9559; + } + shf.l.wrap.b32 %r6752, %r6751, %r6750, 6; + shf.l.wrap.b32 %r6753, %r6750, %r6751, 6; + mov.b64 %rd9587, {%r6753, %r6752}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6754}, %rd9565; + } + { + .reg .b32 %dummy; + mov.b64 {%r6755,%dummy}, %rd9565; + } + shf.l.wrap.b32 %r6756, %r6755, %r6754, 25; + shf.l.wrap.b32 %r6757, %r6754, %r6755, 25; + mov.b64 %rd9588, {%r6757, %r6756}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6758}, %rd9571; + } + { + .reg .b32 %dummy; + mov.b64 {%r6759,%dummy}, %rd9571; + } + shf.l.wrap.b32 %r6760, %r6759, %r6758, 8; + shf.l.wrap.b32 %r6761, %r6758, %r6759, 8; + mov.b64 %rd9589, {%r6761, %r6760}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6762}, %rd9572; + } + { + .reg .b32 %dummy; + mov.b64 {%r6763,%dummy}, %rd9572; + } + shf.l.wrap.b32 %r6764, %r6763, %r6762, 18; + shf.l.wrap.b32 %r6765, %r6762, %r6763, 18; + mov.b64 %rd9590, {%r6765, %r6764}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6766}, %rd9556; + } + { + .reg .b32 %dummy; + mov.b64 {%r6767,%dummy}, %rd9556; + } + shf.l.wrap.b32 %r6768, %r6767, %r6766, 27; + shf.l.wrap.b32 %r6769, %r6766, %r6767, 27; + mov.b64 %rd9591, {%r6769, %r6768}; + { + .reg .b32 %dummy; + mov.b64 {%r6770,%dummy}, %rd9557; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6771}, %rd9557; + } + shf.r.wrap.b32 %r6772, %r6771, %r6770, 28; + shf.r.wrap.b32 %r6773, %r6770, %r6771, 28; + mov.b64 %rd9592, {%r6773, %r6772}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6774}, %rd9563; + } + { + .reg .b32 %dummy; + mov.b64 {%r6775,%dummy}, %rd9563; + } + shf.l.wrap.b32 %r6776, %r6775, %r6774, 10; + shf.l.wrap.b32 %r6777, %r6774, %r6775, 10; + mov.b64 %rd9593, {%r6777, %r6776}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6778}, %rd9569; + } + { + .reg .b32 %dummy; + mov.b64 {%r6779,%dummy}, %rd9569; + } + shf.l.wrap.b32 %r6780, %r6779, %r6778, 15; + shf.l.wrap.b32 %r6781, %r6778, %r6779, 15; + mov.b64 %rd9594, {%r6781, %r6780}; + { + .reg .b32 %dummy; + mov.b64 {%r6782,%dummy}, %rd9575; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6783}, %rd9575; + } + shf.r.wrap.b32 %r6784, %r6783, %r6782, 8; + shf.r.wrap.b32 %r6785, %r6782, %r6783, 8; + mov.b64 %rd9595, {%r6785, %r6784}; + { + .reg .b32 %dummy; + mov.b64 {%r6786,%dummy}, %rd9554; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6787}, %rd9554; + } + shf.r.wrap.b32 %r6788, %r6787, %r6786, 2; + shf.r.wrap.b32 %r6789, %r6786, %r6787, 2; + mov.b64 %rd9596, {%r6789, %r6788}; + { + .reg .b32 %dummy; + mov.b64 {%r6790,%dummy}, %rd9560; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6791}, %rd9560; + } + shf.r.wrap.b32 %r6792, %r6791, %r6790, 9; + shf.r.wrap.b32 %r6793, %r6790, %r6791, 9; + mov.b64 %rd9597, {%r6793, %r6792}; + { + .reg .b32 %dummy; + mov.b64 {%r6794,%dummy}, %rd9566; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6795}, %rd9566; + } + shf.r.wrap.b32 %r6796, %r6795, %r6794, 25; + shf.r.wrap.b32 %r6797, %r6794, %r6795, 25; + mov.b64 %rd9598, {%r6797, %r6796}; + { + .reg .b32 %dummy; + mov.b64 {%r6798,%dummy}, %rd9567; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6799}, %rd9567; + } + shf.r.wrap.b32 %r6800, %r6799, %r6798, 23; + shf.r.wrap.b32 %r6801, %r6798, %r6799, 23; + mov.b64 %rd9599, {%r6801, %r6800}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6802}, %rd9573; + } + { + .reg .b32 %dummy; + mov.b64 {%r6803,%dummy}, %rd9573; + } + shf.l.wrap.b32 %r6804, %r6803, %r6802, 2; + shf.l.wrap.b32 %r6805, %r6802, %r6803, 2; + mov.b64 %rd9600, {%r6805, %r6804}; + not.b64 %rd9601, %rd9577; + and.b64 %rd9602, %rd9578, %rd9601; + xor.b64 %rd9603, %rd9602, %rd9552; + not.b64 %rd9604, %rd9578; + and.b64 %rd9605, %rd9579, %rd9604; + xor.b64 %rd9606, %rd9605, %rd9577; + not.b64 %rd9607, %rd9579; + and.b64 %rd9608, %rd9580, %rd9607; + xor.b64 %rd9609, %rd9608, %rd9578; + not.b64 %rd9610, %rd9580; + and.b64 %rd9611, %rd9552, %rd9610; + xor.b64 %rd9612, %rd9611, %rd9579; + not.b64 %rd9613, %rd9552; + and.b64 %rd9614, %rd9577, %rd9613; + xor.b64 %rd9615, %rd9614, %rd9580; + not.b64 %rd9616, %rd9582; + and.b64 %rd9617, %rd9583, %rd9616; + xor.b64 %rd9618, %rd9617, %rd9581; + not.b64 %rd9619, %rd9583; + and.b64 %rd9620, %rd9584, %rd9619; + xor.b64 %rd9621, %rd9620, %rd9582; + not.b64 %rd9622, %rd9584; + and.b64 %rd9623, %rd9585, %rd9622; + xor.b64 %rd9624, %rd9623, %rd9583; + not.b64 %rd9625, %rd9585; + and.b64 %rd9626, %rd9581, %rd9625; + xor.b64 %rd9627, %rd9626, %rd9584; + not.b64 %rd9628, %rd9581; + and.b64 %rd9629, %rd9582, %rd9628; + xor.b64 %rd9630, %rd9629, %rd9585; + not.b64 %rd9631, %rd9587; + and.b64 %rd9632, %rd9588, %rd9631; + xor.b64 %rd9633, %rd9632, %rd9586; + not.b64 %rd9634, %rd9588; + and.b64 %rd9635, %rd9589, %rd9634; + xor.b64 %rd9636, %rd9635, %rd9587; + not.b64 %rd9637, %rd9589; + and.b64 %rd9638, %rd9590, %rd9637; + xor.b64 %rd9639, %rd9638, %rd9588; + not.b64 %rd9640, %rd9590; + and.b64 %rd9641, %rd9586, %rd9640; + xor.b64 %rd9642, %rd9641, %rd9589; + not.b64 %rd9643, %rd9586; + and.b64 %rd9644, %rd9587, %rd9643; + xor.b64 %rd9645, %rd9644, %rd9590; + not.b64 %rd9646, %rd9592; + and.b64 %rd9647, %rd9593, %rd9646; + xor.b64 %rd9648, %rd9647, %rd9591; + not.b64 %rd9649, %rd9593; + and.b64 %rd9650, %rd9594, %rd9649; + xor.b64 %rd9651, %rd9650, %rd9592; + not.b64 %rd9652, %rd9594; + and.b64 %rd9653, %rd9595, %rd9652; + xor.b64 %rd9654, %rd9653, %rd9593; + not.b64 %rd9655, %rd9595; + and.b64 %rd9656, %rd9591, %rd9655; + xor.b64 %rd9657, %rd9656, %rd9594; + not.b64 %rd9658, %rd9591; + and.b64 %rd9659, %rd9592, %rd9658; + xor.b64 %rd9660, %rd9659, %rd9595; + not.b64 %rd9661, %rd9597; + and.b64 %rd9662, %rd9598, %rd9661; + xor.b64 %rd9663, %rd9662, %rd9596; + not.b64 %rd9664, %rd9598; + and.b64 %rd9665, %rd9599, %rd9664; + xor.b64 %rd9666, %rd9665, %rd9597; + not.b64 %rd9667, %rd9599; + and.b64 %rd9668, %rd9600, %rd9667; + xor.b64 %rd9669, %rd9668, %rd9598; + not.b64 %rd9670, %rd9600; + and.b64 %rd9671, %rd9596, %rd9670; + xor.b64 %rd9672, %rd9671, %rd9599; + not.b64 %rd9673, %rd9596; + and.b64 %rd9674, %rd9597, %rd9673; + xor.b64 %rd9675, %rd9674, %rd9600; + xor.b64 %rd9677, %rd9603, %rd25; + xor.b64 %rd9678, %rd9618, %rd9677; + xor.b64 %rd9679, %rd9678, %rd9633; + xor.b64 %rd9680, %rd9679, %rd9648; + xor.b64 %rd9681, %rd9680, %rd9663; + xor.b64 %rd9682, %rd9621, %rd9606; + xor.b64 %rd9683, %rd9682, %rd9636; + xor.b64 %rd9684, %rd9683, %rd9651; + xor.b64 %rd9685, %rd9684, %rd9666; + xor.b64 %rd9686, %rd9624, %rd9609; + xor.b64 %rd9687, %rd9686, %rd9639; + xor.b64 %rd9688, %rd9687, %rd9654; + xor.b64 %rd9689, %rd9688, %rd9669; + xor.b64 %rd9690, %rd9627, %rd9612; + xor.b64 %rd9691, %rd9690, %rd9642; + xor.b64 %rd9692, %rd9691, %rd9657; + xor.b64 %rd9693, %rd9692, %rd9672; + xor.b64 %rd9694, %rd9630, %rd9615; + xor.b64 %rd9695, %rd9694, %rd9645; + xor.b64 %rd9696, %rd9695, %rd9660; + xor.b64 %rd9697, %rd9696, %rd9675; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6806}, %rd9685; + } + { + .reg .b32 %dummy; + mov.b64 {%r6807,%dummy}, %rd9685; + } + shf.l.wrap.b32 %r6808, %r6807, %r6806, 1; + shf.l.wrap.b32 %r6809, %r6806, %r6807, 1; + mov.b64 %rd9698, {%r6809, %r6808}; + xor.b64 %rd9699, %rd9697, %rd9698; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6810}, %rd9689; + } + { + .reg .b32 %dummy; + mov.b64 {%r6811,%dummy}, %rd9689; + } + shf.l.wrap.b32 %r6812, %r6811, %r6810, 1; + shf.l.wrap.b32 %r6813, %r6810, %r6811, 1; + mov.b64 %rd9700, {%r6813, %r6812}; + xor.b64 %rd9701, %rd9700, %rd9681; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6814}, %rd9693; + } + { + .reg .b32 %dummy; + mov.b64 {%r6815,%dummy}, %rd9693; + } + shf.l.wrap.b32 %r6816, %r6815, %r6814, 1; + shf.l.wrap.b32 %r6817, %r6814, %r6815, 1; + mov.b64 %rd9702, {%r6817, %r6816}; + xor.b64 %rd9703, %rd9702, %rd9685; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6818}, %rd9697; + } + { + .reg .b32 %dummy; + mov.b64 {%r6819,%dummy}, %rd9697; + } + shf.l.wrap.b32 %r6820, %r6819, %r6818, 1; + shf.l.wrap.b32 %r6821, %r6818, %r6819, 1; + mov.b64 %rd9704, {%r6821, %r6820}; + xor.b64 %rd9705, %rd9704, %rd9689; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6822}, %rd9681; + } + { + .reg .b32 %dummy; + mov.b64 {%r6823,%dummy}, %rd9681; + } + shf.l.wrap.b32 %r6824, %r6823, %r6822, 1; + shf.l.wrap.b32 %r6825, %r6822, %r6823, 1; + mov.b64 %rd9706, {%r6825, %r6824}; + xor.b64 %rd9707, %rd9693, %rd9706; + xor.b64 %rd9708, %rd9699, %rd9677; + xor.b64 %rd9709, %rd9701, %rd9606; + xor.b64 %rd9710, %rd9703, %rd9609; + xor.b64 %rd9711, %rd9705, %rd9612; + xor.b64 %rd9712, %rd9707, %rd9615; + xor.b64 %rd9713, %rd9699, %rd9618; + xor.b64 %rd9714, %rd9701, %rd9621; + xor.b64 %rd9715, %rd9703, %rd9624; + xor.b64 %rd9716, %rd9705, %rd9627; + xor.b64 %rd9717, %rd9707, %rd9630; + xor.b64 %rd9718, %rd9699, %rd9633; + xor.b64 %rd9719, %rd9701, %rd9636; + xor.b64 %rd9720, %rd9703, %rd9639; + xor.b64 %rd9721, %rd9705, %rd9642; + xor.b64 %rd9722, %rd9707, %rd9645; + xor.b64 %rd9723, %rd9699, %rd9648; + xor.b64 %rd9724, %rd9701, %rd9651; + xor.b64 %rd9725, %rd9703, %rd9654; + xor.b64 %rd9726, %rd9705, %rd9657; + xor.b64 %rd9727, %rd9707, %rd9660; + xor.b64 %rd9728, %rd9699, %rd9663; + xor.b64 %rd9729, %rd9701, %rd9666; + xor.b64 %rd9730, %rd9703, %rd9669; + xor.b64 %rd9731, %rd9705, %rd9672; + xor.b64 %rd9732, %rd9707, %rd9675; + { + .reg .b32 %dummy; + mov.b64 {%r6826,%dummy}, %rd9714; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6827}, %rd9714; + } + shf.r.wrap.b32 %r6828, %r6827, %r6826, 20; + shf.r.wrap.b32 %r6829, %r6826, %r6827, 20; + mov.b64 %rd9733, {%r6829, %r6828}; + { + .reg .b32 %dummy; + mov.b64 {%r6830,%dummy}, %rd9720; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6831}, %rd9720; + } + shf.r.wrap.b32 %r6832, %r6831, %r6830, 21; + shf.r.wrap.b32 %r6833, %r6830, %r6831, 21; + mov.b64 %rd9734, {%r6833, %r6832}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6834}, %rd9726; + } + { + .reg .b32 %dummy; + mov.b64 {%r6835,%dummy}, %rd9726; + } + shf.l.wrap.b32 %r6836, %r6835, %r6834, 21; + shf.l.wrap.b32 %r6837, %r6834, %r6835, 21; + mov.b64 %rd9735, {%r6837, %r6836}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6838}, %rd9732; + } + { + .reg .b32 %dummy; + mov.b64 {%r6839,%dummy}, %rd9732; + } + shf.l.wrap.b32 %r6840, %r6839, %r6838, 14; + shf.l.wrap.b32 %r6841, %r6838, %r6839, 14; + mov.b64 %rd9736, {%r6841, %r6840}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6842}, %rd9711; + } + { + .reg .b32 %dummy; + mov.b64 {%r6843,%dummy}, %rd9711; + } + shf.l.wrap.b32 %r6844, %r6843, %r6842, 28; + shf.l.wrap.b32 %r6845, %r6842, %r6843, 28; + mov.b64 %rd9737, {%r6845, %r6844}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6846}, %rd9717; + } + { + .reg .b32 %dummy; + mov.b64 {%r6847,%dummy}, %rd9717; + } + shf.l.wrap.b32 %r6848, %r6847, %r6846, 20; + shf.l.wrap.b32 %r6849, %r6846, %r6847, 20; + mov.b64 %rd9738, {%r6849, %r6848}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6850}, %rd9718; + } + { + .reg .b32 %dummy; + mov.b64 {%r6851,%dummy}, %rd9718; + } + shf.l.wrap.b32 %r6852, %r6851, %r6850, 3; + shf.l.wrap.b32 %r6853, %r6850, %r6851, 3; + mov.b64 %rd9739, {%r6853, %r6852}; + { + .reg .b32 %dummy; + mov.b64 {%r6854,%dummy}, %rd9724; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6855}, %rd9724; + } + shf.r.wrap.b32 %r6856, %r6855, %r6854, 19; + shf.r.wrap.b32 %r6857, %r6854, %r6855, 19; + mov.b64 %rd9740, {%r6857, %r6856}; + { + .reg .b32 %dummy; + mov.b64 {%r6858,%dummy}, %rd9730; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6859}, %rd9730; + } + shf.r.wrap.b32 %r6860, %r6859, %r6858, 3; + shf.r.wrap.b32 %r6861, %r6858, %r6859, 3; + mov.b64 %rd9741, {%r6861, %r6860}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6862}, %rd9709; + } + { + .reg .b32 %dummy; + mov.b64 {%r6863,%dummy}, %rd9709; + } + shf.l.wrap.b32 %r6864, %r6863, %r6862, 1; + shf.l.wrap.b32 %r6865, %r6862, %r6863, 1; + mov.b64 %rd9742, {%r6865, %r6864}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6866}, %rd9715; + } + { + .reg .b32 %dummy; + mov.b64 {%r6867,%dummy}, %rd9715; + } + shf.l.wrap.b32 %r6868, %r6867, %r6866, 6; + shf.l.wrap.b32 %r6869, %r6866, %r6867, 6; + mov.b64 %rd9743, {%r6869, %r6868}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6870}, %rd9721; + } + { + .reg .b32 %dummy; + mov.b64 {%r6871,%dummy}, %rd9721; + } + shf.l.wrap.b32 %r6872, %r6871, %r6870, 25; + shf.l.wrap.b32 %r6873, %r6870, %r6871, 25; + mov.b64 %rd9744, {%r6873, %r6872}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6874}, %rd9727; + } + { + .reg .b32 %dummy; + mov.b64 {%r6875,%dummy}, %rd9727; + } + shf.l.wrap.b32 %r6876, %r6875, %r6874, 8; + shf.l.wrap.b32 %r6877, %r6874, %r6875, 8; + mov.b64 %rd9745, {%r6877, %r6876}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6878}, %rd9728; + } + { + .reg .b32 %dummy; + mov.b64 {%r6879,%dummy}, %rd9728; + } + shf.l.wrap.b32 %r6880, %r6879, %r6878, 18; + shf.l.wrap.b32 %r6881, %r6878, %r6879, 18; + mov.b64 %rd9746, {%r6881, %r6880}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6882}, %rd9712; + } + { + .reg .b32 %dummy; + mov.b64 {%r6883,%dummy}, %rd9712; + } + shf.l.wrap.b32 %r6884, %r6883, %r6882, 27; + shf.l.wrap.b32 %r6885, %r6882, %r6883, 27; + mov.b64 %rd9747, {%r6885, %r6884}; + { + .reg .b32 %dummy; + mov.b64 {%r6886,%dummy}, %rd9713; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6887}, %rd9713; + } + shf.r.wrap.b32 %r6888, %r6887, %r6886, 28; + shf.r.wrap.b32 %r6889, %r6886, %r6887, 28; + mov.b64 %rd9748, {%r6889, %r6888}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6890}, %rd9719; + } + { + .reg .b32 %dummy; + mov.b64 {%r6891,%dummy}, %rd9719; + } + shf.l.wrap.b32 %r6892, %r6891, %r6890, 10; + shf.l.wrap.b32 %r6893, %r6890, %r6891, 10; + mov.b64 %rd9749, {%r6893, %r6892}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6894}, %rd9725; + } + { + .reg .b32 %dummy; + mov.b64 {%r6895,%dummy}, %rd9725; + } + shf.l.wrap.b32 %r6896, %r6895, %r6894, 15; + shf.l.wrap.b32 %r6897, %r6894, %r6895, 15; + mov.b64 %rd9750, {%r6897, %r6896}; + { + .reg .b32 %dummy; + mov.b64 {%r6898,%dummy}, %rd9731; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6899}, %rd9731; + } + shf.r.wrap.b32 %r6900, %r6899, %r6898, 8; + shf.r.wrap.b32 %r6901, %r6898, %r6899, 8; + mov.b64 %rd9751, {%r6901, %r6900}; + { + .reg .b32 %dummy; + mov.b64 {%r6902,%dummy}, %rd9710; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6903}, %rd9710; + } + shf.r.wrap.b32 %r6904, %r6903, %r6902, 2; + shf.r.wrap.b32 %r6905, %r6902, %r6903, 2; + mov.b64 %rd9752, {%r6905, %r6904}; + { + .reg .b32 %dummy; + mov.b64 {%r6906,%dummy}, %rd9716; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6907}, %rd9716; + } + shf.r.wrap.b32 %r6908, %r6907, %r6906, 9; + shf.r.wrap.b32 %r6909, %r6906, %r6907, 9; + mov.b64 %rd9753, {%r6909, %r6908}; + { + .reg .b32 %dummy; + mov.b64 {%r6910,%dummy}, %rd9722; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6911}, %rd9722; + } + shf.r.wrap.b32 %r6912, %r6911, %r6910, 25; + shf.r.wrap.b32 %r6913, %r6910, %r6911, 25; + mov.b64 %rd9754, {%r6913, %r6912}; + { + .reg .b32 %dummy; + mov.b64 {%r6914,%dummy}, %rd9723; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6915}, %rd9723; + } + shf.r.wrap.b32 %r6916, %r6915, %r6914, 23; + shf.r.wrap.b32 %r6917, %r6914, %r6915, 23; + mov.b64 %rd9755, {%r6917, %r6916}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6918}, %rd9729; + } + { + .reg .b32 %dummy; + mov.b64 {%r6919,%dummy}, %rd9729; + } + shf.l.wrap.b32 %r6920, %r6919, %r6918, 2; + shf.l.wrap.b32 %r6921, %r6918, %r6919, 2; + mov.b64 %rd9756, {%r6921, %r6920}; + not.b64 %rd9757, %rd9733; + and.b64 %rd9758, %rd9734, %rd9757; + xor.b64 %rd9759, %rd9758, %rd9708; + not.b64 %rd9760, %rd9734; + and.b64 %rd9761, %rd9735, %rd9760; + xor.b64 %rd9762, %rd9761, %rd9733; + not.b64 %rd9763, %rd9735; + and.b64 %rd9764, %rd9736, %rd9763; + xor.b64 %rd9765, %rd9764, %rd9734; + not.b64 %rd9766, %rd9736; + and.b64 %rd9767, %rd9708, %rd9766; + xor.b64 %rd9768, %rd9767, %rd9735; + not.b64 %rd9769, %rd9708; + and.b64 %rd9770, %rd9733, %rd9769; + xor.b64 %rd9771, %rd9770, %rd9736; + not.b64 %rd9772, %rd9738; + and.b64 %rd9773, %rd9739, %rd9772; + xor.b64 %rd9774, %rd9773, %rd9737; + not.b64 %rd9775, %rd9739; + and.b64 %rd9776, %rd9740, %rd9775; + xor.b64 %rd9777, %rd9776, %rd9738; + not.b64 %rd9778, %rd9740; + and.b64 %rd9779, %rd9741, %rd9778; + xor.b64 %rd9780, %rd9779, %rd9739; + not.b64 %rd9781, %rd9741; + and.b64 %rd9782, %rd9737, %rd9781; + xor.b64 %rd9783, %rd9782, %rd9740; + not.b64 %rd9784, %rd9737; + and.b64 %rd9785, %rd9738, %rd9784; + xor.b64 %rd9786, %rd9785, %rd9741; + not.b64 %rd9787, %rd9743; + and.b64 %rd9788, %rd9744, %rd9787; + xor.b64 %rd9789, %rd9788, %rd9742; + not.b64 %rd9790, %rd9744; + and.b64 %rd9791, %rd9745, %rd9790; + xor.b64 %rd9792, %rd9791, %rd9743; + not.b64 %rd9793, %rd9745; + and.b64 %rd9794, %rd9746, %rd9793; + xor.b64 %rd9795, %rd9794, %rd9744; + not.b64 %rd9796, %rd9746; + and.b64 %rd9797, %rd9742, %rd9796; + xor.b64 %rd9798, %rd9797, %rd9745; + not.b64 %rd9799, %rd9742; + and.b64 %rd9800, %rd9743, %rd9799; + xor.b64 %rd9801, %rd9800, %rd9746; + not.b64 %rd9802, %rd9748; + and.b64 %rd9803, %rd9749, %rd9802; + xor.b64 %rd9804, %rd9803, %rd9747; + not.b64 %rd9805, %rd9749; + and.b64 %rd9806, %rd9750, %rd9805; + xor.b64 %rd9807, %rd9806, %rd9748; + not.b64 %rd9808, %rd9750; + and.b64 %rd9809, %rd9751, %rd9808; + xor.b64 %rd9810, %rd9809, %rd9749; + not.b64 %rd9811, %rd9751; + and.b64 %rd9812, %rd9747, %rd9811; + xor.b64 %rd9813, %rd9812, %rd9750; + not.b64 %rd9814, %rd9747; + and.b64 %rd9815, %rd9748, %rd9814; + xor.b64 %rd9816, %rd9815, %rd9751; + not.b64 %rd9817, %rd9753; + and.b64 %rd9818, %rd9754, %rd9817; + xor.b64 %rd9819, %rd9818, %rd9752; + not.b64 %rd9820, %rd9754; + and.b64 %rd9821, %rd9755, %rd9820; + xor.b64 %rd9822, %rd9821, %rd9753; + not.b64 %rd9823, %rd9755; + and.b64 %rd9824, %rd9756, %rd9823; + xor.b64 %rd9825, %rd9824, %rd9754; + not.b64 %rd9826, %rd9756; + and.b64 %rd9827, %rd9752, %rd9826; + xor.b64 %rd9828, %rd9827, %rd9755; + not.b64 %rd9829, %rd9752; + and.b64 %rd9830, %rd9753, %rd9829; + xor.b64 %rd9831, %rd9830, %rd9756; + xor.b64 %rd9833, %rd9759, %rd26; + xor.b64 %rd9834, %rd9774, %rd9833; + xor.b64 %rd9835, %rd9834, %rd9789; + xor.b64 %rd9836, %rd9835, %rd9804; + xor.b64 %rd9837, %rd9836, %rd9819; + xor.b64 %rd9838, %rd9777, %rd9762; + xor.b64 %rd9839, %rd9838, %rd9792; + xor.b64 %rd9840, %rd9839, %rd9807; + xor.b64 %rd9841, %rd9840, %rd9822; + xor.b64 %rd9842, %rd9780, %rd9765; + xor.b64 %rd9843, %rd9842, %rd9795; + xor.b64 %rd9844, %rd9843, %rd9810; + xor.b64 %rd9845, %rd9844, %rd9825; + xor.b64 %rd9846, %rd9783, %rd9768; + xor.b64 %rd9847, %rd9846, %rd9798; + xor.b64 %rd9848, %rd9847, %rd9813; + xor.b64 %rd9849, %rd9848, %rd9828; + xor.b64 %rd9850, %rd9786, %rd9771; + xor.b64 %rd9851, %rd9850, %rd9801; + xor.b64 %rd9852, %rd9851, %rd9816; + xor.b64 %rd9853, %rd9852, %rd9831; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6922}, %rd9841; + } + { + .reg .b32 %dummy; + mov.b64 {%r6923,%dummy}, %rd9841; + } + shf.l.wrap.b32 %r6924, %r6923, %r6922, 1; + shf.l.wrap.b32 %r6925, %r6922, %r6923, 1; + mov.b64 %rd9854, {%r6925, %r6924}; + xor.b64 %rd9855, %rd9853, %rd9854; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6926}, %rd9845; + } + { + .reg .b32 %dummy; + mov.b64 {%r6927,%dummy}, %rd9845; + } + shf.l.wrap.b32 %r6928, %r6927, %r6926, 1; + shf.l.wrap.b32 %r6929, %r6926, %r6927, 1; + mov.b64 %rd9856, {%r6929, %r6928}; + xor.b64 %rd9857, %rd9856, %rd9837; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6930}, %rd9849; + } + { + .reg .b32 %dummy; + mov.b64 {%r6931,%dummy}, %rd9849; + } + shf.l.wrap.b32 %r6932, %r6931, %r6930, 1; + shf.l.wrap.b32 %r6933, %r6930, %r6931, 1; + mov.b64 %rd9858, {%r6933, %r6932}; + xor.b64 %rd9859, %rd9858, %rd9841; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6934}, %rd9853; + } + { + .reg .b32 %dummy; + mov.b64 {%r6935,%dummy}, %rd9853; + } + shf.l.wrap.b32 %r6936, %r6935, %r6934, 1; + shf.l.wrap.b32 %r6937, %r6934, %r6935, 1; + mov.b64 %rd9860, {%r6937, %r6936}; + xor.b64 %rd9861, %rd9860, %rd9845; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6938}, %rd9837; + } + { + .reg .b32 %dummy; + mov.b64 {%r6939,%dummy}, %rd9837; + } + shf.l.wrap.b32 %r6940, %r6939, %r6938, 1; + shf.l.wrap.b32 %r6941, %r6938, %r6939, 1; + mov.b64 %rd9862, {%r6941, %r6940}; + xor.b64 %rd9863, %rd9849, %rd9862; + xor.b64 %rd9864, %rd9855, %rd9833; + xor.b64 %rd9865, %rd9857, %rd9762; + xor.b64 %rd9866, %rd9859, %rd9765; + xor.b64 %rd9867, %rd9861, %rd9768; + xor.b64 %rd9868, %rd9863, %rd9771; + xor.b64 %rd9869, %rd9855, %rd9774; + xor.b64 %rd9870, %rd9857, %rd9777; + xor.b64 %rd9871, %rd9859, %rd9780; + xor.b64 %rd9872, %rd9861, %rd9783; + xor.b64 %rd9873, %rd9863, %rd9786; + xor.b64 %rd9874, %rd9855, %rd9789; + xor.b64 %rd9875, %rd9857, %rd9792; + xor.b64 %rd9876, %rd9859, %rd9795; + xor.b64 %rd9877, %rd9861, %rd9798; + xor.b64 %rd9878, %rd9863, %rd9801; + xor.b64 %rd9879, %rd9855, %rd9804; + xor.b64 %rd9880, %rd9857, %rd9807; + xor.b64 %rd9881, %rd9859, %rd9810; + xor.b64 %rd9882, %rd9861, %rd9813; + xor.b64 %rd9883, %rd9863, %rd9816; + xor.b64 %rd9884, %rd9855, %rd9819; + xor.b64 %rd9885, %rd9857, %rd9822; + xor.b64 %rd9886, %rd9859, %rd9825; + xor.b64 %rd9887, %rd9861, %rd9828; + xor.b64 %rd9888, %rd9863, %rd9831; + { + .reg .b32 %dummy; + mov.b64 {%r6942,%dummy}, %rd9870; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6943}, %rd9870; + } + shf.r.wrap.b32 %r6944, %r6943, %r6942, 20; + shf.r.wrap.b32 %r6945, %r6942, %r6943, 20; + mov.b64 %rd9889, {%r6945, %r6944}; + { + .reg .b32 %dummy; + mov.b64 {%r6946,%dummy}, %rd9876; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6947}, %rd9876; + } + shf.r.wrap.b32 %r6948, %r6947, %r6946, 21; + shf.r.wrap.b32 %r6949, %r6946, %r6947, 21; + mov.b64 %rd9890, {%r6949, %r6948}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6950}, %rd9882; + } + { + .reg .b32 %dummy; + mov.b64 {%r6951,%dummy}, %rd9882; + } + shf.l.wrap.b32 %r6952, %r6951, %r6950, 21; + shf.l.wrap.b32 %r6953, %r6950, %r6951, 21; + mov.b64 %rd9891, {%r6953, %r6952}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6954}, %rd9888; + } + { + .reg .b32 %dummy; + mov.b64 {%r6955,%dummy}, %rd9888; + } + shf.l.wrap.b32 %r6956, %r6955, %r6954, 14; + shf.l.wrap.b32 %r6957, %r6954, %r6955, 14; + mov.b64 %rd9892, {%r6957, %r6956}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6958}, %rd9867; + } + { + .reg .b32 %dummy; + mov.b64 {%r6959,%dummy}, %rd9867; + } + shf.l.wrap.b32 %r6960, %r6959, %r6958, 28; + shf.l.wrap.b32 %r6961, %r6958, %r6959, 28; + mov.b64 %rd9893, {%r6961, %r6960}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6962}, %rd9873; + } + { + .reg .b32 %dummy; + mov.b64 {%r6963,%dummy}, %rd9873; + } + shf.l.wrap.b32 %r6964, %r6963, %r6962, 20; + shf.l.wrap.b32 %r6965, %r6962, %r6963, 20; + mov.b64 %rd9894, {%r6965, %r6964}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6966}, %rd9874; + } + { + .reg .b32 %dummy; + mov.b64 {%r6967,%dummy}, %rd9874; + } + shf.l.wrap.b32 %r6968, %r6967, %r6966, 3; + shf.l.wrap.b32 %r6969, %r6966, %r6967, 3; + mov.b64 %rd9895, {%r6969, %r6968}; + { + .reg .b32 %dummy; + mov.b64 {%r6970,%dummy}, %rd9880; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6971}, %rd9880; + } + shf.r.wrap.b32 %r6972, %r6971, %r6970, 19; + shf.r.wrap.b32 %r6973, %r6970, %r6971, 19; + mov.b64 %rd9896, {%r6973, %r6972}; + { + .reg .b32 %dummy; + mov.b64 {%r6974,%dummy}, %rd9886; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6975}, %rd9886; + } + shf.r.wrap.b32 %r6976, %r6975, %r6974, 3; + shf.r.wrap.b32 %r6977, %r6974, %r6975, 3; + mov.b64 %rd9897, {%r6977, %r6976}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6978}, %rd9865; + } + { + .reg .b32 %dummy; + mov.b64 {%r6979,%dummy}, %rd9865; + } + shf.l.wrap.b32 %r6980, %r6979, %r6978, 1; + shf.l.wrap.b32 %r6981, %r6978, %r6979, 1; + mov.b64 %rd9898, {%r6981, %r6980}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6982}, %rd9871; + } + { + .reg .b32 %dummy; + mov.b64 {%r6983,%dummy}, %rd9871; + } + shf.l.wrap.b32 %r6984, %r6983, %r6982, 6; + shf.l.wrap.b32 %r6985, %r6982, %r6983, 6; + mov.b64 %rd9899, {%r6985, %r6984}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6986}, %rd9877; + } + { + .reg .b32 %dummy; + mov.b64 {%r6987,%dummy}, %rd9877; + } + shf.l.wrap.b32 %r6988, %r6987, %r6986, 25; + shf.l.wrap.b32 %r6989, %r6986, %r6987, 25; + mov.b64 %rd9900, {%r6989, %r6988}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6990}, %rd9883; + } + { + .reg .b32 %dummy; + mov.b64 {%r6991,%dummy}, %rd9883; + } + shf.l.wrap.b32 %r6992, %r6991, %r6990, 8; + shf.l.wrap.b32 %r6993, %r6990, %r6991, 8; + mov.b64 %rd9901, {%r6993, %r6992}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6994}, %rd9884; + } + { + .reg .b32 %dummy; + mov.b64 {%r6995,%dummy}, %rd9884; + } + shf.l.wrap.b32 %r6996, %r6995, %r6994, 18; + shf.l.wrap.b32 %r6997, %r6994, %r6995, 18; + mov.b64 %rd9902, {%r6997, %r6996}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r6998}, %rd9868; + } + { + .reg .b32 %dummy; + mov.b64 {%r6999,%dummy}, %rd9868; + } + shf.l.wrap.b32 %r7000, %r6999, %r6998, 27; + shf.l.wrap.b32 %r7001, %r6998, %r6999, 27; + mov.b64 %rd9903, {%r7001, %r7000}; + { + .reg .b32 %dummy; + mov.b64 {%r7002,%dummy}, %rd9869; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7003}, %rd9869; + } + shf.r.wrap.b32 %r7004, %r7003, %r7002, 28; + shf.r.wrap.b32 %r7005, %r7002, %r7003, 28; + mov.b64 %rd9904, {%r7005, %r7004}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7006}, %rd9875; + } + { + .reg .b32 %dummy; + mov.b64 {%r7007,%dummy}, %rd9875; + } + shf.l.wrap.b32 %r7008, %r7007, %r7006, 10; + shf.l.wrap.b32 %r7009, %r7006, %r7007, 10; + mov.b64 %rd9905, {%r7009, %r7008}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7010}, %rd9881; + } + { + .reg .b32 %dummy; + mov.b64 {%r7011,%dummy}, %rd9881; + } + shf.l.wrap.b32 %r7012, %r7011, %r7010, 15; + shf.l.wrap.b32 %r7013, %r7010, %r7011, 15; + mov.b64 %rd9906, {%r7013, %r7012}; + { + .reg .b32 %dummy; + mov.b64 {%r7014,%dummy}, %rd9887; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7015}, %rd9887; + } + shf.r.wrap.b32 %r7016, %r7015, %r7014, 8; + shf.r.wrap.b32 %r7017, %r7014, %r7015, 8; + mov.b64 %rd9907, {%r7017, %r7016}; + { + .reg .b32 %dummy; + mov.b64 {%r7018,%dummy}, %rd9866; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7019}, %rd9866; + } + shf.r.wrap.b32 %r7020, %r7019, %r7018, 2; + shf.r.wrap.b32 %r7021, %r7018, %r7019, 2; + mov.b64 %rd9908, {%r7021, %r7020}; + { + .reg .b32 %dummy; + mov.b64 {%r7022,%dummy}, %rd9872; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7023}, %rd9872; + } + shf.r.wrap.b32 %r7024, %r7023, %r7022, 9; + shf.r.wrap.b32 %r7025, %r7022, %r7023, 9; + mov.b64 %rd9909, {%r7025, %r7024}; + { + .reg .b32 %dummy; + mov.b64 {%r7026,%dummy}, %rd9878; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7027}, %rd9878; + } + shf.r.wrap.b32 %r7028, %r7027, %r7026, 25; + shf.r.wrap.b32 %r7029, %r7026, %r7027, 25; + mov.b64 %rd9910, {%r7029, %r7028}; + { + .reg .b32 %dummy; + mov.b64 {%r7030,%dummy}, %rd9879; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7031}, %rd9879; + } + shf.r.wrap.b32 %r7032, %r7031, %r7030, 23; + shf.r.wrap.b32 %r7033, %r7030, %r7031, 23; + mov.b64 %rd9911, {%r7033, %r7032}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7034}, %rd9885; + } + { + .reg .b32 %dummy; + mov.b64 {%r7035,%dummy}, %rd9885; + } + shf.l.wrap.b32 %r7036, %r7035, %r7034, 2; + shf.l.wrap.b32 %r7037, %r7034, %r7035, 2; + mov.b64 %rd9912, {%r7037, %r7036}; + not.b64 %rd9913, %rd9889; + and.b64 %rd9914, %rd9890, %rd9913; + xor.b64 %rd9915, %rd9914, %rd9864; + not.b64 %rd9916, %rd9890; + and.b64 %rd9917, %rd9891, %rd9916; + xor.b64 %rd9918, %rd9917, %rd9889; + not.b64 %rd9919, %rd9891; + and.b64 %rd9920, %rd9892, %rd9919; + xor.b64 %rd9921, %rd9920, %rd9890; + not.b64 %rd9922, %rd9892; + and.b64 %rd9923, %rd9864, %rd9922; + xor.b64 %rd9924, %rd9923, %rd9891; + not.b64 %rd9925, %rd9864; + and.b64 %rd9926, %rd9889, %rd9925; + xor.b64 %rd9927, %rd9926, %rd9892; + not.b64 %rd9928, %rd9894; + and.b64 %rd9929, %rd9895, %rd9928; + xor.b64 %rd9930, %rd9929, %rd9893; + not.b64 %rd9931, %rd9895; + and.b64 %rd9932, %rd9896, %rd9931; + xor.b64 %rd9933, %rd9932, %rd9894; + not.b64 %rd9934, %rd9896; + and.b64 %rd9935, %rd9897, %rd9934; + xor.b64 %rd9936, %rd9935, %rd9895; + not.b64 %rd9937, %rd9897; + and.b64 %rd9938, %rd9893, %rd9937; + xor.b64 %rd9939, %rd9938, %rd9896; + not.b64 %rd9940, %rd9893; + and.b64 %rd9941, %rd9894, %rd9940; + xor.b64 %rd9942, %rd9941, %rd9897; + not.b64 %rd9943, %rd9899; + and.b64 %rd9944, %rd9900, %rd9943; + xor.b64 %rd9945, %rd9944, %rd9898; + not.b64 %rd9946, %rd9900; + and.b64 %rd9947, %rd9901, %rd9946; + xor.b64 %rd9948, %rd9947, %rd9899; + not.b64 %rd9949, %rd9901; + and.b64 %rd9950, %rd9902, %rd9949; + xor.b64 %rd9951, %rd9950, %rd9900; + not.b64 %rd9952, %rd9902; + and.b64 %rd9953, %rd9898, %rd9952; + xor.b64 %rd9954, %rd9953, %rd9901; + not.b64 %rd9955, %rd9898; + and.b64 %rd9956, %rd9899, %rd9955; + xor.b64 %rd9957, %rd9956, %rd9902; + not.b64 %rd9958, %rd9904; + and.b64 %rd9959, %rd9905, %rd9958; + xor.b64 %rd9960, %rd9959, %rd9903; + not.b64 %rd9961, %rd9905; + and.b64 %rd9962, %rd9906, %rd9961; + xor.b64 %rd9963, %rd9962, %rd9904; + not.b64 %rd9964, %rd9906; + and.b64 %rd9965, %rd9907, %rd9964; + xor.b64 %rd9966, %rd9965, %rd9905; + not.b64 %rd9967, %rd9907; + and.b64 %rd9968, %rd9903, %rd9967; + xor.b64 %rd9969, %rd9968, %rd9906; + not.b64 %rd9970, %rd9903; + and.b64 %rd9971, %rd9904, %rd9970; + xor.b64 %rd9972, %rd9971, %rd9907; + not.b64 %rd9973, %rd9909; + and.b64 %rd9974, %rd9910, %rd9973; + xor.b64 %rd9975, %rd9974, %rd9908; + not.b64 %rd9976, %rd9910; + and.b64 %rd9977, %rd9911, %rd9976; + xor.b64 %rd9978, %rd9977, %rd9909; + not.b64 %rd9979, %rd9911; + and.b64 %rd9980, %rd9912, %rd9979; + xor.b64 %rd9981, %rd9980, %rd9910; + not.b64 %rd9982, %rd9912; + and.b64 %rd9983, %rd9908, %rd9982; + xor.b64 %rd9984, %rd9983, %rd9911; + not.b64 %rd9985, %rd9908; + and.b64 %rd9986, %rd9909, %rd9985; + xor.b64 %rd9987, %rd9986, %rd9912; + xor.b64 %rd9989, %rd9915, %rd27; + xor.b64 %rd9990, %rd9930, %rd9989; + xor.b64 %rd9991, %rd9990, %rd9945; + xor.b64 %rd9992, %rd9991, %rd9960; + xor.b64 %rd9993, %rd9992, %rd9975; + xor.b64 %rd9994, %rd9933, %rd9918; + xor.b64 %rd9995, %rd9994, %rd9948; + xor.b64 %rd9996, %rd9995, %rd9963; + xor.b64 %rd9997, %rd9996, %rd9978; + xor.b64 %rd9998, %rd9936, %rd9921; + xor.b64 %rd9999, %rd9998, %rd9951; + xor.b64 %rd10000, %rd9999, %rd9966; + xor.b64 %rd10001, %rd10000, %rd9981; + xor.b64 %rd10002, %rd9939, %rd9924; + xor.b64 %rd10003, %rd10002, %rd9954; + xor.b64 %rd10004, %rd10003, %rd9969; + xor.b64 %rd10005, %rd10004, %rd9984; + xor.b64 %rd10006, %rd9942, %rd9927; + xor.b64 %rd10007, %rd10006, %rd9957; + xor.b64 %rd10008, %rd10007, %rd9972; + xor.b64 %rd10009, %rd10008, %rd9987; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7038}, %rd9997; + } + { + .reg .b32 %dummy; + mov.b64 {%r7039,%dummy}, %rd9997; + } + shf.l.wrap.b32 %r7040, %r7039, %r7038, 1; + shf.l.wrap.b32 %r7041, %r7038, %r7039, 1; + mov.b64 %rd10010, {%r7041, %r7040}; + xor.b64 %rd10011, %rd10009, %rd10010; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7042}, %rd10001; + } + { + .reg .b32 %dummy; + mov.b64 {%r7043,%dummy}, %rd10001; + } + shf.l.wrap.b32 %r7044, %r7043, %r7042, 1; + shf.l.wrap.b32 %r7045, %r7042, %r7043, 1; + mov.b64 %rd10012, {%r7045, %r7044}; + xor.b64 %rd10013, %rd10012, %rd9993; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7046}, %rd10005; + } + { + .reg .b32 %dummy; + mov.b64 {%r7047,%dummy}, %rd10005; + } + shf.l.wrap.b32 %r7048, %r7047, %r7046, 1; + shf.l.wrap.b32 %r7049, %r7046, %r7047, 1; + mov.b64 %rd10014, {%r7049, %r7048}; + xor.b64 %rd10015, %rd10014, %rd9997; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7050}, %rd10009; + } + { + .reg .b32 %dummy; + mov.b64 {%r7051,%dummy}, %rd10009; + } + shf.l.wrap.b32 %r7052, %r7051, %r7050, 1; + shf.l.wrap.b32 %r7053, %r7050, %r7051, 1; + mov.b64 %rd10016, {%r7053, %r7052}; + xor.b64 %rd10017, %rd10016, %rd10001; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7054}, %rd9993; + } + { + .reg .b32 %dummy; + mov.b64 {%r7055,%dummy}, %rd9993; + } + shf.l.wrap.b32 %r7056, %r7055, %r7054, 1; + shf.l.wrap.b32 %r7057, %r7054, %r7055, 1; + mov.b64 %rd10018, {%r7057, %r7056}; + xor.b64 %rd10019, %rd10005, %rd10018; + xor.b64 %rd10020, %rd10011, %rd9989; + xor.b64 %rd10021, %rd10013, %rd9918; + xor.b64 %rd10022, %rd10015, %rd9921; + xor.b64 %rd10023, %rd10017, %rd9924; + xor.b64 %rd10024, %rd10019, %rd9927; + xor.b64 %rd10025, %rd10011, %rd9930; + xor.b64 %rd10026, %rd10013, %rd9933; + xor.b64 %rd10027, %rd10015, %rd9936; + xor.b64 %rd10028, %rd10017, %rd9939; + xor.b64 %rd10029, %rd10019, %rd9942; + xor.b64 %rd10030, %rd10011, %rd9945; + xor.b64 %rd10031, %rd10013, %rd9948; + xor.b64 %rd10032, %rd10015, %rd9951; + xor.b64 %rd10033, %rd10017, %rd9954; + xor.b64 %rd10034, %rd10019, %rd9957; + xor.b64 %rd10035, %rd10011, %rd9960; + xor.b64 %rd10036, %rd10013, %rd9963; + xor.b64 %rd10037, %rd10015, %rd9966; + xor.b64 %rd10038, %rd10017, %rd9969; + xor.b64 %rd10039, %rd10019, %rd9972; + xor.b64 %rd10040, %rd10011, %rd9975; + xor.b64 %rd10041, %rd10013, %rd9978; + xor.b64 %rd10042, %rd10015, %rd9981; + xor.b64 %rd10043, %rd10017, %rd9984; + xor.b64 %rd10044, %rd10019, %rd9987; + { + .reg .b32 %dummy; + mov.b64 {%r7058,%dummy}, %rd10026; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7059}, %rd10026; + } + shf.r.wrap.b32 %r7060, %r7059, %r7058, 20; + shf.r.wrap.b32 %r7061, %r7058, %r7059, 20; + mov.b64 %rd10045, {%r7061, %r7060}; + { + .reg .b32 %dummy; + mov.b64 {%r7062,%dummy}, %rd10032; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7063}, %rd10032; + } + shf.r.wrap.b32 %r7064, %r7063, %r7062, 21; + shf.r.wrap.b32 %r7065, %r7062, %r7063, 21; + mov.b64 %rd10046, {%r7065, %r7064}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7066}, %rd10038; + } + { + .reg .b32 %dummy; + mov.b64 {%r7067,%dummy}, %rd10038; + } + shf.l.wrap.b32 %r7068, %r7067, %r7066, 21; + shf.l.wrap.b32 %r7069, %r7066, %r7067, 21; + mov.b64 %rd10047, {%r7069, %r7068}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7070}, %rd10044; + } + { + .reg .b32 %dummy; + mov.b64 {%r7071,%dummy}, %rd10044; + } + shf.l.wrap.b32 %r7072, %r7071, %r7070, 14; + shf.l.wrap.b32 %r7073, %r7070, %r7071, 14; + mov.b64 %rd10048, {%r7073, %r7072}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7074}, %rd10023; + } + { + .reg .b32 %dummy; + mov.b64 {%r7075,%dummy}, %rd10023; + } + shf.l.wrap.b32 %r7076, %r7075, %r7074, 28; + shf.l.wrap.b32 %r7077, %r7074, %r7075, 28; + mov.b64 %rd10049, {%r7077, %r7076}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7078}, %rd10029; + } + { + .reg .b32 %dummy; + mov.b64 {%r7079,%dummy}, %rd10029; + } + shf.l.wrap.b32 %r7080, %r7079, %r7078, 20; + shf.l.wrap.b32 %r7081, %r7078, %r7079, 20; + mov.b64 %rd10050, {%r7081, %r7080}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7082}, %rd10030; + } + { + .reg .b32 %dummy; + mov.b64 {%r7083,%dummy}, %rd10030; + } + shf.l.wrap.b32 %r7084, %r7083, %r7082, 3; + shf.l.wrap.b32 %r7085, %r7082, %r7083, 3; + mov.b64 %rd10051, {%r7085, %r7084}; + { + .reg .b32 %dummy; + mov.b64 {%r7086,%dummy}, %rd10036; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7087}, %rd10036; + } + shf.r.wrap.b32 %r7088, %r7087, %r7086, 19; + shf.r.wrap.b32 %r7089, %r7086, %r7087, 19; + mov.b64 %rd10052, {%r7089, %r7088}; + { + .reg .b32 %dummy; + mov.b64 {%r7090,%dummy}, %rd10042; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7091}, %rd10042; + } + shf.r.wrap.b32 %r7092, %r7091, %r7090, 3; + shf.r.wrap.b32 %r7093, %r7090, %r7091, 3; + mov.b64 %rd10053, {%r7093, %r7092}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7094}, %rd10021; + } + { + .reg .b32 %dummy; + mov.b64 {%r7095,%dummy}, %rd10021; + } + shf.l.wrap.b32 %r7096, %r7095, %r7094, 1; + shf.l.wrap.b32 %r7097, %r7094, %r7095, 1; + mov.b64 %rd10054, {%r7097, %r7096}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7098}, %rd10027; + } + { + .reg .b32 %dummy; + mov.b64 {%r7099,%dummy}, %rd10027; + } + shf.l.wrap.b32 %r7100, %r7099, %r7098, 6; + shf.l.wrap.b32 %r7101, %r7098, %r7099, 6; + mov.b64 %rd10055, {%r7101, %r7100}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7102}, %rd10033; + } + { + .reg .b32 %dummy; + mov.b64 {%r7103,%dummy}, %rd10033; + } + shf.l.wrap.b32 %r7104, %r7103, %r7102, 25; + shf.l.wrap.b32 %r7105, %r7102, %r7103, 25; + mov.b64 %rd10056, {%r7105, %r7104}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7106}, %rd10039; + } + { + .reg .b32 %dummy; + mov.b64 {%r7107,%dummy}, %rd10039; + } + shf.l.wrap.b32 %r7108, %r7107, %r7106, 8; + shf.l.wrap.b32 %r7109, %r7106, %r7107, 8; + mov.b64 %rd10057, {%r7109, %r7108}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7110}, %rd10040; + } + { + .reg .b32 %dummy; + mov.b64 {%r7111,%dummy}, %rd10040; + } + shf.l.wrap.b32 %r7112, %r7111, %r7110, 18; + shf.l.wrap.b32 %r7113, %r7110, %r7111, 18; + mov.b64 %rd10058, {%r7113, %r7112}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7114}, %rd10024; + } + { + .reg .b32 %dummy; + mov.b64 {%r7115,%dummy}, %rd10024; + } + shf.l.wrap.b32 %r7116, %r7115, %r7114, 27; + shf.l.wrap.b32 %r7117, %r7114, %r7115, 27; + mov.b64 %rd10059, {%r7117, %r7116}; + { + .reg .b32 %dummy; + mov.b64 {%r7118,%dummy}, %rd10025; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7119}, %rd10025; + } + shf.r.wrap.b32 %r7120, %r7119, %r7118, 28; + shf.r.wrap.b32 %r7121, %r7118, %r7119, 28; + mov.b64 %rd10060, {%r7121, %r7120}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7122}, %rd10031; + } + { + .reg .b32 %dummy; + mov.b64 {%r7123,%dummy}, %rd10031; + } + shf.l.wrap.b32 %r7124, %r7123, %r7122, 10; + shf.l.wrap.b32 %r7125, %r7122, %r7123, 10; + mov.b64 %rd10061, {%r7125, %r7124}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7126}, %rd10037; + } + { + .reg .b32 %dummy; + mov.b64 {%r7127,%dummy}, %rd10037; + } + shf.l.wrap.b32 %r7128, %r7127, %r7126, 15; + shf.l.wrap.b32 %r7129, %r7126, %r7127, 15; + mov.b64 %rd10062, {%r7129, %r7128}; + { + .reg .b32 %dummy; + mov.b64 {%r7130,%dummy}, %rd10043; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7131}, %rd10043; + } + shf.r.wrap.b32 %r7132, %r7131, %r7130, 8; + shf.r.wrap.b32 %r7133, %r7130, %r7131, 8; + mov.b64 %rd10063, {%r7133, %r7132}; + { + .reg .b32 %dummy; + mov.b64 {%r7134,%dummy}, %rd10022; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7135}, %rd10022; + } + shf.r.wrap.b32 %r7136, %r7135, %r7134, 2; + shf.r.wrap.b32 %r7137, %r7134, %r7135, 2; + mov.b64 %rd10064, {%r7137, %r7136}; + { + .reg .b32 %dummy; + mov.b64 {%r7138,%dummy}, %rd10028; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7139}, %rd10028; + } + shf.r.wrap.b32 %r7140, %r7139, %r7138, 9; + shf.r.wrap.b32 %r7141, %r7138, %r7139, 9; + mov.b64 %rd10065, {%r7141, %r7140}; + { + .reg .b32 %dummy; + mov.b64 {%r7142,%dummy}, %rd10034; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7143}, %rd10034; + } + shf.r.wrap.b32 %r7144, %r7143, %r7142, 25; + shf.r.wrap.b32 %r7145, %r7142, %r7143, 25; + mov.b64 %rd10066, {%r7145, %r7144}; + { + .reg .b32 %dummy; + mov.b64 {%r7146,%dummy}, %rd10035; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7147}, %rd10035; + } + shf.r.wrap.b32 %r7148, %r7147, %r7146, 23; + shf.r.wrap.b32 %r7149, %r7146, %r7147, 23; + mov.b64 %rd10067, {%r7149, %r7148}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7150}, %rd10041; + } + { + .reg .b32 %dummy; + mov.b64 {%r7151,%dummy}, %rd10041; + } + shf.l.wrap.b32 %r7152, %r7151, %r7150, 2; + shf.l.wrap.b32 %r7153, %r7150, %r7151, 2; + mov.b64 %rd10068, {%r7153, %r7152}; + not.b64 %rd10069, %rd10045; + and.b64 %rd10070, %rd10046, %rd10069; + xor.b64 %rd10071, %rd10070, %rd10020; + not.b64 %rd10072, %rd10046; + and.b64 %rd10073, %rd10047, %rd10072; + xor.b64 %rd10074, %rd10073, %rd10045; + not.b64 %rd10075, %rd10047; + and.b64 %rd10076, %rd10048, %rd10075; + xor.b64 %rd10077, %rd10076, %rd10046; + not.b64 %rd10078, %rd10048; + and.b64 %rd10079, %rd10020, %rd10078; + xor.b64 %rd10080, %rd10079, %rd10047; + not.b64 %rd10081, %rd10020; + and.b64 %rd10082, %rd10045, %rd10081; + xor.b64 %rd10083, %rd10082, %rd10048; + not.b64 %rd10084, %rd10050; + and.b64 %rd10085, %rd10051, %rd10084; + xor.b64 %rd10086, %rd10085, %rd10049; + not.b64 %rd10087, %rd10051; + and.b64 %rd10088, %rd10052, %rd10087; + xor.b64 %rd10089, %rd10088, %rd10050; + not.b64 %rd10090, %rd10052; + and.b64 %rd10091, %rd10053, %rd10090; + xor.b64 %rd10092, %rd10091, %rd10051; + not.b64 %rd10093, %rd10053; + and.b64 %rd10094, %rd10049, %rd10093; + xor.b64 %rd10095, %rd10094, %rd10052; + not.b64 %rd10096, %rd10049; + and.b64 %rd10097, %rd10050, %rd10096; + xor.b64 %rd10098, %rd10097, %rd10053; + not.b64 %rd10099, %rd10055; + and.b64 %rd10100, %rd10056, %rd10099; + xor.b64 %rd10101, %rd10100, %rd10054; + not.b64 %rd10102, %rd10056; + and.b64 %rd10103, %rd10057, %rd10102; + xor.b64 %rd10104, %rd10103, %rd10055; + not.b64 %rd10105, %rd10057; + and.b64 %rd10106, %rd10058, %rd10105; + xor.b64 %rd10107, %rd10106, %rd10056; + not.b64 %rd10108, %rd10058; + and.b64 %rd10109, %rd10054, %rd10108; + xor.b64 %rd10110, %rd10109, %rd10057; + not.b64 %rd10111, %rd10054; + and.b64 %rd10112, %rd10055, %rd10111; + xor.b64 %rd10113, %rd10112, %rd10058; + not.b64 %rd10114, %rd10060; + and.b64 %rd10115, %rd10061, %rd10114; + xor.b64 %rd10116, %rd10115, %rd10059; + not.b64 %rd10117, %rd10061; + and.b64 %rd10118, %rd10062, %rd10117; + xor.b64 %rd10119, %rd10118, %rd10060; + not.b64 %rd10120, %rd10062; + and.b64 %rd10121, %rd10063, %rd10120; + xor.b64 %rd10122, %rd10121, %rd10061; + not.b64 %rd10123, %rd10063; + and.b64 %rd10124, %rd10059, %rd10123; + xor.b64 %rd10125, %rd10124, %rd10062; + not.b64 %rd10126, %rd10059; + and.b64 %rd10127, %rd10060, %rd10126; + xor.b64 %rd10128, %rd10127, %rd10063; + not.b64 %rd10129, %rd10065; + and.b64 %rd10130, %rd10066, %rd10129; + xor.b64 %rd10131, %rd10130, %rd10064; + not.b64 %rd10132, %rd10066; + and.b64 %rd10133, %rd10067, %rd10132; + xor.b64 %rd10134, %rd10133, %rd10065; + not.b64 %rd10135, %rd10067; + and.b64 %rd10136, %rd10068, %rd10135; + xor.b64 %rd10137, %rd10136, %rd10066; + not.b64 %rd10138, %rd10068; + and.b64 %rd10139, %rd10064, %rd10138; + xor.b64 %rd10140, %rd10139, %rd10067; + not.b64 %rd10141, %rd10064; + and.b64 %rd10142, %rd10065, %rd10141; + xor.b64 %rd10143, %rd10142, %rd10068; + xor.b64 %rd10145, %rd10071, %rd28; + xor.b64 %rd10146, %rd10086, %rd10145; + xor.b64 %rd10147, %rd10146, %rd10101; + xor.b64 %rd10148, %rd10147, %rd10116; + xor.b64 %rd10149, %rd10148, %rd10131; + xor.b64 %rd10150, %rd10089, %rd10074; + xor.b64 %rd10151, %rd10150, %rd10104; + xor.b64 %rd10152, %rd10151, %rd10119; + xor.b64 %rd10153, %rd10152, %rd10134; + xor.b64 %rd10154, %rd10092, %rd10077; + xor.b64 %rd10155, %rd10154, %rd10107; + xor.b64 %rd10156, %rd10155, %rd10122; + xor.b64 %rd10157, %rd10156, %rd10137; + xor.b64 %rd10158, %rd10095, %rd10080; + xor.b64 %rd10159, %rd10158, %rd10110; + xor.b64 %rd10160, %rd10159, %rd10125; + xor.b64 %rd10161, %rd10160, %rd10140; + xor.b64 %rd10162, %rd10098, %rd10083; + xor.b64 %rd10163, %rd10162, %rd10113; + xor.b64 %rd10164, %rd10163, %rd10128; + xor.b64 %rd10165, %rd10164, %rd10143; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7154}, %rd10153; + } + { + .reg .b32 %dummy; + mov.b64 {%r7155,%dummy}, %rd10153; + } + shf.l.wrap.b32 %r7156, %r7155, %r7154, 1; + shf.l.wrap.b32 %r7157, %r7154, %r7155, 1; + mov.b64 %rd10166, {%r7157, %r7156}; + xor.b64 %rd10167, %rd10165, %rd10166; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7158}, %rd10157; + } + { + .reg .b32 %dummy; + mov.b64 {%r7159,%dummy}, %rd10157; + } + shf.l.wrap.b32 %r7160, %r7159, %r7158, 1; + shf.l.wrap.b32 %r7161, %r7158, %r7159, 1; + mov.b64 %rd10168, {%r7161, %r7160}; + xor.b64 %rd10169, %rd10168, %rd10149; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7162}, %rd10161; + } + { + .reg .b32 %dummy; + mov.b64 {%r7163,%dummy}, %rd10161; + } + shf.l.wrap.b32 %r7164, %r7163, %r7162, 1; + shf.l.wrap.b32 %r7165, %r7162, %r7163, 1; + mov.b64 %rd10170, {%r7165, %r7164}; + xor.b64 %rd10171, %rd10170, %rd10153; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7166}, %rd10165; + } + { + .reg .b32 %dummy; + mov.b64 {%r7167,%dummy}, %rd10165; + } + shf.l.wrap.b32 %r7168, %r7167, %r7166, 1; + shf.l.wrap.b32 %r7169, %r7166, %r7167, 1; + mov.b64 %rd10172, {%r7169, %r7168}; + xor.b64 %rd10173, %rd10172, %rd10157; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7170}, %rd10149; + } + { + .reg .b32 %dummy; + mov.b64 {%r7171,%dummy}, %rd10149; + } + shf.l.wrap.b32 %r7172, %r7171, %r7170, 1; + shf.l.wrap.b32 %r7173, %r7170, %r7171, 1; + mov.b64 %rd10174, {%r7173, %r7172}; + xor.b64 %rd10175, %rd10161, %rd10174; + xor.b64 %rd10176, %rd10167, %rd10145; + xor.b64 %rd10177, %rd10169, %rd10074; + xor.b64 %rd10178, %rd10171, %rd10077; + xor.b64 %rd10179, %rd10173, %rd10080; + xor.b64 %rd10180, %rd10175, %rd10083; + xor.b64 %rd10181, %rd10167, %rd10086; + xor.b64 %rd10182, %rd10169, %rd10089; + xor.b64 %rd10183, %rd10171, %rd10092; + xor.b64 %rd10184, %rd10173, %rd10095; + xor.b64 %rd10185, %rd10175, %rd10098; + xor.b64 %rd10186, %rd10167, %rd10101; + xor.b64 %rd10187, %rd10169, %rd10104; + xor.b64 %rd10188, %rd10171, %rd10107; + xor.b64 %rd10189, %rd10173, %rd10110; + xor.b64 %rd10190, %rd10175, %rd10113; + xor.b64 %rd10191, %rd10167, %rd10116; + xor.b64 %rd10192, %rd10169, %rd10119; + xor.b64 %rd10193, %rd10171, %rd10122; + xor.b64 %rd10194, %rd10173, %rd10125; + xor.b64 %rd10195, %rd10175, %rd10128; + xor.b64 %rd10196, %rd10167, %rd10131; + xor.b64 %rd10197, %rd10169, %rd10134; + xor.b64 %rd10198, %rd10171, %rd10137; + xor.b64 %rd10199, %rd10173, %rd10140; + xor.b64 %rd10200, %rd10175, %rd10143; + { + .reg .b32 %dummy; + mov.b64 {%r7174,%dummy}, %rd10182; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7175}, %rd10182; + } + shf.r.wrap.b32 %r7176, %r7175, %r7174, 20; + shf.r.wrap.b32 %r7177, %r7174, %r7175, 20; + mov.b64 %rd10201, {%r7177, %r7176}; + { + .reg .b32 %dummy; + mov.b64 {%r7178,%dummy}, %rd10188; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7179}, %rd10188; + } + shf.r.wrap.b32 %r7180, %r7179, %r7178, 21; + shf.r.wrap.b32 %r7181, %r7178, %r7179, 21; + mov.b64 %rd10202, {%r7181, %r7180}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7182}, %rd10194; + } + { + .reg .b32 %dummy; + mov.b64 {%r7183,%dummy}, %rd10194; + } + shf.l.wrap.b32 %r7184, %r7183, %r7182, 21; + shf.l.wrap.b32 %r7185, %r7182, %r7183, 21; + mov.b64 %rd10203, {%r7185, %r7184}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7186}, %rd10200; + } + { + .reg .b32 %dummy; + mov.b64 {%r7187,%dummy}, %rd10200; + } + shf.l.wrap.b32 %r7188, %r7187, %r7186, 14; + shf.l.wrap.b32 %r7189, %r7186, %r7187, 14; + mov.b64 %rd10204, {%r7189, %r7188}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7190}, %rd10179; + } + { + .reg .b32 %dummy; + mov.b64 {%r7191,%dummy}, %rd10179; + } + shf.l.wrap.b32 %r7192, %r7191, %r7190, 28; + shf.l.wrap.b32 %r7193, %r7190, %r7191, 28; + mov.b64 %rd10205, {%r7193, %r7192}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7194}, %rd10185; + } + { + .reg .b32 %dummy; + mov.b64 {%r7195,%dummy}, %rd10185; + } + shf.l.wrap.b32 %r7196, %r7195, %r7194, 20; + shf.l.wrap.b32 %r7197, %r7194, %r7195, 20; + mov.b64 %rd10206, {%r7197, %r7196}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7198}, %rd10186; + } + { + .reg .b32 %dummy; + mov.b64 {%r7199,%dummy}, %rd10186; + } + shf.l.wrap.b32 %r7200, %r7199, %r7198, 3; + shf.l.wrap.b32 %r7201, %r7198, %r7199, 3; + mov.b64 %rd10207, {%r7201, %r7200}; + { + .reg .b32 %dummy; + mov.b64 {%r7202,%dummy}, %rd10192; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7203}, %rd10192; + } + shf.r.wrap.b32 %r7204, %r7203, %r7202, 19; + shf.r.wrap.b32 %r7205, %r7202, %r7203, 19; + mov.b64 %rd10208, {%r7205, %r7204}; + { + .reg .b32 %dummy; + mov.b64 {%r7206,%dummy}, %rd10198; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7207}, %rd10198; + } + shf.r.wrap.b32 %r7208, %r7207, %r7206, 3; + shf.r.wrap.b32 %r7209, %r7206, %r7207, 3; + mov.b64 %rd10209, {%r7209, %r7208}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7210}, %rd10177; + } + { + .reg .b32 %dummy; + mov.b64 {%r7211,%dummy}, %rd10177; + } + shf.l.wrap.b32 %r7212, %r7211, %r7210, 1; + shf.l.wrap.b32 %r7213, %r7210, %r7211, 1; + mov.b64 %rd10210, {%r7213, %r7212}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7214}, %rd10183; + } + { + .reg .b32 %dummy; + mov.b64 {%r7215,%dummy}, %rd10183; + } + shf.l.wrap.b32 %r7216, %r7215, %r7214, 6; + shf.l.wrap.b32 %r7217, %r7214, %r7215, 6; + mov.b64 %rd10211, {%r7217, %r7216}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7218}, %rd10189; + } + { + .reg .b32 %dummy; + mov.b64 {%r7219,%dummy}, %rd10189; + } + shf.l.wrap.b32 %r7220, %r7219, %r7218, 25; + shf.l.wrap.b32 %r7221, %r7218, %r7219, 25; + mov.b64 %rd10212, {%r7221, %r7220}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7222}, %rd10195; + } + { + .reg .b32 %dummy; + mov.b64 {%r7223,%dummy}, %rd10195; + } + shf.l.wrap.b32 %r7224, %r7223, %r7222, 8; + shf.l.wrap.b32 %r7225, %r7222, %r7223, 8; + mov.b64 %rd10213, {%r7225, %r7224}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7226}, %rd10196; + } + { + .reg .b32 %dummy; + mov.b64 {%r7227,%dummy}, %rd10196; + } + shf.l.wrap.b32 %r7228, %r7227, %r7226, 18; + shf.l.wrap.b32 %r7229, %r7226, %r7227, 18; + mov.b64 %rd10214, {%r7229, %r7228}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7230}, %rd10180; + } + { + .reg .b32 %dummy; + mov.b64 {%r7231,%dummy}, %rd10180; + } + shf.l.wrap.b32 %r7232, %r7231, %r7230, 27; + shf.l.wrap.b32 %r7233, %r7230, %r7231, 27; + mov.b64 %rd10215, {%r7233, %r7232}; + { + .reg .b32 %dummy; + mov.b64 {%r7234,%dummy}, %rd10181; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7235}, %rd10181; + } + shf.r.wrap.b32 %r7236, %r7235, %r7234, 28; + shf.r.wrap.b32 %r7237, %r7234, %r7235, 28; + mov.b64 %rd10216, {%r7237, %r7236}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7238}, %rd10187; + } + { + .reg .b32 %dummy; + mov.b64 {%r7239,%dummy}, %rd10187; + } + shf.l.wrap.b32 %r7240, %r7239, %r7238, 10; + shf.l.wrap.b32 %r7241, %r7238, %r7239, 10; + mov.b64 %rd10217, {%r7241, %r7240}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7242}, %rd10193; + } + { + .reg .b32 %dummy; + mov.b64 {%r7243,%dummy}, %rd10193; + } + shf.l.wrap.b32 %r7244, %r7243, %r7242, 15; + shf.l.wrap.b32 %r7245, %r7242, %r7243, 15; + mov.b64 %rd10218, {%r7245, %r7244}; + { + .reg .b32 %dummy; + mov.b64 {%r7246,%dummy}, %rd10199; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7247}, %rd10199; + } + shf.r.wrap.b32 %r7248, %r7247, %r7246, 8; + shf.r.wrap.b32 %r7249, %r7246, %r7247, 8; + mov.b64 %rd10219, {%r7249, %r7248}; + { + .reg .b32 %dummy; + mov.b64 {%r7250,%dummy}, %rd10178; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7251}, %rd10178; + } + shf.r.wrap.b32 %r7252, %r7251, %r7250, 2; + shf.r.wrap.b32 %r7253, %r7250, %r7251, 2; + mov.b64 %rd10220, {%r7253, %r7252}; + { + .reg .b32 %dummy; + mov.b64 {%r7254,%dummy}, %rd10184; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7255}, %rd10184; + } + shf.r.wrap.b32 %r7256, %r7255, %r7254, 9; + shf.r.wrap.b32 %r7257, %r7254, %r7255, 9; + mov.b64 %rd10221, {%r7257, %r7256}; + { + .reg .b32 %dummy; + mov.b64 {%r7258,%dummy}, %rd10190; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7259}, %rd10190; + } + shf.r.wrap.b32 %r7260, %r7259, %r7258, 25; + shf.r.wrap.b32 %r7261, %r7258, %r7259, 25; + mov.b64 %rd10222, {%r7261, %r7260}; + { + .reg .b32 %dummy; + mov.b64 {%r7262,%dummy}, %rd10191; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7263}, %rd10191; + } + shf.r.wrap.b32 %r7264, %r7263, %r7262, 23; + shf.r.wrap.b32 %r7265, %r7262, %r7263, 23; + mov.b64 %rd10223, {%r7265, %r7264}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7266}, %rd10197; + } + { + .reg .b32 %dummy; + mov.b64 {%r7267,%dummy}, %rd10197; + } + shf.l.wrap.b32 %r7268, %r7267, %r7266, 2; + shf.l.wrap.b32 %r7269, %r7266, %r7267, 2; + mov.b64 %rd10224, {%r7269, %r7268}; + not.b64 %rd10225, %rd10201; + and.b64 %rd10226, %rd10202, %rd10225; + xor.b64 %rd10227, %rd10226, %rd10176; + not.b64 %rd10228, %rd10202; + and.b64 %rd10229, %rd10203, %rd10228; + xor.b64 %rd10230, %rd10229, %rd10201; + not.b64 %rd10231, %rd10203; + and.b64 %rd10232, %rd10204, %rd10231; + xor.b64 %rd10233, %rd10232, %rd10202; + not.b64 %rd10234, %rd10204; + and.b64 %rd10235, %rd10176, %rd10234; + xor.b64 %rd10236, %rd10235, %rd10203; + not.b64 %rd10237, %rd10176; + and.b64 %rd10238, %rd10201, %rd10237; + xor.b64 %rd10239, %rd10238, %rd10204; + not.b64 %rd10240, %rd10206; + and.b64 %rd10241, %rd10207, %rd10240; + xor.b64 %rd10242, %rd10241, %rd10205; + not.b64 %rd10243, %rd10207; + and.b64 %rd10244, %rd10208, %rd10243; + xor.b64 %rd10245, %rd10244, %rd10206; + not.b64 %rd10246, %rd10208; + and.b64 %rd10247, %rd10209, %rd10246; + xor.b64 %rd10248, %rd10247, %rd10207; + not.b64 %rd10249, %rd10209; + and.b64 %rd10250, %rd10205, %rd10249; + xor.b64 %rd10251, %rd10250, %rd10208; + not.b64 %rd10252, %rd10205; + and.b64 %rd10253, %rd10206, %rd10252; + xor.b64 %rd10254, %rd10253, %rd10209; + not.b64 %rd10255, %rd10211; + and.b64 %rd10256, %rd10212, %rd10255; + xor.b64 %rd10257, %rd10256, %rd10210; + not.b64 %rd10258, %rd10212; + and.b64 %rd10259, %rd10213, %rd10258; + xor.b64 %rd10260, %rd10259, %rd10211; + not.b64 %rd10261, %rd10213; + and.b64 %rd10262, %rd10214, %rd10261; + xor.b64 %rd10263, %rd10262, %rd10212; + not.b64 %rd10264, %rd10214; + and.b64 %rd10265, %rd10210, %rd10264; + xor.b64 %rd10266, %rd10265, %rd10213; + not.b64 %rd10267, %rd10210; + and.b64 %rd10268, %rd10211, %rd10267; + xor.b64 %rd10269, %rd10268, %rd10214; + not.b64 %rd10270, %rd10216; + and.b64 %rd10271, %rd10217, %rd10270; + xor.b64 %rd10272, %rd10271, %rd10215; + not.b64 %rd10273, %rd10217; + and.b64 %rd10274, %rd10218, %rd10273; + xor.b64 %rd10275, %rd10274, %rd10216; + not.b64 %rd10276, %rd10218; + and.b64 %rd10277, %rd10219, %rd10276; + xor.b64 %rd10278, %rd10277, %rd10217; + not.b64 %rd10279, %rd10219; + and.b64 %rd10280, %rd10215, %rd10279; + xor.b64 %rd10281, %rd10280, %rd10218; + not.b64 %rd10282, %rd10215; + and.b64 %rd10283, %rd10216, %rd10282; + xor.b64 %rd10284, %rd10283, %rd10219; + not.b64 %rd10285, %rd10221; + and.b64 %rd10286, %rd10222, %rd10285; + xor.b64 %rd10287, %rd10286, %rd10220; + not.b64 %rd10288, %rd10222; + and.b64 %rd10289, %rd10223, %rd10288; + xor.b64 %rd10290, %rd10289, %rd10221; + not.b64 %rd10291, %rd10223; + and.b64 %rd10292, %rd10224, %rd10291; + xor.b64 %rd10293, %rd10292, %rd10222; + not.b64 %rd10294, %rd10224; + and.b64 %rd10295, %rd10220, %rd10294; + xor.b64 %rd10296, %rd10295, %rd10223; + not.b64 %rd10297, %rd10220; + and.b64 %rd10298, %rd10221, %rd10297; + xor.b64 %rd10299, %rd10298, %rd10224; + xor.b64 %rd10301, %rd10227, %rd29; + xor.b64 %rd10302, %rd10242, %rd10301; + xor.b64 %rd10303, %rd10302, %rd10257; + xor.b64 %rd10304, %rd10303, %rd10272; + xor.b64 %rd10305, %rd10304, %rd10287; + xor.b64 %rd10306, %rd10245, %rd10230; + xor.b64 %rd10307, %rd10306, %rd10260; + xor.b64 %rd10308, %rd10307, %rd10275; + xor.b64 %rd10309, %rd10308, %rd10290; + xor.b64 %rd10310, %rd10248, %rd10233; + xor.b64 %rd10311, %rd10310, %rd10263; + xor.b64 %rd10312, %rd10311, %rd10278; + xor.b64 %rd10313, %rd10312, %rd10293; + xor.b64 %rd10314, %rd10251, %rd10236; + xor.b64 %rd10315, %rd10314, %rd10266; + xor.b64 %rd10316, %rd10315, %rd10281; + xor.b64 %rd10317, %rd10316, %rd10296; + xor.b64 %rd10318, %rd10254, %rd10239; + xor.b64 %rd10319, %rd10318, %rd10269; + xor.b64 %rd10320, %rd10319, %rd10284; + xor.b64 %rd10321, %rd10320, %rd10299; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7270}, %rd10309; + } + { + .reg .b32 %dummy; + mov.b64 {%r7271,%dummy}, %rd10309; + } + shf.l.wrap.b32 %r7272, %r7271, %r7270, 1; + shf.l.wrap.b32 %r7273, %r7270, %r7271, 1; + mov.b64 %rd10322, {%r7273, %r7272}; + xor.b64 %rd10323, %rd10321, %rd10322; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7274}, %rd10313; + } + { + .reg .b32 %dummy; + mov.b64 {%r7275,%dummy}, %rd10313; + } + shf.l.wrap.b32 %r7276, %r7275, %r7274, 1; + shf.l.wrap.b32 %r7277, %r7274, %r7275, 1; + mov.b64 %rd10324, {%r7277, %r7276}; + xor.b64 %rd10325, %rd10324, %rd10305; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7278}, %rd10317; + } + { + .reg .b32 %dummy; + mov.b64 {%r7279,%dummy}, %rd10317; + } + shf.l.wrap.b32 %r7280, %r7279, %r7278, 1; + shf.l.wrap.b32 %r7281, %r7278, %r7279, 1; + mov.b64 %rd10326, {%r7281, %r7280}; + xor.b64 %rd10327, %rd10326, %rd10309; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7282}, %rd10321; + } + { + .reg .b32 %dummy; + mov.b64 {%r7283,%dummy}, %rd10321; + } + shf.l.wrap.b32 %r7284, %r7283, %r7282, 1; + shf.l.wrap.b32 %r7285, %r7282, %r7283, 1; + mov.b64 %rd10328, {%r7285, %r7284}; + xor.b64 %rd10329, %rd10328, %rd10313; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7286}, %rd10305; + } + { + .reg .b32 %dummy; + mov.b64 {%r7287,%dummy}, %rd10305; + } + shf.l.wrap.b32 %r7288, %r7287, %r7286, 1; + shf.l.wrap.b32 %r7289, %r7286, %r7287, 1; + mov.b64 %rd10330, {%r7289, %r7288}; + xor.b64 %rd10331, %rd10317, %rd10330; + xor.b64 %rd10332, %rd10323, %rd10301; + xor.b64 %rd10333, %rd10325, %rd10230; + xor.b64 %rd10334, %rd10327, %rd10233; + xor.b64 %rd10335, %rd10329, %rd10236; + xor.b64 %rd10336, %rd10331, %rd10239; + xor.b64 %rd10337, %rd10323, %rd10242; + xor.b64 %rd10338, %rd10325, %rd10245; + xor.b64 %rd10339, %rd10327, %rd10248; + xor.b64 %rd10340, %rd10329, %rd10251; + xor.b64 %rd10341, %rd10331, %rd10254; + xor.b64 %rd10342, %rd10323, %rd10257; + xor.b64 %rd10343, %rd10325, %rd10260; + xor.b64 %rd10344, %rd10327, %rd10263; + xor.b64 %rd10345, %rd10329, %rd10266; + xor.b64 %rd10346, %rd10331, %rd10269; + xor.b64 %rd10347, %rd10323, %rd10272; + xor.b64 %rd10348, %rd10325, %rd10275; + xor.b64 %rd10349, %rd10327, %rd10278; + xor.b64 %rd10350, %rd10329, %rd10281; + xor.b64 %rd10351, %rd10331, %rd10284; + xor.b64 %rd10352, %rd10323, %rd10287; + xor.b64 %rd10353, %rd10325, %rd10290; + xor.b64 %rd10354, %rd10327, %rd10293; + xor.b64 %rd10355, %rd10329, %rd10296; + xor.b64 %rd10356, %rd10331, %rd10299; + { + .reg .b32 %dummy; + mov.b64 {%r7290,%dummy}, %rd10338; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7291}, %rd10338; + } + shf.r.wrap.b32 %r7292, %r7291, %r7290, 20; + shf.r.wrap.b32 %r7293, %r7290, %r7291, 20; + mov.b64 %rd10357, {%r7293, %r7292}; + { + .reg .b32 %dummy; + mov.b64 {%r7294,%dummy}, %rd10344; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7295}, %rd10344; + } + shf.r.wrap.b32 %r7296, %r7295, %r7294, 21; + shf.r.wrap.b32 %r7297, %r7294, %r7295, 21; + mov.b64 %rd10358, {%r7297, %r7296}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7298}, %rd10350; + } + { + .reg .b32 %dummy; + mov.b64 {%r7299,%dummy}, %rd10350; + } + shf.l.wrap.b32 %r7300, %r7299, %r7298, 21; + shf.l.wrap.b32 %r7301, %r7298, %r7299, 21; + mov.b64 %rd10359, {%r7301, %r7300}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7302}, %rd10356; + } + { + .reg .b32 %dummy; + mov.b64 {%r7303,%dummy}, %rd10356; + } + shf.l.wrap.b32 %r7304, %r7303, %r7302, 14; + shf.l.wrap.b32 %r7305, %r7302, %r7303, 14; + mov.b64 %rd10360, {%r7305, %r7304}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7306}, %rd10335; + } + { + .reg .b32 %dummy; + mov.b64 {%r7307,%dummy}, %rd10335; + } + shf.l.wrap.b32 %r7308, %r7307, %r7306, 28; + shf.l.wrap.b32 %r7309, %r7306, %r7307, 28; + mov.b64 %rd10361, {%r7309, %r7308}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7310}, %rd10341; + } + { + .reg .b32 %dummy; + mov.b64 {%r7311,%dummy}, %rd10341; + } + shf.l.wrap.b32 %r7312, %r7311, %r7310, 20; + shf.l.wrap.b32 %r7313, %r7310, %r7311, 20; + mov.b64 %rd10362, {%r7313, %r7312}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7314}, %rd10342; + } + { + .reg .b32 %dummy; + mov.b64 {%r7315,%dummy}, %rd10342; + } + shf.l.wrap.b32 %r7316, %r7315, %r7314, 3; + shf.l.wrap.b32 %r7317, %r7314, %r7315, 3; + mov.b64 %rd10363, {%r7317, %r7316}; + { + .reg .b32 %dummy; + mov.b64 {%r7318,%dummy}, %rd10348; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7319}, %rd10348; + } + shf.r.wrap.b32 %r7320, %r7319, %r7318, 19; + shf.r.wrap.b32 %r7321, %r7318, %r7319, 19; + mov.b64 %rd10364, {%r7321, %r7320}; + { + .reg .b32 %dummy; + mov.b64 {%r7322,%dummy}, %rd10354; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7323}, %rd10354; + } + shf.r.wrap.b32 %r7324, %r7323, %r7322, 3; + shf.r.wrap.b32 %r7325, %r7322, %r7323, 3; + mov.b64 %rd10365, {%r7325, %r7324}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7326}, %rd10333; + } + { + .reg .b32 %dummy; + mov.b64 {%r7327,%dummy}, %rd10333; + } + shf.l.wrap.b32 %r7328, %r7327, %r7326, 1; + shf.l.wrap.b32 %r7329, %r7326, %r7327, 1; + mov.b64 %rd10366, {%r7329, %r7328}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7330}, %rd10339; + } + { + .reg .b32 %dummy; + mov.b64 {%r7331,%dummy}, %rd10339; + } + shf.l.wrap.b32 %r7332, %r7331, %r7330, 6; + shf.l.wrap.b32 %r7333, %r7330, %r7331, 6; + mov.b64 %rd10367, {%r7333, %r7332}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7334}, %rd10345; + } + { + .reg .b32 %dummy; + mov.b64 {%r7335,%dummy}, %rd10345; + } + shf.l.wrap.b32 %r7336, %r7335, %r7334, 25; + shf.l.wrap.b32 %r7337, %r7334, %r7335, 25; + mov.b64 %rd10368, {%r7337, %r7336}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7338}, %rd10351; + } + { + .reg .b32 %dummy; + mov.b64 {%r7339,%dummy}, %rd10351; + } + shf.l.wrap.b32 %r7340, %r7339, %r7338, 8; + shf.l.wrap.b32 %r7341, %r7338, %r7339, 8; + mov.b64 %rd10369, {%r7341, %r7340}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7342}, %rd10352; + } + { + .reg .b32 %dummy; + mov.b64 {%r7343,%dummy}, %rd10352; + } + shf.l.wrap.b32 %r7344, %r7343, %r7342, 18; + shf.l.wrap.b32 %r7345, %r7342, %r7343, 18; + mov.b64 %rd10370, {%r7345, %r7344}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7346}, %rd10336; + } + { + .reg .b32 %dummy; + mov.b64 {%r7347,%dummy}, %rd10336; + } + shf.l.wrap.b32 %r7348, %r7347, %r7346, 27; + shf.l.wrap.b32 %r7349, %r7346, %r7347, 27; + mov.b64 %rd10371, {%r7349, %r7348}; + { + .reg .b32 %dummy; + mov.b64 {%r7350,%dummy}, %rd10337; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7351}, %rd10337; + } + shf.r.wrap.b32 %r7352, %r7351, %r7350, 28; + shf.r.wrap.b32 %r7353, %r7350, %r7351, 28; + mov.b64 %rd10372, {%r7353, %r7352}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7354}, %rd10343; + } + { + .reg .b32 %dummy; + mov.b64 {%r7355,%dummy}, %rd10343; + } + shf.l.wrap.b32 %r7356, %r7355, %r7354, 10; + shf.l.wrap.b32 %r7357, %r7354, %r7355, 10; + mov.b64 %rd10373, {%r7357, %r7356}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7358}, %rd10349; + } + { + .reg .b32 %dummy; + mov.b64 {%r7359,%dummy}, %rd10349; + } + shf.l.wrap.b32 %r7360, %r7359, %r7358, 15; + shf.l.wrap.b32 %r7361, %r7358, %r7359, 15; + mov.b64 %rd10374, {%r7361, %r7360}; + { + .reg .b32 %dummy; + mov.b64 {%r7362,%dummy}, %rd10355; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7363}, %rd10355; + } + shf.r.wrap.b32 %r7364, %r7363, %r7362, 8; + shf.r.wrap.b32 %r7365, %r7362, %r7363, 8; + mov.b64 %rd10375, {%r7365, %r7364}; + { + .reg .b32 %dummy; + mov.b64 {%r7366,%dummy}, %rd10334; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7367}, %rd10334; + } + shf.r.wrap.b32 %r7368, %r7367, %r7366, 2; + shf.r.wrap.b32 %r7369, %r7366, %r7367, 2; + mov.b64 %rd10376, {%r7369, %r7368}; + { + .reg .b32 %dummy; + mov.b64 {%r7370,%dummy}, %rd10340; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7371}, %rd10340; + } + shf.r.wrap.b32 %r7372, %r7371, %r7370, 9; + shf.r.wrap.b32 %r7373, %r7370, %r7371, 9; + mov.b64 %rd10377, {%r7373, %r7372}; + { + .reg .b32 %dummy; + mov.b64 {%r7374,%dummy}, %rd10346; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7375}, %rd10346; + } + shf.r.wrap.b32 %r7376, %r7375, %r7374, 25; + shf.r.wrap.b32 %r7377, %r7374, %r7375, 25; + mov.b64 %rd10378, {%r7377, %r7376}; + { + .reg .b32 %dummy; + mov.b64 {%r7378,%dummy}, %rd10347; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7379}, %rd10347; + } + shf.r.wrap.b32 %r7380, %r7379, %r7378, 23; + shf.r.wrap.b32 %r7381, %r7378, %r7379, 23; + mov.b64 %rd10379, {%r7381, %r7380}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7382}, %rd10353; + } + { + .reg .b32 %dummy; + mov.b64 {%r7383,%dummy}, %rd10353; + } + shf.l.wrap.b32 %r7384, %r7383, %r7382, 2; + shf.l.wrap.b32 %r7385, %r7382, %r7383, 2; + mov.b64 %rd10380, {%r7385, %r7384}; + not.b64 %rd10381, %rd10357; + and.b64 %rd10382, %rd10358, %rd10381; + xor.b64 %rd10383, %rd10382, %rd10332; + not.b64 %rd10384, %rd10358; + and.b64 %rd10385, %rd10359, %rd10384; + xor.b64 %rd10386, %rd10385, %rd10357; + not.b64 %rd10387, %rd10359; + and.b64 %rd10388, %rd10360, %rd10387; + xor.b64 %rd10389, %rd10388, %rd10358; + not.b64 %rd10390, %rd10360; + and.b64 %rd10391, %rd10332, %rd10390; + xor.b64 %rd10392, %rd10391, %rd10359; + not.b64 %rd10393, %rd10332; + and.b64 %rd10394, %rd10357, %rd10393; + xor.b64 %rd10395, %rd10394, %rd10360; + not.b64 %rd10396, %rd10362; + and.b64 %rd10397, %rd10363, %rd10396; + xor.b64 %rd10398, %rd10397, %rd10361; + not.b64 %rd10399, %rd10363; + and.b64 %rd10400, %rd10364, %rd10399; + xor.b64 %rd10401, %rd10400, %rd10362; + not.b64 %rd10402, %rd10364; + and.b64 %rd10403, %rd10365, %rd10402; + xor.b64 %rd10404, %rd10403, %rd10363; + not.b64 %rd10405, %rd10365; + and.b64 %rd10406, %rd10361, %rd10405; + xor.b64 %rd10407, %rd10406, %rd10364; + not.b64 %rd10408, %rd10361; + and.b64 %rd10409, %rd10362, %rd10408; + xor.b64 %rd10410, %rd10409, %rd10365; + not.b64 %rd10411, %rd10367; + and.b64 %rd10412, %rd10368, %rd10411; + xor.b64 %rd10413, %rd10412, %rd10366; + not.b64 %rd10414, %rd10368; + and.b64 %rd10415, %rd10369, %rd10414; + xor.b64 %rd10416, %rd10415, %rd10367; + not.b64 %rd10417, %rd10369; + and.b64 %rd10418, %rd10370, %rd10417; + xor.b64 %rd10419, %rd10418, %rd10368; + not.b64 %rd10420, %rd10370; + and.b64 %rd10421, %rd10366, %rd10420; + xor.b64 %rd10422, %rd10421, %rd10369; + not.b64 %rd10423, %rd10366; + and.b64 %rd10424, %rd10367, %rd10423; + xor.b64 %rd10425, %rd10424, %rd10370; + not.b64 %rd10426, %rd10372; + and.b64 %rd10427, %rd10373, %rd10426; + xor.b64 %rd10428, %rd10427, %rd10371; + not.b64 %rd10429, %rd10373; + and.b64 %rd10430, %rd10374, %rd10429; + xor.b64 %rd10431, %rd10430, %rd10372; + not.b64 %rd10432, %rd10374; + and.b64 %rd10433, %rd10375, %rd10432; + xor.b64 %rd10434, %rd10433, %rd10373; + not.b64 %rd10435, %rd10375; + and.b64 %rd10436, %rd10371, %rd10435; + xor.b64 %rd10437, %rd10436, %rd10374; + not.b64 %rd10438, %rd10371; + and.b64 %rd10439, %rd10372, %rd10438; + xor.b64 %rd10440, %rd10439, %rd10375; + not.b64 %rd10441, %rd10377; + and.b64 %rd10442, %rd10378, %rd10441; + xor.b64 %rd10443, %rd10442, %rd10376; + not.b64 %rd10444, %rd10378; + and.b64 %rd10445, %rd10379, %rd10444; + xor.b64 %rd10446, %rd10445, %rd10377; + not.b64 %rd10447, %rd10379; + and.b64 %rd10448, %rd10380, %rd10447; + xor.b64 %rd10449, %rd10448, %rd10378; + not.b64 %rd10450, %rd10380; + and.b64 %rd10451, %rd10376, %rd10450; + xor.b64 %rd10452, %rd10451, %rd10379; + not.b64 %rd10453, %rd10376; + and.b64 %rd10454, %rd10377, %rd10453; + xor.b64 %rd10455, %rd10454, %rd10380; + xor.b64 %rd10457, %rd10383, %rd30; + xor.b64 %rd10458, %rd10398, %rd10457; + xor.b64 %rd10459, %rd10458, %rd10413; + xor.b64 %rd10460, %rd10459, %rd10428; + xor.b64 %rd10461, %rd10460, %rd10443; + xor.b64 %rd10462, %rd10401, %rd10386; + xor.b64 %rd10463, %rd10462, %rd10416; + xor.b64 %rd10464, %rd10463, %rd10431; + xor.b64 %rd10465, %rd10464, %rd10446; + xor.b64 %rd10466, %rd10404, %rd10389; + xor.b64 %rd10467, %rd10466, %rd10419; + xor.b64 %rd10468, %rd10467, %rd10434; + xor.b64 %rd10469, %rd10468, %rd10449; + xor.b64 %rd10470, %rd10407, %rd10392; + xor.b64 %rd10471, %rd10470, %rd10422; + xor.b64 %rd10472, %rd10471, %rd10437; + xor.b64 %rd10473, %rd10472, %rd10452; + xor.b64 %rd10474, %rd10410, %rd10395; + xor.b64 %rd10475, %rd10474, %rd10425; + xor.b64 %rd10476, %rd10475, %rd10440; + xor.b64 %rd10477, %rd10476, %rd10455; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7386}, %rd10465; + } + { + .reg .b32 %dummy; + mov.b64 {%r7387,%dummy}, %rd10465; + } + shf.l.wrap.b32 %r7388, %r7387, %r7386, 1; + shf.l.wrap.b32 %r7389, %r7386, %r7387, 1; + mov.b64 %rd10478, {%r7389, %r7388}; + xor.b64 %rd10479, %rd10477, %rd10478; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7390}, %rd10469; + } + { + .reg .b32 %dummy; + mov.b64 {%r7391,%dummy}, %rd10469; + } + shf.l.wrap.b32 %r7392, %r7391, %r7390, 1; + shf.l.wrap.b32 %r7393, %r7390, %r7391, 1; + mov.b64 %rd10480, {%r7393, %r7392}; + xor.b64 %rd10481, %rd10480, %rd10461; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7394}, %rd10473; + } + { + .reg .b32 %dummy; + mov.b64 {%r7395,%dummy}, %rd10473; + } + shf.l.wrap.b32 %r7396, %r7395, %r7394, 1; + shf.l.wrap.b32 %r7397, %r7394, %r7395, 1; + mov.b64 %rd10482, {%r7397, %r7396}; + xor.b64 %rd10483, %rd10482, %rd10465; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7398}, %rd10477; + } + { + .reg .b32 %dummy; + mov.b64 {%r7399,%dummy}, %rd10477; + } + shf.l.wrap.b32 %r7400, %r7399, %r7398, 1; + shf.l.wrap.b32 %r7401, %r7398, %r7399, 1; + mov.b64 %rd10484, {%r7401, %r7400}; + xor.b64 %rd10485, %rd10484, %rd10469; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7402}, %rd10461; + } + { + .reg .b32 %dummy; + mov.b64 {%r7403,%dummy}, %rd10461; + } + shf.l.wrap.b32 %r7404, %r7403, %r7402, 1; + shf.l.wrap.b32 %r7405, %r7402, %r7403, 1; + mov.b64 %rd10486, {%r7405, %r7404}; + xor.b64 %rd10487, %rd10473, %rd10486; + xor.b64 %rd10488, %rd10479, %rd10457; + xor.b64 %rd10489, %rd10481, %rd10386; + xor.b64 %rd10490, %rd10483, %rd10389; + xor.b64 %rd10491, %rd10485, %rd10392; + xor.b64 %rd10492, %rd10487, %rd10395; + xor.b64 %rd10493, %rd10479, %rd10398; + xor.b64 %rd10494, %rd10481, %rd10401; + xor.b64 %rd10495, %rd10483, %rd10404; + xor.b64 %rd10496, %rd10485, %rd10407; + xor.b64 %rd10497, %rd10487, %rd10410; + xor.b64 %rd10498, %rd10479, %rd10413; + xor.b64 %rd10499, %rd10481, %rd10416; + xor.b64 %rd10500, %rd10483, %rd10419; + xor.b64 %rd10501, %rd10485, %rd10422; + xor.b64 %rd10502, %rd10487, %rd10425; + xor.b64 %rd10503, %rd10479, %rd10428; + xor.b64 %rd10504, %rd10481, %rd10431; + xor.b64 %rd10505, %rd10483, %rd10434; + xor.b64 %rd10506, %rd10485, %rd10437; + xor.b64 %rd10507, %rd10487, %rd10440; + xor.b64 %rd10508, %rd10479, %rd10443; + xor.b64 %rd10509, %rd10481, %rd10446; + xor.b64 %rd10510, %rd10483, %rd10449; + xor.b64 %rd10511, %rd10485, %rd10452; + xor.b64 %rd10512, %rd10487, %rd10455; + { + .reg .b32 %dummy; + mov.b64 {%r7406,%dummy}, %rd10494; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7407}, %rd10494; + } + shf.r.wrap.b32 %r7408, %r7407, %r7406, 20; + shf.r.wrap.b32 %r7409, %r7406, %r7407, 20; + mov.b64 %rd10513, {%r7409, %r7408}; + { + .reg .b32 %dummy; + mov.b64 {%r7410,%dummy}, %rd10500; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7411}, %rd10500; + } + shf.r.wrap.b32 %r7412, %r7411, %r7410, 21; + shf.r.wrap.b32 %r7413, %r7410, %r7411, 21; + mov.b64 %rd10514, {%r7413, %r7412}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7414}, %rd10506; + } + { + .reg .b32 %dummy; + mov.b64 {%r7415,%dummy}, %rd10506; + } + shf.l.wrap.b32 %r7416, %r7415, %r7414, 21; + shf.l.wrap.b32 %r7417, %r7414, %r7415, 21; + mov.b64 %rd10515, {%r7417, %r7416}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7418}, %rd10512; + } + { + .reg .b32 %dummy; + mov.b64 {%r7419,%dummy}, %rd10512; + } + shf.l.wrap.b32 %r7420, %r7419, %r7418, 14; + shf.l.wrap.b32 %r7421, %r7418, %r7419, 14; + mov.b64 %rd10516, {%r7421, %r7420}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7422}, %rd10491; + } + { + .reg .b32 %dummy; + mov.b64 {%r7423,%dummy}, %rd10491; + } + shf.l.wrap.b32 %r7424, %r7423, %r7422, 28; + shf.l.wrap.b32 %r7425, %r7422, %r7423, 28; + mov.b64 %rd10517, {%r7425, %r7424}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7426}, %rd10497; + } + { + .reg .b32 %dummy; + mov.b64 {%r7427,%dummy}, %rd10497; + } + shf.l.wrap.b32 %r7428, %r7427, %r7426, 20; + shf.l.wrap.b32 %r7429, %r7426, %r7427, 20; + mov.b64 %rd10518, {%r7429, %r7428}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7430}, %rd10498; + } + { + .reg .b32 %dummy; + mov.b64 {%r7431,%dummy}, %rd10498; + } + shf.l.wrap.b32 %r7432, %r7431, %r7430, 3; + shf.l.wrap.b32 %r7433, %r7430, %r7431, 3; + mov.b64 %rd10519, {%r7433, %r7432}; + { + .reg .b32 %dummy; + mov.b64 {%r7434,%dummy}, %rd10504; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7435}, %rd10504; + } + shf.r.wrap.b32 %r7436, %r7435, %r7434, 19; + shf.r.wrap.b32 %r7437, %r7434, %r7435, 19; + mov.b64 %rd10520, {%r7437, %r7436}; + { + .reg .b32 %dummy; + mov.b64 {%r7438,%dummy}, %rd10510; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7439}, %rd10510; + } + shf.r.wrap.b32 %r7440, %r7439, %r7438, 3; + shf.r.wrap.b32 %r7441, %r7438, %r7439, 3; + mov.b64 %rd10521, {%r7441, %r7440}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7442}, %rd10489; + } + { + .reg .b32 %dummy; + mov.b64 {%r7443,%dummy}, %rd10489; + } + shf.l.wrap.b32 %r7444, %r7443, %r7442, 1; + shf.l.wrap.b32 %r7445, %r7442, %r7443, 1; + mov.b64 %rd10522, {%r7445, %r7444}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7446}, %rd10495; + } + { + .reg .b32 %dummy; + mov.b64 {%r7447,%dummy}, %rd10495; + } + shf.l.wrap.b32 %r7448, %r7447, %r7446, 6; + shf.l.wrap.b32 %r7449, %r7446, %r7447, 6; + mov.b64 %rd10523, {%r7449, %r7448}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7450}, %rd10501; + } + { + .reg .b32 %dummy; + mov.b64 {%r7451,%dummy}, %rd10501; + } + shf.l.wrap.b32 %r7452, %r7451, %r7450, 25; + shf.l.wrap.b32 %r7453, %r7450, %r7451, 25; + mov.b64 %rd10524, {%r7453, %r7452}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7454}, %rd10507; + } + { + .reg .b32 %dummy; + mov.b64 {%r7455,%dummy}, %rd10507; + } + shf.l.wrap.b32 %r7456, %r7455, %r7454, 8; + shf.l.wrap.b32 %r7457, %r7454, %r7455, 8; + mov.b64 %rd10525, {%r7457, %r7456}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7458}, %rd10508; + } + { + .reg .b32 %dummy; + mov.b64 {%r7459,%dummy}, %rd10508; + } + shf.l.wrap.b32 %r7460, %r7459, %r7458, 18; + shf.l.wrap.b32 %r7461, %r7458, %r7459, 18; + mov.b64 %rd10526, {%r7461, %r7460}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7462}, %rd10492; + } + { + .reg .b32 %dummy; + mov.b64 {%r7463,%dummy}, %rd10492; + } + shf.l.wrap.b32 %r7464, %r7463, %r7462, 27; + shf.l.wrap.b32 %r7465, %r7462, %r7463, 27; + mov.b64 %rd10527, {%r7465, %r7464}; + { + .reg .b32 %dummy; + mov.b64 {%r7466,%dummy}, %rd10493; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7467}, %rd10493; + } + shf.r.wrap.b32 %r7468, %r7467, %r7466, 28; + shf.r.wrap.b32 %r7469, %r7466, %r7467, 28; + mov.b64 %rd10528, {%r7469, %r7468}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7470}, %rd10499; + } + { + .reg .b32 %dummy; + mov.b64 {%r7471,%dummy}, %rd10499; + } + shf.l.wrap.b32 %r7472, %r7471, %r7470, 10; + shf.l.wrap.b32 %r7473, %r7470, %r7471, 10; + mov.b64 %rd10529, {%r7473, %r7472}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7474}, %rd10505; + } + { + .reg .b32 %dummy; + mov.b64 {%r7475,%dummy}, %rd10505; + } + shf.l.wrap.b32 %r7476, %r7475, %r7474, 15; + shf.l.wrap.b32 %r7477, %r7474, %r7475, 15; + mov.b64 %rd10530, {%r7477, %r7476}; + { + .reg .b32 %dummy; + mov.b64 {%r7478,%dummy}, %rd10511; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7479}, %rd10511; + } + shf.r.wrap.b32 %r7480, %r7479, %r7478, 8; + shf.r.wrap.b32 %r7481, %r7478, %r7479, 8; + mov.b64 %rd10531, {%r7481, %r7480}; + { + .reg .b32 %dummy; + mov.b64 {%r7482,%dummy}, %rd10490; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7483}, %rd10490; + } + shf.r.wrap.b32 %r7484, %r7483, %r7482, 2; + shf.r.wrap.b32 %r7485, %r7482, %r7483, 2; + mov.b64 %rd10532, {%r7485, %r7484}; + { + .reg .b32 %dummy; + mov.b64 {%r7486,%dummy}, %rd10496; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7487}, %rd10496; + } + shf.r.wrap.b32 %r7488, %r7487, %r7486, 9; + shf.r.wrap.b32 %r7489, %r7486, %r7487, 9; + mov.b64 %rd10533, {%r7489, %r7488}; + { + .reg .b32 %dummy; + mov.b64 {%r7490,%dummy}, %rd10502; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7491}, %rd10502; + } + shf.r.wrap.b32 %r7492, %r7491, %r7490, 25; + shf.r.wrap.b32 %r7493, %r7490, %r7491, 25; + mov.b64 %rd10534, {%r7493, %r7492}; + { + .reg .b32 %dummy; + mov.b64 {%r7494,%dummy}, %rd10503; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7495}, %rd10503; + } + shf.r.wrap.b32 %r7496, %r7495, %r7494, 23; + shf.r.wrap.b32 %r7497, %r7494, %r7495, 23; + mov.b64 %rd10535, {%r7497, %r7496}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7498}, %rd10509; + } + { + .reg .b32 %dummy; + mov.b64 {%r7499,%dummy}, %rd10509; + } + shf.l.wrap.b32 %r7500, %r7499, %r7498, 2; + shf.l.wrap.b32 %r7501, %r7498, %r7499, 2; + mov.b64 %rd10536, {%r7501, %r7500}; + not.b64 %rd10537, %rd10513; + and.b64 %rd10538, %rd10514, %rd10537; + xor.b64 %rd10539, %rd10538, %rd10488; + not.b64 %rd10540, %rd10514; + and.b64 %rd10541, %rd10515, %rd10540; + xor.b64 %rd10542, %rd10541, %rd10513; + not.b64 %rd10543, %rd10515; + and.b64 %rd10544, %rd10516, %rd10543; + xor.b64 %rd10545, %rd10544, %rd10514; + not.b64 %rd10546, %rd10516; + and.b64 %rd10547, %rd10488, %rd10546; + xor.b64 %rd10548, %rd10547, %rd10515; + not.b64 %rd10549, %rd10488; + and.b64 %rd10550, %rd10513, %rd10549; + xor.b64 %rd10551, %rd10550, %rd10516; + not.b64 %rd10552, %rd10518; + and.b64 %rd10553, %rd10519, %rd10552; + xor.b64 %rd10554, %rd10553, %rd10517; + not.b64 %rd10555, %rd10519; + and.b64 %rd10556, %rd10520, %rd10555; + xor.b64 %rd10557, %rd10556, %rd10518; + not.b64 %rd10558, %rd10520; + and.b64 %rd10559, %rd10521, %rd10558; + xor.b64 %rd10560, %rd10559, %rd10519; + not.b64 %rd10561, %rd10521; + and.b64 %rd10562, %rd10517, %rd10561; + xor.b64 %rd10563, %rd10562, %rd10520; + not.b64 %rd10564, %rd10517; + and.b64 %rd10565, %rd10518, %rd10564; + xor.b64 %rd10566, %rd10565, %rd10521; + not.b64 %rd10567, %rd10523; + and.b64 %rd10568, %rd10524, %rd10567; + xor.b64 %rd10569, %rd10568, %rd10522; + not.b64 %rd10570, %rd10524; + and.b64 %rd10571, %rd10525, %rd10570; + xor.b64 %rd10572, %rd10571, %rd10523; + not.b64 %rd10573, %rd10525; + and.b64 %rd10574, %rd10526, %rd10573; + xor.b64 %rd10575, %rd10574, %rd10524; + not.b64 %rd10576, %rd10526; + and.b64 %rd10577, %rd10522, %rd10576; + xor.b64 %rd10578, %rd10577, %rd10525; + not.b64 %rd10579, %rd10522; + and.b64 %rd10580, %rd10523, %rd10579; + xor.b64 %rd10581, %rd10580, %rd10526; + not.b64 %rd10582, %rd10528; + and.b64 %rd10583, %rd10529, %rd10582; + xor.b64 %rd10584, %rd10583, %rd10527; + not.b64 %rd10585, %rd10529; + and.b64 %rd10586, %rd10530, %rd10585; + xor.b64 %rd10587, %rd10586, %rd10528; + not.b64 %rd10588, %rd10530; + and.b64 %rd10589, %rd10531, %rd10588; + xor.b64 %rd10590, %rd10589, %rd10529; + not.b64 %rd10591, %rd10531; + and.b64 %rd10592, %rd10527, %rd10591; + xor.b64 %rd10593, %rd10592, %rd10530; + not.b64 %rd10594, %rd10527; + and.b64 %rd10595, %rd10528, %rd10594; + xor.b64 %rd10596, %rd10595, %rd10531; + not.b64 %rd10597, %rd10533; + and.b64 %rd10598, %rd10534, %rd10597; + xor.b64 %rd10599, %rd10598, %rd10532; + not.b64 %rd10600, %rd10534; + and.b64 %rd10601, %rd10535, %rd10600; + xor.b64 %rd10602, %rd10601, %rd10533; + not.b64 %rd10603, %rd10535; + and.b64 %rd10604, %rd10536, %rd10603; + xor.b64 %rd10605, %rd10604, %rd10534; + not.b64 %rd10606, %rd10536; + and.b64 %rd10607, %rd10532, %rd10606; + xor.b64 %rd10608, %rd10607, %rd10535; + not.b64 %rd10609, %rd10532; + and.b64 %rd10610, %rd10533, %rd10609; + xor.b64 %rd10611, %rd10610, %rd10536; + xor.b64 %rd10613, %rd10539, %rd31; + xor.b64 %rd10614, %rd10554, %rd10613; + xor.b64 %rd10615, %rd10614, %rd10569; + xor.b64 %rd10616, %rd10615, %rd10584; + xor.b64 %rd10617, %rd10616, %rd10599; + xor.b64 %rd10618, %rd10557, %rd10542; + xor.b64 %rd10619, %rd10618, %rd10572; + xor.b64 %rd10620, %rd10619, %rd10587; + xor.b64 %rd10621, %rd10620, %rd10602; + xor.b64 %rd10622, %rd10560, %rd10545; + xor.b64 %rd10623, %rd10622, %rd10575; + xor.b64 %rd10624, %rd10623, %rd10590; + xor.b64 %rd10625, %rd10624, %rd10605; + xor.b64 %rd10626, %rd10563, %rd10548; + xor.b64 %rd10627, %rd10626, %rd10578; + xor.b64 %rd10628, %rd10627, %rd10593; + xor.b64 %rd10629, %rd10628, %rd10608; + xor.b64 %rd10630, %rd10566, %rd10551; + xor.b64 %rd10631, %rd10630, %rd10581; + xor.b64 %rd10632, %rd10631, %rd10596; + xor.b64 %rd10633, %rd10632, %rd10611; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7502}, %rd10621; + } + { + .reg .b32 %dummy; + mov.b64 {%r7503,%dummy}, %rd10621; + } + shf.l.wrap.b32 %r7504, %r7503, %r7502, 1; + shf.l.wrap.b32 %r7505, %r7502, %r7503, 1; + mov.b64 %rd10634, {%r7505, %r7504}; + xor.b64 %rd10635, %rd10633, %rd10634; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7506}, %rd10625; + } + { + .reg .b32 %dummy; + mov.b64 {%r7507,%dummy}, %rd10625; + } + shf.l.wrap.b32 %r7508, %r7507, %r7506, 1; + shf.l.wrap.b32 %r7509, %r7506, %r7507, 1; + mov.b64 %rd10636, {%r7509, %r7508}; + xor.b64 %rd10637, %rd10636, %rd10617; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7510}, %rd10629; + } + { + .reg .b32 %dummy; + mov.b64 {%r7511,%dummy}, %rd10629; + } + shf.l.wrap.b32 %r7512, %r7511, %r7510, 1; + shf.l.wrap.b32 %r7513, %r7510, %r7511, 1; + mov.b64 %rd10638, {%r7513, %r7512}; + xor.b64 %rd10639, %rd10638, %rd10621; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7514}, %rd10633; + } + { + .reg .b32 %dummy; + mov.b64 {%r7515,%dummy}, %rd10633; + } + shf.l.wrap.b32 %r7516, %r7515, %r7514, 1; + shf.l.wrap.b32 %r7517, %r7514, %r7515, 1; + mov.b64 %rd10640, {%r7517, %r7516}; + xor.b64 %rd10641, %rd10640, %rd10625; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7518}, %rd10617; + } + { + .reg .b32 %dummy; + mov.b64 {%r7519,%dummy}, %rd10617; + } + shf.l.wrap.b32 %r7520, %r7519, %r7518, 1; + shf.l.wrap.b32 %r7521, %r7518, %r7519, 1; + mov.b64 %rd10642, {%r7521, %r7520}; + xor.b64 %rd10643, %rd10629, %rd10642; + xor.b64 %rd10644, %rd10635, %rd10613; + xor.b64 %rd10645, %rd10637, %rd10542; + xor.b64 %rd10646, %rd10639, %rd10545; + xor.b64 %rd10647, %rd10641, %rd10548; + xor.b64 %rd10648, %rd10643, %rd10551; + xor.b64 %rd10649, %rd10635, %rd10554; + xor.b64 %rd10650, %rd10637, %rd10557; + xor.b64 %rd10651, %rd10639, %rd10560; + xor.b64 %rd10652, %rd10641, %rd10563; + xor.b64 %rd10653, %rd10643, %rd10566; + xor.b64 %rd10654, %rd10635, %rd10569; + xor.b64 %rd10655, %rd10637, %rd10572; + xor.b64 %rd10656, %rd10639, %rd10575; + xor.b64 %rd10657, %rd10641, %rd10578; + xor.b64 %rd10658, %rd10643, %rd10581; + xor.b64 %rd10659, %rd10635, %rd10584; + xor.b64 %rd10660, %rd10637, %rd10587; + xor.b64 %rd10661, %rd10639, %rd10590; + xor.b64 %rd10662, %rd10641, %rd10593; + xor.b64 %rd10663, %rd10643, %rd10596; + xor.b64 %rd10664, %rd10635, %rd10599; + xor.b64 %rd10665, %rd10637, %rd10602; + xor.b64 %rd10666, %rd10639, %rd10605; + xor.b64 %rd10667, %rd10641, %rd10608; + xor.b64 %rd10668, %rd10643, %rd10611; + { + .reg .b32 %dummy; + mov.b64 {%r7522,%dummy}, %rd10650; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7523}, %rd10650; + } + shf.r.wrap.b32 %r7524, %r7523, %r7522, 20; + shf.r.wrap.b32 %r7525, %r7522, %r7523, 20; + mov.b64 %rd10669, {%r7525, %r7524}; + { + .reg .b32 %dummy; + mov.b64 {%r7526,%dummy}, %rd10656; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7527}, %rd10656; + } + shf.r.wrap.b32 %r7528, %r7527, %r7526, 21; + shf.r.wrap.b32 %r7529, %r7526, %r7527, 21; + mov.b64 %rd10670, {%r7529, %r7528}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7530}, %rd10662; + } + { + .reg .b32 %dummy; + mov.b64 {%r7531,%dummy}, %rd10662; + } + shf.l.wrap.b32 %r7532, %r7531, %r7530, 21; + shf.l.wrap.b32 %r7533, %r7530, %r7531, 21; + mov.b64 %rd10671, {%r7533, %r7532}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7534}, %rd10668; + } + { + .reg .b32 %dummy; + mov.b64 {%r7535,%dummy}, %rd10668; + } + shf.l.wrap.b32 %r7536, %r7535, %r7534, 14; + shf.l.wrap.b32 %r7537, %r7534, %r7535, 14; + mov.b64 %rd10672, {%r7537, %r7536}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7538}, %rd10647; + } + { + .reg .b32 %dummy; + mov.b64 {%r7539,%dummy}, %rd10647; + } + shf.l.wrap.b32 %r7540, %r7539, %r7538, 28; + shf.l.wrap.b32 %r7541, %r7538, %r7539, 28; + mov.b64 %rd10673, {%r7541, %r7540}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7542}, %rd10653; + } + { + .reg .b32 %dummy; + mov.b64 {%r7543,%dummy}, %rd10653; + } + shf.l.wrap.b32 %r7544, %r7543, %r7542, 20; + shf.l.wrap.b32 %r7545, %r7542, %r7543, 20; + mov.b64 %rd10674, {%r7545, %r7544}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7546}, %rd10654; + } + { + .reg .b32 %dummy; + mov.b64 {%r7547,%dummy}, %rd10654; + } + shf.l.wrap.b32 %r7548, %r7547, %r7546, 3; + shf.l.wrap.b32 %r7549, %r7546, %r7547, 3; + mov.b64 %rd10675, {%r7549, %r7548}; + { + .reg .b32 %dummy; + mov.b64 {%r7550,%dummy}, %rd10660; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7551}, %rd10660; + } + shf.r.wrap.b32 %r7552, %r7551, %r7550, 19; + shf.r.wrap.b32 %r7553, %r7550, %r7551, 19; + mov.b64 %rd10676, {%r7553, %r7552}; + { + .reg .b32 %dummy; + mov.b64 {%r7554,%dummy}, %rd10666; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7555}, %rd10666; + } + shf.r.wrap.b32 %r7556, %r7555, %r7554, 3; + shf.r.wrap.b32 %r7557, %r7554, %r7555, 3; + mov.b64 %rd10677, {%r7557, %r7556}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7558}, %rd10645; + } + { + .reg .b32 %dummy; + mov.b64 {%r7559,%dummy}, %rd10645; + } + shf.l.wrap.b32 %r7560, %r7559, %r7558, 1; + shf.l.wrap.b32 %r7561, %r7558, %r7559, 1; + mov.b64 %rd10678, {%r7561, %r7560}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7562}, %rd10651; + } + { + .reg .b32 %dummy; + mov.b64 {%r7563,%dummy}, %rd10651; + } + shf.l.wrap.b32 %r7564, %r7563, %r7562, 6; + shf.l.wrap.b32 %r7565, %r7562, %r7563, 6; + mov.b64 %rd10679, {%r7565, %r7564}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7566}, %rd10657; + } + { + .reg .b32 %dummy; + mov.b64 {%r7567,%dummy}, %rd10657; + } + shf.l.wrap.b32 %r7568, %r7567, %r7566, 25; + shf.l.wrap.b32 %r7569, %r7566, %r7567, 25; + mov.b64 %rd10680, {%r7569, %r7568}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7570}, %rd10663; + } + { + .reg .b32 %dummy; + mov.b64 {%r7571,%dummy}, %rd10663; + } + shf.l.wrap.b32 %r7572, %r7571, %r7570, 8; + shf.l.wrap.b32 %r7573, %r7570, %r7571, 8; + mov.b64 %rd10681, {%r7573, %r7572}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7574}, %rd10664; + } + { + .reg .b32 %dummy; + mov.b64 {%r7575,%dummy}, %rd10664; + } + shf.l.wrap.b32 %r7576, %r7575, %r7574, 18; + shf.l.wrap.b32 %r7577, %r7574, %r7575, 18; + mov.b64 %rd10682, {%r7577, %r7576}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7578}, %rd10648; + } + { + .reg .b32 %dummy; + mov.b64 {%r7579,%dummy}, %rd10648; + } + shf.l.wrap.b32 %r7580, %r7579, %r7578, 27; + shf.l.wrap.b32 %r7581, %r7578, %r7579, 27; + mov.b64 %rd10683, {%r7581, %r7580}; + { + .reg .b32 %dummy; + mov.b64 {%r7582,%dummy}, %rd10649; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7583}, %rd10649; + } + shf.r.wrap.b32 %r7584, %r7583, %r7582, 28; + shf.r.wrap.b32 %r7585, %r7582, %r7583, 28; + mov.b64 %rd10684, {%r7585, %r7584}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7586}, %rd10655; + } + { + .reg .b32 %dummy; + mov.b64 {%r7587,%dummy}, %rd10655; + } + shf.l.wrap.b32 %r7588, %r7587, %r7586, 10; + shf.l.wrap.b32 %r7589, %r7586, %r7587, 10; + mov.b64 %rd10685, {%r7589, %r7588}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7590}, %rd10661; + } + { + .reg .b32 %dummy; + mov.b64 {%r7591,%dummy}, %rd10661; + } + shf.l.wrap.b32 %r7592, %r7591, %r7590, 15; + shf.l.wrap.b32 %r7593, %r7590, %r7591, 15; + mov.b64 %rd10686, {%r7593, %r7592}; + { + .reg .b32 %dummy; + mov.b64 {%r7594,%dummy}, %rd10667; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7595}, %rd10667; + } + shf.r.wrap.b32 %r7596, %r7595, %r7594, 8; + shf.r.wrap.b32 %r7597, %r7594, %r7595, 8; + mov.b64 %rd10687, {%r7597, %r7596}; + { + .reg .b32 %dummy; + mov.b64 {%r7598,%dummy}, %rd10646; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7599}, %rd10646; + } + shf.r.wrap.b32 %r7600, %r7599, %r7598, 2; + shf.r.wrap.b32 %r7601, %r7598, %r7599, 2; + mov.b64 %rd10688, {%r7601, %r7600}; + { + .reg .b32 %dummy; + mov.b64 {%r7602,%dummy}, %rd10652; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7603}, %rd10652; + } + shf.r.wrap.b32 %r7604, %r7603, %r7602, 9; + shf.r.wrap.b32 %r7605, %r7602, %r7603, 9; + mov.b64 %rd10689, {%r7605, %r7604}; + { + .reg .b32 %dummy; + mov.b64 {%r7606,%dummy}, %rd10658; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7607}, %rd10658; + } + shf.r.wrap.b32 %r7608, %r7607, %r7606, 25; + shf.r.wrap.b32 %r7609, %r7606, %r7607, 25; + mov.b64 %rd10690, {%r7609, %r7608}; + { + .reg .b32 %dummy; + mov.b64 {%r7610,%dummy}, %rd10659; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7611}, %rd10659; + } + shf.r.wrap.b32 %r7612, %r7611, %r7610, 23; + shf.r.wrap.b32 %r7613, %r7610, %r7611, 23; + mov.b64 %rd10691, {%r7613, %r7612}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7614}, %rd10665; + } + { + .reg .b32 %dummy; + mov.b64 {%r7615,%dummy}, %rd10665; + } + shf.l.wrap.b32 %r7616, %r7615, %r7614, 2; + shf.l.wrap.b32 %r7617, %r7614, %r7615, 2; + mov.b64 %rd10692, {%r7617, %r7616}; + not.b64 %rd10693, %rd10669; + and.b64 %rd10694, %rd10670, %rd10693; + xor.b64 %rd10695, %rd10694, %rd10644; + not.b64 %rd10696, %rd10670; + and.b64 %rd10697, %rd10671, %rd10696; + xor.b64 %rd10698, %rd10697, %rd10669; + not.b64 %rd10699, %rd10671; + and.b64 %rd10700, %rd10672, %rd10699; + xor.b64 %rd10701, %rd10700, %rd10670; + not.b64 %rd10702, %rd10672; + and.b64 %rd10703, %rd10644, %rd10702; + xor.b64 %rd10704, %rd10703, %rd10671; + not.b64 %rd10705, %rd10644; + and.b64 %rd10706, %rd10669, %rd10705; + xor.b64 %rd10707, %rd10706, %rd10672; + not.b64 %rd10708, %rd10674; + and.b64 %rd10709, %rd10675, %rd10708; + xor.b64 %rd10710, %rd10709, %rd10673; + not.b64 %rd10711, %rd10675; + and.b64 %rd10712, %rd10676, %rd10711; + xor.b64 %rd10713, %rd10712, %rd10674; + not.b64 %rd10714, %rd10676; + and.b64 %rd10715, %rd10677, %rd10714; + xor.b64 %rd10716, %rd10715, %rd10675; + not.b64 %rd10717, %rd10677; + and.b64 %rd10718, %rd10673, %rd10717; + xor.b64 %rd10719, %rd10718, %rd10676; + not.b64 %rd10720, %rd10673; + and.b64 %rd10721, %rd10674, %rd10720; + xor.b64 %rd10722, %rd10721, %rd10677; + not.b64 %rd10723, %rd10679; + and.b64 %rd10724, %rd10680, %rd10723; + xor.b64 %rd10725, %rd10724, %rd10678; + not.b64 %rd10726, %rd10680; + and.b64 %rd10727, %rd10681, %rd10726; + xor.b64 %rd10728, %rd10727, %rd10679; + not.b64 %rd10729, %rd10681; + and.b64 %rd10730, %rd10682, %rd10729; + xor.b64 %rd10731, %rd10730, %rd10680; + not.b64 %rd10732, %rd10682; + and.b64 %rd10733, %rd10678, %rd10732; + xor.b64 %rd10734, %rd10733, %rd10681; + not.b64 %rd10735, %rd10678; + and.b64 %rd10736, %rd10679, %rd10735; + xor.b64 %rd10737, %rd10736, %rd10682; + not.b64 %rd10738, %rd10684; + and.b64 %rd10739, %rd10685, %rd10738; + xor.b64 %rd10740, %rd10739, %rd10683; + not.b64 %rd10741, %rd10685; + and.b64 %rd10742, %rd10686, %rd10741; + xor.b64 %rd10743, %rd10742, %rd10684; + not.b64 %rd10744, %rd10686; + and.b64 %rd10745, %rd10687, %rd10744; + xor.b64 %rd10746, %rd10745, %rd10685; + not.b64 %rd10747, %rd10687; + and.b64 %rd10748, %rd10683, %rd10747; + xor.b64 %rd10749, %rd10748, %rd10686; + not.b64 %rd10750, %rd10683; + and.b64 %rd10751, %rd10684, %rd10750; + xor.b64 %rd10752, %rd10751, %rd10687; + not.b64 %rd10753, %rd10689; + and.b64 %rd10754, %rd10690, %rd10753; + xor.b64 %rd10755, %rd10754, %rd10688; + not.b64 %rd10756, %rd10690; + and.b64 %rd10757, %rd10691, %rd10756; + xor.b64 %rd10758, %rd10757, %rd10689; + not.b64 %rd10759, %rd10691; + and.b64 %rd10760, %rd10692, %rd10759; + xor.b64 %rd10761, %rd10760, %rd10690; + not.b64 %rd10762, %rd10692; + and.b64 %rd10763, %rd10688, %rd10762; + xor.b64 %rd10764, %rd10763, %rd10691; + not.b64 %rd10765, %rd10688; + and.b64 %rd10766, %rd10689, %rd10765; + xor.b64 %rd10767, %rd10766, %rd10692; + xor.b64 %rd10769, %rd10695, %rd32; + xor.b64 %rd10770, %rd10710, %rd10769; + xor.b64 %rd10771, %rd10770, %rd10725; + xor.b64 %rd10772, %rd10771, %rd10740; + xor.b64 %rd10773, %rd10772, %rd10755; + xor.b64 %rd10774, %rd10713, %rd10698; + xor.b64 %rd10775, %rd10774, %rd10728; + xor.b64 %rd10776, %rd10775, %rd10743; + xor.b64 %rd10777, %rd10776, %rd10758; + xor.b64 %rd10778, %rd10716, %rd10701; + xor.b64 %rd10779, %rd10778, %rd10731; + xor.b64 %rd10780, %rd10779, %rd10746; + xor.b64 %rd10781, %rd10780, %rd10761; + xor.b64 %rd10782, %rd10719, %rd10704; + xor.b64 %rd10783, %rd10782, %rd10734; + xor.b64 %rd10784, %rd10783, %rd10749; + xor.b64 %rd10785, %rd10784, %rd10764; + xor.b64 %rd10786, %rd10722, %rd10707; + xor.b64 %rd10787, %rd10786, %rd10737; + xor.b64 %rd10788, %rd10787, %rd10752; + xor.b64 %rd10789, %rd10788, %rd10767; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7618}, %rd10777; + } + { + .reg .b32 %dummy; + mov.b64 {%r7619,%dummy}, %rd10777; + } + shf.l.wrap.b32 %r7620, %r7619, %r7618, 1; + shf.l.wrap.b32 %r7621, %r7618, %r7619, 1; + mov.b64 %rd10790, {%r7621, %r7620}; + xor.b64 %rd10791, %rd10789, %rd10790; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7622}, %rd10781; + } + { + .reg .b32 %dummy; + mov.b64 {%r7623,%dummy}, %rd10781; + } + shf.l.wrap.b32 %r7624, %r7623, %r7622, 1; + shf.l.wrap.b32 %r7625, %r7622, %r7623, 1; + mov.b64 %rd10792, {%r7625, %r7624}; + xor.b64 %rd10793, %rd10792, %rd10773; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7626}, %rd10785; + } + { + .reg .b32 %dummy; + mov.b64 {%r7627,%dummy}, %rd10785; + } + shf.l.wrap.b32 %r7628, %r7627, %r7626, 1; + shf.l.wrap.b32 %r7629, %r7626, %r7627, 1; + mov.b64 %rd10794, {%r7629, %r7628}; + xor.b64 %rd10795, %rd10794, %rd10777; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7630}, %rd10789; + } + { + .reg .b32 %dummy; + mov.b64 {%r7631,%dummy}, %rd10789; + } + shf.l.wrap.b32 %r7632, %r7631, %r7630, 1; + shf.l.wrap.b32 %r7633, %r7630, %r7631, 1; + mov.b64 %rd10796, {%r7633, %r7632}; + xor.b64 %rd10797, %rd10796, %rd10781; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7634}, %rd10773; + } + { + .reg .b32 %dummy; + mov.b64 {%r7635,%dummy}, %rd10773; + } + shf.l.wrap.b32 %r7636, %r7635, %r7634, 1; + shf.l.wrap.b32 %r7637, %r7634, %r7635, 1; + mov.b64 %rd10798, {%r7637, %r7636}; + xor.b64 %rd10799, %rd10785, %rd10798; + xor.b64 %rd10800, %rd10791, %rd10769; + xor.b64 %rd10801, %rd10793, %rd10698; + xor.b64 %rd10802, %rd10795, %rd10701; + xor.b64 %rd10803, %rd10797, %rd10704; + xor.b64 %rd10804, %rd10799, %rd10707; + xor.b64 %rd10805, %rd10791, %rd10710; + xor.b64 %rd10806, %rd10793, %rd10713; + xor.b64 %rd10807, %rd10795, %rd10716; + xor.b64 %rd10808, %rd10797, %rd10719; + xor.b64 %rd10809, %rd10799, %rd10722; + xor.b64 %rd10810, %rd10791, %rd10725; + xor.b64 %rd10811, %rd10793, %rd10728; + xor.b64 %rd10812, %rd10795, %rd10731; + xor.b64 %rd10813, %rd10797, %rd10734; + xor.b64 %rd10814, %rd10799, %rd10737; + xor.b64 %rd10815, %rd10791, %rd10740; + xor.b64 %rd10816, %rd10793, %rd10743; + xor.b64 %rd10817, %rd10795, %rd10746; + xor.b64 %rd10818, %rd10797, %rd10749; + xor.b64 %rd10819, %rd10799, %rd10752; + xor.b64 %rd10820, %rd10791, %rd10755; + xor.b64 %rd10821, %rd10793, %rd10758; + xor.b64 %rd10822, %rd10795, %rd10761; + xor.b64 %rd10823, %rd10797, %rd10764; + xor.b64 %rd10824, %rd10799, %rd10767; + { + .reg .b32 %dummy; + mov.b64 {%r7638,%dummy}, %rd10806; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7639}, %rd10806; + } + shf.r.wrap.b32 %r7640, %r7639, %r7638, 20; + shf.r.wrap.b32 %r7641, %r7638, %r7639, 20; + mov.b64 %rd10825, {%r7641, %r7640}; + { + .reg .b32 %dummy; + mov.b64 {%r7642,%dummy}, %rd10812; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7643}, %rd10812; + } + shf.r.wrap.b32 %r7644, %r7643, %r7642, 21; + shf.r.wrap.b32 %r7645, %r7642, %r7643, 21; + mov.b64 %rd10826, {%r7645, %r7644}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7646}, %rd10818; + } + { + .reg .b32 %dummy; + mov.b64 {%r7647,%dummy}, %rd10818; + } + shf.l.wrap.b32 %r7648, %r7647, %r7646, 21; + shf.l.wrap.b32 %r7649, %r7646, %r7647, 21; + mov.b64 %rd10827, {%r7649, %r7648}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7650}, %rd10824; + } + { + .reg .b32 %dummy; + mov.b64 {%r7651,%dummy}, %rd10824; + } + shf.l.wrap.b32 %r7652, %r7651, %r7650, 14; + shf.l.wrap.b32 %r7653, %r7650, %r7651, 14; + mov.b64 %rd10828, {%r7653, %r7652}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7654}, %rd10803; + } + { + .reg .b32 %dummy; + mov.b64 {%r7655,%dummy}, %rd10803; + } + shf.l.wrap.b32 %r7656, %r7655, %r7654, 28; + shf.l.wrap.b32 %r7657, %r7654, %r7655, 28; + mov.b64 %rd10829, {%r7657, %r7656}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7658}, %rd10809; + } + { + .reg .b32 %dummy; + mov.b64 {%r7659,%dummy}, %rd10809; + } + shf.l.wrap.b32 %r7660, %r7659, %r7658, 20; + shf.l.wrap.b32 %r7661, %r7658, %r7659, 20; + mov.b64 %rd10830, {%r7661, %r7660}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7662}, %rd10810; + } + { + .reg .b32 %dummy; + mov.b64 {%r7663,%dummy}, %rd10810; + } + shf.l.wrap.b32 %r7664, %r7663, %r7662, 3; + shf.l.wrap.b32 %r7665, %r7662, %r7663, 3; + mov.b64 %rd10831, {%r7665, %r7664}; + { + .reg .b32 %dummy; + mov.b64 {%r7666,%dummy}, %rd10816; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7667}, %rd10816; + } + shf.r.wrap.b32 %r7668, %r7667, %r7666, 19; + shf.r.wrap.b32 %r7669, %r7666, %r7667, 19; + mov.b64 %rd10832, {%r7669, %r7668}; + { + .reg .b32 %dummy; + mov.b64 {%r7670,%dummy}, %rd10822; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7671}, %rd10822; + } + shf.r.wrap.b32 %r7672, %r7671, %r7670, 3; + shf.r.wrap.b32 %r7673, %r7670, %r7671, 3; + mov.b64 %rd10833, {%r7673, %r7672}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7674}, %rd10801; + } + { + .reg .b32 %dummy; + mov.b64 {%r7675,%dummy}, %rd10801; + } + shf.l.wrap.b32 %r7676, %r7675, %r7674, 1; + shf.l.wrap.b32 %r7677, %r7674, %r7675, 1; + mov.b64 %rd10834, {%r7677, %r7676}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7678}, %rd10807; + } + { + .reg .b32 %dummy; + mov.b64 {%r7679,%dummy}, %rd10807; + } + shf.l.wrap.b32 %r7680, %r7679, %r7678, 6; + shf.l.wrap.b32 %r7681, %r7678, %r7679, 6; + mov.b64 %rd10835, {%r7681, %r7680}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7682}, %rd10813; + } + { + .reg .b32 %dummy; + mov.b64 {%r7683,%dummy}, %rd10813; + } + shf.l.wrap.b32 %r7684, %r7683, %r7682, 25; + shf.l.wrap.b32 %r7685, %r7682, %r7683, 25; + mov.b64 %rd10836, {%r7685, %r7684}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7686}, %rd10819; + } + { + .reg .b32 %dummy; + mov.b64 {%r7687,%dummy}, %rd10819; + } + shf.l.wrap.b32 %r7688, %r7687, %r7686, 8; + shf.l.wrap.b32 %r7689, %r7686, %r7687, 8; + mov.b64 %rd10837, {%r7689, %r7688}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7690}, %rd10820; + } + { + .reg .b32 %dummy; + mov.b64 {%r7691,%dummy}, %rd10820; + } + shf.l.wrap.b32 %r7692, %r7691, %r7690, 18; + shf.l.wrap.b32 %r7693, %r7690, %r7691, 18; + mov.b64 %rd10838, {%r7693, %r7692}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7694}, %rd10804; + } + { + .reg .b32 %dummy; + mov.b64 {%r7695,%dummy}, %rd10804; + } + shf.l.wrap.b32 %r7696, %r7695, %r7694, 27; + shf.l.wrap.b32 %r7697, %r7694, %r7695, 27; + mov.b64 %rd10839, {%r7697, %r7696}; + { + .reg .b32 %dummy; + mov.b64 {%r7698,%dummy}, %rd10805; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7699}, %rd10805; + } + shf.r.wrap.b32 %r7700, %r7699, %r7698, 28; + shf.r.wrap.b32 %r7701, %r7698, %r7699, 28; + mov.b64 %rd10840, {%r7701, %r7700}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7702}, %rd10811; + } + { + .reg .b32 %dummy; + mov.b64 {%r7703,%dummy}, %rd10811; + } + shf.l.wrap.b32 %r7704, %r7703, %r7702, 10; + shf.l.wrap.b32 %r7705, %r7702, %r7703, 10; + mov.b64 %rd10841, {%r7705, %r7704}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7706}, %rd10817; + } + { + .reg .b32 %dummy; + mov.b64 {%r7707,%dummy}, %rd10817; + } + shf.l.wrap.b32 %r7708, %r7707, %r7706, 15; + shf.l.wrap.b32 %r7709, %r7706, %r7707, 15; + mov.b64 %rd10842, {%r7709, %r7708}; + { + .reg .b32 %dummy; + mov.b64 {%r7710,%dummy}, %rd10823; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7711}, %rd10823; + } + shf.r.wrap.b32 %r7712, %r7711, %r7710, 8; + shf.r.wrap.b32 %r7713, %r7710, %r7711, 8; + mov.b64 %rd10843, {%r7713, %r7712}; + { + .reg .b32 %dummy; + mov.b64 {%r7714,%dummy}, %rd10802; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7715}, %rd10802; + } + shf.r.wrap.b32 %r7716, %r7715, %r7714, 2; + shf.r.wrap.b32 %r7717, %r7714, %r7715, 2; + mov.b64 %rd10844, {%r7717, %r7716}; + { + .reg .b32 %dummy; + mov.b64 {%r7718,%dummy}, %rd10808; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7719}, %rd10808; + } + shf.r.wrap.b32 %r7720, %r7719, %r7718, 9; + shf.r.wrap.b32 %r7721, %r7718, %r7719, 9; + mov.b64 %rd10845, {%r7721, %r7720}; + { + .reg .b32 %dummy; + mov.b64 {%r7722,%dummy}, %rd10814; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7723}, %rd10814; + } + shf.r.wrap.b32 %r7724, %r7723, %r7722, 25; + shf.r.wrap.b32 %r7725, %r7722, %r7723, 25; + mov.b64 %rd10846, {%r7725, %r7724}; + { + .reg .b32 %dummy; + mov.b64 {%r7726,%dummy}, %rd10815; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7727}, %rd10815; + } + shf.r.wrap.b32 %r7728, %r7727, %r7726, 23; + shf.r.wrap.b32 %r7729, %r7726, %r7727, 23; + mov.b64 %rd10847, {%r7729, %r7728}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7730}, %rd10821; + } + { + .reg .b32 %dummy; + mov.b64 {%r7731,%dummy}, %rd10821; + } + shf.l.wrap.b32 %r7732, %r7731, %r7730, 2; + shf.l.wrap.b32 %r7733, %r7730, %r7731, 2; + mov.b64 %rd10848, {%r7733, %r7732}; + not.b64 %rd10849, %rd10825; + and.b64 %rd10850, %rd10826, %rd10849; + xor.b64 %rd10851, %rd10850, %rd10800; + not.b64 %rd10852, %rd10826; + and.b64 %rd10853, %rd10827, %rd10852; + xor.b64 %rd10854, %rd10853, %rd10825; + not.b64 %rd10855, %rd10827; + and.b64 %rd10856, %rd10828, %rd10855; + xor.b64 %rd10857, %rd10856, %rd10826; + not.b64 %rd10858, %rd10828; + and.b64 %rd10859, %rd10800, %rd10858; + xor.b64 %rd10860, %rd10859, %rd10827; + not.b64 %rd10861, %rd10800; + and.b64 %rd10862, %rd10825, %rd10861; + xor.b64 %rd10863, %rd10862, %rd10828; + not.b64 %rd10864, %rd10830; + and.b64 %rd10865, %rd10831, %rd10864; + xor.b64 %rd10866, %rd10865, %rd10829; + not.b64 %rd10867, %rd10831; + and.b64 %rd10868, %rd10832, %rd10867; + xor.b64 %rd10869, %rd10868, %rd10830; + not.b64 %rd10870, %rd10832; + and.b64 %rd10871, %rd10833, %rd10870; + xor.b64 %rd10872, %rd10871, %rd10831; + not.b64 %rd10873, %rd10833; + and.b64 %rd10874, %rd10829, %rd10873; + xor.b64 %rd10875, %rd10874, %rd10832; + not.b64 %rd10876, %rd10829; + and.b64 %rd10877, %rd10830, %rd10876; + xor.b64 %rd10878, %rd10877, %rd10833; + not.b64 %rd10879, %rd10835; + and.b64 %rd10880, %rd10836, %rd10879; + xor.b64 %rd10881, %rd10880, %rd10834; + not.b64 %rd10882, %rd10836; + and.b64 %rd10883, %rd10837, %rd10882; + xor.b64 %rd10884, %rd10883, %rd10835; + not.b64 %rd10885, %rd10837; + and.b64 %rd10886, %rd10838, %rd10885; + xor.b64 %rd10887, %rd10886, %rd10836; + not.b64 %rd10888, %rd10838; + and.b64 %rd10889, %rd10834, %rd10888; + xor.b64 %rd10890, %rd10889, %rd10837; + not.b64 %rd10891, %rd10834; + and.b64 %rd10892, %rd10835, %rd10891; + xor.b64 %rd10893, %rd10892, %rd10838; + not.b64 %rd10894, %rd10840; + and.b64 %rd10895, %rd10841, %rd10894; + xor.b64 %rd10896, %rd10895, %rd10839; + not.b64 %rd10897, %rd10841; + and.b64 %rd10898, %rd10842, %rd10897; + xor.b64 %rd10899, %rd10898, %rd10840; + not.b64 %rd10900, %rd10842; + and.b64 %rd10901, %rd10843, %rd10900; + xor.b64 %rd10902, %rd10901, %rd10841; + not.b64 %rd10903, %rd10843; + and.b64 %rd10904, %rd10839, %rd10903; + xor.b64 %rd10905, %rd10904, %rd10842; + not.b64 %rd10906, %rd10839; + and.b64 %rd10907, %rd10840, %rd10906; + xor.b64 %rd10908, %rd10907, %rd10843; + not.b64 %rd10909, %rd10845; + and.b64 %rd10910, %rd10846, %rd10909; + xor.b64 %rd10911, %rd10910, %rd10844; + not.b64 %rd10912, %rd10846; + and.b64 %rd10913, %rd10847, %rd10912; + xor.b64 %rd10914, %rd10913, %rd10845; + not.b64 %rd10915, %rd10847; + and.b64 %rd10916, %rd10848, %rd10915; + xor.b64 %rd10917, %rd10916, %rd10846; + not.b64 %rd10918, %rd10848; + and.b64 %rd10919, %rd10844, %rd10918; + xor.b64 %rd10920, %rd10919, %rd10847; + not.b64 %rd10921, %rd10844; + and.b64 %rd10922, %rd10845, %rd10921; + xor.b64 %rd10923, %rd10922, %rd10848; + xor.b64 %rd10925, %rd10851, %rd33; + xor.b64 %rd10926, %rd10866, %rd10925; + xor.b64 %rd10927, %rd10926, %rd10881; + xor.b64 %rd10928, %rd10927, %rd10896; + xor.b64 %rd10929, %rd10928, %rd10911; + xor.b64 %rd10930, %rd10869, %rd10854; + xor.b64 %rd10931, %rd10930, %rd10884; + xor.b64 %rd10932, %rd10931, %rd10899; + xor.b64 %rd10933, %rd10932, %rd10914; + xor.b64 %rd10934, %rd10872, %rd10857; + xor.b64 %rd10935, %rd10934, %rd10887; + xor.b64 %rd10936, %rd10935, %rd10902; + xor.b64 %rd10937, %rd10936, %rd10917; + xor.b64 %rd10938, %rd10875, %rd10860; + xor.b64 %rd10939, %rd10938, %rd10890; + xor.b64 %rd10940, %rd10939, %rd10905; + xor.b64 %rd10941, %rd10940, %rd10920; + xor.b64 %rd10942, %rd10878, %rd10863; + xor.b64 %rd10943, %rd10942, %rd10893; + xor.b64 %rd10944, %rd10943, %rd10908; + xor.b64 %rd10945, %rd10944, %rd10923; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7734}, %rd10933; + } + { + .reg .b32 %dummy; + mov.b64 {%r7735,%dummy}, %rd10933; + } + shf.l.wrap.b32 %r7736, %r7735, %r7734, 1; + shf.l.wrap.b32 %r7737, %r7734, %r7735, 1; + mov.b64 %rd10946, {%r7737, %r7736}; + xor.b64 %rd10947, %rd10945, %rd10946; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7738}, %rd10937; + } + { + .reg .b32 %dummy; + mov.b64 {%r7739,%dummy}, %rd10937; + } + shf.l.wrap.b32 %r7740, %r7739, %r7738, 1; + shf.l.wrap.b32 %r7741, %r7738, %r7739, 1; + mov.b64 %rd10948, {%r7741, %r7740}; + xor.b64 %rd10949, %rd10948, %rd10929; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7742}, %rd10941; + } + { + .reg .b32 %dummy; + mov.b64 {%r7743,%dummy}, %rd10941; + } + shf.l.wrap.b32 %r7744, %r7743, %r7742, 1; + shf.l.wrap.b32 %r7745, %r7742, %r7743, 1; + mov.b64 %rd10950, {%r7745, %r7744}; + xor.b64 %rd10951, %rd10950, %rd10933; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7746}, %rd10945; + } + { + .reg .b32 %dummy; + mov.b64 {%r7747,%dummy}, %rd10945; + } + shf.l.wrap.b32 %r7748, %r7747, %r7746, 1; + shf.l.wrap.b32 %r7749, %r7746, %r7747, 1; + mov.b64 %rd10952, {%r7749, %r7748}; + xor.b64 %rd10953, %rd10952, %rd10937; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7750}, %rd10929; + } + { + .reg .b32 %dummy; + mov.b64 {%r7751,%dummy}, %rd10929; + } + shf.l.wrap.b32 %r7752, %r7751, %r7750, 1; + shf.l.wrap.b32 %r7753, %r7750, %r7751, 1; + mov.b64 %rd10954, {%r7753, %r7752}; + xor.b64 %rd10955, %rd10941, %rd10954; + xor.b64 %rd10956, %rd10947, %rd10925; + xor.b64 %rd10957, %rd10949, %rd10854; + xor.b64 %rd10958, %rd10951, %rd10857; + xor.b64 %rd10959, %rd10953, %rd10860; + xor.b64 %rd10960, %rd10955, %rd10863; + xor.b64 %rd10961, %rd10947, %rd10866; + xor.b64 %rd10962, %rd10949, %rd10869; + xor.b64 %rd10963, %rd10951, %rd10872; + xor.b64 %rd10964, %rd10953, %rd10875; + xor.b64 %rd10965, %rd10955, %rd10878; + xor.b64 %rd10966, %rd10947, %rd10881; + xor.b64 %rd10967, %rd10949, %rd10884; + xor.b64 %rd10968, %rd10951, %rd10887; + xor.b64 %rd10969, %rd10953, %rd10890; + xor.b64 %rd10970, %rd10955, %rd10893; + xor.b64 %rd10971, %rd10947, %rd10896; + xor.b64 %rd10972, %rd10949, %rd10899; + xor.b64 %rd10973, %rd10951, %rd10902; + xor.b64 %rd10974, %rd10953, %rd10905; + xor.b64 %rd10975, %rd10955, %rd10908; + xor.b64 %rd10976, %rd10947, %rd10911; + xor.b64 %rd10977, %rd10949, %rd10914; + xor.b64 %rd10978, %rd10951, %rd10917; + xor.b64 %rd10979, %rd10953, %rd10920; + xor.b64 %rd10980, %rd10955, %rd10923; + { + .reg .b32 %dummy; + mov.b64 {%r7754,%dummy}, %rd10962; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7755}, %rd10962; + } + shf.r.wrap.b32 %r7756, %r7755, %r7754, 20; + shf.r.wrap.b32 %r7757, %r7754, %r7755, 20; + mov.b64 %rd10981, {%r7757, %r7756}; + { + .reg .b32 %dummy; + mov.b64 {%r7758,%dummy}, %rd10968; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7759}, %rd10968; + } + shf.r.wrap.b32 %r7760, %r7759, %r7758, 21; + shf.r.wrap.b32 %r7761, %r7758, %r7759, 21; + mov.b64 %rd10982, {%r7761, %r7760}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7762}, %rd10974; + } + { + .reg .b32 %dummy; + mov.b64 {%r7763,%dummy}, %rd10974; + } + shf.l.wrap.b32 %r7764, %r7763, %r7762, 21; + shf.l.wrap.b32 %r7765, %r7762, %r7763, 21; + mov.b64 %rd10983, {%r7765, %r7764}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7766}, %rd10980; + } + { + .reg .b32 %dummy; + mov.b64 {%r7767,%dummy}, %rd10980; + } + shf.l.wrap.b32 %r7768, %r7767, %r7766, 14; + shf.l.wrap.b32 %r7769, %r7766, %r7767, 14; + mov.b64 %rd10984, {%r7769, %r7768}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7770}, %rd10959; + } + { + .reg .b32 %dummy; + mov.b64 {%r7771,%dummy}, %rd10959; + } + shf.l.wrap.b32 %r7772, %r7771, %r7770, 28; + shf.l.wrap.b32 %r7773, %r7770, %r7771, 28; + mov.b64 %rd10985, {%r7773, %r7772}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7774}, %rd10965; + } + { + .reg .b32 %dummy; + mov.b64 {%r7775,%dummy}, %rd10965; + } + shf.l.wrap.b32 %r7776, %r7775, %r7774, 20; + shf.l.wrap.b32 %r7777, %r7774, %r7775, 20; + mov.b64 %rd10986, {%r7777, %r7776}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7778}, %rd10966; + } + { + .reg .b32 %dummy; + mov.b64 {%r7779,%dummy}, %rd10966; + } + shf.l.wrap.b32 %r7780, %r7779, %r7778, 3; + shf.l.wrap.b32 %r7781, %r7778, %r7779, 3; + mov.b64 %rd10987, {%r7781, %r7780}; + { + .reg .b32 %dummy; + mov.b64 {%r7782,%dummy}, %rd10972; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7783}, %rd10972; + } + shf.r.wrap.b32 %r7784, %r7783, %r7782, 19; + shf.r.wrap.b32 %r7785, %r7782, %r7783, 19; + mov.b64 %rd10988, {%r7785, %r7784}; + { + .reg .b32 %dummy; + mov.b64 {%r7786,%dummy}, %rd10978; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7787}, %rd10978; + } + shf.r.wrap.b32 %r7788, %r7787, %r7786, 3; + shf.r.wrap.b32 %r7789, %r7786, %r7787, 3; + mov.b64 %rd10989, {%r7789, %r7788}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7790}, %rd10957; + } + { + .reg .b32 %dummy; + mov.b64 {%r7791,%dummy}, %rd10957; + } + shf.l.wrap.b32 %r7792, %r7791, %r7790, 1; + shf.l.wrap.b32 %r7793, %r7790, %r7791, 1; + mov.b64 %rd10990, {%r7793, %r7792}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7794}, %rd10963; + } + { + .reg .b32 %dummy; + mov.b64 {%r7795,%dummy}, %rd10963; + } + shf.l.wrap.b32 %r7796, %r7795, %r7794, 6; + shf.l.wrap.b32 %r7797, %r7794, %r7795, 6; + mov.b64 %rd10991, {%r7797, %r7796}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7798}, %rd10969; + } + { + .reg .b32 %dummy; + mov.b64 {%r7799,%dummy}, %rd10969; + } + shf.l.wrap.b32 %r7800, %r7799, %r7798, 25; + shf.l.wrap.b32 %r7801, %r7798, %r7799, 25; + mov.b64 %rd10992, {%r7801, %r7800}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7802}, %rd10975; + } + { + .reg .b32 %dummy; + mov.b64 {%r7803,%dummy}, %rd10975; + } + shf.l.wrap.b32 %r7804, %r7803, %r7802, 8; + shf.l.wrap.b32 %r7805, %r7802, %r7803, 8; + mov.b64 %rd10993, {%r7805, %r7804}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7806}, %rd10976; + } + { + .reg .b32 %dummy; + mov.b64 {%r7807,%dummy}, %rd10976; + } + shf.l.wrap.b32 %r7808, %r7807, %r7806, 18; + shf.l.wrap.b32 %r7809, %r7806, %r7807, 18; + mov.b64 %rd10994, {%r7809, %r7808}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7810}, %rd10960; + } + { + .reg .b32 %dummy; + mov.b64 {%r7811,%dummy}, %rd10960; + } + shf.l.wrap.b32 %r7812, %r7811, %r7810, 27; + shf.l.wrap.b32 %r7813, %r7810, %r7811, 27; + mov.b64 %rd10995, {%r7813, %r7812}; + { + .reg .b32 %dummy; + mov.b64 {%r7814,%dummy}, %rd10961; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7815}, %rd10961; + } + shf.r.wrap.b32 %r7816, %r7815, %r7814, 28; + shf.r.wrap.b32 %r7817, %r7814, %r7815, 28; + mov.b64 %rd10996, {%r7817, %r7816}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7818}, %rd10967; + } + { + .reg .b32 %dummy; + mov.b64 {%r7819,%dummy}, %rd10967; + } + shf.l.wrap.b32 %r7820, %r7819, %r7818, 10; + shf.l.wrap.b32 %r7821, %r7818, %r7819, 10; + mov.b64 %rd10997, {%r7821, %r7820}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7822}, %rd10973; + } + { + .reg .b32 %dummy; + mov.b64 {%r7823,%dummy}, %rd10973; + } + shf.l.wrap.b32 %r7824, %r7823, %r7822, 15; + shf.l.wrap.b32 %r7825, %r7822, %r7823, 15; + mov.b64 %rd10998, {%r7825, %r7824}; + { + .reg .b32 %dummy; + mov.b64 {%r7826,%dummy}, %rd10979; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7827}, %rd10979; + } + shf.r.wrap.b32 %r7828, %r7827, %r7826, 8; + shf.r.wrap.b32 %r7829, %r7826, %r7827, 8; + mov.b64 %rd10999, {%r7829, %r7828}; + { + .reg .b32 %dummy; + mov.b64 {%r7830,%dummy}, %rd10958; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7831}, %rd10958; + } + shf.r.wrap.b32 %r7832, %r7831, %r7830, 2; + shf.r.wrap.b32 %r7833, %r7830, %r7831, 2; + mov.b64 %rd11000, {%r7833, %r7832}; + { + .reg .b32 %dummy; + mov.b64 {%r7834,%dummy}, %rd10964; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7835}, %rd10964; + } + shf.r.wrap.b32 %r7836, %r7835, %r7834, 9; + shf.r.wrap.b32 %r7837, %r7834, %r7835, 9; + mov.b64 %rd11001, {%r7837, %r7836}; + { + .reg .b32 %dummy; + mov.b64 {%r7838,%dummy}, %rd10970; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7839}, %rd10970; + } + shf.r.wrap.b32 %r7840, %r7839, %r7838, 25; + shf.r.wrap.b32 %r7841, %r7838, %r7839, 25; + mov.b64 %rd11002, {%r7841, %r7840}; + { + .reg .b32 %dummy; + mov.b64 {%r7842,%dummy}, %rd10971; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7843}, %rd10971; + } + shf.r.wrap.b32 %r7844, %r7843, %r7842, 23; + shf.r.wrap.b32 %r7845, %r7842, %r7843, 23; + mov.b64 %rd11003, {%r7845, %r7844}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7846}, %rd10977; + } + { + .reg .b32 %dummy; + mov.b64 {%r7847,%dummy}, %rd10977; + } + shf.l.wrap.b32 %r7848, %r7847, %r7846, 2; + shf.l.wrap.b32 %r7849, %r7846, %r7847, 2; + mov.b64 %rd11004, {%r7849, %r7848}; + not.b64 %rd11005, %rd10981; + and.b64 %rd11006, %rd10982, %rd11005; + xor.b64 %rd11007, %rd11006, %rd10956; + not.b64 %rd11008, %rd10982; + and.b64 %rd11009, %rd10983, %rd11008; + xor.b64 %rd11010, %rd11009, %rd10981; + not.b64 %rd11011, %rd10983; + and.b64 %rd11012, %rd10984, %rd11011; + xor.b64 %rd11013, %rd11012, %rd10982; + not.b64 %rd11014, %rd10984; + and.b64 %rd11015, %rd10956, %rd11014; + xor.b64 %rd11016, %rd11015, %rd10983; + not.b64 %rd11017, %rd10956; + and.b64 %rd11018, %rd10981, %rd11017; + xor.b64 %rd11019, %rd11018, %rd10984; + not.b64 %rd11020, %rd10986; + and.b64 %rd11021, %rd10987, %rd11020; + xor.b64 %rd11022, %rd11021, %rd10985; + not.b64 %rd11023, %rd10987; + and.b64 %rd11024, %rd10988, %rd11023; + xor.b64 %rd11025, %rd11024, %rd10986; + not.b64 %rd11026, %rd10988; + and.b64 %rd11027, %rd10989, %rd11026; + xor.b64 %rd11028, %rd11027, %rd10987; + not.b64 %rd11029, %rd10989; + and.b64 %rd11030, %rd10985, %rd11029; + xor.b64 %rd11031, %rd11030, %rd10988; + not.b64 %rd11032, %rd10985; + and.b64 %rd11033, %rd10986, %rd11032; + xor.b64 %rd11034, %rd11033, %rd10989; + not.b64 %rd11035, %rd10991; + and.b64 %rd11036, %rd10992, %rd11035; + xor.b64 %rd11037, %rd11036, %rd10990; + not.b64 %rd11038, %rd10992; + and.b64 %rd11039, %rd10993, %rd11038; + xor.b64 %rd11040, %rd11039, %rd10991; + not.b64 %rd11041, %rd10993; + and.b64 %rd11042, %rd10994, %rd11041; + xor.b64 %rd11043, %rd11042, %rd10992; + not.b64 %rd11044, %rd10994; + and.b64 %rd11045, %rd10990, %rd11044; + xor.b64 %rd11046, %rd11045, %rd10993; + not.b64 %rd11047, %rd10990; + and.b64 %rd11048, %rd10991, %rd11047; + xor.b64 %rd11049, %rd11048, %rd10994; + not.b64 %rd11050, %rd10996; + and.b64 %rd11051, %rd10997, %rd11050; + xor.b64 %rd11052, %rd11051, %rd10995; + not.b64 %rd11053, %rd10997; + and.b64 %rd11054, %rd10998, %rd11053; + xor.b64 %rd11055, %rd11054, %rd10996; + not.b64 %rd11056, %rd10998; + and.b64 %rd11057, %rd10999, %rd11056; + xor.b64 %rd11058, %rd11057, %rd10997; + not.b64 %rd11059, %rd10999; + and.b64 %rd11060, %rd10995, %rd11059; + xor.b64 %rd11061, %rd11060, %rd10998; + not.b64 %rd11062, %rd10995; + and.b64 %rd11063, %rd10996, %rd11062; + xor.b64 %rd11064, %rd11063, %rd10999; + not.b64 %rd11065, %rd11001; + and.b64 %rd11066, %rd11002, %rd11065; + xor.b64 %rd11067, %rd11066, %rd11000; + not.b64 %rd11068, %rd11002; + and.b64 %rd11069, %rd11003, %rd11068; + xor.b64 %rd11070, %rd11069, %rd11001; + not.b64 %rd11071, %rd11003; + and.b64 %rd11072, %rd11004, %rd11071; + xor.b64 %rd11073, %rd11072, %rd11002; + not.b64 %rd11074, %rd11004; + and.b64 %rd11075, %rd11000, %rd11074; + xor.b64 %rd11076, %rd11075, %rd11003; + not.b64 %rd11077, %rd11000; + and.b64 %rd11078, %rd11001, %rd11077; + xor.b64 %rd11079, %rd11078, %rd11004; + xor.b64 %rd11081, %rd11007, %rd34; + xor.b64 %rd11082, %rd11022, %rd11081; + xor.b64 %rd11083, %rd11082, %rd11037; + xor.b64 %rd11084, %rd11083, %rd11052; + xor.b64 %rd11085, %rd11084, %rd11067; + xor.b64 %rd11086, %rd11025, %rd11010; + xor.b64 %rd11087, %rd11086, %rd11040; + xor.b64 %rd11088, %rd11087, %rd11055; + xor.b64 %rd11089, %rd11088, %rd11070; + xor.b64 %rd11090, %rd11028, %rd11013; + xor.b64 %rd11091, %rd11090, %rd11043; + xor.b64 %rd11092, %rd11091, %rd11058; + xor.b64 %rd11093, %rd11092, %rd11073; + xor.b64 %rd11094, %rd11031, %rd11016; + xor.b64 %rd11095, %rd11094, %rd11046; + xor.b64 %rd11096, %rd11095, %rd11061; + xor.b64 %rd11097, %rd11096, %rd11076; + xor.b64 %rd11098, %rd11034, %rd11019; + xor.b64 %rd11099, %rd11098, %rd11049; + xor.b64 %rd11100, %rd11099, %rd11064; + xor.b64 %rd11101, %rd11100, %rd11079; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7850}, %rd11089; + } + { + .reg .b32 %dummy; + mov.b64 {%r7851,%dummy}, %rd11089; + } + shf.l.wrap.b32 %r7852, %r7851, %r7850, 1; + shf.l.wrap.b32 %r7853, %r7850, %r7851, 1; + mov.b64 %rd11102, {%r7853, %r7852}; + xor.b64 %rd11103, %rd11101, %rd11102; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7854}, %rd11093; + } + { + .reg .b32 %dummy; + mov.b64 {%r7855,%dummy}, %rd11093; + } + shf.l.wrap.b32 %r7856, %r7855, %r7854, 1; + shf.l.wrap.b32 %r7857, %r7854, %r7855, 1; + mov.b64 %rd11104, {%r7857, %r7856}; + xor.b64 %rd11105, %rd11104, %rd11085; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7858}, %rd11097; + } + { + .reg .b32 %dummy; + mov.b64 {%r7859,%dummy}, %rd11097; + } + shf.l.wrap.b32 %r7860, %r7859, %r7858, 1; + shf.l.wrap.b32 %r7861, %r7858, %r7859, 1; + mov.b64 %rd11106, {%r7861, %r7860}; + xor.b64 %rd11107, %rd11106, %rd11089; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7862}, %rd11101; + } + { + .reg .b32 %dummy; + mov.b64 {%r7863,%dummy}, %rd11101; + } + shf.l.wrap.b32 %r7864, %r7863, %r7862, 1; + shf.l.wrap.b32 %r7865, %r7862, %r7863, 1; + mov.b64 %rd11108, {%r7865, %r7864}; + xor.b64 %rd11109, %rd11108, %rd11093; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7866}, %rd11085; + } + { + .reg .b32 %dummy; + mov.b64 {%r7867,%dummy}, %rd11085; + } + shf.l.wrap.b32 %r7868, %r7867, %r7866, 1; + shf.l.wrap.b32 %r7869, %r7866, %r7867, 1; + mov.b64 %rd11110, {%r7869, %r7868}; + xor.b64 %rd11111, %rd11097, %rd11110; + xor.b64 %rd11112, %rd11103, %rd11081; + xor.b64 %rd11113, %rd11105, %rd11010; + xor.b64 %rd11114, %rd11107, %rd11013; + xor.b64 %rd11115, %rd11109, %rd11016; + xor.b64 %rd11116, %rd11111, %rd11019; + xor.b64 %rd11117, %rd11103, %rd11022; + xor.b64 %rd11118, %rd11105, %rd11025; + xor.b64 %rd11119, %rd11107, %rd11028; + xor.b64 %rd11120, %rd11109, %rd11031; + xor.b64 %rd11121, %rd11111, %rd11034; + xor.b64 %rd11122, %rd11103, %rd11037; + xor.b64 %rd11123, %rd11105, %rd11040; + xor.b64 %rd11124, %rd11107, %rd11043; + xor.b64 %rd11125, %rd11109, %rd11046; + xor.b64 %rd11126, %rd11111, %rd11049; + xor.b64 %rd11127, %rd11103, %rd11052; + xor.b64 %rd11128, %rd11105, %rd11055; + xor.b64 %rd11129, %rd11107, %rd11058; + xor.b64 %rd11130, %rd11109, %rd11061; + xor.b64 %rd11131, %rd11111, %rd11064; + xor.b64 %rd11132, %rd11103, %rd11067; + xor.b64 %rd11133, %rd11105, %rd11070; + xor.b64 %rd11134, %rd11107, %rd11073; + xor.b64 %rd11135, %rd11109, %rd11076; + xor.b64 %rd11136, %rd11111, %rd11079; + { + .reg .b32 %dummy; + mov.b64 {%r7870,%dummy}, %rd11118; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7871}, %rd11118; + } + shf.r.wrap.b32 %r7872, %r7871, %r7870, 20; + shf.r.wrap.b32 %r7873, %r7870, %r7871, 20; + mov.b64 %rd11137, {%r7873, %r7872}; + { + .reg .b32 %dummy; + mov.b64 {%r7874,%dummy}, %rd11124; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7875}, %rd11124; + } + shf.r.wrap.b32 %r7876, %r7875, %r7874, 21; + shf.r.wrap.b32 %r7877, %r7874, %r7875, 21; + mov.b64 %rd11138, {%r7877, %r7876}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7878}, %rd11130; + } + { + .reg .b32 %dummy; + mov.b64 {%r7879,%dummy}, %rd11130; + } + shf.l.wrap.b32 %r7880, %r7879, %r7878, 21; + shf.l.wrap.b32 %r7881, %r7878, %r7879, 21; + mov.b64 %rd11139, {%r7881, %r7880}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7882}, %rd11136; + } + { + .reg .b32 %dummy; + mov.b64 {%r7883,%dummy}, %rd11136; + } + shf.l.wrap.b32 %r7884, %r7883, %r7882, 14; + shf.l.wrap.b32 %r7885, %r7882, %r7883, 14; + mov.b64 %rd11140, {%r7885, %r7884}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7886}, %rd11115; + } + { + .reg .b32 %dummy; + mov.b64 {%r7887,%dummy}, %rd11115; + } + shf.l.wrap.b32 %r7888, %r7887, %r7886, 28; + shf.l.wrap.b32 %r7889, %r7886, %r7887, 28; + mov.b64 %rd11141, {%r7889, %r7888}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7890}, %rd11121; + } + { + .reg .b32 %dummy; + mov.b64 {%r7891,%dummy}, %rd11121; + } + shf.l.wrap.b32 %r7892, %r7891, %r7890, 20; + shf.l.wrap.b32 %r7893, %r7890, %r7891, 20; + mov.b64 %rd11142, {%r7893, %r7892}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7894}, %rd11122; + } + { + .reg .b32 %dummy; + mov.b64 {%r7895,%dummy}, %rd11122; + } + shf.l.wrap.b32 %r7896, %r7895, %r7894, 3; + shf.l.wrap.b32 %r7897, %r7894, %r7895, 3; + mov.b64 %rd11143, {%r7897, %r7896}; + { + .reg .b32 %dummy; + mov.b64 {%r7898,%dummy}, %rd11128; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7899}, %rd11128; + } + shf.r.wrap.b32 %r7900, %r7899, %r7898, 19; + shf.r.wrap.b32 %r7901, %r7898, %r7899, 19; + mov.b64 %rd11144, {%r7901, %r7900}; + { + .reg .b32 %dummy; + mov.b64 {%r7902,%dummy}, %rd11134; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7903}, %rd11134; + } + shf.r.wrap.b32 %r7904, %r7903, %r7902, 3; + shf.r.wrap.b32 %r7905, %r7902, %r7903, 3; + mov.b64 %rd11145, {%r7905, %r7904}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7906}, %rd11113; + } + { + .reg .b32 %dummy; + mov.b64 {%r7907,%dummy}, %rd11113; + } + shf.l.wrap.b32 %r7908, %r7907, %r7906, 1; + shf.l.wrap.b32 %r7909, %r7906, %r7907, 1; + mov.b64 %rd11146, {%r7909, %r7908}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7910}, %rd11119; + } + { + .reg .b32 %dummy; + mov.b64 {%r7911,%dummy}, %rd11119; + } + shf.l.wrap.b32 %r7912, %r7911, %r7910, 6; + shf.l.wrap.b32 %r7913, %r7910, %r7911, 6; + mov.b64 %rd11147, {%r7913, %r7912}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7914}, %rd11125; + } + { + .reg .b32 %dummy; + mov.b64 {%r7915,%dummy}, %rd11125; + } + shf.l.wrap.b32 %r7916, %r7915, %r7914, 25; + shf.l.wrap.b32 %r7917, %r7914, %r7915, 25; + mov.b64 %rd11148, {%r7917, %r7916}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7918}, %rd11131; + } + { + .reg .b32 %dummy; + mov.b64 {%r7919,%dummy}, %rd11131; + } + shf.l.wrap.b32 %r7920, %r7919, %r7918, 8; + shf.l.wrap.b32 %r7921, %r7918, %r7919, 8; + mov.b64 %rd11149, {%r7921, %r7920}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7922}, %rd11132; + } + { + .reg .b32 %dummy; + mov.b64 {%r7923,%dummy}, %rd11132; + } + shf.l.wrap.b32 %r7924, %r7923, %r7922, 18; + shf.l.wrap.b32 %r7925, %r7922, %r7923, 18; + mov.b64 %rd11150, {%r7925, %r7924}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7926}, %rd11116; + } + { + .reg .b32 %dummy; + mov.b64 {%r7927,%dummy}, %rd11116; + } + shf.l.wrap.b32 %r7928, %r7927, %r7926, 27; + shf.l.wrap.b32 %r7929, %r7926, %r7927, 27; + mov.b64 %rd11151, {%r7929, %r7928}; + { + .reg .b32 %dummy; + mov.b64 {%r7930,%dummy}, %rd11117; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7931}, %rd11117; + } + shf.r.wrap.b32 %r7932, %r7931, %r7930, 28; + shf.r.wrap.b32 %r7933, %r7930, %r7931, 28; + mov.b64 %rd11152, {%r7933, %r7932}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7934}, %rd11123; + } + { + .reg .b32 %dummy; + mov.b64 {%r7935,%dummy}, %rd11123; + } + shf.l.wrap.b32 %r7936, %r7935, %r7934, 10; + shf.l.wrap.b32 %r7937, %r7934, %r7935, 10; + mov.b64 %rd11153, {%r7937, %r7936}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7938}, %rd11129; + } + { + .reg .b32 %dummy; + mov.b64 {%r7939,%dummy}, %rd11129; + } + shf.l.wrap.b32 %r7940, %r7939, %r7938, 15; + shf.l.wrap.b32 %r7941, %r7938, %r7939, 15; + mov.b64 %rd11154, {%r7941, %r7940}; + { + .reg .b32 %dummy; + mov.b64 {%r7942,%dummy}, %rd11135; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7943}, %rd11135; + } + shf.r.wrap.b32 %r7944, %r7943, %r7942, 8; + shf.r.wrap.b32 %r7945, %r7942, %r7943, 8; + mov.b64 %rd11155, {%r7945, %r7944}; + { + .reg .b32 %dummy; + mov.b64 {%r7946,%dummy}, %rd11114; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7947}, %rd11114; + } + shf.r.wrap.b32 %r7948, %r7947, %r7946, 2; + shf.r.wrap.b32 %r7949, %r7946, %r7947, 2; + mov.b64 %rd11156, {%r7949, %r7948}; + { + .reg .b32 %dummy; + mov.b64 {%r7950,%dummy}, %rd11120; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7951}, %rd11120; + } + shf.r.wrap.b32 %r7952, %r7951, %r7950, 9; + shf.r.wrap.b32 %r7953, %r7950, %r7951, 9; + mov.b64 %rd11157, {%r7953, %r7952}; + { + .reg .b32 %dummy; + mov.b64 {%r7954,%dummy}, %rd11126; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7955}, %rd11126; + } + shf.r.wrap.b32 %r7956, %r7955, %r7954, 25; + shf.r.wrap.b32 %r7957, %r7954, %r7955, 25; + mov.b64 %rd11158, {%r7957, %r7956}; + { + .reg .b32 %dummy; + mov.b64 {%r7958,%dummy}, %rd11127; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7959}, %rd11127; + } + shf.r.wrap.b32 %r7960, %r7959, %r7958, 23; + shf.r.wrap.b32 %r7961, %r7958, %r7959, 23; + mov.b64 %rd11159, {%r7961, %r7960}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7962}, %rd11133; + } + { + .reg .b32 %dummy; + mov.b64 {%r7963,%dummy}, %rd11133; + } + shf.l.wrap.b32 %r7964, %r7963, %r7962, 2; + shf.l.wrap.b32 %r7965, %r7962, %r7963, 2; + mov.b64 %rd11160, {%r7965, %r7964}; + not.b64 %rd11161, %rd11137; + and.b64 %rd11162, %rd11138, %rd11161; + xor.b64 %rd11163, %rd11162, %rd11112; + not.b64 %rd11164, %rd11138; + and.b64 %rd11165, %rd11139, %rd11164; + xor.b64 %rd11166, %rd11165, %rd11137; + not.b64 %rd11167, %rd11139; + and.b64 %rd11168, %rd11140, %rd11167; + xor.b64 %rd11169, %rd11168, %rd11138; + not.b64 %rd11170, %rd11140; + and.b64 %rd11171, %rd11112, %rd11170; + xor.b64 %rd11172, %rd11171, %rd11139; + not.b64 %rd11173, %rd11112; + and.b64 %rd11174, %rd11137, %rd11173; + xor.b64 %rd11175, %rd11174, %rd11140; + not.b64 %rd11176, %rd11142; + and.b64 %rd11177, %rd11143, %rd11176; + xor.b64 %rd11178, %rd11177, %rd11141; + not.b64 %rd11179, %rd11143; + and.b64 %rd11180, %rd11144, %rd11179; + xor.b64 %rd11181, %rd11180, %rd11142; + not.b64 %rd11182, %rd11144; + and.b64 %rd11183, %rd11145, %rd11182; + xor.b64 %rd11184, %rd11183, %rd11143; + not.b64 %rd11185, %rd11145; + and.b64 %rd11186, %rd11141, %rd11185; + xor.b64 %rd11187, %rd11186, %rd11144; + not.b64 %rd11188, %rd11141; + and.b64 %rd11189, %rd11142, %rd11188; + xor.b64 %rd11190, %rd11189, %rd11145; + not.b64 %rd11191, %rd11147; + and.b64 %rd11192, %rd11148, %rd11191; + xor.b64 %rd11193, %rd11192, %rd11146; + not.b64 %rd11194, %rd11148; + and.b64 %rd11195, %rd11149, %rd11194; + xor.b64 %rd11196, %rd11195, %rd11147; + not.b64 %rd11197, %rd11149; + and.b64 %rd11198, %rd11150, %rd11197; + xor.b64 %rd11199, %rd11198, %rd11148; + not.b64 %rd11200, %rd11150; + and.b64 %rd11201, %rd11146, %rd11200; + xor.b64 %rd11202, %rd11201, %rd11149; + not.b64 %rd11203, %rd11146; + and.b64 %rd11204, %rd11147, %rd11203; + xor.b64 %rd11205, %rd11204, %rd11150; + not.b64 %rd11206, %rd11152; + and.b64 %rd11207, %rd11153, %rd11206; + xor.b64 %rd11208, %rd11207, %rd11151; + not.b64 %rd11209, %rd11153; + and.b64 %rd11210, %rd11154, %rd11209; + xor.b64 %rd11211, %rd11210, %rd11152; + not.b64 %rd11212, %rd11154; + and.b64 %rd11213, %rd11155, %rd11212; + xor.b64 %rd11214, %rd11213, %rd11153; + not.b64 %rd11215, %rd11155; + and.b64 %rd11216, %rd11151, %rd11215; + xor.b64 %rd11217, %rd11216, %rd11154; + not.b64 %rd11218, %rd11151; + and.b64 %rd11219, %rd11152, %rd11218; + xor.b64 %rd11220, %rd11219, %rd11155; + not.b64 %rd11221, %rd11157; + and.b64 %rd11222, %rd11158, %rd11221; + xor.b64 %rd11223, %rd11222, %rd11156; + not.b64 %rd11224, %rd11158; + and.b64 %rd11225, %rd11159, %rd11224; + xor.b64 %rd11226, %rd11225, %rd11157; + not.b64 %rd11227, %rd11159; + and.b64 %rd11228, %rd11160, %rd11227; + xor.b64 %rd11229, %rd11228, %rd11158; + not.b64 %rd11230, %rd11160; + and.b64 %rd11231, %rd11156, %rd11230; + xor.b64 %rd11232, %rd11231, %rd11159; + not.b64 %rd11233, %rd11156; + and.b64 %rd11234, %rd11157, %rd11233; + xor.b64 %rd11235, %rd11234, %rd11160; + xor.b64 %rd11237, %rd11163, %rd35; + xor.b64 %rd11238, %rd11178, %rd11237; + xor.b64 %rd11239, %rd11238, %rd11193; + xor.b64 %rd11240, %rd11239, %rd11208; + xor.b64 %rd11241, %rd11240, %rd11223; + xor.b64 %rd11242, %rd11181, %rd11166; + xor.b64 %rd11243, %rd11242, %rd11196; + xor.b64 %rd11244, %rd11243, %rd11211; + xor.b64 %rd11245, %rd11244, %rd11226; + xor.b64 %rd11246, %rd11184, %rd11169; + xor.b64 %rd11247, %rd11246, %rd11199; + xor.b64 %rd11248, %rd11247, %rd11214; + xor.b64 %rd11249, %rd11248, %rd11229; + xor.b64 %rd11250, %rd11187, %rd11172; + xor.b64 %rd11251, %rd11250, %rd11202; + xor.b64 %rd11252, %rd11251, %rd11217; + xor.b64 %rd11253, %rd11252, %rd11232; + xor.b64 %rd11254, %rd11190, %rd11175; + xor.b64 %rd11255, %rd11254, %rd11205; + xor.b64 %rd11256, %rd11255, %rd11220; + xor.b64 %rd11257, %rd11256, %rd11235; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7966}, %rd11245; + } + { + .reg .b32 %dummy; + mov.b64 {%r7967,%dummy}, %rd11245; + } + shf.l.wrap.b32 %r7968, %r7967, %r7966, 1; + shf.l.wrap.b32 %r7969, %r7966, %r7967, 1; + mov.b64 %rd11258, {%r7969, %r7968}; + xor.b64 %rd11259, %rd11257, %rd11258; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7970}, %rd11249; + } + { + .reg .b32 %dummy; + mov.b64 {%r7971,%dummy}, %rd11249; + } + shf.l.wrap.b32 %r7972, %r7971, %r7970, 1; + shf.l.wrap.b32 %r7973, %r7970, %r7971, 1; + mov.b64 %rd11260, {%r7973, %r7972}; + xor.b64 %rd11261, %rd11260, %rd11241; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7974}, %rd11253; + } + { + .reg .b32 %dummy; + mov.b64 {%r7975,%dummy}, %rd11253; + } + shf.l.wrap.b32 %r7976, %r7975, %r7974, 1; + shf.l.wrap.b32 %r7977, %r7974, %r7975, 1; + mov.b64 %rd11262, {%r7977, %r7976}; + xor.b64 %rd11263, %rd11262, %rd11245; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7978}, %rd11257; + } + { + .reg .b32 %dummy; + mov.b64 {%r7979,%dummy}, %rd11257; + } + shf.l.wrap.b32 %r7980, %r7979, %r7978, 1; + shf.l.wrap.b32 %r7981, %r7978, %r7979, 1; + mov.b64 %rd11264, {%r7981, %r7980}; + xor.b64 %rd11265, %rd11264, %rd11249; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7982}, %rd11241; + } + { + .reg .b32 %dummy; + mov.b64 {%r7983,%dummy}, %rd11241; + } + shf.l.wrap.b32 %r7984, %r7983, %r7982, 1; + shf.l.wrap.b32 %r7985, %r7982, %r7983, 1; + mov.b64 %rd11266, {%r7985, %r7984}; + xor.b64 %rd11267, %rd11253, %rd11266; + xor.b64 %rd11268, %rd11259, %rd11237; + xor.b64 %rd11269, %rd11261, %rd11166; + xor.b64 %rd11270, %rd11263, %rd11169; + xor.b64 %rd11271, %rd11265, %rd11172; + xor.b64 %rd11272, %rd11267, %rd11175; + xor.b64 %rd11273, %rd11259, %rd11178; + xor.b64 %rd11274, %rd11261, %rd11181; + xor.b64 %rd11275, %rd11263, %rd11184; + xor.b64 %rd11276, %rd11265, %rd11187; + xor.b64 %rd11277, %rd11267, %rd11190; + xor.b64 %rd11278, %rd11259, %rd11193; + xor.b64 %rd11279, %rd11261, %rd11196; + xor.b64 %rd11280, %rd11263, %rd11199; + xor.b64 %rd11281, %rd11265, %rd11202; + xor.b64 %rd11282, %rd11267, %rd11205; + xor.b64 %rd11283, %rd11259, %rd11208; + xor.b64 %rd11284, %rd11261, %rd11211; + xor.b64 %rd11285, %rd11263, %rd11214; + xor.b64 %rd11286, %rd11265, %rd11217; + xor.b64 %rd11287, %rd11267, %rd11220; + xor.b64 %rd11288, %rd11259, %rd11223; + xor.b64 %rd11289, %rd11261, %rd11226; + xor.b64 %rd11290, %rd11263, %rd11229; + xor.b64 %rd11291, %rd11265, %rd11232; + xor.b64 %rd11292, %rd11267, %rd11235; + { + .reg .b32 %dummy; + mov.b64 {%r7986,%dummy}, %rd11274; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7987}, %rd11274; + } + shf.r.wrap.b32 %r7988, %r7987, %r7986, 20; + shf.r.wrap.b32 %r7989, %r7986, %r7987, 20; + mov.b64 %rd11293, {%r7989, %r7988}; + { + .reg .b32 %dummy; + mov.b64 {%r7990,%dummy}, %rd11280; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7991}, %rd11280; + } + shf.r.wrap.b32 %r7992, %r7991, %r7990, 21; + shf.r.wrap.b32 %r7993, %r7990, %r7991, 21; + mov.b64 %rd11294, {%r7993, %r7992}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7994}, %rd11286; + } + { + .reg .b32 %dummy; + mov.b64 {%r7995,%dummy}, %rd11286; + } + shf.l.wrap.b32 %r7996, %r7995, %r7994, 21; + shf.l.wrap.b32 %r7997, %r7994, %r7995, 21; + mov.b64 %rd11295, {%r7997, %r7996}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r7998}, %rd11292; + } + { + .reg .b32 %dummy; + mov.b64 {%r7999,%dummy}, %rd11292; + } + shf.l.wrap.b32 %r8000, %r7999, %r7998, 14; + shf.l.wrap.b32 %r8001, %r7998, %r7999, 14; + mov.b64 %rd11296, {%r8001, %r8000}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8002}, %rd11271; + } + { + .reg .b32 %dummy; + mov.b64 {%r8003,%dummy}, %rd11271; + } + shf.l.wrap.b32 %r8004, %r8003, %r8002, 28; + shf.l.wrap.b32 %r8005, %r8002, %r8003, 28; + mov.b64 %rd11297, {%r8005, %r8004}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8006}, %rd11277; + } + { + .reg .b32 %dummy; + mov.b64 {%r8007,%dummy}, %rd11277; + } + shf.l.wrap.b32 %r8008, %r8007, %r8006, 20; + shf.l.wrap.b32 %r8009, %r8006, %r8007, 20; + mov.b64 %rd11298, {%r8009, %r8008}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8010}, %rd11278; + } + { + .reg .b32 %dummy; + mov.b64 {%r8011,%dummy}, %rd11278; + } + shf.l.wrap.b32 %r8012, %r8011, %r8010, 3; + shf.l.wrap.b32 %r8013, %r8010, %r8011, 3; + mov.b64 %rd11299, {%r8013, %r8012}; + { + .reg .b32 %dummy; + mov.b64 {%r8014,%dummy}, %rd11284; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8015}, %rd11284; + } + shf.r.wrap.b32 %r8016, %r8015, %r8014, 19; + shf.r.wrap.b32 %r8017, %r8014, %r8015, 19; + mov.b64 %rd11300, {%r8017, %r8016}; + { + .reg .b32 %dummy; + mov.b64 {%r8018,%dummy}, %rd11290; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8019}, %rd11290; + } + shf.r.wrap.b32 %r8020, %r8019, %r8018, 3; + shf.r.wrap.b32 %r8021, %r8018, %r8019, 3; + mov.b64 %rd11301, {%r8021, %r8020}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8022}, %rd11269; + } + { + .reg .b32 %dummy; + mov.b64 {%r8023,%dummy}, %rd11269; + } + shf.l.wrap.b32 %r8024, %r8023, %r8022, 1; + shf.l.wrap.b32 %r8025, %r8022, %r8023, 1; + mov.b64 %rd11302, {%r8025, %r8024}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8026}, %rd11275; + } + { + .reg .b32 %dummy; + mov.b64 {%r8027,%dummy}, %rd11275; + } + shf.l.wrap.b32 %r8028, %r8027, %r8026, 6; + shf.l.wrap.b32 %r8029, %r8026, %r8027, 6; + mov.b64 %rd11303, {%r8029, %r8028}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8030}, %rd11281; + } + { + .reg .b32 %dummy; + mov.b64 {%r8031,%dummy}, %rd11281; + } + shf.l.wrap.b32 %r8032, %r8031, %r8030, 25; + shf.l.wrap.b32 %r8033, %r8030, %r8031, 25; + mov.b64 %rd11304, {%r8033, %r8032}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8034}, %rd11287; + } + { + .reg .b32 %dummy; + mov.b64 {%r8035,%dummy}, %rd11287; + } + shf.l.wrap.b32 %r8036, %r8035, %r8034, 8; + shf.l.wrap.b32 %r8037, %r8034, %r8035, 8; + mov.b64 %rd11305, {%r8037, %r8036}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8038}, %rd11288; + } + { + .reg .b32 %dummy; + mov.b64 {%r8039,%dummy}, %rd11288; + } + shf.l.wrap.b32 %r8040, %r8039, %r8038, 18; + shf.l.wrap.b32 %r8041, %r8038, %r8039, 18; + mov.b64 %rd11306, {%r8041, %r8040}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8042}, %rd11272; + } + { + .reg .b32 %dummy; + mov.b64 {%r8043,%dummy}, %rd11272; + } + shf.l.wrap.b32 %r8044, %r8043, %r8042, 27; + shf.l.wrap.b32 %r8045, %r8042, %r8043, 27; + mov.b64 %rd11307, {%r8045, %r8044}; + { + .reg .b32 %dummy; + mov.b64 {%r8046,%dummy}, %rd11273; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8047}, %rd11273; + } + shf.r.wrap.b32 %r8048, %r8047, %r8046, 28; + shf.r.wrap.b32 %r8049, %r8046, %r8047, 28; + mov.b64 %rd11308, {%r8049, %r8048}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8050}, %rd11279; + } + { + .reg .b32 %dummy; + mov.b64 {%r8051,%dummy}, %rd11279; + } + shf.l.wrap.b32 %r8052, %r8051, %r8050, 10; + shf.l.wrap.b32 %r8053, %r8050, %r8051, 10; + mov.b64 %rd11309, {%r8053, %r8052}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8054}, %rd11285; + } + { + .reg .b32 %dummy; + mov.b64 {%r8055,%dummy}, %rd11285; + } + shf.l.wrap.b32 %r8056, %r8055, %r8054, 15; + shf.l.wrap.b32 %r8057, %r8054, %r8055, 15; + mov.b64 %rd11310, {%r8057, %r8056}; + { + .reg .b32 %dummy; + mov.b64 {%r8058,%dummy}, %rd11291; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8059}, %rd11291; + } + shf.r.wrap.b32 %r8060, %r8059, %r8058, 8; + shf.r.wrap.b32 %r8061, %r8058, %r8059, 8; + mov.b64 %rd11311, {%r8061, %r8060}; + { + .reg .b32 %dummy; + mov.b64 {%r8062,%dummy}, %rd11270; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8063}, %rd11270; + } + shf.r.wrap.b32 %r8064, %r8063, %r8062, 2; + shf.r.wrap.b32 %r8065, %r8062, %r8063, 2; + mov.b64 %rd11312, {%r8065, %r8064}; + { + .reg .b32 %dummy; + mov.b64 {%r8066,%dummy}, %rd11276; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8067}, %rd11276; + } + shf.r.wrap.b32 %r8068, %r8067, %r8066, 9; + shf.r.wrap.b32 %r8069, %r8066, %r8067, 9; + mov.b64 %rd11313, {%r8069, %r8068}; + { + .reg .b32 %dummy; + mov.b64 {%r8070,%dummy}, %rd11282; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8071}, %rd11282; + } + shf.r.wrap.b32 %r8072, %r8071, %r8070, 25; + shf.r.wrap.b32 %r8073, %r8070, %r8071, 25; + mov.b64 %rd11314, {%r8073, %r8072}; + { + .reg .b32 %dummy; + mov.b64 {%r8074,%dummy}, %rd11283; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8075}, %rd11283; + } + shf.r.wrap.b32 %r8076, %r8075, %r8074, 23; + shf.r.wrap.b32 %r8077, %r8074, %r8075, 23; + mov.b64 %rd11315, {%r8077, %r8076}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8078}, %rd11289; + } + { + .reg .b32 %dummy; + mov.b64 {%r8079,%dummy}, %rd11289; + } + shf.l.wrap.b32 %r8080, %r8079, %r8078, 2; + shf.l.wrap.b32 %r8081, %r8078, %r8079, 2; + mov.b64 %rd11316, {%r8081, %r8080}; + not.b64 %rd11317, %rd11293; + and.b64 %rd11318, %rd11294, %rd11317; + xor.b64 %rd11319, %rd11318, %rd11268; + not.b64 %rd11320, %rd11294; + and.b64 %rd11321, %rd11295, %rd11320; + xor.b64 %rd11322, %rd11321, %rd11293; + not.b64 %rd11323, %rd11295; + and.b64 %rd11324, %rd11296, %rd11323; + xor.b64 %rd11325, %rd11324, %rd11294; + not.b64 %rd11326, %rd11296; + and.b64 %rd11327, %rd11268, %rd11326; + xor.b64 %rd11328, %rd11327, %rd11295; + not.b64 %rd11329, %rd11268; + and.b64 %rd11330, %rd11293, %rd11329; + xor.b64 %rd11331, %rd11330, %rd11296; + not.b64 %rd11332, %rd11298; + and.b64 %rd11333, %rd11299, %rd11332; + xor.b64 %rd11334, %rd11333, %rd11297; + not.b64 %rd11335, %rd11299; + and.b64 %rd11336, %rd11300, %rd11335; + xor.b64 %rd11337, %rd11336, %rd11298; + not.b64 %rd11338, %rd11300; + and.b64 %rd11339, %rd11301, %rd11338; + xor.b64 %rd11340, %rd11339, %rd11299; + not.b64 %rd11341, %rd11301; + and.b64 %rd11342, %rd11297, %rd11341; + xor.b64 %rd11343, %rd11342, %rd11300; + not.b64 %rd11344, %rd11297; + and.b64 %rd11345, %rd11298, %rd11344; + xor.b64 %rd11346, %rd11345, %rd11301; + not.b64 %rd11347, %rd11303; + and.b64 %rd11348, %rd11304, %rd11347; + xor.b64 %rd11349, %rd11348, %rd11302; + not.b64 %rd11350, %rd11304; + and.b64 %rd11351, %rd11305, %rd11350; + xor.b64 %rd11352, %rd11351, %rd11303; + not.b64 %rd11353, %rd11305; + and.b64 %rd11354, %rd11306, %rd11353; + xor.b64 %rd11355, %rd11354, %rd11304; + not.b64 %rd11356, %rd11306; + and.b64 %rd11357, %rd11302, %rd11356; + xor.b64 %rd11358, %rd11357, %rd11305; + not.b64 %rd11359, %rd11302; + and.b64 %rd11360, %rd11303, %rd11359; + xor.b64 %rd11361, %rd11360, %rd11306; + not.b64 %rd11362, %rd11308; + and.b64 %rd11363, %rd11309, %rd11362; + xor.b64 %rd11364, %rd11363, %rd11307; + not.b64 %rd11365, %rd11309; + and.b64 %rd11366, %rd11310, %rd11365; + xor.b64 %rd11367, %rd11366, %rd11308; + not.b64 %rd11368, %rd11310; + and.b64 %rd11369, %rd11311, %rd11368; + xor.b64 %rd11370, %rd11369, %rd11309; + not.b64 %rd11371, %rd11311; + and.b64 %rd11372, %rd11307, %rd11371; + xor.b64 %rd11373, %rd11372, %rd11310; + not.b64 %rd11374, %rd11307; + and.b64 %rd11375, %rd11308, %rd11374; + xor.b64 %rd11376, %rd11375, %rd11311; + not.b64 %rd11377, %rd11313; + and.b64 %rd11378, %rd11314, %rd11377; + xor.b64 %rd11379, %rd11378, %rd11312; + not.b64 %rd11380, %rd11314; + and.b64 %rd11381, %rd11315, %rd11380; + xor.b64 %rd11382, %rd11381, %rd11313; + not.b64 %rd11383, %rd11315; + and.b64 %rd11384, %rd11316, %rd11383; + xor.b64 %rd11385, %rd11384, %rd11314; + not.b64 %rd11386, %rd11316; + and.b64 %rd11387, %rd11312, %rd11386; + xor.b64 %rd11388, %rd11387, %rd11315; + not.b64 %rd11389, %rd11312; + and.b64 %rd11390, %rd11313, %rd11389; + xor.b64 %rd11391, %rd11390, %rd11316; + xor.b64 %rd11393, %rd11319, %rd36; + xor.b64 %rd11394, %rd11334, %rd11393; + xor.b64 %rd11395, %rd11394, %rd11349; + xor.b64 %rd11396, %rd11395, %rd11364; + xor.b64 %rd11397, %rd11396, %rd11379; + xor.b64 %rd11398, %rd11337, %rd11322; + xor.b64 %rd11399, %rd11398, %rd11352; + xor.b64 %rd11400, %rd11399, %rd11367; + xor.b64 %rd11401, %rd11400, %rd11382; + xor.b64 %rd11402, %rd11340, %rd11325; + xor.b64 %rd11403, %rd11402, %rd11355; + xor.b64 %rd11404, %rd11403, %rd11370; + xor.b64 %rd11405, %rd11404, %rd11385; + xor.b64 %rd11406, %rd11343, %rd11328; + xor.b64 %rd11407, %rd11406, %rd11358; + xor.b64 %rd11408, %rd11407, %rd11373; + xor.b64 %rd11409, %rd11408, %rd11388; + xor.b64 %rd11410, %rd11346, %rd11331; + xor.b64 %rd11411, %rd11410, %rd11361; + xor.b64 %rd11412, %rd11411, %rd11376; + xor.b64 %rd11413, %rd11412, %rd11391; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8082}, %rd11401; + } + { + .reg .b32 %dummy; + mov.b64 {%r8083,%dummy}, %rd11401; + } + shf.l.wrap.b32 %r8084, %r8083, %r8082, 1; + shf.l.wrap.b32 %r8085, %r8082, %r8083, 1; + mov.b64 %rd11414, {%r8085, %r8084}; + xor.b64 %rd11415, %rd11413, %rd11414; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8086}, %rd11405; + } + { + .reg .b32 %dummy; + mov.b64 {%r8087,%dummy}, %rd11405; + } + shf.l.wrap.b32 %r8088, %r8087, %r8086, 1; + shf.l.wrap.b32 %r8089, %r8086, %r8087, 1; + mov.b64 %rd11416, {%r8089, %r8088}; + xor.b64 %rd11417, %rd11416, %rd11397; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8090}, %rd11409; + } + { + .reg .b32 %dummy; + mov.b64 {%r8091,%dummy}, %rd11409; + } + shf.l.wrap.b32 %r8092, %r8091, %r8090, 1; + shf.l.wrap.b32 %r8093, %r8090, %r8091, 1; + mov.b64 %rd11418, {%r8093, %r8092}; + xor.b64 %rd11419, %rd11418, %rd11401; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8094}, %rd11413; + } + { + .reg .b32 %dummy; + mov.b64 {%r8095,%dummy}, %rd11413; + } + shf.l.wrap.b32 %r8096, %r8095, %r8094, 1; + shf.l.wrap.b32 %r8097, %r8094, %r8095, 1; + mov.b64 %rd11420, {%r8097, %r8096}; + xor.b64 %rd11421, %rd11420, %rd11405; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8098}, %rd11397; + } + { + .reg .b32 %dummy; + mov.b64 {%r8099,%dummy}, %rd11397; + } + shf.l.wrap.b32 %r8100, %r8099, %r8098, 1; + shf.l.wrap.b32 %r8101, %r8098, %r8099, 1; + mov.b64 %rd11422, {%r8101, %r8100}; + xor.b64 %rd11423, %rd11409, %rd11422; + xor.b64 %rd11424, %rd11415, %rd11393; + xor.b64 %rd11425, %rd11417, %rd11322; + xor.b64 %rd11426, %rd11419, %rd11325; + xor.b64 %rd11427, %rd11421, %rd11328; + xor.b64 %rd11428, %rd11423, %rd11331; + xor.b64 %rd11429, %rd11415, %rd11334; + xor.b64 %rd11430, %rd11417, %rd11337; + xor.b64 %rd11431, %rd11419, %rd11340; + xor.b64 %rd11432, %rd11421, %rd11343; + xor.b64 %rd11433, %rd11423, %rd11346; + xor.b64 %rd11434, %rd11415, %rd11349; + xor.b64 %rd11435, %rd11417, %rd11352; + xor.b64 %rd11436, %rd11419, %rd11355; + xor.b64 %rd11437, %rd11421, %rd11358; + xor.b64 %rd11438, %rd11423, %rd11361; + xor.b64 %rd11439, %rd11415, %rd11364; + xor.b64 %rd11440, %rd11417, %rd11367; + xor.b64 %rd11441, %rd11419, %rd11370; + xor.b64 %rd11442, %rd11421, %rd11373; + xor.b64 %rd11443, %rd11423, %rd11376; + xor.b64 %rd11444, %rd11415, %rd11379; + xor.b64 %rd11445, %rd11417, %rd11382; + xor.b64 %rd11446, %rd11419, %rd11385; + xor.b64 %rd11447, %rd11421, %rd11388; + xor.b64 %rd11448, %rd11423, %rd11391; + { + .reg .b32 %dummy; + mov.b64 {%r8102,%dummy}, %rd11430; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8103}, %rd11430; + } + shf.r.wrap.b32 %r8104, %r8103, %r8102, 20; + shf.r.wrap.b32 %r8105, %r8102, %r8103, 20; + mov.b64 %rd11449, {%r8105, %r8104}; + { + .reg .b32 %dummy; + mov.b64 {%r8106,%dummy}, %rd11436; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8107}, %rd11436; + } + shf.r.wrap.b32 %r8108, %r8107, %r8106, 21; + shf.r.wrap.b32 %r8109, %r8106, %r8107, 21; + mov.b64 %rd11450, {%r8109, %r8108}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8110}, %rd11442; + } + { + .reg .b32 %dummy; + mov.b64 {%r8111,%dummy}, %rd11442; + } + shf.l.wrap.b32 %r8112, %r8111, %r8110, 21; + shf.l.wrap.b32 %r8113, %r8110, %r8111, 21; + mov.b64 %rd11451, {%r8113, %r8112}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8114}, %rd11448; + } + { + .reg .b32 %dummy; + mov.b64 {%r8115,%dummy}, %rd11448; + } + shf.l.wrap.b32 %r8116, %r8115, %r8114, 14; + shf.l.wrap.b32 %r8117, %r8114, %r8115, 14; + mov.b64 %rd11452, {%r8117, %r8116}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8118}, %rd11427; + } + { + .reg .b32 %dummy; + mov.b64 {%r8119,%dummy}, %rd11427; + } + shf.l.wrap.b32 %r8120, %r8119, %r8118, 28; + shf.l.wrap.b32 %r8121, %r8118, %r8119, 28; + mov.b64 %rd11453, {%r8121, %r8120}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8122}, %rd11433; + } + { + .reg .b32 %dummy; + mov.b64 {%r8123,%dummy}, %rd11433; + } + shf.l.wrap.b32 %r8124, %r8123, %r8122, 20; + shf.l.wrap.b32 %r8125, %r8122, %r8123, 20; + mov.b64 %rd11454, {%r8125, %r8124}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8126}, %rd11434; + } + { + .reg .b32 %dummy; + mov.b64 {%r8127,%dummy}, %rd11434; + } + shf.l.wrap.b32 %r8128, %r8127, %r8126, 3; + shf.l.wrap.b32 %r8129, %r8126, %r8127, 3; + mov.b64 %rd11455, {%r8129, %r8128}; + { + .reg .b32 %dummy; + mov.b64 {%r8130,%dummy}, %rd11440; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8131}, %rd11440; + } + shf.r.wrap.b32 %r8132, %r8131, %r8130, 19; + shf.r.wrap.b32 %r8133, %r8130, %r8131, 19; + mov.b64 %rd11456, {%r8133, %r8132}; + { + .reg .b32 %dummy; + mov.b64 {%r8134,%dummy}, %rd11446; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8135}, %rd11446; + } + shf.r.wrap.b32 %r8136, %r8135, %r8134, 3; + shf.r.wrap.b32 %r8137, %r8134, %r8135, 3; + mov.b64 %rd11457, {%r8137, %r8136}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8138}, %rd11425; + } + { + .reg .b32 %dummy; + mov.b64 {%r8139,%dummy}, %rd11425; + } + shf.l.wrap.b32 %r8140, %r8139, %r8138, 1; + shf.l.wrap.b32 %r8141, %r8138, %r8139, 1; + mov.b64 %rd11458, {%r8141, %r8140}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8142}, %rd11431; + } + { + .reg .b32 %dummy; + mov.b64 {%r8143,%dummy}, %rd11431; + } + shf.l.wrap.b32 %r8144, %r8143, %r8142, 6; + shf.l.wrap.b32 %r8145, %r8142, %r8143, 6; + mov.b64 %rd11459, {%r8145, %r8144}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8146}, %rd11437; + } + { + .reg .b32 %dummy; + mov.b64 {%r8147,%dummy}, %rd11437; + } + shf.l.wrap.b32 %r8148, %r8147, %r8146, 25; + shf.l.wrap.b32 %r8149, %r8146, %r8147, 25; + mov.b64 %rd11460, {%r8149, %r8148}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8150}, %rd11443; + } + { + .reg .b32 %dummy; + mov.b64 {%r8151,%dummy}, %rd11443; + } + shf.l.wrap.b32 %r8152, %r8151, %r8150, 8; + shf.l.wrap.b32 %r8153, %r8150, %r8151, 8; + mov.b64 %rd11461, {%r8153, %r8152}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8154}, %rd11444; + } + { + .reg .b32 %dummy; + mov.b64 {%r8155,%dummy}, %rd11444; + } + shf.l.wrap.b32 %r8156, %r8155, %r8154, 18; + shf.l.wrap.b32 %r8157, %r8154, %r8155, 18; + mov.b64 %rd11462, {%r8157, %r8156}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8158}, %rd11428; + } + { + .reg .b32 %dummy; + mov.b64 {%r8159,%dummy}, %rd11428; + } + shf.l.wrap.b32 %r8160, %r8159, %r8158, 27; + shf.l.wrap.b32 %r8161, %r8158, %r8159, 27; + mov.b64 %rd11463, {%r8161, %r8160}; + { + .reg .b32 %dummy; + mov.b64 {%r8162,%dummy}, %rd11429; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8163}, %rd11429; + } + shf.r.wrap.b32 %r8164, %r8163, %r8162, 28; + shf.r.wrap.b32 %r8165, %r8162, %r8163, 28; + mov.b64 %rd11464, {%r8165, %r8164}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8166}, %rd11435; + } + { + .reg .b32 %dummy; + mov.b64 {%r8167,%dummy}, %rd11435; + } + shf.l.wrap.b32 %r8168, %r8167, %r8166, 10; + shf.l.wrap.b32 %r8169, %r8166, %r8167, 10; + mov.b64 %rd11465, {%r8169, %r8168}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8170}, %rd11441; + } + { + .reg .b32 %dummy; + mov.b64 {%r8171,%dummy}, %rd11441; + } + shf.l.wrap.b32 %r8172, %r8171, %r8170, 15; + shf.l.wrap.b32 %r8173, %r8170, %r8171, 15; + mov.b64 %rd11466, {%r8173, %r8172}; + { + .reg .b32 %dummy; + mov.b64 {%r8174,%dummy}, %rd11447; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8175}, %rd11447; + } + shf.r.wrap.b32 %r8176, %r8175, %r8174, 8; + shf.r.wrap.b32 %r8177, %r8174, %r8175, 8; + mov.b64 %rd11467, {%r8177, %r8176}; + { + .reg .b32 %dummy; + mov.b64 {%r8178,%dummy}, %rd11426; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8179}, %rd11426; + } + shf.r.wrap.b32 %r8180, %r8179, %r8178, 2; + shf.r.wrap.b32 %r8181, %r8178, %r8179, 2; + mov.b64 %rd11468, {%r8181, %r8180}; + { + .reg .b32 %dummy; + mov.b64 {%r8182,%dummy}, %rd11432; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8183}, %rd11432; + } + shf.r.wrap.b32 %r8184, %r8183, %r8182, 9; + shf.r.wrap.b32 %r8185, %r8182, %r8183, 9; + mov.b64 %rd11469, {%r8185, %r8184}; + { + .reg .b32 %dummy; + mov.b64 {%r8186,%dummy}, %rd11438; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8187}, %rd11438; + } + shf.r.wrap.b32 %r8188, %r8187, %r8186, 25; + shf.r.wrap.b32 %r8189, %r8186, %r8187, 25; + mov.b64 %rd11470, {%r8189, %r8188}; + { + .reg .b32 %dummy; + mov.b64 {%r8190,%dummy}, %rd11439; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8191}, %rd11439; + } + shf.r.wrap.b32 %r8192, %r8191, %r8190, 23; + shf.r.wrap.b32 %r8193, %r8190, %r8191, 23; + mov.b64 %rd11471, {%r8193, %r8192}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8194}, %rd11445; + } + { + .reg .b32 %dummy; + mov.b64 {%r8195,%dummy}, %rd11445; + } + shf.l.wrap.b32 %r8196, %r8195, %r8194, 2; + shf.l.wrap.b32 %r8197, %r8194, %r8195, 2; + mov.b64 %rd11472, {%r8197, %r8196}; + not.b64 %rd11473, %rd11449; + and.b64 %rd11474, %rd11450, %rd11473; + xor.b64 %rd11475, %rd11474, %rd11424; + not.b64 %rd11476, %rd11450; + and.b64 %rd11477, %rd11451, %rd11476; + xor.b64 %rd11478, %rd11477, %rd11449; + not.b64 %rd11479, %rd11451; + and.b64 %rd11480, %rd11452, %rd11479; + xor.b64 %rd11481, %rd11480, %rd11450; + not.b64 %rd11482, %rd11452; + and.b64 %rd11483, %rd11424, %rd11482; + xor.b64 %rd11484, %rd11483, %rd11451; + not.b64 %rd11485, %rd11424; + and.b64 %rd11486, %rd11449, %rd11485; + xor.b64 %rd11487, %rd11486, %rd11452; + not.b64 %rd11488, %rd11454; + and.b64 %rd11489, %rd11455, %rd11488; + xor.b64 %rd11490, %rd11489, %rd11453; + not.b64 %rd11491, %rd11455; + and.b64 %rd11492, %rd11456, %rd11491; + xor.b64 %rd11493, %rd11492, %rd11454; + not.b64 %rd11494, %rd11456; + and.b64 %rd11495, %rd11457, %rd11494; + xor.b64 %rd11496, %rd11495, %rd11455; + not.b64 %rd11497, %rd11457; + and.b64 %rd11498, %rd11453, %rd11497; + xor.b64 %rd11499, %rd11498, %rd11456; + not.b64 %rd11500, %rd11453; + and.b64 %rd11501, %rd11454, %rd11500; + xor.b64 %rd11502, %rd11501, %rd11457; + not.b64 %rd11503, %rd11459; + and.b64 %rd11504, %rd11460, %rd11503; + xor.b64 %rd11505, %rd11504, %rd11458; + not.b64 %rd11506, %rd11460; + and.b64 %rd11507, %rd11461, %rd11506; + xor.b64 %rd11508, %rd11507, %rd11459; + not.b64 %rd11509, %rd11461; + and.b64 %rd11510, %rd11462, %rd11509; + xor.b64 %rd11511, %rd11510, %rd11460; + not.b64 %rd11512, %rd11462; + and.b64 %rd11513, %rd11458, %rd11512; + xor.b64 %rd11514, %rd11513, %rd11461; + not.b64 %rd11515, %rd11458; + and.b64 %rd11516, %rd11459, %rd11515; + xor.b64 %rd11517, %rd11516, %rd11462; + not.b64 %rd11518, %rd11464; + and.b64 %rd11519, %rd11465, %rd11518; + xor.b64 %rd11520, %rd11519, %rd11463; + not.b64 %rd11521, %rd11465; + and.b64 %rd11522, %rd11466, %rd11521; + xor.b64 %rd11523, %rd11522, %rd11464; + not.b64 %rd11524, %rd11466; + and.b64 %rd11525, %rd11467, %rd11524; + xor.b64 %rd11526, %rd11525, %rd11465; + not.b64 %rd11527, %rd11467; + and.b64 %rd11528, %rd11463, %rd11527; + xor.b64 %rd11529, %rd11528, %rd11466; + not.b64 %rd11530, %rd11463; + and.b64 %rd11531, %rd11464, %rd11530; + xor.b64 %rd11532, %rd11531, %rd11467; + not.b64 %rd11533, %rd11469; + and.b64 %rd11534, %rd11470, %rd11533; + xor.b64 %rd11535, %rd11534, %rd11468; + not.b64 %rd11536, %rd11470; + and.b64 %rd11537, %rd11471, %rd11536; + xor.b64 %rd11538, %rd11537, %rd11469; + not.b64 %rd11539, %rd11471; + and.b64 %rd11540, %rd11472, %rd11539; + xor.b64 %rd11541, %rd11540, %rd11470; + not.b64 %rd11542, %rd11472; + and.b64 %rd11543, %rd11468, %rd11542; + xor.b64 %rd11544, %rd11543, %rd11471; + not.b64 %rd11545, %rd11468; + and.b64 %rd11546, %rd11469, %rd11545; + xor.b64 %rd11547, %rd11546, %rd11472; + xor.b64 %rd11549, %rd11475, %rd37; + xor.b64 %rd11550, %rd11490, %rd11549; + xor.b64 %rd11551, %rd11550, %rd11505; + xor.b64 %rd11552, %rd11551, %rd11520; + xor.b64 %rd11553, %rd11552, %rd11535; + xor.b64 %rd11554, %rd11493, %rd11478; + xor.b64 %rd11555, %rd11554, %rd11508; + xor.b64 %rd11556, %rd11555, %rd11523; + xor.b64 %rd11557, %rd11556, %rd11538; + xor.b64 %rd11558, %rd11496, %rd11481; + xor.b64 %rd11559, %rd11558, %rd11511; + xor.b64 %rd11560, %rd11559, %rd11526; + xor.b64 %rd11561, %rd11560, %rd11541; + xor.b64 %rd11562, %rd11499, %rd11484; + xor.b64 %rd11563, %rd11562, %rd11514; + xor.b64 %rd11564, %rd11563, %rd11529; + xor.b64 %rd11565, %rd11564, %rd11544; + xor.b64 %rd11566, %rd11502, %rd11487; + xor.b64 %rd11567, %rd11566, %rd11517; + xor.b64 %rd11568, %rd11567, %rd11532; + xor.b64 %rd11569, %rd11568, %rd11547; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8198}, %rd11557; + } + { + .reg .b32 %dummy; + mov.b64 {%r8199,%dummy}, %rd11557; + } + shf.l.wrap.b32 %r8200, %r8199, %r8198, 1; + shf.l.wrap.b32 %r8201, %r8198, %r8199, 1; + mov.b64 %rd11570, {%r8201, %r8200}; + xor.b64 %rd11571, %rd11569, %rd11570; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8202}, %rd11561; + } + { + .reg .b32 %dummy; + mov.b64 {%r8203,%dummy}, %rd11561; + } + shf.l.wrap.b32 %r8204, %r8203, %r8202, 1; + shf.l.wrap.b32 %r8205, %r8202, %r8203, 1; + mov.b64 %rd11572, {%r8205, %r8204}; + xor.b64 %rd11573, %rd11572, %rd11553; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8206}, %rd11565; + } + { + .reg .b32 %dummy; + mov.b64 {%r8207,%dummy}, %rd11565; + } + shf.l.wrap.b32 %r8208, %r8207, %r8206, 1; + shf.l.wrap.b32 %r8209, %r8206, %r8207, 1; + mov.b64 %rd11574, {%r8209, %r8208}; + xor.b64 %rd11575, %rd11574, %rd11557; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8210}, %rd11569; + } + { + .reg .b32 %dummy; + mov.b64 {%r8211,%dummy}, %rd11569; + } + shf.l.wrap.b32 %r8212, %r8211, %r8210, 1; + shf.l.wrap.b32 %r8213, %r8210, %r8211, 1; + mov.b64 %rd11576, {%r8213, %r8212}; + xor.b64 %rd11577, %rd11576, %rd11561; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8214}, %rd11553; + } + { + .reg .b32 %dummy; + mov.b64 {%r8215,%dummy}, %rd11553; + } + shf.l.wrap.b32 %r8216, %r8215, %r8214, 1; + shf.l.wrap.b32 %r8217, %r8214, %r8215, 1; + mov.b64 %rd11578, {%r8217, %r8216}; + xor.b64 %rd11579, %rd11565, %rd11578; + xor.b64 %rd11580, %rd11571, %rd11549; + xor.b64 %rd11581, %rd11573, %rd11478; + xor.b64 %rd11582, %rd11575, %rd11481; + xor.b64 %rd11583, %rd11577, %rd11484; + xor.b64 %rd11584, %rd11579, %rd11487; + xor.b64 %rd11585, %rd11571, %rd11490; + xor.b64 %rd11586, %rd11573, %rd11493; + xor.b64 %rd11587, %rd11575, %rd11496; + xor.b64 %rd11588, %rd11577, %rd11499; + xor.b64 %rd11589, %rd11579, %rd11502; + xor.b64 %rd11590, %rd11571, %rd11505; + xor.b64 %rd11591, %rd11573, %rd11508; + xor.b64 %rd11592, %rd11575, %rd11511; + xor.b64 %rd11593, %rd11577, %rd11514; + xor.b64 %rd11594, %rd11579, %rd11517; + xor.b64 %rd11595, %rd11571, %rd11520; + xor.b64 %rd11596, %rd11573, %rd11523; + xor.b64 %rd11597, %rd11575, %rd11526; + xor.b64 %rd11598, %rd11577, %rd11529; + xor.b64 %rd11599, %rd11579, %rd11532; + xor.b64 %rd11600, %rd11571, %rd11535; + xor.b64 %rd11601, %rd11573, %rd11538; + xor.b64 %rd11602, %rd11575, %rd11541; + xor.b64 %rd11603, %rd11577, %rd11544; + xor.b64 %rd11604, %rd11579, %rd11547; + { + .reg .b32 %dummy; + mov.b64 {%r8218,%dummy}, %rd11586; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8219}, %rd11586; + } + shf.r.wrap.b32 %r8220, %r8219, %r8218, 20; + shf.r.wrap.b32 %r8221, %r8218, %r8219, 20; + mov.b64 %rd11605, {%r8221, %r8220}; + { + .reg .b32 %dummy; + mov.b64 {%r8222,%dummy}, %rd11592; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8223}, %rd11592; + } + shf.r.wrap.b32 %r8224, %r8223, %r8222, 21; + shf.r.wrap.b32 %r8225, %r8222, %r8223, 21; + mov.b64 %rd11606, {%r8225, %r8224}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8226}, %rd11598; + } + { + .reg .b32 %dummy; + mov.b64 {%r8227,%dummy}, %rd11598; + } + shf.l.wrap.b32 %r8228, %r8227, %r8226, 21; + shf.l.wrap.b32 %r8229, %r8226, %r8227, 21; + mov.b64 %rd11607, {%r8229, %r8228}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8230}, %rd11604; + } + { + .reg .b32 %dummy; + mov.b64 {%r8231,%dummy}, %rd11604; + } + shf.l.wrap.b32 %r8232, %r8231, %r8230, 14; + shf.l.wrap.b32 %r8233, %r8230, %r8231, 14; + mov.b64 %rd11608, {%r8233, %r8232}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8234}, %rd11583; + } + { + .reg .b32 %dummy; + mov.b64 {%r8235,%dummy}, %rd11583; + } + shf.l.wrap.b32 %r8236, %r8235, %r8234, 28; + shf.l.wrap.b32 %r8237, %r8234, %r8235, 28; + mov.b64 %rd11609, {%r8237, %r8236}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8238}, %rd11589; + } + { + .reg .b32 %dummy; + mov.b64 {%r8239,%dummy}, %rd11589; + } + shf.l.wrap.b32 %r8240, %r8239, %r8238, 20; + shf.l.wrap.b32 %r8241, %r8238, %r8239, 20; + mov.b64 %rd11610, {%r8241, %r8240}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8242}, %rd11590; + } + { + .reg .b32 %dummy; + mov.b64 {%r8243,%dummy}, %rd11590; + } + shf.l.wrap.b32 %r8244, %r8243, %r8242, 3; + shf.l.wrap.b32 %r8245, %r8242, %r8243, 3; + mov.b64 %rd11611, {%r8245, %r8244}; + { + .reg .b32 %dummy; + mov.b64 {%r8246,%dummy}, %rd11596; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8247}, %rd11596; + } + shf.r.wrap.b32 %r8248, %r8247, %r8246, 19; + shf.r.wrap.b32 %r8249, %r8246, %r8247, 19; + mov.b64 %rd11612, {%r8249, %r8248}; + { + .reg .b32 %dummy; + mov.b64 {%r8250,%dummy}, %rd11602; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8251}, %rd11602; + } + shf.r.wrap.b32 %r8252, %r8251, %r8250, 3; + shf.r.wrap.b32 %r8253, %r8250, %r8251, 3; + mov.b64 %rd11613, {%r8253, %r8252}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8254}, %rd11581; + } + { + .reg .b32 %dummy; + mov.b64 {%r8255,%dummy}, %rd11581; + } + shf.l.wrap.b32 %r8256, %r8255, %r8254, 1; + shf.l.wrap.b32 %r8257, %r8254, %r8255, 1; + mov.b64 %rd11614, {%r8257, %r8256}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8258}, %rd11587; + } + { + .reg .b32 %dummy; + mov.b64 {%r8259,%dummy}, %rd11587; + } + shf.l.wrap.b32 %r8260, %r8259, %r8258, 6; + shf.l.wrap.b32 %r8261, %r8258, %r8259, 6; + mov.b64 %rd11615, {%r8261, %r8260}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8262}, %rd11593; + } + { + .reg .b32 %dummy; + mov.b64 {%r8263,%dummy}, %rd11593; + } + shf.l.wrap.b32 %r8264, %r8263, %r8262, 25; + shf.l.wrap.b32 %r8265, %r8262, %r8263, 25; + mov.b64 %rd11616, {%r8265, %r8264}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8266}, %rd11599; + } + { + .reg .b32 %dummy; + mov.b64 {%r8267,%dummy}, %rd11599; + } + shf.l.wrap.b32 %r8268, %r8267, %r8266, 8; + shf.l.wrap.b32 %r8269, %r8266, %r8267, 8; + mov.b64 %rd11617, {%r8269, %r8268}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8270}, %rd11600; + } + { + .reg .b32 %dummy; + mov.b64 {%r8271,%dummy}, %rd11600; + } + shf.l.wrap.b32 %r8272, %r8271, %r8270, 18; + shf.l.wrap.b32 %r8273, %r8270, %r8271, 18; + mov.b64 %rd11618, {%r8273, %r8272}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8274}, %rd11584; + } + { + .reg .b32 %dummy; + mov.b64 {%r8275,%dummy}, %rd11584; + } + shf.l.wrap.b32 %r8276, %r8275, %r8274, 27; + shf.l.wrap.b32 %r8277, %r8274, %r8275, 27; + mov.b64 %rd11619, {%r8277, %r8276}; + { + .reg .b32 %dummy; + mov.b64 {%r8278,%dummy}, %rd11585; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8279}, %rd11585; + } + shf.r.wrap.b32 %r8280, %r8279, %r8278, 28; + shf.r.wrap.b32 %r8281, %r8278, %r8279, 28; + mov.b64 %rd11620, {%r8281, %r8280}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8282}, %rd11591; + } + { + .reg .b32 %dummy; + mov.b64 {%r8283,%dummy}, %rd11591; + } + shf.l.wrap.b32 %r8284, %r8283, %r8282, 10; + shf.l.wrap.b32 %r8285, %r8282, %r8283, 10; + mov.b64 %rd11621, {%r8285, %r8284}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8286}, %rd11597; + } + { + .reg .b32 %dummy; + mov.b64 {%r8287,%dummy}, %rd11597; + } + shf.l.wrap.b32 %r8288, %r8287, %r8286, 15; + shf.l.wrap.b32 %r8289, %r8286, %r8287, 15; + mov.b64 %rd11622, {%r8289, %r8288}; + { + .reg .b32 %dummy; + mov.b64 {%r8290,%dummy}, %rd11603; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8291}, %rd11603; + } + shf.r.wrap.b32 %r8292, %r8291, %r8290, 8; + shf.r.wrap.b32 %r8293, %r8290, %r8291, 8; + mov.b64 %rd11623, {%r8293, %r8292}; + { + .reg .b32 %dummy; + mov.b64 {%r8294,%dummy}, %rd11582; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8295}, %rd11582; + } + shf.r.wrap.b32 %r8296, %r8295, %r8294, 2; + shf.r.wrap.b32 %r8297, %r8294, %r8295, 2; + mov.b64 %rd11624, {%r8297, %r8296}; + { + .reg .b32 %dummy; + mov.b64 {%r8298,%dummy}, %rd11588; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8299}, %rd11588; + } + shf.r.wrap.b32 %r8300, %r8299, %r8298, 9; + shf.r.wrap.b32 %r8301, %r8298, %r8299, 9; + mov.b64 %rd11625, {%r8301, %r8300}; + { + .reg .b32 %dummy; + mov.b64 {%r8302,%dummy}, %rd11594; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8303}, %rd11594; + } + shf.r.wrap.b32 %r8304, %r8303, %r8302, 25; + shf.r.wrap.b32 %r8305, %r8302, %r8303, 25; + mov.b64 %rd11626, {%r8305, %r8304}; + { + .reg .b32 %dummy; + mov.b64 {%r8306,%dummy}, %rd11595; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8307}, %rd11595; + } + shf.r.wrap.b32 %r8308, %r8307, %r8306, 23; + shf.r.wrap.b32 %r8309, %r8306, %r8307, 23; + mov.b64 %rd11627, {%r8309, %r8308}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8310}, %rd11601; + } + { + .reg .b32 %dummy; + mov.b64 {%r8311,%dummy}, %rd11601; + } + shf.l.wrap.b32 %r8312, %r8311, %r8310, 2; + shf.l.wrap.b32 %r8313, %r8310, %r8311, 2; + mov.b64 %rd11628, {%r8313, %r8312}; + not.b64 %rd11629, %rd11605; + and.b64 %rd11630, %rd11606, %rd11629; + xor.b64 %rd11631, %rd11630, %rd11580; + not.b64 %rd11632, %rd11606; + and.b64 %rd11633, %rd11607, %rd11632; + xor.b64 %rd11634, %rd11633, %rd11605; + not.b64 %rd11635, %rd11607; + and.b64 %rd11636, %rd11608, %rd11635; + xor.b64 %rd11637, %rd11636, %rd11606; + not.b64 %rd11638, %rd11608; + and.b64 %rd11639, %rd11580, %rd11638; + xor.b64 %rd11640, %rd11639, %rd11607; + not.b64 %rd11641, %rd11580; + and.b64 %rd11642, %rd11605, %rd11641; + xor.b64 %rd11643, %rd11642, %rd11608; + not.b64 %rd11644, %rd11610; + and.b64 %rd11645, %rd11611, %rd11644; + xor.b64 %rd11646, %rd11645, %rd11609; + not.b64 %rd11647, %rd11611; + and.b64 %rd11648, %rd11612, %rd11647; + xor.b64 %rd11649, %rd11648, %rd11610; + not.b64 %rd11650, %rd11612; + and.b64 %rd11651, %rd11613, %rd11650; + xor.b64 %rd11652, %rd11651, %rd11611; + not.b64 %rd11653, %rd11613; + and.b64 %rd11654, %rd11609, %rd11653; + xor.b64 %rd11655, %rd11654, %rd11612; + not.b64 %rd11656, %rd11609; + and.b64 %rd11657, %rd11610, %rd11656; + xor.b64 %rd11658, %rd11657, %rd11613; + not.b64 %rd11659, %rd11615; + and.b64 %rd11660, %rd11616, %rd11659; + xor.b64 %rd11661, %rd11660, %rd11614; + not.b64 %rd11662, %rd11616; + and.b64 %rd11663, %rd11617, %rd11662; + xor.b64 %rd11664, %rd11663, %rd11615; + not.b64 %rd11665, %rd11617; + and.b64 %rd11666, %rd11618, %rd11665; + xor.b64 %rd11667, %rd11666, %rd11616; + not.b64 %rd11668, %rd11618; + and.b64 %rd11669, %rd11614, %rd11668; + xor.b64 %rd11670, %rd11669, %rd11617; + not.b64 %rd11671, %rd11614; + and.b64 %rd11672, %rd11615, %rd11671; + xor.b64 %rd11673, %rd11672, %rd11618; + not.b64 %rd11674, %rd11620; + and.b64 %rd11675, %rd11621, %rd11674; + xor.b64 %rd11676, %rd11675, %rd11619; + not.b64 %rd11677, %rd11621; + and.b64 %rd11678, %rd11622, %rd11677; + xor.b64 %rd11679, %rd11678, %rd11620; + not.b64 %rd11680, %rd11622; + and.b64 %rd11681, %rd11623, %rd11680; + xor.b64 %rd11682, %rd11681, %rd11621; + not.b64 %rd11683, %rd11623; + and.b64 %rd11684, %rd11619, %rd11683; + xor.b64 %rd11685, %rd11684, %rd11622; + not.b64 %rd11686, %rd11619; + and.b64 %rd11687, %rd11620, %rd11686; + xor.b64 %rd11688, %rd11687, %rd11623; + not.b64 %rd11689, %rd11625; + and.b64 %rd11690, %rd11626, %rd11689; + xor.b64 %rd11691, %rd11690, %rd11624; + not.b64 %rd11692, %rd11626; + and.b64 %rd11693, %rd11627, %rd11692; + xor.b64 %rd11694, %rd11693, %rd11625; + not.b64 %rd11695, %rd11627; + and.b64 %rd11696, %rd11628, %rd11695; + xor.b64 %rd11697, %rd11696, %rd11626; + not.b64 %rd11698, %rd11628; + and.b64 %rd11699, %rd11624, %rd11698; + xor.b64 %rd11700, %rd11699, %rd11627; + not.b64 %rd11701, %rd11624; + and.b64 %rd11702, %rd11625, %rd11701; + xor.b64 %rd11703, %rd11702, %rd11628; + xor.b64 %rd11705, %rd11631, %rd38; + xor.b64 %rd11706, %rd11646, %rd11705; + xor.b64 %rd11707, %rd11706, %rd11661; + xor.b64 %rd11708, %rd11707, %rd11676; + xor.b64 %rd11709, %rd11708, %rd11691; + xor.b64 %rd11710, %rd11649, %rd11634; + xor.b64 %rd11711, %rd11710, %rd11664; + xor.b64 %rd11712, %rd11711, %rd11679; + xor.b64 %rd11713, %rd11712, %rd11694; + xor.b64 %rd11714, %rd11652, %rd11637; + xor.b64 %rd11715, %rd11714, %rd11667; + xor.b64 %rd11716, %rd11715, %rd11682; + xor.b64 %rd11717, %rd11716, %rd11697; + xor.b64 %rd11718, %rd11655, %rd11640; + xor.b64 %rd11719, %rd11718, %rd11670; + xor.b64 %rd11720, %rd11719, %rd11685; + xor.b64 %rd11721, %rd11720, %rd11700; + xor.b64 %rd11722, %rd11658, %rd11643; + xor.b64 %rd11723, %rd11722, %rd11673; + xor.b64 %rd11724, %rd11723, %rd11688; + xor.b64 %rd11725, %rd11724, %rd11703; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8314}, %rd11713; + } + { + .reg .b32 %dummy; + mov.b64 {%r8315,%dummy}, %rd11713; + } + shf.l.wrap.b32 %r8316, %r8315, %r8314, 1; + shf.l.wrap.b32 %r8317, %r8314, %r8315, 1; + mov.b64 %rd11726, {%r8317, %r8316}; + xor.b64 %rd11727, %rd11725, %rd11726; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8318}, %rd11717; + } + { + .reg .b32 %dummy; + mov.b64 {%r8319,%dummy}, %rd11717; + } + shf.l.wrap.b32 %r8320, %r8319, %r8318, 1; + shf.l.wrap.b32 %r8321, %r8318, %r8319, 1; + mov.b64 %rd11728, {%r8321, %r8320}; + xor.b64 %rd11729, %rd11728, %rd11709; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8322}, %rd11721; + } + { + .reg .b32 %dummy; + mov.b64 {%r8323,%dummy}, %rd11721; + } + shf.l.wrap.b32 %r8324, %r8323, %r8322, 1; + shf.l.wrap.b32 %r8325, %r8322, %r8323, 1; + mov.b64 %rd11730, {%r8325, %r8324}; + xor.b64 %rd11731, %rd11730, %rd11713; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8326}, %rd11725; + } + { + .reg .b32 %dummy; + mov.b64 {%r8327,%dummy}, %rd11725; + } + shf.l.wrap.b32 %r8328, %r8327, %r8326, 1; + shf.l.wrap.b32 %r8329, %r8326, %r8327, 1; + mov.b64 %rd11732, {%r8329, %r8328}; + xor.b64 %rd11733, %rd11732, %rd11717; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8330}, %rd11709; + } + { + .reg .b32 %dummy; + mov.b64 {%r8331,%dummy}, %rd11709; + } + shf.l.wrap.b32 %r8332, %r8331, %r8330, 1; + shf.l.wrap.b32 %r8333, %r8330, %r8331, 1; + mov.b64 %rd11734, {%r8333, %r8332}; + xor.b64 %rd11735, %rd11721, %rd11734; + xor.b64 %rd11736, %rd11727, %rd11705; + xor.b64 %rd11737, %rd11729, %rd11634; + xor.b64 %rd11738, %rd11731, %rd11637; + xor.b64 %rd11739, %rd11733, %rd11640; + xor.b64 %rd11740, %rd11735, %rd11643; + xor.b64 %rd11741, %rd11727, %rd11646; + xor.b64 %rd11742, %rd11729, %rd11649; + xor.b64 %rd11743, %rd11731, %rd11652; + xor.b64 %rd11744, %rd11733, %rd11655; + xor.b64 %rd11745, %rd11735, %rd11658; + xor.b64 %rd11746, %rd11727, %rd11661; + xor.b64 %rd11747, %rd11729, %rd11664; + xor.b64 %rd11748, %rd11731, %rd11667; + xor.b64 %rd11749, %rd11733, %rd11670; + xor.b64 %rd11750, %rd11735, %rd11673; + xor.b64 %rd11751, %rd11727, %rd11676; + xor.b64 %rd11752, %rd11729, %rd11679; + xor.b64 %rd11753, %rd11731, %rd11682; + xor.b64 %rd11754, %rd11733, %rd11685; + xor.b64 %rd11755, %rd11735, %rd11688; + xor.b64 %rd11756, %rd11727, %rd11691; + xor.b64 %rd11757, %rd11729, %rd11694; + xor.b64 %rd11758, %rd11731, %rd11697; + xor.b64 %rd11759, %rd11733, %rd11700; + xor.b64 %rd11760, %rd11735, %rd11703; + { + .reg .b32 %dummy; + mov.b64 {%r8334,%dummy}, %rd11742; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8335}, %rd11742; + } + shf.r.wrap.b32 %r8336, %r8335, %r8334, 20; + shf.r.wrap.b32 %r8337, %r8334, %r8335, 20; + mov.b64 %rd11761, {%r8337, %r8336}; + { + .reg .b32 %dummy; + mov.b64 {%r8338,%dummy}, %rd11748; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8339}, %rd11748; + } + shf.r.wrap.b32 %r8340, %r8339, %r8338, 21; + shf.r.wrap.b32 %r8341, %r8338, %r8339, 21; + mov.b64 %rd11762, {%r8341, %r8340}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8342}, %rd11754; + } + { + .reg .b32 %dummy; + mov.b64 {%r8343,%dummy}, %rd11754; + } + shf.l.wrap.b32 %r8344, %r8343, %r8342, 21; + shf.l.wrap.b32 %r8345, %r8342, %r8343, 21; + mov.b64 %rd11763, {%r8345, %r8344}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8346}, %rd11760; + } + { + .reg .b32 %dummy; + mov.b64 {%r8347,%dummy}, %rd11760; + } + shf.l.wrap.b32 %r8348, %r8347, %r8346, 14; + shf.l.wrap.b32 %r8349, %r8346, %r8347, 14; + mov.b64 %rd11764, {%r8349, %r8348}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8350}, %rd11739; + } + { + .reg .b32 %dummy; + mov.b64 {%r8351,%dummy}, %rd11739; + } + shf.l.wrap.b32 %r8352, %r8351, %r8350, 28; + shf.l.wrap.b32 %r8353, %r8350, %r8351, 28; + mov.b64 %rd11765, {%r8353, %r8352}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8354}, %rd11745; + } + { + .reg .b32 %dummy; + mov.b64 {%r8355,%dummy}, %rd11745; + } + shf.l.wrap.b32 %r8356, %r8355, %r8354, 20; + shf.l.wrap.b32 %r8357, %r8354, %r8355, 20; + mov.b64 %rd11766, {%r8357, %r8356}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8358}, %rd11746; + } + { + .reg .b32 %dummy; + mov.b64 {%r8359,%dummy}, %rd11746; + } + shf.l.wrap.b32 %r8360, %r8359, %r8358, 3; + shf.l.wrap.b32 %r8361, %r8358, %r8359, 3; + mov.b64 %rd11767, {%r8361, %r8360}; + { + .reg .b32 %dummy; + mov.b64 {%r8362,%dummy}, %rd11752; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8363}, %rd11752; + } + shf.r.wrap.b32 %r8364, %r8363, %r8362, 19; + shf.r.wrap.b32 %r8365, %r8362, %r8363, 19; + mov.b64 %rd11768, {%r8365, %r8364}; + { + .reg .b32 %dummy; + mov.b64 {%r8366,%dummy}, %rd11758; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8367}, %rd11758; + } + shf.r.wrap.b32 %r8368, %r8367, %r8366, 3; + shf.r.wrap.b32 %r8369, %r8366, %r8367, 3; + mov.b64 %rd11769, {%r8369, %r8368}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8370}, %rd11737; + } + { + .reg .b32 %dummy; + mov.b64 {%r8371,%dummy}, %rd11737; + } + shf.l.wrap.b32 %r8372, %r8371, %r8370, 1; + shf.l.wrap.b32 %r8373, %r8370, %r8371, 1; + mov.b64 %rd11770, {%r8373, %r8372}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8374}, %rd11743; + } + { + .reg .b32 %dummy; + mov.b64 {%r8375,%dummy}, %rd11743; + } + shf.l.wrap.b32 %r8376, %r8375, %r8374, 6; + shf.l.wrap.b32 %r8377, %r8374, %r8375, 6; + mov.b64 %rd11771, {%r8377, %r8376}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8378}, %rd11749; + } + { + .reg .b32 %dummy; + mov.b64 {%r8379,%dummy}, %rd11749; + } + shf.l.wrap.b32 %r8380, %r8379, %r8378, 25; + shf.l.wrap.b32 %r8381, %r8378, %r8379, 25; + mov.b64 %rd11772, {%r8381, %r8380}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8382}, %rd11755; + } + { + .reg .b32 %dummy; + mov.b64 {%r8383,%dummy}, %rd11755; + } + shf.l.wrap.b32 %r8384, %r8383, %r8382, 8; + shf.l.wrap.b32 %r8385, %r8382, %r8383, 8; + mov.b64 %rd11773, {%r8385, %r8384}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8386}, %rd11756; + } + { + .reg .b32 %dummy; + mov.b64 {%r8387,%dummy}, %rd11756; + } + shf.l.wrap.b32 %r8388, %r8387, %r8386, 18; + shf.l.wrap.b32 %r8389, %r8386, %r8387, 18; + mov.b64 %rd11774, {%r8389, %r8388}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8390}, %rd11740; + } + { + .reg .b32 %dummy; + mov.b64 {%r8391,%dummy}, %rd11740; + } + shf.l.wrap.b32 %r8392, %r8391, %r8390, 27; + shf.l.wrap.b32 %r8393, %r8390, %r8391, 27; + mov.b64 %rd11775, {%r8393, %r8392}; + { + .reg .b32 %dummy; + mov.b64 {%r8394,%dummy}, %rd11741; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8395}, %rd11741; + } + shf.r.wrap.b32 %r8396, %r8395, %r8394, 28; + shf.r.wrap.b32 %r8397, %r8394, %r8395, 28; + mov.b64 %rd11776, {%r8397, %r8396}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8398}, %rd11747; + } + { + .reg .b32 %dummy; + mov.b64 {%r8399,%dummy}, %rd11747; + } + shf.l.wrap.b32 %r8400, %r8399, %r8398, 10; + shf.l.wrap.b32 %r8401, %r8398, %r8399, 10; + mov.b64 %rd11777, {%r8401, %r8400}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8402}, %rd11753; + } + { + .reg .b32 %dummy; + mov.b64 {%r8403,%dummy}, %rd11753; + } + shf.l.wrap.b32 %r8404, %r8403, %r8402, 15; + shf.l.wrap.b32 %r8405, %r8402, %r8403, 15; + mov.b64 %rd11778, {%r8405, %r8404}; + { + .reg .b32 %dummy; + mov.b64 {%r8406,%dummy}, %rd11759; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8407}, %rd11759; + } + shf.r.wrap.b32 %r8408, %r8407, %r8406, 8; + shf.r.wrap.b32 %r8409, %r8406, %r8407, 8; + mov.b64 %rd11779, {%r8409, %r8408}; + { + .reg .b32 %dummy; + mov.b64 {%r8410,%dummy}, %rd11738; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8411}, %rd11738; + } + shf.r.wrap.b32 %r8412, %r8411, %r8410, 2; + shf.r.wrap.b32 %r8413, %r8410, %r8411, 2; + mov.b64 %rd11780, {%r8413, %r8412}; + { + .reg .b32 %dummy; + mov.b64 {%r8414,%dummy}, %rd11744; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8415}, %rd11744; + } + shf.r.wrap.b32 %r8416, %r8415, %r8414, 9; + shf.r.wrap.b32 %r8417, %r8414, %r8415, 9; + mov.b64 %rd11781, {%r8417, %r8416}; + { + .reg .b32 %dummy; + mov.b64 {%r8418,%dummy}, %rd11750; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8419}, %rd11750; + } + shf.r.wrap.b32 %r8420, %r8419, %r8418, 25; + shf.r.wrap.b32 %r8421, %r8418, %r8419, 25; + mov.b64 %rd11782, {%r8421, %r8420}; + { + .reg .b32 %dummy; + mov.b64 {%r8422,%dummy}, %rd11751; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8423}, %rd11751; + } + shf.r.wrap.b32 %r8424, %r8423, %r8422, 23; + shf.r.wrap.b32 %r8425, %r8422, %r8423, 23; + mov.b64 %rd11783, {%r8425, %r8424}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8426}, %rd11757; + } + { + .reg .b32 %dummy; + mov.b64 {%r8427,%dummy}, %rd11757; + } + shf.l.wrap.b32 %r8428, %r8427, %r8426, 2; + shf.l.wrap.b32 %r8429, %r8426, %r8427, 2; + mov.b64 %rd11784, {%r8429, %r8428}; + not.b64 %rd11785, %rd11761; + and.b64 %rd11786, %rd11762, %rd11785; + xor.b64 %rd11787, %rd11786, %rd11736; + not.b64 %rd11788, %rd11762; + and.b64 %rd11789, %rd11763, %rd11788; + xor.b64 %rd11790, %rd11789, %rd11761; + st.local.u64 [%rd2+8], %rd11790; + not.b64 %rd11791, %rd11763; + and.b64 %rd11792, %rd11764, %rd11791; + xor.b64 %rd11793, %rd11792, %rd11762; + st.local.u64 [%rd2+16], %rd11793; + not.b64 %rd11794, %rd11764; + and.b64 %rd11795, %rd11736, %rd11794; + xor.b64 %rd11796, %rd11795, %rd11763; + st.local.u64 [%rd2+24], %rd11796; + not.b64 %rd11797, %rd11736; + and.b64 %rd11798, %rd11761, %rd11797; + xor.b64 %rd11799, %rd11798, %rd11764; + st.local.u64 [%rd2+32], %rd11799; + not.b64 %rd11800, %rd11766; + and.b64 %rd11801, %rd11767, %rd11800; + xor.b64 %rd11802, %rd11801, %rd11765; + st.local.u64 [%rd2+40], %rd11802; + not.b64 %rd11803, %rd11767; + and.b64 %rd11804, %rd11768, %rd11803; + xor.b64 %rd11805, %rd11804, %rd11766; + st.local.u64 [%rd2+48], %rd11805; + not.b64 %rd11806, %rd11768; + and.b64 %rd11807, %rd11769, %rd11806; + xor.b64 %rd11808, %rd11807, %rd11767; + st.local.u64 [%rd2+56], %rd11808; + not.b64 %rd11809, %rd11769; + and.b64 %rd11810, %rd11765, %rd11809; + xor.b64 %rd11811, %rd11810, %rd11768; + st.local.u64 [%rd2+64], %rd11811; + not.b64 %rd11812, %rd11765; + and.b64 %rd11813, %rd11766, %rd11812; + xor.b64 %rd11814, %rd11813, %rd11769; + st.local.u64 [%rd2+72], %rd11814; + not.b64 %rd11815, %rd11771; + and.b64 %rd11816, %rd11772, %rd11815; + xor.b64 %rd11817, %rd11816, %rd11770; + st.local.u64 [%rd2+80], %rd11817; + not.b64 %rd11818, %rd11772; + and.b64 %rd11819, %rd11773, %rd11818; + xor.b64 %rd11820, %rd11819, %rd11771; + st.local.u64 [%rd2+88], %rd11820; + not.b64 %rd11821, %rd11773; + and.b64 %rd11822, %rd11774, %rd11821; + xor.b64 %rd11823, %rd11822, %rd11772; + st.local.u64 [%rd2+96], %rd11823; + not.b64 %rd11824, %rd11774; + and.b64 %rd11825, %rd11770, %rd11824; + xor.b64 %rd11826, %rd11825, %rd11773; + st.local.u64 [%rd2+104], %rd11826; + not.b64 %rd11827, %rd11770; + and.b64 %rd11828, %rd11771, %rd11827; + xor.b64 %rd11829, %rd11828, %rd11774; + st.local.u64 [%rd2+112], %rd11829; + not.b64 %rd11830, %rd11776; + and.b64 %rd11831, %rd11777, %rd11830; + xor.b64 %rd11832, %rd11831, %rd11775; + st.local.u64 [%rd2+120], %rd11832; + not.b64 %rd11833, %rd11777; + and.b64 %rd11834, %rd11778, %rd11833; + xor.b64 %rd11835, %rd11834, %rd11776; + st.local.u64 [%rd2+128], %rd11835; + not.b64 %rd11836, %rd11778; + and.b64 %rd11837, %rd11779, %rd11836; + xor.b64 %rd11838, %rd11837, %rd11777; + st.local.u64 [%rd2+136], %rd11838; + not.b64 %rd11839, %rd11779; + and.b64 %rd11840, %rd11775, %rd11839; + xor.b64 %rd11841, %rd11840, %rd11778; + st.local.u64 [%rd2+144], %rd11841; + not.b64 %rd11842, %rd11775; + and.b64 %rd11843, %rd11776, %rd11842; + xor.b64 %rd11844, %rd11843, %rd11779; + st.local.u64 [%rd2+152], %rd11844; + not.b64 %rd11845, %rd11781; + and.b64 %rd11846, %rd11782, %rd11845; + xor.b64 %rd11847, %rd11846, %rd11780; + st.local.u64 [%rd2+160], %rd11847; + not.b64 %rd11848, %rd11782; + and.b64 %rd11849, %rd11783, %rd11848; + xor.b64 %rd11850, %rd11849, %rd11781; + st.local.u64 [%rd2+168], %rd11850; + not.b64 %rd11851, %rd11783; + and.b64 %rd11852, %rd11784, %rd11851; + xor.b64 %rd11853, %rd11852, %rd11782; + st.local.u64 [%rd2+176], %rd11853; + not.b64 %rd11854, %rd11784; + and.b64 %rd11855, %rd11780, %rd11854; + xor.b64 %rd11856, %rd11855, %rd11783; + st.local.u64 [%rd2+184], %rd11856; + not.b64 %rd11857, %rd11780; + and.b64 %rd11858, %rd11781, %rd11857; + xor.b64 %rd11859, %rd11858, %rd11784; + st.local.u64 [%rd2+192], %rd11859; + xor.b64 %rd11861, %rd11787, %rd39; + st.local.u64 [%rd2], %rd11861; + st.local.u64 [%rd2+392], %rd19853; + +$L__BB1_38: + and.b64 %rd93, %rd19853, 63; + shr.u64 %rd94, %rd19853, 6; + setp.eq.s64 %p33, %rd94, 0; + mov.u64 %rd19860, 0; + @%p33 bra $L__BB1_44; + + add.s64 %rd11865, %rd94, -1; + and.b64 %rd19859, %rd94, 3; + setp.lt.u64 %p34, %rd11865, 3; + mov.u64 %rd19860, 0; + mov.u32 %r14034, 0; + @%p34 bra $L__BB1_42; + + sub.s64 %rd19855, %rd94, %rd19859; + +$L__BB1_41: + add.s64 %rd11867, %rd2, %rd19860; + ld.local.u64 %rd11868, [%rd11867]; + ld.local.u8 %rd11869, [%rd11867+200]; + ld.local.u8 %rd11870, [%rd11867+201]; + bfi.b64 %rd11871, %rd11870, %rd11869, 8, 8; + ld.local.u8 %rd11872, [%rd11867+202]; + ld.local.u8 %rd11873, [%rd11867+203]; + bfi.b64 %rd11874, %rd11873, %rd11872, 8, 8; + bfi.b64 %rd11875, %rd11874, %rd11871, 16, 16; + ld.local.u8 %rd11876, [%rd11867+204]; + ld.local.u8 %rd11877, [%rd11867+205]; + bfi.b64 %rd11878, %rd11877, %rd11876, 8, 8; + ld.local.u8 %rd11879, [%rd11867+206]; + ld.local.u8 %rd11880, [%rd11867+207]; + bfi.b64 %rd11881, %rd11880, %rd11879, 8, 8; + bfi.b64 %rd11882, %rd11881, %rd11878, 16, 16; + bfi.b64 %rd11883, %rd11882, %rd11875, 32, 32; + xor.b64 %rd11884, %rd11868, %rd11883; + st.local.u64 [%rd11867], %rd11884; + ld.local.u64 %rd11885, [%rd11867+8]; + ld.local.u8 %rd11886, [%rd11867+208]; + ld.local.u8 %rd11887, [%rd11867+209]; + bfi.b64 %rd11888, %rd11887, %rd11886, 8, 8; + ld.local.u8 %rd11889, [%rd11867+210]; + ld.local.u8 %rd11890, [%rd11867+211]; + bfi.b64 %rd11891, %rd11890, %rd11889, 8, 8; + bfi.b64 %rd11892, %rd11891, %rd11888, 16, 16; + ld.local.u8 %rd11893, [%rd11867+212]; + ld.local.u8 %rd11894, [%rd11867+213]; + bfi.b64 %rd11895, %rd11894, %rd11893, 8, 8; + ld.local.u8 %rd11896, [%rd11867+214]; + ld.local.u8 %rd11897, [%rd11867+215]; + bfi.b64 %rd11898, %rd11897, %rd11896, 8, 8; + bfi.b64 %rd11899, %rd11898, %rd11895, 16, 16; + bfi.b64 %rd11900, %rd11899, %rd11892, 32, 32; + xor.b64 %rd11901, %rd11885, %rd11900; + st.local.u64 [%rd11867+8], %rd11901; + ld.local.u64 %rd11902, [%rd11867+16]; + ld.local.u8 %rd11903, [%rd11867+216]; + ld.local.u8 %rd11904, [%rd11867+217]; + bfi.b64 %rd11905, %rd11904, %rd11903, 8, 8; + ld.local.u8 %rd11906, [%rd11867+218]; + ld.local.u8 %rd11907, [%rd11867+219]; + bfi.b64 %rd11908, %rd11907, %rd11906, 8, 8; + bfi.b64 %rd11909, %rd11908, %rd11905, 16, 16; + ld.local.u8 %rd11910, [%rd11867+220]; + ld.local.u8 %rd11911, [%rd11867+221]; + bfi.b64 %rd11912, %rd11911, %rd11910, 8, 8; + ld.local.u8 %rd11913, [%rd11867+222]; + ld.local.u8 %rd11914, [%rd11867+223]; + bfi.b64 %rd11915, %rd11914, %rd11913, 8, 8; + bfi.b64 %rd11916, %rd11915, %rd11912, 16, 16; + bfi.b64 %rd11917, %rd11916, %rd11909, 32, 32; + xor.b64 %rd11918, %rd11902, %rd11917; + st.local.u64 [%rd11867+16], %rd11918; + ld.local.u64 %rd11919, [%rd11867+24]; + ld.local.u8 %rd11920, [%rd11867+224]; + ld.local.u8 %rd11921, [%rd11867+225]; + bfi.b64 %rd11922, %rd11921, %rd11920, 8, 8; + ld.local.u8 %rd11923, [%rd11867+226]; + ld.local.u8 %rd11924, [%rd11867+227]; + bfi.b64 %rd11925, %rd11924, %rd11923, 8, 8; + bfi.b64 %rd11926, %rd11925, %rd11922, 16, 16; + ld.local.u8 %rd11927, [%rd11867+228]; + ld.local.u8 %rd11928, [%rd11867+229]; + bfi.b64 %rd11929, %rd11928, %rd11927, 8, 8; + ld.local.u8 %rd11930, [%rd11867+230]; + ld.local.u8 %rd11931, [%rd11867+231]; + bfi.b64 %rd11932, %rd11931, %rd11930, 8, 8; + bfi.b64 %rd11933, %rd11932, %rd11929, 16, 16; + bfi.b64 %rd11934, %rd11933, %rd11926, 32, 32; + xor.b64 %rd11935, %rd11919, %rd11934; + st.local.u64 [%rd11867+24], %rd11935; + add.s64 %rd19860, %rd19860, 32; + add.s32 %r14034, %r14034, 4; + add.s64 %rd19855, %rd19855, -4; + setp.ne.s64 %p35, %rd19855, 0; + @%p35 bra $L__BB1_41; + +$L__BB1_42: + setp.eq.s64 %p36, %rd19859, 0; + @%p36 bra $L__BB1_44; + +$L__BB1_43: + .pragma "nounroll"; + add.s64 %rd11936, %rd2, %rd19860; + mul.wide.s32 %rd11937, %r14034, 8; + add.s64 %rd11938, %rd2, %rd11937; + ld.local.u64 %rd11939, [%rd11938]; + ld.local.u8 %rd11940, [%rd11936+200]; + ld.local.u8 %rd11941, [%rd11936+201]; + bfi.b64 %rd11942, %rd11941, %rd11940, 8, 8; + ld.local.u8 %rd11943, [%rd11936+202]; + ld.local.u8 %rd11944, [%rd11936+203]; + bfi.b64 %rd11945, %rd11944, %rd11943, 8, 8; + bfi.b64 %rd11946, %rd11945, %rd11942, 16, 16; + ld.local.u8 %rd11947, [%rd11936+204]; + ld.local.u8 %rd11948, [%rd11936+205]; + bfi.b64 %rd11949, %rd11948, %rd11947, 8, 8; + ld.local.u8 %rd11950, [%rd11936+206]; + ld.local.u8 %rd11951, [%rd11936+207]; + bfi.b64 %rd11952, %rd11951, %rd11950, 8, 8; + bfi.b64 %rd11953, %rd11952, %rd11949, 16, 16; + bfi.b64 %rd11954, %rd11953, %rd11946, 32, 32; + xor.b64 %rd11955, %rd11939, %rd11954; + st.local.u64 [%rd11938], %rd11955; + add.s64 %rd19860, %rd19860, 8; + add.s32 %r14034, %r14034, 1; + add.s64 %rd19859, %rd19859, -1; + setp.ne.s64 %p37, %rd19859, 0; + @%p37 bra $L__BB1_43; + +$L__BB1_44: + setp.eq.s64 %p38, %rd93, 0; + @%p38 bra $L__BB1_46; + + cvt.u32.u64 %r8432, %rd93; + mov.u32 %r8433, -1; + shl.b32 %r8434, %r8433, %r8432; + not.b32 %r8435, %r8434; + cvt.s64.s32 %rd11956, %r8435; + add.s64 %rd11957, %rd2, %rd19860; + ld.local.u8 %rd11958, [%rd11957+200]; + ld.local.u8 %rd11959, [%rd11957+201]; + bfi.b64 %rd11960, %rd11959, %rd11958, 8, 8; + ld.local.u8 %rd11961, [%rd11957+202]; + ld.local.u8 %rd11962, [%rd11957+203]; + bfi.b64 %rd11963, %rd11962, %rd11961, 8, 8; + bfi.b64 %rd11964, %rd11963, %rd11960, 16, 16; + ld.local.u8 %rd11965, [%rd11957+204]; + ld.local.u8 %rd11966, [%rd11957+205]; + bfi.b64 %rd11967, %rd11966, %rd11965, 8, 8; + ld.local.u8 %rd11968, [%rd11957+206]; + ld.local.u8 %rd11969, [%rd11957+207]; + bfi.b64 %rd11970, %rd11969, %rd11968, 8, 8; + bfi.b64 %rd11971, %rd11970, %rd11967, 16, 16; + bfi.b64 %rd11972, %rd11971, %rd11964, 32, 32; + and.b64 %rd11973, %rd11972, %rd11956; + shl.b64 %rd11974, %rd94, 3; + add.s64 %rd11975, %rd2, %rd11974; + ld.local.u64 %rd11976, [%rd11975]; + xor.b64 %rd11977, %rd11976, %rd11973; + st.local.u64 [%rd11975], %rd11977; + +$L__BB1_46: + ld.const.u64 %rd19825, [CUDA_KECCAK_CONSTS+40]; + ld.const.u64 %rd19824, [CUDA_KECCAK_CONSTS+32]; + ld.const.u64 %rd19823, [CUDA_KECCAK_CONSTS+24]; + ld.const.u64 %rd19822, [CUDA_KECCAK_CONSTS+16]; + ld.const.u64 %rd19821, [CUDA_KECCAK_CONSTS+8]; + ld.const.u64 %rd19820, [CUDA_KECCAK_CONSTS]; + add.s64 %rd11978, %rd7997, -1; + shr.u64 %rd11979, %rd11978, 3; + and.b64 %rd11980, %rd11979, 2305843009213693944; + add.s64 %rd11981, %rd2, %rd11980; + ld.local.u64 %rd11982, [%rd11981]; + xor.b64 %rd11983, %rd11982, -9223372036854775808; + st.local.u64 [%rd11981], %rd11983; + ld.local.u64 %rd11984, [%rd2+40]; + ld.local.u64 %rd11985, [%rd2]; + xor.b64 %rd11986, %rd11984, %rd11985; + ld.local.u64 %rd11987, [%rd2+80]; + xor.b64 %rd11988, %rd11986, %rd11987; + ld.local.u64 %rd11989, [%rd2+120]; + xor.b64 %rd11990, %rd11988, %rd11989; + ld.local.u64 %rd11991, [%rd2+160]; + xor.b64 %rd11992, %rd11990, %rd11991; + ld.local.u64 %rd11993, [%rd2+48]; + ld.local.u64 %rd11994, [%rd2+8]; + xor.b64 %rd11995, %rd11993, %rd11994; + ld.local.u64 %rd11996, [%rd2+88]; + xor.b64 %rd11997, %rd11995, %rd11996; + ld.local.u64 %rd11998, [%rd2+128]; + xor.b64 %rd11999, %rd11997, %rd11998; + ld.local.u64 %rd12000, [%rd2+168]; + xor.b64 %rd12001, %rd11999, %rd12000; + ld.local.u64 %rd12002, [%rd2+56]; + ld.local.u64 %rd12003, [%rd2+16]; + xor.b64 %rd12004, %rd12002, %rd12003; + ld.local.u64 %rd12005, [%rd2+96]; + xor.b64 %rd12006, %rd12004, %rd12005; + ld.local.u64 %rd12007, [%rd2+136]; + xor.b64 %rd12008, %rd12006, %rd12007; + ld.local.u64 %rd12009, [%rd2+176]; + xor.b64 %rd12010, %rd12008, %rd12009; + ld.local.u64 %rd12011, [%rd2+64]; + ld.local.u64 %rd12012, [%rd2+24]; + xor.b64 %rd12013, %rd12011, %rd12012; + ld.local.u64 %rd12014, [%rd2+104]; + xor.b64 %rd12015, %rd12013, %rd12014; + ld.local.u64 %rd12016, [%rd2+144]; + xor.b64 %rd12017, %rd12015, %rd12016; + ld.local.u64 %rd12018, [%rd2+184]; + xor.b64 %rd12019, %rd12017, %rd12018; + ld.local.u64 %rd12020, [%rd2+72]; + ld.local.u64 %rd12021, [%rd2+32]; + xor.b64 %rd12022, %rd12020, %rd12021; + ld.local.u64 %rd12023, [%rd2+112]; + xor.b64 %rd12024, %rd12022, %rd12023; + ld.local.u64 %rd12025, [%rd2+152]; + xor.b64 %rd12026, %rd12024, %rd12025; + ld.local.u64 %rd12027, [%rd2+192]; + xor.b64 %rd12028, %rd12026, %rd12027; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8436}, %rd12001; + } + { + .reg .b32 %dummy; + mov.b64 {%r8437,%dummy}, %rd12001; + } + shf.l.wrap.b32 %r8438, %r8437, %r8436, 1; + shf.l.wrap.b32 %r8439, %r8436, %r8437, 1; + mov.b64 %rd12029, {%r8439, %r8438}; + xor.b64 %rd12030, %rd12028, %rd12029; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8440}, %rd12010; + } + { + .reg .b32 %dummy; + mov.b64 {%r8441,%dummy}, %rd12010; + } + shf.l.wrap.b32 %r8442, %r8441, %r8440, 1; + shf.l.wrap.b32 %r8443, %r8440, %r8441, 1; + mov.b64 %rd12031, {%r8443, %r8442}; + xor.b64 %rd12032, %rd12031, %rd11992; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8444}, %rd12019; + } + { + .reg .b32 %dummy; + mov.b64 {%r8445,%dummy}, %rd12019; + } + shf.l.wrap.b32 %r8446, %r8445, %r8444, 1; + shf.l.wrap.b32 %r8447, %r8444, %r8445, 1; + mov.b64 %rd12033, {%r8447, %r8446}; + xor.b64 %rd12034, %rd12033, %rd12001; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8448}, %rd12028; + } + { + .reg .b32 %dummy; + mov.b64 {%r8449,%dummy}, %rd12028; + } + shf.l.wrap.b32 %r8450, %r8449, %r8448, 1; + shf.l.wrap.b32 %r8451, %r8448, %r8449, 1; + mov.b64 %rd12035, {%r8451, %r8450}; + xor.b64 %rd12036, %rd12035, %rd12010; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8452}, %rd11992; + } + { + .reg .b32 %dummy; + mov.b64 {%r8453,%dummy}, %rd11992; + } + shf.l.wrap.b32 %r8454, %r8453, %r8452, 1; + shf.l.wrap.b32 %r8455, %r8452, %r8453, 1; + mov.b64 %rd12037, {%r8455, %r8454}; + xor.b64 %rd12038, %rd12019, %rd12037; + xor.b64 %rd12039, %rd12030, %rd11985; + xor.b64 %rd12040, %rd12032, %rd11994; + xor.b64 %rd12041, %rd12034, %rd12003; + xor.b64 %rd12042, %rd12036, %rd12012; + xor.b64 %rd12043, %rd12038, %rd12021; + xor.b64 %rd12044, %rd12030, %rd11984; + xor.b64 %rd12045, %rd12032, %rd11993; + xor.b64 %rd12046, %rd12034, %rd12002; + xor.b64 %rd12047, %rd12036, %rd12011; + xor.b64 %rd12048, %rd12038, %rd12020; + xor.b64 %rd12049, %rd12030, %rd11987; + xor.b64 %rd12050, %rd12032, %rd11996; + xor.b64 %rd12051, %rd12034, %rd12005; + xor.b64 %rd12052, %rd12036, %rd12014; + xor.b64 %rd12053, %rd12038, %rd12023; + xor.b64 %rd12054, %rd12030, %rd11989; + xor.b64 %rd12055, %rd12032, %rd11998; + xor.b64 %rd12056, %rd12034, %rd12007; + xor.b64 %rd12057, %rd12036, %rd12016; + xor.b64 %rd12058, %rd12038, %rd12025; + xor.b64 %rd12059, %rd12030, %rd11991; + xor.b64 %rd12060, %rd12032, %rd12000; + xor.b64 %rd12061, %rd12034, %rd12009; + xor.b64 %rd12062, %rd12036, %rd12018; + xor.b64 %rd12063, %rd12038, %rd12027; + { + .reg .b32 %dummy; + mov.b64 {%r8456,%dummy}, %rd12045; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8457}, %rd12045; + } + shf.r.wrap.b32 %r8458, %r8457, %r8456, 20; + shf.r.wrap.b32 %r8459, %r8456, %r8457, 20; + mov.b64 %rd12064, {%r8459, %r8458}; + { + .reg .b32 %dummy; + mov.b64 {%r8460,%dummy}, %rd12051; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8461}, %rd12051; + } + shf.r.wrap.b32 %r8462, %r8461, %r8460, 21; + shf.r.wrap.b32 %r8463, %r8460, %r8461, 21; + mov.b64 %rd12065, {%r8463, %r8462}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8464}, %rd12057; + } + { + .reg .b32 %dummy; + mov.b64 {%r8465,%dummy}, %rd12057; + } + shf.l.wrap.b32 %r8466, %r8465, %r8464, 21; + shf.l.wrap.b32 %r8467, %r8464, %r8465, 21; + mov.b64 %rd12066, {%r8467, %r8466}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8468}, %rd12063; + } + { + .reg .b32 %dummy; + mov.b64 {%r8469,%dummy}, %rd12063; + } + shf.l.wrap.b32 %r8470, %r8469, %r8468, 14; + shf.l.wrap.b32 %r8471, %r8468, %r8469, 14; + mov.b64 %rd12067, {%r8471, %r8470}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8472}, %rd12042; + } + { + .reg .b32 %dummy; + mov.b64 {%r8473,%dummy}, %rd12042; + } + shf.l.wrap.b32 %r8474, %r8473, %r8472, 28; + shf.l.wrap.b32 %r8475, %r8472, %r8473, 28; + mov.b64 %rd12068, {%r8475, %r8474}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8476}, %rd12048; + } + { + .reg .b32 %dummy; + mov.b64 {%r8477,%dummy}, %rd12048; + } + shf.l.wrap.b32 %r8478, %r8477, %r8476, 20; + shf.l.wrap.b32 %r8479, %r8476, %r8477, 20; + mov.b64 %rd12069, {%r8479, %r8478}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8480}, %rd12049; + } + { + .reg .b32 %dummy; + mov.b64 {%r8481,%dummy}, %rd12049; + } + shf.l.wrap.b32 %r8482, %r8481, %r8480, 3; + shf.l.wrap.b32 %r8483, %r8480, %r8481, 3; + mov.b64 %rd12070, {%r8483, %r8482}; + { + .reg .b32 %dummy; + mov.b64 {%r8484,%dummy}, %rd12055; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8485}, %rd12055; + } + shf.r.wrap.b32 %r8486, %r8485, %r8484, 19; + shf.r.wrap.b32 %r8487, %r8484, %r8485, 19; + mov.b64 %rd12071, {%r8487, %r8486}; + { + .reg .b32 %dummy; + mov.b64 {%r8488,%dummy}, %rd12061; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8489}, %rd12061; + } + shf.r.wrap.b32 %r8490, %r8489, %r8488, 3; + shf.r.wrap.b32 %r8491, %r8488, %r8489, 3; + mov.b64 %rd12072, {%r8491, %r8490}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8492}, %rd12040; + } + { + .reg .b32 %dummy; + mov.b64 {%r8493,%dummy}, %rd12040; + } + shf.l.wrap.b32 %r8494, %r8493, %r8492, 1; + shf.l.wrap.b32 %r8495, %r8492, %r8493, 1; + mov.b64 %rd12073, {%r8495, %r8494}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8496}, %rd12046; + } + { + .reg .b32 %dummy; + mov.b64 {%r8497,%dummy}, %rd12046; + } + shf.l.wrap.b32 %r8498, %r8497, %r8496, 6; + shf.l.wrap.b32 %r8499, %r8496, %r8497, 6; + mov.b64 %rd12074, {%r8499, %r8498}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8500}, %rd12052; + } + { + .reg .b32 %dummy; + mov.b64 {%r8501,%dummy}, %rd12052; + } + shf.l.wrap.b32 %r8502, %r8501, %r8500, 25; + shf.l.wrap.b32 %r8503, %r8500, %r8501, 25; + mov.b64 %rd12075, {%r8503, %r8502}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8504}, %rd12058; + } + { + .reg .b32 %dummy; + mov.b64 {%r8505,%dummy}, %rd12058; + } + shf.l.wrap.b32 %r8506, %r8505, %r8504, 8; + shf.l.wrap.b32 %r8507, %r8504, %r8505, 8; + mov.b64 %rd12076, {%r8507, %r8506}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8508}, %rd12059; + } + { + .reg .b32 %dummy; + mov.b64 {%r8509,%dummy}, %rd12059; + } + shf.l.wrap.b32 %r8510, %r8509, %r8508, 18; + shf.l.wrap.b32 %r8511, %r8508, %r8509, 18; + mov.b64 %rd12077, {%r8511, %r8510}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8512}, %rd12043; + } + { + .reg .b32 %dummy; + mov.b64 {%r8513,%dummy}, %rd12043; + } + shf.l.wrap.b32 %r8514, %r8513, %r8512, 27; + shf.l.wrap.b32 %r8515, %r8512, %r8513, 27; + mov.b64 %rd12078, {%r8515, %r8514}; + { + .reg .b32 %dummy; + mov.b64 {%r8516,%dummy}, %rd12044; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8517}, %rd12044; + } + shf.r.wrap.b32 %r8518, %r8517, %r8516, 28; + shf.r.wrap.b32 %r8519, %r8516, %r8517, 28; + mov.b64 %rd12079, {%r8519, %r8518}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8520}, %rd12050; + } + { + .reg .b32 %dummy; + mov.b64 {%r8521,%dummy}, %rd12050; + } + shf.l.wrap.b32 %r8522, %r8521, %r8520, 10; + shf.l.wrap.b32 %r8523, %r8520, %r8521, 10; + mov.b64 %rd12080, {%r8523, %r8522}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8524}, %rd12056; + } + { + .reg .b32 %dummy; + mov.b64 {%r8525,%dummy}, %rd12056; + } + shf.l.wrap.b32 %r8526, %r8525, %r8524, 15; + shf.l.wrap.b32 %r8527, %r8524, %r8525, 15; + mov.b64 %rd12081, {%r8527, %r8526}; + { + .reg .b32 %dummy; + mov.b64 {%r8528,%dummy}, %rd12062; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8529}, %rd12062; + } + shf.r.wrap.b32 %r8530, %r8529, %r8528, 8; + shf.r.wrap.b32 %r8531, %r8528, %r8529, 8; + mov.b64 %rd12082, {%r8531, %r8530}; + { + .reg .b32 %dummy; + mov.b64 {%r8532,%dummy}, %rd12041; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8533}, %rd12041; + } + shf.r.wrap.b32 %r8534, %r8533, %r8532, 2; + shf.r.wrap.b32 %r8535, %r8532, %r8533, 2; + mov.b64 %rd12083, {%r8535, %r8534}; + { + .reg .b32 %dummy; + mov.b64 {%r8536,%dummy}, %rd12047; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8537}, %rd12047; + } + shf.r.wrap.b32 %r8538, %r8537, %r8536, 9; + shf.r.wrap.b32 %r8539, %r8536, %r8537, 9; + mov.b64 %rd12084, {%r8539, %r8538}; + { + .reg .b32 %dummy; + mov.b64 {%r8540,%dummy}, %rd12053; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8541}, %rd12053; + } + shf.r.wrap.b32 %r8542, %r8541, %r8540, 25; + shf.r.wrap.b32 %r8543, %r8540, %r8541, 25; + mov.b64 %rd12085, {%r8543, %r8542}; + { + .reg .b32 %dummy; + mov.b64 {%r8544,%dummy}, %rd12054; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8545}, %rd12054; + } + shf.r.wrap.b32 %r8546, %r8545, %r8544, 23; + shf.r.wrap.b32 %r8547, %r8544, %r8545, 23; + mov.b64 %rd12086, {%r8547, %r8546}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8548}, %rd12060; + } + { + .reg .b32 %dummy; + mov.b64 {%r8549,%dummy}, %rd12060; + } + shf.l.wrap.b32 %r8550, %r8549, %r8548, 2; + shf.l.wrap.b32 %r8551, %r8548, %r8549, 2; + mov.b64 %rd12087, {%r8551, %r8550}; + not.b64 %rd12088, %rd12064; + and.b64 %rd12089, %rd12065, %rd12088; + xor.b64 %rd12090, %rd12089, %rd12039; + not.b64 %rd12091, %rd12065; + and.b64 %rd12092, %rd12066, %rd12091; + xor.b64 %rd12093, %rd12092, %rd12064; + not.b64 %rd12094, %rd12066; + and.b64 %rd12095, %rd12067, %rd12094; + xor.b64 %rd12096, %rd12095, %rd12065; + not.b64 %rd12097, %rd12067; + and.b64 %rd12098, %rd12039, %rd12097; + xor.b64 %rd12099, %rd12098, %rd12066; + not.b64 %rd12100, %rd12039; + and.b64 %rd12101, %rd12064, %rd12100; + xor.b64 %rd12102, %rd12101, %rd12067; + not.b64 %rd12103, %rd12069; + and.b64 %rd12104, %rd12070, %rd12103; + xor.b64 %rd12105, %rd12104, %rd12068; + not.b64 %rd12106, %rd12070; + and.b64 %rd12107, %rd12071, %rd12106; + xor.b64 %rd12108, %rd12107, %rd12069; + not.b64 %rd12109, %rd12071; + and.b64 %rd12110, %rd12072, %rd12109; + xor.b64 %rd12111, %rd12110, %rd12070; + not.b64 %rd12112, %rd12072; + and.b64 %rd12113, %rd12068, %rd12112; + xor.b64 %rd12114, %rd12113, %rd12071; + not.b64 %rd12115, %rd12068; + and.b64 %rd12116, %rd12069, %rd12115; + xor.b64 %rd12117, %rd12116, %rd12072; + not.b64 %rd12118, %rd12074; + and.b64 %rd12119, %rd12075, %rd12118; + xor.b64 %rd12120, %rd12119, %rd12073; + not.b64 %rd12121, %rd12075; + and.b64 %rd12122, %rd12076, %rd12121; + xor.b64 %rd12123, %rd12122, %rd12074; + not.b64 %rd12124, %rd12076; + and.b64 %rd12125, %rd12077, %rd12124; + xor.b64 %rd12126, %rd12125, %rd12075; + not.b64 %rd12127, %rd12077; + and.b64 %rd12128, %rd12073, %rd12127; + xor.b64 %rd12129, %rd12128, %rd12076; + not.b64 %rd12130, %rd12073; + and.b64 %rd12131, %rd12074, %rd12130; + xor.b64 %rd12132, %rd12131, %rd12077; + not.b64 %rd12133, %rd12079; + and.b64 %rd12134, %rd12080, %rd12133; + xor.b64 %rd12135, %rd12134, %rd12078; + not.b64 %rd12136, %rd12080; + and.b64 %rd12137, %rd12081, %rd12136; + xor.b64 %rd12138, %rd12137, %rd12079; + not.b64 %rd12139, %rd12081; + and.b64 %rd12140, %rd12082, %rd12139; + xor.b64 %rd12141, %rd12140, %rd12080; + not.b64 %rd12142, %rd12082; + and.b64 %rd12143, %rd12078, %rd12142; + xor.b64 %rd12144, %rd12143, %rd12081; + not.b64 %rd12145, %rd12078; + and.b64 %rd12146, %rd12079, %rd12145; + xor.b64 %rd12147, %rd12146, %rd12082; + not.b64 %rd12148, %rd12084; + and.b64 %rd12149, %rd12085, %rd12148; + xor.b64 %rd12150, %rd12149, %rd12083; + not.b64 %rd12151, %rd12085; + and.b64 %rd12152, %rd12086, %rd12151; + xor.b64 %rd12153, %rd12152, %rd12084; + not.b64 %rd12154, %rd12086; + and.b64 %rd12155, %rd12087, %rd12154; + xor.b64 %rd12156, %rd12155, %rd12085; + not.b64 %rd12157, %rd12087; + and.b64 %rd12158, %rd12083, %rd12157; + xor.b64 %rd12159, %rd12158, %rd12086; + not.b64 %rd12160, %rd12083; + and.b64 %rd12161, %rd12084, %rd12160; + xor.b64 %rd12162, %rd12161, %rd12087; + xor.b64 %rd12164, %rd12090, %rd19820; + xor.b64 %rd12165, %rd12105, %rd12164; + xor.b64 %rd12166, %rd12165, %rd12120; + xor.b64 %rd12167, %rd12166, %rd12135; + xor.b64 %rd12168, %rd12167, %rd12150; + xor.b64 %rd12169, %rd12108, %rd12093; + xor.b64 %rd12170, %rd12169, %rd12123; + xor.b64 %rd12171, %rd12170, %rd12138; + xor.b64 %rd12172, %rd12171, %rd12153; + xor.b64 %rd12173, %rd12111, %rd12096; + xor.b64 %rd12174, %rd12173, %rd12126; + xor.b64 %rd12175, %rd12174, %rd12141; + xor.b64 %rd12176, %rd12175, %rd12156; + xor.b64 %rd12177, %rd12114, %rd12099; + xor.b64 %rd12178, %rd12177, %rd12129; + xor.b64 %rd12179, %rd12178, %rd12144; + xor.b64 %rd12180, %rd12179, %rd12159; + xor.b64 %rd12181, %rd12117, %rd12102; + xor.b64 %rd12182, %rd12181, %rd12132; + xor.b64 %rd12183, %rd12182, %rd12147; + xor.b64 %rd12184, %rd12183, %rd12162; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8552}, %rd12172; + } + { + .reg .b32 %dummy; + mov.b64 {%r8553,%dummy}, %rd12172; + } + shf.l.wrap.b32 %r8554, %r8553, %r8552, 1; + shf.l.wrap.b32 %r8555, %r8552, %r8553, 1; + mov.b64 %rd12185, {%r8555, %r8554}; + xor.b64 %rd12186, %rd12184, %rd12185; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8556}, %rd12176; + } + { + .reg .b32 %dummy; + mov.b64 {%r8557,%dummy}, %rd12176; + } + shf.l.wrap.b32 %r8558, %r8557, %r8556, 1; + shf.l.wrap.b32 %r8559, %r8556, %r8557, 1; + mov.b64 %rd12187, {%r8559, %r8558}; + xor.b64 %rd12188, %rd12187, %rd12168; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8560}, %rd12180; + } + { + .reg .b32 %dummy; + mov.b64 {%r8561,%dummy}, %rd12180; + } + shf.l.wrap.b32 %r8562, %r8561, %r8560, 1; + shf.l.wrap.b32 %r8563, %r8560, %r8561, 1; + mov.b64 %rd12189, {%r8563, %r8562}; + xor.b64 %rd12190, %rd12189, %rd12172; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8564}, %rd12184; + } + { + .reg .b32 %dummy; + mov.b64 {%r8565,%dummy}, %rd12184; + } + shf.l.wrap.b32 %r8566, %r8565, %r8564, 1; + shf.l.wrap.b32 %r8567, %r8564, %r8565, 1; + mov.b64 %rd12191, {%r8567, %r8566}; + xor.b64 %rd12192, %rd12191, %rd12176; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8568}, %rd12168; + } + { + .reg .b32 %dummy; + mov.b64 {%r8569,%dummy}, %rd12168; + } + shf.l.wrap.b32 %r8570, %r8569, %r8568, 1; + shf.l.wrap.b32 %r8571, %r8568, %r8569, 1; + mov.b64 %rd12193, {%r8571, %r8570}; + xor.b64 %rd12194, %rd12180, %rd12193; + xor.b64 %rd12195, %rd12186, %rd12164; + xor.b64 %rd12196, %rd12188, %rd12093; + xor.b64 %rd12197, %rd12190, %rd12096; + xor.b64 %rd12198, %rd12192, %rd12099; + xor.b64 %rd12199, %rd12194, %rd12102; + xor.b64 %rd12200, %rd12186, %rd12105; + xor.b64 %rd12201, %rd12188, %rd12108; + xor.b64 %rd12202, %rd12190, %rd12111; + xor.b64 %rd12203, %rd12192, %rd12114; + xor.b64 %rd12204, %rd12194, %rd12117; + xor.b64 %rd12205, %rd12186, %rd12120; + xor.b64 %rd12206, %rd12188, %rd12123; + xor.b64 %rd12207, %rd12190, %rd12126; + xor.b64 %rd12208, %rd12192, %rd12129; + xor.b64 %rd12209, %rd12194, %rd12132; + xor.b64 %rd12210, %rd12186, %rd12135; + xor.b64 %rd12211, %rd12188, %rd12138; + xor.b64 %rd12212, %rd12190, %rd12141; + xor.b64 %rd12213, %rd12192, %rd12144; + xor.b64 %rd12214, %rd12194, %rd12147; + xor.b64 %rd12215, %rd12186, %rd12150; + xor.b64 %rd12216, %rd12188, %rd12153; + xor.b64 %rd12217, %rd12190, %rd12156; + xor.b64 %rd12218, %rd12192, %rd12159; + xor.b64 %rd12219, %rd12194, %rd12162; + { + .reg .b32 %dummy; + mov.b64 {%r8572,%dummy}, %rd12201; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8573}, %rd12201; + } + shf.r.wrap.b32 %r8574, %r8573, %r8572, 20; + shf.r.wrap.b32 %r8575, %r8572, %r8573, 20; + mov.b64 %rd12220, {%r8575, %r8574}; + { + .reg .b32 %dummy; + mov.b64 {%r8576,%dummy}, %rd12207; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8577}, %rd12207; + } + shf.r.wrap.b32 %r8578, %r8577, %r8576, 21; + shf.r.wrap.b32 %r8579, %r8576, %r8577, 21; + mov.b64 %rd12221, {%r8579, %r8578}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8580}, %rd12213; + } + { + .reg .b32 %dummy; + mov.b64 {%r8581,%dummy}, %rd12213; + } + shf.l.wrap.b32 %r8582, %r8581, %r8580, 21; + shf.l.wrap.b32 %r8583, %r8580, %r8581, 21; + mov.b64 %rd12222, {%r8583, %r8582}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8584}, %rd12219; + } + { + .reg .b32 %dummy; + mov.b64 {%r8585,%dummy}, %rd12219; + } + shf.l.wrap.b32 %r8586, %r8585, %r8584, 14; + shf.l.wrap.b32 %r8587, %r8584, %r8585, 14; + mov.b64 %rd12223, {%r8587, %r8586}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8588}, %rd12198; + } + { + .reg .b32 %dummy; + mov.b64 {%r8589,%dummy}, %rd12198; + } + shf.l.wrap.b32 %r8590, %r8589, %r8588, 28; + shf.l.wrap.b32 %r8591, %r8588, %r8589, 28; + mov.b64 %rd12224, {%r8591, %r8590}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8592}, %rd12204; + } + { + .reg .b32 %dummy; + mov.b64 {%r8593,%dummy}, %rd12204; + } + shf.l.wrap.b32 %r8594, %r8593, %r8592, 20; + shf.l.wrap.b32 %r8595, %r8592, %r8593, 20; + mov.b64 %rd12225, {%r8595, %r8594}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8596}, %rd12205; + } + { + .reg .b32 %dummy; + mov.b64 {%r8597,%dummy}, %rd12205; + } + shf.l.wrap.b32 %r8598, %r8597, %r8596, 3; + shf.l.wrap.b32 %r8599, %r8596, %r8597, 3; + mov.b64 %rd12226, {%r8599, %r8598}; + { + .reg .b32 %dummy; + mov.b64 {%r8600,%dummy}, %rd12211; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8601}, %rd12211; + } + shf.r.wrap.b32 %r8602, %r8601, %r8600, 19; + shf.r.wrap.b32 %r8603, %r8600, %r8601, 19; + mov.b64 %rd12227, {%r8603, %r8602}; + { + .reg .b32 %dummy; + mov.b64 {%r8604,%dummy}, %rd12217; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8605}, %rd12217; + } + shf.r.wrap.b32 %r8606, %r8605, %r8604, 3; + shf.r.wrap.b32 %r8607, %r8604, %r8605, 3; + mov.b64 %rd12228, {%r8607, %r8606}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8608}, %rd12196; + } + { + .reg .b32 %dummy; + mov.b64 {%r8609,%dummy}, %rd12196; + } + shf.l.wrap.b32 %r8610, %r8609, %r8608, 1; + shf.l.wrap.b32 %r8611, %r8608, %r8609, 1; + mov.b64 %rd12229, {%r8611, %r8610}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8612}, %rd12202; + } + { + .reg .b32 %dummy; + mov.b64 {%r8613,%dummy}, %rd12202; + } + shf.l.wrap.b32 %r8614, %r8613, %r8612, 6; + shf.l.wrap.b32 %r8615, %r8612, %r8613, 6; + mov.b64 %rd12230, {%r8615, %r8614}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8616}, %rd12208; + } + { + .reg .b32 %dummy; + mov.b64 {%r8617,%dummy}, %rd12208; + } + shf.l.wrap.b32 %r8618, %r8617, %r8616, 25; + shf.l.wrap.b32 %r8619, %r8616, %r8617, 25; + mov.b64 %rd12231, {%r8619, %r8618}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8620}, %rd12214; + } + { + .reg .b32 %dummy; + mov.b64 {%r8621,%dummy}, %rd12214; + } + shf.l.wrap.b32 %r8622, %r8621, %r8620, 8; + shf.l.wrap.b32 %r8623, %r8620, %r8621, 8; + mov.b64 %rd12232, {%r8623, %r8622}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8624}, %rd12215; + } + { + .reg .b32 %dummy; + mov.b64 {%r8625,%dummy}, %rd12215; + } + shf.l.wrap.b32 %r8626, %r8625, %r8624, 18; + shf.l.wrap.b32 %r8627, %r8624, %r8625, 18; + mov.b64 %rd12233, {%r8627, %r8626}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8628}, %rd12199; + } + { + .reg .b32 %dummy; + mov.b64 {%r8629,%dummy}, %rd12199; + } + shf.l.wrap.b32 %r8630, %r8629, %r8628, 27; + shf.l.wrap.b32 %r8631, %r8628, %r8629, 27; + mov.b64 %rd12234, {%r8631, %r8630}; + { + .reg .b32 %dummy; + mov.b64 {%r8632,%dummy}, %rd12200; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8633}, %rd12200; + } + shf.r.wrap.b32 %r8634, %r8633, %r8632, 28; + shf.r.wrap.b32 %r8635, %r8632, %r8633, 28; + mov.b64 %rd12235, {%r8635, %r8634}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8636}, %rd12206; + } + { + .reg .b32 %dummy; + mov.b64 {%r8637,%dummy}, %rd12206; + } + shf.l.wrap.b32 %r8638, %r8637, %r8636, 10; + shf.l.wrap.b32 %r8639, %r8636, %r8637, 10; + mov.b64 %rd12236, {%r8639, %r8638}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8640}, %rd12212; + } + { + .reg .b32 %dummy; + mov.b64 {%r8641,%dummy}, %rd12212; + } + shf.l.wrap.b32 %r8642, %r8641, %r8640, 15; + shf.l.wrap.b32 %r8643, %r8640, %r8641, 15; + mov.b64 %rd12237, {%r8643, %r8642}; + { + .reg .b32 %dummy; + mov.b64 {%r8644,%dummy}, %rd12218; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8645}, %rd12218; + } + shf.r.wrap.b32 %r8646, %r8645, %r8644, 8; + shf.r.wrap.b32 %r8647, %r8644, %r8645, 8; + mov.b64 %rd12238, {%r8647, %r8646}; + { + .reg .b32 %dummy; + mov.b64 {%r8648,%dummy}, %rd12197; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8649}, %rd12197; + } + shf.r.wrap.b32 %r8650, %r8649, %r8648, 2; + shf.r.wrap.b32 %r8651, %r8648, %r8649, 2; + mov.b64 %rd12239, {%r8651, %r8650}; + { + .reg .b32 %dummy; + mov.b64 {%r8652,%dummy}, %rd12203; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8653}, %rd12203; + } + shf.r.wrap.b32 %r8654, %r8653, %r8652, 9; + shf.r.wrap.b32 %r8655, %r8652, %r8653, 9; + mov.b64 %rd12240, {%r8655, %r8654}; + { + .reg .b32 %dummy; + mov.b64 {%r8656,%dummy}, %rd12209; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8657}, %rd12209; + } + shf.r.wrap.b32 %r8658, %r8657, %r8656, 25; + shf.r.wrap.b32 %r8659, %r8656, %r8657, 25; + mov.b64 %rd12241, {%r8659, %r8658}; + { + .reg .b32 %dummy; + mov.b64 {%r8660,%dummy}, %rd12210; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8661}, %rd12210; + } + shf.r.wrap.b32 %r8662, %r8661, %r8660, 23; + shf.r.wrap.b32 %r8663, %r8660, %r8661, 23; + mov.b64 %rd12242, {%r8663, %r8662}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8664}, %rd12216; + } + { + .reg .b32 %dummy; + mov.b64 {%r8665,%dummy}, %rd12216; + } + shf.l.wrap.b32 %r8666, %r8665, %r8664, 2; + shf.l.wrap.b32 %r8667, %r8664, %r8665, 2; + mov.b64 %rd12243, {%r8667, %r8666}; + not.b64 %rd12244, %rd12220; + and.b64 %rd12245, %rd12221, %rd12244; + xor.b64 %rd12246, %rd12245, %rd12195; + not.b64 %rd12247, %rd12221; + and.b64 %rd12248, %rd12222, %rd12247; + xor.b64 %rd12249, %rd12248, %rd12220; + not.b64 %rd12250, %rd12222; + and.b64 %rd12251, %rd12223, %rd12250; + xor.b64 %rd12252, %rd12251, %rd12221; + not.b64 %rd12253, %rd12223; + and.b64 %rd12254, %rd12195, %rd12253; + xor.b64 %rd12255, %rd12254, %rd12222; + not.b64 %rd12256, %rd12195; + and.b64 %rd12257, %rd12220, %rd12256; + xor.b64 %rd12258, %rd12257, %rd12223; + not.b64 %rd12259, %rd12225; + and.b64 %rd12260, %rd12226, %rd12259; + xor.b64 %rd12261, %rd12260, %rd12224; + not.b64 %rd12262, %rd12226; + and.b64 %rd12263, %rd12227, %rd12262; + xor.b64 %rd12264, %rd12263, %rd12225; + not.b64 %rd12265, %rd12227; + and.b64 %rd12266, %rd12228, %rd12265; + xor.b64 %rd12267, %rd12266, %rd12226; + not.b64 %rd12268, %rd12228; + and.b64 %rd12269, %rd12224, %rd12268; + xor.b64 %rd12270, %rd12269, %rd12227; + not.b64 %rd12271, %rd12224; + and.b64 %rd12272, %rd12225, %rd12271; + xor.b64 %rd12273, %rd12272, %rd12228; + not.b64 %rd12274, %rd12230; + and.b64 %rd12275, %rd12231, %rd12274; + xor.b64 %rd12276, %rd12275, %rd12229; + not.b64 %rd12277, %rd12231; + and.b64 %rd12278, %rd12232, %rd12277; + xor.b64 %rd12279, %rd12278, %rd12230; + not.b64 %rd12280, %rd12232; + and.b64 %rd12281, %rd12233, %rd12280; + xor.b64 %rd12282, %rd12281, %rd12231; + not.b64 %rd12283, %rd12233; + and.b64 %rd12284, %rd12229, %rd12283; + xor.b64 %rd12285, %rd12284, %rd12232; + not.b64 %rd12286, %rd12229; + and.b64 %rd12287, %rd12230, %rd12286; + xor.b64 %rd12288, %rd12287, %rd12233; + not.b64 %rd12289, %rd12235; + and.b64 %rd12290, %rd12236, %rd12289; + xor.b64 %rd12291, %rd12290, %rd12234; + not.b64 %rd12292, %rd12236; + and.b64 %rd12293, %rd12237, %rd12292; + xor.b64 %rd12294, %rd12293, %rd12235; + not.b64 %rd12295, %rd12237; + and.b64 %rd12296, %rd12238, %rd12295; + xor.b64 %rd12297, %rd12296, %rd12236; + not.b64 %rd12298, %rd12238; + and.b64 %rd12299, %rd12234, %rd12298; + xor.b64 %rd12300, %rd12299, %rd12237; + not.b64 %rd12301, %rd12234; + and.b64 %rd12302, %rd12235, %rd12301; + xor.b64 %rd12303, %rd12302, %rd12238; + not.b64 %rd12304, %rd12240; + and.b64 %rd12305, %rd12241, %rd12304; + xor.b64 %rd12306, %rd12305, %rd12239; + not.b64 %rd12307, %rd12241; + and.b64 %rd12308, %rd12242, %rd12307; + xor.b64 %rd12309, %rd12308, %rd12240; + not.b64 %rd12310, %rd12242; + and.b64 %rd12311, %rd12243, %rd12310; + xor.b64 %rd12312, %rd12311, %rd12241; + not.b64 %rd12313, %rd12243; + and.b64 %rd12314, %rd12239, %rd12313; + xor.b64 %rd12315, %rd12314, %rd12242; + not.b64 %rd12316, %rd12239; + and.b64 %rd12317, %rd12240, %rd12316; + xor.b64 %rd12318, %rd12317, %rd12243; + xor.b64 %rd12320, %rd12246, %rd19821; + xor.b64 %rd12321, %rd12261, %rd12320; + xor.b64 %rd12322, %rd12321, %rd12276; + xor.b64 %rd12323, %rd12322, %rd12291; + xor.b64 %rd12324, %rd12323, %rd12306; + xor.b64 %rd12325, %rd12264, %rd12249; + xor.b64 %rd12326, %rd12325, %rd12279; + xor.b64 %rd12327, %rd12326, %rd12294; + xor.b64 %rd12328, %rd12327, %rd12309; + xor.b64 %rd12329, %rd12267, %rd12252; + xor.b64 %rd12330, %rd12329, %rd12282; + xor.b64 %rd12331, %rd12330, %rd12297; + xor.b64 %rd12332, %rd12331, %rd12312; + xor.b64 %rd12333, %rd12270, %rd12255; + xor.b64 %rd12334, %rd12333, %rd12285; + xor.b64 %rd12335, %rd12334, %rd12300; + xor.b64 %rd12336, %rd12335, %rd12315; + xor.b64 %rd12337, %rd12273, %rd12258; + xor.b64 %rd12338, %rd12337, %rd12288; + xor.b64 %rd12339, %rd12338, %rd12303; + xor.b64 %rd12340, %rd12339, %rd12318; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8668}, %rd12328; + } + { + .reg .b32 %dummy; + mov.b64 {%r8669,%dummy}, %rd12328; + } + shf.l.wrap.b32 %r8670, %r8669, %r8668, 1; + shf.l.wrap.b32 %r8671, %r8668, %r8669, 1; + mov.b64 %rd12341, {%r8671, %r8670}; + xor.b64 %rd12342, %rd12340, %rd12341; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8672}, %rd12332; + } + { + .reg .b32 %dummy; + mov.b64 {%r8673,%dummy}, %rd12332; + } + shf.l.wrap.b32 %r8674, %r8673, %r8672, 1; + shf.l.wrap.b32 %r8675, %r8672, %r8673, 1; + mov.b64 %rd12343, {%r8675, %r8674}; + xor.b64 %rd12344, %rd12343, %rd12324; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8676}, %rd12336; + } + { + .reg .b32 %dummy; + mov.b64 {%r8677,%dummy}, %rd12336; + } + shf.l.wrap.b32 %r8678, %r8677, %r8676, 1; + shf.l.wrap.b32 %r8679, %r8676, %r8677, 1; + mov.b64 %rd12345, {%r8679, %r8678}; + xor.b64 %rd12346, %rd12345, %rd12328; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8680}, %rd12340; + } + { + .reg .b32 %dummy; + mov.b64 {%r8681,%dummy}, %rd12340; + } + shf.l.wrap.b32 %r8682, %r8681, %r8680, 1; + shf.l.wrap.b32 %r8683, %r8680, %r8681, 1; + mov.b64 %rd12347, {%r8683, %r8682}; + xor.b64 %rd12348, %rd12347, %rd12332; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8684}, %rd12324; + } + { + .reg .b32 %dummy; + mov.b64 {%r8685,%dummy}, %rd12324; + } + shf.l.wrap.b32 %r8686, %r8685, %r8684, 1; + shf.l.wrap.b32 %r8687, %r8684, %r8685, 1; + mov.b64 %rd12349, {%r8687, %r8686}; + xor.b64 %rd12350, %rd12336, %rd12349; + xor.b64 %rd12351, %rd12342, %rd12320; + xor.b64 %rd12352, %rd12344, %rd12249; + xor.b64 %rd12353, %rd12346, %rd12252; + xor.b64 %rd12354, %rd12348, %rd12255; + xor.b64 %rd12355, %rd12350, %rd12258; + xor.b64 %rd12356, %rd12342, %rd12261; + xor.b64 %rd12357, %rd12344, %rd12264; + xor.b64 %rd12358, %rd12346, %rd12267; + xor.b64 %rd12359, %rd12348, %rd12270; + xor.b64 %rd12360, %rd12350, %rd12273; + xor.b64 %rd12361, %rd12342, %rd12276; + xor.b64 %rd12362, %rd12344, %rd12279; + xor.b64 %rd12363, %rd12346, %rd12282; + xor.b64 %rd12364, %rd12348, %rd12285; + xor.b64 %rd12365, %rd12350, %rd12288; + xor.b64 %rd12366, %rd12342, %rd12291; + xor.b64 %rd12367, %rd12344, %rd12294; + xor.b64 %rd12368, %rd12346, %rd12297; + xor.b64 %rd12369, %rd12348, %rd12300; + xor.b64 %rd12370, %rd12350, %rd12303; + xor.b64 %rd12371, %rd12342, %rd12306; + xor.b64 %rd12372, %rd12344, %rd12309; + xor.b64 %rd12373, %rd12346, %rd12312; + xor.b64 %rd12374, %rd12348, %rd12315; + xor.b64 %rd12375, %rd12350, %rd12318; + { + .reg .b32 %dummy; + mov.b64 {%r8688,%dummy}, %rd12357; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8689}, %rd12357; + } + shf.r.wrap.b32 %r8690, %r8689, %r8688, 20; + shf.r.wrap.b32 %r8691, %r8688, %r8689, 20; + mov.b64 %rd12376, {%r8691, %r8690}; + { + .reg .b32 %dummy; + mov.b64 {%r8692,%dummy}, %rd12363; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8693}, %rd12363; + } + shf.r.wrap.b32 %r8694, %r8693, %r8692, 21; + shf.r.wrap.b32 %r8695, %r8692, %r8693, 21; + mov.b64 %rd12377, {%r8695, %r8694}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8696}, %rd12369; + } + { + .reg .b32 %dummy; + mov.b64 {%r8697,%dummy}, %rd12369; + } + shf.l.wrap.b32 %r8698, %r8697, %r8696, 21; + shf.l.wrap.b32 %r8699, %r8696, %r8697, 21; + mov.b64 %rd12378, {%r8699, %r8698}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8700}, %rd12375; + } + { + .reg .b32 %dummy; + mov.b64 {%r8701,%dummy}, %rd12375; + } + shf.l.wrap.b32 %r8702, %r8701, %r8700, 14; + shf.l.wrap.b32 %r8703, %r8700, %r8701, 14; + mov.b64 %rd12379, {%r8703, %r8702}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8704}, %rd12354; + } + { + .reg .b32 %dummy; + mov.b64 {%r8705,%dummy}, %rd12354; + } + shf.l.wrap.b32 %r8706, %r8705, %r8704, 28; + shf.l.wrap.b32 %r8707, %r8704, %r8705, 28; + mov.b64 %rd12380, {%r8707, %r8706}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8708}, %rd12360; + } + { + .reg .b32 %dummy; + mov.b64 {%r8709,%dummy}, %rd12360; + } + shf.l.wrap.b32 %r8710, %r8709, %r8708, 20; + shf.l.wrap.b32 %r8711, %r8708, %r8709, 20; + mov.b64 %rd12381, {%r8711, %r8710}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8712}, %rd12361; + } + { + .reg .b32 %dummy; + mov.b64 {%r8713,%dummy}, %rd12361; + } + shf.l.wrap.b32 %r8714, %r8713, %r8712, 3; + shf.l.wrap.b32 %r8715, %r8712, %r8713, 3; + mov.b64 %rd12382, {%r8715, %r8714}; + { + .reg .b32 %dummy; + mov.b64 {%r8716,%dummy}, %rd12367; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8717}, %rd12367; + } + shf.r.wrap.b32 %r8718, %r8717, %r8716, 19; + shf.r.wrap.b32 %r8719, %r8716, %r8717, 19; + mov.b64 %rd12383, {%r8719, %r8718}; + { + .reg .b32 %dummy; + mov.b64 {%r8720,%dummy}, %rd12373; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8721}, %rd12373; + } + shf.r.wrap.b32 %r8722, %r8721, %r8720, 3; + shf.r.wrap.b32 %r8723, %r8720, %r8721, 3; + mov.b64 %rd12384, {%r8723, %r8722}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8724}, %rd12352; + } + { + .reg .b32 %dummy; + mov.b64 {%r8725,%dummy}, %rd12352; + } + shf.l.wrap.b32 %r8726, %r8725, %r8724, 1; + shf.l.wrap.b32 %r8727, %r8724, %r8725, 1; + mov.b64 %rd12385, {%r8727, %r8726}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8728}, %rd12358; + } + { + .reg .b32 %dummy; + mov.b64 {%r8729,%dummy}, %rd12358; + } + shf.l.wrap.b32 %r8730, %r8729, %r8728, 6; + shf.l.wrap.b32 %r8731, %r8728, %r8729, 6; + mov.b64 %rd12386, {%r8731, %r8730}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8732}, %rd12364; + } + { + .reg .b32 %dummy; + mov.b64 {%r8733,%dummy}, %rd12364; + } + shf.l.wrap.b32 %r8734, %r8733, %r8732, 25; + shf.l.wrap.b32 %r8735, %r8732, %r8733, 25; + mov.b64 %rd12387, {%r8735, %r8734}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8736}, %rd12370; + } + { + .reg .b32 %dummy; + mov.b64 {%r8737,%dummy}, %rd12370; + } + shf.l.wrap.b32 %r8738, %r8737, %r8736, 8; + shf.l.wrap.b32 %r8739, %r8736, %r8737, 8; + mov.b64 %rd12388, {%r8739, %r8738}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8740}, %rd12371; + } + { + .reg .b32 %dummy; + mov.b64 {%r8741,%dummy}, %rd12371; + } + shf.l.wrap.b32 %r8742, %r8741, %r8740, 18; + shf.l.wrap.b32 %r8743, %r8740, %r8741, 18; + mov.b64 %rd12389, {%r8743, %r8742}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8744}, %rd12355; + } + { + .reg .b32 %dummy; + mov.b64 {%r8745,%dummy}, %rd12355; + } + shf.l.wrap.b32 %r8746, %r8745, %r8744, 27; + shf.l.wrap.b32 %r8747, %r8744, %r8745, 27; + mov.b64 %rd12390, {%r8747, %r8746}; + { + .reg .b32 %dummy; + mov.b64 {%r8748,%dummy}, %rd12356; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8749}, %rd12356; + } + shf.r.wrap.b32 %r8750, %r8749, %r8748, 28; + shf.r.wrap.b32 %r8751, %r8748, %r8749, 28; + mov.b64 %rd12391, {%r8751, %r8750}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8752}, %rd12362; + } + { + .reg .b32 %dummy; + mov.b64 {%r8753,%dummy}, %rd12362; + } + shf.l.wrap.b32 %r8754, %r8753, %r8752, 10; + shf.l.wrap.b32 %r8755, %r8752, %r8753, 10; + mov.b64 %rd12392, {%r8755, %r8754}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8756}, %rd12368; + } + { + .reg .b32 %dummy; + mov.b64 {%r8757,%dummy}, %rd12368; + } + shf.l.wrap.b32 %r8758, %r8757, %r8756, 15; + shf.l.wrap.b32 %r8759, %r8756, %r8757, 15; + mov.b64 %rd12393, {%r8759, %r8758}; + { + .reg .b32 %dummy; + mov.b64 {%r8760,%dummy}, %rd12374; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8761}, %rd12374; + } + shf.r.wrap.b32 %r8762, %r8761, %r8760, 8; + shf.r.wrap.b32 %r8763, %r8760, %r8761, 8; + mov.b64 %rd12394, {%r8763, %r8762}; + { + .reg .b32 %dummy; + mov.b64 {%r8764,%dummy}, %rd12353; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8765}, %rd12353; + } + shf.r.wrap.b32 %r8766, %r8765, %r8764, 2; + shf.r.wrap.b32 %r8767, %r8764, %r8765, 2; + mov.b64 %rd12395, {%r8767, %r8766}; + { + .reg .b32 %dummy; + mov.b64 {%r8768,%dummy}, %rd12359; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8769}, %rd12359; + } + shf.r.wrap.b32 %r8770, %r8769, %r8768, 9; + shf.r.wrap.b32 %r8771, %r8768, %r8769, 9; + mov.b64 %rd12396, {%r8771, %r8770}; + { + .reg .b32 %dummy; + mov.b64 {%r8772,%dummy}, %rd12365; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8773}, %rd12365; + } + shf.r.wrap.b32 %r8774, %r8773, %r8772, 25; + shf.r.wrap.b32 %r8775, %r8772, %r8773, 25; + mov.b64 %rd12397, {%r8775, %r8774}; + { + .reg .b32 %dummy; + mov.b64 {%r8776,%dummy}, %rd12366; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8777}, %rd12366; + } + shf.r.wrap.b32 %r8778, %r8777, %r8776, 23; + shf.r.wrap.b32 %r8779, %r8776, %r8777, 23; + mov.b64 %rd12398, {%r8779, %r8778}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8780}, %rd12372; + } + { + .reg .b32 %dummy; + mov.b64 {%r8781,%dummy}, %rd12372; + } + shf.l.wrap.b32 %r8782, %r8781, %r8780, 2; + shf.l.wrap.b32 %r8783, %r8780, %r8781, 2; + mov.b64 %rd12399, {%r8783, %r8782}; + not.b64 %rd12400, %rd12376; + and.b64 %rd12401, %rd12377, %rd12400; + xor.b64 %rd12402, %rd12401, %rd12351; + not.b64 %rd12403, %rd12377; + and.b64 %rd12404, %rd12378, %rd12403; + xor.b64 %rd12405, %rd12404, %rd12376; + not.b64 %rd12406, %rd12378; + and.b64 %rd12407, %rd12379, %rd12406; + xor.b64 %rd12408, %rd12407, %rd12377; + not.b64 %rd12409, %rd12379; + and.b64 %rd12410, %rd12351, %rd12409; + xor.b64 %rd12411, %rd12410, %rd12378; + not.b64 %rd12412, %rd12351; + and.b64 %rd12413, %rd12376, %rd12412; + xor.b64 %rd12414, %rd12413, %rd12379; + not.b64 %rd12415, %rd12381; + and.b64 %rd12416, %rd12382, %rd12415; + xor.b64 %rd12417, %rd12416, %rd12380; + not.b64 %rd12418, %rd12382; + and.b64 %rd12419, %rd12383, %rd12418; + xor.b64 %rd12420, %rd12419, %rd12381; + not.b64 %rd12421, %rd12383; + and.b64 %rd12422, %rd12384, %rd12421; + xor.b64 %rd12423, %rd12422, %rd12382; + not.b64 %rd12424, %rd12384; + and.b64 %rd12425, %rd12380, %rd12424; + xor.b64 %rd12426, %rd12425, %rd12383; + not.b64 %rd12427, %rd12380; + and.b64 %rd12428, %rd12381, %rd12427; + xor.b64 %rd12429, %rd12428, %rd12384; + not.b64 %rd12430, %rd12386; + and.b64 %rd12431, %rd12387, %rd12430; + xor.b64 %rd12432, %rd12431, %rd12385; + not.b64 %rd12433, %rd12387; + and.b64 %rd12434, %rd12388, %rd12433; + xor.b64 %rd12435, %rd12434, %rd12386; + not.b64 %rd12436, %rd12388; + and.b64 %rd12437, %rd12389, %rd12436; + xor.b64 %rd12438, %rd12437, %rd12387; + not.b64 %rd12439, %rd12389; + and.b64 %rd12440, %rd12385, %rd12439; + xor.b64 %rd12441, %rd12440, %rd12388; + not.b64 %rd12442, %rd12385; + and.b64 %rd12443, %rd12386, %rd12442; + xor.b64 %rd12444, %rd12443, %rd12389; + not.b64 %rd12445, %rd12391; + and.b64 %rd12446, %rd12392, %rd12445; + xor.b64 %rd12447, %rd12446, %rd12390; + not.b64 %rd12448, %rd12392; + and.b64 %rd12449, %rd12393, %rd12448; + xor.b64 %rd12450, %rd12449, %rd12391; + not.b64 %rd12451, %rd12393; + and.b64 %rd12452, %rd12394, %rd12451; + xor.b64 %rd12453, %rd12452, %rd12392; + not.b64 %rd12454, %rd12394; + and.b64 %rd12455, %rd12390, %rd12454; + xor.b64 %rd12456, %rd12455, %rd12393; + not.b64 %rd12457, %rd12390; + and.b64 %rd12458, %rd12391, %rd12457; + xor.b64 %rd12459, %rd12458, %rd12394; + not.b64 %rd12460, %rd12396; + and.b64 %rd12461, %rd12397, %rd12460; + xor.b64 %rd12462, %rd12461, %rd12395; + not.b64 %rd12463, %rd12397; + and.b64 %rd12464, %rd12398, %rd12463; + xor.b64 %rd12465, %rd12464, %rd12396; + not.b64 %rd12466, %rd12398; + and.b64 %rd12467, %rd12399, %rd12466; + xor.b64 %rd12468, %rd12467, %rd12397; + not.b64 %rd12469, %rd12399; + and.b64 %rd12470, %rd12395, %rd12469; + xor.b64 %rd12471, %rd12470, %rd12398; + not.b64 %rd12472, %rd12395; + and.b64 %rd12473, %rd12396, %rd12472; + xor.b64 %rd12474, %rd12473, %rd12399; + xor.b64 %rd12476, %rd12402, %rd19822; + xor.b64 %rd12477, %rd12417, %rd12476; + xor.b64 %rd12478, %rd12477, %rd12432; + xor.b64 %rd12479, %rd12478, %rd12447; + xor.b64 %rd12480, %rd12479, %rd12462; + xor.b64 %rd12481, %rd12420, %rd12405; + xor.b64 %rd12482, %rd12481, %rd12435; + xor.b64 %rd12483, %rd12482, %rd12450; + xor.b64 %rd12484, %rd12483, %rd12465; + xor.b64 %rd12485, %rd12423, %rd12408; + xor.b64 %rd12486, %rd12485, %rd12438; + xor.b64 %rd12487, %rd12486, %rd12453; + xor.b64 %rd12488, %rd12487, %rd12468; + xor.b64 %rd12489, %rd12426, %rd12411; + xor.b64 %rd12490, %rd12489, %rd12441; + xor.b64 %rd12491, %rd12490, %rd12456; + xor.b64 %rd12492, %rd12491, %rd12471; + xor.b64 %rd12493, %rd12429, %rd12414; + xor.b64 %rd12494, %rd12493, %rd12444; + xor.b64 %rd12495, %rd12494, %rd12459; + xor.b64 %rd12496, %rd12495, %rd12474; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8784}, %rd12484; + } + { + .reg .b32 %dummy; + mov.b64 {%r8785,%dummy}, %rd12484; + } + shf.l.wrap.b32 %r8786, %r8785, %r8784, 1; + shf.l.wrap.b32 %r8787, %r8784, %r8785, 1; + mov.b64 %rd12497, {%r8787, %r8786}; + xor.b64 %rd12498, %rd12496, %rd12497; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8788}, %rd12488; + } + { + .reg .b32 %dummy; + mov.b64 {%r8789,%dummy}, %rd12488; + } + shf.l.wrap.b32 %r8790, %r8789, %r8788, 1; + shf.l.wrap.b32 %r8791, %r8788, %r8789, 1; + mov.b64 %rd12499, {%r8791, %r8790}; + xor.b64 %rd12500, %rd12499, %rd12480; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8792}, %rd12492; + } + { + .reg .b32 %dummy; + mov.b64 {%r8793,%dummy}, %rd12492; + } + shf.l.wrap.b32 %r8794, %r8793, %r8792, 1; + shf.l.wrap.b32 %r8795, %r8792, %r8793, 1; + mov.b64 %rd12501, {%r8795, %r8794}; + xor.b64 %rd12502, %rd12501, %rd12484; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8796}, %rd12496; + } + { + .reg .b32 %dummy; + mov.b64 {%r8797,%dummy}, %rd12496; + } + shf.l.wrap.b32 %r8798, %r8797, %r8796, 1; + shf.l.wrap.b32 %r8799, %r8796, %r8797, 1; + mov.b64 %rd12503, {%r8799, %r8798}; + xor.b64 %rd12504, %rd12503, %rd12488; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8800}, %rd12480; + } + { + .reg .b32 %dummy; + mov.b64 {%r8801,%dummy}, %rd12480; + } + shf.l.wrap.b32 %r8802, %r8801, %r8800, 1; + shf.l.wrap.b32 %r8803, %r8800, %r8801, 1; + mov.b64 %rd12505, {%r8803, %r8802}; + xor.b64 %rd12506, %rd12492, %rd12505; + xor.b64 %rd12507, %rd12498, %rd12476; + xor.b64 %rd12508, %rd12500, %rd12405; + xor.b64 %rd12509, %rd12502, %rd12408; + xor.b64 %rd12510, %rd12504, %rd12411; + xor.b64 %rd12511, %rd12506, %rd12414; + xor.b64 %rd12512, %rd12498, %rd12417; + xor.b64 %rd12513, %rd12500, %rd12420; + xor.b64 %rd12514, %rd12502, %rd12423; + xor.b64 %rd12515, %rd12504, %rd12426; + xor.b64 %rd12516, %rd12506, %rd12429; + xor.b64 %rd12517, %rd12498, %rd12432; + xor.b64 %rd12518, %rd12500, %rd12435; + xor.b64 %rd12519, %rd12502, %rd12438; + xor.b64 %rd12520, %rd12504, %rd12441; + xor.b64 %rd12521, %rd12506, %rd12444; + xor.b64 %rd12522, %rd12498, %rd12447; + xor.b64 %rd12523, %rd12500, %rd12450; + xor.b64 %rd12524, %rd12502, %rd12453; + xor.b64 %rd12525, %rd12504, %rd12456; + xor.b64 %rd12526, %rd12506, %rd12459; + xor.b64 %rd12527, %rd12498, %rd12462; + xor.b64 %rd12528, %rd12500, %rd12465; + xor.b64 %rd12529, %rd12502, %rd12468; + xor.b64 %rd12530, %rd12504, %rd12471; + xor.b64 %rd12531, %rd12506, %rd12474; + { + .reg .b32 %dummy; + mov.b64 {%r8804,%dummy}, %rd12513; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8805}, %rd12513; + } + shf.r.wrap.b32 %r8806, %r8805, %r8804, 20; + shf.r.wrap.b32 %r8807, %r8804, %r8805, 20; + mov.b64 %rd12532, {%r8807, %r8806}; + { + .reg .b32 %dummy; + mov.b64 {%r8808,%dummy}, %rd12519; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8809}, %rd12519; + } + shf.r.wrap.b32 %r8810, %r8809, %r8808, 21; + shf.r.wrap.b32 %r8811, %r8808, %r8809, 21; + mov.b64 %rd12533, {%r8811, %r8810}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8812}, %rd12525; + } + { + .reg .b32 %dummy; + mov.b64 {%r8813,%dummy}, %rd12525; + } + shf.l.wrap.b32 %r8814, %r8813, %r8812, 21; + shf.l.wrap.b32 %r8815, %r8812, %r8813, 21; + mov.b64 %rd12534, {%r8815, %r8814}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8816}, %rd12531; + } + { + .reg .b32 %dummy; + mov.b64 {%r8817,%dummy}, %rd12531; + } + shf.l.wrap.b32 %r8818, %r8817, %r8816, 14; + shf.l.wrap.b32 %r8819, %r8816, %r8817, 14; + mov.b64 %rd12535, {%r8819, %r8818}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8820}, %rd12510; + } + { + .reg .b32 %dummy; + mov.b64 {%r8821,%dummy}, %rd12510; + } + shf.l.wrap.b32 %r8822, %r8821, %r8820, 28; + shf.l.wrap.b32 %r8823, %r8820, %r8821, 28; + mov.b64 %rd12536, {%r8823, %r8822}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8824}, %rd12516; + } + { + .reg .b32 %dummy; + mov.b64 {%r8825,%dummy}, %rd12516; + } + shf.l.wrap.b32 %r8826, %r8825, %r8824, 20; + shf.l.wrap.b32 %r8827, %r8824, %r8825, 20; + mov.b64 %rd12537, {%r8827, %r8826}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8828}, %rd12517; + } + { + .reg .b32 %dummy; + mov.b64 {%r8829,%dummy}, %rd12517; + } + shf.l.wrap.b32 %r8830, %r8829, %r8828, 3; + shf.l.wrap.b32 %r8831, %r8828, %r8829, 3; + mov.b64 %rd12538, {%r8831, %r8830}; + { + .reg .b32 %dummy; + mov.b64 {%r8832,%dummy}, %rd12523; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8833}, %rd12523; + } + shf.r.wrap.b32 %r8834, %r8833, %r8832, 19; + shf.r.wrap.b32 %r8835, %r8832, %r8833, 19; + mov.b64 %rd12539, {%r8835, %r8834}; + { + .reg .b32 %dummy; + mov.b64 {%r8836,%dummy}, %rd12529; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8837}, %rd12529; + } + shf.r.wrap.b32 %r8838, %r8837, %r8836, 3; + shf.r.wrap.b32 %r8839, %r8836, %r8837, 3; + mov.b64 %rd12540, {%r8839, %r8838}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8840}, %rd12508; + } + { + .reg .b32 %dummy; + mov.b64 {%r8841,%dummy}, %rd12508; + } + shf.l.wrap.b32 %r8842, %r8841, %r8840, 1; + shf.l.wrap.b32 %r8843, %r8840, %r8841, 1; + mov.b64 %rd12541, {%r8843, %r8842}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8844}, %rd12514; + } + { + .reg .b32 %dummy; + mov.b64 {%r8845,%dummy}, %rd12514; + } + shf.l.wrap.b32 %r8846, %r8845, %r8844, 6; + shf.l.wrap.b32 %r8847, %r8844, %r8845, 6; + mov.b64 %rd12542, {%r8847, %r8846}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8848}, %rd12520; + } + { + .reg .b32 %dummy; + mov.b64 {%r8849,%dummy}, %rd12520; + } + shf.l.wrap.b32 %r8850, %r8849, %r8848, 25; + shf.l.wrap.b32 %r8851, %r8848, %r8849, 25; + mov.b64 %rd12543, {%r8851, %r8850}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8852}, %rd12526; + } + { + .reg .b32 %dummy; + mov.b64 {%r8853,%dummy}, %rd12526; + } + shf.l.wrap.b32 %r8854, %r8853, %r8852, 8; + shf.l.wrap.b32 %r8855, %r8852, %r8853, 8; + mov.b64 %rd12544, {%r8855, %r8854}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8856}, %rd12527; + } + { + .reg .b32 %dummy; + mov.b64 {%r8857,%dummy}, %rd12527; + } + shf.l.wrap.b32 %r8858, %r8857, %r8856, 18; + shf.l.wrap.b32 %r8859, %r8856, %r8857, 18; + mov.b64 %rd12545, {%r8859, %r8858}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8860}, %rd12511; + } + { + .reg .b32 %dummy; + mov.b64 {%r8861,%dummy}, %rd12511; + } + shf.l.wrap.b32 %r8862, %r8861, %r8860, 27; + shf.l.wrap.b32 %r8863, %r8860, %r8861, 27; + mov.b64 %rd12546, {%r8863, %r8862}; + { + .reg .b32 %dummy; + mov.b64 {%r8864,%dummy}, %rd12512; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8865}, %rd12512; + } + shf.r.wrap.b32 %r8866, %r8865, %r8864, 28; + shf.r.wrap.b32 %r8867, %r8864, %r8865, 28; + mov.b64 %rd12547, {%r8867, %r8866}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8868}, %rd12518; + } + { + .reg .b32 %dummy; + mov.b64 {%r8869,%dummy}, %rd12518; + } + shf.l.wrap.b32 %r8870, %r8869, %r8868, 10; + shf.l.wrap.b32 %r8871, %r8868, %r8869, 10; + mov.b64 %rd12548, {%r8871, %r8870}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8872}, %rd12524; + } + { + .reg .b32 %dummy; + mov.b64 {%r8873,%dummy}, %rd12524; + } + shf.l.wrap.b32 %r8874, %r8873, %r8872, 15; + shf.l.wrap.b32 %r8875, %r8872, %r8873, 15; + mov.b64 %rd12549, {%r8875, %r8874}; + { + .reg .b32 %dummy; + mov.b64 {%r8876,%dummy}, %rd12530; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8877}, %rd12530; + } + shf.r.wrap.b32 %r8878, %r8877, %r8876, 8; + shf.r.wrap.b32 %r8879, %r8876, %r8877, 8; + mov.b64 %rd12550, {%r8879, %r8878}; + { + .reg .b32 %dummy; + mov.b64 {%r8880,%dummy}, %rd12509; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8881}, %rd12509; + } + shf.r.wrap.b32 %r8882, %r8881, %r8880, 2; + shf.r.wrap.b32 %r8883, %r8880, %r8881, 2; + mov.b64 %rd12551, {%r8883, %r8882}; + { + .reg .b32 %dummy; + mov.b64 {%r8884,%dummy}, %rd12515; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8885}, %rd12515; + } + shf.r.wrap.b32 %r8886, %r8885, %r8884, 9; + shf.r.wrap.b32 %r8887, %r8884, %r8885, 9; + mov.b64 %rd12552, {%r8887, %r8886}; + { + .reg .b32 %dummy; + mov.b64 {%r8888,%dummy}, %rd12521; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8889}, %rd12521; + } + shf.r.wrap.b32 %r8890, %r8889, %r8888, 25; + shf.r.wrap.b32 %r8891, %r8888, %r8889, 25; + mov.b64 %rd12553, {%r8891, %r8890}; + { + .reg .b32 %dummy; + mov.b64 {%r8892,%dummy}, %rd12522; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8893}, %rd12522; + } + shf.r.wrap.b32 %r8894, %r8893, %r8892, 23; + shf.r.wrap.b32 %r8895, %r8892, %r8893, 23; + mov.b64 %rd12554, {%r8895, %r8894}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8896}, %rd12528; + } + { + .reg .b32 %dummy; + mov.b64 {%r8897,%dummy}, %rd12528; + } + shf.l.wrap.b32 %r8898, %r8897, %r8896, 2; + shf.l.wrap.b32 %r8899, %r8896, %r8897, 2; + mov.b64 %rd12555, {%r8899, %r8898}; + not.b64 %rd12556, %rd12532; + and.b64 %rd12557, %rd12533, %rd12556; + xor.b64 %rd12558, %rd12557, %rd12507; + not.b64 %rd12559, %rd12533; + and.b64 %rd12560, %rd12534, %rd12559; + xor.b64 %rd12561, %rd12560, %rd12532; + not.b64 %rd12562, %rd12534; + and.b64 %rd12563, %rd12535, %rd12562; + xor.b64 %rd12564, %rd12563, %rd12533; + not.b64 %rd12565, %rd12535; + and.b64 %rd12566, %rd12507, %rd12565; + xor.b64 %rd12567, %rd12566, %rd12534; + not.b64 %rd12568, %rd12507; + and.b64 %rd12569, %rd12532, %rd12568; + xor.b64 %rd12570, %rd12569, %rd12535; + not.b64 %rd12571, %rd12537; + and.b64 %rd12572, %rd12538, %rd12571; + xor.b64 %rd12573, %rd12572, %rd12536; + not.b64 %rd12574, %rd12538; + and.b64 %rd12575, %rd12539, %rd12574; + xor.b64 %rd12576, %rd12575, %rd12537; + not.b64 %rd12577, %rd12539; + and.b64 %rd12578, %rd12540, %rd12577; + xor.b64 %rd12579, %rd12578, %rd12538; + not.b64 %rd12580, %rd12540; + and.b64 %rd12581, %rd12536, %rd12580; + xor.b64 %rd12582, %rd12581, %rd12539; + not.b64 %rd12583, %rd12536; + and.b64 %rd12584, %rd12537, %rd12583; + xor.b64 %rd12585, %rd12584, %rd12540; + not.b64 %rd12586, %rd12542; + and.b64 %rd12587, %rd12543, %rd12586; + xor.b64 %rd12588, %rd12587, %rd12541; + not.b64 %rd12589, %rd12543; + and.b64 %rd12590, %rd12544, %rd12589; + xor.b64 %rd12591, %rd12590, %rd12542; + not.b64 %rd12592, %rd12544; + and.b64 %rd12593, %rd12545, %rd12592; + xor.b64 %rd12594, %rd12593, %rd12543; + not.b64 %rd12595, %rd12545; + and.b64 %rd12596, %rd12541, %rd12595; + xor.b64 %rd12597, %rd12596, %rd12544; + not.b64 %rd12598, %rd12541; + and.b64 %rd12599, %rd12542, %rd12598; + xor.b64 %rd12600, %rd12599, %rd12545; + not.b64 %rd12601, %rd12547; + and.b64 %rd12602, %rd12548, %rd12601; + xor.b64 %rd12603, %rd12602, %rd12546; + not.b64 %rd12604, %rd12548; + and.b64 %rd12605, %rd12549, %rd12604; + xor.b64 %rd12606, %rd12605, %rd12547; + not.b64 %rd12607, %rd12549; + and.b64 %rd12608, %rd12550, %rd12607; + xor.b64 %rd12609, %rd12608, %rd12548; + not.b64 %rd12610, %rd12550; + and.b64 %rd12611, %rd12546, %rd12610; + xor.b64 %rd12612, %rd12611, %rd12549; + not.b64 %rd12613, %rd12546; + and.b64 %rd12614, %rd12547, %rd12613; + xor.b64 %rd12615, %rd12614, %rd12550; + not.b64 %rd12616, %rd12552; + and.b64 %rd12617, %rd12553, %rd12616; + xor.b64 %rd12618, %rd12617, %rd12551; + not.b64 %rd12619, %rd12553; + and.b64 %rd12620, %rd12554, %rd12619; + xor.b64 %rd12621, %rd12620, %rd12552; + not.b64 %rd12622, %rd12554; + and.b64 %rd12623, %rd12555, %rd12622; + xor.b64 %rd12624, %rd12623, %rd12553; + not.b64 %rd12625, %rd12555; + and.b64 %rd12626, %rd12551, %rd12625; + xor.b64 %rd12627, %rd12626, %rd12554; + not.b64 %rd12628, %rd12551; + and.b64 %rd12629, %rd12552, %rd12628; + xor.b64 %rd12630, %rd12629, %rd12555; + xor.b64 %rd12631, %rd12558, %rd19823; + xor.b64 %rd12632, %rd12573, %rd12631; + xor.b64 %rd12633, %rd12632, %rd12588; + xor.b64 %rd12634, %rd12633, %rd12603; + xor.b64 %rd12635, %rd12634, %rd12618; + xor.b64 %rd12636, %rd12576, %rd12561; + xor.b64 %rd12637, %rd12636, %rd12591; + xor.b64 %rd12638, %rd12637, %rd12606; + xor.b64 %rd12639, %rd12638, %rd12621; + xor.b64 %rd12640, %rd12579, %rd12564; + xor.b64 %rd12641, %rd12640, %rd12594; + xor.b64 %rd12642, %rd12641, %rd12609; + xor.b64 %rd12643, %rd12642, %rd12624; + xor.b64 %rd12644, %rd12582, %rd12567; + xor.b64 %rd12645, %rd12644, %rd12597; + xor.b64 %rd12646, %rd12645, %rd12612; + xor.b64 %rd12647, %rd12646, %rd12627; + xor.b64 %rd12648, %rd12585, %rd12570; + xor.b64 %rd12649, %rd12648, %rd12600; + xor.b64 %rd12650, %rd12649, %rd12615; + xor.b64 %rd12651, %rd12650, %rd12630; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8900}, %rd12639; + } + { + .reg .b32 %dummy; + mov.b64 {%r8901,%dummy}, %rd12639; + } + shf.l.wrap.b32 %r8902, %r8901, %r8900, 1; + shf.l.wrap.b32 %r8903, %r8900, %r8901, 1; + mov.b64 %rd12652, {%r8903, %r8902}; + xor.b64 %rd12653, %rd12651, %rd12652; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8904}, %rd12643; + } + { + .reg .b32 %dummy; + mov.b64 {%r8905,%dummy}, %rd12643; + } + shf.l.wrap.b32 %r8906, %r8905, %r8904, 1; + shf.l.wrap.b32 %r8907, %r8904, %r8905, 1; + mov.b64 %rd12654, {%r8907, %r8906}; + xor.b64 %rd12655, %rd12654, %rd12635; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8908}, %rd12647; + } + { + .reg .b32 %dummy; + mov.b64 {%r8909,%dummy}, %rd12647; + } + shf.l.wrap.b32 %r8910, %r8909, %r8908, 1; + shf.l.wrap.b32 %r8911, %r8908, %r8909, 1; + mov.b64 %rd12656, {%r8911, %r8910}; + xor.b64 %rd12657, %rd12656, %rd12639; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8912}, %rd12651; + } + { + .reg .b32 %dummy; + mov.b64 {%r8913,%dummy}, %rd12651; + } + shf.l.wrap.b32 %r8914, %r8913, %r8912, 1; + shf.l.wrap.b32 %r8915, %r8912, %r8913, 1; + mov.b64 %rd12658, {%r8915, %r8914}; + xor.b64 %rd12659, %rd12658, %rd12643; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8916}, %rd12635; + } + { + .reg .b32 %dummy; + mov.b64 {%r8917,%dummy}, %rd12635; + } + shf.l.wrap.b32 %r8918, %r8917, %r8916, 1; + shf.l.wrap.b32 %r8919, %r8916, %r8917, 1; + mov.b64 %rd12660, {%r8919, %r8918}; + xor.b64 %rd12661, %rd12647, %rd12660; + xor.b64 %rd12662, %rd12653, %rd12631; + xor.b64 %rd12663, %rd12655, %rd12561; + xor.b64 %rd12664, %rd12657, %rd12564; + xor.b64 %rd12665, %rd12659, %rd12567; + xor.b64 %rd12666, %rd12661, %rd12570; + xor.b64 %rd12667, %rd12653, %rd12573; + xor.b64 %rd12668, %rd12655, %rd12576; + xor.b64 %rd12669, %rd12657, %rd12579; + xor.b64 %rd12670, %rd12659, %rd12582; + xor.b64 %rd12671, %rd12661, %rd12585; + xor.b64 %rd12672, %rd12653, %rd12588; + xor.b64 %rd12673, %rd12655, %rd12591; + xor.b64 %rd12674, %rd12657, %rd12594; + xor.b64 %rd12675, %rd12659, %rd12597; + xor.b64 %rd12676, %rd12661, %rd12600; + xor.b64 %rd12677, %rd12653, %rd12603; + xor.b64 %rd12678, %rd12655, %rd12606; + xor.b64 %rd12679, %rd12657, %rd12609; + xor.b64 %rd12680, %rd12659, %rd12612; + xor.b64 %rd12681, %rd12661, %rd12615; + xor.b64 %rd12682, %rd12653, %rd12618; + xor.b64 %rd12683, %rd12655, %rd12621; + xor.b64 %rd12684, %rd12657, %rd12624; + xor.b64 %rd12685, %rd12659, %rd12627; + xor.b64 %rd12686, %rd12661, %rd12630; + { + .reg .b32 %dummy; + mov.b64 {%r8920,%dummy}, %rd12668; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8921}, %rd12668; + } + shf.r.wrap.b32 %r8922, %r8921, %r8920, 20; + shf.r.wrap.b32 %r8923, %r8920, %r8921, 20; + mov.b64 %rd12687, {%r8923, %r8922}; + { + .reg .b32 %dummy; + mov.b64 {%r8924,%dummy}, %rd12674; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8925}, %rd12674; + } + shf.r.wrap.b32 %r8926, %r8925, %r8924, 21; + shf.r.wrap.b32 %r8927, %r8924, %r8925, 21; + mov.b64 %rd12688, {%r8927, %r8926}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8928}, %rd12680; + } + { + .reg .b32 %dummy; + mov.b64 {%r8929,%dummy}, %rd12680; + } + shf.l.wrap.b32 %r8930, %r8929, %r8928, 21; + shf.l.wrap.b32 %r8931, %r8928, %r8929, 21; + mov.b64 %rd12689, {%r8931, %r8930}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8932}, %rd12686; + } + { + .reg .b32 %dummy; + mov.b64 {%r8933,%dummy}, %rd12686; + } + shf.l.wrap.b32 %r8934, %r8933, %r8932, 14; + shf.l.wrap.b32 %r8935, %r8932, %r8933, 14; + mov.b64 %rd12690, {%r8935, %r8934}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8936}, %rd12665; + } + { + .reg .b32 %dummy; + mov.b64 {%r8937,%dummy}, %rd12665; + } + shf.l.wrap.b32 %r8938, %r8937, %r8936, 28; + shf.l.wrap.b32 %r8939, %r8936, %r8937, 28; + mov.b64 %rd12691, {%r8939, %r8938}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8940}, %rd12671; + } + { + .reg .b32 %dummy; + mov.b64 {%r8941,%dummy}, %rd12671; + } + shf.l.wrap.b32 %r8942, %r8941, %r8940, 20; + shf.l.wrap.b32 %r8943, %r8940, %r8941, 20; + mov.b64 %rd12692, {%r8943, %r8942}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8944}, %rd12672; + } + { + .reg .b32 %dummy; + mov.b64 {%r8945,%dummy}, %rd12672; + } + shf.l.wrap.b32 %r8946, %r8945, %r8944, 3; + shf.l.wrap.b32 %r8947, %r8944, %r8945, 3; + mov.b64 %rd12693, {%r8947, %r8946}; + { + .reg .b32 %dummy; + mov.b64 {%r8948,%dummy}, %rd12678; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8949}, %rd12678; + } + shf.r.wrap.b32 %r8950, %r8949, %r8948, 19; + shf.r.wrap.b32 %r8951, %r8948, %r8949, 19; + mov.b64 %rd12694, {%r8951, %r8950}; + { + .reg .b32 %dummy; + mov.b64 {%r8952,%dummy}, %rd12684; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8953}, %rd12684; + } + shf.r.wrap.b32 %r8954, %r8953, %r8952, 3; + shf.r.wrap.b32 %r8955, %r8952, %r8953, 3; + mov.b64 %rd12695, {%r8955, %r8954}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8956}, %rd12663; + } + { + .reg .b32 %dummy; + mov.b64 {%r8957,%dummy}, %rd12663; + } + shf.l.wrap.b32 %r8958, %r8957, %r8956, 1; + shf.l.wrap.b32 %r8959, %r8956, %r8957, 1; + mov.b64 %rd12696, {%r8959, %r8958}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8960}, %rd12669; + } + { + .reg .b32 %dummy; + mov.b64 {%r8961,%dummy}, %rd12669; + } + shf.l.wrap.b32 %r8962, %r8961, %r8960, 6; + shf.l.wrap.b32 %r8963, %r8960, %r8961, 6; + mov.b64 %rd12697, {%r8963, %r8962}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8964}, %rd12675; + } + { + .reg .b32 %dummy; + mov.b64 {%r8965,%dummy}, %rd12675; + } + shf.l.wrap.b32 %r8966, %r8965, %r8964, 25; + shf.l.wrap.b32 %r8967, %r8964, %r8965, 25; + mov.b64 %rd12698, {%r8967, %r8966}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8968}, %rd12681; + } + { + .reg .b32 %dummy; + mov.b64 {%r8969,%dummy}, %rd12681; + } + shf.l.wrap.b32 %r8970, %r8969, %r8968, 8; + shf.l.wrap.b32 %r8971, %r8968, %r8969, 8; + mov.b64 %rd12699, {%r8971, %r8970}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8972}, %rd12682; + } + { + .reg .b32 %dummy; + mov.b64 {%r8973,%dummy}, %rd12682; + } + shf.l.wrap.b32 %r8974, %r8973, %r8972, 18; + shf.l.wrap.b32 %r8975, %r8972, %r8973, 18; + mov.b64 %rd12700, {%r8975, %r8974}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8976}, %rd12666; + } + { + .reg .b32 %dummy; + mov.b64 {%r8977,%dummy}, %rd12666; + } + shf.l.wrap.b32 %r8978, %r8977, %r8976, 27; + shf.l.wrap.b32 %r8979, %r8976, %r8977, 27; + mov.b64 %rd12701, {%r8979, %r8978}; + { + .reg .b32 %dummy; + mov.b64 {%r8980,%dummy}, %rd12667; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8981}, %rd12667; + } + shf.r.wrap.b32 %r8982, %r8981, %r8980, 28; + shf.r.wrap.b32 %r8983, %r8980, %r8981, 28; + mov.b64 %rd12702, {%r8983, %r8982}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8984}, %rd12673; + } + { + .reg .b32 %dummy; + mov.b64 {%r8985,%dummy}, %rd12673; + } + shf.l.wrap.b32 %r8986, %r8985, %r8984, 10; + shf.l.wrap.b32 %r8987, %r8984, %r8985, 10; + mov.b64 %rd12703, {%r8987, %r8986}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8988}, %rd12679; + } + { + .reg .b32 %dummy; + mov.b64 {%r8989,%dummy}, %rd12679; + } + shf.l.wrap.b32 %r8990, %r8989, %r8988, 15; + shf.l.wrap.b32 %r8991, %r8988, %r8989, 15; + mov.b64 %rd12704, {%r8991, %r8990}; + { + .reg .b32 %dummy; + mov.b64 {%r8992,%dummy}, %rd12685; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8993}, %rd12685; + } + shf.r.wrap.b32 %r8994, %r8993, %r8992, 8; + shf.r.wrap.b32 %r8995, %r8992, %r8993, 8; + mov.b64 %rd12705, {%r8995, %r8994}; + { + .reg .b32 %dummy; + mov.b64 {%r8996,%dummy}, %rd12664; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r8997}, %rd12664; + } + shf.r.wrap.b32 %r8998, %r8997, %r8996, 2; + shf.r.wrap.b32 %r8999, %r8996, %r8997, 2; + mov.b64 %rd12706, {%r8999, %r8998}; + { + .reg .b32 %dummy; + mov.b64 {%r9000,%dummy}, %rd12670; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9001}, %rd12670; + } + shf.r.wrap.b32 %r9002, %r9001, %r9000, 9; + shf.r.wrap.b32 %r9003, %r9000, %r9001, 9; + mov.b64 %rd12707, {%r9003, %r9002}; + { + .reg .b32 %dummy; + mov.b64 {%r9004,%dummy}, %rd12676; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9005}, %rd12676; + } + shf.r.wrap.b32 %r9006, %r9005, %r9004, 25; + shf.r.wrap.b32 %r9007, %r9004, %r9005, 25; + mov.b64 %rd12708, {%r9007, %r9006}; + { + .reg .b32 %dummy; + mov.b64 {%r9008,%dummy}, %rd12677; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9009}, %rd12677; + } + shf.r.wrap.b32 %r9010, %r9009, %r9008, 23; + shf.r.wrap.b32 %r9011, %r9008, %r9009, 23; + mov.b64 %rd12709, {%r9011, %r9010}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9012}, %rd12683; + } + { + .reg .b32 %dummy; + mov.b64 {%r9013,%dummy}, %rd12683; + } + shf.l.wrap.b32 %r9014, %r9013, %r9012, 2; + shf.l.wrap.b32 %r9015, %r9012, %r9013, 2; + mov.b64 %rd12710, {%r9015, %r9014}; + not.b64 %rd12711, %rd12687; + and.b64 %rd12712, %rd12688, %rd12711; + xor.b64 %rd12713, %rd12712, %rd12662; + not.b64 %rd12714, %rd12688; + and.b64 %rd12715, %rd12689, %rd12714; + xor.b64 %rd12716, %rd12715, %rd12687; + not.b64 %rd12717, %rd12689; + and.b64 %rd12718, %rd12690, %rd12717; + xor.b64 %rd12719, %rd12718, %rd12688; + not.b64 %rd12720, %rd12690; + and.b64 %rd12721, %rd12662, %rd12720; + xor.b64 %rd12722, %rd12721, %rd12689; + not.b64 %rd12723, %rd12662; + and.b64 %rd12724, %rd12687, %rd12723; + xor.b64 %rd12725, %rd12724, %rd12690; + not.b64 %rd12726, %rd12692; + and.b64 %rd12727, %rd12693, %rd12726; + xor.b64 %rd12728, %rd12727, %rd12691; + not.b64 %rd12729, %rd12693; + and.b64 %rd12730, %rd12694, %rd12729; + xor.b64 %rd12731, %rd12730, %rd12692; + not.b64 %rd12732, %rd12694; + and.b64 %rd12733, %rd12695, %rd12732; + xor.b64 %rd12734, %rd12733, %rd12693; + not.b64 %rd12735, %rd12695; + and.b64 %rd12736, %rd12691, %rd12735; + xor.b64 %rd12737, %rd12736, %rd12694; + not.b64 %rd12738, %rd12691; + and.b64 %rd12739, %rd12692, %rd12738; + xor.b64 %rd12740, %rd12739, %rd12695; + not.b64 %rd12741, %rd12697; + and.b64 %rd12742, %rd12698, %rd12741; + xor.b64 %rd12743, %rd12742, %rd12696; + not.b64 %rd12744, %rd12698; + and.b64 %rd12745, %rd12699, %rd12744; + xor.b64 %rd12746, %rd12745, %rd12697; + not.b64 %rd12747, %rd12699; + and.b64 %rd12748, %rd12700, %rd12747; + xor.b64 %rd12749, %rd12748, %rd12698; + not.b64 %rd12750, %rd12700; + and.b64 %rd12751, %rd12696, %rd12750; + xor.b64 %rd12752, %rd12751, %rd12699; + not.b64 %rd12753, %rd12696; + and.b64 %rd12754, %rd12697, %rd12753; + xor.b64 %rd12755, %rd12754, %rd12700; + not.b64 %rd12756, %rd12702; + and.b64 %rd12757, %rd12703, %rd12756; + xor.b64 %rd12758, %rd12757, %rd12701; + not.b64 %rd12759, %rd12703; + and.b64 %rd12760, %rd12704, %rd12759; + xor.b64 %rd12761, %rd12760, %rd12702; + not.b64 %rd12762, %rd12704; + and.b64 %rd12763, %rd12705, %rd12762; + xor.b64 %rd12764, %rd12763, %rd12703; + not.b64 %rd12765, %rd12705; + and.b64 %rd12766, %rd12701, %rd12765; + xor.b64 %rd12767, %rd12766, %rd12704; + not.b64 %rd12768, %rd12701; + and.b64 %rd12769, %rd12702, %rd12768; + xor.b64 %rd12770, %rd12769, %rd12705; + not.b64 %rd12771, %rd12707; + and.b64 %rd12772, %rd12708, %rd12771; + xor.b64 %rd12773, %rd12772, %rd12706; + not.b64 %rd12774, %rd12708; + and.b64 %rd12775, %rd12709, %rd12774; + xor.b64 %rd12776, %rd12775, %rd12707; + not.b64 %rd12777, %rd12709; + and.b64 %rd12778, %rd12710, %rd12777; + xor.b64 %rd12779, %rd12778, %rd12708; + not.b64 %rd12780, %rd12710; + and.b64 %rd12781, %rd12706, %rd12780; + xor.b64 %rd12782, %rd12781, %rd12709; + not.b64 %rd12783, %rd12706; + and.b64 %rd12784, %rd12707, %rd12783; + xor.b64 %rd12785, %rd12784, %rd12710; + xor.b64 %rd12786, %rd12713, %rd19824; + xor.b64 %rd12787, %rd12728, %rd12786; + xor.b64 %rd12788, %rd12787, %rd12743; + xor.b64 %rd12789, %rd12788, %rd12758; + xor.b64 %rd12790, %rd12789, %rd12773; + xor.b64 %rd12791, %rd12731, %rd12716; + xor.b64 %rd12792, %rd12791, %rd12746; + xor.b64 %rd12793, %rd12792, %rd12761; + xor.b64 %rd12794, %rd12793, %rd12776; + xor.b64 %rd12795, %rd12734, %rd12719; + xor.b64 %rd12796, %rd12795, %rd12749; + xor.b64 %rd12797, %rd12796, %rd12764; + xor.b64 %rd12798, %rd12797, %rd12779; + xor.b64 %rd12799, %rd12737, %rd12722; + xor.b64 %rd12800, %rd12799, %rd12752; + xor.b64 %rd12801, %rd12800, %rd12767; + xor.b64 %rd12802, %rd12801, %rd12782; + xor.b64 %rd12803, %rd12740, %rd12725; + xor.b64 %rd12804, %rd12803, %rd12755; + xor.b64 %rd12805, %rd12804, %rd12770; + xor.b64 %rd12806, %rd12805, %rd12785; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9016}, %rd12794; + } + { + .reg .b32 %dummy; + mov.b64 {%r9017,%dummy}, %rd12794; + } + shf.l.wrap.b32 %r9018, %r9017, %r9016, 1; + shf.l.wrap.b32 %r9019, %r9016, %r9017, 1; + mov.b64 %rd12807, {%r9019, %r9018}; + xor.b64 %rd12808, %rd12806, %rd12807; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9020}, %rd12798; + } + { + .reg .b32 %dummy; + mov.b64 {%r9021,%dummy}, %rd12798; + } + shf.l.wrap.b32 %r9022, %r9021, %r9020, 1; + shf.l.wrap.b32 %r9023, %r9020, %r9021, 1; + mov.b64 %rd12809, {%r9023, %r9022}; + xor.b64 %rd12810, %rd12809, %rd12790; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9024}, %rd12802; + } + { + .reg .b32 %dummy; + mov.b64 {%r9025,%dummy}, %rd12802; + } + shf.l.wrap.b32 %r9026, %r9025, %r9024, 1; + shf.l.wrap.b32 %r9027, %r9024, %r9025, 1; + mov.b64 %rd12811, {%r9027, %r9026}; + xor.b64 %rd12812, %rd12811, %rd12794; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9028}, %rd12806; + } + { + .reg .b32 %dummy; + mov.b64 {%r9029,%dummy}, %rd12806; + } + shf.l.wrap.b32 %r9030, %r9029, %r9028, 1; + shf.l.wrap.b32 %r9031, %r9028, %r9029, 1; + mov.b64 %rd12813, {%r9031, %r9030}; + xor.b64 %rd12814, %rd12813, %rd12798; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9032}, %rd12790; + } + { + .reg .b32 %dummy; + mov.b64 {%r9033,%dummy}, %rd12790; + } + shf.l.wrap.b32 %r9034, %r9033, %r9032, 1; + shf.l.wrap.b32 %r9035, %r9032, %r9033, 1; + mov.b64 %rd12815, {%r9035, %r9034}; + xor.b64 %rd12816, %rd12802, %rd12815; + xor.b64 %rd12817, %rd12808, %rd12786; + xor.b64 %rd12818, %rd12810, %rd12716; + xor.b64 %rd12819, %rd12812, %rd12719; + xor.b64 %rd12820, %rd12814, %rd12722; + xor.b64 %rd12821, %rd12816, %rd12725; + xor.b64 %rd12822, %rd12808, %rd12728; + xor.b64 %rd12823, %rd12810, %rd12731; + xor.b64 %rd12824, %rd12812, %rd12734; + xor.b64 %rd12825, %rd12814, %rd12737; + xor.b64 %rd12826, %rd12816, %rd12740; + xor.b64 %rd12827, %rd12808, %rd12743; + xor.b64 %rd12828, %rd12810, %rd12746; + xor.b64 %rd12829, %rd12812, %rd12749; + xor.b64 %rd12830, %rd12814, %rd12752; + xor.b64 %rd12831, %rd12816, %rd12755; + xor.b64 %rd12832, %rd12808, %rd12758; + xor.b64 %rd12833, %rd12810, %rd12761; + xor.b64 %rd12834, %rd12812, %rd12764; + xor.b64 %rd12835, %rd12814, %rd12767; + xor.b64 %rd12836, %rd12816, %rd12770; + xor.b64 %rd12837, %rd12808, %rd12773; + xor.b64 %rd12838, %rd12810, %rd12776; + xor.b64 %rd12839, %rd12812, %rd12779; + xor.b64 %rd12840, %rd12814, %rd12782; + xor.b64 %rd12841, %rd12816, %rd12785; + { + .reg .b32 %dummy; + mov.b64 {%r9036,%dummy}, %rd12823; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9037}, %rd12823; + } + shf.r.wrap.b32 %r9038, %r9037, %r9036, 20; + shf.r.wrap.b32 %r9039, %r9036, %r9037, 20; + mov.b64 %rd12842, {%r9039, %r9038}; + { + .reg .b32 %dummy; + mov.b64 {%r9040,%dummy}, %rd12829; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9041}, %rd12829; + } + shf.r.wrap.b32 %r9042, %r9041, %r9040, 21; + shf.r.wrap.b32 %r9043, %r9040, %r9041, 21; + mov.b64 %rd12843, {%r9043, %r9042}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9044}, %rd12835; + } + { + .reg .b32 %dummy; + mov.b64 {%r9045,%dummy}, %rd12835; + } + shf.l.wrap.b32 %r9046, %r9045, %r9044, 21; + shf.l.wrap.b32 %r9047, %r9044, %r9045, 21; + mov.b64 %rd12844, {%r9047, %r9046}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9048}, %rd12841; + } + { + .reg .b32 %dummy; + mov.b64 {%r9049,%dummy}, %rd12841; + } + shf.l.wrap.b32 %r9050, %r9049, %r9048, 14; + shf.l.wrap.b32 %r9051, %r9048, %r9049, 14; + mov.b64 %rd12845, {%r9051, %r9050}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9052}, %rd12820; + } + { + .reg .b32 %dummy; + mov.b64 {%r9053,%dummy}, %rd12820; + } + shf.l.wrap.b32 %r9054, %r9053, %r9052, 28; + shf.l.wrap.b32 %r9055, %r9052, %r9053, 28; + mov.b64 %rd12846, {%r9055, %r9054}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9056}, %rd12826; + } + { + .reg .b32 %dummy; + mov.b64 {%r9057,%dummy}, %rd12826; + } + shf.l.wrap.b32 %r9058, %r9057, %r9056, 20; + shf.l.wrap.b32 %r9059, %r9056, %r9057, 20; + mov.b64 %rd12847, {%r9059, %r9058}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9060}, %rd12827; + } + { + .reg .b32 %dummy; + mov.b64 {%r9061,%dummy}, %rd12827; + } + shf.l.wrap.b32 %r9062, %r9061, %r9060, 3; + shf.l.wrap.b32 %r9063, %r9060, %r9061, 3; + mov.b64 %rd12848, {%r9063, %r9062}; + { + .reg .b32 %dummy; + mov.b64 {%r9064,%dummy}, %rd12833; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9065}, %rd12833; + } + shf.r.wrap.b32 %r9066, %r9065, %r9064, 19; + shf.r.wrap.b32 %r9067, %r9064, %r9065, 19; + mov.b64 %rd12849, {%r9067, %r9066}; + { + .reg .b32 %dummy; + mov.b64 {%r9068,%dummy}, %rd12839; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9069}, %rd12839; + } + shf.r.wrap.b32 %r9070, %r9069, %r9068, 3; + shf.r.wrap.b32 %r9071, %r9068, %r9069, 3; + mov.b64 %rd12850, {%r9071, %r9070}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9072}, %rd12818; + } + { + .reg .b32 %dummy; + mov.b64 {%r9073,%dummy}, %rd12818; + } + shf.l.wrap.b32 %r9074, %r9073, %r9072, 1; + shf.l.wrap.b32 %r9075, %r9072, %r9073, 1; + mov.b64 %rd12851, {%r9075, %r9074}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9076}, %rd12824; + } + { + .reg .b32 %dummy; + mov.b64 {%r9077,%dummy}, %rd12824; + } + shf.l.wrap.b32 %r9078, %r9077, %r9076, 6; + shf.l.wrap.b32 %r9079, %r9076, %r9077, 6; + mov.b64 %rd12852, {%r9079, %r9078}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9080}, %rd12830; + } + { + .reg .b32 %dummy; + mov.b64 {%r9081,%dummy}, %rd12830; + } + shf.l.wrap.b32 %r9082, %r9081, %r9080, 25; + shf.l.wrap.b32 %r9083, %r9080, %r9081, 25; + mov.b64 %rd12853, {%r9083, %r9082}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9084}, %rd12836; + } + { + .reg .b32 %dummy; + mov.b64 {%r9085,%dummy}, %rd12836; + } + shf.l.wrap.b32 %r9086, %r9085, %r9084, 8; + shf.l.wrap.b32 %r9087, %r9084, %r9085, 8; + mov.b64 %rd12854, {%r9087, %r9086}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9088}, %rd12837; + } + { + .reg .b32 %dummy; + mov.b64 {%r9089,%dummy}, %rd12837; + } + shf.l.wrap.b32 %r9090, %r9089, %r9088, 18; + shf.l.wrap.b32 %r9091, %r9088, %r9089, 18; + mov.b64 %rd12855, {%r9091, %r9090}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9092}, %rd12821; + } + { + .reg .b32 %dummy; + mov.b64 {%r9093,%dummy}, %rd12821; + } + shf.l.wrap.b32 %r9094, %r9093, %r9092, 27; + shf.l.wrap.b32 %r9095, %r9092, %r9093, 27; + mov.b64 %rd12856, {%r9095, %r9094}; + { + .reg .b32 %dummy; + mov.b64 {%r9096,%dummy}, %rd12822; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9097}, %rd12822; + } + shf.r.wrap.b32 %r9098, %r9097, %r9096, 28; + shf.r.wrap.b32 %r9099, %r9096, %r9097, 28; + mov.b64 %rd12857, {%r9099, %r9098}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9100}, %rd12828; + } + { + .reg .b32 %dummy; + mov.b64 {%r9101,%dummy}, %rd12828; + } + shf.l.wrap.b32 %r9102, %r9101, %r9100, 10; + shf.l.wrap.b32 %r9103, %r9100, %r9101, 10; + mov.b64 %rd12858, {%r9103, %r9102}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9104}, %rd12834; + } + { + .reg .b32 %dummy; + mov.b64 {%r9105,%dummy}, %rd12834; + } + shf.l.wrap.b32 %r9106, %r9105, %r9104, 15; + shf.l.wrap.b32 %r9107, %r9104, %r9105, 15; + mov.b64 %rd12859, {%r9107, %r9106}; + { + .reg .b32 %dummy; + mov.b64 {%r9108,%dummy}, %rd12840; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9109}, %rd12840; + } + shf.r.wrap.b32 %r9110, %r9109, %r9108, 8; + shf.r.wrap.b32 %r9111, %r9108, %r9109, 8; + mov.b64 %rd12860, {%r9111, %r9110}; + { + .reg .b32 %dummy; + mov.b64 {%r9112,%dummy}, %rd12819; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9113}, %rd12819; + } + shf.r.wrap.b32 %r9114, %r9113, %r9112, 2; + shf.r.wrap.b32 %r9115, %r9112, %r9113, 2; + mov.b64 %rd12861, {%r9115, %r9114}; + { + .reg .b32 %dummy; + mov.b64 {%r9116,%dummy}, %rd12825; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9117}, %rd12825; + } + shf.r.wrap.b32 %r9118, %r9117, %r9116, 9; + shf.r.wrap.b32 %r9119, %r9116, %r9117, 9; + mov.b64 %rd12862, {%r9119, %r9118}; + { + .reg .b32 %dummy; + mov.b64 {%r9120,%dummy}, %rd12831; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9121}, %rd12831; + } + shf.r.wrap.b32 %r9122, %r9121, %r9120, 25; + shf.r.wrap.b32 %r9123, %r9120, %r9121, 25; + mov.b64 %rd12863, {%r9123, %r9122}; + { + .reg .b32 %dummy; + mov.b64 {%r9124,%dummy}, %rd12832; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9125}, %rd12832; + } + shf.r.wrap.b32 %r9126, %r9125, %r9124, 23; + shf.r.wrap.b32 %r9127, %r9124, %r9125, 23; + mov.b64 %rd12864, {%r9127, %r9126}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9128}, %rd12838; + } + { + .reg .b32 %dummy; + mov.b64 {%r9129,%dummy}, %rd12838; + } + shf.l.wrap.b32 %r9130, %r9129, %r9128, 2; + shf.l.wrap.b32 %r9131, %r9128, %r9129, 2; + mov.b64 %rd12865, {%r9131, %r9130}; + not.b64 %rd12866, %rd12842; + and.b64 %rd12867, %rd12843, %rd12866; + xor.b64 %rd12868, %rd12867, %rd12817; + not.b64 %rd12869, %rd12843; + and.b64 %rd12870, %rd12844, %rd12869; + xor.b64 %rd12871, %rd12870, %rd12842; + not.b64 %rd12872, %rd12844; + and.b64 %rd12873, %rd12845, %rd12872; + xor.b64 %rd12874, %rd12873, %rd12843; + not.b64 %rd12875, %rd12845; + and.b64 %rd12876, %rd12817, %rd12875; + xor.b64 %rd12877, %rd12876, %rd12844; + not.b64 %rd12878, %rd12817; + and.b64 %rd12879, %rd12842, %rd12878; + xor.b64 %rd12880, %rd12879, %rd12845; + not.b64 %rd12881, %rd12847; + and.b64 %rd12882, %rd12848, %rd12881; + xor.b64 %rd12883, %rd12882, %rd12846; + not.b64 %rd12884, %rd12848; + and.b64 %rd12885, %rd12849, %rd12884; + xor.b64 %rd12886, %rd12885, %rd12847; + not.b64 %rd12887, %rd12849; + and.b64 %rd12888, %rd12850, %rd12887; + xor.b64 %rd12889, %rd12888, %rd12848; + not.b64 %rd12890, %rd12850; + and.b64 %rd12891, %rd12846, %rd12890; + xor.b64 %rd12892, %rd12891, %rd12849; + not.b64 %rd12893, %rd12846; + and.b64 %rd12894, %rd12847, %rd12893; + xor.b64 %rd12895, %rd12894, %rd12850; + not.b64 %rd12896, %rd12852; + and.b64 %rd12897, %rd12853, %rd12896; + xor.b64 %rd12898, %rd12897, %rd12851; + not.b64 %rd12899, %rd12853; + and.b64 %rd12900, %rd12854, %rd12899; + xor.b64 %rd12901, %rd12900, %rd12852; + not.b64 %rd12902, %rd12854; + and.b64 %rd12903, %rd12855, %rd12902; + xor.b64 %rd12904, %rd12903, %rd12853; + not.b64 %rd12905, %rd12855; + and.b64 %rd12906, %rd12851, %rd12905; + xor.b64 %rd12907, %rd12906, %rd12854; + not.b64 %rd12908, %rd12851; + and.b64 %rd12909, %rd12852, %rd12908; + xor.b64 %rd12910, %rd12909, %rd12855; + not.b64 %rd12911, %rd12857; + and.b64 %rd12912, %rd12858, %rd12911; + xor.b64 %rd12913, %rd12912, %rd12856; + not.b64 %rd12914, %rd12858; + and.b64 %rd12915, %rd12859, %rd12914; + xor.b64 %rd12916, %rd12915, %rd12857; + not.b64 %rd12917, %rd12859; + and.b64 %rd12918, %rd12860, %rd12917; + xor.b64 %rd12919, %rd12918, %rd12858; + not.b64 %rd12920, %rd12860; + and.b64 %rd12921, %rd12856, %rd12920; + xor.b64 %rd12922, %rd12921, %rd12859; + not.b64 %rd12923, %rd12856; + and.b64 %rd12924, %rd12857, %rd12923; + xor.b64 %rd12925, %rd12924, %rd12860; + not.b64 %rd12926, %rd12862; + and.b64 %rd12927, %rd12863, %rd12926; + xor.b64 %rd12928, %rd12927, %rd12861; + not.b64 %rd12929, %rd12863; + and.b64 %rd12930, %rd12864, %rd12929; + xor.b64 %rd12931, %rd12930, %rd12862; + not.b64 %rd12932, %rd12864; + and.b64 %rd12933, %rd12865, %rd12932; + xor.b64 %rd12934, %rd12933, %rd12863; + not.b64 %rd12935, %rd12865; + and.b64 %rd12936, %rd12861, %rd12935; + xor.b64 %rd12937, %rd12936, %rd12864; + not.b64 %rd12938, %rd12861; + and.b64 %rd12939, %rd12862, %rd12938; + xor.b64 %rd12940, %rd12939, %rd12865; + xor.b64 %rd12941, %rd12868, %rd19825; + xor.b64 %rd12942, %rd12883, %rd12941; + xor.b64 %rd12943, %rd12942, %rd12898; + xor.b64 %rd12944, %rd12943, %rd12913; + xor.b64 %rd12945, %rd12944, %rd12928; + xor.b64 %rd12946, %rd12886, %rd12871; + xor.b64 %rd12947, %rd12946, %rd12901; + xor.b64 %rd12948, %rd12947, %rd12916; + xor.b64 %rd12949, %rd12948, %rd12931; + xor.b64 %rd12950, %rd12889, %rd12874; + xor.b64 %rd12951, %rd12950, %rd12904; + xor.b64 %rd12952, %rd12951, %rd12919; + xor.b64 %rd12953, %rd12952, %rd12934; + xor.b64 %rd12954, %rd12892, %rd12877; + xor.b64 %rd12955, %rd12954, %rd12907; + xor.b64 %rd12956, %rd12955, %rd12922; + xor.b64 %rd12957, %rd12956, %rd12937; + xor.b64 %rd12958, %rd12895, %rd12880; + xor.b64 %rd12959, %rd12958, %rd12910; + xor.b64 %rd12960, %rd12959, %rd12925; + xor.b64 %rd12961, %rd12960, %rd12940; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9132}, %rd12949; + } + { + .reg .b32 %dummy; + mov.b64 {%r9133,%dummy}, %rd12949; + } + shf.l.wrap.b32 %r9134, %r9133, %r9132, 1; + shf.l.wrap.b32 %r9135, %r9132, %r9133, 1; + mov.b64 %rd12962, {%r9135, %r9134}; + xor.b64 %rd12963, %rd12961, %rd12962; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9136}, %rd12953; + } + { + .reg .b32 %dummy; + mov.b64 {%r9137,%dummy}, %rd12953; + } + shf.l.wrap.b32 %r9138, %r9137, %r9136, 1; + shf.l.wrap.b32 %r9139, %r9136, %r9137, 1; + mov.b64 %rd12964, {%r9139, %r9138}; + xor.b64 %rd12965, %rd12964, %rd12945; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9140}, %rd12957; + } + { + .reg .b32 %dummy; + mov.b64 {%r9141,%dummy}, %rd12957; + } + shf.l.wrap.b32 %r9142, %r9141, %r9140, 1; + shf.l.wrap.b32 %r9143, %r9140, %r9141, 1; + mov.b64 %rd12966, {%r9143, %r9142}; + xor.b64 %rd12967, %rd12966, %rd12949; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9144}, %rd12961; + } + { + .reg .b32 %dummy; + mov.b64 {%r9145,%dummy}, %rd12961; + } + shf.l.wrap.b32 %r9146, %r9145, %r9144, 1; + shf.l.wrap.b32 %r9147, %r9144, %r9145, 1; + mov.b64 %rd12968, {%r9147, %r9146}; + xor.b64 %rd12969, %rd12968, %rd12953; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9148}, %rd12945; + } + { + .reg .b32 %dummy; + mov.b64 {%r9149,%dummy}, %rd12945; + } + shf.l.wrap.b32 %r9150, %r9149, %r9148, 1; + shf.l.wrap.b32 %r9151, %r9148, %r9149, 1; + mov.b64 %rd12970, {%r9151, %r9150}; + xor.b64 %rd12971, %rd12957, %rd12970; + xor.b64 %rd12972, %rd12963, %rd12941; + xor.b64 %rd12973, %rd12965, %rd12871; + xor.b64 %rd12974, %rd12967, %rd12874; + xor.b64 %rd12975, %rd12969, %rd12877; + xor.b64 %rd12976, %rd12971, %rd12880; + xor.b64 %rd12977, %rd12963, %rd12883; + xor.b64 %rd12978, %rd12965, %rd12886; + xor.b64 %rd12979, %rd12967, %rd12889; + xor.b64 %rd12980, %rd12969, %rd12892; + xor.b64 %rd12981, %rd12971, %rd12895; + xor.b64 %rd12982, %rd12963, %rd12898; + xor.b64 %rd12983, %rd12965, %rd12901; + xor.b64 %rd12984, %rd12967, %rd12904; + xor.b64 %rd12985, %rd12969, %rd12907; + xor.b64 %rd12986, %rd12971, %rd12910; + xor.b64 %rd12987, %rd12963, %rd12913; + xor.b64 %rd12988, %rd12965, %rd12916; + xor.b64 %rd12989, %rd12967, %rd12919; + xor.b64 %rd12990, %rd12969, %rd12922; + xor.b64 %rd12991, %rd12971, %rd12925; + xor.b64 %rd12992, %rd12963, %rd12928; + xor.b64 %rd12993, %rd12965, %rd12931; + xor.b64 %rd12994, %rd12967, %rd12934; + xor.b64 %rd12995, %rd12969, %rd12937; + xor.b64 %rd12996, %rd12971, %rd12940; + { + .reg .b32 %dummy; + mov.b64 {%r9152,%dummy}, %rd12978; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9153}, %rd12978; + } + shf.r.wrap.b32 %r9154, %r9153, %r9152, 20; + shf.r.wrap.b32 %r9155, %r9152, %r9153, 20; + mov.b64 %rd12997, {%r9155, %r9154}; + { + .reg .b32 %dummy; + mov.b64 {%r9156,%dummy}, %rd12984; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9157}, %rd12984; + } + shf.r.wrap.b32 %r9158, %r9157, %r9156, 21; + shf.r.wrap.b32 %r9159, %r9156, %r9157, 21; + mov.b64 %rd12998, {%r9159, %r9158}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9160}, %rd12990; + } + { + .reg .b32 %dummy; + mov.b64 {%r9161,%dummy}, %rd12990; + } + shf.l.wrap.b32 %r9162, %r9161, %r9160, 21; + shf.l.wrap.b32 %r9163, %r9160, %r9161, 21; + mov.b64 %rd12999, {%r9163, %r9162}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9164}, %rd12996; + } + { + .reg .b32 %dummy; + mov.b64 {%r9165,%dummy}, %rd12996; + } + shf.l.wrap.b32 %r9166, %r9165, %r9164, 14; + shf.l.wrap.b32 %r9167, %r9164, %r9165, 14; + mov.b64 %rd13000, {%r9167, %r9166}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9168}, %rd12975; + } + { + .reg .b32 %dummy; + mov.b64 {%r9169,%dummy}, %rd12975; + } + shf.l.wrap.b32 %r9170, %r9169, %r9168, 28; + shf.l.wrap.b32 %r9171, %r9168, %r9169, 28; + mov.b64 %rd13001, {%r9171, %r9170}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9172}, %rd12981; + } + { + .reg .b32 %dummy; + mov.b64 {%r9173,%dummy}, %rd12981; + } + shf.l.wrap.b32 %r9174, %r9173, %r9172, 20; + shf.l.wrap.b32 %r9175, %r9172, %r9173, 20; + mov.b64 %rd13002, {%r9175, %r9174}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9176}, %rd12982; + } + { + .reg .b32 %dummy; + mov.b64 {%r9177,%dummy}, %rd12982; + } + shf.l.wrap.b32 %r9178, %r9177, %r9176, 3; + shf.l.wrap.b32 %r9179, %r9176, %r9177, 3; + mov.b64 %rd13003, {%r9179, %r9178}; + { + .reg .b32 %dummy; + mov.b64 {%r9180,%dummy}, %rd12988; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9181}, %rd12988; + } + shf.r.wrap.b32 %r9182, %r9181, %r9180, 19; + shf.r.wrap.b32 %r9183, %r9180, %r9181, 19; + mov.b64 %rd13004, {%r9183, %r9182}; + { + .reg .b32 %dummy; + mov.b64 {%r9184,%dummy}, %rd12994; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9185}, %rd12994; + } + shf.r.wrap.b32 %r9186, %r9185, %r9184, 3; + shf.r.wrap.b32 %r9187, %r9184, %r9185, 3; + mov.b64 %rd13005, {%r9187, %r9186}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9188}, %rd12973; + } + { + .reg .b32 %dummy; + mov.b64 {%r9189,%dummy}, %rd12973; + } + shf.l.wrap.b32 %r9190, %r9189, %r9188, 1; + shf.l.wrap.b32 %r9191, %r9188, %r9189, 1; + mov.b64 %rd13006, {%r9191, %r9190}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9192}, %rd12979; + } + { + .reg .b32 %dummy; + mov.b64 {%r9193,%dummy}, %rd12979; + } + shf.l.wrap.b32 %r9194, %r9193, %r9192, 6; + shf.l.wrap.b32 %r9195, %r9192, %r9193, 6; + mov.b64 %rd13007, {%r9195, %r9194}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9196}, %rd12985; + } + { + .reg .b32 %dummy; + mov.b64 {%r9197,%dummy}, %rd12985; + } + shf.l.wrap.b32 %r9198, %r9197, %r9196, 25; + shf.l.wrap.b32 %r9199, %r9196, %r9197, 25; + mov.b64 %rd13008, {%r9199, %r9198}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9200}, %rd12991; + } + { + .reg .b32 %dummy; + mov.b64 {%r9201,%dummy}, %rd12991; + } + shf.l.wrap.b32 %r9202, %r9201, %r9200, 8; + shf.l.wrap.b32 %r9203, %r9200, %r9201, 8; + mov.b64 %rd13009, {%r9203, %r9202}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9204}, %rd12992; + } + { + .reg .b32 %dummy; + mov.b64 {%r9205,%dummy}, %rd12992; + } + shf.l.wrap.b32 %r9206, %r9205, %r9204, 18; + shf.l.wrap.b32 %r9207, %r9204, %r9205, 18; + mov.b64 %rd13010, {%r9207, %r9206}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9208}, %rd12976; + } + { + .reg .b32 %dummy; + mov.b64 {%r9209,%dummy}, %rd12976; + } + shf.l.wrap.b32 %r9210, %r9209, %r9208, 27; + shf.l.wrap.b32 %r9211, %r9208, %r9209, 27; + mov.b64 %rd13011, {%r9211, %r9210}; + { + .reg .b32 %dummy; + mov.b64 {%r9212,%dummy}, %rd12977; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9213}, %rd12977; + } + shf.r.wrap.b32 %r9214, %r9213, %r9212, 28; + shf.r.wrap.b32 %r9215, %r9212, %r9213, 28; + mov.b64 %rd13012, {%r9215, %r9214}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9216}, %rd12983; + } + { + .reg .b32 %dummy; + mov.b64 {%r9217,%dummy}, %rd12983; + } + shf.l.wrap.b32 %r9218, %r9217, %r9216, 10; + shf.l.wrap.b32 %r9219, %r9216, %r9217, 10; + mov.b64 %rd13013, {%r9219, %r9218}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9220}, %rd12989; + } + { + .reg .b32 %dummy; + mov.b64 {%r9221,%dummy}, %rd12989; + } + shf.l.wrap.b32 %r9222, %r9221, %r9220, 15; + shf.l.wrap.b32 %r9223, %r9220, %r9221, 15; + mov.b64 %rd13014, {%r9223, %r9222}; + { + .reg .b32 %dummy; + mov.b64 {%r9224,%dummy}, %rd12995; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9225}, %rd12995; + } + shf.r.wrap.b32 %r9226, %r9225, %r9224, 8; + shf.r.wrap.b32 %r9227, %r9224, %r9225, 8; + mov.b64 %rd13015, {%r9227, %r9226}; + { + .reg .b32 %dummy; + mov.b64 {%r9228,%dummy}, %rd12974; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9229}, %rd12974; + } + shf.r.wrap.b32 %r9230, %r9229, %r9228, 2; + shf.r.wrap.b32 %r9231, %r9228, %r9229, 2; + mov.b64 %rd13016, {%r9231, %r9230}; + { + .reg .b32 %dummy; + mov.b64 {%r9232,%dummy}, %rd12980; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9233}, %rd12980; + } + shf.r.wrap.b32 %r9234, %r9233, %r9232, 9; + shf.r.wrap.b32 %r9235, %r9232, %r9233, 9; + mov.b64 %rd13017, {%r9235, %r9234}; + { + .reg .b32 %dummy; + mov.b64 {%r9236,%dummy}, %rd12986; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9237}, %rd12986; + } + shf.r.wrap.b32 %r9238, %r9237, %r9236, 25; + shf.r.wrap.b32 %r9239, %r9236, %r9237, 25; + mov.b64 %rd13018, {%r9239, %r9238}; + { + .reg .b32 %dummy; + mov.b64 {%r9240,%dummy}, %rd12987; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9241}, %rd12987; + } + shf.r.wrap.b32 %r9242, %r9241, %r9240, 23; + shf.r.wrap.b32 %r9243, %r9240, %r9241, 23; + mov.b64 %rd13019, {%r9243, %r9242}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9244}, %rd12993; + } + { + .reg .b32 %dummy; + mov.b64 {%r9245,%dummy}, %rd12993; + } + shf.l.wrap.b32 %r9246, %r9245, %r9244, 2; + shf.l.wrap.b32 %r9247, %r9244, %r9245, 2; + mov.b64 %rd13020, {%r9247, %r9246}; + not.b64 %rd13021, %rd12997; + and.b64 %rd13022, %rd12998, %rd13021; + xor.b64 %rd13023, %rd13022, %rd12972; + not.b64 %rd13024, %rd12998; + and.b64 %rd13025, %rd12999, %rd13024; + xor.b64 %rd13026, %rd13025, %rd12997; + not.b64 %rd13027, %rd12999; + and.b64 %rd13028, %rd13000, %rd13027; + xor.b64 %rd13029, %rd13028, %rd12998; + not.b64 %rd13030, %rd13000; + and.b64 %rd13031, %rd12972, %rd13030; + xor.b64 %rd13032, %rd13031, %rd12999; + not.b64 %rd13033, %rd12972; + and.b64 %rd13034, %rd12997, %rd13033; + xor.b64 %rd13035, %rd13034, %rd13000; + not.b64 %rd13036, %rd13002; + and.b64 %rd13037, %rd13003, %rd13036; + xor.b64 %rd13038, %rd13037, %rd13001; + not.b64 %rd13039, %rd13003; + and.b64 %rd13040, %rd13004, %rd13039; + xor.b64 %rd13041, %rd13040, %rd13002; + not.b64 %rd13042, %rd13004; + and.b64 %rd13043, %rd13005, %rd13042; + xor.b64 %rd13044, %rd13043, %rd13003; + not.b64 %rd13045, %rd13005; + and.b64 %rd13046, %rd13001, %rd13045; + xor.b64 %rd13047, %rd13046, %rd13004; + not.b64 %rd13048, %rd13001; + and.b64 %rd13049, %rd13002, %rd13048; + xor.b64 %rd13050, %rd13049, %rd13005; + not.b64 %rd13051, %rd13007; + and.b64 %rd13052, %rd13008, %rd13051; + xor.b64 %rd13053, %rd13052, %rd13006; + not.b64 %rd13054, %rd13008; + and.b64 %rd13055, %rd13009, %rd13054; + xor.b64 %rd13056, %rd13055, %rd13007; + not.b64 %rd13057, %rd13009; + and.b64 %rd13058, %rd13010, %rd13057; + xor.b64 %rd13059, %rd13058, %rd13008; + not.b64 %rd13060, %rd13010; + and.b64 %rd13061, %rd13006, %rd13060; + xor.b64 %rd13062, %rd13061, %rd13009; + not.b64 %rd13063, %rd13006; + and.b64 %rd13064, %rd13007, %rd13063; + xor.b64 %rd13065, %rd13064, %rd13010; + not.b64 %rd13066, %rd13012; + and.b64 %rd13067, %rd13013, %rd13066; + xor.b64 %rd13068, %rd13067, %rd13011; + not.b64 %rd13069, %rd13013; + and.b64 %rd13070, %rd13014, %rd13069; + xor.b64 %rd13071, %rd13070, %rd13012; + not.b64 %rd13072, %rd13014; + and.b64 %rd13073, %rd13015, %rd13072; + xor.b64 %rd13074, %rd13073, %rd13013; + not.b64 %rd13075, %rd13015; + and.b64 %rd13076, %rd13011, %rd13075; + xor.b64 %rd13077, %rd13076, %rd13014; + not.b64 %rd13078, %rd13011; + and.b64 %rd13079, %rd13012, %rd13078; + xor.b64 %rd13080, %rd13079, %rd13015; + not.b64 %rd13081, %rd13017; + and.b64 %rd13082, %rd13018, %rd13081; + xor.b64 %rd13083, %rd13082, %rd13016; + not.b64 %rd13084, %rd13018; + and.b64 %rd13085, %rd13019, %rd13084; + xor.b64 %rd13086, %rd13085, %rd13017; + not.b64 %rd13087, %rd13019; + and.b64 %rd13088, %rd13020, %rd13087; + xor.b64 %rd13089, %rd13088, %rd13018; + not.b64 %rd13090, %rd13020; + and.b64 %rd13091, %rd13016, %rd13090; + xor.b64 %rd13092, %rd13091, %rd13019; + not.b64 %rd13093, %rd13016; + and.b64 %rd13094, %rd13017, %rd13093; + xor.b64 %rd13095, %rd13094, %rd13020; + xor.b64 %rd13096, %rd13023, %rd22; + xor.b64 %rd13097, %rd13038, %rd13096; + xor.b64 %rd13098, %rd13097, %rd13053; + xor.b64 %rd13099, %rd13098, %rd13068; + xor.b64 %rd13100, %rd13099, %rd13083; + xor.b64 %rd13101, %rd13041, %rd13026; + xor.b64 %rd13102, %rd13101, %rd13056; + xor.b64 %rd13103, %rd13102, %rd13071; + xor.b64 %rd13104, %rd13103, %rd13086; + xor.b64 %rd13105, %rd13044, %rd13029; + xor.b64 %rd13106, %rd13105, %rd13059; + xor.b64 %rd13107, %rd13106, %rd13074; + xor.b64 %rd13108, %rd13107, %rd13089; + xor.b64 %rd13109, %rd13047, %rd13032; + xor.b64 %rd13110, %rd13109, %rd13062; + xor.b64 %rd13111, %rd13110, %rd13077; + xor.b64 %rd13112, %rd13111, %rd13092; + xor.b64 %rd13113, %rd13050, %rd13035; + xor.b64 %rd13114, %rd13113, %rd13065; + xor.b64 %rd13115, %rd13114, %rd13080; + xor.b64 %rd13116, %rd13115, %rd13095; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9248}, %rd13104; + } + { + .reg .b32 %dummy; + mov.b64 {%r9249,%dummy}, %rd13104; + } + shf.l.wrap.b32 %r9250, %r9249, %r9248, 1; + shf.l.wrap.b32 %r9251, %r9248, %r9249, 1; + mov.b64 %rd13117, {%r9251, %r9250}; + xor.b64 %rd13118, %rd13116, %rd13117; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9252}, %rd13108; + } + { + .reg .b32 %dummy; + mov.b64 {%r9253,%dummy}, %rd13108; + } + shf.l.wrap.b32 %r9254, %r9253, %r9252, 1; + shf.l.wrap.b32 %r9255, %r9252, %r9253, 1; + mov.b64 %rd13119, {%r9255, %r9254}; + xor.b64 %rd13120, %rd13119, %rd13100; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9256}, %rd13112; + } + { + .reg .b32 %dummy; + mov.b64 {%r9257,%dummy}, %rd13112; + } + shf.l.wrap.b32 %r9258, %r9257, %r9256, 1; + shf.l.wrap.b32 %r9259, %r9256, %r9257, 1; + mov.b64 %rd13121, {%r9259, %r9258}; + xor.b64 %rd13122, %rd13121, %rd13104; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9260}, %rd13116; + } + { + .reg .b32 %dummy; + mov.b64 {%r9261,%dummy}, %rd13116; + } + shf.l.wrap.b32 %r9262, %r9261, %r9260, 1; + shf.l.wrap.b32 %r9263, %r9260, %r9261, 1; + mov.b64 %rd13123, {%r9263, %r9262}; + xor.b64 %rd13124, %rd13123, %rd13108; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9264}, %rd13100; + } + { + .reg .b32 %dummy; + mov.b64 {%r9265,%dummy}, %rd13100; + } + shf.l.wrap.b32 %r9266, %r9265, %r9264, 1; + shf.l.wrap.b32 %r9267, %r9264, %r9265, 1; + mov.b64 %rd13125, {%r9267, %r9266}; + xor.b64 %rd13126, %rd13112, %rd13125; + xor.b64 %rd13127, %rd13118, %rd13096; + xor.b64 %rd13128, %rd13120, %rd13026; + xor.b64 %rd13129, %rd13122, %rd13029; + xor.b64 %rd13130, %rd13124, %rd13032; + xor.b64 %rd13131, %rd13126, %rd13035; + xor.b64 %rd13132, %rd13118, %rd13038; + xor.b64 %rd13133, %rd13120, %rd13041; + xor.b64 %rd13134, %rd13122, %rd13044; + xor.b64 %rd13135, %rd13124, %rd13047; + xor.b64 %rd13136, %rd13126, %rd13050; + xor.b64 %rd13137, %rd13118, %rd13053; + xor.b64 %rd13138, %rd13120, %rd13056; + xor.b64 %rd13139, %rd13122, %rd13059; + xor.b64 %rd13140, %rd13124, %rd13062; + xor.b64 %rd13141, %rd13126, %rd13065; + xor.b64 %rd13142, %rd13118, %rd13068; + xor.b64 %rd13143, %rd13120, %rd13071; + xor.b64 %rd13144, %rd13122, %rd13074; + xor.b64 %rd13145, %rd13124, %rd13077; + xor.b64 %rd13146, %rd13126, %rd13080; + xor.b64 %rd13147, %rd13118, %rd13083; + xor.b64 %rd13148, %rd13120, %rd13086; + xor.b64 %rd13149, %rd13122, %rd13089; + xor.b64 %rd13150, %rd13124, %rd13092; + xor.b64 %rd13151, %rd13126, %rd13095; + { + .reg .b32 %dummy; + mov.b64 {%r9268,%dummy}, %rd13133; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9269}, %rd13133; + } + shf.r.wrap.b32 %r9270, %r9269, %r9268, 20; + shf.r.wrap.b32 %r9271, %r9268, %r9269, 20; + mov.b64 %rd13152, {%r9271, %r9270}; + { + .reg .b32 %dummy; + mov.b64 {%r9272,%dummy}, %rd13139; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9273}, %rd13139; + } + shf.r.wrap.b32 %r9274, %r9273, %r9272, 21; + shf.r.wrap.b32 %r9275, %r9272, %r9273, 21; + mov.b64 %rd13153, {%r9275, %r9274}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9276}, %rd13145; + } + { + .reg .b32 %dummy; + mov.b64 {%r9277,%dummy}, %rd13145; + } + shf.l.wrap.b32 %r9278, %r9277, %r9276, 21; + shf.l.wrap.b32 %r9279, %r9276, %r9277, 21; + mov.b64 %rd13154, {%r9279, %r9278}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9280}, %rd13151; + } + { + .reg .b32 %dummy; + mov.b64 {%r9281,%dummy}, %rd13151; + } + shf.l.wrap.b32 %r9282, %r9281, %r9280, 14; + shf.l.wrap.b32 %r9283, %r9280, %r9281, 14; + mov.b64 %rd13155, {%r9283, %r9282}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9284}, %rd13130; + } + { + .reg .b32 %dummy; + mov.b64 {%r9285,%dummy}, %rd13130; + } + shf.l.wrap.b32 %r9286, %r9285, %r9284, 28; + shf.l.wrap.b32 %r9287, %r9284, %r9285, 28; + mov.b64 %rd13156, {%r9287, %r9286}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9288}, %rd13136; + } + { + .reg .b32 %dummy; + mov.b64 {%r9289,%dummy}, %rd13136; + } + shf.l.wrap.b32 %r9290, %r9289, %r9288, 20; + shf.l.wrap.b32 %r9291, %r9288, %r9289, 20; + mov.b64 %rd13157, {%r9291, %r9290}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9292}, %rd13137; + } + { + .reg .b32 %dummy; + mov.b64 {%r9293,%dummy}, %rd13137; + } + shf.l.wrap.b32 %r9294, %r9293, %r9292, 3; + shf.l.wrap.b32 %r9295, %r9292, %r9293, 3; + mov.b64 %rd13158, {%r9295, %r9294}; + { + .reg .b32 %dummy; + mov.b64 {%r9296,%dummy}, %rd13143; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9297}, %rd13143; + } + shf.r.wrap.b32 %r9298, %r9297, %r9296, 19; + shf.r.wrap.b32 %r9299, %r9296, %r9297, 19; + mov.b64 %rd13159, {%r9299, %r9298}; + { + .reg .b32 %dummy; + mov.b64 {%r9300,%dummy}, %rd13149; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9301}, %rd13149; + } + shf.r.wrap.b32 %r9302, %r9301, %r9300, 3; + shf.r.wrap.b32 %r9303, %r9300, %r9301, 3; + mov.b64 %rd13160, {%r9303, %r9302}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9304}, %rd13128; + } + { + .reg .b32 %dummy; + mov.b64 {%r9305,%dummy}, %rd13128; + } + shf.l.wrap.b32 %r9306, %r9305, %r9304, 1; + shf.l.wrap.b32 %r9307, %r9304, %r9305, 1; + mov.b64 %rd13161, {%r9307, %r9306}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9308}, %rd13134; + } + { + .reg .b32 %dummy; + mov.b64 {%r9309,%dummy}, %rd13134; + } + shf.l.wrap.b32 %r9310, %r9309, %r9308, 6; + shf.l.wrap.b32 %r9311, %r9308, %r9309, 6; + mov.b64 %rd13162, {%r9311, %r9310}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9312}, %rd13140; + } + { + .reg .b32 %dummy; + mov.b64 {%r9313,%dummy}, %rd13140; + } + shf.l.wrap.b32 %r9314, %r9313, %r9312, 25; + shf.l.wrap.b32 %r9315, %r9312, %r9313, 25; + mov.b64 %rd13163, {%r9315, %r9314}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9316}, %rd13146; + } + { + .reg .b32 %dummy; + mov.b64 {%r9317,%dummy}, %rd13146; + } + shf.l.wrap.b32 %r9318, %r9317, %r9316, 8; + shf.l.wrap.b32 %r9319, %r9316, %r9317, 8; + mov.b64 %rd13164, {%r9319, %r9318}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9320}, %rd13147; + } + { + .reg .b32 %dummy; + mov.b64 {%r9321,%dummy}, %rd13147; + } + shf.l.wrap.b32 %r9322, %r9321, %r9320, 18; + shf.l.wrap.b32 %r9323, %r9320, %r9321, 18; + mov.b64 %rd13165, {%r9323, %r9322}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9324}, %rd13131; + } + { + .reg .b32 %dummy; + mov.b64 {%r9325,%dummy}, %rd13131; + } + shf.l.wrap.b32 %r9326, %r9325, %r9324, 27; + shf.l.wrap.b32 %r9327, %r9324, %r9325, 27; + mov.b64 %rd13166, {%r9327, %r9326}; + { + .reg .b32 %dummy; + mov.b64 {%r9328,%dummy}, %rd13132; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9329}, %rd13132; + } + shf.r.wrap.b32 %r9330, %r9329, %r9328, 28; + shf.r.wrap.b32 %r9331, %r9328, %r9329, 28; + mov.b64 %rd13167, {%r9331, %r9330}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9332}, %rd13138; + } + { + .reg .b32 %dummy; + mov.b64 {%r9333,%dummy}, %rd13138; + } + shf.l.wrap.b32 %r9334, %r9333, %r9332, 10; + shf.l.wrap.b32 %r9335, %r9332, %r9333, 10; + mov.b64 %rd13168, {%r9335, %r9334}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9336}, %rd13144; + } + { + .reg .b32 %dummy; + mov.b64 {%r9337,%dummy}, %rd13144; + } + shf.l.wrap.b32 %r9338, %r9337, %r9336, 15; + shf.l.wrap.b32 %r9339, %r9336, %r9337, 15; + mov.b64 %rd13169, {%r9339, %r9338}; + { + .reg .b32 %dummy; + mov.b64 {%r9340,%dummy}, %rd13150; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9341}, %rd13150; + } + shf.r.wrap.b32 %r9342, %r9341, %r9340, 8; + shf.r.wrap.b32 %r9343, %r9340, %r9341, 8; + mov.b64 %rd13170, {%r9343, %r9342}; + { + .reg .b32 %dummy; + mov.b64 {%r9344,%dummy}, %rd13129; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9345}, %rd13129; + } + shf.r.wrap.b32 %r9346, %r9345, %r9344, 2; + shf.r.wrap.b32 %r9347, %r9344, %r9345, 2; + mov.b64 %rd13171, {%r9347, %r9346}; + { + .reg .b32 %dummy; + mov.b64 {%r9348,%dummy}, %rd13135; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9349}, %rd13135; + } + shf.r.wrap.b32 %r9350, %r9349, %r9348, 9; + shf.r.wrap.b32 %r9351, %r9348, %r9349, 9; + mov.b64 %rd13172, {%r9351, %r9350}; + { + .reg .b32 %dummy; + mov.b64 {%r9352,%dummy}, %rd13141; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9353}, %rd13141; + } + shf.r.wrap.b32 %r9354, %r9353, %r9352, 25; + shf.r.wrap.b32 %r9355, %r9352, %r9353, 25; + mov.b64 %rd13173, {%r9355, %r9354}; + { + .reg .b32 %dummy; + mov.b64 {%r9356,%dummy}, %rd13142; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9357}, %rd13142; + } + shf.r.wrap.b32 %r9358, %r9357, %r9356, 23; + shf.r.wrap.b32 %r9359, %r9356, %r9357, 23; + mov.b64 %rd13174, {%r9359, %r9358}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9360}, %rd13148; + } + { + .reg .b32 %dummy; + mov.b64 {%r9361,%dummy}, %rd13148; + } + shf.l.wrap.b32 %r9362, %r9361, %r9360, 2; + shf.l.wrap.b32 %r9363, %r9360, %r9361, 2; + mov.b64 %rd13175, {%r9363, %r9362}; + not.b64 %rd13176, %rd13152; + and.b64 %rd13177, %rd13153, %rd13176; + xor.b64 %rd13178, %rd13177, %rd13127; + not.b64 %rd13179, %rd13153; + and.b64 %rd13180, %rd13154, %rd13179; + xor.b64 %rd13181, %rd13180, %rd13152; + not.b64 %rd13182, %rd13154; + and.b64 %rd13183, %rd13155, %rd13182; + xor.b64 %rd13184, %rd13183, %rd13153; + not.b64 %rd13185, %rd13155; + and.b64 %rd13186, %rd13127, %rd13185; + xor.b64 %rd13187, %rd13186, %rd13154; + not.b64 %rd13188, %rd13127; + and.b64 %rd13189, %rd13152, %rd13188; + xor.b64 %rd13190, %rd13189, %rd13155; + not.b64 %rd13191, %rd13157; + and.b64 %rd13192, %rd13158, %rd13191; + xor.b64 %rd13193, %rd13192, %rd13156; + not.b64 %rd13194, %rd13158; + and.b64 %rd13195, %rd13159, %rd13194; + xor.b64 %rd13196, %rd13195, %rd13157; + not.b64 %rd13197, %rd13159; + and.b64 %rd13198, %rd13160, %rd13197; + xor.b64 %rd13199, %rd13198, %rd13158; + not.b64 %rd13200, %rd13160; + and.b64 %rd13201, %rd13156, %rd13200; + xor.b64 %rd13202, %rd13201, %rd13159; + not.b64 %rd13203, %rd13156; + and.b64 %rd13204, %rd13157, %rd13203; + xor.b64 %rd13205, %rd13204, %rd13160; + not.b64 %rd13206, %rd13162; + and.b64 %rd13207, %rd13163, %rd13206; + xor.b64 %rd13208, %rd13207, %rd13161; + not.b64 %rd13209, %rd13163; + and.b64 %rd13210, %rd13164, %rd13209; + xor.b64 %rd13211, %rd13210, %rd13162; + not.b64 %rd13212, %rd13164; + and.b64 %rd13213, %rd13165, %rd13212; + xor.b64 %rd13214, %rd13213, %rd13163; + not.b64 %rd13215, %rd13165; + and.b64 %rd13216, %rd13161, %rd13215; + xor.b64 %rd13217, %rd13216, %rd13164; + not.b64 %rd13218, %rd13161; + and.b64 %rd13219, %rd13162, %rd13218; + xor.b64 %rd13220, %rd13219, %rd13165; + not.b64 %rd13221, %rd13167; + and.b64 %rd13222, %rd13168, %rd13221; + xor.b64 %rd13223, %rd13222, %rd13166; + not.b64 %rd13224, %rd13168; + and.b64 %rd13225, %rd13169, %rd13224; + xor.b64 %rd13226, %rd13225, %rd13167; + not.b64 %rd13227, %rd13169; + and.b64 %rd13228, %rd13170, %rd13227; + xor.b64 %rd13229, %rd13228, %rd13168; + not.b64 %rd13230, %rd13170; + and.b64 %rd13231, %rd13166, %rd13230; + xor.b64 %rd13232, %rd13231, %rd13169; + not.b64 %rd13233, %rd13166; + and.b64 %rd13234, %rd13167, %rd13233; + xor.b64 %rd13235, %rd13234, %rd13170; + not.b64 %rd13236, %rd13172; + and.b64 %rd13237, %rd13173, %rd13236; + xor.b64 %rd13238, %rd13237, %rd13171; + not.b64 %rd13239, %rd13173; + and.b64 %rd13240, %rd13174, %rd13239; + xor.b64 %rd13241, %rd13240, %rd13172; + not.b64 %rd13242, %rd13174; + and.b64 %rd13243, %rd13175, %rd13242; + xor.b64 %rd13244, %rd13243, %rd13173; + not.b64 %rd13245, %rd13175; + and.b64 %rd13246, %rd13171, %rd13245; + xor.b64 %rd13247, %rd13246, %rd13174; + not.b64 %rd13248, %rd13171; + and.b64 %rd13249, %rd13172, %rd13248; + xor.b64 %rd13250, %rd13249, %rd13175; + xor.b64 %rd13251, %rd13178, %rd23; + xor.b64 %rd13252, %rd13193, %rd13251; + xor.b64 %rd13253, %rd13252, %rd13208; + xor.b64 %rd13254, %rd13253, %rd13223; + xor.b64 %rd13255, %rd13254, %rd13238; + xor.b64 %rd13256, %rd13196, %rd13181; + xor.b64 %rd13257, %rd13256, %rd13211; + xor.b64 %rd13258, %rd13257, %rd13226; + xor.b64 %rd13259, %rd13258, %rd13241; + xor.b64 %rd13260, %rd13199, %rd13184; + xor.b64 %rd13261, %rd13260, %rd13214; + xor.b64 %rd13262, %rd13261, %rd13229; + xor.b64 %rd13263, %rd13262, %rd13244; + xor.b64 %rd13264, %rd13202, %rd13187; + xor.b64 %rd13265, %rd13264, %rd13217; + xor.b64 %rd13266, %rd13265, %rd13232; + xor.b64 %rd13267, %rd13266, %rd13247; + xor.b64 %rd13268, %rd13205, %rd13190; + xor.b64 %rd13269, %rd13268, %rd13220; + xor.b64 %rd13270, %rd13269, %rd13235; + xor.b64 %rd13271, %rd13270, %rd13250; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9364}, %rd13259; + } + { + .reg .b32 %dummy; + mov.b64 {%r9365,%dummy}, %rd13259; + } + shf.l.wrap.b32 %r9366, %r9365, %r9364, 1; + shf.l.wrap.b32 %r9367, %r9364, %r9365, 1; + mov.b64 %rd13272, {%r9367, %r9366}; + xor.b64 %rd13273, %rd13271, %rd13272; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9368}, %rd13263; + } + { + .reg .b32 %dummy; + mov.b64 {%r9369,%dummy}, %rd13263; + } + shf.l.wrap.b32 %r9370, %r9369, %r9368, 1; + shf.l.wrap.b32 %r9371, %r9368, %r9369, 1; + mov.b64 %rd13274, {%r9371, %r9370}; + xor.b64 %rd13275, %rd13274, %rd13255; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9372}, %rd13267; + } + { + .reg .b32 %dummy; + mov.b64 {%r9373,%dummy}, %rd13267; + } + shf.l.wrap.b32 %r9374, %r9373, %r9372, 1; + shf.l.wrap.b32 %r9375, %r9372, %r9373, 1; + mov.b64 %rd13276, {%r9375, %r9374}; + xor.b64 %rd13277, %rd13276, %rd13259; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9376}, %rd13271; + } + { + .reg .b32 %dummy; + mov.b64 {%r9377,%dummy}, %rd13271; + } + shf.l.wrap.b32 %r9378, %r9377, %r9376, 1; + shf.l.wrap.b32 %r9379, %r9376, %r9377, 1; + mov.b64 %rd13278, {%r9379, %r9378}; + xor.b64 %rd13279, %rd13278, %rd13263; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9380}, %rd13255; + } + { + .reg .b32 %dummy; + mov.b64 {%r9381,%dummy}, %rd13255; + } + shf.l.wrap.b32 %r9382, %r9381, %r9380, 1; + shf.l.wrap.b32 %r9383, %r9380, %r9381, 1; + mov.b64 %rd13280, {%r9383, %r9382}; + xor.b64 %rd13281, %rd13267, %rd13280; + xor.b64 %rd13282, %rd13273, %rd13251; + xor.b64 %rd13283, %rd13275, %rd13181; + xor.b64 %rd13284, %rd13277, %rd13184; + xor.b64 %rd13285, %rd13279, %rd13187; + xor.b64 %rd13286, %rd13281, %rd13190; + xor.b64 %rd13287, %rd13273, %rd13193; + xor.b64 %rd13288, %rd13275, %rd13196; + xor.b64 %rd13289, %rd13277, %rd13199; + xor.b64 %rd13290, %rd13279, %rd13202; + xor.b64 %rd13291, %rd13281, %rd13205; + xor.b64 %rd13292, %rd13273, %rd13208; + xor.b64 %rd13293, %rd13275, %rd13211; + xor.b64 %rd13294, %rd13277, %rd13214; + xor.b64 %rd13295, %rd13279, %rd13217; + xor.b64 %rd13296, %rd13281, %rd13220; + xor.b64 %rd13297, %rd13273, %rd13223; + xor.b64 %rd13298, %rd13275, %rd13226; + xor.b64 %rd13299, %rd13277, %rd13229; + xor.b64 %rd13300, %rd13279, %rd13232; + xor.b64 %rd13301, %rd13281, %rd13235; + xor.b64 %rd13302, %rd13273, %rd13238; + xor.b64 %rd13303, %rd13275, %rd13241; + xor.b64 %rd13304, %rd13277, %rd13244; + xor.b64 %rd13305, %rd13279, %rd13247; + xor.b64 %rd13306, %rd13281, %rd13250; + { + .reg .b32 %dummy; + mov.b64 {%r9384,%dummy}, %rd13288; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9385}, %rd13288; + } + shf.r.wrap.b32 %r9386, %r9385, %r9384, 20; + shf.r.wrap.b32 %r9387, %r9384, %r9385, 20; + mov.b64 %rd13307, {%r9387, %r9386}; + { + .reg .b32 %dummy; + mov.b64 {%r9388,%dummy}, %rd13294; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9389}, %rd13294; + } + shf.r.wrap.b32 %r9390, %r9389, %r9388, 21; + shf.r.wrap.b32 %r9391, %r9388, %r9389, 21; + mov.b64 %rd13308, {%r9391, %r9390}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9392}, %rd13300; + } + { + .reg .b32 %dummy; + mov.b64 {%r9393,%dummy}, %rd13300; + } + shf.l.wrap.b32 %r9394, %r9393, %r9392, 21; + shf.l.wrap.b32 %r9395, %r9392, %r9393, 21; + mov.b64 %rd13309, {%r9395, %r9394}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9396}, %rd13306; + } + { + .reg .b32 %dummy; + mov.b64 {%r9397,%dummy}, %rd13306; + } + shf.l.wrap.b32 %r9398, %r9397, %r9396, 14; + shf.l.wrap.b32 %r9399, %r9396, %r9397, 14; + mov.b64 %rd13310, {%r9399, %r9398}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9400}, %rd13285; + } + { + .reg .b32 %dummy; + mov.b64 {%r9401,%dummy}, %rd13285; + } + shf.l.wrap.b32 %r9402, %r9401, %r9400, 28; + shf.l.wrap.b32 %r9403, %r9400, %r9401, 28; + mov.b64 %rd13311, {%r9403, %r9402}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9404}, %rd13291; + } + { + .reg .b32 %dummy; + mov.b64 {%r9405,%dummy}, %rd13291; + } + shf.l.wrap.b32 %r9406, %r9405, %r9404, 20; + shf.l.wrap.b32 %r9407, %r9404, %r9405, 20; + mov.b64 %rd13312, {%r9407, %r9406}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9408}, %rd13292; + } + { + .reg .b32 %dummy; + mov.b64 {%r9409,%dummy}, %rd13292; + } + shf.l.wrap.b32 %r9410, %r9409, %r9408, 3; + shf.l.wrap.b32 %r9411, %r9408, %r9409, 3; + mov.b64 %rd13313, {%r9411, %r9410}; + { + .reg .b32 %dummy; + mov.b64 {%r9412,%dummy}, %rd13298; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9413}, %rd13298; + } + shf.r.wrap.b32 %r9414, %r9413, %r9412, 19; + shf.r.wrap.b32 %r9415, %r9412, %r9413, 19; + mov.b64 %rd13314, {%r9415, %r9414}; + { + .reg .b32 %dummy; + mov.b64 {%r9416,%dummy}, %rd13304; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9417}, %rd13304; + } + shf.r.wrap.b32 %r9418, %r9417, %r9416, 3; + shf.r.wrap.b32 %r9419, %r9416, %r9417, 3; + mov.b64 %rd13315, {%r9419, %r9418}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9420}, %rd13283; + } + { + .reg .b32 %dummy; + mov.b64 {%r9421,%dummy}, %rd13283; + } + shf.l.wrap.b32 %r9422, %r9421, %r9420, 1; + shf.l.wrap.b32 %r9423, %r9420, %r9421, 1; + mov.b64 %rd13316, {%r9423, %r9422}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9424}, %rd13289; + } + { + .reg .b32 %dummy; + mov.b64 {%r9425,%dummy}, %rd13289; + } + shf.l.wrap.b32 %r9426, %r9425, %r9424, 6; + shf.l.wrap.b32 %r9427, %r9424, %r9425, 6; + mov.b64 %rd13317, {%r9427, %r9426}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9428}, %rd13295; + } + { + .reg .b32 %dummy; + mov.b64 {%r9429,%dummy}, %rd13295; + } + shf.l.wrap.b32 %r9430, %r9429, %r9428, 25; + shf.l.wrap.b32 %r9431, %r9428, %r9429, 25; + mov.b64 %rd13318, {%r9431, %r9430}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9432}, %rd13301; + } + { + .reg .b32 %dummy; + mov.b64 {%r9433,%dummy}, %rd13301; + } + shf.l.wrap.b32 %r9434, %r9433, %r9432, 8; + shf.l.wrap.b32 %r9435, %r9432, %r9433, 8; + mov.b64 %rd13319, {%r9435, %r9434}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9436}, %rd13302; + } + { + .reg .b32 %dummy; + mov.b64 {%r9437,%dummy}, %rd13302; + } + shf.l.wrap.b32 %r9438, %r9437, %r9436, 18; + shf.l.wrap.b32 %r9439, %r9436, %r9437, 18; + mov.b64 %rd13320, {%r9439, %r9438}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9440}, %rd13286; + } + { + .reg .b32 %dummy; + mov.b64 {%r9441,%dummy}, %rd13286; + } + shf.l.wrap.b32 %r9442, %r9441, %r9440, 27; + shf.l.wrap.b32 %r9443, %r9440, %r9441, 27; + mov.b64 %rd13321, {%r9443, %r9442}; + { + .reg .b32 %dummy; + mov.b64 {%r9444,%dummy}, %rd13287; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9445}, %rd13287; + } + shf.r.wrap.b32 %r9446, %r9445, %r9444, 28; + shf.r.wrap.b32 %r9447, %r9444, %r9445, 28; + mov.b64 %rd13322, {%r9447, %r9446}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9448}, %rd13293; + } + { + .reg .b32 %dummy; + mov.b64 {%r9449,%dummy}, %rd13293; + } + shf.l.wrap.b32 %r9450, %r9449, %r9448, 10; + shf.l.wrap.b32 %r9451, %r9448, %r9449, 10; + mov.b64 %rd13323, {%r9451, %r9450}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9452}, %rd13299; + } + { + .reg .b32 %dummy; + mov.b64 {%r9453,%dummy}, %rd13299; + } + shf.l.wrap.b32 %r9454, %r9453, %r9452, 15; + shf.l.wrap.b32 %r9455, %r9452, %r9453, 15; + mov.b64 %rd13324, {%r9455, %r9454}; + { + .reg .b32 %dummy; + mov.b64 {%r9456,%dummy}, %rd13305; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9457}, %rd13305; + } + shf.r.wrap.b32 %r9458, %r9457, %r9456, 8; + shf.r.wrap.b32 %r9459, %r9456, %r9457, 8; + mov.b64 %rd13325, {%r9459, %r9458}; + { + .reg .b32 %dummy; + mov.b64 {%r9460,%dummy}, %rd13284; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9461}, %rd13284; + } + shf.r.wrap.b32 %r9462, %r9461, %r9460, 2; + shf.r.wrap.b32 %r9463, %r9460, %r9461, 2; + mov.b64 %rd13326, {%r9463, %r9462}; + { + .reg .b32 %dummy; + mov.b64 {%r9464,%dummy}, %rd13290; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9465}, %rd13290; + } + shf.r.wrap.b32 %r9466, %r9465, %r9464, 9; + shf.r.wrap.b32 %r9467, %r9464, %r9465, 9; + mov.b64 %rd13327, {%r9467, %r9466}; + { + .reg .b32 %dummy; + mov.b64 {%r9468,%dummy}, %rd13296; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9469}, %rd13296; + } + shf.r.wrap.b32 %r9470, %r9469, %r9468, 25; + shf.r.wrap.b32 %r9471, %r9468, %r9469, 25; + mov.b64 %rd13328, {%r9471, %r9470}; + { + .reg .b32 %dummy; + mov.b64 {%r9472,%dummy}, %rd13297; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9473}, %rd13297; + } + shf.r.wrap.b32 %r9474, %r9473, %r9472, 23; + shf.r.wrap.b32 %r9475, %r9472, %r9473, 23; + mov.b64 %rd13329, {%r9475, %r9474}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9476}, %rd13303; + } + { + .reg .b32 %dummy; + mov.b64 {%r9477,%dummy}, %rd13303; + } + shf.l.wrap.b32 %r9478, %r9477, %r9476, 2; + shf.l.wrap.b32 %r9479, %r9476, %r9477, 2; + mov.b64 %rd13330, {%r9479, %r9478}; + not.b64 %rd13331, %rd13307; + and.b64 %rd13332, %rd13308, %rd13331; + xor.b64 %rd13333, %rd13332, %rd13282; + not.b64 %rd13334, %rd13308; + and.b64 %rd13335, %rd13309, %rd13334; + xor.b64 %rd13336, %rd13335, %rd13307; + not.b64 %rd13337, %rd13309; + and.b64 %rd13338, %rd13310, %rd13337; + xor.b64 %rd13339, %rd13338, %rd13308; + not.b64 %rd13340, %rd13310; + and.b64 %rd13341, %rd13282, %rd13340; + xor.b64 %rd13342, %rd13341, %rd13309; + not.b64 %rd13343, %rd13282; + and.b64 %rd13344, %rd13307, %rd13343; + xor.b64 %rd13345, %rd13344, %rd13310; + not.b64 %rd13346, %rd13312; + and.b64 %rd13347, %rd13313, %rd13346; + xor.b64 %rd13348, %rd13347, %rd13311; + not.b64 %rd13349, %rd13313; + and.b64 %rd13350, %rd13314, %rd13349; + xor.b64 %rd13351, %rd13350, %rd13312; + not.b64 %rd13352, %rd13314; + and.b64 %rd13353, %rd13315, %rd13352; + xor.b64 %rd13354, %rd13353, %rd13313; + not.b64 %rd13355, %rd13315; + and.b64 %rd13356, %rd13311, %rd13355; + xor.b64 %rd13357, %rd13356, %rd13314; + not.b64 %rd13358, %rd13311; + and.b64 %rd13359, %rd13312, %rd13358; + xor.b64 %rd13360, %rd13359, %rd13315; + not.b64 %rd13361, %rd13317; + and.b64 %rd13362, %rd13318, %rd13361; + xor.b64 %rd13363, %rd13362, %rd13316; + not.b64 %rd13364, %rd13318; + and.b64 %rd13365, %rd13319, %rd13364; + xor.b64 %rd13366, %rd13365, %rd13317; + not.b64 %rd13367, %rd13319; + and.b64 %rd13368, %rd13320, %rd13367; + xor.b64 %rd13369, %rd13368, %rd13318; + not.b64 %rd13370, %rd13320; + and.b64 %rd13371, %rd13316, %rd13370; + xor.b64 %rd13372, %rd13371, %rd13319; + not.b64 %rd13373, %rd13316; + and.b64 %rd13374, %rd13317, %rd13373; + xor.b64 %rd13375, %rd13374, %rd13320; + not.b64 %rd13376, %rd13322; + and.b64 %rd13377, %rd13323, %rd13376; + xor.b64 %rd13378, %rd13377, %rd13321; + not.b64 %rd13379, %rd13323; + and.b64 %rd13380, %rd13324, %rd13379; + xor.b64 %rd13381, %rd13380, %rd13322; + not.b64 %rd13382, %rd13324; + and.b64 %rd13383, %rd13325, %rd13382; + xor.b64 %rd13384, %rd13383, %rd13323; + not.b64 %rd13385, %rd13325; + and.b64 %rd13386, %rd13321, %rd13385; + xor.b64 %rd13387, %rd13386, %rd13324; + not.b64 %rd13388, %rd13321; + and.b64 %rd13389, %rd13322, %rd13388; + xor.b64 %rd13390, %rd13389, %rd13325; + not.b64 %rd13391, %rd13327; + and.b64 %rd13392, %rd13328, %rd13391; + xor.b64 %rd13393, %rd13392, %rd13326; + not.b64 %rd13394, %rd13328; + and.b64 %rd13395, %rd13329, %rd13394; + xor.b64 %rd13396, %rd13395, %rd13327; + not.b64 %rd13397, %rd13329; + and.b64 %rd13398, %rd13330, %rd13397; + xor.b64 %rd13399, %rd13398, %rd13328; + not.b64 %rd13400, %rd13330; + and.b64 %rd13401, %rd13326, %rd13400; + xor.b64 %rd13402, %rd13401, %rd13329; + not.b64 %rd13403, %rd13326; + and.b64 %rd13404, %rd13327, %rd13403; + xor.b64 %rd13405, %rd13404, %rd13330; + xor.b64 %rd13406, %rd13333, %rd24; + xor.b64 %rd13407, %rd13348, %rd13406; + xor.b64 %rd13408, %rd13407, %rd13363; + xor.b64 %rd13409, %rd13408, %rd13378; + xor.b64 %rd13410, %rd13409, %rd13393; + xor.b64 %rd13411, %rd13351, %rd13336; + xor.b64 %rd13412, %rd13411, %rd13366; + xor.b64 %rd13413, %rd13412, %rd13381; + xor.b64 %rd13414, %rd13413, %rd13396; + xor.b64 %rd13415, %rd13354, %rd13339; + xor.b64 %rd13416, %rd13415, %rd13369; + xor.b64 %rd13417, %rd13416, %rd13384; + xor.b64 %rd13418, %rd13417, %rd13399; + xor.b64 %rd13419, %rd13357, %rd13342; + xor.b64 %rd13420, %rd13419, %rd13372; + xor.b64 %rd13421, %rd13420, %rd13387; + xor.b64 %rd13422, %rd13421, %rd13402; + xor.b64 %rd13423, %rd13360, %rd13345; + xor.b64 %rd13424, %rd13423, %rd13375; + xor.b64 %rd13425, %rd13424, %rd13390; + xor.b64 %rd13426, %rd13425, %rd13405; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9480}, %rd13414; + } + { + .reg .b32 %dummy; + mov.b64 {%r9481,%dummy}, %rd13414; + } + shf.l.wrap.b32 %r9482, %r9481, %r9480, 1; + shf.l.wrap.b32 %r9483, %r9480, %r9481, 1; + mov.b64 %rd13427, {%r9483, %r9482}; + xor.b64 %rd13428, %rd13426, %rd13427; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9484}, %rd13418; + } + { + .reg .b32 %dummy; + mov.b64 {%r9485,%dummy}, %rd13418; + } + shf.l.wrap.b32 %r9486, %r9485, %r9484, 1; + shf.l.wrap.b32 %r9487, %r9484, %r9485, 1; + mov.b64 %rd13429, {%r9487, %r9486}; + xor.b64 %rd13430, %rd13429, %rd13410; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9488}, %rd13422; + } + { + .reg .b32 %dummy; + mov.b64 {%r9489,%dummy}, %rd13422; + } + shf.l.wrap.b32 %r9490, %r9489, %r9488, 1; + shf.l.wrap.b32 %r9491, %r9488, %r9489, 1; + mov.b64 %rd13431, {%r9491, %r9490}; + xor.b64 %rd13432, %rd13431, %rd13414; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9492}, %rd13426; + } + { + .reg .b32 %dummy; + mov.b64 {%r9493,%dummy}, %rd13426; + } + shf.l.wrap.b32 %r9494, %r9493, %r9492, 1; + shf.l.wrap.b32 %r9495, %r9492, %r9493, 1; + mov.b64 %rd13433, {%r9495, %r9494}; + xor.b64 %rd13434, %rd13433, %rd13418; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9496}, %rd13410; + } + { + .reg .b32 %dummy; + mov.b64 {%r9497,%dummy}, %rd13410; + } + shf.l.wrap.b32 %r9498, %r9497, %r9496, 1; + shf.l.wrap.b32 %r9499, %r9496, %r9497, 1; + mov.b64 %rd13435, {%r9499, %r9498}; + xor.b64 %rd13436, %rd13422, %rd13435; + xor.b64 %rd13437, %rd13428, %rd13406; + xor.b64 %rd13438, %rd13430, %rd13336; + xor.b64 %rd13439, %rd13432, %rd13339; + xor.b64 %rd13440, %rd13434, %rd13342; + xor.b64 %rd13441, %rd13436, %rd13345; + xor.b64 %rd13442, %rd13428, %rd13348; + xor.b64 %rd13443, %rd13430, %rd13351; + xor.b64 %rd13444, %rd13432, %rd13354; + xor.b64 %rd13445, %rd13434, %rd13357; + xor.b64 %rd13446, %rd13436, %rd13360; + xor.b64 %rd13447, %rd13428, %rd13363; + xor.b64 %rd13448, %rd13430, %rd13366; + xor.b64 %rd13449, %rd13432, %rd13369; + xor.b64 %rd13450, %rd13434, %rd13372; + xor.b64 %rd13451, %rd13436, %rd13375; + xor.b64 %rd13452, %rd13428, %rd13378; + xor.b64 %rd13453, %rd13430, %rd13381; + xor.b64 %rd13454, %rd13432, %rd13384; + xor.b64 %rd13455, %rd13434, %rd13387; + xor.b64 %rd13456, %rd13436, %rd13390; + xor.b64 %rd13457, %rd13428, %rd13393; + xor.b64 %rd13458, %rd13430, %rd13396; + xor.b64 %rd13459, %rd13432, %rd13399; + xor.b64 %rd13460, %rd13434, %rd13402; + xor.b64 %rd13461, %rd13436, %rd13405; + { + .reg .b32 %dummy; + mov.b64 {%r9500,%dummy}, %rd13443; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9501}, %rd13443; + } + shf.r.wrap.b32 %r9502, %r9501, %r9500, 20; + shf.r.wrap.b32 %r9503, %r9500, %r9501, 20; + mov.b64 %rd13462, {%r9503, %r9502}; + { + .reg .b32 %dummy; + mov.b64 {%r9504,%dummy}, %rd13449; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9505}, %rd13449; + } + shf.r.wrap.b32 %r9506, %r9505, %r9504, 21; + shf.r.wrap.b32 %r9507, %r9504, %r9505, 21; + mov.b64 %rd13463, {%r9507, %r9506}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9508}, %rd13455; + } + { + .reg .b32 %dummy; + mov.b64 {%r9509,%dummy}, %rd13455; + } + shf.l.wrap.b32 %r9510, %r9509, %r9508, 21; + shf.l.wrap.b32 %r9511, %r9508, %r9509, 21; + mov.b64 %rd13464, {%r9511, %r9510}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9512}, %rd13461; + } + { + .reg .b32 %dummy; + mov.b64 {%r9513,%dummy}, %rd13461; + } + shf.l.wrap.b32 %r9514, %r9513, %r9512, 14; + shf.l.wrap.b32 %r9515, %r9512, %r9513, 14; + mov.b64 %rd13465, {%r9515, %r9514}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9516}, %rd13440; + } + { + .reg .b32 %dummy; + mov.b64 {%r9517,%dummy}, %rd13440; + } + shf.l.wrap.b32 %r9518, %r9517, %r9516, 28; + shf.l.wrap.b32 %r9519, %r9516, %r9517, 28; + mov.b64 %rd13466, {%r9519, %r9518}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9520}, %rd13446; + } + { + .reg .b32 %dummy; + mov.b64 {%r9521,%dummy}, %rd13446; + } + shf.l.wrap.b32 %r9522, %r9521, %r9520, 20; + shf.l.wrap.b32 %r9523, %r9520, %r9521, 20; + mov.b64 %rd13467, {%r9523, %r9522}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9524}, %rd13447; + } + { + .reg .b32 %dummy; + mov.b64 {%r9525,%dummy}, %rd13447; + } + shf.l.wrap.b32 %r9526, %r9525, %r9524, 3; + shf.l.wrap.b32 %r9527, %r9524, %r9525, 3; + mov.b64 %rd13468, {%r9527, %r9526}; + { + .reg .b32 %dummy; + mov.b64 {%r9528,%dummy}, %rd13453; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9529}, %rd13453; + } + shf.r.wrap.b32 %r9530, %r9529, %r9528, 19; + shf.r.wrap.b32 %r9531, %r9528, %r9529, 19; + mov.b64 %rd13469, {%r9531, %r9530}; + { + .reg .b32 %dummy; + mov.b64 {%r9532,%dummy}, %rd13459; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9533}, %rd13459; + } + shf.r.wrap.b32 %r9534, %r9533, %r9532, 3; + shf.r.wrap.b32 %r9535, %r9532, %r9533, 3; + mov.b64 %rd13470, {%r9535, %r9534}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9536}, %rd13438; + } + { + .reg .b32 %dummy; + mov.b64 {%r9537,%dummy}, %rd13438; + } + shf.l.wrap.b32 %r9538, %r9537, %r9536, 1; + shf.l.wrap.b32 %r9539, %r9536, %r9537, 1; + mov.b64 %rd13471, {%r9539, %r9538}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9540}, %rd13444; + } + { + .reg .b32 %dummy; + mov.b64 {%r9541,%dummy}, %rd13444; + } + shf.l.wrap.b32 %r9542, %r9541, %r9540, 6; + shf.l.wrap.b32 %r9543, %r9540, %r9541, 6; + mov.b64 %rd13472, {%r9543, %r9542}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9544}, %rd13450; + } + { + .reg .b32 %dummy; + mov.b64 {%r9545,%dummy}, %rd13450; + } + shf.l.wrap.b32 %r9546, %r9545, %r9544, 25; + shf.l.wrap.b32 %r9547, %r9544, %r9545, 25; + mov.b64 %rd13473, {%r9547, %r9546}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9548}, %rd13456; + } + { + .reg .b32 %dummy; + mov.b64 {%r9549,%dummy}, %rd13456; + } + shf.l.wrap.b32 %r9550, %r9549, %r9548, 8; + shf.l.wrap.b32 %r9551, %r9548, %r9549, 8; + mov.b64 %rd13474, {%r9551, %r9550}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9552}, %rd13457; + } + { + .reg .b32 %dummy; + mov.b64 {%r9553,%dummy}, %rd13457; + } + shf.l.wrap.b32 %r9554, %r9553, %r9552, 18; + shf.l.wrap.b32 %r9555, %r9552, %r9553, 18; + mov.b64 %rd13475, {%r9555, %r9554}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9556}, %rd13441; + } + { + .reg .b32 %dummy; + mov.b64 {%r9557,%dummy}, %rd13441; + } + shf.l.wrap.b32 %r9558, %r9557, %r9556, 27; + shf.l.wrap.b32 %r9559, %r9556, %r9557, 27; + mov.b64 %rd13476, {%r9559, %r9558}; + { + .reg .b32 %dummy; + mov.b64 {%r9560,%dummy}, %rd13442; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9561}, %rd13442; + } + shf.r.wrap.b32 %r9562, %r9561, %r9560, 28; + shf.r.wrap.b32 %r9563, %r9560, %r9561, 28; + mov.b64 %rd13477, {%r9563, %r9562}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9564}, %rd13448; + } + { + .reg .b32 %dummy; + mov.b64 {%r9565,%dummy}, %rd13448; + } + shf.l.wrap.b32 %r9566, %r9565, %r9564, 10; + shf.l.wrap.b32 %r9567, %r9564, %r9565, 10; + mov.b64 %rd13478, {%r9567, %r9566}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9568}, %rd13454; + } + { + .reg .b32 %dummy; + mov.b64 {%r9569,%dummy}, %rd13454; + } + shf.l.wrap.b32 %r9570, %r9569, %r9568, 15; + shf.l.wrap.b32 %r9571, %r9568, %r9569, 15; + mov.b64 %rd13479, {%r9571, %r9570}; + { + .reg .b32 %dummy; + mov.b64 {%r9572,%dummy}, %rd13460; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9573}, %rd13460; + } + shf.r.wrap.b32 %r9574, %r9573, %r9572, 8; + shf.r.wrap.b32 %r9575, %r9572, %r9573, 8; + mov.b64 %rd13480, {%r9575, %r9574}; + { + .reg .b32 %dummy; + mov.b64 {%r9576,%dummy}, %rd13439; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9577}, %rd13439; + } + shf.r.wrap.b32 %r9578, %r9577, %r9576, 2; + shf.r.wrap.b32 %r9579, %r9576, %r9577, 2; + mov.b64 %rd13481, {%r9579, %r9578}; + { + .reg .b32 %dummy; + mov.b64 {%r9580,%dummy}, %rd13445; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9581}, %rd13445; + } + shf.r.wrap.b32 %r9582, %r9581, %r9580, 9; + shf.r.wrap.b32 %r9583, %r9580, %r9581, 9; + mov.b64 %rd13482, {%r9583, %r9582}; + { + .reg .b32 %dummy; + mov.b64 {%r9584,%dummy}, %rd13451; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9585}, %rd13451; + } + shf.r.wrap.b32 %r9586, %r9585, %r9584, 25; + shf.r.wrap.b32 %r9587, %r9584, %r9585, 25; + mov.b64 %rd13483, {%r9587, %r9586}; + { + .reg .b32 %dummy; + mov.b64 {%r9588,%dummy}, %rd13452; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9589}, %rd13452; + } + shf.r.wrap.b32 %r9590, %r9589, %r9588, 23; + shf.r.wrap.b32 %r9591, %r9588, %r9589, 23; + mov.b64 %rd13484, {%r9591, %r9590}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9592}, %rd13458; + } + { + .reg .b32 %dummy; + mov.b64 {%r9593,%dummy}, %rd13458; + } + shf.l.wrap.b32 %r9594, %r9593, %r9592, 2; + shf.l.wrap.b32 %r9595, %r9592, %r9593, 2; + mov.b64 %rd13485, {%r9595, %r9594}; + not.b64 %rd13486, %rd13462; + and.b64 %rd13487, %rd13463, %rd13486; + xor.b64 %rd13488, %rd13487, %rd13437; + not.b64 %rd13489, %rd13463; + and.b64 %rd13490, %rd13464, %rd13489; + xor.b64 %rd13491, %rd13490, %rd13462; + not.b64 %rd13492, %rd13464; + and.b64 %rd13493, %rd13465, %rd13492; + xor.b64 %rd13494, %rd13493, %rd13463; + not.b64 %rd13495, %rd13465; + and.b64 %rd13496, %rd13437, %rd13495; + xor.b64 %rd13497, %rd13496, %rd13464; + not.b64 %rd13498, %rd13437; + and.b64 %rd13499, %rd13462, %rd13498; + xor.b64 %rd13500, %rd13499, %rd13465; + not.b64 %rd13501, %rd13467; + and.b64 %rd13502, %rd13468, %rd13501; + xor.b64 %rd13503, %rd13502, %rd13466; + not.b64 %rd13504, %rd13468; + and.b64 %rd13505, %rd13469, %rd13504; + xor.b64 %rd13506, %rd13505, %rd13467; + not.b64 %rd13507, %rd13469; + and.b64 %rd13508, %rd13470, %rd13507; + xor.b64 %rd13509, %rd13508, %rd13468; + not.b64 %rd13510, %rd13470; + and.b64 %rd13511, %rd13466, %rd13510; + xor.b64 %rd13512, %rd13511, %rd13469; + not.b64 %rd13513, %rd13466; + and.b64 %rd13514, %rd13467, %rd13513; + xor.b64 %rd13515, %rd13514, %rd13470; + not.b64 %rd13516, %rd13472; + and.b64 %rd13517, %rd13473, %rd13516; + xor.b64 %rd13518, %rd13517, %rd13471; + not.b64 %rd13519, %rd13473; + and.b64 %rd13520, %rd13474, %rd13519; + xor.b64 %rd13521, %rd13520, %rd13472; + not.b64 %rd13522, %rd13474; + and.b64 %rd13523, %rd13475, %rd13522; + xor.b64 %rd13524, %rd13523, %rd13473; + not.b64 %rd13525, %rd13475; + and.b64 %rd13526, %rd13471, %rd13525; + xor.b64 %rd13527, %rd13526, %rd13474; + not.b64 %rd13528, %rd13471; + and.b64 %rd13529, %rd13472, %rd13528; + xor.b64 %rd13530, %rd13529, %rd13475; + not.b64 %rd13531, %rd13477; + and.b64 %rd13532, %rd13478, %rd13531; + xor.b64 %rd13533, %rd13532, %rd13476; + not.b64 %rd13534, %rd13478; + and.b64 %rd13535, %rd13479, %rd13534; + xor.b64 %rd13536, %rd13535, %rd13477; + not.b64 %rd13537, %rd13479; + and.b64 %rd13538, %rd13480, %rd13537; + xor.b64 %rd13539, %rd13538, %rd13478; + not.b64 %rd13540, %rd13480; + and.b64 %rd13541, %rd13476, %rd13540; + xor.b64 %rd13542, %rd13541, %rd13479; + not.b64 %rd13543, %rd13476; + and.b64 %rd13544, %rd13477, %rd13543; + xor.b64 %rd13545, %rd13544, %rd13480; + not.b64 %rd13546, %rd13482; + and.b64 %rd13547, %rd13483, %rd13546; + xor.b64 %rd13548, %rd13547, %rd13481; + not.b64 %rd13549, %rd13483; + and.b64 %rd13550, %rd13484, %rd13549; + xor.b64 %rd13551, %rd13550, %rd13482; + not.b64 %rd13552, %rd13484; + and.b64 %rd13553, %rd13485, %rd13552; + xor.b64 %rd13554, %rd13553, %rd13483; + not.b64 %rd13555, %rd13485; + and.b64 %rd13556, %rd13481, %rd13555; + xor.b64 %rd13557, %rd13556, %rd13484; + not.b64 %rd13558, %rd13481; + and.b64 %rd13559, %rd13482, %rd13558; + xor.b64 %rd13560, %rd13559, %rd13485; + xor.b64 %rd13561, %rd13488, %rd25; + xor.b64 %rd13562, %rd13503, %rd13561; + xor.b64 %rd13563, %rd13562, %rd13518; + xor.b64 %rd13564, %rd13563, %rd13533; + xor.b64 %rd13565, %rd13564, %rd13548; + xor.b64 %rd13566, %rd13506, %rd13491; + xor.b64 %rd13567, %rd13566, %rd13521; + xor.b64 %rd13568, %rd13567, %rd13536; + xor.b64 %rd13569, %rd13568, %rd13551; + xor.b64 %rd13570, %rd13509, %rd13494; + xor.b64 %rd13571, %rd13570, %rd13524; + xor.b64 %rd13572, %rd13571, %rd13539; + xor.b64 %rd13573, %rd13572, %rd13554; + xor.b64 %rd13574, %rd13512, %rd13497; + xor.b64 %rd13575, %rd13574, %rd13527; + xor.b64 %rd13576, %rd13575, %rd13542; + xor.b64 %rd13577, %rd13576, %rd13557; + xor.b64 %rd13578, %rd13515, %rd13500; + xor.b64 %rd13579, %rd13578, %rd13530; + xor.b64 %rd13580, %rd13579, %rd13545; + xor.b64 %rd13581, %rd13580, %rd13560; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9596}, %rd13569; + } + { + .reg .b32 %dummy; + mov.b64 {%r9597,%dummy}, %rd13569; + } + shf.l.wrap.b32 %r9598, %r9597, %r9596, 1; + shf.l.wrap.b32 %r9599, %r9596, %r9597, 1; + mov.b64 %rd13582, {%r9599, %r9598}; + xor.b64 %rd13583, %rd13581, %rd13582; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9600}, %rd13573; + } + { + .reg .b32 %dummy; + mov.b64 {%r9601,%dummy}, %rd13573; + } + shf.l.wrap.b32 %r9602, %r9601, %r9600, 1; + shf.l.wrap.b32 %r9603, %r9600, %r9601, 1; + mov.b64 %rd13584, {%r9603, %r9602}; + xor.b64 %rd13585, %rd13584, %rd13565; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9604}, %rd13577; + } + { + .reg .b32 %dummy; + mov.b64 {%r9605,%dummy}, %rd13577; + } + shf.l.wrap.b32 %r9606, %r9605, %r9604, 1; + shf.l.wrap.b32 %r9607, %r9604, %r9605, 1; + mov.b64 %rd13586, {%r9607, %r9606}; + xor.b64 %rd13587, %rd13586, %rd13569; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9608}, %rd13581; + } + { + .reg .b32 %dummy; + mov.b64 {%r9609,%dummy}, %rd13581; + } + shf.l.wrap.b32 %r9610, %r9609, %r9608, 1; + shf.l.wrap.b32 %r9611, %r9608, %r9609, 1; + mov.b64 %rd13588, {%r9611, %r9610}; + xor.b64 %rd13589, %rd13588, %rd13573; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9612}, %rd13565; + } + { + .reg .b32 %dummy; + mov.b64 {%r9613,%dummy}, %rd13565; + } + shf.l.wrap.b32 %r9614, %r9613, %r9612, 1; + shf.l.wrap.b32 %r9615, %r9612, %r9613, 1; + mov.b64 %rd13590, {%r9615, %r9614}; + xor.b64 %rd13591, %rd13577, %rd13590; + xor.b64 %rd13592, %rd13583, %rd13561; + xor.b64 %rd13593, %rd13585, %rd13491; + xor.b64 %rd13594, %rd13587, %rd13494; + xor.b64 %rd13595, %rd13589, %rd13497; + xor.b64 %rd13596, %rd13591, %rd13500; + xor.b64 %rd13597, %rd13583, %rd13503; + xor.b64 %rd13598, %rd13585, %rd13506; + xor.b64 %rd13599, %rd13587, %rd13509; + xor.b64 %rd13600, %rd13589, %rd13512; + xor.b64 %rd13601, %rd13591, %rd13515; + xor.b64 %rd13602, %rd13583, %rd13518; + xor.b64 %rd13603, %rd13585, %rd13521; + xor.b64 %rd13604, %rd13587, %rd13524; + xor.b64 %rd13605, %rd13589, %rd13527; + xor.b64 %rd13606, %rd13591, %rd13530; + xor.b64 %rd13607, %rd13583, %rd13533; + xor.b64 %rd13608, %rd13585, %rd13536; + xor.b64 %rd13609, %rd13587, %rd13539; + xor.b64 %rd13610, %rd13589, %rd13542; + xor.b64 %rd13611, %rd13591, %rd13545; + xor.b64 %rd13612, %rd13583, %rd13548; + xor.b64 %rd13613, %rd13585, %rd13551; + xor.b64 %rd13614, %rd13587, %rd13554; + xor.b64 %rd13615, %rd13589, %rd13557; + xor.b64 %rd13616, %rd13591, %rd13560; + { + .reg .b32 %dummy; + mov.b64 {%r9616,%dummy}, %rd13598; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9617}, %rd13598; + } + shf.r.wrap.b32 %r9618, %r9617, %r9616, 20; + shf.r.wrap.b32 %r9619, %r9616, %r9617, 20; + mov.b64 %rd13617, {%r9619, %r9618}; + { + .reg .b32 %dummy; + mov.b64 {%r9620,%dummy}, %rd13604; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9621}, %rd13604; + } + shf.r.wrap.b32 %r9622, %r9621, %r9620, 21; + shf.r.wrap.b32 %r9623, %r9620, %r9621, 21; + mov.b64 %rd13618, {%r9623, %r9622}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9624}, %rd13610; + } + { + .reg .b32 %dummy; + mov.b64 {%r9625,%dummy}, %rd13610; + } + shf.l.wrap.b32 %r9626, %r9625, %r9624, 21; + shf.l.wrap.b32 %r9627, %r9624, %r9625, 21; + mov.b64 %rd13619, {%r9627, %r9626}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9628}, %rd13616; + } + { + .reg .b32 %dummy; + mov.b64 {%r9629,%dummy}, %rd13616; + } + shf.l.wrap.b32 %r9630, %r9629, %r9628, 14; + shf.l.wrap.b32 %r9631, %r9628, %r9629, 14; + mov.b64 %rd13620, {%r9631, %r9630}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9632}, %rd13595; + } + { + .reg .b32 %dummy; + mov.b64 {%r9633,%dummy}, %rd13595; + } + shf.l.wrap.b32 %r9634, %r9633, %r9632, 28; + shf.l.wrap.b32 %r9635, %r9632, %r9633, 28; + mov.b64 %rd13621, {%r9635, %r9634}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9636}, %rd13601; + } + { + .reg .b32 %dummy; + mov.b64 {%r9637,%dummy}, %rd13601; + } + shf.l.wrap.b32 %r9638, %r9637, %r9636, 20; + shf.l.wrap.b32 %r9639, %r9636, %r9637, 20; + mov.b64 %rd13622, {%r9639, %r9638}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9640}, %rd13602; + } + { + .reg .b32 %dummy; + mov.b64 {%r9641,%dummy}, %rd13602; + } + shf.l.wrap.b32 %r9642, %r9641, %r9640, 3; + shf.l.wrap.b32 %r9643, %r9640, %r9641, 3; + mov.b64 %rd13623, {%r9643, %r9642}; + { + .reg .b32 %dummy; + mov.b64 {%r9644,%dummy}, %rd13608; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9645}, %rd13608; + } + shf.r.wrap.b32 %r9646, %r9645, %r9644, 19; + shf.r.wrap.b32 %r9647, %r9644, %r9645, 19; + mov.b64 %rd13624, {%r9647, %r9646}; + { + .reg .b32 %dummy; + mov.b64 {%r9648,%dummy}, %rd13614; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9649}, %rd13614; + } + shf.r.wrap.b32 %r9650, %r9649, %r9648, 3; + shf.r.wrap.b32 %r9651, %r9648, %r9649, 3; + mov.b64 %rd13625, {%r9651, %r9650}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9652}, %rd13593; + } + { + .reg .b32 %dummy; + mov.b64 {%r9653,%dummy}, %rd13593; + } + shf.l.wrap.b32 %r9654, %r9653, %r9652, 1; + shf.l.wrap.b32 %r9655, %r9652, %r9653, 1; + mov.b64 %rd13626, {%r9655, %r9654}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9656}, %rd13599; + } + { + .reg .b32 %dummy; + mov.b64 {%r9657,%dummy}, %rd13599; + } + shf.l.wrap.b32 %r9658, %r9657, %r9656, 6; + shf.l.wrap.b32 %r9659, %r9656, %r9657, 6; + mov.b64 %rd13627, {%r9659, %r9658}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9660}, %rd13605; + } + { + .reg .b32 %dummy; + mov.b64 {%r9661,%dummy}, %rd13605; + } + shf.l.wrap.b32 %r9662, %r9661, %r9660, 25; + shf.l.wrap.b32 %r9663, %r9660, %r9661, 25; + mov.b64 %rd13628, {%r9663, %r9662}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9664}, %rd13611; + } + { + .reg .b32 %dummy; + mov.b64 {%r9665,%dummy}, %rd13611; + } + shf.l.wrap.b32 %r9666, %r9665, %r9664, 8; + shf.l.wrap.b32 %r9667, %r9664, %r9665, 8; + mov.b64 %rd13629, {%r9667, %r9666}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9668}, %rd13612; + } + { + .reg .b32 %dummy; + mov.b64 {%r9669,%dummy}, %rd13612; + } + shf.l.wrap.b32 %r9670, %r9669, %r9668, 18; + shf.l.wrap.b32 %r9671, %r9668, %r9669, 18; + mov.b64 %rd13630, {%r9671, %r9670}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9672}, %rd13596; + } + { + .reg .b32 %dummy; + mov.b64 {%r9673,%dummy}, %rd13596; + } + shf.l.wrap.b32 %r9674, %r9673, %r9672, 27; + shf.l.wrap.b32 %r9675, %r9672, %r9673, 27; + mov.b64 %rd13631, {%r9675, %r9674}; + { + .reg .b32 %dummy; + mov.b64 {%r9676,%dummy}, %rd13597; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9677}, %rd13597; + } + shf.r.wrap.b32 %r9678, %r9677, %r9676, 28; + shf.r.wrap.b32 %r9679, %r9676, %r9677, 28; + mov.b64 %rd13632, {%r9679, %r9678}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9680}, %rd13603; + } + { + .reg .b32 %dummy; + mov.b64 {%r9681,%dummy}, %rd13603; + } + shf.l.wrap.b32 %r9682, %r9681, %r9680, 10; + shf.l.wrap.b32 %r9683, %r9680, %r9681, 10; + mov.b64 %rd13633, {%r9683, %r9682}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9684}, %rd13609; + } + { + .reg .b32 %dummy; + mov.b64 {%r9685,%dummy}, %rd13609; + } + shf.l.wrap.b32 %r9686, %r9685, %r9684, 15; + shf.l.wrap.b32 %r9687, %r9684, %r9685, 15; + mov.b64 %rd13634, {%r9687, %r9686}; + { + .reg .b32 %dummy; + mov.b64 {%r9688,%dummy}, %rd13615; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9689}, %rd13615; + } + shf.r.wrap.b32 %r9690, %r9689, %r9688, 8; + shf.r.wrap.b32 %r9691, %r9688, %r9689, 8; + mov.b64 %rd13635, {%r9691, %r9690}; + { + .reg .b32 %dummy; + mov.b64 {%r9692,%dummy}, %rd13594; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9693}, %rd13594; + } + shf.r.wrap.b32 %r9694, %r9693, %r9692, 2; + shf.r.wrap.b32 %r9695, %r9692, %r9693, 2; + mov.b64 %rd13636, {%r9695, %r9694}; + { + .reg .b32 %dummy; + mov.b64 {%r9696,%dummy}, %rd13600; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9697}, %rd13600; + } + shf.r.wrap.b32 %r9698, %r9697, %r9696, 9; + shf.r.wrap.b32 %r9699, %r9696, %r9697, 9; + mov.b64 %rd13637, {%r9699, %r9698}; + { + .reg .b32 %dummy; + mov.b64 {%r9700,%dummy}, %rd13606; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9701}, %rd13606; + } + shf.r.wrap.b32 %r9702, %r9701, %r9700, 25; + shf.r.wrap.b32 %r9703, %r9700, %r9701, 25; + mov.b64 %rd13638, {%r9703, %r9702}; + { + .reg .b32 %dummy; + mov.b64 {%r9704,%dummy}, %rd13607; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9705}, %rd13607; + } + shf.r.wrap.b32 %r9706, %r9705, %r9704, 23; + shf.r.wrap.b32 %r9707, %r9704, %r9705, 23; + mov.b64 %rd13639, {%r9707, %r9706}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9708}, %rd13613; + } + { + .reg .b32 %dummy; + mov.b64 {%r9709,%dummy}, %rd13613; + } + shf.l.wrap.b32 %r9710, %r9709, %r9708, 2; + shf.l.wrap.b32 %r9711, %r9708, %r9709, 2; + mov.b64 %rd13640, {%r9711, %r9710}; + not.b64 %rd13641, %rd13617; + and.b64 %rd13642, %rd13618, %rd13641; + xor.b64 %rd13643, %rd13642, %rd13592; + not.b64 %rd13644, %rd13618; + and.b64 %rd13645, %rd13619, %rd13644; + xor.b64 %rd13646, %rd13645, %rd13617; + not.b64 %rd13647, %rd13619; + and.b64 %rd13648, %rd13620, %rd13647; + xor.b64 %rd13649, %rd13648, %rd13618; + not.b64 %rd13650, %rd13620; + and.b64 %rd13651, %rd13592, %rd13650; + xor.b64 %rd13652, %rd13651, %rd13619; + not.b64 %rd13653, %rd13592; + and.b64 %rd13654, %rd13617, %rd13653; + xor.b64 %rd13655, %rd13654, %rd13620; + not.b64 %rd13656, %rd13622; + and.b64 %rd13657, %rd13623, %rd13656; + xor.b64 %rd13658, %rd13657, %rd13621; + not.b64 %rd13659, %rd13623; + and.b64 %rd13660, %rd13624, %rd13659; + xor.b64 %rd13661, %rd13660, %rd13622; + not.b64 %rd13662, %rd13624; + and.b64 %rd13663, %rd13625, %rd13662; + xor.b64 %rd13664, %rd13663, %rd13623; + not.b64 %rd13665, %rd13625; + and.b64 %rd13666, %rd13621, %rd13665; + xor.b64 %rd13667, %rd13666, %rd13624; + not.b64 %rd13668, %rd13621; + and.b64 %rd13669, %rd13622, %rd13668; + xor.b64 %rd13670, %rd13669, %rd13625; + not.b64 %rd13671, %rd13627; + and.b64 %rd13672, %rd13628, %rd13671; + xor.b64 %rd13673, %rd13672, %rd13626; + not.b64 %rd13674, %rd13628; + and.b64 %rd13675, %rd13629, %rd13674; + xor.b64 %rd13676, %rd13675, %rd13627; + not.b64 %rd13677, %rd13629; + and.b64 %rd13678, %rd13630, %rd13677; + xor.b64 %rd13679, %rd13678, %rd13628; + not.b64 %rd13680, %rd13630; + and.b64 %rd13681, %rd13626, %rd13680; + xor.b64 %rd13682, %rd13681, %rd13629; + not.b64 %rd13683, %rd13626; + and.b64 %rd13684, %rd13627, %rd13683; + xor.b64 %rd13685, %rd13684, %rd13630; + not.b64 %rd13686, %rd13632; + and.b64 %rd13687, %rd13633, %rd13686; + xor.b64 %rd13688, %rd13687, %rd13631; + not.b64 %rd13689, %rd13633; + and.b64 %rd13690, %rd13634, %rd13689; + xor.b64 %rd13691, %rd13690, %rd13632; + not.b64 %rd13692, %rd13634; + and.b64 %rd13693, %rd13635, %rd13692; + xor.b64 %rd13694, %rd13693, %rd13633; + not.b64 %rd13695, %rd13635; + and.b64 %rd13696, %rd13631, %rd13695; + xor.b64 %rd13697, %rd13696, %rd13634; + not.b64 %rd13698, %rd13631; + and.b64 %rd13699, %rd13632, %rd13698; + xor.b64 %rd13700, %rd13699, %rd13635; + not.b64 %rd13701, %rd13637; + and.b64 %rd13702, %rd13638, %rd13701; + xor.b64 %rd13703, %rd13702, %rd13636; + not.b64 %rd13704, %rd13638; + and.b64 %rd13705, %rd13639, %rd13704; + xor.b64 %rd13706, %rd13705, %rd13637; + not.b64 %rd13707, %rd13639; + and.b64 %rd13708, %rd13640, %rd13707; + xor.b64 %rd13709, %rd13708, %rd13638; + not.b64 %rd13710, %rd13640; + and.b64 %rd13711, %rd13636, %rd13710; + xor.b64 %rd13712, %rd13711, %rd13639; + not.b64 %rd13713, %rd13636; + and.b64 %rd13714, %rd13637, %rd13713; + xor.b64 %rd13715, %rd13714, %rd13640; + xor.b64 %rd13716, %rd13643, %rd26; + xor.b64 %rd13717, %rd13658, %rd13716; + xor.b64 %rd13718, %rd13717, %rd13673; + xor.b64 %rd13719, %rd13718, %rd13688; + xor.b64 %rd13720, %rd13719, %rd13703; + xor.b64 %rd13721, %rd13661, %rd13646; + xor.b64 %rd13722, %rd13721, %rd13676; + xor.b64 %rd13723, %rd13722, %rd13691; + xor.b64 %rd13724, %rd13723, %rd13706; + xor.b64 %rd13725, %rd13664, %rd13649; + xor.b64 %rd13726, %rd13725, %rd13679; + xor.b64 %rd13727, %rd13726, %rd13694; + xor.b64 %rd13728, %rd13727, %rd13709; + xor.b64 %rd13729, %rd13667, %rd13652; + xor.b64 %rd13730, %rd13729, %rd13682; + xor.b64 %rd13731, %rd13730, %rd13697; + xor.b64 %rd13732, %rd13731, %rd13712; + xor.b64 %rd13733, %rd13670, %rd13655; + xor.b64 %rd13734, %rd13733, %rd13685; + xor.b64 %rd13735, %rd13734, %rd13700; + xor.b64 %rd13736, %rd13735, %rd13715; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9712}, %rd13724; + } + { + .reg .b32 %dummy; + mov.b64 {%r9713,%dummy}, %rd13724; + } + shf.l.wrap.b32 %r9714, %r9713, %r9712, 1; + shf.l.wrap.b32 %r9715, %r9712, %r9713, 1; + mov.b64 %rd13737, {%r9715, %r9714}; + xor.b64 %rd13738, %rd13736, %rd13737; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9716}, %rd13728; + } + { + .reg .b32 %dummy; + mov.b64 {%r9717,%dummy}, %rd13728; + } + shf.l.wrap.b32 %r9718, %r9717, %r9716, 1; + shf.l.wrap.b32 %r9719, %r9716, %r9717, 1; + mov.b64 %rd13739, {%r9719, %r9718}; + xor.b64 %rd13740, %rd13739, %rd13720; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9720}, %rd13732; + } + { + .reg .b32 %dummy; + mov.b64 {%r9721,%dummy}, %rd13732; + } + shf.l.wrap.b32 %r9722, %r9721, %r9720, 1; + shf.l.wrap.b32 %r9723, %r9720, %r9721, 1; + mov.b64 %rd13741, {%r9723, %r9722}; + xor.b64 %rd13742, %rd13741, %rd13724; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9724}, %rd13736; + } + { + .reg .b32 %dummy; + mov.b64 {%r9725,%dummy}, %rd13736; + } + shf.l.wrap.b32 %r9726, %r9725, %r9724, 1; + shf.l.wrap.b32 %r9727, %r9724, %r9725, 1; + mov.b64 %rd13743, {%r9727, %r9726}; + xor.b64 %rd13744, %rd13743, %rd13728; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9728}, %rd13720; + } + { + .reg .b32 %dummy; + mov.b64 {%r9729,%dummy}, %rd13720; + } + shf.l.wrap.b32 %r9730, %r9729, %r9728, 1; + shf.l.wrap.b32 %r9731, %r9728, %r9729, 1; + mov.b64 %rd13745, {%r9731, %r9730}; + xor.b64 %rd13746, %rd13732, %rd13745; + xor.b64 %rd13747, %rd13738, %rd13716; + xor.b64 %rd13748, %rd13740, %rd13646; + xor.b64 %rd13749, %rd13742, %rd13649; + xor.b64 %rd13750, %rd13744, %rd13652; + xor.b64 %rd13751, %rd13746, %rd13655; + xor.b64 %rd13752, %rd13738, %rd13658; + xor.b64 %rd13753, %rd13740, %rd13661; + xor.b64 %rd13754, %rd13742, %rd13664; + xor.b64 %rd13755, %rd13744, %rd13667; + xor.b64 %rd13756, %rd13746, %rd13670; + xor.b64 %rd13757, %rd13738, %rd13673; + xor.b64 %rd13758, %rd13740, %rd13676; + xor.b64 %rd13759, %rd13742, %rd13679; + xor.b64 %rd13760, %rd13744, %rd13682; + xor.b64 %rd13761, %rd13746, %rd13685; + xor.b64 %rd13762, %rd13738, %rd13688; + xor.b64 %rd13763, %rd13740, %rd13691; + xor.b64 %rd13764, %rd13742, %rd13694; + xor.b64 %rd13765, %rd13744, %rd13697; + xor.b64 %rd13766, %rd13746, %rd13700; + xor.b64 %rd13767, %rd13738, %rd13703; + xor.b64 %rd13768, %rd13740, %rd13706; + xor.b64 %rd13769, %rd13742, %rd13709; + xor.b64 %rd13770, %rd13744, %rd13712; + xor.b64 %rd13771, %rd13746, %rd13715; + { + .reg .b32 %dummy; + mov.b64 {%r9732,%dummy}, %rd13753; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9733}, %rd13753; + } + shf.r.wrap.b32 %r9734, %r9733, %r9732, 20; + shf.r.wrap.b32 %r9735, %r9732, %r9733, 20; + mov.b64 %rd13772, {%r9735, %r9734}; + { + .reg .b32 %dummy; + mov.b64 {%r9736,%dummy}, %rd13759; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9737}, %rd13759; + } + shf.r.wrap.b32 %r9738, %r9737, %r9736, 21; + shf.r.wrap.b32 %r9739, %r9736, %r9737, 21; + mov.b64 %rd13773, {%r9739, %r9738}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9740}, %rd13765; + } + { + .reg .b32 %dummy; + mov.b64 {%r9741,%dummy}, %rd13765; + } + shf.l.wrap.b32 %r9742, %r9741, %r9740, 21; + shf.l.wrap.b32 %r9743, %r9740, %r9741, 21; + mov.b64 %rd13774, {%r9743, %r9742}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9744}, %rd13771; + } + { + .reg .b32 %dummy; + mov.b64 {%r9745,%dummy}, %rd13771; + } + shf.l.wrap.b32 %r9746, %r9745, %r9744, 14; + shf.l.wrap.b32 %r9747, %r9744, %r9745, 14; + mov.b64 %rd13775, {%r9747, %r9746}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9748}, %rd13750; + } + { + .reg .b32 %dummy; + mov.b64 {%r9749,%dummy}, %rd13750; + } + shf.l.wrap.b32 %r9750, %r9749, %r9748, 28; + shf.l.wrap.b32 %r9751, %r9748, %r9749, 28; + mov.b64 %rd13776, {%r9751, %r9750}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9752}, %rd13756; + } + { + .reg .b32 %dummy; + mov.b64 {%r9753,%dummy}, %rd13756; + } + shf.l.wrap.b32 %r9754, %r9753, %r9752, 20; + shf.l.wrap.b32 %r9755, %r9752, %r9753, 20; + mov.b64 %rd13777, {%r9755, %r9754}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9756}, %rd13757; + } + { + .reg .b32 %dummy; + mov.b64 {%r9757,%dummy}, %rd13757; + } + shf.l.wrap.b32 %r9758, %r9757, %r9756, 3; + shf.l.wrap.b32 %r9759, %r9756, %r9757, 3; + mov.b64 %rd13778, {%r9759, %r9758}; + { + .reg .b32 %dummy; + mov.b64 {%r9760,%dummy}, %rd13763; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9761}, %rd13763; + } + shf.r.wrap.b32 %r9762, %r9761, %r9760, 19; + shf.r.wrap.b32 %r9763, %r9760, %r9761, 19; + mov.b64 %rd13779, {%r9763, %r9762}; + { + .reg .b32 %dummy; + mov.b64 {%r9764,%dummy}, %rd13769; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9765}, %rd13769; + } + shf.r.wrap.b32 %r9766, %r9765, %r9764, 3; + shf.r.wrap.b32 %r9767, %r9764, %r9765, 3; + mov.b64 %rd13780, {%r9767, %r9766}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9768}, %rd13748; + } + { + .reg .b32 %dummy; + mov.b64 {%r9769,%dummy}, %rd13748; + } + shf.l.wrap.b32 %r9770, %r9769, %r9768, 1; + shf.l.wrap.b32 %r9771, %r9768, %r9769, 1; + mov.b64 %rd13781, {%r9771, %r9770}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9772}, %rd13754; + } + { + .reg .b32 %dummy; + mov.b64 {%r9773,%dummy}, %rd13754; + } + shf.l.wrap.b32 %r9774, %r9773, %r9772, 6; + shf.l.wrap.b32 %r9775, %r9772, %r9773, 6; + mov.b64 %rd13782, {%r9775, %r9774}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9776}, %rd13760; + } + { + .reg .b32 %dummy; + mov.b64 {%r9777,%dummy}, %rd13760; + } + shf.l.wrap.b32 %r9778, %r9777, %r9776, 25; + shf.l.wrap.b32 %r9779, %r9776, %r9777, 25; + mov.b64 %rd13783, {%r9779, %r9778}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9780}, %rd13766; + } + { + .reg .b32 %dummy; + mov.b64 {%r9781,%dummy}, %rd13766; + } + shf.l.wrap.b32 %r9782, %r9781, %r9780, 8; + shf.l.wrap.b32 %r9783, %r9780, %r9781, 8; + mov.b64 %rd13784, {%r9783, %r9782}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9784}, %rd13767; + } + { + .reg .b32 %dummy; + mov.b64 {%r9785,%dummy}, %rd13767; + } + shf.l.wrap.b32 %r9786, %r9785, %r9784, 18; + shf.l.wrap.b32 %r9787, %r9784, %r9785, 18; + mov.b64 %rd13785, {%r9787, %r9786}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9788}, %rd13751; + } + { + .reg .b32 %dummy; + mov.b64 {%r9789,%dummy}, %rd13751; + } + shf.l.wrap.b32 %r9790, %r9789, %r9788, 27; + shf.l.wrap.b32 %r9791, %r9788, %r9789, 27; + mov.b64 %rd13786, {%r9791, %r9790}; + { + .reg .b32 %dummy; + mov.b64 {%r9792,%dummy}, %rd13752; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9793}, %rd13752; + } + shf.r.wrap.b32 %r9794, %r9793, %r9792, 28; + shf.r.wrap.b32 %r9795, %r9792, %r9793, 28; + mov.b64 %rd13787, {%r9795, %r9794}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9796}, %rd13758; + } + { + .reg .b32 %dummy; + mov.b64 {%r9797,%dummy}, %rd13758; + } + shf.l.wrap.b32 %r9798, %r9797, %r9796, 10; + shf.l.wrap.b32 %r9799, %r9796, %r9797, 10; + mov.b64 %rd13788, {%r9799, %r9798}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9800}, %rd13764; + } + { + .reg .b32 %dummy; + mov.b64 {%r9801,%dummy}, %rd13764; + } + shf.l.wrap.b32 %r9802, %r9801, %r9800, 15; + shf.l.wrap.b32 %r9803, %r9800, %r9801, 15; + mov.b64 %rd13789, {%r9803, %r9802}; + { + .reg .b32 %dummy; + mov.b64 {%r9804,%dummy}, %rd13770; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9805}, %rd13770; + } + shf.r.wrap.b32 %r9806, %r9805, %r9804, 8; + shf.r.wrap.b32 %r9807, %r9804, %r9805, 8; + mov.b64 %rd13790, {%r9807, %r9806}; + { + .reg .b32 %dummy; + mov.b64 {%r9808,%dummy}, %rd13749; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9809}, %rd13749; + } + shf.r.wrap.b32 %r9810, %r9809, %r9808, 2; + shf.r.wrap.b32 %r9811, %r9808, %r9809, 2; + mov.b64 %rd13791, {%r9811, %r9810}; + { + .reg .b32 %dummy; + mov.b64 {%r9812,%dummy}, %rd13755; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9813}, %rd13755; + } + shf.r.wrap.b32 %r9814, %r9813, %r9812, 9; + shf.r.wrap.b32 %r9815, %r9812, %r9813, 9; + mov.b64 %rd13792, {%r9815, %r9814}; + { + .reg .b32 %dummy; + mov.b64 {%r9816,%dummy}, %rd13761; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9817}, %rd13761; + } + shf.r.wrap.b32 %r9818, %r9817, %r9816, 25; + shf.r.wrap.b32 %r9819, %r9816, %r9817, 25; + mov.b64 %rd13793, {%r9819, %r9818}; + { + .reg .b32 %dummy; + mov.b64 {%r9820,%dummy}, %rd13762; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9821}, %rd13762; + } + shf.r.wrap.b32 %r9822, %r9821, %r9820, 23; + shf.r.wrap.b32 %r9823, %r9820, %r9821, 23; + mov.b64 %rd13794, {%r9823, %r9822}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9824}, %rd13768; + } + { + .reg .b32 %dummy; + mov.b64 {%r9825,%dummy}, %rd13768; + } + shf.l.wrap.b32 %r9826, %r9825, %r9824, 2; + shf.l.wrap.b32 %r9827, %r9824, %r9825, 2; + mov.b64 %rd13795, {%r9827, %r9826}; + not.b64 %rd13796, %rd13772; + and.b64 %rd13797, %rd13773, %rd13796; + xor.b64 %rd13798, %rd13797, %rd13747; + not.b64 %rd13799, %rd13773; + and.b64 %rd13800, %rd13774, %rd13799; + xor.b64 %rd13801, %rd13800, %rd13772; + not.b64 %rd13802, %rd13774; + and.b64 %rd13803, %rd13775, %rd13802; + xor.b64 %rd13804, %rd13803, %rd13773; + not.b64 %rd13805, %rd13775; + and.b64 %rd13806, %rd13747, %rd13805; + xor.b64 %rd13807, %rd13806, %rd13774; + not.b64 %rd13808, %rd13747; + and.b64 %rd13809, %rd13772, %rd13808; + xor.b64 %rd13810, %rd13809, %rd13775; + not.b64 %rd13811, %rd13777; + and.b64 %rd13812, %rd13778, %rd13811; + xor.b64 %rd13813, %rd13812, %rd13776; + not.b64 %rd13814, %rd13778; + and.b64 %rd13815, %rd13779, %rd13814; + xor.b64 %rd13816, %rd13815, %rd13777; + not.b64 %rd13817, %rd13779; + and.b64 %rd13818, %rd13780, %rd13817; + xor.b64 %rd13819, %rd13818, %rd13778; + not.b64 %rd13820, %rd13780; + and.b64 %rd13821, %rd13776, %rd13820; + xor.b64 %rd13822, %rd13821, %rd13779; + not.b64 %rd13823, %rd13776; + and.b64 %rd13824, %rd13777, %rd13823; + xor.b64 %rd13825, %rd13824, %rd13780; + not.b64 %rd13826, %rd13782; + and.b64 %rd13827, %rd13783, %rd13826; + xor.b64 %rd13828, %rd13827, %rd13781; + not.b64 %rd13829, %rd13783; + and.b64 %rd13830, %rd13784, %rd13829; + xor.b64 %rd13831, %rd13830, %rd13782; + not.b64 %rd13832, %rd13784; + and.b64 %rd13833, %rd13785, %rd13832; + xor.b64 %rd13834, %rd13833, %rd13783; + not.b64 %rd13835, %rd13785; + and.b64 %rd13836, %rd13781, %rd13835; + xor.b64 %rd13837, %rd13836, %rd13784; + not.b64 %rd13838, %rd13781; + and.b64 %rd13839, %rd13782, %rd13838; + xor.b64 %rd13840, %rd13839, %rd13785; + not.b64 %rd13841, %rd13787; + and.b64 %rd13842, %rd13788, %rd13841; + xor.b64 %rd13843, %rd13842, %rd13786; + not.b64 %rd13844, %rd13788; + and.b64 %rd13845, %rd13789, %rd13844; + xor.b64 %rd13846, %rd13845, %rd13787; + not.b64 %rd13847, %rd13789; + and.b64 %rd13848, %rd13790, %rd13847; + xor.b64 %rd13849, %rd13848, %rd13788; + not.b64 %rd13850, %rd13790; + and.b64 %rd13851, %rd13786, %rd13850; + xor.b64 %rd13852, %rd13851, %rd13789; + not.b64 %rd13853, %rd13786; + and.b64 %rd13854, %rd13787, %rd13853; + xor.b64 %rd13855, %rd13854, %rd13790; + not.b64 %rd13856, %rd13792; + and.b64 %rd13857, %rd13793, %rd13856; + xor.b64 %rd13858, %rd13857, %rd13791; + not.b64 %rd13859, %rd13793; + and.b64 %rd13860, %rd13794, %rd13859; + xor.b64 %rd13861, %rd13860, %rd13792; + not.b64 %rd13862, %rd13794; + and.b64 %rd13863, %rd13795, %rd13862; + xor.b64 %rd13864, %rd13863, %rd13793; + not.b64 %rd13865, %rd13795; + and.b64 %rd13866, %rd13791, %rd13865; + xor.b64 %rd13867, %rd13866, %rd13794; + not.b64 %rd13868, %rd13791; + and.b64 %rd13869, %rd13792, %rd13868; + xor.b64 %rd13870, %rd13869, %rd13795; + xor.b64 %rd13871, %rd13798, %rd27; + xor.b64 %rd13872, %rd13813, %rd13871; + xor.b64 %rd13873, %rd13872, %rd13828; + xor.b64 %rd13874, %rd13873, %rd13843; + xor.b64 %rd13875, %rd13874, %rd13858; + xor.b64 %rd13876, %rd13816, %rd13801; + xor.b64 %rd13877, %rd13876, %rd13831; + xor.b64 %rd13878, %rd13877, %rd13846; + xor.b64 %rd13879, %rd13878, %rd13861; + xor.b64 %rd13880, %rd13819, %rd13804; + xor.b64 %rd13881, %rd13880, %rd13834; + xor.b64 %rd13882, %rd13881, %rd13849; + xor.b64 %rd13883, %rd13882, %rd13864; + xor.b64 %rd13884, %rd13822, %rd13807; + xor.b64 %rd13885, %rd13884, %rd13837; + xor.b64 %rd13886, %rd13885, %rd13852; + xor.b64 %rd13887, %rd13886, %rd13867; + xor.b64 %rd13888, %rd13825, %rd13810; + xor.b64 %rd13889, %rd13888, %rd13840; + xor.b64 %rd13890, %rd13889, %rd13855; + xor.b64 %rd13891, %rd13890, %rd13870; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9828}, %rd13879; + } + { + .reg .b32 %dummy; + mov.b64 {%r9829,%dummy}, %rd13879; + } + shf.l.wrap.b32 %r9830, %r9829, %r9828, 1; + shf.l.wrap.b32 %r9831, %r9828, %r9829, 1; + mov.b64 %rd13892, {%r9831, %r9830}; + xor.b64 %rd13893, %rd13891, %rd13892; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9832}, %rd13883; + } + { + .reg .b32 %dummy; + mov.b64 {%r9833,%dummy}, %rd13883; + } + shf.l.wrap.b32 %r9834, %r9833, %r9832, 1; + shf.l.wrap.b32 %r9835, %r9832, %r9833, 1; + mov.b64 %rd13894, {%r9835, %r9834}; + xor.b64 %rd13895, %rd13894, %rd13875; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9836}, %rd13887; + } + { + .reg .b32 %dummy; + mov.b64 {%r9837,%dummy}, %rd13887; + } + shf.l.wrap.b32 %r9838, %r9837, %r9836, 1; + shf.l.wrap.b32 %r9839, %r9836, %r9837, 1; + mov.b64 %rd13896, {%r9839, %r9838}; + xor.b64 %rd13897, %rd13896, %rd13879; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9840}, %rd13891; + } + { + .reg .b32 %dummy; + mov.b64 {%r9841,%dummy}, %rd13891; + } + shf.l.wrap.b32 %r9842, %r9841, %r9840, 1; + shf.l.wrap.b32 %r9843, %r9840, %r9841, 1; + mov.b64 %rd13898, {%r9843, %r9842}; + xor.b64 %rd13899, %rd13898, %rd13883; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9844}, %rd13875; + } + { + .reg .b32 %dummy; + mov.b64 {%r9845,%dummy}, %rd13875; + } + shf.l.wrap.b32 %r9846, %r9845, %r9844, 1; + shf.l.wrap.b32 %r9847, %r9844, %r9845, 1; + mov.b64 %rd13900, {%r9847, %r9846}; + xor.b64 %rd13901, %rd13887, %rd13900; + xor.b64 %rd13902, %rd13893, %rd13871; + xor.b64 %rd13903, %rd13895, %rd13801; + xor.b64 %rd13904, %rd13897, %rd13804; + xor.b64 %rd13905, %rd13899, %rd13807; + xor.b64 %rd13906, %rd13901, %rd13810; + xor.b64 %rd13907, %rd13893, %rd13813; + xor.b64 %rd13908, %rd13895, %rd13816; + xor.b64 %rd13909, %rd13897, %rd13819; + xor.b64 %rd13910, %rd13899, %rd13822; + xor.b64 %rd13911, %rd13901, %rd13825; + xor.b64 %rd13912, %rd13893, %rd13828; + xor.b64 %rd13913, %rd13895, %rd13831; + xor.b64 %rd13914, %rd13897, %rd13834; + xor.b64 %rd13915, %rd13899, %rd13837; + xor.b64 %rd13916, %rd13901, %rd13840; + xor.b64 %rd13917, %rd13893, %rd13843; + xor.b64 %rd13918, %rd13895, %rd13846; + xor.b64 %rd13919, %rd13897, %rd13849; + xor.b64 %rd13920, %rd13899, %rd13852; + xor.b64 %rd13921, %rd13901, %rd13855; + xor.b64 %rd13922, %rd13893, %rd13858; + xor.b64 %rd13923, %rd13895, %rd13861; + xor.b64 %rd13924, %rd13897, %rd13864; + xor.b64 %rd13925, %rd13899, %rd13867; + xor.b64 %rd13926, %rd13901, %rd13870; + { + .reg .b32 %dummy; + mov.b64 {%r9848,%dummy}, %rd13908; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9849}, %rd13908; + } + shf.r.wrap.b32 %r9850, %r9849, %r9848, 20; + shf.r.wrap.b32 %r9851, %r9848, %r9849, 20; + mov.b64 %rd13927, {%r9851, %r9850}; + { + .reg .b32 %dummy; + mov.b64 {%r9852,%dummy}, %rd13914; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9853}, %rd13914; + } + shf.r.wrap.b32 %r9854, %r9853, %r9852, 21; + shf.r.wrap.b32 %r9855, %r9852, %r9853, 21; + mov.b64 %rd13928, {%r9855, %r9854}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9856}, %rd13920; + } + { + .reg .b32 %dummy; + mov.b64 {%r9857,%dummy}, %rd13920; + } + shf.l.wrap.b32 %r9858, %r9857, %r9856, 21; + shf.l.wrap.b32 %r9859, %r9856, %r9857, 21; + mov.b64 %rd13929, {%r9859, %r9858}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9860}, %rd13926; + } + { + .reg .b32 %dummy; + mov.b64 {%r9861,%dummy}, %rd13926; + } + shf.l.wrap.b32 %r9862, %r9861, %r9860, 14; + shf.l.wrap.b32 %r9863, %r9860, %r9861, 14; + mov.b64 %rd13930, {%r9863, %r9862}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9864}, %rd13905; + } + { + .reg .b32 %dummy; + mov.b64 {%r9865,%dummy}, %rd13905; + } + shf.l.wrap.b32 %r9866, %r9865, %r9864, 28; + shf.l.wrap.b32 %r9867, %r9864, %r9865, 28; + mov.b64 %rd13931, {%r9867, %r9866}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9868}, %rd13911; + } + { + .reg .b32 %dummy; + mov.b64 {%r9869,%dummy}, %rd13911; + } + shf.l.wrap.b32 %r9870, %r9869, %r9868, 20; + shf.l.wrap.b32 %r9871, %r9868, %r9869, 20; + mov.b64 %rd13932, {%r9871, %r9870}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9872}, %rd13912; + } + { + .reg .b32 %dummy; + mov.b64 {%r9873,%dummy}, %rd13912; + } + shf.l.wrap.b32 %r9874, %r9873, %r9872, 3; + shf.l.wrap.b32 %r9875, %r9872, %r9873, 3; + mov.b64 %rd13933, {%r9875, %r9874}; + { + .reg .b32 %dummy; + mov.b64 {%r9876,%dummy}, %rd13918; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9877}, %rd13918; + } + shf.r.wrap.b32 %r9878, %r9877, %r9876, 19; + shf.r.wrap.b32 %r9879, %r9876, %r9877, 19; + mov.b64 %rd13934, {%r9879, %r9878}; + { + .reg .b32 %dummy; + mov.b64 {%r9880,%dummy}, %rd13924; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9881}, %rd13924; + } + shf.r.wrap.b32 %r9882, %r9881, %r9880, 3; + shf.r.wrap.b32 %r9883, %r9880, %r9881, 3; + mov.b64 %rd13935, {%r9883, %r9882}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9884}, %rd13903; + } + { + .reg .b32 %dummy; + mov.b64 {%r9885,%dummy}, %rd13903; + } + shf.l.wrap.b32 %r9886, %r9885, %r9884, 1; + shf.l.wrap.b32 %r9887, %r9884, %r9885, 1; + mov.b64 %rd13936, {%r9887, %r9886}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9888}, %rd13909; + } + { + .reg .b32 %dummy; + mov.b64 {%r9889,%dummy}, %rd13909; + } + shf.l.wrap.b32 %r9890, %r9889, %r9888, 6; + shf.l.wrap.b32 %r9891, %r9888, %r9889, 6; + mov.b64 %rd13937, {%r9891, %r9890}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9892}, %rd13915; + } + { + .reg .b32 %dummy; + mov.b64 {%r9893,%dummy}, %rd13915; + } + shf.l.wrap.b32 %r9894, %r9893, %r9892, 25; + shf.l.wrap.b32 %r9895, %r9892, %r9893, 25; + mov.b64 %rd13938, {%r9895, %r9894}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9896}, %rd13921; + } + { + .reg .b32 %dummy; + mov.b64 {%r9897,%dummy}, %rd13921; + } + shf.l.wrap.b32 %r9898, %r9897, %r9896, 8; + shf.l.wrap.b32 %r9899, %r9896, %r9897, 8; + mov.b64 %rd13939, {%r9899, %r9898}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9900}, %rd13922; + } + { + .reg .b32 %dummy; + mov.b64 {%r9901,%dummy}, %rd13922; + } + shf.l.wrap.b32 %r9902, %r9901, %r9900, 18; + shf.l.wrap.b32 %r9903, %r9900, %r9901, 18; + mov.b64 %rd13940, {%r9903, %r9902}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9904}, %rd13906; + } + { + .reg .b32 %dummy; + mov.b64 {%r9905,%dummy}, %rd13906; + } + shf.l.wrap.b32 %r9906, %r9905, %r9904, 27; + shf.l.wrap.b32 %r9907, %r9904, %r9905, 27; + mov.b64 %rd13941, {%r9907, %r9906}; + { + .reg .b32 %dummy; + mov.b64 {%r9908,%dummy}, %rd13907; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9909}, %rd13907; + } + shf.r.wrap.b32 %r9910, %r9909, %r9908, 28; + shf.r.wrap.b32 %r9911, %r9908, %r9909, 28; + mov.b64 %rd13942, {%r9911, %r9910}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9912}, %rd13913; + } + { + .reg .b32 %dummy; + mov.b64 {%r9913,%dummy}, %rd13913; + } + shf.l.wrap.b32 %r9914, %r9913, %r9912, 10; + shf.l.wrap.b32 %r9915, %r9912, %r9913, 10; + mov.b64 %rd13943, {%r9915, %r9914}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9916}, %rd13919; + } + { + .reg .b32 %dummy; + mov.b64 {%r9917,%dummy}, %rd13919; + } + shf.l.wrap.b32 %r9918, %r9917, %r9916, 15; + shf.l.wrap.b32 %r9919, %r9916, %r9917, 15; + mov.b64 %rd13944, {%r9919, %r9918}; + { + .reg .b32 %dummy; + mov.b64 {%r9920,%dummy}, %rd13925; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9921}, %rd13925; + } + shf.r.wrap.b32 %r9922, %r9921, %r9920, 8; + shf.r.wrap.b32 %r9923, %r9920, %r9921, 8; + mov.b64 %rd13945, {%r9923, %r9922}; + { + .reg .b32 %dummy; + mov.b64 {%r9924,%dummy}, %rd13904; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9925}, %rd13904; + } + shf.r.wrap.b32 %r9926, %r9925, %r9924, 2; + shf.r.wrap.b32 %r9927, %r9924, %r9925, 2; + mov.b64 %rd13946, {%r9927, %r9926}; + { + .reg .b32 %dummy; + mov.b64 {%r9928,%dummy}, %rd13910; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9929}, %rd13910; + } + shf.r.wrap.b32 %r9930, %r9929, %r9928, 9; + shf.r.wrap.b32 %r9931, %r9928, %r9929, 9; + mov.b64 %rd13947, {%r9931, %r9930}; + { + .reg .b32 %dummy; + mov.b64 {%r9932,%dummy}, %rd13916; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9933}, %rd13916; + } + shf.r.wrap.b32 %r9934, %r9933, %r9932, 25; + shf.r.wrap.b32 %r9935, %r9932, %r9933, 25; + mov.b64 %rd13948, {%r9935, %r9934}; + { + .reg .b32 %dummy; + mov.b64 {%r9936,%dummy}, %rd13917; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9937}, %rd13917; + } + shf.r.wrap.b32 %r9938, %r9937, %r9936, 23; + shf.r.wrap.b32 %r9939, %r9936, %r9937, 23; + mov.b64 %rd13949, {%r9939, %r9938}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9940}, %rd13923; + } + { + .reg .b32 %dummy; + mov.b64 {%r9941,%dummy}, %rd13923; + } + shf.l.wrap.b32 %r9942, %r9941, %r9940, 2; + shf.l.wrap.b32 %r9943, %r9940, %r9941, 2; + mov.b64 %rd13950, {%r9943, %r9942}; + not.b64 %rd13951, %rd13927; + and.b64 %rd13952, %rd13928, %rd13951; + xor.b64 %rd13953, %rd13952, %rd13902; + not.b64 %rd13954, %rd13928; + and.b64 %rd13955, %rd13929, %rd13954; + xor.b64 %rd13956, %rd13955, %rd13927; + not.b64 %rd13957, %rd13929; + and.b64 %rd13958, %rd13930, %rd13957; + xor.b64 %rd13959, %rd13958, %rd13928; + not.b64 %rd13960, %rd13930; + and.b64 %rd13961, %rd13902, %rd13960; + xor.b64 %rd13962, %rd13961, %rd13929; + not.b64 %rd13963, %rd13902; + and.b64 %rd13964, %rd13927, %rd13963; + xor.b64 %rd13965, %rd13964, %rd13930; + not.b64 %rd13966, %rd13932; + and.b64 %rd13967, %rd13933, %rd13966; + xor.b64 %rd13968, %rd13967, %rd13931; + not.b64 %rd13969, %rd13933; + and.b64 %rd13970, %rd13934, %rd13969; + xor.b64 %rd13971, %rd13970, %rd13932; + not.b64 %rd13972, %rd13934; + and.b64 %rd13973, %rd13935, %rd13972; + xor.b64 %rd13974, %rd13973, %rd13933; + not.b64 %rd13975, %rd13935; + and.b64 %rd13976, %rd13931, %rd13975; + xor.b64 %rd13977, %rd13976, %rd13934; + not.b64 %rd13978, %rd13931; + and.b64 %rd13979, %rd13932, %rd13978; + xor.b64 %rd13980, %rd13979, %rd13935; + not.b64 %rd13981, %rd13937; + and.b64 %rd13982, %rd13938, %rd13981; + xor.b64 %rd13983, %rd13982, %rd13936; + not.b64 %rd13984, %rd13938; + and.b64 %rd13985, %rd13939, %rd13984; + xor.b64 %rd13986, %rd13985, %rd13937; + not.b64 %rd13987, %rd13939; + and.b64 %rd13988, %rd13940, %rd13987; + xor.b64 %rd13989, %rd13988, %rd13938; + not.b64 %rd13990, %rd13940; + and.b64 %rd13991, %rd13936, %rd13990; + xor.b64 %rd13992, %rd13991, %rd13939; + not.b64 %rd13993, %rd13936; + and.b64 %rd13994, %rd13937, %rd13993; + xor.b64 %rd13995, %rd13994, %rd13940; + not.b64 %rd13996, %rd13942; + and.b64 %rd13997, %rd13943, %rd13996; + xor.b64 %rd13998, %rd13997, %rd13941; + not.b64 %rd13999, %rd13943; + and.b64 %rd14000, %rd13944, %rd13999; + xor.b64 %rd14001, %rd14000, %rd13942; + not.b64 %rd14002, %rd13944; + and.b64 %rd14003, %rd13945, %rd14002; + xor.b64 %rd14004, %rd14003, %rd13943; + not.b64 %rd14005, %rd13945; + and.b64 %rd14006, %rd13941, %rd14005; + xor.b64 %rd14007, %rd14006, %rd13944; + not.b64 %rd14008, %rd13941; + and.b64 %rd14009, %rd13942, %rd14008; + xor.b64 %rd14010, %rd14009, %rd13945; + not.b64 %rd14011, %rd13947; + and.b64 %rd14012, %rd13948, %rd14011; + xor.b64 %rd14013, %rd14012, %rd13946; + not.b64 %rd14014, %rd13948; + and.b64 %rd14015, %rd13949, %rd14014; + xor.b64 %rd14016, %rd14015, %rd13947; + not.b64 %rd14017, %rd13949; + and.b64 %rd14018, %rd13950, %rd14017; + xor.b64 %rd14019, %rd14018, %rd13948; + not.b64 %rd14020, %rd13950; + and.b64 %rd14021, %rd13946, %rd14020; + xor.b64 %rd14022, %rd14021, %rd13949; + not.b64 %rd14023, %rd13946; + and.b64 %rd14024, %rd13947, %rd14023; + xor.b64 %rd14025, %rd14024, %rd13950; + xor.b64 %rd14026, %rd13953, %rd28; + xor.b64 %rd14027, %rd13968, %rd14026; + xor.b64 %rd14028, %rd14027, %rd13983; + xor.b64 %rd14029, %rd14028, %rd13998; + xor.b64 %rd14030, %rd14029, %rd14013; + xor.b64 %rd14031, %rd13971, %rd13956; + xor.b64 %rd14032, %rd14031, %rd13986; + xor.b64 %rd14033, %rd14032, %rd14001; + xor.b64 %rd14034, %rd14033, %rd14016; + xor.b64 %rd14035, %rd13974, %rd13959; + xor.b64 %rd14036, %rd14035, %rd13989; + xor.b64 %rd14037, %rd14036, %rd14004; + xor.b64 %rd14038, %rd14037, %rd14019; + xor.b64 %rd14039, %rd13977, %rd13962; + xor.b64 %rd14040, %rd14039, %rd13992; + xor.b64 %rd14041, %rd14040, %rd14007; + xor.b64 %rd14042, %rd14041, %rd14022; + xor.b64 %rd14043, %rd13980, %rd13965; + xor.b64 %rd14044, %rd14043, %rd13995; + xor.b64 %rd14045, %rd14044, %rd14010; + xor.b64 %rd14046, %rd14045, %rd14025; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9944}, %rd14034; + } + { + .reg .b32 %dummy; + mov.b64 {%r9945,%dummy}, %rd14034; + } + shf.l.wrap.b32 %r9946, %r9945, %r9944, 1; + shf.l.wrap.b32 %r9947, %r9944, %r9945, 1; + mov.b64 %rd14047, {%r9947, %r9946}; + xor.b64 %rd14048, %rd14046, %rd14047; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9948}, %rd14038; + } + { + .reg .b32 %dummy; + mov.b64 {%r9949,%dummy}, %rd14038; + } + shf.l.wrap.b32 %r9950, %r9949, %r9948, 1; + shf.l.wrap.b32 %r9951, %r9948, %r9949, 1; + mov.b64 %rd14049, {%r9951, %r9950}; + xor.b64 %rd14050, %rd14049, %rd14030; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9952}, %rd14042; + } + { + .reg .b32 %dummy; + mov.b64 {%r9953,%dummy}, %rd14042; + } + shf.l.wrap.b32 %r9954, %r9953, %r9952, 1; + shf.l.wrap.b32 %r9955, %r9952, %r9953, 1; + mov.b64 %rd14051, {%r9955, %r9954}; + xor.b64 %rd14052, %rd14051, %rd14034; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9956}, %rd14046; + } + { + .reg .b32 %dummy; + mov.b64 {%r9957,%dummy}, %rd14046; + } + shf.l.wrap.b32 %r9958, %r9957, %r9956, 1; + shf.l.wrap.b32 %r9959, %r9956, %r9957, 1; + mov.b64 %rd14053, {%r9959, %r9958}; + xor.b64 %rd14054, %rd14053, %rd14038; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9960}, %rd14030; + } + { + .reg .b32 %dummy; + mov.b64 {%r9961,%dummy}, %rd14030; + } + shf.l.wrap.b32 %r9962, %r9961, %r9960, 1; + shf.l.wrap.b32 %r9963, %r9960, %r9961, 1; + mov.b64 %rd14055, {%r9963, %r9962}; + xor.b64 %rd14056, %rd14042, %rd14055; + xor.b64 %rd14057, %rd14048, %rd14026; + xor.b64 %rd14058, %rd14050, %rd13956; + xor.b64 %rd14059, %rd14052, %rd13959; + xor.b64 %rd14060, %rd14054, %rd13962; + xor.b64 %rd14061, %rd14056, %rd13965; + xor.b64 %rd14062, %rd14048, %rd13968; + xor.b64 %rd14063, %rd14050, %rd13971; + xor.b64 %rd14064, %rd14052, %rd13974; + xor.b64 %rd14065, %rd14054, %rd13977; + xor.b64 %rd14066, %rd14056, %rd13980; + xor.b64 %rd14067, %rd14048, %rd13983; + xor.b64 %rd14068, %rd14050, %rd13986; + xor.b64 %rd14069, %rd14052, %rd13989; + xor.b64 %rd14070, %rd14054, %rd13992; + xor.b64 %rd14071, %rd14056, %rd13995; + xor.b64 %rd14072, %rd14048, %rd13998; + xor.b64 %rd14073, %rd14050, %rd14001; + xor.b64 %rd14074, %rd14052, %rd14004; + xor.b64 %rd14075, %rd14054, %rd14007; + xor.b64 %rd14076, %rd14056, %rd14010; + xor.b64 %rd14077, %rd14048, %rd14013; + xor.b64 %rd14078, %rd14050, %rd14016; + xor.b64 %rd14079, %rd14052, %rd14019; + xor.b64 %rd14080, %rd14054, %rd14022; + xor.b64 %rd14081, %rd14056, %rd14025; + { + .reg .b32 %dummy; + mov.b64 {%r9964,%dummy}, %rd14063; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9965}, %rd14063; + } + shf.r.wrap.b32 %r9966, %r9965, %r9964, 20; + shf.r.wrap.b32 %r9967, %r9964, %r9965, 20; + mov.b64 %rd14082, {%r9967, %r9966}; + { + .reg .b32 %dummy; + mov.b64 {%r9968,%dummy}, %rd14069; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9969}, %rd14069; + } + shf.r.wrap.b32 %r9970, %r9969, %r9968, 21; + shf.r.wrap.b32 %r9971, %r9968, %r9969, 21; + mov.b64 %rd14083, {%r9971, %r9970}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9972}, %rd14075; + } + { + .reg .b32 %dummy; + mov.b64 {%r9973,%dummy}, %rd14075; + } + shf.l.wrap.b32 %r9974, %r9973, %r9972, 21; + shf.l.wrap.b32 %r9975, %r9972, %r9973, 21; + mov.b64 %rd14084, {%r9975, %r9974}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9976}, %rd14081; + } + { + .reg .b32 %dummy; + mov.b64 {%r9977,%dummy}, %rd14081; + } + shf.l.wrap.b32 %r9978, %r9977, %r9976, 14; + shf.l.wrap.b32 %r9979, %r9976, %r9977, 14; + mov.b64 %rd14085, {%r9979, %r9978}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9980}, %rd14060; + } + { + .reg .b32 %dummy; + mov.b64 {%r9981,%dummy}, %rd14060; + } + shf.l.wrap.b32 %r9982, %r9981, %r9980, 28; + shf.l.wrap.b32 %r9983, %r9980, %r9981, 28; + mov.b64 %rd14086, {%r9983, %r9982}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9984}, %rd14066; + } + { + .reg .b32 %dummy; + mov.b64 {%r9985,%dummy}, %rd14066; + } + shf.l.wrap.b32 %r9986, %r9985, %r9984, 20; + shf.l.wrap.b32 %r9987, %r9984, %r9985, 20; + mov.b64 %rd14087, {%r9987, %r9986}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9988}, %rd14067; + } + { + .reg .b32 %dummy; + mov.b64 {%r9989,%dummy}, %rd14067; + } + shf.l.wrap.b32 %r9990, %r9989, %r9988, 3; + shf.l.wrap.b32 %r9991, %r9988, %r9989, 3; + mov.b64 %rd14088, {%r9991, %r9990}; + { + .reg .b32 %dummy; + mov.b64 {%r9992,%dummy}, %rd14073; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9993}, %rd14073; + } + shf.r.wrap.b32 %r9994, %r9993, %r9992, 19; + shf.r.wrap.b32 %r9995, %r9992, %r9993, 19; + mov.b64 %rd14089, {%r9995, %r9994}; + { + .reg .b32 %dummy; + mov.b64 {%r9996,%dummy}, %rd14079; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r9997}, %rd14079; + } + shf.r.wrap.b32 %r9998, %r9997, %r9996, 3; + shf.r.wrap.b32 %r9999, %r9996, %r9997, 3; + mov.b64 %rd14090, {%r9999, %r9998}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10000}, %rd14058; + } + { + .reg .b32 %dummy; + mov.b64 {%r10001,%dummy}, %rd14058; + } + shf.l.wrap.b32 %r10002, %r10001, %r10000, 1; + shf.l.wrap.b32 %r10003, %r10000, %r10001, 1; + mov.b64 %rd14091, {%r10003, %r10002}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10004}, %rd14064; + } + { + .reg .b32 %dummy; + mov.b64 {%r10005,%dummy}, %rd14064; + } + shf.l.wrap.b32 %r10006, %r10005, %r10004, 6; + shf.l.wrap.b32 %r10007, %r10004, %r10005, 6; + mov.b64 %rd14092, {%r10007, %r10006}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10008}, %rd14070; + } + { + .reg .b32 %dummy; + mov.b64 {%r10009,%dummy}, %rd14070; + } + shf.l.wrap.b32 %r10010, %r10009, %r10008, 25; + shf.l.wrap.b32 %r10011, %r10008, %r10009, 25; + mov.b64 %rd14093, {%r10011, %r10010}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10012}, %rd14076; + } + { + .reg .b32 %dummy; + mov.b64 {%r10013,%dummy}, %rd14076; + } + shf.l.wrap.b32 %r10014, %r10013, %r10012, 8; + shf.l.wrap.b32 %r10015, %r10012, %r10013, 8; + mov.b64 %rd14094, {%r10015, %r10014}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10016}, %rd14077; + } + { + .reg .b32 %dummy; + mov.b64 {%r10017,%dummy}, %rd14077; + } + shf.l.wrap.b32 %r10018, %r10017, %r10016, 18; + shf.l.wrap.b32 %r10019, %r10016, %r10017, 18; + mov.b64 %rd14095, {%r10019, %r10018}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10020}, %rd14061; + } + { + .reg .b32 %dummy; + mov.b64 {%r10021,%dummy}, %rd14061; + } + shf.l.wrap.b32 %r10022, %r10021, %r10020, 27; + shf.l.wrap.b32 %r10023, %r10020, %r10021, 27; + mov.b64 %rd14096, {%r10023, %r10022}; + { + .reg .b32 %dummy; + mov.b64 {%r10024,%dummy}, %rd14062; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10025}, %rd14062; + } + shf.r.wrap.b32 %r10026, %r10025, %r10024, 28; + shf.r.wrap.b32 %r10027, %r10024, %r10025, 28; + mov.b64 %rd14097, {%r10027, %r10026}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10028}, %rd14068; + } + { + .reg .b32 %dummy; + mov.b64 {%r10029,%dummy}, %rd14068; + } + shf.l.wrap.b32 %r10030, %r10029, %r10028, 10; + shf.l.wrap.b32 %r10031, %r10028, %r10029, 10; + mov.b64 %rd14098, {%r10031, %r10030}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10032}, %rd14074; + } + { + .reg .b32 %dummy; + mov.b64 {%r10033,%dummy}, %rd14074; + } + shf.l.wrap.b32 %r10034, %r10033, %r10032, 15; + shf.l.wrap.b32 %r10035, %r10032, %r10033, 15; + mov.b64 %rd14099, {%r10035, %r10034}; + { + .reg .b32 %dummy; + mov.b64 {%r10036,%dummy}, %rd14080; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10037}, %rd14080; + } + shf.r.wrap.b32 %r10038, %r10037, %r10036, 8; + shf.r.wrap.b32 %r10039, %r10036, %r10037, 8; + mov.b64 %rd14100, {%r10039, %r10038}; + { + .reg .b32 %dummy; + mov.b64 {%r10040,%dummy}, %rd14059; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10041}, %rd14059; + } + shf.r.wrap.b32 %r10042, %r10041, %r10040, 2; + shf.r.wrap.b32 %r10043, %r10040, %r10041, 2; + mov.b64 %rd14101, {%r10043, %r10042}; + { + .reg .b32 %dummy; + mov.b64 {%r10044,%dummy}, %rd14065; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10045}, %rd14065; + } + shf.r.wrap.b32 %r10046, %r10045, %r10044, 9; + shf.r.wrap.b32 %r10047, %r10044, %r10045, 9; + mov.b64 %rd14102, {%r10047, %r10046}; + { + .reg .b32 %dummy; + mov.b64 {%r10048,%dummy}, %rd14071; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10049}, %rd14071; + } + shf.r.wrap.b32 %r10050, %r10049, %r10048, 25; + shf.r.wrap.b32 %r10051, %r10048, %r10049, 25; + mov.b64 %rd14103, {%r10051, %r10050}; + { + .reg .b32 %dummy; + mov.b64 {%r10052,%dummy}, %rd14072; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10053}, %rd14072; + } + shf.r.wrap.b32 %r10054, %r10053, %r10052, 23; + shf.r.wrap.b32 %r10055, %r10052, %r10053, 23; + mov.b64 %rd14104, {%r10055, %r10054}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10056}, %rd14078; + } + { + .reg .b32 %dummy; + mov.b64 {%r10057,%dummy}, %rd14078; + } + shf.l.wrap.b32 %r10058, %r10057, %r10056, 2; + shf.l.wrap.b32 %r10059, %r10056, %r10057, 2; + mov.b64 %rd14105, {%r10059, %r10058}; + not.b64 %rd14106, %rd14082; + and.b64 %rd14107, %rd14083, %rd14106; + xor.b64 %rd14108, %rd14107, %rd14057; + not.b64 %rd14109, %rd14083; + and.b64 %rd14110, %rd14084, %rd14109; + xor.b64 %rd14111, %rd14110, %rd14082; + not.b64 %rd14112, %rd14084; + and.b64 %rd14113, %rd14085, %rd14112; + xor.b64 %rd14114, %rd14113, %rd14083; + not.b64 %rd14115, %rd14085; + and.b64 %rd14116, %rd14057, %rd14115; + xor.b64 %rd14117, %rd14116, %rd14084; + not.b64 %rd14118, %rd14057; + and.b64 %rd14119, %rd14082, %rd14118; + xor.b64 %rd14120, %rd14119, %rd14085; + not.b64 %rd14121, %rd14087; + and.b64 %rd14122, %rd14088, %rd14121; + xor.b64 %rd14123, %rd14122, %rd14086; + not.b64 %rd14124, %rd14088; + and.b64 %rd14125, %rd14089, %rd14124; + xor.b64 %rd14126, %rd14125, %rd14087; + not.b64 %rd14127, %rd14089; + and.b64 %rd14128, %rd14090, %rd14127; + xor.b64 %rd14129, %rd14128, %rd14088; + not.b64 %rd14130, %rd14090; + and.b64 %rd14131, %rd14086, %rd14130; + xor.b64 %rd14132, %rd14131, %rd14089; + not.b64 %rd14133, %rd14086; + and.b64 %rd14134, %rd14087, %rd14133; + xor.b64 %rd14135, %rd14134, %rd14090; + not.b64 %rd14136, %rd14092; + and.b64 %rd14137, %rd14093, %rd14136; + xor.b64 %rd14138, %rd14137, %rd14091; + not.b64 %rd14139, %rd14093; + and.b64 %rd14140, %rd14094, %rd14139; + xor.b64 %rd14141, %rd14140, %rd14092; + not.b64 %rd14142, %rd14094; + and.b64 %rd14143, %rd14095, %rd14142; + xor.b64 %rd14144, %rd14143, %rd14093; + not.b64 %rd14145, %rd14095; + and.b64 %rd14146, %rd14091, %rd14145; + xor.b64 %rd14147, %rd14146, %rd14094; + not.b64 %rd14148, %rd14091; + and.b64 %rd14149, %rd14092, %rd14148; + xor.b64 %rd14150, %rd14149, %rd14095; + not.b64 %rd14151, %rd14097; + and.b64 %rd14152, %rd14098, %rd14151; + xor.b64 %rd14153, %rd14152, %rd14096; + not.b64 %rd14154, %rd14098; + and.b64 %rd14155, %rd14099, %rd14154; + xor.b64 %rd14156, %rd14155, %rd14097; + not.b64 %rd14157, %rd14099; + and.b64 %rd14158, %rd14100, %rd14157; + xor.b64 %rd14159, %rd14158, %rd14098; + not.b64 %rd14160, %rd14100; + and.b64 %rd14161, %rd14096, %rd14160; + xor.b64 %rd14162, %rd14161, %rd14099; + not.b64 %rd14163, %rd14096; + and.b64 %rd14164, %rd14097, %rd14163; + xor.b64 %rd14165, %rd14164, %rd14100; + not.b64 %rd14166, %rd14102; + and.b64 %rd14167, %rd14103, %rd14166; + xor.b64 %rd14168, %rd14167, %rd14101; + not.b64 %rd14169, %rd14103; + and.b64 %rd14170, %rd14104, %rd14169; + xor.b64 %rd14171, %rd14170, %rd14102; + not.b64 %rd14172, %rd14104; + and.b64 %rd14173, %rd14105, %rd14172; + xor.b64 %rd14174, %rd14173, %rd14103; + not.b64 %rd14175, %rd14105; + and.b64 %rd14176, %rd14101, %rd14175; + xor.b64 %rd14177, %rd14176, %rd14104; + not.b64 %rd14178, %rd14101; + and.b64 %rd14179, %rd14102, %rd14178; + xor.b64 %rd14180, %rd14179, %rd14105; + xor.b64 %rd14181, %rd14108, %rd29; + xor.b64 %rd14182, %rd14123, %rd14181; + xor.b64 %rd14183, %rd14182, %rd14138; + xor.b64 %rd14184, %rd14183, %rd14153; + xor.b64 %rd14185, %rd14184, %rd14168; + xor.b64 %rd14186, %rd14126, %rd14111; + xor.b64 %rd14187, %rd14186, %rd14141; + xor.b64 %rd14188, %rd14187, %rd14156; + xor.b64 %rd14189, %rd14188, %rd14171; + xor.b64 %rd14190, %rd14129, %rd14114; + xor.b64 %rd14191, %rd14190, %rd14144; + xor.b64 %rd14192, %rd14191, %rd14159; + xor.b64 %rd14193, %rd14192, %rd14174; + xor.b64 %rd14194, %rd14132, %rd14117; + xor.b64 %rd14195, %rd14194, %rd14147; + xor.b64 %rd14196, %rd14195, %rd14162; + xor.b64 %rd14197, %rd14196, %rd14177; + xor.b64 %rd14198, %rd14135, %rd14120; + xor.b64 %rd14199, %rd14198, %rd14150; + xor.b64 %rd14200, %rd14199, %rd14165; + xor.b64 %rd14201, %rd14200, %rd14180; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10060}, %rd14189; + } + { + .reg .b32 %dummy; + mov.b64 {%r10061,%dummy}, %rd14189; + } + shf.l.wrap.b32 %r10062, %r10061, %r10060, 1; + shf.l.wrap.b32 %r10063, %r10060, %r10061, 1; + mov.b64 %rd14202, {%r10063, %r10062}; + xor.b64 %rd14203, %rd14201, %rd14202; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10064}, %rd14193; + } + { + .reg .b32 %dummy; + mov.b64 {%r10065,%dummy}, %rd14193; + } + shf.l.wrap.b32 %r10066, %r10065, %r10064, 1; + shf.l.wrap.b32 %r10067, %r10064, %r10065, 1; + mov.b64 %rd14204, {%r10067, %r10066}; + xor.b64 %rd14205, %rd14204, %rd14185; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10068}, %rd14197; + } + { + .reg .b32 %dummy; + mov.b64 {%r10069,%dummy}, %rd14197; + } + shf.l.wrap.b32 %r10070, %r10069, %r10068, 1; + shf.l.wrap.b32 %r10071, %r10068, %r10069, 1; + mov.b64 %rd14206, {%r10071, %r10070}; + xor.b64 %rd14207, %rd14206, %rd14189; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10072}, %rd14201; + } + { + .reg .b32 %dummy; + mov.b64 {%r10073,%dummy}, %rd14201; + } + shf.l.wrap.b32 %r10074, %r10073, %r10072, 1; + shf.l.wrap.b32 %r10075, %r10072, %r10073, 1; + mov.b64 %rd14208, {%r10075, %r10074}; + xor.b64 %rd14209, %rd14208, %rd14193; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10076}, %rd14185; + } + { + .reg .b32 %dummy; + mov.b64 {%r10077,%dummy}, %rd14185; + } + shf.l.wrap.b32 %r10078, %r10077, %r10076, 1; + shf.l.wrap.b32 %r10079, %r10076, %r10077, 1; + mov.b64 %rd14210, {%r10079, %r10078}; + xor.b64 %rd14211, %rd14197, %rd14210; + xor.b64 %rd14212, %rd14203, %rd14181; + xor.b64 %rd14213, %rd14205, %rd14111; + xor.b64 %rd14214, %rd14207, %rd14114; + xor.b64 %rd14215, %rd14209, %rd14117; + xor.b64 %rd14216, %rd14211, %rd14120; + xor.b64 %rd14217, %rd14203, %rd14123; + xor.b64 %rd14218, %rd14205, %rd14126; + xor.b64 %rd14219, %rd14207, %rd14129; + xor.b64 %rd14220, %rd14209, %rd14132; + xor.b64 %rd14221, %rd14211, %rd14135; + xor.b64 %rd14222, %rd14203, %rd14138; + xor.b64 %rd14223, %rd14205, %rd14141; + xor.b64 %rd14224, %rd14207, %rd14144; + xor.b64 %rd14225, %rd14209, %rd14147; + xor.b64 %rd14226, %rd14211, %rd14150; + xor.b64 %rd14227, %rd14203, %rd14153; + xor.b64 %rd14228, %rd14205, %rd14156; + xor.b64 %rd14229, %rd14207, %rd14159; + xor.b64 %rd14230, %rd14209, %rd14162; + xor.b64 %rd14231, %rd14211, %rd14165; + xor.b64 %rd14232, %rd14203, %rd14168; + xor.b64 %rd14233, %rd14205, %rd14171; + xor.b64 %rd14234, %rd14207, %rd14174; + xor.b64 %rd14235, %rd14209, %rd14177; + xor.b64 %rd14236, %rd14211, %rd14180; + { + .reg .b32 %dummy; + mov.b64 {%r10080,%dummy}, %rd14218; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10081}, %rd14218; + } + shf.r.wrap.b32 %r10082, %r10081, %r10080, 20; + shf.r.wrap.b32 %r10083, %r10080, %r10081, 20; + mov.b64 %rd14237, {%r10083, %r10082}; + { + .reg .b32 %dummy; + mov.b64 {%r10084,%dummy}, %rd14224; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10085}, %rd14224; + } + shf.r.wrap.b32 %r10086, %r10085, %r10084, 21; + shf.r.wrap.b32 %r10087, %r10084, %r10085, 21; + mov.b64 %rd14238, {%r10087, %r10086}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10088}, %rd14230; + } + { + .reg .b32 %dummy; + mov.b64 {%r10089,%dummy}, %rd14230; + } + shf.l.wrap.b32 %r10090, %r10089, %r10088, 21; + shf.l.wrap.b32 %r10091, %r10088, %r10089, 21; + mov.b64 %rd14239, {%r10091, %r10090}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10092}, %rd14236; + } + { + .reg .b32 %dummy; + mov.b64 {%r10093,%dummy}, %rd14236; + } + shf.l.wrap.b32 %r10094, %r10093, %r10092, 14; + shf.l.wrap.b32 %r10095, %r10092, %r10093, 14; + mov.b64 %rd14240, {%r10095, %r10094}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10096}, %rd14215; + } + { + .reg .b32 %dummy; + mov.b64 {%r10097,%dummy}, %rd14215; + } + shf.l.wrap.b32 %r10098, %r10097, %r10096, 28; + shf.l.wrap.b32 %r10099, %r10096, %r10097, 28; + mov.b64 %rd14241, {%r10099, %r10098}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10100}, %rd14221; + } + { + .reg .b32 %dummy; + mov.b64 {%r10101,%dummy}, %rd14221; + } + shf.l.wrap.b32 %r10102, %r10101, %r10100, 20; + shf.l.wrap.b32 %r10103, %r10100, %r10101, 20; + mov.b64 %rd14242, {%r10103, %r10102}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10104}, %rd14222; + } + { + .reg .b32 %dummy; + mov.b64 {%r10105,%dummy}, %rd14222; + } + shf.l.wrap.b32 %r10106, %r10105, %r10104, 3; + shf.l.wrap.b32 %r10107, %r10104, %r10105, 3; + mov.b64 %rd14243, {%r10107, %r10106}; + { + .reg .b32 %dummy; + mov.b64 {%r10108,%dummy}, %rd14228; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10109}, %rd14228; + } + shf.r.wrap.b32 %r10110, %r10109, %r10108, 19; + shf.r.wrap.b32 %r10111, %r10108, %r10109, 19; + mov.b64 %rd14244, {%r10111, %r10110}; + { + .reg .b32 %dummy; + mov.b64 {%r10112,%dummy}, %rd14234; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10113}, %rd14234; + } + shf.r.wrap.b32 %r10114, %r10113, %r10112, 3; + shf.r.wrap.b32 %r10115, %r10112, %r10113, 3; + mov.b64 %rd14245, {%r10115, %r10114}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10116}, %rd14213; + } + { + .reg .b32 %dummy; + mov.b64 {%r10117,%dummy}, %rd14213; + } + shf.l.wrap.b32 %r10118, %r10117, %r10116, 1; + shf.l.wrap.b32 %r10119, %r10116, %r10117, 1; + mov.b64 %rd14246, {%r10119, %r10118}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10120}, %rd14219; + } + { + .reg .b32 %dummy; + mov.b64 {%r10121,%dummy}, %rd14219; + } + shf.l.wrap.b32 %r10122, %r10121, %r10120, 6; + shf.l.wrap.b32 %r10123, %r10120, %r10121, 6; + mov.b64 %rd14247, {%r10123, %r10122}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10124}, %rd14225; + } + { + .reg .b32 %dummy; + mov.b64 {%r10125,%dummy}, %rd14225; + } + shf.l.wrap.b32 %r10126, %r10125, %r10124, 25; + shf.l.wrap.b32 %r10127, %r10124, %r10125, 25; + mov.b64 %rd14248, {%r10127, %r10126}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10128}, %rd14231; + } + { + .reg .b32 %dummy; + mov.b64 {%r10129,%dummy}, %rd14231; + } + shf.l.wrap.b32 %r10130, %r10129, %r10128, 8; + shf.l.wrap.b32 %r10131, %r10128, %r10129, 8; + mov.b64 %rd14249, {%r10131, %r10130}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10132}, %rd14232; + } + { + .reg .b32 %dummy; + mov.b64 {%r10133,%dummy}, %rd14232; + } + shf.l.wrap.b32 %r10134, %r10133, %r10132, 18; + shf.l.wrap.b32 %r10135, %r10132, %r10133, 18; + mov.b64 %rd14250, {%r10135, %r10134}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10136}, %rd14216; + } + { + .reg .b32 %dummy; + mov.b64 {%r10137,%dummy}, %rd14216; + } + shf.l.wrap.b32 %r10138, %r10137, %r10136, 27; + shf.l.wrap.b32 %r10139, %r10136, %r10137, 27; + mov.b64 %rd14251, {%r10139, %r10138}; + { + .reg .b32 %dummy; + mov.b64 {%r10140,%dummy}, %rd14217; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10141}, %rd14217; + } + shf.r.wrap.b32 %r10142, %r10141, %r10140, 28; + shf.r.wrap.b32 %r10143, %r10140, %r10141, 28; + mov.b64 %rd14252, {%r10143, %r10142}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10144}, %rd14223; + } + { + .reg .b32 %dummy; + mov.b64 {%r10145,%dummy}, %rd14223; + } + shf.l.wrap.b32 %r10146, %r10145, %r10144, 10; + shf.l.wrap.b32 %r10147, %r10144, %r10145, 10; + mov.b64 %rd14253, {%r10147, %r10146}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10148}, %rd14229; + } + { + .reg .b32 %dummy; + mov.b64 {%r10149,%dummy}, %rd14229; + } + shf.l.wrap.b32 %r10150, %r10149, %r10148, 15; + shf.l.wrap.b32 %r10151, %r10148, %r10149, 15; + mov.b64 %rd14254, {%r10151, %r10150}; + { + .reg .b32 %dummy; + mov.b64 {%r10152,%dummy}, %rd14235; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10153}, %rd14235; + } + shf.r.wrap.b32 %r10154, %r10153, %r10152, 8; + shf.r.wrap.b32 %r10155, %r10152, %r10153, 8; + mov.b64 %rd14255, {%r10155, %r10154}; + { + .reg .b32 %dummy; + mov.b64 {%r10156,%dummy}, %rd14214; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10157}, %rd14214; + } + shf.r.wrap.b32 %r10158, %r10157, %r10156, 2; + shf.r.wrap.b32 %r10159, %r10156, %r10157, 2; + mov.b64 %rd14256, {%r10159, %r10158}; + { + .reg .b32 %dummy; + mov.b64 {%r10160,%dummy}, %rd14220; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10161}, %rd14220; + } + shf.r.wrap.b32 %r10162, %r10161, %r10160, 9; + shf.r.wrap.b32 %r10163, %r10160, %r10161, 9; + mov.b64 %rd14257, {%r10163, %r10162}; + { + .reg .b32 %dummy; + mov.b64 {%r10164,%dummy}, %rd14226; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10165}, %rd14226; + } + shf.r.wrap.b32 %r10166, %r10165, %r10164, 25; + shf.r.wrap.b32 %r10167, %r10164, %r10165, 25; + mov.b64 %rd14258, {%r10167, %r10166}; + { + .reg .b32 %dummy; + mov.b64 {%r10168,%dummy}, %rd14227; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10169}, %rd14227; + } + shf.r.wrap.b32 %r10170, %r10169, %r10168, 23; + shf.r.wrap.b32 %r10171, %r10168, %r10169, 23; + mov.b64 %rd14259, {%r10171, %r10170}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10172}, %rd14233; + } + { + .reg .b32 %dummy; + mov.b64 {%r10173,%dummy}, %rd14233; + } + shf.l.wrap.b32 %r10174, %r10173, %r10172, 2; + shf.l.wrap.b32 %r10175, %r10172, %r10173, 2; + mov.b64 %rd14260, {%r10175, %r10174}; + not.b64 %rd14261, %rd14237; + and.b64 %rd14262, %rd14238, %rd14261; + xor.b64 %rd14263, %rd14262, %rd14212; + not.b64 %rd14264, %rd14238; + and.b64 %rd14265, %rd14239, %rd14264; + xor.b64 %rd14266, %rd14265, %rd14237; + not.b64 %rd14267, %rd14239; + and.b64 %rd14268, %rd14240, %rd14267; + xor.b64 %rd14269, %rd14268, %rd14238; + not.b64 %rd14270, %rd14240; + and.b64 %rd14271, %rd14212, %rd14270; + xor.b64 %rd14272, %rd14271, %rd14239; + not.b64 %rd14273, %rd14212; + and.b64 %rd14274, %rd14237, %rd14273; + xor.b64 %rd14275, %rd14274, %rd14240; + not.b64 %rd14276, %rd14242; + and.b64 %rd14277, %rd14243, %rd14276; + xor.b64 %rd14278, %rd14277, %rd14241; + not.b64 %rd14279, %rd14243; + and.b64 %rd14280, %rd14244, %rd14279; + xor.b64 %rd14281, %rd14280, %rd14242; + not.b64 %rd14282, %rd14244; + and.b64 %rd14283, %rd14245, %rd14282; + xor.b64 %rd14284, %rd14283, %rd14243; + not.b64 %rd14285, %rd14245; + and.b64 %rd14286, %rd14241, %rd14285; + xor.b64 %rd14287, %rd14286, %rd14244; + not.b64 %rd14288, %rd14241; + and.b64 %rd14289, %rd14242, %rd14288; + xor.b64 %rd14290, %rd14289, %rd14245; + not.b64 %rd14291, %rd14247; + and.b64 %rd14292, %rd14248, %rd14291; + xor.b64 %rd14293, %rd14292, %rd14246; + not.b64 %rd14294, %rd14248; + and.b64 %rd14295, %rd14249, %rd14294; + xor.b64 %rd14296, %rd14295, %rd14247; + not.b64 %rd14297, %rd14249; + and.b64 %rd14298, %rd14250, %rd14297; + xor.b64 %rd14299, %rd14298, %rd14248; + not.b64 %rd14300, %rd14250; + and.b64 %rd14301, %rd14246, %rd14300; + xor.b64 %rd14302, %rd14301, %rd14249; + not.b64 %rd14303, %rd14246; + and.b64 %rd14304, %rd14247, %rd14303; + xor.b64 %rd14305, %rd14304, %rd14250; + not.b64 %rd14306, %rd14252; + and.b64 %rd14307, %rd14253, %rd14306; + xor.b64 %rd14308, %rd14307, %rd14251; + not.b64 %rd14309, %rd14253; + and.b64 %rd14310, %rd14254, %rd14309; + xor.b64 %rd14311, %rd14310, %rd14252; + not.b64 %rd14312, %rd14254; + and.b64 %rd14313, %rd14255, %rd14312; + xor.b64 %rd14314, %rd14313, %rd14253; + not.b64 %rd14315, %rd14255; + and.b64 %rd14316, %rd14251, %rd14315; + xor.b64 %rd14317, %rd14316, %rd14254; + not.b64 %rd14318, %rd14251; + and.b64 %rd14319, %rd14252, %rd14318; + xor.b64 %rd14320, %rd14319, %rd14255; + not.b64 %rd14321, %rd14257; + and.b64 %rd14322, %rd14258, %rd14321; + xor.b64 %rd14323, %rd14322, %rd14256; + not.b64 %rd14324, %rd14258; + and.b64 %rd14325, %rd14259, %rd14324; + xor.b64 %rd14326, %rd14325, %rd14257; + not.b64 %rd14327, %rd14259; + and.b64 %rd14328, %rd14260, %rd14327; + xor.b64 %rd14329, %rd14328, %rd14258; + not.b64 %rd14330, %rd14260; + and.b64 %rd14331, %rd14256, %rd14330; + xor.b64 %rd14332, %rd14331, %rd14259; + not.b64 %rd14333, %rd14256; + and.b64 %rd14334, %rd14257, %rd14333; + xor.b64 %rd14335, %rd14334, %rd14260; + xor.b64 %rd14336, %rd14263, %rd30; + xor.b64 %rd14337, %rd14278, %rd14336; + xor.b64 %rd14338, %rd14337, %rd14293; + xor.b64 %rd14339, %rd14338, %rd14308; + xor.b64 %rd14340, %rd14339, %rd14323; + xor.b64 %rd14341, %rd14281, %rd14266; + xor.b64 %rd14342, %rd14341, %rd14296; + xor.b64 %rd14343, %rd14342, %rd14311; + xor.b64 %rd14344, %rd14343, %rd14326; + xor.b64 %rd14345, %rd14284, %rd14269; + xor.b64 %rd14346, %rd14345, %rd14299; + xor.b64 %rd14347, %rd14346, %rd14314; + xor.b64 %rd14348, %rd14347, %rd14329; + xor.b64 %rd14349, %rd14287, %rd14272; + xor.b64 %rd14350, %rd14349, %rd14302; + xor.b64 %rd14351, %rd14350, %rd14317; + xor.b64 %rd14352, %rd14351, %rd14332; + xor.b64 %rd14353, %rd14290, %rd14275; + xor.b64 %rd14354, %rd14353, %rd14305; + xor.b64 %rd14355, %rd14354, %rd14320; + xor.b64 %rd14356, %rd14355, %rd14335; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10176}, %rd14344; + } + { + .reg .b32 %dummy; + mov.b64 {%r10177,%dummy}, %rd14344; + } + shf.l.wrap.b32 %r10178, %r10177, %r10176, 1; + shf.l.wrap.b32 %r10179, %r10176, %r10177, 1; + mov.b64 %rd14357, {%r10179, %r10178}; + xor.b64 %rd14358, %rd14356, %rd14357; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10180}, %rd14348; + } + { + .reg .b32 %dummy; + mov.b64 {%r10181,%dummy}, %rd14348; + } + shf.l.wrap.b32 %r10182, %r10181, %r10180, 1; + shf.l.wrap.b32 %r10183, %r10180, %r10181, 1; + mov.b64 %rd14359, {%r10183, %r10182}; + xor.b64 %rd14360, %rd14359, %rd14340; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10184}, %rd14352; + } + { + .reg .b32 %dummy; + mov.b64 {%r10185,%dummy}, %rd14352; + } + shf.l.wrap.b32 %r10186, %r10185, %r10184, 1; + shf.l.wrap.b32 %r10187, %r10184, %r10185, 1; + mov.b64 %rd14361, {%r10187, %r10186}; + xor.b64 %rd14362, %rd14361, %rd14344; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10188}, %rd14356; + } + { + .reg .b32 %dummy; + mov.b64 {%r10189,%dummy}, %rd14356; + } + shf.l.wrap.b32 %r10190, %r10189, %r10188, 1; + shf.l.wrap.b32 %r10191, %r10188, %r10189, 1; + mov.b64 %rd14363, {%r10191, %r10190}; + xor.b64 %rd14364, %rd14363, %rd14348; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10192}, %rd14340; + } + { + .reg .b32 %dummy; + mov.b64 {%r10193,%dummy}, %rd14340; + } + shf.l.wrap.b32 %r10194, %r10193, %r10192, 1; + shf.l.wrap.b32 %r10195, %r10192, %r10193, 1; + mov.b64 %rd14365, {%r10195, %r10194}; + xor.b64 %rd14366, %rd14352, %rd14365; + xor.b64 %rd14367, %rd14358, %rd14336; + xor.b64 %rd14368, %rd14360, %rd14266; + xor.b64 %rd14369, %rd14362, %rd14269; + xor.b64 %rd14370, %rd14364, %rd14272; + xor.b64 %rd14371, %rd14366, %rd14275; + xor.b64 %rd14372, %rd14358, %rd14278; + xor.b64 %rd14373, %rd14360, %rd14281; + xor.b64 %rd14374, %rd14362, %rd14284; + xor.b64 %rd14375, %rd14364, %rd14287; + xor.b64 %rd14376, %rd14366, %rd14290; + xor.b64 %rd14377, %rd14358, %rd14293; + xor.b64 %rd14378, %rd14360, %rd14296; + xor.b64 %rd14379, %rd14362, %rd14299; + xor.b64 %rd14380, %rd14364, %rd14302; + xor.b64 %rd14381, %rd14366, %rd14305; + xor.b64 %rd14382, %rd14358, %rd14308; + xor.b64 %rd14383, %rd14360, %rd14311; + xor.b64 %rd14384, %rd14362, %rd14314; + xor.b64 %rd14385, %rd14364, %rd14317; + xor.b64 %rd14386, %rd14366, %rd14320; + xor.b64 %rd14387, %rd14358, %rd14323; + xor.b64 %rd14388, %rd14360, %rd14326; + xor.b64 %rd14389, %rd14362, %rd14329; + xor.b64 %rd14390, %rd14364, %rd14332; + xor.b64 %rd14391, %rd14366, %rd14335; + { + .reg .b32 %dummy; + mov.b64 {%r10196,%dummy}, %rd14373; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10197}, %rd14373; + } + shf.r.wrap.b32 %r10198, %r10197, %r10196, 20; + shf.r.wrap.b32 %r10199, %r10196, %r10197, 20; + mov.b64 %rd14392, {%r10199, %r10198}; + { + .reg .b32 %dummy; + mov.b64 {%r10200,%dummy}, %rd14379; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10201}, %rd14379; + } + shf.r.wrap.b32 %r10202, %r10201, %r10200, 21; + shf.r.wrap.b32 %r10203, %r10200, %r10201, 21; + mov.b64 %rd14393, {%r10203, %r10202}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10204}, %rd14385; + } + { + .reg .b32 %dummy; + mov.b64 {%r10205,%dummy}, %rd14385; + } + shf.l.wrap.b32 %r10206, %r10205, %r10204, 21; + shf.l.wrap.b32 %r10207, %r10204, %r10205, 21; + mov.b64 %rd14394, {%r10207, %r10206}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10208}, %rd14391; + } + { + .reg .b32 %dummy; + mov.b64 {%r10209,%dummy}, %rd14391; + } + shf.l.wrap.b32 %r10210, %r10209, %r10208, 14; + shf.l.wrap.b32 %r10211, %r10208, %r10209, 14; + mov.b64 %rd14395, {%r10211, %r10210}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10212}, %rd14370; + } + { + .reg .b32 %dummy; + mov.b64 {%r10213,%dummy}, %rd14370; + } + shf.l.wrap.b32 %r10214, %r10213, %r10212, 28; + shf.l.wrap.b32 %r10215, %r10212, %r10213, 28; + mov.b64 %rd14396, {%r10215, %r10214}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10216}, %rd14376; + } + { + .reg .b32 %dummy; + mov.b64 {%r10217,%dummy}, %rd14376; + } + shf.l.wrap.b32 %r10218, %r10217, %r10216, 20; + shf.l.wrap.b32 %r10219, %r10216, %r10217, 20; + mov.b64 %rd14397, {%r10219, %r10218}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10220}, %rd14377; + } + { + .reg .b32 %dummy; + mov.b64 {%r10221,%dummy}, %rd14377; + } + shf.l.wrap.b32 %r10222, %r10221, %r10220, 3; + shf.l.wrap.b32 %r10223, %r10220, %r10221, 3; + mov.b64 %rd14398, {%r10223, %r10222}; + { + .reg .b32 %dummy; + mov.b64 {%r10224,%dummy}, %rd14383; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10225}, %rd14383; + } + shf.r.wrap.b32 %r10226, %r10225, %r10224, 19; + shf.r.wrap.b32 %r10227, %r10224, %r10225, 19; + mov.b64 %rd14399, {%r10227, %r10226}; + { + .reg .b32 %dummy; + mov.b64 {%r10228,%dummy}, %rd14389; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10229}, %rd14389; + } + shf.r.wrap.b32 %r10230, %r10229, %r10228, 3; + shf.r.wrap.b32 %r10231, %r10228, %r10229, 3; + mov.b64 %rd14400, {%r10231, %r10230}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10232}, %rd14368; + } + { + .reg .b32 %dummy; + mov.b64 {%r10233,%dummy}, %rd14368; + } + shf.l.wrap.b32 %r10234, %r10233, %r10232, 1; + shf.l.wrap.b32 %r10235, %r10232, %r10233, 1; + mov.b64 %rd14401, {%r10235, %r10234}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10236}, %rd14374; + } + { + .reg .b32 %dummy; + mov.b64 {%r10237,%dummy}, %rd14374; + } + shf.l.wrap.b32 %r10238, %r10237, %r10236, 6; + shf.l.wrap.b32 %r10239, %r10236, %r10237, 6; + mov.b64 %rd14402, {%r10239, %r10238}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10240}, %rd14380; + } + { + .reg .b32 %dummy; + mov.b64 {%r10241,%dummy}, %rd14380; + } + shf.l.wrap.b32 %r10242, %r10241, %r10240, 25; + shf.l.wrap.b32 %r10243, %r10240, %r10241, 25; + mov.b64 %rd14403, {%r10243, %r10242}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10244}, %rd14386; + } + { + .reg .b32 %dummy; + mov.b64 {%r10245,%dummy}, %rd14386; + } + shf.l.wrap.b32 %r10246, %r10245, %r10244, 8; + shf.l.wrap.b32 %r10247, %r10244, %r10245, 8; + mov.b64 %rd14404, {%r10247, %r10246}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10248}, %rd14387; + } + { + .reg .b32 %dummy; + mov.b64 {%r10249,%dummy}, %rd14387; + } + shf.l.wrap.b32 %r10250, %r10249, %r10248, 18; + shf.l.wrap.b32 %r10251, %r10248, %r10249, 18; + mov.b64 %rd14405, {%r10251, %r10250}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10252}, %rd14371; + } + { + .reg .b32 %dummy; + mov.b64 {%r10253,%dummy}, %rd14371; + } + shf.l.wrap.b32 %r10254, %r10253, %r10252, 27; + shf.l.wrap.b32 %r10255, %r10252, %r10253, 27; + mov.b64 %rd14406, {%r10255, %r10254}; + { + .reg .b32 %dummy; + mov.b64 {%r10256,%dummy}, %rd14372; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10257}, %rd14372; + } + shf.r.wrap.b32 %r10258, %r10257, %r10256, 28; + shf.r.wrap.b32 %r10259, %r10256, %r10257, 28; + mov.b64 %rd14407, {%r10259, %r10258}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10260}, %rd14378; + } + { + .reg .b32 %dummy; + mov.b64 {%r10261,%dummy}, %rd14378; + } + shf.l.wrap.b32 %r10262, %r10261, %r10260, 10; + shf.l.wrap.b32 %r10263, %r10260, %r10261, 10; + mov.b64 %rd14408, {%r10263, %r10262}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10264}, %rd14384; + } + { + .reg .b32 %dummy; + mov.b64 {%r10265,%dummy}, %rd14384; + } + shf.l.wrap.b32 %r10266, %r10265, %r10264, 15; + shf.l.wrap.b32 %r10267, %r10264, %r10265, 15; + mov.b64 %rd14409, {%r10267, %r10266}; + { + .reg .b32 %dummy; + mov.b64 {%r10268,%dummy}, %rd14390; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10269}, %rd14390; + } + shf.r.wrap.b32 %r10270, %r10269, %r10268, 8; + shf.r.wrap.b32 %r10271, %r10268, %r10269, 8; + mov.b64 %rd14410, {%r10271, %r10270}; + { + .reg .b32 %dummy; + mov.b64 {%r10272,%dummy}, %rd14369; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10273}, %rd14369; + } + shf.r.wrap.b32 %r10274, %r10273, %r10272, 2; + shf.r.wrap.b32 %r10275, %r10272, %r10273, 2; + mov.b64 %rd14411, {%r10275, %r10274}; + { + .reg .b32 %dummy; + mov.b64 {%r10276,%dummy}, %rd14375; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10277}, %rd14375; + } + shf.r.wrap.b32 %r10278, %r10277, %r10276, 9; + shf.r.wrap.b32 %r10279, %r10276, %r10277, 9; + mov.b64 %rd14412, {%r10279, %r10278}; + { + .reg .b32 %dummy; + mov.b64 {%r10280,%dummy}, %rd14381; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10281}, %rd14381; + } + shf.r.wrap.b32 %r10282, %r10281, %r10280, 25; + shf.r.wrap.b32 %r10283, %r10280, %r10281, 25; + mov.b64 %rd14413, {%r10283, %r10282}; + { + .reg .b32 %dummy; + mov.b64 {%r10284,%dummy}, %rd14382; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10285}, %rd14382; + } + shf.r.wrap.b32 %r10286, %r10285, %r10284, 23; + shf.r.wrap.b32 %r10287, %r10284, %r10285, 23; + mov.b64 %rd14414, {%r10287, %r10286}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10288}, %rd14388; + } + { + .reg .b32 %dummy; + mov.b64 {%r10289,%dummy}, %rd14388; + } + shf.l.wrap.b32 %r10290, %r10289, %r10288, 2; + shf.l.wrap.b32 %r10291, %r10288, %r10289, 2; + mov.b64 %rd14415, {%r10291, %r10290}; + not.b64 %rd14416, %rd14392; + and.b64 %rd14417, %rd14393, %rd14416; + xor.b64 %rd14418, %rd14417, %rd14367; + not.b64 %rd14419, %rd14393; + and.b64 %rd14420, %rd14394, %rd14419; + xor.b64 %rd14421, %rd14420, %rd14392; + not.b64 %rd14422, %rd14394; + and.b64 %rd14423, %rd14395, %rd14422; + xor.b64 %rd14424, %rd14423, %rd14393; + not.b64 %rd14425, %rd14395; + and.b64 %rd14426, %rd14367, %rd14425; + xor.b64 %rd14427, %rd14426, %rd14394; + not.b64 %rd14428, %rd14367; + and.b64 %rd14429, %rd14392, %rd14428; + xor.b64 %rd14430, %rd14429, %rd14395; + not.b64 %rd14431, %rd14397; + and.b64 %rd14432, %rd14398, %rd14431; + xor.b64 %rd14433, %rd14432, %rd14396; + not.b64 %rd14434, %rd14398; + and.b64 %rd14435, %rd14399, %rd14434; + xor.b64 %rd14436, %rd14435, %rd14397; + not.b64 %rd14437, %rd14399; + and.b64 %rd14438, %rd14400, %rd14437; + xor.b64 %rd14439, %rd14438, %rd14398; + not.b64 %rd14440, %rd14400; + and.b64 %rd14441, %rd14396, %rd14440; + xor.b64 %rd14442, %rd14441, %rd14399; + not.b64 %rd14443, %rd14396; + and.b64 %rd14444, %rd14397, %rd14443; + xor.b64 %rd14445, %rd14444, %rd14400; + not.b64 %rd14446, %rd14402; + and.b64 %rd14447, %rd14403, %rd14446; + xor.b64 %rd14448, %rd14447, %rd14401; + not.b64 %rd14449, %rd14403; + and.b64 %rd14450, %rd14404, %rd14449; + xor.b64 %rd14451, %rd14450, %rd14402; + not.b64 %rd14452, %rd14404; + and.b64 %rd14453, %rd14405, %rd14452; + xor.b64 %rd14454, %rd14453, %rd14403; + not.b64 %rd14455, %rd14405; + and.b64 %rd14456, %rd14401, %rd14455; + xor.b64 %rd14457, %rd14456, %rd14404; + not.b64 %rd14458, %rd14401; + and.b64 %rd14459, %rd14402, %rd14458; + xor.b64 %rd14460, %rd14459, %rd14405; + not.b64 %rd14461, %rd14407; + and.b64 %rd14462, %rd14408, %rd14461; + xor.b64 %rd14463, %rd14462, %rd14406; + not.b64 %rd14464, %rd14408; + and.b64 %rd14465, %rd14409, %rd14464; + xor.b64 %rd14466, %rd14465, %rd14407; + not.b64 %rd14467, %rd14409; + and.b64 %rd14468, %rd14410, %rd14467; + xor.b64 %rd14469, %rd14468, %rd14408; + not.b64 %rd14470, %rd14410; + and.b64 %rd14471, %rd14406, %rd14470; + xor.b64 %rd14472, %rd14471, %rd14409; + not.b64 %rd14473, %rd14406; + and.b64 %rd14474, %rd14407, %rd14473; + xor.b64 %rd14475, %rd14474, %rd14410; + not.b64 %rd14476, %rd14412; + and.b64 %rd14477, %rd14413, %rd14476; + xor.b64 %rd14478, %rd14477, %rd14411; + not.b64 %rd14479, %rd14413; + and.b64 %rd14480, %rd14414, %rd14479; + xor.b64 %rd14481, %rd14480, %rd14412; + not.b64 %rd14482, %rd14414; + and.b64 %rd14483, %rd14415, %rd14482; + xor.b64 %rd14484, %rd14483, %rd14413; + not.b64 %rd14485, %rd14415; + and.b64 %rd14486, %rd14411, %rd14485; + xor.b64 %rd14487, %rd14486, %rd14414; + not.b64 %rd14488, %rd14411; + and.b64 %rd14489, %rd14412, %rd14488; + xor.b64 %rd14490, %rd14489, %rd14415; + xor.b64 %rd14491, %rd14418, %rd31; + xor.b64 %rd14492, %rd14433, %rd14491; + xor.b64 %rd14493, %rd14492, %rd14448; + xor.b64 %rd14494, %rd14493, %rd14463; + xor.b64 %rd14495, %rd14494, %rd14478; + xor.b64 %rd14496, %rd14436, %rd14421; + xor.b64 %rd14497, %rd14496, %rd14451; + xor.b64 %rd14498, %rd14497, %rd14466; + xor.b64 %rd14499, %rd14498, %rd14481; + xor.b64 %rd14500, %rd14439, %rd14424; + xor.b64 %rd14501, %rd14500, %rd14454; + xor.b64 %rd14502, %rd14501, %rd14469; + xor.b64 %rd14503, %rd14502, %rd14484; + xor.b64 %rd14504, %rd14442, %rd14427; + xor.b64 %rd14505, %rd14504, %rd14457; + xor.b64 %rd14506, %rd14505, %rd14472; + xor.b64 %rd14507, %rd14506, %rd14487; + xor.b64 %rd14508, %rd14445, %rd14430; + xor.b64 %rd14509, %rd14508, %rd14460; + xor.b64 %rd14510, %rd14509, %rd14475; + xor.b64 %rd14511, %rd14510, %rd14490; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10292}, %rd14499; + } + { + .reg .b32 %dummy; + mov.b64 {%r10293,%dummy}, %rd14499; + } + shf.l.wrap.b32 %r10294, %r10293, %r10292, 1; + shf.l.wrap.b32 %r10295, %r10292, %r10293, 1; + mov.b64 %rd14512, {%r10295, %r10294}; + xor.b64 %rd14513, %rd14511, %rd14512; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10296}, %rd14503; + } + { + .reg .b32 %dummy; + mov.b64 {%r10297,%dummy}, %rd14503; + } + shf.l.wrap.b32 %r10298, %r10297, %r10296, 1; + shf.l.wrap.b32 %r10299, %r10296, %r10297, 1; + mov.b64 %rd14514, {%r10299, %r10298}; + xor.b64 %rd14515, %rd14514, %rd14495; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10300}, %rd14507; + } + { + .reg .b32 %dummy; + mov.b64 {%r10301,%dummy}, %rd14507; + } + shf.l.wrap.b32 %r10302, %r10301, %r10300, 1; + shf.l.wrap.b32 %r10303, %r10300, %r10301, 1; + mov.b64 %rd14516, {%r10303, %r10302}; + xor.b64 %rd14517, %rd14516, %rd14499; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10304}, %rd14511; + } + { + .reg .b32 %dummy; + mov.b64 {%r10305,%dummy}, %rd14511; + } + shf.l.wrap.b32 %r10306, %r10305, %r10304, 1; + shf.l.wrap.b32 %r10307, %r10304, %r10305, 1; + mov.b64 %rd14518, {%r10307, %r10306}; + xor.b64 %rd14519, %rd14518, %rd14503; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10308}, %rd14495; + } + { + .reg .b32 %dummy; + mov.b64 {%r10309,%dummy}, %rd14495; + } + shf.l.wrap.b32 %r10310, %r10309, %r10308, 1; + shf.l.wrap.b32 %r10311, %r10308, %r10309, 1; + mov.b64 %rd14520, {%r10311, %r10310}; + xor.b64 %rd14521, %rd14507, %rd14520; + xor.b64 %rd14522, %rd14513, %rd14491; + xor.b64 %rd14523, %rd14515, %rd14421; + xor.b64 %rd14524, %rd14517, %rd14424; + xor.b64 %rd14525, %rd14519, %rd14427; + xor.b64 %rd14526, %rd14521, %rd14430; + xor.b64 %rd14527, %rd14513, %rd14433; + xor.b64 %rd14528, %rd14515, %rd14436; + xor.b64 %rd14529, %rd14517, %rd14439; + xor.b64 %rd14530, %rd14519, %rd14442; + xor.b64 %rd14531, %rd14521, %rd14445; + xor.b64 %rd14532, %rd14513, %rd14448; + xor.b64 %rd14533, %rd14515, %rd14451; + xor.b64 %rd14534, %rd14517, %rd14454; + xor.b64 %rd14535, %rd14519, %rd14457; + xor.b64 %rd14536, %rd14521, %rd14460; + xor.b64 %rd14537, %rd14513, %rd14463; + xor.b64 %rd14538, %rd14515, %rd14466; + xor.b64 %rd14539, %rd14517, %rd14469; + xor.b64 %rd14540, %rd14519, %rd14472; + xor.b64 %rd14541, %rd14521, %rd14475; + xor.b64 %rd14542, %rd14513, %rd14478; + xor.b64 %rd14543, %rd14515, %rd14481; + xor.b64 %rd14544, %rd14517, %rd14484; + xor.b64 %rd14545, %rd14519, %rd14487; + xor.b64 %rd14546, %rd14521, %rd14490; + { + .reg .b32 %dummy; + mov.b64 {%r10312,%dummy}, %rd14528; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10313}, %rd14528; + } + shf.r.wrap.b32 %r10314, %r10313, %r10312, 20; + shf.r.wrap.b32 %r10315, %r10312, %r10313, 20; + mov.b64 %rd14547, {%r10315, %r10314}; + { + .reg .b32 %dummy; + mov.b64 {%r10316,%dummy}, %rd14534; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10317}, %rd14534; + } + shf.r.wrap.b32 %r10318, %r10317, %r10316, 21; + shf.r.wrap.b32 %r10319, %r10316, %r10317, 21; + mov.b64 %rd14548, {%r10319, %r10318}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10320}, %rd14540; + } + { + .reg .b32 %dummy; + mov.b64 {%r10321,%dummy}, %rd14540; + } + shf.l.wrap.b32 %r10322, %r10321, %r10320, 21; + shf.l.wrap.b32 %r10323, %r10320, %r10321, 21; + mov.b64 %rd14549, {%r10323, %r10322}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10324}, %rd14546; + } + { + .reg .b32 %dummy; + mov.b64 {%r10325,%dummy}, %rd14546; + } + shf.l.wrap.b32 %r10326, %r10325, %r10324, 14; + shf.l.wrap.b32 %r10327, %r10324, %r10325, 14; + mov.b64 %rd14550, {%r10327, %r10326}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10328}, %rd14525; + } + { + .reg .b32 %dummy; + mov.b64 {%r10329,%dummy}, %rd14525; + } + shf.l.wrap.b32 %r10330, %r10329, %r10328, 28; + shf.l.wrap.b32 %r10331, %r10328, %r10329, 28; + mov.b64 %rd14551, {%r10331, %r10330}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10332}, %rd14531; + } + { + .reg .b32 %dummy; + mov.b64 {%r10333,%dummy}, %rd14531; + } + shf.l.wrap.b32 %r10334, %r10333, %r10332, 20; + shf.l.wrap.b32 %r10335, %r10332, %r10333, 20; + mov.b64 %rd14552, {%r10335, %r10334}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10336}, %rd14532; + } + { + .reg .b32 %dummy; + mov.b64 {%r10337,%dummy}, %rd14532; + } + shf.l.wrap.b32 %r10338, %r10337, %r10336, 3; + shf.l.wrap.b32 %r10339, %r10336, %r10337, 3; + mov.b64 %rd14553, {%r10339, %r10338}; + { + .reg .b32 %dummy; + mov.b64 {%r10340,%dummy}, %rd14538; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10341}, %rd14538; + } + shf.r.wrap.b32 %r10342, %r10341, %r10340, 19; + shf.r.wrap.b32 %r10343, %r10340, %r10341, 19; + mov.b64 %rd14554, {%r10343, %r10342}; + { + .reg .b32 %dummy; + mov.b64 {%r10344,%dummy}, %rd14544; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10345}, %rd14544; + } + shf.r.wrap.b32 %r10346, %r10345, %r10344, 3; + shf.r.wrap.b32 %r10347, %r10344, %r10345, 3; + mov.b64 %rd14555, {%r10347, %r10346}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10348}, %rd14523; + } + { + .reg .b32 %dummy; + mov.b64 {%r10349,%dummy}, %rd14523; + } + shf.l.wrap.b32 %r10350, %r10349, %r10348, 1; + shf.l.wrap.b32 %r10351, %r10348, %r10349, 1; + mov.b64 %rd14556, {%r10351, %r10350}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10352}, %rd14529; + } + { + .reg .b32 %dummy; + mov.b64 {%r10353,%dummy}, %rd14529; + } + shf.l.wrap.b32 %r10354, %r10353, %r10352, 6; + shf.l.wrap.b32 %r10355, %r10352, %r10353, 6; + mov.b64 %rd14557, {%r10355, %r10354}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10356}, %rd14535; + } + { + .reg .b32 %dummy; + mov.b64 {%r10357,%dummy}, %rd14535; + } + shf.l.wrap.b32 %r10358, %r10357, %r10356, 25; + shf.l.wrap.b32 %r10359, %r10356, %r10357, 25; + mov.b64 %rd14558, {%r10359, %r10358}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10360}, %rd14541; + } + { + .reg .b32 %dummy; + mov.b64 {%r10361,%dummy}, %rd14541; + } + shf.l.wrap.b32 %r10362, %r10361, %r10360, 8; + shf.l.wrap.b32 %r10363, %r10360, %r10361, 8; + mov.b64 %rd14559, {%r10363, %r10362}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10364}, %rd14542; + } + { + .reg .b32 %dummy; + mov.b64 {%r10365,%dummy}, %rd14542; + } + shf.l.wrap.b32 %r10366, %r10365, %r10364, 18; + shf.l.wrap.b32 %r10367, %r10364, %r10365, 18; + mov.b64 %rd14560, {%r10367, %r10366}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10368}, %rd14526; + } + { + .reg .b32 %dummy; + mov.b64 {%r10369,%dummy}, %rd14526; + } + shf.l.wrap.b32 %r10370, %r10369, %r10368, 27; + shf.l.wrap.b32 %r10371, %r10368, %r10369, 27; + mov.b64 %rd14561, {%r10371, %r10370}; + { + .reg .b32 %dummy; + mov.b64 {%r10372,%dummy}, %rd14527; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10373}, %rd14527; + } + shf.r.wrap.b32 %r10374, %r10373, %r10372, 28; + shf.r.wrap.b32 %r10375, %r10372, %r10373, 28; + mov.b64 %rd14562, {%r10375, %r10374}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10376}, %rd14533; + } + { + .reg .b32 %dummy; + mov.b64 {%r10377,%dummy}, %rd14533; + } + shf.l.wrap.b32 %r10378, %r10377, %r10376, 10; + shf.l.wrap.b32 %r10379, %r10376, %r10377, 10; + mov.b64 %rd14563, {%r10379, %r10378}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10380}, %rd14539; + } + { + .reg .b32 %dummy; + mov.b64 {%r10381,%dummy}, %rd14539; + } + shf.l.wrap.b32 %r10382, %r10381, %r10380, 15; + shf.l.wrap.b32 %r10383, %r10380, %r10381, 15; + mov.b64 %rd14564, {%r10383, %r10382}; + { + .reg .b32 %dummy; + mov.b64 {%r10384,%dummy}, %rd14545; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10385}, %rd14545; + } + shf.r.wrap.b32 %r10386, %r10385, %r10384, 8; + shf.r.wrap.b32 %r10387, %r10384, %r10385, 8; + mov.b64 %rd14565, {%r10387, %r10386}; + { + .reg .b32 %dummy; + mov.b64 {%r10388,%dummy}, %rd14524; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10389}, %rd14524; + } + shf.r.wrap.b32 %r10390, %r10389, %r10388, 2; + shf.r.wrap.b32 %r10391, %r10388, %r10389, 2; + mov.b64 %rd14566, {%r10391, %r10390}; + { + .reg .b32 %dummy; + mov.b64 {%r10392,%dummy}, %rd14530; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10393}, %rd14530; + } + shf.r.wrap.b32 %r10394, %r10393, %r10392, 9; + shf.r.wrap.b32 %r10395, %r10392, %r10393, 9; + mov.b64 %rd14567, {%r10395, %r10394}; + { + .reg .b32 %dummy; + mov.b64 {%r10396,%dummy}, %rd14536; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10397}, %rd14536; + } + shf.r.wrap.b32 %r10398, %r10397, %r10396, 25; + shf.r.wrap.b32 %r10399, %r10396, %r10397, 25; + mov.b64 %rd14568, {%r10399, %r10398}; + { + .reg .b32 %dummy; + mov.b64 {%r10400,%dummy}, %rd14537; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10401}, %rd14537; + } + shf.r.wrap.b32 %r10402, %r10401, %r10400, 23; + shf.r.wrap.b32 %r10403, %r10400, %r10401, 23; + mov.b64 %rd14569, {%r10403, %r10402}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10404}, %rd14543; + } + { + .reg .b32 %dummy; + mov.b64 {%r10405,%dummy}, %rd14543; + } + shf.l.wrap.b32 %r10406, %r10405, %r10404, 2; + shf.l.wrap.b32 %r10407, %r10404, %r10405, 2; + mov.b64 %rd14570, {%r10407, %r10406}; + not.b64 %rd14571, %rd14547; + and.b64 %rd14572, %rd14548, %rd14571; + xor.b64 %rd14573, %rd14572, %rd14522; + not.b64 %rd14574, %rd14548; + and.b64 %rd14575, %rd14549, %rd14574; + xor.b64 %rd14576, %rd14575, %rd14547; + not.b64 %rd14577, %rd14549; + and.b64 %rd14578, %rd14550, %rd14577; + xor.b64 %rd14579, %rd14578, %rd14548; + not.b64 %rd14580, %rd14550; + and.b64 %rd14581, %rd14522, %rd14580; + xor.b64 %rd14582, %rd14581, %rd14549; + not.b64 %rd14583, %rd14522; + and.b64 %rd14584, %rd14547, %rd14583; + xor.b64 %rd14585, %rd14584, %rd14550; + not.b64 %rd14586, %rd14552; + and.b64 %rd14587, %rd14553, %rd14586; + xor.b64 %rd14588, %rd14587, %rd14551; + not.b64 %rd14589, %rd14553; + and.b64 %rd14590, %rd14554, %rd14589; + xor.b64 %rd14591, %rd14590, %rd14552; + not.b64 %rd14592, %rd14554; + and.b64 %rd14593, %rd14555, %rd14592; + xor.b64 %rd14594, %rd14593, %rd14553; + not.b64 %rd14595, %rd14555; + and.b64 %rd14596, %rd14551, %rd14595; + xor.b64 %rd14597, %rd14596, %rd14554; + not.b64 %rd14598, %rd14551; + and.b64 %rd14599, %rd14552, %rd14598; + xor.b64 %rd14600, %rd14599, %rd14555; + not.b64 %rd14601, %rd14557; + and.b64 %rd14602, %rd14558, %rd14601; + xor.b64 %rd14603, %rd14602, %rd14556; + not.b64 %rd14604, %rd14558; + and.b64 %rd14605, %rd14559, %rd14604; + xor.b64 %rd14606, %rd14605, %rd14557; + not.b64 %rd14607, %rd14559; + and.b64 %rd14608, %rd14560, %rd14607; + xor.b64 %rd14609, %rd14608, %rd14558; + not.b64 %rd14610, %rd14560; + and.b64 %rd14611, %rd14556, %rd14610; + xor.b64 %rd14612, %rd14611, %rd14559; + not.b64 %rd14613, %rd14556; + and.b64 %rd14614, %rd14557, %rd14613; + xor.b64 %rd14615, %rd14614, %rd14560; + not.b64 %rd14616, %rd14562; + and.b64 %rd14617, %rd14563, %rd14616; + xor.b64 %rd14618, %rd14617, %rd14561; + not.b64 %rd14619, %rd14563; + and.b64 %rd14620, %rd14564, %rd14619; + xor.b64 %rd14621, %rd14620, %rd14562; + not.b64 %rd14622, %rd14564; + and.b64 %rd14623, %rd14565, %rd14622; + xor.b64 %rd14624, %rd14623, %rd14563; + not.b64 %rd14625, %rd14565; + and.b64 %rd14626, %rd14561, %rd14625; + xor.b64 %rd14627, %rd14626, %rd14564; + not.b64 %rd14628, %rd14561; + and.b64 %rd14629, %rd14562, %rd14628; + xor.b64 %rd14630, %rd14629, %rd14565; + not.b64 %rd14631, %rd14567; + and.b64 %rd14632, %rd14568, %rd14631; + xor.b64 %rd14633, %rd14632, %rd14566; + not.b64 %rd14634, %rd14568; + and.b64 %rd14635, %rd14569, %rd14634; + xor.b64 %rd14636, %rd14635, %rd14567; + not.b64 %rd14637, %rd14569; + and.b64 %rd14638, %rd14570, %rd14637; + xor.b64 %rd14639, %rd14638, %rd14568; + not.b64 %rd14640, %rd14570; + and.b64 %rd14641, %rd14566, %rd14640; + xor.b64 %rd14642, %rd14641, %rd14569; + not.b64 %rd14643, %rd14566; + and.b64 %rd14644, %rd14567, %rd14643; + xor.b64 %rd14645, %rd14644, %rd14570; + xor.b64 %rd14646, %rd14573, %rd32; + xor.b64 %rd14647, %rd14588, %rd14646; + xor.b64 %rd14648, %rd14647, %rd14603; + xor.b64 %rd14649, %rd14648, %rd14618; + xor.b64 %rd14650, %rd14649, %rd14633; + xor.b64 %rd14651, %rd14591, %rd14576; + xor.b64 %rd14652, %rd14651, %rd14606; + xor.b64 %rd14653, %rd14652, %rd14621; + xor.b64 %rd14654, %rd14653, %rd14636; + xor.b64 %rd14655, %rd14594, %rd14579; + xor.b64 %rd14656, %rd14655, %rd14609; + xor.b64 %rd14657, %rd14656, %rd14624; + xor.b64 %rd14658, %rd14657, %rd14639; + xor.b64 %rd14659, %rd14597, %rd14582; + xor.b64 %rd14660, %rd14659, %rd14612; + xor.b64 %rd14661, %rd14660, %rd14627; + xor.b64 %rd14662, %rd14661, %rd14642; + xor.b64 %rd14663, %rd14600, %rd14585; + xor.b64 %rd14664, %rd14663, %rd14615; + xor.b64 %rd14665, %rd14664, %rd14630; + xor.b64 %rd14666, %rd14665, %rd14645; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10408}, %rd14654; + } + { + .reg .b32 %dummy; + mov.b64 {%r10409,%dummy}, %rd14654; + } + shf.l.wrap.b32 %r10410, %r10409, %r10408, 1; + shf.l.wrap.b32 %r10411, %r10408, %r10409, 1; + mov.b64 %rd14667, {%r10411, %r10410}; + xor.b64 %rd14668, %rd14666, %rd14667; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10412}, %rd14658; + } + { + .reg .b32 %dummy; + mov.b64 {%r10413,%dummy}, %rd14658; + } + shf.l.wrap.b32 %r10414, %r10413, %r10412, 1; + shf.l.wrap.b32 %r10415, %r10412, %r10413, 1; + mov.b64 %rd14669, {%r10415, %r10414}; + xor.b64 %rd14670, %rd14669, %rd14650; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10416}, %rd14662; + } + { + .reg .b32 %dummy; + mov.b64 {%r10417,%dummy}, %rd14662; + } + shf.l.wrap.b32 %r10418, %r10417, %r10416, 1; + shf.l.wrap.b32 %r10419, %r10416, %r10417, 1; + mov.b64 %rd14671, {%r10419, %r10418}; + xor.b64 %rd14672, %rd14671, %rd14654; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10420}, %rd14666; + } + { + .reg .b32 %dummy; + mov.b64 {%r10421,%dummy}, %rd14666; + } + shf.l.wrap.b32 %r10422, %r10421, %r10420, 1; + shf.l.wrap.b32 %r10423, %r10420, %r10421, 1; + mov.b64 %rd14673, {%r10423, %r10422}; + xor.b64 %rd14674, %rd14673, %rd14658; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10424}, %rd14650; + } + { + .reg .b32 %dummy; + mov.b64 {%r10425,%dummy}, %rd14650; + } + shf.l.wrap.b32 %r10426, %r10425, %r10424, 1; + shf.l.wrap.b32 %r10427, %r10424, %r10425, 1; + mov.b64 %rd14675, {%r10427, %r10426}; + xor.b64 %rd14676, %rd14662, %rd14675; + xor.b64 %rd14677, %rd14668, %rd14646; + xor.b64 %rd14678, %rd14670, %rd14576; + xor.b64 %rd14679, %rd14672, %rd14579; + xor.b64 %rd14680, %rd14674, %rd14582; + xor.b64 %rd14681, %rd14676, %rd14585; + xor.b64 %rd14682, %rd14668, %rd14588; + xor.b64 %rd14683, %rd14670, %rd14591; + xor.b64 %rd14684, %rd14672, %rd14594; + xor.b64 %rd14685, %rd14674, %rd14597; + xor.b64 %rd14686, %rd14676, %rd14600; + xor.b64 %rd14687, %rd14668, %rd14603; + xor.b64 %rd14688, %rd14670, %rd14606; + xor.b64 %rd14689, %rd14672, %rd14609; + xor.b64 %rd14690, %rd14674, %rd14612; + xor.b64 %rd14691, %rd14676, %rd14615; + xor.b64 %rd14692, %rd14668, %rd14618; + xor.b64 %rd14693, %rd14670, %rd14621; + xor.b64 %rd14694, %rd14672, %rd14624; + xor.b64 %rd14695, %rd14674, %rd14627; + xor.b64 %rd14696, %rd14676, %rd14630; + xor.b64 %rd14697, %rd14668, %rd14633; + xor.b64 %rd14698, %rd14670, %rd14636; + xor.b64 %rd14699, %rd14672, %rd14639; + xor.b64 %rd14700, %rd14674, %rd14642; + xor.b64 %rd14701, %rd14676, %rd14645; + { + .reg .b32 %dummy; + mov.b64 {%r10428,%dummy}, %rd14683; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10429}, %rd14683; + } + shf.r.wrap.b32 %r10430, %r10429, %r10428, 20; + shf.r.wrap.b32 %r10431, %r10428, %r10429, 20; + mov.b64 %rd14702, {%r10431, %r10430}; + { + .reg .b32 %dummy; + mov.b64 {%r10432,%dummy}, %rd14689; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10433}, %rd14689; + } + shf.r.wrap.b32 %r10434, %r10433, %r10432, 21; + shf.r.wrap.b32 %r10435, %r10432, %r10433, 21; + mov.b64 %rd14703, {%r10435, %r10434}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10436}, %rd14695; + } + { + .reg .b32 %dummy; + mov.b64 {%r10437,%dummy}, %rd14695; + } + shf.l.wrap.b32 %r10438, %r10437, %r10436, 21; + shf.l.wrap.b32 %r10439, %r10436, %r10437, 21; + mov.b64 %rd14704, {%r10439, %r10438}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10440}, %rd14701; + } + { + .reg .b32 %dummy; + mov.b64 {%r10441,%dummy}, %rd14701; + } + shf.l.wrap.b32 %r10442, %r10441, %r10440, 14; + shf.l.wrap.b32 %r10443, %r10440, %r10441, 14; + mov.b64 %rd14705, {%r10443, %r10442}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10444}, %rd14680; + } + { + .reg .b32 %dummy; + mov.b64 {%r10445,%dummy}, %rd14680; + } + shf.l.wrap.b32 %r10446, %r10445, %r10444, 28; + shf.l.wrap.b32 %r10447, %r10444, %r10445, 28; + mov.b64 %rd14706, {%r10447, %r10446}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10448}, %rd14686; + } + { + .reg .b32 %dummy; + mov.b64 {%r10449,%dummy}, %rd14686; + } + shf.l.wrap.b32 %r10450, %r10449, %r10448, 20; + shf.l.wrap.b32 %r10451, %r10448, %r10449, 20; + mov.b64 %rd14707, {%r10451, %r10450}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10452}, %rd14687; + } + { + .reg .b32 %dummy; + mov.b64 {%r10453,%dummy}, %rd14687; + } + shf.l.wrap.b32 %r10454, %r10453, %r10452, 3; + shf.l.wrap.b32 %r10455, %r10452, %r10453, 3; + mov.b64 %rd14708, {%r10455, %r10454}; + { + .reg .b32 %dummy; + mov.b64 {%r10456,%dummy}, %rd14693; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10457}, %rd14693; + } + shf.r.wrap.b32 %r10458, %r10457, %r10456, 19; + shf.r.wrap.b32 %r10459, %r10456, %r10457, 19; + mov.b64 %rd14709, {%r10459, %r10458}; + { + .reg .b32 %dummy; + mov.b64 {%r10460,%dummy}, %rd14699; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10461}, %rd14699; + } + shf.r.wrap.b32 %r10462, %r10461, %r10460, 3; + shf.r.wrap.b32 %r10463, %r10460, %r10461, 3; + mov.b64 %rd14710, {%r10463, %r10462}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10464}, %rd14678; + } + { + .reg .b32 %dummy; + mov.b64 {%r10465,%dummy}, %rd14678; + } + shf.l.wrap.b32 %r10466, %r10465, %r10464, 1; + shf.l.wrap.b32 %r10467, %r10464, %r10465, 1; + mov.b64 %rd14711, {%r10467, %r10466}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10468}, %rd14684; + } + { + .reg .b32 %dummy; + mov.b64 {%r10469,%dummy}, %rd14684; + } + shf.l.wrap.b32 %r10470, %r10469, %r10468, 6; + shf.l.wrap.b32 %r10471, %r10468, %r10469, 6; + mov.b64 %rd14712, {%r10471, %r10470}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10472}, %rd14690; + } + { + .reg .b32 %dummy; + mov.b64 {%r10473,%dummy}, %rd14690; + } + shf.l.wrap.b32 %r10474, %r10473, %r10472, 25; + shf.l.wrap.b32 %r10475, %r10472, %r10473, 25; + mov.b64 %rd14713, {%r10475, %r10474}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10476}, %rd14696; + } + { + .reg .b32 %dummy; + mov.b64 {%r10477,%dummy}, %rd14696; + } + shf.l.wrap.b32 %r10478, %r10477, %r10476, 8; + shf.l.wrap.b32 %r10479, %r10476, %r10477, 8; + mov.b64 %rd14714, {%r10479, %r10478}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10480}, %rd14697; + } + { + .reg .b32 %dummy; + mov.b64 {%r10481,%dummy}, %rd14697; + } + shf.l.wrap.b32 %r10482, %r10481, %r10480, 18; + shf.l.wrap.b32 %r10483, %r10480, %r10481, 18; + mov.b64 %rd14715, {%r10483, %r10482}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10484}, %rd14681; + } + { + .reg .b32 %dummy; + mov.b64 {%r10485,%dummy}, %rd14681; + } + shf.l.wrap.b32 %r10486, %r10485, %r10484, 27; + shf.l.wrap.b32 %r10487, %r10484, %r10485, 27; + mov.b64 %rd14716, {%r10487, %r10486}; + { + .reg .b32 %dummy; + mov.b64 {%r10488,%dummy}, %rd14682; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10489}, %rd14682; + } + shf.r.wrap.b32 %r10490, %r10489, %r10488, 28; + shf.r.wrap.b32 %r10491, %r10488, %r10489, 28; + mov.b64 %rd14717, {%r10491, %r10490}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10492}, %rd14688; + } + { + .reg .b32 %dummy; + mov.b64 {%r10493,%dummy}, %rd14688; + } + shf.l.wrap.b32 %r10494, %r10493, %r10492, 10; + shf.l.wrap.b32 %r10495, %r10492, %r10493, 10; + mov.b64 %rd14718, {%r10495, %r10494}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10496}, %rd14694; + } + { + .reg .b32 %dummy; + mov.b64 {%r10497,%dummy}, %rd14694; + } + shf.l.wrap.b32 %r10498, %r10497, %r10496, 15; + shf.l.wrap.b32 %r10499, %r10496, %r10497, 15; + mov.b64 %rd14719, {%r10499, %r10498}; + { + .reg .b32 %dummy; + mov.b64 {%r10500,%dummy}, %rd14700; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10501}, %rd14700; + } + shf.r.wrap.b32 %r10502, %r10501, %r10500, 8; + shf.r.wrap.b32 %r10503, %r10500, %r10501, 8; + mov.b64 %rd14720, {%r10503, %r10502}; + { + .reg .b32 %dummy; + mov.b64 {%r10504,%dummy}, %rd14679; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10505}, %rd14679; + } + shf.r.wrap.b32 %r10506, %r10505, %r10504, 2; + shf.r.wrap.b32 %r10507, %r10504, %r10505, 2; + mov.b64 %rd14721, {%r10507, %r10506}; + { + .reg .b32 %dummy; + mov.b64 {%r10508,%dummy}, %rd14685; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10509}, %rd14685; + } + shf.r.wrap.b32 %r10510, %r10509, %r10508, 9; + shf.r.wrap.b32 %r10511, %r10508, %r10509, 9; + mov.b64 %rd14722, {%r10511, %r10510}; + { + .reg .b32 %dummy; + mov.b64 {%r10512,%dummy}, %rd14691; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10513}, %rd14691; + } + shf.r.wrap.b32 %r10514, %r10513, %r10512, 25; + shf.r.wrap.b32 %r10515, %r10512, %r10513, 25; + mov.b64 %rd14723, {%r10515, %r10514}; + { + .reg .b32 %dummy; + mov.b64 {%r10516,%dummy}, %rd14692; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10517}, %rd14692; + } + shf.r.wrap.b32 %r10518, %r10517, %r10516, 23; + shf.r.wrap.b32 %r10519, %r10516, %r10517, 23; + mov.b64 %rd14724, {%r10519, %r10518}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10520}, %rd14698; + } + { + .reg .b32 %dummy; + mov.b64 {%r10521,%dummy}, %rd14698; + } + shf.l.wrap.b32 %r10522, %r10521, %r10520, 2; + shf.l.wrap.b32 %r10523, %r10520, %r10521, 2; + mov.b64 %rd14725, {%r10523, %r10522}; + not.b64 %rd14726, %rd14702; + and.b64 %rd14727, %rd14703, %rd14726; + xor.b64 %rd14728, %rd14727, %rd14677; + not.b64 %rd14729, %rd14703; + and.b64 %rd14730, %rd14704, %rd14729; + xor.b64 %rd14731, %rd14730, %rd14702; + not.b64 %rd14732, %rd14704; + and.b64 %rd14733, %rd14705, %rd14732; + xor.b64 %rd14734, %rd14733, %rd14703; + not.b64 %rd14735, %rd14705; + and.b64 %rd14736, %rd14677, %rd14735; + xor.b64 %rd14737, %rd14736, %rd14704; + not.b64 %rd14738, %rd14677; + and.b64 %rd14739, %rd14702, %rd14738; + xor.b64 %rd14740, %rd14739, %rd14705; + not.b64 %rd14741, %rd14707; + and.b64 %rd14742, %rd14708, %rd14741; + xor.b64 %rd14743, %rd14742, %rd14706; + not.b64 %rd14744, %rd14708; + and.b64 %rd14745, %rd14709, %rd14744; + xor.b64 %rd14746, %rd14745, %rd14707; + not.b64 %rd14747, %rd14709; + and.b64 %rd14748, %rd14710, %rd14747; + xor.b64 %rd14749, %rd14748, %rd14708; + not.b64 %rd14750, %rd14710; + and.b64 %rd14751, %rd14706, %rd14750; + xor.b64 %rd14752, %rd14751, %rd14709; + not.b64 %rd14753, %rd14706; + and.b64 %rd14754, %rd14707, %rd14753; + xor.b64 %rd14755, %rd14754, %rd14710; + not.b64 %rd14756, %rd14712; + and.b64 %rd14757, %rd14713, %rd14756; + xor.b64 %rd14758, %rd14757, %rd14711; + not.b64 %rd14759, %rd14713; + and.b64 %rd14760, %rd14714, %rd14759; + xor.b64 %rd14761, %rd14760, %rd14712; + not.b64 %rd14762, %rd14714; + and.b64 %rd14763, %rd14715, %rd14762; + xor.b64 %rd14764, %rd14763, %rd14713; + not.b64 %rd14765, %rd14715; + and.b64 %rd14766, %rd14711, %rd14765; + xor.b64 %rd14767, %rd14766, %rd14714; + not.b64 %rd14768, %rd14711; + and.b64 %rd14769, %rd14712, %rd14768; + xor.b64 %rd14770, %rd14769, %rd14715; + not.b64 %rd14771, %rd14717; + and.b64 %rd14772, %rd14718, %rd14771; + xor.b64 %rd14773, %rd14772, %rd14716; + not.b64 %rd14774, %rd14718; + and.b64 %rd14775, %rd14719, %rd14774; + xor.b64 %rd14776, %rd14775, %rd14717; + not.b64 %rd14777, %rd14719; + and.b64 %rd14778, %rd14720, %rd14777; + xor.b64 %rd14779, %rd14778, %rd14718; + not.b64 %rd14780, %rd14720; + and.b64 %rd14781, %rd14716, %rd14780; + xor.b64 %rd14782, %rd14781, %rd14719; + not.b64 %rd14783, %rd14716; + and.b64 %rd14784, %rd14717, %rd14783; + xor.b64 %rd14785, %rd14784, %rd14720; + not.b64 %rd14786, %rd14722; + and.b64 %rd14787, %rd14723, %rd14786; + xor.b64 %rd14788, %rd14787, %rd14721; + not.b64 %rd14789, %rd14723; + and.b64 %rd14790, %rd14724, %rd14789; + xor.b64 %rd14791, %rd14790, %rd14722; + not.b64 %rd14792, %rd14724; + and.b64 %rd14793, %rd14725, %rd14792; + xor.b64 %rd14794, %rd14793, %rd14723; + not.b64 %rd14795, %rd14725; + and.b64 %rd14796, %rd14721, %rd14795; + xor.b64 %rd14797, %rd14796, %rd14724; + not.b64 %rd14798, %rd14721; + and.b64 %rd14799, %rd14722, %rd14798; + xor.b64 %rd14800, %rd14799, %rd14725; + xor.b64 %rd14801, %rd14728, %rd33; + xor.b64 %rd14802, %rd14743, %rd14801; + xor.b64 %rd14803, %rd14802, %rd14758; + xor.b64 %rd14804, %rd14803, %rd14773; + xor.b64 %rd14805, %rd14804, %rd14788; + xor.b64 %rd14806, %rd14746, %rd14731; + xor.b64 %rd14807, %rd14806, %rd14761; + xor.b64 %rd14808, %rd14807, %rd14776; + xor.b64 %rd14809, %rd14808, %rd14791; + xor.b64 %rd14810, %rd14749, %rd14734; + xor.b64 %rd14811, %rd14810, %rd14764; + xor.b64 %rd14812, %rd14811, %rd14779; + xor.b64 %rd14813, %rd14812, %rd14794; + xor.b64 %rd14814, %rd14752, %rd14737; + xor.b64 %rd14815, %rd14814, %rd14767; + xor.b64 %rd14816, %rd14815, %rd14782; + xor.b64 %rd14817, %rd14816, %rd14797; + xor.b64 %rd14818, %rd14755, %rd14740; + xor.b64 %rd14819, %rd14818, %rd14770; + xor.b64 %rd14820, %rd14819, %rd14785; + xor.b64 %rd14821, %rd14820, %rd14800; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10524}, %rd14809; + } + { + .reg .b32 %dummy; + mov.b64 {%r10525,%dummy}, %rd14809; + } + shf.l.wrap.b32 %r10526, %r10525, %r10524, 1; + shf.l.wrap.b32 %r10527, %r10524, %r10525, 1; + mov.b64 %rd14822, {%r10527, %r10526}; + xor.b64 %rd14823, %rd14821, %rd14822; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10528}, %rd14813; + } + { + .reg .b32 %dummy; + mov.b64 {%r10529,%dummy}, %rd14813; + } + shf.l.wrap.b32 %r10530, %r10529, %r10528, 1; + shf.l.wrap.b32 %r10531, %r10528, %r10529, 1; + mov.b64 %rd14824, {%r10531, %r10530}; + xor.b64 %rd14825, %rd14824, %rd14805; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10532}, %rd14817; + } + { + .reg .b32 %dummy; + mov.b64 {%r10533,%dummy}, %rd14817; + } + shf.l.wrap.b32 %r10534, %r10533, %r10532, 1; + shf.l.wrap.b32 %r10535, %r10532, %r10533, 1; + mov.b64 %rd14826, {%r10535, %r10534}; + xor.b64 %rd14827, %rd14826, %rd14809; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10536}, %rd14821; + } + { + .reg .b32 %dummy; + mov.b64 {%r10537,%dummy}, %rd14821; + } + shf.l.wrap.b32 %r10538, %r10537, %r10536, 1; + shf.l.wrap.b32 %r10539, %r10536, %r10537, 1; + mov.b64 %rd14828, {%r10539, %r10538}; + xor.b64 %rd14829, %rd14828, %rd14813; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10540}, %rd14805; + } + { + .reg .b32 %dummy; + mov.b64 {%r10541,%dummy}, %rd14805; + } + shf.l.wrap.b32 %r10542, %r10541, %r10540, 1; + shf.l.wrap.b32 %r10543, %r10540, %r10541, 1; + mov.b64 %rd14830, {%r10543, %r10542}; + xor.b64 %rd14831, %rd14817, %rd14830; + xor.b64 %rd14832, %rd14823, %rd14801; + xor.b64 %rd14833, %rd14825, %rd14731; + xor.b64 %rd14834, %rd14827, %rd14734; + xor.b64 %rd14835, %rd14829, %rd14737; + xor.b64 %rd14836, %rd14831, %rd14740; + xor.b64 %rd14837, %rd14823, %rd14743; + xor.b64 %rd14838, %rd14825, %rd14746; + xor.b64 %rd14839, %rd14827, %rd14749; + xor.b64 %rd14840, %rd14829, %rd14752; + xor.b64 %rd14841, %rd14831, %rd14755; + xor.b64 %rd14842, %rd14823, %rd14758; + xor.b64 %rd14843, %rd14825, %rd14761; + xor.b64 %rd14844, %rd14827, %rd14764; + xor.b64 %rd14845, %rd14829, %rd14767; + xor.b64 %rd14846, %rd14831, %rd14770; + xor.b64 %rd14847, %rd14823, %rd14773; + xor.b64 %rd14848, %rd14825, %rd14776; + xor.b64 %rd14849, %rd14827, %rd14779; + xor.b64 %rd14850, %rd14829, %rd14782; + xor.b64 %rd14851, %rd14831, %rd14785; + xor.b64 %rd14852, %rd14823, %rd14788; + xor.b64 %rd14853, %rd14825, %rd14791; + xor.b64 %rd14854, %rd14827, %rd14794; + xor.b64 %rd14855, %rd14829, %rd14797; + xor.b64 %rd14856, %rd14831, %rd14800; + { + .reg .b32 %dummy; + mov.b64 {%r10544,%dummy}, %rd14838; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10545}, %rd14838; + } + shf.r.wrap.b32 %r10546, %r10545, %r10544, 20; + shf.r.wrap.b32 %r10547, %r10544, %r10545, 20; + mov.b64 %rd14857, {%r10547, %r10546}; + { + .reg .b32 %dummy; + mov.b64 {%r10548,%dummy}, %rd14844; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10549}, %rd14844; + } + shf.r.wrap.b32 %r10550, %r10549, %r10548, 21; + shf.r.wrap.b32 %r10551, %r10548, %r10549, 21; + mov.b64 %rd14858, {%r10551, %r10550}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10552}, %rd14850; + } + { + .reg .b32 %dummy; + mov.b64 {%r10553,%dummy}, %rd14850; + } + shf.l.wrap.b32 %r10554, %r10553, %r10552, 21; + shf.l.wrap.b32 %r10555, %r10552, %r10553, 21; + mov.b64 %rd14859, {%r10555, %r10554}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10556}, %rd14856; + } + { + .reg .b32 %dummy; + mov.b64 {%r10557,%dummy}, %rd14856; + } + shf.l.wrap.b32 %r10558, %r10557, %r10556, 14; + shf.l.wrap.b32 %r10559, %r10556, %r10557, 14; + mov.b64 %rd14860, {%r10559, %r10558}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10560}, %rd14835; + } + { + .reg .b32 %dummy; + mov.b64 {%r10561,%dummy}, %rd14835; + } + shf.l.wrap.b32 %r10562, %r10561, %r10560, 28; + shf.l.wrap.b32 %r10563, %r10560, %r10561, 28; + mov.b64 %rd14861, {%r10563, %r10562}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10564}, %rd14841; + } + { + .reg .b32 %dummy; + mov.b64 {%r10565,%dummy}, %rd14841; + } + shf.l.wrap.b32 %r10566, %r10565, %r10564, 20; + shf.l.wrap.b32 %r10567, %r10564, %r10565, 20; + mov.b64 %rd14862, {%r10567, %r10566}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10568}, %rd14842; + } + { + .reg .b32 %dummy; + mov.b64 {%r10569,%dummy}, %rd14842; + } + shf.l.wrap.b32 %r10570, %r10569, %r10568, 3; + shf.l.wrap.b32 %r10571, %r10568, %r10569, 3; + mov.b64 %rd14863, {%r10571, %r10570}; + { + .reg .b32 %dummy; + mov.b64 {%r10572,%dummy}, %rd14848; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10573}, %rd14848; + } + shf.r.wrap.b32 %r10574, %r10573, %r10572, 19; + shf.r.wrap.b32 %r10575, %r10572, %r10573, 19; + mov.b64 %rd14864, {%r10575, %r10574}; + { + .reg .b32 %dummy; + mov.b64 {%r10576,%dummy}, %rd14854; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10577}, %rd14854; + } + shf.r.wrap.b32 %r10578, %r10577, %r10576, 3; + shf.r.wrap.b32 %r10579, %r10576, %r10577, 3; + mov.b64 %rd14865, {%r10579, %r10578}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10580}, %rd14833; + } + { + .reg .b32 %dummy; + mov.b64 {%r10581,%dummy}, %rd14833; + } + shf.l.wrap.b32 %r10582, %r10581, %r10580, 1; + shf.l.wrap.b32 %r10583, %r10580, %r10581, 1; + mov.b64 %rd14866, {%r10583, %r10582}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10584}, %rd14839; + } + { + .reg .b32 %dummy; + mov.b64 {%r10585,%dummy}, %rd14839; + } + shf.l.wrap.b32 %r10586, %r10585, %r10584, 6; + shf.l.wrap.b32 %r10587, %r10584, %r10585, 6; + mov.b64 %rd14867, {%r10587, %r10586}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10588}, %rd14845; + } + { + .reg .b32 %dummy; + mov.b64 {%r10589,%dummy}, %rd14845; + } + shf.l.wrap.b32 %r10590, %r10589, %r10588, 25; + shf.l.wrap.b32 %r10591, %r10588, %r10589, 25; + mov.b64 %rd14868, {%r10591, %r10590}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10592}, %rd14851; + } + { + .reg .b32 %dummy; + mov.b64 {%r10593,%dummy}, %rd14851; + } + shf.l.wrap.b32 %r10594, %r10593, %r10592, 8; + shf.l.wrap.b32 %r10595, %r10592, %r10593, 8; + mov.b64 %rd14869, {%r10595, %r10594}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10596}, %rd14852; + } + { + .reg .b32 %dummy; + mov.b64 {%r10597,%dummy}, %rd14852; + } + shf.l.wrap.b32 %r10598, %r10597, %r10596, 18; + shf.l.wrap.b32 %r10599, %r10596, %r10597, 18; + mov.b64 %rd14870, {%r10599, %r10598}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10600}, %rd14836; + } + { + .reg .b32 %dummy; + mov.b64 {%r10601,%dummy}, %rd14836; + } + shf.l.wrap.b32 %r10602, %r10601, %r10600, 27; + shf.l.wrap.b32 %r10603, %r10600, %r10601, 27; + mov.b64 %rd14871, {%r10603, %r10602}; + { + .reg .b32 %dummy; + mov.b64 {%r10604,%dummy}, %rd14837; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10605}, %rd14837; + } + shf.r.wrap.b32 %r10606, %r10605, %r10604, 28; + shf.r.wrap.b32 %r10607, %r10604, %r10605, 28; + mov.b64 %rd14872, {%r10607, %r10606}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10608}, %rd14843; + } + { + .reg .b32 %dummy; + mov.b64 {%r10609,%dummy}, %rd14843; + } + shf.l.wrap.b32 %r10610, %r10609, %r10608, 10; + shf.l.wrap.b32 %r10611, %r10608, %r10609, 10; + mov.b64 %rd14873, {%r10611, %r10610}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10612}, %rd14849; + } + { + .reg .b32 %dummy; + mov.b64 {%r10613,%dummy}, %rd14849; + } + shf.l.wrap.b32 %r10614, %r10613, %r10612, 15; + shf.l.wrap.b32 %r10615, %r10612, %r10613, 15; + mov.b64 %rd14874, {%r10615, %r10614}; + { + .reg .b32 %dummy; + mov.b64 {%r10616,%dummy}, %rd14855; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10617}, %rd14855; + } + shf.r.wrap.b32 %r10618, %r10617, %r10616, 8; + shf.r.wrap.b32 %r10619, %r10616, %r10617, 8; + mov.b64 %rd14875, {%r10619, %r10618}; + { + .reg .b32 %dummy; + mov.b64 {%r10620,%dummy}, %rd14834; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10621}, %rd14834; + } + shf.r.wrap.b32 %r10622, %r10621, %r10620, 2; + shf.r.wrap.b32 %r10623, %r10620, %r10621, 2; + mov.b64 %rd14876, {%r10623, %r10622}; + { + .reg .b32 %dummy; + mov.b64 {%r10624,%dummy}, %rd14840; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10625}, %rd14840; + } + shf.r.wrap.b32 %r10626, %r10625, %r10624, 9; + shf.r.wrap.b32 %r10627, %r10624, %r10625, 9; + mov.b64 %rd14877, {%r10627, %r10626}; + { + .reg .b32 %dummy; + mov.b64 {%r10628,%dummy}, %rd14846; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10629}, %rd14846; + } + shf.r.wrap.b32 %r10630, %r10629, %r10628, 25; + shf.r.wrap.b32 %r10631, %r10628, %r10629, 25; + mov.b64 %rd14878, {%r10631, %r10630}; + { + .reg .b32 %dummy; + mov.b64 {%r10632,%dummy}, %rd14847; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10633}, %rd14847; + } + shf.r.wrap.b32 %r10634, %r10633, %r10632, 23; + shf.r.wrap.b32 %r10635, %r10632, %r10633, 23; + mov.b64 %rd14879, {%r10635, %r10634}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10636}, %rd14853; + } + { + .reg .b32 %dummy; + mov.b64 {%r10637,%dummy}, %rd14853; + } + shf.l.wrap.b32 %r10638, %r10637, %r10636, 2; + shf.l.wrap.b32 %r10639, %r10636, %r10637, 2; + mov.b64 %rd14880, {%r10639, %r10638}; + not.b64 %rd14881, %rd14857; + and.b64 %rd14882, %rd14858, %rd14881; + xor.b64 %rd14883, %rd14882, %rd14832; + not.b64 %rd14884, %rd14858; + and.b64 %rd14885, %rd14859, %rd14884; + xor.b64 %rd14886, %rd14885, %rd14857; + not.b64 %rd14887, %rd14859; + and.b64 %rd14888, %rd14860, %rd14887; + xor.b64 %rd14889, %rd14888, %rd14858; + not.b64 %rd14890, %rd14860; + and.b64 %rd14891, %rd14832, %rd14890; + xor.b64 %rd14892, %rd14891, %rd14859; + not.b64 %rd14893, %rd14832; + and.b64 %rd14894, %rd14857, %rd14893; + xor.b64 %rd14895, %rd14894, %rd14860; + not.b64 %rd14896, %rd14862; + and.b64 %rd14897, %rd14863, %rd14896; + xor.b64 %rd14898, %rd14897, %rd14861; + not.b64 %rd14899, %rd14863; + and.b64 %rd14900, %rd14864, %rd14899; + xor.b64 %rd14901, %rd14900, %rd14862; + not.b64 %rd14902, %rd14864; + and.b64 %rd14903, %rd14865, %rd14902; + xor.b64 %rd14904, %rd14903, %rd14863; + not.b64 %rd14905, %rd14865; + and.b64 %rd14906, %rd14861, %rd14905; + xor.b64 %rd14907, %rd14906, %rd14864; + not.b64 %rd14908, %rd14861; + and.b64 %rd14909, %rd14862, %rd14908; + xor.b64 %rd14910, %rd14909, %rd14865; + not.b64 %rd14911, %rd14867; + and.b64 %rd14912, %rd14868, %rd14911; + xor.b64 %rd14913, %rd14912, %rd14866; + not.b64 %rd14914, %rd14868; + and.b64 %rd14915, %rd14869, %rd14914; + xor.b64 %rd14916, %rd14915, %rd14867; + not.b64 %rd14917, %rd14869; + and.b64 %rd14918, %rd14870, %rd14917; + xor.b64 %rd14919, %rd14918, %rd14868; + not.b64 %rd14920, %rd14870; + and.b64 %rd14921, %rd14866, %rd14920; + xor.b64 %rd14922, %rd14921, %rd14869; + not.b64 %rd14923, %rd14866; + and.b64 %rd14924, %rd14867, %rd14923; + xor.b64 %rd14925, %rd14924, %rd14870; + not.b64 %rd14926, %rd14872; + and.b64 %rd14927, %rd14873, %rd14926; + xor.b64 %rd14928, %rd14927, %rd14871; + not.b64 %rd14929, %rd14873; + and.b64 %rd14930, %rd14874, %rd14929; + xor.b64 %rd14931, %rd14930, %rd14872; + not.b64 %rd14932, %rd14874; + and.b64 %rd14933, %rd14875, %rd14932; + xor.b64 %rd14934, %rd14933, %rd14873; + not.b64 %rd14935, %rd14875; + and.b64 %rd14936, %rd14871, %rd14935; + xor.b64 %rd14937, %rd14936, %rd14874; + not.b64 %rd14938, %rd14871; + and.b64 %rd14939, %rd14872, %rd14938; + xor.b64 %rd14940, %rd14939, %rd14875; + not.b64 %rd14941, %rd14877; + and.b64 %rd14942, %rd14878, %rd14941; + xor.b64 %rd14943, %rd14942, %rd14876; + not.b64 %rd14944, %rd14878; + and.b64 %rd14945, %rd14879, %rd14944; + xor.b64 %rd14946, %rd14945, %rd14877; + not.b64 %rd14947, %rd14879; + and.b64 %rd14948, %rd14880, %rd14947; + xor.b64 %rd14949, %rd14948, %rd14878; + not.b64 %rd14950, %rd14880; + and.b64 %rd14951, %rd14876, %rd14950; + xor.b64 %rd14952, %rd14951, %rd14879; + not.b64 %rd14953, %rd14876; + and.b64 %rd14954, %rd14877, %rd14953; + xor.b64 %rd14955, %rd14954, %rd14880; + xor.b64 %rd14956, %rd14883, %rd34; + xor.b64 %rd14957, %rd14898, %rd14956; + xor.b64 %rd14958, %rd14957, %rd14913; + xor.b64 %rd14959, %rd14958, %rd14928; + xor.b64 %rd14960, %rd14959, %rd14943; + xor.b64 %rd14961, %rd14901, %rd14886; + xor.b64 %rd14962, %rd14961, %rd14916; + xor.b64 %rd14963, %rd14962, %rd14931; + xor.b64 %rd14964, %rd14963, %rd14946; + xor.b64 %rd14965, %rd14904, %rd14889; + xor.b64 %rd14966, %rd14965, %rd14919; + xor.b64 %rd14967, %rd14966, %rd14934; + xor.b64 %rd14968, %rd14967, %rd14949; + xor.b64 %rd14969, %rd14907, %rd14892; + xor.b64 %rd14970, %rd14969, %rd14922; + xor.b64 %rd14971, %rd14970, %rd14937; + xor.b64 %rd14972, %rd14971, %rd14952; + xor.b64 %rd14973, %rd14910, %rd14895; + xor.b64 %rd14974, %rd14973, %rd14925; + xor.b64 %rd14975, %rd14974, %rd14940; + xor.b64 %rd14976, %rd14975, %rd14955; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10640}, %rd14964; + } + { + .reg .b32 %dummy; + mov.b64 {%r10641,%dummy}, %rd14964; + } + shf.l.wrap.b32 %r10642, %r10641, %r10640, 1; + shf.l.wrap.b32 %r10643, %r10640, %r10641, 1; + mov.b64 %rd14977, {%r10643, %r10642}; + xor.b64 %rd14978, %rd14976, %rd14977; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10644}, %rd14968; + } + { + .reg .b32 %dummy; + mov.b64 {%r10645,%dummy}, %rd14968; + } + shf.l.wrap.b32 %r10646, %r10645, %r10644, 1; + shf.l.wrap.b32 %r10647, %r10644, %r10645, 1; + mov.b64 %rd14979, {%r10647, %r10646}; + xor.b64 %rd14980, %rd14979, %rd14960; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10648}, %rd14972; + } + { + .reg .b32 %dummy; + mov.b64 {%r10649,%dummy}, %rd14972; + } + shf.l.wrap.b32 %r10650, %r10649, %r10648, 1; + shf.l.wrap.b32 %r10651, %r10648, %r10649, 1; + mov.b64 %rd14981, {%r10651, %r10650}; + xor.b64 %rd14982, %rd14981, %rd14964; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10652}, %rd14976; + } + { + .reg .b32 %dummy; + mov.b64 {%r10653,%dummy}, %rd14976; + } + shf.l.wrap.b32 %r10654, %r10653, %r10652, 1; + shf.l.wrap.b32 %r10655, %r10652, %r10653, 1; + mov.b64 %rd14983, {%r10655, %r10654}; + xor.b64 %rd14984, %rd14983, %rd14968; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10656}, %rd14960; + } + { + .reg .b32 %dummy; + mov.b64 {%r10657,%dummy}, %rd14960; + } + shf.l.wrap.b32 %r10658, %r10657, %r10656, 1; + shf.l.wrap.b32 %r10659, %r10656, %r10657, 1; + mov.b64 %rd14985, {%r10659, %r10658}; + xor.b64 %rd14986, %rd14972, %rd14985; + xor.b64 %rd14987, %rd14978, %rd14956; + xor.b64 %rd14988, %rd14980, %rd14886; + xor.b64 %rd14989, %rd14982, %rd14889; + xor.b64 %rd14990, %rd14984, %rd14892; + xor.b64 %rd14991, %rd14986, %rd14895; + xor.b64 %rd14992, %rd14978, %rd14898; + xor.b64 %rd14993, %rd14980, %rd14901; + xor.b64 %rd14994, %rd14982, %rd14904; + xor.b64 %rd14995, %rd14984, %rd14907; + xor.b64 %rd14996, %rd14986, %rd14910; + xor.b64 %rd14997, %rd14978, %rd14913; + xor.b64 %rd14998, %rd14980, %rd14916; + xor.b64 %rd14999, %rd14982, %rd14919; + xor.b64 %rd15000, %rd14984, %rd14922; + xor.b64 %rd15001, %rd14986, %rd14925; + xor.b64 %rd15002, %rd14978, %rd14928; + xor.b64 %rd15003, %rd14980, %rd14931; + xor.b64 %rd15004, %rd14982, %rd14934; + xor.b64 %rd15005, %rd14984, %rd14937; + xor.b64 %rd15006, %rd14986, %rd14940; + xor.b64 %rd15007, %rd14978, %rd14943; + xor.b64 %rd15008, %rd14980, %rd14946; + xor.b64 %rd15009, %rd14982, %rd14949; + xor.b64 %rd15010, %rd14984, %rd14952; + xor.b64 %rd15011, %rd14986, %rd14955; + { + .reg .b32 %dummy; + mov.b64 {%r10660,%dummy}, %rd14993; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10661}, %rd14993; + } + shf.r.wrap.b32 %r10662, %r10661, %r10660, 20; + shf.r.wrap.b32 %r10663, %r10660, %r10661, 20; + mov.b64 %rd15012, {%r10663, %r10662}; + { + .reg .b32 %dummy; + mov.b64 {%r10664,%dummy}, %rd14999; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10665}, %rd14999; + } + shf.r.wrap.b32 %r10666, %r10665, %r10664, 21; + shf.r.wrap.b32 %r10667, %r10664, %r10665, 21; + mov.b64 %rd15013, {%r10667, %r10666}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10668}, %rd15005; + } + { + .reg .b32 %dummy; + mov.b64 {%r10669,%dummy}, %rd15005; + } + shf.l.wrap.b32 %r10670, %r10669, %r10668, 21; + shf.l.wrap.b32 %r10671, %r10668, %r10669, 21; + mov.b64 %rd15014, {%r10671, %r10670}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10672}, %rd15011; + } + { + .reg .b32 %dummy; + mov.b64 {%r10673,%dummy}, %rd15011; + } + shf.l.wrap.b32 %r10674, %r10673, %r10672, 14; + shf.l.wrap.b32 %r10675, %r10672, %r10673, 14; + mov.b64 %rd15015, {%r10675, %r10674}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10676}, %rd14990; + } + { + .reg .b32 %dummy; + mov.b64 {%r10677,%dummy}, %rd14990; + } + shf.l.wrap.b32 %r10678, %r10677, %r10676, 28; + shf.l.wrap.b32 %r10679, %r10676, %r10677, 28; + mov.b64 %rd15016, {%r10679, %r10678}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10680}, %rd14996; + } + { + .reg .b32 %dummy; + mov.b64 {%r10681,%dummy}, %rd14996; + } + shf.l.wrap.b32 %r10682, %r10681, %r10680, 20; + shf.l.wrap.b32 %r10683, %r10680, %r10681, 20; + mov.b64 %rd15017, {%r10683, %r10682}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10684}, %rd14997; + } + { + .reg .b32 %dummy; + mov.b64 {%r10685,%dummy}, %rd14997; + } + shf.l.wrap.b32 %r10686, %r10685, %r10684, 3; + shf.l.wrap.b32 %r10687, %r10684, %r10685, 3; + mov.b64 %rd15018, {%r10687, %r10686}; + { + .reg .b32 %dummy; + mov.b64 {%r10688,%dummy}, %rd15003; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10689}, %rd15003; + } + shf.r.wrap.b32 %r10690, %r10689, %r10688, 19; + shf.r.wrap.b32 %r10691, %r10688, %r10689, 19; + mov.b64 %rd15019, {%r10691, %r10690}; + { + .reg .b32 %dummy; + mov.b64 {%r10692,%dummy}, %rd15009; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10693}, %rd15009; + } + shf.r.wrap.b32 %r10694, %r10693, %r10692, 3; + shf.r.wrap.b32 %r10695, %r10692, %r10693, 3; + mov.b64 %rd15020, {%r10695, %r10694}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10696}, %rd14988; + } + { + .reg .b32 %dummy; + mov.b64 {%r10697,%dummy}, %rd14988; + } + shf.l.wrap.b32 %r10698, %r10697, %r10696, 1; + shf.l.wrap.b32 %r10699, %r10696, %r10697, 1; + mov.b64 %rd15021, {%r10699, %r10698}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10700}, %rd14994; + } + { + .reg .b32 %dummy; + mov.b64 {%r10701,%dummy}, %rd14994; + } + shf.l.wrap.b32 %r10702, %r10701, %r10700, 6; + shf.l.wrap.b32 %r10703, %r10700, %r10701, 6; + mov.b64 %rd15022, {%r10703, %r10702}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10704}, %rd15000; + } + { + .reg .b32 %dummy; + mov.b64 {%r10705,%dummy}, %rd15000; + } + shf.l.wrap.b32 %r10706, %r10705, %r10704, 25; + shf.l.wrap.b32 %r10707, %r10704, %r10705, 25; + mov.b64 %rd15023, {%r10707, %r10706}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10708}, %rd15006; + } + { + .reg .b32 %dummy; + mov.b64 {%r10709,%dummy}, %rd15006; + } + shf.l.wrap.b32 %r10710, %r10709, %r10708, 8; + shf.l.wrap.b32 %r10711, %r10708, %r10709, 8; + mov.b64 %rd15024, {%r10711, %r10710}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10712}, %rd15007; + } + { + .reg .b32 %dummy; + mov.b64 {%r10713,%dummy}, %rd15007; + } + shf.l.wrap.b32 %r10714, %r10713, %r10712, 18; + shf.l.wrap.b32 %r10715, %r10712, %r10713, 18; + mov.b64 %rd15025, {%r10715, %r10714}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10716}, %rd14991; + } + { + .reg .b32 %dummy; + mov.b64 {%r10717,%dummy}, %rd14991; + } + shf.l.wrap.b32 %r10718, %r10717, %r10716, 27; + shf.l.wrap.b32 %r10719, %r10716, %r10717, 27; + mov.b64 %rd15026, {%r10719, %r10718}; + { + .reg .b32 %dummy; + mov.b64 {%r10720,%dummy}, %rd14992; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10721}, %rd14992; + } + shf.r.wrap.b32 %r10722, %r10721, %r10720, 28; + shf.r.wrap.b32 %r10723, %r10720, %r10721, 28; + mov.b64 %rd15027, {%r10723, %r10722}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10724}, %rd14998; + } + { + .reg .b32 %dummy; + mov.b64 {%r10725,%dummy}, %rd14998; + } + shf.l.wrap.b32 %r10726, %r10725, %r10724, 10; + shf.l.wrap.b32 %r10727, %r10724, %r10725, 10; + mov.b64 %rd15028, {%r10727, %r10726}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10728}, %rd15004; + } + { + .reg .b32 %dummy; + mov.b64 {%r10729,%dummy}, %rd15004; + } + shf.l.wrap.b32 %r10730, %r10729, %r10728, 15; + shf.l.wrap.b32 %r10731, %r10728, %r10729, 15; + mov.b64 %rd15029, {%r10731, %r10730}; + { + .reg .b32 %dummy; + mov.b64 {%r10732,%dummy}, %rd15010; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10733}, %rd15010; + } + shf.r.wrap.b32 %r10734, %r10733, %r10732, 8; + shf.r.wrap.b32 %r10735, %r10732, %r10733, 8; + mov.b64 %rd15030, {%r10735, %r10734}; + { + .reg .b32 %dummy; + mov.b64 {%r10736,%dummy}, %rd14989; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10737}, %rd14989; + } + shf.r.wrap.b32 %r10738, %r10737, %r10736, 2; + shf.r.wrap.b32 %r10739, %r10736, %r10737, 2; + mov.b64 %rd15031, {%r10739, %r10738}; + { + .reg .b32 %dummy; + mov.b64 {%r10740,%dummy}, %rd14995; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10741}, %rd14995; + } + shf.r.wrap.b32 %r10742, %r10741, %r10740, 9; + shf.r.wrap.b32 %r10743, %r10740, %r10741, 9; + mov.b64 %rd15032, {%r10743, %r10742}; + { + .reg .b32 %dummy; + mov.b64 {%r10744,%dummy}, %rd15001; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10745}, %rd15001; + } + shf.r.wrap.b32 %r10746, %r10745, %r10744, 25; + shf.r.wrap.b32 %r10747, %r10744, %r10745, 25; + mov.b64 %rd15033, {%r10747, %r10746}; + { + .reg .b32 %dummy; + mov.b64 {%r10748,%dummy}, %rd15002; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10749}, %rd15002; + } + shf.r.wrap.b32 %r10750, %r10749, %r10748, 23; + shf.r.wrap.b32 %r10751, %r10748, %r10749, 23; + mov.b64 %rd15034, {%r10751, %r10750}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10752}, %rd15008; + } + { + .reg .b32 %dummy; + mov.b64 {%r10753,%dummy}, %rd15008; + } + shf.l.wrap.b32 %r10754, %r10753, %r10752, 2; + shf.l.wrap.b32 %r10755, %r10752, %r10753, 2; + mov.b64 %rd15035, {%r10755, %r10754}; + not.b64 %rd15036, %rd15012; + and.b64 %rd15037, %rd15013, %rd15036; + xor.b64 %rd15038, %rd15037, %rd14987; + not.b64 %rd15039, %rd15013; + and.b64 %rd15040, %rd15014, %rd15039; + xor.b64 %rd15041, %rd15040, %rd15012; + not.b64 %rd15042, %rd15014; + and.b64 %rd15043, %rd15015, %rd15042; + xor.b64 %rd15044, %rd15043, %rd15013; + not.b64 %rd15045, %rd15015; + and.b64 %rd15046, %rd14987, %rd15045; + xor.b64 %rd15047, %rd15046, %rd15014; + not.b64 %rd15048, %rd14987; + and.b64 %rd15049, %rd15012, %rd15048; + xor.b64 %rd15050, %rd15049, %rd15015; + not.b64 %rd15051, %rd15017; + and.b64 %rd15052, %rd15018, %rd15051; + xor.b64 %rd15053, %rd15052, %rd15016; + not.b64 %rd15054, %rd15018; + and.b64 %rd15055, %rd15019, %rd15054; + xor.b64 %rd15056, %rd15055, %rd15017; + not.b64 %rd15057, %rd15019; + and.b64 %rd15058, %rd15020, %rd15057; + xor.b64 %rd15059, %rd15058, %rd15018; + not.b64 %rd15060, %rd15020; + and.b64 %rd15061, %rd15016, %rd15060; + xor.b64 %rd15062, %rd15061, %rd15019; + not.b64 %rd15063, %rd15016; + and.b64 %rd15064, %rd15017, %rd15063; + xor.b64 %rd15065, %rd15064, %rd15020; + not.b64 %rd15066, %rd15022; + and.b64 %rd15067, %rd15023, %rd15066; + xor.b64 %rd15068, %rd15067, %rd15021; + not.b64 %rd15069, %rd15023; + and.b64 %rd15070, %rd15024, %rd15069; + xor.b64 %rd15071, %rd15070, %rd15022; + not.b64 %rd15072, %rd15024; + and.b64 %rd15073, %rd15025, %rd15072; + xor.b64 %rd15074, %rd15073, %rd15023; + not.b64 %rd15075, %rd15025; + and.b64 %rd15076, %rd15021, %rd15075; + xor.b64 %rd15077, %rd15076, %rd15024; + not.b64 %rd15078, %rd15021; + and.b64 %rd15079, %rd15022, %rd15078; + xor.b64 %rd15080, %rd15079, %rd15025; + not.b64 %rd15081, %rd15027; + and.b64 %rd15082, %rd15028, %rd15081; + xor.b64 %rd15083, %rd15082, %rd15026; + not.b64 %rd15084, %rd15028; + and.b64 %rd15085, %rd15029, %rd15084; + xor.b64 %rd15086, %rd15085, %rd15027; + not.b64 %rd15087, %rd15029; + and.b64 %rd15088, %rd15030, %rd15087; + xor.b64 %rd15089, %rd15088, %rd15028; + not.b64 %rd15090, %rd15030; + and.b64 %rd15091, %rd15026, %rd15090; + xor.b64 %rd15092, %rd15091, %rd15029; + not.b64 %rd15093, %rd15026; + and.b64 %rd15094, %rd15027, %rd15093; + xor.b64 %rd15095, %rd15094, %rd15030; + not.b64 %rd15096, %rd15032; + and.b64 %rd15097, %rd15033, %rd15096; + xor.b64 %rd15098, %rd15097, %rd15031; + not.b64 %rd15099, %rd15033; + and.b64 %rd15100, %rd15034, %rd15099; + xor.b64 %rd15101, %rd15100, %rd15032; + not.b64 %rd15102, %rd15034; + and.b64 %rd15103, %rd15035, %rd15102; + xor.b64 %rd15104, %rd15103, %rd15033; + not.b64 %rd15105, %rd15035; + and.b64 %rd15106, %rd15031, %rd15105; + xor.b64 %rd15107, %rd15106, %rd15034; + not.b64 %rd15108, %rd15031; + and.b64 %rd15109, %rd15032, %rd15108; + xor.b64 %rd15110, %rd15109, %rd15035; + xor.b64 %rd15111, %rd15038, %rd35; + xor.b64 %rd15112, %rd15053, %rd15111; + xor.b64 %rd15113, %rd15112, %rd15068; + xor.b64 %rd15114, %rd15113, %rd15083; + xor.b64 %rd15115, %rd15114, %rd15098; + xor.b64 %rd15116, %rd15056, %rd15041; + xor.b64 %rd15117, %rd15116, %rd15071; + xor.b64 %rd15118, %rd15117, %rd15086; + xor.b64 %rd15119, %rd15118, %rd15101; + xor.b64 %rd15120, %rd15059, %rd15044; + xor.b64 %rd15121, %rd15120, %rd15074; + xor.b64 %rd15122, %rd15121, %rd15089; + xor.b64 %rd15123, %rd15122, %rd15104; + xor.b64 %rd15124, %rd15062, %rd15047; + xor.b64 %rd15125, %rd15124, %rd15077; + xor.b64 %rd15126, %rd15125, %rd15092; + xor.b64 %rd15127, %rd15126, %rd15107; + xor.b64 %rd15128, %rd15065, %rd15050; + xor.b64 %rd15129, %rd15128, %rd15080; + xor.b64 %rd15130, %rd15129, %rd15095; + xor.b64 %rd15131, %rd15130, %rd15110; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10756}, %rd15119; + } + { + .reg .b32 %dummy; + mov.b64 {%r10757,%dummy}, %rd15119; + } + shf.l.wrap.b32 %r10758, %r10757, %r10756, 1; + shf.l.wrap.b32 %r10759, %r10756, %r10757, 1; + mov.b64 %rd15132, {%r10759, %r10758}; + xor.b64 %rd15133, %rd15131, %rd15132; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10760}, %rd15123; + } + { + .reg .b32 %dummy; + mov.b64 {%r10761,%dummy}, %rd15123; + } + shf.l.wrap.b32 %r10762, %r10761, %r10760, 1; + shf.l.wrap.b32 %r10763, %r10760, %r10761, 1; + mov.b64 %rd15134, {%r10763, %r10762}; + xor.b64 %rd15135, %rd15134, %rd15115; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10764}, %rd15127; + } + { + .reg .b32 %dummy; + mov.b64 {%r10765,%dummy}, %rd15127; + } + shf.l.wrap.b32 %r10766, %r10765, %r10764, 1; + shf.l.wrap.b32 %r10767, %r10764, %r10765, 1; + mov.b64 %rd15136, {%r10767, %r10766}; + xor.b64 %rd15137, %rd15136, %rd15119; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10768}, %rd15131; + } + { + .reg .b32 %dummy; + mov.b64 {%r10769,%dummy}, %rd15131; + } + shf.l.wrap.b32 %r10770, %r10769, %r10768, 1; + shf.l.wrap.b32 %r10771, %r10768, %r10769, 1; + mov.b64 %rd15138, {%r10771, %r10770}; + xor.b64 %rd15139, %rd15138, %rd15123; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10772}, %rd15115; + } + { + .reg .b32 %dummy; + mov.b64 {%r10773,%dummy}, %rd15115; + } + shf.l.wrap.b32 %r10774, %r10773, %r10772, 1; + shf.l.wrap.b32 %r10775, %r10772, %r10773, 1; + mov.b64 %rd15140, {%r10775, %r10774}; + xor.b64 %rd15141, %rd15127, %rd15140; + xor.b64 %rd15142, %rd15133, %rd15111; + xor.b64 %rd15143, %rd15135, %rd15041; + xor.b64 %rd15144, %rd15137, %rd15044; + xor.b64 %rd15145, %rd15139, %rd15047; + xor.b64 %rd15146, %rd15141, %rd15050; + xor.b64 %rd15147, %rd15133, %rd15053; + xor.b64 %rd15148, %rd15135, %rd15056; + xor.b64 %rd15149, %rd15137, %rd15059; + xor.b64 %rd15150, %rd15139, %rd15062; + xor.b64 %rd15151, %rd15141, %rd15065; + xor.b64 %rd15152, %rd15133, %rd15068; + xor.b64 %rd15153, %rd15135, %rd15071; + xor.b64 %rd15154, %rd15137, %rd15074; + xor.b64 %rd15155, %rd15139, %rd15077; + xor.b64 %rd15156, %rd15141, %rd15080; + xor.b64 %rd15157, %rd15133, %rd15083; + xor.b64 %rd15158, %rd15135, %rd15086; + xor.b64 %rd15159, %rd15137, %rd15089; + xor.b64 %rd15160, %rd15139, %rd15092; + xor.b64 %rd15161, %rd15141, %rd15095; + xor.b64 %rd15162, %rd15133, %rd15098; + xor.b64 %rd15163, %rd15135, %rd15101; + xor.b64 %rd15164, %rd15137, %rd15104; + xor.b64 %rd15165, %rd15139, %rd15107; + xor.b64 %rd15166, %rd15141, %rd15110; + { + .reg .b32 %dummy; + mov.b64 {%r10776,%dummy}, %rd15148; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10777}, %rd15148; + } + shf.r.wrap.b32 %r10778, %r10777, %r10776, 20; + shf.r.wrap.b32 %r10779, %r10776, %r10777, 20; + mov.b64 %rd15167, {%r10779, %r10778}; + { + .reg .b32 %dummy; + mov.b64 {%r10780,%dummy}, %rd15154; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10781}, %rd15154; + } + shf.r.wrap.b32 %r10782, %r10781, %r10780, 21; + shf.r.wrap.b32 %r10783, %r10780, %r10781, 21; + mov.b64 %rd15168, {%r10783, %r10782}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10784}, %rd15160; + } + { + .reg .b32 %dummy; + mov.b64 {%r10785,%dummy}, %rd15160; + } + shf.l.wrap.b32 %r10786, %r10785, %r10784, 21; + shf.l.wrap.b32 %r10787, %r10784, %r10785, 21; + mov.b64 %rd15169, {%r10787, %r10786}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10788}, %rd15166; + } + { + .reg .b32 %dummy; + mov.b64 {%r10789,%dummy}, %rd15166; + } + shf.l.wrap.b32 %r10790, %r10789, %r10788, 14; + shf.l.wrap.b32 %r10791, %r10788, %r10789, 14; + mov.b64 %rd15170, {%r10791, %r10790}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10792}, %rd15145; + } + { + .reg .b32 %dummy; + mov.b64 {%r10793,%dummy}, %rd15145; + } + shf.l.wrap.b32 %r10794, %r10793, %r10792, 28; + shf.l.wrap.b32 %r10795, %r10792, %r10793, 28; + mov.b64 %rd15171, {%r10795, %r10794}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10796}, %rd15151; + } + { + .reg .b32 %dummy; + mov.b64 {%r10797,%dummy}, %rd15151; + } + shf.l.wrap.b32 %r10798, %r10797, %r10796, 20; + shf.l.wrap.b32 %r10799, %r10796, %r10797, 20; + mov.b64 %rd15172, {%r10799, %r10798}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10800}, %rd15152; + } + { + .reg .b32 %dummy; + mov.b64 {%r10801,%dummy}, %rd15152; + } + shf.l.wrap.b32 %r10802, %r10801, %r10800, 3; + shf.l.wrap.b32 %r10803, %r10800, %r10801, 3; + mov.b64 %rd15173, {%r10803, %r10802}; + { + .reg .b32 %dummy; + mov.b64 {%r10804,%dummy}, %rd15158; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10805}, %rd15158; + } + shf.r.wrap.b32 %r10806, %r10805, %r10804, 19; + shf.r.wrap.b32 %r10807, %r10804, %r10805, 19; + mov.b64 %rd15174, {%r10807, %r10806}; + { + .reg .b32 %dummy; + mov.b64 {%r10808,%dummy}, %rd15164; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10809}, %rd15164; + } + shf.r.wrap.b32 %r10810, %r10809, %r10808, 3; + shf.r.wrap.b32 %r10811, %r10808, %r10809, 3; + mov.b64 %rd15175, {%r10811, %r10810}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10812}, %rd15143; + } + { + .reg .b32 %dummy; + mov.b64 {%r10813,%dummy}, %rd15143; + } + shf.l.wrap.b32 %r10814, %r10813, %r10812, 1; + shf.l.wrap.b32 %r10815, %r10812, %r10813, 1; + mov.b64 %rd15176, {%r10815, %r10814}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10816}, %rd15149; + } + { + .reg .b32 %dummy; + mov.b64 {%r10817,%dummy}, %rd15149; + } + shf.l.wrap.b32 %r10818, %r10817, %r10816, 6; + shf.l.wrap.b32 %r10819, %r10816, %r10817, 6; + mov.b64 %rd15177, {%r10819, %r10818}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10820}, %rd15155; + } + { + .reg .b32 %dummy; + mov.b64 {%r10821,%dummy}, %rd15155; + } + shf.l.wrap.b32 %r10822, %r10821, %r10820, 25; + shf.l.wrap.b32 %r10823, %r10820, %r10821, 25; + mov.b64 %rd15178, {%r10823, %r10822}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10824}, %rd15161; + } + { + .reg .b32 %dummy; + mov.b64 {%r10825,%dummy}, %rd15161; + } + shf.l.wrap.b32 %r10826, %r10825, %r10824, 8; + shf.l.wrap.b32 %r10827, %r10824, %r10825, 8; + mov.b64 %rd15179, {%r10827, %r10826}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10828}, %rd15162; + } + { + .reg .b32 %dummy; + mov.b64 {%r10829,%dummy}, %rd15162; + } + shf.l.wrap.b32 %r10830, %r10829, %r10828, 18; + shf.l.wrap.b32 %r10831, %r10828, %r10829, 18; + mov.b64 %rd15180, {%r10831, %r10830}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10832}, %rd15146; + } + { + .reg .b32 %dummy; + mov.b64 {%r10833,%dummy}, %rd15146; + } + shf.l.wrap.b32 %r10834, %r10833, %r10832, 27; + shf.l.wrap.b32 %r10835, %r10832, %r10833, 27; + mov.b64 %rd15181, {%r10835, %r10834}; + { + .reg .b32 %dummy; + mov.b64 {%r10836,%dummy}, %rd15147; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10837}, %rd15147; + } + shf.r.wrap.b32 %r10838, %r10837, %r10836, 28; + shf.r.wrap.b32 %r10839, %r10836, %r10837, 28; + mov.b64 %rd15182, {%r10839, %r10838}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10840}, %rd15153; + } + { + .reg .b32 %dummy; + mov.b64 {%r10841,%dummy}, %rd15153; + } + shf.l.wrap.b32 %r10842, %r10841, %r10840, 10; + shf.l.wrap.b32 %r10843, %r10840, %r10841, 10; + mov.b64 %rd15183, {%r10843, %r10842}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10844}, %rd15159; + } + { + .reg .b32 %dummy; + mov.b64 {%r10845,%dummy}, %rd15159; + } + shf.l.wrap.b32 %r10846, %r10845, %r10844, 15; + shf.l.wrap.b32 %r10847, %r10844, %r10845, 15; + mov.b64 %rd15184, {%r10847, %r10846}; + { + .reg .b32 %dummy; + mov.b64 {%r10848,%dummy}, %rd15165; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10849}, %rd15165; + } + shf.r.wrap.b32 %r10850, %r10849, %r10848, 8; + shf.r.wrap.b32 %r10851, %r10848, %r10849, 8; + mov.b64 %rd15185, {%r10851, %r10850}; + { + .reg .b32 %dummy; + mov.b64 {%r10852,%dummy}, %rd15144; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10853}, %rd15144; + } + shf.r.wrap.b32 %r10854, %r10853, %r10852, 2; + shf.r.wrap.b32 %r10855, %r10852, %r10853, 2; + mov.b64 %rd15186, {%r10855, %r10854}; + { + .reg .b32 %dummy; + mov.b64 {%r10856,%dummy}, %rd15150; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10857}, %rd15150; + } + shf.r.wrap.b32 %r10858, %r10857, %r10856, 9; + shf.r.wrap.b32 %r10859, %r10856, %r10857, 9; + mov.b64 %rd15187, {%r10859, %r10858}; + { + .reg .b32 %dummy; + mov.b64 {%r10860,%dummy}, %rd15156; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10861}, %rd15156; + } + shf.r.wrap.b32 %r10862, %r10861, %r10860, 25; + shf.r.wrap.b32 %r10863, %r10860, %r10861, 25; + mov.b64 %rd15188, {%r10863, %r10862}; + { + .reg .b32 %dummy; + mov.b64 {%r10864,%dummy}, %rd15157; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10865}, %rd15157; + } + shf.r.wrap.b32 %r10866, %r10865, %r10864, 23; + shf.r.wrap.b32 %r10867, %r10864, %r10865, 23; + mov.b64 %rd15189, {%r10867, %r10866}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10868}, %rd15163; + } + { + .reg .b32 %dummy; + mov.b64 {%r10869,%dummy}, %rd15163; + } + shf.l.wrap.b32 %r10870, %r10869, %r10868, 2; + shf.l.wrap.b32 %r10871, %r10868, %r10869, 2; + mov.b64 %rd15190, {%r10871, %r10870}; + not.b64 %rd15191, %rd15167; + and.b64 %rd15192, %rd15168, %rd15191; + xor.b64 %rd15193, %rd15192, %rd15142; + not.b64 %rd15194, %rd15168; + and.b64 %rd15195, %rd15169, %rd15194; + xor.b64 %rd15196, %rd15195, %rd15167; + not.b64 %rd15197, %rd15169; + and.b64 %rd15198, %rd15170, %rd15197; + xor.b64 %rd15199, %rd15198, %rd15168; + not.b64 %rd15200, %rd15170; + and.b64 %rd15201, %rd15142, %rd15200; + xor.b64 %rd15202, %rd15201, %rd15169; + not.b64 %rd15203, %rd15142; + and.b64 %rd15204, %rd15167, %rd15203; + xor.b64 %rd15205, %rd15204, %rd15170; + not.b64 %rd15206, %rd15172; + and.b64 %rd15207, %rd15173, %rd15206; + xor.b64 %rd15208, %rd15207, %rd15171; + not.b64 %rd15209, %rd15173; + and.b64 %rd15210, %rd15174, %rd15209; + xor.b64 %rd15211, %rd15210, %rd15172; + not.b64 %rd15212, %rd15174; + and.b64 %rd15213, %rd15175, %rd15212; + xor.b64 %rd15214, %rd15213, %rd15173; + not.b64 %rd15215, %rd15175; + and.b64 %rd15216, %rd15171, %rd15215; + xor.b64 %rd15217, %rd15216, %rd15174; + not.b64 %rd15218, %rd15171; + and.b64 %rd15219, %rd15172, %rd15218; + xor.b64 %rd15220, %rd15219, %rd15175; + not.b64 %rd15221, %rd15177; + and.b64 %rd15222, %rd15178, %rd15221; + xor.b64 %rd15223, %rd15222, %rd15176; + not.b64 %rd15224, %rd15178; + and.b64 %rd15225, %rd15179, %rd15224; + xor.b64 %rd15226, %rd15225, %rd15177; + not.b64 %rd15227, %rd15179; + and.b64 %rd15228, %rd15180, %rd15227; + xor.b64 %rd15229, %rd15228, %rd15178; + not.b64 %rd15230, %rd15180; + and.b64 %rd15231, %rd15176, %rd15230; + xor.b64 %rd15232, %rd15231, %rd15179; + not.b64 %rd15233, %rd15176; + and.b64 %rd15234, %rd15177, %rd15233; + xor.b64 %rd15235, %rd15234, %rd15180; + not.b64 %rd15236, %rd15182; + and.b64 %rd15237, %rd15183, %rd15236; + xor.b64 %rd15238, %rd15237, %rd15181; + not.b64 %rd15239, %rd15183; + and.b64 %rd15240, %rd15184, %rd15239; + xor.b64 %rd15241, %rd15240, %rd15182; + not.b64 %rd15242, %rd15184; + and.b64 %rd15243, %rd15185, %rd15242; + xor.b64 %rd15244, %rd15243, %rd15183; + not.b64 %rd15245, %rd15185; + and.b64 %rd15246, %rd15181, %rd15245; + xor.b64 %rd15247, %rd15246, %rd15184; + not.b64 %rd15248, %rd15181; + and.b64 %rd15249, %rd15182, %rd15248; + xor.b64 %rd15250, %rd15249, %rd15185; + not.b64 %rd15251, %rd15187; + and.b64 %rd15252, %rd15188, %rd15251; + xor.b64 %rd15253, %rd15252, %rd15186; + not.b64 %rd15254, %rd15188; + and.b64 %rd15255, %rd15189, %rd15254; + xor.b64 %rd15256, %rd15255, %rd15187; + not.b64 %rd15257, %rd15189; + and.b64 %rd15258, %rd15190, %rd15257; + xor.b64 %rd15259, %rd15258, %rd15188; + not.b64 %rd15260, %rd15190; + and.b64 %rd15261, %rd15186, %rd15260; + xor.b64 %rd15262, %rd15261, %rd15189; + not.b64 %rd15263, %rd15186; + and.b64 %rd15264, %rd15187, %rd15263; + xor.b64 %rd15265, %rd15264, %rd15190; + xor.b64 %rd15266, %rd15193, %rd36; + xor.b64 %rd15267, %rd15208, %rd15266; + xor.b64 %rd15268, %rd15267, %rd15223; + xor.b64 %rd15269, %rd15268, %rd15238; + xor.b64 %rd15270, %rd15269, %rd15253; + xor.b64 %rd15271, %rd15211, %rd15196; + xor.b64 %rd15272, %rd15271, %rd15226; + xor.b64 %rd15273, %rd15272, %rd15241; + xor.b64 %rd15274, %rd15273, %rd15256; + xor.b64 %rd15275, %rd15214, %rd15199; + xor.b64 %rd15276, %rd15275, %rd15229; + xor.b64 %rd15277, %rd15276, %rd15244; + xor.b64 %rd15278, %rd15277, %rd15259; + xor.b64 %rd15279, %rd15217, %rd15202; + xor.b64 %rd15280, %rd15279, %rd15232; + xor.b64 %rd15281, %rd15280, %rd15247; + xor.b64 %rd15282, %rd15281, %rd15262; + xor.b64 %rd15283, %rd15220, %rd15205; + xor.b64 %rd15284, %rd15283, %rd15235; + xor.b64 %rd15285, %rd15284, %rd15250; + xor.b64 %rd15286, %rd15285, %rd15265; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10872}, %rd15274; + } + { + .reg .b32 %dummy; + mov.b64 {%r10873,%dummy}, %rd15274; + } + shf.l.wrap.b32 %r10874, %r10873, %r10872, 1; + shf.l.wrap.b32 %r10875, %r10872, %r10873, 1; + mov.b64 %rd15287, {%r10875, %r10874}; + xor.b64 %rd15288, %rd15286, %rd15287; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10876}, %rd15278; + } + { + .reg .b32 %dummy; + mov.b64 {%r10877,%dummy}, %rd15278; + } + shf.l.wrap.b32 %r10878, %r10877, %r10876, 1; + shf.l.wrap.b32 %r10879, %r10876, %r10877, 1; + mov.b64 %rd15289, {%r10879, %r10878}; + xor.b64 %rd15290, %rd15289, %rd15270; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10880}, %rd15282; + } + { + .reg .b32 %dummy; + mov.b64 {%r10881,%dummy}, %rd15282; + } + shf.l.wrap.b32 %r10882, %r10881, %r10880, 1; + shf.l.wrap.b32 %r10883, %r10880, %r10881, 1; + mov.b64 %rd15291, {%r10883, %r10882}; + xor.b64 %rd15292, %rd15291, %rd15274; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10884}, %rd15286; + } + { + .reg .b32 %dummy; + mov.b64 {%r10885,%dummy}, %rd15286; + } + shf.l.wrap.b32 %r10886, %r10885, %r10884, 1; + shf.l.wrap.b32 %r10887, %r10884, %r10885, 1; + mov.b64 %rd15293, {%r10887, %r10886}; + xor.b64 %rd15294, %rd15293, %rd15278; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10888}, %rd15270; + } + { + .reg .b32 %dummy; + mov.b64 {%r10889,%dummy}, %rd15270; + } + shf.l.wrap.b32 %r10890, %r10889, %r10888, 1; + shf.l.wrap.b32 %r10891, %r10888, %r10889, 1; + mov.b64 %rd15295, {%r10891, %r10890}; + xor.b64 %rd15296, %rd15282, %rd15295; + xor.b64 %rd15297, %rd15288, %rd15266; + xor.b64 %rd15298, %rd15290, %rd15196; + xor.b64 %rd15299, %rd15292, %rd15199; + xor.b64 %rd15300, %rd15294, %rd15202; + xor.b64 %rd15301, %rd15296, %rd15205; + xor.b64 %rd15302, %rd15288, %rd15208; + xor.b64 %rd15303, %rd15290, %rd15211; + xor.b64 %rd15304, %rd15292, %rd15214; + xor.b64 %rd15305, %rd15294, %rd15217; + xor.b64 %rd15306, %rd15296, %rd15220; + xor.b64 %rd15307, %rd15288, %rd15223; + xor.b64 %rd15308, %rd15290, %rd15226; + xor.b64 %rd15309, %rd15292, %rd15229; + xor.b64 %rd15310, %rd15294, %rd15232; + xor.b64 %rd15311, %rd15296, %rd15235; + xor.b64 %rd15312, %rd15288, %rd15238; + xor.b64 %rd15313, %rd15290, %rd15241; + xor.b64 %rd15314, %rd15292, %rd15244; + xor.b64 %rd15315, %rd15294, %rd15247; + xor.b64 %rd15316, %rd15296, %rd15250; + xor.b64 %rd15317, %rd15288, %rd15253; + xor.b64 %rd15318, %rd15290, %rd15256; + xor.b64 %rd15319, %rd15292, %rd15259; + xor.b64 %rd15320, %rd15294, %rd15262; + xor.b64 %rd15321, %rd15296, %rd15265; + { + .reg .b32 %dummy; + mov.b64 {%r10892,%dummy}, %rd15303; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10893}, %rd15303; + } + shf.r.wrap.b32 %r10894, %r10893, %r10892, 20; + shf.r.wrap.b32 %r10895, %r10892, %r10893, 20; + mov.b64 %rd15322, {%r10895, %r10894}; + { + .reg .b32 %dummy; + mov.b64 {%r10896,%dummy}, %rd15309; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10897}, %rd15309; + } + shf.r.wrap.b32 %r10898, %r10897, %r10896, 21; + shf.r.wrap.b32 %r10899, %r10896, %r10897, 21; + mov.b64 %rd15323, {%r10899, %r10898}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10900}, %rd15315; + } + { + .reg .b32 %dummy; + mov.b64 {%r10901,%dummy}, %rd15315; + } + shf.l.wrap.b32 %r10902, %r10901, %r10900, 21; + shf.l.wrap.b32 %r10903, %r10900, %r10901, 21; + mov.b64 %rd15324, {%r10903, %r10902}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10904}, %rd15321; + } + { + .reg .b32 %dummy; + mov.b64 {%r10905,%dummy}, %rd15321; + } + shf.l.wrap.b32 %r10906, %r10905, %r10904, 14; + shf.l.wrap.b32 %r10907, %r10904, %r10905, 14; + mov.b64 %rd15325, {%r10907, %r10906}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10908}, %rd15300; + } + { + .reg .b32 %dummy; + mov.b64 {%r10909,%dummy}, %rd15300; + } + shf.l.wrap.b32 %r10910, %r10909, %r10908, 28; + shf.l.wrap.b32 %r10911, %r10908, %r10909, 28; + mov.b64 %rd15326, {%r10911, %r10910}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10912}, %rd15306; + } + { + .reg .b32 %dummy; + mov.b64 {%r10913,%dummy}, %rd15306; + } + shf.l.wrap.b32 %r10914, %r10913, %r10912, 20; + shf.l.wrap.b32 %r10915, %r10912, %r10913, 20; + mov.b64 %rd15327, {%r10915, %r10914}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10916}, %rd15307; + } + { + .reg .b32 %dummy; + mov.b64 {%r10917,%dummy}, %rd15307; + } + shf.l.wrap.b32 %r10918, %r10917, %r10916, 3; + shf.l.wrap.b32 %r10919, %r10916, %r10917, 3; + mov.b64 %rd15328, {%r10919, %r10918}; + { + .reg .b32 %dummy; + mov.b64 {%r10920,%dummy}, %rd15313; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10921}, %rd15313; + } + shf.r.wrap.b32 %r10922, %r10921, %r10920, 19; + shf.r.wrap.b32 %r10923, %r10920, %r10921, 19; + mov.b64 %rd15329, {%r10923, %r10922}; + { + .reg .b32 %dummy; + mov.b64 {%r10924,%dummy}, %rd15319; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10925}, %rd15319; + } + shf.r.wrap.b32 %r10926, %r10925, %r10924, 3; + shf.r.wrap.b32 %r10927, %r10924, %r10925, 3; + mov.b64 %rd15330, {%r10927, %r10926}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10928}, %rd15298; + } + { + .reg .b32 %dummy; + mov.b64 {%r10929,%dummy}, %rd15298; + } + shf.l.wrap.b32 %r10930, %r10929, %r10928, 1; + shf.l.wrap.b32 %r10931, %r10928, %r10929, 1; + mov.b64 %rd15331, {%r10931, %r10930}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10932}, %rd15304; + } + { + .reg .b32 %dummy; + mov.b64 {%r10933,%dummy}, %rd15304; + } + shf.l.wrap.b32 %r10934, %r10933, %r10932, 6; + shf.l.wrap.b32 %r10935, %r10932, %r10933, 6; + mov.b64 %rd15332, {%r10935, %r10934}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10936}, %rd15310; + } + { + .reg .b32 %dummy; + mov.b64 {%r10937,%dummy}, %rd15310; + } + shf.l.wrap.b32 %r10938, %r10937, %r10936, 25; + shf.l.wrap.b32 %r10939, %r10936, %r10937, 25; + mov.b64 %rd15333, {%r10939, %r10938}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10940}, %rd15316; + } + { + .reg .b32 %dummy; + mov.b64 {%r10941,%dummy}, %rd15316; + } + shf.l.wrap.b32 %r10942, %r10941, %r10940, 8; + shf.l.wrap.b32 %r10943, %r10940, %r10941, 8; + mov.b64 %rd15334, {%r10943, %r10942}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10944}, %rd15317; + } + { + .reg .b32 %dummy; + mov.b64 {%r10945,%dummy}, %rd15317; + } + shf.l.wrap.b32 %r10946, %r10945, %r10944, 18; + shf.l.wrap.b32 %r10947, %r10944, %r10945, 18; + mov.b64 %rd15335, {%r10947, %r10946}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10948}, %rd15301; + } + { + .reg .b32 %dummy; + mov.b64 {%r10949,%dummy}, %rd15301; + } + shf.l.wrap.b32 %r10950, %r10949, %r10948, 27; + shf.l.wrap.b32 %r10951, %r10948, %r10949, 27; + mov.b64 %rd15336, {%r10951, %r10950}; + { + .reg .b32 %dummy; + mov.b64 {%r10952,%dummy}, %rd15302; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10953}, %rd15302; + } + shf.r.wrap.b32 %r10954, %r10953, %r10952, 28; + shf.r.wrap.b32 %r10955, %r10952, %r10953, 28; + mov.b64 %rd15337, {%r10955, %r10954}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10956}, %rd15308; + } + { + .reg .b32 %dummy; + mov.b64 {%r10957,%dummy}, %rd15308; + } + shf.l.wrap.b32 %r10958, %r10957, %r10956, 10; + shf.l.wrap.b32 %r10959, %r10956, %r10957, 10; + mov.b64 %rd15338, {%r10959, %r10958}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10960}, %rd15314; + } + { + .reg .b32 %dummy; + mov.b64 {%r10961,%dummy}, %rd15314; + } + shf.l.wrap.b32 %r10962, %r10961, %r10960, 15; + shf.l.wrap.b32 %r10963, %r10960, %r10961, 15; + mov.b64 %rd15339, {%r10963, %r10962}; + { + .reg .b32 %dummy; + mov.b64 {%r10964,%dummy}, %rd15320; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10965}, %rd15320; + } + shf.r.wrap.b32 %r10966, %r10965, %r10964, 8; + shf.r.wrap.b32 %r10967, %r10964, %r10965, 8; + mov.b64 %rd15340, {%r10967, %r10966}; + { + .reg .b32 %dummy; + mov.b64 {%r10968,%dummy}, %rd15299; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10969}, %rd15299; + } + shf.r.wrap.b32 %r10970, %r10969, %r10968, 2; + shf.r.wrap.b32 %r10971, %r10968, %r10969, 2; + mov.b64 %rd15341, {%r10971, %r10970}; + { + .reg .b32 %dummy; + mov.b64 {%r10972,%dummy}, %rd15305; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10973}, %rd15305; + } + shf.r.wrap.b32 %r10974, %r10973, %r10972, 9; + shf.r.wrap.b32 %r10975, %r10972, %r10973, 9; + mov.b64 %rd15342, {%r10975, %r10974}; + { + .reg .b32 %dummy; + mov.b64 {%r10976,%dummy}, %rd15311; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10977}, %rd15311; + } + shf.r.wrap.b32 %r10978, %r10977, %r10976, 25; + shf.r.wrap.b32 %r10979, %r10976, %r10977, 25; + mov.b64 %rd15343, {%r10979, %r10978}; + { + .reg .b32 %dummy; + mov.b64 {%r10980,%dummy}, %rd15312; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10981}, %rd15312; + } + shf.r.wrap.b32 %r10982, %r10981, %r10980, 23; + shf.r.wrap.b32 %r10983, %r10980, %r10981, 23; + mov.b64 %rd15344, {%r10983, %r10982}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10984}, %rd15318; + } + { + .reg .b32 %dummy; + mov.b64 {%r10985,%dummy}, %rd15318; + } + shf.l.wrap.b32 %r10986, %r10985, %r10984, 2; + shf.l.wrap.b32 %r10987, %r10984, %r10985, 2; + mov.b64 %rd15345, {%r10987, %r10986}; + not.b64 %rd15346, %rd15322; + and.b64 %rd15347, %rd15323, %rd15346; + xor.b64 %rd15348, %rd15347, %rd15297; + not.b64 %rd15349, %rd15323; + and.b64 %rd15350, %rd15324, %rd15349; + xor.b64 %rd15351, %rd15350, %rd15322; + not.b64 %rd15352, %rd15324; + and.b64 %rd15353, %rd15325, %rd15352; + xor.b64 %rd15354, %rd15353, %rd15323; + not.b64 %rd15355, %rd15325; + and.b64 %rd15356, %rd15297, %rd15355; + xor.b64 %rd15357, %rd15356, %rd15324; + not.b64 %rd15358, %rd15297; + and.b64 %rd15359, %rd15322, %rd15358; + xor.b64 %rd15360, %rd15359, %rd15325; + not.b64 %rd15361, %rd15327; + and.b64 %rd15362, %rd15328, %rd15361; + xor.b64 %rd15363, %rd15362, %rd15326; + not.b64 %rd15364, %rd15328; + and.b64 %rd15365, %rd15329, %rd15364; + xor.b64 %rd15366, %rd15365, %rd15327; + not.b64 %rd15367, %rd15329; + and.b64 %rd15368, %rd15330, %rd15367; + xor.b64 %rd15369, %rd15368, %rd15328; + not.b64 %rd15370, %rd15330; + and.b64 %rd15371, %rd15326, %rd15370; + xor.b64 %rd15372, %rd15371, %rd15329; + not.b64 %rd15373, %rd15326; + and.b64 %rd15374, %rd15327, %rd15373; + xor.b64 %rd15375, %rd15374, %rd15330; + not.b64 %rd15376, %rd15332; + and.b64 %rd15377, %rd15333, %rd15376; + xor.b64 %rd15378, %rd15377, %rd15331; + not.b64 %rd15379, %rd15333; + and.b64 %rd15380, %rd15334, %rd15379; + xor.b64 %rd15381, %rd15380, %rd15332; + not.b64 %rd15382, %rd15334; + and.b64 %rd15383, %rd15335, %rd15382; + xor.b64 %rd15384, %rd15383, %rd15333; + not.b64 %rd15385, %rd15335; + and.b64 %rd15386, %rd15331, %rd15385; + xor.b64 %rd15387, %rd15386, %rd15334; + not.b64 %rd15388, %rd15331; + and.b64 %rd15389, %rd15332, %rd15388; + xor.b64 %rd15390, %rd15389, %rd15335; + not.b64 %rd15391, %rd15337; + and.b64 %rd15392, %rd15338, %rd15391; + xor.b64 %rd15393, %rd15392, %rd15336; + not.b64 %rd15394, %rd15338; + and.b64 %rd15395, %rd15339, %rd15394; + xor.b64 %rd15396, %rd15395, %rd15337; + not.b64 %rd15397, %rd15339; + and.b64 %rd15398, %rd15340, %rd15397; + xor.b64 %rd15399, %rd15398, %rd15338; + not.b64 %rd15400, %rd15340; + and.b64 %rd15401, %rd15336, %rd15400; + xor.b64 %rd15402, %rd15401, %rd15339; + not.b64 %rd15403, %rd15336; + and.b64 %rd15404, %rd15337, %rd15403; + xor.b64 %rd15405, %rd15404, %rd15340; + not.b64 %rd15406, %rd15342; + and.b64 %rd15407, %rd15343, %rd15406; + xor.b64 %rd15408, %rd15407, %rd15341; + not.b64 %rd15409, %rd15343; + and.b64 %rd15410, %rd15344, %rd15409; + xor.b64 %rd15411, %rd15410, %rd15342; + not.b64 %rd15412, %rd15344; + and.b64 %rd15413, %rd15345, %rd15412; + xor.b64 %rd15414, %rd15413, %rd15343; + not.b64 %rd15415, %rd15345; + and.b64 %rd15416, %rd15341, %rd15415; + xor.b64 %rd15417, %rd15416, %rd15344; + not.b64 %rd15418, %rd15341; + and.b64 %rd15419, %rd15342, %rd15418; + xor.b64 %rd15420, %rd15419, %rd15345; + xor.b64 %rd15421, %rd15348, %rd37; + xor.b64 %rd15422, %rd15363, %rd15421; + xor.b64 %rd15423, %rd15422, %rd15378; + xor.b64 %rd15424, %rd15423, %rd15393; + xor.b64 %rd15425, %rd15424, %rd15408; + xor.b64 %rd15426, %rd15366, %rd15351; + xor.b64 %rd15427, %rd15426, %rd15381; + xor.b64 %rd15428, %rd15427, %rd15396; + xor.b64 %rd15429, %rd15428, %rd15411; + xor.b64 %rd15430, %rd15369, %rd15354; + xor.b64 %rd15431, %rd15430, %rd15384; + xor.b64 %rd15432, %rd15431, %rd15399; + xor.b64 %rd15433, %rd15432, %rd15414; + xor.b64 %rd15434, %rd15372, %rd15357; + xor.b64 %rd15435, %rd15434, %rd15387; + xor.b64 %rd15436, %rd15435, %rd15402; + xor.b64 %rd15437, %rd15436, %rd15417; + xor.b64 %rd15438, %rd15375, %rd15360; + xor.b64 %rd15439, %rd15438, %rd15390; + xor.b64 %rd15440, %rd15439, %rd15405; + xor.b64 %rd15441, %rd15440, %rd15420; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10988}, %rd15429; + } + { + .reg .b32 %dummy; + mov.b64 {%r10989,%dummy}, %rd15429; + } + shf.l.wrap.b32 %r10990, %r10989, %r10988, 1; + shf.l.wrap.b32 %r10991, %r10988, %r10989, 1; + mov.b64 %rd15442, {%r10991, %r10990}; + xor.b64 %rd15443, %rd15441, %rd15442; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10992}, %rd15433; + } + { + .reg .b32 %dummy; + mov.b64 {%r10993,%dummy}, %rd15433; + } + shf.l.wrap.b32 %r10994, %r10993, %r10992, 1; + shf.l.wrap.b32 %r10995, %r10992, %r10993, 1; + mov.b64 %rd15444, {%r10995, %r10994}; + xor.b64 %rd15445, %rd15444, %rd15425; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r10996}, %rd15437; + } + { + .reg .b32 %dummy; + mov.b64 {%r10997,%dummy}, %rd15437; + } + shf.l.wrap.b32 %r10998, %r10997, %r10996, 1; + shf.l.wrap.b32 %r10999, %r10996, %r10997, 1; + mov.b64 %rd15446, {%r10999, %r10998}; + xor.b64 %rd15447, %rd15446, %rd15429; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11000}, %rd15441; + } + { + .reg .b32 %dummy; + mov.b64 {%r11001,%dummy}, %rd15441; + } + shf.l.wrap.b32 %r11002, %r11001, %r11000, 1; + shf.l.wrap.b32 %r11003, %r11000, %r11001, 1; + mov.b64 %rd15448, {%r11003, %r11002}; + xor.b64 %rd15449, %rd15448, %rd15433; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11004}, %rd15425; + } + { + .reg .b32 %dummy; + mov.b64 {%r11005,%dummy}, %rd15425; + } + shf.l.wrap.b32 %r11006, %r11005, %r11004, 1; + shf.l.wrap.b32 %r11007, %r11004, %r11005, 1; + mov.b64 %rd15450, {%r11007, %r11006}; + xor.b64 %rd15451, %rd15437, %rd15450; + xor.b64 %rd15452, %rd15443, %rd15421; + xor.b64 %rd15453, %rd15445, %rd15351; + xor.b64 %rd15454, %rd15447, %rd15354; + xor.b64 %rd15455, %rd15449, %rd15357; + xor.b64 %rd15456, %rd15451, %rd15360; + xor.b64 %rd15457, %rd15443, %rd15363; + xor.b64 %rd15458, %rd15445, %rd15366; + xor.b64 %rd15459, %rd15447, %rd15369; + xor.b64 %rd15460, %rd15449, %rd15372; + xor.b64 %rd15461, %rd15451, %rd15375; + xor.b64 %rd15462, %rd15443, %rd15378; + xor.b64 %rd15463, %rd15445, %rd15381; + xor.b64 %rd15464, %rd15447, %rd15384; + xor.b64 %rd15465, %rd15449, %rd15387; + xor.b64 %rd15466, %rd15451, %rd15390; + xor.b64 %rd15467, %rd15443, %rd15393; + xor.b64 %rd15468, %rd15445, %rd15396; + xor.b64 %rd15469, %rd15447, %rd15399; + xor.b64 %rd15470, %rd15449, %rd15402; + xor.b64 %rd15471, %rd15451, %rd15405; + xor.b64 %rd15472, %rd15443, %rd15408; + xor.b64 %rd15473, %rd15445, %rd15411; + xor.b64 %rd15474, %rd15447, %rd15414; + xor.b64 %rd15475, %rd15449, %rd15417; + xor.b64 %rd15476, %rd15451, %rd15420; + { + .reg .b32 %dummy; + mov.b64 {%r11008,%dummy}, %rd15458; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11009}, %rd15458; + } + shf.r.wrap.b32 %r11010, %r11009, %r11008, 20; + shf.r.wrap.b32 %r11011, %r11008, %r11009, 20; + mov.b64 %rd15477, {%r11011, %r11010}; + { + .reg .b32 %dummy; + mov.b64 {%r11012,%dummy}, %rd15464; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11013}, %rd15464; + } + shf.r.wrap.b32 %r11014, %r11013, %r11012, 21; + shf.r.wrap.b32 %r11015, %r11012, %r11013, 21; + mov.b64 %rd15478, {%r11015, %r11014}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11016}, %rd15470; + } + { + .reg .b32 %dummy; + mov.b64 {%r11017,%dummy}, %rd15470; + } + shf.l.wrap.b32 %r11018, %r11017, %r11016, 21; + shf.l.wrap.b32 %r11019, %r11016, %r11017, 21; + mov.b64 %rd15479, {%r11019, %r11018}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11020}, %rd15476; + } + { + .reg .b32 %dummy; + mov.b64 {%r11021,%dummy}, %rd15476; + } + shf.l.wrap.b32 %r11022, %r11021, %r11020, 14; + shf.l.wrap.b32 %r11023, %r11020, %r11021, 14; + mov.b64 %rd15480, {%r11023, %r11022}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11024}, %rd15455; + } + { + .reg .b32 %dummy; + mov.b64 {%r11025,%dummy}, %rd15455; + } + shf.l.wrap.b32 %r11026, %r11025, %r11024, 28; + shf.l.wrap.b32 %r11027, %r11024, %r11025, 28; + mov.b64 %rd15481, {%r11027, %r11026}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11028}, %rd15461; + } + { + .reg .b32 %dummy; + mov.b64 {%r11029,%dummy}, %rd15461; + } + shf.l.wrap.b32 %r11030, %r11029, %r11028, 20; + shf.l.wrap.b32 %r11031, %r11028, %r11029, 20; + mov.b64 %rd15482, {%r11031, %r11030}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11032}, %rd15462; + } + { + .reg .b32 %dummy; + mov.b64 {%r11033,%dummy}, %rd15462; + } + shf.l.wrap.b32 %r11034, %r11033, %r11032, 3; + shf.l.wrap.b32 %r11035, %r11032, %r11033, 3; + mov.b64 %rd15483, {%r11035, %r11034}; + { + .reg .b32 %dummy; + mov.b64 {%r11036,%dummy}, %rd15468; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11037}, %rd15468; + } + shf.r.wrap.b32 %r11038, %r11037, %r11036, 19; + shf.r.wrap.b32 %r11039, %r11036, %r11037, 19; + mov.b64 %rd15484, {%r11039, %r11038}; + { + .reg .b32 %dummy; + mov.b64 {%r11040,%dummy}, %rd15474; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11041}, %rd15474; + } + shf.r.wrap.b32 %r11042, %r11041, %r11040, 3; + shf.r.wrap.b32 %r11043, %r11040, %r11041, 3; + mov.b64 %rd15485, {%r11043, %r11042}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11044}, %rd15453; + } + { + .reg .b32 %dummy; + mov.b64 {%r11045,%dummy}, %rd15453; + } + shf.l.wrap.b32 %r11046, %r11045, %r11044, 1; + shf.l.wrap.b32 %r11047, %r11044, %r11045, 1; + mov.b64 %rd15486, {%r11047, %r11046}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11048}, %rd15459; + } + { + .reg .b32 %dummy; + mov.b64 {%r11049,%dummy}, %rd15459; + } + shf.l.wrap.b32 %r11050, %r11049, %r11048, 6; + shf.l.wrap.b32 %r11051, %r11048, %r11049, 6; + mov.b64 %rd15487, {%r11051, %r11050}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11052}, %rd15465; + } + { + .reg .b32 %dummy; + mov.b64 {%r11053,%dummy}, %rd15465; + } + shf.l.wrap.b32 %r11054, %r11053, %r11052, 25; + shf.l.wrap.b32 %r11055, %r11052, %r11053, 25; + mov.b64 %rd15488, {%r11055, %r11054}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11056}, %rd15471; + } + { + .reg .b32 %dummy; + mov.b64 {%r11057,%dummy}, %rd15471; + } + shf.l.wrap.b32 %r11058, %r11057, %r11056, 8; + shf.l.wrap.b32 %r11059, %r11056, %r11057, 8; + mov.b64 %rd15489, {%r11059, %r11058}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11060}, %rd15472; + } + { + .reg .b32 %dummy; + mov.b64 {%r11061,%dummy}, %rd15472; + } + shf.l.wrap.b32 %r11062, %r11061, %r11060, 18; + shf.l.wrap.b32 %r11063, %r11060, %r11061, 18; + mov.b64 %rd15490, {%r11063, %r11062}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11064}, %rd15456; + } + { + .reg .b32 %dummy; + mov.b64 {%r11065,%dummy}, %rd15456; + } + shf.l.wrap.b32 %r11066, %r11065, %r11064, 27; + shf.l.wrap.b32 %r11067, %r11064, %r11065, 27; + mov.b64 %rd15491, {%r11067, %r11066}; + { + .reg .b32 %dummy; + mov.b64 {%r11068,%dummy}, %rd15457; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11069}, %rd15457; + } + shf.r.wrap.b32 %r11070, %r11069, %r11068, 28; + shf.r.wrap.b32 %r11071, %r11068, %r11069, 28; + mov.b64 %rd15492, {%r11071, %r11070}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11072}, %rd15463; + } + { + .reg .b32 %dummy; + mov.b64 {%r11073,%dummy}, %rd15463; + } + shf.l.wrap.b32 %r11074, %r11073, %r11072, 10; + shf.l.wrap.b32 %r11075, %r11072, %r11073, 10; + mov.b64 %rd15493, {%r11075, %r11074}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11076}, %rd15469; + } + { + .reg .b32 %dummy; + mov.b64 {%r11077,%dummy}, %rd15469; + } + shf.l.wrap.b32 %r11078, %r11077, %r11076, 15; + shf.l.wrap.b32 %r11079, %r11076, %r11077, 15; + mov.b64 %rd15494, {%r11079, %r11078}; + { + .reg .b32 %dummy; + mov.b64 {%r11080,%dummy}, %rd15475; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11081}, %rd15475; + } + shf.r.wrap.b32 %r11082, %r11081, %r11080, 8; + shf.r.wrap.b32 %r11083, %r11080, %r11081, 8; + mov.b64 %rd15495, {%r11083, %r11082}; + { + .reg .b32 %dummy; + mov.b64 {%r11084,%dummy}, %rd15454; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11085}, %rd15454; + } + shf.r.wrap.b32 %r11086, %r11085, %r11084, 2; + shf.r.wrap.b32 %r11087, %r11084, %r11085, 2; + mov.b64 %rd15496, {%r11087, %r11086}; + { + .reg .b32 %dummy; + mov.b64 {%r11088,%dummy}, %rd15460; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11089}, %rd15460; + } + shf.r.wrap.b32 %r11090, %r11089, %r11088, 9; + shf.r.wrap.b32 %r11091, %r11088, %r11089, 9; + mov.b64 %rd15497, {%r11091, %r11090}; + { + .reg .b32 %dummy; + mov.b64 {%r11092,%dummy}, %rd15466; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11093}, %rd15466; + } + shf.r.wrap.b32 %r11094, %r11093, %r11092, 25; + shf.r.wrap.b32 %r11095, %r11092, %r11093, 25; + mov.b64 %rd15498, {%r11095, %r11094}; + { + .reg .b32 %dummy; + mov.b64 {%r11096,%dummy}, %rd15467; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11097}, %rd15467; + } + shf.r.wrap.b32 %r11098, %r11097, %r11096, 23; + shf.r.wrap.b32 %r11099, %r11096, %r11097, 23; + mov.b64 %rd15499, {%r11099, %r11098}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11100}, %rd15473; + } + { + .reg .b32 %dummy; + mov.b64 {%r11101,%dummy}, %rd15473; + } + shf.l.wrap.b32 %r11102, %r11101, %r11100, 2; + shf.l.wrap.b32 %r11103, %r11100, %r11101, 2; + mov.b64 %rd15500, {%r11103, %r11102}; + not.b64 %rd15501, %rd15477; + and.b64 %rd15502, %rd15478, %rd15501; + xor.b64 %rd15503, %rd15502, %rd15452; + not.b64 %rd15504, %rd15478; + and.b64 %rd15505, %rd15479, %rd15504; + xor.b64 %rd15506, %rd15505, %rd15477; + not.b64 %rd15507, %rd15479; + and.b64 %rd15508, %rd15480, %rd15507; + xor.b64 %rd15509, %rd15508, %rd15478; + not.b64 %rd15510, %rd15480; + and.b64 %rd15511, %rd15452, %rd15510; + xor.b64 %rd15512, %rd15511, %rd15479; + not.b64 %rd15513, %rd15452; + and.b64 %rd15514, %rd15477, %rd15513; + xor.b64 %rd15515, %rd15514, %rd15480; + not.b64 %rd15516, %rd15482; + and.b64 %rd15517, %rd15483, %rd15516; + xor.b64 %rd15518, %rd15517, %rd15481; + not.b64 %rd15519, %rd15483; + and.b64 %rd15520, %rd15484, %rd15519; + xor.b64 %rd15521, %rd15520, %rd15482; + not.b64 %rd15522, %rd15484; + and.b64 %rd15523, %rd15485, %rd15522; + xor.b64 %rd15524, %rd15523, %rd15483; + not.b64 %rd15525, %rd15485; + and.b64 %rd15526, %rd15481, %rd15525; + xor.b64 %rd15527, %rd15526, %rd15484; + not.b64 %rd15528, %rd15481; + and.b64 %rd15529, %rd15482, %rd15528; + xor.b64 %rd15530, %rd15529, %rd15485; + not.b64 %rd15531, %rd15487; + and.b64 %rd15532, %rd15488, %rd15531; + xor.b64 %rd15533, %rd15532, %rd15486; + not.b64 %rd15534, %rd15488; + and.b64 %rd15535, %rd15489, %rd15534; + xor.b64 %rd15536, %rd15535, %rd15487; + not.b64 %rd15537, %rd15489; + and.b64 %rd15538, %rd15490, %rd15537; + xor.b64 %rd15539, %rd15538, %rd15488; + not.b64 %rd15540, %rd15490; + and.b64 %rd15541, %rd15486, %rd15540; + xor.b64 %rd15542, %rd15541, %rd15489; + not.b64 %rd15543, %rd15486; + and.b64 %rd15544, %rd15487, %rd15543; + xor.b64 %rd15545, %rd15544, %rd15490; + not.b64 %rd15546, %rd15492; + and.b64 %rd15547, %rd15493, %rd15546; + xor.b64 %rd15548, %rd15547, %rd15491; + not.b64 %rd15549, %rd15493; + and.b64 %rd15550, %rd15494, %rd15549; + xor.b64 %rd15551, %rd15550, %rd15492; + not.b64 %rd15552, %rd15494; + and.b64 %rd15553, %rd15495, %rd15552; + xor.b64 %rd15554, %rd15553, %rd15493; + not.b64 %rd15555, %rd15495; + and.b64 %rd15556, %rd15491, %rd15555; + xor.b64 %rd15557, %rd15556, %rd15494; + not.b64 %rd15558, %rd15491; + and.b64 %rd15559, %rd15492, %rd15558; + xor.b64 %rd15560, %rd15559, %rd15495; + not.b64 %rd15561, %rd15497; + and.b64 %rd15562, %rd15498, %rd15561; + xor.b64 %rd15563, %rd15562, %rd15496; + not.b64 %rd15564, %rd15498; + and.b64 %rd15565, %rd15499, %rd15564; + xor.b64 %rd15566, %rd15565, %rd15497; + not.b64 %rd15567, %rd15499; + and.b64 %rd15568, %rd15500, %rd15567; + xor.b64 %rd15569, %rd15568, %rd15498; + not.b64 %rd15570, %rd15500; + and.b64 %rd15571, %rd15496, %rd15570; + xor.b64 %rd15572, %rd15571, %rd15499; + not.b64 %rd15573, %rd15496; + and.b64 %rd15574, %rd15497, %rd15573; + xor.b64 %rd15575, %rd15574, %rd15500; + xor.b64 %rd15576, %rd15503, %rd38; + xor.b64 %rd15577, %rd15518, %rd15576; + xor.b64 %rd15578, %rd15577, %rd15533; + xor.b64 %rd15579, %rd15578, %rd15548; + xor.b64 %rd15580, %rd15579, %rd15563; + xor.b64 %rd15581, %rd15521, %rd15506; + xor.b64 %rd15582, %rd15581, %rd15536; + xor.b64 %rd15583, %rd15582, %rd15551; + xor.b64 %rd15584, %rd15583, %rd15566; + xor.b64 %rd15585, %rd15524, %rd15509; + xor.b64 %rd15586, %rd15585, %rd15539; + xor.b64 %rd15587, %rd15586, %rd15554; + xor.b64 %rd15588, %rd15587, %rd15569; + xor.b64 %rd15589, %rd15527, %rd15512; + xor.b64 %rd15590, %rd15589, %rd15542; + xor.b64 %rd15591, %rd15590, %rd15557; + xor.b64 %rd15592, %rd15591, %rd15572; + xor.b64 %rd15593, %rd15530, %rd15515; + xor.b64 %rd15594, %rd15593, %rd15545; + xor.b64 %rd15595, %rd15594, %rd15560; + xor.b64 %rd15596, %rd15595, %rd15575; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11104}, %rd15584; + } + { + .reg .b32 %dummy; + mov.b64 {%r11105,%dummy}, %rd15584; + } + shf.l.wrap.b32 %r11106, %r11105, %r11104, 1; + shf.l.wrap.b32 %r11107, %r11104, %r11105, 1; + mov.b64 %rd15597, {%r11107, %r11106}; + xor.b64 %rd15598, %rd15596, %rd15597; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11108}, %rd15588; + } + { + .reg .b32 %dummy; + mov.b64 {%r11109,%dummy}, %rd15588; + } + shf.l.wrap.b32 %r11110, %r11109, %r11108, 1; + shf.l.wrap.b32 %r11111, %r11108, %r11109, 1; + mov.b64 %rd15599, {%r11111, %r11110}; + xor.b64 %rd15600, %rd15599, %rd15580; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11112}, %rd15592; + } + { + .reg .b32 %dummy; + mov.b64 {%r11113,%dummy}, %rd15592; + } + shf.l.wrap.b32 %r11114, %r11113, %r11112, 1; + shf.l.wrap.b32 %r11115, %r11112, %r11113, 1; + mov.b64 %rd15601, {%r11115, %r11114}; + xor.b64 %rd15602, %rd15601, %rd15584; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11116}, %rd15596; + } + { + .reg .b32 %dummy; + mov.b64 {%r11117,%dummy}, %rd15596; + } + shf.l.wrap.b32 %r11118, %r11117, %r11116, 1; + shf.l.wrap.b32 %r11119, %r11116, %r11117, 1; + mov.b64 %rd15603, {%r11119, %r11118}; + xor.b64 %rd15604, %rd15603, %rd15588; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11120}, %rd15580; + } + { + .reg .b32 %dummy; + mov.b64 {%r11121,%dummy}, %rd15580; + } + shf.l.wrap.b32 %r11122, %r11121, %r11120, 1; + shf.l.wrap.b32 %r11123, %r11120, %r11121, 1; + mov.b64 %rd15605, {%r11123, %r11122}; + xor.b64 %rd15606, %rd15592, %rd15605; + xor.b64 %rd15607, %rd15598, %rd15576; + xor.b64 %rd15608, %rd15600, %rd15506; + xor.b64 %rd15609, %rd15602, %rd15509; + xor.b64 %rd15610, %rd15604, %rd15512; + xor.b64 %rd15611, %rd15606, %rd15515; + xor.b64 %rd15612, %rd15598, %rd15518; + xor.b64 %rd15613, %rd15600, %rd15521; + xor.b64 %rd15614, %rd15602, %rd15524; + xor.b64 %rd15615, %rd15604, %rd15527; + xor.b64 %rd15616, %rd15606, %rd15530; + xor.b64 %rd15617, %rd15598, %rd15533; + xor.b64 %rd15618, %rd15600, %rd15536; + xor.b64 %rd15619, %rd15602, %rd15539; + xor.b64 %rd15620, %rd15604, %rd15542; + xor.b64 %rd15621, %rd15606, %rd15545; + xor.b64 %rd15622, %rd15598, %rd15548; + xor.b64 %rd15623, %rd15600, %rd15551; + xor.b64 %rd15624, %rd15602, %rd15554; + xor.b64 %rd15625, %rd15604, %rd15557; + xor.b64 %rd15626, %rd15606, %rd15560; + xor.b64 %rd15627, %rd15598, %rd15563; + xor.b64 %rd15628, %rd15600, %rd15566; + xor.b64 %rd15629, %rd15602, %rd15569; + xor.b64 %rd15630, %rd15604, %rd15572; + xor.b64 %rd15631, %rd15606, %rd15575; + { + .reg .b32 %dummy; + mov.b64 {%r11124,%dummy}, %rd15613; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11125}, %rd15613; + } + shf.r.wrap.b32 %r11126, %r11125, %r11124, 20; + shf.r.wrap.b32 %r11127, %r11124, %r11125, 20; + mov.b64 %rd15632, {%r11127, %r11126}; + { + .reg .b32 %dummy; + mov.b64 {%r11128,%dummy}, %rd15619; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11129}, %rd15619; + } + shf.r.wrap.b32 %r11130, %r11129, %r11128, 21; + shf.r.wrap.b32 %r11131, %r11128, %r11129, 21; + mov.b64 %rd15633, {%r11131, %r11130}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11132}, %rd15625; + } + { + .reg .b32 %dummy; + mov.b64 {%r11133,%dummy}, %rd15625; + } + shf.l.wrap.b32 %r11134, %r11133, %r11132, 21; + shf.l.wrap.b32 %r11135, %r11132, %r11133, 21; + mov.b64 %rd15634, {%r11135, %r11134}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11136}, %rd15631; + } + { + .reg .b32 %dummy; + mov.b64 {%r11137,%dummy}, %rd15631; + } + shf.l.wrap.b32 %r11138, %r11137, %r11136, 14; + shf.l.wrap.b32 %r11139, %r11136, %r11137, 14; + mov.b64 %rd15635, {%r11139, %r11138}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11140}, %rd15610; + } + { + .reg .b32 %dummy; + mov.b64 {%r11141,%dummy}, %rd15610; + } + shf.l.wrap.b32 %r11142, %r11141, %r11140, 28; + shf.l.wrap.b32 %r11143, %r11140, %r11141, 28; + mov.b64 %rd15636, {%r11143, %r11142}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11144}, %rd15616; + } + { + .reg .b32 %dummy; + mov.b64 {%r11145,%dummy}, %rd15616; + } + shf.l.wrap.b32 %r11146, %r11145, %r11144, 20; + shf.l.wrap.b32 %r11147, %r11144, %r11145, 20; + mov.b64 %rd15637, {%r11147, %r11146}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11148}, %rd15617; + } + { + .reg .b32 %dummy; + mov.b64 {%r11149,%dummy}, %rd15617; + } + shf.l.wrap.b32 %r11150, %r11149, %r11148, 3; + shf.l.wrap.b32 %r11151, %r11148, %r11149, 3; + mov.b64 %rd15638, {%r11151, %r11150}; + { + .reg .b32 %dummy; + mov.b64 {%r11152,%dummy}, %rd15623; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11153}, %rd15623; + } + shf.r.wrap.b32 %r11154, %r11153, %r11152, 19; + shf.r.wrap.b32 %r11155, %r11152, %r11153, 19; + mov.b64 %rd15639, {%r11155, %r11154}; + { + .reg .b32 %dummy; + mov.b64 {%r11156,%dummy}, %rd15629; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11157}, %rd15629; + } + shf.r.wrap.b32 %r11158, %r11157, %r11156, 3; + shf.r.wrap.b32 %r11159, %r11156, %r11157, 3; + mov.b64 %rd15640, {%r11159, %r11158}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11160}, %rd15608; + } + { + .reg .b32 %dummy; + mov.b64 {%r11161,%dummy}, %rd15608; + } + shf.l.wrap.b32 %r11162, %r11161, %r11160, 1; + shf.l.wrap.b32 %r11163, %r11160, %r11161, 1; + mov.b64 %rd15641, {%r11163, %r11162}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11164}, %rd15614; + } + { + .reg .b32 %dummy; + mov.b64 {%r11165,%dummy}, %rd15614; + } + shf.l.wrap.b32 %r11166, %r11165, %r11164, 6; + shf.l.wrap.b32 %r11167, %r11164, %r11165, 6; + mov.b64 %rd15642, {%r11167, %r11166}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11168}, %rd15620; + } + { + .reg .b32 %dummy; + mov.b64 {%r11169,%dummy}, %rd15620; + } + shf.l.wrap.b32 %r11170, %r11169, %r11168, 25; + shf.l.wrap.b32 %r11171, %r11168, %r11169, 25; + mov.b64 %rd15643, {%r11171, %r11170}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11172}, %rd15626; + } + { + .reg .b32 %dummy; + mov.b64 {%r11173,%dummy}, %rd15626; + } + shf.l.wrap.b32 %r11174, %r11173, %r11172, 8; + shf.l.wrap.b32 %r11175, %r11172, %r11173, 8; + mov.b64 %rd15644, {%r11175, %r11174}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11176}, %rd15627; + } + { + .reg .b32 %dummy; + mov.b64 {%r11177,%dummy}, %rd15627; + } + shf.l.wrap.b32 %r11178, %r11177, %r11176, 18; + shf.l.wrap.b32 %r11179, %r11176, %r11177, 18; + mov.b64 %rd15645, {%r11179, %r11178}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11180}, %rd15611; + } + { + .reg .b32 %dummy; + mov.b64 {%r11181,%dummy}, %rd15611; + } + shf.l.wrap.b32 %r11182, %r11181, %r11180, 27; + shf.l.wrap.b32 %r11183, %r11180, %r11181, 27; + mov.b64 %rd15646, {%r11183, %r11182}; + { + .reg .b32 %dummy; + mov.b64 {%r11184,%dummy}, %rd15612; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11185}, %rd15612; + } + shf.r.wrap.b32 %r11186, %r11185, %r11184, 28; + shf.r.wrap.b32 %r11187, %r11184, %r11185, 28; + mov.b64 %rd15647, {%r11187, %r11186}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11188}, %rd15618; + } + { + .reg .b32 %dummy; + mov.b64 {%r11189,%dummy}, %rd15618; + } + shf.l.wrap.b32 %r11190, %r11189, %r11188, 10; + shf.l.wrap.b32 %r11191, %r11188, %r11189, 10; + mov.b64 %rd15648, {%r11191, %r11190}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11192}, %rd15624; + } + { + .reg .b32 %dummy; + mov.b64 {%r11193,%dummy}, %rd15624; + } + shf.l.wrap.b32 %r11194, %r11193, %r11192, 15; + shf.l.wrap.b32 %r11195, %r11192, %r11193, 15; + mov.b64 %rd15649, {%r11195, %r11194}; + { + .reg .b32 %dummy; + mov.b64 {%r11196,%dummy}, %rd15630; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11197}, %rd15630; + } + shf.r.wrap.b32 %r11198, %r11197, %r11196, 8; + shf.r.wrap.b32 %r11199, %r11196, %r11197, 8; + mov.b64 %rd15650, {%r11199, %r11198}; + { + .reg .b32 %dummy; + mov.b64 {%r11200,%dummy}, %rd15609; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11201}, %rd15609; + } + shf.r.wrap.b32 %r11202, %r11201, %r11200, 2; + shf.r.wrap.b32 %r11203, %r11200, %r11201, 2; + mov.b64 %rd15651, {%r11203, %r11202}; + { + .reg .b32 %dummy; + mov.b64 {%r11204,%dummy}, %rd15615; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11205}, %rd15615; + } + shf.r.wrap.b32 %r11206, %r11205, %r11204, 9; + shf.r.wrap.b32 %r11207, %r11204, %r11205, 9; + mov.b64 %rd15652, {%r11207, %r11206}; + { + .reg .b32 %dummy; + mov.b64 {%r11208,%dummy}, %rd15621; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11209}, %rd15621; + } + shf.r.wrap.b32 %r11210, %r11209, %r11208, 25; + shf.r.wrap.b32 %r11211, %r11208, %r11209, 25; + mov.b64 %rd15653, {%r11211, %r11210}; + { + .reg .b32 %dummy; + mov.b64 {%r11212,%dummy}, %rd15622; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11213}, %rd15622; + } + shf.r.wrap.b32 %r11214, %r11213, %r11212, 23; + shf.r.wrap.b32 %r11215, %r11212, %r11213, 23; + mov.b64 %rd15654, {%r11215, %r11214}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11216}, %rd15628; + } + { + .reg .b32 %dummy; + mov.b64 {%r11217,%dummy}, %rd15628; + } + shf.l.wrap.b32 %r11218, %r11217, %r11216, 2; + shf.l.wrap.b32 %r11219, %r11216, %r11217, 2; + mov.b64 %rd15655, {%r11219, %r11218}; + not.b64 %rd15656, %rd15632; + and.b64 %rd15657, %rd15633, %rd15656; + xor.b64 %rd15658, %rd15657, %rd15607; + not.b64 %rd15659, %rd15633; + and.b64 %rd15660, %rd15634, %rd15659; + xor.b64 %rd15661, %rd15660, %rd15632; + st.local.u64 [%rd2+8], %rd15661; + not.b64 %rd15662, %rd15634; + and.b64 %rd15663, %rd15635, %rd15662; + xor.b64 %rd15664, %rd15663, %rd15633; + st.local.u64 [%rd2+16], %rd15664; + not.b64 %rd15665, %rd15635; + and.b64 %rd15666, %rd15607, %rd15665; + xor.b64 %rd15667, %rd15666, %rd15634; + st.local.u64 [%rd2+24], %rd15667; + not.b64 %rd15668, %rd15607; + and.b64 %rd15669, %rd15632, %rd15668; + xor.b64 %rd15670, %rd15669, %rd15635; + st.local.u64 [%rd2+32], %rd15670; + not.b64 %rd15671, %rd15637; + and.b64 %rd15672, %rd15638, %rd15671; + xor.b64 %rd15673, %rd15672, %rd15636; + st.local.u64 [%rd2+40], %rd15673; + not.b64 %rd15674, %rd15638; + and.b64 %rd15675, %rd15639, %rd15674; + xor.b64 %rd15676, %rd15675, %rd15637; + st.local.u64 [%rd2+48], %rd15676; + not.b64 %rd15677, %rd15639; + and.b64 %rd15678, %rd15640, %rd15677; + xor.b64 %rd15679, %rd15678, %rd15638; + st.local.u64 [%rd2+56], %rd15679; + not.b64 %rd15680, %rd15640; + and.b64 %rd15681, %rd15636, %rd15680; + xor.b64 %rd15682, %rd15681, %rd15639; + st.local.u64 [%rd2+64], %rd15682; + not.b64 %rd15683, %rd15636; + and.b64 %rd15684, %rd15637, %rd15683; + xor.b64 %rd15685, %rd15684, %rd15640; + st.local.u64 [%rd2+72], %rd15685; + not.b64 %rd15686, %rd15642; + and.b64 %rd15687, %rd15643, %rd15686; + xor.b64 %rd15688, %rd15687, %rd15641; + st.local.u64 [%rd2+80], %rd15688; + not.b64 %rd15689, %rd15643; + and.b64 %rd15690, %rd15644, %rd15689; + xor.b64 %rd15691, %rd15690, %rd15642; + st.local.u64 [%rd2+88], %rd15691; + not.b64 %rd15692, %rd15644; + and.b64 %rd15693, %rd15645, %rd15692; + xor.b64 %rd15694, %rd15693, %rd15643; + st.local.u64 [%rd2+96], %rd15694; + not.b64 %rd15695, %rd15645; + and.b64 %rd15696, %rd15641, %rd15695; + xor.b64 %rd15697, %rd15696, %rd15644; + st.local.u64 [%rd2+104], %rd15697; + not.b64 %rd15698, %rd15641; + and.b64 %rd15699, %rd15642, %rd15698; + xor.b64 %rd15700, %rd15699, %rd15645; + st.local.u64 [%rd2+112], %rd15700; + not.b64 %rd15701, %rd15647; + and.b64 %rd15702, %rd15648, %rd15701; + xor.b64 %rd15703, %rd15702, %rd15646; + st.local.u64 [%rd2+120], %rd15703; + not.b64 %rd15704, %rd15648; + and.b64 %rd15705, %rd15649, %rd15704; + xor.b64 %rd15706, %rd15705, %rd15647; + st.local.u64 [%rd2+128], %rd15706; + not.b64 %rd15707, %rd15649; + and.b64 %rd15708, %rd15650, %rd15707; + xor.b64 %rd15709, %rd15708, %rd15648; + st.local.u64 [%rd2+136], %rd15709; + not.b64 %rd15710, %rd15650; + and.b64 %rd15711, %rd15646, %rd15710; + xor.b64 %rd15712, %rd15711, %rd15649; + st.local.u64 [%rd2+144], %rd15712; + not.b64 %rd15713, %rd15646; + and.b64 %rd15714, %rd15647, %rd15713; + xor.b64 %rd15715, %rd15714, %rd15650; + st.local.u64 [%rd2+152], %rd15715; + not.b64 %rd15716, %rd15652; + and.b64 %rd15717, %rd15653, %rd15716; + xor.b64 %rd15718, %rd15717, %rd15651; + st.local.u64 [%rd2+160], %rd15718; + not.b64 %rd15719, %rd15653; + and.b64 %rd15720, %rd15654, %rd15719; + xor.b64 %rd15721, %rd15720, %rd15652; + st.local.u64 [%rd2+168], %rd15721; + not.b64 %rd15722, %rd15654; + and.b64 %rd15723, %rd15655, %rd15722; + xor.b64 %rd15724, %rd15723, %rd15653; + st.local.u64 [%rd2+176], %rd15724; + not.b64 %rd15725, %rd15655; + and.b64 %rd15726, %rd15651, %rd15725; + xor.b64 %rd15727, %rd15726, %rd15654; + st.local.u64 [%rd2+184], %rd15727; + not.b64 %rd15728, %rd15651; + and.b64 %rd15729, %rd15652, %rd15728; + xor.b64 %rd15730, %rd15729, %rd15655; + st.local.u64 [%rd2+192], %rd15730; + xor.b64 %rd15731, %rd15658, %rd39; + st.local.u64 [%rd2], %rd15731; + shr.u64 %rd130, %rd7997, 6; + setp.eq.s64 %p39, %rd130, 0; + @%p39 bra $L__BB1_52; + + add.s64 %rd15732, %rd130, -1; + and.b64 %rd19862, %rd130, 3; + setp.lt.u64 %p40, %rd15732, 3; + mov.u32 %r14037, 0; + @%p40 bra $L__BB1_50; + + sub.s64 %rd19861, %rd130, %rd19862; + +$L__BB1_49: + mul.wide.s32 %rd15733, %r14037, 8; + add.s64 %rd15734, %rd2, %rd15733; + ld.local.u8 %rd15735, [%rd15734]; + ld.local.u8 %rd15736, [%rd15734+1]; + bfi.b64 %rd15737, %rd15736, %rd15735, 8, 8; + ld.local.u8 %rd15738, [%rd15734+2]; + ld.local.u8 %rd15739, [%rd15734+3]; + bfi.b64 %rd15740, %rd15739, %rd15738, 8, 8; + bfi.b64 %rd15741, %rd15740, %rd15737, 16, 16; + ld.local.u8 %rd15742, [%rd15734+4]; + ld.local.u8 %rd15743, [%rd15734+5]; + bfi.b64 %rd15744, %rd15743, %rd15742, 8, 8; + ld.local.u8 %rd15745, [%rd15734+6]; + ld.local.u8 %rd15746, [%rd15734+7]; + bfi.b64 %rd15747, %rd15746, %rd15745, 8, 8; + bfi.b64 %rd15748, %rd15747, %rd15744, 16, 16; + bfi.b64 %rd15749, %rd15748, %rd15741, 32, 32; + st.local.u8 [%rd15734+200], %rd15737; + shr.u64 %rd15750, %rd15737, 8; + st.local.u8 [%rd15734+201], %rd15750; + shr.u64 %rd15751, %rd15741, 24; + st.local.u8 [%rd15734+203], %rd15751; + shr.u64 %rd15752, %rd15741, 16; + st.local.u8 [%rd15734+202], %rd15752; + shr.u64 %rd15753, %rd15749, 56; + st.local.u8 [%rd15734+207], %rd15753; + shr.u64 %rd15754, %rd15749, 48; + st.local.u8 [%rd15734+206], %rd15754; + shr.u64 %rd15755, %rd15749, 40; + st.local.u8 [%rd15734+205], %rd15755; + shr.u64 %rd15756, %rd15749, 32; + st.local.u8 [%rd15734+204], %rd15756; + ld.local.u8 %rd15757, [%rd15734+8]; + ld.local.u8 %rd15758, [%rd15734+9]; + bfi.b64 %rd15759, %rd15758, %rd15757, 8, 8; + ld.local.u8 %rd15760, [%rd15734+10]; + ld.local.u8 %rd15761, [%rd15734+11]; + bfi.b64 %rd15762, %rd15761, %rd15760, 8, 8; + bfi.b64 %rd15763, %rd15762, %rd15759, 16, 16; + ld.local.u8 %rd15764, [%rd15734+12]; + ld.local.u8 %rd15765, [%rd15734+13]; + bfi.b64 %rd15766, %rd15765, %rd15764, 8, 8; + ld.local.u8 %rd15767, [%rd15734+14]; + ld.local.u8 %rd15768, [%rd15734+15]; + bfi.b64 %rd15769, %rd15768, %rd15767, 8, 8; + bfi.b64 %rd15770, %rd15769, %rd15766, 16, 16; + bfi.b64 %rd15771, %rd15770, %rd15763, 32, 32; + st.local.u8 [%rd15734+208], %rd15759; + shr.u64 %rd15772, %rd15759, 8; + st.local.u8 [%rd15734+209], %rd15772; + shr.u64 %rd15773, %rd15763, 24; + st.local.u8 [%rd15734+211], %rd15773; + shr.u64 %rd15774, %rd15763, 16; + st.local.u8 [%rd15734+210], %rd15774; + shr.u64 %rd15775, %rd15771, 56; + st.local.u8 [%rd15734+215], %rd15775; + shr.u64 %rd15776, %rd15771, 48; + st.local.u8 [%rd15734+214], %rd15776; + shr.u64 %rd15777, %rd15771, 40; + st.local.u8 [%rd15734+213], %rd15777; + shr.u64 %rd15778, %rd15771, 32; + st.local.u8 [%rd15734+212], %rd15778; + ld.local.u8 %rd15779, [%rd15734+16]; + ld.local.u8 %rd15780, [%rd15734+17]; + bfi.b64 %rd15781, %rd15780, %rd15779, 8, 8; + ld.local.u8 %rd15782, [%rd15734+18]; + ld.local.u8 %rd15783, [%rd15734+19]; + bfi.b64 %rd15784, %rd15783, %rd15782, 8, 8; + bfi.b64 %rd15785, %rd15784, %rd15781, 16, 16; + ld.local.u8 %rd15786, [%rd15734+20]; + ld.local.u8 %rd15787, [%rd15734+21]; + bfi.b64 %rd15788, %rd15787, %rd15786, 8, 8; + ld.local.u8 %rd15789, [%rd15734+22]; + ld.local.u8 %rd15790, [%rd15734+23]; + bfi.b64 %rd15791, %rd15790, %rd15789, 8, 8; + bfi.b64 %rd15792, %rd15791, %rd15788, 16, 16; + bfi.b64 %rd15793, %rd15792, %rd15785, 32, 32; + st.local.u8 [%rd15734+216], %rd15781; + shr.u64 %rd15794, %rd15781, 8; + st.local.u8 [%rd15734+217], %rd15794; + shr.u64 %rd15795, %rd15785, 24; + st.local.u8 [%rd15734+219], %rd15795; + shr.u64 %rd15796, %rd15785, 16; + st.local.u8 [%rd15734+218], %rd15796; + shr.u64 %rd15797, %rd15793, 56; + st.local.u8 [%rd15734+223], %rd15797; + shr.u64 %rd15798, %rd15793, 48; + st.local.u8 [%rd15734+222], %rd15798; + shr.u64 %rd15799, %rd15793, 40; + st.local.u8 [%rd15734+221], %rd15799; + shr.u64 %rd15800, %rd15793, 32; + st.local.u8 [%rd15734+220], %rd15800; + ld.local.u8 %rd15801, [%rd15734+24]; + ld.local.u8 %rd15802, [%rd15734+25]; + bfi.b64 %rd15803, %rd15802, %rd15801, 8, 8; + ld.local.u8 %rd15804, [%rd15734+26]; + ld.local.u8 %rd15805, [%rd15734+27]; + bfi.b64 %rd15806, %rd15805, %rd15804, 8, 8; + bfi.b64 %rd15807, %rd15806, %rd15803, 16, 16; + ld.local.u8 %rd15808, [%rd15734+28]; + ld.local.u8 %rd15809, [%rd15734+29]; + bfi.b64 %rd15810, %rd15809, %rd15808, 8, 8; + ld.local.u8 %rd15811, [%rd15734+30]; + ld.local.u8 %rd15812, [%rd15734+31]; + bfi.b64 %rd15813, %rd15812, %rd15811, 8, 8; + bfi.b64 %rd15814, %rd15813, %rd15810, 16, 16; + bfi.b64 %rd15815, %rd15814, %rd15807, 32, 32; + st.local.u8 [%rd15734+224], %rd15803; + shr.u64 %rd15816, %rd15803, 8; + st.local.u8 [%rd15734+225], %rd15816; + shr.u64 %rd15817, %rd15807, 24; + st.local.u8 [%rd15734+227], %rd15817; + shr.u64 %rd15818, %rd15807, 16; + st.local.u8 [%rd15734+226], %rd15818; + shr.u64 %rd15819, %rd15815, 56; + st.local.u8 [%rd15734+231], %rd15819; + shr.u64 %rd15820, %rd15815, 48; + st.local.u8 [%rd15734+230], %rd15820; + shr.u64 %rd15821, %rd15815, 40; + st.local.u8 [%rd15734+229], %rd15821; + shr.u64 %rd15822, %rd15815, 32; + st.local.u8 [%rd15734+228], %rd15822; + add.s32 %r14037, %r14037, 4; + add.s64 %rd19861, %rd19861, -4; + setp.ne.s64 %p41, %rd19861, 0; + @%p41 bra $L__BB1_49; + +$L__BB1_50: + setp.eq.s64 %p42, %rd19862, 0; + @%p42 bra $L__BB1_52; + +$L__BB1_51: + .pragma "nounroll"; + mul.wide.s32 %rd15823, %r14037, 8; + add.s64 %rd15824, %rd2, %rd15823; + ld.local.u8 %rd15825, [%rd15824]; + ld.local.u8 %rd15826, [%rd15824+1]; + bfi.b64 %rd15827, %rd15826, %rd15825, 8, 8; + ld.local.u8 %rd15828, [%rd15824+2]; + ld.local.u8 %rd15829, [%rd15824+3]; + bfi.b64 %rd15830, %rd15829, %rd15828, 8, 8; + bfi.b64 %rd15831, %rd15830, %rd15827, 16, 16; + ld.local.u8 %rd15832, [%rd15824+4]; + ld.local.u8 %rd15833, [%rd15824+5]; + bfi.b64 %rd15834, %rd15833, %rd15832, 8, 8; + ld.local.u8 %rd15835, [%rd15824+6]; + ld.local.u8 %rd15836, [%rd15824+7]; + bfi.b64 %rd15837, %rd15836, %rd15835, 8, 8; + bfi.b64 %rd15838, %rd15837, %rd15834, 16, 16; + bfi.b64 %rd15839, %rd15838, %rd15831, 32, 32; + shl.b32 %r11222, %r14037, 3; + cvt.s64.s32 %rd15840, %r11222; + add.s64 %rd15841, %rd2, %rd15840; + st.local.u8 [%rd15841+200], %rd15827; + shr.u64 %rd15842, %rd15827, 8; + st.local.u8 [%rd15841+201], %rd15842; + shr.u64 %rd15843, %rd15831, 24; + st.local.u8 [%rd15841+203], %rd15843; + shr.u64 %rd15844, %rd15831, 16; + st.local.u8 [%rd15841+202], %rd15844; + shr.u64 %rd15845, %rd15839, 56; + st.local.u8 [%rd15841+207], %rd15845; + shr.u64 %rd15846, %rd15839, 48; + st.local.u8 [%rd15841+206], %rd15846; + shr.u64 %rd15847, %rd15839, 40; + st.local.u8 [%rd15841+205], %rd15847; + shr.u64 %rd15848, %rd15839, 32; + st.local.u8 [%rd15841+204], %rd15848; + add.s32 %r14037, %r14037, 1; + add.s64 %rd19862, %rd19862, -1; + setp.ne.s64 %p43, %rd19862, 0; + @%p43 bra $L__BB1_51; + +$L__BB1_52: + st.local.u64 [%rd2+392], %rd7997; + ld.const.u64 %rd137, [digestbitlen]; + setp.eq.s64 %p44, %rd137, 0; + @%p44 bra $L__BB1_73; + + mov.u64 %rd19864, 0; + mov.u64 %rd19866, %rd7997; + +$L__BB1_54: + setp.ne.s64 %p45, %rd19866, 0; + @%p45 bra $L__BB1_64; + + ld.const.u64 %rd19819, [CUDA_KECCAK_CONSTS+88]; + ld.const.u64 %rd19818, [CUDA_KECCAK_CONSTS+80]; + ld.const.u64 %rd19817, [CUDA_KECCAK_CONSTS+72]; + ld.const.u64 %rd19816, [CUDA_KECCAK_CONSTS+64]; + ld.const.u64 %rd19815, [CUDA_KECCAK_CONSTS+56]; + ld.const.u64 %rd19814, [CUDA_KECCAK_CONSTS+48]; + ld.const.u64 %rd19813, [CUDA_KECCAK_CONSTS+40]; + ld.const.u64 %rd19812, [CUDA_KECCAK_CONSTS+32]; + ld.const.u64 %rd19811, [CUDA_KECCAK_CONSTS+24]; + ld.const.u64 %rd19810, [CUDA_KECCAK_CONSTS+16]; + ld.const.u64 %rd19809, [CUDA_KECCAK_CONSTS+8]; + ld.const.u64 %rd19808, [CUDA_KECCAK_CONSTS]; + ld.local.u64 %rd15851, [%rd2+40]; + ld.local.u64 %rd15852, [%rd2]; + xor.b64 %rd15853, %rd15851, %rd15852; + ld.local.u64 %rd15854, [%rd2+80]; + xor.b64 %rd15855, %rd15853, %rd15854; + ld.local.u64 %rd15856, [%rd2+120]; + xor.b64 %rd15857, %rd15855, %rd15856; + ld.local.u64 %rd15858, [%rd2+160]; + xor.b64 %rd15859, %rd15857, %rd15858; + ld.local.u64 %rd15860, [%rd2+48]; + ld.local.u64 %rd15861, [%rd2+8]; + xor.b64 %rd15862, %rd15860, %rd15861; + ld.local.u64 %rd15863, [%rd2+88]; + xor.b64 %rd15864, %rd15862, %rd15863; + ld.local.u64 %rd15865, [%rd2+128]; + xor.b64 %rd15866, %rd15864, %rd15865; + ld.local.u64 %rd15867, [%rd2+168]; + xor.b64 %rd15868, %rd15866, %rd15867; + ld.local.u64 %rd15869, [%rd2+56]; + ld.local.u64 %rd15870, [%rd2+16]; + xor.b64 %rd15871, %rd15869, %rd15870; + ld.local.u64 %rd15872, [%rd2+96]; + xor.b64 %rd15873, %rd15871, %rd15872; + ld.local.u64 %rd15874, [%rd2+136]; + xor.b64 %rd15875, %rd15873, %rd15874; + ld.local.u64 %rd15876, [%rd2+176]; + xor.b64 %rd15877, %rd15875, %rd15876; + ld.local.u64 %rd15878, [%rd2+64]; + ld.local.u64 %rd15879, [%rd2+24]; + xor.b64 %rd15880, %rd15878, %rd15879; + ld.local.u64 %rd15881, [%rd2+104]; + xor.b64 %rd15882, %rd15880, %rd15881; + ld.local.u64 %rd15883, [%rd2+144]; + xor.b64 %rd15884, %rd15882, %rd15883; + ld.local.u64 %rd15885, [%rd2+184]; + xor.b64 %rd15886, %rd15884, %rd15885; + ld.local.u64 %rd15887, [%rd2+72]; + ld.local.u64 %rd15888, [%rd2+32]; + xor.b64 %rd15889, %rd15887, %rd15888; + ld.local.u64 %rd15890, [%rd2+112]; + xor.b64 %rd15891, %rd15889, %rd15890; + ld.local.u64 %rd15892, [%rd2+152]; + xor.b64 %rd15893, %rd15891, %rd15892; + ld.local.u64 %rd15894, [%rd2+192]; + xor.b64 %rd15895, %rd15893, %rd15894; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11223}, %rd15868; + } + { + .reg .b32 %dummy; + mov.b64 {%r11224,%dummy}, %rd15868; + } + shf.l.wrap.b32 %r11225, %r11224, %r11223, 1; + shf.l.wrap.b32 %r11226, %r11223, %r11224, 1; + mov.b64 %rd15896, {%r11226, %r11225}; + xor.b64 %rd15897, %rd15895, %rd15896; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11227}, %rd15877; + } + { + .reg .b32 %dummy; + mov.b64 {%r11228,%dummy}, %rd15877; + } + shf.l.wrap.b32 %r11229, %r11228, %r11227, 1; + shf.l.wrap.b32 %r11230, %r11227, %r11228, 1; + mov.b64 %rd15898, {%r11230, %r11229}; + xor.b64 %rd15899, %rd15898, %rd15859; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11231}, %rd15886; + } + { + .reg .b32 %dummy; + mov.b64 {%r11232,%dummy}, %rd15886; + } + shf.l.wrap.b32 %r11233, %r11232, %r11231, 1; + shf.l.wrap.b32 %r11234, %r11231, %r11232, 1; + mov.b64 %rd15900, {%r11234, %r11233}; + xor.b64 %rd15901, %rd15900, %rd15868; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11235}, %rd15895; + } + { + .reg .b32 %dummy; + mov.b64 {%r11236,%dummy}, %rd15895; + } + shf.l.wrap.b32 %r11237, %r11236, %r11235, 1; + shf.l.wrap.b32 %r11238, %r11235, %r11236, 1; + mov.b64 %rd15902, {%r11238, %r11237}; + xor.b64 %rd15903, %rd15902, %rd15877; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11239}, %rd15859; + } + { + .reg .b32 %dummy; + mov.b64 {%r11240,%dummy}, %rd15859; + } + shf.l.wrap.b32 %r11241, %r11240, %r11239, 1; + shf.l.wrap.b32 %r11242, %r11239, %r11240, 1; + mov.b64 %rd15904, {%r11242, %r11241}; + xor.b64 %rd15905, %rd15886, %rd15904; + xor.b64 %rd15906, %rd15897, %rd15852; + xor.b64 %rd15907, %rd15899, %rd15861; + xor.b64 %rd15908, %rd15901, %rd15870; + xor.b64 %rd15909, %rd15903, %rd15879; + xor.b64 %rd15910, %rd15905, %rd15888; + xor.b64 %rd15911, %rd15897, %rd15851; + xor.b64 %rd15912, %rd15899, %rd15860; + xor.b64 %rd15913, %rd15901, %rd15869; + xor.b64 %rd15914, %rd15903, %rd15878; + xor.b64 %rd15915, %rd15905, %rd15887; + xor.b64 %rd15916, %rd15897, %rd15854; + xor.b64 %rd15917, %rd15899, %rd15863; + xor.b64 %rd15918, %rd15901, %rd15872; + xor.b64 %rd15919, %rd15903, %rd15881; + xor.b64 %rd15920, %rd15905, %rd15890; + xor.b64 %rd15921, %rd15897, %rd15856; + xor.b64 %rd15922, %rd15899, %rd15865; + xor.b64 %rd15923, %rd15901, %rd15874; + xor.b64 %rd15924, %rd15903, %rd15883; + xor.b64 %rd15925, %rd15905, %rd15892; + xor.b64 %rd15926, %rd15897, %rd15858; + xor.b64 %rd15927, %rd15899, %rd15867; + xor.b64 %rd15928, %rd15901, %rd15876; + xor.b64 %rd15929, %rd15903, %rd15885; + xor.b64 %rd15930, %rd15905, %rd15894; + { + .reg .b32 %dummy; + mov.b64 {%r11243,%dummy}, %rd15912; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11244}, %rd15912; + } + shf.r.wrap.b32 %r11245, %r11244, %r11243, 20; + shf.r.wrap.b32 %r11246, %r11243, %r11244, 20; + mov.b64 %rd15931, {%r11246, %r11245}; + { + .reg .b32 %dummy; + mov.b64 {%r11247,%dummy}, %rd15918; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11248}, %rd15918; + } + shf.r.wrap.b32 %r11249, %r11248, %r11247, 21; + shf.r.wrap.b32 %r11250, %r11247, %r11248, 21; + mov.b64 %rd15932, {%r11250, %r11249}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11251}, %rd15924; + } + { + .reg .b32 %dummy; + mov.b64 {%r11252,%dummy}, %rd15924; + } + shf.l.wrap.b32 %r11253, %r11252, %r11251, 21; + shf.l.wrap.b32 %r11254, %r11251, %r11252, 21; + mov.b64 %rd15933, {%r11254, %r11253}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11255}, %rd15930; + } + { + .reg .b32 %dummy; + mov.b64 {%r11256,%dummy}, %rd15930; + } + shf.l.wrap.b32 %r11257, %r11256, %r11255, 14; + shf.l.wrap.b32 %r11258, %r11255, %r11256, 14; + mov.b64 %rd15934, {%r11258, %r11257}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11259}, %rd15909; + } + { + .reg .b32 %dummy; + mov.b64 {%r11260,%dummy}, %rd15909; + } + shf.l.wrap.b32 %r11261, %r11260, %r11259, 28; + shf.l.wrap.b32 %r11262, %r11259, %r11260, 28; + mov.b64 %rd15935, {%r11262, %r11261}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11263}, %rd15915; + } + { + .reg .b32 %dummy; + mov.b64 {%r11264,%dummy}, %rd15915; + } + shf.l.wrap.b32 %r11265, %r11264, %r11263, 20; + shf.l.wrap.b32 %r11266, %r11263, %r11264, 20; + mov.b64 %rd15936, {%r11266, %r11265}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11267}, %rd15916; + } + { + .reg .b32 %dummy; + mov.b64 {%r11268,%dummy}, %rd15916; + } + shf.l.wrap.b32 %r11269, %r11268, %r11267, 3; + shf.l.wrap.b32 %r11270, %r11267, %r11268, 3; + mov.b64 %rd15937, {%r11270, %r11269}; + { + .reg .b32 %dummy; + mov.b64 {%r11271,%dummy}, %rd15922; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11272}, %rd15922; + } + shf.r.wrap.b32 %r11273, %r11272, %r11271, 19; + shf.r.wrap.b32 %r11274, %r11271, %r11272, 19; + mov.b64 %rd15938, {%r11274, %r11273}; + { + .reg .b32 %dummy; + mov.b64 {%r11275,%dummy}, %rd15928; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11276}, %rd15928; + } + shf.r.wrap.b32 %r11277, %r11276, %r11275, 3; + shf.r.wrap.b32 %r11278, %r11275, %r11276, 3; + mov.b64 %rd15939, {%r11278, %r11277}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11279}, %rd15907; + } + { + .reg .b32 %dummy; + mov.b64 {%r11280,%dummy}, %rd15907; + } + shf.l.wrap.b32 %r11281, %r11280, %r11279, 1; + shf.l.wrap.b32 %r11282, %r11279, %r11280, 1; + mov.b64 %rd15940, {%r11282, %r11281}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11283}, %rd15913; + } + { + .reg .b32 %dummy; + mov.b64 {%r11284,%dummy}, %rd15913; + } + shf.l.wrap.b32 %r11285, %r11284, %r11283, 6; + shf.l.wrap.b32 %r11286, %r11283, %r11284, 6; + mov.b64 %rd15941, {%r11286, %r11285}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11287}, %rd15919; + } + { + .reg .b32 %dummy; + mov.b64 {%r11288,%dummy}, %rd15919; + } + shf.l.wrap.b32 %r11289, %r11288, %r11287, 25; + shf.l.wrap.b32 %r11290, %r11287, %r11288, 25; + mov.b64 %rd15942, {%r11290, %r11289}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11291}, %rd15925; + } + { + .reg .b32 %dummy; + mov.b64 {%r11292,%dummy}, %rd15925; + } + shf.l.wrap.b32 %r11293, %r11292, %r11291, 8; + shf.l.wrap.b32 %r11294, %r11291, %r11292, 8; + mov.b64 %rd15943, {%r11294, %r11293}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11295}, %rd15926; + } + { + .reg .b32 %dummy; + mov.b64 {%r11296,%dummy}, %rd15926; + } + shf.l.wrap.b32 %r11297, %r11296, %r11295, 18; + shf.l.wrap.b32 %r11298, %r11295, %r11296, 18; + mov.b64 %rd15944, {%r11298, %r11297}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11299}, %rd15910; + } + { + .reg .b32 %dummy; + mov.b64 {%r11300,%dummy}, %rd15910; + } + shf.l.wrap.b32 %r11301, %r11300, %r11299, 27; + shf.l.wrap.b32 %r11302, %r11299, %r11300, 27; + mov.b64 %rd15945, {%r11302, %r11301}; + { + .reg .b32 %dummy; + mov.b64 {%r11303,%dummy}, %rd15911; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11304}, %rd15911; + } + shf.r.wrap.b32 %r11305, %r11304, %r11303, 28; + shf.r.wrap.b32 %r11306, %r11303, %r11304, 28; + mov.b64 %rd15946, {%r11306, %r11305}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11307}, %rd15917; + } + { + .reg .b32 %dummy; + mov.b64 {%r11308,%dummy}, %rd15917; + } + shf.l.wrap.b32 %r11309, %r11308, %r11307, 10; + shf.l.wrap.b32 %r11310, %r11307, %r11308, 10; + mov.b64 %rd15947, {%r11310, %r11309}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11311}, %rd15923; + } + { + .reg .b32 %dummy; + mov.b64 {%r11312,%dummy}, %rd15923; + } + shf.l.wrap.b32 %r11313, %r11312, %r11311, 15; + shf.l.wrap.b32 %r11314, %r11311, %r11312, 15; + mov.b64 %rd15948, {%r11314, %r11313}; + { + .reg .b32 %dummy; + mov.b64 {%r11315,%dummy}, %rd15929; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11316}, %rd15929; + } + shf.r.wrap.b32 %r11317, %r11316, %r11315, 8; + shf.r.wrap.b32 %r11318, %r11315, %r11316, 8; + mov.b64 %rd15949, {%r11318, %r11317}; + { + .reg .b32 %dummy; + mov.b64 {%r11319,%dummy}, %rd15908; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11320}, %rd15908; + } + shf.r.wrap.b32 %r11321, %r11320, %r11319, 2; + shf.r.wrap.b32 %r11322, %r11319, %r11320, 2; + mov.b64 %rd15950, {%r11322, %r11321}; + { + .reg .b32 %dummy; + mov.b64 {%r11323,%dummy}, %rd15914; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11324}, %rd15914; + } + shf.r.wrap.b32 %r11325, %r11324, %r11323, 9; + shf.r.wrap.b32 %r11326, %r11323, %r11324, 9; + mov.b64 %rd15951, {%r11326, %r11325}; + { + .reg .b32 %dummy; + mov.b64 {%r11327,%dummy}, %rd15920; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11328}, %rd15920; + } + shf.r.wrap.b32 %r11329, %r11328, %r11327, 25; + shf.r.wrap.b32 %r11330, %r11327, %r11328, 25; + mov.b64 %rd15952, {%r11330, %r11329}; + { + .reg .b32 %dummy; + mov.b64 {%r11331,%dummy}, %rd15921; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11332}, %rd15921; + } + shf.r.wrap.b32 %r11333, %r11332, %r11331, 23; + shf.r.wrap.b32 %r11334, %r11331, %r11332, 23; + mov.b64 %rd15953, {%r11334, %r11333}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11335}, %rd15927; + } + { + .reg .b32 %dummy; + mov.b64 {%r11336,%dummy}, %rd15927; + } + shf.l.wrap.b32 %r11337, %r11336, %r11335, 2; + shf.l.wrap.b32 %r11338, %r11335, %r11336, 2; + mov.b64 %rd15954, {%r11338, %r11337}; + not.b64 %rd15955, %rd15931; + and.b64 %rd15956, %rd15932, %rd15955; + xor.b64 %rd15957, %rd15956, %rd15906; + not.b64 %rd15958, %rd15932; + and.b64 %rd15959, %rd15933, %rd15958; + xor.b64 %rd15960, %rd15959, %rd15931; + not.b64 %rd15961, %rd15933; + and.b64 %rd15962, %rd15934, %rd15961; + xor.b64 %rd15963, %rd15962, %rd15932; + not.b64 %rd15964, %rd15934; + and.b64 %rd15965, %rd15906, %rd15964; + xor.b64 %rd15966, %rd15965, %rd15933; + not.b64 %rd15967, %rd15906; + and.b64 %rd15968, %rd15931, %rd15967; + xor.b64 %rd15969, %rd15968, %rd15934; + not.b64 %rd15970, %rd15936; + and.b64 %rd15971, %rd15937, %rd15970; + xor.b64 %rd15972, %rd15971, %rd15935; + not.b64 %rd15973, %rd15937; + and.b64 %rd15974, %rd15938, %rd15973; + xor.b64 %rd15975, %rd15974, %rd15936; + not.b64 %rd15976, %rd15938; + and.b64 %rd15977, %rd15939, %rd15976; + xor.b64 %rd15978, %rd15977, %rd15937; + not.b64 %rd15979, %rd15939; + and.b64 %rd15980, %rd15935, %rd15979; + xor.b64 %rd15981, %rd15980, %rd15938; + not.b64 %rd15982, %rd15935; + and.b64 %rd15983, %rd15936, %rd15982; + xor.b64 %rd15984, %rd15983, %rd15939; + not.b64 %rd15985, %rd15941; + and.b64 %rd15986, %rd15942, %rd15985; + xor.b64 %rd15987, %rd15986, %rd15940; + not.b64 %rd15988, %rd15942; + and.b64 %rd15989, %rd15943, %rd15988; + xor.b64 %rd15990, %rd15989, %rd15941; + not.b64 %rd15991, %rd15943; + and.b64 %rd15992, %rd15944, %rd15991; + xor.b64 %rd15993, %rd15992, %rd15942; + not.b64 %rd15994, %rd15944; + and.b64 %rd15995, %rd15940, %rd15994; + xor.b64 %rd15996, %rd15995, %rd15943; + not.b64 %rd15997, %rd15940; + and.b64 %rd15998, %rd15941, %rd15997; + xor.b64 %rd15999, %rd15998, %rd15944; + not.b64 %rd16000, %rd15946; + and.b64 %rd16001, %rd15947, %rd16000; + xor.b64 %rd16002, %rd16001, %rd15945; + not.b64 %rd16003, %rd15947; + and.b64 %rd16004, %rd15948, %rd16003; + xor.b64 %rd16005, %rd16004, %rd15946; + not.b64 %rd16006, %rd15948; + and.b64 %rd16007, %rd15949, %rd16006; + xor.b64 %rd16008, %rd16007, %rd15947; + not.b64 %rd16009, %rd15949; + and.b64 %rd16010, %rd15945, %rd16009; + xor.b64 %rd16011, %rd16010, %rd15948; + not.b64 %rd16012, %rd15945; + and.b64 %rd16013, %rd15946, %rd16012; + xor.b64 %rd16014, %rd16013, %rd15949; + not.b64 %rd16015, %rd15951; + and.b64 %rd16016, %rd15952, %rd16015; + xor.b64 %rd16017, %rd16016, %rd15950; + not.b64 %rd16018, %rd15952; + and.b64 %rd16019, %rd15953, %rd16018; + xor.b64 %rd16020, %rd16019, %rd15951; + not.b64 %rd16021, %rd15953; + and.b64 %rd16022, %rd15954, %rd16021; + xor.b64 %rd16023, %rd16022, %rd15952; + not.b64 %rd16024, %rd15954; + and.b64 %rd16025, %rd15950, %rd16024; + xor.b64 %rd16026, %rd16025, %rd15953; + not.b64 %rd16027, %rd15950; + and.b64 %rd16028, %rd15951, %rd16027; + xor.b64 %rd16029, %rd16028, %rd15954; + xor.b64 %rd16030, %rd15957, %rd19808; + xor.b64 %rd16031, %rd15972, %rd16030; + xor.b64 %rd16032, %rd16031, %rd15987; + xor.b64 %rd16033, %rd16032, %rd16002; + xor.b64 %rd16034, %rd16033, %rd16017; + xor.b64 %rd16035, %rd15975, %rd15960; + xor.b64 %rd16036, %rd16035, %rd15990; + xor.b64 %rd16037, %rd16036, %rd16005; + xor.b64 %rd16038, %rd16037, %rd16020; + xor.b64 %rd16039, %rd15978, %rd15963; + xor.b64 %rd16040, %rd16039, %rd15993; + xor.b64 %rd16041, %rd16040, %rd16008; + xor.b64 %rd16042, %rd16041, %rd16023; + xor.b64 %rd16043, %rd15981, %rd15966; + xor.b64 %rd16044, %rd16043, %rd15996; + xor.b64 %rd16045, %rd16044, %rd16011; + xor.b64 %rd16046, %rd16045, %rd16026; + xor.b64 %rd16047, %rd15984, %rd15969; + xor.b64 %rd16048, %rd16047, %rd15999; + xor.b64 %rd16049, %rd16048, %rd16014; + xor.b64 %rd16050, %rd16049, %rd16029; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11339}, %rd16038; + } + { + .reg .b32 %dummy; + mov.b64 {%r11340,%dummy}, %rd16038; + } + shf.l.wrap.b32 %r11341, %r11340, %r11339, 1; + shf.l.wrap.b32 %r11342, %r11339, %r11340, 1; + mov.b64 %rd16051, {%r11342, %r11341}; + xor.b64 %rd16052, %rd16050, %rd16051; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11343}, %rd16042; + } + { + .reg .b32 %dummy; + mov.b64 {%r11344,%dummy}, %rd16042; + } + shf.l.wrap.b32 %r11345, %r11344, %r11343, 1; + shf.l.wrap.b32 %r11346, %r11343, %r11344, 1; + mov.b64 %rd16053, {%r11346, %r11345}; + xor.b64 %rd16054, %rd16053, %rd16034; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11347}, %rd16046; + } + { + .reg .b32 %dummy; + mov.b64 {%r11348,%dummy}, %rd16046; + } + shf.l.wrap.b32 %r11349, %r11348, %r11347, 1; + shf.l.wrap.b32 %r11350, %r11347, %r11348, 1; + mov.b64 %rd16055, {%r11350, %r11349}; + xor.b64 %rd16056, %rd16055, %rd16038; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11351}, %rd16050; + } + { + .reg .b32 %dummy; + mov.b64 {%r11352,%dummy}, %rd16050; + } + shf.l.wrap.b32 %r11353, %r11352, %r11351, 1; + shf.l.wrap.b32 %r11354, %r11351, %r11352, 1; + mov.b64 %rd16057, {%r11354, %r11353}; + xor.b64 %rd16058, %rd16057, %rd16042; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11355}, %rd16034; + } + { + .reg .b32 %dummy; + mov.b64 {%r11356,%dummy}, %rd16034; + } + shf.l.wrap.b32 %r11357, %r11356, %r11355, 1; + shf.l.wrap.b32 %r11358, %r11355, %r11356, 1; + mov.b64 %rd16059, {%r11358, %r11357}; + xor.b64 %rd16060, %rd16046, %rd16059; + xor.b64 %rd16061, %rd16052, %rd16030; + xor.b64 %rd16062, %rd16054, %rd15960; + xor.b64 %rd16063, %rd16056, %rd15963; + xor.b64 %rd16064, %rd16058, %rd15966; + xor.b64 %rd16065, %rd16060, %rd15969; + xor.b64 %rd16066, %rd16052, %rd15972; + xor.b64 %rd16067, %rd16054, %rd15975; + xor.b64 %rd16068, %rd16056, %rd15978; + xor.b64 %rd16069, %rd16058, %rd15981; + xor.b64 %rd16070, %rd16060, %rd15984; + xor.b64 %rd16071, %rd16052, %rd15987; + xor.b64 %rd16072, %rd16054, %rd15990; + xor.b64 %rd16073, %rd16056, %rd15993; + xor.b64 %rd16074, %rd16058, %rd15996; + xor.b64 %rd16075, %rd16060, %rd15999; + xor.b64 %rd16076, %rd16052, %rd16002; + xor.b64 %rd16077, %rd16054, %rd16005; + xor.b64 %rd16078, %rd16056, %rd16008; + xor.b64 %rd16079, %rd16058, %rd16011; + xor.b64 %rd16080, %rd16060, %rd16014; + xor.b64 %rd16081, %rd16052, %rd16017; + xor.b64 %rd16082, %rd16054, %rd16020; + xor.b64 %rd16083, %rd16056, %rd16023; + xor.b64 %rd16084, %rd16058, %rd16026; + xor.b64 %rd16085, %rd16060, %rd16029; + { + .reg .b32 %dummy; + mov.b64 {%r11359,%dummy}, %rd16067; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11360}, %rd16067; + } + shf.r.wrap.b32 %r11361, %r11360, %r11359, 20; + shf.r.wrap.b32 %r11362, %r11359, %r11360, 20; + mov.b64 %rd16086, {%r11362, %r11361}; + { + .reg .b32 %dummy; + mov.b64 {%r11363,%dummy}, %rd16073; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11364}, %rd16073; + } + shf.r.wrap.b32 %r11365, %r11364, %r11363, 21; + shf.r.wrap.b32 %r11366, %r11363, %r11364, 21; + mov.b64 %rd16087, {%r11366, %r11365}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11367}, %rd16079; + } + { + .reg .b32 %dummy; + mov.b64 {%r11368,%dummy}, %rd16079; + } + shf.l.wrap.b32 %r11369, %r11368, %r11367, 21; + shf.l.wrap.b32 %r11370, %r11367, %r11368, 21; + mov.b64 %rd16088, {%r11370, %r11369}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11371}, %rd16085; + } + { + .reg .b32 %dummy; + mov.b64 {%r11372,%dummy}, %rd16085; + } + shf.l.wrap.b32 %r11373, %r11372, %r11371, 14; + shf.l.wrap.b32 %r11374, %r11371, %r11372, 14; + mov.b64 %rd16089, {%r11374, %r11373}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11375}, %rd16064; + } + { + .reg .b32 %dummy; + mov.b64 {%r11376,%dummy}, %rd16064; + } + shf.l.wrap.b32 %r11377, %r11376, %r11375, 28; + shf.l.wrap.b32 %r11378, %r11375, %r11376, 28; + mov.b64 %rd16090, {%r11378, %r11377}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11379}, %rd16070; + } + { + .reg .b32 %dummy; + mov.b64 {%r11380,%dummy}, %rd16070; + } + shf.l.wrap.b32 %r11381, %r11380, %r11379, 20; + shf.l.wrap.b32 %r11382, %r11379, %r11380, 20; + mov.b64 %rd16091, {%r11382, %r11381}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11383}, %rd16071; + } + { + .reg .b32 %dummy; + mov.b64 {%r11384,%dummy}, %rd16071; + } + shf.l.wrap.b32 %r11385, %r11384, %r11383, 3; + shf.l.wrap.b32 %r11386, %r11383, %r11384, 3; + mov.b64 %rd16092, {%r11386, %r11385}; + { + .reg .b32 %dummy; + mov.b64 {%r11387,%dummy}, %rd16077; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11388}, %rd16077; + } + shf.r.wrap.b32 %r11389, %r11388, %r11387, 19; + shf.r.wrap.b32 %r11390, %r11387, %r11388, 19; + mov.b64 %rd16093, {%r11390, %r11389}; + { + .reg .b32 %dummy; + mov.b64 {%r11391,%dummy}, %rd16083; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11392}, %rd16083; + } + shf.r.wrap.b32 %r11393, %r11392, %r11391, 3; + shf.r.wrap.b32 %r11394, %r11391, %r11392, 3; + mov.b64 %rd16094, {%r11394, %r11393}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11395}, %rd16062; + } + { + .reg .b32 %dummy; + mov.b64 {%r11396,%dummy}, %rd16062; + } + shf.l.wrap.b32 %r11397, %r11396, %r11395, 1; + shf.l.wrap.b32 %r11398, %r11395, %r11396, 1; + mov.b64 %rd16095, {%r11398, %r11397}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11399}, %rd16068; + } + { + .reg .b32 %dummy; + mov.b64 {%r11400,%dummy}, %rd16068; + } + shf.l.wrap.b32 %r11401, %r11400, %r11399, 6; + shf.l.wrap.b32 %r11402, %r11399, %r11400, 6; + mov.b64 %rd16096, {%r11402, %r11401}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11403}, %rd16074; + } + { + .reg .b32 %dummy; + mov.b64 {%r11404,%dummy}, %rd16074; + } + shf.l.wrap.b32 %r11405, %r11404, %r11403, 25; + shf.l.wrap.b32 %r11406, %r11403, %r11404, 25; + mov.b64 %rd16097, {%r11406, %r11405}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11407}, %rd16080; + } + { + .reg .b32 %dummy; + mov.b64 {%r11408,%dummy}, %rd16080; + } + shf.l.wrap.b32 %r11409, %r11408, %r11407, 8; + shf.l.wrap.b32 %r11410, %r11407, %r11408, 8; + mov.b64 %rd16098, {%r11410, %r11409}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11411}, %rd16081; + } + { + .reg .b32 %dummy; + mov.b64 {%r11412,%dummy}, %rd16081; + } + shf.l.wrap.b32 %r11413, %r11412, %r11411, 18; + shf.l.wrap.b32 %r11414, %r11411, %r11412, 18; + mov.b64 %rd16099, {%r11414, %r11413}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11415}, %rd16065; + } + { + .reg .b32 %dummy; + mov.b64 {%r11416,%dummy}, %rd16065; + } + shf.l.wrap.b32 %r11417, %r11416, %r11415, 27; + shf.l.wrap.b32 %r11418, %r11415, %r11416, 27; + mov.b64 %rd16100, {%r11418, %r11417}; + { + .reg .b32 %dummy; + mov.b64 {%r11419,%dummy}, %rd16066; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11420}, %rd16066; + } + shf.r.wrap.b32 %r11421, %r11420, %r11419, 28; + shf.r.wrap.b32 %r11422, %r11419, %r11420, 28; + mov.b64 %rd16101, {%r11422, %r11421}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11423}, %rd16072; + } + { + .reg .b32 %dummy; + mov.b64 {%r11424,%dummy}, %rd16072; + } + shf.l.wrap.b32 %r11425, %r11424, %r11423, 10; + shf.l.wrap.b32 %r11426, %r11423, %r11424, 10; + mov.b64 %rd16102, {%r11426, %r11425}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11427}, %rd16078; + } + { + .reg .b32 %dummy; + mov.b64 {%r11428,%dummy}, %rd16078; + } + shf.l.wrap.b32 %r11429, %r11428, %r11427, 15; + shf.l.wrap.b32 %r11430, %r11427, %r11428, 15; + mov.b64 %rd16103, {%r11430, %r11429}; + { + .reg .b32 %dummy; + mov.b64 {%r11431,%dummy}, %rd16084; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11432}, %rd16084; + } + shf.r.wrap.b32 %r11433, %r11432, %r11431, 8; + shf.r.wrap.b32 %r11434, %r11431, %r11432, 8; + mov.b64 %rd16104, {%r11434, %r11433}; + { + .reg .b32 %dummy; + mov.b64 {%r11435,%dummy}, %rd16063; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11436}, %rd16063; + } + shf.r.wrap.b32 %r11437, %r11436, %r11435, 2; + shf.r.wrap.b32 %r11438, %r11435, %r11436, 2; + mov.b64 %rd16105, {%r11438, %r11437}; + { + .reg .b32 %dummy; + mov.b64 {%r11439,%dummy}, %rd16069; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11440}, %rd16069; + } + shf.r.wrap.b32 %r11441, %r11440, %r11439, 9; + shf.r.wrap.b32 %r11442, %r11439, %r11440, 9; + mov.b64 %rd16106, {%r11442, %r11441}; + { + .reg .b32 %dummy; + mov.b64 {%r11443,%dummy}, %rd16075; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11444}, %rd16075; + } + shf.r.wrap.b32 %r11445, %r11444, %r11443, 25; + shf.r.wrap.b32 %r11446, %r11443, %r11444, 25; + mov.b64 %rd16107, {%r11446, %r11445}; + { + .reg .b32 %dummy; + mov.b64 {%r11447,%dummy}, %rd16076; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11448}, %rd16076; + } + shf.r.wrap.b32 %r11449, %r11448, %r11447, 23; + shf.r.wrap.b32 %r11450, %r11447, %r11448, 23; + mov.b64 %rd16108, {%r11450, %r11449}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11451}, %rd16082; + } + { + .reg .b32 %dummy; + mov.b64 {%r11452,%dummy}, %rd16082; + } + shf.l.wrap.b32 %r11453, %r11452, %r11451, 2; + shf.l.wrap.b32 %r11454, %r11451, %r11452, 2; + mov.b64 %rd16109, {%r11454, %r11453}; + not.b64 %rd16110, %rd16086; + and.b64 %rd16111, %rd16087, %rd16110; + xor.b64 %rd16112, %rd16111, %rd16061; + not.b64 %rd16113, %rd16087; + and.b64 %rd16114, %rd16088, %rd16113; + xor.b64 %rd16115, %rd16114, %rd16086; + not.b64 %rd16116, %rd16088; + and.b64 %rd16117, %rd16089, %rd16116; + xor.b64 %rd16118, %rd16117, %rd16087; + not.b64 %rd16119, %rd16089; + and.b64 %rd16120, %rd16061, %rd16119; + xor.b64 %rd16121, %rd16120, %rd16088; + not.b64 %rd16122, %rd16061; + and.b64 %rd16123, %rd16086, %rd16122; + xor.b64 %rd16124, %rd16123, %rd16089; + not.b64 %rd16125, %rd16091; + and.b64 %rd16126, %rd16092, %rd16125; + xor.b64 %rd16127, %rd16126, %rd16090; + not.b64 %rd16128, %rd16092; + and.b64 %rd16129, %rd16093, %rd16128; + xor.b64 %rd16130, %rd16129, %rd16091; + not.b64 %rd16131, %rd16093; + and.b64 %rd16132, %rd16094, %rd16131; + xor.b64 %rd16133, %rd16132, %rd16092; + not.b64 %rd16134, %rd16094; + and.b64 %rd16135, %rd16090, %rd16134; + xor.b64 %rd16136, %rd16135, %rd16093; + not.b64 %rd16137, %rd16090; + and.b64 %rd16138, %rd16091, %rd16137; + xor.b64 %rd16139, %rd16138, %rd16094; + not.b64 %rd16140, %rd16096; + and.b64 %rd16141, %rd16097, %rd16140; + xor.b64 %rd16142, %rd16141, %rd16095; + not.b64 %rd16143, %rd16097; + and.b64 %rd16144, %rd16098, %rd16143; + xor.b64 %rd16145, %rd16144, %rd16096; + not.b64 %rd16146, %rd16098; + and.b64 %rd16147, %rd16099, %rd16146; + xor.b64 %rd16148, %rd16147, %rd16097; + not.b64 %rd16149, %rd16099; + and.b64 %rd16150, %rd16095, %rd16149; + xor.b64 %rd16151, %rd16150, %rd16098; + not.b64 %rd16152, %rd16095; + and.b64 %rd16153, %rd16096, %rd16152; + xor.b64 %rd16154, %rd16153, %rd16099; + not.b64 %rd16155, %rd16101; + and.b64 %rd16156, %rd16102, %rd16155; + xor.b64 %rd16157, %rd16156, %rd16100; + not.b64 %rd16158, %rd16102; + and.b64 %rd16159, %rd16103, %rd16158; + xor.b64 %rd16160, %rd16159, %rd16101; + not.b64 %rd16161, %rd16103; + and.b64 %rd16162, %rd16104, %rd16161; + xor.b64 %rd16163, %rd16162, %rd16102; + not.b64 %rd16164, %rd16104; + and.b64 %rd16165, %rd16100, %rd16164; + xor.b64 %rd16166, %rd16165, %rd16103; + not.b64 %rd16167, %rd16100; + and.b64 %rd16168, %rd16101, %rd16167; + xor.b64 %rd16169, %rd16168, %rd16104; + not.b64 %rd16170, %rd16106; + and.b64 %rd16171, %rd16107, %rd16170; + xor.b64 %rd16172, %rd16171, %rd16105; + not.b64 %rd16173, %rd16107; + and.b64 %rd16174, %rd16108, %rd16173; + xor.b64 %rd16175, %rd16174, %rd16106; + not.b64 %rd16176, %rd16108; + and.b64 %rd16177, %rd16109, %rd16176; + xor.b64 %rd16178, %rd16177, %rd16107; + not.b64 %rd16179, %rd16109; + and.b64 %rd16180, %rd16105, %rd16179; + xor.b64 %rd16181, %rd16180, %rd16108; + not.b64 %rd16182, %rd16105; + and.b64 %rd16183, %rd16106, %rd16182; + xor.b64 %rd16184, %rd16183, %rd16109; + xor.b64 %rd16185, %rd16112, %rd19809; + xor.b64 %rd16186, %rd16127, %rd16185; + xor.b64 %rd16187, %rd16186, %rd16142; + xor.b64 %rd16188, %rd16187, %rd16157; + xor.b64 %rd16189, %rd16188, %rd16172; + xor.b64 %rd16190, %rd16130, %rd16115; + xor.b64 %rd16191, %rd16190, %rd16145; + xor.b64 %rd16192, %rd16191, %rd16160; + xor.b64 %rd16193, %rd16192, %rd16175; + xor.b64 %rd16194, %rd16133, %rd16118; + xor.b64 %rd16195, %rd16194, %rd16148; + xor.b64 %rd16196, %rd16195, %rd16163; + xor.b64 %rd16197, %rd16196, %rd16178; + xor.b64 %rd16198, %rd16136, %rd16121; + xor.b64 %rd16199, %rd16198, %rd16151; + xor.b64 %rd16200, %rd16199, %rd16166; + xor.b64 %rd16201, %rd16200, %rd16181; + xor.b64 %rd16202, %rd16139, %rd16124; + xor.b64 %rd16203, %rd16202, %rd16154; + xor.b64 %rd16204, %rd16203, %rd16169; + xor.b64 %rd16205, %rd16204, %rd16184; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11455}, %rd16193; + } + { + .reg .b32 %dummy; + mov.b64 {%r11456,%dummy}, %rd16193; + } + shf.l.wrap.b32 %r11457, %r11456, %r11455, 1; + shf.l.wrap.b32 %r11458, %r11455, %r11456, 1; + mov.b64 %rd16206, {%r11458, %r11457}; + xor.b64 %rd16207, %rd16205, %rd16206; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11459}, %rd16197; + } + { + .reg .b32 %dummy; + mov.b64 {%r11460,%dummy}, %rd16197; + } + shf.l.wrap.b32 %r11461, %r11460, %r11459, 1; + shf.l.wrap.b32 %r11462, %r11459, %r11460, 1; + mov.b64 %rd16208, {%r11462, %r11461}; + xor.b64 %rd16209, %rd16208, %rd16189; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11463}, %rd16201; + } + { + .reg .b32 %dummy; + mov.b64 {%r11464,%dummy}, %rd16201; + } + shf.l.wrap.b32 %r11465, %r11464, %r11463, 1; + shf.l.wrap.b32 %r11466, %r11463, %r11464, 1; + mov.b64 %rd16210, {%r11466, %r11465}; + xor.b64 %rd16211, %rd16210, %rd16193; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11467}, %rd16205; + } + { + .reg .b32 %dummy; + mov.b64 {%r11468,%dummy}, %rd16205; + } + shf.l.wrap.b32 %r11469, %r11468, %r11467, 1; + shf.l.wrap.b32 %r11470, %r11467, %r11468, 1; + mov.b64 %rd16212, {%r11470, %r11469}; + xor.b64 %rd16213, %rd16212, %rd16197; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11471}, %rd16189; + } + { + .reg .b32 %dummy; + mov.b64 {%r11472,%dummy}, %rd16189; + } + shf.l.wrap.b32 %r11473, %r11472, %r11471, 1; + shf.l.wrap.b32 %r11474, %r11471, %r11472, 1; + mov.b64 %rd16214, {%r11474, %r11473}; + xor.b64 %rd16215, %rd16201, %rd16214; + xor.b64 %rd16216, %rd16207, %rd16185; + xor.b64 %rd16217, %rd16209, %rd16115; + xor.b64 %rd16218, %rd16211, %rd16118; + xor.b64 %rd16219, %rd16213, %rd16121; + xor.b64 %rd16220, %rd16215, %rd16124; + xor.b64 %rd16221, %rd16207, %rd16127; + xor.b64 %rd16222, %rd16209, %rd16130; + xor.b64 %rd16223, %rd16211, %rd16133; + xor.b64 %rd16224, %rd16213, %rd16136; + xor.b64 %rd16225, %rd16215, %rd16139; + xor.b64 %rd16226, %rd16207, %rd16142; + xor.b64 %rd16227, %rd16209, %rd16145; + xor.b64 %rd16228, %rd16211, %rd16148; + xor.b64 %rd16229, %rd16213, %rd16151; + xor.b64 %rd16230, %rd16215, %rd16154; + xor.b64 %rd16231, %rd16207, %rd16157; + xor.b64 %rd16232, %rd16209, %rd16160; + xor.b64 %rd16233, %rd16211, %rd16163; + xor.b64 %rd16234, %rd16213, %rd16166; + xor.b64 %rd16235, %rd16215, %rd16169; + xor.b64 %rd16236, %rd16207, %rd16172; + xor.b64 %rd16237, %rd16209, %rd16175; + xor.b64 %rd16238, %rd16211, %rd16178; + xor.b64 %rd16239, %rd16213, %rd16181; + xor.b64 %rd16240, %rd16215, %rd16184; + { + .reg .b32 %dummy; + mov.b64 {%r11475,%dummy}, %rd16222; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11476}, %rd16222; + } + shf.r.wrap.b32 %r11477, %r11476, %r11475, 20; + shf.r.wrap.b32 %r11478, %r11475, %r11476, 20; + mov.b64 %rd16241, {%r11478, %r11477}; + { + .reg .b32 %dummy; + mov.b64 {%r11479,%dummy}, %rd16228; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11480}, %rd16228; + } + shf.r.wrap.b32 %r11481, %r11480, %r11479, 21; + shf.r.wrap.b32 %r11482, %r11479, %r11480, 21; + mov.b64 %rd16242, {%r11482, %r11481}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11483}, %rd16234; + } + { + .reg .b32 %dummy; + mov.b64 {%r11484,%dummy}, %rd16234; + } + shf.l.wrap.b32 %r11485, %r11484, %r11483, 21; + shf.l.wrap.b32 %r11486, %r11483, %r11484, 21; + mov.b64 %rd16243, {%r11486, %r11485}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11487}, %rd16240; + } + { + .reg .b32 %dummy; + mov.b64 {%r11488,%dummy}, %rd16240; + } + shf.l.wrap.b32 %r11489, %r11488, %r11487, 14; + shf.l.wrap.b32 %r11490, %r11487, %r11488, 14; + mov.b64 %rd16244, {%r11490, %r11489}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11491}, %rd16219; + } + { + .reg .b32 %dummy; + mov.b64 {%r11492,%dummy}, %rd16219; + } + shf.l.wrap.b32 %r11493, %r11492, %r11491, 28; + shf.l.wrap.b32 %r11494, %r11491, %r11492, 28; + mov.b64 %rd16245, {%r11494, %r11493}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11495}, %rd16225; + } + { + .reg .b32 %dummy; + mov.b64 {%r11496,%dummy}, %rd16225; + } + shf.l.wrap.b32 %r11497, %r11496, %r11495, 20; + shf.l.wrap.b32 %r11498, %r11495, %r11496, 20; + mov.b64 %rd16246, {%r11498, %r11497}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11499}, %rd16226; + } + { + .reg .b32 %dummy; + mov.b64 {%r11500,%dummy}, %rd16226; + } + shf.l.wrap.b32 %r11501, %r11500, %r11499, 3; + shf.l.wrap.b32 %r11502, %r11499, %r11500, 3; + mov.b64 %rd16247, {%r11502, %r11501}; + { + .reg .b32 %dummy; + mov.b64 {%r11503,%dummy}, %rd16232; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11504}, %rd16232; + } + shf.r.wrap.b32 %r11505, %r11504, %r11503, 19; + shf.r.wrap.b32 %r11506, %r11503, %r11504, 19; + mov.b64 %rd16248, {%r11506, %r11505}; + { + .reg .b32 %dummy; + mov.b64 {%r11507,%dummy}, %rd16238; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11508}, %rd16238; + } + shf.r.wrap.b32 %r11509, %r11508, %r11507, 3; + shf.r.wrap.b32 %r11510, %r11507, %r11508, 3; + mov.b64 %rd16249, {%r11510, %r11509}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11511}, %rd16217; + } + { + .reg .b32 %dummy; + mov.b64 {%r11512,%dummy}, %rd16217; + } + shf.l.wrap.b32 %r11513, %r11512, %r11511, 1; + shf.l.wrap.b32 %r11514, %r11511, %r11512, 1; + mov.b64 %rd16250, {%r11514, %r11513}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11515}, %rd16223; + } + { + .reg .b32 %dummy; + mov.b64 {%r11516,%dummy}, %rd16223; + } + shf.l.wrap.b32 %r11517, %r11516, %r11515, 6; + shf.l.wrap.b32 %r11518, %r11515, %r11516, 6; + mov.b64 %rd16251, {%r11518, %r11517}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11519}, %rd16229; + } + { + .reg .b32 %dummy; + mov.b64 {%r11520,%dummy}, %rd16229; + } + shf.l.wrap.b32 %r11521, %r11520, %r11519, 25; + shf.l.wrap.b32 %r11522, %r11519, %r11520, 25; + mov.b64 %rd16252, {%r11522, %r11521}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11523}, %rd16235; + } + { + .reg .b32 %dummy; + mov.b64 {%r11524,%dummy}, %rd16235; + } + shf.l.wrap.b32 %r11525, %r11524, %r11523, 8; + shf.l.wrap.b32 %r11526, %r11523, %r11524, 8; + mov.b64 %rd16253, {%r11526, %r11525}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11527}, %rd16236; + } + { + .reg .b32 %dummy; + mov.b64 {%r11528,%dummy}, %rd16236; + } + shf.l.wrap.b32 %r11529, %r11528, %r11527, 18; + shf.l.wrap.b32 %r11530, %r11527, %r11528, 18; + mov.b64 %rd16254, {%r11530, %r11529}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11531}, %rd16220; + } + { + .reg .b32 %dummy; + mov.b64 {%r11532,%dummy}, %rd16220; + } + shf.l.wrap.b32 %r11533, %r11532, %r11531, 27; + shf.l.wrap.b32 %r11534, %r11531, %r11532, 27; + mov.b64 %rd16255, {%r11534, %r11533}; + { + .reg .b32 %dummy; + mov.b64 {%r11535,%dummy}, %rd16221; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11536}, %rd16221; + } + shf.r.wrap.b32 %r11537, %r11536, %r11535, 28; + shf.r.wrap.b32 %r11538, %r11535, %r11536, 28; + mov.b64 %rd16256, {%r11538, %r11537}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11539}, %rd16227; + } + { + .reg .b32 %dummy; + mov.b64 {%r11540,%dummy}, %rd16227; + } + shf.l.wrap.b32 %r11541, %r11540, %r11539, 10; + shf.l.wrap.b32 %r11542, %r11539, %r11540, 10; + mov.b64 %rd16257, {%r11542, %r11541}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11543}, %rd16233; + } + { + .reg .b32 %dummy; + mov.b64 {%r11544,%dummy}, %rd16233; + } + shf.l.wrap.b32 %r11545, %r11544, %r11543, 15; + shf.l.wrap.b32 %r11546, %r11543, %r11544, 15; + mov.b64 %rd16258, {%r11546, %r11545}; + { + .reg .b32 %dummy; + mov.b64 {%r11547,%dummy}, %rd16239; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11548}, %rd16239; + } + shf.r.wrap.b32 %r11549, %r11548, %r11547, 8; + shf.r.wrap.b32 %r11550, %r11547, %r11548, 8; + mov.b64 %rd16259, {%r11550, %r11549}; + { + .reg .b32 %dummy; + mov.b64 {%r11551,%dummy}, %rd16218; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11552}, %rd16218; + } + shf.r.wrap.b32 %r11553, %r11552, %r11551, 2; + shf.r.wrap.b32 %r11554, %r11551, %r11552, 2; + mov.b64 %rd16260, {%r11554, %r11553}; + { + .reg .b32 %dummy; + mov.b64 {%r11555,%dummy}, %rd16224; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11556}, %rd16224; + } + shf.r.wrap.b32 %r11557, %r11556, %r11555, 9; + shf.r.wrap.b32 %r11558, %r11555, %r11556, 9; + mov.b64 %rd16261, {%r11558, %r11557}; + { + .reg .b32 %dummy; + mov.b64 {%r11559,%dummy}, %rd16230; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11560}, %rd16230; + } + shf.r.wrap.b32 %r11561, %r11560, %r11559, 25; + shf.r.wrap.b32 %r11562, %r11559, %r11560, 25; + mov.b64 %rd16262, {%r11562, %r11561}; + { + .reg .b32 %dummy; + mov.b64 {%r11563,%dummy}, %rd16231; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11564}, %rd16231; + } + shf.r.wrap.b32 %r11565, %r11564, %r11563, 23; + shf.r.wrap.b32 %r11566, %r11563, %r11564, 23; + mov.b64 %rd16263, {%r11566, %r11565}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11567}, %rd16237; + } + { + .reg .b32 %dummy; + mov.b64 {%r11568,%dummy}, %rd16237; + } + shf.l.wrap.b32 %r11569, %r11568, %r11567, 2; + shf.l.wrap.b32 %r11570, %r11567, %r11568, 2; + mov.b64 %rd16264, {%r11570, %r11569}; + not.b64 %rd16265, %rd16241; + and.b64 %rd16266, %rd16242, %rd16265; + xor.b64 %rd16267, %rd16266, %rd16216; + not.b64 %rd16268, %rd16242; + and.b64 %rd16269, %rd16243, %rd16268; + xor.b64 %rd16270, %rd16269, %rd16241; + not.b64 %rd16271, %rd16243; + and.b64 %rd16272, %rd16244, %rd16271; + xor.b64 %rd16273, %rd16272, %rd16242; + not.b64 %rd16274, %rd16244; + and.b64 %rd16275, %rd16216, %rd16274; + xor.b64 %rd16276, %rd16275, %rd16243; + not.b64 %rd16277, %rd16216; + and.b64 %rd16278, %rd16241, %rd16277; + xor.b64 %rd16279, %rd16278, %rd16244; + not.b64 %rd16280, %rd16246; + and.b64 %rd16281, %rd16247, %rd16280; + xor.b64 %rd16282, %rd16281, %rd16245; + not.b64 %rd16283, %rd16247; + and.b64 %rd16284, %rd16248, %rd16283; + xor.b64 %rd16285, %rd16284, %rd16246; + not.b64 %rd16286, %rd16248; + and.b64 %rd16287, %rd16249, %rd16286; + xor.b64 %rd16288, %rd16287, %rd16247; + not.b64 %rd16289, %rd16249; + and.b64 %rd16290, %rd16245, %rd16289; + xor.b64 %rd16291, %rd16290, %rd16248; + not.b64 %rd16292, %rd16245; + and.b64 %rd16293, %rd16246, %rd16292; + xor.b64 %rd16294, %rd16293, %rd16249; + not.b64 %rd16295, %rd16251; + and.b64 %rd16296, %rd16252, %rd16295; + xor.b64 %rd16297, %rd16296, %rd16250; + not.b64 %rd16298, %rd16252; + and.b64 %rd16299, %rd16253, %rd16298; + xor.b64 %rd16300, %rd16299, %rd16251; + not.b64 %rd16301, %rd16253; + and.b64 %rd16302, %rd16254, %rd16301; + xor.b64 %rd16303, %rd16302, %rd16252; + not.b64 %rd16304, %rd16254; + and.b64 %rd16305, %rd16250, %rd16304; + xor.b64 %rd16306, %rd16305, %rd16253; + not.b64 %rd16307, %rd16250; + and.b64 %rd16308, %rd16251, %rd16307; + xor.b64 %rd16309, %rd16308, %rd16254; + not.b64 %rd16310, %rd16256; + and.b64 %rd16311, %rd16257, %rd16310; + xor.b64 %rd16312, %rd16311, %rd16255; + not.b64 %rd16313, %rd16257; + and.b64 %rd16314, %rd16258, %rd16313; + xor.b64 %rd16315, %rd16314, %rd16256; + not.b64 %rd16316, %rd16258; + and.b64 %rd16317, %rd16259, %rd16316; + xor.b64 %rd16318, %rd16317, %rd16257; + not.b64 %rd16319, %rd16259; + and.b64 %rd16320, %rd16255, %rd16319; + xor.b64 %rd16321, %rd16320, %rd16258; + not.b64 %rd16322, %rd16255; + and.b64 %rd16323, %rd16256, %rd16322; + xor.b64 %rd16324, %rd16323, %rd16259; + not.b64 %rd16325, %rd16261; + and.b64 %rd16326, %rd16262, %rd16325; + xor.b64 %rd16327, %rd16326, %rd16260; + not.b64 %rd16328, %rd16262; + and.b64 %rd16329, %rd16263, %rd16328; + xor.b64 %rd16330, %rd16329, %rd16261; + not.b64 %rd16331, %rd16263; + and.b64 %rd16332, %rd16264, %rd16331; + xor.b64 %rd16333, %rd16332, %rd16262; + not.b64 %rd16334, %rd16264; + and.b64 %rd16335, %rd16260, %rd16334; + xor.b64 %rd16336, %rd16335, %rd16263; + not.b64 %rd16337, %rd16260; + and.b64 %rd16338, %rd16261, %rd16337; + xor.b64 %rd16339, %rd16338, %rd16264; + xor.b64 %rd16340, %rd16267, %rd19810; + xor.b64 %rd16341, %rd16282, %rd16340; + xor.b64 %rd16342, %rd16341, %rd16297; + xor.b64 %rd16343, %rd16342, %rd16312; + xor.b64 %rd16344, %rd16343, %rd16327; + xor.b64 %rd16345, %rd16285, %rd16270; + xor.b64 %rd16346, %rd16345, %rd16300; + xor.b64 %rd16347, %rd16346, %rd16315; + xor.b64 %rd16348, %rd16347, %rd16330; + xor.b64 %rd16349, %rd16288, %rd16273; + xor.b64 %rd16350, %rd16349, %rd16303; + xor.b64 %rd16351, %rd16350, %rd16318; + xor.b64 %rd16352, %rd16351, %rd16333; + xor.b64 %rd16353, %rd16291, %rd16276; + xor.b64 %rd16354, %rd16353, %rd16306; + xor.b64 %rd16355, %rd16354, %rd16321; + xor.b64 %rd16356, %rd16355, %rd16336; + xor.b64 %rd16357, %rd16294, %rd16279; + xor.b64 %rd16358, %rd16357, %rd16309; + xor.b64 %rd16359, %rd16358, %rd16324; + xor.b64 %rd16360, %rd16359, %rd16339; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11571}, %rd16348; + } + { + .reg .b32 %dummy; + mov.b64 {%r11572,%dummy}, %rd16348; + } + shf.l.wrap.b32 %r11573, %r11572, %r11571, 1; + shf.l.wrap.b32 %r11574, %r11571, %r11572, 1; + mov.b64 %rd16361, {%r11574, %r11573}; + xor.b64 %rd16362, %rd16360, %rd16361; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11575}, %rd16352; + } + { + .reg .b32 %dummy; + mov.b64 {%r11576,%dummy}, %rd16352; + } + shf.l.wrap.b32 %r11577, %r11576, %r11575, 1; + shf.l.wrap.b32 %r11578, %r11575, %r11576, 1; + mov.b64 %rd16363, {%r11578, %r11577}; + xor.b64 %rd16364, %rd16363, %rd16344; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11579}, %rd16356; + } + { + .reg .b32 %dummy; + mov.b64 {%r11580,%dummy}, %rd16356; + } + shf.l.wrap.b32 %r11581, %r11580, %r11579, 1; + shf.l.wrap.b32 %r11582, %r11579, %r11580, 1; + mov.b64 %rd16365, {%r11582, %r11581}; + xor.b64 %rd16366, %rd16365, %rd16348; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11583}, %rd16360; + } + { + .reg .b32 %dummy; + mov.b64 {%r11584,%dummy}, %rd16360; + } + shf.l.wrap.b32 %r11585, %r11584, %r11583, 1; + shf.l.wrap.b32 %r11586, %r11583, %r11584, 1; + mov.b64 %rd16367, {%r11586, %r11585}; + xor.b64 %rd16368, %rd16367, %rd16352; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11587}, %rd16344; + } + { + .reg .b32 %dummy; + mov.b64 {%r11588,%dummy}, %rd16344; + } + shf.l.wrap.b32 %r11589, %r11588, %r11587, 1; + shf.l.wrap.b32 %r11590, %r11587, %r11588, 1; + mov.b64 %rd16369, {%r11590, %r11589}; + xor.b64 %rd16370, %rd16356, %rd16369; + xor.b64 %rd16371, %rd16362, %rd16340; + xor.b64 %rd16372, %rd16364, %rd16270; + xor.b64 %rd16373, %rd16366, %rd16273; + xor.b64 %rd16374, %rd16368, %rd16276; + xor.b64 %rd16375, %rd16370, %rd16279; + xor.b64 %rd16376, %rd16362, %rd16282; + xor.b64 %rd16377, %rd16364, %rd16285; + xor.b64 %rd16378, %rd16366, %rd16288; + xor.b64 %rd16379, %rd16368, %rd16291; + xor.b64 %rd16380, %rd16370, %rd16294; + xor.b64 %rd16381, %rd16362, %rd16297; + xor.b64 %rd16382, %rd16364, %rd16300; + xor.b64 %rd16383, %rd16366, %rd16303; + xor.b64 %rd16384, %rd16368, %rd16306; + xor.b64 %rd16385, %rd16370, %rd16309; + xor.b64 %rd16386, %rd16362, %rd16312; + xor.b64 %rd16387, %rd16364, %rd16315; + xor.b64 %rd16388, %rd16366, %rd16318; + xor.b64 %rd16389, %rd16368, %rd16321; + xor.b64 %rd16390, %rd16370, %rd16324; + xor.b64 %rd16391, %rd16362, %rd16327; + xor.b64 %rd16392, %rd16364, %rd16330; + xor.b64 %rd16393, %rd16366, %rd16333; + xor.b64 %rd16394, %rd16368, %rd16336; + xor.b64 %rd16395, %rd16370, %rd16339; + { + .reg .b32 %dummy; + mov.b64 {%r11591,%dummy}, %rd16377; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11592}, %rd16377; + } + shf.r.wrap.b32 %r11593, %r11592, %r11591, 20; + shf.r.wrap.b32 %r11594, %r11591, %r11592, 20; + mov.b64 %rd16396, {%r11594, %r11593}; + { + .reg .b32 %dummy; + mov.b64 {%r11595,%dummy}, %rd16383; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11596}, %rd16383; + } + shf.r.wrap.b32 %r11597, %r11596, %r11595, 21; + shf.r.wrap.b32 %r11598, %r11595, %r11596, 21; + mov.b64 %rd16397, {%r11598, %r11597}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11599}, %rd16389; + } + { + .reg .b32 %dummy; + mov.b64 {%r11600,%dummy}, %rd16389; + } + shf.l.wrap.b32 %r11601, %r11600, %r11599, 21; + shf.l.wrap.b32 %r11602, %r11599, %r11600, 21; + mov.b64 %rd16398, {%r11602, %r11601}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11603}, %rd16395; + } + { + .reg .b32 %dummy; + mov.b64 {%r11604,%dummy}, %rd16395; + } + shf.l.wrap.b32 %r11605, %r11604, %r11603, 14; + shf.l.wrap.b32 %r11606, %r11603, %r11604, 14; + mov.b64 %rd16399, {%r11606, %r11605}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11607}, %rd16374; + } + { + .reg .b32 %dummy; + mov.b64 {%r11608,%dummy}, %rd16374; + } + shf.l.wrap.b32 %r11609, %r11608, %r11607, 28; + shf.l.wrap.b32 %r11610, %r11607, %r11608, 28; + mov.b64 %rd16400, {%r11610, %r11609}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11611}, %rd16380; + } + { + .reg .b32 %dummy; + mov.b64 {%r11612,%dummy}, %rd16380; + } + shf.l.wrap.b32 %r11613, %r11612, %r11611, 20; + shf.l.wrap.b32 %r11614, %r11611, %r11612, 20; + mov.b64 %rd16401, {%r11614, %r11613}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11615}, %rd16381; + } + { + .reg .b32 %dummy; + mov.b64 {%r11616,%dummy}, %rd16381; + } + shf.l.wrap.b32 %r11617, %r11616, %r11615, 3; + shf.l.wrap.b32 %r11618, %r11615, %r11616, 3; + mov.b64 %rd16402, {%r11618, %r11617}; + { + .reg .b32 %dummy; + mov.b64 {%r11619,%dummy}, %rd16387; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11620}, %rd16387; + } + shf.r.wrap.b32 %r11621, %r11620, %r11619, 19; + shf.r.wrap.b32 %r11622, %r11619, %r11620, 19; + mov.b64 %rd16403, {%r11622, %r11621}; + { + .reg .b32 %dummy; + mov.b64 {%r11623,%dummy}, %rd16393; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11624}, %rd16393; + } + shf.r.wrap.b32 %r11625, %r11624, %r11623, 3; + shf.r.wrap.b32 %r11626, %r11623, %r11624, 3; + mov.b64 %rd16404, {%r11626, %r11625}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11627}, %rd16372; + } + { + .reg .b32 %dummy; + mov.b64 {%r11628,%dummy}, %rd16372; + } + shf.l.wrap.b32 %r11629, %r11628, %r11627, 1; + shf.l.wrap.b32 %r11630, %r11627, %r11628, 1; + mov.b64 %rd16405, {%r11630, %r11629}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11631}, %rd16378; + } + { + .reg .b32 %dummy; + mov.b64 {%r11632,%dummy}, %rd16378; + } + shf.l.wrap.b32 %r11633, %r11632, %r11631, 6; + shf.l.wrap.b32 %r11634, %r11631, %r11632, 6; + mov.b64 %rd16406, {%r11634, %r11633}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11635}, %rd16384; + } + { + .reg .b32 %dummy; + mov.b64 {%r11636,%dummy}, %rd16384; + } + shf.l.wrap.b32 %r11637, %r11636, %r11635, 25; + shf.l.wrap.b32 %r11638, %r11635, %r11636, 25; + mov.b64 %rd16407, {%r11638, %r11637}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11639}, %rd16390; + } + { + .reg .b32 %dummy; + mov.b64 {%r11640,%dummy}, %rd16390; + } + shf.l.wrap.b32 %r11641, %r11640, %r11639, 8; + shf.l.wrap.b32 %r11642, %r11639, %r11640, 8; + mov.b64 %rd16408, {%r11642, %r11641}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11643}, %rd16391; + } + { + .reg .b32 %dummy; + mov.b64 {%r11644,%dummy}, %rd16391; + } + shf.l.wrap.b32 %r11645, %r11644, %r11643, 18; + shf.l.wrap.b32 %r11646, %r11643, %r11644, 18; + mov.b64 %rd16409, {%r11646, %r11645}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11647}, %rd16375; + } + { + .reg .b32 %dummy; + mov.b64 {%r11648,%dummy}, %rd16375; + } + shf.l.wrap.b32 %r11649, %r11648, %r11647, 27; + shf.l.wrap.b32 %r11650, %r11647, %r11648, 27; + mov.b64 %rd16410, {%r11650, %r11649}; + { + .reg .b32 %dummy; + mov.b64 {%r11651,%dummy}, %rd16376; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11652}, %rd16376; + } + shf.r.wrap.b32 %r11653, %r11652, %r11651, 28; + shf.r.wrap.b32 %r11654, %r11651, %r11652, 28; + mov.b64 %rd16411, {%r11654, %r11653}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11655}, %rd16382; + } + { + .reg .b32 %dummy; + mov.b64 {%r11656,%dummy}, %rd16382; + } + shf.l.wrap.b32 %r11657, %r11656, %r11655, 10; + shf.l.wrap.b32 %r11658, %r11655, %r11656, 10; + mov.b64 %rd16412, {%r11658, %r11657}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11659}, %rd16388; + } + { + .reg .b32 %dummy; + mov.b64 {%r11660,%dummy}, %rd16388; + } + shf.l.wrap.b32 %r11661, %r11660, %r11659, 15; + shf.l.wrap.b32 %r11662, %r11659, %r11660, 15; + mov.b64 %rd16413, {%r11662, %r11661}; + { + .reg .b32 %dummy; + mov.b64 {%r11663,%dummy}, %rd16394; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11664}, %rd16394; + } + shf.r.wrap.b32 %r11665, %r11664, %r11663, 8; + shf.r.wrap.b32 %r11666, %r11663, %r11664, 8; + mov.b64 %rd16414, {%r11666, %r11665}; + { + .reg .b32 %dummy; + mov.b64 {%r11667,%dummy}, %rd16373; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11668}, %rd16373; + } + shf.r.wrap.b32 %r11669, %r11668, %r11667, 2; + shf.r.wrap.b32 %r11670, %r11667, %r11668, 2; + mov.b64 %rd16415, {%r11670, %r11669}; + { + .reg .b32 %dummy; + mov.b64 {%r11671,%dummy}, %rd16379; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11672}, %rd16379; + } + shf.r.wrap.b32 %r11673, %r11672, %r11671, 9; + shf.r.wrap.b32 %r11674, %r11671, %r11672, 9; + mov.b64 %rd16416, {%r11674, %r11673}; + { + .reg .b32 %dummy; + mov.b64 {%r11675,%dummy}, %rd16385; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11676}, %rd16385; + } + shf.r.wrap.b32 %r11677, %r11676, %r11675, 25; + shf.r.wrap.b32 %r11678, %r11675, %r11676, 25; + mov.b64 %rd16417, {%r11678, %r11677}; + { + .reg .b32 %dummy; + mov.b64 {%r11679,%dummy}, %rd16386; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11680}, %rd16386; + } + shf.r.wrap.b32 %r11681, %r11680, %r11679, 23; + shf.r.wrap.b32 %r11682, %r11679, %r11680, 23; + mov.b64 %rd16418, {%r11682, %r11681}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11683}, %rd16392; + } + { + .reg .b32 %dummy; + mov.b64 {%r11684,%dummy}, %rd16392; + } + shf.l.wrap.b32 %r11685, %r11684, %r11683, 2; + shf.l.wrap.b32 %r11686, %r11683, %r11684, 2; + mov.b64 %rd16419, {%r11686, %r11685}; + not.b64 %rd16420, %rd16396; + and.b64 %rd16421, %rd16397, %rd16420; + xor.b64 %rd16422, %rd16421, %rd16371; + not.b64 %rd16423, %rd16397; + and.b64 %rd16424, %rd16398, %rd16423; + xor.b64 %rd16425, %rd16424, %rd16396; + not.b64 %rd16426, %rd16398; + and.b64 %rd16427, %rd16399, %rd16426; + xor.b64 %rd16428, %rd16427, %rd16397; + not.b64 %rd16429, %rd16399; + and.b64 %rd16430, %rd16371, %rd16429; + xor.b64 %rd16431, %rd16430, %rd16398; + not.b64 %rd16432, %rd16371; + and.b64 %rd16433, %rd16396, %rd16432; + xor.b64 %rd16434, %rd16433, %rd16399; + not.b64 %rd16435, %rd16401; + and.b64 %rd16436, %rd16402, %rd16435; + xor.b64 %rd16437, %rd16436, %rd16400; + not.b64 %rd16438, %rd16402; + and.b64 %rd16439, %rd16403, %rd16438; + xor.b64 %rd16440, %rd16439, %rd16401; + not.b64 %rd16441, %rd16403; + and.b64 %rd16442, %rd16404, %rd16441; + xor.b64 %rd16443, %rd16442, %rd16402; + not.b64 %rd16444, %rd16404; + and.b64 %rd16445, %rd16400, %rd16444; + xor.b64 %rd16446, %rd16445, %rd16403; + not.b64 %rd16447, %rd16400; + and.b64 %rd16448, %rd16401, %rd16447; + xor.b64 %rd16449, %rd16448, %rd16404; + not.b64 %rd16450, %rd16406; + and.b64 %rd16451, %rd16407, %rd16450; + xor.b64 %rd16452, %rd16451, %rd16405; + not.b64 %rd16453, %rd16407; + and.b64 %rd16454, %rd16408, %rd16453; + xor.b64 %rd16455, %rd16454, %rd16406; + not.b64 %rd16456, %rd16408; + and.b64 %rd16457, %rd16409, %rd16456; + xor.b64 %rd16458, %rd16457, %rd16407; + not.b64 %rd16459, %rd16409; + and.b64 %rd16460, %rd16405, %rd16459; + xor.b64 %rd16461, %rd16460, %rd16408; + not.b64 %rd16462, %rd16405; + and.b64 %rd16463, %rd16406, %rd16462; + xor.b64 %rd16464, %rd16463, %rd16409; + not.b64 %rd16465, %rd16411; + and.b64 %rd16466, %rd16412, %rd16465; + xor.b64 %rd16467, %rd16466, %rd16410; + not.b64 %rd16468, %rd16412; + and.b64 %rd16469, %rd16413, %rd16468; + xor.b64 %rd16470, %rd16469, %rd16411; + not.b64 %rd16471, %rd16413; + and.b64 %rd16472, %rd16414, %rd16471; + xor.b64 %rd16473, %rd16472, %rd16412; + not.b64 %rd16474, %rd16414; + and.b64 %rd16475, %rd16410, %rd16474; + xor.b64 %rd16476, %rd16475, %rd16413; + not.b64 %rd16477, %rd16410; + and.b64 %rd16478, %rd16411, %rd16477; + xor.b64 %rd16479, %rd16478, %rd16414; + not.b64 %rd16480, %rd16416; + and.b64 %rd16481, %rd16417, %rd16480; + xor.b64 %rd16482, %rd16481, %rd16415; + not.b64 %rd16483, %rd16417; + and.b64 %rd16484, %rd16418, %rd16483; + xor.b64 %rd16485, %rd16484, %rd16416; + not.b64 %rd16486, %rd16418; + and.b64 %rd16487, %rd16419, %rd16486; + xor.b64 %rd16488, %rd16487, %rd16417; + not.b64 %rd16489, %rd16419; + and.b64 %rd16490, %rd16415, %rd16489; + xor.b64 %rd16491, %rd16490, %rd16418; + not.b64 %rd16492, %rd16415; + and.b64 %rd16493, %rd16416, %rd16492; + xor.b64 %rd16494, %rd16493, %rd16419; + xor.b64 %rd16495, %rd16422, %rd19811; + xor.b64 %rd16496, %rd16437, %rd16495; + xor.b64 %rd16497, %rd16496, %rd16452; + xor.b64 %rd16498, %rd16497, %rd16467; + xor.b64 %rd16499, %rd16498, %rd16482; + xor.b64 %rd16500, %rd16440, %rd16425; + xor.b64 %rd16501, %rd16500, %rd16455; + xor.b64 %rd16502, %rd16501, %rd16470; + xor.b64 %rd16503, %rd16502, %rd16485; + xor.b64 %rd16504, %rd16443, %rd16428; + xor.b64 %rd16505, %rd16504, %rd16458; + xor.b64 %rd16506, %rd16505, %rd16473; + xor.b64 %rd16507, %rd16506, %rd16488; + xor.b64 %rd16508, %rd16446, %rd16431; + xor.b64 %rd16509, %rd16508, %rd16461; + xor.b64 %rd16510, %rd16509, %rd16476; + xor.b64 %rd16511, %rd16510, %rd16491; + xor.b64 %rd16512, %rd16449, %rd16434; + xor.b64 %rd16513, %rd16512, %rd16464; + xor.b64 %rd16514, %rd16513, %rd16479; + xor.b64 %rd16515, %rd16514, %rd16494; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11687}, %rd16503; + } + { + .reg .b32 %dummy; + mov.b64 {%r11688,%dummy}, %rd16503; + } + shf.l.wrap.b32 %r11689, %r11688, %r11687, 1; + shf.l.wrap.b32 %r11690, %r11687, %r11688, 1; + mov.b64 %rd16516, {%r11690, %r11689}; + xor.b64 %rd16517, %rd16515, %rd16516; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11691}, %rd16507; + } + { + .reg .b32 %dummy; + mov.b64 {%r11692,%dummy}, %rd16507; + } + shf.l.wrap.b32 %r11693, %r11692, %r11691, 1; + shf.l.wrap.b32 %r11694, %r11691, %r11692, 1; + mov.b64 %rd16518, {%r11694, %r11693}; + xor.b64 %rd16519, %rd16518, %rd16499; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11695}, %rd16511; + } + { + .reg .b32 %dummy; + mov.b64 {%r11696,%dummy}, %rd16511; + } + shf.l.wrap.b32 %r11697, %r11696, %r11695, 1; + shf.l.wrap.b32 %r11698, %r11695, %r11696, 1; + mov.b64 %rd16520, {%r11698, %r11697}; + xor.b64 %rd16521, %rd16520, %rd16503; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11699}, %rd16515; + } + { + .reg .b32 %dummy; + mov.b64 {%r11700,%dummy}, %rd16515; + } + shf.l.wrap.b32 %r11701, %r11700, %r11699, 1; + shf.l.wrap.b32 %r11702, %r11699, %r11700, 1; + mov.b64 %rd16522, {%r11702, %r11701}; + xor.b64 %rd16523, %rd16522, %rd16507; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11703}, %rd16499; + } + { + .reg .b32 %dummy; + mov.b64 {%r11704,%dummy}, %rd16499; + } + shf.l.wrap.b32 %r11705, %r11704, %r11703, 1; + shf.l.wrap.b32 %r11706, %r11703, %r11704, 1; + mov.b64 %rd16524, {%r11706, %r11705}; + xor.b64 %rd16525, %rd16511, %rd16524; + xor.b64 %rd16526, %rd16517, %rd16495; + xor.b64 %rd16527, %rd16519, %rd16425; + xor.b64 %rd16528, %rd16521, %rd16428; + xor.b64 %rd16529, %rd16523, %rd16431; + xor.b64 %rd16530, %rd16525, %rd16434; + xor.b64 %rd16531, %rd16517, %rd16437; + xor.b64 %rd16532, %rd16519, %rd16440; + xor.b64 %rd16533, %rd16521, %rd16443; + xor.b64 %rd16534, %rd16523, %rd16446; + xor.b64 %rd16535, %rd16525, %rd16449; + xor.b64 %rd16536, %rd16517, %rd16452; + xor.b64 %rd16537, %rd16519, %rd16455; + xor.b64 %rd16538, %rd16521, %rd16458; + xor.b64 %rd16539, %rd16523, %rd16461; + xor.b64 %rd16540, %rd16525, %rd16464; + xor.b64 %rd16541, %rd16517, %rd16467; + xor.b64 %rd16542, %rd16519, %rd16470; + xor.b64 %rd16543, %rd16521, %rd16473; + xor.b64 %rd16544, %rd16523, %rd16476; + xor.b64 %rd16545, %rd16525, %rd16479; + xor.b64 %rd16546, %rd16517, %rd16482; + xor.b64 %rd16547, %rd16519, %rd16485; + xor.b64 %rd16548, %rd16521, %rd16488; + xor.b64 %rd16549, %rd16523, %rd16491; + xor.b64 %rd16550, %rd16525, %rd16494; + { + .reg .b32 %dummy; + mov.b64 {%r11707,%dummy}, %rd16532; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11708}, %rd16532; + } + shf.r.wrap.b32 %r11709, %r11708, %r11707, 20; + shf.r.wrap.b32 %r11710, %r11707, %r11708, 20; + mov.b64 %rd16551, {%r11710, %r11709}; + { + .reg .b32 %dummy; + mov.b64 {%r11711,%dummy}, %rd16538; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11712}, %rd16538; + } + shf.r.wrap.b32 %r11713, %r11712, %r11711, 21; + shf.r.wrap.b32 %r11714, %r11711, %r11712, 21; + mov.b64 %rd16552, {%r11714, %r11713}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11715}, %rd16544; + } + { + .reg .b32 %dummy; + mov.b64 {%r11716,%dummy}, %rd16544; + } + shf.l.wrap.b32 %r11717, %r11716, %r11715, 21; + shf.l.wrap.b32 %r11718, %r11715, %r11716, 21; + mov.b64 %rd16553, {%r11718, %r11717}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11719}, %rd16550; + } + { + .reg .b32 %dummy; + mov.b64 {%r11720,%dummy}, %rd16550; + } + shf.l.wrap.b32 %r11721, %r11720, %r11719, 14; + shf.l.wrap.b32 %r11722, %r11719, %r11720, 14; + mov.b64 %rd16554, {%r11722, %r11721}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11723}, %rd16529; + } + { + .reg .b32 %dummy; + mov.b64 {%r11724,%dummy}, %rd16529; + } + shf.l.wrap.b32 %r11725, %r11724, %r11723, 28; + shf.l.wrap.b32 %r11726, %r11723, %r11724, 28; + mov.b64 %rd16555, {%r11726, %r11725}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11727}, %rd16535; + } + { + .reg .b32 %dummy; + mov.b64 {%r11728,%dummy}, %rd16535; + } + shf.l.wrap.b32 %r11729, %r11728, %r11727, 20; + shf.l.wrap.b32 %r11730, %r11727, %r11728, 20; + mov.b64 %rd16556, {%r11730, %r11729}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11731}, %rd16536; + } + { + .reg .b32 %dummy; + mov.b64 {%r11732,%dummy}, %rd16536; + } + shf.l.wrap.b32 %r11733, %r11732, %r11731, 3; + shf.l.wrap.b32 %r11734, %r11731, %r11732, 3; + mov.b64 %rd16557, {%r11734, %r11733}; + { + .reg .b32 %dummy; + mov.b64 {%r11735,%dummy}, %rd16542; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11736}, %rd16542; + } + shf.r.wrap.b32 %r11737, %r11736, %r11735, 19; + shf.r.wrap.b32 %r11738, %r11735, %r11736, 19; + mov.b64 %rd16558, {%r11738, %r11737}; + { + .reg .b32 %dummy; + mov.b64 {%r11739,%dummy}, %rd16548; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11740}, %rd16548; + } + shf.r.wrap.b32 %r11741, %r11740, %r11739, 3; + shf.r.wrap.b32 %r11742, %r11739, %r11740, 3; + mov.b64 %rd16559, {%r11742, %r11741}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11743}, %rd16527; + } + { + .reg .b32 %dummy; + mov.b64 {%r11744,%dummy}, %rd16527; + } + shf.l.wrap.b32 %r11745, %r11744, %r11743, 1; + shf.l.wrap.b32 %r11746, %r11743, %r11744, 1; + mov.b64 %rd16560, {%r11746, %r11745}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11747}, %rd16533; + } + { + .reg .b32 %dummy; + mov.b64 {%r11748,%dummy}, %rd16533; + } + shf.l.wrap.b32 %r11749, %r11748, %r11747, 6; + shf.l.wrap.b32 %r11750, %r11747, %r11748, 6; + mov.b64 %rd16561, {%r11750, %r11749}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11751}, %rd16539; + } + { + .reg .b32 %dummy; + mov.b64 {%r11752,%dummy}, %rd16539; + } + shf.l.wrap.b32 %r11753, %r11752, %r11751, 25; + shf.l.wrap.b32 %r11754, %r11751, %r11752, 25; + mov.b64 %rd16562, {%r11754, %r11753}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11755}, %rd16545; + } + { + .reg .b32 %dummy; + mov.b64 {%r11756,%dummy}, %rd16545; + } + shf.l.wrap.b32 %r11757, %r11756, %r11755, 8; + shf.l.wrap.b32 %r11758, %r11755, %r11756, 8; + mov.b64 %rd16563, {%r11758, %r11757}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11759}, %rd16546; + } + { + .reg .b32 %dummy; + mov.b64 {%r11760,%dummy}, %rd16546; + } + shf.l.wrap.b32 %r11761, %r11760, %r11759, 18; + shf.l.wrap.b32 %r11762, %r11759, %r11760, 18; + mov.b64 %rd16564, {%r11762, %r11761}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11763}, %rd16530; + } + { + .reg .b32 %dummy; + mov.b64 {%r11764,%dummy}, %rd16530; + } + shf.l.wrap.b32 %r11765, %r11764, %r11763, 27; + shf.l.wrap.b32 %r11766, %r11763, %r11764, 27; + mov.b64 %rd16565, {%r11766, %r11765}; + { + .reg .b32 %dummy; + mov.b64 {%r11767,%dummy}, %rd16531; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11768}, %rd16531; + } + shf.r.wrap.b32 %r11769, %r11768, %r11767, 28; + shf.r.wrap.b32 %r11770, %r11767, %r11768, 28; + mov.b64 %rd16566, {%r11770, %r11769}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11771}, %rd16537; + } + { + .reg .b32 %dummy; + mov.b64 {%r11772,%dummy}, %rd16537; + } + shf.l.wrap.b32 %r11773, %r11772, %r11771, 10; + shf.l.wrap.b32 %r11774, %r11771, %r11772, 10; + mov.b64 %rd16567, {%r11774, %r11773}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11775}, %rd16543; + } + { + .reg .b32 %dummy; + mov.b64 {%r11776,%dummy}, %rd16543; + } + shf.l.wrap.b32 %r11777, %r11776, %r11775, 15; + shf.l.wrap.b32 %r11778, %r11775, %r11776, 15; + mov.b64 %rd16568, {%r11778, %r11777}; + { + .reg .b32 %dummy; + mov.b64 {%r11779,%dummy}, %rd16549; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11780}, %rd16549; + } + shf.r.wrap.b32 %r11781, %r11780, %r11779, 8; + shf.r.wrap.b32 %r11782, %r11779, %r11780, 8; + mov.b64 %rd16569, {%r11782, %r11781}; + { + .reg .b32 %dummy; + mov.b64 {%r11783,%dummy}, %rd16528; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11784}, %rd16528; + } + shf.r.wrap.b32 %r11785, %r11784, %r11783, 2; + shf.r.wrap.b32 %r11786, %r11783, %r11784, 2; + mov.b64 %rd16570, {%r11786, %r11785}; + { + .reg .b32 %dummy; + mov.b64 {%r11787,%dummy}, %rd16534; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11788}, %rd16534; + } + shf.r.wrap.b32 %r11789, %r11788, %r11787, 9; + shf.r.wrap.b32 %r11790, %r11787, %r11788, 9; + mov.b64 %rd16571, {%r11790, %r11789}; + { + .reg .b32 %dummy; + mov.b64 {%r11791,%dummy}, %rd16540; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11792}, %rd16540; + } + shf.r.wrap.b32 %r11793, %r11792, %r11791, 25; + shf.r.wrap.b32 %r11794, %r11791, %r11792, 25; + mov.b64 %rd16572, {%r11794, %r11793}; + { + .reg .b32 %dummy; + mov.b64 {%r11795,%dummy}, %rd16541; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11796}, %rd16541; + } + shf.r.wrap.b32 %r11797, %r11796, %r11795, 23; + shf.r.wrap.b32 %r11798, %r11795, %r11796, 23; + mov.b64 %rd16573, {%r11798, %r11797}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11799}, %rd16547; + } + { + .reg .b32 %dummy; + mov.b64 {%r11800,%dummy}, %rd16547; + } + shf.l.wrap.b32 %r11801, %r11800, %r11799, 2; + shf.l.wrap.b32 %r11802, %r11799, %r11800, 2; + mov.b64 %rd16574, {%r11802, %r11801}; + not.b64 %rd16575, %rd16551; + and.b64 %rd16576, %rd16552, %rd16575; + xor.b64 %rd16577, %rd16576, %rd16526; + not.b64 %rd16578, %rd16552; + and.b64 %rd16579, %rd16553, %rd16578; + xor.b64 %rd16580, %rd16579, %rd16551; + not.b64 %rd16581, %rd16553; + and.b64 %rd16582, %rd16554, %rd16581; + xor.b64 %rd16583, %rd16582, %rd16552; + not.b64 %rd16584, %rd16554; + and.b64 %rd16585, %rd16526, %rd16584; + xor.b64 %rd16586, %rd16585, %rd16553; + not.b64 %rd16587, %rd16526; + and.b64 %rd16588, %rd16551, %rd16587; + xor.b64 %rd16589, %rd16588, %rd16554; + not.b64 %rd16590, %rd16556; + and.b64 %rd16591, %rd16557, %rd16590; + xor.b64 %rd16592, %rd16591, %rd16555; + not.b64 %rd16593, %rd16557; + and.b64 %rd16594, %rd16558, %rd16593; + xor.b64 %rd16595, %rd16594, %rd16556; + not.b64 %rd16596, %rd16558; + and.b64 %rd16597, %rd16559, %rd16596; + xor.b64 %rd16598, %rd16597, %rd16557; + not.b64 %rd16599, %rd16559; + and.b64 %rd16600, %rd16555, %rd16599; + xor.b64 %rd16601, %rd16600, %rd16558; + not.b64 %rd16602, %rd16555; + and.b64 %rd16603, %rd16556, %rd16602; + xor.b64 %rd16604, %rd16603, %rd16559; + not.b64 %rd16605, %rd16561; + and.b64 %rd16606, %rd16562, %rd16605; + xor.b64 %rd16607, %rd16606, %rd16560; + not.b64 %rd16608, %rd16562; + and.b64 %rd16609, %rd16563, %rd16608; + xor.b64 %rd16610, %rd16609, %rd16561; + not.b64 %rd16611, %rd16563; + and.b64 %rd16612, %rd16564, %rd16611; + xor.b64 %rd16613, %rd16612, %rd16562; + not.b64 %rd16614, %rd16564; + and.b64 %rd16615, %rd16560, %rd16614; + xor.b64 %rd16616, %rd16615, %rd16563; + not.b64 %rd16617, %rd16560; + and.b64 %rd16618, %rd16561, %rd16617; + xor.b64 %rd16619, %rd16618, %rd16564; + not.b64 %rd16620, %rd16566; + and.b64 %rd16621, %rd16567, %rd16620; + xor.b64 %rd16622, %rd16621, %rd16565; + not.b64 %rd16623, %rd16567; + and.b64 %rd16624, %rd16568, %rd16623; + xor.b64 %rd16625, %rd16624, %rd16566; + not.b64 %rd16626, %rd16568; + and.b64 %rd16627, %rd16569, %rd16626; + xor.b64 %rd16628, %rd16627, %rd16567; + not.b64 %rd16629, %rd16569; + and.b64 %rd16630, %rd16565, %rd16629; + xor.b64 %rd16631, %rd16630, %rd16568; + not.b64 %rd16632, %rd16565; + and.b64 %rd16633, %rd16566, %rd16632; + xor.b64 %rd16634, %rd16633, %rd16569; + not.b64 %rd16635, %rd16571; + and.b64 %rd16636, %rd16572, %rd16635; + xor.b64 %rd16637, %rd16636, %rd16570; + not.b64 %rd16638, %rd16572; + and.b64 %rd16639, %rd16573, %rd16638; + xor.b64 %rd16640, %rd16639, %rd16571; + not.b64 %rd16641, %rd16573; + and.b64 %rd16642, %rd16574, %rd16641; + xor.b64 %rd16643, %rd16642, %rd16572; + not.b64 %rd16644, %rd16574; + and.b64 %rd16645, %rd16570, %rd16644; + xor.b64 %rd16646, %rd16645, %rd16573; + not.b64 %rd16647, %rd16570; + and.b64 %rd16648, %rd16571, %rd16647; + xor.b64 %rd16649, %rd16648, %rd16574; + xor.b64 %rd16650, %rd16577, %rd19812; + xor.b64 %rd16651, %rd16592, %rd16650; + xor.b64 %rd16652, %rd16651, %rd16607; + xor.b64 %rd16653, %rd16652, %rd16622; + xor.b64 %rd16654, %rd16653, %rd16637; + xor.b64 %rd16655, %rd16595, %rd16580; + xor.b64 %rd16656, %rd16655, %rd16610; + xor.b64 %rd16657, %rd16656, %rd16625; + xor.b64 %rd16658, %rd16657, %rd16640; + xor.b64 %rd16659, %rd16598, %rd16583; + xor.b64 %rd16660, %rd16659, %rd16613; + xor.b64 %rd16661, %rd16660, %rd16628; + xor.b64 %rd16662, %rd16661, %rd16643; + xor.b64 %rd16663, %rd16601, %rd16586; + xor.b64 %rd16664, %rd16663, %rd16616; + xor.b64 %rd16665, %rd16664, %rd16631; + xor.b64 %rd16666, %rd16665, %rd16646; + xor.b64 %rd16667, %rd16604, %rd16589; + xor.b64 %rd16668, %rd16667, %rd16619; + xor.b64 %rd16669, %rd16668, %rd16634; + xor.b64 %rd16670, %rd16669, %rd16649; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11803}, %rd16658; + } + { + .reg .b32 %dummy; + mov.b64 {%r11804,%dummy}, %rd16658; + } + shf.l.wrap.b32 %r11805, %r11804, %r11803, 1; + shf.l.wrap.b32 %r11806, %r11803, %r11804, 1; + mov.b64 %rd16671, {%r11806, %r11805}; + xor.b64 %rd16672, %rd16670, %rd16671; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11807}, %rd16662; + } + { + .reg .b32 %dummy; + mov.b64 {%r11808,%dummy}, %rd16662; + } + shf.l.wrap.b32 %r11809, %r11808, %r11807, 1; + shf.l.wrap.b32 %r11810, %r11807, %r11808, 1; + mov.b64 %rd16673, {%r11810, %r11809}; + xor.b64 %rd16674, %rd16673, %rd16654; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11811}, %rd16666; + } + { + .reg .b32 %dummy; + mov.b64 {%r11812,%dummy}, %rd16666; + } + shf.l.wrap.b32 %r11813, %r11812, %r11811, 1; + shf.l.wrap.b32 %r11814, %r11811, %r11812, 1; + mov.b64 %rd16675, {%r11814, %r11813}; + xor.b64 %rd16676, %rd16675, %rd16658; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11815}, %rd16670; + } + { + .reg .b32 %dummy; + mov.b64 {%r11816,%dummy}, %rd16670; + } + shf.l.wrap.b32 %r11817, %r11816, %r11815, 1; + shf.l.wrap.b32 %r11818, %r11815, %r11816, 1; + mov.b64 %rd16677, {%r11818, %r11817}; + xor.b64 %rd16678, %rd16677, %rd16662; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11819}, %rd16654; + } + { + .reg .b32 %dummy; + mov.b64 {%r11820,%dummy}, %rd16654; + } + shf.l.wrap.b32 %r11821, %r11820, %r11819, 1; + shf.l.wrap.b32 %r11822, %r11819, %r11820, 1; + mov.b64 %rd16679, {%r11822, %r11821}; + xor.b64 %rd16680, %rd16666, %rd16679; + xor.b64 %rd16681, %rd16672, %rd16650; + xor.b64 %rd16682, %rd16674, %rd16580; + xor.b64 %rd16683, %rd16676, %rd16583; + xor.b64 %rd16684, %rd16678, %rd16586; + xor.b64 %rd16685, %rd16680, %rd16589; + xor.b64 %rd16686, %rd16672, %rd16592; + xor.b64 %rd16687, %rd16674, %rd16595; + xor.b64 %rd16688, %rd16676, %rd16598; + xor.b64 %rd16689, %rd16678, %rd16601; + xor.b64 %rd16690, %rd16680, %rd16604; + xor.b64 %rd16691, %rd16672, %rd16607; + xor.b64 %rd16692, %rd16674, %rd16610; + xor.b64 %rd16693, %rd16676, %rd16613; + xor.b64 %rd16694, %rd16678, %rd16616; + xor.b64 %rd16695, %rd16680, %rd16619; + xor.b64 %rd16696, %rd16672, %rd16622; + xor.b64 %rd16697, %rd16674, %rd16625; + xor.b64 %rd16698, %rd16676, %rd16628; + xor.b64 %rd16699, %rd16678, %rd16631; + xor.b64 %rd16700, %rd16680, %rd16634; + xor.b64 %rd16701, %rd16672, %rd16637; + xor.b64 %rd16702, %rd16674, %rd16640; + xor.b64 %rd16703, %rd16676, %rd16643; + xor.b64 %rd16704, %rd16678, %rd16646; + xor.b64 %rd16705, %rd16680, %rd16649; + { + .reg .b32 %dummy; + mov.b64 {%r11823,%dummy}, %rd16687; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11824}, %rd16687; + } + shf.r.wrap.b32 %r11825, %r11824, %r11823, 20; + shf.r.wrap.b32 %r11826, %r11823, %r11824, 20; + mov.b64 %rd16706, {%r11826, %r11825}; + { + .reg .b32 %dummy; + mov.b64 {%r11827,%dummy}, %rd16693; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11828}, %rd16693; + } + shf.r.wrap.b32 %r11829, %r11828, %r11827, 21; + shf.r.wrap.b32 %r11830, %r11827, %r11828, 21; + mov.b64 %rd16707, {%r11830, %r11829}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11831}, %rd16699; + } + { + .reg .b32 %dummy; + mov.b64 {%r11832,%dummy}, %rd16699; + } + shf.l.wrap.b32 %r11833, %r11832, %r11831, 21; + shf.l.wrap.b32 %r11834, %r11831, %r11832, 21; + mov.b64 %rd16708, {%r11834, %r11833}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11835}, %rd16705; + } + { + .reg .b32 %dummy; + mov.b64 {%r11836,%dummy}, %rd16705; + } + shf.l.wrap.b32 %r11837, %r11836, %r11835, 14; + shf.l.wrap.b32 %r11838, %r11835, %r11836, 14; + mov.b64 %rd16709, {%r11838, %r11837}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11839}, %rd16684; + } + { + .reg .b32 %dummy; + mov.b64 {%r11840,%dummy}, %rd16684; + } + shf.l.wrap.b32 %r11841, %r11840, %r11839, 28; + shf.l.wrap.b32 %r11842, %r11839, %r11840, 28; + mov.b64 %rd16710, {%r11842, %r11841}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11843}, %rd16690; + } + { + .reg .b32 %dummy; + mov.b64 {%r11844,%dummy}, %rd16690; + } + shf.l.wrap.b32 %r11845, %r11844, %r11843, 20; + shf.l.wrap.b32 %r11846, %r11843, %r11844, 20; + mov.b64 %rd16711, {%r11846, %r11845}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11847}, %rd16691; + } + { + .reg .b32 %dummy; + mov.b64 {%r11848,%dummy}, %rd16691; + } + shf.l.wrap.b32 %r11849, %r11848, %r11847, 3; + shf.l.wrap.b32 %r11850, %r11847, %r11848, 3; + mov.b64 %rd16712, {%r11850, %r11849}; + { + .reg .b32 %dummy; + mov.b64 {%r11851,%dummy}, %rd16697; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11852}, %rd16697; + } + shf.r.wrap.b32 %r11853, %r11852, %r11851, 19; + shf.r.wrap.b32 %r11854, %r11851, %r11852, 19; + mov.b64 %rd16713, {%r11854, %r11853}; + { + .reg .b32 %dummy; + mov.b64 {%r11855,%dummy}, %rd16703; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11856}, %rd16703; + } + shf.r.wrap.b32 %r11857, %r11856, %r11855, 3; + shf.r.wrap.b32 %r11858, %r11855, %r11856, 3; + mov.b64 %rd16714, {%r11858, %r11857}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11859}, %rd16682; + } + { + .reg .b32 %dummy; + mov.b64 {%r11860,%dummy}, %rd16682; + } + shf.l.wrap.b32 %r11861, %r11860, %r11859, 1; + shf.l.wrap.b32 %r11862, %r11859, %r11860, 1; + mov.b64 %rd16715, {%r11862, %r11861}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11863}, %rd16688; + } + { + .reg .b32 %dummy; + mov.b64 {%r11864,%dummy}, %rd16688; + } + shf.l.wrap.b32 %r11865, %r11864, %r11863, 6; + shf.l.wrap.b32 %r11866, %r11863, %r11864, 6; + mov.b64 %rd16716, {%r11866, %r11865}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11867}, %rd16694; + } + { + .reg .b32 %dummy; + mov.b64 {%r11868,%dummy}, %rd16694; + } + shf.l.wrap.b32 %r11869, %r11868, %r11867, 25; + shf.l.wrap.b32 %r11870, %r11867, %r11868, 25; + mov.b64 %rd16717, {%r11870, %r11869}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11871}, %rd16700; + } + { + .reg .b32 %dummy; + mov.b64 {%r11872,%dummy}, %rd16700; + } + shf.l.wrap.b32 %r11873, %r11872, %r11871, 8; + shf.l.wrap.b32 %r11874, %r11871, %r11872, 8; + mov.b64 %rd16718, {%r11874, %r11873}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11875}, %rd16701; + } + { + .reg .b32 %dummy; + mov.b64 {%r11876,%dummy}, %rd16701; + } + shf.l.wrap.b32 %r11877, %r11876, %r11875, 18; + shf.l.wrap.b32 %r11878, %r11875, %r11876, 18; + mov.b64 %rd16719, {%r11878, %r11877}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11879}, %rd16685; + } + { + .reg .b32 %dummy; + mov.b64 {%r11880,%dummy}, %rd16685; + } + shf.l.wrap.b32 %r11881, %r11880, %r11879, 27; + shf.l.wrap.b32 %r11882, %r11879, %r11880, 27; + mov.b64 %rd16720, {%r11882, %r11881}; + { + .reg .b32 %dummy; + mov.b64 {%r11883,%dummy}, %rd16686; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11884}, %rd16686; + } + shf.r.wrap.b32 %r11885, %r11884, %r11883, 28; + shf.r.wrap.b32 %r11886, %r11883, %r11884, 28; + mov.b64 %rd16721, {%r11886, %r11885}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11887}, %rd16692; + } + { + .reg .b32 %dummy; + mov.b64 {%r11888,%dummy}, %rd16692; + } + shf.l.wrap.b32 %r11889, %r11888, %r11887, 10; + shf.l.wrap.b32 %r11890, %r11887, %r11888, 10; + mov.b64 %rd16722, {%r11890, %r11889}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11891}, %rd16698; + } + { + .reg .b32 %dummy; + mov.b64 {%r11892,%dummy}, %rd16698; + } + shf.l.wrap.b32 %r11893, %r11892, %r11891, 15; + shf.l.wrap.b32 %r11894, %r11891, %r11892, 15; + mov.b64 %rd16723, {%r11894, %r11893}; + { + .reg .b32 %dummy; + mov.b64 {%r11895,%dummy}, %rd16704; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11896}, %rd16704; + } + shf.r.wrap.b32 %r11897, %r11896, %r11895, 8; + shf.r.wrap.b32 %r11898, %r11895, %r11896, 8; + mov.b64 %rd16724, {%r11898, %r11897}; + { + .reg .b32 %dummy; + mov.b64 {%r11899,%dummy}, %rd16683; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11900}, %rd16683; + } + shf.r.wrap.b32 %r11901, %r11900, %r11899, 2; + shf.r.wrap.b32 %r11902, %r11899, %r11900, 2; + mov.b64 %rd16725, {%r11902, %r11901}; + { + .reg .b32 %dummy; + mov.b64 {%r11903,%dummy}, %rd16689; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11904}, %rd16689; + } + shf.r.wrap.b32 %r11905, %r11904, %r11903, 9; + shf.r.wrap.b32 %r11906, %r11903, %r11904, 9; + mov.b64 %rd16726, {%r11906, %r11905}; + { + .reg .b32 %dummy; + mov.b64 {%r11907,%dummy}, %rd16695; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11908}, %rd16695; + } + shf.r.wrap.b32 %r11909, %r11908, %r11907, 25; + shf.r.wrap.b32 %r11910, %r11907, %r11908, 25; + mov.b64 %rd16727, {%r11910, %r11909}; + { + .reg .b32 %dummy; + mov.b64 {%r11911,%dummy}, %rd16696; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11912}, %rd16696; + } + shf.r.wrap.b32 %r11913, %r11912, %r11911, 23; + shf.r.wrap.b32 %r11914, %r11911, %r11912, 23; + mov.b64 %rd16728, {%r11914, %r11913}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11915}, %rd16702; + } + { + .reg .b32 %dummy; + mov.b64 {%r11916,%dummy}, %rd16702; + } + shf.l.wrap.b32 %r11917, %r11916, %r11915, 2; + shf.l.wrap.b32 %r11918, %r11915, %r11916, 2; + mov.b64 %rd16729, {%r11918, %r11917}; + not.b64 %rd16730, %rd16706; + and.b64 %rd16731, %rd16707, %rd16730; + xor.b64 %rd16732, %rd16731, %rd16681; + not.b64 %rd16733, %rd16707; + and.b64 %rd16734, %rd16708, %rd16733; + xor.b64 %rd16735, %rd16734, %rd16706; + not.b64 %rd16736, %rd16708; + and.b64 %rd16737, %rd16709, %rd16736; + xor.b64 %rd16738, %rd16737, %rd16707; + not.b64 %rd16739, %rd16709; + and.b64 %rd16740, %rd16681, %rd16739; + xor.b64 %rd16741, %rd16740, %rd16708; + not.b64 %rd16742, %rd16681; + and.b64 %rd16743, %rd16706, %rd16742; + xor.b64 %rd16744, %rd16743, %rd16709; + not.b64 %rd16745, %rd16711; + and.b64 %rd16746, %rd16712, %rd16745; + xor.b64 %rd16747, %rd16746, %rd16710; + not.b64 %rd16748, %rd16712; + and.b64 %rd16749, %rd16713, %rd16748; + xor.b64 %rd16750, %rd16749, %rd16711; + not.b64 %rd16751, %rd16713; + and.b64 %rd16752, %rd16714, %rd16751; + xor.b64 %rd16753, %rd16752, %rd16712; + not.b64 %rd16754, %rd16714; + and.b64 %rd16755, %rd16710, %rd16754; + xor.b64 %rd16756, %rd16755, %rd16713; + not.b64 %rd16757, %rd16710; + and.b64 %rd16758, %rd16711, %rd16757; + xor.b64 %rd16759, %rd16758, %rd16714; + not.b64 %rd16760, %rd16716; + and.b64 %rd16761, %rd16717, %rd16760; + xor.b64 %rd16762, %rd16761, %rd16715; + not.b64 %rd16763, %rd16717; + and.b64 %rd16764, %rd16718, %rd16763; + xor.b64 %rd16765, %rd16764, %rd16716; + not.b64 %rd16766, %rd16718; + and.b64 %rd16767, %rd16719, %rd16766; + xor.b64 %rd16768, %rd16767, %rd16717; + not.b64 %rd16769, %rd16719; + and.b64 %rd16770, %rd16715, %rd16769; + xor.b64 %rd16771, %rd16770, %rd16718; + not.b64 %rd16772, %rd16715; + and.b64 %rd16773, %rd16716, %rd16772; + xor.b64 %rd16774, %rd16773, %rd16719; + not.b64 %rd16775, %rd16721; + and.b64 %rd16776, %rd16722, %rd16775; + xor.b64 %rd16777, %rd16776, %rd16720; + not.b64 %rd16778, %rd16722; + and.b64 %rd16779, %rd16723, %rd16778; + xor.b64 %rd16780, %rd16779, %rd16721; + not.b64 %rd16781, %rd16723; + and.b64 %rd16782, %rd16724, %rd16781; + xor.b64 %rd16783, %rd16782, %rd16722; + not.b64 %rd16784, %rd16724; + and.b64 %rd16785, %rd16720, %rd16784; + xor.b64 %rd16786, %rd16785, %rd16723; + not.b64 %rd16787, %rd16720; + and.b64 %rd16788, %rd16721, %rd16787; + xor.b64 %rd16789, %rd16788, %rd16724; + not.b64 %rd16790, %rd16726; + and.b64 %rd16791, %rd16727, %rd16790; + xor.b64 %rd16792, %rd16791, %rd16725; + not.b64 %rd16793, %rd16727; + and.b64 %rd16794, %rd16728, %rd16793; + xor.b64 %rd16795, %rd16794, %rd16726; + not.b64 %rd16796, %rd16728; + and.b64 %rd16797, %rd16729, %rd16796; + xor.b64 %rd16798, %rd16797, %rd16727; + not.b64 %rd16799, %rd16729; + and.b64 %rd16800, %rd16725, %rd16799; + xor.b64 %rd16801, %rd16800, %rd16728; + not.b64 %rd16802, %rd16725; + and.b64 %rd16803, %rd16726, %rd16802; + xor.b64 %rd16804, %rd16803, %rd16729; + xor.b64 %rd16805, %rd16732, %rd19813; + xor.b64 %rd16806, %rd16747, %rd16805; + xor.b64 %rd16807, %rd16806, %rd16762; + xor.b64 %rd16808, %rd16807, %rd16777; + xor.b64 %rd16809, %rd16808, %rd16792; + xor.b64 %rd16810, %rd16750, %rd16735; + xor.b64 %rd16811, %rd16810, %rd16765; + xor.b64 %rd16812, %rd16811, %rd16780; + xor.b64 %rd16813, %rd16812, %rd16795; + xor.b64 %rd16814, %rd16753, %rd16738; + xor.b64 %rd16815, %rd16814, %rd16768; + xor.b64 %rd16816, %rd16815, %rd16783; + xor.b64 %rd16817, %rd16816, %rd16798; + xor.b64 %rd16818, %rd16756, %rd16741; + xor.b64 %rd16819, %rd16818, %rd16771; + xor.b64 %rd16820, %rd16819, %rd16786; + xor.b64 %rd16821, %rd16820, %rd16801; + xor.b64 %rd16822, %rd16759, %rd16744; + xor.b64 %rd16823, %rd16822, %rd16774; + xor.b64 %rd16824, %rd16823, %rd16789; + xor.b64 %rd16825, %rd16824, %rd16804; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11919}, %rd16813; + } + { + .reg .b32 %dummy; + mov.b64 {%r11920,%dummy}, %rd16813; + } + shf.l.wrap.b32 %r11921, %r11920, %r11919, 1; + shf.l.wrap.b32 %r11922, %r11919, %r11920, 1; + mov.b64 %rd16826, {%r11922, %r11921}; + xor.b64 %rd16827, %rd16825, %rd16826; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11923}, %rd16817; + } + { + .reg .b32 %dummy; + mov.b64 {%r11924,%dummy}, %rd16817; + } + shf.l.wrap.b32 %r11925, %r11924, %r11923, 1; + shf.l.wrap.b32 %r11926, %r11923, %r11924, 1; + mov.b64 %rd16828, {%r11926, %r11925}; + xor.b64 %rd16829, %rd16828, %rd16809; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11927}, %rd16821; + } + { + .reg .b32 %dummy; + mov.b64 {%r11928,%dummy}, %rd16821; + } + shf.l.wrap.b32 %r11929, %r11928, %r11927, 1; + shf.l.wrap.b32 %r11930, %r11927, %r11928, 1; + mov.b64 %rd16830, {%r11930, %r11929}; + xor.b64 %rd16831, %rd16830, %rd16813; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11931}, %rd16825; + } + { + .reg .b32 %dummy; + mov.b64 {%r11932,%dummy}, %rd16825; + } + shf.l.wrap.b32 %r11933, %r11932, %r11931, 1; + shf.l.wrap.b32 %r11934, %r11931, %r11932, 1; + mov.b64 %rd16832, {%r11934, %r11933}; + xor.b64 %rd16833, %rd16832, %rd16817; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11935}, %rd16809; + } + { + .reg .b32 %dummy; + mov.b64 {%r11936,%dummy}, %rd16809; + } + shf.l.wrap.b32 %r11937, %r11936, %r11935, 1; + shf.l.wrap.b32 %r11938, %r11935, %r11936, 1; + mov.b64 %rd16834, {%r11938, %r11937}; + xor.b64 %rd16835, %rd16821, %rd16834; + xor.b64 %rd16836, %rd16827, %rd16805; + xor.b64 %rd16837, %rd16829, %rd16735; + xor.b64 %rd16838, %rd16831, %rd16738; + xor.b64 %rd16839, %rd16833, %rd16741; + xor.b64 %rd16840, %rd16835, %rd16744; + xor.b64 %rd16841, %rd16827, %rd16747; + xor.b64 %rd16842, %rd16829, %rd16750; + xor.b64 %rd16843, %rd16831, %rd16753; + xor.b64 %rd16844, %rd16833, %rd16756; + xor.b64 %rd16845, %rd16835, %rd16759; + xor.b64 %rd16846, %rd16827, %rd16762; + xor.b64 %rd16847, %rd16829, %rd16765; + xor.b64 %rd16848, %rd16831, %rd16768; + xor.b64 %rd16849, %rd16833, %rd16771; + xor.b64 %rd16850, %rd16835, %rd16774; + xor.b64 %rd16851, %rd16827, %rd16777; + xor.b64 %rd16852, %rd16829, %rd16780; + xor.b64 %rd16853, %rd16831, %rd16783; + xor.b64 %rd16854, %rd16833, %rd16786; + xor.b64 %rd16855, %rd16835, %rd16789; + xor.b64 %rd16856, %rd16827, %rd16792; + xor.b64 %rd16857, %rd16829, %rd16795; + xor.b64 %rd16858, %rd16831, %rd16798; + xor.b64 %rd16859, %rd16833, %rd16801; + xor.b64 %rd16860, %rd16835, %rd16804; + { + .reg .b32 %dummy; + mov.b64 {%r11939,%dummy}, %rd16842; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11940}, %rd16842; + } + shf.r.wrap.b32 %r11941, %r11940, %r11939, 20; + shf.r.wrap.b32 %r11942, %r11939, %r11940, 20; + mov.b64 %rd16861, {%r11942, %r11941}; + { + .reg .b32 %dummy; + mov.b64 {%r11943,%dummy}, %rd16848; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11944}, %rd16848; + } + shf.r.wrap.b32 %r11945, %r11944, %r11943, 21; + shf.r.wrap.b32 %r11946, %r11943, %r11944, 21; + mov.b64 %rd16862, {%r11946, %r11945}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11947}, %rd16854; + } + { + .reg .b32 %dummy; + mov.b64 {%r11948,%dummy}, %rd16854; + } + shf.l.wrap.b32 %r11949, %r11948, %r11947, 21; + shf.l.wrap.b32 %r11950, %r11947, %r11948, 21; + mov.b64 %rd16863, {%r11950, %r11949}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11951}, %rd16860; + } + { + .reg .b32 %dummy; + mov.b64 {%r11952,%dummy}, %rd16860; + } + shf.l.wrap.b32 %r11953, %r11952, %r11951, 14; + shf.l.wrap.b32 %r11954, %r11951, %r11952, 14; + mov.b64 %rd16864, {%r11954, %r11953}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11955}, %rd16839; + } + { + .reg .b32 %dummy; + mov.b64 {%r11956,%dummy}, %rd16839; + } + shf.l.wrap.b32 %r11957, %r11956, %r11955, 28; + shf.l.wrap.b32 %r11958, %r11955, %r11956, 28; + mov.b64 %rd16865, {%r11958, %r11957}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11959}, %rd16845; + } + { + .reg .b32 %dummy; + mov.b64 {%r11960,%dummy}, %rd16845; + } + shf.l.wrap.b32 %r11961, %r11960, %r11959, 20; + shf.l.wrap.b32 %r11962, %r11959, %r11960, 20; + mov.b64 %rd16866, {%r11962, %r11961}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11963}, %rd16846; + } + { + .reg .b32 %dummy; + mov.b64 {%r11964,%dummy}, %rd16846; + } + shf.l.wrap.b32 %r11965, %r11964, %r11963, 3; + shf.l.wrap.b32 %r11966, %r11963, %r11964, 3; + mov.b64 %rd16867, {%r11966, %r11965}; + { + .reg .b32 %dummy; + mov.b64 {%r11967,%dummy}, %rd16852; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11968}, %rd16852; + } + shf.r.wrap.b32 %r11969, %r11968, %r11967, 19; + shf.r.wrap.b32 %r11970, %r11967, %r11968, 19; + mov.b64 %rd16868, {%r11970, %r11969}; + { + .reg .b32 %dummy; + mov.b64 {%r11971,%dummy}, %rd16858; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11972}, %rd16858; + } + shf.r.wrap.b32 %r11973, %r11972, %r11971, 3; + shf.r.wrap.b32 %r11974, %r11971, %r11972, 3; + mov.b64 %rd16869, {%r11974, %r11973}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11975}, %rd16837; + } + { + .reg .b32 %dummy; + mov.b64 {%r11976,%dummy}, %rd16837; + } + shf.l.wrap.b32 %r11977, %r11976, %r11975, 1; + shf.l.wrap.b32 %r11978, %r11975, %r11976, 1; + mov.b64 %rd16870, {%r11978, %r11977}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11979}, %rd16843; + } + { + .reg .b32 %dummy; + mov.b64 {%r11980,%dummy}, %rd16843; + } + shf.l.wrap.b32 %r11981, %r11980, %r11979, 6; + shf.l.wrap.b32 %r11982, %r11979, %r11980, 6; + mov.b64 %rd16871, {%r11982, %r11981}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11983}, %rd16849; + } + { + .reg .b32 %dummy; + mov.b64 {%r11984,%dummy}, %rd16849; + } + shf.l.wrap.b32 %r11985, %r11984, %r11983, 25; + shf.l.wrap.b32 %r11986, %r11983, %r11984, 25; + mov.b64 %rd16872, {%r11986, %r11985}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11987}, %rd16855; + } + { + .reg .b32 %dummy; + mov.b64 {%r11988,%dummy}, %rd16855; + } + shf.l.wrap.b32 %r11989, %r11988, %r11987, 8; + shf.l.wrap.b32 %r11990, %r11987, %r11988, 8; + mov.b64 %rd16873, {%r11990, %r11989}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11991}, %rd16856; + } + { + .reg .b32 %dummy; + mov.b64 {%r11992,%dummy}, %rd16856; + } + shf.l.wrap.b32 %r11993, %r11992, %r11991, 18; + shf.l.wrap.b32 %r11994, %r11991, %r11992, 18; + mov.b64 %rd16874, {%r11994, %r11993}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r11995}, %rd16840; + } + { + .reg .b32 %dummy; + mov.b64 {%r11996,%dummy}, %rd16840; + } + shf.l.wrap.b32 %r11997, %r11996, %r11995, 27; + shf.l.wrap.b32 %r11998, %r11995, %r11996, 27; + mov.b64 %rd16875, {%r11998, %r11997}; + { + .reg .b32 %dummy; + mov.b64 {%r11999,%dummy}, %rd16841; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12000}, %rd16841; + } + shf.r.wrap.b32 %r12001, %r12000, %r11999, 28; + shf.r.wrap.b32 %r12002, %r11999, %r12000, 28; + mov.b64 %rd16876, {%r12002, %r12001}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12003}, %rd16847; + } + { + .reg .b32 %dummy; + mov.b64 {%r12004,%dummy}, %rd16847; + } + shf.l.wrap.b32 %r12005, %r12004, %r12003, 10; + shf.l.wrap.b32 %r12006, %r12003, %r12004, 10; + mov.b64 %rd16877, {%r12006, %r12005}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12007}, %rd16853; + } + { + .reg .b32 %dummy; + mov.b64 {%r12008,%dummy}, %rd16853; + } + shf.l.wrap.b32 %r12009, %r12008, %r12007, 15; + shf.l.wrap.b32 %r12010, %r12007, %r12008, 15; + mov.b64 %rd16878, {%r12010, %r12009}; + { + .reg .b32 %dummy; + mov.b64 {%r12011,%dummy}, %rd16859; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12012}, %rd16859; + } + shf.r.wrap.b32 %r12013, %r12012, %r12011, 8; + shf.r.wrap.b32 %r12014, %r12011, %r12012, 8; + mov.b64 %rd16879, {%r12014, %r12013}; + { + .reg .b32 %dummy; + mov.b64 {%r12015,%dummy}, %rd16838; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12016}, %rd16838; + } + shf.r.wrap.b32 %r12017, %r12016, %r12015, 2; + shf.r.wrap.b32 %r12018, %r12015, %r12016, 2; + mov.b64 %rd16880, {%r12018, %r12017}; + { + .reg .b32 %dummy; + mov.b64 {%r12019,%dummy}, %rd16844; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12020}, %rd16844; + } + shf.r.wrap.b32 %r12021, %r12020, %r12019, 9; + shf.r.wrap.b32 %r12022, %r12019, %r12020, 9; + mov.b64 %rd16881, {%r12022, %r12021}; + { + .reg .b32 %dummy; + mov.b64 {%r12023,%dummy}, %rd16850; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12024}, %rd16850; + } + shf.r.wrap.b32 %r12025, %r12024, %r12023, 25; + shf.r.wrap.b32 %r12026, %r12023, %r12024, 25; + mov.b64 %rd16882, {%r12026, %r12025}; + { + .reg .b32 %dummy; + mov.b64 {%r12027,%dummy}, %rd16851; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12028}, %rd16851; + } + shf.r.wrap.b32 %r12029, %r12028, %r12027, 23; + shf.r.wrap.b32 %r12030, %r12027, %r12028, 23; + mov.b64 %rd16883, {%r12030, %r12029}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12031}, %rd16857; + } + { + .reg .b32 %dummy; + mov.b64 {%r12032,%dummy}, %rd16857; + } + shf.l.wrap.b32 %r12033, %r12032, %r12031, 2; + shf.l.wrap.b32 %r12034, %r12031, %r12032, 2; + mov.b64 %rd16884, {%r12034, %r12033}; + not.b64 %rd16885, %rd16861; + and.b64 %rd16886, %rd16862, %rd16885; + xor.b64 %rd16887, %rd16886, %rd16836; + not.b64 %rd16888, %rd16862; + and.b64 %rd16889, %rd16863, %rd16888; + xor.b64 %rd16890, %rd16889, %rd16861; + not.b64 %rd16891, %rd16863; + and.b64 %rd16892, %rd16864, %rd16891; + xor.b64 %rd16893, %rd16892, %rd16862; + not.b64 %rd16894, %rd16864; + and.b64 %rd16895, %rd16836, %rd16894; + xor.b64 %rd16896, %rd16895, %rd16863; + not.b64 %rd16897, %rd16836; + and.b64 %rd16898, %rd16861, %rd16897; + xor.b64 %rd16899, %rd16898, %rd16864; + not.b64 %rd16900, %rd16866; + and.b64 %rd16901, %rd16867, %rd16900; + xor.b64 %rd16902, %rd16901, %rd16865; + not.b64 %rd16903, %rd16867; + and.b64 %rd16904, %rd16868, %rd16903; + xor.b64 %rd16905, %rd16904, %rd16866; + not.b64 %rd16906, %rd16868; + and.b64 %rd16907, %rd16869, %rd16906; + xor.b64 %rd16908, %rd16907, %rd16867; + not.b64 %rd16909, %rd16869; + and.b64 %rd16910, %rd16865, %rd16909; + xor.b64 %rd16911, %rd16910, %rd16868; + not.b64 %rd16912, %rd16865; + and.b64 %rd16913, %rd16866, %rd16912; + xor.b64 %rd16914, %rd16913, %rd16869; + not.b64 %rd16915, %rd16871; + and.b64 %rd16916, %rd16872, %rd16915; + xor.b64 %rd16917, %rd16916, %rd16870; + not.b64 %rd16918, %rd16872; + and.b64 %rd16919, %rd16873, %rd16918; + xor.b64 %rd16920, %rd16919, %rd16871; + not.b64 %rd16921, %rd16873; + and.b64 %rd16922, %rd16874, %rd16921; + xor.b64 %rd16923, %rd16922, %rd16872; + not.b64 %rd16924, %rd16874; + and.b64 %rd16925, %rd16870, %rd16924; + xor.b64 %rd16926, %rd16925, %rd16873; + not.b64 %rd16927, %rd16870; + and.b64 %rd16928, %rd16871, %rd16927; + xor.b64 %rd16929, %rd16928, %rd16874; + not.b64 %rd16930, %rd16876; + and.b64 %rd16931, %rd16877, %rd16930; + xor.b64 %rd16932, %rd16931, %rd16875; + not.b64 %rd16933, %rd16877; + and.b64 %rd16934, %rd16878, %rd16933; + xor.b64 %rd16935, %rd16934, %rd16876; + not.b64 %rd16936, %rd16878; + and.b64 %rd16937, %rd16879, %rd16936; + xor.b64 %rd16938, %rd16937, %rd16877; + not.b64 %rd16939, %rd16879; + and.b64 %rd16940, %rd16875, %rd16939; + xor.b64 %rd16941, %rd16940, %rd16878; + not.b64 %rd16942, %rd16875; + and.b64 %rd16943, %rd16876, %rd16942; + xor.b64 %rd16944, %rd16943, %rd16879; + not.b64 %rd16945, %rd16881; + and.b64 %rd16946, %rd16882, %rd16945; + xor.b64 %rd16947, %rd16946, %rd16880; + not.b64 %rd16948, %rd16882; + and.b64 %rd16949, %rd16883, %rd16948; + xor.b64 %rd16950, %rd16949, %rd16881; + not.b64 %rd16951, %rd16883; + and.b64 %rd16952, %rd16884, %rd16951; + xor.b64 %rd16953, %rd16952, %rd16882; + not.b64 %rd16954, %rd16884; + and.b64 %rd16955, %rd16880, %rd16954; + xor.b64 %rd16956, %rd16955, %rd16883; + not.b64 %rd16957, %rd16880; + and.b64 %rd16958, %rd16881, %rd16957; + xor.b64 %rd16959, %rd16958, %rd16884; + xor.b64 %rd16960, %rd16887, %rd19814; + xor.b64 %rd16961, %rd16902, %rd16960; + xor.b64 %rd16962, %rd16961, %rd16917; + xor.b64 %rd16963, %rd16962, %rd16932; + xor.b64 %rd16964, %rd16963, %rd16947; + xor.b64 %rd16965, %rd16905, %rd16890; + xor.b64 %rd16966, %rd16965, %rd16920; + xor.b64 %rd16967, %rd16966, %rd16935; + xor.b64 %rd16968, %rd16967, %rd16950; + xor.b64 %rd16969, %rd16908, %rd16893; + xor.b64 %rd16970, %rd16969, %rd16923; + xor.b64 %rd16971, %rd16970, %rd16938; + xor.b64 %rd16972, %rd16971, %rd16953; + xor.b64 %rd16973, %rd16911, %rd16896; + xor.b64 %rd16974, %rd16973, %rd16926; + xor.b64 %rd16975, %rd16974, %rd16941; + xor.b64 %rd16976, %rd16975, %rd16956; + xor.b64 %rd16977, %rd16914, %rd16899; + xor.b64 %rd16978, %rd16977, %rd16929; + xor.b64 %rd16979, %rd16978, %rd16944; + xor.b64 %rd16980, %rd16979, %rd16959; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12035}, %rd16968; + } + { + .reg .b32 %dummy; + mov.b64 {%r12036,%dummy}, %rd16968; + } + shf.l.wrap.b32 %r12037, %r12036, %r12035, 1; + shf.l.wrap.b32 %r12038, %r12035, %r12036, 1; + mov.b64 %rd16981, {%r12038, %r12037}; + xor.b64 %rd16982, %rd16980, %rd16981; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12039}, %rd16972; + } + { + .reg .b32 %dummy; + mov.b64 {%r12040,%dummy}, %rd16972; + } + shf.l.wrap.b32 %r12041, %r12040, %r12039, 1; + shf.l.wrap.b32 %r12042, %r12039, %r12040, 1; + mov.b64 %rd16983, {%r12042, %r12041}; + xor.b64 %rd16984, %rd16983, %rd16964; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12043}, %rd16976; + } + { + .reg .b32 %dummy; + mov.b64 {%r12044,%dummy}, %rd16976; + } + shf.l.wrap.b32 %r12045, %r12044, %r12043, 1; + shf.l.wrap.b32 %r12046, %r12043, %r12044, 1; + mov.b64 %rd16985, {%r12046, %r12045}; + xor.b64 %rd16986, %rd16985, %rd16968; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12047}, %rd16980; + } + { + .reg .b32 %dummy; + mov.b64 {%r12048,%dummy}, %rd16980; + } + shf.l.wrap.b32 %r12049, %r12048, %r12047, 1; + shf.l.wrap.b32 %r12050, %r12047, %r12048, 1; + mov.b64 %rd16987, {%r12050, %r12049}; + xor.b64 %rd16988, %rd16987, %rd16972; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12051}, %rd16964; + } + { + .reg .b32 %dummy; + mov.b64 {%r12052,%dummy}, %rd16964; + } + shf.l.wrap.b32 %r12053, %r12052, %r12051, 1; + shf.l.wrap.b32 %r12054, %r12051, %r12052, 1; + mov.b64 %rd16989, {%r12054, %r12053}; + xor.b64 %rd16990, %rd16976, %rd16989; + xor.b64 %rd16991, %rd16982, %rd16960; + xor.b64 %rd16992, %rd16984, %rd16890; + xor.b64 %rd16993, %rd16986, %rd16893; + xor.b64 %rd16994, %rd16988, %rd16896; + xor.b64 %rd16995, %rd16990, %rd16899; + xor.b64 %rd16996, %rd16982, %rd16902; + xor.b64 %rd16997, %rd16984, %rd16905; + xor.b64 %rd16998, %rd16986, %rd16908; + xor.b64 %rd16999, %rd16988, %rd16911; + xor.b64 %rd17000, %rd16990, %rd16914; + xor.b64 %rd17001, %rd16982, %rd16917; + xor.b64 %rd17002, %rd16984, %rd16920; + xor.b64 %rd17003, %rd16986, %rd16923; + xor.b64 %rd17004, %rd16988, %rd16926; + xor.b64 %rd17005, %rd16990, %rd16929; + xor.b64 %rd17006, %rd16982, %rd16932; + xor.b64 %rd17007, %rd16984, %rd16935; + xor.b64 %rd17008, %rd16986, %rd16938; + xor.b64 %rd17009, %rd16988, %rd16941; + xor.b64 %rd17010, %rd16990, %rd16944; + xor.b64 %rd17011, %rd16982, %rd16947; + xor.b64 %rd17012, %rd16984, %rd16950; + xor.b64 %rd17013, %rd16986, %rd16953; + xor.b64 %rd17014, %rd16988, %rd16956; + xor.b64 %rd17015, %rd16990, %rd16959; + { + .reg .b32 %dummy; + mov.b64 {%r12055,%dummy}, %rd16997; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12056}, %rd16997; + } + shf.r.wrap.b32 %r12057, %r12056, %r12055, 20; + shf.r.wrap.b32 %r12058, %r12055, %r12056, 20; + mov.b64 %rd17016, {%r12058, %r12057}; + { + .reg .b32 %dummy; + mov.b64 {%r12059,%dummy}, %rd17003; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12060}, %rd17003; + } + shf.r.wrap.b32 %r12061, %r12060, %r12059, 21; + shf.r.wrap.b32 %r12062, %r12059, %r12060, 21; + mov.b64 %rd17017, {%r12062, %r12061}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12063}, %rd17009; + } + { + .reg .b32 %dummy; + mov.b64 {%r12064,%dummy}, %rd17009; + } + shf.l.wrap.b32 %r12065, %r12064, %r12063, 21; + shf.l.wrap.b32 %r12066, %r12063, %r12064, 21; + mov.b64 %rd17018, {%r12066, %r12065}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12067}, %rd17015; + } + { + .reg .b32 %dummy; + mov.b64 {%r12068,%dummy}, %rd17015; + } + shf.l.wrap.b32 %r12069, %r12068, %r12067, 14; + shf.l.wrap.b32 %r12070, %r12067, %r12068, 14; + mov.b64 %rd17019, {%r12070, %r12069}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12071}, %rd16994; + } + { + .reg .b32 %dummy; + mov.b64 {%r12072,%dummy}, %rd16994; + } + shf.l.wrap.b32 %r12073, %r12072, %r12071, 28; + shf.l.wrap.b32 %r12074, %r12071, %r12072, 28; + mov.b64 %rd17020, {%r12074, %r12073}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12075}, %rd17000; + } + { + .reg .b32 %dummy; + mov.b64 {%r12076,%dummy}, %rd17000; + } + shf.l.wrap.b32 %r12077, %r12076, %r12075, 20; + shf.l.wrap.b32 %r12078, %r12075, %r12076, 20; + mov.b64 %rd17021, {%r12078, %r12077}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12079}, %rd17001; + } + { + .reg .b32 %dummy; + mov.b64 {%r12080,%dummy}, %rd17001; + } + shf.l.wrap.b32 %r12081, %r12080, %r12079, 3; + shf.l.wrap.b32 %r12082, %r12079, %r12080, 3; + mov.b64 %rd17022, {%r12082, %r12081}; + { + .reg .b32 %dummy; + mov.b64 {%r12083,%dummy}, %rd17007; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12084}, %rd17007; + } + shf.r.wrap.b32 %r12085, %r12084, %r12083, 19; + shf.r.wrap.b32 %r12086, %r12083, %r12084, 19; + mov.b64 %rd17023, {%r12086, %r12085}; + { + .reg .b32 %dummy; + mov.b64 {%r12087,%dummy}, %rd17013; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12088}, %rd17013; + } + shf.r.wrap.b32 %r12089, %r12088, %r12087, 3; + shf.r.wrap.b32 %r12090, %r12087, %r12088, 3; + mov.b64 %rd17024, {%r12090, %r12089}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12091}, %rd16992; + } + { + .reg .b32 %dummy; + mov.b64 {%r12092,%dummy}, %rd16992; + } + shf.l.wrap.b32 %r12093, %r12092, %r12091, 1; + shf.l.wrap.b32 %r12094, %r12091, %r12092, 1; + mov.b64 %rd17025, {%r12094, %r12093}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12095}, %rd16998; + } + { + .reg .b32 %dummy; + mov.b64 {%r12096,%dummy}, %rd16998; + } + shf.l.wrap.b32 %r12097, %r12096, %r12095, 6; + shf.l.wrap.b32 %r12098, %r12095, %r12096, 6; + mov.b64 %rd17026, {%r12098, %r12097}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12099}, %rd17004; + } + { + .reg .b32 %dummy; + mov.b64 {%r12100,%dummy}, %rd17004; + } + shf.l.wrap.b32 %r12101, %r12100, %r12099, 25; + shf.l.wrap.b32 %r12102, %r12099, %r12100, 25; + mov.b64 %rd17027, {%r12102, %r12101}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12103}, %rd17010; + } + { + .reg .b32 %dummy; + mov.b64 {%r12104,%dummy}, %rd17010; + } + shf.l.wrap.b32 %r12105, %r12104, %r12103, 8; + shf.l.wrap.b32 %r12106, %r12103, %r12104, 8; + mov.b64 %rd17028, {%r12106, %r12105}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12107}, %rd17011; + } + { + .reg .b32 %dummy; + mov.b64 {%r12108,%dummy}, %rd17011; + } + shf.l.wrap.b32 %r12109, %r12108, %r12107, 18; + shf.l.wrap.b32 %r12110, %r12107, %r12108, 18; + mov.b64 %rd17029, {%r12110, %r12109}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12111}, %rd16995; + } + { + .reg .b32 %dummy; + mov.b64 {%r12112,%dummy}, %rd16995; + } + shf.l.wrap.b32 %r12113, %r12112, %r12111, 27; + shf.l.wrap.b32 %r12114, %r12111, %r12112, 27; + mov.b64 %rd17030, {%r12114, %r12113}; + { + .reg .b32 %dummy; + mov.b64 {%r12115,%dummy}, %rd16996; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12116}, %rd16996; + } + shf.r.wrap.b32 %r12117, %r12116, %r12115, 28; + shf.r.wrap.b32 %r12118, %r12115, %r12116, 28; + mov.b64 %rd17031, {%r12118, %r12117}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12119}, %rd17002; + } + { + .reg .b32 %dummy; + mov.b64 {%r12120,%dummy}, %rd17002; + } + shf.l.wrap.b32 %r12121, %r12120, %r12119, 10; + shf.l.wrap.b32 %r12122, %r12119, %r12120, 10; + mov.b64 %rd17032, {%r12122, %r12121}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12123}, %rd17008; + } + { + .reg .b32 %dummy; + mov.b64 {%r12124,%dummy}, %rd17008; + } + shf.l.wrap.b32 %r12125, %r12124, %r12123, 15; + shf.l.wrap.b32 %r12126, %r12123, %r12124, 15; + mov.b64 %rd17033, {%r12126, %r12125}; + { + .reg .b32 %dummy; + mov.b64 {%r12127,%dummy}, %rd17014; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12128}, %rd17014; + } + shf.r.wrap.b32 %r12129, %r12128, %r12127, 8; + shf.r.wrap.b32 %r12130, %r12127, %r12128, 8; + mov.b64 %rd17034, {%r12130, %r12129}; + { + .reg .b32 %dummy; + mov.b64 {%r12131,%dummy}, %rd16993; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12132}, %rd16993; + } + shf.r.wrap.b32 %r12133, %r12132, %r12131, 2; + shf.r.wrap.b32 %r12134, %r12131, %r12132, 2; + mov.b64 %rd17035, {%r12134, %r12133}; + { + .reg .b32 %dummy; + mov.b64 {%r12135,%dummy}, %rd16999; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12136}, %rd16999; + } + shf.r.wrap.b32 %r12137, %r12136, %r12135, 9; + shf.r.wrap.b32 %r12138, %r12135, %r12136, 9; + mov.b64 %rd17036, {%r12138, %r12137}; + { + .reg .b32 %dummy; + mov.b64 {%r12139,%dummy}, %rd17005; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12140}, %rd17005; + } + shf.r.wrap.b32 %r12141, %r12140, %r12139, 25; + shf.r.wrap.b32 %r12142, %r12139, %r12140, 25; + mov.b64 %rd17037, {%r12142, %r12141}; + { + .reg .b32 %dummy; + mov.b64 {%r12143,%dummy}, %rd17006; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12144}, %rd17006; + } + shf.r.wrap.b32 %r12145, %r12144, %r12143, 23; + shf.r.wrap.b32 %r12146, %r12143, %r12144, 23; + mov.b64 %rd17038, {%r12146, %r12145}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12147}, %rd17012; + } + { + .reg .b32 %dummy; + mov.b64 {%r12148,%dummy}, %rd17012; + } + shf.l.wrap.b32 %r12149, %r12148, %r12147, 2; + shf.l.wrap.b32 %r12150, %r12147, %r12148, 2; + mov.b64 %rd17039, {%r12150, %r12149}; + not.b64 %rd17040, %rd17016; + and.b64 %rd17041, %rd17017, %rd17040; + xor.b64 %rd17042, %rd17041, %rd16991; + not.b64 %rd17043, %rd17017; + and.b64 %rd17044, %rd17018, %rd17043; + xor.b64 %rd17045, %rd17044, %rd17016; + not.b64 %rd17046, %rd17018; + and.b64 %rd17047, %rd17019, %rd17046; + xor.b64 %rd17048, %rd17047, %rd17017; + not.b64 %rd17049, %rd17019; + and.b64 %rd17050, %rd16991, %rd17049; + xor.b64 %rd17051, %rd17050, %rd17018; + not.b64 %rd17052, %rd16991; + and.b64 %rd17053, %rd17016, %rd17052; + xor.b64 %rd17054, %rd17053, %rd17019; + not.b64 %rd17055, %rd17021; + and.b64 %rd17056, %rd17022, %rd17055; + xor.b64 %rd17057, %rd17056, %rd17020; + not.b64 %rd17058, %rd17022; + and.b64 %rd17059, %rd17023, %rd17058; + xor.b64 %rd17060, %rd17059, %rd17021; + not.b64 %rd17061, %rd17023; + and.b64 %rd17062, %rd17024, %rd17061; + xor.b64 %rd17063, %rd17062, %rd17022; + not.b64 %rd17064, %rd17024; + and.b64 %rd17065, %rd17020, %rd17064; + xor.b64 %rd17066, %rd17065, %rd17023; + not.b64 %rd17067, %rd17020; + and.b64 %rd17068, %rd17021, %rd17067; + xor.b64 %rd17069, %rd17068, %rd17024; + not.b64 %rd17070, %rd17026; + and.b64 %rd17071, %rd17027, %rd17070; + xor.b64 %rd17072, %rd17071, %rd17025; + not.b64 %rd17073, %rd17027; + and.b64 %rd17074, %rd17028, %rd17073; + xor.b64 %rd17075, %rd17074, %rd17026; + not.b64 %rd17076, %rd17028; + and.b64 %rd17077, %rd17029, %rd17076; + xor.b64 %rd17078, %rd17077, %rd17027; + not.b64 %rd17079, %rd17029; + and.b64 %rd17080, %rd17025, %rd17079; + xor.b64 %rd17081, %rd17080, %rd17028; + not.b64 %rd17082, %rd17025; + and.b64 %rd17083, %rd17026, %rd17082; + xor.b64 %rd17084, %rd17083, %rd17029; + not.b64 %rd17085, %rd17031; + and.b64 %rd17086, %rd17032, %rd17085; + xor.b64 %rd17087, %rd17086, %rd17030; + not.b64 %rd17088, %rd17032; + and.b64 %rd17089, %rd17033, %rd17088; + xor.b64 %rd17090, %rd17089, %rd17031; + not.b64 %rd17091, %rd17033; + and.b64 %rd17092, %rd17034, %rd17091; + xor.b64 %rd17093, %rd17092, %rd17032; + not.b64 %rd17094, %rd17034; + and.b64 %rd17095, %rd17030, %rd17094; + xor.b64 %rd17096, %rd17095, %rd17033; + not.b64 %rd17097, %rd17030; + and.b64 %rd17098, %rd17031, %rd17097; + xor.b64 %rd17099, %rd17098, %rd17034; + not.b64 %rd17100, %rd17036; + and.b64 %rd17101, %rd17037, %rd17100; + xor.b64 %rd17102, %rd17101, %rd17035; + not.b64 %rd17103, %rd17037; + and.b64 %rd17104, %rd17038, %rd17103; + xor.b64 %rd17105, %rd17104, %rd17036; + not.b64 %rd17106, %rd17038; + and.b64 %rd17107, %rd17039, %rd17106; + xor.b64 %rd17108, %rd17107, %rd17037; + not.b64 %rd17109, %rd17039; + and.b64 %rd17110, %rd17035, %rd17109; + xor.b64 %rd17111, %rd17110, %rd17038; + not.b64 %rd17112, %rd17035; + and.b64 %rd17113, %rd17036, %rd17112; + xor.b64 %rd17114, %rd17113, %rd17039; + xor.b64 %rd17115, %rd17042, %rd19815; + xor.b64 %rd17116, %rd17057, %rd17115; + xor.b64 %rd17117, %rd17116, %rd17072; + xor.b64 %rd17118, %rd17117, %rd17087; + xor.b64 %rd17119, %rd17118, %rd17102; + xor.b64 %rd17120, %rd17060, %rd17045; + xor.b64 %rd17121, %rd17120, %rd17075; + xor.b64 %rd17122, %rd17121, %rd17090; + xor.b64 %rd17123, %rd17122, %rd17105; + xor.b64 %rd17124, %rd17063, %rd17048; + xor.b64 %rd17125, %rd17124, %rd17078; + xor.b64 %rd17126, %rd17125, %rd17093; + xor.b64 %rd17127, %rd17126, %rd17108; + xor.b64 %rd17128, %rd17066, %rd17051; + xor.b64 %rd17129, %rd17128, %rd17081; + xor.b64 %rd17130, %rd17129, %rd17096; + xor.b64 %rd17131, %rd17130, %rd17111; + xor.b64 %rd17132, %rd17069, %rd17054; + xor.b64 %rd17133, %rd17132, %rd17084; + xor.b64 %rd17134, %rd17133, %rd17099; + xor.b64 %rd17135, %rd17134, %rd17114; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12151}, %rd17123; + } + { + .reg .b32 %dummy; + mov.b64 {%r12152,%dummy}, %rd17123; + } + shf.l.wrap.b32 %r12153, %r12152, %r12151, 1; + shf.l.wrap.b32 %r12154, %r12151, %r12152, 1; + mov.b64 %rd17136, {%r12154, %r12153}; + xor.b64 %rd17137, %rd17135, %rd17136; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12155}, %rd17127; + } + { + .reg .b32 %dummy; + mov.b64 {%r12156,%dummy}, %rd17127; + } + shf.l.wrap.b32 %r12157, %r12156, %r12155, 1; + shf.l.wrap.b32 %r12158, %r12155, %r12156, 1; + mov.b64 %rd17138, {%r12158, %r12157}; + xor.b64 %rd17139, %rd17138, %rd17119; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12159}, %rd17131; + } + { + .reg .b32 %dummy; + mov.b64 {%r12160,%dummy}, %rd17131; + } + shf.l.wrap.b32 %r12161, %r12160, %r12159, 1; + shf.l.wrap.b32 %r12162, %r12159, %r12160, 1; + mov.b64 %rd17140, {%r12162, %r12161}; + xor.b64 %rd17141, %rd17140, %rd17123; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12163}, %rd17135; + } + { + .reg .b32 %dummy; + mov.b64 {%r12164,%dummy}, %rd17135; + } + shf.l.wrap.b32 %r12165, %r12164, %r12163, 1; + shf.l.wrap.b32 %r12166, %r12163, %r12164, 1; + mov.b64 %rd17142, {%r12166, %r12165}; + xor.b64 %rd17143, %rd17142, %rd17127; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12167}, %rd17119; + } + { + .reg .b32 %dummy; + mov.b64 {%r12168,%dummy}, %rd17119; + } + shf.l.wrap.b32 %r12169, %r12168, %r12167, 1; + shf.l.wrap.b32 %r12170, %r12167, %r12168, 1; + mov.b64 %rd17144, {%r12170, %r12169}; + xor.b64 %rd17145, %rd17131, %rd17144; + xor.b64 %rd17146, %rd17137, %rd17115; + xor.b64 %rd17147, %rd17139, %rd17045; + xor.b64 %rd17148, %rd17141, %rd17048; + xor.b64 %rd17149, %rd17143, %rd17051; + xor.b64 %rd17150, %rd17145, %rd17054; + xor.b64 %rd17151, %rd17137, %rd17057; + xor.b64 %rd17152, %rd17139, %rd17060; + xor.b64 %rd17153, %rd17141, %rd17063; + xor.b64 %rd17154, %rd17143, %rd17066; + xor.b64 %rd17155, %rd17145, %rd17069; + xor.b64 %rd17156, %rd17137, %rd17072; + xor.b64 %rd17157, %rd17139, %rd17075; + xor.b64 %rd17158, %rd17141, %rd17078; + xor.b64 %rd17159, %rd17143, %rd17081; + xor.b64 %rd17160, %rd17145, %rd17084; + xor.b64 %rd17161, %rd17137, %rd17087; + xor.b64 %rd17162, %rd17139, %rd17090; + xor.b64 %rd17163, %rd17141, %rd17093; + xor.b64 %rd17164, %rd17143, %rd17096; + xor.b64 %rd17165, %rd17145, %rd17099; + xor.b64 %rd17166, %rd17137, %rd17102; + xor.b64 %rd17167, %rd17139, %rd17105; + xor.b64 %rd17168, %rd17141, %rd17108; + xor.b64 %rd17169, %rd17143, %rd17111; + xor.b64 %rd17170, %rd17145, %rd17114; + { + .reg .b32 %dummy; + mov.b64 {%r12171,%dummy}, %rd17152; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12172}, %rd17152; + } + shf.r.wrap.b32 %r12173, %r12172, %r12171, 20; + shf.r.wrap.b32 %r12174, %r12171, %r12172, 20; + mov.b64 %rd17171, {%r12174, %r12173}; + { + .reg .b32 %dummy; + mov.b64 {%r12175,%dummy}, %rd17158; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12176}, %rd17158; + } + shf.r.wrap.b32 %r12177, %r12176, %r12175, 21; + shf.r.wrap.b32 %r12178, %r12175, %r12176, 21; + mov.b64 %rd17172, {%r12178, %r12177}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12179}, %rd17164; + } + { + .reg .b32 %dummy; + mov.b64 {%r12180,%dummy}, %rd17164; + } + shf.l.wrap.b32 %r12181, %r12180, %r12179, 21; + shf.l.wrap.b32 %r12182, %r12179, %r12180, 21; + mov.b64 %rd17173, {%r12182, %r12181}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12183}, %rd17170; + } + { + .reg .b32 %dummy; + mov.b64 {%r12184,%dummy}, %rd17170; + } + shf.l.wrap.b32 %r12185, %r12184, %r12183, 14; + shf.l.wrap.b32 %r12186, %r12183, %r12184, 14; + mov.b64 %rd17174, {%r12186, %r12185}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12187}, %rd17149; + } + { + .reg .b32 %dummy; + mov.b64 {%r12188,%dummy}, %rd17149; + } + shf.l.wrap.b32 %r12189, %r12188, %r12187, 28; + shf.l.wrap.b32 %r12190, %r12187, %r12188, 28; + mov.b64 %rd17175, {%r12190, %r12189}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12191}, %rd17155; + } + { + .reg .b32 %dummy; + mov.b64 {%r12192,%dummy}, %rd17155; + } + shf.l.wrap.b32 %r12193, %r12192, %r12191, 20; + shf.l.wrap.b32 %r12194, %r12191, %r12192, 20; + mov.b64 %rd17176, {%r12194, %r12193}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12195}, %rd17156; + } + { + .reg .b32 %dummy; + mov.b64 {%r12196,%dummy}, %rd17156; + } + shf.l.wrap.b32 %r12197, %r12196, %r12195, 3; + shf.l.wrap.b32 %r12198, %r12195, %r12196, 3; + mov.b64 %rd17177, {%r12198, %r12197}; + { + .reg .b32 %dummy; + mov.b64 {%r12199,%dummy}, %rd17162; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12200}, %rd17162; + } + shf.r.wrap.b32 %r12201, %r12200, %r12199, 19; + shf.r.wrap.b32 %r12202, %r12199, %r12200, 19; + mov.b64 %rd17178, {%r12202, %r12201}; + { + .reg .b32 %dummy; + mov.b64 {%r12203,%dummy}, %rd17168; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12204}, %rd17168; + } + shf.r.wrap.b32 %r12205, %r12204, %r12203, 3; + shf.r.wrap.b32 %r12206, %r12203, %r12204, 3; + mov.b64 %rd17179, {%r12206, %r12205}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12207}, %rd17147; + } + { + .reg .b32 %dummy; + mov.b64 {%r12208,%dummy}, %rd17147; + } + shf.l.wrap.b32 %r12209, %r12208, %r12207, 1; + shf.l.wrap.b32 %r12210, %r12207, %r12208, 1; + mov.b64 %rd17180, {%r12210, %r12209}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12211}, %rd17153; + } + { + .reg .b32 %dummy; + mov.b64 {%r12212,%dummy}, %rd17153; + } + shf.l.wrap.b32 %r12213, %r12212, %r12211, 6; + shf.l.wrap.b32 %r12214, %r12211, %r12212, 6; + mov.b64 %rd17181, {%r12214, %r12213}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12215}, %rd17159; + } + { + .reg .b32 %dummy; + mov.b64 {%r12216,%dummy}, %rd17159; + } + shf.l.wrap.b32 %r12217, %r12216, %r12215, 25; + shf.l.wrap.b32 %r12218, %r12215, %r12216, 25; + mov.b64 %rd17182, {%r12218, %r12217}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12219}, %rd17165; + } + { + .reg .b32 %dummy; + mov.b64 {%r12220,%dummy}, %rd17165; + } + shf.l.wrap.b32 %r12221, %r12220, %r12219, 8; + shf.l.wrap.b32 %r12222, %r12219, %r12220, 8; + mov.b64 %rd17183, {%r12222, %r12221}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12223}, %rd17166; + } + { + .reg .b32 %dummy; + mov.b64 {%r12224,%dummy}, %rd17166; + } + shf.l.wrap.b32 %r12225, %r12224, %r12223, 18; + shf.l.wrap.b32 %r12226, %r12223, %r12224, 18; + mov.b64 %rd17184, {%r12226, %r12225}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12227}, %rd17150; + } + { + .reg .b32 %dummy; + mov.b64 {%r12228,%dummy}, %rd17150; + } + shf.l.wrap.b32 %r12229, %r12228, %r12227, 27; + shf.l.wrap.b32 %r12230, %r12227, %r12228, 27; + mov.b64 %rd17185, {%r12230, %r12229}; + { + .reg .b32 %dummy; + mov.b64 {%r12231,%dummy}, %rd17151; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12232}, %rd17151; + } + shf.r.wrap.b32 %r12233, %r12232, %r12231, 28; + shf.r.wrap.b32 %r12234, %r12231, %r12232, 28; + mov.b64 %rd17186, {%r12234, %r12233}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12235}, %rd17157; + } + { + .reg .b32 %dummy; + mov.b64 {%r12236,%dummy}, %rd17157; + } + shf.l.wrap.b32 %r12237, %r12236, %r12235, 10; + shf.l.wrap.b32 %r12238, %r12235, %r12236, 10; + mov.b64 %rd17187, {%r12238, %r12237}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12239}, %rd17163; + } + { + .reg .b32 %dummy; + mov.b64 {%r12240,%dummy}, %rd17163; + } + shf.l.wrap.b32 %r12241, %r12240, %r12239, 15; + shf.l.wrap.b32 %r12242, %r12239, %r12240, 15; + mov.b64 %rd17188, {%r12242, %r12241}; + { + .reg .b32 %dummy; + mov.b64 {%r12243,%dummy}, %rd17169; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12244}, %rd17169; + } + shf.r.wrap.b32 %r12245, %r12244, %r12243, 8; + shf.r.wrap.b32 %r12246, %r12243, %r12244, 8; + mov.b64 %rd17189, {%r12246, %r12245}; + { + .reg .b32 %dummy; + mov.b64 {%r12247,%dummy}, %rd17148; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12248}, %rd17148; + } + shf.r.wrap.b32 %r12249, %r12248, %r12247, 2; + shf.r.wrap.b32 %r12250, %r12247, %r12248, 2; + mov.b64 %rd17190, {%r12250, %r12249}; + { + .reg .b32 %dummy; + mov.b64 {%r12251,%dummy}, %rd17154; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12252}, %rd17154; + } + shf.r.wrap.b32 %r12253, %r12252, %r12251, 9; + shf.r.wrap.b32 %r12254, %r12251, %r12252, 9; + mov.b64 %rd17191, {%r12254, %r12253}; + { + .reg .b32 %dummy; + mov.b64 {%r12255,%dummy}, %rd17160; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12256}, %rd17160; + } + shf.r.wrap.b32 %r12257, %r12256, %r12255, 25; + shf.r.wrap.b32 %r12258, %r12255, %r12256, 25; + mov.b64 %rd17192, {%r12258, %r12257}; + { + .reg .b32 %dummy; + mov.b64 {%r12259,%dummy}, %rd17161; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12260}, %rd17161; + } + shf.r.wrap.b32 %r12261, %r12260, %r12259, 23; + shf.r.wrap.b32 %r12262, %r12259, %r12260, 23; + mov.b64 %rd17193, {%r12262, %r12261}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12263}, %rd17167; + } + { + .reg .b32 %dummy; + mov.b64 {%r12264,%dummy}, %rd17167; + } + shf.l.wrap.b32 %r12265, %r12264, %r12263, 2; + shf.l.wrap.b32 %r12266, %r12263, %r12264, 2; + mov.b64 %rd17194, {%r12266, %r12265}; + not.b64 %rd17195, %rd17171; + and.b64 %rd17196, %rd17172, %rd17195; + xor.b64 %rd17197, %rd17196, %rd17146; + not.b64 %rd17198, %rd17172; + and.b64 %rd17199, %rd17173, %rd17198; + xor.b64 %rd17200, %rd17199, %rd17171; + not.b64 %rd17201, %rd17173; + and.b64 %rd17202, %rd17174, %rd17201; + xor.b64 %rd17203, %rd17202, %rd17172; + not.b64 %rd17204, %rd17174; + and.b64 %rd17205, %rd17146, %rd17204; + xor.b64 %rd17206, %rd17205, %rd17173; + not.b64 %rd17207, %rd17146; + and.b64 %rd17208, %rd17171, %rd17207; + xor.b64 %rd17209, %rd17208, %rd17174; + not.b64 %rd17210, %rd17176; + and.b64 %rd17211, %rd17177, %rd17210; + xor.b64 %rd17212, %rd17211, %rd17175; + not.b64 %rd17213, %rd17177; + and.b64 %rd17214, %rd17178, %rd17213; + xor.b64 %rd17215, %rd17214, %rd17176; + not.b64 %rd17216, %rd17178; + and.b64 %rd17217, %rd17179, %rd17216; + xor.b64 %rd17218, %rd17217, %rd17177; + not.b64 %rd17219, %rd17179; + and.b64 %rd17220, %rd17175, %rd17219; + xor.b64 %rd17221, %rd17220, %rd17178; + not.b64 %rd17222, %rd17175; + and.b64 %rd17223, %rd17176, %rd17222; + xor.b64 %rd17224, %rd17223, %rd17179; + not.b64 %rd17225, %rd17181; + and.b64 %rd17226, %rd17182, %rd17225; + xor.b64 %rd17227, %rd17226, %rd17180; + not.b64 %rd17228, %rd17182; + and.b64 %rd17229, %rd17183, %rd17228; + xor.b64 %rd17230, %rd17229, %rd17181; + not.b64 %rd17231, %rd17183; + and.b64 %rd17232, %rd17184, %rd17231; + xor.b64 %rd17233, %rd17232, %rd17182; + not.b64 %rd17234, %rd17184; + and.b64 %rd17235, %rd17180, %rd17234; + xor.b64 %rd17236, %rd17235, %rd17183; + not.b64 %rd17237, %rd17180; + and.b64 %rd17238, %rd17181, %rd17237; + xor.b64 %rd17239, %rd17238, %rd17184; + not.b64 %rd17240, %rd17186; + and.b64 %rd17241, %rd17187, %rd17240; + xor.b64 %rd17242, %rd17241, %rd17185; + not.b64 %rd17243, %rd17187; + and.b64 %rd17244, %rd17188, %rd17243; + xor.b64 %rd17245, %rd17244, %rd17186; + not.b64 %rd17246, %rd17188; + and.b64 %rd17247, %rd17189, %rd17246; + xor.b64 %rd17248, %rd17247, %rd17187; + not.b64 %rd17249, %rd17189; + and.b64 %rd17250, %rd17185, %rd17249; + xor.b64 %rd17251, %rd17250, %rd17188; + not.b64 %rd17252, %rd17185; + and.b64 %rd17253, %rd17186, %rd17252; + xor.b64 %rd17254, %rd17253, %rd17189; + not.b64 %rd17255, %rd17191; + and.b64 %rd17256, %rd17192, %rd17255; + xor.b64 %rd17257, %rd17256, %rd17190; + not.b64 %rd17258, %rd17192; + and.b64 %rd17259, %rd17193, %rd17258; + xor.b64 %rd17260, %rd17259, %rd17191; + not.b64 %rd17261, %rd17193; + and.b64 %rd17262, %rd17194, %rd17261; + xor.b64 %rd17263, %rd17262, %rd17192; + not.b64 %rd17264, %rd17194; + and.b64 %rd17265, %rd17190, %rd17264; + xor.b64 %rd17266, %rd17265, %rd17193; + not.b64 %rd17267, %rd17190; + and.b64 %rd17268, %rd17191, %rd17267; + xor.b64 %rd17269, %rd17268, %rd17194; + xor.b64 %rd17270, %rd17197, %rd19816; + xor.b64 %rd17271, %rd17212, %rd17270; + xor.b64 %rd17272, %rd17271, %rd17227; + xor.b64 %rd17273, %rd17272, %rd17242; + xor.b64 %rd17274, %rd17273, %rd17257; + xor.b64 %rd17275, %rd17215, %rd17200; + xor.b64 %rd17276, %rd17275, %rd17230; + xor.b64 %rd17277, %rd17276, %rd17245; + xor.b64 %rd17278, %rd17277, %rd17260; + xor.b64 %rd17279, %rd17218, %rd17203; + xor.b64 %rd17280, %rd17279, %rd17233; + xor.b64 %rd17281, %rd17280, %rd17248; + xor.b64 %rd17282, %rd17281, %rd17263; + xor.b64 %rd17283, %rd17221, %rd17206; + xor.b64 %rd17284, %rd17283, %rd17236; + xor.b64 %rd17285, %rd17284, %rd17251; + xor.b64 %rd17286, %rd17285, %rd17266; + xor.b64 %rd17287, %rd17224, %rd17209; + xor.b64 %rd17288, %rd17287, %rd17239; + xor.b64 %rd17289, %rd17288, %rd17254; + xor.b64 %rd17290, %rd17289, %rd17269; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12267}, %rd17278; + } + { + .reg .b32 %dummy; + mov.b64 {%r12268,%dummy}, %rd17278; + } + shf.l.wrap.b32 %r12269, %r12268, %r12267, 1; + shf.l.wrap.b32 %r12270, %r12267, %r12268, 1; + mov.b64 %rd17291, {%r12270, %r12269}; + xor.b64 %rd17292, %rd17290, %rd17291; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12271}, %rd17282; + } + { + .reg .b32 %dummy; + mov.b64 {%r12272,%dummy}, %rd17282; + } + shf.l.wrap.b32 %r12273, %r12272, %r12271, 1; + shf.l.wrap.b32 %r12274, %r12271, %r12272, 1; + mov.b64 %rd17293, {%r12274, %r12273}; + xor.b64 %rd17294, %rd17293, %rd17274; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12275}, %rd17286; + } + { + .reg .b32 %dummy; + mov.b64 {%r12276,%dummy}, %rd17286; + } + shf.l.wrap.b32 %r12277, %r12276, %r12275, 1; + shf.l.wrap.b32 %r12278, %r12275, %r12276, 1; + mov.b64 %rd17295, {%r12278, %r12277}; + xor.b64 %rd17296, %rd17295, %rd17278; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12279}, %rd17290; + } + { + .reg .b32 %dummy; + mov.b64 {%r12280,%dummy}, %rd17290; + } + shf.l.wrap.b32 %r12281, %r12280, %r12279, 1; + shf.l.wrap.b32 %r12282, %r12279, %r12280, 1; + mov.b64 %rd17297, {%r12282, %r12281}; + xor.b64 %rd17298, %rd17297, %rd17282; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12283}, %rd17274; + } + { + .reg .b32 %dummy; + mov.b64 {%r12284,%dummy}, %rd17274; + } + shf.l.wrap.b32 %r12285, %r12284, %r12283, 1; + shf.l.wrap.b32 %r12286, %r12283, %r12284, 1; + mov.b64 %rd17299, {%r12286, %r12285}; + xor.b64 %rd17300, %rd17286, %rd17299; + xor.b64 %rd17301, %rd17292, %rd17270; + xor.b64 %rd17302, %rd17294, %rd17200; + xor.b64 %rd17303, %rd17296, %rd17203; + xor.b64 %rd17304, %rd17298, %rd17206; + xor.b64 %rd17305, %rd17300, %rd17209; + xor.b64 %rd17306, %rd17292, %rd17212; + xor.b64 %rd17307, %rd17294, %rd17215; + xor.b64 %rd17308, %rd17296, %rd17218; + xor.b64 %rd17309, %rd17298, %rd17221; + xor.b64 %rd17310, %rd17300, %rd17224; + xor.b64 %rd17311, %rd17292, %rd17227; + xor.b64 %rd17312, %rd17294, %rd17230; + xor.b64 %rd17313, %rd17296, %rd17233; + xor.b64 %rd17314, %rd17298, %rd17236; + xor.b64 %rd17315, %rd17300, %rd17239; + xor.b64 %rd17316, %rd17292, %rd17242; + xor.b64 %rd17317, %rd17294, %rd17245; + xor.b64 %rd17318, %rd17296, %rd17248; + xor.b64 %rd17319, %rd17298, %rd17251; + xor.b64 %rd17320, %rd17300, %rd17254; + xor.b64 %rd17321, %rd17292, %rd17257; + xor.b64 %rd17322, %rd17294, %rd17260; + xor.b64 %rd17323, %rd17296, %rd17263; + xor.b64 %rd17324, %rd17298, %rd17266; + xor.b64 %rd17325, %rd17300, %rd17269; + { + .reg .b32 %dummy; + mov.b64 {%r12287,%dummy}, %rd17307; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12288}, %rd17307; + } + shf.r.wrap.b32 %r12289, %r12288, %r12287, 20; + shf.r.wrap.b32 %r12290, %r12287, %r12288, 20; + mov.b64 %rd17326, {%r12290, %r12289}; + { + .reg .b32 %dummy; + mov.b64 {%r12291,%dummy}, %rd17313; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12292}, %rd17313; + } + shf.r.wrap.b32 %r12293, %r12292, %r12291, 21; + shf.r.wrap.b32 %r12294, %r12291, %r12292, 21; + mov.b64 %rd17327, {%r12294, %r12293}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12295}, %rd17319; + } + { + .reg .b32 %dummy; + mov.b64 {%r12296,%dummy}, %rd17319; + } + shf.l.wrap.b32 %r12297, %r12296, %r12295, 21; + shf.l.wrap.b32 %r12298, %r12295, %r12296, 21; + mov.b64 %rd17328, {%r12298, %r12297}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12299}, %rd17325; + } + { + .reg .b32 %dummy; + mov.b64 {%r12300,%dummy}, %rd17325; + } + shf.l.wrap.b32 %r12301, %r12300, %r12299, 14; + shf.l.wrap.b32 %r12302, %r12299, %r12300, 14; + mov.b64 %rd17329, {%r12302, %r12301}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12303}, %rd17304; + } + { + .reg .b32 %dummy; + mov.b64 {%r12304,%dummy}, %rd17304; + } + shf.l.wrap.b32 %r12305, %r12304, %r12303, 28; + shf.l.wrap.b32 %r12306, %r12303, %r12304, 28; + mov.b64 %rd17330, {%r12306, %r12305}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12307}, %rd17310; + } + { + .reg .b32 %dummy; + mov.b64 {%r12308,%dummy}, %rd17310; + } + shf.l.wrap.b32 %r12309, %r12308, %r12307, 20; + shf.l.wrap.b32 %r12310, %r12307, %r12308, 20; + mov.b64 %rd17331, {%r12310, %r12309}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12311}, %rd17311; + } + { + .reg .b32 %dummy; + mov.b64 {%r12312,%dummy}, %rd17311; + } + shf.l.wrap.b32 %r12313, %r12312, %r12311, 3; + shf.l.wrap.b32 %r12314, %r12311, %r12312, 3; + mov.b64 %rd17332, {%r12314, %r12313}; + { + .reg .b32 %dummy; + mov.b64 {%r12315,%dummy}, %rd17317; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12316}, %rd17317; + } + shf.r.wrap.b32 %r12317, %r12316, %r12315, 19; + shf.r.wrap.b32 %r12318, %r12315, %r12316, 19; + mov.b64 %rd17333, {%r12318, %r12317}; + { + .reg .b32 %dummy; + mov.b64 {%r12319,%dummy}, %rd17323; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12320}, %rd17323; + } + shf.r.wrap.b32 %r12321, %r12320, %r12319, 3; + shf.r.wrap.b32 %r12322, %r12319, %r12320, 3; + mov.b64 %rd17334, {%r12322, %r12321}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12323}, %rd17302; + } + { + .reg .b32 %dummy; + mov.b64 {%r12324,%dummy}, %rd17302; + } + shf.l.wrap.b32 %r12325, %r12324, %r12323, 1; + shf.l.wrap.b32 %r12326, %r12323, %r12324, 1; + mov.b64 %rd17335, {%r12326, %r12325}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12327}, %rd17308; + } + { + .reg .b32 %dummy; + mov.b64 {%r12328,%dummy}, %rd17308; + } + shf.l.wrap.b32 %r12329, %r12328, %r12327, 6; + shf.l.wrap.b32 %r12330, %r12327, %r12328, 6; + mov.b64 %rd17336, {%r12330, %r12329}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12331}, %rd17314; + } + { + .reg .b32 %dummy; + mov.b64 {%r12332,%dummy}, %rd17314; + } + shf.l.wrap.b32 %r12333, %r12332, %r12331, 25; + shf.l.wrap.b32 %r12334, %r12331, %r12332, 25; + mov.b64 %rd17337, {%r12334, %r12333}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12335}, %rd17320; + } + { + .reg .b32 %dummy; + mov.b64 {%r12336,%dummy}, %rd17320; + } + shf.l.wrap.b32 %r12337, %r12336, %r12335, 8; + shf.l.wrap.b32 %r12338, %r12335, %r12336, 8; + mov.b64 %rd17338, {%r12338, %r12337}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12339}, %rd17321; + } + { + .reg .b32 %dummy; + mov.b64 {%r12340,%dummy}, %rd17321; + } + shf.l.wrap.b32 %r12341, %r12340, %r12339, 18; + shf.l.wrap.b32 %r12342, %r12339, %r12340, 18; + mov.b64 %rd17339, {%r12342, %r12341}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12343}, %rd17305; + } + { + .reg .b32 %dummy; + mov.b64 {%r12344,%dummy}, %rd17305; + } + shf.l.wrap.b32 %r12345, %r12344, %r12343, 27; + shf.l.wrap.b32 %r12346, %r12343, %r12344, 27; + mov.b64 %rd17340, {%r12346, %r12345}; + { + .reg .b32 %dummy; + mov.b64 {%r12347,%dummy}, %rd17306; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12348}, %rd17306; + } + shf.r.wrap.b32 %r12349, %r12348, %r12347, 28; + shf.r.wrap.b32 %r12350, %r12347, %r12348, 28; + mov.b64 %rd17341, {%r12350, %r12349}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12351}, %rd17312; + } + { + .reg .b32 %dummy; + mov.b64 {%r12352,%dummy}, %rd17312; + } + shf.l.wrap.b32 %r12353, %r12352, %r12351, 10; + shf.l.wrap.b32 %r12354, %r12351, %r12352, 10; + mov.b64 %rd17342, {%r12354, %r12353}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12355}, %rd17318; + } + { + .reg .b32 %dummy; + mov.b64 {%r12356,%dummy}, %rd17318; + } + shf.l.wrap.b32 %r12357, %r12356, %r12355, 15; + shf.l.wrap.b32 %r12358, %r12355, %r12356, 15; + mov.b64 %rd17343, {%r12358, %r12357}; + { + .reg .b32 %dummy; + mov.b64 {%r12359,%dummy}, %rd17324; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12360}, %rd17324; + } + shf.r.wrap.b32 %r12361, %r12360, %r12359, 8; + shf.r.wrap.b32 %r12362, %r12359, %r12360, 8; + mov.b64 %rd17344, {%r12362, %r12361}; + { + .reg .b32 %dummy; + mov.b64 {%r12363,%dummy}, %rd17303; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12364}, %rd17303; + } + shf.r.wrap.b32 %r12365, %r12364, %r12363, 2; + shf.r.wrap.b32 %r12366, %r12363, %r12364, 2; + mov.b64 %rd17345, {%r12366, %r12365}; + { + .reg .b32 %dummy; + mov.b64 {%r12367,%dummy}, %rd17309; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12368}, %rd17309; + } + shf.r.wrap.b32 %r12369, %r12368, %r12367, 9; + shf.r.wrap.b32 %r12370, %r12367, %r12368, 9; + mov.b64 %rd17346, {%r12370, %r12369}; + { + .reg .b32 %dummy; + mov.b64 {%r12371,%dummy}, %rd17315; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12372}, %rd17315; + } + shf.r.wrap.b32 %r12373, %r12372, %r12371, 25; + shf.r.wrap.b32 %r12374, %r12371, %r12372, 25; + mov.b64 %rd17347, {%r12374, %r12373}; + { + .reg .b32 %dummy; + mov.b64 {%r12375,%dummy}, %rd17316; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12376}, %rd17316; + } + shf.r.wrap.b32 %r12377, %r12376, %r12375, 23; + shf.r.wrap.b32 %r12378, %r12375, %r12376, 23; + mov.b64 %rd17348, {%r12378, %r12377}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12379}, %rd17322; + } + { + .reg .b32 %dummy; + mov.b64 {%r12380,%dummy}, %rd17322; + } + shf.l.wrap.b32 %r12381, %r12380, %r12379, 2; + shf.l.wrap.b32 %r12382, %r12379, %r12380, 2; + mov.b64 %rd17349, {%r12382, %r12381}; + not.b64 %rd17350, %rd17326; + and.b64 %rd17351, %rd17327, %rd17350; + xor.b64 %rd17352, %rd17351, %rd17301; + not.b64 %rd17353, %rd17327; + and.b64 %rd17354, %rd17328, %rd17353; + xor.b64 %rd17355, %rd17354, %rd17326; + not.b64 %rd17356, %rd17328; + and.b64 %rd17357, %rd17329, %rd17356; + xor.b64 %rd17358, %rd17357, %rd17327; + not.b64 %rd17359, %rd17329; + and.b64 %rd17360, %rd17301, %rd17359; + xor.b64 %rd17361, %rd17360, %rd17328; + not.b64 %rd17362, %rd17301; + and.b64 %rd17363, %rd17326, %rd17362; + xor.b64 %rd17364, %rd17363, %rd17329; + not.b64 %rd17365, %rd17331; + and.b64 %rd17366, %rd17332, %rd17365; + xor.b64 %rd17367, %rd17366, %rd17330; + not.b64 %rd17368, %rd17332; + and.b64 %rd17369, %rd17333, %rd17368; + xor.b64 %rd17370, %rd17369, %rd17331; + not.b64 %rd17371, %rd17333; + and.b64 %rd17372, %rd17334, %rd17371; + xor.b64 %rd17373, %rd17372, %rd17332; + not.b64 %rd17374, %rd17334; + and.b64 %rd17375, %rd17330, %rd17374; + xor.b64 %rd17376, %rd17375, %rd17333; + not.b64 %rd17377, %rd17330; + and.b64 %rd17378, %rd17331, %rd17377; + xor.b64 %rd17379, %rd17378, %rd17334; + not.b64 %rd17380, %rd17336; + and.b64 %rd17381, %rd17337, %rd17380; + xor.b64 %rd17382, %rd17381, %rd17335; + not.b64 %rd17383, %rd17337; + and.b64 %rd17384, %rd17338, %rd17383; + xor.b64 %rd17385, %rd17384, %rd17336; + not.b64 %rd17386, %rd17338; + and.b64 %rd17387, %rd17339, %rd17386; + xor.b64 %rd17388, %rd17387, %rd17337; + not.b64 %rd17389, %rd17339; + and.b64 %rd17390, %rd17335, %rd17389; + xor.b64 %rd17391, %rd17390, %rd17338; + not.b64 %rd17392, %rd17335; + and.b64 %rd17393, %rd17336, %rd17392; + xor.b64 %rd17394, %rd17393, %rd17339; + not.b64 %rd17395, %rd17341; + and.b64 %rd17396, %rd17342, %rd17395; + xor.b64 %rd17397, %rd17396, %rd17340; + not.b64 %rd17398, %rd17342; + and.b64 %rd17399, %rd17343, %rd17398; + xor.b64 %rd17400, %rd17399, %rd17341; + not.b64 %rd17401, %rd17343; + and.b64 %rd17402, %rd17344, %rd17401; + xor.b64 %rd17403, %rd17402, %rd17342; + not.b64 %rd17404, %rd17344; + and.b64 %rd17405, %rd17340, %rd17404; + xor.b64 %rd17406, %rd17405, %rd17343; + not.b64 %rd17407, %rd17340; + and.b64 %rd17408, %rd17341, %rd17407; + xor.b64 %rd17409, %rd17408, %rd17344; + not.b64 %rd17410, %rd17346; + and.b64 %rd17411, %rd17347, %rd17410; + xor.b64 %rd17412, %rd17411, %rd17345; + not.b64 %rd17413, %rd17347; + and.b64 %rd17414, %rd17348, %rd17413; + xor.b64 %rd17415, %rd17414, %rd17346; + not.b64 %rd17416, %rd17348; + and.b64 %rd17417, %rd17349, %rd17416; + xor.b64 %rd17418, %rd17417, %rd17347; + not.b64 %rd17419, %rd17349; + and.b64 %rd17420, %rd17345, %rd17419; + xor.b64 %rd17421, %rd17420, %rd17348; + not.b64 %rd17422, %rd17345; + and.b64 %rd17423, %rd17346, %rd17422; + xor.b64 %rd17424, %rd17423, %rd17349; + xor.b64 %rd17425, %rd17352, %rd19817; + xor.b64 %rd17426, %rd17367, %rd17425; + xor.b64 %rd17427, %rd17426, %rd17382; + xor.b64 %rd17428, %rd17427, %rd17397; + xor.b64 %rd17429, %rd17428, %rd17412; + xor.b64 %rd17430, %rd17370, %rd17355; + xor.b64 %rd17431, %rd17430, %rd17385; + xor.b64 %rd17432, %rd17431, %rd17400; + xor.b64 %rd17433, %rd17432, %rd17415; + xor.b64 %rd17434, %rd17373, %rd17358; + xor.b64 %rd17435, %rd17434, %rd17388; + xor.b64 %rd17436, %rd17435, %rd17403; + xor.b64 %rd17437, %rd17436, %rd17418; + xor.b64 %rd17438, %rd17376, %rd17361; + xor.b64 %rd17439, %rd17438, %rd17391; + xor.b64 %rd17440, %rd17439, %rd17406; + xor.b64 %rd17441, %rd17440, %rd17421; + xor.b64 %rd17442, %rd17379, %rd17364; + xor.b64 %rd17443, %rd17442, %rd17394; + xor.b64 %rd17444, %rd17443, %rd17409; + xor.b64 %rd17445, %rd17444, %rd17424; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12383}, %rd17433; + } + { + .reg .b32 %dummy; + mov.b64 {%r12384,%dummy}, %rd17433; + } + shf.l.wrap.b32 %r12385, %r12384, %r12383, 1; + shf.l.wrap.b32 %r12386, %r12383, %r12384, 1; + mov.b64 %rd17446, {%r12386, %r12385}; + xor.b64 %rd17447, %rd17445, %rd17446; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12387}, %rd17437; + } + { + .reg .b32 %dummy; + mov.b64 {%r12388,%dummy}, %rd17437; + } + shf.l.wrap.b32 %r12389, %r12388, %r12387, 1; + shf.l.wrap.b32 %r12390, %r12387, %r12388, 1; + mov.b64 %rd17448, {%r12390, %r12389}; + xor.b64 %rd17449, %rd17448, %rd17429; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12391}, %rd17441; + } + { + .reg .b32 %dummy; + mov.b64 {%r12392,%dummy}, %rd17441; + } + shf.l.wrap.b32 %r12393, %r12392, %r12391, 1; + shf.l.wrap.b32 %r12394, %r12391, %r12392, 1; + mov.b64 %rd17450, {%r12394, %r12393}; + xor.b64 %rd17451, %rd17450, %rd17433; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12395}, %rd17445; + } + { + .reg .b32 %dummy; + mov.b64 {%r12396,%dummy}, %rd17445; + } + shf.l.wrap.b32 %r12397, %r12396, %r12395, 1; + shf.l.wrap.b32 %r12398, %r12395, %r12396, 1; + mov.b64 %rd17452, {%r12398, %r12397}; + xor.b64 %rd17453, %rd17452, %rd17437; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12399}, %rd17429; + } + { + .reg .b32 %dummy; + mov.b64 {%r12400,%dummy}, %rd17429; + } + shf.l.wrap.b32 %r12401, %r12400, %r12399, 1; + shf.l.wrap.b32 %r12402, %r12399, %r12400, 1; + mov.b64 %rd17454, {%r12402, %r12401}; + xor.b64 %rd17455, %rd17441, %rd17454; + xor.b64 %rd17456, %rd17447, %rd17425; + xor.b64 %rd17457, %rd17449, %rd17355; + xor.b64 %rd17458, %rd17451, %rd17358; + xor.b64 %rd17459, %rd17453, %rd17361; + xor.b64 %rd17460, %rd17455, %rd17364; + xor.b64 %rd17461, %rd17447, %rd17367; + xor.b64 %rd17462, %rd17449, %rd17370; + xor.b64 %rd17463, %rd17451, %rd17373; + xor.b64 %rd17464, %rd17453, %rd17376; + xor.b64 %rd17465, %rd17455, %rd17379; + xor.b64 %rd17466, %rd17447, %rd17382; + xor.b64 %rd17467, %rd17449, %rd17385; + xor.b64 %rd17468, %rd17451, %rd17388; + xor.b64 %rd17469, %rd17453, %rd17391; + xor.b64 %rd17470, %rd17455, %rd17394; + xor.b64 %rd17471, %rd17447, %rd17397; + xor.b64 %rd17472, %rd17449, %rd17400; + xor.b64 %rd17473, %rd17451, %rd17403; + xor.b64 %rd17474, %rd17453, %rd17406; + xor.b64 %rd17475, %rd17455, %rd17409; + xor.b64 %rd17476, %rd17447, %rd17412; + xor.b64 %rd17477, %rd17449, %rd17415; + xor.b64 %rd17478, %rd17451, %rd17418; + xor.b64 %rd17479, %rd17453, %rd17421; + xor.b64 %rd17480, %rd17455, %rd17424; + { + .reg .b32 %dummy; + mov.b64 {%r12403,%dummy}, %rd17462; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12404}, %rd17462; + } + shf.r.wrap.b32 %r12405, %r12404, %r12403, 20; + shf.r.wrap.b32 %r12406, %r12403, %r12404, 20; + mov.b64 %rd17481, {%r12406, %r12405}; + { + .reg .b32 %dummy; + mov.b64 {%r12407,%dummy}, %rd17468; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12408}, %rd17468; + } + shf.r.wrap.b32 %r12409, %r12408, %r12407, 21; + shf.r.wrap.b32 %r12410, %r12407, %r12408, 21; + mov.b64 %rd17482, {%r12410, %r12409}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12411}, %rd17474; + } + { + .reg .b32 %dummy; + mov.b64 {%r12412,%dummy}, %rd17474; + } + shf.l.wrap.b32 %r12413, %r12412, %r12411, 21; + shf.l.wrap.b32 %r12414, %r12411, %r12412, 21; + mov.b64 %rd17483, {%r12414, %r12413}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12415}, %rd17480; + } + { + .reg .b32 %dummy; + mov.b64 {%r12416,%dummy}, %rd17480; + } + shf.l.wrap.b32 %r12417, %r12416, %r12415, 14; + shf.l.wrap.b32 %r12418, %r12415, %r12416, 14; + mov.b64 %rd17484, {%r12418, %r12417}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12419}, %rd17459; + } + { + .reg .b32 %dummy; + mov.b64 {%r12420,%dummy}, %rd17459; + } + shf.l.wrap.b32 %r12421, %r12420, %r12419, 28; + shf.l.wrap.b32 %r12422, %r12419, %r12420, 28; + mov.b64 %rd17485, {%r12422, %r12421}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12423}, %rd17465; + } + { + .reg .b32 %dummy; + mov.b64 {%r12424,%dummy}, %rd17465; + } + shf.l.wrap.b32 %r12425, %r12424, %r12423, 20; + shf.l.wrap.b32 %r12426, %r12423, %r12424, 20; + mov.b64 %rd17486, {%r12426, %r12425}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12427}, %rd17466; + } + { + .reg .b32 %dummy; + mov.b64 {%r12428,%dummy}, %rd17466; + } + shf.l.wrap.b32 %r12429, %r12428, %r12427, 3; + shf.l.wrap.b32 %r12430, %r12427, %r12428, 3; + mov.b64 %rd17487, {%r12430, %r12429}; + { + .reg .b32 %dummy; + mov.b64 {%r12431,%dummy}, %rd17472; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12432}, %rd17472; + } + shf.r.wrap.b32 %r12433, %r12432, %r12431, 19; + shf.r.wrap.b32 %r12434, %r12431, %r12432, 19; + mov.b64 %rd17488, {%r12434, %r12433}; + { + .reg .b32 %dummy; + mov.b64 {%r12435,%dummy}, %rd17478; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12436}, %rd17478; + } + shf.r.wrap.b32 %r12437, %r12436, %r12435, 3; + shf.r.wrap.b32 %r12438, %r12435, %r12436, 3; + mov.b64 %rd17489, {%r12438, %r12437}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12439}, %rd17457; + } + { + .reg .b32 %dummy; + mov.b64 {%r12440,%dummy}, %rd17457; + } + shf.l.wrap.b32 %r12441, %r12440, %r12439, 1; + shf.l.wrap.b32 %r12442, %r12439, %r12440, 1; + mov.b64 %rd17490, {%r12442, %r12441}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12443}, %rd17463; + } + { + .reg .b32 %dummy; + mov.b64 {%r12444,%dummy}, %rd17463; + } + shf.l.wrap.b32 %r12445, %r12444, %r12443, 6; + shf.l.wrap.b32 %r12446, %r12443, %r12444, 6; + mov.b64 %rd17491, {%r12446, %r12445}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12447}, %rd17469; + } + { + .reg .b32 %dummy; + mov.b64 {%r12448,%dummy}, %rd17469; + } + shf.l.wrap.b32 %r12449, %r12448, %r12447, 25; + shf.l.wrap.b32 %r12450, %r12447, %r12448, 25; + mov.b64 %rd17492, {%r12450, %r12449}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12451}, %rd17475; + } + { + .reg .b32 %dummy; + mov.b64 {%r12452,%dummy}, %rd17475; + } + shf.l.wrap.b32 %r12453, %r12452, %r12451, 8; + shf.l.wrap.b32 %r12454, %r12451, %r12452, 8; + mov.b64 %rd17493, {%r12454, %r12453}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12455}, %rd17476; + } + { + .reg .b32 %dummy; + mov.b64 {%r12456,%dummy}, %rd17476; + } + shf.l.wrap.b32 %r12457, %r12456, %r12455, 18; + shf.l.wrap.b32 %r12458, %r12455, %r12456, 18; + mov.b64 %rd17494, {%r12458, %r12457}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12459}, %rd17460; + } + { + .reg .b32 %dummy; + mov.b64 {%r12460,%dummy}, %rd17460; + } + shf.l.wrap.b32 %r12461, %r12460, %r12459, 27; + shf.l.wrap.b32 %r12462, %r12459, %r12460, 27; + mov.b64 %rd17495, {%r12462, %r12461}; + { + .reg .b32 %dummy; + mov.b64 {%r12463,%dummy}, %rd17461; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12464}, %rd17461; + } + shf.r.wrap.b32 %r12465, %r12464, %r12463, 28; + shf.r.wrap.b32 %r12466, %r12463, %r12464, 28; + mov.b64 %rd17496, {%r12466, %r12465}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12467}, %rd17467; + } + { + .reg .b32 %dummy; + mov.b64 {%r12468,%dummy}, %rd17467; + } + shf.l.wrap.b32 %r12469, %r12468, %r12467, 10; + shf.l.wrap.b32 %r12470, %r12467, %r12468, 10; + mov.b64 %rd17497, {%r12470, %r12469}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12471}, %rd17473; + } + { + .reg .b32 %dummy; + mov.b64 {%r12472,%dummy}, %rd17473; + } + shf.l.wrap.b32 %r12473, %r12472, %r12471, 15; + shf.l.wrap.b32 %r12474, %r12471, %r12472, 15; + mov.b64 %rd17498, {%r12474, %r12473}; + { + .reg .b32 %dummy; + mov.b64 {%r12475,%dummy}, %rd17479; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12476}, %rd17479; + } + shf.r.wrap.b32 %r12477, %r12476, %r12475, 8; + shf.r.wrap.b32 %r12478, %r12475, %r12476, 8; + mov.b64 %rd17499, {%r12478, %r12477}; + { + .reg .b32 %dummy; + mov.b64 {%r12479,%dummy}, %rd17458; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12480}, %rd17458; + } + shf.r.wrap.b32 %r12481, %r12480, %r12479, 2; + shf.r.wrap.b32 %r12482, %r12479, %r12480, 2; + mov.b64 %rd17500, {%r12482, %r12481}; + { + .reg .b32 %dummy; + mov.b64 {%r12483,%dummy}, %rd17464; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12484}, %rd17464; + } + shf.r.wrap.b32 %r12485, %r12484, %r12483, 9; + shf.r.wrap.b32 %r12486, %r12483, %r12484, 9; + mov.b64 %rd17501, {%r12486, %r12485}; + { + .reg .b32 %dummy; + mov.b64 {%r12487,%dummy}, %rd17470; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12488}, %rd17470; + } + shf.r.wrap.b32 %r12489, %r12488, %r12487, 25; + shf.r.wrap.b32 %r12490, %r12487, %r12488, 25; + mov.b64 %rd17502, {%r12490, %r12489}; + { + .reg .b32 %dummy; + mov.b64 {%r12491,%dummy}, %rd17471; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12492}, %rd17471; + } + shf.r.wrap.b32 %r12493, %r12492, %r12491, 23; + shf.r.wrap.b32 %r12494, %r12491, %r12492, 23; + mov.b64 %rd17503, {%r12494, %r12493}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12495}, %rd17477; + } + { + .reg .b32 %dummy; + mov.b64 {%r12496,%dummy}, %rd17477; + } + shf.l.wrap.b32 %r12497, %r12496, %r12495, 2; + shf.l.wrap.b32 %r12498, %r12495, %r12496, 2; + mov.b64 %rd17504, {%r12498, %r12497}; + not.b64 %rd17505, %rd17481; + and.b64 %rd17506, %rd17482, %rd17505; + xor.b64 %rd17507, %rd17506, %rd17456; + not.b64 %rd17508, %rd17482; + and.b64 %rd17509, %rd17483, %rd17508; + xor.b64 %rd17510, %rd17509, %rd17481; + not.b64 %rd17511, %rd17483; + and.b64 %rd17512, %rd17484, %rd17511; + xor.b64 %rd17513, %rd17512, %rd17482; + not.b64 %rd17514, %rd17484; + and.b64 %rd17515, %rd17456, %rd17514; + xor.b64 %rd17516, %rd17515, %rd17483; + not.b64 %rd17517, %rd17456; + and.b64 %rd17518, %rd17481, %rd17517; + xor.b64 %rd17519, %rd17518, %rd17484; + not.b64 %rd17520, %rd17486; + and.b64 %rd17521, %rd17487, %rd17520; + xor.b64 %rd17522, %rd17521, %rd17485; + not.b64 %rd17523, %rd17487; + and.b64 %rd17524, %rd17488, %rd17523; + xor.b64 %rd17525, %rd17524, %rd17486; + not.b64 %rd17526, %rd17488; + and.b64 %rd17527, %rd17489, %rd17526; + xor.b64 %rd17528, %rd17527, %rd17487; + not.b64 %rd17529, %rd17489; + and.b64 %rd17530, %rd17485, %rd17529; + xor.b64 %rd17531, %rd17530, %rd17488; + not.b64 %rd17532, %rd17485; + and.b64 %rd17533, %rd17486, %rd17532; + xor.b64 %rd17534, %rd17533, %rd17489; + not.b64 %rd17535, %rd17491; + and.b64 %rd17536, %rd17492, %rd17535; + xor.b64 %rd17537, %rd17536, %rd17490; + not.b64 %rd17538, %rd17492; + and.b64 %rd17539, %rd17493, %rd17538; + xor.b64 %rd17540, %rd17539, %rd17491; + not.b64 %rd17541, %rd17493; + and.b64 %rd17542, %rd17494, %rd17541; + xor.b64 %rd17543, %rd17542, %rd17492; + not.b64 %rd17544, %rd17494; + and.b64 %rd17545, %rd17490, %rd17544; + xor.b64 %rd17546, %rd17545, %rd17493; + not.b64 %rd17547, %rd17490; + and.b64 %rd17548, %rd17491, %rd17547; + xor.b64 %rd17549, %rd17548, %rd17494; + not.b64 %rd17550, %rd17496; + and.b64 %rd17551, %rd17497, %rd17550; + xor.b64 %rd17552, %rd17551, %rd17495; + not.b64 %rd17553, %rd17497; + and.b64 %rd17554, %rd17498, %rd17553; + xor.b64 %rd17555, %rd17554, %rd17496; + not.b64 %rd17556, %rd17498; + and.b64 %rd17557, %rd17499, %rd17556; + xor.b64 %rd17558, %rd17557, %rd17497; + not.b64 %rd17559, %rd17499; + and.b64 %rd17560, %rd17495, %rd17559; + xor.b64 %rd17561, %rd17560, %rd17498; + not.b64 %rd17562, %rd17495; + and.b64 %rd17563, %rd17496, %rd17562; + xor.b64 %rd17564, %rd17563, %rd17499; + not.b64 %rd17565, %rd17501; + and.b64 %rd17566, %rd17502, %rd17565; + xor.b64 %rd17567, %rd17566, %rd17500; + not.b64 %rd17568, %rd17502; + and.b64 %rd17569, %rd17503, %rd17568; + xor.b64 %rd17570, %rd17569, %rd17501; + not.b64 %rd17571, %rd17503; + and.b64 %rd17572, %rd17504, %rd17571; + xor.b64 %rd17573, %rd17572, %rd17502; + not.b64 %rd17574, %rd17504; + and.b64 %rd17575, %rd17500, %rd17574; + xor.b64 %rd17576, %rd17575, %rd17503; + not.b64 %rd17577, %rd17500; + and.b64 %rd17578, %rd17501, %rd17577; + xor.b64 %rd17579, %rd17578, %rd17504; + xor.b64 %rd17580, %rd17507, %rd19818; + xor.b64 %rd17581, %rd17522, %rd17580; + xor.b64 %rd17582, %rd17581, %rd17537; + xor.b64 %rd17583, %rd17582, %rd17552; + xor.b64 %rd17584, %rd17583, %rd17567; + xor.b64 %rd17585, %rd17525, %rd17510; + xor.b64 %rd17586, %rd17585, %rd17540; + xor.b64 %rd17587, %rd17586, %rd17555; + xor.b64 %rd17588, %rd17587, %rd17570; + xor.b64 %rd17589, %rd17528, %rd17513; + xor.b64 %rd17590, %rd17589, %rd17543; + xor.b64 %rd17591, %rd17590, %rd17558; + xor.b64 %rd17592, %rd17591, %rd17573; + xor.b64 %rd17593, %rd17531, %rd17516; + xor.b64 %rd17594, %rd17593, %rd17546; + xor.b64 %rd17595, %rd17594, %rd17561; + xor.b64 %rd17596, %rd17595, %rd17576; + xor.b64 %rd17597, %rd17534, %rd17519; + xor.b64 %rd17598, %rd17597, %rd17549; + xor.b64 %rd17599, %rd17598, %rd17564; + xor.b64 %rd17600, %rd17599, %rd17579; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12499}, %rd17588; + } + { + .reg .b32 %dummy; + mov.b64 {%r12500,%dummy}, %rd17588; + } + shf.l.wrap.b32 %r12501, %r12500, %r12499, 1; + shf.l.wrap.b32 %r12502, %r12499, %r12500, 1; + mov.b64 %rd17601, {%r12502, %r12501}; + xor.b64 %rd17602, %rd17600, %rd17601; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12503}, %rd17592; + } + { + .reg .b32 %dummy; + mov.b64 {%r12504,%dummy}, %rd17592; + } + shf.l.wrap.b32 %r12505, %r12504, %r12503, 1; + shf.l.wrap.b32 %r12506, %r12503, %r12504, 1; + mov.b64 %rd17603, {%r12506, %r12505}; + xor.b64 %rd17604, %rd17603, %rd17584; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12507}, %rd17596; + } + { + .reg .b32 %dummy; + mov.b64 {%r12508,%dummy}, %rd17596; + } + shf.l.wrap.b32 %r12509, %r12508, %r12507, 1; + shf.l.wrap.b32 %r12510, %r12507, %r12508, 1; + mov.b64 %rd17605, {%r12510, %r12509}; + xor.b64 %rd17606, %rd17605, %rd17588; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12511}, %rd17600; + } + { + .reg .b32 %dummy; + mov.b64 {%r12512,%dummy}, %rd17600; + } + shf.l.wrap.b32 %r12513, %r12512, %r12511, 1; + shf.l.wrap.b32 %r12514, %r12511, %r12512, 1; + mov.b64 %rd17607, {%r12514, %r12513}; + xor.b64 %rd17608, %rd17607, %rd17592; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12515}, %rd17584; + } + { + .reg .b32 %dummy; + mov.b64 {%r12516,%dummy}, %rd17584; + } + shf.l.wrap.b32 %r12517, %r12516, %r12515, 1; + shf.l.wrap.b32 %r12518, %r12515, %r12516, 1; + mov.b64 %rd17609, {%r12518, %r12517}; + xor.b64 %rd17610, %rd17596, %rd17609; + xor.b64 %rd17611, %rd17602, %rd17580; + xor.b64 %rd17612, %rd17604, %rd17510; + xor.b64 %rd17613, %rd17606, %rd17513; + xor.b64 %rd17614, %rd17608, %rd17516; + xor.b64 %rd17615, %rd17610, %rd17519; + xor.b64 %rd17616, %rd17602, %rd17522; + xor.b64 %rd17617, %rd17604, %rd17525; + xor.b64 %rd17618, %rd17606, %rd17528; + xor.b64 %rd17619, %rd17608, %rd17531; + xor.b64 %rd17620, %rd17610, %rd17534; + xor.b64 %rd17621, %rd17602, %rd17537; + xor.b64 %rd17622, %rd17604, %rd17540; + xor.b64 %rd17623, %rd17606, %rd17543; + xor.b64 %rd17624, %rd17608, %rd17546; + xor.b64 %rd17625, %rd17610, %rd17549; + xor.b64 %rd17626, %rd17602, %rd17552; + xor.b64 %rd17627, %rd17604, %rd17555; + xor.b64 %rd17628, %rd17606, %rd17558; + xor.b64 %rd17629, %rd17608, %rd17561; + xor.b64 %rd17630, %rd17610, %rd17564; + xor.b64 %rd17631, %rd17602, %rd17567; + xor.b64 %rd17632, %rd17604, %rd17570; + xor.b64 %rd17633, %rd17606, %rd17573; + xor.b64 %rd17634, %rd17608, %rd17576; + xor.b64 %rd17635, %rd17610, %rd17579; + { + .reg .b32 %dummy; + mov.b64 {%r12519,%dummy}, %rd17617; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12520}, %rd17617; + } + shf.r.wrap.b32 %r12521, %r12520, %r12519, 20; + shf.r.wrap.b32 %r12522, %r12519, %r12520, 20; + mov.b64 %rd17636, {%r12522, %r12521}; + { + .reg .b32 %dummy; + mov.b64 {%r12523,%dummy}, %rd17623; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12524}, %rd17623; + } + shf.r.wrap.b32 %r12525, %r12524, %r12523, 21; + shf.r.wrap.b32 %r12526, %r12523, %r12524, 21; + mov.b64 %rd17637, {%r12526, %r12525}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12527}, %rd17629; + } + { + .reg .b32 %dummy; + mov.b64 {%r12528,%dummy}, %rd17629; + } + shf.l.wrap.b32 %r12529, %r12528, %r12527, 21; + shf.l.wrap.b32 %r12530, %r12527, %r12528, 21; + mov.b64 %rd17638, {%r12530, %r12529}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12531}, %rd17635; + } + { + .reg .b32 %dummy; + mov.b64 {%r12532,%dummy}, %rd17635; + } + shf.l.wrap.b32 %r12533, %r12532, %r12531, 14; + shf.l.wrap.b32 %r12534, %r12531, %r12532, 14; + mov.b64 %rd17639, {%r12534, %r12533}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12535}, %rd17614; + } + { + .reg .b32 %dummy; + mov.b64 {%r12536,%dummy}, %rd17614; + } + shf.l.wrap.b32 %r12537, %r12536, %r12535, 28; + shf.l.wrap.b32 %r12538, %r12535, %r12536, 28; + mov.b64 %rd17640, {%r12538, %r12537}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12539}, %rd17620; + } + { + .reg .b32 %dummy; + mov.b64 {%r12540,%dummy}, %rd17620; + } + shf.l.wrap.b32 %r12541, %r12540, %r12539, 20; + shf.l.wrap.b32 %r12542, %r12539, %r12540, 20; + mov.b64 %rd17641, {%r12542, %r12541}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12543}, %rd17621; + } + { + .reg .b32 %dummy; + mov.b64 {%r12544,%dummy}, %rd17621; + } + shf.l.wrap.b32 %r12545, %r12544, %r12543, 3; + shf.l.wrap.b32 %r12546, %r12543, %r12544, 3; + mov.b64 %rd17642, {%r12546, %r12545}; + { + .reg .b32 %dummy; + mov.b64 {%r12547,%dummy}, %rd17627; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12548}, %rd17627; + } + shf.r.wrap.b32 %r12549, %r12548, %r12547, 19; + shf.r.wrap.b32 %r12550, %r12547, %r12548, 19; + mov.b64 %rd17643, {%r12550, %r12549}; + { + .reg .b32 %dummy; + mov.b64 {%r12551,%dummy}, %rd17633; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12552}, %rd17633; + } + shf.r.wrap.b32 %r12553, %r12552, %r12551, 3; + shf.r.wrap.b32 %r12554, %r12551, %r12552, 3; + mov.b64 %rd17644, {%r12554, %r12553}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12555}, %rd17612; + } + { + .reg .b32 %dummy; + mov.b64 {%r12556,%dummy}, %rd17612; + } + shf.l.wrap.b32 %r12557, %r12556, %r12555, 1; + shf.l.wrap.b32 %r12558, %r12555, %r12556, 1; + mov.b64 %rd17645, {%r12558, %r12557}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12559}, %rd17618; + } + { + .reg .b32 %dummy; + mov.b64 {%r12560,%dummy}, %rd17618; + } + shf.l.wrap.b32 %r12561, %r12560, %r12559, 6; + shf.l.wrap.b32 %r12562, %r12559, %r12560, 6; + mov.b64 %rd17646, {%r12562, %r12561}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12563}, %rd17624; + } + { + .reg .b32 %dummy; + mov.b64 {%r12564,%dummy}, %rd17624; + } + shf.l.wrap.b32 %r12565, %r12564, %r12563, 25; + shf.l.wrap.b32 %r12566, %r12563, %r12564, 25; + mov.b64 %rd17647, {%r12566, %r12565}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12567}, %rd17630; + } + { + .reg .b32 %dummy; + mov.b64 {%r12568,%dummy}, %rd17630; + } + shf.l.wrap.b32 %r12569, %r12568, %r12567, 8; + shf.l.wrap.b32 %r12570, %r12567, %r12568, 8; + mov.b64 %rd17648, {%r12570, %r12569}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12571}, %rd17631; + } + { + .reg .b32 %dummy; + mov.b64 {%r12572,%dummy}, %rd17631; + } + shf.l.wrap.b32 %r12573, %r12572, %r12571, 18; + shf.l.wrap.b32 %r12574, %r12571, %r12572, 18; + mov.b64 %rd17649, {%r12574, %r12573}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12575}, %rd17615; + } + { + .reg .b32 %dummy; + mov.b64 {%r12576,%dummy}, %rd17615; + } + shf.l.wrap.b32 %r12577, %r12576, %r12575, 27; + shf.l.wrap.b32 %r12578, %r12575, %r12576, 27; + mov.b64 %rd17650, {%r12578, %r12577}; + { + .reg .b32 %dummy; + mov.b64 {%r12579,%dummy}, %rd17616; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12580}, %rd17616; + } + shf.r.wrap.b32 %r12581, %r12580, %r12579, 28; + shf.r.wrap.b32 %r12582, %r12579, %r12580, 28; + mov.b64 %rd17651, {%r12582, %r12581}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12583}, %rd17622; + } + { + .reg .b32 %dummy; + mov.b64 {%r12584,%dummy}, %rd17622; + } + shf.l.wrap.b32 %r12585, %r12584, %r12583, 10; + shf.l.wrap.b32 %r12586, %r12583, %r12584, 10; + mov.b64 %rd17652, {%r12586, %r12585}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12587}, %rd17628; + } + { + .reg .b32 %dummy; + mov.b64 {%r12588,%dummy}, %rd17628; + } + shf.l.wrap.b32 %r12589, %r12588, %r12587, 15; + shf.l.wrap.b32 %r12590, %r12587, %r12588, 15; + mov.b64 %rd17653, {%r12590, %r12589}; + { + .reg .b32 %dummy; + mov.b64 {%r12591,%dummy}, %rd17634; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12592}, %rd17634; + } + shf.r.wrap.b32 %r12593, %r12592, %r12591, 8; + shf.r.wrap.b32 %r12594, %r12591, %r12592, 8; + mov.b64 %rd17654, {%r12594, %r12593}; + { + .reg .b32 %dummy; + mov.b64 {%r12595,%dummy}, %rd17613; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12596}, %rd17613; + } + shf.r.wrap.b32 %r12597, %r12596, %r12595, 2; + shf.r.wrap.b32 %r12598, %r12595, %r12596, 2; + mov.b64 %rd17655, {%r12598, %r12597}; + { + .reg .b32 %dummy; + mov.b64 {%r12599,%dummy}, %rd17619; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12600}, %rd17619; + } + shf.r.wrap.b32 %r12601, %r12600, %r12599, 9; + shf.r.wrap.b32 %r12602, %r12599, %r12600, 9; + mov.b64 %rd17656, {%r12602, %r12601}; + { + .reg .b32 %dummy; + mov.b64 {%r12603,%dummy}, %rd17625; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12604}, %rd17625; + } + shf.r.wrap.b32 %r12605, %r12604, %r12603, 25; + shf.r.wrap.b32 %r12606, %r12603, %r12604, 25; + mov.b64 %rd17657, {%r12606, %r12605}; + { + .reg .b32 %dummy; + mov.b64 {%r12607,%dummy}, %rd17626; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12608}, %rd17626; + } + shf.r.wrap.b32 %r12609, %r12608, %r12607, 23; + shf.r.wrap.b32 %r12610, %r12607, %r12608, 23; + mov.b64 %rd17658, {%r12610, %r12609}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12611}, %rd17632; + } + { + .reg .b32 %dummy; + mov.b64 {%r12612,%dummy}, %rd17632; + } + shf.l.wrap.b32 %r12613, %r12612, %r12611, 2; + shf.l.wrap.b32 %r12614, %r12611, %r12612, 2; + mov.b64 %rd17659, {%r12614, %r12613}; + not.b64 %rd17660, %rd17636; + and.b64 %rd17661, %rd17637, %rd17660; + xor.b64 %rd17662, %rd17661, %rd17611; + not.b64 %rd17663, %rd17637; + and.b64 %rd17664, %rd17638, %rd17663; + xor.b64 %rd17665, %rd17664, %rd17636; + not.b64 %rd17666, %rd17638; + and.b64 %rd17667, %rd17639, %rd17666; + xor.b64 %rd17668, %rd17667, %rd17637; + not.b64 %rd17669, %rd17639; + and.b64 %rd17670, %rd17611, %rd17669; + xor.b64 %rd17671, %rd17670, %rd17638; + not.b64 %rd17672, %rd17611; + and.b64 %rd17673, %rd17636, %rd17672; + xor.b64 %rd17674, %rd17673, %rd17639; + not.b64 %rd17675, %rd17641; + and.b64 %rd17676, %rd17642, %rd17675; + xor.b64 %rd17677, %rd17676, %rd17640; + not.b64 %rd17678, %rd17642; + and.b64 %rd17679, %rd17643, %rd17678; + xor.b64 %rd17680, %rd17679, %rd17641; + not.b64 %rd17681, %rd17643; + and.b64 %rd17682, %rd17644, %rd17681; + xor.b64 %rd17683, %rd17682, %rd17642; + not.b64 %rd17684, %rd17644; + and.b64 %rd17685, %rd17640, %rd17684; + xor.b64 %rd17686, %rd17685, %rd17643; + not.b64 %rd17687, %rd17640; + and.b64 %rd17688, %rd17641, %rd17687; + xor.b64 %rd17689, %rd17688, %rd17644; + not.b64 %rd17690, %rd17646; + and.b64 %rd17691, %rd17647, %rd17690; + xor.b64 %rd17692, %rd17691, %rd17645; + not.b64 %rd17693, %rd17647; + and.b64 %rd17694, %rd17648, %rd17693; + xor.b64 %rd17695, %rd17694, %rd17646; + not.b64 %rd17696, %rd17648; + and.b64 %rd17697, %rd17649, %rd17696; + xor.b64 %rd17698, %rd17697, %rd17647; + not.b64 %rd17699, %rd17649; + and.b64 %rd17700, %rd17645, %rd17699; + xor.b64 %rd17701, %rd17700, %rd17648; + not.b64 %rd17702, %rd17645; + and.b64 %rd17703, %rd17646, %rd17702; + xor.b64 %rd17704, %rd17703, %rd17649; + not.b64 %rd17705, %rd17651; + and.b64 %rd17706, %rd17652, %rd17705; + xor.b64 %rd17707, %rd17706, %rd17650; + not.b64 %rd17708, %rd17652; + and.b64 %rd17709, %rd17653, %rd17708; + xor.b64 %rd17710, %rd17709, %rd17651; + not.b64 %rd17711, %rd17653; + and.b64 %rd17712, %rd17654, %rd17711; + xor.b64 %rd17713, %rd17712, %rd17652; + not.b64 %rd17714, %rd17654; + and.b64 %rd17715, %rd17650, %rd17714; + xor.b64 %rd17716, %rd17715, %rd17653; + not.b64 %rd17717, %rd17650; + and.b64 %rd17718, %rd17651, %rd17717; + xor.b64 %rd17719, %rd17718, %rd17654; + not.b64 %rd17720, %rd17656; + and.b64 %rd17721, %rd17657, %rd17720; + xor.b64 %rd17722, %rd17721, %rd17655; + not.b64 %rd17723, %rd17657; + and.b64 %rd17724, %rd17658, %rd17723; + xor.b64 %rd17725, %rd17724, %rd17656; + not.b64 %rd17726, %rd17658; + and.b64 %rd17727, %rd17659, %rd17726; + xor.b64 %rd17728, %rd17727, %rd17657; + not.b64 %rd17729, %rd17659; + and.b64 %rd17730, %rd17655, %rd17729; + xor.b64 %rd17731, %rd17730, %rd17658; + not.b64 %rd17732, %rd17655; + and.b64 %rd17733, %rd17656, %rd17732; + xor.b64 %rd17734, %rd17733, %rd17659; + xor.b64 %rd17735, %rd17662, %rd19819; + xor.b64 %rd17736, %rd17677, %rd17735; + xor.b64 %rd17737, %rd17736, %rd17692; + xor.b64 %rd17738, %rd17737, %rd17707; + xor.b64 %rd17739, %rd17738, %rd17722; + xor.b64 %rd17740, %rd17680, %rd17665; + xor.b64 %rd17741, %rd17740, %rd17695; + xor.b64 %rd17742, %rd17741, %rd17710; + xor.b64 %rd17743, %rd17742, %rd17725; + xor.b64 %rd17744, %rd17683, %rd17668; + xor.b64 %rd17745, %rd17744, %rd17698; + xor.b64 %rd17746, %rd17745, %rd17713; + xor.b64 %rd17747, %rd17746, %rd17728; + xor.b64 %rd17748, %rd17686, %rd17671; + xor.b64 %rd17749, %rd17748, %rd17701; + xor.b64 %rd17750, %rd17749, %rd17716; + xor.b64 %rd17751, %rd17750, %rd17731; + xor.b64 %rd17752, %rd17689, %rd17674; + xor.b64 %rd17753, %rd17752, %rd17704; + xor.b64 %rd17754, %rd17753, %rd17719; + xor.b64 %rd17755, %rd17754, %rd17734; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12615}, %rd17743; + } + { + .reg .b32 %dummy; + mov.b64 {%r12616,%dummy}, %rd17743; + } + shf.l.wrap.b32 %r12617, %r12616, %r12615, 1; + shf.l.wrap.b32 %r12618, %r12615, %r12616, 1; + mov.b64 %rd17756, {%r12618, %r12617}; + xor.b64 %rd17757, %rd17755, %rd17756; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12619}, %rd17747; + } + { + .reg .b32 %dummy; + mov.b64 {%r12620,%dummy}, %rd17747; + } + shf.l.wrap.b32 %r12621, %r12620, %r12619, 1; + shf.l.wrap.b32 %r12622, %r12619, %r12620, 1; + mov.b64 %rd17758, {%r12622, %r12621}; + xor.b64 %rd17759, %rd17758, %rd17739; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12623}, %rd17751; + } + { + .reg .b32 %dummy; + mov.b64 {%r12624,%dummy}, %rd17751; + } + shf.l.wrap.b32 %r12625, %r12624, %r12623, 1; + shf.l.wrap.b32 %r12626, %r12623, %r12624, 1; + mov.b64 %rd17760, {%r12626, %r12625}; + xor.b64 %rd17761, %rd17760, %rd17743; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12627}, %rd17755; + } + { + .reg .b32 %dummy; + mov.b64 {%r12628,%dummy}, %rd17755; + } + shf.l.wrap.b32 %r12629, %r12628, %r12627, 1; + shf.l.wrap.b32 %r12630, %r12627, %r12628, 1; + mov.b64 %rd17762, {%r12630, %r12629}; + xor.b64 %rd17763, %rd17762, %rd17747; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12631}, %rd17739; + } + { + .reg .b32 %dummy; + mov.b64 {%r12632,%dummy}, %rd17739; + } + shf.l.wrap.b32 %r12633, %r12632, %r12631, 1; + shf.l.wrap.b32 %r12634, %r12631, %r12632, 1; + mov.b64 %rd17764, {%r12634, %r12633}; + xor.b64 %rd17765, %rd17751, %rd17764; + xor.b64 %rd17766, %rd17757, %rd17735; + xor.b64 %rd17767, %rd17759, %rd17665; + xor.b64 %rd17768, %rd17761, %rd17668; + xor.b64 %rd17769, %rd17763, %rd17671; + xor.b64 %rd17770, %rd17765, %rd17674; + xor.b64 %rd17771, %rd17757, %rd17677; + xor.b64 %rd17772, %rd17759, %rd17680; + xor.b64 %rd17773, %rd17761, %rd17683; + xor.b64 %rd17774, %rd17763, %rd17686; + xor.b64 %rd17775, %rd17765, %rd17689; + xor.b64 %rd17776, %rd17757, %rd17692; + xor.b64 %rd17777, %rd17759, %rd17695; + xor.b64 %rd17778, %rd17761, %rd17698; + xor.b64 %rd17779, %rd17763, %rd17701; + xor.b64 %rd17780, %rd17765, %rd17704; + xor.b64 %rd17781, %rd17757, %rd17707; + xor.b64 %rd17782, %rd17759, %rd17710; + xor.b64 %rd17783, %rd17761, %rd17713; + xor.b64 %rd17784, %rd17763, %rd17716; + xor.b64 %rd17785, %rd17765, %rd17719; + xor.b64 %rd17786, %rd17757, %rd17722; + xor.b64 %rd17787, %rd17759, %rd17725; + xor.b64 %rd17788, %rd17761, %rd17728; + xor.b64 %rd17789, %rd17763, %rd17731; + xor.b64 %rd17790, %rd17765, %rd17734; + { + .reg .b32 %dummy; + mov.b64 {%r12635,%dummy}, %rd17772; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12636}, %rd17772; + } + shf.r.wrap.b32 %r12637, %r12636, %r12635, 20; + shf.r.wrap.b32 %r12638, %r12635, %r12636, 20; + mov.b64 %rd17791, {%r12638, %r12637}; + { + .reg .b32 %dummy; + mov.b64 {%r12639,%dummy}, %rd17778; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12640}, %rd17778; + } + shf.r.wrap.b32 %r12641, %r12640, %r12639, 21; + shf.r.wrap.b32 %r12642, %r12639, %r12640, 21; + mov.b64 %rd17792, {%r12642, %r12641}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12643}, %rd17784; + } + { + .reg .b32 %dummy; + mov.b64 {%r12644,%dummy}, %rd17784; + } + shf.l.wrap.b32 %r12645, %r12644, %r12643, 21; + shf.l.wrap.b32 %r12646, %r12643, %r12644, 21; + mov.b64 %rd17793, {%r12646, %r12645}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12647}, %rd17790; + } + { + .reg .b32 %dummy; + mov.b64 {%r12648,%dummy}, %rd17790; + } + shf.l.wrap.b32 %r12649, %r12648, %r12647, 14; + shf.l.wrap.b32 %r12650, %r12647, %r12648, 14; + mov.b64 %rd17794, {%r12650, %r12649}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12651}, %rd17769; + } + { + .reg .b32 %dummy; + mov.b64 {%r12652,%dummy}, %rd17769; + } + shf.l.wrap.b32 %r12653, %r12652, %r12651, 28; + shf.l.wrap.b32 %r12654, %r12651, %r12652, 28; + mov.b64 %rd17795, {%r12654, %r12653}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12655}, %rd17775; + } + { + .reg .b32 %dummy; + mov.b64 {%r12656,%dummy}, %rd17775; + } + shf.l.wrap.b32 %r12657, %r12656, %r12655, 20; + shf.l.wrap.b32 %r12658, %r12655, %r12656, 20; + mov.b64 %rd17796, {%r12658, %r12657}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12659}, %rd17776; + } + { + .reg .b32 %dummy; + mov.b64 {%r12660,%dummy}, %rd17776; + } + shf.l.wrap.b32 %r12661, %r12660, %r12659, 3; + shf.l.wrap.b32 %r12662, %r12659, %r12660, 3; + mov.b64 %rd17797, {%r12662, %r12661}; + { + .reg .b32 %dummy; + mov.b64 {%r12663,%dummy}, %rd17782; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12664}, %rd17782; + } + shf.r.wrap.b32 %r12665, %r12664, %r12663, 19; + shf.r.wrap.b32 %r12666, %r12663, %r12664, 19; + mov.b64 %rd17798, {%r12666, %r12665}; + { + .reg .b32 %dummy; + mov.b64 {%r12667,%dummy}, %rd17788; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12668}, %rd17788; + } + shf.r.wrap.b32 %r12669, %r12668, %r12667, 3; + shf.r.wrap.b32 %r12670, %r12667, %r12668, 3; + mov.b64 %rd17799, {%r12670, %r12669}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12671}, %rd17767; + } + { + .reg .b32 %dummy; + mov.b64 {%r12672,%dummy}, %rd17767; + } + shf.l.wrap.b32 %r12673, %r12672, %r12671, 1; + shf.l.wrap.b32 %r12674, %r12671, %r12672, 1; + mov.b64 %rd17800, {%r12674, %r12673}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12675}, %rd17773; + } + { + .reg .b32 %dummy; + mov.b64 {%r12676,%dummy}, %rd17773; + } + shf.l.wrap.b32 %r12677, %r12676, %r12675, 6; + shf.l.wrap.b32 %r12678, %r12675, %r12676, 6; + mov.b64 %rd17801, {%r12678, %r12677}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12679}, %rd17779; + } + { + .reg .b32 %dummy; + mov.b64 {%r12680,%dummy}, %rd17779; + } + shf.l.wrap.b32 %r12681, %r12680, %r12679, 25; + shf.l.wrap.b32 %r12682, %r12679, %r12680, 25; + mov.b64 %rd17802, {%r12682, %r12681}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12683}, %rd17785; + } + { + .reg .b32 %dummy; + mov.b64 {%r12684,%dummy}, %rd17785; + } + shf.l.wrap.b32 %r12685, %r12684, %r12683, 8; + shf.l.wrap.b32 %r12686, %r12683, %r12684, 8; + mov.b64 %rd17803, {%r12686, %r12685}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12687}, %rd17786; + } + { + .reg .b32 %dummy; + mov.b64 {%r12688,%dummy}, %rd17786; + } + shf.l.wrap.b32 %r12689, %r12688, %r12687, 18; + shf.l.wrap.b32 %r12690, %r12687, %r12688, 18; + mov.b64 %rd17804, {%r12690, %r12689}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12691}, %rd17770; + } + { + .reg .b32 %dummy; + mov.b64 {%r12692,%dummy}, %rd17770; + } + shf.l.wrap.b32 %r12693, %r12692, %r12691, 27; + shf.l.wrap.b32 %r12694, %r12691, %r12692, 27; + mov.b64 %rd17805, {%r12694, %r12693}; + { + .reg .b32 %dummy; + mov.b64 {%r12695,%dummy}, %rd17771; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12696}, %rd17771; + } + shf.r.wrap.b32 %r12697, %r12696, %r12695, 28; + shf.r.wrap.b32 %r12698, %r12695, %r12696, 28; + mov.b64 %rd17806, {%r12698, %r12697}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12699}, %rd17777; + } + { + .reg .b32 %dummy; + mov.b64 {%r12700,%dummy}, %rd17777; + } + shf.l.wrap.b32 %r12701, %r12700, %r12699, 10; + shf.l.wrap.b32 %r12702, %r12699, %r12700, 10; + mov.b64 %rd17807, {%r12702, %r12701}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12703}, %rd17783; + } + { + .reg .b32 %dummy; + mov.b64 {%r12704,%dummy}, %rd17783; + } + shf.l.wrap.b32 %r12705, %r12704, %r12703, 15; + shf.l.wrap.b32 %r12706, %r12703, %r12704, 15; + mov.b64 %rd17808, {%r12706, %r12705}; + { + .reg .b32 %dummy; + mov.b64 {%r12707,%dummy}, %rd17789; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12708}, %rd17789; + } + shf.r.wrap.b32 %r12709, %r12708, %r12707, 8; + shf.r.wrap.b32 %r12710, %r12707, %r12708, 8; + mov.b64 %rd17809, {%r12710, %r12709}; + { + .reg .b32 %dummy; + mov.b64 {%r12711,%dummy}, %rd17768; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12712}, %rd17768; + } + shf.r.wrap.b32 %r12713, %r12712, %r12711, 2; + shf.r.wrap.b32 %r12714, %r12711, %r12712, 2; + mov.b64 %rd17810, {%r12714, %r12713}; + { + .reg .b32 %dummy; + mov.b64 {%r12715,%dummy}, %rd17774; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12716}, %rd17774; + } + shf.r.wrap.b32 %r12717, %r12716, %r12715, 9; + shf.r.wrap.b32 %r12718, %r12715, %r12716, 9; + mov.b64 %rd17811, {%r12718, %r12717}; + { + .reg .b32 %dummy; + mov.b64 {%r12719,%dummy}, %rd17780; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12720}, %rd17780; + } + shf.r.wrap.b32 %r12721, %r12720, %r12719, 25; + shf.r.wrap.b32 %r12722, %r12719, %r12720, 25; + mov.b64 %rd17812, {%r12722, %r12721}; + { + .reg .b32 %dummy; + mov.b64 {%r12723,%dummy}, %rd17781; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12724}, %rd17781; + } + shf.r.wrap.b32 %r12725, %r12724, %r12723, 23; + shf.r.wrap.b32 %r12726, %r12723, %r12724, 23; + mov.b64 %rd17813, {%r12726, %r12725}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12727}, %rd17787; + } + { + .reg .b32 %dummy; + mov.b64 {%r12728,%dummy}, %rd17787; + } + shf.l.wrap.b32 %r12729, %r12728, %r12727, 2; + shf.l.wrap.b32 %r12730, %r12727, %r12728, 2; + mov.b64 %rd17814, {%r12730, %r12729}; + not.b64 %rd17815, %rd17791; + and.b64 %rd17816, %rd17792, %rd17815; + xor.b64 %rd17817, %rd17816, %rd17766; + not.b64 %rd17818, %rd17792; + and.b64 %rd17819, %rd17793, %rd17818; + xor.b64 %rd17820, %rd17819, %rd17791; + not.b64 %rd17821, %rd17793; + and.b64 %rd17822, %rd17794, %rd17821; + xor.b64 %rd17823, %rd17822, %rd17792; + not.b64 %rd17824, %rd17794; + and.b64 %rd17825, %rd17766, %rd17824; + xor.b64 %rd17826, %rd17825, %rd17793; + not.b64 %rd17827, %rd17766; + and.b64 %rd17828, %rd17791, %rd17827; + xor.b64 %rd17829, %rd17828, %rd17794; + not.b64 %rd17830, %rd17796; + and.b64 %rd17831, %rd17797, %rd17830; + xor.b64 %rd17832, %rd17831, %rd17795; + not.b64 %rd17833, %rd17797; + and.b64 %rd17834, %rd17798, %rd17833; + xor.b64 %rd17835, %rd17834, %rd17796; + not.b64 %rd17836, %rd17798; + and.b64 %rd17837, %rd17799, %rd17836; + xor.b64 %rd17838, %rd17837, %rd17797; + not.b64 %rd17839, %rd17799; + and.b64 %rd17840, %rd17795, %rd17839; + xor.b64 %rd17841, %rd17840, %rd17798; + not.b64 %rd17842, %rd17795; + and.b64 %rd17843, %rd17796, %rd17842; + xor.b64 %rd17844, %rd17843, %rd17799; + not.b64 %rd17845, %rd17801; + and.b64 %rd17846, %rd17802, %rd17845; + xor.b64 %rd17847, %rd17846, %rd17800; + not.b64 %rd17848, %rd17802; + and.b64 %rd17849, %rd17803, %rd17848; + xor.b64 %rd17850, %rd17849, %rd17801; + not.b64 %rd17851, %rd17803; + and.b64 %rd17852, %rd17804, %rd17851; + xor.b64 %rd17853, %rd17852, %rd17802; + not.b64 %rd17854, %rd17804; + and.b64 %rd17855, %rd17800, %rd17854; + xor.b64 %rd17856, %rd17855, %rd17803; + not.b64 %rd17857, %rd17800; + and.b64 %rd17858, %rd17801, %rd17857; + xor.b64 %rd17859, %rd17858, %rd17804; + not.b64 %rd17860, %rd17806; + and.b64 %rd17861, %rd17807, %rd17860; + xor.b64 %rd17862, %rd17861, %rd17805; + not.b64 %rd17863, %rd17807; + and.b64 %rd17864, %rd17808, %rd17863; + xor.b64 %rd17865, %rd17864, %rd17806; + not.b64 %rd17866, %rd17808; + and.b64 %rd17867, %rd17809, %rd17866; + xor.b64 %rd17868, %rd17867, %rd17807; + not.b64 %rd17869, %rd17809; + and.b64 %rd17870, %rd17805, %rd17869; + xor.b64 %rd17871, %rd17870, %rd17808; + not.b64 %rd17872, %rd17805; + and.b64 %rd17873, %rd17806, %rd17872; + xor.b64 %rd17874, %rd17873, %rd17809; + not.b64 %rd17875, %rd17811; + and.b64 %rd17876, %rd17812, %rd17875; + xor.b64 %rd17877, %rd17876, %rd17810; + not.b64 %rd17878, %rd17812; + and.b64 %rd17879, %rd17813, %rd17878; + xor.b64 %rd17880, %rd17879, %rd17811; + not.b64 %rd17881, %rd17813; + and.b64 %rd17882, %rd17814, %rd17881; + xor.b64 %rd17883, %rd17882, %rd17812; + not.b64 %rd17884, %rd17814; + and.b64 %rd17885, %rd17810, %rd17884; + xor.b64 %rd17886, %rd17885, %rd17813; + not.b64 %rd17887, %rd17810; + and.b64 %rd17888, %rd17811, %rd17887; + xor.b64 %rd17889, %rd17888, %rd17814; + xor.b64 %rd17890, %rd17817, %rd28; + xor.b64 %rd17891, %rd17832, %rd17890; + xor.b64 %rd17892, %rd17891, %rd17847; + xor.b64 %rd17893, %rd17892, %rd17862; + xor.b64 %rd17894, %rd17893, %rd17877; + xor.b64 %rd17895, %rd17835, %rd17820; + xor.b64 %rd17896, %rd17895, %rd17850; + xor.b64 %rd17897, %rd17896, %rd17865; + xor.b64 %rd17898, %rd17897, %rd17880; + xor.b64 %rd17899, %rd17838, %rd17823; + xor.b64 %rd17900, %rd17899, %rd17853; + xor.b64 %rd17901, %rd17900, %rd17868; + xor.b64 %rd17902, %rd17901, %rd17883; + xor.b64 %rd17903, %rd17841, %rd17826; + xor.b64 %rd17904, %rd17903, %rd17856; + xor.b64 %rd17905, %rd17904, %rd17871; + xor.b64 %rd17906, %rd17905, %rd17886; + xor.b64 %rd17907, %rd17844, %rd17829; + xor.b64 %rd17908, %rd17907, %rd17859; + xor.b64 %rd17909, %rd17908, %rd17874; + xor.b64 %rd17910, %rd17909, %rd17889; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12731}, %rd17898; + } + { + .reg .b32 %dummy; + mov.b64 {%r12732,%dummy}, %rd17898; + } + shf.l.wrap.b32 %r12733, %r12732, %r12731, 1; + shf.l.wrap.b32 %r12734, %r12731, %r12732, 1; + mov.b64 %rd17911, {%r12734, %r12733}; + xor.b64 %rd17912, %rd17910, %rd17911; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12735}, %rd17902; + } + { + .reg .b32 %dummy; + mov.b64 {%r12736,%dummy}, %rd17902; + } + shf.l.wrap.b32 %r12737, %r12736, %r12735, 1; + shf.l.wrap.b32 %r12738, %r12735, %r12736, 1; + mov.b64 %rd17913, {%r12738, %r12737}; + xor.b64 %rd17914, %rd17913, %rd17894; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12739}, %rd17906; + } + { + .reg .b32 %dummy; + mov.b64 {%r12740,%dummy}, %rd17906; + } + shf.l.wrap.b32 %r12741, %r12740, %r12739, 1; + shf.l.wrap.b32 %r12742, %r12739, %r12740, 1; + mov.b64 %rd17915, {%r12742, %r12741}; + xor.b64 %rd17916, %rd17915, %rd17898; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12743}, %rd17910; + } + { + .reg .b32 %dummy; + mov.b64 {%r12744,%dummy}, %rd17910; + } + shf.l.wrap.b32 %r12745, %r12744, %r12743, 1; + shf.l.wrap.b32 %r12746, %r12743, %r12744, 1; + mov.b64 %rd17917, {%r12746, %r12745}; + xor.b64 %rd17918, %rd17917, %rd17902; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12747}, %rd17894; + } + { + .reg .b32 %dummy; + mov.b64 {%r12748,%dummy}, %rd17894; + } + shf.l.wrap.b32 %r12749, %r12748, %r12747, 1; + shf.l.wrap.b32 %r12750, %r12747, %r12748, 1; + mov.b64 %rd17919, {%r12750, %r12749}; + xor.b64 %rd17920, %rd17906, %rd17919; + xor.b64 %rd17921, %rd17912, %rd17890; + xor.b64 %rd17922, %rd17914, %rd17820; + xor.b64 %rd17923, %rd17916, %rd17823; + xor.b64 %rd17924, %rd17918, %rd17826; + xor.b64 %rd17925, %rd17920, %rd17829; + xor.b64 %rd17926, %rd17912, %rd17832; + xor.b64 %rd17927, %rd17914, %rd17835; + xor.b64 %rd17928, %rd17916, %rd17838; + xor.b64 %rd17929, %rd17918, %rd17841; + xor.b64 %rd17930, %rd17920, %rd17844; + xor.b64 %rd17931, %rd17912, %rd17847; + xor.b64 %rd17932, %rd17914, %rd17850; + xor.b64 %rd17933, %rd17916, %rd17853; + xor.b64 %rd17934, %rd17918, %rd17856; + xor.b64 %rd17935, %rd17920, %rd17859; + xor.b64 %rd17936, %rd17912, %rd17862; + xor.b64 %rd17937, %rd17914, %rd17865; + xor.b64 %rd17938, %rd17916, %rd17868; + xor.b64 %rd17939, %rd17918, %rd17871; + xor.b64 %rd17940, %rd17920, %rd17874; + xor.b64 %rd17941, %rd17912, %rd17877; + xor.b64 %rd17942, %rd17914, %rd17880; + xor.b64 %rd17943, %rd17916, %rd17883; + xor.b64 %rd17944, %rd17918, %rd17886; + xor.b64 %rd17945, %rd17920, %rd17889; + { + .reg .b32 %dummy; + mov.b64 {%r12751,%dummy}, %rd17927; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12752}, %rd17927; + } + shf.r.wrap.b32 %r12753, %r12752, %r12751, 20; + shf.r.wrap.b32 %r12754, %r12751, %r12752, 20; + mov.b64 %rd17946, {%r12754, %r12753}; + { + .reg .b32 %dummy; + mov.b64 {%r12755,%dummy}, %rd17933; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12756}, %rd17933; + } + shf.r.wrap.b32 %r12757, %r12756, %r12755, 21; + shf.r.wrap.b32 %r12758, %r12755, %r12756, 21; + mov.b64 %rd17947, {%r12758, %r12757}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12759}, %rd17939; + } + { + .reg .b32 %dummy; + mov.b64 {%r12760,%dummy}, %rd17939; + } + shf.l.wrap.b32 %r12761, %r12760, %r12759, 21; + shf.l.wrap.b32 %r12762, %r12759, %r12760, 21; + mov.b64 %rd17948, {%r12762, %r12761}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12763}, %rd17945; + } + { + .reg .b32 %dummy; + mov.b64 {%r12764,%dummy}, %rd17945; + } + shf.l.wrap.b32 %r12765, %r12764, %r12763, 14; + shf.l.wrap.b32 %r12766, %r12763, %r12764, 14; + mov.b64 %rd17949, {%r12766, %r12765}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12767}, %rd17924; + } + { + .reg .b32 %dummy; + mov.b64 {%r12768,%dummy}, %rd17924; + } + shf.l.wrap.b32 %r12769, %r12768, %r12767, 28; + shf.l.wrap.b32 %r12770, %r12767, %r12768, 28; + mov.b64 %rd17950, {%r12770, %r12769}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12771}, %rd17930; + } + { + .reg .b32 %dummy; + mov.b64 {%r12772,%dummy}, %rd17930; + } + shf.l.wrap.b32 %r12773, %r12772, %r12771, 20; + shf.l.wrap.b32 %r12774, %r12771, %r12772, 20; + mov.b64 %rd17951, {%r12774, %r12773}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12775}, %rd17931; + } + { + .reg .b32 %dummy; + mov.b64 {%r12776,%dummy}, %rd17931; + } + shf.l.wrap.b32 %r12777, %r12776, %r12775, 3; + shf.l.wrap.b32 %r12778, %r12775, %r12776, 3; + mov.b64 %rd17952, {%r12778, %r12777}; + { + .reg .b32 %dummy; + mov.b64 {%r12779,%dummy}, %rd17937; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12780}, %rd17937; + } + shf.r.wrap.b32 %r12781, %r12780, %r12779, 19; + shf.r.wrap.b32 %r12782, %r12779, %r12780, 19; + mov.b64 %rd17953, {%r12782, %r12781}; + { + .reg .b32 %dummy; + mov.b64 {%r12783,%dummy}, %rd17943; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12784}, %rd17943; + } + shf.r.wrap.b32 %r12785, %r12784, %r12783, 3; + shf.r.wrap.b32 %r12786, %r12783, %r12784, 3; + mov.b64 %rd17954, {%r12786, %r12785}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12787}, %rd17922; + } + { + .reg .b32 %dummy; + mov.b64 {%r12788,%dummy}, %rd17922; + } + shf.l.wrap.b32 %r12789, %r12788, %r12787, 1; + shf.l.wrap.b32 %r12790, %r12787, %r12788, 1; + mov.b64 %rd17955, {%r12790, %r12789}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12791}, %rd17928; + } + { + .reg .b32 %dummy; + mov.b64 {%r12792,%dummy}, %rd17928; + } + shf.l.wrap.b32 %r12793, %r12792, %r12791, 6; + shf.l.wrap.b32 %r12794, %r12791, %r12792, 6; + mov.b64 %rd17956, {%r12794, %r12793}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12795}, %rd17934; + } + { + .reg .b32 %dummy; + mov.b64 {%r12796,%dummy}, %rd17934; + } + shf.l.wrap.b32 %r12797, %r12796, %r12795, 25; + shf.l.wrap.b32 %r12798, %r12795, %r12796, 25; + mov.b64 %rd17957, {%r12798, %r12797}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12799}, %rd17940; + } + { + .reg .b32 %dummy; + mov.b64 {%r12800,%dummy}, %rd17940; + } + shf.l.wrap.b32 %r12801, %r12800, %r12799, 8; + shf.l.wrap.b32 %r12802, %r12799, %r12800, 8; + mov.b64 %rd17958, {%r12802, %r12801}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12803}, %rd17941; + } + { + .reg .b32 %dummy; + mov.b64 {%r12804,%dummy}, %rd17941; + } + shf.l.wrap.b32 %r12805, %r12804, %r12803, 18; + shf.l.wrap.b32 %r12806, %r12803, %r12804, 18; + mov.b64 %rd17959, {%r12806, %r12805}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12807}, %rd17925; + } + { + .reg .b32 %dummy; + mov.b64 {%r12808,%dummy}, %rd17925; + } + shf.l.wrap.b32 %r12809, %r12808, %r12807, 27; + shf.l.wrap.b32 %r12810, %r12807, %r12808, 27; + mov.b64 %rd17960, {%r12810, %r12809}; + { + .reg .b32 %dummy; + mov.b64 {%r12811,%dummy}, %rd17926; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12812}, %rd17926; + } + shf.r.wrap.b32 %r12813, %r12812, %r12811, 28; + shf.r.wrap.b32 %r12814, %r12811, %r12812, 28; + mov.b64 %rd17961, {%r12814, %r12813}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12815}, %rd17932; + } + { + .reg .b32 %dummy; + mov.b64 {%r12816,%dummy}, %rd17932; + } + shf.l.wrap.b32 %r12817, %r12816, %r12815, 10; + shf.l.wrap.b32 %r12818, %r12815, %r12816, 10; + mov.b64 %rd17962, {%r12818, %r12817}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12819}, %rd17938; + } + { + .reg .b32 %dummy; + mov.b64 {%r12820,%dummy}, %rd17938; + } + shf.l.wrap.b32 %r12821, %r12820, %r12819, 15; + shf.l.wrap.b32 %r12822, %r12819, %r12820, 15; + mov.b64 %rd17963, {%r12822, %r12821}; + { + .reg .b32 %dummy; + mov.b64 {%r12823,%dummy}, %rd17944; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12824}, %rd17944; + } + shf.r.wrap.b32 %r12825, %r12824, %r12823, 8; + shf.r.wrap.b32 %r12826, %r12823, %r12824, 8; + mov.b64 %rd17964, {%r12826, %r12825}; + { + .reg .b32 %dummy; + mov.b64 {%r12827,%dummy}, %rd17923; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12828}, %rd17923; + } + shf.r.wrap.b32 %r12829, %r12828, %r12827, 2; + shf.r.wrap.b32 %r12830, %r12827, %r12828, 2; + mov.b64 %rd17965, {%r12830, %r12829}; + { + .reg .b32 %dummy; + mov.b64 {%r12831,%dummy}, %rd17929; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12832}, %rd17929; + } + shf.r.wrap.b32 %r12833, %r12832, %r12831, 9; + shf.r.wrap.b32 %r12834, %r12831, %r12832, 9; + mov.b64 %rd17966, {%r12834, %r12833}; + { + .reg .b32 %dummy; + mov.b64 {%r12835,%dummy}, %rd17935; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12836}, %rd17935; + } + shf.r.wrap.b32 %r12837, %r12836, %r12835, 25; + shf.r.wrap.b32 %r12838, %r12835, %r12836, 25; + mov.b64 %rd17967, {%r12838, %r12837}; + { + .reg .b32 %dummy; + mov.b64 {%r12839,%dummy}, %rd17936; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12840}, %rd17936; + } + shf.r.wrap.b32 %r12841, %r12840, %r12839, 23; + shf.r.wrap.b32 %r12842, %r12839, %r12840, 23; + mov.b64 %rd17968, {%r12842, %r12841}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12843}, %rd17942; + } + { + .reg .b32 %dummy; + mov.b64 {%r12844,%dummy}, %rd17942; + } + shf.l.wrap.b32 %r12845, %r12844, %r12843, 2; + shf.l.wrap.b32 %r12846, %r12843, %r12844, 2; + mov.b64 %rd17969, {%r12846, %r12845}; + not.b64 %rd17970, %rd17946; + and.b64 %rd17971, %rd17947, %rd17970; + xor.b64 %rd17972, %rd17971, %rd17921; + not.b64 %rd17973, %rd17947; + and.b64 %rd17974, %rd17948, %rd17973; + xor.b64 %rd17975, %rd17974, %rd17946; + not.b64 %rd17976, %rd17948; + and.b64 %rd17977, %rd17949, %rd17976; + xor.b64 %rd17978, %rd17977, %rd17947; + not.b64 %rd17979, %rd17949; + and.b64 %rd17980, %rd17921, %rd17979; + xor.b64 %rd17981, %rd17980, %rd17948; + not.b64 %rd17982, %rd17921; + and.b64 %rd17983, %rd17946, %rd17982; + xor.b64 %rd17984, %rd17983, %rd17949; + not.b64 %rd17985, %rd17951; + and.b64 %rd17986, %rd17952, %rd17985; + xor.b64 %rd17987, %rd17986, %rd17950; + not.b64 %rd17988, %rd17952; + and.b64 %rd17989, %rd17953, %rd17988; + xor.b64 %rd17990, %rd17989, %rd17951; + not.b64 %rd17991, %rd17953; + and.b64 %rd17992, %rd17954, %rd17991; + xor.b64 %rd17993, %rd17992, %rd17952; + not.b64 %rd17994, %rd17954; + and.b64 %rd17995, %rd17950, %rd17994; + xor.b64 %rd17996, %rd17995, %rd17953; + not.b64 %rd17997, %rd17950; + and.b64 %rd17998, %rd17951, %rd17997; + xor.b64 %rd17999, %rd17998, %rd17954; + not.b64 %rd18000, %rd17956; + and.b64 %rd18001, %rd17957, %rd18000; + xor.b64 %rd18002, %rd18001, %rd17955; + not.b64 %rd18003, %rd17957; + and.b64 %rd18004, %rd17958, %rd18003; + xor.b64 %rd18005, %rd18004, %rd17956; + not.b64 %rd18006, %rd17958; + and.b64 %rd18007, %rd17959, %rd18006; + xor.b64 %rd18008, %rd18007, %rd17957; + not.b64 %rd18009, %rd17959; + and.b64 %rd18010, %rd17955, %rd18009; + xor.b64 %rd18011, %rd18010, %rd17958; + not.b64 %rd18012, %rd17955; + and.b64 %rd18013, %rd17956, %rd18012; + xor.b64 %rd18014, %rd18013, %rd17959; + not.b64 %rd18015, %rd17961; + and.b64 %rd18016, %rd17962, %rd18015; + xor.b64 %rd18017, %rd18016, %rd17960; + not.b64 %rd18018, %rd17962; + and.b64 %rd18019, %rd17963, %rd18018; + xor.b64 %rd18020, %rd18019, %rd17961; + not.b64 %rd18021, %rd17963; + and.b64 %rd18022, %rd17964, %rd18021; + xor.b64 %rd18023, %rd18022, %rd17962; + not.b64 %rd18024, %rd17964; + and.b64 %rd18025, %rd17960, %rd18024; + xor.b64 %rd18026, %rd18025, %rd17963; + not.b64 %rd18027, %rd17960; + and.b64 %rd18028, %rd17961, %rd18027; + xor.b64 %rd18029, %rd18028, %rd17964; + not.b64 %rd18030, %rd17966; + and.b64 %rd18031, %rd17967, %rd18030; + xor.b64 %rd18032, %rd18031, %rd17965; + not.b64 %rd18033, %rd17967; + and.b64 %rd18034, %rd17968, %rd18033; + xor.b64 %rd18035, %rd18034, %rd17966; + not.b64 %rd18036, %rd17968; + and.b64 %rd18037, %rd17969, %rd18036; + xor.b64 %rd18038, %rd18037, %rd17967; + not.b64 %rd18039, %rd17969; + and.b64 %rd18040, %rd17965, %rd18039; + xor.b64 %rd18041, %rd18040, %rd17968; + not.b64 %rd18042, %rd17965; + and.b64 %rd18043, %rd17966, %rd18042; + xor.b64 %rd18044, %rd18043, %rd17969; + xor.b64 %rd18045, %rd17972, %rd29; + xor.b64 %rd18046, %rd17987, %rd18045; + xor.b64 %rd18047, %rd18046, %rd18002; + xor.b64 %rd18048, %rd18047, %rd18017; + xor.b64 %rd18049, %rd18048, %rd18032; + xor.b64 %rd18050, %rd17990, %rd17975; + xor.b64 %rd18051, %rd18050, %rd18005; + xor.b64 %rd18052, %rd18051, %rd18020; + xor.b64 %rd18053, %rd18052, %rd18035; + xor.b64 %rd18054, %rd17993, %rd17978; + xor.b64 %rd18055, %rd18054, %rd18008; + xor.b64 %rd18056, %rd18055, %rd18023; + xor.b64 %rd18057, %rd18056, %rd18038; + xor.b64 %rd18058, %rd17996, %rd17981; + xor.b64 %rd18059, %rd18058, %rd18011; + xor.b64 %rd18060, %rd18059, %rd18026; + xor.b64 %rd18061, %rd18060, %rd18041; + xor.b64 %rd18062, %rd17999, %rd17984; + xor.b64 %rd18063, %rd18062, %rd18014; + xor.b64 %rd18064, %rd18063, %rd18029; + xor.b64 %rd18065, %rd18064, %rd18044; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12847}, %rd18053; + } + { + .reg .b32 %dummy; + mov.b64 {%r12848,%dummy}, %rd18053; + } + shf.l.wrap.b32 %r12849, %r12848, %r12847, 1; + shf.l.wrap.b32 %r12850, %r12847, %r12848, 1; + mov.b64 %rd18066, {%r12850, %r12849}; + xor.b64 %rd18067, %rd18065, %rd18066; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12851}, %rd18057; + } + { + .reg .b32 %dummy; + mov.b64 {%r12852,%dummy}, %rd18057; + } + shf.l.wrap.b32 %r12853, %r12852, %r12851, 1; + shf.l.wrap.b32 %r12854, %r12851, %r12852, 1; + mov.b64 %rd18068, {%r12854, %r12853}; + xor.b64 %rd18069, %rd18068, %rd18049; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12855}, %rd18061; + } + { + .reg .b32 %dummy; + mov.b64 {%r12856,%dummy}, %rd18061; + } + shf.l.wrap.b32 %r12857, %r12856, %r12855, 1; + shf.l.wrap.b32 %r12858, %r12855, %r12856, 1; + mov.b64 %rd18070, {%r12858, %r12857}; + xor.b64 %rd18071, %rd18070, %rd18053; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12859}, %rd18065; + } + { + .reg .b32 %dummy; + mov.b64 {%r12860,%dummy}, %rd18065; + } + shf.l.wrap.b32 %r12861, %r12860, %r12859, 1; + shf.l.wrap.b32 %r12862, %r12859, %r12860, 1; + mov.b64 %rd18072, {%r12862, %r12861}; + xor.b64 %rd18073, %rd18072, %rd18057; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12863}, %rd18049; + } + { + .reg .b32 %dummy; + mov.b64 {%r12864,%dummy}, %rd18049; + } + shf.l.wrap.b32 %r12865, %r12864, %r12863, 1; + shf.l.wrap.b32 %r12866, %r12863, %r12864, 1; + mov.b64 %rd18074, {%r12866, %r12865}; + xor.b64 %rd18075, %rd18061, %rd18074; + xor.b64 %rd18076, %rd18067, %rd18045; + xor.b64 %rd18077, %rd18069, %rd17975; + xor.b64 %rd18078, %rd18071, %rd17978; + xor.b64 %rd18079, %rd18073, %rd17981; + xor.b64 %rd18080, %rd18075, %rd17984; + xor.b64 %rd18081, %rd18067, %rd17987; + xor.b64 %rd18082, %rd18069, %rd17990; + xor.b64 %rd18083, %rd18071, %rd17993; + xor.b64 %rd18084, %rd18073, %rd17996; + xor.b64 %rd18085, %rd18075, %rd17999; + xor.b64 %rd18086, %rd18067, %rd18002; + xor.b64 %rd18087, %rd18069, %rd18005; + xor.b64 %rd18088, %rd18071, %rd18008; + xor.b64 %rd18089, %rd18073, %rd18011; + xor.b64 %rd18090, %rd18075, %rd18014; + xor.b64 %rd18091, %rd18067, %rd18017; + xor.b64 %rd18092, %rd18069, %rd18020; + xor.b64 %rd18093, %rd18071, %rd18023; + xor.b64 %rd18094, %rd18073, %rd18026; + xor.b64 %rd18095, %rd18075, %rd18029; + xor.b64 %rd18096, %rd18067, %rd18032; + xor.b64 %rd18097, %rd18069, %rd18035; + xor.b64 %rd18098, %rd18071, %rd18038; + xor.b64 %rd18099, %rd18073, %rd18041; + xor.b64 %rd18100, %rd18075, %rd18044; + { + .reg .b32 %dummy; + mov.b64 {%r12867,%dummy}, %rd18082; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12868}, %rd18082; + } + shf.r.wrap.b32 %r12869, %r12868, %r12867, 20; + shf.r.wrap.b32 %r12870, %r12867, %r12868, 20; + mov.b64 %rd18101, {%r12870, %r12869}; + { + .reg .b32 %dummy; + mov.b64 {%r12871,%dummy}, %rd18088; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12872}, %rd18088; + } + shf.r.wrap.b32 %r12873, %r12872, %r12871, 21; + shf.r.wrap.b32 %r12874, %r12871, %r12872, 21; + mov.b64 %rd18102, {%r12874, %r12873}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12875}, %rd18094; + } + { + .reg .b32 %dummy; + mov.b64 {%r12876,%dummy}, %rd18094; + } + shf.l.wrap.b32 %r12877, %r12876, %r12875, 21; + shf.l.wrap.b32 %r12878, %r12875, %r12876, 21; + mov.b64 %rd18103, {%r12878, %r12877}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12879}, %rd18100; + } + { + .reg .b32 %dummy; + mov.b64 {%r12880,%dummy}, %rd18100; + } + shf.l.wrap.b32 %r12881, %r12880, %r12879, 14; + shf.l.wrap.b32 %r12882, %r12879, %r12880, 14; + mov.b64 %rd18104, {%r12882, %r12881}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12883}, %rd18079; + } + { + .reg .b32 %dummy; + mov.b64 {%r12884,%dummy}, %rd18079; + } + shf.l.wrap.b32 %r12885, %r12884, %r12883, 28; + shf.l.wrap.b32 %r12886, %r12883, %r12884, 28; + mov.b64 %rd18105, {%r12886, %r12885}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12887}, %rd18085; + } + { + .reg .b32 %dummy; + mov.b64 {%r12888,%dummy}, %rd18085; + } + shf.l.wrap.b32 %r12889, %r12888, %r12887, 20; + shf.l.wrap.b32 %r12890, %r12887, %r12888, 20; + mov.b64 %rd18106, {%r12890, %r12889}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12891}, %rd18086; + } + { + .reg .b32 %dummy; + mov.b64 {%r12892,%dummy}, %rd18086; + } + shf.l.wrap.b32 %r12893, %r12892, %r12891, 3; + shf.l.wrap.b32 %r12894, %r12891, %r12892, 3; + mov.b64 %rd18107, {%r12894, %r12893}; + { + .reg .b32 %dummy; + mov.b64 {%r12895,%dummy}, %rd18092; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12896}, %rd18092; + } + shf.r.wrap.b32 %r12897, %r12896, %r12895, 19; + shf.r.wrap.b32 %r12898, %r12895, %r12896, 19; + mov.b64 %rd18108, {%r12898, %r12897}; + { + .reg .b32 %dummy; + mov.b64 {%r12899,%dummy}, %rd18098; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12900}, %rd18098; + } + shf.r.wrap.b32 %r12901, %r12900, %r12899, 3; + shf.r.wrap.b32 %r12902, %r12899, %r12900, 3; + mov.b64 %rd18109, {%r12902, %r12901}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12903}, %rd18077; + } + { + .reg .b32 %dummy; + mov.b64 {%r12904,%dummy}, %rd18077; + } + shf.l.wrap.b32 %r12905, %r12904, %r12903, 1; + shf.l.wrap.b32 %r12906, %r12903, %r12904, 1; + mov.b64 %rd18110, {%r12906, %r12905}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12907}, %rd18083; + } + { + .reg .b32 %dummy; + mov.b64 {%r12908,%dummy}, %rd18083; + } + shf.l.wrap.b32 %r12909, %r12908, %r12907, 6; + shf.l.wrap.b32 %r12910, %r12907, %r12908, 6; + mov.b64 %rd18111, {%r12910, %r12909}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12911}, %rd18089; + } + { + .reg .b32 %dummy; + mov.b64 {%r12912,%dummy}, %rd18089; + } + shf.l.wrap.b32 %r12913, %r12912, %r12911, 25; + shf.l.wrap.b32 %r12914, %r12911, %r12912, 25; + mov.b64 %rd18112, {%r12914, %r12913}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12915}, %rd18095; + } + { + .reg .b32 %dummy; + mov.b64 {%r12916,%dummy}, %rd18095; + } + shf.l.wrap.b32 %r12917, %r12916, %r12915, 8; + shf.l.wrap.b32 %r12918, %r12915, %r12916, 8; + mov.b64 %rd18113, {%r12918, %r12917}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12919}, %rd18096; + } + { + .reg .b32 %dummy; + mov.b64 {%r12920,%dummy}, %rd18096; + } + shf.l.wrap.b32 %r12921, %r12920, %r12919, 18; + shf.l.wrap.b32 %r12922, %r12919, %r12920, 18; + mov.b64 %rd18114, {%r12922, %r12921}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12923}, %rd18080; + } + { + .reg .b32 %dummy; + mov.b64 {%r12924,%dummy}, %rd18080; + } + shf.l.wrap.b32 %r12925, %r12924, %r12923, 27; + shf.l.wrap.b32 %r12926, %r12923, %r12924, 27; + mov.b64 %rd18115, {%r12926, %r12925}; + { + .reg .b32 %dummy; + mov.b64 {%r12927,%dummy}, %rd18081; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12928}, %rd18081; + } + shf.r.wrap.b32 %r12929, %r12928, %r12927, 28; + shf.r.wrap.b32 %r12930, %r12927, %r12928, 28; + mov.b64 %rd18116, {%r12930, %r12929}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12931}, %rd18087; + } + { + .reg .b32 %dummy; + mov.b64 {%r12932,%dummy}, %rd18087; + } + shf.l.wrap.b32 %r12933, %r12932, %r12931, 10; + shf.l.wrap.b32 %r12934, %r12931, %r12932, 10; + mov.b64 %rd18117, {%r12934, %r12933}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12935}, %rd18093; + } + { + .reg .b32 %dummy; + mov.b64 {%r12936,%dummy}, %rd18093; + } + shf.l.wrap.b32 %r12937, %r12936, %r12935, 15; + shf.l.wrap.b32 %r12938, %r12935, %r12936, 15; + mov.b64 %rd18118, {%r12938, %r12937}; + { + .reg .b32 %dummy; + mov.b64 {%r12939,%dummy}, %rd18099; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12940}, %rd18099; + } + shf.r.wrap.b32 %r12941, %r12940, %r12939, 8; + shf.r.wrap.b32 %r12942, %r12939, %r12940, 8; + mov.b64 %rd18119, {%r12942, %r12941}; + { + .reg .b32 %dummy; + mov.b64 {%r12943,%dummy}, %rd18078; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12944}, %rd18078; + } + shf.r.wrap.b32 %r12945, %r12944, %r12943, 2; + shf.r.wrap.b32 %r12946, %r12943, %r12944, 2; + mov.b64 %rd18120, {%r12946, %r12945}; + { + .reg .b32 %dummy; + mov.b64 {%r12947,%dummy}, %rd18084; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12948}, %rd18084; + } + shf.r.wrap.b32 %r12949, %r12948, %r12947, 9; + shf.r.wrap.b32 %r12950, %r12947, %r12948, 9; + mov.b64 %rd18121, {%r12950, %r12949}; + { + .reg .b32 %dummy; + mov.b64 {%r12951,%dummy}, %rd18090; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12952}, %rd18090; + } + shf.r.wrap.b32 %r12953, %r12952, %r12951, 25; + shf.r.wrap.b32 %r12954, %r12951, %r12952, 25; + mov.b64 %rd18122, {%r12954, %r12953}; + { + .reg .b32 %dummy; + mov.b64 {%r12955,%dummy}, %rd18091; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12956}, %rd18091; + } + shf.r.wrap.b32 %r12957, %r12956, %r12955, 23; + shf.r.wrap.b32 %r12958, %r12955, %r12956, 23; + mov.b64 %rd18123, {%r12958, %r12957}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12959}, %rd18097; + } + { + .reg .b32 %dummy; + mov.b64 {%r12960,%dummy}, %rd18097; + } + shf.l.wrap.b32 %r12961, %r12960, %r12959, 2; + shf.l.wrap.b32 %r12962, %r12959, %r12960, 2; + mov.b64 %rd18124, {%r12962, %r12961}; + not.b64 %rd18125, %rd18101; + and.b64 %rd18126, %rd18102, %rd18125; + xor.b64 %rd18127, %rd18126, %rd18076; + not.b64 %rd18128, %rd18102; + and.b64 %rd18129, %rd18103, %rd18128; + xor.b64 %rd18130, %rd18129, %rd18101; + not.b64 %rd18131, %rd18103; + and.b64 %rd18132, %rd18104, %rd18131; + xor.b64 %rd18133, %rd18132, %rd18102; + not.b64 %rd18134, %rd18104; + and.b64 %rd18135, %rd18076, %rd18134; + xor.b64 %rd18136, %rd18135, %rd18103; + not.b64 %rd18137, %rd18076; + and.b64 %rd18138, %rd18101, %rd18137; + xor.b64 %rd18139, %rd18138, %rd18104; + not.b64 %rd18140, %rd18106; + and.b64 %rd18141, %rd18107, %rd18140; + xor.b64 %rd18142, %rd18141, %rd18105; + not.b64 %rd18143, %rd18107; + and.b64 %rd18144, %rd18108, %rd18143; + xor.b64 %rd18145, %rd18144, %rd18106; + not.b64 %rd18146, %rd18108; + and.b64 %rd18147, %rd18109, %rd18146; + xor.b64 %rd18148, %rd18147, %rd18107; + not.b64 %rd18149, %rd18109; + and.b64 %rd18150, %rd18105, %rd18149; + xor.b64 %rd18151, %rd18150, %rd18108; + not.b64 %rd18152, %rd18105; + and.b64 %rd18153, %rd18106, %rd18152; + xor.b64 %rd18154, %rd18153, %rd18109; + not.b64 %rd18155, %rd18111; + and.b64 %rd18156, %rd18112, %rd18155; + xor.b64 %rd18157, %rd18156, %rd18110; + not.b64 %rd18158, %rd18112; + and.b64 %rd18159, %rd18113, %rd18158; + xor.b64 %rd18160, %rd18159, %rd18111; + not.b64 %rd18161, %rd18113; + and.b64 %rd18162, %rd18114, %rd18161; + xor.b64 %rd18163, %rd18162, %rd18112; + not.b64 %rd18164, %rd18114; + and.b64 %rd18165, %rd18110, %rd18164; + xor.b64 %rd18166, %rd18165, %rd18113; + not.b64 %rd18167, %rd18110; + and.b64 %rd18168, %rd18111, %rd18167; + xor.b64 %rd18169, %rd18168, %rd18114; + not.b64 %rd18170, %rd18116; + and.b64 %rd18171, %rd18117, %rd18170; + xor.b64 %rd18172, %rd18171, %rd18115; + not.b64 %rd18173, %rd18117; + and.b64 %rd18174, %rd18118, %rd18173; + xor.b64 %rd18175, %rd18174, %rd18116; + not.b64 %rd18176, %rd18118; + and.b64 %rd18177, %rd18119, %rd18176; + xor.b64 %rd18178, %rd18177, %rd18117; + not.b64 %rd18179, %rd18119; + and.b64 %rd18180, %rd18115, %rd18179; + xor.b64 %rd18181, %rd18180, %rd18118; + not.b64 %rd18182, %rd18115; + and.b64 %rd18183, %rd18116, %rd18182; + xor.b64 %rd18184, %rd18183, %rd18119; + not.b64 %rd18185, %rd18121; + and.b64 %rd18186, %rd18122, %rd18185; + xor.b64 %rd18187, %rd18186, %rd18120; + not.b64 %rd18188, %rd18122; + and.b64 %rd18189, %rd18123, %rd18188; + xor.b64 %rd18190, %rd18189, %rd18121; + not.b64 %rd18191, %rd18123; + and.b64 %rd18192, %rd18124, %rd18191; + xor.b64 %rd18193, %rd18192, %rd18122; + not.b64 %rd18194, %rd18124; + and.b64 %rd18195, %rd18120, %rd18194; + xor.b64 %rd18196, %rd18195, %rd18123; + not.b64 %rd18197, %rd18120; + and.b64 %rd18198, %rd18121, %rd18197; + xor.b64 %rd18199, %rd18198, %rd18124; + xor.b64 %rd18200, %rd18127, %rd30; + xor.b64 %rd18201, %rd18142, %rd18200; + xor.b64 %rd18202, %rd18201, %rd18157; + xor.b64 %rd18203, %rd18202, %rd18172; + xor.b64 %rd18204, %rd18203, %rd18187; + xor.b64 %rd18205, %rd18145, %rd18130; + xor.b64 %rd18206, %rd18205, %rd18160; + xor.b64 %rd18207, %rd18206, %rd18175; + xor.b64 %rd18208, %rd18207, %rd18190; + xor.b64 %rd18209, %rd18148, %rd18133; + xor.b64 %rd18210, %rd18209, %rd18163; + xor.b64 %rd18211, %rd18210, %rd18178; + xor.b64 %rd18212, %rd18211, %rd18193; + xor.b64 %rd18213, %rd18151, %rd18136; + xor.b64 %rd18214, %rd18213, %rd18166; + xor.b64 %rd18215, %rd18214, %rd18181; + xor.b64 %rd18216, %rd18215, %rd18196; + xor.b64 %rd18217, %rd18154, %rd18139; + xor.b64 %rd18218, %rd18217, %rd18169; + xor.b64 %rd18219, %rd18218, %rd18184; + xor.b64 %rd18220, %rd18219, %rd18199; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12963}, %rd18208; + } + { + .reg .b32 %dummy; + mov.b64 {%r12964,%dummy}, %rd18208; + } + shf.l.wrap.b32 %r12965, %r12964, %r12963, 1; + shf.l.wrap.b32 %r12966, %r12963, %r12964, 1; + mov.b64 %rd18221, {%r12966, %r12965}; + xor.b64 %rd18222, %rd18220, %rd18221; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12967}, %rd18212; + } + { + .reg .b32 %dummy; + mov.b64 {%r12968,%dummy}, %rd18212; + } + shf.l.wrap.b32 %r12969, %r12968, %r12967, 1; + shf.l.wrap.b32 %r12970, %r12967, %r12968, 1; + mov.b64 %rd18223, {%r12970, %r12969}; + xor.b64 %rd18224, %rd18223, %rd18204; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12971}, %rd18216; + } + { + .reg .b32 %dummy; + mov.b64 {%r12972,%dummy}, %rd18216; + } + shf.l.wrap.b32 %r12973, %r12972, %r12971, 1; + shf.l.wrap.b32 %r12974, %r12971, %r12972, 1; + mov.b64 %rd18225, {%r12974, %r12973}; + xor.b64 %rd18226, %rd18225, %rd18208; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12975}, %rd18220; + } + { + .reg .b32 %dummy; + mov.b64 {%r12976,%dummy}, %rd18220; + } + shf.l.wrap.b32 %r12977, %r12976, %r12975, 1; + shf.l.wrap.b32 %r12978, %r12975, %r12976, 1; + mov.b64 %rd18227, {%r12978, %r12977}; + xor.b64 %rd18228, %rd18227, %rd18212; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12979}, %rd18204; + } + { + .reg .b32 %dummy; + mov.b64 {%r12980,%dummy}, %rd18204; + } + shf.l.wrap.b32 %r12981, %r12980, %r12979, 1; + shf.l.wrap.b32 %r12982, %r12979, %r12980, 1; + mov.b64 %rd18229, {%r12982, %r12981}; + xor.b64 %rd18230, %rd18216, %rd18229; + xor.b64 %rd18231, %rd18222, %rd18200; + xor.b64 %rd18232, %rd18224, %rd18130; + xor.b64 %rd18233, %rd18226, %rd18133; + xor.b64 %rd18234, %rd18228, %rd18136; + xor.b64 %rd18235, %rd18230, %rd18139; + xor.b64 %rd18236, %rd18222, %rd18142; + xor.b64 %rd18237, %rd18224, %rd18145; + xor.b64 %rd18238, %rd18226, %rd18148; + xor.b64 %rd18239, %rd18228, %rd18151; + xor.b64 %rd18240, %rd18230, %rd18154; + xor.b64 %rd18241, %rd18222, %rd18157; + xor.b64 %rd18242, %rd18224, %rd18160; + xor.b64 %rd18243, %rd18226, %rd18163; + xor.b64 %rd18244, %rd18228, %rd18166; + xor.b64 %rd18245, %rd18230, %rd18169; + xor.b64 %rd18246, %rd18222, %rd18172; + xor.b64 %rd18247, %rd18224, %rd18175; + xor.b64 %rd18248, %rd18226, %rd18178; + xor.b64 %rd18249, %rd18228, %rd18181; + xor.b64 %rd18250, %rd18230, %rd18184; + xor.b64 %rd18251, %rd18222, %rd18187; + xor.b64 %rd18252, %rd18224, %rd18190; + xor.b64 %rd18253, %rd18226, %rd18193; + xor.b64 %rd18254, %rd18228, %rd18196; + xor.b64 %rd18255, %rd18230, %rd18199; + { + .reg .b32 %dummy; + mov.b64 {%r12983,%dummy}, %rd18237; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12984}, %rd18237; + } + shf.r.wrap.b32 %r12985, %r12984, %r12983, 20; + shf.r.wrap.b32 %r12986, %r12983, %r12984, 20; + mov.b64 %rd18256, {%r12986, %r12985}; + { + .reg .b32 %dummy; + mov.b64 {%r12987,%dummy}, %rd18243; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12988}, %rd18243; + } + shf.r.wrap.b32 %r12989, %r12988, %r12987, 21; + shf.r.wrap.b32 %r12990, %r12987, %r12988, 21; + mov.b64 %rd18257, {%r12990, %r12989}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12991}, %rd18249; + } + { + .reg .b32 %dummy; + mov.b64 {%r12992,%dummy}, %rd18249; + } + shf.l.wrap.b32 %r12993, %r12992, %r12991, 21; + shf.l.wrap.b32 %r12994, %r12991, %r12992, 21; + mov.b64 %rd18258, {%r12994, %r12993}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12995}, %rd18255; + } + { + .reg .b32 %dummy; + mov.b64 {%r12996,%dummy}, %rd18255; + } + shf.l.wrap.b32 %r12997, %r12996, %r12995, 14; + shf.l.wrap.b32 %r12998, %r12995, %r12996, 14; + mov.b64 %rd18259, {%r12998, %r12997}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r12999}, %rd18234; + } + { + .reg .b32 %dummy; + mov.b64 {%r13000,%dummy}, %rd18234; + } + shf.l.wrap.b32 %r13001, %r13000, %r12999, 28; + shf.l.wrap.b32 %r13002, %r12999, %r13000, 28; + mov.b64 %rd18260, {%r13002, %r13001}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13003}, %rd18240; + } + { + .reg .b32 %dummy; + mov.b64 {%r13004,%dummy}, %rd18240; + } + shf.l.wrap.b32 %r13005, %r13004, %r13003, 20; + shf.l.wrap.b32 %r13006, %r13003, %r13004, 20; + mov.b64 %rd18261, {%r13006, %r13005}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13007}, %rd18241; + } + { + .reg .b32 %dummy; + mov.b64 {%r13008,%dummy}, %rd18241; + } + shf.l.wrap.b32 %r13009, %r13008, %r13007, 3; + shf.l.wrap.b32 %r13010, %r13007, %r13008, 3; + mov.b64 %rd18262, {%r13010, %r13009}; + { + .reg .b32 %dummy; + mov.b64 {%r13011,%dummy}, %rd18247; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13012}, %rd18247; + } + shf.r.wrap.b32 %r13013, %r13012, %r13011, 19; + shf.r.wrap.b32 %r13014, %r13011, %r13012, 19; + mov.b64 %rd18263, {%r13014, %r13013}; + { + .reg .b32 %dummy; + mov.b64 {%r13015,%dummy}, %rd18253; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13016}, %rd18253; + } + shf.r.wrap.b32 %r13017, %r13016, %r13015, 3; + shf.r.wrap.b32 %r13018, %r13015, %r13016, 3; + mov.b64 %rd18264, {%r13018, %r13017}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13019}, %rd18232; + } + { + .reg .b32 %dummy; + mov.b64 {%r13020,%dummy}, %rd18232; + } + shf.l.wrap.b32 %r13021, %r13020, %r13019, 1; + shf.l.wrap.b32 %r13022, %r13019, %r13020, 1; + mov.b64 %rd18265, {%r13022, %r13021}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13023}, %rd18238; + } + { + .reg .b32 %dummy; + mov.b64 {%r13024,%dummy}, %rd18238; + } + shf.l.wrap.b32 %r13025, %r13024, %r13023, 6; + shf.l.wrap.b32 %r13026, %r13023, %r13024, 6; + mov.b64 %rd18266, {%r13026, %r13025}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13027}, %rd18244; + } + { + .reg .b32 %dummy; + mov.b64 {%r13028,%dummy}, %rd18244; + } + shf.l.wrap.b32 %r13029, %r13028, %r13027, 25; + shf.l.wrap.b32 %r13030, %r13027, %r13028, 25; + mov.b64 %rd18267, {%r13030, %r13029}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13031}, %rd18250; + } + { + .reg .b32 %dummy; + mov.b64 {%r13032,%dummy}, %rd18250; + } + shf.l.wrap.b32 %r13033, %r13032, %r13031, 8; + shf.l.wrap.b32 %r13034, %r13031, %r13032, 8; + mov.b64 %rd18268, {%r13034, %r13033}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13035}, %rd18251; + } + { + .reg .b32 %dummy; + mov.b64 {%r13036,%dummy}, %rd18251; + } + shf.l.wrap.b32 %r13037, %r13036, %r13035, 18; + shf.l.wrap.b32 %r13038, %r13035, %r13036, 18; + mov.b64 %rd18269, {%r13038, %r13037}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13039}, %rd18235; + } + { + .reg .b32 %dummy; + mov.b64 {%r13040,%dummy}, %rd18235; + } + shf.l.wrap.b32 %r13041, %r13040, %r13039, 27; + shf.l.wrap.b32 %r13042, %r13039, %r13040, 27; + mov.b64 %rd18270, {%r13042, %r13041}; + { + .reg .b32 %dummy; + mov.b64 {%r13043,%dummy}, %rd18236; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13044}, %rd18236; + } + shf.r.wrap.b32 %r13045, %r13044, %r13043, 28; + shf.r.wrap.b32 %r13046, %r13043, %r13044, 28; + mov.b64 %rd18271, {%r13046, %r13045}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13047}, %rd18242; + } + { + .reg .b32 %dummy; + mov.b64 {%r13048,%dummy}, %rd18242; + } + shf.l.wrap.b32 %r13049, %r13048, %r13047, 10; + shf.l.wrap.b32 %r13050, %r13047, %r13048, 10; + mov.b64 %rd18272, {%r13050, %r13049}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13051}, %rd18248; + } + { + .reg .b32 %dummy; + mov.b64 {%r13052,%dummy}, %rd18248; + } + shf.l.wrap.b32 %r13053, %r13052, %r13051, 15; + shf.l.wrap.b32 %r13054, %r13051, %r13052, 15; + mov.b64 %rd18273, {%r13054, %r13053}; + { + .reg .b32 %dummy; + mov.b64 {%r13055,%dummy}, %rd18254; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13056}, %rd18254; + } + shf.r.wrap.b32 %r13057, %r13056, %r13055, 8; + shf.r.wrap.b32 %r13058, %r13055, %r13056, 8; + mov.b64 %rd18274, {%r13058, %r13057}; + { + .reg .b32 %dummy; + mov.b64 {%r13059,%dummy}, %rd18233; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13060}, %rd18233; + } + shf.r.wrap.b32 %r13061, %r13060, %r13059, 2; + shf.r.wrap.b32 %r13062, %r13059, %r13060, 2; + mov.b64 %rd18275, {%r13062, %r13061}; + { + .reg .b32 %dummy; + mov.b64 {%r13063,%dummy}, %rd18239; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13064}, %rd18239; + } + shf.r.wrap.b32 %r13065, %r13064, %r13063, 9; + shf.r.wrap.b32 %r13066, %r13063, %r13064, 9; + mov.b64 %rd18276, {%r13066, %r13065}; + { + .reg .b32 %dummy; + mov.b64 {%r13067,%dummy}, %rd18245; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13068}, %rd18245; + } + shf.r.wrap.b32 %r13069, %r13068, %r13067, 25; + shf.r.wrap.b32 %r13070, %r13067, %r13068, 25; + mov.b64 %rd18277, {%r13070, %r13069}; + { + .reg .b32 %dummy; + mov.b64 {%r13071,%dummy}, %rd18246; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13072}, %rd18246; + } + shf.r.wrap.b32 %r13073, %r13072, %r13071, 23; + shf.r.wrap.b32 %r13074, %r13071, %r13072, 23; + mov.b64 %rd18278, {%r13074, %r13073}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13075}, %rd18252; + } + { + .reg .b32 %dummy; + mov.b64 {%r13076,%dummy}, %rd18252; + } + shf.l.wrap.b32 %r13077, %r13076, %r13075, 2; + shf.l.wrap.b32 %r13078, %r13075, %r13076, 2; + mov.b64 %rd18279, {%r13078, %r13077}; + not.b64 %rd18280, %rd18256; + and.b64 %rd18281, %rd18257, %rd18280; + xor.b64 %rd18282, %rd18281, %rd18231; + not.b64 %rd18283, %rd18257; + and.b64 %rd18284, %rd18258, %rd18283; + xor.b64 %rd18285, %rd18284, %rd18256; + not.b64 %rd18286, %rd18258; + and.b64 %rd18287, %rd18259, %rd18286; + xor.b64 %rd18288, %rd18287, %rd18257; + not.b64 %rd18289, %rd18259; + and.b64 %rd18290, %rd18231, %rd18289; + xor.b64 %rd18291, %rd18290, %rd18258; + not.b64 %rd18292, %rd18231; + and.b64 %rd18293, %rd18256, %rd18292; + xor.b64 %rd18294, %rd18293, %rd18259; + not.b64 %rd18295, %rd18261; + and.b64 %rd18296, %rd18262, %rd18295; + xor.b64 %rd18297, %rd18296, %rd18260; + not.b64 %rd18298, %rd18262; + and.b64 %rd18299, %rd18263, %rd18298; + xor.b64 %rd18300, %rd18299, %rd18261; + not.b64 %rd18301, %rd18263; + and.b64 %rd18302, %rd18264, %rd18301; + xor.b64 %rd18303, %rd18302, %rd18262; + not.b64 %rd18304, %rd18264; + and.b64 %rd18305, %rd18260, %rd18304; + xor.b64 %rd18306, %rd18305, %rd18263; + not.b64 %rd18307, %rd18260; + and.b64 %rd18308, %rd18261, %rd18307; + xor.b64 %rd18309, %rd18308, %rd18264; + not.b64 %rd18310, %rd18266; + and.b64 %rd18311, %rd18267, %rd18310; + xor.b64 %rd18312, %rd18311, %rd18265; + not.b64 %rd18313, %rd18267; + and.b64 %rd18314, %rd18268, %rd18313; + xor.b64 %rd18315, %rd18314, %rd18266; + not.b64 %rd18316, %rd18268; + and.b64 %rd18317, %rd18269, %rd18316; + xor.b64 %rd18318, %rd18317, %rd18267; + not.b64 %rd18319, %rd18269; + and.b64 %rd18320, %rd18265, %rd18319; + xor.b64 %rd18321, %rd18320, %rd18268; + not.b64 %rd18322, %rd18265; + and.b64 %rd18323, %rd18266, %rd18322; + xor.b64 %rd18324, %rd18323, %rd18269; + not.b64 %rd18325, %rd18271; + and.b64 %rd18326, %rd18272, %rd18325; + xor.b64 %rd18327, %rd18326, %rd18270; + not.b64 %rd18328, %rd18272; + and.b64 %rd18329, %rd18273, %rd18328; + xor.b64 %rd18330, %rd18329, %rd18271; + not.b64 %rd18331, %rd18273; + and.b64 %rd18332, %rd18274, %rd18331; + xor.b64 %rd18333, %rd18332, %rd18272; + not.b64 %rd18334, %rd18274; + and.b64 %rd18335, %rd18270, %rd18334; + xor.b64 %rd18336, %rd18335, %rd18273; + not.b64 %rd18337, %rd18270; + and.b64 %rd18338, %rd18271, %rd18337; + xor.b64 %rd18339, %rd18338, %rd18274; + not.b64 %rd18340, %rd18276; + and.b64 %rd18341, %rd18277, %rd18340; + xor.b64 %rd18342, %rd18341, %rd18275; + not.b64 %rd18343, %rd18277; + and.b64 %rd18344, %rd18278, %rd18343; + xor.b64 %rd18345, %rd18344, %rd18276; + not.b64 %rd18346, %rd18278; + and.b64 %rd18347, %rd18279, %rd18346; + xor.b64 %rd18348, %rd18347, %rd18277; + not.b64 %rd18349, %rd18279; + and.b64 %rd18350, %rd18275, %rd18349; + xor.b64 %rd18351, %rd18350, %rd18278; + not.b64 %rd18352, %rd18275; + and.b64 %rd18353, %rd18276, %rd18352; + xor.b64 %rd18354, %rd18353, %rd18279; + xor.b64 %rd18355, %rd18282, %rd31; + xor.b64 %rd18356, %rd18297, %rd18355; + xor.b64 %rd18357, %rd18356, %rd18312; + xor.b64 %rd18358, %rd18357, %rd18327; + xor.b64 %rd18359, %rd18358, %rd18342; + xor.b64 %rd18360, %rd18300, %rd18285; + xor.b64 %rd18361, %rd18360, %rd18315; + xor.b64 %rd18362, %rd18361, %rd18330; + xor.b64 %rd18363, %rd18362, %rd18345; + xor.b64 %rd18364, %rd18303, %rd18288; + xor.b64 %rd18365, %rd18364, %rd18318; + xor.b64 %rd18366, %rd18365, %rd18333; + xor.b64 %rd18367, %rd18366, %rd18348; + xor.b64 %rd18368, %rd18306, %rd18291; + xor.b64 %rd18369, %rd18368, %rd18321; + xor.b64 %rd18370, %rd18369, %rd18336; + xor.b64 %rd18371, %rd18370, %rd18351; + xor.b64 %rd18372, %rd18309, %rd18294; + xor.b64 %rd18373, %rd18372, %rd18324; + xor.b64 %rd18374, %rd18373, %rd18339; + xor.b64 %rd18375, %rd18374, %rd18354; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13079}, %rd18363; + } + { + .reg .b32 %dummy; + mov.b64 {%r13080,%dummy}, %rd18363; + } + shf.l.wrap.b32 %r13081, %r13080, %r13079, 1; + shf.l.wrap.b32 %r13082, %r13079, %r13080, 1; + mov.b64 %rd18376, {%r13082, %r13081}; + xor.b64 %rd18377, %rd18375, %rd18376; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13083}, %rd18367; + } + { + .reg .b32 %dummy; + mov.b64 {%r13084,%dummy}, %rd18367; + } + shf.l.wrap.b32 %r13085, %r13084, %r13083, 1; + shf.l.wrap.b32 %r13086, %r13083, %r13084, 1; + mov.b64 %rd18378, {%r13086, %r13085}; + xor.b64 %rd18379, %rd18378, %rd18359; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13087}, %rd18371; + } + { + .reg .b32 %dummy; + mov.b64 {%r13088,%dummy}, %rd18371; + } + shf.l.wrap.b32 %r13089, %r13088, %r13087, 1; + shf.l.wrap.b32 %r13090, %r13087, %r13088, 1; + mov.b64 %rd18380, {%r13090, %r13089}; + xor.b64 %rd18381, %rd18380, %rd18363; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13091}, %rd18375; + } + { + .reg .b32 %dummy; + mov.b64 {%r13092,%dummy}, %rd18375; + } + shf.l.wrap.b32 %r13093, %r13092, %r13091, 1; + shf.l.wrap.b32 %r13094, %r13091, %r13092, 1; + mov.b64 %rd18382, {%r13094, %r13093}; + xor.b64 %rd18383, %rd18382, %rd18367; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13095}, %rd18359; + } + { + .reg .b32 %dummy; + mov.b64 {%r13096,%dummy}, %rd18359; + } + shf.l.wrap.b32 %r13097, %r13096, %r13095, 1; + shf.l.wrap.b32 %r13098, %r13095, %r13096, 1; + mov.b64 %rd18384, {%r13098, %r13097}; + xor.b64 %rd18385, %rd18371, %rd18384; + xor.b64 %rd18386, %rd18377, %rd18355; + xor.b64 %rd18387, %rd18379, %rd18285; + xor.b64 %rd18388, %rd18381, %rd18288; + xor.b64 %rd18389, %rd18383, %rd18291; + xor.b64 %rd18390, %rd18385, %rd18294; + xor.b64 %rd18391, %rd18377, %rd18297; + xor.b64 %rd18392, %rd18379, %rd18300; + xor.b64 %rd18393, %rd18381, %rd18303; + xor.b64 %rd18394, %rd18383, %rd18306; + xor.b64 %rd18395, %rd18385, %rd18309; + xor.b64 %rd18396, %rd18377, %rd18312; + xor.b64 %rd18397, %rd18379, %rd18315; + xor.b64 %rd18398, %rd18381, %rd18318; + xor.b64 %rd18399, %rd18383, %rd18321; + xor.b64 %rd18400, %rd18385, %rd18324; + xor.b64 %rd18401, %rd18377, %rd18327; + xor.b64 %rd18402, %rd18379, %rd18330; + xor.b64 %rd18403, %rd18381, %rd18333; + xor.b64 %rd18404, %rd18383, %rd18336; + xor.b64 %rd18405, %rd18385, %rd18339; + xor.b64 %rd18406, %rd18377, %rd18342; + xor.b64 %rd18407, %rd18379, %rd18345; + xor.b64 %rd18408, %rd18381, %rd18348; + xor.b64 %rd18409, %rd18383, %rd18351; + xor.b64 %rd18410, %rd18385, %rd18354; + { + .reg .b32 %dummy; + mov.b64 {%r13099,%dummy}, %rd18392; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13100}, %rd18392; + } + shf.r.wrap.b32 %r13101, %r13100, %r13099, 20; + shf.r.wrap.b32 %r13102, %r13099, %r13100, 20; + mov.b64 %rd18411, {%r13102, %r13101}; + { + .reg .b32 %dummy; + mov.b64 {%r13103,%dummy}, %rd18398; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13104}, %rd18398; + } + shf.r.wrap.b32 %r13105, %r13104, %r13103, 21; + shf.r.wrap.b32 %r13106, %r13103, %r13104, 21; + mov.b64 %rd18412, {%r13106, %r13105}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13107}, %rd18404; + } + { + .reg .b32 %dummy; + mov.b64 {%r13108,%dummy}, %rd18404; + } + shf.l.wrap.b32 %r13109, %r13108, %r13107, 21; + shf.l.wrap.b32 %r13110, %r13107, %r13108, 21; + mov.b64 %rd18413, {%r13110, %r13109}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13111}, %rd18410; + } + { + .reg .b32 %dummy; + mov.b64 {%r13112,%dummy}, %rd18410; + } + shf.l.wrap.b32 %r13113, %r13112, %r13111, 14; + shf.l.wrap.b32 %r13114, %r13111, %r13112, 14; + mov.b64 %rd18414, {%r13114, %r13113}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13115}, %rd18389; + } + { + .reg .b32 %dummy; + mov.b64 {%r13116,%dummy}, %rd18389; + } + shf.l.wrap.b32 %r13117, %r13116, %r13115, 28; + shf.l.wrap.b32 %r13118, %r13115, %r13116, 28; + mov.b64 %rd18415, {%r13118, %r13117}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13119}, %rd18395; + } + { + .reg .b32 %dummy; + mov.b64 {%r13120,%dummy}, %rd18395; + } + shf.l.wrap.b32 %r13121, %r13120, %r13119, 20; + shf.l.wrap.b32 %r13122, %r13119, %r13120, 20; + mov.b64 %rd18416, {%r13122, %r13121}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13123}, %rd18396; + } + { + .reg .b32 %dummy; + mov.b64 {%r13124,%dummy}, %rd18396; + } + shf.l.wrap.b32 %r13125, %r13124, %r13123, 3; + shf.l.wrap.b32 %r13126, %r13123, %r13124, 3; + mov.b64 %rd18417, {%r13126, %r13125}; + { + .reg .b32 %dummy; + mov.b64 {%r13127,%dummy}, %rd18402; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13128}, %rd18402; + } + shf.r.wrap.b32 %r13129, %r13128, %r13127, 19; + shf.r.wrap.b32 %r13130, %r13127, %r13128, 19; + mov.b64 %rd18418, {%r13130, %r13129}; + { + .reg .b32 %dummy; + mov.b64 {%r13131,%dummy}, %rd18408; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13132}, %rd18408; + } + shf.r.wrap.b32 %r13133, %r13132, %r13131, 3; + shf.r.wrap.b32 %r13134, %r13131, %r13132, 3; + mov.b64 %rd18419, {%r13134, %r13133}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13135}, %rd18387; + } + { + .reg .b32 %dummy; + mov.b64 {%r13136,%dummy}, %rd18387; + } + shf.l.wrap.b32 %r13137, %r13136, %r13135, 1; + shf.l.wrap.b32 %r13138, %r13135, %r13136, 1; + mov.b64 %rd18420, {%r13138, %r13137}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13139}, %rd18393; + } + { + .reg .b32 %dummy; + mov.b64 {%r13140,%dummy}, %rd18393; + } + shf.l.wrap.b32 %r13141, %r13140, %r13139, 6; + shf.l.wrap.b32 %r13142, %r13139, %r13140, 6; + mov.b64 %rd18421, {%r13142, %r13141}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13143}, %rd18399; + } + { + .reg .b32 %dummy; + mov.b64 {%r13144,%dummy}, %rd18399; + } + shf.l.wrap.b32 %r13145, %r13144, %r13143, 25; + shf.l.wrap.b32 %r13146, %r13143, %r13144, 25; + mov.b64 %rd18422, {%r13146, %r13145}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13147}, %rd18405; + } + { + .reg .b32 %dummy; + mov.b64 {%r13148,%dummy}, %rd18405; + } + shf.l.wrap.b32 %r13149, %r13148, %r13147, 8; + shf.l.wrap.b32 %r13150, %r13147, %r13148, 8; + mov.b64 %rd18423, {%r13150, %r13149}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13151}, %rd18406; + } + { + .reg .b32 %dummy; + mov.b64 {%r13152,%dummy}, %rd18406; + } + shf.l.wrap.b32 %r13153, %r13152, %r13151, 18; + shf.l.wrap.b32 %r13154, %r13151, %r13152, 18; + mov.b64 %rd18424, {%r13154, %r13153}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13155}, %rd18390; + } + { + .reg .b32 %dummy; + mov.b64 {%r13156,%dummy}, %rd18390; + } + shf.l.wrap.b32 %r13157, %r13156, %r13155, 27; + shf.l.wrap.b32 %r13158, %r13155, %r13156, 27; + mov.b64 %rd18425, {%r13158, %r13157}; + { + .reg .b32 %dummy; + mov.b64 {%r13159,%dummy}, %rd18391; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13160}, %rd18391; + } + shf.r.wrap.b32 %r13161, %r13160, %r13159, 28; + shf.r.wrap.b32 %r13162, %r13159, %r13160, 28; + mov.b64 %rd18426, {%r13162, %r13161}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13163}, %rd18397; + } + { + .reg .b32 %dummy; + mov.b64 {%r13164,%dummy}, %rd18397; + } + shf.l.wrap.b32 %r13165, %r13164, %r13163, 10; + shf.l.wrap.b32 %r13166, %r13163, %r13164, 10; + mov.b64 %rd18427, {%r13166, %r13165}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13167}, %rd18403; + } + { + .reg .b32 %dummy; + mov.b64 {%r13168,%dummy}, %rd18403; + } + shf.l.wrap.b32 %r13169, %r13168, %r13167, 15; + shf.l.wrap.b32 %r13170, %r13167, %r13168, 15; + mov.b64 %rd18428, {%r13170, %r13169}; + { + .reg .b32 %dummy; + mov.b64 {%r13171,%dummy}, %rd18409; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13172}, %rd18409; + } + shf.r.wrap.b32 %r13173, %r13172, %r13171, 8; + shf.r.wrap.b32 %r13174, %r13171, %r13172, 8; + mov.b64 %rd18429, {%r13174, %r13173}; + { + .reg .b32 %dummy; + mov.b64 {%r13175,%dummy}, %rd18388; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13176}, %rd18388; + } + shf.r.wrap.b32 %r13177, %r13176, %r13175, 2; + shf.r.wrap.b32 %r13178, %r13175, %r13176, 2; + mov.b64 %rd18430, {%r13178, %r13177}; + { + .reg .b32 %dummy; + mov.b64 {%r13179,%dummy}, %rd18394; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13180}, %rd18394; + } + shf.r.wrap.b32 %r13181, %r13180, %r13179, 9; + shf.r.wrap.b32 %r13182, %r13179, %r13180, 9; + mov.b64 %rd18431, {%r13182, %r13181}; + { + .reg .b32 %dummy; + mov.b64 {%r13183,%dummy}, %rd18400; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13184}, %rd18400; + } + shf.r.wrap.b32 %r13185, %r13184, %r13183, 25; + shf.r.wrap.b32 %r13186, %r13183, %r13184, 25; + mov.b64 %rd18432, {%r13186, %r13185}; + { + .reg .b32 %dummy; + mov.b64 {%r13187,%dummy}, %rd18401; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13188}, %rd18401; + } + shf.r.wrap.b32 %r13189, %r13188, %r13187, 23; + shf.r.wrap.b32 %r13190, %r13187, %r13188, 23; + mov.b64 %rd18433, {%r13190, %r13189}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13191}, %rd18407; + } + { + .reg .b32 %dummy; + mov.b64 {%r13192,%dummy}, %rd18407; + } + shf.l.wrap.b32 %r13193, %r13192, %r13191, 2; + shf.l.wrap.b32 %r13194, %r13191, %r13192, 2; + mov.b64 %rd18434, {%r13194, %r13193}; + not.b64 %rd18435, %rd18411; + and.b64 %rd18436, %rd18412, %rd18435; + xor.b64 %rd18437, %rd18436, %rd18386; + not.b64 %rd18438, %rd18412; + and.b64 %rd18439, %rd18413, %rd18438; + xor.b64 %rd18440, %rd18439, %rd18411; + not.b64 %rd18441, %rd18413; + and.b64 %rd18442, %rd18414, %rd18441; + xor.b64 %rd18443, %rd18442, %rd18412; + not.b64 %rd18444, %rd18414; + and.b64 %rd18445, %rd18386, %rd18444; + xor.b64 %rd18446, %rd18445, %rd18413; + not.b64 %rd18447, %rd18386; + and.b64 %rd18448, %rd18411, %rd18447; + xor.b64 %rd18449, %rd18448, %rd18414; + not.b64 %rd18450, %rd18416; + and.b64 %rd18451, %rd18417, %rd18450; + xor.b64 %rd18452, %rd18451, %rd18415; + not.b64 %rd18453, %rd18417; + and.b64 %rd18454, %rd18418, %rd18453; + xor.b64 %rd18455, %rd18454, %rd18416; + not.b64 %rd18456, %rd18418; + and.b64 %rd18457, %rd18419, %rd18456; + xor.b64 %rd18458, %rd18457, %rd18417; + not.b64 %rd18459, %rd18419; + and.b64 %rd18460, %rd18415, %rd18459; + xor.b64 %rd18461, %rd18460, %rd18418; + not.b64 %rd18462, %rd18415; + and.b64 %rd18463, %rd18416, %rd18462; + xor.b64 %rd18464, %rd18463, %rd18419; + not.b64 %rd18465, %rd18421; + and.b64 %rd18466, %rd18422, %rd18465; + xor.b64 %rd18467, %rd18466, %rd18420; + not.b64 %rd18468, %rd18422; + and.b64 %rd18469, %rd18423, %rd18468; + xor.b64 %rd18470, %rd18469, %rd18421; + not.b64 %rd18471, %rd18423; + and.b64 %rd18472, %rd18424, %rd18471; + xor.b64 %rd18473, %rd18472, %rd18422; + not.b64 %rd18474, %rd18424; + and.b64 %rd18475, %rd18420, %rd18474; + xor.b64 %rd18476, %rd18475, %rd18423; + not.b64 %rd18477, %rd18420; + and.b64 %rd18478, %rd18421, %rd18477; + xor.b64 %rd18479, %rd18478, %rd18424; + not.b64 %rd18480, %rd18426; + and.b64 %rd18481, %rd18427, %rd18480; + xor.b64 %rd18482, %rd18481, %rd18425; + not.b64 %rd18483, %rd18427; + and.b64 %rd18484, %rd18428, %rd18483; + xor.b64 %rd18485, %rd18484, %rd18426; + not.b64 %rd18486, %rd18428; + and.b64 %rd18487, %rd18429, %rd18486; + xor.b64 %rd18488, %rd18487, %rd18427; + not.b64 %rd18489, %rd18429; + and.b64 %rd18490, %rd18425, %rd18489; + xor.b64 %rd18491, %rd18490, %rd18428; + not.b64 %rd18492, %rd18425; + and.b64 %rd18493, %rd18426, %rd18492; + xor.b64 %rd18494, %rd18493, %rd18429; + not.b64 %rd18495, %rd18431; + and.b64 %rd18496, %rd18432, %rd18495; + xor.b64 %rd18497, %rd18496, %rd18430; + not.b64 %rd18498, %rd18432; + and.b64 %rd18499, %rd18433, %rd18498; + xor.b64 %rd18500, %rd18499, %rd18431; + not.b64 %rd18501, %rd18433; + and.b64 %rd18502, %rd18434, %rd18501; + xor.b64 %rd18503, %rd18502, %rd18432; + not.b64 %rd18504, %rd18434; + and.b64 %rd18505, %rd18430, %rd18504; + xor.b64 %rd18506, %rd18505, %rd18433; + not.b64 %rd18507, %rd18430; + and.b64 %rd18508, %rd18431, %rd18507; + xor.b64 %rd18509, %rd18508, %rd18434; + xor.b64 %rd18510, %rd18437, %rd32; + xor.b64 %rd18511, %rd18452, %rd18510; + xor.b64 %rd18512, %rd18511, %rd18467; + xor.b64 %rd18513, %rd18512, %rd18482; + xor.b64 %rd18514, %rd18513, %rd18497; + xor.b64 %rd18515, %rd18455, %rd18440; + xor.b64 %rd18516, %rd18515, %rd18470; + xor.b64 %rd18517, %rd18516, %rd18485; + xor.b64 %rd18518, %rd18517, %rd18500; + xor.b64 %rd18519, %rd18458, %rd18443; + xor.b64 %rd18520, %rd18519, %rd18473; + xor.b64 %rd18521, %rd18520, %rd18488; + xor.b64 %rd18522, %rd18521, %rd18503; + xor.b64 %rd18523, %rd18461, %rd18446; + xor.b64 %rd18524, %rd18523, %rd18476; + xor.b64 %rd18525, %rd18524, %rd18491; + xor.b64 %rd18526, %rd18525, %rd18506; + xor.b64 %rd18527, %rd18464, %rd18449; + xor.b64 %rd18528, %rd18527, %rd18479; + xor.b64 %rd18529, %rd18528, %rd18494; + xor.b64 %rd18530, %rd18529, %rd18509; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13195}, %rd18518; + } + { + .reg .b32 %dummy; + mov.b64 {%r13196,%dummy}, %rd18518; + } + shf.l.wrap.b32 %r13197, %r13196, %r13195, 1; + shf.l.wrap.b32 %r13198, %r13195, %r13196, 1; + mov.b64 %rd18531, {%r13198, %r13197}; + xor.b64 %rd18532, %rd18530, %rd18531; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13199}, %rd18522; + } + { + .reg .b32 %dummy; + mov.b64 {%r13200,%dummy}, %rd18522; + } + shf.l.wrap.b32 %r13201, %r13200, %r13199, 1; + shf.l.wrap.b32 %r13202, %r13199, %r13200, 1; + mov.b64 %rd18533, {%r13202, %r13201}; + xor.b64 %rd18534, %rd18533, %rd18514; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13203}, %rd18526; + } + { + .reg .b32 %dummy; + mov.b64 {%r13204,%dummy}, %rd18526; + } + shf.l.wrap.b32 %r13205, %r13204, %r13203, 1; + shf.l.wrap.b32 %r13206, %r13203, %r13204, 1; + mov.b64 %rd18535, {%r13206, %r13205}; + xor.b64 %rd18536, %rd18535, %rd18518; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13207}, %rd18530; + } + { + .reg .b32 %dummy; + mov.b64 {%r13208,%dummy}, %rd18530; + } + shf.l.wrap.b32 %r13209, %r13208, %r13207, 1; + shf.l.wrap.b32 %r13210, %r13207, %r13208, 1; + mov.b64 %rd18537, {%r13210, %r13209}; + xor.b64 %rd18538, %rd18537, %rd18522; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13211}, %rd18514; + } + { + .reg .b32 %dummy; + mov.b64 {%r13212,%dummy}, %rd18514; + } + shf.l.wrap.b32 %r13213, %r13212, %r13211, 1; + shf.l.wrap.b32 %r13214, %r13211, %r13212, 1; + mov.b64 %rd18539, {%r13214, %r13213}; + xor.b64 %rd18540, %rd18526, %rd18539; + xor.b64 %rd18541, %rd18532, %rd18510; + xor.b64 %rd18542, %rd18534, %rd18440; + xor.b64 %rd18543, %rd18536, %rd18443; + xor.b64 %rd18544, %rd18538, %rd18446; + xor.b64 %rd18545, %rd18540, %rd18449; + xor.b64 %rd18546, %rd18532, %rd18452; + xor.b64 %rd18547, %rd18534, %rd18455; + xor.b64 %rd18548, %rd18536, %rd18458; + xor.b64 %rd18549, %rd18538, %rd18461; + xor.b64 %rd18550, %rd18540, %rd18464; + xor.b64 %rd18551, %rd18532, %rd18467; + xor.b64 %rd18552, %rd18534, %rd18470; + xor.b64 %rd18553, %rd18536, %rd18473; + xor.b64 %rd18554, %rd18538, %rd18476; + xor.b64 %rd18555, %rd18540, %rd18479; + xor.b64 %rd18556, %rd18532, %rd18482; + xor.b64 %rd18557, %rd18534, %rd18485; + xor.b64 %rd18558, %rd18536, %rd18488; + xor.b64 %rd18559, %rd18538, %rd18491; + xor.b64 %rd18560, %rd18540, %rd18494; + xor.b64 %rd18561, %rd18532, %rd18497; + xor.b64 %rd18562, %rd18534, %rd18500; + xor.b64 %rd18563, %rd18536, %rd18503; + xor.b64 %rd18564, %rd18538, %rd18506; + xor.b64 %rd18565, %rd18540, %rd18509; + { + .reg .b32 %dummy; + mov.b64 {%r13215,%dummy}, %rd18547; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13216}, %rd18547; + } + shf.r.wrap.b32 %r13217, %r13216, %r13215, 20; + shf.r.wrap.b32 %r13218, %r13215, %r13216, 20; + mov.b64 %rd18566, {%r13218, %r13217}; + { + .reg .b32 %dummy; + mov.b64 {%r13219,%dummy}, %rd18553; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13220}, %rd18553; + } + shf.r.wrap.b32 %r13221, %r13220, %r13219, 21; + shf.r.wrap.b32 %r13222, %r13219, %r13220, 21; + mov.b64 %rd18567, {%r13222, %r13221}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13223}, %rd18559; + } + { + .reg .b32 %dummy; + mov.b64 {%r13224,%dummy}, %rd18559; + } + shf.l.wrap.b32 %r13225, %r13224, %r13223, 21; + shf.l.wrap.b32 %r13226, %r13223, %r13224, 21; + mov.b64 %rd18568, {%r13226, %r13225}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13227}, %rd18565; + } + { + .reg .b32 %dummy; + mov.b64 {%r13228,%dummy}, %rd18565; + } + shf.l.wrap.b32 %r13229, %r13228, %r13227, 14; + shf.l.wrap.b32 %r13230, %r13227, %r13228, 14; + mov.b64 %rd18569, {%r13230, %r13229}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13231}, %rd18544; + } + { + .reg .b32 %dummy; + mov.b64 {%r13232,%dummy}, %rd18544; + } + shf.l.wrap.b32 %r13233, %r13232, %r13231, 28; + shf.l.wrap.b32 %r13234, %r13231, %r13232, 28; + mov.b64 %rd18570, {%r13234, %r13233}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13235}, %rd18550; + } + { + .reg .b32 %dummy; + mov.b64 {%r13236,%dummy}, %rd18550; + } + shf.l.wrap.b32 %r13237, %r13236, %r13235, 20; + shf.l.wrap.b32 %r13238, %r13235, %r13236, 20; + mov.b64 %rd18571, {%r13238, %r13237}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13239}, %rd18551; + } + { + .reg .b32 %dummy; + mov.b64 {%r13240,%dummy}, %rd18551; + } + shf.l.wrap.b32 %r13241, %r13240, %r13239, 3; + shf.l.wrap.b32 %r13242, %r13239, %r13240, 3; + mov.b64 %rd18572, {%r13242, %r13241}; + { + .reg .b32 %dummy; + mov.b64 {%r13243,%dummy}, %rd18557; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13244}, %rd18557; + } + shf.r.wrap.b32 %r13245, %r13244, %r13243, 19; + shf.r.wrap.b32 %r13246, %r13243, %r13244, 19; + mov.b64 %rd18573, {%r13246, %r13245}; + { + .reg .b32 %dummy; + mov.b64 {%r13247,%dummy}, %rd18563; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13248}, %rd18563; + } + shf.r.wrap.b32 %r13249, %r13248, %r13247, 3; + shf.r.wrap.b32 %r13250, %r13247, %r13248, 3; + mov.b64 %rd18574, {%r13250, %r13249}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13251}, %rd18542; + } + { + .reg .b32 %dummy; + mov.b64 {%r13252,%dummy}, %rd18542; + } + shf.l.wrap.b32 %r13253, %r13252, %r13251, 1; + shf.l.wrap.b32 %r13254, %r13251, %r13252, 1; + mov.b64 %rd18575, {%r13254, %r13253}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13255}, %rd18548; + } + { + .reg .b32 %dummy; + mov.b64 {%r13256,%dummy}, %rd18548; + } + shf.l.wrap.b32 %r13257, %r13256, %r13255, 6; + shf.l.wrap.b32 %r13258, %r13255, %r13256, 6; + mov.b64 %rd18576, {%r13258, %r13257}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13259}, %rd18554; + } + { + .reg .b32 %dummy; + mov.b64 {%r13260,%dummy}, %rd18554; + } + shf.l.wrap.b32 %r13261, %r13260, %r13259, 25; + shf.l.wrap.b32 %r13262, %r13259, %r13260, 25; + mov.b64 %rd18577, {%r13262, %r13261}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13263}, %rd18560; + } + { + .reg .b32 %dummy; + mov.b64 {%r13264,%dummy}, %rd18560; + } + shf.l.wrap.b32 %r13265, %r13264, %r13263, 8; + shf.l.wrap.b32 %r13266, %r13263, %r13264, 8; + mov.b64 %rd18578, {%r13266, %r13265}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13267}, %rd18561; + } + { + .reg .b32 %dummy; + mov.b64 {%r13268,%dummy}, %rd18561; + } + shf.l.wrap.b32 %r13269, %r13268, %r13267, 18; + shf.l.wrap.b32 %r13270, %r13267, %r13268, 18; + mov.b64 %rd18579, {%r13270, %r13269}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13271}, %rd18545; + } + { + .reg .b32 %dummy; + mov.b64 {%r13272,%dummy}, %rd18545; + } + shf.l.wrap.b32 %r13273, %r13272, %r13271, 27; + shf.l.wrap.b32 %r13274, %r13271, %r13272, 27; + mov.b64 %rd18580, {%r13274, %r13273}; + { + .reg .b32 %dummy; + mov.b64 {%r13275,%dummy}, %rd18546; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13276}, %rd18546; + } + shf.r.wrap.b32 %r13277, %r13276, %r13275, 28; + shf.r.wrap.b32 %r13278, %r13275, %r13276, 28; + mov.b64 %rd18581, {%r13278, %r13277}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13279}, %rd18552; + } + { + .reg .b32 %dummy; + mov.b64 {%r13280,%dummy}, %rd18552; + } + shf.l.wrap.b32 %r13281, %r13280, %r13279, 10; + shf.l.wrap.b32 %r13282, %r13279, %r13280, 10; + mov.b64 %rd18582, {%r13282, %r13281}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13283}, %rd18558; + } + { + .reg .b32 %dummy; + mov.b64 {%r13284,%dummy}, %rd18558; + } + shf.l.wrap.b32 %r13285, %r13284, %r13283, 15; + shf.l.wrap.b32 %r13286, %r13283, %r13284, 15; + mov.b64 %rd18583, {%r13286, %r13285}; + { + .reg .b32 %dummy; + mov.b64 {%r13287,%dummy}, %rd18564; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13288}, %rd18564; + } + shf.r.wrap.b32 %r13289, %r13288, %r13287, 8; + shf.r.wrap.b32 %r13290, %r13287, %r13288, 8; + mov.b64 %rd18584, {%r13290, %r13289}; + { + .reg .b32 %dummy; + mov.b64 {%r13291,%dummy}, %rd18543; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13292}, %rd18543; + } + shf.r.wrap.b32 %r13293, %r13292, %r13291, 2; + shf.r.wrap.b32 %r13294, %r13291, %r13292, 2; + mov.b64 %rd18585, {%r13294, %r13293}; + { + .reg .b32 %dummy; + mov.b64 {%r13295,%dummy}, %rd18549; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13296}, %rd18549; + } + shf.r.wrap.b32 %r13297, %r13296, %r13295, 9; + shf.r.wrap.b32 %r13298, %r13295, %r13296, 9; + mov.b64 %rd18586, {%r13298, %r13297}; + { + .reg .b32 %dummy; + mov.b64 {%r13299,%dummy}, %rd18555; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13300}, %rd18555; + } + shf.r.wrap.b32 %r13301, %r13300, %r13299, 25; + shf.r.wrap.b32 %r13302, %r13299, %r13300, 25; + mov.b64 %rd18587, {%r13302, %r13301}; + { + .reg .b32 %dummy; + mov.b64 {%r13303,%dummy}, %rd18556; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13304}, %rd18556; + } + shf.r.wrap.b32 %r13305, %r13304, %r13303, 23; + shf.r.wrap.b32 %r13306, %r13303, %r13304, 23; + mov.b64 %rd18588, {%r13306, %r13305}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13307}, %rd18562; + } + { + .reg .b32 %dummy; + mov.b64 {%r13308,%dummy}, %rd18562; + } + shf.l.wrap.b32 %r13309, %r13308, %r13307, 2; + shf.l.wrap.b32 %r13310, %r13307, %r13308, 2; + mov.b64 %rd18589, {%r13310, %r13309}; + not.b64 %rd18590, %rd18566; + and.b64 %rd18591, %rd18567, %rd18590; + xor.b64 %rd18592, %rd18591, %rd18541; + not.b64 %rd18593, %rd18567; + and.b64 %rd18594, %rd18568, %rd18593; + xor.b64 %rd18595, %rd18594, %rd18566; + not.b64 %rd18596, %rd18568; + and.b64 %rd18597, %rd18569, %rd18596; + xor.b64 %rd18598, %rd18597, %rd18567; + not.b64 %rd18599, %rd18569; + and.b64 %rd18600, %rd18541, %rd18599; + xor.b64 %rd18601, %rd18600, %rd18568; + not.b64 %rd18602, %rd18541; + and.b64 %rd18603, %rd18566, %rd18602; + xor.b64 %rd18604, %rd18603, %rd18569; + not.b64 %rd18605, %rd18571; + and.b64 %rd18606, %rd18572, %rd18605; + xor.b64 %rd18607, %rd18606, %rd18570; + not.b64 %rd18608, %rd18572; + and.b64 %rd18609, %rd18573, %rd18608; + xor.b64 %rd18610, %rd18609, %rd18571; + not.b64 %rd18611, %rd18573; + and.b64 %rd18612, %rd18574, %rd18611; + xor.b64 %rd18613, %rd18612, %rd18572; + not.b64 %rd18614, %rd18574; + and.b64 %rd18615, %rd18570, %rd18614; + xor.b64 %rd18616, %rd18615, %rd18573; + not.b64 %rd18617, %rd18570; + and.b64 %rd18618, %rd18571, %rd18617; + xor.b64 %rd18619, %rd18618, %rd18574; + not.b64 %rd18620, %rd18576; + and.b64 %rd18621, %rd18577, %rd18620; + xor.b64 %rd18622, %rd18621, %rd18575; + not.b64 %rd18623, %rd18577; + and.b64 %rd18624, %rd18578, %rd18623; + xor.b64 %rd18625, %rd18624, %rd18576; + not.b64 %rd18626, %rd18578; + and.b64 %rd18627, %rd18579, %rd18626; + xor.b64 %rd18628, %rd18627, %rd18577; + not.b64 %rd18629, %rd18579; + and.b64 %rd18630, %rd18575, %rd18629; + xor.b64 %rd18631, %rd18630, %rd18578; + not.b64 %rd18632, %rd18575; + and.b64 %rd18633, %rd18576, %rd18632; + xor.b64 %rd18634, %rd18633, %rd18579; + not.b64 %rd18635, %rd18581; + and.b64 %rd18636, %rd18582, %rd18635; + xor.b64 %rd18637, %rd18636, %rd18580; + not.b64 %rd18638, %rd18582; + and.b64 %rd18639, %rd18583, %rd18638; + xor.b64 %rd18640, %rd18639, %rd18581; + not.b64 %rd18641, %rd18583; + and.b64 %rd18642, %rd18584, %rd18641; + xor.b64 %rd18643, %rd18642, %rd18582; + not.b64 %rd18644, %rd18584; + and.b64 %rd18645, %rd18580, %rd18644; + xor.b64 %rd18646, %rd18645, %rd18583; + not.b64 %rd18647, %rd18580; + and.b64 %rd18648, %rd18581, %rd18647; + xor.b64 %rd18649, %rd18648, %rd18584; + not.b64 %rd18650, %rd18586; + and.b64 %rd18651, %rd18587, %rd18650; + xor.b64 %rd18652, %rd18651, %rd18585; + not.b64 %rd18653, %rd18587; + and.b64 %rd18654, %rd18588, %rd18653; + xor.b64 %rd18655, %rd18654, %rd18586; + not.b64 %rd18656, %rd18588; + and.b64 %rd18657, %rd18589, %rd18656; + xor.b64 %rd18658, %rd18657, %rd18587; + not.b64 %rd18659, %rd18589; + and.b64 %rd18660, %rd18585, %rd18659; + xor.b64 %rd18661, %rd18660, %rd18588; + not.b64 %rd18662, %rd18585; + and.b64 %rd18663, %rd18586, %rd18662; + xor.b64 %rd18664, %rd18663, %rd18589; + xor.b64 %rd18665, %rd18592, %rd33; + xor.b64 %rd18666, %rd18607, %rd18665; + xor.b64 %rd18667, %rd18666, %rd18622; + xor.b64 %rd18668, %rd18667, %rd18637; + xor.b64 %rd18669, %rd18668, %rd18652; + xor.b64 %rd18670, %rd18610, %rd18595; + xor.b64 %rd18671, %rd18670, %rd18625; + xor.b64 %rd18672, %rd18671, %rd18640; + xor.b64 %rd18673, %rd18672, %rd18655; + xor.b64 %rd18674, %rd18613, %rd18598; + xor.b64 %rd18675, %rd18674, %rd18628; + xor.b64 %rd18676, %rd18675, %rd18643; + xor.b64 %rd18677, %rd18676, %rd18658; + xor.b64 %rd18678, %rd18616, %rd18601; + xor.b64 %rd18679, %rd18678, %rd18631; + xor.b64 %rd18680, %rd18679, %rd18646; + xor.b64 %rd18681, %rd18680, %rd18661; + xor.b64 %rd18682, %rd18619, %rd18604; + xor.b64 %rd18683, %rd18682, %rd18634; + xor.b64 %rd18684, %rd18683, %rd18649; + xor.b64 %rd18685, %rd18684, %rd18664; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13311}, %rd18673; + } + { + .reg .b32 %dummy; + mov.b64 {%r13312,%dummy}, %rd18673; + } + shf.l.wrap.b32 %r13313, %r13312, %r13311, 1; + shf.l.wrap.b32 %r13314, %r13311, %r13312, 1; + mov.b64 %rd18686, {%r13314, %r13313}; + xor.b64 %rd18687, %rd18685, %rd18686; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13315}, %rd18677; + } + { + .reg .b32 %dummy; + mov.b64 {%r13316,%dummy}, %rd18677; + } + shf.l.wrap.b32 %r13317, %r13316, %r13315, 1; + shf.l.wrap.b32 %r13318, %r13315, %r13316, 1; + mov.b64 %rd18688, {%r13318, %r13317}; + xor.b64 %rd18689, %rd18688, %rd18669; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13319}, %rd18681; + } + { + .reg .b32 %dummy; + mov.b64 {%r13320,%dummy}, %rd18681; + } + shf.l.wrap.b32 %r13321, %r13320, %r13319, 1; + shf.l.wrap.b32 %r13322, %r13319, %r13320, 1; + mov.b64 %rd18690, {%r13322, %r13321}; + xor.b64 %rd18691, %rd18690, %rd18673; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13323}, %rd18685; + } + { + .reg .b32 %dummy; + mov.b64 {%r13324,%dummy}, %rd18685; + } + shf.l.wrap.b32 %r13325, %r13324, %r13323, 1; + shf.l.wrap.b32 %r13326, %r13323, %r13324, 1; + mov.b64 %rd18692, {%r13326, %r13325}; + xor.b64 %rd18693, %rd18692, %rd18677; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13327}, %rd18669; + } + { + .reg .b32 %dummy; + mov.b64 {%r13328,%dummy}, %rd18669; + } + shf.l.wrap.b32 %r13329, %r13328, %r13327, 1; + shf.l.wrap.b32 %r13330, %r13327, %r13328, 1; + mov.b64 %rd18694, {%r13330, %r13329}; + xor.b64 %rd18695, %rd18681, %rd18694; + xor.b64 %rd18696, %rd18687, %rd18665; + xor.b64 %rd18697, %rd18689, %rd18595; + xor.b64 %rd18698, %rd18691, %rd18598; + xor.b64 %rd18699, %rd18693, %rd18601; + xor.b64 %rd18700, %rd18695, %rd18604; + xor.b64 %rd18701, %rd18687, %rd18607; + xor.b64 %rd18702, %rd18689, %rd18610; + xor.b64 %rd18703, %rd18691, %rd18613; + xor.b64 %rd18704, %rd18693, %rd18616; + xor.b64 %rd18705, %rd18695, %rd18619; + xor.b64 %rd18706, %rd18687, %rd18622; + xor.b64 %rd18707, %rd18689, %rd18625; + xor.b64 %rd18708, %rd18691, %rd18628; + xor.b64 %rd18709, %rd18693, %rd18631; + xor.b64 %rd18710, %rd18695, %rd18634; + xor.b64 %rd18711, %rd18687, %rd18637; + xor.b64 %rd18712, %rd18689, %rd18640; + xor.b64 %rd18713, %rd18691, %rd18643; + xor.b64 %rd18714, %rd18693, %rd18646; + xor.b64 %rd18715, %rd18695, %rd18649; + xor.b64 %rd18716, %rd18687, %rd18652; + xor.b64 %rd18717, %rd18689, %rd18655; + xor.b64 %rd18718, %rd18691, %rd18658; + xor.b64 %rd18719, %rd18693, %rd18661; + xor.b64 %rd18720, %rd18695, %rd18664; + { + .reg .b32 %dummy; + mov.b64 {%r13331,%dummy}, %rd18702; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13332}, %rd18702; + } + shf.r.wrap.b32 %r13333, %r13332, %r13331, 20; + shf.r.wrap.b32 %r13334, %r13331, %r13332, 20; + mov.b64 %rd18721, {%r13334, %r13333}; + { + .reg .b32 %dummy; + mov.b64 {%r13335,%dummy}, %rd18708; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13336}, %rd18708; + } + shf.r.wrap.b32 %r13337, %r13336, %r13335, 21; + shf.r.wrap.b32 %r13338, %r13335, %r13336, 21; + mov.b64 %rd18722, {%r13338, %r13337}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13339}, %rd18714; + } + { + .reg .b32 %dummy; + mov.b64 {%r13340,%dummy}, %rd18714; + } + shf.l.wrap.b32 %r13341, %r13340, %r13339, 21; + shf.l.wrap.b32 %r13342, %r13339, %r13340, 21; + mov.b64 %rd18723, {%r13342, %r13341}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13343}, %rd18720; + } + { + .reg .b32 %dummy; + mov.b64 {%r13344,%dummy}, %rd18720; + } + shf.l.wrap.b32 %r13345, %r13344, %r13343, 14; + shf.l.wrap.b32 %r13346, %r13343, %r13344, 14; + mov.b64 %rd18724, {%r13346, %r13345}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13347}, %rd18699; + } + { + .reg .b32 %dummy; + mov.b64 {%r13348,%dummy}, %rd18699; + } + shf.l.wrap.b32 %r13349, %r13348, %r13347, 28; + shf.l.wrap.b32 %r13350, %r13347, %r13348, 28; + mov.b64 %rd18725, {%r13350, %r13349}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13351}, %rd18705; + } + { + .reg .b32 %dummy; + mov.b64 {%r13352,%dummy}, %rd18705; + } + shf.l.wrap.b32 %r13353, %r13352, %r13351, 20; + shf.l.wrap.b32 %r13354, %r13351, %r13352, 20; + mov.b64 %rd18726, {%r13354, %r13353}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13355}, %rd18706; + } + { + .reg .b32 %dummy; + mov.b64 {%r13356,%dummy}, %rd18706; + } + shf.l.wrap.b32 %r13357, %r13356, %r13355, 3; + shf.l.wrap.b32 %r13358, %r13355, %r13356, 3; + mov.b64 %rd18727, {%r13358, %r13357}; + { + .reg .b32 %dummy; + mov.b64 {%r13359,%dummy}, %rd18712; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13360}, %rd18712; + } + shf.r.wrap.b32 %r13361, %r13360, %r13359, 19; + shf.r.wrap.b32 %r13362, %r13359, %r13360, 19; + mov.b64 %rd18728, {%r13362, %r13361}; + { + .reg .b32 %dummy; + mov.b64 {%r13363,%dummy}, %rd18718; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13364}, %rd18718; + } + shf.r.wrap.b32 %r13365, %r13364, %r13363, 3; + shf.r.wrap.b32 %r13366, %r13363, %r13364, 3; + mov.b64 %rd18729, {%r13366, %r13365}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13367}, %rd18697; + } + { + .reg .b32 %dummy; + mov.b64 {%r13368,%dummy}, %rd18697; + } + shf.l.wrap.b32 %r13369, %r13368, %r13367, 1; + shf.l.wrap.b32 %r13370, %r13367, %r13368, 1; + mov.b64 %rd18730, {%r13370, %r13369}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13371}, %rd18703; + } + { + .reg .b32 %dummy; + mov.b64 {%r13372,%dummy}, %rd18703; + } + shf.l.wrap.b32 %r13373, %r13372, %r13371, 6; + shf.l.wrap.b32 %r13374, %r13371, %r13372, 6; + mov.b64 %rd18731, {%r13374, %r13373}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13375}, %rd18709; + } + { + .reg .b32 %dummy; + mov.b64 {%r13376,%dummy}, %rd18709; + } + shf.l.wrap.b32 %r13377, %r13376, %r13375, 25; + shf.l.wrap.b32 %r13378, %r13375, %r13376, 25; + mov.b64 %rd18732, {%r13378, %r13377}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13379}, %rd18715; + } + { + .reg .b32 %dummy; + mov.b64 {%r13380,%dummy}, %rd18715; + } + shf.l.wrap.b32 %r13381, %r13380, %r13379, 8; + shf.l.wrap.b32 %r13382, %r13379, %r13380, 8; + mov.b64 %rd18733, {%r13382, %r13381}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13383}, %rd18716; + } + { + .reg .b32 %dummy; + mov.b64 {%r13384,%dummy}, %rd18716; + } + shf.l.wrap.b32 %r13385, %r13384, %r13383, 18; + shf.l.wrap.b32 %r13386, %r13383, %r13384, 18; + mov.b64 %rd18734, {%r13386, %r13385}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13387}, %rd18700; + } + { + .reg .b32 %dummy; + mov.b64 {%r13388,%dummy}, %rd18700; + } + shf.l.wrap.b32 %r13389, %r13388, %r13387, 27; + shf.l.wrap.b32 %r13390, %r13387, %r13388, 27; + mov.b64 %rd18735, {%r13390, %r13389}; + { + .reg .b32 %dummy; + mov.b64 {%r13391,%dummy}, %rd18701; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13392}, %rd18701; + } + shf.r.wrap.b32 %r13393, %r13392, %r13391, 28; + shf.r.wrap.b32 %r13394, %r13391, %r13392, 28; + mov.b64 %rd18736, {%r13394, %r13393}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13395}, %rd18707; + } + { + .reg .b32 %dummy; + mov.b64 {%r13396,%dummy}, %rd18707; + } + shf.l.wrap.b32 %r13397, %r13396, %r13395, 10; + shf.l.wrap.b32 %r13398, %r13395, %r13396, 10; + mov.b64 %rd18737, {%r13398, %r13397}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13399}, %rd18713; + } + { + .reg .b32 %dummy; + mov.b64 {%r13400,%dummy}, %rd18713; + } + shf.l.wrap.b32 %r13401, %r13400, %r13399, 15; + shf.l.wrap.b32 %r13402, %r13399, %r13400, 15; + mov.b64 %rd18738, {%r13402, %r13401}; + { + .reg .b32 %dummy; + mov.b64 {%r13403,%dummy}, %rd18719; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13404}, %rd18719; + } + shf.r.wrap.b32 %r13405, %r13404, %r13403, 8; + shf.r.wrap.b32 %r13406, %r13403, %r13404, 8; + mov.b64 %rd18739, {%r13406, %r13405}; + { + .reg .b32 %dummy; + mov.b64 {%r13407,%dummy}, %rd18698; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13408}, %rd18698; + } + shf.r.wrap.b32 %r13409, %r13408, %r13407, 2; + shf.r.wrap.b32 %r13410, %r13407, %r13408, 2; + mov.b64 %rd18740, {%r13410, %r13409}; + { + .reg .b32 %dummy; + mov.b64 {%r13411,%dummy}, %rd18704; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13412}, %rd18704; + } + shf.r.wrap.b32 %r13413, %r13412, %r13411, 9; + shf.r.wrap.b32 %r13414, %r13411, %r13412, 9; + mov.b64 %rd18741, {%r13414, %r13413}; + { + .reg .b32 %dummy; + mov.b64 {%r13415,%dummy}, %rd18710; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13416}, %rd18710; + } + shf.r.wrap.b32 %r13417, %r13416, %r13415, 25; + shf.r.wrap.b32 %r13418, %r13415, %r13416, 25; + mov.b64 %rd18742, {%r13418, %r13417}; + { + .reg .b32 %dummy; + mov.b64 {%r13419,%dummy}, %rd18711; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13420}, %rd18711; + } + shf.r.wrap.b32 %r13421, %r13420, %r13419, 23; + shf.r.wrap.b32 %r13422, %r13419, %r13420, 23; + mov.b64 %rd18743, {%r13422, %r13421}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13423}, %rd18717; + } + { + .reg .b32 %dummy; + mov.b64 {%r13424,%dummy}, %rd18717; + } + shf.l.wrap.b32 %r13425, %r13424, %r13423, 2; + shf.l.wrap.b32 %r13426, %r13423, %r13424, 2; + mov.b64 %rd18744, {%r13426, %r13425}; + not.b64 %rd18745, %rd18721; + and.b64 %rd18746, %rd18722, %rd18745; + xor.b64 %rd18747, %rd18746, %rd18696; + not.b64 %rd18748, %rd18722; + and.b64 %rd18749, %rd18723, %rd18748; + xor.b64 %rd18750, %rd18749, %rd18721; + not.b64 %rd18751, %rd18723; + and.b64 %rd18752, %rd18724, %rd18751; + xor.b64 %rd18753, %rd18752, %rd18722; + not.b64 %rd18754, %rd18724; + and.b64 %rd18755, %rd18696, %rd18754; + xor.b64 %rd18756, %rd18755, %rd18723; + not.b64 %rd18757, %rd18696; + and.b64 %rd18758, %rd18721, %rd18757; + xor.b64 %rd18759, %rd18758, %rd18724; + not.b64 %rd18760, %rd18726; + and.b64 %rd18761, %rd18727, %rd18760; + xor.b64 %rd18762, %rd18761, %rd18725; + not.b64 %rd18763, %rd18727; + and.b64 %rd18764, %rd18728, %rd18763; + xor.b64 %rd18765, %rd18764, %rd18726; + not.b64 %rd18766, %rd18728; + and.b64 %rd18767, %rd18729, %rd18766; + xor.b64 %rd18768, %rd18767, %rd18727; + not.b64 %rd18769, %rd18729; + and.b64 %rd18770, %rd18725, %rd18769; + xor.b64 %rd18771, %rd18770, %rd18728; + not.b64 %rd18772, %rd18725; + and.b64 %rd18773, %rd18726, %rd18772; + xor.b64 %rd18774, %rd18773, %rd18729; + not.b64 %rd18775, %rd18731; + and.b64 %rd18776, %rd18732, %rd18775; + xor.b64 %rd18777, %rd18776, %rd18730; + not.b64 %rd18778, %rd18732; + and.b64 %rd18779, %rd18733, %rd18778; + xor.b64 %rd18780, %rd18779, %rd18731; + not.b64 %rd18781, %rd18733; + and.b64 %rd18782, %rd18734, %rd18781; + xor.b64 %rd18783, %rd18782, %rd18732; + not.b64 %rd18784, %rd18734; + and.b64 %rd18785, %rd18730, %rd18784; + xor.b64 %rd18786, %rd18785, %rd18733; + not.b64 %rd18787, %rd18730; + and.b64 %rd18788, %rd18731, %rd18787; + xor.b64 %rd18789, %rd18788, %rd18734; + not.b64 %rd18790, %rd18736; + and.b64 %rd18791, %rd18737, %rd18790; + xor.b64 %rd18792, %rd18791, %rd18735; + not.b64 %rd18793, %rd18737; + and.b64 %rd18794, %rd18738, %rd18793; + xor.b64 %rd18795, %rd18794, %rd18736; + not.b64 %rd18796, %rd18738; + and.b64 %rd18797, %rd18739, %rd18796; + xor.b64 %rd18798, %rd18797, %rd18737; + not.b64 %rd18799, %rd18739; + and.b64 %rd18800, %rd18735, %rd18799; + xor.b64 %rd18801, %rd18800, %rd18738; + not.b64 %rd18802, %rd18735; + and.b64 %rd18803, %rd18736, %rd18802; + xor.b64 %rd18804, %rd18803, %rd18739; + not.b64 %rd18805, %rd18741; + and.b64 %rd18806, %rd18742, %rd18805; + xor.b64 %rd18807, %rd18806, %rd18740; + not.b64 %rd18808, %rd18742; + and.b64 %rd18809, %rd18743, %rd18808; + xor.b64 %rd18810, %rd18809, %rd18741; + not.b64 %rd18811, %rd18743; + and.b64 %rd18812, %rd18744, %rd18811; + xor.b64 %rd18813, %rd18812, %rd18742; + not.b64 %rd18814, %rd18744; + and.b64 %rd18815, %rd18740, %rd18814; + xor.b64 %rd18816, %rd18815, %rd18743; + not.b64 %rd18817, %rd18740; + and.b64 %rd18818, %rd18741, %rd18817; + xor.b64 %rd18819, %rd18818, %rd18744; + xor.b64 %rd18820, %rd18747, %rd34; + xor.b64 %rd18821, %rd18762, %rd18820; + xor.b64 %rd18822, %rd18821, %rd18777; + xor.b64 %rd18823, %rd18822, %rd18792; + xor.b64 %rd18824, %rd18823, %rd18807; + xor.b64 %rd18825, %rd18765, %rd18750; + xor.b64 %rd18826, %rd18825, %rd18780; + xor.b64 %rd18827, %rd18826, %rd18795; + xor.b64 %rd18828, %rd18827, %rd18810; + xor.b64 %rd18829, %rd18768, %rd18753; + xor.b64 %rd18830, %rd18829, %rd18783; + xor.b64 %rd18831, %rd18830, %rd18798; + xor.b64 %rd18832, %rd18831, %rd18813; + xor.b64 %rd18833, %rd18771, %rd18756; + xor.b64 %rd18834, %rd18833, %rd18786; + xor.b64 %rd18835, %rd18834, %rd18801; + xor.b64 %rd18836, %rd18835, %rd18816; + xor.b64 %rd18837, %rd18774, %rd18759; + xor.b64 %rd18838, %rd18837, %rd18789; + xor.b64 %rd18839, %rd18838, %rd18804; + xor.b64 %rd18840, %rd18839, %rd18819; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13427}, %rd18828; + } + { + .reg .b32 %dummy; + mov.b64 {%r13428,%dummy}, %rd18828; + } + shf.l.wrap.b32 %r13429, %r13428, %r13427, 1; + shf.l.wrap.b32 %r13430, %r13427, %r13428, 1; + mov.b64 %rd18841, {%r13430, %r13429}; + xor.b64 %rd18842, %rd18840, %rd18841; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13431}, %rd18832; + } + { + .reg .b32 %dummy; + mov.b64 {%r13432,%dummy}, %rd18832; + } + shf.l.wrap.b32 %r13433, %r13432, %r13431, 1; + shf.l.wrap.b32 %r13434, %r13431, %r13432, 1; + mov.b64 %rd18843, {%r13434, %r13433}; + xor.b64 %rd18844, %rd18843, %rd18824; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13435}, %rd18836; + } + { + .reg .b32 %dummy; + mov.b64 {%r13436,%dummy}, %rd18836; + } + shf.l.wrap.b32 %r13437, %r13436, %r13435, 1; + shf.l.wrap.b32 %r13438, %r13435, %r13436, 1; + mov.b64 %rd18845, {%r13438, %r13437}; + xor.b64 %rd18846, %rd18845, %rd18828; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13439}, %rd18840; + } + { + .reg .b32 %dummy; + mov.b64 {%r13440,%dummy}, %rd18840; + } + shf.l.wrap.b32 %r13441, %r13440, %r13439, 1; + shf.l.wrap.b32 %r13442, %r13439, %r13440, 1; + mov.b64 %rd18847, {%r13442, %r13441}; + xor.b64 %rd18848, %rd18847, %rd18832; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13443}, %rd18824; + } + { + .reg .b32 %dummy; + mov.b64 {%r13444,%dummy}, %rd18824; + } + shf.l.wrap.b32 %r13445, %r13444, %r13443, 1; + shf.l.wrap.b32 %r13446, %r13443, %r13444, 1; + mov.b64 %rd18849, {%r13446, %r13445}; + xor.b64 %rd18850, %rd18836, %rd18849; + xor.b64 %rd18851, %rd18842, %rd18820; + xor.b64 %rd18852, %rd18844, %rd18750; + xor.b64 %rd18853, %rd18846, %rd18753; + xor.b64 %rd18854, %rd18848, %rd18756; + xor.b64 %rd18855, %rd18850, %rd18759; + xor.b64 %rd18856, %rd18842, %rd18762; + xor.b64 %rd18857, %rd18844, %rd18765; + xor.b64 %rd18858, %rd18846, %rd18768; + xor.b64 %rd18859, %rd18848, %rd18771; + xor.b64 %rd18860, %rd18850, %rd18774; + xor.b64 %rd18861, %rd18842, %rd18777; + xor.b64 %rd18862, %rd18844, %rd18780; + xor.b64 %rd18863, %rd18846, %rd18783; + xor.b64 %rd18864, %rd18848, %rd18786; + xor.b64 %rd18865, %rd18850, %rd18789; + xor.b64 %rd18866, %rd18842, %rd18792; + xor.b64 %rd18867, %rd18844, %rd18795; + xor.b64 %rd18868, %rd18846, %rd18798; + xor.b64 %rd18869, %rd18848, %rd18801; + xor.b64 %rd18870, %rd18850, %rd18804; + xor.b64 %rd18871, %rd18842, %rd18807; + xor.b64 %rd18872, %rd18844, %rd18810; + xor.b64 %rd18873, %rd18846, %rd18813; + xor.b64 %rd18874, %rd18848, %rd18816; + xor.b64 %rd18875, %rd18850, %rd18819; + { + .reg .b32 %dummy; + mov.b64 {%r13447,%dummy}, %rd18857; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13448}, %rd18857; + } + shf.r.wrap.b32 %r13449, %r13448, %r13447, 20; + shf.r.wrap.b32 %r13450, %r13447, %r13448, 20; + mov.b64 %rd18876, {%r13450, %r13449}; + { + .reg .b32 %dummy; + mov.b64 {%r13451,%dummy}, %rd18863; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13452}, %rd18863; + } + shf.r.wrap.b32 %r13453, %r13452, %r13451, 21; + shf.r.wrap.b32 %r13454, %r13451, %r13452, 21; + mov.b64 %rd18877, {%r13454, %r13453}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13455}, %rd18869; + } + { + .reg .b32 %dummy; + mov.b64 {%r13456,%dummy}, %rd18869; + } + shf.l.wrap.b32 %r13457, %r13456, %r13455, 21; + shf.l.wrap.b32 %r13458, %r13455, %r13456, 21; + mov.b64 %rd18878, {%r13458, %r13457}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13459}, %rd18875; + } + { + .reg .b32 %dummy; + mov.b64 {%r13460,%dummy}, %rd18875; + } + shf.l.wrap.b32 %r13461, %r13460, %r13459, 14; + shf.l.wrap.b32 %r13462, %r13459, %r13460, 14; + mov.b64 %rd18879, {%r13462, %r13461}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13463}, %rd18854; + } + { + .reg .b32 %dummy; + mov.b64 {%r13464,%dummy}, %rd18854; + } + shf.l.wrap.b32 %r13465, %r13464, %r13463, 28; + shf.l.wrap.b32 %r13466, %r13463, %r13464, 28; + mov.b64 %rd18880, {%r13466, %r13465}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13467}, %rd18860; + } + { + .reg .b32 %dummy; + mov.b64 {%r13468,%dummy}, %rd18860; + } + shf.l.wrap.b32 %r13469, %r13468, %r13467, 20; + shf.l.wrap.b32 %r13470, %r13467, %r13468, 20; + mov.b64 %rd18881, {%r13470, %r13469}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13471}, %rd18861; + } + { + .reg .b32 %dummy; + mov.b64 {%r13472,%dummy}, %rd18861; + } + shf.l.wrap.b32 %r13473, %r13472, %r13471, 3; + shf.l.wrap.b32 %r13474, %r13471, %r13472, 3; + mov.b64 %rd18882, {%r13474, %r13473}; + { + .reg .b32 %dummy; + mov.b64 {%r13475,%dummy}, %rd18867; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13476}, %rd18867; + } + shf.r.wrap.b32 %r13477, %r13476, %r13475, 19; + shf.r.wrap.b32 %r13478, %r13475, %r13476, 19; + mov.b64 %rd18883, {%r13478, %r13477}; + { + .reg .b32 %dummy; + mov.b64 {%r13479,%dummy}, %rd18873; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13480}, %rd18873; + } + shf.r.wrap.b32 %r13481, %r13480, %r13479, 3; + shf.r.wrap.b32 %r13482, %r13479, %r13480, 3; + mov.b64 %rd18884, {%r13482, %r13481}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13483}, %rd18852; + } + { + .reg .b32 %dummy; + mov.b64 {%r13484,%dummy}, %rd18852; + } + shf.l.wrap.b32 %r13485, %r13484, %r13483, 1; + shf.l.wrap.b32 %r13486, %r13483, %r13484, 1; + mov.b64 %rd18885, {%r13486, %r13485}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13487}, %rd18858; + } + { + .reg .b32 %dummy; + mov.b64 {%r13488,%dummy}, %rd18858; + } + shf.l.wrap.b32 %r13489, %r13488, %r13487, 6; + shf.l.wrap.b32 %r13490, %r13487, %r13488, 6; + mov.b64 %rd18886, {%r13490, %r13489}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13491}, %rd18864; + } + { + .reg .b32 %dummy; + mov.b64 {%r13492,%dummy}, %rd18864; + } + shf.l.wrap.b32 %r13493, %r13492, %r13491, 25; + shf.l.wrap.b32 %r13494, %r13491, %r13492, 25; + mov.b64 %rd18887, {%r13494, %r13493}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13495}, %rd18870; + } + { + .reg .b32 %dummy; + mov.b64 {%r13496,%dummy}, %rd18870; + } + shf.l.wrap.b32 %r13497, %r13496, %r13495, 8; + shf.l.wrap.b32 %r13498, %r13495, %r13496, 8; + mov.b64 %rd18888, {%r13498, %r13497}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13499}, %rd18871; + } + { + .reg .b32 %dummy; + mov.b64 {%r13500,%dummy}, %rd18871; + } + shf.l.wrap.b32 %r13501, %r13500, %r13499, 18; + shf.l.wrap.b32 %r13502, %r13499, %r13500, 18; + mov.b64 %rd18889, {%r13502, %r13501}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13503}, %rd18855; + } + { + .reg .b32 %dummy; + mov.b64 {%r13504,%dummy}, %rd18855; + } + shf.l.wrap.b32 %r13505, %r13504, %r13503, 27; + shf.l.wrap.b32 %r13506, %r13503, %r13504, 27; + mov.b64 %rd18890, {%r13506, %r13505}; + { + .reg .b32 %dummy; + mov.b64 {%r13507,%dummy}, %rd18856; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13508}, %rd18856; + } + shf.r.wrap.b32 %r13509, %r13508, %r13507, 28; + shf.r.wrap.b32 %r13510, %r13507, %r13508, 28; + mov.b64 %rd18891, {%r13510, %r13509}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13511}, %rd18862; + } + { + .reg .b32 %dummy; + mov.b64 {%r13512,%dummy}, %rd18862; + } + shf.l.wrap.b32 %r13513, %r13512, %r13511, 10; + shf.l.wrap.b32 %r13514, %r13511, %r13512, 10; + mov.b64 %rd18892, {%r13514, %r13513}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13515}, %rd18868; + } + { + .reg .b32 %dummy; + mov.b64 {%r13516,%dummy}, %rd18868; + } + shf.l.wrap.b32 %r13517, %r13516, %r13515, 15; + shf.l.wrap.b32 %r13518, %r13515, %r13516, 15; + mov.b64 %rd18893, {%r13518, %r13517}; + { + .reg .b32 %dummy; + mov.b64 {%r13519,%dummy}, %rd18874; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13520}, %rd18874; + } + shf.r.wrap.b32 %r13521, %r13520, %r13519, 8; + shf.r.wrap.b32 %r13522, %r13519, %r13520, 8; + mov.b64 %rd18894, {%r13522, %r13521}; + { + .reg .b32 %dummy; + mov.b64 {%r13523,%dummy}, %rd18853; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13524}, %rd18853; + } + shf.r.wrap.b32 %r13525, %r13524, %r13523, 2; + shf.r.wrap.b32 %r13526, %r13523, %r13524, 2; + mov.b64 %rd18895, {%r13526, %r13525}; + { + .reg .b32 %dummy; + mov.b64 {%r13527,%dummy}, %rd18859; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13528}, %rd18859; + } + shf.r.wrap.b32 %r13529, %r13528, %r13527, 9; + shf.r.wrap.b32 %r13530, %r13527, %r13528, 9; + mov.b64 %rd18896, {%r13530, %r13529}; + { + .reg .b32 %dummy; + mov.b64 {%r13531,%dummy}, %rd18865; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13532}, %rd18865; + } + shf.r.wrap.b32 %r13533, %r13532, %r13531, 25; + shf.r.wrap.b32 %r13534, %r13531, %r13532, 25; + mov.b64 %rd18897, {%r13534, %r13533}; + { + .reg .b32 %dummy; + mov.b64 {%r13535,%dummy}, %rd18866; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13536}, %rd18866; + } + shf.r.wrap.b32 %r13537, %r13536, %r13535, 23; + shf.r.wrap.b32 %r13538, %r13535, %r13536, 23; + mov.b64 %rd18898, {%r13538, %r13537}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13539}, %rd18872; + } + { + .reg .b32 %dummy; + mov.b64 {%r13540,%dummy}, %rd18872; + } + shf.l.wrap.b32 %r13541, %r13540, %r13539, 2; + shf.l.wrap.b32 %r13542, %r13539, %r13540, 2; + mov.b64 %rd18899, {%r13542, %r13541}; + not.b64 %rd18900, %rd18876; + and.b64 %rd18901, %rd18877, %rd18900; + xor.b64 %rd18902, %rd18901, %rd18851; + not.b64 %rd18903, %rd18877; + and.b64 %rd18904, %rd18878, %rd18903; + xor.b64 %rd18905, %rd18904, %rd18876; + not.b64 %rd18906, %rd18878; + and.b64 %rd18907, %rd18879, %rd18906; + xor.b64 %rd18908, %rd18907, %rd18877; + not.b64 %rd18909, %rd18879; + and.b64 %rd18910, %rd18851, %rd18909; + xor.b64 %rd18911, %rd18910, %rd18878; + not.b64 %rd18912, %rd18851; + and.b64 %rd18913, %rd18876, %rd18912; + xor.b64 %rd18914, %rd18913, %rd18879; + not.b64 %rd18915, %rd18881; + and.b64 %rd18916, %rd18882, %rd18915; + xor.b64 %rd18917, %rd18916, %rd18880; + not.b64 %rd18918, %rd18882; + and.b64 %rd18919, %rd18883, %rd18918; + xor.b64 %rd18920, %rd18919, %rd18881; + not.b64 %rd18921, %rd18883; + and.b64 %rd18922, %rd18884, %rd18921; + xor.b64 %rd18923, %rd18922, %rd18882; + not.b64 %rd18924, %rd18884; + and.b64 %rd18925, %rd18880, %rd18924; + xor.b64 %rd18926, %rd18925, %rd18883; + not.b64 %rd18927, %rd18880; + and.b64 %rd18928, %rd18881, %rd18927; + xor.b64 %rd18929, %rd18928, %rd18884; + not.b64 %rd18930, %rd18886; + and.b64 %rd18931, %rd18887, %rd18930; + xor.b64 %rd18932, %rd18931, %rd18885; + not.b64 %rd18933, %rd18887; + and.b64 %rd18934, %rd18888, %rd18933; + xor.b64 %rd18935, %rd18934, %rd18886; + not.b64 %rd18936, %rd18888; + and.b64 %rd18937, %rd18889, %rd18936; + xor.b64 %rd18938, %rd18937, %rd18887; + not.b64 %rd18939, %rd18889; + and.b64 %rd18940, %rd18885, %rd18939; + xor.b64 %rd18941, %rd18940, %rd18888; + not.b64 %rd18942, %rd18885; + and.b64 %rd18943, %rd18886, %rd18942; + xor.b64 %rd18944, %rd18943, %rd18889; + not.b64 %rd18945, %rd18891; + and.b64 %rd18946, %rd18892, %rd18945; + xor.b64 %rd18947, %rd18946, %rd18890; + not.b64 %rd18948, %rd18892; + and.b64 %rd18949, %rd18893, %rd18948; + xor.b64 %rd18950, %rd18949, %rd18891; + not.b64 %rd18951, %rd18893; + and.b64 %rd18952, %rd18894, %rd18951; + xor.b64 %rd18953, %rd18952, %rd18892; + not.b64 %rd18954, %rd18894; + and.b64 %rd18955, %rd18890, %rd18954; + xor.b64 %rd18956, %rd18955, %rd18893; + not.b64 %rd18957, %rd18890; + and.b64 %rd18958, %rd18891, %rd18957; + xor.b64 %rd18959, %rd18958, %rd18894; + not.b64 %rd18960, %rd18896; + and.b64 %rd18961, %rd18897, %rd18960; + xor.b64 %rd18962, %rd18961, %rd18895; + not.b64 %rd18963, %rd18897; + and.b64 %rd18964, %rd18898, %rd18963; + xor.b64 %rd18965, %rd18964, %rd18896; + not.b64 %rd18966, %rd18898; + and.b64 %rd18967, %rd18899, %rd18966; + xor.b64 %rd18968, %rd18967, %rd18897; + not.b64 %rd18969, %rd18899; + and.b64 %rd18970, %rd18895, %rd18969; + xor.b64 %rd18971, %rd18970, %rd18898; + not.b64 %rd18972, %rd18895; + and.b64 %rd18973, %rd18896, %rd18972; + xor.b64 %rd18974, %rd18973, %rd18899; + xor.b64 %rd18975, %rd18902, %rd35; + xor.b64 %rd18976, %rd18917, %rd18975; + xor.b64 %rd18977, %rd18976, %rd18932; + xor.b64 %rd18978, %rd18977, %rd18947; + xor.b64 %rd18979, %rd18978, %rd18962; + xor.b64 %rd18980, %rd18920, %rd18905; + xor.b64 %rd18981, %rd18980, %rd18935; + xor.b64 %rd18982, %rd18981, %rd18950; + xor.b64 %rd18983, %rd18982, %rd18965; + xor.b64 %rd18984, %rd18923, %rd18908; + xor.b64 %rd18985, %rd18984, %rd18938; + xor.b64 %rd18986, %rd18985, %rd18953; + xor.b64 %rd18987, %rd18986, %rd18968; + xor.b64 %rd18988, %rd18926, %rd18911; + xor.b64 %rd18989, %rd18988, %rd18941; + xor.b64 %rd18990, %rd18989, %rd18956; + xor.b64 %rd18991, %rd18990, %rd18971; + xor.b64 %rd18992, %rd18929, %rd18914; + xor.b64 %rd18993, %rd18992, %rd18944; + xor.b64 %rd18994, %rd18993, %rd18959; + xor.b64 %rd18995, %rd18994, %rd18974; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13543}, %rd18983; + } + { + .reg .b32 %dummy; + mov.b64 {%r13544,%dummy}, %rd18983; + } + shf.l.wrap.b32 %r13545, %r13544, %r13543, 1; + shf.l.wrap.b32 %r13546, %r13543, %r13544, 1; + mov.b64 %rd18996, {%r13546, %r13545}; + xor.b64 %rd18997, %rd18995, %rd18996; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13547}, %rd18987; + } + { + .reg .b32 %dummy; + mov.b64 {%r13548,%dummy}, %rd18987; + } + shf.l.wrap.b32 %r13549, %r13548, %r13547, 1; + shf.l.wrap.b32 %r13550, %r13547, %r13548, 1; + mov.b64 %rd18998, {%r13550, %r13549}; + xor.b64 %rd18999, %rd18998, %rd18979; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13551}, %rd18991; + } + { + .reg .b32 %dummy; + mov.b64 {%r13552,%dummy}, %rd18991; + } + shf.l.wrap.b32 %r13553, %r13552, %r13551, 1; + shf.l.wrap.b32 %r13554, %r13551, %r13552, 1; + mov.b64 %rd19000, {%r13554, %r13553}; + xor.b64 %rd19001, %rd19000, %rd18983; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13555}, %rd18995; + } + { + .reg .b32 %dummy; + mov.b64 {%r13556,%dummy}, %rd18995; + } + shf.l.wrap.b32 %r13557, %r13556, %r13555, 1; + shf.l.wrap.b32 %r13558, %r13555, %r13556, 1; + mov.b64 %rd19002, {%r13558, %r13557}; + xor.b64 %rd19003, %rd19002, %rd18987; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13559}, %rd18979; + } + { + .reg .b32 %dummy; + mov.b64 {%r13560,%dummy}, %rd18979; + } + shf.l.wrap.b32 %r13561, %r13560, %r13559, 1; + shf.l.wrap.b32 %r13562, %r13559, %r13560, 1; + mov.b64 %rd19004, {%r13562, %r13561}; + xor.b64 %rd19005, %rd18991, %rd19004; + xor.b64 %rd19006, %rd18997, %rd18975; + xor.b64 %rd19007, %rd18999, %rd18905; + xor.b64 %rd19008, %rd19001, %rd18908; + xor.b64 %rd19009, %rd19003, %rd18911; + xor.b64 %rd19010, %rd19005, %rd18914; + xor.b64 %rd19011, %rd18997, %rd18917; + xor.b64 %rd19012, %rd18999, %rd18920; + xor.b64 %rd19013, %rd19001, %rd18923; + xor.b64 %rd19014, %rd19003, %rd18926; + xor.b64 %rd19015, %rd19005, %rd18929; + xor.b64 %rd19016, %rd18997, %rd18932; + xor.b64 %rd19017, %rd18999, %rd18935; + xor.b64 %rd19018, %rd19001, %rd18938; + xor.b64 %rd19019, %rd19003, %rd18941; + xor.b64 %rd19020, %rd19005, %rd18944; + xor.b64 %rd19021, %rd18997, %rd18947; + xor.b64 %rd19022, %rd18999, %rd18950; + xor.b64 %rd19023, %rd19001, %rd18953; + xor.b64 %rd19024, %rd19003, %rd18956; + xor.b64 %rd19025, %rd19005, %rd18959; + xor.b64 %rd19026, %rd18997, %rd18962; + xor.b64 %rd19027, %rd18999, %rd18965; + xor.b64 %rd19028, %rd19001, %rd18968; + xor.b64 %rd19029, %rd19003, %rd18971; + xor.b64 %rd19030, %rd19005, %rd18974; + { + .reg .b32 %dummy; + mov.b64 {%r13563,%dummy}, %rd19012; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13564}, %rd19012; + } + shf.r.wrap.b32 %r13565, %r13564, %r13563, 20; + shf.r.wrap.b32 %r13566, %r13563, %r13564, 20; + mov.b64 %rd19031, {%r13566, %r13565}; + { + .reg .b32 %dummy; + mov.b64 {%r13567,%dummy}, %rd19018; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13568}, %rd19018; + } + shf.r.wrap.b32 %r13569, %r13568, %r13567, 21; + shf.r.wrap.b32 %r13570, %r13567, %r13568, 21; + mov.b64 %rd19032, {%r13570, %r13569}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13571}, %rd19024; + } + { + .reg .b32 %dummy; + mov.b64 {%r13572,%dummy}, %rd19024; + } + shf.l.wrap.b32 %r13573, %r13572, %r13571, 21; + shf.l.wrap.b32 %r13574, %r13571, %r13572, 21; + mov.b64 %rd19033, {%r13574, %r13573}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13575}, %rd19030; + } + { + .reg .b32 %dummy; + mov.b64 {%r13576,%dummy}, %rd19030; + } + shf.l.wrap.b32 %r13577, %r13576, %r13575, 14; + shf.l.wrap.b32 %r13578, %r13575, %r13576, 14; + mov.b64 %rd19034, {%r13578, %r13577}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13579}, %rd19009; + } + { + .reg .b32 %dummy; + mov.b64 {%r13580,%dummy}, %rd19009; + } + shf.l.wrap.b32 %r13581, %r13580, %r13579, 28; + shf.l.wrap.b32 %r13582, %r13579, %r13580, 28; + mov.b64 %rd19035, {%r13582, %r13581}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13583}, %rd19015; + } + { + .reg .b32 %dummy; + mov.b64 {%r13584,%dummy}, %rd19015; + } + shf.l.wrap.b32 %r13585, %r13584, %r13583, 20; + shf.l.wrap.b32 %r13586, %r13583, %r13584, 20; + mov.b64 %rd19036, {%r13586, %r13585}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13587}, %rd19016; + } + { + .reg .b32 %dummy; + mov.b64 {%r13588,%dummy}, %rd19016; + } + shf.l.wrap.b32 %r13589, %r13588, %r13587, 3; + shf.l.wrap.b32 %r13590, %r13587, %r13588, 3; + mov.b64 %rd19037, {%r13590, %r13589}; + { + .reg .b32 %dummy; + mov.b64 {%r13591,%dummy}, %rd19022; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13592}, %rd19022; + } + shf.r.wrap.b32 %r13593, %r13592, %r13591, 19; + shf.r.wrap.b32 %r13594, %r13591, %r13592, 19; + mov.b64 %rd19038, {%r13594, %r13593}; + { + .reg .b32 %dummy; + mov.b64 {%r13595,%dummy}, %rd19028; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13596}, %rd19028; + } + shf.r.wrap.b32 %r13597, %r13596, %r13595, 3; + shf.r.wrap.b32 %r13598, %r13595, %r13596, 3; + mov.b64 %rd19039, {%r13598, %r13597}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13599}, %rd19007; + } + { + .reg .b32 %dummy; + mov.b64 {%r13600,%dummy}, %rd19007; + } + shf.l.wrap.b32 %r13601, %r13600, %r13599, 1; + shf.l.wrap.b32 %r13602, %r13599, %r13600, 1; + mov.b64 %rd19040, {%r13602, %r13601}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13603}, %rd19013; + } + { + .reg .b32 %dummy; + mov.b64 {%r13604,%dummy}, %rd19013; + } + shf.l.wrap.b32 %r13605, %r13604, %r13603, 6; + shf.l.wrap.b32 %r13606, %r13603, %r13604, 6; + mov.b64 %rd19041, {%r13606, %r13605}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13607}, %rd19019; + } + { + .reg .b32 %dummy; + mov.b64 {%r13608,%dummy}, %rd19019; + } + shf.l.wrap.b32 %r13609, %r13608, %r13607, 25; + shf.l.wrap.b32 %r13610, %r13607, %r13608, 25; + mov.b64 %rd19042, {%r13610, %r13609}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13611}, %rd19025; + } + { + .reg .b32 %dummy; + mov.b64 {%r13612,%dummy}, %rd19025; + } + shf.l.wrap.b32 %r13613, %r13612, %r13611, 8; + shf.l.wrap.b32 %r13614, %r13611, %r13612, 8; + mov.b64 %rd19043, {%r13614, %r13613}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13615}, %rd19026; + } + { + .reg .b32 %dummy; + mov.b64 {%r13616,%dummy}, %rd19026; + } + shf.l.wrap.b32 %r13617, %r13616, %r13615, 18; + shf.l.wrap.b32 %r13618, %r13615, %r13616, 18; + mov.b64 %rd19044, {%r13618, %r13617}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13619}, %rd19010; + } + { + .reg .b32 %dummy; + mov.b64 {%r13620,%dummy}, %rd19010; + } + shf.l.wrap.b32 %r13621, %r13620, %r13619, 27; + shf.l.wrap.b32 %r13622, %r13619, %r13620, 27; + mov.b64 %rd19045, {%r13622, %r13621}; + { + .reg .b32 %dummy; + mov.b64 {%r13623,%dummy}, %rd19011; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13624}, %rd19011; + } + shf.r.wrap.b32 %r13625, %r13624, %r13623, 28; + shf.r.wrap.b32 %r13626, %r13623, %r13624, 28; + mov.b64 %rd19046, {%r13626, %r13625}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13627}, %rd19017; + } + { + .reg .b32 %dummy; + mov.b64 {%r13628,%dummy}, %rd19017; + } + shf.l.wrap.b32 %r13629, %r13628, %r13627, 10; + shf.l.wrap.b32 %r13630, %r13627, %r13628, 10; + mov.b64 %rd19047, {%r13630, %r13629}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13631}, %rd19023; + } + { + .reg .b32 %dummy; + mov.b64 {%r13632,%dummy}, %rd19023; + } + shf.l.wrap.b32 %r13633, %r13632, %r13631, 15; + shf.l.wrap.b32 %r13634, %r13631, %r13632, 15; + mov.b64 %rd19048, {%r13634, %r13633}; + { + .reg .b32 %dummy; + mov.b64 {%r13635,%dummy}, %rd19029; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13636}, %rd19029; + } + shf.r.wrap.b32 %r13637, %r13636, %r13635, 8; + shf.r.wrap.b32 %r13638, %r13635, %r13636, 8; + mov.b64 %rd19049, {%r13638, %r13637}; + { + .reg .b32 %dummy; + mov.b64 {%r13639,%dummy}, %rd19008; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13640}, %rd19008; + } + shf.r.wrap.b32 %r13641, %r13640, %r13639, 2; + shf.r.wrap.b32 %r13642, %r13639, %r13640, 2; + mov.b64 %rd19050, {%r13642, %r13641}; + { + .reg .b32 %dummy; + mov.b64 {%r13643,%dummy}, %rd19014; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13644}, %rd19014; + } + shf.r.wrap.b32 %r13645, %r13644, %r13643, 9; + shf.r.wrap.b32 %r13646, %r13643, %r13644, 9; + mov.b64 %rd19051, {%r13646, %r13645}; + { + .reg .b32 %dummy; + mov.b64 {%r13647,%dummy}, %rd19020; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13648}, %rd19020; + } + shf.r.wrap.b32 %r13649, %r13648, %r13647, 25; + shf.r.wrap.b32 %r13650, %r13647, %r13648, 25; + mov.b64 %rd19052, {%r13650, %r13649}; + { + .reg .b32 %dummy; + mov.b64 {%r13651,%dummy}, %rd19021; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13652}, %rd19021; + } + shf.r.wrap.b32 %r13653, %r13652, %r13651, 23; + shf.r.wrap.b32 %r13654, %r13651, %r13652, 23; + mov.b64 %rd19053, {%r13654, %r13653}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13655}, %rd19027; + } + { + .reg .b32 %dummy; + mov.b64 {%r13656,%dummy}, %rd19027; + } + shf.l.wrap.b32 %r13657, %r13656, %r13655, 2; + shf.l.wrap.b32 %r13658, %r13655, %r13656, 2; + mov.b64 %rd19054, {%r13658, %r13657}; + not.b64 %rd19055, %rd19031; + and.b64 %rd19056, %rd19032, %rd19055; + xor.b64 %rd19057, %rd19056, %rd19006; + not.b64 %rd19058, %rd19032; + and.b64 %rd19059, %rd19033, %rd19058; + xor.b64 %rd19060, %rd19059, %rd19031; + not.b64 %rd19061, %rd19033; + and.b64 %rd19062, %rd19034, %rd19061; + xor.b64 %rd19063, %rd19062, %rd19032; + not.b64 %rd19064, %rd19034; + and.b64 %rd19065, %rd19006, %rd19064; + xor.b64 %rd19066, %rd19065, %rd19033; + not.b64 %rd19067, %rd19006; + and.b64 %rd19068, %rd19031, %rd19067; + xor.b64 %rd19069, %rd19068, %rd19034; + not.b64 %rd19070, %rd19036; + and.b64 %rd19071, %rd19037, %rd19070; + xor.b64 %rd19072, %rd19071, %rd19035; + not.b64 %rd19073, %rd19037; + and.b64 %rd19074, %rd19038, %rd19073; + xor.b64 %rd19075, %rd19074, %rd19036; + not.b64 %rd19076, %rd19038; + and.b64 %rd19077, %rd19039, %rd19076; + xor.b64 %rd19078, %rd19077, %rd19037; + not.b64 %rd19079, %rd19039; + and.b64 %rd19080, %rd19035, %rd19079; + xor.b64 %rd19081, %rd19080, %rd19038; + not.b64 %rd19082, %rd19035; + and.b64 %rd19083, %rd19036, %rd19082; + xor.b64 %rd19084, %rd19083, %rd19039; + not.b64 %rd19085, %rd19041; + and.b64 %rd19086, %rd19042, %rd19085; + xor.b64 %rd19087, %rd19086, %rd19040; + not.b64 %rd19088, %rd19042; + and.b64 %rd19089, %rd19043, %rd19088; + xor.b64 %rd19090, %rd19089, %rd19041; + not.b64 %rd19091, %rd19043; + and.b64 %rd19092, %rd19044, %rd19091; + xor.b64 %rd19093, %rd19092, %rd19042; + not.b64 %rd19094, %rd19044; + and.b64 %rd19095, %rd19040, %rd19094; + xor.b64 %rd19096, %rd19095, %rd19043; + not.b64 %rd19097, %rd19040; + and.b64 %rd19098, %rd19041, %rd19097; + xor.b64 %rd19099, %rd19098, %rd19044; + not.b64 %rd19100, %rd19046; + and.b64 %rd19101, %rd19047, %rd19100; + xor.b64 %rd19102, %rd19101, %rd19045; + not.b64 %rd19103, %rd19047; + and.b64 %rd19104, %rd19048, %rd19103; + xor.b64 %rd19105, %rd19104, %rd19046; + not.b64 %rd19106, %rd19048; + and.b64 %rd19107, %rd19049, %rd19106; + xor.b64 %rd19108, %rd19107, %rd19047; + not.b64 %rd19109, %rd19049; + and.b64 %rd19110, %rd19045, %rd19109; + xor.b64 %rd19111, %rd19110, %rd19048; + not.b64 %rd19112, %rd19045; + and.b64 %rd19113, %rd19046, %rd19112; + xor.b64 %rd19114, %rd19113, %rd19049; + not.b64 %rd19115, %rd19051; + and.b64 %rd19116, %rd19052, %rd19115; + xor.b64 %rd19117, %rd19116, %rd19050; + not.b64 %rd19118, %rd19052; + and.b64 %rd19119, %rd19053, %rd19118; + xor.b64 %rd19120, %rd19119, %rd19051; + not.b64 %rd19121, %rd19053; + and.b64 %rd19122, %rd19054, %rd19121; + xor.b64 %rd19123, %rd19122, %rd19052; + not.b64 %rd19124, %rd19054; + and.b64 %rd19125, %rd19050, %rd19124; + xor.b64 %rd19126, %rd19125, %rd19053; + not.b64 %rd19127, %rd19050; + and.b64 %rd19128, %rd19051, %rd19127; + xor.b64 %rd19129, %rd19128, %rd19054; + xor.b64 %rd19130, %rd19057, %rd36; + xor.b64 %rd19131, %rd19072, %rd19130; + xor.b64 %rd19132, %rd19131, %rd19087; + xor.b64 %rd19133, %rd19132, %rd19102; + xor.b64 %rd19134, %rd19133, %rd19117; + xor.b64 %rd19135, %rd19075, %rd19060; + xor.b64 %rd19136, %rd19135, %rd19090; + xor.b64 %rd19137, %rd19136, %rd19105; + xor.b64 %rd19138, %rd19137, %rd19120; + xor.b64 %rd19139, %rd19078, %rd19063; + xor.b64 %rd19140, %rd19139, %rd19093; + xor.b64 %rd19141, %rd19140, %rd19108; + xor.b64 %rd19142, %rd19141, %rd19123; + xor.b64 %rd19143, %rd19081, %rd19066; + xor.b64 %rd19144, %rd19143, %rd19096; + xor.b64 %rd19145, %rd19144, %rd19111; + xor.b64 %rd19146, %rd19145, %rd19126; + xor.b64 %rd19147, %rd19084, %rd19069; + xor.b64 %rd19148, %rd19147, %rd19099; + xor.b64 %rd19149, %rd19148, %rd19114; + xor.b64 %rd19150, %rd19149, %rd19129; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13659}, %rd19138; + } + { + .reg .b32 %dummy; + mov.b64 {%r13660,%dummy}, %rd19138; + } + shf.l.wrap.b32 %r13661, %r13660, %r13659, 1; + shf.l.wrap.b32 %r13662, %r13659, %r13660, 1; + mov.b64 %rd19151, {%r13662, %r13661}; + xor.b64 %rd19152, %rd19150, %rd19151; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13663}, %rd19142; + } + { + .reg .b32 %dummy; + mov.b64 {%r13664,%dummy}, %rd19142; + } + shf.l.wrap.b32 %r13665, %r13664, %r13663, 1; + shf.l.wrap.b32 %r13666, %r13663, %r13664, 1; + mov.b64 %rd19153, {%r13666, %r13665}; + xor.b64 %rd19154, %rd19153, %rd19134; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13667}, %rd19146; + } + { + .reg .b32 %dummy; + mov.b64 {%r13668,%dummy}, %rd19146; + } + shf.l.wrap.b32 %r13669, %r13668, %r13667, 1; + shf.l.wrap.b32 %r13670, %r13667, %r13668, 1; + mov.b64 %rd19155, {%r13670, %r13669}; + xor.b64 %rd19156, %rd19155, %rd19138; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13671}, %rd19150; + } + { + .reg .b32 %dummy; + mov.b64 {%r13672,%dummy}, %rd19150; + } + shf.l.wrap.b32 %r13673, %r13672, %r13671, 1; + shf.l.wrap.b32 %r13674, %r13671, %r13672, 1; + mov.b64 %rd19157, {%r13674, %r13673}; + xor.b64 %rd19158, %rd19157, %rd19142; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13675}, %rd19134; + } + { + .reg .b32 %dummy; + mov.b64 {%r13676,%dummy}, %rd19134; + } + shf.l.wrap.b32 %r13677, %r13676, %r13675, 1; + shf.l.wrap.b32 %r13678, %r13675, %r13676, 1; + mov.b64 %rd19159, {%r13678, %r13677}; + xor.b64 %rd19160, %rd19146, %rd19159; + xor.b64 %rd19161, %rd19152, %rd19130; + xor.b64 %rd19162, %rd19154, %rd19060; + xor.b64 %rd19163, %rd19156, %rd19063; + xor.b64 %rd19164, %rd19158, %rd19066; + xor.b64 %rd19165, %rd19160, %rd19069; + xor.b64 %rd19166, %rd19152, %rd19072; + xor.b64 %rd19167, %rd19154, %rd19075; + xor.b64 %rd19168, %rd19156, %rd19078; + xor.b64 %rd19169, %rd19158, %rd19081; + xor.b64 %rd19170, %rd19160, %rd19084; + xor.b64 %rd19171, %rd19152, %rd19087; + xor.b64 %rd19172, %rd19154, %rd19090; + xor.b64 %rd19173, %rd19156, %rd19093; + xor.b64 %rd19174, %rd19158, %rd19096; + xor.b64 %rd19175, %rd19160, %rd19099; + xor.b64 %rd19176, %rd19152, %rd19102; + xor.b64 %rd19177, %rd19154, %rd19105; + xor.b64 %rd19178, %rd19156, %rd19108; + xor.b64 %rd19179, %rd19158, %rd19111; + xor.b64 %rd19180, %rd19160, %rd19114; + xor.b64 %rd19181, %rd19152, %rd19117; + xor.b64 %rd19182, %rd19154, %rd19120; + xor.b64 %rd19183, %rd19156, %rd19123; + xor.b64 %rd19184, %rd19158, %rd19126; + xor.b64 %rd19185, %rd19160, %rd19129; + { + .reg .b32 %dummy; + mov.b64 {%r13679,%dummy}, %rd19167; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13680}, %rd19167; + } + shf.r.wrap.b32 %r13681, %r13680, %r13679, 20; + shf.r.wrap.b32 %r13682, %r13679, %r13680, 20; + mov.b64 %rd19186, {%r13682, %r13681}; + { + .reg .b32 %dummy; + mov.b64 {%r13683,%dummy}, %rd19173; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13684}, %rd19173; + } + shf.r.wrap.b32 %r13685, %r13684, %r13683, 21; + shf.r.wrap.b32 %r13686, %r13683, %r13684, 21; + mov.b64 %rd19187, {%r13686, %r13685}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13687}, %rd19179; + } + { + .reg .b32 %dummy; + mov.b64 {%r13688,%dummy}, %rd19179; + } + shf.l.wrap.b32 %r13689, %r13688, %r13687, 21; + shf.l.wrap.b32 %r13690, %r13687, %r13688, 21; + mov.b64 %rd19188, {%r13690, %r13689}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13691}, %rd19185; + } + { + .reg .b32 %dummy; + mov.b64 {%r13692,%dummy}, %rd19185; + } + shf.l.wrap.b32 %r13693, %r13692, %r13691, 14; + shf.l.wrap.b32 %r13694, %r13691, %r13692, 14; + mov.b64 %rd19189, {%r13694, %r13693}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13695}, %rd19164; + } + { + .reg .b32 %dummy; + mov.b64 {%r13696,%dummy}, %rd19164; + } + shf.l.wrap.b32 %r13697, %r13696, %r13695, 28; + shf.l.wrap.b32 %r13698, %r13695, %r13696, 28; + mov.b64 %rd19190, {%r13698, %r13697}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13699}, %rd19170; + } + { + .reg .b32 %dummy; + mov.b64 {%r13700,%dummy}, %rd19170; + } + shf.l.wrap.b32 %r13701, %r13700, %r13699, 20; + shf.l.wrap.b32 %r13702, %r13699, %r13700, 20; + mov.b64 %rd19191, {%r13702, %r13701}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13703}, %rd19171; + } + { + .reg .b32 %dummy; + mov.b64 {%r13704,%dummy}, %rd19171; + } + shf.l.wrap.b32 %r13705, %r13704, %r13703, 3; + shf.l.wrap.b32 %r13706, %r13703, %r13704, 3; + mov.b64 %rd19192, {%r13706, %r13705}; + { + .reg .b32 %dummy; + mov.b64 {%r13707,%dummy}, %rd19177; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13708}, %rd19177; + } + shf.r.wrap.b32 %r13709, %r13708, %r13707, 19; + shf.r.wrap.b32 %r13710, %r13707, %r13708, 19; + mov.b64 %rd19193, {%r13710, %r13709}; + { + .reg .b32 %dummy; + mov.b64 {%r13711,%dummy}, %rd19183; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13712}, %rd19183; + } + shf.r.wrap.b32 %r13713, %r13712, %r13711, 3; + shf.r.wrap.b32 %r13714, %r13711, %r13712, 3; + mov.b64 %rd19194, {%r13714, %r13713}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13715}, %rd19162; + } + { + .reg .b32 %dummy; + mov.b64 {%r13716,%dummy}, %rd19162; + } + shf.l.wrap.b32 %r13717, %r13716, %r13715, 1; + shf.l.wrap.b32 %r13718, %r13715, %r13716, 1; + mov.b64 %rd19195, {%r13718, %r13717}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13719}, %rd19168; + } + { + .reg .b32 %dummy; + mov.b64 {%r13720,%dummy}, %rd19168; + } + shf.l.wrap.b32 %r13721, %r13720, %r13719, 6; + shf.l.wrap.b32 %r13722, %r13719, %r13720, 6; + mov.b64 %rd19196, {%r13722, %r13721}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13723}, %rd19174; + } + { + .reg .b32 %dummy; + mov.b64 {%r13724,%dummy}, %rd19174; + } + shf.l.wrap.b32 %r13725, %r13724, %r13723, 25; + shf.l.wrap.b32 %r13726, %r13723, %r13724, 25; + mov.b64 %rd19197, {%r13726, %r13725}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13727}, %rd19180; + } + { + .reg .b32 %dummy; + mov.b64 {%r13728,%dummy}, %rd19180; + } + shf.l.wrap.b32 %r13729, %r13728, %r13727, 8; + shf.l.wrap.b32 %r13730, %r13727, %r13728, 8; + mov.b64 %rd19198, {%r13730, %r13729}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13731}, %rd19181; + } + { + .reg .b32 %dummy; + mov.b64 {%r13732,%dummy}, %rd19181; + } + shf.l.wrap.b32 %r13733, %r13732, %r13731, 18; + shf.l.wrap.b32 %r13734, %r13731, %r13732, 18; + mov.b64 %rd19199, {%r13734, %r13733}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13735}, %rd19165; + } + { + .reg .b32 %dummy; + mov.b64 {%r13736,%dummy}, %rd19165; + } + shf.l.wrap.b32 %r13737, %r13736, %r13735, 27; + shf.l.wrap.b32 %r13738, %r13735, %r13736, 27; + mov.b64 %rd19200, {%r13738, %r13737}; + { + .reg .b32 %dummy; + mov.b64 {%r13739,%dummy}, %rd19166; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13740}, %rd19166; + } + shf.r.wrap.b32 %r13741, %r13740, %r13739, 28; + shf.r.wrap.b32 %r13742, %r13739, %r13740, 28; + mov.b64 %rd19201, {%r13742, %r13741}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13743}, %rd19172; + } + { + .reg .b32 %dummy; + mov.b64 {%r13744,%dummy}, %rd19172; + } + shf.l.wrap.b32 %r13745, %r13744, %r13743, 10; + shf.l.wrap.b32 %r13746, %r13743, %r13744, 10; + mov.b64 %rd19202, {%r13746, %r13745}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13747}, %rd19178; + } + { + .reg .b32 %dummy; + mov.b64 {%r13748,%dummy}, %rd19178; + } + shf.l.wrap.b32 %r13749, %r13748, %r13747, 15; + shf.l.wrap.b32 %r13750, %r13747, %r13748, 15; + mov.b64 %rd19203, {%r13750, %r13749}; + { + .reg .b32 %dummy; + mov.b64 {%r13751,%dummy}, %rd19184; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13752}, %rd19184; + } + shf.r.wrap.b32 %r13753, %r13752, %r13751, 8; + shf.r.wrap.b32 %r13754, %r13751, %r13752, 8; + mov.b64 %rd19204, {%r13754, %r13753}; + { + .reg .b32 %dummy; + mov.b64 {%r13755,%dummy}, %rd19163; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13756}, %rd19163; + } + shf.r.wrap.b32 %r13757, %r13756, %r13755, 2; + shf.r.wrap.b32 %r13758, %r13755, %r13756, 2; + mov.b64 %rd19205, {%r13758, %r13757}; + { + .reg .b32 %dummy; + mov.b64 {%r13759,%dummy}, %rd19169; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13760}, %rd19169; + } + shf.r.wrap.b32 %r13761, %r13760, %r13759, 9; + shf.r.wrap.b32 %r13762, %r13759, %r13760, 9; + mov.b64 %rd19206, {%r13762, %r13761}; + { + .reg .b32 %dummy; + mov.b64 {%r13763,%dummy}, %rd19175; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13764}, %rd19175; + } + shf.r.wrap.b32 %r13765, %r13764, %r13763, 25; + shf.r.wrap.b32 %r13766, %r13763, %r13764, 25; + mov.b64 %rd19207, {%r13766, %r13765}; + { + .reg .b32 %dummy; + mov.b64 {%r13767,%dummy}, %rd19176; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13768}, %rd19176; + } + shf.r.wrap.b32 %r13769, %r13768, %r13767, 23; + shf.r.wrap.b32 %r13770, %r13767, %r13768, 23; + mov.b64 %rd19208, {%r13770, %r13769}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13771}, %rd19182; + } + { + .reg .b32 %dummy; + mov.b64 {%r13772,%dummy}, %rd19182; + } + shf.l.wrap.b32 %r13773, %r13772, %r13771, 2; + shf.l.wrap.b32 %r13774, %r13771, %r13772, 2; + mov.b64 %rd19209, {%r13774, %r13773}; + not.b64 %rd19210, %rd19186; + and.b64 %rd19211, %rd19187, %rd19210; + xor.b64 %rd19212, %rd19211, %rd19161; + not.b64 %rd19213, %rd19187; + and.b64 %rd19214, %rd19188, %rd19213; + xor.b64 %rd19215, %rd19214, %rd19186; + not.b64 %rd19216, %rd19188; + and.b64 %rd19217, %rd19189, %rd19216; + xor.b64 %rd19218, %rd19217, %rd19187; + not.b64 %rd19219, %rd19189; + and.b64 %rd19220, %rd19161, %rd19219; + xor.b64 %rd19221, %rd19220, %rd19188; + not.b64 %rd19222, %rd19161; + and.b64 %rd19223, %rd19186, %rd19222; + xor.b64 %rd19224, %rd19223, %rd19189; + not.b64 %rd19225, %rd19191; + and.b64 %rd19226, %rd19192, %rd19225; + xor.b64 %rd19227, %rd19226, %rd19190; + not.b64 %rd19228, %rd19192; + and.b64 %rd19229, %rd19193, %rd19228; + xor.b64 %rd19230, %rd19229, %rd19191; + not.b64 %rd19231, %rd19193; + and.b64 %rd19232, %rd19194, %rd19231; + xor.b64 %rd19233, %rd19232, %rd19192; + not.b64 %rd19234, %rd19194; + and.b64 %rd19235, %rd19190, %rd19234; + xor.b64 %rd19236, %rd19235, %rd19193; + not.b64 %rd19237, %rd19190; + and.b64 %rd19238, %rd19191, %rd19237; + xor.b64 %rd19239, %rd19238, %rd19194; + not.b64 %rd19240, %rd19196; + and.b64 %rd19241, %rd19197, %rd19240; + xor.b64 %rd19242, %rd19241, %rd19195; + not.b64 %rd19243, %rd19197; + and.b64 %rd19244, %rd19198, %rd19243; + xor.b64 %rd19245, %rd19244, %rd19196; + not.b64 %rd19246, %rd19198; + and.b64 %rd19247, %rd19199, %rd19246; + xor.b64 %rd19248, %rd19247, %rd19197; + not.b64 %rd19249, %rd19199; + and.b64 %rd19250, %rd19195, %rd19249; + xor.b64 %rd19251, %rd19250, %rd19198; + not.b64 %rd19252, %rd19195; + and.b64 %rd19253, %rd19196, %rd19252; + xor.b64 %rd19254, %rd19253, %rd19199; + not.b64 %rd19255, %rd19201; + and.b64 %rd19256, %rd19202, %rd19255; + xor.b64 %rd19257, %rd19256, %rd19200; + not.b64 %rd19258, %rd19202; + and.b64 %rd19259, %rd19203, %rd19258; + xor.b64 %rd19260, %rd19259, %rd19201; + not.b64 %rd19261, %rd19203; + and.b64 %rd19262, %rd19204, %rd19261; + xor.b64 %rd19263, %rd19262, %rd19202; + not.b64 %rd19264, %rd19204; + and.b64 %rd19265, %rd19200, %rd19264; + xor.b64 %rd19266, %rd19265, %rd19203; + not.b64 %rd19267, %rd19200; + and.b64 %rd19268, %rd19201, %rd19267; + xor.b64 %rd19269, %rd19268, %rd19204; + not.b64 %rd19270, %rd19206; + and.b64 %rd19271, %rd19207, %rd19270; + xor.b64 %rd19272, %rd19271, %rd19205; + not.b64 %rd19273, %rd19207; + and.b64 %rd19274, %rd19208, %rd19273; + xor.b64 %rd19275, %rd19274, %rd19206; + not.b64 %rd19276, %rd19208; + and.b64 %rd19277, %rd19209, %rd19276; + xor.b64 %rd19278, %rd19277, %rd19207; + not.b64 %rd19279, %rd19209; + and.b64 %rd19280, %rd19205, %rd19279; + xor.b64 %rd19281, %rd19280, %rd19208; + not.b64 %rd19282, %rd19205; + and.b64 %rd19283, %rd19206, %rd19282; + xor.b64 %rd19284, %rd19283, %rd19209; + xor.b64 %rd19285, %rd19212, %rd37; + xor.b64 %rd19286, %rd19227, %rd19285; + xor.b64 %rd19287, %rd19286, %rd19242; + xor.b64 %rd19288, %rd19287, %rd19257; + xor.b64 %rd19289, %rd19288, %rd19272; + xor.b64 %rd19290, %rd19230, %rd19215; + xor.b64 %rd19291, %rd19290, %rd19245; + xor.b64 %rd19292, %rd19291, %rd19260; + xor.b64 %rd19293, %rd19292, %rd19275; + xor.b64 %rd19294, %rd19233, %rd19218; + xor.b64 %rd19295, %rd19294, %rd19248; + xor.b64 %rd19296, %rd19295, %rd19263; + xor.b64 %rd19297, %rd19296, %rd19278; + xor.b64 %rd19298, %rd19236, %rd19221; + xor.b64 %rd19299, %rd19298, %rd19251; + xor.b64 %rd19300, %rd19299, %rd19266; + xor.b64 %rd19301, %rd19300, %rd19281; + xor.b64 %rd19302, %rd19239, %rd19224; + xor.b64 %rd19303, %rd19302, %rd19254; + xor.b64 %rd19304, %rd19303, %rd19269; + xor.b64 %rd19305, %rd19304, %rd19284; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13775}, %rd19293; + } + { + .reg .b32 %dummy; + mov.b64 {%r13776,%dummy}, %rd19293; + } + shf.l.wrap.b32 %r13777, %r13776, %r13775, 1; + shf.l.wrap.b32 %r13778, %r13775, %r13776, 1; + mov.b64 %rd19306, {%r13778, %r13777}; + xor.b64 %rd19307, %rd19305, %rd19306; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13779}, %rd19297; + } + { + .reg .b32 %dummy; + mov.b64 {%r13780,%dummy}, %rd19297; + } + shf.l.wrap.b32 %r13781, %r13780, %r13779, 1; + shf.l.wrap.b32 %r13782, %r13779, %r13780, 1; + mov.b64 %rd19308, {%r13782, %r13781}; + xor.b64 %rd19309, %rd19308, %rd19289; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13783}, %rd19301; + } + { + .reg .b32 %dummy; + mov.b64 {%r13784,%dummy}, %rd19301; + } + shf.l.wrap.b32 %r13785, %r13784, %r13783, 1; + shf.l.wrap.b32 %r13786, %r13783, %r13784, 1; + mov.b64 %rd19310, {%r13786, %r13785}; + xor.b64 %rd19311, %rd19310, %rd19293; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13787}, %rd19305; + } + { + .reg .b32 %dummy; + mov.b64 {%r13788,%dummy}, %rd19305; + } + shf.l.wrap.b32 %r13789, %r13788, %r13787, 1; + shf.l.wrap.b32 %r13790, %r13787, %r13788, 1; + mov.b64 %rd19312, {%r13790, %r13789}; + xor.b64 %rd19313, %rd19312, %rd19297; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13791}, %rd19289; + } + { + .reg .b32 %dummy; + mov.b64 {%r13792,%dummy}, %rd19289; + } + shf.l.wrap.b32 %r13793, %r13792, %r13791, 1; + shf.l.wrap.b32 %r13794, %r13791, %r13792, 1; + mov.b64 %rd19314, {%r13794, %r13793}; + xor.b64 %rd19315, %rd19301, %rd19314; + xor.b64 %rd19316, %rd19307, %rd19285; + xor.b64 %rd19317, %rd19309, %rd19215; + xor.b64 %rd19318, %rd19311, %rd19218; + xor.b64 %rd19319, %rd19313, %rd19221; + xor.b64 %rd19320, %rd19315, %rd19224; + xor.b64 %rd19321, %rd19307, %rd19227; + xor.b64 %rd19322, %rd19309, %rd19230; + xor.b64 %rd19323, %rd19311, %rd19233; + xor.b64 %rd19324, %rd19313, %rd19236; + xor.b64 %rd19325, %rd19315, %rd19239; + xor.b64 %rd19326, %rd19307, %rd19242; + xor.b64 %rd19327, %rd19309, %rd19245; + xor.b64 %rd19328, %rd19311, %rd19248; + xor.b64 %rd19329, %rd19313, %rd19251; + xor.b64 %rd19330, %rd19315, %rd19254; + xor.b64 %rd19331, %rd19307, %rd19257; + xor.b64 %rd19332, %rd19309, %rd19260; + xor.b64 %rd19333, %rd19311, %rd19263; + xor.b64 %rd19334, %rd19313, %rd19266; + xor.b64 %rd19335, %rd19315, %rd19269; + xor.b64 %rd19336, %rd19307, %rd19272; + xor.b64 %rd19337, %rd19309, %rd19275; + xor.b64 %rd19338, %rd19311, %rd19278; + xor.b64 %rd19339, %rd19313, %rd19281; + xor.b64 %rd19340, %rd19315, %rd19284; + { + .reg .b32 %dummy; + mov.b64 {%r13795,%dummy}, %rd19322; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13796}, %rd19322; + } + shf.r.wrap.b32 %r13797, %r13796, %r13795, 20; + shf.r.wrap.b32 %r13798, %r13795, %r13796, 20; + mov.b64 %rd19341, {%r13798, %r13797}; + { + .reg .b32 %dummy; + mov.b64 {%r13799,%dummy}, %rd19328; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13800}, %rd19328; + } + shf.r.wrap.b32 %r13801, %r13800, %r13799, 21; + shf.r.wrap.b32 %r13802, %r13799, %r13800, 21; + mov.b64 %rd19342, {%r13802, %r13801}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13803}, %rd19334; + } + { + .reg .b32 %dummy; + mov.b64 {%r13804,%dummy}, %rd19334; + } + shf.l.wrap.b32 %r13805, %r13804, %r13803, 21; + shf.l.wrap.b32 %r13806, %r13803, %r13804, 21; + mov.b64 %rd19343, {%r13806, %r13805}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13807}, %rd19340; + } + { + .reg .b32 %dummy; + mov.b64 {%r13808,%dummy}, %rd19340; + } + shf.l.wrap.b32 %r13809, %r13808, %r13807, 14; + shf.l.wrap.b32 %r13810, %r13807, %r13808, 14; + mov.b64 %rd19344, {%r13810, %r13809}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13811}, %rd19319; + } + { + .reg .b32 %dummy; + mov.b64 {%r13812,%dummy}, %rd19319; + } + shf.l.wrap.b32 %r13813, %r13812, %r13811, 28; + shf.l.wrap.b32 %r13814, %r13811, %r13812, 28; + mov.b64 %rd19345, {%r13814, %r13813}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13815}, %rd19325; + } + { + .reg .b32 %dummy; + mov.b64 {%r13816,%dummy}, %rd19325; + } + shf.l.wrap.b32 %r13817, %r13816, %r13815, 20; + shf.l.wrap.b32 %r13818, %r13815, %r13816, 20; + mov.b64 %rd19346, {%r13818, %r13817}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13819}, %rd19326; + } + { + .reg .b32 %dummy; + mov.b64 {%r13820,%dummy}, %rd19326; + } + shf.l.wrap.b32 %r13821, %r13820, %r13819, 3; + shf.l.wrap.b32 %r13822, %r13819, %r13820, 3; + mov.b64 %rd19347, {%r13822, %r13821}; + { + .reg .b32 %dummy; + mov.b64 {%r13823,%dummy}, %rd19332; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13824}, %rd19332; + } + shf.r.wrap.b32 %r13825, %r13824, %r13823, 19; + shf.r.wrap.b32 %r13826, %r13823, %r13824, 19; + mov.b64 %rd19348, {%r13826, %r13825}; + { + .reg .b32 %dummy; + mov.b64 {%r13827,%dummy}, %rd19338; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13828}, %rd19338; + } + shf.r.wrap.b32 %r13829, %r13828, %r13827, 3; + shf.r.wrap.b32 %r13830, %r13827, %r13828, 3; + mov.b64 %rd19349, {%r13830, %r13829}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13831}, %rd19317; + } + { + .reg .b32 %dummy; + mov.b64 {%r13832,%dummy}, %rd19317; + } + shf.l.wrap.b32 %r13833, %r13832, %r13831, 1; + shf.l.wrap.b32 %r13834, %r13831, %r13832, 1; + mov.b64 %rd19350, {%r13834, %r13833}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13835}, %rd19323; + } + { + .reg .b32 %dummy; + mov.b64 {%r13836,%dummy}, %rd19323; + } + shf.l.wrap.b32 %r13837, %r13836, %r13835, 6; + shf.l.wrap.b32 %r13838, %r13835, %r13836, 6; + mov.b64 %rd19351, {%r13838, %r13837}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13839}, %rd19329; + } + { + .reg .b32 %dummy; + mov.b64 {%r13840,%dummy}, %rd19329; + } + shf.l.wrap.b32 %r13841, %r13840, %r13839, 25; + shf.l.wrap.b32 %r13842, %r13839, %r13840, 25; + mov.b64 %rd19352, {%r13842, %r13841}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13843}, %rd19335; + } + { + .reg .b32 %dummy; + mov.b64 {%r13844,%dummy}, %rd19335; + } + shf.l.wrap.b32 %r13845, %r13844, %r13843, 8; + shf.l.wrap.b32 %r13846, %r13843, %r13844, 8; + mov.b64 %rd19353, {%r13846, %r13845}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13847}, %rd19336; + } + { + .reg .b32 %dummy; + mov.b64 {%r13848,%dummy}, %rd19336; + } + shf.l.wrap.b32 %r13849, %r13848, %r13847, 18; + shf.l.wrap.b32 %r13850, %r13847, %r13848, 18; + mov.b64 %rd19354, {%r13850, %r13849}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13851}, %rd19320; + } + { + .reg .b32 %dummy; + mov.b64 {%r13852,%dummy}, %rd19320; + } + shf.l.wrap.b32 %r13853, %r13852, %r13851, 27; + shf.l.wrap.b32 %r13854, %r13851, %r13852, 27; + mov.b64 %rd19355, {%r13854, %r13853}; + { + .reg .b32 %dummy; + mov.b64 {%r13855,%dummy}, %rd19321; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13856}, %rd19321; + } + shf.r.wrap.b32 %r13857, %r13856, %r13855, 28; + shf.r.wrap.b32 %r13858, %r13855, %r13856, 28; + mov.b64 %rd19356, {%r13858, %r13857}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13859}, %rd19327; + } + { + .reg .b32 %dummy; + mov.b64 {%r13860,%dummy}, %rd19327; + } + shf.l.wrap.b32 %r13861, %r13860, %r13859, 10; + shf.l.wrap.b32 %r13862, %r13859, %r13860, 10; + mov.b64 %rd19357, {%r13862, %r13861}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13863}, %rd19333; + } + { + .reg .b32 %dummy; + mov.b64 {%r13864,%dummy}, %rd19333; + } + shf.l.wrap.b32 %r13865, %r13864, %r13863, 15; + shf.l.wrap.b32 %r13866, %r13863, %r13864, 15; + mov.b64 %rd19358, {%r13866, %r13865}; + { + .reg .b32 %dummy; + mov.b64 {%r13867,%dummy}, %rd19339; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13868}, %rd19339; + } + shf.r.wrap.b32 %r13869, %r13868, %r13867, 8; + shf.r.wrap.b32 %r13870, %r13867, %r13868, 8; + mov.b64 %rd19359, {%r13870, %r13869}; + { + .reg .b32 %dummy; + mov.b64 {%r13871,%dummy}, %rd19318; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13872}, %rd19318; + } + shf.r.wrap.b32 %r13873, %r13872, %r13871, 2; + shf.r.wrap.b32 %r13874, %r13871, %r13872, 2; + mov.b64 %rd19360, {%r13874, %r13873}; + { + .reg .b32 %dummy; + mov.b64 {%r13875,%dummy}, %rd19324; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13876}, %rd19324; + } + shf.r.wrap.b32 %r13877, %r13876, %r13875, 9; + shf.r.wrap.b32 %r13878, %r13875, %r13876, 9; + mov.b64 %rd19361, {%r13878, %r13877}; + { + .reg .b32 %dummy; + mov.b64 {%r13879,%dummy}, %rd19330; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13880}, %rd19330; + } + shf.r.wrap.b32 %r13881, %r13880, %r13879, 25; + shf.r.wrap.b32 %r13882, %r13879, %r13880, 25; + mov.b64 %rd19362, {%r13882, %r13881}; + { + .reg .b32 %dummy; + mov.b64 {%r13883,%dummy}, %rd19331; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13884}, %rd19331; + } + shf.r.wrap.b32 %r13885, %r13884, %r13883, 23; + shf.r.wrap.b32 %r13886, %r13883, %r13884, 23; + mov.b64 %rd19363, {%r13886, %r13885}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13887}, %rd19337; + } + { + .reg .b32 %dummy; + mov.b64 {%r13888,%dummy}, %rd19337; + } + shf.l.wrap.b32 %r13889, %r13888, %r13887, 2; + shf.l.wrap.b32 %r13890, %r13887, %r13888, 2; + mov.b64 %rd19364, {%r13890, %r13889}; + not.b64 %rd19365, %rd19341; + and.b64 %rd19366, %rd19342, %rd19365; + xor.b64 %rd19367, %rd19366, %rd19316; + not.b64 %rd19368, %rd19342; + and.b64 %rd19369, %rd19343, %rd19368; + xor.b64 %rd19370, %rd19369, %rd19341; + not.b64 %rd19371, %rd19343; + and.b64 %rd19372, %rd19344, %rd19371; + xor.b64 %rd19373, %rd19372, %rd19342; + not.b64 %rd19374, %rd19344; + and.b64 %rd19375, %rd19316, %rd19374; + xor.b64 %rd19376, %rd19375, %rd19343; + not.b64 %rd19377, %rd19316; + and.b64 %rd19378, %rd19341, %rd19377; + xor.b64 %rd19379, %rd19378, %rd19344; + not.b64 %rd19380, %rd19346; + and.b64 %rd19381, %rd19347, %rd19380; + xor.b64 %rd19382, %rd19381, %rd19345; + not.b64 %rd19383, %rd19347; + and.b64 %rd19384, %rd19348, %rd19383; + xor.b64 %rd19385, %rd19384, %rd19346; + not.b64 %rd19386, %rd19348; + and.b64 %rd19387, %rd19349, %rd19386; + xor.b64 %rd19388, %rd19387, %rd19347; + not.b64 %rd19389, %rd19349; + and.b64 %rd19390, %rd19345, %rd19389; + xor.b64 %rd19391, %rd19390, %rd19348; + not.b64 %rd19392, %rd19345; + and.b64 %rd19393, %rd19346, %rd19392; + xor.b64 %rd19394, %rd19393, %rd19349; + not.b64 %rd19395, %rd19351; + and.b64 %rd19396, %rd19352, %rd19395; + xor.b64 %rd19397, %rd19396, %rd19350; + not.b64 %rd19398, %rd19352; + and.b64 %rd19399, %rd19353, %rd19398; + xor.b64 %rd19400, %rd19399, %rd19351; + not.b64 %rd19401, %rd19353; + and.b64 %rd19402, %rd19354, %rd19401; + xor.b64 %rd19403, %rd19402, %rd19352; + not.b64 %rd19404, %rd19354; + and.b64 %rd19405, %rd19350, %rd19404; + xor.b64 %rd19406, %rd19405, %rd19353; + not.b64 %rd19407, %rd19350; + and.b64 %rd19408, %rd19351, %rd19407; + xor.b64 %rd19409, %rd19408, %rd19354; + not.b64 %rd19410, %rd19356; + and.b64 %rd19411, %rd19357, %rd19410; + xor.b64 %rd19412, %rd19411, %rd19355; + not.b64 %rd19413, %rd19357; + and.b64 %rd19414, %rd19358, %rd19413; + xor.b64 %rd19415, %rd19414, %rd19356; + not.b64 %rd19416, %rd19358; + and.b64 %rd19417, %rd19359, %rd19416; + xor.b64 %rd19418, %rd19417, %rd19357; + not.b64 %rd19419, %rd19359; + and.b64 %rd19420, %rd19355, %rd19419; + xor.b64 %rd19421, %rd19420, %rd19358; + not.b64 %rd19422, %rd19355; + and.b64 %rd19423, %rd19356, %rd19422; + xor.b64 %rd19424, %rd19423, %rd19359; + not.b64 %rd19425, %rd19361; + and.b64 %rd19426, %rd19362, %rd19425; + xor.b64 %rd19427, %rd19426, %rd19360; + not.b64 %rd19428, %rd19362; + and.b64 %rd19429, %rd19363, %rd19428; + xor.b64 %rd19430, %rd19429, %rd19361; + not.b64 %rd19431, %rd19363; + and.b64 %rd19432, %rd19364, %rd19431; + xor.b64 %rd19433, %rd19432, %rd19362; + not.b64 %rd19434, %rd19364; + and.b64 %rd19435, %rd19360, %rd19434; + xor.b64 %rd19436, %rd19435, %rd19363; + not.b64 %rd19437, %rd19360; + and.b64 %rd19438, %rd19361, %rd19437; + xor.b64 %rd19439, %rd19438, %rd19364; + xor.b64 %rd19440, %rd19367, %rd38; + xor.b64 %rd19441, %rd19382, %rd19440; + xor.b64 %rd19442, %rd19441, %rd19397; + xor.b64 %rd19443, %rd19442, %rd19412; + xor.b64 %rd19444, %rd19443, %rd19427; + xor.b64 %rd19445, %rd19385, %rd19370; + xor.b64 %rd19446, %rd19445, %rd19400; + xor.b64 %rd19447, %rd19446, %rd19415; + xor.b64 %rd19448, %rd19447, %rd19430; + xor.b64 %rd19449, %rd19388, %rd19373; + xor.b64 %rd19450, %rd19449, %rd19403; + xor.b64 %rd19451, %rd19450, %rd19418; + xor.b64 %rd19452, %rd19451, %rd19433; + xor.b64 %rd19453, %rd19391, %rd19376; + xor.b64 %rd19454, %rd19453, %rd19406; + xor.b64 %rd19455, %rd19454, %rd19421; + xor.b64 %rd19456, %rd19455, %rd19436; + xor.b64 %rd19457, %rd19394, %rd19379; + xor.b64 %rd19458, %rd19457, %rd19409; + xor.b64 %rd19459, %rd19458, %rd19424; + xor.b64 %rd19460, %rd19459, %rd19439; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13891}, %rd19448; + } + { + .reg .b32 %dummy; + mov.b64 {%r13892,%dummy}, %rd19448; + } + shf.l.wrap.b32 %r13893, %r13892, %r13891, 1; + shf.l.wrap.b32 %r13894, %r13891, %r13892, 1; + mov.b64 %rd19461, {%r13894, %r13893}; + xor.b64 %rd19462, %rd19460, %rd19461; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13895}, %rd19452; + } + { + .reg .b32 %dummy; + mov.b64 {%r13896,%dummy}, %rd19452; + } + shf.l.wrap.b32 %r13897, %r13896, %r13895, 1; + shf.l.wrap.b32 %r13898, %r13895, %r13896, 1; + mov.b64 %rd19463, {%r13898, %r13897}; + xor.b64 %rd19464, %rd19463, %rd19444; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13899}, %rd19456; + } + { + .reg .b32 %dummy; + mov.b64 {%r13900,%dummy}, %rd19456; + } + shf.l.wrap.b32 %r13901, %r13900, %r13899, 1; + shf.l.wrap.b32 %r13902, %r13899, %r13900, 1; + mov.b64 %rd19465, {%r13902, %r13901}; + xor.b64 %rd19466, %rd19465, %rd19448; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13903}, %rd19460; + } + { + .reg .b32 %dummy; + mov.b64 {%r13904,%dummy}, %rd19460; + } + shf.l.wrap.b32 %r13905, %r13904, %r13903, 1; + shf.l.wrap.b32 %r13906, %r13903, %r13904, 1; + mov.b64 %rd19467, {%r13906, %r13905}; + xor.b64 %rd19468, %rd19467, %rd19452; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13907}, %rd19444; + } + { + .reg .b32 %dummy; + mov.b64 {%r13908,%dummy}, %rd19444; + } + shf.l.wrap.b32 %r13909, %r13908, %r13907, 1; + shf.l.wrap.b32 %r13910, %r13907, %r13908, 1; + mov.b64 %rd19469, {%r13910, %r13909}; + xor.b64 %rd19470, %rd19456, %rd19469; + xor.b64 %rd19471, %rd19462, %rd19440; + xor.b64 %rd19472, %rd19464, %rd19370; + xor.b64 %rd19473, %rd19466, %rd19373; + xor.b64 %rd19474, %rd19468, %rd19376; + xor.b64 %rd19475, %rd19470, %rd19379; + xor.b64 %rd19476, %rd19462, %rd19382; + xor.b64 %rd19477, %rd19464, %rd19385; + xor.b64 %rd19478, %rd19466, %rd19388; + xor.b64 %rd19479, %rd19468, %rd19391; + xor.b64 %rd19480, %rd19470, %rd19394; + xor.b64 %rd19481, %rd19462, %rd19397; + xor.b64 %rd19482, %rd19464, %rd19400; + xor.b64 %rd19483, %rd19466, %rd19403; + xor.b64 %rd19484, %rd19468, %rd19406; + xor.b64 %rd19485, %rd19470, %rd19409; + xor.b64 %rd19486, %rd19462, %rd19412; + xor.b64 %rd19487, %rd19464, %rd19415; + xor.b64 %rd19488, %rd19466, %rd19418; + xor.b64 %rd19489, %rd19468, %rd19421; + xor.b64 %rd19490, %rd19470, %rd19424; + xor.b64 %rd19491, %rd19462, %rd19427; + xor.b64 %rd19492, %rd19464, %rd19430; + xor.b64 %rd19493, %rd19466, %rd19433; + xor.b64 %rd19494, %rd19468, %rd19436; + xor.b64 %rd19495, %rd19470, %rd19439; + { + .reg .b32 %dummy; + mov.b64 {%r13911,%dummy}, %rd19477; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13912}, %rd19477; + } + shf.r.wrap.b32 %r13913, %r13912, %r13911, 20; + shf.r.wrap.b32 %r13914, %r13911, %r13912, 20; + mov.b64 %rd19496, {%r13914, %r13913}; + { + .reg .b32 %dummy; + mov.b64 {%r13915,%dummy}, %rd19483; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13916}, %rd19483; + } + shf.r.wrap.b32 %r13917, %r13916, %r13915, 21; + shf.r.wrap.b32 %r13918, %r13915, %r13916, 21; + mov.b64 %rd19497, {%r13918, %r13917}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13919}, %rd19489; + } + { + .reg .b32 %dummy; + mov.b64 {%r13920,%dummy}, %rd19489; + } + shf.l.wrap.b32 %r13921, %r13920, %r13919, 21; + shf.l.wrap.b32 %r13922, %r13919, %r13920, 21; + mov.b64 %rd19498, {%r13922, %r13921}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13923}, %rd19495; + } + { + .reg .b32 %dummy; + mov.b64 {%r13924,%dummy}, %rd19495; + } + shf.l.wrap.b32 %r13925, %r13924, %r13923, 14; + shf.l.wrap.b32 %r13926, %r13923, %r13924, 14; + mov.b64 %rd19499, {%r13926, %r13925}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13927}, %rd19474; + } + { + .reg .b32 %dummy; + mov.b64 {%r13928,%dummy}, %rd19474; + } + shf.l.wrap.b32 %r13929, %r13928, %r13927, 28; + shf.l.wrap.b32 %r13930, %r13927, %r13928, 28; + mov.b64 %rd19500, {%r13930, %r13929}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13931}, %rd19480; + } + { + .reg .b32 %dummy; + mov.b64 {%r13932,%dummy}, %rd19480; + } + shf.l.wrap.b32 %r13933, %r13932, %r13931, 20; + shf.l.wrap.b32 %r13934, %r13931, %r13932, 20; + mov.b64 %rd19501, {%r13934, %r13933}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13935}, %rd19481; + } + { + .reg .b32 %dummy; + mov.b64 {%r13936,%dummy}, %rd19481; + } + shf.l.wrap.b32 %r13937, %r13936, %r13935, 3; + shf.l.wrap.b32 %r13938, %r13935, %r13936, 3; + mov.b64 %rd19502, {%r13938, %r13937}; + { + .reg .b32 %dummy; + mov.b64 {%r13939,%dummy}, %rd19487; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13940}, %rd19487; + } + shf.r.wrap.b32 %r13941, %r13940, %r13939, 19; + shf.r.wrap.b32 %r13942, %r13939, %r13940, 19; + mov.b64 %rd19503, {%r13942, %r13941}; + { + .reg .b32 %dummy; + mov.b64 {%r13943,%dummy}, %rd19493; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13944}, %rd19493; + } + shf.r.wrap.b32 %r13945, %r13944, %r13943, 3; + shf.r.wrap.b32 %r13946, %r13943, %r13944, 3; + mov.b64 %rd19504, {%r13946, %r13945}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13947}, %rd19472; + } + { + .reg .b32 %dummy; + mov.b64 {%r13948,%dummy}, %rd19472; + } + shf.l.wrap.b32 %r13949, %r13948, %r13947, 1; + shf.l.wrap.b32 %r13950, %r13947, %r13948, 1; + mov.b64 %rd19505, {%r13950, %r13949}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13951}, %rd19478; + } + { + .reg .b32 %dummy; + mov.b64 {%r13952,%dummy}, %rd19478; + } + shf.l.wrap.b32 %r13953, %r13952, %r13951, 6; + shf.l.wrap.b32 %r13954, %r13951, %r13952, 6; + mov.b64 %rd19506, {%r13954, %r13953}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13955}, %rd19484; + } + { + .reg .b32 %dummy; + mov.b64 {%r13956,%dummy}, %rd19484; + } + shf.l.wrap.b32 %r13957, %r13956, %r13955, 25; + shf.l.wrap.b32 %r13958, %r13955, %r13956, 25; + mov.b64 %rd19507, {%r13958, %r13957}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13959}, %rd19490; + } + { + .reg .b32 %dummy; + mov.b64 {%r13960,%dummy}, %rd19490; + } + shf.l.wrap.b32 %r13961, %r13960, %r13959, 8; + shf.l.wrap.b32 %r13962, %r13959, %r13960, 8; + mov.b64 %rd19508, {%r13962, %r13961}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13963}, %rd19491; + } + { + .reg .b32 %dummy; + mov.b64 {%r13964,%dummy}, %rd19491; + } + shf.l.wrap.b32 %r13965, %r13964, %r13963, 18; + shf.l.wrap.b32 %r13966, %r13963, %r13964, 18; + mov.b64 %rd19509, {%r13966, %r13965}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13967}, %rd19475; + } + { + .reg .b32 %dummy; + mov.b64 {%r13968,%dummy}, %rd19475; + } + shf.l.wrap.b32 %r13969, %r13968, %r13967, 27; + shf.l.wrap.b32 %r13970, %r13967, %r13968, 27; + mov.b64 %rd19510, {%r13970, %r13969}; + { + .reg .b32 %dummy; + mov.b64 {%r13971,%dummy}, %rd19476; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13972}, %rd19476; + } + shf.r.wrap.b32 %r13973, %r13972, %r13971, 28; + shf.r.wrap.b32 %r13974, %r13971, %r13972, 28; + mov.b64 %rd19511, {%r13974, %r13973}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13975}, %rd19482; + } + { + .reg .b32 %dummy; + mov.b64 {%r13976,%dummy}, %rd19482; + } + shf.l.wrap.b32 %r13977, %r13976, %r13975, 10; + shf.l.wrap.b32 %r13978, %r13975, %r13976, 10; + mov.b64 %rd19512, {%r13978, %r13977}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13979}, %rd19488; + } + { + .reg .b32 %dummy; + mov.b64 {%r13980,%dummy}, %rd19488; + } + shf.l.wrap.b32 %r13981, %r13980, %r13979, 15; + shf.l.wrap.b32 %r13982, %r13979, %r13980, 15; + mov.b64 %rd19513, {%r13982, %r13981}; + { + .reg .b32 %dummy; + mov.b64 {%r13983,%dummy}, %rd19494; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13984}, %rd19494; + } + shf.r.wrap.b32 %r13985, %r13984, %r13983, 8; + shf.r.wrap.b32 %r13986, %r13983, %r13984, 8; + mov.b64 %rd19514, {%r13986, %r13985}; + { + .reg .b32 %dummy; + mov.b64 {%r13987,%dummy}, %rd19473; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13988}, %rd19473; + } + shf.r.wrap.b32 %r13989, %r13988, %r13987, 2; + shf.r.wrap.b32 %r13990, %r13987, %r13988, 2; + mov.b64 %rd19515, {%r13990, %r13989}; + { + .reg .b32 %dummy; + mov.b64 {%r13991,%dummy}, %rd19479; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13992}, %rd19479; + } + shf.r.wrap.b32 %r13993, %r13992, %r13991, 9; + shf.r.wrap.b32 %r13994, %r13991, %r13992, 9; + mov.b64 %rd19516, {%r13994, %r13993}; + { + .reg .b32 %dummy; + mov.b64 {%r13995,%dummy}, %rd19485; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r13996}, %rd19485; + } + shf.r.wrap.b32 %r13997, %r13996, %r13995, 25; + shf.r.wrap.b32 %r13998, %r13995, %r13996, 25; + mov.b64 %rd19517, {%r13998, %r13997}; + { + .reg .b32 %dummy; + mov.b64 {%r13999,%dummy}, %rd19486; + } + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r14000}, %rd19486; + } + shf.r.wrap.b32 %r14001, %r14000, %r13999, 23; + shf.r.wrap.b32 %r14002, %r13999, %r14000, 23; + mov.b64 %rd19518, {%r14002, %r14001}; + { + .reg .b32 %dummy; + mov.b64 {%dummy,%r14003}, %rd19492; + } + { + .reg .b32 %dummy; + mov.b64 {%r14004,%dummy}, %rd19492; + } + shf.l.wrap.b32 %r14005, %r14004, %r14003, 2; + shf.l.wrap.b32 %r14006, %r14003, %r14004, 2; + mov.b64 %rd19519, {%r14006, %r14005}; + not.b64 %rd19520, %rd19496; + and.b64 %rd19521, %rd19497, %rd19520; + xor.b64 %rd19522, %rd19521, %rd19471; + not.b64 %rd19523, %rd19497; + and.b64 %rd19524, %rd19498, %rd19523; + xor.b64 %rd19525, %rd19524, %rd19496; + st.local.u64 [%rd2+8], %rd19525; + not.b64 %rd19526, %rd19498; + and.b64 %rd19527, %rd19499, %rd19526; + xor.b64 %rd19528, %rd19527, %rd19497; + st.local.u64 [%rd2+16], %rd19528; + not.b64 %rd19529, %rd19499; + and.b64 %rd19530, %rd19471, %rd19529; + xor.b64 %rd19531, %rd19530, %rd19498; + st.local.u64 [%rd2+24], %rd19531; + not.b64 %rd19532, %rd19471; + and.b64 %rd19533, %rd19496, %rd19532; + xor.b64 %rd19534, %rd19533, %rd19499; + st.local.u64 [%rd2+32], %rd19534; + not.b64 %rd19535, %rd19501; + and.b64 %rd19536, %rd19502, %rd19535; + xor.b64 %rd19537, %rd19536, %rd19500; + st.local.u64 [%rd2+40], %rd19537; + not.b64 %rd19538, %rd19502; + and.b64 %rd19539, %rd19503, %rd19538; + xor.b64 %rd19540, %rd19539, %rd19501; + st.local.u64 [%rd2+48], %rd19540; + not.b64 %rd19541, %rd19503; + and.b64 %rd19542, %rd19504, %rd19541; + xor.b64 %rd19543, %rd19542, %rd19502; + st.local.u64 [%rd2+56], %rd19543; + not.b64 %rd19544, %rd19504; + and.b64 %rd19545, %rd19500, %rd19544; + xor.b64 %rd19546, %rd19545, %rd19503; + st.local.u64 [%rd2+64], %rd19546; + not.b64 %rd19547, %rd19500; + and.b64 %rd19548, %rd19501, %rd19547; + xor.b64 %rd19549, %rd19548, %rd19504; + st.local.u64 [%rd2+72], %rd19549; + not.b64 %rd19550, %rd19506; + and.b64 %rd19551, %rd19507, %rd19550; + xor.b64 %rd19552, %rd19551, %rd19505; + st.local.u64 [%rd2+80], %rd19552; + not.b64 %rd19553, %rd19507; + and.b64 %rd19554, %rd19508, %rd19553; + xor.b64 %rd19555, %rd19554, %rd19506; + st.local.u64 [%rd2+88], %rd19555; + not.b64 %rd19556, %rd19508; + and.b64 %rd19557, %rd19509, %rd19556; + xor.b64 %rd19558, %rd19557, %rd19507; + st.local.u64 [%rd2+96], %rd19558; + not.b64 %rd19559, %rd19509; + and.b64 %rd19560, %rd19505, %rd19559; + xor.b64 %rd19561, %rd19560, %rd19508; + st.local.u64 [%rd2+104], %rd19561; + not.b64 %rd19562, %rd19505; + and.b64 %rd19563, %rd19506, %rd19562; + xor.b64 %rd19564, %rd19563, %rd19509; + st.local.u64 [%rd2+112], %rd19564; + not.b64 %rd19565, %rd19511; + and.b64 %rd19566, %rd19512, %rd19565; + xor.b64 %rd19567, %rd19566, %rd19510; + st.local.u64 [%rd2+120], %rd19567; + not.b64 %rd19568, %rd19512; + and.b64 %rd19569, %rd19513, %rd19568; + xor.b64 %rd19570, %rd19569, %rd19511; + st.local.u64 [%rd2+128], %rd19570; + not.b64 %rd19571, %rd19513; + and.b64 %rd19572, %rd19514, %rd19571; + xor.b64 %rd19573, %rd19572, %rd19512; + st.local.u64 [%rd2+136], %rd19573; + not.b64 %rd19574, %rd19514; + and.b64 %rd19575, %rd19510, %rd19574; + xor.b64 %rd19576, %rd19575, %rd19513; + st.local.u64 [%rd2+144], %rd19576; + not.b64 %rd19577, %rd19510; + and.b64 %rd19578, %rd19511, %rd19577; + xor.b64 %rd19579, %rd19578, %rd19514; + st.local.u64 [%rd2+152], %rd19579; + not.b64 %rd19580, %rd19516; + and.b64 %rd19581, %rd19517, %rd19580; + xor.b64 %rd19582, %rd19581, %rd19515; + st.local.u64 [%rd2+160], %rd19582; + not.b64 %rd19583, %rd19517; + and.b64 %rd19584, %rd19518, %rd19583; + xor.b64 %rd19585, %rd19584, %rd19516; + st.local.u64 [%rd2+168], %rd19585; + not.b64 %rd19586, %rd19518; + and.b64 %rd19587, %rd19519, %rd19586; + xor.b64 %rd19588, %rd19587, %rd19517; + st.local.u64 [%rd2+176], %rd19588; + not.b64 %rd19589, %rd19519; + and.b64 %rd19590, %rd19515, %rd19589; + xor.b64 %rd19591, %rd19590, %rd19518; + st.local.u64 [%rd2+184], %rd19591; + not.b64 %rd19592, %rd19515; + and.b64 %rd19593, %rd19516, %rd19592; + xor.b64 %rd19594, %rd19593, %rd19519; + st.local.u64 [%rd2+192], %rd19594; + xor.b64 %rd19595, %rd19522, %rd39; + st.local.u64 [%rd2], %rd19595; + @%p39 bra $L__BB1_63; + + add.s64 %rd19597, %rd130, -1; + setp.lt.u64 %p47, %rd19597, 3; + mov.u32 %r14040, 0; + @%p47 bra $L__BB1_59; + + bfe.u64 %rd19599, %rd7997, 6, 2; + sub.s64 %rd19865, %rd130, %rd19599; + +$L__BB1_58: + mul.wide.s32 %rd19600, %r14040, 8; + add.s64 %rd19601, %rd2, %rd19600; + ld.local.u8 %rd19602, [%rd19601]; + ld.local.u8 %rd19603, [%rd19601+1]; + bfi.b64 %rd19604, %rd19603, %rd19602, 8, 8; + ld.local.u8 %rd19605, [%rd19601+2]; + ld.local.u8 %rd19606, [%rd19601+3]; + bfi.b64 %rd19607, %rd19606, %rd19605, 8, 8; + bfi.b64 %rd19608, %rd19607, %rd19604, 16, 16; + ld.local.u8 %rd19609, [%rd19601+4]; + ld.local.u8 %rd19610, [%rd19601+5]; + bfi.b64 %rd19611, %rd19610, %rd19609, 8, 8; + ld.local.u8 %rd19612, [%rd19601+6]; + ld.local.u8 %rd19613, [%rd19601+7]; + bfi.b64 %rd19614, %rd19613, %rd19612, 8, 8; + bfi.b64 %rd19615, %rd19614, %rd19611, 16, 16; + bfi.b64 %rd19616, %rd19615, %rd19608, 32, 32; + st.local.u8 [%rd19601+200], %rd19604; + shr.u64 %rd19617, %rd19604, 8; + st.local.u8 [%rd19601+201], %rd19617; + shr.u64 %rd19618, %rd19608, 24; + st.local.u8 [%rd19601+203], %rd19618; + shr.u64 %rd19619, %rd19608, 16; + st.local.u8 [%rd19601+202], %rd19619; + shr.u64 %rd19620, %rd19616, 56; + st.local.u8 [%rd19601+207], %rd19620; + shr.u64 %rd19621, %rd19616, 48; + st.local.u8 [%rd19601+206], %rd19621; + shr.u64 %rd19622, %rd19616, 40; + st.local.u8 [%rd19601+205], %rd19622; + shr.u64 %rd19623, %rd19616, 32; + st.local.u8 [%rd19601+204], %rd19623; + ld.local.u8 %rd19624, [%rd19601+8]; + ld.local.u8 %rd19625, [%rd19601+9]; + bfi.b64 %rd19626, %rd19625, %rd19624, 8, 8; + ld.local.u8 %rd19627, [%rd19601+10]; + ld.local.u8 %rd19628, [%rd19601+11]; + bfi.b64 %rd19629, %rd19628, %rd19627, 8, 8; + bfi.b64 %rd19630, %rd19629, %rd19626, 16, 16; + ld.local.u8 %rd19631, [%rd19601+12]; + ld.local.u8 %rd19632, [%rd19601+13]; + bfi.b64 %rd19633, %rd19632, %rd19631, 8, 8; + ld.local.u8 %rd19634, [%rd19601+14]; + ld.local.u8 %rd19635, [%rd19601+15]; + bfi.b64 %rd19636, %rd19635, %rd19634, 8, 8; + bfi.b64 %rd19637, %rd19636, %rd19633, 16, 16; + bfi.b64 %rd19638, %rd19637, %rd19630, 32, 32; + st.local.u8 [%rd19601+208], %rd19626; + shr.u64 %rd19639, %rd19626, 8; + st.local.u8 [%rd19601+209], %rd19639; + shr.u64 %rd19640, %rd19630, 24; + st.local.u8 [%rd19601+211], %rd19640; + shr.u64 %rd19641, %rd19630, 16; + st.local.u8 [%rd19601+210], %rd19641; + shr.u64 %rd19642, %rd19638, 56; + st.local.u8 [%rd19601+215], %rd19642; + shr.u64 %rd19643, %rd19638, 48; + st.local.u8 [%rd19601+214], %rd19643; + shr.u64 %rd19644, %rd19638, 40; + st.local.u8 [%rd19601+213], %rd19644; + shr.u64 %rd19645, %rd19638, 32; + st.local.u8 [%rd19601+212], %rd19645; + ld.local.u8 %rd19646, [%rd19601+16]; + ld.local.u8 %rd19647, [%rd19601+17]; + bfi.b64 %rd19648, %rd19647, %rd19646, 8, 8; + ld.local.u8 %rd19649, [%rd19601+18]; + ld.local.u8 %rd19650, [%rd19601+19]; + bfi.b64 %rd19651, %rd19650, %rd19649, 8, 8; + bfi.b64 %rd19652, %rd19651, %rd19648, 16, 16; + ld.local.u8 %rd19653, [%rd19601+20]; + ld.local.u8 %rd19654, [%rd19601+21]; + bfi.b64 %rd19655, %rd19654, %rd19653, 8, 8; + ld.local.u8 %rd19656, [%rd19601+22]; + ld.local.u8 %rd19657, [%rd19601+23]; + bfi.b64 %rd19658, %rd19657, %rd19656, 8, 8; + bfi.b64 %rd19659, %rd19658, %rd19655, 16, 16; + bfi.b64 %rd19660, %rd19659, %rd19652, 32, 32; + st.local.u8 [%rd19601+216], %rd19648; + shr.u64 %rd19661, %rd19648, 8; + st.local.u8 [%rd19601+217], %rd19661; + shr.u64 %rd19662, %rd19652, 24; + st.local.u8 [%rd19601+219], %rd19662; + shr.u64 %rd19663, %rd19652, 16; + st.local.u8 [%rd19601+218], %rd19663; + shr.u64 %rd19664, %rd19660, 56; + st.local.u8 [%rd19601+223], %rd19664; + shr.u64 %rd19665, %rd19660, 48; + st.local.u8 [%rd19601+222], %rd19665; + shr.u64 %rd19666, %rd19660, 40; + st.local.u8 [%rd19601+221], %rd19666; + shr.u64 %rd19667, %rd19660, 32; + st.local.u8 [%rd19601+220], %rd19667; + ld.local.u8 %rd19668, [%rd19601+24]; + ld.local.u8 %rd19669, [%rd19601+25]; + bfi.b64 %rd19670, %rd19669, %rd19668, 8, 8; + ld.local.u8 %rd19671, [%rd19601+26]; + ld.local.u8 %rd19672, [%rd19601+27]; + bfi.b64 %rd19673, %rd19672, %rd19671, 8, 8; + bfi.b64 %rd19674, %rd19673, %rd19670, 16, 16; + ld.local.u8 %rd19675, [%rd19601+28]; + ld.local.u8 %rd19676, [%rd19601+29]; + bfi.b64 %rd19677, %rd19676, %rd19675, 8, 8; + ld.local.u8 %rd19678, [%rd19601+30]; + ld.local.u8 %rd19679, [%rd19601+31]; + bfi.b64 %rd19680, %rd19679, %rd19678, 8, 8; + bfi.b64 %rd19681, %rd19680, %rd19677, 16, 16; + bfi.b64 %rd19682, %rd19681, %rd19674, 32, 32; + st.local.u8 [%rd19601+224], %rd19670; + shr.u64 %rd19683, %rd19670, 8; + st.local.u8 [%rd19601+225], %rd19683; + shr.u64 %rd19684, %rd19674, 24; + st.local.u8 [%rd19601+227], %rd19684; + shr.u64 %rd19685, %rd19674, 16; + st.local.u8 [%rd19601+226], %rd19685; + shr.u64 %rd19686, %rd19682, 56; + st.local.u8 [%rd19601+231], %rd19686; + shr.u64 %rd19687, %rd19682, 48; + st.local.u8 [%rd19601+230], %rd19687; + shr.u64 %rd19688, %rd19682, 40; + st.local.u8 [%rd19601+229], %rd19688; + shr.u64 %rd19689, %rd19682, 32; + st.local.u8 [%rd19601+228], %rd19689; + add.s32 %r14040, %r14040, 4; + add.s64 %rd19865, %rd19865, -4; + setp.ne.s64 %p48, %rd19865, 0; + @%p48 bra $L__BB1_58; + +$L__BB1_59: + bfe.u64 %rd19690, %rd7997, 6, 2; + setp.eq.s64 %p49, %rd19690, 0; + @%p49 bra $L__BB1_63; + + setp.eq.s64 %p50, %rd19690, 1; + mul.wide.s32 %rd19692, %r14040, 8; + add.s64 %rd147, %rd2, %rd19692; + ld.local.u8 %rd19693, [%rd147]; + ld.local.u8 %rd19694, [%rd147+1]; + bfi.b64 %rd19695, %rd19694, %rd19693, 8, 8; + ld.local.u8 %rd19696, [%rd147+2]; + ld.local.u8 %rd19697, [%rd147+3]; + bfi.b64 %rd19698, %rd19697, %rd19696, 8, 8; + bfi.b64 %rd19699, %rd19698, %rd19695, 16, 16; + ld.local.u8 %rd19700, [%rd147+4]; + ld.local.u8 %rd19701, [%rd147+5]; + bfi.b64 %rd19702, %rd19701, %rd19700, 8, 8; + ld.local.u8 %rd19703, [%rd147+6]; + ld.local.u8 %rd19704, [%rd147+7]; + bfi.b64 %rd19705, %rd19704, %rd19703, 8, 8; + bfi.b64 %rd19706, %rd19705, %rd19702, 16, 16; + bfi.b64 %rd19707, %rd19706, %rd19699, 32, 32; + shl.b32 %r14009, %r14040, 3; + cvt.s64.s32 %rd19708, %r14009; + add.s64 %rd19709, %rd2, %rd19708; + add.s64 %rd148, %rd19709, 200; + st.local.u8 [%rd19709+200], %rd19695; + shr.u64 %rd19710, %rd19695, 8; + st.local.u8 [%rd19709+201], %rd19710; + shr.u64 %rd19711, %rd19699, 24; + st.local.u8 [%rd19709+203], %rd19711; + shr.u64 %rd19712, %rd19699, 16; + st.local.u8 [%rd19709+202], %rd19712; + shr.u64 %rd19713, %rd19707, 56; + st.local.u8 [%rd19709+207], %rd19713; + shr.u64 %rd19714, %rd19707, 48; + st.local.u8 [%rd19709+206], %rd19714; + shr.u64 %rd19715, %rd19707, 40; + st.local.u8 [%rd19709+205], %rd19715; + shr.u64 %rd19716, %rd19707, 32; + st.local.u8 [%rd19709+204], %rd19716; + @%p50 bra $L__BB1_63; + + setp.eq.s64 %p51, %rd19690, 2; + ld.local.u8 %rd19718, [%rd147+8]; + ld.local.u8 %rd19719, [%rd147+9]; + bfi.b64 %rd19720, %rd19719, %rd19718, 8, 8; + ld.local.u8 %rd19721, [%rd147+10]; + ld.local.u8 %rd19722, [%rd147+11]; + bfi.b64 %rd19723, %rd19722, %rd19721, 8, 8; + bfi.b64 %rd19724, %rd19723, %rd19720, 16, 16; + ld.local.u8 %rd19725, [%rd147+12]; + ld.local.u8 %rd19726, [%rd147+13]; + bfi.b64 %rd19727, %rd19726, %rd19725, 8, 8; + ld.local.u8 %rd19728, [%rd147+14]; + ld.local.u8 %rd19729, [%rd147+15]; + bfi.b64 %rd19730, %rd19729, %rd19728, 8, 8; + bfi.b64 %rd19731, %rd19730, %rd19727, 16, 16; + bfi.b64 %rd19732, %rd19731, %rd19724, 32, 32; + st.local.u8 [%rd148+8], %rd19720; + shr.u64 %rd19733, %rd19720, 8; + st.local.u8 [%rd148+9], %rd19733; + shr.u64 %rd19734, %rd19724, 24; + st.local.u8 [%rd148+11], %rd19734; + shr.u64 %rd19735, %rd19724, 16; + st.local.u8 [%rd148+10], %rd19735; + shr.u64 %rd19736, %rd19732, 56; + st.local.u8 [%rd148+15], %rd19736; + shr.u64 %rd19737, %rd19732, 48; + st.local.u8 [%rd148+14], %rd19737; + shr.u64 %rd19738, %rd19732, 40; + st.local.u8 [%rd148+13], %rd19738; + shr.u64 %rd19739, %rd19732, 32; + st.local.u8 [%rd148+12], %rd19739; + @%p51 bra $L__BB1_63; + + ld.local.u8 %rd19740, [%rd147+16]; + ld.local.u8 %rd19741, [%rd147+17]; + bfi.b64 %rd19742, %rd19741, %rd19740, 8, 8; + ld.local.u8 %rd19743, [%rd147+18]; + ld.local.u8 %rd19744, [%rd147+19]; + bfi.b64 %rd19745, %rd19744, %rd19743, 8, 8; + bfi.b64 %rd19746, %rd19745, %rd19742, 16, 16; + ld.local.u8 %rd19747, [%rd147+20]; + ld.local.u8 %rd19748, [%rd147+21]; + bfi.b64 %rd19749, %rd19748, %rd19747, 8, 8; + ld.local.u8 %rd19750, [%rd147+22]; + ld.local.u8 %rd19751, [%rd147+23]; + bfi.b64 %rd19752, %rd19751, %rd19750, 8, 8; + bfi.b64 %rd19753, %rd19752, %rd19749, 16, 16; + bfi.b64 %rd19754, %rd19753, %rd19746, 32, 32; + st.local.u8 [%rd148+16], %rd19742; + shr.u64 %rd19755, %rd19742, 8; + st.local.u8 [%rd148+17], %rd19755; + shr.u64 %rd19756, %rd19746, 24; + st.local.u8 [%rd148+19], %rd19756; + shr.u64 %rd19757, %rd19746, 16; + st.local.u8 [%rd148+18], %rd19757; + shr.u64 %rd19758, %rd19754, 56; + st.local.u8 [%rd148+23], %rd19758; + shr.u64 %rd19759, %rd19754, 48; + st.local.u8 [%rd148+22], %rd19759; + shr.u64 %rd19760, %rd19754, 40; + st.local.u8 [%rd148+21], %rd19760; + shr.u64 %rd19761, %rd19754, 32; + st.local.u8 [%rd148+20], %rd19761; + +$L__BB1_63: + st.local.u64 [%rd2+392], %rd7997; + mov.u64 %rd19866, %rd7997; + +$L__BB1_64: + sub.s64 %rd19762, %rd137, %rd19864; + min.u64 %rd150, %rd19866, %rd19762; + shr.u64 %rd151, %rd150, 3; + setp.eq.s64 %p52, %rd151, 0; + @%p52 bra $L__BB1_72; + + add.s64 %rd19763, %rd151, -1; + setp.lt.u64 %p53, %rd19763, 3; + mov.u32 %r14042, 0; + @%p53 bra $L__BB1_68; + + bfe.u64 %rd19765, %rd150, 3, 2; + sub.s64 %rd19867, %rd151, %rd19765; + add.u64 %rd153, %SPL, 64; + shr.u64 %rd19767, %rd19864, 3; + cvt.u32.u64 %r14012, %rd19767; + mov.u32 %r14013, 31; + sub.s32 %r16, %r14013, %r14012; + shr.u64 %rd19768, %rd19866, 3; + sub.s64 %rd154, %rd14, %rd19768; + +$L__BB1_67: + cvt.s64.s32 %rd19769, %r14042; + add.s64 %rd19770, %rd154, %rd19769; + add.s64 %rd19771, %rd2, %rd19770; + ld.local.u8 %rs48, [%rd19771+200]; + sub.s32 %r14014, %r16, %r14042; + cvt.s64.s32 %rd19772, %r14014; + add.s64 %rd19773, %rd153, %rd19772; + st.local.u8 [%rd19773], %rs48; + ld.local.u8 %rs49, [%rd19771+201]; + st.local.u8 [%rd19773+-1], %rs49; + ld.local.u8 %rs50, [%rd19771+202]; + st.local.u8 [%rd19773+-2], %rs50; + ld.local.u8 %rs51, [%rd19771+203]; + st.local.u8 [%rd19773+-3], %rs51; + add.s32 %r14042, %r14042, 4; + add.s64 %rd19867, %rd19867, -4; + setp.ne.s64 %p54, %rd19867, 0; + @%p54 bra $L__BB1_67; + +$L__BB1_68: + bfe.u64 %rd19774, %rd150, 3, 2; + setp.eq.s64 %p55, %rd19774, 0; + @%p55 bra $L__BB1_72; + + cvt.s64.s32 %rd19775, %r14042; + shr.u64 %rd19776, %rd19866, 3; + sub.s64 %rd19777, %rd14, %rd19776; + add.s64 %rd19778, %rd19777, %rd19775; + add.s64 %rd19779, %rd2, %rd19778; + add.s64 %rd157, %rd19779, 200; + ld.local.u8 %rs52, [%rd19779+200]; + shr.u64 %rd19780, %rd19864, 3; + cvt.u32.u64 %r14015, %rd19780; + mov.u32 %r14016, 31; + sub.s32 %r14017, %r14016, %r14015; + sub.s32 %r14018, %r14017, %r14042; + cvt.s64.s32 %rd19781, %r14018; + add.u64 %rd19783, %SPL, 64; + add.s64 %rd158, %rd19783, %rd19781; + st.local.u8 [%rd158], %rs52; + setp.eq.s64 %p56, %rd19774, 1; + @%p56 bra $L__BB1_72; + + ld.local.u8 %rs53, [%rd157+1]; + st.local.u8 [%rd158+-1], %rs53; + setp.eq.s64 %p57, %rd19774, 2; + @%p57 bra $L__BB1_72; + + ld.local.u8 %rs54, [%rd157+2]; + st.local.u8 [%rd158+-2], %rs54; + +$L__BB1_72: + sub.s64 %rd19866, %rd19866, %rd150; + st.local.u64 [%rd2+392], %rd19866; + add.s64 %rd19864, %rd150, %rd19864; + setp.lt.u64 %p58, %rd19864, %rd137; + @%p58 bra $L__BB1_54; + +$L__BB1_73: + ld.param.u64 %rd19795, [kernel_lilypad_pow_param_2]; + add.u64 %rd19786, %SP, 64; + { // callseq 1, 0 + .reg .b32 temp_param_reg; + .param .b64 param0; + st.param.b64 [param0+0], %rd19786; + .param .b64 param1; + st.param.b64 [param1+0], %rd19795; + .param .b32 retval0; + call.uni (retval0), + _ZN39_INTERNAL_467e079b_9_keccak_cu_bbb2fa6e15hashbelowtargetEPKyS1_, + ( + param0, + param1 + ); + ld.param.b32 %r14019, [retval0+0]; + } // callseq 1 + cvt.u16.u32 %rs55, %r14019; + setp.eq.s16 %p59, %rs55, 0; + @%p59 bra $L__BB1_75; + + mov.u32 %r14031, %tid.x; + mov.u32 %r14030, %ntid.x; + mov.u32 %r14029, %ctaid.x; + mad.lo.s32 %r14028, %r14029, %r14030, %r14031; + cvt.u64.u32 %rd19807, %r14028; + add.s64 %rd19806, %rd173, %rd19807; + shr.u64 %rd19805, %rd19806, 16; + mov.u32 %r14027, %ctaid.x; + mov.u32 %r14026, %tid.x; + mov.u32 %r14025, %ntid.x; + mad.lo.s32 %r14024, %r14027, %r14025, %r14026; + cvt.u64.u32 %rd19804, %r14024; + add.s64 %rd19803, %rd173, %rd19804; + shr.u64 %rd19802, %rd19803, 32; + shr.u64 %rd19801, %rd19803, 40; + mov.u32 %r14023, %ctaid.x; + mov.u32 %r14022, %tid.x; + mov.u32 %r14021, %ntid.x; + mad.lo.s32 %r14020, %r14023, %r14021, %r14022; + cvt.u64.u32 %rd19800, %r14020; + add.s64 %rd19799, %rd173, %rd19800; + shr.u64 %rd19798, %rd19799, 48; + shr.u64 %rd19797, %rd19799, 56; + ld.param.u64 %rd19796, [kernel_lilypad_pow_param_4]; + cvta.to.global.u64 %rd19787, %rd19796; + st.global.u8 [%rd19787], %rs1; + st.global.u8 [%rd19787+1], %rs2; + st.global.u8 [%rd19787+2], %rs3; + st.global.u8 [%rd19787+3], %rs4; + st.global.u8 [%rd19787+4], %rs5; + st.global.u8 [%rd19787+5], %rs6; + st.global.u8 [%rd19787+6], %rs7; + st.global.u8 [%rd19787+7], %rs8; + st.global.u8 [%rd19787+8], %rs9; + st.global.u8 [%rd19787+9], %rs10; + st.global.u8 [%rd19787+10], %rs11; + st.global.u8 [%rd19787+11], %rs12; + st.global.u8 [%rd19787+12], %rs13; + st.global.u8 [%rd19787+13], %rs14; + st.global.u8 [%rd19787+14], %rs15; + st.global.u8 [%rd19787+15], %rs16; + st.global.u8 [%rd19787+16], %rs17; + st.global.u8 [%rd19787+17], %rs18; + st.global.u8 [%rd19787+18], %rs19; + st.global.u8 [%rd19787+19], %rs20; + st.global.u8 [%rd19787+20], %rs21; + st.global.u8 [%rd19787+21], %rs22; + st.global.u8 [%rd19787+22], %rs23; + st.global.u8 [%rd19787+23], %rs24; + st.global.u8 [%rd19787+24], %rd19797; + st.global.u8 [%rd19787+25], %rd19798; + st.global.u8 [%rd19787+26], %rd19801; + st.global.u8 [%rd19787+27], %rd19802; + shr.u64 %rd19792, %rd19799, 24; + st.global.u8 [%rd19787+28], %rd19792; + st.global.u8 [%rd19787+29], %rd19805; + shr.u64 %rd19794, %rd19799, 8; + st.global.u8 [%rd19787+30], %rd19794; + st.global.u8 [%rd19787+31], %rd19799; + +$L__BB1_75: + { // callseq 2, 0 + .reg .b32 temp_param_reg; + .param .b64 param0; + st.param.b64 [param0+0], %rd8; + call.uni + free, + ( + param0 + ); + } // callseq 2 + +$L__BB1_76: + ret; + +} + diff --git a/pkg/resourceprovider/cudaworker.go b/pkg/resourceprovider/cudaworker.go new file mode 100644 index 00000000..a54ad460 --- /dev/null +++ b/pkg/resourceprovider/cudaworker.go @@ -0,0 +1,237 @@ +//go:build cuda +// +build cuda + +package resourceprovider + +import ( + "context" + _ "embed" + "math/big" + "os" + "slices" + "sync/atomic" + "time" + "unsafe" + + "github.com/ethereum/go-ethereum/common/math" + "github.com/ethereum/go-ethereum/crypto" + "github.com/holiman/uint256" + "github.com/pkg/errors" + "github.com/rs/zerolog/log" + "gorgonia.org/cu" +) + +//go:embed cudaminer/keccak.ptx +var keccakPtx string + +const entry_point = "kernel_lilypad_pow" + +var MaybeCudaOrCpu = NewGpuWorker + +func DefaultWorkerNum() int { + return 20 // different on different device +} + +type GpuWorker struct { + cfg *WorkerConfig + state atomic.Int32 + entryFn cu.Function + cuCtx *cu.Ctx + + quit chan chan struct{} +} + +func NewGpuWorker(cfg *WorkerConfig) (Worker, error) { + if GetGpuNumber() == 0 { + log.Warn().Msg("No gpu found and fallback to cpu") + return NewCpuWorker(cfg) + } + //TODO use first gpu for now, plan to support multiple gpu in future + cuCtx, err := setupGPU() + if err != nil { + return nil, err + } + + fs, err := os.CreateTemp(os.TempDir(), "*") + if err != nil { + return nil, err + } + + //TODO no LoadData for Cuda Ctx, maybe cu author forget or just not supprot + _, err = fs.WriteString(keccakPtx) + if err != nil { + return nil, err + } + myModule, err := cuCtx.Load(fs.Name()) + if err != nil { + return nil, err + } + + entryFn, err := myModule.Function(entry_point) + if err != nil { + return nil, err + } + + return &GpuWorker{ + cfg: cfg, + cuCtx: cuCtx, + entryFn: entryFn, + quit: make(chan chan struct{}, 1), + }, nil +} +func (w *GpuWorker) Stop() { + if w.state.Load() == 0 { + return + } + + //cancel previous task + waitChan := make(chan struct{}) + select { + case w.quit <- waitChan: // may already exit + <-waitChan + default: + } +} +func (w *GpuWorker) FindSolution(ctx context.Context, task *Task) { + w.state.Store(1) + defer w.state.Store(0) + + nonce := task.From.Clone() + startTime := time.Now() + + hashesCompleted := uint64(0) + ticker := time.NewTicker(time.Second * hashUpdateSecs) + defer ticker.Stop() + batch_size := uint64(w.cfg.gridSize * w.cfg.blockSize) +OUT: + for { + select { + case <-ctx.Done(): + break OUT + case respCh := <-w.quit: + respCh <- struct{}{} + return + case <-ticker.C: + w.cfg.updateHashes <- hashesCompleted + hashesCompleted = 0 + default: + // Non-blocking select to fall through + } + + if nonce.Cmp(task.End) >= 0 { + return + } + + maybeNonce, err := kernel_lilypad_pow_with_ctx(w.cuCtx, w.entryFn, task.Challenge, nonce.ToBig(), task.Difficulty.ToBig(), w.cfg.gridSize, w.cfg.blockSize) + if err != nil { + log.Err(err).Msg("InvokeGpu fail") + continue + } + hashesCompleted += batch_size + nonce = nonce.Add(nonce, uint256.NewInt(batch_size)) + + if maybeNonce.Int64() == 0 { + continue + } + + data, err := formatMinerArgs(task.Challenge, maybeNonce) + if err != nil { + log.Err(err).Msg("Generate hash data") + continue + } + result := crypto.Keccak256Hash(data) + + hashNumber := new(uint256.Int).SetBytes(result[:]) + // Check if the hash is below the target difficulty + if hashNumber.Cmp(task.Difficulty) == -1 { + log.Info().Int("WorkerID", w.cfg.id).Str("Elapsed Time", time.Since(startTime).String()). + Str("challenge", new(big.Int).SetBytes(task.Challenge[:]).String()). + Str("Nonce", maybeNonce.String()). + Str("HashNumber", hashNumber.String()). + Msg("Success!") + select { + case w.cfg.resultCh <- TaskResult{ + Id: task.Id, + Nonce: uint256.MustFromBig(maybeNonce), + }: + default: //avoid deadlock + } + } else { + log.Error().Msg("This branch should never happen, only when cuda algo may have error") + } + } +} + +func GetGpuNumber() int { + devices, err := cu.NumDevices() + if err != nil { + log.Warn().Msgf("Cannot detect gpu numbers %v", err) + return 0 + } + return devices +} + +func setupGPU() (*cu.Ctx, error) { + devices, _ := cu.NumDevices() + + if devices == 0 { + return nil, errors.Errorf("NoDevice") + } + + dev := cu.Device(0) + return cu.NewContext(dev, cu.SchedAuto), nil +} + +func kernel_lilypad_pow_with_ctx(cuCtx *cu.Ctx, fn cu.Function, challenge [32]byte, startNonce *big.Int, difficulty *big.Int, grid, block int) (*big.Int, error) { + dIn1, err := cuCtx.MemAllocManaged(32, cu.AttachGlobal) + if err != nil { + return nil, err + } + + dIn2, err := cuCtx.MemAllocManaged(32, cu.AttachGlobal) + if err != nil { + return nil, err + } + + dIn3, err := cuCtx.MemAllocManaged(32, cu.AttachGlobal) + if err != nil { + return nil, err + } + + dOut, err := cuCtx.MemAllocManaged(32, cu.AttachGlobal) + if err != nil { + return nil, err + } + + cuCtx.MemcpyHtoD(dIn1, unsafe.Pointer(&challenge[0]), 32) + + startNonceBytes := math.U256Bytes(startNonce) + slices.Reverse(startNonceBytes) + cuCtx.MemcpyHtoD(dIn2, unsafe.Pointer(&startNonceBytes[0]), 32) + + difficutyBytes := math.U256Bytes(difficulty) + slices.Reverse(difficutyBytes) //to big + cuCtx.MemcpyHtoD(dIn3, unsafe.Pointer(&difficutyBytes[0]), 32) + + batch_size := int64(grid * block) + args := []unsafe.Pointer{ + unsafe.Pointer(&dIn1), + unsafe.Pointer(&dIn2), + unsafe.Pointer(&dIn3), + unsafe.Pointer(&batch_size), + unsafe.Pointer(&dOut), + } + + cuCtx.LaunchKernel(fn, grid, 1, 1, block, 1, 1, 1, cu.Stream{}, args) + cuCtx.Synchronize() + + hOut := make([]byte, 32) + cuCtx.MemcpyDtoH(unsafe.Pointer(&hOut[0]), dOut, 32) + + cuCtx.MemFree(dIn1) + cuCtx.MemFree(dIn2) + cuCtx.MemFree(dIn3) + cuCtx.MemFree(dIn2) + cuCtx.MemFree(dOut) + return new(big.Int).SetBytes(hOut), nil +} diff --git a/pkg/resourceprovider/cpuminer.go b/pkg/resourceprovider/minerctl.go similarity index 57% rename from pkg/resourceprovider/cpuminer.go rename to pkg/resourceprovider/minerctl.go index e55c1425..a6e7862a 100644 --- a/pkg/resourceprovider/cpuminer.go +++ b/pkg/resourceprovider/minerctl.go @@ -3,8 +3,10 @@ package resourceprovider import ( "context" "math/big" + "sync" "time" + "github.com/ethereum/go-ethereum/accounts/abi" "github.com/google/uuid" lru "github.com/hashicorp/golang-lru/v2" "github.com/holiman/uint256" @@ -25,36 +27,63 @@ const ( ) type SubmitWork func(nonce *big.Int) +type Worker interface { + FindSolution(ctx context.Context, task *Task) + Stop() +} + +type WorkerConfig struct { + id int + updateHashes chan uint64 + resultCh chan TaskResult + + //cuda + gridSize int + blockSize int +} + +type Task struct { + Id uuid.UUID + Challenge [32]byte + Difficulty *uint256.Int + From *uint256.Int + End *uint256.Int +} + +type TaskResult struct { + Id uuid.UUID + Nonce *uint256.Int +} -type CPUMiner struct { +type MinerController struct { submit SubmitWork - runningWorkers []*Worker + runningWorkers []Worker - numWorkers int + powCfg ResourceProviderPowOptions task chan Task updateHashes chan uint64 } -func NewCpuMiner(nodeId string, numWorkers int, task chan Task, submit SubmitWork) *CPUMiner { - return &CPUMiner{ - numWorkers: numWorkers, +func NewMinerController(nodeId string, powCfg ResourceProviderPowOptions, task chan Task, submit SubmitWork) *MinerController { + return &MinerController{ + powCfg: powCfg, task: task, updateHashes: make(chan uint64), submit: submit, } } -func (m *CPUMiner) Start(ctx context.Context) { +func (m *MinerController) Start(ctx context.Context) { go m.miningWorkerController(ctx) go m.speedMonitor(ctx) } // speedMonitor handles tracking the number of hashes per second the mining // process is performing. It must be run as a goroutine. -func (m *CPUMiner) speedMonitor(ctx context.Context) { +func (m *MinerController) speedMonitor(ctx context.Context) { log.Debug().Msg("CPU miner speed monitor started") var hashesPerSec float64 var totalHashes uint64 @@ -89,25 +118,53 @@ out: log.Debug().Msgf("CPU miner speed monitor done") } -func (m *CPUMiner) miningWorkerController(ctx context.Context) { - resultCh := make(chan TaskResult) - launchWorkers := func(numWorkers int) { - for i := 0; i < numWorkers; i++ { - w := NewWorker(i, m.updateHashes, resultCh) +func (m *MinerController) miningWorkerController(ctx context.Context) { + numworkers := m.powCfg.NumWorkers + if numworkers == 0 { + numworkers = DefaultWorkerNum() + } + + resultCh := make(chan TaskResult, numworkers*2) //avoid lock worker if have much work to submit + launchWorkers := func(powCfg ResourceProviderPowOptions) error { + for i := 0; i < numworkers; i++ { + wCfg := &WorkerConfig{ + id: i, + updateHashes: m.updateHashes, + resultCh: resultCh, + + gridSize: powCfg.CudaGridSize, + blockSize: powCfg.CudaBlockSize, + } + + w, err := MaybeCudaOrCpu(wCfg) + if err != nil { + return err + } + m.runningWorkers = append(m.runningWorkers, w) } + return nil } maxUint256 := new(uint256.Int).Sub(uint256.NewInt(0), uint256.NewInt(1)) - noncePerWorker := new(uint256.Int).Div(maxUint256, uint256.NewInt(uint64(m.numWorkers))) + noncePerWorker := new(uint256.Int).Div(maxUint256, uint256.NewInt(uint64(numworkers))) // Launch the current number of workers by default. - launchWorkers(m.numWorkers) + err := launchWorkers(m.powCfg) + if err != nil { + log.Err(err).Msg("Cannt create worker") + } stopWrokers := func() { - for _, w := range m.runningWorkers { - w.Stop() + var wg sync.WaitGroup + for _, worker := range m.runningWorkers { + wg.Add(1) + go func(w Worker) { + defer wg.Done() + w.Stop() + }(worker) } + wg.Wait() } spawNewWork := func(allTask *Task) { @@ -115,7 +172,7 @@ func (m *CPUMiner) miningWorkerController(ctx context.Context) { w.Stop() from := new(uint256.Int).Mul(noncePerWorker, uint256.NewInt(uint64(i))) end := new(uint256.Int).Mul(noncePerWorker, uint256.NewInt(uint64(i+1))) - go w.Solve(ctx, &Task{ + go w.FindSolution(ctx, &Task{ Id: allTask.Id, Challenge: allTask.Challenge, Difficulty: allTask.Difficulty, @@ -147,3 +204,29 @@ out: } } } + +func formatMinerArgs(challenge [32]byte, nonce *big.Int) ([]byte, error) { + //todo use nonce in replace instead of building from scratch for better performance + // keccak256(abi.encodePacked(lastChallenge, msg.sender, nodeId)); + bytes32Ty, _ := abi.NewType("bytes32", "", nil) + uint256Ty, _ := abi.NewType("uint256", "", nil) + + arguments := abi.Arguments{ + { + Type: bytes32Ty, + }, + { + Type: uint256Ty, + }, + } + + bytes, err := arguments.Pack( + challenge, + nonce, + ) + if err != nil { + return nil, err + } + + return bytes, nil +} diff --git a/pkg/resourceprovider/others_os.go b/pkg/resourceprovider/others_os.go new file mode 100644 index 00000000..acd434c3 --- /dev/null +++ b/pkg/resourceprovider/others_os.go @@ -0,0 +1,12 @@ +//go:build !cuda +// +build !cuda + +package resourceprovider + +import "runtime" + +var MaybeCudaOrCpu = NewCpuWorker + +func DefaultWorkerNum() int { + return runtime.NumCPU() * 2 +} diff --git a/pkg/resourceprovider/resourceprovider.go b/pkg/resourceprovider/resourceprovider.go index 1b25d049..d854201e 100644 --- a/pkg/resourceprovider/resourceprovider.go +++ b/pkg/resourceprovider/resourceprovider.go @@ -3,9 +3,10 @@ package resourceprovider import ( "context" "encoding/hex" + "fmt" "math/big" - "runtime" + "github.com/ethereum/go-ethereum/common" "github.com/google/uuid" "github.com/holiman/uint256" "github.com/lilypad-tech/lilypad/pkg/data" @@ -52,7 +53,11 @@ type ResourceProviderOfferOptions struct { // this configures the pow we will keep track of type ResourceProviderPowOptions struct { - EnablePow bool + EnablePow bool + NumWorkers int + + CudaGridSize int + CudaBlockSize int } type ResourceProviderOptions struct { @@ -118,9 +123,6 @@ func (resourceProvider *ResourceProvider) StartMineLoop(ctx context.Context) err } }) - numWorkers := runtime.NumCPU() * 2 - log.Info().Msgf("Listen to new pow round signal, %d workers read to work", numWorkers) - submitWork := func(nonce *big.Int) { txId, submission, err := resourceProvider.web3SDK.SubmitWork(ctx, nonce, nodeId) if err != nil { @@ -134,7 +136,25 @@ func (resourceProvider *ResourceProvider) StartMineLoop(ctx context.Context) err Msgf("Mine and submit successfully") } - miner := NewCpuMiner(nodeId, numWorkers, taskCh, submitWork) + log.Info().Msgf("Listen to new pow round signal, %d workers read to work", resourceProvider.options.Pow.NumWorkers) + miner := NewMinerController(nodeId, resourceProvider.options.Pow, taskCh, submitWork) go miner.Start(ctx) return nil } + +func TriggerNewPowRound(ctx context.Context, web3SDK *web3.Web3SDK) (common.Hash, error) { + tx, err := web3SDK.Contracts.Pow.TriggerNewPowRound(web3SDK.TransactOpts) + if err != nil { + return common.Hash{}, err + } + + receipt, err := web3SDK.WaitTx(ctx, tx) + if err != nil { + return common.Hash{}, err + } + + if receipt.Status != 1 { + return tx.Hash(), fmt.Errorf("trigger new pow round") + } + return tx.Hash(), nil +} diff --git a/pkg/resourceprovider/worker.go b/pkg/resourceprovider/worker.go deleted file mode 100644 index 04770759..00000000 --- a/pkg/resourceprovider/worker.go +++ /dev/null @@ -1,174 +0,0 @@ -package resourceprovider - -import ( - "context" - "fmt" - "math/big" - "sync/atomic" - "time" - - "github.com/ethereum/go-ethereum/accounts/abi" - "github.com/ethereum/go-ethereum/common" - "github.com/ethereum/go-ethereum/crypto" - "github.com/google/uuid" - "github.com/holiman/uint256" - "github.com/lilypad-tech/lilypad/pkg/web3" - "github.com/rs/zerolog/log" -) - -var ( - bigOne = uint256.NewInt(1) -) - -type Task struct { - Id uuid.UUID - Challenge [32]byte - Difficulty *uint256.Int - From *uint256.Int - End *uint256.Int -} - -type TaskResult struct { - Id uuid.UUID - Nonce *uint256.Int -} - -type Worker struct { - id int - state atomic.Int32 - - updateHashes chan uint64 - - resultCh chan TaskResult - quit chan chan struct{} -} - -func NewWorker(id int, updateHashes chan uint64, resultCh chan TaskResult) *Worker { - return &Worker{ - id: id, - updateHashes: updateHashes, - resultCh: resultCh, - quit: make(chan chan struct{}, 1), - } -} -func (w *Worker) Stop() { - if w.state.Load() == 0 { - return - } - - //cancel previous task - waitChan := make(chan struct{}) - select { - case w.quit <- waitChan: // may already exit - <-waitChan - default: - } -} - -func (w *Worker) Solve(ctx context.Context, task *Task) { - w.state.Store(1) - defer w.state.Store(0) - - nonce := task.From.Clone() - startTime := time.Now() - - hashesCompleted := uint64(0) - ticker := time.NewTicker(time.Second * hashUpdateSecs) - defer ticker.Stop() - -OUT: - for { - select { - case <-ctx.Done(): - break OUT - case respCh := <-w.quit: - respCh <- struct{}{} - return - case <-ticker.C: - w.updateHashes <- hashesCompleted - hashesCompleted = 0 - default: - // Non-blocking select to fall through - } - - if nonce.Cmp(task.End) >= 0 { - return - } - hashNumber, err := calculateHashNumber(task.Challenge, nonce.ToBig()) - if err != nil { - log.Err(err).Msg("Calculate hash number") - return - } - hashesCompleted += 1 - - // Check if the hash is below the target difficulty - if hashNumber.Cmp(task.Difficulty) == -1 { - log.Info().Str("Elapsed Time", time.Since(startTime).String()). - Str("challenge", new(big.Int).SetBytes(task.Challenge[:]).String()). - Str("Nonce", nonce.String()). - Str("HashNumber", hashNumber.String()). - Msg("Success!") - w.resultCh <- TaskResult{ - Id: task.Id, - Nonce: nonce.Clone(), - } - } - - nonce.Add(nonce, bigOne) - } -} - -func calculateHashNumber(challenge [32]byte, nonce *big.Int) (*uint256.Int, error) { - data, err := formatMinerArgs(challenge, nonce) - if err != nil { - return nil, err - } - - // Calculate Keccak-256 hash - hashResult := crypto.Keccak256(data) - - return new(uint256.Int).SetBytes(hashResult), nil -} - -func formatMinerArgs(challenge [32]byte, nonce *big.Int) ([]byte, error) { - //todo use nonce in replace instead of building from scratch for better performance - // keccak256(abi.encodePacked(lastChallenge, msg.sender, nodeId)); - bytes32Ty, _ := abi.NewType("bytes32", "", nil) - uint256Ty, _ := abi.NewType("uint256", "", nil) - - arguments := abi.Arguments{ - { - Type: bytes32Ty, - }, - { - Type: uint256Ty, - }, - } - - bytes, err := arguments.Pack( - challenge, - nonce, - ) - if err != nil { - return nil, err - } - - return bytes, nil -} - -func TriggerNewPowRound(ctx context.Context, web3SDK *web3.Web3SDK) (common.Hash, error) { - tx, err := web3SDK.Contracts.Pow.TriggerNewPowRound(web3SDK.TransactOpts) - if err != nil { - return common.Hash{}, err - } - - receipt, err := web3SDK.WaitTx(ctx, tx) - if err != nil { - return common.Hash{}, err - } - - if receipt.Status != 1 { - return tx.Hash(), fmt.Errorf("trigger new pow round") - } - return tx.Hash(), nil -}