From 1a30a43a2dcf0cd0b15039d66946037b5cdbc316 Mon Sep 17 00:00:00 2001
From: James Osborn <osborn@alcf.anl.gov>
Date: Mon, 12 Dec 2022 14:21:38 -0600
Subject: [PATCH] add optional QIOlite reader

---
 build/configBase.nims    |   3 +-
 qex.nimble               |   2 +-
 src/io/reader.nim        | 286 +--------------------------------------
 src/io/readerQio.nim     | 282 ++++++++++++++++++++++++++++++++++++++
 src/io/readerQiolite.nim | 225 ++++++++++++++++++++++++++++++
 src/rng/milcrng.nim      |   1 -
 6 files changed, 516 insertions(+), 283 deletions(-)
 create mode 100644 src/io/readerQio.nim
 create mode 100644 src/io/readerQiolite.nim

diff --git a/build/configBase.nims b/build/configBase.nims
index 34afd1bd..5287f704 100644
--- a/build/configBase.nims
+++ b/build/configBase.nims
@@ -116,7 +116,8 @@ proc getNimFlags*(fo: flagsOpts): seq[string] =
   verbosity ~ buildVerbosity
   nimcache ~ nimcache
   warning[SmallLshouldNotBeUsed] ~ off
-  styleCheck ~ usages
+  if (NimMajor, NimMinor) >= (1, 6):
+    styleCheck ~ usages
   embedsrc ~ ""
   #exceptions ~ quirky
   #if not nimargs.any(proc (x:string):bool = x.startswith("--mm:")):
diff --git a/qex.nimble b/qex.nimble
index c6766fe2..c5c267b8 100644
--- a/qex.nimble
+++ b/qex.nimble
@@ -30,7 +30,7 @@ requires "mdevolve >= 1.0.0"
 if primmeDir != "":
   requires "primme >= 3.0.0"
 
-requires "https://github.com/usqcd-software/qiolite"
+requires "https://github.com/usqcd-software/qiolite#head"
 
 # Helpers
 
diff --git a/src/io/reader.nim b/src/io/reader.nim
index 3c8f368f..f5d2fdd0 100644
--- a/src/io/reader.nim
+++ b/src/io/reader.nim
@@ -1,282 +1,8 @@
-import base
-import layout
-import qio
-import strutils
-import macros
-import field
-#import comms
 import os
-#import stdUtils
-import iocommon
 
-proc getFileLattice*(s:string):seq[int] =
-  if not fileExists(s):
-    warn "file not found: ", s
-    return
-  var ql:QIO_Layout
-  ql.this_node = myRank.cint
-  ql.number_of_nodes = nRanks.cint
-  var qs = QIO_string_create()
-  var qr = QIO_open_read(qs, s, ql.addr, nil, nil)
-  let nd = QIO_get_reader_latdim(qr)
-  #echo nd
-  let ls = QIO_get_reader_latsize(qr)
-  result.newSeq(nd)
-  for i in 0..<nd:
-    result[i] = ls[i]
-  discard QIO_close_read(qr)
-
-type Reader*[V:static[int]] = ref object
-  layout*:Layout[V]
-  latsize*:seq[cint]
-  status*:int
-  fileName*:string
-  fileMetadata*:string
-  recordInfoValid:bool
-  recordMd:string
-  recordInfo:QIO_RecordInfo
-  nioranklist*: seq[int32]
-  iogeom*: seq[int32]
-  iorank*: seq[int32]
-  qr:ptr QIO_Reader
-  verb*: cint
-
-proc IOverb*(level: int) =
-  discard QIO_verbose(level.cint)
-
-proc ioLayout(x: Layout): Layout =
-  var ioLayoutV{.global.}: Layout
-  result = ioLayoutV
-  if not x.isNil: ioLayoutV = x
-template setLayout(x: Reader): untyped =
-  discard ioLayout(x.layout)
-template getLayout(x: static[int]): untyped =
-  ioLayout(Layout[x](nil))
-
-var readnodes = -1
-var riorank: ptr seq[int32]
-proc getNumReadRanks*(): int = readnodes
-proc setNumReadRanks*(n: int) =
-  readnodes = n
-proc ioReadRank(node: cint): cint =
-  riorank[][node]
-proc ioMasterRank(): cint = 0
-
-proc toString(qs:ptr QIO_String):string =
-  let n = qs.length.int - 2  # seems to have 2 byte padding
-  result = spaces(n)
-  for i in 0..<n:
-    result[i] = qs.string[i]
-
-proc open(rd: var Reader; ql: var QIO_Layout) =
-  let nd = rd.layout.nDim.cint
-  ql.latdim = nd
-  rd.latsize.newSeq(nd)
-  for i in 0..<nd:
-    rd.latsize[i] = rd.layout.physGeom[i].cint
-  ql.latsize = rd.latsize[0].addr
-  ql.volume = rd.layout.physVol.csize_t
-  ql.sites_on_node = rd.layout.nSites.csize_t
-  ql.this_node = rd.layout.myRank.cint
-  ql.number_of_nodes = rd.layout.nRanks.cint
-
-  rd.nioranklist = listNumIoRanks(rd.layout.rankGeom)
-  if readnodes<=0:
-    let n = rd.nioranklist.len
-    let k = min(n-1, (n div 2)+1)
-    readnodes = rd.niorankList[k]
-  var rdnodes = getClosestNumRanks(rd.nioranklist, int32 readnodes)
-  echo "Read num nodes: ", rdnodes
-  rd.iogeom = getIoGeom(rd.layout.rankGeom, rdnodes)
-  echo "Read geom: ", rd.iogeom
-  rd.iorank = getIoRanks(rd.layout, rd.iogeom)
-  echo "Read IO ranks: ", rd.iorank
-  riorank = addr rd.iorank
-
-  var fs: QIO_Filesystem
-  fs.my_io_node = ioReadRank
-  fs.master_io_node = ioMasterRank
-
-  var iflag: QIO_Iflag
-  #iflag.serpar = QIO_SERIAL;
-  iflag.serpar = QIO_PARALLEL
-  #//iflag.volfmt = QIO_UNKNOWN;
-  iflag.volfmt = QIO_SINGLEFILE
-
-  var qioMd = QIO_string_create()
-  rd.qr = QIO_open_read(qioMd, cstring rd.fileName, ql.addr, fs.addr, iflag.addr)
-  rd.fileMetadata = toString(qioMd)
-  QIO_string_destroy(qioMd)
-  rd.recordInfoValid = false
-  if rd.qr==nil: rd.status = -1
-
-template newReader*[V: static[int]](l: Layout[V]; fn: string): untyped =
-  var rd: Reader[V]
-  if not fileExists(fn):
-    warn "file not found: ", fn
-  else:
-    proc ioNodeNumber2(x:ptr ConstInt):cint =
-      rankIndex(getLayout(V), cast[ptr cint](x)).rank.cint
-    proc ioNodeIndex2(x:ptr ConstInt):cint =
-      rankIndex(getLayout(V), cast[ptr cint](x)).index.cint
-    proc ioGetCoords2(x:ptr cint; node: cint; index: cint) =
-      getLayout(V).coord(x, (node,index))
-    proc ioNumSites2(node: cint):cint =
-      return getLayout(V).nSites.cint
-    var ql: QIO_Layout
-    ql.node_number = ioNodeNumber2
-    ql.node_index = ioNodeIndex2
-    ql.get_coords = ioGetCoords2
-    ql.num_sites = ioNumSites2
-    rd.new()
-    rd.layout = l
-    rd.fileName = fn
-    setLayout(rd)
-    open(rd, ql)
-  rd
-
-proc close*(rd: var Reader) =
-  rd.setLayout()
-  riorank = addr rd.iorank
-  rd.status = QIO_close_read(rd.qr)
-
-proc readRecordInfo(rd: var Reader) =
-  rd.setLayout()
-  riorank = addr rd.iorank
-  var qioMd = QIO_string_create()
-  rd.status = QIO_read_record_info(rd.qr, rd.recordInfo.addr, qioMd)
-  rd.recordMd = toString(qioMd)
-  QIO_string_destroy(qioMd);
-  rd.recordInfoValid = true
-
-proc nextRecord*(rd: var Reader) =
-  rd.setLayout()
-  riorank = addr rd.iorank
-  rd.status = QIO_next_record(rd.qr)
-  rd.recordInfoValid = false
-
-proc recordMetadata*(rd: var Reader):string =
-  rd.setLayout()
-  riorank = addr rd.iorank
-  if not rd.recordInfoValid: rd.readRecordInfo()
-  return rd.recordMd
-
-template recGet(f,t:untyped):untyped =
-  proc f*(r:var Reader):auto =
-    r.setLayout()
-    if not r.recordInfoValid: r.readRecordInfo()
-    let x = `"QIO_get_" f`(r.recordInfo.addr)
-    return t(x)
-recGet(record_date, `$`)
-recGet(datatype, `$`)
-recGet(precision, `$`)
-recGet(colors, int)
-recGet(spins, int)
-recGet(typesize, int)
-recGet(datacount, int)
-
-proc getWordSize(rd: var Reader):int =
-  #var result = 0
-  let c = precision(rd)
-  case c:
-    of "F": result = 4
-    of "D": result = 8
-    else: result = 0
-  #result
-
-import qioInternal
-
-proc put[T](buf: cstring; index: csize_t, count: cint; arg: pointer) =
-  type srcT = cArray[IOtype(T)]
-  type destT1 = cArray[T]
-  type destT = cArray[ptr destT1]
-  let src = cast[ptr srcT](buf)
-  let dest = cast[ptr destT](arg)
-  let vi = index div simdLength(T).csize_t
-  let vl = int(index mod simdLength(T).csize_t)
-  #let vlm = 1 shl vl
-  for i in 0..<count:
-    #masked(dest[i][vi], vlm) := src[i]
-    mindexed(dest[i][vi], asSimd(vl)) := src[i]
-
-proc putP[T](buf: cstring; index: csize_t, count: cint; arg: pointer) =
-  type srcT = cArray[IOtypeP(T)]
-  type destT1 = cArray[T]
-  type destT = cArray[ptr destT1]
-  let src = cast[ptr srcT](buf)
-  let dest = cast[ptr destT](arg)
-  let vi = index div simdLength(T).csize_t
-  let vl = int(index mod simdLength(T).csize_t)
-  #let vlm = 1 shl vl
-  for i in 0..<count:
-    #masked(dest[i][vi], vlm) := src[i]
-    mindexed(dest[i][vi], asSimd(vl)) := src[i]
-
-proc read[T](rd: var Reader, v: var openArray[ptr T]) =
-  rd.setLayout()
-  riorank = addr rd.iorank
-  var qioMd = QIO_string_create()
-
-  rd.readRecordInfo()
-  let recWordSize = getWordSize(rd)
-  var size = sizeOf(v[0][]) div rd.layout.nSitesInner
-  var vsize = size*v.len
-  var wordSize = sizeOf(numberType(v[0][]))
-  if wordSize==recWordSize:
-    rd.status = QIO_read(rd.qr, rd.recordInfo.addr, qioMd, put[T],
-                         vsize.csize_t, wordSize.cint, v[0].addr)
-  else:
-    vsize = (recWordSize*vsize) div wordSize
-    wordSize = recWordSize
-    rd.status = QIO_read(rd.qr, rd.recordInfo.addr, qioMd, putP[T],
-                         vsize.csize_t, wordSize.cint, v[0].addr)
-
-  rd.recordMd = toString(qioMd)
-  QIO_string_destroy(qioMd);
-  rd.recordInfoValid = true
-
-proc read*[V:static[int],T](rd: var Reader[V]; v: openArray[Field[V,T]]) =
-  let nv = v.len
-  var f = newSeq[type(v[0][0].addr)](nv)
-  for i in 0..<nv: f[i] = v[i][0].addr
-  rd.read(f)
-
-proc read*(rd: var Reader; v: Field) =
-  var f = @[ v[0].addr ]
-  rd.read(f)
-
-when isMainModule:
-  qexInit()
-  echo "rank ", myRank, "/", nRanks
-  var lat = [8,8,8,8]
-  var lo = newLayout(lat)
-  var g:array[4,type(lo.ColorMatrix())]
-  for i in 0..<4: g[i] = lo.ColorMatrix()
-  #var fn = "l88.scidac"
-  const fn = "testlat0.bin"
-  if not fileExists(fn):
-    echo "gauge file not found: ", fn
-    qexExit()
-  var rd = lo.newReader(fn)
-  echo rd.fileMetadata
-  echo rd.recordMetadata
-  echo rd.recordDate
-  echo rd.datatype
-  echo rd.precision
-  echo rd.colors
-  echo rd.spins
-  echo rd.typesize
-  echo rd.datacount
-  rd.read(g)
-  #rd.nextRecord()
-  #echo rd.status
-  #echo rd.recordMetadata
-  rd.close()
-  var tr:type(g[0][0][0,0])
-  for i in 0..<4:
-    for x in g[i].all:
-      for c in 0..<g[i][0].ncols:
-        #echo g[i][x][c,c]
-        tr += g[i][x][c,c]
-  echo tr
-  qexFinalize()
+when defined(QioReader) or not defined(QioliteReader):
+  import readerQio
+  export readerQio
+else:
+  import readerQiolite
+  export readerQiolite
diff --git a/src/io/readerQio.nim b/src/io/readerQio.nim
new file mode 100644
index 00000000..3c8f368f
--- /dev/null
+++ b/src/io/readerQio.nim
@@ -0,0 +1,282 @@
+import base
+import layout
+import qio
+import strutils
+import macros
+import field
+#import comms
+import os
+#import stdUtils
+import iocommon
+
+proc getFileLattice*(s:string):seq[int] =
+  if not fileExists(s):
+    warn "file not found: ", s
+    return
+  var ql:QIO_Layout
+  ql.this_node = myRank.cint
+  ql.number_of_nodes = nRanks.cint
+  var qs = QIO_string_create()
+  var qr = QIO_open_read(qs, s, ql.addr, nil, nil)
+  let nd = QIO_get_reader_latdim(qr)
+  #echo nd
+  let ls = QIO_get_reader_latsize(qr)
+  result.newSeq(nd)
+  for i in 0..<nd:
+    result[i] = ls[i]
+  discard QIO_close_read(qr)
+
+type Reader*[V:static[int]] = ref object
+  layout*:Layout[V]
+  latsize*:seq[cint]
+  status*:int
+  fileName*:string
+  fileMetadata*:string
+  recordInfoValid:bool
+  recordMd:string
+  recordInfo:QIO_RecordInfo
+  nioranklist*: seq[int32]
+  iogeom*: seq[int32]
+  iorank*: seq[int32]
+  qr:ptr QIO_Reader
+  verb*: cint
+
+proc IOverb*(level: int) =
+  discard QIO_verbose(level.cint)
+
+proc ioLayout(x: Layout): Layout =
+  var ioLayoutV{.global.}: Layout
+  result = ioLayoutV
+  if not x.isNil: ioLayoutV = x
+template setLayout(x: Reader): untyped =
+  discard ioLayout(x.layout)
+template getLayout(x: static[int]): untyped =
+  ioLayout(Layout[x](nil))
+
+var readnodes = -1
+var riorank: ptr seq[int32]
+proc getNumReadRanks*(): int = readnodes
+proc setNumReadRanks*(n: int) =
+  readnodes = n
+proc ioReadRank(node: cint): cint =
+  riorank[][node]
+proc ioMasterRank(): cint = 0
+
+proc toString(qs:ptr QIO_String):string =
+  let n = qs.length.int - 2  # seems to have 2 byte padding
+  result = spaces(n)
+  for i in 0..<n:
+    result[i] = qs.string[i]
+
+proc open(rd: var Reader; ql: var QIO_Layout) =
+  let nd = rd.layout.nDim.cint
+  ql.latdim = nd
+  rd.latsize.newSeq(nd)
+  for i in 0..<nd:
+    rd.latsize[i] = rd.layout.physGeom[i].cint
+  ql.latsize = rd.latsize[0].addr
+  ql.volume = rd.layout.physVol.csize_t
+  ql.sites_on_node = rd.layout.nSites.csize_t
+  ql.this_node = rd.layout.myRank.cint
+  ql.number_of_nodes = rd.layout.nRanks.cint
+
+  rd.nioranklist = listNumIoRanks(rd.layout.rankGeom)
+  if readnodes<=0:
+    let n = rd.nioranklist.len
+    let k = min(n-1, (n div 2)+1)
+    readnodes = rd.niorankList[k]
+  var rdnodes = getClosestNumRanks(rd.nioranklist, int32 readnodes)
+  echo "Read num nodes: ", rdnodes
+  rd.iogeom = getIoGeom(rd.layout.rankGeom, rdnodes)
+  echo "Read geom: ", rd.iogeom
+  rd.iorank = getIoRanks(rd.layout, rd.iogeom)
+  echo "Read IO ranks: ", rd.iorank
+  riorank = addr rd.iorank
+
+  var fs: QIO_Filesystem
+  fs.my_io_node = ioReadRank
+  fs.master_io_node = ioMasterRank
+
+  var iflag: QIO_Iflag
+  #iflag.serpar = QIO_SERIAL;
+  iflag.serpar = QIO_PARALLEL
+  #//iflag.volfmt = QIO_UNKNOWN;
+  iflag.volfmt = QIO_SINGLEFILE
+
+  var qioMd = QIO_string_create()
+  rd.qr = QIO_open_read(qioMd, cstring rd.fileName, ql.addr, fs.addr, iflag.addr)
+  rd.fileMetadata = toString(qioMd)
+  QIO_string_destroy(qioMd)
+  rd.recordInfoValid = false
+  if rd.qr==nil: rd.status = -1
+
+template newReader*[V: static[int]](l: Layout[V]; fn: string): untyped =
+  var rd: Reader[V]
+  if not fileExists(fn):
+    warn "file not found: ", fn
+  else:
+    proc ioNodeNumber2(x:ptr ConstInt):cint =
+      rankIndex(getLayout(V), cast[ptr cint](x)).rank.cint
+    proc ioNodeIndex2(x:ptr ConstInt):cint =
+      rankIndex(getLayout(V), cast[ptr cint](x)).index.cint
+    proc ioGetCoords2(x:ptr cint; node: cint; index: cint) =
+      getLayout(V).coord(x, (node,index))
+    proc ioNumSites2(node: cint):cint =
+      return getLayout(V).nSites.cint
+    var ql: QIO_Layout
+    ql.node_number = ioNodeNumber2
+    ql.node_index = ioNodeIndex2
+    ql.get_coords = ioGetCoords2
+    ql.num_sites = ioNumSites2
+    rd.new()
+    rd.layout = l
+    rd.fileName = fn
+    setLayout(rd)
+    open(rd, ql)
+  rd
+
+proc close*(rd: var Reader) =
+  rd.setLayout()
+  riorank = addr rd.iorank
+  rd.status = QIO_close_read(rd.qr)
+
+proc readRecordInfo(rd: var Reader) =
+  rd.setLayout()
+  riorank = addr rd.iorank
+  var qioMd = QIO_string_create()
+  rd.status = QIO_read_record_info(rd.qr, rd.recordInfo.addr, qioMd)
+  rd.recordMd = toString(qioMd)
+  QIO_string_destroy(qioMd);
+  rd.recordInfoValid = true
+
+proc nextRecord*(rd: var Reader) =
+  rd.setLayout()
+  riorank = addr rd.iorank
+  rd.status = QIO_next_record(rd.qr)
+  rd.recordInfoValid = false
+
+proc recordMetadata*(rd: var Reader):string =
+  rd.setLayout()
+  riorank = addr rd.iorank
+  if not rd.recordInfoValid: rd.readRecordInfo()
+  return rd.recordMd
+
+template recGet(f,t:untyped):untyped =
+  proc f*(r:var Reader):auto =
+    r.setLayout()
+    if not r.recordInfoValid: r.readRecordInfo()
+    let x = `"QIO_get_" f`(r.recordInfo.addr)
+    return t(x)
+recGet(record_date, `$`)
+recGet(datatype, `$`)
+recGet(precision, `$`)
+recGet(colors, int)
+recGet(spins, int)
+recGet(typesize, int)
+recGet(datacount, int)
+
+proc getWordSize(rd: var Reader):int =
+  #var result = 0
+  let c = precision(rd)
+  case c:
+    of "F": result = 4
+    of "D": result = 8
+    else: result = 0
+  #result
+
+import qioInternal
+
+proc put[T](buf: cstring; index: csize_t, count: cint; arg: pointer) =
+  type srcT = cArray[IOtype(T)]
+  type destT1 = cArray[T]
+  type destT = cArray[ptr destT1]
+  let src = cast[ptr srcT](buf)
+  let dest = cast[ptr destT](arg)
+  let vi = index div simdLength(T).csize_t
+  let vl = int(index mod simdLength(T).csize_t)
+  #let vlm = 1 shl vl
+  for i in 0..<count:
+    #masked(dest[i][vi], vlm) := src[i]
+    mindexed(dest[i][vi], asSimd(vl)) := src[i]
+
+proc putP[T](buf: cstring; index: csize_t, count: cint; arg: pointer) =
+  type srcT = cArray[IOtypeP(T)]
+  type destT1 = cArray[T]
+  type destT = cArray[ptr destT1]
+  let src = cast[ptr srcT](buf)
+  let dest = cast[ptr destT](arg)
+  let vi = index div simdLength(T).csize_t
+  let vl = int(index mod simdLength(T).csize_t)
+  #let vlm = 1 shl vl
+  for i in 0..<count:
+    #masked(dest[i][vi], vlm) := src[i]
+    mindexed(dest[i][vi], asSimd(vl)) := src[i]
+
+proc read[T](rd: var Reader, v: var openArray[ptr T]) =
+  rd.setLayout()
+  riorank = addr rd.iorank
+  var qioMd = QIO_string_create()
+
+  rd.readRecordInfo()
+  let recWordSize = getWordSize(rd)
+  var size = sizeOf(v[0][]) div rd.layout.nSitesInner
+  var vsize = size*v.len
+  var wordSize = sizeOf(numberType(v[0][]))
+  if wordSize==recWordSize:
+    rd.status = QIO_read(rd.qr, rd.recordInfo.addr, qioMd, put[T],
+                         vsize.csize_t, wordSize.cint, v[0].addr)
+  else:
+    vsize = (recWordSize*vsize) div wordSize
+    wordSize = recWordSize
+    rd.status = QIO_read(rd.qr, rd.recordInfo.addr, qioMd, putP[T],
+                         vsize.csize_t, wordSize.cint, v[0].addr)
+
+  rd.recordMd = toString(qioMd)
+  QIO_string_destroy(qioMd);
+  rd.recordInfoValid = true
+
+proc read*[V:static[int],T](rd: var Reader[V]; v: openArray[Field[V,T]]) =
+  let nv = v.len
+  var f = newSeq[type(v[0][0].addr)](nv)
+  for i in 0..<nv: f[i] = v[i][0].addr
+  rd.read(f)
+
+proc read*(rd: var Reader; v: Field) =
+  var f = @[ v[0].addr ]
+  rd.read(f)
+
+when isMainModule:
+  qexInit()
+  echo "rank ", myRank, "/", nRanks
+  var lat = [8,8,8,8]
+  var lo = newLayout(lat)
+  var g:array[4,type(lo.ColorMatrix())]
+  for i in 0..<4: g[i] = lo.ColorMatrix()
+  #var fn = "l88.scidac"
+  const fn = "testlat0.bin"
+  if not fileExists(fn):
+    echo "gauge file not found: ", fn
+    qexExit()
+  var rd = lo.newReader(fn)
+  echo rd.fileMetadata
+  echo rd.recordMetadata
+  echo rd.recordDate
+  echo rd.datatype
+  echo rd.precision
+  echo rd.colors
+  echo rd.spins
+  echo rd.typesize
+  echo rd.datacount
+  rd.read(g)
+  #rd.nextRecord()
+  #echo rd.status
+  #echo rd.recordMetadata
+  rd.close()
+  var tr:type(g[0][0][0,0])
+  for i in 0..<4:
+    for x in g[i].all:
+      for c in 0..<g[i][0].ncols:
+        #echo g[i][x][c,c]
+        tr += g[i][x][c,c]
+  echo tr
+  qexFinalize()
diff --git a/src/io/readerQiolite.nim b/src/io/readerQiolite.nim
new file mode 100644
index 00000000..38760393
--- /dev/null
+++ b/src/io/readerQiolite.nim
@@ -0,0 +1,225 @@
+import base
+import layout
+import strutils, strformat
+import macros
+import field
+import os
+import scidacio
+
+var verb = 0
+
+proc getFileLattice*(fn: string): seq[int] =
+  if not fileExists(fn):
+    warn "file not found: ", fn
+    return
+  var sr = newScidacReader(fn)
+  result = sr.lattice
+  sr.close
+
+type Reader*[V:static[int]] = ref object
+  layout*: Layout[V]
+  latsize*: seq[int]
+  status*: int
+  fileName*: string
+  fileMetadata*: string
+  #recordInfoValid:bool
+  #recordMd:string
+  #nioranklist*: seq[int32]
+  #iogeom*: seq[int32]
+  #iorank*: seq[int32]
+  verb*: int
+  sr*: ScidacReader
+
+proc IOverb*(level: int) =
+  verb = level
+
+proc newReader*[V: static[int]](l: Layout[V]; fn: string): Reader[V] =
+  if not fileExists(fn):
+    warn "file not found: ", fn
+    return
+  var sr = newScidacReader(fn, verb)
+  result.new
+  result.layout = l
+  result.latsize = sr.lattice
+  result.status = 0
+  result.fileName = fn
+  result.fileMetadata = sr.fileMd
+  result.verb = verb
+  result.sr = sr
+
+proc close*(rd: var Reader) =
+  rd.sr.close
+
+proc nextRecord*(rd: var Reader) =
+  rd.sr.nextRecord
+
+proc recordMetadata*(rd: var Reader): string =
+  rd.sr.recordMd
+
+#template recGet(f,t:untyped):untyped =
+#  proc f*(r: var Reader): auto =
+    #if not r.recordInfoValid: r.readRecordInfo()
+    #let x = `"QIO_get_" f`(r.recordInfo.addr)
+    #return t(x)
+#recGet(record_date, `$`)
+#recGet(datatype, `$`)
+#recGet(precision, `$`)
+#recGet(colors, int)
+#recGet(spins, int)
+#recGet(typesize, int)
+#recGet(datacount, int)
+proc recordDate*(rd: var Reader): string =
+  rd.sr.record.date
+
+proc hyperindex(x: seq[cint]; subl,offs: seq[int]): int =
+  let n = x.len
+  for i in countdown(n-1,0):
+    result = result*subl[i] + (x[i]-offs[i])
+
+import qioInternal
+
+proc put[T](buf: cstring; index: int, count: int; arg: openArray[ptr T]) =
+  type srcT = cArray[IOtype(T)]
+  type destT1 = cArray[T]
+  type destT = cArray[ptr destT1]
+  let src = cast[ptr srcT](buf)
+  let dest = cast[ptr destT](arg)
+  let vi = index div simdLength(T)
+  let vl = index mod simdLength(T)
+  #let vlm = 1 shl vl
+  for i in 0..<count:
+    #masked(dest[i][vi], vlm) := src[i]
+    #mindexed(dest[i][vi], asSimd(vl)) := src[i]
+    dest[][i][][vi][asSimd(vl)] = src[][i]
+
+proc putP[T](buf: cstring; index: int, count: int; arg: openArray[ptr T]) =
+  type srcT = cArray[IOtypeP(T)]
+  type destT1 = cArray[T]
+  type destT = cArray[ptr destT1]
+  let src = cast[ptr srcT](buf)
+  let dest = cast[ptr destT](arg)
+  let vi = index div simdLength(T)
+  let vl = index mod simdLength(T)
+  #let vlm = 1 shl vl
+  for i in 0..<count:
+    #masked(dest[i][vi], vlm) := src[i]
+    #mindexed(dest[i][vi], asSimd(vl)) := src[i]
+    dest[i][][vi][asSimd(vl)] = src[i]
+
+template `&`[T](x: seq[T]): untyped = cast[ptr T](unsafeaddr x[0])
+template `+`(x: ptr char, i: SomeInteger): untyped =
+  cast[ptr char](cast[ByteAddress](x) + ByteAddress(i))
+
+proc read[T](rd: var Reader, v: openArray[ptr T]) =
+  let t0 = getMonoTime()
+  template r: untyped = rd.sr.record
+  let nsites = rd.layout.nsites
+  let objcount = r.datacount
+  let datum_size = objcount * r.typesize
+  #echo "datum_size: ", datum_size, "  objcount: ", objcount
+  let nbytes = nsites * datum_size
+  let ioprec = if rd.sr.record.precision=="F": 1 else: 2
+  let tprec = if sizeof(numberType(T)) == 4: 1 else: 2
+  let ioelem = r.typesize / (4*ioprec)
+  let telem = sizeof(IOtype(T)) / (4*tprec)
+  #let telem = numNumbers IOtype(T)
+  if ioelem != telem:
+    qexError &"field data elements {telem} != reader data elements {ioelem}"
+  let nd = rd.layout.nDim
+  var x = newSeq[cint](nd)
+  let this_node = rd.layout.myrank
+  var sublattice = newSeq[int](nd)
+  var hypermin = rd.sr.lattice
+  var hypermax = newSeq[int](nd)
+  for i in 0 ..< nsites:
+    rd.layout.coord(x, this_node, i)
+    #echo i, " ", x
+    for j in 0..<nd:
+      hypermin[j] = min(hypermin[j], x[j])
+      hypermax[j] = max(hypermax[j], x[j])
+  #echo hypermin, " ", hypermax
+  for j in 0..<nd:
+    sublattice[j] = hypermax[j] - hypermin[j] + 1
+  var buf = create(char, nbytes)
+  let t1 = getMonoTime()
+  rd.sr.readBinary(buf, sublattice, hypermin)
+  let t2 = getMonoTime()
+  for i in countup(0'i32, nsites.int32-1):
+    rd.layout.coord(x, this_node, i)
+    let j = hyperindex(x, sublattice, hypermin)
+    let tbuf = cast[cstring](buf + j*datum_size.int)
+    if ioprec==tprec:
+      put(tbuf, i, objcount, v)
+    else:
+      putP(tbuf, i, objcount, v)
+  dealloc(buf)
+  let t3 = getMonoTime()
+  rd.sr.finishReadBinary
+  let t4 = getMonoTime()
+  if verb > 0:
+    echo &"read seconds layout: {t1-t0} read: {t2-t1} put: {t3-t2} finish: {t4-t3}"
+
+proc check(rd: Reader, l: Layout) =
+  var err = false
+  let nd = l.nDim
+  if rd.sr.lattice.len != nd:
+    err = true
+  else:
+    for i in 0..<nd:
+      if rd.sr.lattice[i] != l.physGeom[i]:
+        err = true
+  if err:
+    qexError &"field lattice {l.physGeom} != reader lattice {rd.sr.lattice}"
+
+proc read*[V:static[int],T](rd: var Reader[V]; v: openArray[Field[V,T]]) =
+  let nv = v.len
+  if rd.sr.record.datacount != nv:
+    qexError &"field count {nv} != reader count {rd.sr.record.datacount}"
+  for i in 0..<nv:
+    check(rd, v[i].l)
+  var f = newSeq[type(v[0][0].addr)](nv)
+  for i in 0..<nv: f[i] = v[i][0].addr
+  rd.read(f)
+
+proc read*(rd: var Reader; v: Field) =
+  if rd.sr.record.datacount != 1:
+    qexError &"field count 1 != reader count {rd.sr.record.datacount}"
+  check(rd, v.l)
+  var f = @[ v[0].addr ]
+  rd.read(f)
+
+when isMainModule:
+  qexInit()
+  echo "rank ", myRank, "/", nRanks
+  var lat = [8,8,8,8]
+  var lo = newLayout(lat)
+  var g:array[4,type(lo.ColorMatrix())]
+  for i in 0..<4: g[i] = lo.ColorMatrix()
+  #var fn = "l88.scidac"
+  const fn = "testlat0.bin"
+  if not fileExists(fn):
+    echo "gauge file not found: ", fn
+    qexExit()
+  var rd = lo.newReader(fn)
+  echo rd.fileMetadata
+  echo rd.recordMetadata
+  echo rd.recordDate
+  echo rd.datatype
+  echo rd.precision
+  echo rd.colors
+  echo rd.spins
+  echo rd.typesize
+  echo rd.datacount
+  rd.read(g)
+  #rd.nextRecord()
+  #echo rd.status
+  #echo rd.recordMetadata
+  rd.close()
+  var tr:type(g[0][0][0,0])
+  for i in 0..<4:
+    for x in g[i].all:
+      for c in 0..<g[i][0].ncols:
+        #echo g[i][x][c,c]
+        tr += g[i][x][c,c]
+  echo tr
+  qexFinalize()
diff --git a/src/rng/milcrng.nim b/src/rng/milcrng.nim
index 61c9b7d6..0dcb2c53 100644
--- a/src/rng/milcrng.nim
+++ b/src/rng/milcrng.nim
@@ -158,7 +158,6 @@ proc gaussian*(prn: var RngMilc6): float32 =
     r = sqrt(-2.0 * ln(v + TINY))
     result = r * cos(p)
 
-import maths/types
 # Only needed for non-vectorized RNGs.
 template gaussian*(x: var auto, r: MaskedObj[RngMilc6]) =
   mixin gaussian