From 1f16423f341ab90530cd60b7e3d32a901000c189 Mon Sep 17 00:00:00 2001 From: Evgeny Mankov Date: Fri, 18 Oct 2024 17:34:21 +0100 Subject: [PATCH] [HIPIFY][BLAS] Sync with `CUDA 12.6.2` - Step 2 - BlasLt API + Updated the regenerated `hipify-perl` and `BLAS` `CUDA2HIP` docs accordingly --- bin/hipify-perl | 64 +++++++++++++++++++ docs/tables/CUBLAS_API_supported_by_HIP.md | 32 ++++++++++ .../CUBLAS_API_supported_by_HIP_and_ROC.md | 32 ++++++++++ docs/tables/CUBLAS_API_supported_by_ROC.md | 32 ++++++++++ src/CUDA2HIP_BLAS_API_types.cpp | 64 +++++++++++++++++++ 5 files changed, 224 insertions(+) diff --git a/bin/hipify-perl b/bin/hipify-perl index 0c743d8c..a5d3d574 100755 --- a/bin/hipify-perl +++ b/bin/hipify-perl @@ -12510,9 +12510,20 @@ sub warnHipOnlyUnsupportedFunctions { "CUBLASLT_MATMUL_TILE_96x64", "CUBLASLT_MATMUL_TILE_96x128", "CUBLASLT_MATMUL_TILE_8x8", + "CUBLASLT_MATMUL_TILE_8x768", + "CUBLASLT_MATMUL_TILE_8x704", + "CUBLASLT_MATMUL_TILE_8x640", "CUBLASLT_MATMUL_TILE_8x64", + "CUBLASLT_MATMUL_TILE_8x576", + "CUBLASLT_MATMUL_TILE_8x512", + "CUBLASLT_MATMUL_TILE_8x448", + "CUBLASLT_MATMUL_TILE_8x384", + "CUBLASLT_MATMUL_TILE_8x320", "CUBLASLT_MATMUL_TILE_8x32", + "CUBLASLT_MATMUL_TILE_8x256", + "CUBLASLT_MATMUL_TILE_8x192", "CUBLASLT_MATMUL_TILE_8x16", + "CUBLASLT_MATMUL_TILE_8x128", "CUBLASLT_MATMUL_TILE_64x96", "CUBLASLT_MATMUL_TILE_64x8", "CUBLASLT_MATMUL_TILE_64x64", @@ -12530,10 +12541,31 @@ sub warnHipOnlyUnsupportedFunctions { "CUBLASLT_MATMUL_TILE_256x64", "CUBLASLT_MATMUL_TILE_256x32", "CUBLASLT_MATMUL_TILE_256x128", + "CUBLASLT_MATMUL_TILE_24x64", + "CUBLASLT_MATMUL_TILE_24x576", + "CUBLASLT_MATMUL_TILE_24x512", + "CUBLASLT_MATMUL_TILE_24x448", + "CUBLASLT_MATMUL_TILE_24x384", + "CUBLASLT_MATMUL_TILE_24x320", + "CUBLASLT_MATMUL_TILE_24x256", + "CUBLASLT_MATMUL_TILE_24x192", + "CUBLASLT_MATMUL_TILE_24x128", "CUBLASLT_MATMUL_TILE_192x128", "CUBLASLT_MATMUL_TILE_16x8", + "CUBLASLT_MATMUL_TILE_16x768", + "CUBLASLT_MATMUL_TILE_16x704", + "CUBLASLT_MATMUL_TILE_16x640", + "CUBLASLT_MATMUL_TILE_16x64", + "CUBLASLT_MATMUL_TILE_16x576", + "CUBLASLT_MATMUL_TILE_16x512", + "CUBLASLT_MATMUL_TILE_16x448", + "CUBLASLT_MATMUL_TILE_16x384", + "CUBLASLT_MATMUL_TILE_16x320", "CUBLASLT_MATMUL_TILE_16x32", + "CUBLASLT_MATMUL_TILE_16x256", + "CUBLASLT_MATMUL_TILE_16x192", "CUBLASLT_MATMUL_TILE_16x16", + "CUBLASLT_MATMUL_TILE_16x128", "CUBLASLT_MATMUL_TILE_160x128", "CUBLASLT_MATMUL_TILE_128x96", "CUBLASLT_MATMUL_TILE_128x64", @@ -13965,9 +13997,20 @@ sub warnRocOnlyUnsupportedFunctions { "CUBLASLT_MATMUL_TILE_96x64", "CUBLASLT_MATMUL_TILE_96x128", "CUBLASLT_MATMUL_TILE_8x8", + "CUBLASLT_MATMUL_TILE_8x768", + "CUBLASLT_MATMUL_TILE_8x704", + "CUBLASLT_MATMUL_TILE_8x640", "CUBLASLT_MATMUL_TILE_8x64", + "CUBLASLT_MATMUL_TILE_8x576", + "CUBLASLT_MATMUL_TILE_8x512", + "CUBLASLT_MATMUL_TILE_8x448", + "CUBLASLT_MATMUL_TILE_8x384", + "CUBLASLT_MATMUL_TILE_8x320", "CUBLASLT_MATMUL_TILE_8x32", + "CUBLASLT_MATMUL_TILE_8x256", + "CUBLASLT_MATMUL_TILE_8x192", "CUBLASLT_MATMUL_TILE_8x16", + "CUBLASLT_MATMUL_TILE_8x128", "CUBLASLT_MATMUL_TILE_64x96", "CUBLASLT_MATMUL_TILE_64x8", "CUBLASLT_MATMUL_TILE_64x64", @@ -13985,10 +14028,31 @@ sub warnRocOnlyUnsupportedFunctions { "CUBLASLT_MATMUL_TILE_256x64", "CUBLASLT_MATMUL_TILE_256x32", "CUBLASLT_MATMUL_TILE_256x128", + "CUBLASLT_MATMUL_TILE_24x64", + "CUBLASLT_MATMUL_TILE_24x576", + "CUBLASLT_MATMUL_TILE_24x512", + "CUBLASLT_MATMUL_TILE_24x448", + "CUBLASLT_MATMUL_TILE_24x384", + "CUBLASLT_MATMUL_TILE_24x320", + "CUBLASLT_MATMUL_TILE_24x256", + "CUBLASLT_MATMUL_TILE_24x192", + "CUBLASLT_MATMUL_TILE_24x128", "CUBLASLT_MATMUL_TILE_192x128", "CUBLASLT_MATMUL_TILE_16x8", + "CUBLASLT_MATMUL_TILE_16x768", + "CUBLASLT_MATMUL_TILE_16x704", + "CUBLASLT_MATMUL_TILE_16x640", + "CUBLASLT_MATMUL_TILE_16x64", + "CUBLASLT_MATMUL_TILE_16x576", + "CUBLASLT_MATMUL_TILE_16x512", + "CUBLASLT_MATMUL_TILE_16x448", + "CUBLASLT_MATMUL_TILE_16x384", + "CUBLASLT_MATMUL_TILE_16x320", "CUBLASLT_MATMUL_TILE_16x32", + "CUBLASLT_MATMUL_TILE_16x256", + "CUBLASLT_MATMUL_TILE_16x192", "CUBLASLT_MATMUL_TILE_16x16", + "CUBLASLT_MATMUL_TILE_16x128", "CUBLASLT_MATMUL_TILE_160x128", "CUBLASLT_MATMUL_TILE_128x96", "CUBLASLT_MATMUL_TILE_128x64", diff --git a/docs/tables/CUBLAS_API_supported_by_HIP.md b/docs/tables/CUBLAS_API_supported_by_HIP.md index 8a6cc80e..0f3b9ad5 100644 --- a/docs/tables/CUBLAS_API_supported_by_HIP.md +++ b/docs/tables/CUBLAS_API_supported_by_HIP.md @@ -330,10 +330,31 @@ |`CUBLASLT_MATMUL_TILE_128x64`|10.1| | | | | | | | | | |`CUBLASLT_MATMUL_TILE_128x96`|11.8| | | | | | | | | | |`CUBLASLT_MATMUL_TILE_160x128`|11.3| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_16x128`|12.6| | | | | | | | | | |`CUBLASLT_MATMUL_TILE_16x16`|10.1| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_16x192`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_16x256`|12.6| | | | | | | | | | |`CUBLASLT_MATMUL_TILE_16x32`|10.1| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_16x320`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_16x384`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_16x448`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_16x512`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_16x576`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_16x64`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_16x640`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_16x704`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_16x768`|12.6| | | | | | | | | | |`CUBLASLT_MATMUL_TILE_16x8`|10.1| | | | | | | | | | |`CUBLASLT_MATMUL_TILE_192x128`|11.3| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_24x128`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_24x192`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_24x256`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_24x320`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_24x384`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_24x448`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_24x512`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_24x576`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_24x64`|12.6| | | | | | | | | | |`CUBLASLT_MATMUL_TILE_256x128`|10.1| | | | | | | | | | |`CUBLASLT_MATMUL_TILE_256x32`|12.1| | | | | | | | | | |`CUBLASLT_MATMUL_TILE_256x64`|10.1| | | | | | | | | | @@ -351,9 +372,20 @@ |`CUBLASLT_MATMUL_TILE_64x64`|10.1| | | | | | | | | | |`CUBLASLT_MATMUL_TILE_64x8`|10.1| | | | | | | | | | |`CUBLASLT_MATMUL_TILE_64x96`|11.3| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_8x128`|12.6| | | | | | | | | | |`CUBLASLT_MATMUL_TILE_8x16`|10.1| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_8x192`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_8x256`|12.6| | | | | | | | | | |`CUBLASLT_MATMUL_TILE_8x32`|10.1| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_8x320`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_8x384`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_8x448`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_8x512`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_8x576`|12.6| | | | | | | | | | |`CUBLASLT_MATMUL_TILE_8x64`|10.1| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_8x640`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_8x704`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_8x768`|12.6| | | | | | | | | | |`CUBLASLT_MATMUL_TILE_8x8`|10.1| | | | | | | | | | |`CUBLASLT_MATMUL_TILE_96x128`|11.3| | | | | | | | | | |`CUBLASLT_MATMUL_TILE_96x64`|11.3| | | | | | | | | | diff --git a/docs/tables/CUBLAS_API_supported_by_HIP_and_ROC.md b/docs/tables/CUBLAS_API_supported_by_HIP_and_ROC.md index eacf573f..b6636f5c 100644 --- a/docs/tables/CUBLAS_API_supported_by_HIP_and_ROC.md +++ b/docs/tables/CUBLAS_API_supported_by_HIP_and_ROC.md @@ -330,10 +330,31 @@ |`CUBLASLT_MATMUL_TILE_128x64`|10.1| | | | | | | | | | | | | | | | |`CUBLASLT_MATMUL_TILE_128x96`|11.8| | | | | | | | | | | | | | | | |`CUBLASLT_MATMUL_TILE_160x128`|11.3| | | | | | | | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_16x128`|12.6| | | | | | | | | | | | | | | | |`CUBLASLT_MATMUL_TILE_16x16`|10.1| | | | | | | | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_16x192`|12.6| | | | | | | | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_16x256`|12.6| | | | | | | | | | | | | | | | |`CUBLASLT_MATMUL_TILE_16x32`|10.1| | | | | | | | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_16x320`|12.6| | | | | | | | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_16x384`|12.6| | | | | | | | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_16x448`|12.6| | | | | | | | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_16x512`|12.6| | | | | | | | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_16x576`|12.6| | | | | | | | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_16x64`|12.6| | | | | | | | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_16x640`|12.6| | | | | | | | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_16x704`|12.6| | | | | | | | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_16x768`|12.6| | | | | | | | | | | | | | | | |`CUBLASLT_MATMUL_TILE_16x8`|10.1| | | | | | | | | | | | | | | | |`CUBLASLT_MATMUL_TILE_192x128`|11.3| | | | | | | | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_24x128`|12.6| | | | | | | | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_24x192`|12.6| | | | | | | | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_24x256`|12.6| | | | | | | | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_24x320`|12.6| | | | | | | | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_24x384`|12.6| | | | | | | | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_24x448`|12.6| | | | | | | | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_24x512`|12.6| | | | | | | | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_24x576`|12.6| | | | | | | | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_24x64`|12.6| | | | | | | | | | | | | | | | |`CUBLASLT_MATMUL_TILE_256x128`|10.1| | | | | | | | | | | | | | | | |`CUBLASLT_MATMUL_TILE_256x32`|12.1| | | | | | | | | | | | | | | | |`CUBLASLT_MATMUL_TILE_256x64`|10.1| | | | | | | | | | | | | | | | @@ -351,9 +372,20 @@ |`CUBLASLT_MATMUL_TILE_64x64`|10.1| | | | | | | | | | | | | | | | |`CUBLASLT_MATMUL_TILE_64x8`|10.1| | | | | | | | | | | | | | | | |`CUBLASLT_MATMUL_TILE_64x96`|11.3| | | | | | | | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_8x128`|12.6| | | | | | | | | | | | | | | | |`CUBLASLT_MATMUL_TILE_8x16`|10.1| | | | | | | | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_8x192`|12.6| | | | | | | | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_8x256`|12.6| | | | | | | | | | | | | | | | |`CUBLASLT_MATMUL_TILE_8x32`|10.1| | | | | | | | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_8x320`|12.6| | | | | | | | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_8x384`|12.6| | | | | | | | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_8x448`|12.6| | | | | | | | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_8x512`|12.6| | | | | | | | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_8x576`|12.6| | | | | | | | | | | | | | | | |`CUBLASLT_MATMUL_TILE_8x64`|10.1| | | | | | | | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_8x640`|12.6| | | | | | | | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_8x704`|12.6| | | | | | | | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_8x768`|12.6| | | | | | | | | | | | | | | | |`CUBLASLT_MATMUL_TILE_8x8`|10.1| | | | | | | | | | | | | | | | |`CUBLASLT_MATMUL_TILE_96x128`|11.3| | | | | | | | | | | | | | | | |`CUBLASLT_MATMUL_TILE_96x64`|11.3| | | | | | | | | | | | | | | | diff --git a/docs/tables/CUBLAS_API_supported_by_ROC.md b/docs/tables/CUBLAS_API_supported_by_ROC.md index 20cbb3db..7584c23e 100644 --- a/docs/tables/CUBLAS_API_supported_by_ROC.md +++ b/docs/tables/CUBLAS_API_supported_by_ROC.md @@ -330,10 +330,31 @@ |`CUBLASLT_MATMUL_TILE_128x64`|10.1| | | | | | | | | | |`CUBLASLT_MATMUL_TILE_128x96`|11.8| | | | | | | | | | |`CUBLASLT_MATMUL_TILE_160x128`|11.3| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_16x128`|12.6| | | | | | | | | | |`CUBLASLT_MATMUL_TILE_16x16`|10.1| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_16x192`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_16x256`|12.6| | | | | | | | | | |`CUBLASLT_MATMUL_TILE_16x32`|10.1| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_16x320`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_16x384`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_16x448`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_16x512`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_16x576`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_16x64`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_16x640`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_16x704`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_16x768`|12.6| | | | | | | | | | |`CUBLASLT_MATMUL_TILE_16x8`|10.1| | | | | | | | | | |`CUBLASLT_MATMUL_TILE_192x128`|11.3| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_24x128`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_24x192`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_24x256`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_24x320`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_24x384`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_24x448`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_24x512`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_24x576`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_24x64`|12.6| | | | | | | | | | |`CUBLASLT_MATMUL_TILE_256x128`|10.1| | | | | | | | | | |`CUBLASLT_MATMUL_TILE_256x32`|12.1| | | | | | | | | | |`CUBLASLT_MATMUL_TILE_256x64`|10.1| | | | | | | | | | @@ -351,9 +372,20 @@ |`CUBLASLT_MATMUL_TILE_64x64`|10.1| | | | | | | | | | |`CUBLASLT_MATMUL_TILE_64x8`|10.1| | | | | | | | | | |`CUBLASLT_MATMUL_TILE_64x96`|11.3| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_8x128`|12.6| | | | | | | | | | |`CUBLASLT_MATMUL_TILE_8x16`|10.1| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_8x192`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_8x256`|12.6| | | | | | | | | | |`CUBLASLT_MATMUL_TILE_8x32`|10.1| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_8x320`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_8x384`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_8x448`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_8x512`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_8x576`|12.6| | | | | | | | | | |`CUBLASLT_MATMUL_TILE_8x64`|10.1| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_8x640`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_8x704`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_8x768`|12.6| | | | | | | | | | |`CUBLASLT_MATMUL_TILE_8x8`|10.1| | | | | | | | | | |`CUBLASLT_MATMUL_TILE_96x128`|11.3| | | | | | | | | | |`CUBLASLT_MATMUL_TILE_96x64`|11.3| | | | | | | | | | diff --git a/src/CUDA2HIP_BLAS_API_types.cpp b/src/CUDA2HIP_BLAS_API_types.cpp index 78321ce6..4eb7b5cf 100644 --- a/src/CUDA2HIP_BLAS_API_types.cpp +++ b/src/CUDA2HIP_BLAS_API_types.cpp @@ -240,6 +240,38 @@ const std::map CUDA_BLAS_TYPE_NAME_MAP { {"CUBLASLT_MATMUL_TILE_128x96", {"HIPBLASLT_MATMUL_TILE_128x96", "", CONV_NUMERIC_LITERAL, API_BLAS, SEC::BLAS_LT_DATA_TYPES, UNSUPPORTED}}, {"CUBLASLT_MATMUL_TILE_32x256", {"HIPBLASLT_MATMUL_TILE_32x256", "", CONV_NUMERIC_LITERAL, API_BLAS, SEC::BLAS_LT_DATA_TYPES, UNSUPPORTED}}, {"CUBLASLT_MATMUL_TILE_256x32", {"HIPBLASLT_MATMUL_TILE_256x32", "", CONV_NUMERIC_LITERAL, API_BLAS, SEC::BLAS_LT_DATA_TYPES, UNSUPPORTED}}, + {"CUBLASLT_MATMUL_TILE_8x128", {"HIPBLASLT_MATMUL_TILE_8x128", "", CONV_NUMERIC_LITERAL, API_BLAS, SEC::BLAS_LT_DATA_TYPES, UNSUPPORTED}}, + {"CUBLASLT_MATMUL_TILE_8x192", {"HIPBLASLT_MATMUL_TILE_8x192", "", CONV_NUMERIC_LITERAL, API_BLAS, SEC::BLAS_LT_DATA_TYPES, UNSUPPORTED}}, + {"CUBLASLT_MATMUL_TILE_8x256", {"HIPBLASLT_MATMUL_TILE_8x256", "", CONV_NUMERIC_LITERAL, API_BLAS, SEC::BLAS_LT_DATA_TYPES, UNSUPPORTED}}, + {"CUBLASLT_MATMUL_TILE_8x320", {"HIPBLASLT_MATMUL_TILE_8x320", "", CONV_NUMERIC_LITERAL, API_BLAS, SEC::BLAS_LT_DATA_TYPES, UNSUPPORTED}}, + {"CUBLASLT_MATMUL_TILE_8x384", {"HIPBLASLT_MATMUL_TILE_8x384", "", CONV_NUMERIC_LITERAL, API_BLAS, SEC::BLAS_LT_DATA_TYPES, UNSUPPORTED}}, + {"CUBLASLT_MATMUL_TILE_8x448", {"HIPBLASLT_MATMUL_TILE_8x448", "", CONV_NUMERIC_LITERAL, API_BLAS, SEC::BLAS_LT_DATA_TYPES, UNSUPPORTED}}, + {"CUBLASLT_MATMUL_TILE_8x512", {"HIPBLASLT_MATMUL_TILE_8x512", "", CONV_NUMERIC_LITERAL, API_BLAS, SEC::BLAS_LT_DATA_TYPES, UNSUPPORTED}}, + {"CUBLASLT_MATMUL_TILE_8x576", {"HIPBLASLT_MATMUL_TILE_8x576", "", CONV_NUMERIC_LITERAL, API_BLAS, SEC::BLAS_LT_DATA_TYPES, UNSUPPORTED}}, + {"CUBLASLT_MATMUL_TILE_8x640", {"HIPBLASLT_MATMUL_TILE_8x640", "", CONV_NUMERIC_LITERAL, API_BLAS, SEC::BLAS_LT_DATA_TYPES, UNSUPPORTED}}, + {"CUBLASLT_MATMUL_TILE_8x704", {"HIPBLASLT_MATMUL_TILE_8x704", "", CONV_NUMERIC_LITERAL, API_BLAS, SEC::BLAS_LT_DATA_TYPES, UNSUPPORTED}}, + {"CUBLASLT_MATMUL_TILE_8x768", {"HIPBLASLT_MATMUL_TILE_8x768", "", CONV_NUMERIC_LITERAL, API_BLAS, SEC::BLAS_LT_DATA_TYPES, UNSUPPORTED}}, + {"CUBLASLT_MATMUL_TILE_16x64", {"HIPBLASLT_MATMUL_TILE_16x64", "", CONV_NUMERIC_LITERAL, API_BLAS, SEC::BLAS_LT_DATA_TYPES, UNSUPPORTED}}, + {"CUBLASLT_MATMUL_TILE_16x128", {"HIPBLASLT_MATMUL_TILE_16x128", "", CONV_NUMERIC_LITERAL, API_BLAS, SEC::BLAS_LT_DATA_TYPES, UNSUPPORTED}}, + {"CUBLASLT_MATMUL_TILE_16x192", {"HIPBLASLT_MATMUL_TILE_16x192", "", CONV_NUMERIC_LITERAL, API_BLAS, SEC::BLAS_LT_DATA_TYPES, UNSUPPORTED}}, + {"CUBLASLT_MATMUL_TILE_16x256", {"HIPBLASLT_MATMUL_TILE_16x256", "", CONV_NUMERIC_LITERAL, API_BLAS, SEC::BLAS_LT_DATA_TYPES, UNSUPPORTED}}, + {"CUBLASLT_MATMUL_TILE_16x320", {"HIPBLASLT_MATMUL_TILE_16x320", "", CONV_NUMERIC_LITERAL, API_BLAS, SEC::BLAS_LT_DATA_TYPES, UNSUPPORTED}}, + {"CUBLASLT_MATMUL_TILE_16x384", {"HIPBLASLT_MATMUL_TILE_16x384", "", CONV_NUMERIC_LITERAL, API_BLAS, SEC::BLAS_LT_DATA_TYPES, UNSUPPORTED}}, + {"CUBLASLT_MATMUL_TILE_16x448", {"HIPBLASLT_MATMUL_TILE_16x448", "", CONV_NUMERIC_LITERAL, API_BLAS, SEC::BLAS_LT_DATA_TYPES, UNSUPPORTED}}, + {"CUBLASLT_MATMUL_TILE_16x512", {"HIPBLASLT_MATMUL_TILE_16x512", "", CONV_NUMERIC_LITERAL, API_BLAS, SEC::BLAS_LT_DATA_TYPES, UNSUPPORTED}}, + {"CUBLASLT_MATMUL_TILE_16x576", {"HIPBLASLT_MATMUL_TILE_16x576", "", CONV_NUMERIC_LITERAL, API_BLAS, SEC::BLAS_LT_DATA_TYPES, UNSUPPORTED}}, + {"CUBLASLT_MATMUL_TILE_16x640", {"HIPBLASLT_MATMUL_TILE_16x640", "", CONV_NUMERIC_LITERAL, API_BLAS, SEC::BLAS_LT_DATA_TYPES, UNSUPPORTED}}, + {"CUBLASLT_MATMUL_TILE_16x704", {"HIPBLASLT_MATMUL_TILE_16x704", "", CONV_NUMERIC_LITERAL, API_BLAS, SEC::BLAS_LT_DATA_TYPES, UNSUPPORTED}}, + {"CUBLASLT_MATMUL_TILE_16x768", {"HIPBLASLT_MATMUL_TILE_16x768", "", CONV_NUMERIC_LITERAL, API_BLAS, SEC::BLAS_LT_DATA_TYPES, UNSUPPORTED}}, + {"CUBLASLT_MATMUL_TILE_24x64", {"HIPBLASLT_MATMUL_TILE_24x64", "", CONV_NUMERIC_LITERAL, API_BLAS, SEC::BLAS_LT_DATA_TYPES, UNSUPPORTED}}, + {"CUBLASLT_MATMUL_TILE_24x128", {"HIPBLASLT_MATMUL_TILE_24x128", "", CONV_NUMERIC_LITERAL, API_BLAS, SEC::BLAS_LT_DATA_TYPES, UNSUPPORTED}}, + {"CUBLASLT_MATMUL_TILE_24x192", {"HIPBLASLT_MATMUL_TILE_24x192", "", CONV_NUMERIC_LITERAL, API_BLAS, SEC::BLAS_LT_DATA_TYPES, UNSUPPORTED}}, + {"CUBLASLT_MATMUL_TILE_24x256", {"HIPBLASLT_MATMUL_TILE_24x256", "", CONV_NUMERIC_LITERAL, API_BLAS, SEC::BLAS_LT_DATA_TYPES, UNSUPPORTED}}, + {"CUBLASLT_MATMUL_TILE_24x320", {"HIPBLASLT_MATMUL_TILE_24x320", "", CONV_NUMERIC_LITERAL, API_BLAS, SEC::BLAS_LT_DATA_TYPES, UNSUPPORTED}}, + {"CUBLASLT_MATMUL_TILE_24x384", {"HIPBLASLT_MATMUL_TILE_24x384", "", CONV_NUMERIC_LITERAL, API_BLAS, SEC::BLAS_LT_DATA_TYPES, UNSUPPORTED}}, + {"CUBLASLT_MATMUL_TILE_24x448", {"HIPBLASLT_MATMUL_TILE_24x448", "", CONV_NUMERIC_LITERAL, API_BLAS, SEC::BLAS_LT_DATA_TYPES, UNSUPPORTED}}, + {"CUBLASLT_MATMUL_TILE_24x512", {"HIPBLASLT_MATMUL_TILE_24x512", "", CONV_NUMERIC_LITERAL, API_BLAS, SEC::BLAS_LT_DATA_TYPES, UNSUPPORTED}}, + {"CUBLASLT_MATMUL_TILE_24x576", {"HIPBLASLT_MATMUL_TILE_24x576", "", CONV_NUMERIC_LITERAL, API_BLAS, SEC::BLAS_LT_DATA_TYPES, UNSUPPORTED}}, {"CUBLASLT_MATMUL_TILE_END", {"HIPBLASLT_MATMUL_TILE_END", "", CONV_NUMERIC_LITERAL, API_BLAS, SEC::BLAS_LT_DATA_TYPES, UNSUPPORTED}}, {"cublasLtMatmulStages_t", {"hipblasLtMatmulStages_t", "", CONV_TYPE, API_BLAS, SEC::BLAS_LT_DATA_TYPES, UNSUPPORTED}}, {"CUBLASLT_MATMUL_STAGES_UNDEFINED", {"HIPBLASLT_MATMUL_STAGES_UNDEFINED", "", CONV_NUMERIC_LITERAL, API_BLAS, SEC::BLAS_LT_DATA_TYPES, UNSUPPORTED}}, @@ -897,6 +929,38 @@ const std::map CUDA_BLAS_TYPE_NAME_VER_MAP { {"CUBLASLT_ALGO_CONFIG_INNER_SHAPE_ID", {CUDA_118, CUDA_0, CUDA_0 }}, {"CUBLASLT_ALGO_CONFIG_CLUSTER_SHAPE_ID", {CUDA_118, CUDA_0, CUDA_0 }}, {"cublasLtLoggerCallback_t", {CUDA_110, CUDA_0, CUDA_0 }}, // A: CUDA_VERSION 11003, CUBLAS_VERSION 11200, CUBLAS_VER_MAJOR 11 CUBLAS_VER_MINOR 2 + {"CUBLASLT_MATMUL_TILE_8x128", {CUDA_126, CUDA_0, CUDA_0 }}, // A: CUDA_VERSION 12062, CUBLAS_VERSION 120603, CUBLAS_VER_MAJOR 12 CUBLAS_VER_MINOR 6 CUBLAS_VER_PATCH 3 + {"CUBLASLT_MATMUL_TILE_8x192", {CUDA_126, CUDA_0, CUDA_0 }}, // A: CUDA_VERSION 12062, CUBLAS_VERSION 120603, CUBLAS_VER_MAJOR 12 CUBLAS_VER_MINOR 6 CUBLAS_VER_PATCH 3 + {"CUBLASLT_MATMUL_TILE_8x256", {CUDA_126, CUDA_0, CUDA_0 }}, // A: CUDA_VERSION 12062, CUBLAS_VERSION 120603, CUBLAS_VER_MAJOR 12 CUBLAS_VER_MINOR 6 CUBLAS_VER_PATCH 3 + {"CUBLASLT_MATMUL_TILE_8x320", {CUDA_126, CUDA_0, CUDA_0 }}, // A: CUDA_VERSION 12062, CUBLAS_VERSION 120603, CUBLAS_VER_MAJOR 12 CUBLAS_VER_MINOR 6 CUBLAS_VER_PATCH 3 + {"CUBLASLT_MATMUL_TILE_8x384", {CUDA_126, CUDA_0, CUDA_0 }}, // A: CUDA_VERSION 12062, CUBLAS_VERSION 120603, CUBLAS_VER_MAJOR 12 CUBLAS_VER_MINOR 6 CUBLAS_VER_PATCH 3 + {"CUBLASLT_MATMUL_TILE_8x448", {CUDA_126, CUDA_0, CUDA_0 }}, // A: CUDA_VERSION 12062, CUBLAS_VERSION 120603, CUBLAS_VER_MAJOR 12 CUBLAS_VER_MINOR 6 CUBLAS_VER_PATCH 3 + {"CUBLASLT_MATMUL_TILE_8x512", {CUDA_126, CUDA_0, CUDA_0 }}, // A: CUDA_VERSION 12062, CUBLAS_VERSION 120603, CUBLAS_VER_MAJOR 12 CUBLAS_VER_MINOR 6 CUBLAS_VER_PATCH 3 + {"CUBLASLT_MATMUL_TILE_8x576", {CUDA_126, CUDA_0, CUDA_0 }}, // A: CUDA_VERSION 12062, CUBLAS_VERSION 120603, CUBLAS_VER_MAJOR 12 CUBLAS_VER_MINOR 6 CUBLAS_VER_PATCH 3 + {"CUBLASLT_MATMUL_TILE_8x640", {CUDA_126, CUDA_0, CUDA_0 }}, // A: CUDA_VERSION 12062, CUBLAS_VERSION 120603, CUBLAS_VER_MAJOR 12 CUBLAS_VER_MINOR 6 CUBLAS_VER_PATCH 3 + {"CUBLASLT_MATMUL_TILE_8x704", {CUDA_126, CUDA_0, CUDA_0 }}, // A: CUDA_VERSION 12062, CUBLAS_VERSION 120603, CUBLAS_VER_MAJOR 12 CUBLAS_VER_MINOR 6 CUBLAS_VER_PATCH 3 + {"CUBLASLT_MATMUL_TILE_8x768", {CUDA_126, CUDA_0, CUDA_0 }}, // A: CUDA_VERSION 12062, CUBLAS_VERSION 120603, CUBLAS_VER_MAJOR 12 CUBLAS_VER_MINOR 6 CUBLAS_VER_PATCH 3 + {"CUBLASLT_MATMUL_TILE_16x64", {CUDA_126, CUDA_0, CUDA_0 }}, // A: CUDA_VERSION 12062, CUBLAS_VERSION 120603, CUBLAS_VER_MAJOR 12 CUBLAS_VER_MINOR 6 CUBLAS_VER_PATCH 3 + {"CUBLASLT_MATMUL_TILE_16x128", {CUDA_126, CUDA_0, CUDA_0 }}, // A: CUDA_VERSION 12062, CUBLAS_VERSION 120603, CUBLAS_VER_MAJOR 12 CUBLAS_VER_MINOR 6 CUBLAS_VER_PATCH 3 + {"CUBLASLT_MATMUL_TILE_16x192", {CUDA_126, CUDA_0, CUDA_0 }}, // A: CUDA_VERSION 12062, CUBLAS_VERSION 120603, CUBLAS_VER_MAJOR 12 CUBLAS_VER_MINOR 6 CUBLAS_VER_PATCH 3 + {"CUBLASLT_MATMUL_TILE_16x256", {CUDA_126, CUDA_0, CUDA_0 }}, // A: CUDA_VERSION 12062, CUBLAS_VERSION 120603, CUBLAS_VER_MAJOR 12 CUBLAS_VER_MINOR 6 CUBLAS_VER_PATCH 3 + {"CUBLASLT_MATMUL_TILE_16x320", {CUDA_126, CUDA_0, CUDA_0 }}, // A: CUDA_VERSION 12062, CUBLAS_VERSION 120603, CUBLAS_VER_MAJOR 12 CUBLAS_VER_MINOR 6 CUBLAS_VER_PATCH 3 + {"CUBLASLT_MATMUL_TILE_16x384", {CUDA_126, CUDA_0, CUDA_0 }}, // A: CUDA_VERSION 12062, CUBLAS_VERSION 120603, CUBLAS_VER_MAJOR 12 CUBLAS_VER_MINOR 6 CUBLAS_VER_PATCH 3 + {"CUBLASLT_MATMUL_TILE_16x448", {CUDA_126, CUDA_0, CUDA_0 }}, // A: CUDA_VERSION 12062, CUBLAS_VERSION 120603, CUBLAS_VER_MAJOR 12 CUBLAS_VER_MINOR 6 CUBLAS_VER_PATCH 3 + {"CUBLASLT_MATMUL_TILE_16x512", {CUDA_126, CUDA_0, CUDA_0 }}, // A: CUDA_VERSION 12062, CUBLAS_VERSION 120603, CUBLAS_VER_MAJOR 12 CUBLAS_VER_MINOR 6 CUBLAS_VER_PATCH 3 + {"CUBLASLT_MATMUL_TILE_16x576", {CUDA_126, CUDA_0, CUDA_0 }}, // A: CUDA_VERSION 12062, CUBLAS_VERSION 120603, CUBLAS_VER_MAJOR 12 CUBLAS_VER_MINOR 6 CUBLAS_VER_PATCH 3 + {"CUBLASLT_MATMUL_TILE_16x640", {CUDA_126, CUDA_0, CUDA_0 }}, // A: CUDA_VERSION 12062, CUBLAS_VERSION 120603, CUBLAS_VER_MAJOR 12 CUBLAS_VER_MINOR 6 CUBLAS_VER_PATCH 3 + {"CUBLASLT_MATMUL_TILE_16x704", {CUDA_126, CUDA_0, CUDA_0 }}, // A: CUDA_VERSION 12062, CUBLAS_VERSION 120603, CUBLAS_VER_MAJOR 12 CUBLAS_VER_MINOR 6 CUBLAS_VER_PATCH 3 + {"CUBLASLT_MATMUL_TILE_16x768", {CUDA_126, CUDA_0, CUDA_0 }}, // A: CUDA_VERSION 12062, CUBLAS_VERSION 120603, CUBLAS_VER_MAJOR 12 CUBLAS_VER_MINOR 6 CUBLAS_VER_PATCH 3 + {"CUBLASLT_MATMUL_TILE_24x64", {CUDA_126, CUDA_0, CUDA_0 }}, // A: CUDA_VERSION 12062, CUBLAS_VERSION 120603, CUBLAS_VER_MAJOR 12 CUBLAS_VER_MINOR 6 CUBLAS_VER_PATCH 3 + {"CUBLASLT_MATMUL_TILE_24x128", {CUDA_126, CUDA_0, CUDA_0 }}, // A: CUDA_VERSION 12062, CUBLAS_VERSION 120603, CUBLAS_VER_MAJOR 12 CUBLAS_VER_MINOR 6 CUBLAS_VER_PATCH 3 + {"CUBLASLT_MATMUL_TILE_24x192", {CUDA_126, CUDA_0, CUDA_0 }}, // A: CUDA_VERSION 12062, CUBLAS_VERSION 120603, CUBLAS_VER_MAJOR 12 CUBLAS_VER_MINOR 6 CUBLAS_VER_PATCH 3 + {"CUBLASLT_MATMUL_TILE_24x256", {CUDA_126, CUDA_0, CUDA_0 }}, // A: CUDA_VERSION 12062, CUBLAS_VERSION 120603, CUBLAS_VER_MAJOR 12 CUBLAS_VER_MINOR 6 CUBLAS_VER_PATCH 3 + {"CUBLASLT_MATMUL_TILE_24x320", {CUDA_126, CUDA_0, CUDA_0 }}, // A: CUDA_VERSION 12062, CUBLAS_VERSION 120603, CUBLAS_VER_MAJOR 12 CUBLAS_VER_MINOR 6 CUBLAS_VER_PATCH 3 + {"CUBLASLT_MATMUL_TILE_24x384", {CUDA_126, CUDA_0, CUDA_0 }}, // A: CUDA_VERSION 12062, CUBLAS_VERSION 120603, CUBLAS_VER_MAJOR 12 CUBLAS_VER_MINOR 6 CUBLAS_VER_PATCH 3 + {"CUBLASLT_MATMUL_TILE_24x448", {CUDA_126, CUDA_0, CUDA_0 }}, // A: CUDA_VERSION 12062, CUBLAS_VERSION 120603, CUBLAS_VER_MAJOR 12 CUBLAS_VER_MINOR 6 CUBLAS_VER_PATCH 3 + {"CUBLASLT_MATMUL_TILE_24x512", {CUDA_126, CUDA_0, CUDA_0 }}, // A: CUDA_VERSION 12062, CUBLAS_VERSION 120603, CUBLAS_VER_MAJOR 12 CUBLAS_VER_MINOR 6 CUBLAS_VER_PATCH 3 + {"CUBLASLT_MATMUL_TILE_24x576", {CUDA_126, CUDA_0, CUDA_0 }}, // A: CUDA_VERSION 12062, CUBLAS_VERSION 120603, CUBLAS_VER_MAJOR 12 CUBLAS_VER_MINOR 6 CUBLAS_VER_PATCH 3 }; const std::map HIP_BLAS_TYPE_NAME_VER_MAP {