diff --git a/bin/hipify-perl b/bin/hipify-perl index a5d3d574..36e0be46 100755 --- a/bin/hipify-perl +++ b/bin/hipify-perl @@ -12532,15 +12532,51 @@ sub warnHipOnlyUnsupportedFunctions { "CUBLASLT_MATMUL_TILE_64x256", "CUBLASLT_MATMUL_TILE_64x128", "CUBLASLT_MATMUL_TILE_512x64", + "CUBLASLT_MATMUL_TILE_48x768", + "CUBLASLT_MATMUL_TILE_48x704", + "CUBLASLT_MATMUL_TILE_48x640", + "CUBLASLT_MATMUL_TILE_48x64", + "CUBLASLT_MATMUL_TILE_48x576", + "CUBLASLT_MATMUL_TILE_48x512", + "CUBLASLT_MATMUL_TILE_48x448", + "CUBLASLT_MATMUL_TILE_48x384", + "CUBLASLT_MATMUL_TILE_48x320", + "CUBLASLT_MATMUL_TILE_48x256", + "CUBLASLT_MATMUL_TILE_48x192", + "CUBLASLT_MATMUL_TILE_48x128", + "CUBLASLT_MATMUL_TILE_40x768", + "CUBLASLT_MATMUL_TILE_40x704", + "CUBLASLT_MATMUL_TILE_40x640", + "CUBLASLT_MATMUL_TILE_40x64", + "CUBLASLT_MATMUL_TILE_40x576", + "CUBLASLT_MATMUL_TILE_40x512", + "CUBLASLT_MATMUL_TILE_40x448", + "CUBLASLT_MATMUL_TILE_40x384", + "CUBLASLT_MATMUL_TILE_40x320", + "CUBLASLT_MATMUL_TILE_40x256", + "CUBLASLT_MATMUL_TILE_40x192", + "CUBLASLT_MATMUL_TILE_40x128", "CUBLASLT_MATMUL_TILE_32x8", + "CUBLASLT_MATMUL_TILE_32x768", + "CUBLASLT_MATMUL_TILE_32x704", + "CUBLASLT_MATMUL_TILE_32x640", "CUBLASLT_MATMUL_TILE_32x64", + "CUBLASLT_MATMUL_TILE_32x576", + "CUBLASLT_MATMUL_TILE_32x512", + "CUBLASLT_MATMUL_TILE_32x448", + "CUBLASLT_MATMUL_TILE_32x384", + "CUBLASLT_MATMUL_TILE_32x320", "CUBLASLT_MATMUL_TILE_32x32", "CUBLASLT_MATMUL_TILE_32x256", + "CUBLASLT_MATMUL_TILE_32x192", "CUBLASLT_MATMUL_TILE_32x16", "CUBLASLT_MATMUL_TILE_32x128", "CUBLASLT_MATMUL_TILE_256x64", "CUBLASLT_MATMUL_TILE_256x32", "CUBLASLT_MATMUL_TILE_256x128", + "CUBLASLT_MATMUL_TILE_24x768", + "CUBLASLT_MATMUL_TILE_24x704", + "CUBLASLT_MATMUL_TILE_24x640", "CUBLASLT_MATMUL_TILE_24x64", "CUBLASLT_MATMUL_TILE_24x576", "CUBLASLT_MATMUL_TILE_24x512", @@ -14019,15 +14055,51 @@ sub warnRocOnlyUnsupportedFunctions { "CUBLASLT_MATMUL_TILE_64x256", "CUBLASLT_MATMUL_TILE_64x128", "CUBLASLT_MATMUL_TILE_512x64", + "CUBLASLT_MATMUL_TILE_48x768", + "CUBLASLT_MATMUL_TILE_48x704", + "CUBLASLT_MATMUL_TILE_48x640", + "CUBLASLT_MATMUL_TILE_48x64", + "CUBLASLT_MATMUL_TILE_48x576", + "CUBLASLT_MATMUL_TILE_48x512", + "CUBLASLT_MATMUL_TILE_48x448", + "CUBLASLT_MATMUL_TILE_48x384", + "CUBLASLT_MATMUL_TILE_48x320", + "CUBLASLT_MATMUL_TILE_48x256", + "CUBLASLT_MATMUL_TILE_48x192", + "CUBLASLT_MATMUL_TILE_48x128", + "CUBLASLT_MATMUL_TILE_40x768", + "CUBLASLT_MATMUL_TILE_40x704", + "CUBLASLT_MATMUL_TILE_40x640", + "CUBLASLT_MATMUL_TILE_40x64", + "CUBLASLT_MATMUL_TILE_40x576", + "CUBLASLT_MATMUL_TILE_40x512", + "CUBLASLT_MATMUL_TILE_40x448", + "CUBLASLT_MATMUL_TILE_40x384", + "CUBLASLT_MATMUL_TILE_40x320", + "CUBLASLT_MATMUL_TILE_40x256", + "CUBLASLT_MATMUL_TILE_40x192", + "CUBLASLT_MATMUL_TILE_40x128", "CUBLASLT_MATMUL_TILE_32x8", + "CUBLASLT_MATMUL_TILE_32x768", + "CUBLASLT_MATMUL_TILE_32x704", + "CUBLASLT_MATMUL_TILE_32x640", "CUBLASLT_MATMUL_TILE_32x64", + "CUBLASLT_MATMUL_TILE_32x576", + "CUBLASLT_MATMUL_TILE_32x512", + "CUBLASLT_MATMUL_TILE_32x448", + "CUBLASLT_MATMUL_TILE_32x384", + "CUBLASLT_MATMUL_TILE_32x320", "CUBLASLT_MATMUL_TILE_32x32", "CUBLASLT_MATMUL_TILE_32x256", + "CUBLASLT_MATMUL_TILE_32x192", "CUBLASLT_MATMUL_TILE_32x16", "CUBLASLT_MATMUL_TILE_32x128", "CUBLASLT_MATMUL_TILE_256x64", "CUBLASLT_MATMUL_TILE_256x32", "CUBLASLT_MATMUL_TILE_256x128", + "CUBLASLT_MATMUL_TILE_24x768", + "CUBLASLT_MATMUL_TILE_24x704", + "CUBLASLT_MATMUL_TILE_24x640", "CUBLASLT_MATMUL_TILE_24x64", "CUBLASLT_MATMUL_TILE_24x576", "CUBLASLT_MATMUL_TILE_24x512", diff --git a/docs/tables/CUBLAS_API_supported_by_HIP.md b/docs/tables/CUBLAS_API_supported_by_HIP.md index 0f3b9ad5..dee598fd 100644 --- a/docs/tables/CUBLAS_API_supported_by_HIP.md +++ b/docs/tables/CUBLAS_API_supported_by_HIP.md @@ -355,15 +355,51 @@ |`CUBLASLT_MATMUL_TILE_24x512`|12.6| | | | | | | | | | |`CUBLASLT_MATMUL_TILE_24x576`|12.6| | | | | | | | | | |`CUBLASLT_MATMUL_TILE_24x64`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_24x640`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_24x704`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_24x768`|12.6| | | | | | | | | | |`CUBLASLT_MATMUL_TILE_256x128`|10.1| | | | | | | | | | |`CUBLASLT_MATMUL_TILE_256x32`|12.1| | | | | | | | | | |`CUBLASLT_MATMUL_TILE_256x64`|10.1| | | | | | | | | | |`CUBLASLT_MATMUL_TILE_32x128`|10.1| | | | | | | | | | |`CUBLASLT_MATMUL_TILE_32x16`|10.1| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_32x192`|12.6| | | | | | | | | | |`CUBLASLT_MATMUL_TILE_32x256`|12.1| | | | | | | | | | |`CUBLASLT_MATMUL_TILE_32x32`|10.1| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_32x320`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_32x384`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_32x448`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_32x512`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_32x576`|12.6| | | | | | | | | | |`CUBLASLT_MATMUL_TILE_32x64`|10.1| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_32x640`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_32x704`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_32x768`|12.6| | | | | | | | | | |`CUBLASLT_MATMUL_TILE_32x8`|10.1| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_40x128`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_40x192`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_40x256`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_40x320`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_40x384`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_40x448`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_40x512`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_40x576`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_40x64`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_40x640`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_40x704`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_40x768`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_48x128`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_48x192`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_48x256`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_48x320`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_48x384`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_48x448`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_48x512`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_48x576`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_48x64`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_48x640`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_48x704`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_48x768`|12.6| | | | | | | | | | |`CUBLASLT_MATMUL_TILE_512x64`|10.1| | | | | | | | | | |`CUBLASLT_MATMUL_TILE_64x128`|10.1| | | | | | | | | | |`CUBLASLT_MATMUL_TILE_64x256`|10.1| | | | | | | | | | 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 b6636f5c..8b692547 100644 --- a/docs/tables/CUBLAS_API_supported_by_HIP_and_ROC.md +++ b/docs/tables/CUBLAS_API_supported_by_HIP_and_ROC.md @@ -355,15 +355,51 @@ |`CUBLASLT_MATMUL_TILE_24x512`|12.6| | | | | | | | | | | | | | | | |`CUBLASLT_MATMUL_TILE_24x576`|12.6| | | | | | | | | | | | | | | | |`CUBLASLT_MATMUL_TILE_24x64`|12.6| | | | | | | | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_24x640`|12.6| | | | | | | | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_24x704`|12.6| | | | | | | | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_24x768`|12.6| | | | | | | | | | | | | | | | |`CUBLASLT_MATMUL_TILE_256x128`|10.1| | | | | | | | | | | | | | | | |`CUBLASLT_MATMUL_TILE_256x32`|12.1| | | | | | | | | | | | | | | | |`CUBLASLT_MATMUL_TILE_256x64`|10.1| | | | | | | | | | | | | | | | |`CUBLASLT_MATMUL_TILE_32x128`|10.1| | | | | | | | | | | | | | | | |`CUBLASLT_MATMUL_TILE_32x16`|10.1| | | | | | | | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_32x192`|12.6| | | | | | | | | | | | | | | | |`CUBLASLT_MATMUL_TILE_32x256`|12.1| | | | | | | | | | | | | | | | |`CUBLASLT_MATMUL_TILE_32x32`|10.1| | | | | | | | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_32x320`|12.6| | | | | | | | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_32x384`|12.6| | | | | | | | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_32x448`|12.6| | | | | | | | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_32x512`|12.6| | | | | | | | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_32x576`|12.6| | | | | | | | | | | | | | | | |`CUBLASLT_MATMUL_TILE_32x64`|10.1| | | | | | | | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_32x640`|12.6| | | | | | | | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_32x704`|12.6| | | | | | | | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_32x768`|12.6| | | | | | | | | | | | | | | | |`CUBLASLT_MATMUL_TILE_32x8`|10.1| | | | | | | | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_40x128`|12.6| | | | | | | | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_40x192`|12.6| | | | | | | | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_40x256`|12.6| | | | | | | | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_40x320`|12.6| | | | | | | | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_40x384`|12.6| | | | | | | | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_40x448`|12.6| | | | | | | | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_40x512`|12.6| | | | | | | | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_40x576`|12.6| | | | | | | | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_40x64`|12.6| | | | | | | | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_40x640`|12.6| | | | | | | | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_40x704`|12.6| | | | | | | | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_40x768`|12.6| | | | | | | | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_48x128`|12.6| | | | | | | | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_48x192`|12.6| | | | | | | | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_48x256`|12.6| | | | | | | | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_48x320`|12.6| | | | | | | | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_48x384`|12.6| | | | | | | | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_48x448`|12.6| | | | | | | | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_48x512`|12.6| | | | | | | | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_48x576`|12.6| | | | | | | | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_48x64`|12.6| | | | | | | | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_48x640`|12.6| | | | | | | | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_48x704`|12.6| | | | | | | | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_48x768`|12.6| | | | | | | | | | | | | | | | |`CUBLASLT_MATMUL_TILE_512x64`|10.1| | | | | | | | | | | | | | | | |`CUBLASLT_MATMUL_TILE_64x128`|10.1| | | | | | | | | | | | | | | | |`CUBLASLT_MATMUL_TILE_64x256`|10.1| | | | | | | | | | | | | | | | diff --git a/docs/tables/CUBLAS_API_supported_by_ROC.md b/docs/tables/CUBLAS_API_supported_by_ROC.md index 7584c23e..f0a40e7c 100644 --- a/docs/tables/CUBLAS_API_supported_by_ROC.md +++ b/docs/tables/CUBLAS_API_supported_by_ROC.md @@ -355,15 +355,51 @@ |`CUBLASLT_MATMUL_TILE_24x512`|12.6| | | | | | | | | | |`CUBLASLT_MATMUL_TILE_24x576`|12.6| | | | | | | | | | |`CUBLASLT_MATMUL_TILE_24x64`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_24x640`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_24x704`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_24x768`|12.6| | | | | | | | | | |`CUBLASLT_MATMUL_TILE_256x128`|10.1| | | | | | | | | | |`CUBLASLT_MATMUL_TILE_256x32`|12.1| | | | | | | | | | |`CUBLASLT_MATMUL_TILE_256x64`|10.1| | | | | | | | | | |`CUBLASLT_MATMUL_TILE_32x128`|10.1| | | | | | | | | | |`CUBLASLT_MATMUL_TILE_32x16`|10.1| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_32x192`|12.6| | | | | | | | | | |`CUBLASLT_MATMUL_TILE_32x256`|12.1| | | | | | | | | | |`CUBLASLT_MATMUL_TILE_32x32`|10.1| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_32x320`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_32x384`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_32x448`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_32x512`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_32x576`|12.6| | | | | | | | | | |`CUBLASLT_MATMUL_TILE_32x64`|10.1| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_32x640`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_32x704`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_32x768`|12.6| | | | | | | | | | |`CUBLASLT_MATMUL_TILE_32x8`|10.1| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_40x128`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_40x192`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_40x256`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_40x320`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_40x384`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_40x448`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_40x512`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_40x576`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_40x64`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_40x640`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_40x704`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_40x768`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_48x128`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_48x192`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_48x256`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_48x320`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_48x384`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_48x448`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_48x512`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_48x576`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_48x64`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_48x640`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_48x704`|12.6| | | | | | | | | | +|`CUBLASLT_MATMUL_TILE_48x768`|12.6| | | | | | | | | | |`CUBLASLT_MATMUL_TILE_512x64`|10.1| | | | | | | | | | |`CUBLASLT_MATMUL_TILE_64x128`|10.1| | | | | | | | | | |`CUBLASLT_MATMUL_TILE_64x256`|10.1| | | | | | | | | | diff --git a/src/CUDA2HIP_BLAS_API_types.cpp b/src/CUDA2HIP_BLAS_API_types.cpp index 4eb7b5cf..23c90533 100644 --- a/src/CUDA2HIP_BLAS_API_types.cpp +++ b/src/CUDA2HIP_BLAS_API_types.cpp @@ -272,6 +272,42 @@ const std::map CUDA_BLAS_TYPE_NAME_MAP { {"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_24x640", {"HIPBLASLT_MATMUL_TILE_24x640", "", CONV_NUMERIC_LITERAL, API_BLAS, SEC::BLAS_LT_DATA_TYPES, UNSUPPORTED}}, + {"CUBLASLT_MATMUL_TILE_24x704", {"HIPBLASLT_MATMUL_TILE_24x704", "", CONV_NUMERIC_LITERAL, API_BLAS, SEC::BLAS_LT_DATA_TYPES, UNSUPPORTED}}, + {"CUBLASLT_MATMUL_TILE_24x768", {"HIPBLASLT_MATMUL_TILE_24x768", "", CONV_NUMERIC_LITERAL, API_BLAS, SEC::BLAS_LT_DATA_TYPES, UNSUPPORTED}}, + {"CUBLASLT_MATMUL_TILE_32x192", {"HIPBLASLT_MATMUL_TILE_32x192", "", CONV_NUMERIC_LITERAL, API_BLAS, SEC::BLAS_LT_DATA_TYPES, UNSUPPORTED}}, + {"CUBLASLT_MATMUL_TILE_32x320", {"HIPBLASLT_MATMUL_TILE_32x320", "", CONV_NUMERIC_LITERAL, API_BLAS, SEC::BLAS_LT_DATA_TYPES, UNSUPPORTED}}, + {"CUBLASLT_MATMUL_TILE_32x384", {"HIPBLASLT_MATMUL_TILE_32x384", "", CONV_NUMERIC_LITERAL, API_BLAS, SEC::BLAS_LT_DATA_TYPES, UNSUPPORTED}}, + {"CUBLASLT_MATMUL_TILE_32x448", {"HIPBLASLT_MATMUL_TILE_32x448", "", CONV_NUMERIC_LITERAL, API_BLAS, SEC::BLAS_LT_DATA_TYPES, UNSUPPORTED}}, + {"CUBLASLT_MATMUL_TILE_32x512", {"HIPBLASLT_MATMUL_TILE_32x512", "", CONV_NUMERIC_LITERAL, API_BLAS, SEC::BLAS_LT_DATA_TYPES, UNSUPPORTED}}, + {"CUBLASLT_MATMUL_TILE_32x576", {"HIPBLASLT_MATMUL_TILE_32x576", "", CONV_NUMERIC_LITERAL, API_BLAS, SEC::BLAS_LT_DATA_TYPES, UNSUPPORTED}}, + {"CUBLASLT_MATMUL_TILE_32x640", {"HIPBLASLT_MATMUL_TILE_32x640", "", CONV_NUMERIC_LITERAL, API_BLAS, SEC::BLAS_LT_DATA_TYPES, UNSUPPORTED}}, + {"CUBLASLT_MATMUL_TILE_32x704", {"HIPBLASLT_MATMUL_TILE_32x704", "", CONV_NUMERIC_LITERAL, API_BLAS, SEC::BLAS_LT_DATA_TYPES, UNSUPPORTED}}, + {"CUBLASLT_MATMUL_TILE_32x768", {"HIPBLASLT_MATMUL_TILE_32x768", "", CONV_NUMERIC_LITERAL, API_BLAS, SEC::BLAS_LT_DATA_TYPES, UNSUPPORTED}}, + {"CUBLASLT_MATMUL_TILE_40x64", {"HIPBLASLT_MATMUL_TILE_40x64", "", CONV_NUMERIC_LITERAL, API_BLAS, SEC::BLAS_LT_DATA_TYPES, UNSUPPORTED}}, + {"CUBLASLT_MATMUL_TILE_40x128", {"HIPBLASLT_MATMUL_TILE_40x128", "", CONV_NUMERIC_LITERAL, API_BLAS, SEC::BLAS_LT_DATA_TYPES, UNSUPPORTED}}, + {"CUBLASLT_MATMUL_TILE_40x192", {"HIPBLASLT_MATMUL_TILE_40x192", "", CONV_NUMERIC_LITERAL, API_BLAS, SEC::BLAS_LT_DATA_TYPES, UNSUPPORTED}}, + {"CUBLASLT_MATMUL_TILE_40x256", {"HIPBLASLT_MATMUL_TILE_40x256", "", CONV_NUMERIC_LITERAL, API_BLAS, SEC::BLAS_LT_DATA_TYPES, UNSUPPORTED}}, + {"CUBLASLT_MATMUL_TILE_40x320", {"HIPBLASLT_MATMUL_TILE_40x320", "", CONV_NUMERIC_LITERAL, API_BLAS, SEC::BLAS_LT_DATA_TYPES, UNSUPPORTED}}, + {"CUBLASLT_MATMUL_TILE_40x384", {"HIPBLASLT_MATMUL_TILE_40x384", "", CONV_NUMERIC_LITERAL, API_BLAS, SEC::BLAS_LT_DATA_TYPES, UNSUPPORTED}}, + {"CUBLASLT_MATMUL_TILE_40x448", {"HIPBLASLT_MATMUL_TILE_40x448", "", CONV_NUMERIC_LITERAL, API_BLAS, SEC::BLAS_LT_DATA_TYPES, UNSUPPORTED}}, + {"CUBLASLT_MATMUL_TILE_40x512", {"HIPBLASLT_MATMUL_TILE_40x512", "", CONV_NUMERIC_LITERAL, API_BLAS, SEC::BLAS_LT_DATA_TYPES, UNSUPPORTED}}, + {"CUBLASLT_MATMUL_TILE_40x576", {"HIPBLASLT_MATMUL_TILE_40x576", "", CONV_NUMERIC_LITERAL, API_BLAS, SEC::BLAS_LT_DATA_TYPES, UNSUPPORTED}}, + {"CUBLASLT_MATMUL_TILE_40x640", {"HIPBLASLT_MATMUL_TILE_40x640", "", CONV_NUMERIC_LITERAL, API_BLAS, SEC::BLAS_LT_DATA_TYPES, UNSUPPORTED}}, + {"CUBLASLT_MATMUL_TILE_40x704", {"HIPBLASLT_MATMUL_TILE_40x704", "", CONV_NUMERIC_LITERAL, API_BLAS, SEC::BLAS_LT_DATA_TYPES, UNSUPPORTED}}, + {"CUBLASLT_MATMUL_TILE_40x768", {"HIPBLASLT_MATMUL_TILE_40x768", "", CONV_NUMERIC_LITERAL, API_BLAS, SEC::BLAS_LT_DATA_TYPES, UNSUPPORTED}}, + {"CUBLASLT_MATMUL_TILE_48x64", {"HIPBLASLT_MATMUL_TILE_48x64", "", CONV_NUMERIC_LITERAL, API_BLAS, SEC::BLAS_LT_DATA_TYPES, UNSUPPORTED}}, + {"CUBLASLT_MATMUL_TILE_48x128", {"HIPBLASLT_MATMUL_TILE_48x128", "", CONV_NUMERIC_LITERAL, API_BLAS, SEC::BLAS_LT_DATA_TYPES, UNSUPPORTED}}, + {"CUBLASLT_MATMUL_TILE_48x192", {"HIPBLASLT_MATMUL_TILE_48x192", "", CONV_NUMERIC_LITERAL, API_BLAS, SEC::BLAS_LT_DATA_TYPES, UNSUPPORTED}}, + {"CUBLASLT_MATMUL_TILE_48x256", {"HIPBLASLT_MATMUL_TILE_48x256", "", CONV_NUMERIC_LITERAL, API_BLAS, SEC::BLAS_LT_DATA_TYPES, UNSUPPORTED}}, + {"CUBLASLT_MATMUL_TILE_48x320", {"HIPBLASLT_MATMUL_TILE_48x320", "", CONV_NUMERIC_LITERAL, API_BLAS, SEC::BLAS_LT_DATA_TYPES, UNSUPPORTED}}, + {"CUBLASLT_MATMUL_TILE_48x384", {"HIPBLASLT_MATMUL_TILE_48x384", "", CONV_NUMERIC_LITERAL, API_BLAS, SEC::BLAS_LT_DATA_TYPES, UNSUPPORTED}}, + {"CUBLASLT_MATMUL_TILE_48x448", {"HIPBLASLT_MATMUL_TILE_48x448", "", CONV_NUMERIC_LITERAL, API_BLAS, SEC::BLAS_LT_DATA_TYPES, UNSUPPORTED}}, + {"CUBLASLT_MATMUL_TILE_48x512", {"HIPBLASLT_MATMUL_TILE_48x512", "", CONV_NUMERIC_LITERAL, API_BLAS, SEC::BLAS_LT_DATA_TYPES, UNSUPPORTED}}, + {"CUBLASLT_MATMUL_TILE_48x576", {"HIPBLASLT_MATMUL_TILE_48x576", "", CONV_NUMERIC_LITERAL, API_BLAS, SEC::BLAS_LT_DATA_TYPES, UNSUPPORTED}}, + {"CUBLASLT_MATMUL_TILE_48x640", {"HIPBLASLT_MATMUL_TILE_48x640", "", CONV_NUMERIC_LITERAL, API_BLAS, SEC::BLAS_LT_DATA_TYPES, UNSUPPORTED}}, + {"CUBLASLT_MATMUL_TILE_48x704", {"HIPBLASLT_MATMUL_TILE_48x704", "", CONV_NUMERIC_LITERAL, API_BLAS, SEC::BLAS_LT_DATA_TYPES, UNSUPPORTED}}, + {"CUBLASLT_MATMUL_TILE_48x768", {"HIPBLASLT_MATMUL_TILE_48x768", "", 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}}, @@ -961,6 +997,42 @@ const std::map CUDA_BLAS_TYPE_NAME_VER_MAP { {"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 + {"CUBLASLT_MATMUL_TILE_24x640", {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_24x704", {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_24x768", {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_32x192", {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_32x320", {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_32x384", {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_32x448", {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_32x512", {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_32x576", {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_32x640", {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_32x704", {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_32x768", {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_40x64", {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_40x128", {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_40x192", {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_40x256", {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_40x320", {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_40x384", {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_40x448", {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_40x512", {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_40x576", {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_40x640", {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_40x704", {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_40x768", {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_48x64", {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_48x128", {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_48x192", {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_48x256", {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_48x320", {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_48x384", {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_48x448", {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_48x512", {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_48x576", {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_48x640", {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_48x704", {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_48x768", {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 {