[HIPIFY][6.3.0][BLAS][fix] Sync with hipBLAS and rocBLAS - Step 1…

…6 - final for `hipBLAS` + `hipblasGemm(Strided(BatchedEx}?)?_v2_64` support + Updated synthetic tests, the regenerated `hipify-perl`, and `BLAS` `CUDA2HIP` documentation + Minor fixes and annotation of `BLAS` synthetic tests + [FIX] `cublasGemm(Strided(Batched)?)?Ex` mark as API-changed since 11.0 + [TODO] [feature] `CUDA VERSION` detection by `HIPIFY` itself to support different CUDA/HIP signatures on hipification based on the CUDA version + [IMP] `rocblas_gemm(_strided(_batched)?)?_ex_64` are not yet supported [REASON] `compute_type` argument of the `rocblas_computetype` type is needed instead of the `rocblas_datatype` type + [TODO] File a ticket for `rocblas_gemm(_strided(_batched)?)?_ex_64` with `compute_type` argument of the `rocblas_computetype` type instead of the `rocblas_datatype` type
ROCm · Nov 12, 2024 · b46b461 · b46b461
1 parent fbd55d4
commit b46b461
Show file tree

Hide file tree

Showing 9 changed files with 512 additions and 433 deletions.
diff --git a/bin/hipify-perl b/bin/hipify-perl
@@ -4719,8 +4719,11 @@ sub simpleSubstitutions {
     subst("cublasDznrm2_v2", "hipblasDznrm2_v2", "library");
     subst("cublasDznrm2_v2_64", "hipblasDznrm2_v2_64", "library");
     subst("cublasGemmBatchedEx", "hipblasGemmBatchedEx_v2", "library");
+    subst("cublasGemmBatchedEx_64", "hipblasGemmBatchedEx_v2_64", "library");
     subst("cublasGemmEx", "hipblasGemmEx_v2", "library");
+    subst("cublasGemmEx_64", "hipblasGemmEx_v2_64", "library");
     subst("cublasGemmStridedBatchedEx", "hipblasGemmStridedBatchedEx_v2", "library");
+    subst("cublasGemmStridedBatchedEx_64", "hipblasGemmStridedBatchedEx_v2_64", "library");
     subst("cublasGetAtomicsMode", "hipblasGetAtomicsMode", "library");
     subst("cublasGetMathMode", "hipblasGetMathMode", "library");
     subst("cublasGetMatrix", "hipblasGetMatrix", "library");
@@ -11771,11 +11774,8 @@ sub warnHipOnlyUnsupportedFunctions {
         "cublasGetLoggerCallback",
         "cublasGetError",
         "cublasGetCudartVersion",
-        "cublasGemmStridedBatchedEx_64",
         "cublasGemmGroupedBatchedEx_64",
         "cublasGemmGroupedBatchedEx",
-        "cublasGemmEx_64",
-        "cublasGemmBatchedEx_64",
         "cublasFree",
         "cublasDtrttp",
         "cublasDtpttr",

diff --git a/docs/tables/CUBLAS_API_supported_by_HIP.md b/docs/tables/CUBLAS_API_supported_by_HIP.md
@@ -1868,12 +1868,12 @@
 |`cublasDtrsmBatched`| | | | |`hipblasDtrsmBatched`|3.2.0| | | | |
 |`cublasDtrsmBatched_64`|12.0| | | |`hipblasDtrsmBatched_64`|6.3.0| | | |6.3.0|
 |`cublasDtrttp`| | | | | | | | | | |
-|`cublasGemmBatchedEx`|9.1| | | |`hipblasGemmBatchedEx_v2`|6.0.0| | | | |
-|`cublasGemmBatchedEx_64`|12.0| | | | | | | | | |
-|`cublasGemmEx`|8.0| | | |`hipblasGemmEx_v2`|6.0.0| | | | |
-|`cublasGemmEx_64`|12.0| | | | | | | | | |
-|`cublasGemmStridedBatchedEx`|9.1| | | |`hipblasGemmStridedBatchedEx_v2`|6.0.0| | | | |
-|`cublasGemmStridedBatchedEx_64`|12.0| | | | | | | | | |
+|`cublasGemmBatchedEx`|9.1| |11.0| |`hipblasGemmBatchedEx_v2`|6.0.0| | | | |
+|`cublasGemmBatchedEx_64`|12.0| | | |`hipblasGemmBatchedEx_v2_64`|6.3.0| | | |6.3.0|
+|`cublasGemmEx`|8.0| |11.0| |`hipblasGemmEx_v2`|6.0.0| | | | |
+|`cublasGemmEx_64`|12.0| | | |`hipblasGemmEx_v2_64`|6.3.0| | | |6.3.0|
+|`cublasGemmStridedBatchedEx`|9.1| |11.0| |`hipblasGemmStridedBatchedEx_v2`|6.0.0| | | | |
+|`cublasGemmStridedBatchedEx_64`|12.0| | | |`hipblasGemmStridedBatchedEx_v2_64`|6.3.0| | | |6.3.0|
 |`cublasIamaxEx`|10.1| | | | | | | | | |
 |`cublasIamaxEx_64`|12.0| | | | | | | | | |
 |`cublasIaminEx`|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
@@ -1868,12 +1868,12 @@
 |`cublasDtrsmBatched`| | | | |`hipblasDtrsmBatched`|3.2.0| | | | |`rocblas_dtrsm_batched`|3.5.0| | | | |
 |`cublasDtrsmBatched_64`|12.0| | | |`hipblasDtrsmBatched_64`|6.3.0| | | |6.3.0|`rocblas_dtrsm_batched_64`|6.2.0| | | | |
 |`cublasDtrttp`| | | | | | | | | | | | | | | | |
-|`cublasGemmBatchedEx`|9.1| | | |`hipblasGemmBatchedEx_v2`|6.0.0| | | | |`rocblas_gemm_batched_ex`|3.5.0| | | | |
-|`cublasGemmBatchedEx_64`|12.0| | | | | | | | | | | | | | | |
-|`cublasGemmEx`|8.0| | | |`hipblasGemmEx_v2`|6.0.0| | | | |`rocblas_gemm_ex`|1.8.2| | | | |
-|`cublasGemmEx_64`|12.0| | | | | | | | | | | | | | | |
-|`cublasGemmStridedBatchedEx`|9.1| | | |`hipblasGemmStridedBatchedEx_v2`|6.0.0| | | | |`rocblas_gemm_strided_batched_ex`|1.9.0| | | | |
-|`cublasGemmStridedBatchedEx_64`|12.0| | | | | | | | | | | | | | | |
+|`cublasGemmBatchedEx`|9.1| |11.0| |`hipblasGemmBatchedEx_v2`|6.0.0| | | | |`rocblas_gemm_batched_ex`|3.5.0| | | | |
+|`cublasGemmBatchedEx_64`|12.0| | | |`hipblasGemmBatchedEx_v2_64`|6.3.0| | | |6.3.0| | | | | | |
+|`cublasGemmEx`|8.0| |11.0| |`hipblasGemmEx_v2`|6.0.0| | | | |`rocblas_gemm_ex`|1.8.2| | | | |
+|`cublasGemmEx_64`|12.0| | | |`hipblasGemmEx_v2_64`|6.3.0| | | |6.3.0| | | | | | |
+|`cublasGemmStridedBatchedEx`|9.1| |11.0| |`hipblasGemmStridedBatchedEx_v2`|6.0.0| | | | |`rocblas_gemm_strided_batched_ex`|1.9.0| | | | |
+|`cublasGemmStridedBatchedEx_64`|12.0| | | |`hipblasGemmStridedBatchedEx_v2_64`|6.3.0| | | |6.3.0| | | | | | |
 |`cublasIamaxEx`|10.1| | | | | | | | | | | | | | | |
 |`cublasIamaxEx_64`|12.0| | | | | | | | | | | | | | | |
 |`cublasIaminEx`|10.1| | | | | | | | | | | | | | | |

diff --git a/docs/tables/CUBLAS_API_supported_by_ROC.md b/docs/tables/CUBLAS_API_supported_by_ROC.md
@@ -1868,11 +1868,11 @@
 |`cublasDtrsmBatched`| | | | |`rocblas_dtrsm_batched`|3.5.0| | | | |
 |`cublasDtrsmBatched_64`|12.0| | | |`rocblas_dtrsm_batched_64`|6.2.0| | | | |
 |`cublasDtrttp`| | | | | | | | | | |
-|`cublasGemmBatchedEx`|9.1| | | |`rocblas_gemm_batched_ex`|3.5.0| | | | |
+|`cublasGemmBatchedEx`|9.1| |11.0| |`rocblas_gemm_batched_ex`|3.5.0| | | | |
 |`cublasGemmBatchedEx_64`|12.0| | | | | | | | | |
-|`cublasGemmEx`|8.0| | | |`rocblas_gemm_ex`|1.8.2| | | | |
+|`cublasGemmEx`|8.0| |11.0| |`rocblas_gemm_ex`|1.8.2| | | | |
 |`cublasGemmEx_64`|12.0| | | | | | | | | |
-|`cublasGemmStridedBatchedEx`|9.1| | | |`rocblas_gemm_strided_batched_ex`|1.9.0| | | | |
+|`cublasGemmStridedBatchedEx`|9.1| |11.0| |`rocblas_gemm_strided_batched_ex`|1.9.0| | | | |
 |`cublasGemmStridedBatchedEx_64`|12.0| | | | | | | | | |
 |`cublasIamaxEx`|10.1| | | | | | | | | |
 |`cublasIamaxEx_64`|12.0| | | | | | | | | |

diff --git a/src/CUDA2HIP_BLAS_API_functions.cpp b/src/CUDA2HIP_BLAS_API_functions.cpp
@@ -835,11 +835,11 @@ const std::map<llvm::StringRef, hipCounter> CUDA_BLAS_FUNCTION_MAP {
   {"cublasSgemmEx",                                        {"hipblasSgemmEx",                                            "",                                                   CONV_LIB_FUNC, API_BLAS, SEC::BLAS_EXT, UNSUPPORTED}},
   {"cublasSgemmEx_64",                                     {"hipblasSgemmEx_64",                                         "",                                                   CONV_LIB_FUNC, API_BLAS, SEC::BLAS_EXT, UNSUPPORTED}},
   {"cublasGemmEx",                                         {"hipblasGemmEx_v2",                                          "rocblas_gemm_ex",                                    CONV_LIB_FUNC, API_BLAS, SEC::BLAS_EXT}},
-  {"cublasGemmEx_64",                                      {"hipblasGemmEx_64",                                          "",                                                   CONV_LIB_FUNC, API_BLAS, SEC::BLAS_EXT, UNSUPPORTED}},
+  {"cublasGemmEx_64",                                      {"hipblasGemmEx_v2_64",                                       "rocblas_gemm_ex_64",                                 CONV_LIB_FUNC, API_BLAS, SEC::BLAS_EXT, ROC_UNSUPPORTED}},
   {"cublasGemmBatchedEx",                                  {"hipblasGemmBatchedEx_v2",                                   "rocblas_gemm_batched_ex",                            CONV_LIB_FUNC, API_BLAS, SEC::BLAS_EXT}},
-  {"cublasGemmBatchedEx_64",                               {"hipblasGemmBatchedEx_64",                                   "",                                                   CONV_LIB_FUNC, API_BLAS, SEC::BLAS_EXT, UNSUPPORTED}},
+  {"cublasGemmBatchedEx_64",                               {"hipblasGemmBatchedEx_v2_64",                                "rocblas_gemm_batched_ex_64",                         CONV_LIB_FUNC, API_BLAS, SEC::BLAS_EXT, ROC_UNSUPPORTED}},
   {"cublasGemmStridedBatchedEx",                           {"hipblasGemmStridedBatchedEx_v2",                            "rocblas_gemm_strided_batched_ex",                    CONV_LIB_FUNC, API_BLAS, SEC::BLAS_EXT}},
-  {"cublasGemmStridedBatchedEx_64",                        {"hipblasGemmStridedBatchedEx_64",                            "",                                                   CONV_LIB_FUNC, API_BLAS, SEC::BLAS_EXT, UNSUPPORTED}},
+  {"cublasGemmStridedBatchedEx_64",                        {"hipblasGemmStridedBatchedEx_v2_64",                         "rocblas_gemm_strided_batched_ex_64",                 CONV_LIB_FUNC, API_BLAS, SEC::BLAS_EXT, ROC_UNSUPPORTED}},
   // IO in Int8 complex/cuComplex, computation in cuComplex
   {"cublasCgemmEx",                                        {"hipblasCgemmEx",                                            "",                                                   CONV_LIB_FUNC, API_BLAS, SEC::BLAS_EXT, UNSUPPORTED}},
   {"cublasCgemmEx_64",                                     {"hipblasCgemmEx_64",                                         "",                                                   CONV_LIB_FUNC, API_BLAS, SEC::BLAS_EXT, UNSUPPORTED}},
@@ -2082,6 +2082,9 @@ const std::map<llvm::StringRef, hipAPIversions> HIP_BLAS_FUNCTION_VER_MAP {
   {"hipblasDdgmm_64",                                      {HIP_6030, HIP_0,    HIP_0,  HIP_LATEST}},
   {"hipblasCdgmm_v2_64",                                   {HIP_6030, HIP_0,    HIP_0,  HIP_LATEST}},
   {"hipblasZdgmm_v2_64",                                   {HIP_6030, HIP_0,    HIP_0,  HIP_LATEST}},
+  {"hipblasGemmEx_v2_64",                                  {HIP_6030, HIP_0,    HIP_0,  HIP_LATEST}},
+  {"hipblasGemmBatchedEx_v2_64",                           {HIP_6030, HIP_0,    HIP_0,  HIP_LATEST}},
+  {"hipblasGemmStridedBatchedEx_v2_64",                    {HIP_6030, HIP_0,    HIP_0,  HIP_LATEST}},
 
   {"rocblas_status_to_string",                             {HIP_3050, HIP_0,    HIP_0   }},
   {"rocblas_sscal",                                        {HIP_1050, HIP_0,    HIP_0   }},
@@ -2521,6 +2524,9 @@ const std::map<llvm::StringRef, hipAPIversions> HIP_BLAS_FUNCTION_VER_MAP {
   {"rocblas_ddgmm_64",                                     {HIP_6030, HIP_0,    HIP_0,  HIP_LATEST}},
   {"rocblas_cdgmm_64",                                     {HIP_6030, HIP_0,    HIP_0,  HIP_LATEST}},
   {"rocblas_zdgmm_64",                                     {HIP_6030, HIP_0,    HIP_0,  HIP_LATEST}},
+  {"rocblas_gemm_ex_64",                                   {HIP_6030, HIP_0,    HIP_0,  HIP_LATEST}},
+  {"rocblas_gemm_batched_ex_64",                           {HIP_6030, HIP_0,    HIP_0,  HIP_LATEST}},
+  {"rocblas_gemm_strided_batched_ex_64",                   {HIP_6030, HIP_0,    HIP_0,  HIP_LATEST}},
 };
 
 const std::map<llvm::StringRef, hipAPIChangedVersions> HIP_BLAS_FUNCTION_CHANGED_VER_MAP {
@@ -2534,6 +2540,9 @@ const std::map<llvm::StringRef, hipAPIChangedVersions> HIP_BLAS_FUNCTION_CHANGED
 
 const std::map<llvm::StringRef, cudaAPIChangedVersions> CUDA_BLAS_FUNCTION_CHANGED_VER_MAP {
   {"cublasLtMatmulDescCreate",                             {CUDA_110}},
+  {"cublasGemmEx",                                         {CUDA_110}},
+  {"cublasGemmBatchedEx",                                  {CUDA_110}},
+  {"cublasGemmStridedBatchedEx",                           {CUDA_110}},
 };
 
 const std::map<unsigned int, llvm::StringRef> CUDA_BLAS_API_SECTION_MAP {