Merge pull request #1676 from emankov/HIPIFY

[HIPIFY][rocBLAS] 64-bit functions support - Step 5
ROCm · Oct 2, 2024 · e1fe277 · e1fe277
2 parents eebaece + 81c200a
commit e1fe277
Show file tree

Hide file tree

Showing 5 changed files with 65 additions and 32 deletions.
diff --git a/bin/hipify-perl b/bin/hipify-perl
@@ -1571,7 +1571,9 @@ sub rocSubstitutions {
     subst("cublasCgeru", "rocblas_cgeru", "library");
     subst("cublasCgeru_v2", "rocblas_cgeru", "library");
     subst("cublasChbmv", "rocblas_chbmv", "library");
+    subst("cublasChbmv_64", "rocblas_chbmv_64", "library");
     subst("cublasChbmv_v2", "rocblas_chbmv", "library");
+    subst("cublasChbmv_v2_64", "rocblas_chbmv_64", "library");
     subst("cublasChemm", "rocblas_chemm", "library");
     subst("cublasChemm_v2", "rocblas_chemm", "library");
     subst("cublasChemv", "rocblas_chemv", "library");
@@ -1704,7 +1706,9 @@ sub rocSubstitutions {
     subst("cublasDrotmg", "rocblas_drotmg", "library");
     subst("cublasDrotmg_v2", "rocblas_drotmg", "library");
     subst("cublasDsbmv", "rocblas_dsbmv", "library");
+    subst("cublasDsbmv_64", "rocblas_dsbmv_64", "library");
     subst("cublasDsbmv_v2", "rocblas_dsbmv", "library");
+    subst("cublasDsbmv_v2_64", "rocblas_dsbmv_64", "library");
     subst("cublasDscal", "rocblas_dscal", "library");
     subst("cublasDscal_64", "rocblas_dscal_64", "library");
     subst("cublasDscal_v2", "rocblas_dscal", "library");
@@ -1892,7 +1896,9 @@ sub rocSubstitutions {
     subst("cublasSrotmg", "rocblas_srotmg", "library");
     subst("cublasSrotmg_v2", "rocblas_srotmg", "library");
     subst("cublasSsbmv", "rocblas_ssbmv", "library");
+    subst("cublasSsbmv_64", "rocblas_ssbmv_64", "library");
     subst("cublasSsbmv_v2", "rocblas_ssbmv", "library");
+    subst("cublasSsbmv_v2_64", "rocblas_ssbmv_64", "library");
     subst("cublasSscal", "rocblas_sscal", "library");
     subst("cublasSscal_64", "rocblas_sscal_64", "library");
     subst("cublasSscal_v2", "rocblas_sscal", "library");
@@ -1992,7 +1998,9 @@ sub rocSubstitutions {
     subst("cublasZgeru", "rocblas_zgeru", "library");
     subst("cublasZgeru_v2", "rocblas_zgeru", "library");
     subst("cublasZhbmv", "rocblas_zhbmv", "library");
+    subst("cublasZhbmv_64", "rocblas_zhbmv_64", "library");
     subst("cublasZhbmv_v2", "rocblas_zhbmv", "library");
+    subst("cublasZhbmv_v2_64", "rocblas_zhbmv_64", "library");
     subst("cublasZhemm", "rocblas_zhemm", "library");
     subst("cublasZhemm_v2", "rocblas_zhemm", "library");
     subst("cublasZhemv", "rocblas_zhemv", "library");
@@ -12518,8 +12526,6 @@ sub warnRocOnlyUnsupportedFunctions {
         "cublasZhemv_64",
         "cublasZhemm_v2_64",
         "cublasZhemm_64",
-        "cublasZhbmv_v2_64",
-        "cublasZhbmv_64",
         "cublasZgetrsBatched",
         "cublasZgetriBatched",
         "cublasZgetrfBatched",
@@ -12579,8 +12585,6 @@ sub warnRocOnlyUnsupportedFunctions {
         "cublasSspr2_64",
         "cublasSspmv_v2_64",
         "cublasSspmv_64",
-        "cublasSsbmv_v2_64",
-        "cublasSsbmv_64",
         "cublasSmatinvBatched",
         "cublasShutdown",
         "cublasSgetrsBatched",
@@ -12740,8 +12744,6 @@ sub warnRocOnlyUnsupportedFunctions {
         "cublasDspr2_64",
         "cublasDspmv_v2_64",
         "cublasDspmv_64",
-        "cublasDsbmv_v2_64",
-        "cublasDsbmv_64",
         "cublasDmatinvBatched",
         "cublasDgetrsBatched",
         "cublasDgetriBatched",
@@ -12820,8 +12822,6 @@ sub warnRocOnlyUnsupportedFunctions {
         "cublasChemv_64",
         "cublasChemm_v2_64",
         "cublasChemm_64",
-        "cublasChbmv_v2_64",
-        "cublasChbmv_64",
         "cublasCgetrsBatched",
         "cublasCgetriBatched",
         "cublasCgetrfBatched",

diff --git a/docs/tables/CUBLAS_API_supported_by_HIP_and_ROC.md b/docs/tables/CUBLAS_API_supported_by_HIP_and_ROC.md
@@ -739,9 +739,9 @@
 |`cublasCgeru_v2`| | | | |`hipblasCgeru_v2`|6.0.0| | | | |`rocblas_cgeru`|3.5.0| | | | |
 |`cublasCgeru_v2_64`|12.0| | | |`hipblasCgeru_v2_64`|6.2.0| | | | | | | | | | |
 |`cublasChbmv`| | | | |`hipblasChbmv_v2`|6.0.0| | | | |`rocblas_chbmv`|3.5.0| | | | |
-|`cublasChbmv_64`|12.0| | | |`hipblasChbmv_v2_64`|6.2.0| | | | | | | | | | |
+|`cublasChbmv_64`|12.0| | | |`hipblasChbmv_v2_64`|6.2.0| | | | |`rocblas_chbmv_64`|6.2.0| | | | |
 |`cublasChbmv_v2`| | | | |`hipblasChbmv_v2`|6.0.0| | | | |`rocblas_chbmv`|3.5.0| | | | |
-|`cublasChbmv_v2_64`|12.0| | | |`hipblasChbmv_v2_64`|6.2.0| | | | | | | | | | |
+|`cublasChbmv_v2_64`|12.0| | | |`hipblasChbmv_v2_64`|6.2.0| | | | |`rocblas_chbmv_64`|6.2.0| | | | |
 |`cublasChemv`| | | | |`hipblasChemv_v2`|6.0.0| | | | |`rocblas_chemv`|1.5.0| | | | |
 |`cublasChemv_64`|12.0| | | |`hipblasChemv_v2_64`|6.2.0| | | | | | | | | | |
 |`cublasChemv_v2`| | | | |`hipblasChemv_v2`|6.0.0| | | | |`rocblas_chemv`|1.5.0| | | | |
@@ -815,9 +815,9 @@
 |`cublasDger_v2`| | | | |`hipblasDger`|1.8.2| | | | |`rocblas_dger`|1.5.0| | | | |
 |`cublasDger_v2_64`|12.0| | | |`hipblasDger_64`|6.2.0| | | | | | | | | | |
 |`cublasDsbmv`| | | | |`hipblasDsbmv`|3.5.0| | | | |`rocblas_dsbmv`|3.5.0| | | | |
-|`cublasDsbmv_64`|12.0| | | |`hipblasDsbmv_64`|6.2.0| | | | | | | | | | |
+|`cublasDsbmv_64`|12.0| | | |`hipblasDsbmv_64`|6.2.0| | | | |`rocblas_dsbmv_64`|6.2.0| | | | |
 |`cublasDsbmv_v2`| | | | |`hipblasDsbmv`|3.5.0| | | | |`rocblas_dsbmv`|3.5.0| | | | |
-|`cublasDsbmv_v2_64`|12.0| | | |`hipblasDsbmv_64`|6.2.0| | | | | | | | | | |
+|`cublasDsbmv_v2_64`|12.0| | | |`hipblasDsbmv_64`|6.2.0| | | | |`rocblas_dsbmv_64`|6.2.0| | | | |
 |`cublasDspmv`| | | | |`hipblasDspmv`|3.5.0| | | | |`rocblas_dspmv`|3.5.0| | | | |
 |`cublasDspmv_64`|12.0| | | |`hipblasDspmv_64`|6.2.0| | | | | | | | | | |
 |`cublasDspmv_v2`| | | | |`hipblasDspmv`|3.5.0| | | | |`rocblas_dspmv`|3.5.0| | | | |
@@ -879,9 +879,9 @@
 |`cublasSger_v2`| | | | |`hipblasSger`|1.8.2| | | | |`rocblas_sger`|1.5.0| | | | |
 |`cublasSger_v2_64`|12.0| | | |`hipblasSger_64`|6.2.0| | | | | | | | | | |
 |`cublasSsbmv`| | | | |`hipblasSsbmv`|3.5.0| | | | |`rocblas_ssbmv`|3.5.0| | | | |
-|`cublasSsbmv_64`|12.0| | | |`hipblasSsbmv_64`|6.2.0| | | | | | | | | | |
+|`cublasSsbmv_64`|12.0| | | |`hipblasSsbmv_64`|6.2.0| | | | |`rocblas_ssbmv_64`|6.2.0| | | | |
 |`cublasSsbmv_v2`| | | | |`hipblasSsbmv`|3.5.0| | | | |`rocblas_ssbmv`|3.5.0| | | | |
-|`cublasSsbmv_v2_64`|12.0| | | |`hipblasSsbmv_64`|6.2.0| | | | | | | | | | |
+|`cublasSsbmv_v2_64`|12.0| | | |`hipblasSsbmv_64`|6.2.0| | | | |`rocblas_ssbmv_64`|6.2.0| | | | |
 |`cublasSspmv`| | | | |`hipblasSspmv`|3.5.0| | | | |`rocblas_sspmv`|3.5.0| | | | |
 |`cublasSspmv_64`|12.0| | | |`hipblasSspmv_64`|6.2.0| | | | | | | | | | |
 |`cublasSspmv_v2`| | | | |`hipblasSspmv`|3.5.0| | | | |`rocblas_sspmv`|3.5.0| | | | |
@@ -947,9 +947,9 @@
 |`cublasZgeru_v2`| | | | |`hipblasZgeru_v2`|6.0.0| | | | |`rocblas_zgeru`|3.5.0| | | | |
 |`cublasZgeru_v2_64`|12.0| | | |`hipblasZgeru_v2_64`|6.2.0| | | | | | | | | | |
 |`cublasZhbmv`| | | | |`hipblasZhbmv_v2`|6.0.0| | | | |`rocblas_zhbmv`|3.5.0| | | | |
-|`cublasZhbmv_64`|12.0| | | |`hipblasZhbmv_v2_64`|6.2.0| | | | | | | | | | |
+|`cublasZhbmv_64`|12.0| | | |`hipblasZhbmv_v2_64`|6.2.0| | | | |`rocblas_zhbmv_64`|6.2.0| | | | |
 |`cublasZhbmv_v2`| | | | |`hipblasZhbmv_v2`|6.0.0| | | | |`rocblas_zhbmv`|3.5.0| | | | |
-|`cublasZhbmv_v2_64`|12.0| | | |`hipblasZhbmv_v2_64`|6.2.0| | | | | | | | | | |
+|`cublasZhbmv_v2_64`|12.0| | | |`hipblasZhbmv_v2_64`|6.2.0| | | | |`rocblas_zhbmv_64`|6.2.0| | | | |
 |`cublasZhemv`| | | | |`hipblasZhemv_v2`|6.0.0| | | | |`rocblas_zhemv`|1.5.0| | | | |
 |`cublasZhemv_64`|12.0| | | |`hipblasZhemv_v2_64`|6.2.0| | | | | | | | | | |
 |`cublasZhemv_v2`| | | | |`hipblasZhemv_v2`|6.0.0| | | | |`rocblas_zhemv`|1.5.0| | | | |

diff --git a/docs/tables/CUBLAS_API_supported_by_ROC.md b/docs/tables/CUBLAS_API_supported_by_ROC.md
@@ -739,9 +739,9 @@
 |`cublasCgeru_v2`| | | | |`rocblas_cgeru`|3.5.0| | | | |
 |`cublasCgeru_v2_64`|12.0| | | | | | | | | |
 |`cublasChbmv`| | | | |`rocblas_chbmv`|3.5.0| | | | |
-|`cublasChbmv_64`|12.0| | | | | | | | | |
+|`cublasChbmv_64`|12.0| | | |`rocblas_chbmv_64`|6.2.0| | | | |
 |`cublasChbmv_v2`| | | | |`rocblas_chbmv`|3.5.0| | | | |
-|`cublasChbmv_v2_64`|12.0| | | | | | | | | |
+|`cublasChbmv_v2_64`|12.0| | | |`rocblas_chbmv_64`|6.2.0| | | | |
 |`cublasChemv`| | | | |`rocblas_chemv`|1.5.0| | | | |
 |`cublasChemv_64`|12.0| | | | | | | | | |
 |`cublasChemv_v2`| | | | |`rocblas_chemv`|1.5.0| | | | |
@@ -815,9 +815,9 @@
 |`cublasDger_v2`| | | | |`rocblas_dger`|1.5.0| | | | |
 |`cublasDger_v2_64`|12.0| | | | | | | | | |
 |`cublasDsbmv`| | | | |`rocblas_dsbmv`|3.5.0| | | | |
-|`cublasDsbmv_64`|12.0| | | | | | | | | |
+|`cublasDsbmv_64`|12.0| | | |`rocblas_dsbmv_64`|6.2.0| | | | |
 |`cublasDsbmv_v2`| | | | |`rocblas_dsbmv`|3.5.0| | | | |
-|`cublasDsbmv_v2_64`|12.0| | | | | | | | | |
+|`cublasDsbmv_v2_64`|12.0| | | |`rocblas_dsbmv_64`|6.2.0| | | | |
 |`cublasDspmv`| | | | |`rocblas_dspmv`|3.5.0| | | | |
 |`cublasDspmv_64`|12.0| | | | | | | | | |
 |`cublasDspmv_v2`| | | | |`rocblas_dspmv`|3.5.0| | | | |
@@ -879,9 +879,9 @@
 |`cublasSger_v2`| | | | |`rocblas_sger`|1.5.0| | | | |
 |`cublasSger_v2_64`|12.0| | | | | | | | | |
 |`cublasSsbmv`| | | | |`rocblas_ssbmv`|3.5.0| | | | |
-|`cublasSsbmv_64`|12.0| | | | | | | | | |
+|`cublasSsbmv_64`|12.0| | | |`rocblas_ssbmv_64`|6.2.0| | | | |
 |`cublasSsbmv_v2`| | | | |`rocblas_ssbmv`|3.5.0| | | | |
-|`cublasSsbmv_v2_64`|12.0| | | | | | | | | |
+|`cublasSsbmv_v2_64`|12.0| | | |`rocblas_ssbmv_64`|6.2.0| | | | |
 |`cublasSspmv`| | | | |`rocblas_sspmv`|3.5.0| | | | |
 |`cublasSspmv_64`|12.0| | | | | | | | | |
 |`cublasSspmv_v2`| | | | |`rocblas_sspmv`|3.5.0| | | | |
@@ -947,9 +947,9 @@
 |`cublasZgeru_v2`| | | | |`rocblas_zgeru`|3.5.0| | | | |
 |`cublasZgeru_v2_64`|12.0| | | | | | | | | |
 |`cublasZhbmv`| | | | |`rocblas_zhbmv`|3.5.0| | | | |
-|`cublasZhbmv_64`|12.0| | | | | | | | | |
+|`cublasZhbmv_64`|12.0| | | |`rocblas_zhbmv_64`|6.2.0| | | | |
 |`cublasZhbmv_v2`| | | | |`rocblas_zhbmv`|3.5.0| | | | |
-|`cublasZhbmv_v2_64`|12.0| | | | | | | | | |
+|`cublasZhbmv_v2_64`|12.0| | | |`rocblas_zhbmv_64`|6.2.0| | | | |
 |`cublasZhemv`| | | | |`rocblas_zhemv`|1.5.0| | | | |
 |`cublasZhemv_64`|12.0| | | | | | | | | |
 |`cublasZhemv_v2`| | | | |`rocblas_zhemv`|1.5.0| | | | |

diff --git a/src/CUDA2HIP_BLAS_API_functions.cpp b/src/CUDA2HIP_BLAS_API_functions.cpp
@@ -316,13 +316,13 @@ const std::map<llvm::StringRef, hipCounter> CUDA_BLAS_FUNCTION_MAP {
 
   // SBMV/HBMV
   {"cublasSsbmv",                                          {"hipblasSsbmv",                                              "rocblas_ssbmv",                                      CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_2, HIP_SUPPORTED_V2_ONLY}},
-  {"cublasSsbmv_64",                                       {"hipblasSsbmv_64",                                           "",                                                   CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_2, ROC_UNSUPPORTED}},
+  {"cublasSsbmv_64",                                       {"hipblasSsbmv_64",                                           "rocblas_ssbmv_64",                                   CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_2}},
   {"cublasDsbmv",                                          {"hipblasDsbmv",                                              "rocblas_dsbmv",                                      CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_2, HIP_SUPPORTED_V2_ONLY}},
-  {"cublasDsbmv_64",                                       {"hipblasDsbmv_64",                                           "",                                                   CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_2, ROC_UNSUPPORTED}},
+  {"cublasDsbmv_64",                                       {"hipblasDsbmv_64",                                           "rocblas_dsbmv_64",                                   CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_2}},
   {"cublasChbmv",                                          {"hipblasChbmv_v2",                                           "rocblas_chbmv",                                      CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_2, HIP_SUPPORTED_V2_ONLY}},
-  {"cublasChbmv_64",                                       {"hipblasChbmv_v2_64",                                        "",                                                   CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_2, ROC_UNSUPPORTED}},
+  {"cublasChbmv_64",                                       {"hipblasChbmv_v2_64",                                        "rocblas_chbmv_64",                                   CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_2}},
   {"cublasZhbmv",                                          {"hipblasZhbmv_v2",                                           "rocblas_zhbmv",                                      CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_2, HIP_SUPPORTED_V2_ONLY}},
-  {"cublasZhbmv_64",                                       {"hipblasZhbmv_v2_64",                                        "",                                                   CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_2, ROC_UNSUPPORTED}},
+  {"cublasZhbmv_64",                                       {"hipblasZhbmv_v2_64",                                        "rocblas_zhbmv_64",                                   CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_2}},
 
   // SPMV/HPMV
   {"cublasSspmv",                                          {"hipblasSspmv",                                              "rocblas_sspmv",                                      CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_2, HIP_SUPPORTED_V2_ONLY}},
@@ -734,13 +734,13 @@ const std::map<llvm::StringRef, hipCounter> CUDA_BLAS_FUNCTION_MAP {
 
   // SBMV/HBMV
   {"cublasSsbmv_v2",                                       {"hipblasSsbmv",                                              "rocblas_ssbmv",                                      CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_2}},
-  {"cublasSsbmv_v2_64",                                    {"hipblasSsbmv_64",                                           "",                                                   CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_2, ROC_UNSUPPORTED}},
+  {"cublasSsbmv_v2_64",                                    {"hipblasSsbmv_64",                                           "rocblas_ssbmv_64",                                   CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_2}},
   {"cublasDsbmv_v2",                                       {"hipblasDsbmv",                                              "rocblas_dsbmv",                                      CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_2}},
-  {"cublasDsbmv_v2_64",                                    {"hipblasDsbmv_64",                                           "",                                                   CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_2, ROC_UNSUPPORTED}},
+  {"cublasDsbmv_v2_64",                                    {"hipblasDsbmv_64",                                           "rocblas_dsbmv_64",                                   CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_2}},
   {"cublasChbmv_v2",                                       {"hipblasChbmv_v2",                                           "rocblas_chbmv",                                      CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_2}},
-  {"cublasChbmv_v2_64",                                    {"hipblasChbmv_v2_64",                                        "",                                                   CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_2, ROC_UNSUPPORTED}},
+  {"cublasChbmv_v2_64",                                    {"hipblasChbmv_v2_64",                                        "rocblas_chbmv_64",                                   CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_2}},
   {"cublasZhbmv_v2",                                       {"hipblasZhbmv_v2",                                           "rocblas_zhbmv",                                      CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_2}},
-  {"cublasZhbmv_v2_64",                                    {"hipblasZhbmv_v2_64",                                        "",                                                   CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_2, ROC_UNSUPPORTED}},
+  {"cublasZhbmv_v2_64",                                    {"hipblasZhbmv_v2_64",                                        "rocblas_zhbmv_64",                                   CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_2}},
 
   // SPMV/HPMV
   {"cublasSspmv_v2",                                       {"hipblasSspmv",                                              "rocblas_sspmv",                                      CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_2}},
@@ -2339,6 +2339,10 @@ const std::map<llvm::StringRef, hipAPIversions> HIP_BLAS_FUNCTION_VER_MAP {
   {"rocblas_hssgemv_strided_batched_64",                   {HIP_6020, HIP_0,    HIP_0   }},
   {"rocblas_tstgemv_strided_batched_64",                   {HIP_6020, HIP_0,    HIP_0   }},
   {"rocblas_tssgemv_strided_batched_64",                   {HIP_6020, HIP_0,    HIP_0   }},
+  {"rocblas_ssbmv_64",                                     {HIP_6020, HIP_0,    HIP_0   }},
+  {"rocblas_dsbmv_64",                                     {HIP_6020, HIP_0,    HIP_0   }},
+  {"rocblas_chbmv_64",                                     {HIP_6020, HIP_0,    HIP_0   }},
+  {"rocblas_zhbmv_64",                                     {HIP_6020, HIP_0,    HIP_0   }},
 };
 
 const std::map<llvm::StringRef, hipAPIChangedVersions> HIP_BLAS_FUNCTION_CHANGED_VER_MAP {