Merge pull request #1693 from emankov/HIPIFY

[HIPIFY][rocBLAS] 64-bit functions support - Step 13
ROCm · Oct 9, 2024 · c0a57bb · c0a57bb
2 parents 1d3fb9f + 24965a7
commit c0a57bb
Show file tree

Hide file tree

Showing 5 changed files with 64 additions and 32 deletions.
diff --git a/bin/hipify-perl b/bin/hipify-perl
@@ -1656,7 +1656,9 @@ sub rocSubstitutions {
     subst("cublasCtbsv", "rocblas_ctbsv", "library");
     subst("cublasCtbsv_v2", "rocblas_ctbsv", "library");
     subst("cublasCtpmv", "rocblas_ctpmv", "library");
+    subst("cublasCtpmv_64", "rocblas_ctpmv_64", "library");
     subst("cublasCtpmv_v2", "rocblas_ctpmv", "library");
+    subst("cublasCtpmv_v2_64", "rocblas_ctpmv_64", "library");
     subst("cublasCtpsv", "rocblas_ctpsv", "library");
     subst("cublasCtpsv_v2", "rocblas_ctpsv", "library");
     subst("cublasCtrmm", "rocblas_ctrmm", "library");
@@ -1776,7 +1778,9 @@ sub rocSubstitutions {
     subst("cublasDtbsv", "rocblas_dtbsv", "library");
     subst("cublasDtbsv_v2", "rocblas_dtbsv", "library");
     subst("cublasDtpmv", "rocblas_dtpmv", "library");
+    subst("cublasDtpmv_64", "rocblas_dtpmv_64", "library");
     subst("cublasDtpmv_v2", "rocblas_dtpmv", "library");
+    subst("cublasDtpmv_v2_64", "rocblas_dtpmv_64", "library");
     subst("cublasDtpsv", "rocblas_dtpsv", "library");
     subst("cublasDtpsv_v2", "rocblas_dtpsv", "library");
     subst("cublasDtrmm", "rocblas_dtrmm", "library");
@@ -1980,7 +1984,9 @@ sub rocSubstitutions {
     subst("cublasStbsv", "rocblas_stbsv", "library");
     subst("cublasStbsv_v2", "rocblas_stbsv", "library");
     subst("cublasStpmv", "rocblas_stpmv", "library");
+    subst("cublasStpmv_64", "rocblas_stpmv_64", "library");
     subst("cublasStpmv_v2", "rocblas_stpmv", "library");
+    subst("cublasStpmv_v2_64", "rocblas_stpmv_64", "library");
     subst("cublasStpsv", "rocblas_stpsv", "library");
     subst("cublasStpsv_v2", "rocblas_stpsv", "library");
     subst("cublasStrmm", "rocblas_strmm", "library");
@@ -2121,7 +2127,9 @@ sub rocSubstitutions {
     subst("cublasZtbsv", "rocblas_ztbsv", "library");
     subst("cublasZtbsv_v2", "rocblas_ztbsv", "library");
     subst("cublasZtpmv", "rocblas_ztpmv", "library");
+    subst("cublasZtpmv_64", "rocblas_ztpmv_64", "library");
     subst("cublasZtpmv_v2", "rocblas_ztpmv", "library");
+    subst("cublasZtpmv_v2_64", "rocblas_ztpmv_64", "library");
     subst("cublasZtpsv", "rocblas_ztpsv", "library");
     subst("cublasZtpsv_v2", "rocblas_ztpsv", "library");
     subst("cublasZtrmm", "rocblas_ztrmm", "library");
@@ -12615,8 +12623,6 @@ sub warnRocOnlyUnsupportedFunctions {
         "cublasZtpttr",
         "cublasZtpsv_v2_64",
         "cublasZtpsv_64",
-        "cublasZtpmv_v2_64",
-        "cublasZtpmv_64",
         "cublasZtbsv_v2_64",
         "cublasZtbsv_64",
         "cublasZtbmv_v2_64",
@@ -12668,8 +12674,6 @@ sub warnRocOnlyUnsupportedFunctions {
         "cublasStpttr",
         "cublasStpsv_v2_64",
         "cublasStpsv_64",
-        "cublasStpmv_v2_64",
-        "cublasStpmv_64",
         "cublasStbsv_v2_64",
         "cublasStbsv_64",
         "cublasStbmv_v2_64",
@@ -12813,8 +12817,6 @@ sub warnRocOnlyUnsupportedFunctions {
         "cublasDtpttr",
         "cublasDtpsv_v2_64",
         "cublasDtpsv_64",
-        "cublasDtpmv_v2_64",
-        "cublasDtpmv_64",
         "cublasDtbsv_v2_64",
         "cublasDtbsv_64",
         "cublasDtbmv_v2_64",
@@ -12853,8 +12855,6 @@ sub warnRocOnlyUnsupportedFunctions {
         "cublasCtpttr",
         "cublasCtpsv_v2_64",
         "cublasCtpsv_64",
-        "cublasCtpmv_v2_64",
-        "cublasCtpmv_64",
         "cublasCtbsv_v2_64",
         "cublasCtbsv_64",
         "cublasCtbmv_v2_64",

diff --git a/docs/tables/CUBLAS_API_supported_by_HIP_and_ROC.md b/docs/tables/CUBLAS_API_supported_by_HIP_and_ROC.md
@@ -787,9 +787,9 @@
 |`cublasCtbsv_v2`| | | | |`hipblasCtbsv_v2`|6.0.0| | | | |`rocblas_ctbsv`|3.5.0| | | | |
 |`cublasCtbsv_v2_64`|12.0| | | |`hipblasCtbsv_v2_64`|6.2.0| | | | | | | | | | |
 |`cublasCtpmv`| | | | |`hipblasCtpmv_v2`|6.0.0| | | | |`rocblas_ctpmv`|3.5.0| | | | |
-|`cublasCtpmv_64`|12.0| | | |`hipblasCtpmv_v2_64`|6.2.0| | | | | | | | | | |
+|`cublasCtpmv_64`|12.0| | | |`hipblasCtpmv_v2_64`|6.2.0| | | | |`rocblas_ctpmv_64`|6.2.0| | | | |
 |`cublasCtpmv_v2`| | | | |`hipblasCtpmv_v2`|6.0.0| | | | |`rocblas_ctpmv`|3.5.0| | | | |
-|`cublasCtpmv_v2_64`|12.0| | | |`hipblasCtpmv_v2_64`|6.2.0| | | | | | | | | | |
+|`cublasCtpmv_v2_64`|12.0| | | |`hipblasCtpmv_v2_64`|6.2.0| | | | |`rocblas_ctpmv_64`|6.2.0| | | | |
 |`cublasCtpsv`| | | | |`hipblasCtpsv_v2`|6.0.0| | | | |`rocblas_ctpsv`|3.5.0| | | | |
 |`cublasCtpsv_64`|12.0| | | |`hipblasCtpsv_v2_64`|6.2.0| | | | | | | | | | |
 |`cublasCtpsv_v2`| | | | |`hipblasCtpsv_v2`|6.0.0| | | | |`rocblas_ctpsv`|3.5.0| | | | |
@@ -851,9 +851,9 @@
 |`cublasDtbsv_v2`| | | | |`hipblasDtbsv`|3.6.0| | | | |`rocblas_dtbsv`|3.5.0| | | | |
 |`cublasDtbsv_v2_64`|12.0| | | |`hipblasDtbsv_64`|6.2.0| | | | | | | | | | |
 |`cublasDtpmv`| | | | |`hipblasDtpmv`|3.5.0| | | | |`rocblas_dtpmv`|3.5.0| | | | |
-|`cublasDtpmv_64`|12.0| | | |`hipblasDtpmv_64`|6.2.0| | | | | | | | | | |
+|`cublasDtpmv_64`|12.0| | | |`hipblasDtpmv_64`|6.2.0| | | | |`rocblas_dtpmv_64`|6.2.0| | | | |
 |`cublasDtpmv_v2`| | | | |`hipblasDtpmv`|3.5.0| | | | |`rocblas_dtpmv`|3.5.0| | | | |
-|`cublasDtpmv_v2_64`|12.0| | | |`hipblasDtpmv_64`|6.2.0| | | | | | | | | | |
+|`cublasDtpmv_v2_64`|12.0| | | |`hipblasDtpmv_64`|6.2.0| | | | |`rocblas_dtpmv_64`|6.2.0| | | | |
 |`cublasDtpsv`| | | | |`hipblasDtpsv`|3.5.0| | | | |`rocblas_dtpsv`|3.5.0| | | | |
 |`cublasDtpsv_64`|12.0| | | |`hipblasDtpsv_64`|6.2.0| | | | | | | | | | |
 |`cublasDtpsv_v2`| | | | |`hipblasDtpsv`|3.5.0| | | | |`rocblas_dtpsv`|3.5.0| | | | |
@@ -915,9 +915,9 @@
 |`cublasStbsv_v2`| | | | |`hipblasStbsv`|3.6.0| | | | |`rocblas_stbsv`|3.5.0| | | | |
 |`cublasStbsv_v2_64`|12.0| | | |`hipblasStbsv_64`|6.2.0| | | | | | | | | | |
 |`cublasStpmv`| | | | |`hipblasStpmv`|3.5.0| | | | |`rocblas_stpmv`|3.5.0| | | | |
-|`cublasStpmv_64`|12.0| | | |`hipblasStpmv_64`|6.2.0| | | | | | | | | | |
+|`cublasStpmv_64`|12.0| | | |`hipblasStpmv_64`|6.2.0| | | | |`rocblas_stpmv_64`|6.2.0| | | | |
 |`cublasStpmv_v2`| | | | |`hipblasStpmv`|3.5.0| | | | |`rocblas_stpmv`|3.5.0| | | | |
-|`cublasStpmv_v2_64`|12.0| | | |`hipblasStpmv_64`|6.2.0| | | | | | | | | | |
+|`cublasStpmv_v2_64`|12.0| | | |`hipblasStpmv_64`|6.2.0| | | | |`rocblas_stpmv_64`|6.2.0| | | | |
 |`cublasStpsv`| | | | |`hipblasStpsv`|3.5.0| | | | |`rocblas_stpsv`|3.5.0| | | | |
 |`cublasStpsv_64`|12.0| | | |`hipblasStpsv_64`|6.2.0| | | | | | | | | | |
 |`cublasStpsv_v2`| | | | |`hipblasStpsv`|3.5.0| | | | |`rocblas_stpsv`|3.5.0| | | | |
@@ -995,9 +995,9 @@
 |`cublasZtbsv_v2`| | | | |`hipblasZtbsv_v2`|6.0.0| | | | |`rocblas_ztbsv`|3.5.0| | | | |
 |`cublasZtbsv_v2_64`|12.0| | | |`hipblasZtbsv_v2_64`|6.2.0| | | | | | | | | | |
 |`cublasZtpmv`| | | | |`hipblasZtpmv_v2`|6.0.0| | | | |`rocblas_ztpmv`|3.5.0| | | | |
-|`cublasZtpmv_64`|12.0| | | |`hipblasZtpmv_v2_64`|6.2.0| | | | | | | | | | |
+|`cublasZtpmv_64`|12.0| | | |`hipblasZtpmv_v2_64`|6.2.0| | | | |`rocblas_ztpmv_64`|6.2.0| | | | |
 |`cublasZtpmv_v2`| | | | |`hipblasZtpmv_v2`|6.0.0| | | | |`rocblas_ztpmv`|3.5.0| | | | |
-|`cublasZtpmv_v2_64`|12.0| | | |`hipblasZtpmv_v2_64`|6.2.0| | | | | | | | | | |
+|`cublasZtpmv_v2_64`|12.0| | | |`hipblasZtpmv_v2_64`|6.2.0| | | | |`rocblas_ztpmv_64`|6.2.0| | | | |
 |`cublasZtpsv`| | | | |`hipblasZtpsv_v2`|6.0.0| | | | |`rocblas_ztpsv`|3.5.0| | | | |
 |`cublasZtpsv_64`|12.0| | | |`hipblasZtpsv_v2_64`|6.2.0| | | | | | | | | | |
 |`cublasZtpsv_v2`| | | | |`hipblasZtpsv_v2`|6.0.0| | | | |`rocblas_ztpsv`|3.5.0| | | | |

diff --git a/docs/tables/CUBLAS_API_supported_by_ROC.md b/docs/tables/CUBLAS_API_supported_by_ROC.md
@@ -787,9 +787,9 @@
 |`cublasCtbsv_v2`| | | | |`rocblas_ctbsv`|3.5.0| | | | |
 |`cublasCtbsv_v2_64`|12.0| | | | | | | | | |
 |`cublasCtpmv`| | | | |`rocblas_ctpmv`|3.5.0| | | | |
-|`cublasCtpmv_64`|12.0| | | | | | | | | |
+|`cublasCtpmv_64`|12.0| | | |`rocblas_ctpmv_64`|6.2.0| | | | |
 |`cublasCtpmv_v2`| | | | |`rocblas_ctpmv`|3.5.0| | | | |
-|`cublasCtpmv_v2_64`|12.0| | | | | | | | | |
+|`cublasCtpmv_v2_64`|12.0| | | |`rocblas_ctpmv_64`|6.2.0| | | | |
 |`cublasCtpsv`| | | | |`rocblas_ctpsv`|3.5.0| | | | |
 |`cublasCtpsv_64`|12.0| | | | | | | | | |
 |`cublasCtpsv_v2`| | | | |`rocblas_ctpsv`|3.5.0| | | | |
@@ -851,9 +851,9 @@
 |`cublasDtbsv_v2`| | | | |`rocblas_dtbsv`|3.5.0| | | | |
 |`cublasDtbsv_v2_64`|12.0| | | | | | | | | |
 |`cublasDtpmv`| | | | |`rocblas_dtpmv`|3.5.0| | | | |
-|`cublasDtpmv_64`|12.0| | | | | | | | | |
+|`cublasDtpmv_64`|12.0| | | |`rocblas_dtpmv_64`|6.2.0| | | | |
 |`cublasDtpmv_v2`| | | | |`rocblas_dtpmv`|3.5.0| | | | |
-|`cublasDtpmv_v2_64`|12.0| | | | | | | | | |
+|`cublasDtpmv_v2_64`|12.0| | | |`rocblas_dtpmv_64`|6.2.0| | | | |
 |`cublasDtpsv`| | | | |`rocblas_dtpsv`|3.5.0| | | | |
 |`cublasDtpsv_64`|12.0| | | | | | | | | |
 |`cublasDtpsv_v2`| | | | |`rocblas_dtpsv`|3.5.0| | | | |
@@ -915,9 +915,9 @@
 |`cublasStbsv_v2`| | | | |`rocblas_stbsv`|3.5.0| | | | |
 |`cublasStbsv_v2_64`|12.0| | | | | | | | | |
 |`cublasStpmv`| | | | |`rocblas_stpmv`|3.5.0| | | | |
-|`cublasStpmv_64`|12.0| | | | | | | | | |
+|`cublasStpmv_64`|12.0| | | |`rocblas_stpmv_64`|6.2.0| | | | |
 |`cublasStpmv_v2`| | | | |`rocblas_stpmv`|3.5.0| | | | |
-|`cublasStpmv_v2_64`|12.0| | | | | | | | | |
+|`cublasStpmv_v2_64`|12.0| | | |`rocblas_stpmv_64`|6.2.0| | | | |
 |`cublasStpsv`| | | | |`rocblas_stpsv`|3.5.0| | | | |
 |`cublasStpsv_64`|12.0| | | | | | | | | |
 |`cublasStpsv_v2`| | | | |`rocblas_stpsv`|3.5.0| | | | |
@@ -995,9 +995,9 @@
 |`cublasZtbsv_v2`| | | | |`rocblas_ztbsv`|3.5.0| | | | |
 |`cublasZtbsv_v2_64`|12.0| | | | | | | | | |
 |`cublasZtpmv`| | | | |`rocblas_ztpmv`|3.5.0| | | | |
-|`cublasZtpmv_64`|12.0| | | | | | | | | |
+|`cublasZtpmv_64`|12.0| | | |`rocblas_ztpmv_64`|6.2.0| | | | |
 |`cublasZtpmv_v2`| | | | |`rocblas_ztpmv`|3.5.0| | | | |
-|`cublasZtpmv_v2_64`|12.0| | | | | | | | | |
+|`cublasZtpmv_v2_64`|12.0| | | |`rocblas_ztpmv_64`|6.2.0| | | | |
 |`cublasZtpsv`| | | | |`rocblas_ztpsv`|3.5.0| | | | |
 |`cublasZtpsv_64`|12.0| | | | | | | | | |
 |`cublasZtpsv_v2`| | | | |`rocblas_ztpsv`|3.5.0| | | | |

diff --git a/src/CUDA2HIP_BLAS_API_functions.cpp b/src/CUDA2HIP_BLAS_API_functions.cpp
@@ -262,13 +262,13 @@ const std::map<llvm::StringRef, hipCounter> CUDA_BLAS_FUNCTION_MAP {
 
   // TPMV
   {"cublasStpmv",                                          {"hipblasStpmv",                                              "rocblas_stpmv",                                      CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_2, HIP_SUPPORTED_V2_ONLY}},
-  {"cublasStpmv_64",                                       {"hipblasStpmv_64",                                           "",                                                   CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_2, ROC_UNSUPPORTED}},
+  {"cublasStpmv_64",                                       {"hipblasStpmv_64",                                           "rocblas_stpmv_64",                                   CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_2}},
   {"cublasDtpmv",                                          {"hipblasDtpmv",                                              "rocblas_dtpmv",                                      CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_2, HIP_SUPPORTED_V2_ONLY}},
-  {"cublasDtpmv_64",                                       {"hipblasDtpmv_64",                                           "",                                                   CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_2, ROC_UNSUPPORTED}},
+  {"cublasDtpmv_64",                                       {"hipblasDtpmv_64",                                           "rocblas_dtpmv_64",                                   CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_2}},
   {"cublasCtpmv",                                          {"hipblasCtpmv_v2",                                           "rocblas_ctpmv",                                      CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_2, HIP_SUPPORTED_V2_ONLY}},
-  {"cublasCtpmv_64",                                       {"hipblasCtpmv_v2_64",                                        "",                                                   CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_2, ROC_UNSUPPORTED}},
+  {"cublasCtpmv_64",                                       {"hipblasCtpmv_v2_64",                                        "rocblas_ctpmv_64",                                   CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_2}},
   {"cublasZtpmv",                                          {"hipblasZtpmv_v2",                                           "rocblas_ztpmv",                                      CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_2, HIP_SUPPORTED_V2_ONLY}},
-  {"cublasZtpmv_64",                                       {"hipblasZtpmv_v2_64",                                        "",                                                   CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_2, ROC_UNSUPPORTED}},
+  {"cublasZtpmv_64",                                       {"hipblasZtpmv_v2_64",                                        "rocblas_ztpmv_64",                                   CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_2}},
 
   // TRSV
   {"cublasStrsv",                                          {"hipblasStrsv",                                              "rocblas_strsv",                                      CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_2, HIP_SUPPORTED_V2_ONLY}},
@@ -680,13 +680,13 @@ const std::map<llvm::StringRef, hipCounter> CUDA_BLAS_FUNCTION_MAP {
 
   // TPMV
   {"cublasStpmv_v2",                                       {"hipblasStpmv",                                              "rocblas_stpmv",                                      CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_2}},
-  {"cublasStpmv_v2_64",                                    {"hipblasStpmv_64",                                           "",                                                   CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_2, ROC_UNSUPPORTED}},
+  {"cublasStpmv_v2_64",                                    {"hipblasStpmv_64",                                           "rocblas_stpmv_64",                                   CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_2}},
   {"cublasDtpmv_v2",                                       {"hipblasDtpmv",                                              "rocblas_dtpmv",                                      CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_2}},
-  {"cublasDtpmv_v2_64",                                    {"hipblasDtpmv_64",                                           "",                                                   CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_2, ROC_UNSUPPORTED}},
+  {"cublasDtpmv_v2_64",                                    {"hipblasDtpmv_64",                                           "rocblas_dtpmv_64",                                   CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_2}},
   {"cublasCtpmv_v2",                                       {"hipblasCtpmv_v2",                                           "rocblas_ctpmv",                                      CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_2}},
-  {"cublasCtpmv_v2_64",                                    {"hipblasCtpmv_v2_64",                                        "",                                                   CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_2, ROC_UNSUPPORTED}},
+  {"cublasCtpmv_v2_64",                                    {"hipblasCtpmv_v2_64",                                        "rocblas_ctpmv_64",                                   CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_2}},
   {"cublasZtpmv_v2",                                       {"hipblasZtpmv_v2",                                           "rocblas_ztpmv",                                      CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_2}},
-  {"cublasZtpmv_v2_64",                                    {"hipblasZtpmv_v2_64",                                        "",                                                   CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_2, ROC_UNSUPPORTED}},
+  {"cublasZtpmv_v2_64",                                    {"hipblasZtpmv_v2_64",                                        "rocblas_ztpmv_64",                                   CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_2}},
 
   // TRSV
   {"cublasStrsv_v2",                                       {"hipblasStrsv",                                              "rocblas_strsv",                                      CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_2}},
@@ -2377,6 +2377,10 @@ const std::map<llvm::StringRef, hipAPIversions> HIP_BLAS_FUNCTION_VER_MAP {
   {"rocblas_dtrmv_64",                                     {HIP_6020, HIP_0,    HIP_0   }},
   {"rocblas_ctrmv_64",                                     {HIP_6020, HIP_0,    HIP_0   }},
   {"rocblas_ztrmv_64",                                     {HIP_6020, HIP_0,    HIP_0   }},
+  {"rocblas_stpmv_64",                                     {HIP_6020, HIP_0,    HIP_0   }},
+  {"rocblas_dtpmv_64",                                     {HIP_6020, HIP_0,    HIP_0   }},
+  {"rocblas_ctpmv_64",                                     {HIP_6020, HIP_0,    HIP_0   }},
+  {"rocblas_ztpmv_64",                                     {HIP_6020, HIP_0,    HIP_0   }},
 };
 
 const std::map<llvm::StringRef, hipAPIChangedVersions> HIP_BLAS_FUNCTION_CHANGED_VER_MAP {