[HIPIFY][BLAS][6.2.0] cuBLAS support - Step 4 - 64-bit functions

+ Updated synthetic tests, the regenerated `hipify-perl`, and `BLAS` `CUDA2HIP` documentation
emankov · Jun 25, 2024 · 4f2a0bc · 4f2a0bc
1 parent 06d011b
commit 4f2a0bc
Show file tree

Hide file tree

Showing 5 changed files with 40 additions and 16 deletions.
diff --git a/bin/hipify-perl b/bin/hipify-perl
@@ -3809,6 +3809,7 @@ sub simpleSubstitutions {
     subst("cublasCgemv", "hipblasCgemv_v2", "library");
     subst("cublasCgemvBatched", "hipblasCgemvBatched_v2", "library");
     subst("cublasCgemvStridedBatched", "hipblasCgemvStridedBatched_v2", "library");
+    subst("cublasCgemvStridedBatched_64", "hipblasCgemvStridedBatched_v2_64", "library");
     subst("cublasCgemv_v2", "hipblasCgemv_v2", "library");
     subst("cublasCgeqrfBatched", "hipblasCgeqrfBatched_v2", "library");
     subst("cublasCgerc", "hipblasCgerc_v2", "library");
@@ -3923,6 +3924,7 @@ sub simpleSubstitutions {
     subst("cublasDgemv", "hipblasDgemv", "library");
     subst("cublasDgemvBatched", "hipblasDgemvBatched", "library");
     subst("cublasDgemvStridedBatched", "hipblasDgemvStridedBatched", "library");
+    subst("cublasDgemvStridedBatched_64", "hipblasDgemvStridedBatched_64", "library");
     subst("cublasDgemv_v2", "hipblasDgemv", "library");
     subst("cublasDgeqrfBatched", "hipblasDgeqrfBatched", "library");
     subst("cublasDger", "hipblasDger", "library");
@@ -4120,6 +4122,7 @@ sub simpleSubstitutions {
     subst("cublasSgemv", "hipblasSgemv", "library");
     subst("cublasSgemvBatched", "hipblasSgemvBatched", "library");
     subst("cublasSgemvStridedBatched", "hipblasSgemvStridedBatched", "library");
+    subst("cublasSgemvStridedBatched_64", "hipblasSgemvStridedBatched_64", "library");
     subst("cublasSgemv_v2", "hipblasSgemv", "library");
     subst("cublasSgeqrfBatched", "hipblasSgeqrfBatched", "library");
     subst("cublasSger", "hipblasSger", "library");
@@ -4225,6 +4228,7 @@ sub simpleSubstitutions {
     subst("cublasZgemv", "hipblasZgemv_v2", "library");
     subst("cublasZgemvBatched", "hipblasZgemvBatched_v2", "library");
     subst("cublasZgemvStridedBatched", "hipblasZgemvStridedBatched_v2", "library");
+    subst("cublasZgemvStridedBatched_64", "hipblasZgemvStridedBatched_v2_64", "library");
     subst("cublasZgemv_v2", "hipblasZgemv_v2", "library");
     subst("cublasZgeqrfBatched", "hipblasZgeqrfBatched_v2", "library");
     subst("cublasZgerc", "hipblasZgerc_v2", "library");
@@ -11378,7 +11382,6 @@ sub warnHipOnlyUnsupportedFunctions {
         "cublasZgeru_64",
         "cublasZgerc_v2_64",
         "cublasZgerc_64",
-        "cublasZgemvStridedBatched_64",
         "cublasZgemm_v2_64",
         "cublasZgemm_64",
         "cublasZgemmStridedBatched_64",
@@ -11443,7 +11446,6 @@ sub warnHipOnlyUnsupportedFunctions {
         "cublasShutdown",
         "cublasSger_v2_64",
         "cublasSger_64",
-        "cublasSgemvStridedBatched_64",
         "cublasSgemm_v2_64",
         "cublasSgemm_64",
         "cublasSgemmStridedBatched_64",
@@ -11587,7 +11589,6 @@ sub warnHipOnlyUnsupportedFunctions {
         "cublasDmatinvBatched",
         "cublasDger_v2_64",
         "cublasDger_64",
-        "cublasDgemvStridedBatched_64",
         "cublasDgemm_v2_64",
         "cublasDgemm_64",
         "cublasDgemmStridedBatched_64",
@@ -11665,7 +11666,6 @@ sub warnHipOnlyUnsupportedFunctions {
         "cublasCgeru_64",
         "cublasCgerc_v2_64",
         "cublasCgerc_64",
-        "cublasCgemvStridedBatched_64",
         "cublasCgemm_v2_64",
         "cublasCgemm_64",
         "cublasCgemmStridedBatched_64",

diff --git a/docs/tables/CUBLAS_API_supported_by_HIP.md b/docs/tables/CUBLAS_API_supported_by_HIP.md
@@ -1034,7 +1034,7 @@
 |`cublasCgemvBatched`|11.6| | | |`hipblasCgemvBatched_v2`|6.0.0| | | | |
 |`cublasCgemvBatched_64`|12.0| | | |`hipblasCgemvBatched_v2_64`|6.2.0| | | |6.2.0|
 |`cublasCgemvStridedBatched`|11.6| | | |`hipblasCgemvStridedBatched_v2`|6.0.0| | | | |
-|`cublasCgemvStridedBatched_64`|12.0| | | | | | | | | |
+|`cublasCgemvStridedBatched_64`|12.0| | | |`hipblasCgemvStridedBatched_v2_64`|6.2.0| | | |6.2.0|
 |`cublasChemm`| | | | |`hipblasChemm_v2`|6.0.0| | | | |
 |`cublasChemm_64`|12.0| | | | | | | | | |
 |`cublasChemm_v2`| | | | |`hipblasChemm_v2`|6.0.0| | | | |
@@ -1084,7 +1084,7 @@
 |`cublasDgemvBatched`|11.6| | | |`hipblasDgemvBatched`|3.0.0| | | | |
 |`cublasDgemvBatched_64`|12.0| | | |`hipblasDgemvBatched_64`|6.2.0| | | |6.2.0|
 |`cublasDgemvStridedBatched`|11.6| | | |`hipblasDgemvStridedBatched`|3.0.0| | | | |
-|`cublasDgemvStridedBatched_64`|12.0| | | | | | | | | |
+|`cublasDgemvStridedBatched_64`|12.0| | | |`hipblasDgemvStridedBatched_64`|6.2.0| | | |6.2.0|
 |`cublasDsymm`| | | | |`hipblasDsymm`|3.6.0| | | | |
 |`cublasDsymm_64`|12.0| | | | | | | | | |
 |`cublasDsymm_v2`| | | | |`hipblasDsymm`|3.6.0| | | | |
@@ -1136,7 +1136,7 @@
 |`cublasSgemvBatched`|11.6| | | |`hipblasSgemvBatched`|1.6.0| | | | |
 |`cublasSgemvBatched_64`|12.0| | | |`hipblasSgemvBatched_64`|6.2.0| | | |6.2.0|
 |`cublasSgemvStridedBatched`|11.6| | | |`hipblasSgemvStridedBatched`|3.0.0| | | | |
-|`cublasSgemvStridedBatched_64`|12.0| | | | | | | | | |
+|`cublasSgemvStridedBatched_64`|12.0| | | |`hipblasSgemvStridedBatched_64`|6.2.0| | | |6.2.0|
 |`cublasSsymm`| | | | |`hipblasSsymm`|3.6.0| | | | |
 |`cublasSsymm_64`|12.0| | | | | | | | | |
 |`cublasSsymm_v2`| | | | |`hipblasSsymm`|3.6.0| | | | |
@@ -1180,7 +1180,7 @@
 |`cublasZgemvBatched`|11.6| | | |`hipblasZgemvBatched_v2`|6.0.0| | | | |
 |`cublasZgemvBatched_64`|12.0| | | |`hipblasZgemvBatched_v2_64`|6.2.0| | | |6.2.0|
 |`cublasZgemvStridedBatched`|11.6| | | |`hipblasZgemvStridedBatched_v2`|6.0.0| | | | |
-|`cublasZgemvStridedBatched_64`|12.0| | | | | | | | | |
+|`cublasZgemvStridedBatched_64`|12.0| | | |`hipblasZgemvStridedBatched_v2_64`|6.2.0| | | |6.2.0|
 |`cublasZhemm`| | | | |`hipblasZhemm_v2`|6.0.0| | | | |
 |`cublasZhemm_64`|12.0| | | | | | | | | |
 |`cublasZhemm_v2`| | | | |`hipblasZhemm_v2`|6.0.0| | | | |

diff --git a/docs/tables/CUBLAS_API_supported_by_HIP_and_ROC.md b/docs/tables/CUBLAS_API_supported_by_HIP_and_ROC.md
@@ -1034,7 +1034,7 @@
 |`cublasCgemvBatched`|11.6| | | |`hipblasCgemvBatched_v2`|6.0.0| | | | |`rocblas_cgemv_batched`|3.5.0| | | | |
 |`cublasCgemvBatched_64`|12.0| | | |`hipblasCgemvBatched_v2_64`|6.2.0| | | |6.2.0| | | | | | |
 |`cublasCgemvStridedBatched`|11.6| | | |`hipblasCgemvStridedBatched_v2`|6.0.0| | | | |`rocblas_cgemv_strided_batched`|3.5.0| | | | |
-|`cublasCgemvStridedBatched_64`|12.0| | | | | | | | | | | | | | | |
+|`cublasCgemvStridedBatched_64`|12.0| | | |`hipblasCgemvStridedBatched_v2_64`|6.2.0| | | |6.2.0| | | | | | |
 |`cublasChemm`| | | | |`hipblasChemm_v2`|6.0.0| | | | |`rocblas_chemm`|3.5.0| | | | |
 |`cublasChemm_64`|12.0| | | | | | | | | | | | | | | |
 |`cublasChemm_v2`| | | | |`hipblasChemm_v2`|6.0.0| | | | |`rocblas_chemm`|3.5.0| | | | |
@@ -1084,7 +1084,7 @@
 |`cublasDgemvBatched`|11.6| | | |`hipblasDgemvBatched`|3.0.0| | | | | | | | | | |
 |`cublasDgemvBatched_64`|12.0| | | |`hipblasDgemvBatched_64`|6.2.0| | | |6.2.0| | | | | | |
 |`cublasDgemvStridedBatched`|11.6| | | |`hipblasDgemvStridedBatched`|3.0.0| | | | | | | | | | |
-|`cublasDgemvStridedBatched_64`|12.0| | | | | | | | | | | | | | | |
+|`cublasDgemvStridedBatched_64`|12.0| | | |`hipblasDgemvStridedBatched_64`|6.2.0| | | |6.2.0| | | | | | |
 |`cublasDsymm`| | | | |`hipblasDsymm`|3.6.0| | | | |`rocblas_dsymm`|3.5.0| | | | |
 |`cublasDsymm_64`|12.0| | | | | | | | | | | | | | | |
 |`cublasDsymm_v2`| | | | |`hipblasDsymm`|3.6.0| | | | |`rocblas_dsymm`|3.5.0| | | | |
@@ -1136,7 +1136,7 @@
 |`cublasSgemvBatched`|11.6| | | |`hipblasSgemvBatched`|1.6.0| | | | | | | | | | |
 |`cublasSgemvBatched_64`|12.0| | | |`hipblasSgemvBatched_64`|6.2.0| | | |6.2.0| | | | | | |
 |`cublasSgemvStridedBatched`|11.6| | | |`hipblasSgemvStridedBatched`|3.0.0| | | | | | | | | | |
-|`cublasSgemvStridedBatched_64`|12.0| | | | | | | | | | | | | | | |
+|`cublasSgemvStridedBatched_64`|12.0| | | |`hipblasSgemvStridedBatched_64`|6.2.0| | | |6.2.0| | | | | | |
 |`cublasSsymm`| | | | |`hipblasSsymm`|3.6.0| | | | |`rocblas_ssymm`|3.5.0| | | | |
 |`cublasSsymm_64`|12.0| | | | | | | | | | | | | | | |
 |`cublasSsymm_v2`| | | | |`hipblasSsymm`|3.6.0| | | | |`rocblas_ssymm`|3.5.0| | | | |
@@ -1180,7 +1180,7 @@
 |`cublasZgemvBatched`|11.6| | | |`hipblasZgemvBatched_v2`|6.0.0| | | | |`rocblas_zgemv_batched`|3.5.0| | | | |
 |`cublasZgemvBatched_64`|12.0| | | |`hipblasZgemvBatched_v2_64`|6.2.0| | | |6.2.0| | | | | | |
 |`cublasZgemvStridedBatched`|11.6| | | |`hipblasZgemvStridedBatched_v2`|6.0.0| | | | |`rocblas_zgemv_strided_batched`|3.5.0| | | | |
-|`cublasZgemvStridedBatched_64`|12.0| | | | | | | | | | | | | | | |
+|`cublasZgemvStridedBatched_64`|12.0| | | |`hipblasZgemvStridedBatched_v2_64`|6.2.0| | | |6.2.0| | | | | | |
 |`cublasZhemm`| | | | |`hipblasZhemm_v2`|6.0.0| | | | |`rocblas_zhemm`|3.5.0| | | | |
 |`cublasZhemm_64`|12.0| | | | | | | | | | | | | | | |
 |`cublasZhemm_v2`| | | | |`hipblasZhemm_v2`|6.0.0| | | | |`rocblas_zhemm`|3.5.0| | | | |

diff --git a/src/CUDA2HIP_BLAS_API_functions.cpp b/src/CUDA2HIP_BLAS_API_functions.cpp
@@ -459,13 +459,13 @@ const std::map<llvm::StringRef, hipCounter> CUDA_BLAS_FUNCTION_MAP {
   {"cublasTSSgemvBatched",                                 {"hipblasTSSgemvBatched",                                     "rocblas_tssgemv_batched",                            CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_3, HIP_UNSUPPORTED}},
   {"cublasTSSgemvBatched_64",                              {"hipblasTSSgemvBatched_64",                                  "",                                                   CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_3, UNSUPPORTED}},
   {"cublasSgemvStridedBatched",                            {"hipblasSgemvStridedBatched",                                "",                                                   CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_3, ROC_UNSUPPORTED}},
-  {"cublasSgemvStridedBatched_64",                         {"hipblasSgemvStridedBatched_64",                             "",                                                   CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_3, UNSUPPORTED}},
+  {"cublasSgemvStridedBatched_64",                         {"hipblasSgemvStridedBatched_64",                             "",                                                   CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_3, ROC_UNSUPPORTED}},
   {"cublasDgemvStridedBatched",                            {"hipblasDgemvStridedBatched",                                "",                                                   CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_3, ROC_UNSUPPORTED}},
-  {"cublasDgemvStridedBatched_64",                         {"hipblasDgemvStridedBatched_64",                             "",                                                   CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_3, UNSUPPORTED}},
+  {"cublasDgemvStridedBatched_64",                         {"hipblasDgemvStridedBatched_64",                             "",                                                   CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_3, ROC_UNSUPPORTED}},
   {"cublasCgemvStridedBatched",                            {"hipblasCgemvStridedBatched_v2",                             "rocblas_cgemv_strided_batched",                      CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_3}},
-  {"cublasCgemvStridedBatched_64",                         {"hipblasCgemvStridedBatched_64",                             "",                                                   CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_3, UNSUPPORTED}},
+  {"cublasCgemvStridedBatched_64",                         {"hipblasCgemvStridedBatched_v2_64",                          "",                                                   CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_3, ROC_UNSUPPORTED}},
   {"cublasZgemvStridedBatched",                            {"hipblasZgemvStridedBatched_v2",                             "rocblas_zgemv_strided_batched",                      CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_3}},
-  {"cublasZgemvStridedBatched_64",                         {"hipblasZgemvStridedBatched_64",                             "",                                                   CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_3, UNSUPPORTED}},
+  {"cublasZgemvStridedBatched_64",                         {"hipblasZgemvStridedBatched_v2_64",                          "",                                                   CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_3, ROC_UNSUPPORTED}},
   {"cublasHSHgemvStridedBatched",                          {"hipblasHSHgemvStridedBatched",                              "rocblas_hshgemv_strided_batched",                    CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_3, HIP_UNSUPPORTED}},
   {"cublasHSHgemvStridedBatched_64",                       {"hipblasHSHgemvStridedBatched_64",                           "",                                                   CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_3, UNSUPPORTED}},
   {"cublasHSSgemvStridedBatched",                          {"hipblasHSSgemvStridedBatched",                              "rocblas_hssgemv_strided_batched",                    CONV_LIB_FUNC, API_BLAS, SEC::BLAS_LEVEL_3, HIP_UNSUPPORTED}},
@@ -2070,6 +2070,10 @@ const std::map<llvm::StringRef, hipAPIversions> HIP_BLAS_FUNCTION_VER_MAP {
   {"hipblasDgemvStridedBatched",                           {HIP_3000, HIP_0,    HIP_0   }},
   {"hipblasSgemvBatched",                                  {HIP_1060, HIP_0,    HIP_0   }},
   {"hipblasDgemvBatched",                                  {HIP_3000, HIP_0,    HIP_0   }},
+  {"hipblasSgemvStridedBatched_64",                        {HIP_6020, HIP_0,    HIP_0,  HIP_LATEST}},
+  {"hipblasDgemvStridedBatched_64",                        {HIP_6020, HIP_0,    HIP_0,  HIP_LATEST}},
+  {"hipblasCgemvStridedBatched_v2_64",                     {HIP_6020, HIP_0,    HIP_0,  HIP_LATEST}},
+  {"hipblasZgemvStridedBatched_v2_64",                     {HIP_6020, HIP_0,    HIP_0,  HIP_LATEST}},
 
   {"rocblas_status_to_string",                             {HIP_3050, HIP_0,    HIP_0   }},
   {"rocblas_sscal",                                        {HIP_1050, HIP_0,    HIP_0   }},

diff --git a/tests/unit_tests/synthetic/libraries/cublas2hipblas_v2.cu b/tests/unit_tests/synthetic/libraries/cublas2hipblas_v2.cu
@@ -2313,6 +2313,26 @@ int main() {
   // HIP: HIPBLAS_EXPORT hipblasStatus_t hipblasZgemvBatched_v2_64(hipblasHandle_t handle, hipblasOperation_t trans, int64_t m, int64_t n, const hipDoubleComplex* alpha, const hipDoubleComplex* const AP[], int64_t lda, const hipDoubleComplex* const x[], int64_t incx, const hipDoubleComplex* beta, hipDoubleComplex* const y[], int64_t incy, int64_t batchCount);
   // CHECK: blasStatus = hipblasZgemvBatched_v2_64(blasHandle, blasOperation, m_64, n_64, &dcomplexa, dcomplexAarray_const, lda_64, dcomplexXarray_const, incx_64, &dcomplexb, dcomplexYarray, incy_64, batchCount_64);
   blasStatus = cublasZgemvBatched_64(blasHandle, blasOperation, m_64, n_64, &dcomplexa, dcomplexAarray_const, lda_64, dcomplexXarray_const, incx_64, &dcomplexb, dcomplexYarray, incy_64, batchCount_64);
+
+  // CUDA: CUBLASAPI cublasStatus_t CUBLASWINAPI cublasSgemvStridedBatched_64(cublasHandle_t handle, cublasOperation_t trans, int64_t m, int64_t n, const float* alpha, const float* A, int64_t lda, long long int strideA, const float* x, int64_t incx, long long int stridex, const float* beta, float* y, int64_t incy, long long int stridey, int64_t batchCount);
+  // HIP: HIPBLAS_EXPORT hipblasStatus_t hipblasSgemvStridedBatched_64(hipblasHandle_t handle, hipblasOperation_t transA, int64_t m, int64_t n, const float* alpha, const float* AP, int64_t lda, hipblasStride strideA, const float* x, int64_t incx, hipblasStride stridex, const float* beta, float* y, int64_t incy, hipblasStride stridey, int64_t batchCount);
+  // CHECK: blasStatus = hipblasSgemvStridedBatched_64(blasHandle, blasOperation, m_64, n_64, &fa, &fA, lda, strideA, &fx, incx_64, strideX, &fb, &fy, incy_64, strideY, batchCount);
+  blasStatus = cublasSgemvStridedBatched_64(blasHandle, blasOperation, m_64, n_64, &fa, &fA, lda, strideA, &fx, incx_64, strideX, &fb, &fy, incy_64, strideY, batchCount);
+
+  // CUDA: CUBLASAPI cublasStatus_t CUBLASWINAPI cublasDgemvStridedBatched_64(cublasHandle_t handle, cublasOperation_t trans, int64_t m, int64_t n, const double* alpha, const double* A, int64_t lda, long long int strideA, const double* x, int64_t incx, long long int stridex, const double* beta, double* y, int64_t incy, long long int stridey, int64_t batchCount);
+  // HIP: HIPBLAS_EXPORT hipblasStatus_t hipblasDgemvStridedBatched_64(hipblasHandle_t handle, hipblasOperation_t transA, int64_t m, int64_t n, const double* alpha, const double* AP, int64_t lda, hipblasStride strideA, const double* x, int64_t incx, hipblasStride stridex, const double* beta, double* y, int64_t incy, hipblasStride stridey, int64_t batchCount);
+  // CHECK: blasStatus = hipblasDgemvStridedBatched_64(blasHandle, blasOperation, m_64, n_64, &da, &dA, lda, strideA, &dx, incx_64, strideX, &db, &dy, incy_64, strideY, batchCount);
+  blasStatus = cublasDgemvStridedBatched_64(blasHandle, blasOperation, m_64, n_64, &da, &dA, lda, strideA, &dx, incx_64, strideX, &db, &dy, incy_64, strideY, batchCount);
+
+  // CUDA: CUBLASAPI cublasStatus_t CUBLASWINAPI cublasCgemvStridedBatched_64(cublasHandle_t handle, cublasOperation_t trans, int64_t m, int64_t n, const cuComplex* alpha, const cuComplex* A, int64_t lda, long long int strideA, const cuComplex* x, int64_t incx, long long int stridex, const cuComplex* beta, cuComplex* y, int64_t incy, long long int stridey, int64_t batchCount);
+  // HIP: HIPBLAS_EXPORT hipblasStatus_t hipblasCgemvStridedBatched_v2_64(hipblasHandle_t handle, hipblasOperation_t transA, int64_t m, int64_t n,const hipComplex* alpha, const hipComplex* AP, int64_t lda,hipblasStride strideA, const hipComplex* x, int64_t incx, hipblasStride stridex, const hipComplex* beta, hipComplex* y, int64_t incy, hipblasStride stridey, int64_t batchCount);
+  // CHECK: blasStatus = hipblasCgemvStridedBatched_v2_64(blasHandle, blasOperation, m_64, n_64, &complexa, &complexA, lda, strideA, &complexx, incx_64, strideX, &complexb, &complexy, incy_64, strideY, batchCount);
+  blasStatus = cublasCgemvStridedBatched_64(blasHandle, blasOperation, m_64, n_64, &complexa, &complexA, lda, strideA, &complexx, incx_64, strideX, &complexb, &complexy, incy_64, strideY, batchCount);
+
+  // CUDA: CUBLASAPI cublasStatus_t CUBLASWINAPI cublasZgemvStridedBatched_64(cublasHandle_t handle, cublasOperation_t trans, int64_t m, int64_t n, const cuDoubleComplex* alpha, const cuDoubleComplex* A, int64_t lda, long long int strideA, const cuDoubleComplex* x, int64_t incx, long long int stridex, const cuDoubleComplex* beta, cuDoubleComplex* y, int64_t incy, long long int stridey, int64_t batchCount);
+  // HIP: HIPBLAS_EXPORT hipblasStatus_t hipblasZgemvStridedBatched_v2_64(hipblasHandle_t handle, hipblasOperation_t transA, int64_t m, int64_t n, const hipDoubleComplex* alpha, const hipDoubleComplex* AP, int64_t lda, hipblasStride strideA, const hipDoubleComplex* x, int64_t incx, hipblasStride stridex, const hipDoubleComplex* beta, hipDoubleComplex* y, int64_t incy, hipblasStride stridey, int64_t batchCount);
+  // CHECK: blasStatus = hipblasZgemvStridedBatched_v2_64(blasHandle, blasOperation, m_64, n_64, &dcomplexa, &dcomplexA, lda, strideA, &dcomplexx, incx_64, strideX, &dcomplexb, &dcomplexy, incy_64, strideY, batchCount);
+  blasStatus = cublasZgemvStridedBatched_64(blasHandle, blasOperation, m_64, n_64, &dcomplexa, &dcomplexA, lda, strideA, &dcomplexx, incx_64, strideX, &dcomplexb, &dcomplexy, incy_64, strideY, batchCount);
 #endif
 
   return 0;