DPCT1046

Intel® DPC++ Compatibility Tool Developer Guide and Reference

Download PDF

ID 768918

Date 6/24/2024

Version

Public

A newer version of this document is available. Customers should click here to go to the newest version.

Visible to Intel only — GUID: GUID-C15A4969-D20A-4FDD-A888-2F0C3189996D

View Details

DPCT1046

NOTE:

This diagnostic message is no longer generated by Intel® DPC++ Compatibility Tool 2023.0 or later.

Message

The <original API name> was not migrated because <reason>. You need to adjust the code.

Detailed Help

Not all data type combinations are supported by mkl::blas::gemm().

This may be due to one of the following reasons:

Not all values of parameters could be evaluated in migration.
The combination of matrix data type and scalar type is unsupported.

Use a supported data type to rewrite the code.

Suggestions to Fix

Please refer to the gemm topic of the Intel® oneAPI Math Kernel Library (oneMKL) - Data Parallel C++ Developer Reference for supported data types to fix the code manually.

For example, this original CUDA* code:

 void foo(cublasHandle_t handle, float alpha, float beta, void *a, void *b,
          void *c, cudaDataType_t type) {
   cublasSgemmEx(handle, CUBLAS_OP_C, CUBLAS_OP_C, 2, 2, 2, &alpha, a, type, 2,
                 b, type, 2, &beta, c, type, 2);
 }

results in the following migrated SYCL* code:

 void foo(dpct::queue_ptr handle, float alpha, float beta, void *a, void *b,
          void *c, dpct::library_data_t type) {
   /*
   DPCT1046:0: The cublasSgemmEx was not migrated because not all values of
   parameters could be evaluated in migration. You need to adjust the code.
   */
   cublasSgemmEx(handle, oneapi::mkl::transpose::conjtrans,
                 oneapi::mkl::transpose::conjtrans, 2, 2, 2, &alpha, a, type, 2,
                 b, type, 2, &beta, c, type, 2);
 }

which is rewritten to:

 void foo(dpct::queue_ptr handle, float alpha, float beta, void *a, void *b,
          void *c, dpct::library_data_t type) {
   switch (type) {
     case dpct::library_data_t::real_bfloat16: {
       oneapi::mkl::blas::column_major::gemm(
           *handle, oneapi::mkl::transpose::conjtrans,
           oneapi::mkl::transpose::conjtrans, 2, 2, 2, alpha,
           (oneapi::mkl::bfloat16 *)a, 2, (oneapi::mkl::bfloat16 *)b, 2, beta,
           (oneapi::mkl::bfloat16 *)c, 2);
       break;
     }
     case dpct::library_data_t::real_half: {
       oneapi::mkl::blas::column_major::gemm(
           *handle, oneapi::mkl::transpose::conjtrans,
           oneapi::mkl::transpose::conjtrans, 2, 2, 2, alpha, (sycl::half *)a, 2,
           (sycl::half *)b, 2, beta, (sycl::half *)c, 2);
       break;
     }
     case dpct::library_data_t::real_float: {
       oneapi::mkl::blas::column_major::gemm(
           *handle, oneapi::mkl::transpose::conjtrans,
           oneapi::mkl::transpose::conjtrans, 2, 2, 2, alpha, (float *)a, 2,
           (float *)b, 2, beta, (float *)c, 2);
       break;
     }
     default:
       throw std::runtime_error("the data type is unsupported");
   }
 }

Select Your Language

Using Intel.com Search

Quick Links

Recent Searches

Advanced Search

Only search in

Intel® DPC++ Compatibility Tool Developer Guide and Reference

DPCT1046

Message

Detailed Help

Suggestions to Fix