DPCT1047

Intel® DPC++ Compatibility Tool Developer Guide and Reference

Download PDF

ID 768918

Date 10/31/2024

Version

Public

Visible to Intel only — GUID: GUID-A682C1B8-695E-4718-91FF-77DF811B7E8F

View Details

DPCT1047

Message

The meaning of <parameter name> in the <API name> is different from the <API name>. You may need to check the migrated code.

Detailed Help

In cuBLAS* and cuSolver* getrf API, the LU factorization is done as P*A=L*U; in Intel® oneAPI Math Kernel Library API, it is done as A=P*L*U. The result of the matrix P may be different.

Suggestions to Fix

If the matrix P is only used in library API, ignore this warning. If P is used other ways, you may need to adjust the value of P.

For example, this original CUDA* code:

 void foo(cublasHandle_t handle, float **a_array, float **b_array, int *p_array,
          int *info_array) {
   cublasSgetrfBatched(handle, 2, a_array, 2, p_array, info_array, 1);
   cublasSgetrsBatched(handle, CUBLAS_OP_N, 2, 2, a_array, 2, p_array, b_array,
                       2, info_array, 1);
 }

results in the following migrated SYCL* code:

 void foo(dpct::queue_ptr handle, float **a_array, float **b_array, int *p_array,
          int *info_array) {
   /*
   DPCT1047:0: The meaning of p_array in the dpct::getrf_batch_wrapper is
   different from the cublasSgetrfBatched. You may need to check the migrated
   code.
   */
   dpct::getrf_batch_wrapper(*handle, 2, a_array, 2, p_array, info_array, 1);
   dpct::getrs_batch_wrapper(*handle, oneapi::mkl::transpose::nontrans, 2, 2,
                             const_cast<float const **>(a_array), 2, p_array,
                             b_array, 2, info_array, 1);
 }

which is rewritten to:

 void foo(dpct::queue_ptr handle, float **a_array, float **b_array, int *p_array,
          int *info_array) {
   // cublas/cusolver API getrs need the output of cublas/cusolver getrf as input.
   // MKL/dpct helper API getrs need the output of MKL/dpct helper getrf as input.
   // In this case, matrix P is only used as a temporary data between library API
   // invocations, so the warning can be ignored.
   dpct::getrf_batch_wrapper(*handle, 2, a_array, 2, p_array, info_array, 1);
   dpct::getrs_batch_wrapper(*handle, oneapi::mkl::transpose::nontrans, 2, 2,
                             const_cast<float const **>(a_array), 2, p_array,
                             b_array, 2, info_array, 1);
 }

Select Your Language

Using Intel.com Search

Quick Links

Recent Searches

Advanced Search

Only search in

Intel® DPC++ Compatibility Tool Developer Guide and Reference

DPCT1047

Message

Detailed Help

Suggestions to Fix