FLA_Trsm_external_gpu.c File Reference

(r12600)

Functions
FLA_Error	FLA_Trsm_external_gpu (FLA_Side side, FLA_Uplo uplo, FLA_Trans trans, FLA_Diag diag, FLA_Obj alpha, FLA_Obj A, void *A_gpu, FLA_Obj B, void *B_gpu)

---

Function Documentation

FLA_Error FLA_Trsm_external_gpu	(	FLA_Side	side,
		FLA_Uplo	uplo,
		FLA_Trans	trans,
		FLA_Diag	diag,
		FLA_Obj	alpha,
		FLA_Obj	A,
		void *	A_gpu,
		FLA_Obj	B,
		void *	B_gpu
	)

References FLA_Check_error_level(), FLA_Obj_datatype(), FLA_Obj_has_zero_dim(), FLA_Obj_length(), FLA_Obj_width(), FLA_Param_map_flame_to_netlib_diag(), FLA_Param_map_flame_to_netlib_side(), FLA_Param_map_flame_to_netlib_trans(), FLA_Param_map_flame_to_netlib_uplo(), and FLA_Trsm_check().

Referenced by FLASH_Queue_exec_task_gpu().

{
  FLA_Datatype datatype;
  int          m_B, n_B;
  int          ldim_A;
  int          ldim_B;
  char         blas_side; 
  char         blas_uplo;
  char         blas_trans;
  char         blas_diag;

  if ( FLA_Check_error_level() == FLA_FULL_ERROR_CHECKING ) 
    FLA_Trsm_check( side, uplo, trans, diag, alpha, A, B );

  if ( FLA_Obj_has_zero_dim( B ) ) return FLA_SUCCESS;

  datatype = FLA_Obj_datatype( A );

  ldim_A   = FLA_Obj_length( A );

  m_B      = FLA_Obj_length( B );
  n_B      = FLA_Obj_width( B );
  ldim_B   = FLA_Obj_length( B );

  FLA_Param_map_flame_to_netlib_side( side, &blas_side );
  FLA_Param_map_flame_to_netlib_uplo( uplo, &blas_uplo );
  FLA_Param_map_flame_to_netlib_trans( trans, &blas_trans );
  FLA_Param_map_flame_to_netlib_diag( diag, &blas_diag );


  switch( datatype ){

  case FLA_FLOAT:
  {
    float *buff_alpha = ( float * ) FLA_FLOAT_PTR( alpha );

    cublasStrsm( blas_side,
                 blas_uplo,
                 blas_trans,
                 blas_diag,
                 m_B,
                 n_B,
                 *buff_alpha,
                 ( float * ) A_gpu, ldim_A,
                 ( float * ) B_gpu, ldim_B );
    
    break;
  }

  case FLA_DOUBLE:
  {
    double *buff_alpha = ( double * ) FLA_DOUBLE_PTR( alpha );

    cublasDtrsm( blas_side,
                 blas_uplo,
                 blas_trans,
                 blas_diag,
                 m_B,
                 n_B,
                 *buff_alpha,
                 ( double * ) A_gpu, ldim_A,
                 ( double * ) B_gpu, ldim_B );

    break;
  }

  case FLA_COMPLEX:
  {
    cuComplex *buff_alpha = ( cuComplex * ) FLA_COMPLEX_PTR( alpha );

    cublasCtrsm( blas_side,
                 blas_uplo,
                 blas_trans,
                 blas_diag,
                 m_B,
                 n_B,
                 *buff_alpha,
                 ( cuComplex * ) A_gpu, ldim_A,
                 ( cuComplex * ) B_gpu, ldim_B );

    break;
  }

  case FLA_DOUBLE_COMPLEX:
  {
    cuDoubleComplex *buff_alpha = ( cuDoubleComplex * ) FLA_DOUBLE_COMPLEX_PTR( alpha );

    cublasZtrsm( blas_side,
                 blas_uplo,
                 blas_trans,
                 blas_diag,
                 m_B,
                 n_B,
                 *buff_alpha,
                 ( cuDoubleComplex * ) A_gpu, ldim_A,
                 ( cuDoubleComplex * ) B_gpu, ldim_B );

    break;
  }

  }

  return FLA_SUCCESS;
}