Esempi in C++ (Cpp) per LIBXSMM_GEMM_DECLARE_FLAGS

Esempio n. 1

File: libxsmm_gemm.c Progetto: danielpeter/libxsmm

LIBXSMM_API_DEFINITION void libxsmm_blas_dgemm(const char* transa, const char* transb,
  const libxsmm_blasint* m, const libxsmm_blasint* n, const libxsmm_blasint* k,
  const double* alpha, const double* a, const libxsmm_blasint* lda,
  const double* b, const libxsmm_blasint* ldb,
  const double* beta, double* c, const libxsmm_blasint* ldc)
{
  LIBXSMM_GEMM_DECLARE_FLAGS(flags, transa, transb, m, n, k, a, b, c);
  LIBXSMM_BLAS_DGEMM(flags, *m, *n, *k,
    0 != alpha ? *alpha : ((double)LIBXSMM_ALPHA),
    a, *(lda ? lda : LIBXSMM_LD(m, k)), b, *(ldb ? ldb : LIBXSMM_LD(k, n)),
    0 != beta ? *beta : ((double)LIBXSMM_BETA),
    c, *(ldc ? ldc : LIBXSMM_LD(m, n)));
}

Esempio n. 2

File: libxsmm_gemm.c Progetto: danielpeter/libxsmm

LIBXSMM_API_DEFINITION void LIBXSMM_FSYMBOL(__wrap_dgemm)(
  const char* transa, const char* transb,
  const libxsmm_blasint* m, const libxsmm_blasint* n, const libxsmm_blasint* k,
  const double* alpha, const double* a, const libxsmm_blasint* lda,
  const double* b, const libxsmm_blasint* ldb,
  const double* beta, double* c, const libxsmm_blasint* ldc)
{
  const int tm = internal_gemm_tile[0/*DP*/][0/*M*/];
  const int tn = internal_gemm_tile[0/*DP*/][1/*N*/];
  const int tk = internal_gemm_tile[0/*DP*/][2/*K*/];
  LIBXSMM_GEMM_DECLARE_FLAGS(flags, transa, transb, m, n, k, a, b, c);
  LIBXSMM_EXT_GEMM_XGEMM(LIBXSMM_EXT_GEMM_FOR_INIT, LIBXSMM_EXT_GEMM_FOR_LOOP_BEGIN,
    LIBXSMM_EXT_GEMM_FOR_LOOP_BODY, LIBXSMM_EXT_GEMM_FOR_LOOP_END,
    double, flags, internal_gemm_nt, tm, tn, tk, *m, *n, *k,
    0 != alpha ? *alpha : ((double)LIBXSMM_ALPHA),
    a, *(lda ? lda : LIBXSMM_LD(m, k)), b, *(ldb ? ldb : LIBXSMM_LD(k, n)),
    0 != beta ? *beta : ((double)LIBXSMM_BETA),
    c, *(ldc ? ldc : LIBXSMM_LD(m, n)));
}

Esempio n. 3

File: libxsmm_ext_gemm.c Progetto: michel2323/libxsmm

LIBXSMM_API_DEFINITION void libxsmm_sgemm_omp(const char* transa, const char* transb,
  const libxsmm_blasint* m, const libxsmm_blasint* n, const libxsmm_blasint* k,
  const float* alpha, const float* a, const libxsmm_blasint* lda,
  const float* b, const libxsmm_blasint* ldb,
  const float* beta, float* c, const libxsmm_blasint* ldc)
{
  LIBXSMM_GEMM_DESCRIPTOR_DIM_TYPE tm, tn, tk;
  LIBXSMM_GEMM_DECLARE_FLAGS(flags, transa, transb);
  LIBXSMM_INIT
  tm = libxsmm_gemm_tile[1/*SP*/][0/*M*/];
  tn = libxsmm_gemm_tile[1/*SP*/][1/*N*/];
  tk = libxsmm_gemm_tile[1/*SP*/][2/*K*/];
  assert(0 < tm && 0 < tn && 0 < tk && 0 < libxsmm_nt);
#if defined(_OPENMP)
  if (0 != libxsmm_mt) { /* enable OpenMP support */
    if (0 == LIBXSMM_MOD2(libxsmm_mt, 2)) { /* even: enable internal parallelization */
# if defined(LIBXSMM_EXT_TASKS)
      if (0 == libxsmm_tasks)
# endif
      {
        LIBXSMM_TILED_XGEMM(LIBXSMM_EXT_FOR_PARALLEL, LIBXSMM_NOOP, LIBXSMM_EXT_FOR_SINGLE,
          LIBXSMM_GEMM_COLLAPSE, LIBXSMM_EXT_FOR_LOOP, LIBXSMM_EXT_FOR_KERNEL, LIBXSMM_EXT_FOR_SYNC,
          LIBXSMM_EXT_MIN_NTASKS, LIBXSMM_EXT_OVERHEAD, libxsmm_nt,
          float, flags | LIBXSMM_GEMM_FLAG_F32PREC, tm, tn, tk, *m, *n, *k,
          0 != alpha ? *alpha : ((float)LIBXSMM_ALPHA),
          a, *(lda ? lda : LIBXSMM_LD(m, k)), b, *(ldb ? ldb : LIBXSMM_LD(k, n)),
          0 != beta ? *beta : ((float)LIBXSMM_BETA),
          c, *(ldc ? ldc : LIBXSMM_LD(m, n)));
      }
# if defined(LIBXSMM_EXT_TASKS)
      else {
        LIBXSMM_TILED_XGEMM(LIBXSMM_EXT_TSK_PARALLEL, LIBXSMM_EXT_SINGLE, LIBXSMM_NOOP,
          LIBXSMM_GEMM_COLLAPSE, LIBXSMM_EXT_TSK_LOOP, LIBXSMM_EXT_TSK_KERNEL_VARS, LIBXSMM_NOOP,
          LIBXSMM_EXT_MIN_NTASKS, LIBXSMM_EXT_OVERHEAD, libxsmm_nt,
          float, flags | LIBXSMM_GEMM_FLAG_F32PREC, tm, tn, tk, *m, *n, *k,
          0 != alpha ? *alpha : ((float)LIBXSMM_ALPHA),
          a, *(lda ? lda : LIBXSMM_LD(m, k)), b, *(ldb ? ldb : LIBXSMM_LD(k, n)),
          0 != beta ? *beta : ((float)LIBXSMM_BETA),
          c, *(ldc ? ldc : LIBXSMM_LD(m, n)));
      }
# endif
    }