diff options
Diffstat (limited to 'libgfortran/generated/matmul_i4.c')
| -rw-r--r-- | libgfortran/generated/matmul_i4.c | 38 |
1 files changed, 38 insertions, 0 deletions
diff --git a/libgfortran/generated/matmul_i4.c b/libgfortran/generated/matmul_i4.c index 475b214ae79..f8f0cdb96cc 100644 --- a/libgfortran/generated/matmul_i4.c +++ b/libgfortran/generated/matmul_i4.c @@ -1734,6 +1734,24 @@ matmul_i4_avx512f (gfc_array_i4 * const restrict retarray, #endif /* HAVE_AVX512F */ +/* AMD-specifix funtions with AVX128 and FMA3/FMA4. */ + +#if defined(HAVE_AVX) && defined(HAVE_FMA3) && defined(HAVE_AVX128) +void +matmul_i4_avx128_fma3 (gfc_array_i4 * const restrict retarray, + gfc_array_i4 * const restrict a, gfc_array_i4 * const restrict b, int try_blas, + int blas_limit, blas_call gemm) __attribute__((__target__("avx,fma"))); +internal_proto(matmul_i4_avx128_fma3); +#endif + +#if defined(HAVE_AVX) && defined(HAVE_FMA4) && defined(HAVE_AVX128) +void +matmul_i4_avx128_fma4 (gfc_array_i4 * const restrict retarray, + gfc_array_i4 * const restrict a, gfc_array_i4 * const restrict b, int try_blas, + int blas_limit, blas_call gemm) __attribute__((__target__("avx,fma4"))); +internal_proto(matmul_i4_avx128_fma4); +#endif + /* Function to fall back to if there is no special processor-specific version. */ static void matmul_i4_vanilla (gfc_array_i4 * const restrict retarray, @@ -2332,6 +2350,26 @@ void matmul_i4 (gfc_array_i4 * const restrict retarray, } #endif /* HAVE_AVX */ } + else if (__cpu_model.__cpu_vendor == VENDOR_AMD) + { +#if defined(HAVE_AVX) && defined(HAVE_FMA3) && defined(HAVE_AVX128) + if ((__cpu_model.__cpu_features[0] & (1 << FEATURE_AVX)) + && (__cpu_model.__cpu_features[0] & (1 << FEATURE_FMA))) + { + matmul_fn = matmul_i4_avx128_fma3; + goto store; + } +#endif +#if defined(HAVE_AVX) && defined(HAVE_FMA4) && defined(HAVE_AVX128) + if ((__cpu_model.__cpu_features[0] & (1 << FEATURE_AVX)) + && (__cpu_model.__cpu_features[0] & (1 << FEATURE_FMA4))) + { + matmul_fn = matmul_i4_avx128_fma4; + goto store; + } +#endif + + } store: __atomic_store_n (&matmul_p, matmul_fn, __ATOMIC_RELAXED); } |