diff options
Diffstat (limited to 'avx2_gemm.h')
-rw-r--r-- | avx2_gemm.h | 15 |
1 files changed, 3 insertions, 12 deletions
diff --git a/avx2_gemm.h b/avx2_gemm.h index c5ca0bc..a03ff09 100644 --- a/avx2_gemm.h +++ b/avx2_gemm.h @@ -80,7 +80,7 @@ struct AVX2_16bit { avx2::SelectColumnsOfB((const __m256i*)input, (__m256i*)output, rows * 2, cols_begin, cols_end); } - INTGEMM_MULTIPLY16(__m256i, INTGEMM_AVX2, OnAVX2) + INTGEMM_MULTIPLY16(__m256i, INTGEMM_AVX2, CPUType::CPU_AVX2) constexpr static const char *const kName = "16-bit INTGEMM_AVX2"; @@ -163,22 +163,13 @@ struct AVX2_8bit { static const Index kBTileRow = 32; static const Index kBTileCol = 8; -/* - INTGEMM_AVX2 static void PrepareB(const float *input, int8_t *output, float quant_mult, Index rows, Index cols) { - PrepareBFor8(input, output, avx2::QuantizeTile8(quant_mult), rows, cols); - }*/ - INTGEMM_PREPARE_B_8(INTGEMM_AVX2, avx2::QuantizeTile8) INTGEMM_AVX2 static void SelectColumnsB(const int8_t *input, int8_t *output, Index rows, const Index *cols_begin, const Index *cols_end) { avx2::SelectColumnsOfB((const __m256i*)input, (__m256i*)output, rows, cols_begin, cols_end); } -/* - INTGEMM_AVX2 static void Multiply(const int8_t *A, const int8_t *B, float *C, float unquant_mult, Index A_rows, Index width, Index B_cols) { - //Multiply8_SSE2OrAVX2<Multiply8_AVXAVX2, __m256i, __m256>(A, B, C, unquant_mult, A_rows, width, B_cols); - Multiply8_SSE2OrAVX2__m256i<JustUnquantizeC>(A, B, JustUnquantizeC(C, unquant_mult), A_rows, width, B_cols); - }*/ - INTGEMM_MULTIPLY8(__m256i, INTGEMM_AVX2, OnAVX2) + + INTGEMM_MULTIPLY8(__m256i, INTGEMM_AVX2, CPUType::CPU_AVX2) constexpr static const char *const kName = "8-bit INTGEMM_AVX2"; |