diff options
author | Mateusz Chudyk <mateuszchudyk@gmail.com> | 2019-05-28 21:00:14 +0300 |
---|---|---|
committer | Mateusz Chudyk <mateuszchudyk@gmail.com> | 2019-05-28 21:07:00 +0300 |
commit | eef4b641f37f0c2208ac63edde191bacf9d3e4f6 (patch) | |
tree | c3a3ef2aa5c409cd098728cd6e704a7d91a4ac79 /ssse3_gemm.h | |
parent | 93785f4292c002a5bb4bbdfd7358785562166500 (diff) |
Outline quantize function
Diffstat (limited to 'ssse3_gemm.h')
-rw-r--r-- | ssse3_gemm.h | 6 |
1 files changed, 3 insertions, 3 deletions
diff --git a/ssse3_gemm.h b/ssse3_gemm.h index cf5d2c1..f7abbda 100644 --- a/ssse3_gemm.h +++ b/ssse3_gemm.h @@ -1,4 +1,5 @@ #pragma once + #include "types.h" #include <cstdint> #include <stdint.h> @@ -11,10 +12,9 @@ namespace intgemm { namespace ssse3 { -// Same implementation as AVX512, just width. Grabs 4 32-bit values. -//TODO duplicated function requires the removal of the annonymous namespace + INTGEMM_SSSE3 inline __m128i QuantizerGrab(const float *input, const __m128 quant_mult_reg) { - return _mm_cvtps_epi32(_mm_mul_ps(*reinterpret_cast<const __m128*>(input), quant_mult_reg)); + return quantize(*reinterpret_cast<const __m128*>(input), quant_mult_reg); } INTGEMM_SELECT_COL_B(INTGEMM_SSSE3, __m128i) |