diff options
| author | Kenneth Heafield <github@kheafield.com> | 2018-06-17 21:54:41 +0300 |
|---|---|---|
| committer | Kenneth Heafield <github@kheafield.com> | 2018-06-17 21:54:41 +0300 |
| commit | 8c29dd6d9e5e0821a92a3e2133f1b10ae4007ae2 (patch) | |
| tree | 195ea5f013f653442e7b3a32b50e0d1b4f40d616 /avx512_gemm.cc | |
| parent | bc1ab9d4a360cd2078d067cca9fdb6fedf78158e (diff) | |
Genericize the prepareb code
Diffstat (limited to 'avx512_gemm.cc')
| -rw-r--r-- | avx512_gemm.cc | 4 |
1 files changed, 1 insertions, 3 deletions
diff --git a/avx512_gemm.cc b/avx512_gemm.cc index 89c300c..ea2b85d 100644 --- a/avx512_gemm.cc +++ b/avx512_gemm.cc @@ -20,10 +20,8 @@ namespace { // Load from memory, multiply, and convert to int32_t. inline __m512i QuantizerGrab(const float *input, const __m512 quant_mult_reg) { - // Load 16 floats - __m512 val = _mm512_load_ps(input); // Multiply each by the quantization factor. - val = _mm512_mul_ps(val, quant_mult_reg); + __m512 val = _mm512_mul_ps(*reinterpret_cast<const __m512*>(input), quant_mult_reg); // Cast to 32-bit int return _mm512_cvtps_epi32(val); } |