vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

avx bf16 vec dot

Browse source
This commit is contained in:
Eve 2024-11-01 22:21:21 -04:00
parent 34b9f0de6b
commit dca0deb3d8
1 changed files with 5 additions and 1 deletions
Download patch file Download diff file Expand all files Collapse all files

6
ggml/src/ggml.c
View file

@ -2245,8 +2245,12 @@ static void ggml_vec_dot_bf16(int n, float * restrict s, size_t bs, ggml_bf16_t
sumf += (ggml_float)_mm512_reduce_add_ps(c2);
#undef LOAD
#elif defined(__AVX2__)
#elif defined(__AVX2__) || defined(__AVX__)
#if defined(__AVX2__)
#define LOAD(p) _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_cvtepu16_epi32(_mm_loadu_si128((const __m128i *)(p))), 16))
#else
#define LOAD(p) _mm256_castsi256_ps(_mm256_insertf128_si256(_mm256_castsi128_si256(_mm_slli_epi32(_mm_cvtepu16_epi32(_mm_loadu_si128((const __m128i *)(p))), 16)), (_mm_slli_epi32(_mm_cvtepu16_epi32(_mm_bsrli_si128(_mm_loadu_si128((const __m128i *)(p)), 8)), 16)), 1))
#endif
__m256 c1 = _mm256_setzero_ps();
__m256 c2 = _mm256_setzero_ps();
__m256 c3 = _mm256_setzero_ps();