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

update avx2

Browse source
This commit is contained in:
Eddie-Wang1120 2024-06-20 00:12:58 +08:00
parent fa9a742b46
commit 230396bc5b
1 changed files with 20 additions and 11 deletions
Download patch file Download diff file Expand all files Collapse all files

31
ggml-quants.c
View file

@ -3806,18 +3806,27 @@ void ggml_vec_dot_q2_2_q8_0(int n, float * restrict s, size_t bs, const void * r
const __m256 d = _mm256_set1_ps( GGML_FP16_TO_FP32(y[i].d) ); const __m256 d = _mm256_set1_ps( GGML_FP16_TO_FP32(y[i].d) );
__m256i xq8 = _mm256_set_epi32( __m128i xq8b = _mm_loadu_si64(x[i].qs);
(int)q22_grid[x[i].qs[7]], __m256i xq8 = MM256_SET_M128I(xq8b, xq8b);
(int)q22_grid[x[i].qs[6]], __m256i xq8l = _mm256_shuffle_epi8(xq8, _mm256_set_epi8(5, -1, 5, -1, 5, -1, 5, -1,
(int)q22_grid[x[i].qs[5]], 4, -1, 4, -1, 4, -1, 4, -1,
(int)q22_grid[x[i].qs[4]], 1, -1, 1, -1, 1, -1, 1, -1,
(int)q22_grid[x[i].qs[3]], 0, -1, 0, -1, 0, -1, 0, -1));
(int)q22_grid[x[i].qs[2]], __m256i xq8h = _mm256_shuffle_epi8(xq8, _mm256_set_epi8(7, -1, 7, -1, 7, -1, 7, -1,
(int)q22_grid[x[i].qs[1]], 6, -1, 6, -1, 6, -1, 6, -1,
(int)q22_grid[x[i].qs[0]] 3, -1, 3, -1, 3, -1, 3, -1,
); 2, -1, 2, -1, 2, -1, 2, -1));
__m256i shift = _mm256_set_epi16(64, 16, 4, 1,
64, 16, 4, 1,
64, 16, 4, 1,
64, 16, 4, 1);
xq8l = _mm256_mullo_epi16(xq8l, shift);
xq8h = _mm256_mullo_epi16(xq8h, shift);
xq8l = _mm256_srai_epi16(xq8l, 14);
xq8h = _mm256_srai_epi16(xq8h, 14);
xq8 = _mm256_packs_epi16(xq8l, xq8h);
__m256i yq8 = _mm256_loadu_si256((const __m256i*)(y[i].qs)); __m256i yq8 = _mm256_lddqu_si256((const __m256i*)(y[i].qs));
const __m256 q = mul_sum_i8_pairs_float(xq8, yq8); const __m256 q = mul_sum_i8_pairs_float(xq8, yq8);
acc = _mm256_fmadd_ps( d, q, acc ); acc = _mm256_fmadd_ps( d, q, acc );