Mirrors/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2024-12-28 15:18:26 +01:00
Code Issues Packages Projects Releases Wiki Activity

WIP

Browse Source
This commit is contained in:
Iwan Kawrakow 2024-03-01 18:48:08 +02:00
parent 160acecaba
commit e43e81a5d7
1 changed files with 6 additions and 6 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

12
ggml-quants.c
View File

@ -10118,8 +10118,8 @@ void ggml_vec_dot_iq3_s_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const v
const __m256i mask1 = _mm256_loadu_si256((const __m256i*)k_mask1);
const __m256i mask2 = _mm256_loadu_si256((const __m256i*)k_mask2);
const __m256i idx_mask = _mm256_set1_epi16(256);
const __m256i idx_shift = _mm256_set_epi16(1, 2, 3, 4, 5, 6, 7, 8, 1, 2, 3, 4, 5, 6, 7, 8);
const __m256i idx_mask = _mm256_set1_epi32(256);
const __m256i idx_shift = _mm256_set_epi32(1, 2, 3, 4, 5, 6, 7, 8);
const __m256i idx_mult = _mm256_set1_epi32(IQ3S_MULTIPLIER);
const __m256i m1 = _mm256_set1_epi32(0x01010101);
const __m256i m7 = _mm256_set1_epi32(0x07070707);
@ -10139,10 +10139,10 @@ void ggml_vec_dot_iq3_s_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const v
const __m256i q8_2 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
const __m128i idx_l_8 = _mm_loadu_si128((const __m128i*)qs); qs += 16;
const __m256i idx_l_16 = _mm256_cvtepu8_epi16(idx_l_8);
const __m256i idx_h_16 = _mm256_set_m128i(_mm_set1_epi16(qh[ib32+1]), _mm_set1_epi16(qh[ib32+0]));
const __m256i idx_16 = _mm256_or_si256(idx_l_16, _mm256_and_si256(shift_left_epi16(idx_h_16, idx_shift), idx_mask));
const __m256i idx_32_l = _mm256_cvtepi16_epi32(_mm256_castsi256_si128(idx_16));
const __m256i idx_32_h = _mm256_cvtepi16_epi32(_mm256_extractf128_si256(idx_16, 1));
const __m256i idx_h_l = _mm256_and_si256(_mm256_sllv_epi32(_mm256_set1_epi32(qh[ib32+0]), idx_shift), idx_mask);
const __m256i idx_h_h = _mm256_and_si256(_mm256_sllv_epi32(_mm256_set1_epi32(qh[ib32+1]), idx_shift), idx_mask);
const __m256i idx_32_l = _mm256_or_si256(idx_h_l, _mm256_cvtepi16_epi32(_mm256_castsi256_si128(idx_l_16)));
const __m256i idx_32_h = _mm256_or_si256(idx_h_h, _mm256_cvtepi16_epi32(_mm256_extractf128_si256(idx_l_16, 1)));
const __m256i idx_l = _mm256_add_epi32(_mm256_and_si256(_mm256_mullo_epi32(idx_mult, idx_32_l), m15), m1);
const __m256i q2_1 = _mm256_or_si256(_mm256_slli_epi32(_mm256_and_si256(_mm256_srli_epi32(idx_l, 1), m7), 1), m1);
const __m256i idx_h = _mm256_add_epi32(_mm256_and_si256(_mm256_mullo_epi32(idx_mult, idx_32_h), m15), m1);