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

256b version, also slow. i tried :)

Browse source
This commit is contained in:
Eve 2024-11-02 16:36:56 -04:00
parent f8dd133ce4
commit 1335c78639
1 changed files with 3 additions and 3 deletions
Download patch file Download diff file Expand all files Collapse all files

6
ggml/src/ggml-quants.c
View file

@ -4237,9 +4237,9 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, size_t bs, const void * r
p_2 = _mm_add_epi32(_mm_cvtepi16_epi32(_mm_bsrli_si128(p_2, 8)), _mm_cvtepi16_epi32(p_2));
// TODO: if f16c
const __m128 del = _mm_cvtph_ps(_mm_set_epi16(0, 0, 0 , 0, x[ib + 1].d, y[ib + 1].d, x[ib].d, y[ib].d));
const __m128 del_mul = _mm_mul_ps(del, _mm_permute_ps(del, 0xB1)); // x.d*y.d
const __m256 deltas = _mm256_set_m128(_mm_permute_ps(del_mul, 0xFF), _mm_permute_ps(del_mul, 0x00));
const __m256 del = _mm256_cvtph_ps(_mm_set_epi16(0 , 0, x[ib + 1].d, y[ib + 1].d, 0, 0, x[ib].d, y[ib].d));
const __m256 del_mul = _mm256_mul_ps(del, _mm256_permute_ps(del, 0x01)); // x.d*y.d
const __m256 deltas = _mm256_permute_ps(del_mul, 0x00);
// TODO: may be room to optimize here?
//const __m256 deltas = _mm256_set_m128(_mm_set1_ps(GGML_FP16_TO_FP32(x[ib + 1].d) * GGML_FP16_TO_FP32(y[ib + 1].d)),