ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-04-26 09:29:27 +00:00
Code Issues Packages Projects Releases Wiki Activity

iq4_ks_r4: slightly better Zen4

Browse Source
This commit is contained in:
Iwan Kawrakow
2024-12-18 19:03:16 +02:00
parent 00452d4683
commit 6b1a1b9880
1 changed files with 10 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files

16
ggml/src/iqk/iqk_mul_mat.cpp
View File

@@ -3117,7 +3117,13 @@ static void mul_mat_iq4_ks_r4_q8_k(int n, const void * vx, size_t bx, const Data
#endif
int nbl = n / QK_K;
using helper_t = union { __m256i vec; uint32_t val[8]; };
#ifndef HAVE_FANCY_SIMD
helper_t h, h_shift;
#else
using helper512_t = union { __m512i vec; uint64_t val[8]; };
helper_t h;
helper512_t h_shift;
#endif
__m256 acc[nrc_y] = {};
__m256i isum[nrc_y] = {};
__m256i qx[4];
@@ -3148,16 +3154,14 @@ static void mul_mat_iq4_ks_r4_q8_k(int n, const void * vx, size_t bx, const Data
}
}
#else
//h_shift.vec = _mm256_slli_epi16(_mm256_and_si256(scales, _mm256_set1_epi8(1)), 1);
h_shift.vec = _mm256_add_epi8(_mm256_set1_epi8(-64), _mm256_slli_epi16(_mm256_and_si256(scales, _mm256_set1_epi8(1)), 1));
auto shift = _mm256_add_epi8(_mm256_set1_epi8(-64), _mm256_slli_epi16(_mm256_and_si256(scales, _mm256_set1_epi8(1)), 1));
h_shift.vec = _mm512_mullo_epi16(_mm512_cvtepi8_epi16(shift), _mm512_cvtepi8_epi16(h.vec));
#endif
for (int ib = 0; ib < QK_K/32; ++ib) {
#ifdef HAVE_FANCY_SIMD
auto iscales = _mm256_cvtepi8_epi32(_mm_set1_epi32(h.val[ib]));
//auto ishifts = _mm256_add_epi32(_mm256_set1_epi32(-64), _mm256_cvtepi8_epi32(_mm_set1_epi32(h_shift.val[ib])));
auto ishifts = _mm256_cvtepi8_epi32(_mm_set1_epi32(h_shift.val[ib]));
auto scales = _mm256_cvtepi32_ps(iscales);
auto scales_m = _mm256_mul_ps(scales, _mm256_cvtepi32_ps(ishifts));
auto ishifts = _mm256_cvtepi16_epi32(_mm_set1_epi64x(h_shift.val[ib]));
auto scales_m = _mm256_cvtepi32_ps(ishifts);
for (int iy = 0; iy < nrc_y; ++iy) {
float m8 = ((const float *)q8.y[iy][ibl].bsums)[ib];
acc[iy] = _mm256_fmadd_ps(scales_m, _mm256_set1_ps(m8), acc[iy]);