ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-02-25 07:34:10 +00:00
Code Issues Packages Projects Releases Wiki Activity

Slightly faster AVX2 implementation for q4_k_r4

Browse Source
This commit is contained in:
Iwan Kawrakow
2025-01-29 15:32:46 +02:00
parent 4a73c25002
commit 118baf3f73
1 changed files with 16 additions and 8 deletions
Download Patch File Download Diff File Expand all files Collapse all files

24
ggml/src/iqk/iqk_mul_mat.cpp
View File

@@ -4456,14 +4456,22 @@ static void mul_mat_q4_k_r4_q8_k_avx2(int n, const void * vx, size_t bx, const D
auto hbits = MM256_SET_M128I(hbits128, _mm_slli_epi16(hbits128, 4));
hd.vec = _mm256_or_si256(_mm256_and_si256(lbits, mf), _mm256_and_si256(hbits, m3));
auto mins = _mm256_or_si256(_mm256_and_si256(_mm256_srli_epi16(lbits, 4), mf), _mm256_and_si256(_mm256_srli_epi16(hbits, 2), m3));
auto shuffle = _mm256_set1_epi64x(0x0000000400000000);
auto c1 = _mm256_mul_ps(m4, _mm256_cvtepi32_ps(_mm256_cvtepi8_epi32(_mm256_castsi256_si128(_mm256_permutevar8x32_epi32(mins, shuffle)))));
shuffle = _mm256_add_epi32(shuffle, _mm256_set1_epi32(1));
auto c2 = _mm256_mul_ps(m4, _mm256_cvtepi32_ps(_mm256_cvtepi8_epi32(_mm256_castsi256_si128(_mm256_permutevar8x32_epi32(mins, shuffle)))));
shuffle = _mm256_add_epi32(shuffle, _mm256_set1_epi32(1));
auto c3 = _mm256_mul_ps(m4, _mm256_cvtepi32_ps(_mm256_cvtepi8_epi32(_mm256_castsi256_si128(_mm256_permutevar8x32_epi32(mins, shuffle)))));
shuffle = _mm256_add_epi32(shuffle, _mm256_set1_epi32(1));
auto c4 = _mm256_mul_ps(m4, _mm256_cvtepi32_ps(_mm256_cvtepi8_epi32(_mm256_castsi256_si128(_mm256_permutevar8x32_epi32(mins, shuffle)))));
auto mins_l = _mm256_castsi256_si128(mins);
auto mins_h = _mm256_extracti128_si256(mins, 1);
auto aux1 = _mm_unpacklo_epi32(mins_l, mins_h);
auto aux2 = _mm_unpackhi_epi32(mins_l, mins_h);
auto c1 = _mm256_mul_ps(m4, _mm256_cvtepi32_ps(_mm256_cvtepi8_epi32(aux1)));
auto c2 = _mm256_mul_ps(m4, _mm256_cvtepi32_ps(_mm256_cvtepi8_epi32(_mm_shuffle_epi32(aux1, 0xee))));
auto c3 = _mm256_mul_ps(m4, _mm256_cvtepi32_ps(_mm256_cvtepi8_epi32(aux2)));
auto c4 = _mm256_mul_ps(m4, _mm256_cvtepi32_ps(_mm256_cvtepi8_epi32(_mm_shuffle_epi32(aux2, 0xee))));
//auto shuffle = _mm256_set1_epi64x(0x0000000400000000);
//auto c1 = _mm256_mul_ps(m4, _mm256_cvtepi32_ps(_mm256_cvtepi8_epi32(_mm256_castsi256_si128(_mm256_permutevar8x32_epi32(mins, shuffle)))));
//shuffle = _mm256_add_epi32(shuffle, _mm256_set1_epi32(1));
//auto c2 = _mm256_mul_ps(m4, _mm256_cvtepi32_ps(_mm256_cvtepi8_epi32(_mm256_castsi256_si128(_mm256_permutevar8x32_epi32(mins, shuffle)))));
//shuffle = _mm256_add_epi32(shuffle, _mm256_set1_epi32(1));
//auto c3 = _mm256_mul_ps(m4, _mm256_cvtepi32_ps(_mm256_cvtepi8_epi32(_mm256_castsi256_si128(_mm256_permutevar8x32_epi32(mins, shuffle)))));
//shuffle = _mm256_add_epi32(shuffle, _mm256_set1_epi32(1));
//auto c4 = _mm256_mul_ps(m4, _mm256_cvtepi32_ps(_mm256_cvtepi8_epi32(_mm256_castsi256_si128(_mm256_permutevar8x32_epi32(mins, shuffle)))));
for (int iy = 0; iy < nrc_y; ++iy) {
auto bs = _mm256_loadu_ps((const float *)q8.y[iy][ibl].bsums);
acc[iy] = _mm256_fmadd_ps(c1, _mm256_shuffle_ps(bs, bs, 0x00), acc[iy]);