ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-01-26 17:20:01 +00:00
Code Issues Packages Projects Releases Wiki Activity

q6_0: this is slightly better

Browse Source
This commit is contained in:
Iwan Kawrakow
2024-10-02 18:07:55 +03:00
parent 9d1552a4fc
commit 38eb7fa499
1 changed files with 4 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
ggml/src/iqk/iqk_mul_mat.cpp
View File

@@ -3231,12 +3231,12 @@ struct Q5_1_Dequantizer {
struct Q6_0_1_Dequantizer {
Dequantizer4bit b4;
const __m256i mh = _mm256_set1_epi8(0x30);
const __m128i shift = _mm_set_epi64x(0, 4);
const __m256i shift1 = _mm256_set_epi64x(0, 2, 0, 4);
const __m256i shift2 = _mm256_set_epi64x(2, 0, 0, 0);
inline __m256i dequant(const block_q6_0 * x) const {
uint64_t aux64; std::memcpy(&aux64, x->qh, 8);
auto h128 = _mm_sllv_epi64(_mm_set1_epi64x(aux64), shift);
auto h256 = MM256_SET_M128I(_mm_srli_epi16(h128, 2), h128);
return _mm256_or_si256(b4.dequant(x->qs), _mm256_and_si256(h256, mh));
auto h256 = _mm256_sllv_epi64(_mm256_set1_epi64x(aux64), shift1);
return _mm256_or_si256(b4.dequant(x->qs), _mm256_and_si256(_mm256_srlv_epi64(h256, shift2), mh));
}
};