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

A better packNibbles implementation using AVX512

Browse source
This commit is contained in:
MeouSker77 2023-04-22 17:37:03 +08:00
parent 857308d1e8
commit bde28f2056
1 changed files with 6 additions and 0 deletions
Download patch file Download diff file Expand all files Collapse all files

6
ggml.c
View file

@ -517,6 +517,11 @@ static inline __m256 mul_sum_i8_pairs_float(const __m256i x, const __m256i y) {
static inline __m128i packNibbles( __m256i bytes )
{
// Move bits within 16-bit lanes from 0000_abcd_0000_efgh into 0000_0000_abcd_efgh
#if __AVX512F__
const __m256i bytes_srli_4 = _mm256_srli_epi16(bytes, 4); // 0000_0000_abcd_0000
bytes = _mm256_or_si256(bytes, bytes_srli_4); // 0000_abcd_abcd_efgh
return _mm256_cvtepi16_epi8(bytes); // abcd_efgh
#else
const __m256i lowByte = _mm256_set1_epi16( 0xFF );
__m256i high = _mm256_andnot_si256( lowByte, bytes );
__m256i low = _mm256_and_si256( lowByte, bytes );
@ -527,6 +532,7 @@ static inline __m128i packNibbles( __m256i bytes )
__m128i r0 = _mm256_castsi256_si128( bytes );
__m128i r1 = _mm256_extracti128_si256( bytes, 1 );
return _mm_packus_epi16( r0, r1 );
#endif
}
#else
static inline __m128i packNibbles( __m128i bytes1, __m128i bytes2 )