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

q4_0c: avoid _mm512_loadu_epi64 instruction

Browse source 
Not supported on some GCC versions
This commit is contained in:
Håkon H. Hitland 2023-05-04 09:53:55 +02:00
parent d53f76760d
commit 76692c90cd
1 changed files with 3 additions and 3 deletions
Download patch file Download diff file Expand all files Collapse all files

6
ggml.c
View file

@ -2855,13 +2855,13 @@ static inline __m512 dot_q4_0c_fourblocks_avx512(
) { ) {
// load quantized bytes // load quantized bytes
// TODO: change back to aligned loads // TODO: change back to aligned loads
const __m512i xqs0123 = _mm512_loadu_epi64( xqs ); const __m512i xqs0123 = _mm512_loadu_si512( xqs );
const __m512i low_nibble_mask = _mm512_set1_epi8( 0xf ); const __m512i low_nibble_mask = _mm512_set1_epi8( 0xf );
const __m512i xqs01 = _mm512_and_si512( low_nibble_mask, xqs0123 ); const __m512i xqs01 = _mm512_and_si512( low_nibble_mask, xqs0123 );
// TODO: try srlv/i? // TODO: try srlv/i?
const __m512i xqs23 = _mm512_and_si512( low_nibble_mask, _mm512_srli_epi32( xqs0123, 4 ) ); const __m512i xqs23 = _mm512_and_si512( low_nibble_mask, _mm512_srli_epi32( xqs0123, 4 ) );
const __m512i yqs01 = _mm512_loadu_epi64( yqs ); const __m512i yqs01 = _mm512_loadu_si512( yqs );
const __m512i yqs23 = _mm512_loadu_epi64( yqs + 2*QK8_0C ); const __m512i yqs23 = _mm512_loadu_si512( yqs + 2*QK8_0C );
// load scales // load scales
const __m512i scale_mask0 = _mm512_set_epi32(1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0); const __m512i scale_mask0 = _mm512_set_epi32(1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0);