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

ggml : fix AVX paths for Q8_0 quantization

Browse source
This commit is contained in:
Georgi Gerganov 2023-04-22 10:49:17 +03:00
parent 76b6b267e6
commit 2c358eca92
1 changed files with 4 additions and 2 deletions
Download patch file Download diff file Expand all files Collapse all files

6
ggml.c
View file

@ -1428,7 +1428,9 @@ static void quantize_row_q8_0(const float * restrict x, void * restrict vy, int
#if defined(__AVX2__)
// Compute the sum of the quants and set y[i].s
y[i].s = d * hsum_i32_8(_mm256_add_epi32(_mm256_add_epi32(i0, i1), _mm256_add_epi32(i2, i3)));
//y[i].s = d * hsum_i32_8(_mm256_add_epi32(_mm256_add_epi32(i0, i1), _mm256_add_epi32(i2, i3)));
y[i].s0 = d * hsum_i32_8(_mm256_add_epi32(i0, i1));
y[i].s1 = d * hsum_i32_8(_mm256_add_epi32(i2, i3));
// Convert int32 to int16
i0 = _mm256_packs_epi32( i0, i1 ); // 0, 1, 2, 3, 8, 9, 10, 11, 4, 5, 6, 7, 12, 13, 14, 15
@ -2657,7 +2659,7 @@ static void ggml_vec_dot_q4_1_q8_0(const int n, float * restrict s, const void *
const float * d0 = &x[i].d;
const float * d1 = &y[i].d;
summs += x[i].m * y[i].s;
summs += x[i].m * (y[i].s0 + y[i].s1);
const __m256 d0v = _mm256_broadcast_ss( d0 );
const __m256 d1v = _mm256_broadcast_ss( d1 );