root/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2025-01-12 19:50:17 +00:00
Code Issues Actions 5 Packages Projects Releases Wiki Activity

Regroup q4_1 dot addition for better numerics.

Browse Source
This commit is contained in:
Matvey Soloviev 2023-03-23 04:56:21 +01:00
parent 580991bbed
commit 4aeee216fd
1 changed files with 3 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

5
ggml.c
View File

@ -1848,12 +1848,13 @@ inline static void ggml_vec_dot_q4_1(const int n, float * restrict s, const void
// Apply the scales, and accumulate
// acc += d0*m1*x + d1*m0*y
acc = _mm256_fmadd_ps( cross_scales, sums, acc );
__m256 delta = _mm256_mul_ps( cross_scales, sums );
// Convert int32_t to float
__m256 p = _mm256_cvtepi32_ps( i32 );
// acc += d0*d1*x*y
acc = _mm256_fmadd_ps( scale_01, p, acc );
delta = _mm256_fmadd_ps( scale_01, p, delta );
acc = _mm256_add_ps( acc, delta );
// acc_offset += m0*m1 (avoid reloading from RAM)
acc_offset = _mm_fmadd_ss( _mm256_castps256_ps128( m0v ), _mm256_castps256_ps128( m1v ), acc_offset );