ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-02-24 07:04:11 +00:00
Code Issues Packages Projects Releases Wiki Activity

iq1_bn(NEON): works now, but very slow

Browse Source 
Basically 2X slower tan q4_0.
This commit is contained in:
Iwan Kawrakow
2024-06-17 13:04:24 +02:00
parent 2f403d4c93
commit 8222c9f3d1
1 changed files with 1 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
iqk_mul_mat.cpp
View File

@@ -4088,7 +4088,7 @@ static void mul_mat_iq1bn_q8_K64(int n, const void * vx, size_t bx, const DataIn
int32x4_t sumi = vdupq_n_s32(0);
for (int j = 0; j < 4; ++j) {
auto tmp = vmulq_s8(q.val[j], vreinterpretq_s8_u8(signs.val[j]));
tmp = vmulq_s8(q.val[j], v.val[j]);
tmp = vmulq_s8(tmp, v.val[j]);
sumi = ggml_vdotq_s32(sumi, m1, tmp);
}
accd[iy] = vfmaq_f32(accd[iy], vdupq_n_f32(q8.scale(iy, i)), vcvtq_f32_s32(sumi));