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

metal : block_q_n_dot_y for q5_1 (broken)

Browse source
This commit is contained in:
Jhen 2023-10-17 13:45:08 +08:00
parent fce44a7628
commit 79d4732c70
1 changed files with 12 additions and 1 deletions
Download patch file Download diff file Expand all files Collapse all files

13
ggml-metal.metal
View file

@ -466,7 +466,18 @@ inline float block_q_n_dot_y(device const block_q5_0 * qb_curr, float sumy, thre
// we assume that the yl's have been multiplied with the appropriate scale factor
// that corresponds to the missing bit shifts (1, 1/16, 1/256, 1/4096)
inline float block_q_n_dot_y(device const block_q5_1 * qb_curr, float sumy, thread float * yl, int il) {
// TODO
float d = qb_curr->d;
float m = qb_curr->m;
float2 acc = 0.f;
device const uint16_t * qs = ((device const uint16_t *)qb_curr + 4 + il/2);
const uint32_t qh = *((device const uint32_t *)qb_curr->qh);
for (int i = 0; i < 8; i+=2) {
acc[0] += yl[i + 0] * ((qs[i / 2] & 0x000F) | ((qh >> (i+0+il ) << 4 ) & 0x0010))
+ yl[i + 1] * ((qs[i / 2] & 0x0F00) | ((qh >> (i+1+il ) << 12) & 0x1000));
acc[1] += yl[i + 8] * ((qs[i / 2] & 0x00F0) | ((qh >> (i+0+il+QK5_1/2) << 4 ) & 0x0010))
+ yl[i + 9] * ((qs[i / 2] & 0xF000) | ((qh >> (i+1+il+QK5_1/2) << 12) & 0x1000));
}
return d * (acc[0] + acc[1]) + sumy * m;
}
// putting them in the kernel cause a significant performance penalty