Mirrors/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2024-12-27 22:59:24 +01:00
Code Issues Packages Projects Releases Wiki Activity

iq1_s: make Metal work with new version

Browse Source
This commit is contained in:
Iwan Kawrakow 2024-03-11 13:03:49 +01:00
parent 4fba3e00c6
commit da4528bce3
1 changed files with 7 additions and 6 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

13
ggml-metal.metal
View File

@ -4377,7 +4377,7 @@ void kernel_mul_mv_iq1_s_f32_impl(
+ yl[j+16] * (grid3[j] & 0xf) + yl[j+20] * (grid3[j] >> 4)
+ yl[j+24] * (grid4[j] & 0xf) + yl[j+28] * (grid4[j] >> 4);
}
sumf[row] += (float)dh[0] * (sum - sumy) * (2*(qh[0] >> 12) + 1);
sumf[row] += (float)dh[0] * (sum + sumy * (qh[0] & 0x8000 ? -1 - IQ1S_DELTA : -1 + IQ1S_DELTA)) * (2*((qh[0] >> 12) & 7) + 1);
dh += nb*sizeof(block_iq1_s)/2;
qs += nb*sizeof(block_iq1_s);
@ -5076,14 +5076,15 @@ void dequantize_iq1_s(device const block_iq1_s * xb, short il, thread type4x4 &
const float d = xb->d;
device const uint8_t * qs = xb->qs + 4*ib32 + 2*il;
device const uint16_t * qh = xb->qh;
const float dl = d * (2*(qh[ib32] >> 12) + 1);
const float dl = d * (2*((qh[ib32] >> 12) & 7) + 1);
const float ml = dl * (qh[ib32] & 0x8000 ? -1 - IQ1S_DELTA : -1 + IQ1S_DELTA);
constant uint8_t * grid1 = (constant uint8_t *)(iq1s_grid_gpu + (qs[0] | (((qh[ib32] >> (6*il+0)) & 7) << 8)));
constant uint8_t * grid2 = (constant uint8_t *)(iq1s_grid_gpu + (qs[1] | (((qh[ib32] >> (6*il+3)) & 7) << 8)));
for (int i = 0; i < 4; ++i) {
reg[0][i] = dl * (grid1[i] & 0xf) - dl;
reg[1][i] = dl * (grid1[i] >> 4) - dl;
reg[2][i] = dl * (grid2[i] & 0xf) - dl;
reg[3][i] = dl * (grid2[i] >> 4) - dl;
reg[0][i] = dl * (grid1[i] & 0xf) + ml;
reg[1][i] = dl * (grid1[i] >> 4) + ml;
reg[2][i] = dl * (grid2[i] & 0xf) + ml;
reg[3][i] = dl * (grid2[i] >> 4) + ml;
}
}