ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-02-21 21:54:10 +00:00
Code Issues Packages Projects Releases Wiki Activity

bitnet: remove iq1_bn lookup table storing +/- signs

Browse Source 
The AVX2 implementation was the only one left using it, so
I decided to see if we can get a performant implementation
using the 0,1,2 lookup table. Turns out we can, and it is
even slightly faster than the sign based table. We now
get PP-512 = 275 t/s and TG-128 = 57.7 t/s with 16 threads
on the Ryzen-7950X.

With only one lookup table left for iq1_bn, I renamed it to
iq1bn_grid_u16.
This commit is contained in:
Iwan Kawrakow
2024-06-25 18:19:11 +03:00
parent 8b436a84c5
commit 753dbaeeb0
6 changed files with 33 additions and 239 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
iqk-quantize.cpp
View File

@@ -190,7 +190,7 @@ void dequantize_row_iq1_bn(const block_iq1_bn * x, float * y, int64_t k) {
auto ql = x[i].ql;
for (int k = 0; k < QK_IQ1BN/8; ++k) {
uint16_t idx = ql[k] | ((qh[k/2] << (8 - 4*(k%2))) & 0x0f00);
uint16_t val = extra & 1 ? 0xaaaa - iq1bn_grid_zzz[idx] : iq1bn_grid_zzz[idx];
uint16_t val = extra & 1 ? 0xaaaa - iq1bn_grid_u16[idx] : iq1bn_grid_u16[idx];
aux32[0] = val | (val << 14);
aux32[1] = (aux32[0] >> 4) & 0x03030303;
aux32[0] &= 0x03030303;
@@ -270,8 +270,8 @@ void ggml_vec_dot_iq1_bn_q8_K64(int n, float * s, size_t bs, const void * vx, si
for (int j = 0; j < QK_IQ1BN/16; ++j) {
uint16_t idx1 = ql[2*j+0] | ((qh[j] << 8) & 0x0f00);
uint16_t idx2 = ql[2*j+1] | ((qh[j] << 4) & 0x0f00);
uint16_t val1 = extra & 1 ? k_magic - iq1bn_grid_zzz[idx1] : iq1bn_grid_zzz[idx1];
uint16_t val2 = extra & 2 ? k_magic - iq1bn_grid_zzz[idx2] : iq1bn_grid_zzz[idx2];
uint16_t val1 = extra & 1 ? k_magic - iq1bn_grid_u16[idx1] : iq1bn_grid_u16[idx1];
uint16_t val2 = extra & 2 ? k_magic - iq1bn_grid_u16[idx2] : iq1bn_grid_u16[idx2];
extra >>= 2;
aux32[0] = val1 | (val1 << 14);
aux32[1] = (aux32[0] >> 4) & 0x03030303;