ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-02-23 22:54:10 +00:00
Code Issues Packages Projects Releases Wiki Activity

Use bperm trick for iq3_k gemv -> 4.5% gain

Browse Source
This commit is contained in:
Iwan Kawrakow
2025-08-21 15:49:00 +03:00
parent 91b20d4bec
commit 90379f3d51
1 changed files with 14 additions and 27 deletions
Download Patch File Download Diff File Expand all files Collapse all files

41
ggml/src/ggml-cuda/iqk_mmvq.cu
View File

@@ -1111,50 +1111,37 @@ __device__ __forceinline__ void vec_dot_iq3_ks_q8_1(
const uint16_t * ql = (const uint16_t *)bq3->qs + 16*ib128 + 4*il8;
const uint16_t * qh = (const uint16_t *)bq3->qh + 4*il8;
int32_t aux32;
const uint8_t * aux8 = (const uint8_t *)&aux32;
uint16_t extra = bq3->extra >> 4*ib128;
uint16_t extra_v = extra >> 8;
uint32_t extra_v = uint32_t(extra >> 8) * 0x01010101;
const uint16_t * values1 = iq3k_table + ((extra_v << 6) & 0x40);
const uint16_t * values2 = iq3k_table + ((extra_v << 5) & 0x40);
const uint16_t * values3 = iq3k_table + ((extra_v << 4) & 0x40);
const uint16_t * values4 = iq3k_table + ((extra_v << 3) & 0x40);
uint32_t extra32_1 = ((extra_v << 3) & 0x08080808) | ((extra_v << 5) & 0x80808080);
uint32_t extra32_2 = ((extra_v << 2) & 0x08080808) | ((extra_v << 4) & 0x80808080);
const int * q8;
int sumi[4] = {0, 0, 0, 0};
int v;
for (int i = 0; i < 2; ++i) {
uint32_t vl = ql[2*i+0] | (ql[2*i+1] << 16);
uint32_t vh = ((qh[2*i+0] | (qh[2*i+1] << 16)) >> 4*ib128) << 2;
uint32_t vh = ((qh[2*i+0] | (qh[2*i+1] << 16)) >> 4*ib128);
uint32_t val1 = ((vl >> 0) & 0x33333333) | extra32_1 | ((vh << 2) & 0x04040404) | ((vh << 4) & 0x40404040);
uint32_t val2 = ((vl >> 2) & 0x33333333) | extra32_2 | ((vh << 1) & 0x04040404) | ((vh << 3) & 0x40404040);
int2 v1 = get_int_from_table_16(val1, iq3nl_values);
int2 v2 = get_int_from_table_16(val2, iq3nl_values);
q8 = (const int *)bq8_1[4*ib128+0].qs + 2*il8;
aux32 = (vl & 0x03030303) | (vh & 0x04040404);
v = int_from_table_2(aux8, values1);
sumi[0] = ggml_cuda_dp4a(v, q8[i], sumi[0]);
vl >>= 2; vh >>= 1;
sumi[0] = ggml_cuda_dp4a(v1.x, q8[i], sumi[0]);
q8 += sizeof(block_q8_1)/4;
aux32 = (vl & 0x03030303) | (vh & 0x04040404);
v = int_from_table_2(aux8, values2);
sumi[1] = ggml_cuda_dp4a(v, q8[i], sumi[1]);
vl >>= 2; vh >>= 1;
sumi[1] = ggml_cuda_dp4a(v2.x, q8[i], sumi[1]);
q8 += sizeof(block_q8_1)/4;
aux32 = (vl & 0x03030303) | (vh & 0x04040404);
v = int_from_table_2(aux8, values3);
sumi[2] = ggml_cuda_dp4a(v, q8[i], sumi[2]);
vl >>= 2; vh >>= 1;
sumi[2] = ggml_cuda_dp4a(v1.y, q8[i], sumi[2]);
q8 += sizeof(block_q8_1)/4;
aux32 = (vl & 0x03030303) | (vh & 0x04040404);
v = int_from_table_2(aux8, values4);
sumi[3] = ggml_cuda_dp4a(v, q8[i], sumi[3]);
sumi[3] = ggml_cuda_dp4a(v2.y, q8[i], sumi[3]);
}
const uint16_t * sl16 = (const uint16_t *)bq3->scales;
aux32 = __vsub4(((sl16[0] | (sl16[1] << 16)) >> 4*ib128) & 0x0f0f0f0f, 0x10101010);
int32_t aux32 = __vsub4(((sl16[0] | (sl16[1] << 16)) >> 4*ib128) & 0x0f0f0f0f, 0x10101010);
const int8_t * a8 = (const int8_t *)&aux32;
*result += d * (__low2float(bq8_1[4*ib128+0].ds) * (a8[0] + ((extra << 4) & 0x10)) * sumi[0] +
__low2float(bq8_1[4*ib128+1].ds) * (a8[1] + ((extra << 3) & 0x10)) * sumi[1] +