ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-02-28 09:04:10 +00:00
Code Issues Packages Projects Releases Wiki Activity

Use bperm trick for iq2_k_r4 gemv -> ~7% gain

Browse Source
This commit is contained in:
Iwan Kawrakow
2025-08-21 18:58:19 +03:00
parent 9cf9172afe
commit 01eee24f0f
1 changed files with 23 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

26
ggml/src/ggml-cuda/iqk_mmvq.cu
View File

@@ -1011,12 +1011,31 @@ __device__ __forceinline__ void vec_dot_iq2_k_r4_q8_1(
int is = ib16%2;
const int * scales_l = (const int *)bq2->scales;
const int * all_values = (const int *)iq2k_table;
int scales = __vsub4(((scales_l[2*(ib32%4)+is] >> 4*(ib32/4)) & 0x0f0f0f0f), 0x08080808);
const int8_t * s8 = (const int8_t *)&scales;
int2 val1;
const int * q2 = (const int *)bq2->qs + 8*ib32 + 4*is;
#ifdef __CUDA_ARCH__
#pragma unroll
for (int i = 0; i < 4; ++i) {
uint32_t extra32 = uint32_t((bq2->extra[i+4*is] >> ib32) & 1) * 0x04040404;
extra32 |= (extra32 << 4);
uint32_t val1 = ((q2[i] >> 0) & 0x33333333) | extra32;
uint32_t val2 = ((q2[i] >> 2) & 0x33333333) | extra32;
int2 v1 = get_int_from_table_8(val1, iq2nl_values);
int2 v2 = get_int_from_table_8(val2, iq2nl_values);
int sumi = 0;
sumi = ggml_cuda_dp4a(v1.x, q8[0], ggml_cuda_dp4a(v2.x, q8[1], sumi));
sumi = ggml_cuda_dp4a(v1.y, q8[2], ggml_cuda_dp4a(v2.y, q8[3], sumi));
const float d = __half2float(bq2->d[i]) * d8;
result[i] += d * sumi * s8[i];
}
#else
const int * all_values = (const int *)iq2k_table;
int2 val1;
int aux32[2];
#pragma unroll
for (int i = 0; i < 4; ++i) {
@@ -1035,6 +1054,7 @@ __device__ __forceinline__ void vec_dot_iq2_k_r4_q8_1(
const float d = __half2float(bq2->d[i]) * d8;
result[i] += d * sumi1 * s8[i];
}
#endif
}
#define VDR_IQ3_K_Q8_1_MMVQ 4