ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-04-29 10:51:51 +00:00
Code Issues Packages Projects Releases Wiki Activity

q8_k_r16: iq2_s now uses q8_k_r16 on Zen4+

Browse Source
This commit is contained in:
Iwan Kawrakow
2025-08-20 11:43:33 +03:00
parent f11612ba59
commit 669f774fb1
1 changed files with 20 additions and 7 deletions
Download Patch File Download Diff File Expand all files Collapse all files

27
ggml/src/iqk/iqk_gemm_iquants.cpp
View File

@@ -2279,14 +2279,21 @@ void iqk_convert_iq2_s_q8_0_r8(int n, const void * vx, size_t bx, void * vy, int
} }
void iqk_convert_iq2_s_q8_k_r8(int n, const void * vx, size_t bx, void * vy, int nrc_x) { void iqk_convert_iq2_s_q8_k_r8(int n, const void * vx, size_t bx, void * vy, int nrc_x) {
#ifdef HAVE_FANCY_SIMD
constexpr int k_nr = 16;
using block_q8_k_r = block_q8_k_r16;
#else
constexpr int k_nr = 8;
using block_q8_k_r = block_q8_k_r8;
#endif
GGML_ASSERT(n%QK_K == 0); GGML_ASSERT(n%QK_K == 0);
GGML_ASSERT(nrc_x%8 == 0); GGML_ASSERT(nrc_x%k_nr == 0);
int nb = n/QK_K; int nb = n/QK_K;
const block_iq2_s * x8[8]; const block_iq2_s * x8[k_nr];
block_q8_k_r8 * y = (block_q8_k_r8 *)vy; block_q8_k_r * y = (block_q8_k_r *)vy;
uint32_t block[8]; uint32_t block[8];
@@ -2295,17 +2302,23 @@ void iqk_convert_iq2_s_q8_k_r8(int n, const void * vx, size_t bx, void * vy, int
SignHelper sh; SignHelper sh;
for (int ix = 0; ix < nrc_x; ix += 8) { for (int ix = 0; ix < nrc_x; ix += k_nr) {
for (int k = 0; k < 8; ++k) x8[k] = (const block_iq2_s *)((const char *)vx + (ix + k)*bx); for (int k = 0; k < k_nr; ++k) x8[k] = (const block_iq2_s *)((const char *)vx + (ix + k)*bx);
for (int i = 0; i < nb; ++i) { for (int i = 0; i < nb; ++i) {
for (int k = 0; k < 8; ++k) { for (int k = 0; k < k_nr; ++k) {
float d = 0.125f * GGML_FP16_TO_FP32(x8[k][i].d); float d = 0.125f * GGML_FP16_TO_FP32(x8[k][i].d);
helper.vec = DequantizerIQ2S::make_scales(x8[k][i].scales); helper.vec = DequantizerIQ2S::make_scales(x8[k][i].scales);
DequantizerIQ2S::prepare(x8[k][i].qs+ 0, x8[k][i].qh+0, (const uint16_t *)(x8[k][i].qs + QK_K/8) + 0, sh, qx+0); DequantizerIQ2S::prepare(x8[k][i].qs+ 0, x8[k][i].qh+0, (const uint16_t *)(x8[k][i].qs + QK_K/8) + 0, sh, qx+0);
DequantizerIQ2S::prepare(x8[k][i].qs+16, x8[k][i].qh+4, (const uint16_t *)(x8[k][i].qs + QK_K/8) + 8, sh, qx+4); DequantizerIQ2S::prepare(x8[k][i].qs+16, x8[k][i].qh+4, (const uint16_t *)(x8[k][i].qs + QK_K/8) + 8, sh, qx+4);
float dnew = convert_to_q8_k_r8(k, 1.f/124, qx, helper.val, block, y[i].qs); float dnew = convert_to_q8_k_r8<k_nr>(k, 1.f/124, qx, helper.val, block, y[i].qs);
y[i].d[k] = GGML_FP32_TO_FP16(d*dnew); y[i].d[k] = GGML_FP32_TO_FP16(d*dnew);
} }
#ifdef HAVE_FANCY_SIMD
for (int l = 0; l < 64; ++l) {
auto v = _mm512_xor_si512(_mm512_loadu_si512((const __m512i *)y[i].qs + l), _mm512_set1_epi8(-128));
_mm512_storeu_si512((__m512i *)y[i].qs + l, v);
}
#endif
} }
y += nb; y += nb;
} }