ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-03-03 10:30:27 +00:00
Code Issues Packages Projects Releases Wiki Activity

q8_k_r16: iq3_xxs now uses q8_k_r16 on Zen4+

Browse Source
This commit is contained in:
Iwan Kawrakow
2025-08-20 11:49:35 +03:00
parent 669f774fb1
commit bdb266baf1
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

@@ -2444,14 +2444,21 @@ static void mul_mat_iq2_s_q8_2_X4(int n, const void * vx, size_t bx, const DataI
}
void iqk_convert_iq3_xxs_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(nrc_x%8 == 0);
GGML_ASSERT(nrc_x%k_nr == 0);
int nb = n/QK_K;
const block_iq3_xxs * x8[8];
const block_iq3_xxs * x8[k_nr];
block_q8_k_r8 * y = (block_q8_k_r8 *)vy;
block_q8_k_r * y = (block_q8_k_r *)vy;
int16_t ls[16];
EvenSignHelper esh;
@@ -2460,10 +2467,10 @@ void iqk_convert_iq3_xxs_q8_k_r8(int n, const void * vx, size_t bx, void * vy, i
uint32_t block[8];
uint32_t aux32;
for (int ix = 0; ix < nrc_x; ix += 8) {
for (int k = 0; k < 8; ++k) x8[k] = (const block_iq3_xxs *)((const char *)vx + (ix + k)*bx);
for (int ix = 0; ix < nrc_x; ix += k_nr) {
for (int k = 0; k < k_nr; ++k) x8[k] = (const block_iq3_xxs *)((const char *)vx + (ix + k)*bx);
for (int i = 0; i < nb; ++i) {
for (int k = 0; k < 8; ++k) {
for (int k = 0; k < k_nr; ++k) {
float d = 0.25f * GGML_FP16_TO_FP32(x8[k][i].d);
auto qs = x8[k][i].qs;
auto sas = qs + QK_K/4;
@@ -2476,9 +2483,15 @@ void iqk_convert_iq3_xxs_q8_k_r8(int n, const void * vx, size_t bx, void * vy, i
esh.sign_value(aux32, values[ib32]);
qs += 8;
}
float dnew = convert_to_q8_k_r8(k, 1.f/124, values, ls, block, y[i].qs);
float dnew = convert_to_q8_k_r8<k_nr>(k, 1.f/124, values, ls, block, y[i].qs);
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;
}