ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-05-02 20:31:45 +00:00
Code Issues Packages Projects Releases Wiki Activity

q8_0_r4: Zen4 matrix-vector specialization

Browse Source
This commit is contained in:
Iwan Kawrakow
2024-12-03 06:52:12 +02:00
parent fc781beec0
commit 4559dc5931
1 changed files with 79 additions and 44 deletions
Download Patch File Download Diff File Expand all files Collapse all files

35
ggml/src/iqk/iqk_mul_mat.cpp
View File

@@ -2290,6 +2290,40 @@ static void mul_mat_q8_0_r4_q8_1(int n, const void * vx, size_t bx, const DataIn
Q8<nrc_y, block_q8_1_x4> q8(info);
int nb = n / QK8_0;
GGML_ASSERT(nb%4 == 0);
if constexpr (nrc_y == 1) {
auto m127 = _mm256_set1_epi8(127);
auto m1 = _mm256_set1_epi16(1);
__m256 acc[nrc_y] = {};
for (int ix = 0; ix < nrc_x; ix += 4) {
const block_q8_0_x4 * iq8 = (const block_q8_0_x4 *)((const char *)vx + ix*bx);
for (int ib4 = 0; ib4 < nb/4; ++ib4) {
for (int k = 0; k < 4; ++k) {
auto scales128 = _mm_cvtph_ps(_mm_loadl_epi64((const __m128i *)iq8[4*ib4+k].d));
auto scales = _mm256_set_m128(scales128, scales128);
auto scales_m = _mm256_mul_ps(scales, _mm256_set1_ps(-63.5f));
auto q1 = _mm256_add_epi8(_mm256_loadu_si256((const __m256i *)iq8[4*ib4+k].qs+0), m127);
auto q2 = _mm256_add_epi8(_mm256_loadu_si256((const __m256i *)iq8[4*ib4+k].qs+1), m127);
auto q3 = _mm256_add_epi8(_mm256_loadu_si256((const __m256i *)iq8[4*ib4+k].qs+2), m127);
auto q4 = _mm256_add_epi8(_mm256_loadu_si256((const __m256i *)iq8[4*ib4+k].qs+3), m127);
for (int iy = 0; iy < nrc_y; ++iy) {
auto y = _mm256_loadu_si256((const __m256i*)q8.y[iy][ib4].qs+k);
auto sumi1 = _mm256_add_epi32(_mm256_madd_epi16(m1, _mm256_maddubs_epi16(q1, _mm256_shuffle_epi32(y, 0x00))),
_mm256_madd_epi16(m1, _mm256_maddubs_epi16(q2, _mm256_shuffle_epi32(y, 0x55))));
auto sumi2 = _mm256_add_epi32(_mm256_madd_epi16(m1, _mm256_maddubs_epi16(q3, _mm256_shuffle_epi32(y, 0xaa))),
_mm256_madd_epi16(m1, _mm256_maddubs_epi16(q4, _mm256_shuffle_epi32(y, 0xff))));
auto d4d8 = _mm256_mul_ps(scales, _mm256_set1_ps(GGML_FP16_TO_FP32(q8.y[iy][ib4].d[k])));
acc[iy] = _mm256_fmadd_ps(d4d8, _mm256_cvtepi32_ps(_mm256_add_epi32(sumi1, sumi2)), acc[iy]);
acc[iy] = _mm256_fmadd_ps(scales_m, _mm256_set1_ps(GGML_FP16_TO_FP32(q8.y[iy][ib4].d[k+4])), acc[iy]);
}
}
}
for (int iy = 0; iy < nrc_y; ++iy) {
auto sum = _mm_add_ps(_mm256_castps256_ps128(acc[iy]), _mm256_extractf128_ps(acc[iy], 1));
info.store(ix, iy, sum);
acc[iy] = _mm256_setzero_ps();
}
}
} else {
__m512 acc[2*nrc_y] = {};
__m512i qx[4];
auto m127 = _mm512_set1_epi8(127);
@@ -2340,6 +2374,7 @@ static void mul_mat_q8_0_r4_q8_1(int n, const void * vx, size_t bx, const DataIn
}
}
}
}
#else
template <int nrc_y>
static void mul_mat_q8_0_r4_q8_1(int n, const void * vx, size_t bx, const DataInfo& info, int nrc_x) {