ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-02-25 07:34:10 +00:00
Code Issues Packages Projects Releases Wiki Activity

q6_k_r4: slightly faster Zen4

Browse Source 
238.3 t/s -> 243.2 t/s
This commit is contained in:
Iwan Kawrakow
2024-12-10 12:43:20 +02:00
parent 8de44659ca
commit 408178aa8a
1 changed files with 19 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

22
ggml/src/iqk/iqk_mul_mat.cpp
View File

@@ -3259,7 +3259,9 @@ static void mul_mat_q6_k_r4_q8_k(int n, const void * vx, size_t bx, const DataIn
auto m3 = _mm256_set1_epi8(0x30);
static const uint8_t k_shuff[32] = {0, 1, 8, 9, 2, 3, 10, 11, 4, 5, 12, 13, 6, 7, 14, 15, 0, 1, 8, 9, 2, 3, 10, 11, 4, 5, 12, 13, 6, 7, 14, 15};
auto shuff = _mm256_loadu_si256((const __m256i *)k_shuff);
#ifndef HAVE_FANCY_SIMD
#ifdef HAVE_FANCY_SIMD
__m256 d4s[nrc_y];
#else
auto m1 = _mm256_set1_epi16(1);
#endif
int nbl = n / QK_K;
@@ -3270,11 +3272,19 @@ static void mul_mat_q6_k_r4_q8_k(int n, const void * vx, size_t bx, const DataIn
for (int ibl = 0; ibl < nbl; ++ibl) { // Block of 256
auto dl = _mm_cvtph_ps(_mm_loadl_epi64((const __m128i *)iq6[ibl].d));
auto d4 = _mm256_set_m128(dl, dl);
#ifdef HAVE_FANCY_SIMD
for (int iy = 0; iy < nrc_y; ++iy) {
d4s[iy] = _mm256_mul_ps(d4, _mm256_set1_ps(q8.scale(iy, ibl)));
}
#else
if constexpr (nrc_y == 1) {
d4 = _mm256_mul_ps(d4, _mm256_set1_ps(q8.scale(0, ibl)));
}
#endif
{
#ifndef HAVE_FANCY_SIMD
auto min = _mm256_mul_ps(d4, _mm256_set1_ps(-32.f));
#endif
auto t1 = _mm256_shuffle_epi8(_mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i *)iq6[ibl].scales+0)), shuff); // blocks 0, 1, 2, 3 for each row
auto t2 = _mm256_shuffle_epi8(_mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i *)iq6[ibl].scales+1)), shuff); // blocks 4, 5, 6, 7 for each row
auto t3 = _mm256_shuffle_epi8(_mm256_cvtepi8_epi16(_mm_loadu_si128((const __m128i *)iq6[ibl].scales+2)), shuff); // blocks 8, 9, 10, 11 for each row
@@ -3291,23 +3301,28 @@ static void mul_mat_q6_k_r4_q8_k(int n, const void * vx, size_t bx, const DataIn
sumi = _mm256_dpwssd_epi32(sumi, s2, _mm256_shuffle_epi32(bsums, 0x55));
sumi = _mm256_dpwssd_epi32(sumi, s3, _mm256_shuffle_epi32(bsums, 0xaa));
sumi = _mm256_dpwssd_epi32(sumi, s4, _mm256_shuffle_epi32(bsums, 0xff));
acc[iy] = _mm256_fmadd_ps(_mm256_mul_ps(d4s[iy], _mm256_set1_ps(-32.f)), _mm256_cvtepi32_ps(sumi), acc[iy]);
#else
sumi = _mm256_add_epi32(sumi, _mm256_madd_epi16(s1, _mm256_shuffle_epi32(bsums, 0x00)));
sumi = _mm256_add_epi32(sumi, _mm256_madd_epi16(s2, _mm256_shuffle_epi32(bsums, 0x55)));
sumi = _mm256_add_epi32(sumi, _mm256_madd_epi16(s3, _mm256_shuffle_epi32(bsums, 0xaa)));
sumi = _mm256_add_epi32(sumi, _mm256_madd_epi16(s4, _mm256_shuffle_epi32(bsums, 0xff)));
#endif
if constexpr (nrc_y == 1) {
acc[iy] = _mm256_fmadd_ps(min, _mm256_cvtepi32_ps(sumi), acc[iy]);
} else {
acc[iy] = _mm256_fmadd_ps(_mm256_mul_ps(min, _mm256_set1_ps(q8.scale(iy, ibl))), _mm256_cvtepi32_ps(sumi), acc[iy]);
}
#endif
}
}
const uint32_t * scales = (const uint32_t *)iq6[ibl].scales;
for (int ib = 0; ib < QK_K/32; ++ib) {
auto iscales = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i *)(scales + 2*ib)));
#ifdef HAVE_FANCY_SIMD
auto scales = _mm256_cvtepi32_ps(iscales);
#else
auto scales = _mm256_mul_ps(d4, _mm256_cvtepi32_ps(iscales));
#endif
auto lbits1 = _mm256_loadu_si256((const __m256i *)iq6[ibl].ql+2*ib+0);
auto lbits2 = _mm256_loadu_si256((const __m256i *)iq6[ibl].ql+2*ib+1);
auto hbits = _mm256_loadu_si256((const __m256i *)iq6[ibl].qh+ib);
@@ -3323,6 +3338,7 @@ static void mul_mat_q6_k_r4_q8_k(int n, const void * vx, size_t bx, const DataIn
sumi = _mm256_dpbusd_epi32(sumi, qx[1], _mm256_shuffle_epi32(y, 0x55));
sumi = _mm256_dpbusd_epi32(sumi, qx[2], _mm256_shuffle_epi32(y, 0xaa));
sumi = _mm256_dpbusd_epi32(sumi, qx[3], _mm256_shuffle_epi32(y, 0xff));
acc[iy] = _mm256_fmadd_ps(_mm256_mul_ps(scales, d4s[iy]), _mm256_cvtepi32_ps(sumi), acc[iy]);
#else
auto sumi1 = _mm256_add_epi16(_mm256_maddubs_epi16(qx[0], _mm256_shuffle_epi32(y, 0x00)),
_mm256_maddubs_epi16(qx[1], _mm256_shuffle_epi32(y, 0x55)));
@@ -3330,12 +3346,12 @@ static void mul_mat_q6_k_r4_q8_k(int n, const void * vx, size_t bx, const DataIn
_mm256_maddubs_epi16(qx[3], _mm256_shuffle_epi32(y, 0xff)));
// Quants are in 0...63, so we can add at most 4 as int16_t to be sure of no int16_t overflow
auto sumi = _mm256_add_epi32(_mm256_madd_epi16(m1, sumi1), _mm256_madd_epi16(m1, sumi2));
#endif
if constexpr (nrc_y == 1) {
acc[iy] = _mm256_fmadd_ps(scales, _mm256_cvtepi32_ps(sumi), acc[iy]);
} else {
acc[iy] = _mm256_fmadd_ps(_mm256_mul_ps(scales, _mm256_set1_ps(q8.scale(iy, ibl))), _mm256_cvtepi32_ps(sumi), acc[iy]);
}
#endif
}
}
}