ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-03-06 12:00:29 +00:00
Code Issues Packages Projects Releases Wiki Activity

iq3_k_r4: AVX2

Browse Source
This commit is contained in:
Iwan Kawrakow
2024-12-16 19:51:51 +02:00
parent d73efc787f
commit 751fbaaa8c
1 changed files with 12 additions and 14 deletions
Download Patch File Download Diff File Expand all files Collapse all files

26
ggml/src/iqk/iqk_mul_mat.cpp
View File

@@ -3969,11 +3969,10 @@ static void mul_mat_iq3_k_r4_q8_k(int n, const void * vx, size_t bx, const DataI
auto shift_shuffle = _mm256_set_epi64x(0x0707070706060606, 0x0505050504040404, 0x0303030302020202, 0x0101010100000000);
auto values128 = _mm_loadu_si128((const __m128i *)iq3nl_values);
auto values = MM256_SET_M128I(values128, values128);
#ifdef HAVE_FANCY_SIMD
values = _mm256_add_epi8(values, _mm256_set1_epi8(64));
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);
#else
#ifndef HAVE_FANCY_SIMD
auto s_shuffle = _mm256_set_epi64x(0x0f0e0f0e0d0c0d0c, 0x0b0a0b0a09080908, 0x0706070605040504, 0x0302030201000100);
#endif
int nbl = n / QK_K;
@@ -3997,7 +3996,6 @@ static void mul_mat_iq3_k_r4_q8_k(int n, const void * vx, size_t bx, const DataI
_mm256_storeu_si256((__m256i *)stored_scales+0, i8scales1);
_mm256_storeu_si256((__m256i *)stored_scales+1, i8scales2);
__m256i isum[nrc_y] = {};
#ifdef HAVE_FANCY_SIMD
{
auto t1 = _mm256_shuffle_epi8(_mm256_cvtepi8_epi16(_mm256_extracti128_si256(i8scales1, 0)), shuff); // blocks 0, 1, 2, 3 for each row
auto t2 = _mm256_shuffle_epi8(_mm256_cvtepi8_epi16(_mm256_extracti128_si256(i8scales1, 1)), shuff); // blocks 4, 5, 6, 7 for each row
@@ -4009,13 +4007,19 @@ static void mul_mat_iq3_k_r4_q8_k(int n, const void * vx, size_t bx, const DataI
auto s4 = _mm256_mullo_epi16(_mm256_set1_epi16(-64), MM256_SET_M128I(_mm256_extracti128_si256(t4, 1), _mm256_extracti128_si256(t2, 1))); // blocks 6, 7, 14, 15
for (int iy = 0; iy < nrc_y; ++iy) {
auto bsums = q8.load_bsums(iy, ibl);
#ifdef HAVE_FANCY_SIMD
isum[iy] = _mm256_dpwssd_epi32(isum[iy], s1, _mm256_shuffle_epi32(bsums, 0x00));
isum[iy] = _mm256_dpwssd_epi32(isum[iy], s2, _mm256_shuffle_epi32(bsums, 0x55));
isum[iy] = _mm256_dpwssd_epi32(isum[iy], s3, _mm256_shuffle_epi32(bsums, 0xaa));
isum[iy] = _mm256_dpwssd_epi32(isum[iy], s4, _mm256_shuffle_epi32(bsums, 0xff));
#else
isum[iy] = _mm256_add_epi32(isum[iy], _mm256_madd_epi16(s1, _mm256_shuffle_epi32(bsums, 0x00)));
isum[iy] = _mm256_add_epi32(isum[iy], _mm256_madd_epi16(s2, _mm256_shuffle_epi32(bsums, 0x55)));
isum[iy] = _mm256_add_epi32(isum[iy], _mm256_madd_epi16(s3, _mm256_shuffle_epi32(bsums, 0xaa)));
isum[iy] = _mm256_add_epi32(isum[iy], _mm256_madd_epi16(s4, _mm256_shuffle_epi32(bsums, 0xff)));
#endif
}
}
#endif
for (int ib = 0; ib < QK_K/32; ++ib) {
#ifdef HAVE_FANCY_SIMD
auto scales = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i *)(stored_scales + ib)));
@@ -4035,12 +4039,6 @@ static void mul_mat_iq3_k_r4_q8_k(int n, const void * vx, size_t bx, const DataI
qx[1] = _mm256_shuffle_epi8(values, _mm256_add_epi8(qx[1], shift));
qx[2] = _mm256_shuffle_epi8(values, _mm256_add_epi8(qx[2], shift));
qx[3] = _mm256_shuffle_epi8(values, _mm256_add_epi8(qx[3], shift));
#ifndef HAVE_FANCY_SIMD
auto s1 = _mm256_sign_epi8(qx[0], qx[0]);
auto s2 = _mm256_sign_epi8(qx[1], qx[1]);
auto s3 = _mm256_sign_epi8(qx[2], qx[2]);
auto s4 = _mm256_sign_epi8(qx[3], qx[3]);
#endif
for (int iy = 0; iy < nrc_y; ++iy) {
auto y = _mm256_loadu_si256((const __m256i*)q8.y[iy][ibl].qs+ib);
#ifdef HAVE_FANCY_SIMD
@@ -4051,10 +4049,10 @@ static void mul_mat_iq3_k_r4_q8_k(int n, const void * vx, size_t bx, const DataI
sumi = _mm256_dpbusd_epi32(sumi, qx[3], _mm256_shuffle_epi32(y, 0xff));
isum[iy] = _mm256_add_epi32(isum[iy], _mm256_mullo_epi32(scales, sumi));
#else
auto sumi1 = _mm256_maddubs_epi16(s1, _mm256_sign_epi8(_mm256_shuffle_epi32(y, 0x00), qx[0]));
auto sumi2 = _mm256_maddubs_epi16(s2, _mm256_sign_epi8(_mm256_shuffle_epi32(y, 0x55), qx[1]));
auto sumi3 = _mm256_maddubs_epi16(s3, _mm256_sign_epi8(_mm256_shuffle_epi32(y, 0xaa), qx[2]));
auto sumi4 = _mm256_maddubs_epi16(s4, _mm256_sign_epi8(_mm256_shuffle_epi32(y, 0xff), qx[3]));
auto sumi1 = _mm256_maddubs_epi16(qx[0], _mm256_shuffle_epi32(y, 0x00));
auto sumi2 = _mm256_maddubs_epi16(qx[1], _mm256_shuffle_epi32(y, 0x55));
auto sumi3 = _mm256_maddubs_epi16(qx[2], _mm256_shuffle_epi32(y, 0xaa));
auto sumi4 = _mm256_maddubs_epi16(qx[3], _mm256_shuffle_epi32(y, 0xff));
isum[iy] = _mm256_add_epi32(isum[iy], _mm256_add_epi32(_mm256_madd_epi16(scales, sumi1), _mm256_madd_epi16(scales, sumi2)));
isum[iy] = _mm256_add_epi32(isum[iy], _mm256_add_epi32(_mm256_madd_epi16(scales, sumi3), _mm256_madd_epi16(scales, sumi4)));
#endif