ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-02-24 07:04:11 +00:00
Code Issues Packages Projects Releases Wiki Activity

iq1_kt: slightly faster GEMV

Browse Source 
18.6 t/s -> 19.4 t/s
This commit is contained in:
Iwan Kawrakow
2025-07-16 11:22:09 +03:00
parent 572ed1e71c
commit 6d1ddf1c26
1 changed files with 57 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files

66
ggml/src/iqk/iqk_gemm_ktquants.cpp
View File

@@ -199,6 +199,56 @@ struct Trellis3 {
#else
auto dot = _mm256_maddubs_epi16(aux[i], _mm256_set1_epi32(0x01010101));
aux[i] = _mm256_add_epi32(offset, _mm256_madd_epi16(dot, _mm256_set1_epi16(1)));
#endif
}
for (int k = 0; k < 4; ++k) {
auto v1 = _mm256_packs_epi32(aux[4*k+0], aux[4*k+1]);
auto v2 = _mm256_packs_epi32(aux[4*k+2], aux[4*k+3]);
result[k] = _mm256_permutevar8x32_epi32(_mm256_packs_epi16(v1, v2), shuffle);
}
if constexpr (is_abs) {
for (int k = 0; k < 4; ++k) {
result[k] = _mm256_sign_epi8(result[k], result[k]);
}
}
}
IQK_ALWAYS_INLINE inline void next_128(__m256i val, __m256i * result) const {
// Even though we only have 16 vector registers nn AVX2, this is still faster
__m256i aux[16];
__m256i tmp[2];
tmp[0] = _mm256_cvtepu16_epi32(_mm256_castsi256_si128(val));
tmp[1] = _mm256_cvtepu16_epi32(_mm256_extracti128_si256(val, 1));
for (int k = 0; k < 2; ++k) {
auto vl = _mm256_castsi256_si128(tmp[k]);
auto v = MM256_SET_M128I(vl, vl);
aux[8*k+0] = _mm256_shuffle_epi32(v, 0x00);
aux[8*k+1] = _mm256_shuffle_epi32(v, 0x55);
aux[8*k+2] = _mm256_shuffle_epi32(v, 0xaa);
aux[8*k+3] = _mm256_shuffle_epi32(v, 0xff);
auto vh = _mm256_extracti128_si256(tmp[k], 1);
v = MM256_SET_M128I(vh, vh);
aux[8*k+4] = _mm256_shuffle_epi32(v, 0x00);
aux[8*k+5] = _mm256_shuffle_epi32(v, 0x55);
aux[8*k+6] = _mm256_shuffle_epi32(v, 0xaa);
aux[8*k+7] = _mm256_shuffle_epi32(v, 0xff);
}
for (int i = 0; i < 16; ++i) {
aux[i] = _mm256_mullo_epi32(aux[i], mka);
}
auto mask = _mm256_set1_epi32(0x3f3f3f3f);
for (int i = 0; i < 16; ++i) {
aux[i] = _mm256_and_si256(aux[i], mask);
}
auto offset = _mm256_set1_epi32(-126);
#if defined(__AVX512F__) && defined(__AVX512VNNI__) && defined(__AVX512VL__)
auto m1 = _mm256_set1_epi32(0x01010101);
#endif
for (int i = 0; i < 16; ++i) {
#if defined(__AVX512F__) && defined(__AVX512VNNI__) && defined(__AVX512VL__)
aux[i] = _mm256_dpbusd_epi32(offset, aux[i], m1);
#else
auto dot = _mm256_maddubs_epi16(aux[i], _mm256_set1_epi32(0x01010101));
aux[i] = _mm256_add_epi32(offset, _mm256_madd_epi16(dot, _mm256_set1_epi16(1)));
#endif
}
for (int k = 0; k < 4; ++k) {
@@ -560,7 +610,6 @@ void mul_mat_iq1_kt_q8_2_x4_T(int n, const void * vx, size_t bx, const DataInfo&
};
__m256i idx[2];
uint16_t index[16];
for (int ix = 0; ix < nrc_x; ++ix) {
const float * dptr = (const float *)((const char*)vx + ix*bx);
@@ -582,15 +631,14 @@ void mul_mat_iq1_kt_q8_2_x4_T(int n, const void * vx, size_t bx, const DataInfo&
auto qh16 = _mm256_cvtepu8_epi16(_mm_loadu_si128((const __m128i *)x[i].qh));
idx[0] = _mm256_or_si256(_mm256_cvtepu8_epi16(qs8l), _mm256_and_si256(_mm256_set1_epi16(0xf00), _mm256_slli_epi16(qh16, 8)));
idx[1] = _mm256_or_si256(_mm256_cvtepu8_epi16(qs8h), _mm256_and_si256(_mm256_set1_epi16(0xf00), _mm256_slli_epi16(qh16, 4)));
//idx[0] = _mm256_add_epi16(idx[0], _mm256_set1_epi16(4096));
//idx[1] = _mm256_add_epi16(idx[1], _mm256_set1_epi16(4096));
auto sh32 = _mm256_and_si256(_mm256_srli_epi32(_mm256_cvtepu8_epi32(sh), 4), _mm256_set1_epi32(0xf));
sh32 = _mm256_and_si256(_mm256_mullo_epi32(sh32, _mm256_set1_epi32(0x01020408)), _mm256_set1_epi8(8));
idx[0] = _mm256_add_epi16(idx[0], _mm256_slli_epi16(_mm256_cvtepu8_epi16(_mm256_castsi256_si128(sh32)), 9));
idx[1] = _mm256_add_epi16(idx[1], _mm256_slli_epi16(_mm256_cvtepu8_epi16(_mm256_extracti128_si256(sh32, 1)), 9));
idx[0] = _mm256_add_epi16(idx[0], _mm256_set1_epi16(4096));
idx[1] = _mm256_add_epi16(idx[1], _mm256_set1_epi16(4096));
auto sh32 = _mm256_and_si256(_mm256_cvtepu8_epi32(sh), _mm256_set1_epi32(0xf0));
sh32 = _mm256_and_si256(_mm256_mullo_epi32(sh32, _mm256_set1_epi32(0x01020408)), _mm256_set1_epi8(-128));
idx[0] = _mm256_add_epi16(idx[0], _mm256_slli_epi16(_mm256_cvtepu8_epi16(_mm256_castsi256_si128(sh32)), 5));
idx[1] = _mm256_add_epi16(idx[1], _mm256_slli_epi16(_mm256_cvtepu8_epi16(_mm256_extracti128_si256(sh32, 1)), 5));
for (int i128 = 0; i128 < 2; ++i128) {
_mm256_storeu_si256((__m256i *)index, idx[i128]);
trellis.next_128(index, 4096, xv);
trellis.next_128(idx[i128], xv);
for (int iy = 0; iy < nrc_y; ++iy) {
const block_q8_2_x4& yb = y[iy][2*i+i128];
auto dy4 = _mm_castsi128_ps(_mm_slli_epi32(_mm_cvtepu16_epi32(_mm_loadl_epi64((const __m128i *)yb.d)), 16));