ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-03-14 15:57:37 +00:00
Code Issues Packages Projects Releases Wiki Activity

Slightly faster IQ4_XS_R4 on AVX2 (#143)

Browse Source 
* iq4_xs_r4: slightly faster and correct AVX2 implementation

* Minor

* Delete unused stuff

---------

Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
This commit is contained in:
Kawrakow
2024-12-16 14:40:10 +01:00
committed by GitHub
parent 3283fbeb82
commit 93cfa3e662
1 changed files with 20 additions and 13 deletions
Download Patch File Download Diff File Expand all files Collapse all files

33
ggml/src/iqk/iqk_mul_mat.cpp
View File

@@ -2935,7 +2935,7 @@ static void mul_mat_iq4_xs_r4_q8_k_avx2(int n, const void * vx, size_t bx, const
Q8<nrc_y, block_q8_K> q8(info);
auto m4 = _mm256_set1_epi8(0xf);
#ifndef HAVE_FANCY_SIMD
auto m1 = _mm256_set1_epi16(1);
auto s_shuffle = _mm256_set_epi64x(0x0f0e0f0e0d0c0d0c, 0x0b0a0b0a09080908, 0x0706070605040504, 0x0302030201000100);
auto values128 = _mm_loadu_si128((const __m128i *)iq4k_values);
auto values = MM256_SET_M128I(values128, values128);
#else
@@ -2945,6 +2945,7 @@ static void mul_mat_iq4_xs_r4_q8_k_avx2(int n, const void * vx, size_t bx, const
using helper_t = union { __m256i vec; uint32_t val[8]; };
helper_t h;
__m256 acc[nrc_y] = {};
__m256i isum[nrc_y] = {};
__m256i qx[4];
for (int ix = 0; ix < nrc_x; ix += 4) {
const block_iq4_xs_r4 * iq4 = (const block_iq4_xs_r4 *)((const char *)vx + (ix+0)*bx);
@@ -2958,10 +2959,16 @@ static void mul_mat_iq4_xs_r4_q8_k_avx2(int n, const void * vx, size_t bx, const
auto sh = _mm256_and_si256(MM256_SET_M128I(shbits, _mm_slli_epi16(shbits, 4)), _mm256_set1_epi8(0x30));
h.vec = _mm256_sub_epi8(_mm256_or_si256(sl, sh), _mm256_set1_epi8(32));
for (int ib = 0; ib < QK_K/32; ++ib) {
#ifdef HAVE_FANCY_SIMD
auto iscales = _mm256_cvtepi8_epi32(_mm_set1_epi32(h.val[ib]));
auto scales = _mm256_mul_ps(d4, _mm256_cvtepi32_ps(iscales));
#ifdef HAVE_FANCY_SIMD
auto scales_m = _mm256_mul_ps(scales, _mm256_set1_ps(-64.f));
for (int iy = 0; iy < nrc_y; ++iy) {
float m8 = ((const float *)q8.y[iy][ibl].bsums)[ib];
acc[iy] = _mm256_fmadd_ps(scales_m, _mm256_set1_ps(m8), acc[iy]);
}
#else
auto iscales = _mm256_shuffle_epi8(_mm256_cvtepi8_epi16(_mm_set1_epi32(h.val[ib])), s_shuffle);
#endif
auto bits1 = _mm256_loadu_si256((const __m256i *)iq4[ibl].qs+2*ib+0);
auto bits2 = _mm256_loadu_si256((const __m256i *)iq4[ibl].qs+2*ib+1);
@@ -2983,21 +2990,21 @@ static void mul_mat_iq4_xs_r4_q8_k_avx2(int n, const void * vx, size_t bx, const
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));
float d8 = q8.scale(iy, ibl);
float m8 = ((const float *)q8.y[iy][ibl].bsums)[ib];
acc[iy] = _mm256_fmadd_ps(_mm256_mul_ps(scales, _mm256_set1_ps(d8)), _mm256_cvtepi32_ps(sumi), acc[iy]);
acc[iy] = _mm256_fmadd_ps(scales_m, _mm256_set1_ps(m8), acc[iy]);
isum[iy] = _mm256_add_epi32(isum[iy], _mm256_mullo_epi32(iscales, sumi));
#else
// This is wrong! (it may overflow the int16_t range
auto sumi1 = _mm256_add_epi16(_mm256_maddubs_epi16(s1, _mm256_sign_epi8(_mm256_shuffle_epi32(y, 0x00), qx[0])),
_mm256_maddubs_epi16(s2, _mm256_sign_epi8(_mm256_shuffle_epi32(y, 0x55), qx[1])));
auto sumi2 = _mm256_add_epi16(_mm256_maddubs_epi16(s3, _mm256_sign_epi8(_mm256_shuffle_epi32(y, 0xaa), qx[2])),
_mm256_maddubs_epi16(s4, _mm256_sign_epi8(_mm256_shuffle_epi32(y, 0xff), qx[3])));
auto sumi = _mm256_add_epi32(_mm256_madd_epi16(m1, sumi1), _mm256_madd_epi16(m1, sumi2));
acc[iy] = _mm256_fmadd_ps(_mm256_mul_ps(scales, _mm256_set1_ps(q8.scale(iy, ibl))), _mm256_cvtepi32_ps(sumi), acc[iy]);
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]));
isum[iy] = _mm256_add_epi32(isum[iy], _mm256_add_epi32(_mm256_madd_epi16(iscales, sumi1), _mm256_madd_epi16(iscales, sumi2)));
isum[iy] = _mm256_add_epi32(isum[iy], _mm256_add_epi32(_mm256_madd_epi16(iscales, sumi3), _mm256_madd_epi16(iscales, sumi4)));
#endif
}
}
for (int iy = 0; iy < nrc_y; ++iy) {
acc[iy] = _mm256_fmadd_ps(_mm256_mul_ps(d4, _mm256_set1_ps(q8.scale(iy, ibl))), _mm256_cvtepi32_ps(isum[iy]), acc[iy]);
isum[iy] = _mm256_setzero_si256();
}
}
for (int iy = 0; iy < nrc_y; ++iy) {
auto sum = _mm_add_ps(_mm256_castps256_ps128(acc[iy]), _mm256_extractf128_ps(acc[iy], 1));