From 8bfd5cff314222c624340ccea9d20bfb1bff16ac Mon Sep 17 00:00:00 2001 From: Iwan Kawrakow Date: Mon, 16 Dec 2024 16:38:14 +0200 Subject: [PATCH] iq4_k_r4: slightly better AVX2 227 t/s -> 249 t/s --- ggml/src/iqk/iqk_mul_mat.cpp | 22 ++++++++++------------ 1 file changed, 10 insertions(+), 12 deletions(-) diff --git a/ggml/src/iqk/iqk_mul_mat.cpp b/ggml/src/iqk/iqk_mul_mat.cpp index e5cb7d52..ff508aaf 100644 --- a/ggml/src/iqk/iqk_mul_mat.cpp +++ b/ggml/src/iqk/iqk_mul_mat.cpp @@ -3963,7 +3963,6 @@ static void mul_mat_iq4_k_r4_q8_k(int n, const void * vx, size_t bx, const DataI auto m30 = _mm256_set1_epi8(0x30); auto m32 = _mm256_set1_epi8(32); auto ms = _mm256_set1_epi8(4); - //auto shift_shuffle = _mm256_set_epi64x(0x0303030302020202, 0x0101010100000000, 0x0303030302020202, 0x0101010100000000); auto shift_shuffle = _mm256_set_epi64x(0x0707070706060606, 0x0505050504040404, 0x0303030302020202, 0x0101010100000000); #ifdef HAVE_FANCY_SIMD auto values = load_iq4nl_values_256(); @@ -3971,14 +3970,14 @@ static void mul_mat_iq4_k_r4_q8_k(int n, const void * vx, size_t bx, const DataI 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 - 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); #endif int nbl = n / QK_K; __m256 acc[nrc_y] = {}; __m256i qx[4]; - int8_t stored_scales[64]; + uint64_t stored_scales[8]; for (int ix = 0; ix < nrc_x; ix += 4) { const block_iq4_k_r4 * iq4 = (const block_iq4_k_r4 *)((const char *)vx + (ix+0)*bx); for (int ibl = 0; ibl < nbl; ++ibl) { // Block of 256 @@ -3989,10 +3988,6 @@ static void mul_mat_iq4_k_r4_q8_k(int n, const void * vx, size_t bx, const DataI 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 auto slbits = _mm256_loadu_si256((const __m256i *)iq4[ibl].scales_l); auto sl1 = _mm256_and_si256(slbits, m4); @@ -4024,12 +4019,13 @@ static void mul_mat_iq4_k_r4_q8_k(int n, const void * vx, size_t bx, const DataI } } #endif + __m256i isum[nrc_y] = {}; for (int ib = 0; ib < QK_K/32; ++ib) { - auto iscales = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i *)(stored_scales + 8*ib))); #ifdef HAVE_FANCY_SIMD + auto iscales = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i *)(stored_scales + ib))); auto scales = _mm256_cvtepi32_ps(iscales); #else - auto scales = _mm256_mul_ps(d4, _mm256_cvtepi32_ps(iscales)); + auto scales = _mm256_shuffle_epi8(_mm256_cvtepi8_epi16(_mm_set1_epi64x(stored_scales[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); @@ -4059,12 +4055,14 @@ static void mul_mat_iq4_k_r4_q8_k(int n, const void * vx, size_t bx, const DataI 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 sumi = _mm256_add_epi32(_mm256_add_epi32(_mm256_madd_epi16(m1, sumi1), _mm256_madd_epi16(m1, sumi2)), - _mm256_add_epi32(_mm256_madd_epi16(m1, sumi3), _mm256_madd_epi16(m1, sumi4))); - acc[iy] = _mm256_fmadd_ps(_mm256_mul_ps(scales, _mm256_set1_ps(q8.scale(iy, ibl))), _mm256_cvtepi32_ps(sumi), acc[iy]); + 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 } } + 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]); + } } for (int iy = 0; iy < nrc_y; ++iy) { auto sum = _mm_add_ps(_mm256_castps256_ps128(acc[iy]), _mm256_extractf128_ps(acc[iy], 1));