From 8db83dac1ddec2f7fc13cf2dbdc620bcb2ed00e9 Mon Sep 17 00:00:00 2001 From: Iwan Kawrakow Date: Mon, 9 Jun 2025 09:00:56 +0300 Subject: [PATCH] New iq2_kt: AVX2 dequantize --- ggml/src/ggml.c | 11 ++- ggml/src/iqk/iqk_gemm_ktquants.cpp | 119 ++++++++++++++++++++++++++++- ggml/src/iqk/iqk_mul_mat.cpp | 4 +- 3 files changed, 126 insertions(+), 8 deletions(-) diff --git a/ggml/src/ggml.c b/ggml/src/ggml.c index 9375963f..edf25e35 100644 --- a/ggml/src/ggml.c +++ b/ggml/src/ggml.c @@ -1596,11 +1596,16 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = { .from_float = quantize_row_iq2_kt, .from_float_ref = (ggml_from_float_t)quantize_row_iq2_kt_ref, .vec_dot = vec_dot_iq2_kt_q8_k, -#ifdef __ARM_NEON - .vec_dot_type = GGML_TYPE_F16, +#if defined __AVX2__ + .vec_dot_type = GGML_TYPE_Q8_2_X4, #else - .vec_dot_type = GGML_TYPE_F32, + .vec_dot_type = GGML_TYPE_Q8_0_X4, #endif +//#ifdef __ARM_NEON +// .vec_dot_type = GGML_TYPE_F16, +//#else +// .vec_dot_type = GGML_TYPE_F32, +//#endif .nrows = 1, .row_meta_size = 4, }, diff --git a/ggml/src/iqk/iqk_gemm_ktquants.cpp b/ggml/src/iqk/iqk_gemm_ktquants.cpp index 965ecc2d..5c971b31 100644 --- a/ggml/src/iqk/iqk_gemm_ktquants.cpp +++ b/ggml/src/iqk/iqk_gemm_ktquants.cpp @@ -98,6 +98,7 @@ struct Trellis2 { }; +template struct Trellis3 { constexpr static uint32_t ka = 89226354; constexpr static uint32_t kb = 64248484; @@ -107,14 +108,26 @@ struct Trellis3 { constexpr static uint32_t kb2 = kb1*ka+kb; constexpr static uint32_t ka3 = ka2*ka; constexpr static uint32_t kb3 = kb2*ka+kb; - const __m256i mka = _mm256_setr_epi32(ka, ka1, ka2, ka3, ka, ka1, ka2, ka3); - const __m256i mkb = _mm256_setr_epi32(kb, kb1, kb2, kb3, kb, kb1, kb2, kb3); + constexpr static uint32_t ka4 = ka3*ka; + constexpr static uint32_t kb4 = kb3*ka+kb; + constexpr static uint32_t ka5 = ka4*ka; + constexpr static uint32_t kb5 = kb4*ka+kb; + constexpr static uint32_t ka6 = ka5*ka; + constexpr static uint32_t kb6 = kb5*ka+kb; + constexpr static uint32_t ka7 = ka6*ka; + constexpr static uint32_t kb7 = kb6*ka+kb; + const __m256i mka = is_8 ? _mm256_setr_epi32(ka, ka1, ka2, ka3, ka4, ka5, ka6, ka7) : _mm256_setr_epi32(ka, ka1, ka2, ka3, ka, ka1, ka2, ka3); + const __m256i mkb = is_8 ? _mm256_setr_epi32(kb, kb1, kb2, kb3, kb4, kb5, kb6, kb7) : _mm256_setr_epi32(kb, kb1, kb2, kb3, kb, kb1, kb2, kb3); const __m256i shuffle = _mm256_set_epi32(7, 3, 6, 2, 5, 1, 4, 0); inline __m256i next8(uint32_t val1, uint32_t val2) const { __m256i mval = MM256_SET_M128I(_mm_set1_epi32(val2), _mm_set1_epi32(val1)); return _mm256_add_epi32(_mm256_mullo_epi32(mval, mka), mkb); } + inline __m256i next8(uint32_t val) const { + __m256i mval = _mm256_set1_epi32(val); + return _mm256_add_epi32(_mm256_mullo_epi32(mval, mka), mkb); + } inline __m256 gen8(uint32_t val1, uint32_t val2) const { auto v8 = _mm256_and_si256(next8(val1, val2), _mm256_set1_epi32(0x3f3f3f3f)); #ifdef HAVE_FANCY_SIMD @@ -122,6 +135,16 @@ struct Trellis3 { #else auto dot = _mm256_maddubs_epi16(v8, _mm256_set1_epi32(0x01010101)); auto i8 = _mm256_add_epi32(_mm256_set1_epi32(-126), _mm256_madd_epi16(dot, _mm256_set1_epi16(1))); +#endif + return _mm256_cvtepi32_ps(i8); + } + inline __m256 gen8(uint32_t val) const { + auto v8 = _mm256_and_si256(next8(val), _mm256_set1_epi32(0x3f3f3f3f)); +#ifdef HAVE_FANCY_SIMD + auto i8 = _mm256_dpbusd_epi32(_mm256_set1_epi32(-126), _mm256_set1_epi32(0x01010101), v8); +#else + auto dot = _mm256_maddubs_epi16(v8, _mm256_set1_epi32(0x01010101)); + auto i8 = _mm256_add_epi32(_mm256_set1_epi32(-126), _mm256_madd_epi16(dot, _mm256_set1_epi16(1))); #endif return _mm256_cvtepi32_ps(i8); } @@ -144,6 +167,34 @@ struct Trellis3 { // 4, 5, 6, 7, 12, 13, 14, 15, 20, 21, 22, 23, 28, 29, 30, 31 return _mm256_permutevar8x32_epi32(aux[0], shuffle); } + template + inline void next64(const uint32_t * val, __m256i * result) const { + const __m256i offset = is_unsigned ? _mm256_setzero_si256() : _mm256_set1_epi32(-126); + auto vka3 = _mm256_set1_epi32(ka3), vkb3 = _mm256_set1_epi32(kb3); + __m256i aux[8]; + for (int i = 0; i < 4; ++i) { + auto i8_1 = next8(val[2*i+0], val[2*i+1]); + auto i8_2 = _mm256_add_epi32(_mm256_mullo_epi32(i8_1, vka3), vkb3); + i8_1 = _mm256_and_si256(i8_1, _mm256_set1_epi32(0x3f3f3f3f)); + i8_2 = _mm256_and_si256(i8_2, _mm256_set1_epi32(0x3f3f3f3f)); +#ifdef HAVE_FANCY_SIMD + aux[i+0] = _mm256_dpbusd_epi32(offset, _mm256_set1_epi32(0x01010101), i8_1); + aux[i+4] = _mm256_dpbusd_epi32(offset, _mm256_set1_epi32(0x01010101), i8_2); +#else + auto dot1 = _mm256_maddubs_epi16(i8_1, _mm256_set1_epi32(0x01010101)); + auto dot2 = _mm256_maddubs_epi16(i8_2, _mm256_set1_epi32(0x01010101)); + aux[i+0] = _mm256_add_epi32(offset, _mm256_madd_epi16(dot1, _mm256_set1_epi16(1))); + aux[i+4] = _mm256_add_epi32(offset, _mm256_madd_epi16(dot2, _mm256_set1_epi16(1))); +#endif + } + for (int k = 0; k < 2; ++k) { + aux[4*k+0] = _mm256_packs_epi32(aux[4*k+0], aux[4*k+1]); // 0, 1, 2, 3, 8, 9, 10, 11, 4, 5, 6, 7, 12, 13, 14, 15 + aux[4*k+2] = _mm256_packs_epi32(aux[4*k+2], aux[4*k+3]); // 16, 17, 18, 19, 24, 25, 26, 27, 20, 21, 22, 23, 28, 29, 30, 31 + aux[4*k+0] = _mm256_packs_epi16(aux[4*k+0], aux[4*k+2]); // 0, 1, 2, 3, 8, 9, 10, 11, 16, 17, 18, 19, 24, 25, 26, 27 + // 4, 5, 6, 7, 12, 13, 14, 15, 20, 21, 22, 23, 28, 29, 30, 31 + result[k] = _mm256_permutevar8x32_epi32(aux[4*k+0], shuffle); + } + } }; void iqk_dequantize_iq2_kt(int n, const void * vx, size_t bx, float * y, size_t stride_y, int nrc_x) { @@ -185,6 +236,60 @@ void iqk_dequantize_iq2_kt(int n, const void * vx, size_t bx, float * y, size_t } } +void iqk_dequantize_iq2_kt_q80_r8(int n, const void * vx, size_t bx, void * vy, int nrc_x) { + GGML_ASSERT(n%QK_K == 0); + GGML_ASSERT(nrc_x%8 == 0); + const int nb = n/QK_K; + + Trellis3 trellis; + + auto shifts = _mm_set_epi32(0, 0, 4, 0); + auto values = _mm_loadu_si128((const __m128i *)iq4k_values); + + block_q8_0_r8 * y = (block_q8_0_r8 *)vy; + + const block_iq2_kt * x8[8]; + float dkt[8]; + float ls[8]; + float ls_all[64]; + uint32_t idx[8]; + + for (int ix = 0; ix < nrc_x; ix += 8) { + for (int k = 0; k < 8; ++k) { + const float * dptr = (const float *)((const char*)vx + (ix+k)*bx); + dkt[k] = dptr[0]; + x8[k] = (const block_iq2_kt *)(dptr + 1); + } + auto vd = _mm256_mul_ps(_mm256_set1_ps(1.05f), _mm256_loadu_ps(dkt)); + + for (int i = 0; i < nb; ++i) { + for (int k = 0; k < 8; ++k) { + auto s8 = _mm_set1_epi32(*(const uint32_t *)x8[k][i].scales); + s8 = _mm_and_si128(_mm_srlv_epi32(s8, shifts), _mm_set1_epi8(0xf)); + s8 = _mm_shuffle_epi8(values, s8); + auto s32 = _mm256_cvtepi8_epi32(s8); + _mm256_storeu_ps(ls_all + 8*k, _mm256_cvtepi32_ps(s32)); + } + for (int ib = 0; ib < QK_K/32; ++ib) { + for (int k = 0; k < 8; ++k) ls[k] = ls_all[8*k+ib]; + auto scales = _mm256_mul_ps(vd, _mm256_loadu_ps(ls)); + _mm_storeu_si128((__m128i *)y[ib].d, _mm256_cvtps_ph(scales, _MM_FROUND_TO_NEAREST_INT)); + for (int j = 0; j < 4; ++j) { + for (int k = 0; k < 8; ++k) { + const uint16_t * ql = (const uint16_t *)x8[k][i].ql; + idx[k] = ql[4*ib+j] + 4096; + } + __m256i packed[2]; + trellis.next64(idx, packed); + _mm256_storeu_si256((__m256i *)y[ib].qs+2*j+0, packed[0]); + _mm256_storeu_si256((__m256i *)y[ib].qs+2*j+1, packed[1]); + } + } + y += 8; // = QK_K/32; + } + } +} + template void mul_mat_iq2_kt_F32_T(int n, const void * vx, size_t bx, const DataInfo& info, int nrc_x) { assert(n%QK_K == 0); @@ -655,6 +760,14 @@ bool iqk_set_kernels_ktquants(int ne00, int typeA, int typeB, std::array= 32 ? GGML_TYPE_Q8_0_R8 : type; case GGML_TYPE_IQ4_NL : return nrc_y >= 32 ? GGML_TYPE_Q8_0_R8 : type; case GGML_TYPE_Q8_0 : return nrc_y >= 32 ? GGML_TYPE_Q8_0_R8 : type; - case GGML_TYPE_IQ2_KT : return nrc_y >= 32 ? GGML_TYPE_F32 : type; - case GGML_TYPE_IQ3_KT : return nrc_y >= 32 ? GGML_TYPE_F32 : type; + case GGML_TYPE_IQ2_KT : return nrc_y >= 32 ? GGML_TYPE_Q8_0_R8 : type; + case GGML_TYPE_IQ3_KT : return nrc_y >= 32 ? GGML_TYPE_Q8_0_R8 : type; case GGML_TYPE_IQ4_KT : return nrc_y >= 32 ? GGML_TYPE_Q8_0_R8 : type; default: break; }