From d1788b72678e5375bba3630f54390aa0f37e43f6 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Sat, 26 Jul 2025 20:03:43 +0300 Subject: [PATCH] ggml : repack block_iq4_nlx8 ggml-ci --- ggml/src/ggml-cpu/arch-fallback.h | 14 + ggml/src/ggml-cpu/arch/x86/repack.cpp | 1695 +++++++++++++------------ ggml/src/ggml-cpu/repack.cpp | 251 +++- ggml/src/ggml-cpu/repack.h | 11 + 4 files changed, 1103 insertions(+), 868 deletions(-) diff --git a/ggml/src/ggml-cpu/arch-fallback.h b/ggml/src/ggml-cpu/arch-fallback.h index 10e5342516a9c..845af7d2f0da3 100644 --- a/ggml/src/ggml-cpu/arch-fallback.h +++ b/ggml/src/ggml-cpu/arch-fallback.h @@ -38,16 +38,20 @@ #define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0 #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0 +#define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0 #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0 #define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0 #define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0 #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0 +#define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0 #elif defined(__aarch64__) || defined(__arm__) || defined(_M_ARM) || defined(_M_ARM64) // repack.cpp #define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8 #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K +#define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0 #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K +#define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0 #elif defined(__x86_64__) || defined(__i386__) || defined(_M_IX86) || defined(_M_X64) // repack.cpp #define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4 @@ -73,11 +77,13 @@ #define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0 #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0 +#define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0 #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0 #define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0 #define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0 #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0 +#define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0 #elif defined(__loongarch64) // quants.c #define quantize_row_q8_K_generic quantize_row_q8_K @@ -93,11 +99,13 @@ #define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0 #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0 +#define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0 #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0 #define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0 #define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0 #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0 +#define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0 #elif defined(__riscv) // quants.c #define quantize_row_q8_K_generic quantize_row_q8_K @@ -120,10 +128,12 @@ #define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0 #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0 +#define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0 #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0 #define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0 #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0 +#define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0 #elif defined(__s390x__) // quants.c #define quantize_row_q8_K_generic quantize_row_q8_K @@ -148,11 +158,13 @@ #define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0 #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0 +#define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0 #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0 #define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0 #define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0 #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0 +#define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0 #elif defined(__wasm__) // quants.c #define ggml_vec_dot_q4_1_q8_1_generic ggml_vec_dot_q4_1_q8_1 @@ -176,9 +188,11 @@ #define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0 #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0 +#define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0 #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0 #define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0 #define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0 #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0 +#define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0 #endif diff --git a/ggml/src/ggml-cpu/arch/x86/repack.cpp b/ggml/src/ggml-cpu/arch/x86/repack.cpp index 1982cfef9951f..e08a10772a86b 100644 --- a/ggml/src/ggml-cpu/arch/x86/repack.cpp +++ b/ggml/src/ggml-cpu/arch/x86/repack.cpp @@ -511,38 +511,34 @@ void ggml_quantize_mat_q8_K_4x8(const float * GGML_RESTRICT x, void * GGML_RESTR #endif } -void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) { +// +// GEMV/GEMM templates +// + +#if defined(__AVX2__) || defined(__AVX512F__) + +// GEMV for 8x blocks of 32 4-bit quants with a single scale factor per block +template +static void gemv_q4_b32_8x8_q8_0_lut_avx(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc, __m256i signextendlut) { + static_assert( + std::is_same_v || + std::is_same_v, + "Unsupported block type"); + const int qk = QK8_0; const int nb = n / qk; - const int ncols_interleaved = 8; - const int blocklen = 8; - assert (n % qk == 0); - assert (nc % ncols_interleaved == 0); - - UNUSED(s); UNUSED(bs); - UNUSED(vx); - UNUSED(vy); - UNUSED(nr); - UNUSED(nc); - UNUSED(nb); - UNUSED(ncols_interleaved); - UNUSED(blocklen); -#if defined(__AVX2__) - // Lookup table to convert signed nibbles to signed bytes - __m256i signextendlut = _mm256_castsi128_si256(_mm_set_epi8(-1, -2, -3, -4, -5, -6, -7, -8, 7, 6, 5, 4, 3, 2, 1, 0)); - signextendlut = _mm256_permute2f128_si256(signextendlut, signextendlut, 0); __m128i changemask = _mm_set_epi8(15, 14, 7, 6, 13, 12, 5, 4, 11, 10, 3, 2, 9, 8, 1, 0); __m256i finalpermutemask = _mm256_set_epi32(7, 5, 3, 1, 6, 4, 2, 0); // Permute mask used for easier vector processing at later stages const __m256i m4b = _mm256_set1_epi8(0x0F); - int64_t b_nb = n / QK4_0; + int64_t b_nb = n / 32; - const block_q4_0x8 * b_ptr_start = (const block_q4_0x8 *)vx; + const block_tx8 * b_ptr_start = (const block_tx8 *)vx; const block_q8_0 * a_ptr_start = (const block_q8_0 *)vy; // Process Q8_0 blocks one by one @@ -551,17 +547,17 @@ void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo // Pointers to LHS blocks of block_q8_0 format const block_q8_0 * a_ptr = a_ptr_start + (y * nb); - // Take group of eight block_q4_0x8 structures at each pass of the loop and perform dot product operation + // Take group of eight blocks at each pass of the loop and perform dot product operation for (int64_t x = 0; x < nc / 8; x++) { // Pointers to RHS blocks - const block_q4_0x8 * b_ptr = b_ptr_start + (x * b_nb); + const block_tx8 * b_ptr = b_ptr_start + (x * b_nb); // Master FP accumulator __m256 acc_row = _mm256_setzero_ps(); for (int64_t b = 0; b < nb; b++) { - // Load 8 blocks of Q4_0 interleaved as 8 bytes (B0 - B7) + // Load 8 blocks of 32 interleaved as 8 bytes (B0 - B7) const __m256i rhs_raw_vec_0123_0 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs)); const __m256i rhs_raw_vec_4567_0 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs) + 1); const __m256i rhs_raw_vec_0123_1 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs) + 2); @@ -578,8 +574,13 @@ void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo const __m256i rhs_vec_0123_3 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_1, 4), m4b)); // B0(24-31) B1(24-31) B2(24-31) B3(24-31) const __m256i rhs_vec_4567_3 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_1, 4), m4b)); // B4(24-31) B5(24-31) B6(24-31) B7(24-31) - // Load the scale values for the 8 blocks interleaved in block_q4_0x8 - const __m256 col_scale_f32 = GGML_F32Cx8_REARRANGE_LOAD(b_ptr[b].d, changemask); + // Load the scale values for the 8 blocks interleaved in block_tx8 + __m256 col_scale_f32; + if constexpr ( + std::is_same_v || + std::is_same_v) { + col_scale_f32 = GGML_F32Cx8_REARRANGE_LOAD(b_ptr[b].d, changemask); + } // Load and convert to FP32 scale from block_q8_0 const __m256 row_scale_f32 = _mm256_set1_ps(GGML_CPU_FP16_TO_FP32(a_ptr[b].d)); @@ -612,578 +613,159 @@ void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo iacc = mul_sum_i8_pairs_acc_int32x8(iacc, _mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_3, 177) ,rhs_vec_4567_3, 170), _mm256_shuffle_epi32(lhs_vec_1, 255)); // Accumulated values multipled with appropriate scales - acc_row = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc), _mm256_mul_ps(col_scale_f32, row_scale_f32), acc_row); - } - - // Accumulated output values permuted so as to be stored in appropriate order post accumulation - acc_row = _mm256_permutevar8x32_ps(acc_row, finalpermutemask); - _mm256_storeu_ps(s + (y * nr + x * 8), acc_row); - } - } - return; - -#endif - ggml_gemv_q4_0_8x8_q8_0_generic(n, s, bs, vx, vy, nr, nc); -} - -void ggml_gemv_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) { - const int qk = QK_K; - const int nb = n / qk; - const int ncols_interleaved = 8; - const int blocklen = 8; - static const uint32_t kmask1 = 0x3f3f3f3f; - static const uint32_t kmask2 = 0x0f0f0f0f; - static const uint32_t kmask3 = 0x03030303; - - assert (n % qk == 0); - assert (nc % ncols_interleaved == 0); - - UNUSED(s); - UNUSED(bs); - UNUSED(vx); - UNUSED(vy); - UNUSED(nr); - UNUSED(nc); - UNUSED(nb); - UNUSED(ncols_interleaved); - UNUSED(blocklen); - -#if defined(__AVX2__) - // Lookup table to convert signed nibbles to signed bytes - __m256i signextendlut = _mm256_castsi128_si256(_mm_set_epi8(-1, -2, -3, -4, -5, -6, -7, -8, 7, 6, 5, 4, 3, 2, 1, 0)); - signextendlut = _mm256_permute2f128_si256(signextendlut, signextendlut, 0); - // Shuffle masks to rearrange delta and scale values to multiply with appropriate scales - __m128i deltamask = _mm_set_epi8(15, 14, 7, 6, 13, 12, 5, 4, 11, 10, 3, 2, 9, 8, 1, 0); - __m128i scalemask = _mm_set_epi8(7, 7, 3, 3, 6, 6, 2, 2, 5, 5, 1, 1, 4, 4, 0, 0); - // Permute mask used for easier vector processing at later stages - __m256i finalpermutemask = _mm256_set_epi32(7, 5, 3, 1, 6, 4, 2, 0); - - // Mask to extract nibbles from bytes - const __m256i m4b = _mm256_set1_epi8(0x0F); - - int64_t b_nb = n / QK_K; - - const block_q4_Kx8 * b_ptr_start = (const block_q4_Kx8 *)vx; - const block_q8_K * a_ptr_start = (const block_q8_K *)vy; - - // Process Q8_K blocks one by one - for (int64_t y = 0; y < nr; y++) { - - // Pointers to LHS blocks of block_q8_K format - const block_q8_K * a_ptr = a_ptr_start + (y * nb); - - // Take group of eight interleaved block_q4_K structures at each pass of the loop and perform dot product operation - for (int64_t x = 0; x < nc / 8; x++) { - - // Pointers to RHS blocks - const block_q4_Kx8 * b_ptr = b_ptr_start + (x * b_nb); - - // Master FP accumulators - __m256 acc_row = _mm256_setzero_ps(); - __m256 acc_min_rows = _mm256_setzero_ps(); - - for (int64_t b = 0; b < nb; b++) { - - // Load and convert to FP32 scale from block_q8_K - const __m256 row_scale_f32 = _mm256_set1_ps((a_ptr[b].d)); - - // Load the scale values for the 8 blocks interleaved in block_q4_Kx8 - // col_scale_f32 rearranged so as to multiply with appropriate quants - const __m256 col_scale_f32 = GGML_F32Cx8_REARRANGE_LOAD(b_ptr[b].d, deltamask); - const __m256 col_dmin_f32 = GGML_F32Cx8_LOAD(b_ptr[b].dmin); - - __m256i iacc_b = _mm256_setzero_si256(); - __m256i iacc_min_b = _mm256_setzero_si256(); - - const __m256i q8sums = _mm256_loadu_si256((const __m256i * )(a_ptr[b].bsums)); - __m256i q8s = _mm256_castsi128_si256(_mm_hadd_epi16(_mm256_castsi256_si128(q8sums), _mm256_extracti128_si256(q8sums, 1))); - q8s = _mm256_permute2f128_si256(q8s, q8s, 0); - - // Processes two sub blocks from each Q4_K in each iteration - for (int sb = 0; sb < QK_K / 64; sb++) { - - // Load the eight block_q4_K for two sub blocks quantized values interleaved with each other in chunks of eight - B0,B1 ....B6,B7 - const __m256i rhs_raw_vec_0123_0 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + sb * 256)); - const __m256i rhs_raw_vec_4567_0 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 32 + sb * 256)); - const __m256i rhs_raw_vec_0123_1 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 64 + sb * 256)); - const __m256i rhs_raw_vec_4567_1 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 96 + sb * 256)); - const __m256i rhs_raw_vec_0123_2 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 128 + sb * 256)); - const __m256i rhs_raw_vec_4567_2 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 160 + sb * 256)); - const __m256i rhs_raw_vec_0123_3 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 192 + sb * 256)); - const __m256i rhs_raw_vec_4567_3 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 224 + sb * 256)); - - // 4-bit -> 8-bit - // Values of the first sub block of eight block_q4_K structures for the sb loop - const __m256i rhs_vec_0123_00 = _mm256_and_si256(rhs_raw_vec_0123_0, m4b); - const __m256i rhs_vec_4567_00 = _mm256_and_si256(rhs_raw_vec_4567_0, m4b); - const __m256i rhs_vec_0123_01 = _mm256_and_si256(rhs_raw_vec_0123_1, m4b); - const __m256i rhs_vec_4567_01 = _mm256_and_si256(rhs_raw_vec_4567_1, m4b); - const __m256i rhs_vec_0123_02 = _mm256_and_si256(rhs_raw_vec_0123_2, m4b); - const __m256i rhs_vec_4567_02 = _mm256_and_si256(rhs_raw_vec_4567_2, m4b); - const __m256i rhs_vec_0123_03 = _mm256_and_si256(rhs_raw_vec_0123_3, m4b); - const __m256i rhs_vec_4567_03 = _mm256_and_si256(rhs_raw_vec_4567_3, m4b); - - // Values of the second sub block of eight block_q4_K structures when sb = 1 - const __m256i rhs_vec_0123_10 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_0, 4), m4b); - const __m256i rhs_vec_4567_10 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_0, 4), m4b); - const __m256i rhs_vec_0123_11 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_1, 4), m4b); - const __m256i rhs_vec_4567_11 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_1, 4), m4b); - const __m256i rhs_vec_0123_12 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_2, 4), m4b); - const __m256i rhs_vec_4567_12 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_2, 4), m4b); - const __m256i rhs_vec_0123_13 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_3, 4), m4b); - const __m256i rhs_vec_4567_13 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_3, 4), m4b); - - uint32_t utmp_0[4], utmp_1[4]; - - // Scales and Mins of corresponding sub blocks from different Q8_K structures are stored together - // The below block is for eg to extract first sub block's scales and mins from different Q4_K structures for the sb loop - memcpy(utmp_0, b_ptr[b].scales + 24 * sb, 12); - utmp_0[3] = ((utmp_0[2] >> 4) & kmask2) | (((utmp_0[1] >> 6) & kmask3) << 4); - const uint32_t uaux_0 = utmp_0[1] & kmask1; - utmp_0[1] = (utmp_0[2] & kmask2) | (((utmp_0[0] >> 6) & kmask3) << 4); - utmp_0[2] = uaux_0; - utmp_0[0] &= kmask1; - - // The below block is for eg to extract second sub block's scales and mins from different Q4_K structures for the sb loop - memcpy(utmp_1, b_ptr[b].scales + 12 + sb * 24, 12); - utmp_1[3] = ((utmp_1[2] >> 4) & kmask2) | (((utmp_1[1] >> 6) & kmask3) << 4); - const uint32_t uaux_1 = utmp_1[1] & kmask1; - utmp_1[1] = (utmp_1[2] & kmask2) | (((utmp_1[0] >> 6) & kmask3) << 4); - utmp_1[2] = uaux_1; - utmp_1[0] &= kmask1; - - // Scales of first sub block in the sb loop - const __m128i mins_and_scales_0 = _mm_set_epi32(utmp_0[3], utmp_0[2], utmp_0[1], utmp_0[0]); - __m128i scales_rearrange_0 = _mm_shuffle_epi8(mins_and_scales_0, scalemask); - __m256i scales_0 = _mm256_cvtepu8_epi16(scales_rearrange_0); - - // Scales of second sub block in the sb loop - __m128i mins_and_scales_1 = _mm_set_epi32(utmp_1[3], utmp_1[2], utmp_1[1], utmp_1[0]); - __m128i scales_rearrange_1 = _mm_shuffle_epi8(mins_and_scales_1, scalemask); - __m256i scales_1 = _mm256_cvtepu8_epi16(scales_rearrange_1); - - // Mins of first and second sub block of Q4_K block are arranged side by side - __m256i mins_01 = _mm256_cvtepu8_epi16(_mm_unpacklo_epi8(_mm_shuffle_epi32(mins_and_scales_0, 78), _mm_shuffle_epi32(mins_and_scales_1, 78))); - - // Load the two sub block values corresponding to sb in block_q8_K in batches of 16 bytes and replicate the same across 256 bit vector - __m256i lhs_vec_00 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + sb * 64))); - __m256i lhs_vec_01 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 16 + sb * 64))); - __m256i lhs_vec_10 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 32 + sb * 64))); - __m256i lhs_vec_11 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 48 + sb * 64))); - - lhs_vec_00 = _mm256_permute2f128_si256(lhs_vec_00, lhs_vec_00, 0); - lhs_vec_01 = _mm256_permute2f128_si256(lhs_vec_01, lhs_vec_01, 0); - lhs_vec_10 = _mm256_permute2f128_si256(lhs_vec_10, lhs_vec_10, 0); - lhs_vec_11 = _mm256_permute2f128_si256(lhs_vec_11, lhs_vec_11, 0); - - // Dot product done within 32 bit lanes and accumulated in the same vector - // First done for first sub block and thenn for second sub block in each sb - // B0(0-3) B4(0-3) B1(0-3) B5(0-3) B2(0-3) B6(0-3) B3(0-3) B7(0-3) with A0(0-3) - // B0(4-7) B4(4-7) B1(4-7) B5(4-7) B2(4-7) B6(4-7) B3(4-7) B7(4-7) with A0(4-7) - // ........................................................................... - // B0(28-31) B4(28-31) B1(28-31) B5(28-31) B2(28-31) B6(28-31) B3(28-31) B7(28-31) with A0(28-31) - - - __m256i iacc_0 = _mm256_setzero_si256(); - __m256i iacc_1 = _mm256_setzero_si256(); - - iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_00 ,_mm256_shuffle_epi32(rhs_vec_4567_00, 177), 170), _mm256_shuffle_epi32(lhs_vec_00, 0))); - iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_00, 177) ,rhs_vec_4567_00, 170), _mm256_shuffle_epi32(lhs_vec_00, 85))); - - iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_01 ,_mm256_shuffle_epi32(rhs_vec_4567_01, 177), 170), _mm256_shuffle_epi32(lhs_vec_00, 170))); - iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_01, 177) ,rhs_vec_4567_01, 170), _mm256_shuffle_epi32(lhs_vec_00, 255))); - - iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_02 ,_mm256_shuffle_epi32(rhs_vec_4567_02, 177), 170), _mm256_shuffle_epi32(lhs_vec_01, 0))); - iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_02, 177) ,rhs_vec_4567_02, 170), _mm256_shuffle_epi32(lhs_vec_01, 85))); - - iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_03 ,_mm256_shuffle_epi32(rhs_vec_4567_03, 177), 170), _mm256_shuffle_epi32(lhs_vec_01, 170))); - iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_03, 177) ,rhs_vec_4567_03, 170), _mm256_shuffle_epi32(lhs_vec_01, 255))); - - iacc_0 = _mm256_madd_epi16(iacc_0, scales_0); - - iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_10 ,_mm256_shuffle_epi32(rhs_vec_4567_10, 177), 170), _mm256_shuffle_epi32(lhs_vec_10, 0))); - iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_10, 177) ,rhs_vec_4567_10, 170), _mm256_shuffle_epi32(lhs_vec_10, 85))); - - iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_11 ,_mm256_shuffle_epi32(rhs_vec_4567_11, 177), 170), _mm256_shuffle_epi32(lhs_vec_10, 170))); - iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_11, 177) ,rhs_vec_4567_11, 170), _mm256_shuffle_epi32(lhs_vec_10, 255))); - - iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_12 ,_mm256_shuffle_epi32(rhs_vec_4567_12, 177), 170), _mm256_shuffle_epi32(lhs_vec_11, 0))); - iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_12, 177) ,rhs_vec_4567_12, 170), _mm256_shuffle_epi32(lhs_vec_11, 85))); - - iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_13 ,_mm256_shuffle_epi32(rhs_vec_4567_13, 177), 170), _mm256_shuffle_epi32(lhs_vec_11, 170))); - iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_13, 177) ,rhs_vec_4567_13, 170), _mm256_shuffle_epi32(lhs_vec_11, 255))); - - iacc_1 = _mm256_madd_epi16(iacc_1, scales_1); - - // Accumulate the iacc value for one sb - __m256i iacc_sb = _mm256_add_epi32(iacc_0, iacc_1); - - // Broadcast the bsums of the two sub blocks of the iteration of Q8_K across the vector - // Multiply-Add with corresponding mins of Q4_Kx8 with bsums - __m256i q8s_sb = _mm256_shuffle_epi32(q8s, 0); - __m256i iacc_min_sb = _mm256_madd_epi16(q8s_sb, mins_01); - q8s = _mm256_bsrli_epi128(q8s, 4); - - // Accumulate for the complete block - iacc_b = _mm256_add_epi32(iacc_b, iacc_sb); - iacc_min_b = _mm256_add_epi32(iacc_min_b, iacc_min_sb); - } - - // Multiply-Add with scale values for the complete super block - acc_row = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_b), _mm256_mul_ps(col_scale_f32, row_scale_f32), acc_row); - acc_min_rows = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_min_b), _mm256_mul_ps(col_dmin_f32, row_scale_f32), acc_min_rows); - - } - - // Accumulated output values permuted so as to be stored in appropriate order post accumulation - acc_row = _mm256_permutevar8x32_ps(acc_row, finalpermutemask); - _mm256_storeu_ps(s + (y * nr + x * 8), _mm256_sub_ps(acc_row, acc_min_rows)); - } - } - -#else - UNUSED(kmask1); - UNUSED(kmask2); - UNUSED(kmask3); - ggml_gemv_q4_K_8x8_q8_K_generic(n, s, bs, vx, vy, nr, nc); -#endif -} - -void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) { - const int qk = QK8_0; - const int nb = n / qk; - const int ncols_interleaved = 8; - const int blocklen = 8; - - assert (n % qk == 0); - assert (nr % 4 == 0); - assert (nc % ncols_interleaved == 0); - - UNUSED(s); - UNUSED(bs); - UNUSED(vx); - UNUSED(vy); - UNUSED(nr); - UNUSED(nc); - UNUSED(nb); - UNUSED(ncols_interleaved); - UNUSED(blocklen); - -#if defined(__AVX2__) || defined(__AVX512F__) - { - const block_q4_0x8 * b_ptr_start = (const block_q4_0x8 *)vx; - const block_q8_0x4 * a_ptr_start = (const block_q8_0x4 *)vy; - int64_t b_nb = n / QK4_0; - int64_t y = 0; - // Mask to mask out nibbles from packed bytes - const __m256i m4b = _mm256_set1_epi8(0x0F); - const __m128i loadMask = _mm_blend_epi32(_mm_setzero_si128(), _mm_set1_epi32(0xFFFFFFFF), 3); - // Lookup table to convert signed nibbles to signed bytes - __m256i signextendlut = _mm256_castsi128_si256(_mm_set_epi8(-1, -2, -3, -4, -5, -6, -7, -8, 7, 6, 5, 4, 3, 2, 1, 0)); - signextendlut = _mm256_permute2f128_si256(signextendlut, signextendlut, 0); - // Permute mask used for easier vector processing at later stages - __m256i requiredOrder = _mm256_set_epi32(3, 2, 1, 0, 7, 6, 5, 4); - int64_t xstart = 0; - int anr = nr - nr%16; // Used to align nr with boundary of 16 - #ifdef __AVX512F__ - int anc = nc - nc%16; // Used to align nc with boundary of 16 - // Mask to mask out nibbles from packed bytes expanded to 512 bit length - const __m512i m4bexpanded = _mm512_set1_epi8(0x0F); - // Lookup table to convert signed nibbles to signed bytes expanded to 512 bit length - __m512i signextendlutexpanded = _mm512_inserti32x8(_mm512_castsi256_si512(signextendlut), signextendlut, 1); - - // Take group of four block_q8_0x4 structures at each pass of the loop and perform dot product operation - for (; y < anr / 4; y += 4) { - - const block_q8_0x4 * a_ptrs[4]; - - a_ptrs[0] = a_ptr_start + (y * nb); - for (int i = 0; i < 3; ++i) { - a_ptrs[i + 1] = a_ptrs[i] + nb; - } - - // Take group of two block_q4_0x8 structures at each pass of the loop and perform dot product operation - for (int64_t x = 0; x < anc / 8; x += 2) { - - const block_q4_0x8 * b_ptr_0 = b_ptr_start + ((x) * b_nb); - const block_q4_0x8 * b_ptr_1 = b_ptr_start + ((x + 1) * b_nb); - - // Master FP accumulators - __m512 acc_rows[16]; - for (int i = 0; i < 16; i++) { - acc_rows[i] = _mm512_setzero_ps(); - } - - for (int64_t b = 0; b < nb; b++) { - // Load the sixteen block_q4_0 quantized values interleaved with each other in chunks of eight - B0,B1 ....BE,BF - const __m256i rhs_raw_mat_0123_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs)); - const __m256i rhs_raw_mat_4567_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs + 32)); - const __m256i rhs_raw_mat_0123_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs + 64)); - const __m256i rhs_raw_mat_4567_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs + 96)); - - const __m256i rhs_raw_mat_89AB_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs)); - const __m256i rhs_raw_mat_CDEF_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs + 32)); - const __m256i rhs_raw_mat_89AB_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs + 64)); - const __m256i rhs_raw_mat_CDEF_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs + 96)); - - // Save the values in the following vectors in the formats B0B1B4B5B8B9BCBD, B2B3B6B7BABBBEBF for further processing and storing of values - const __m256i rhs_raw_mat_0145_0 = _mm256_blend_epi32(rhs_raw_mat_0123_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_0, requiredOrder), 240); - const __m256i rhs_raw_mat_2367_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_0, requiredOrder), rhs_raw_mat_4567_0, 240); - const __m256i rhs_raw_mat_0145_1 = _mm256_blend_epi32(rhs_raw_mat_0123_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_1, requiredOrder), 240); - const __m256i rhs_raw_mat_2367_1 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_1, requiredOrder), rhs_raw_mat_4567_1, 240); - - const __m256i rhs_raw_mat_89CD_0 = _mm256_blend_epi32(rhs_raw_mat_89AB_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_CDEF_0, requiredOrder), 240); - const __m256i rhs_raw_mat_ABEF_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_89AB_0, requiredOrder), rhs_raw_mat_CDEF_0, 240); - const __m256i rhs_raw_mat_89CD_1 = _mm256_blend_epi32(rhs_raw_mat_89AB_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_CDEF_1, requiredOrder), 240); - const __m256i rhs_raw_mat_ABEF_1 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_89AB_1, requiredOrder), rhs_raw_mat_CDEF_1, 240); - - const __m512i rhs_raw_mat_014589CD_0 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_0), rhs_raw_mat_89CD_0, 1); - const __m512i rhs_raw_mat_2367ABEF_0 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_0), rhs_raw_mat_ABEF_0, 1); - const __m512i rhs_raw_mat_014589CD_1 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_1), rhs_raw_mat_89CD_1, 1); - const __m512i rhs_raw_mat_2367ABEF_1 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_1), rhs_raw_mat_ABEF_1, 1); - - // 4-bit -> 8-bit - Sign is maintained - const __m512i rhs_mat_014589CD_0 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_014589CD_0, m4bexpanded)); //B0(0-7) B1(0-7) B4(0-7) B5(0-7) B8(0-7) B9(0-7) BC(0-7) BD(0-7) - const __m512i rhs_mat_2367ABEF_0 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_2367ABEF_0, m4bexpanded)); //B2(0-7) B3(0-7) B6(0-7) B7(0-7) BA(0-7) BB(0-7) BE(0-7) BF(0-7) - - const __m512i rhs_mat_014589CD_1 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_014589CD_1, m4bexpanded)); //B0(8-15) B1(8-15) B4(8-15) B5(8-15) B8(8-15) B9(8-15) BC(8-15) BD(8-15) - const __m512i rhs_mat_2367ABEF_1 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_2367ABEF_1, m4bexpanded)); //B2(8-15) B3(8-15) B6(8-15) B7(8-15) BA(8-15) BB(8-15) BE(8-15) BF(8-15) - - const __m512i rhs_mat_014589CD_2 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_0, 4), m4bexpanded)); //B0(16-23) B1(16-23) B4(16-23) B5(16-23) B8(16-23) B9(16-23) BC(16-23) BD(16-23) - const __m512i rhs_mat_2367ABEF_2 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_0, 4), m4bexpanded)); //B2(16-23) B3(16-23) B6(16-23) B7(16-23) BA(16-23) BB(16-23) BE(16-23) BF(16-23) - - const __m512i rhs_mat_014589CD_3 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_1, 4), m4bexpanded)); //B0(24-31) B1(24-31) B4(24-31) B5(24-31) B8(24-31) B9(24-31) BC(24-31) BD(24-31) - const __m512i rhs_mat_2367ABEF_3 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_1, 4), m4bexpanded)); //B2(24-31) B3(24-31) B6(24-31) B7(24-31) BA(24-31) BB(24-31) BE(24-31) BF(24-31) - - // Shuffle pattern one - right side input - const __m512i rhs_mat_014589CD_0_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_0, (_MM_PERM_ENUM)136); //B0(0-3) B1(0-3) B0(0-3) B1(0-3) B4(0-3) B5(0-3) B4(0-3) B5(0-3) B8(0-3) B9(0-3) B8(0-3) B9(0-3) BC(0-3) BD(0-3) BC(0-3) BD(0-3) - const __m512i rhs_mat_2367ABEF_0_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_0, (_MM_PERM_ENUM)136); //B2(0-3) B3(0-3) B2(0-3) B3(0-3) B6(0-3) B7(0-3) B6(0-3) B7(0-3) BA(0-3) BB(0-3) BA(0-3) BB(0-3) BE(0-3) BF(0-3) BE(0-3) BF(0-3) - - const __m512i rhs_mat_014589CD_1_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_1, (_MM_PERM_ENUM)136); //B0(8-11) B1(8-11) B0(8-11) B1(8-11) B4(8-11) B5(8-11) B4(8-11) B5(8-11) B8(8-11) B9(8-11) B8(8-11) B9(8-11) BC(8-11) BD(8-11) BC(8-11) BD(8-11) - const __m512i rhs_mat_2367ABEF_1_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_1, (_MM_PERM_ENUM)136); //B2(8-11) B3(8-11) B2(8-11) B3(8-11) B6(8-11) B7(8-11) B6(8-11) B7(8-11) BA(8-11) BB(8-11) BA(8-11) BB(8-11) BE(8-11) BF(8-11) BE(8-11) BF(8-11) - - const __m512i rhs_mat_014589CD_2_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_2, (_MM_PERM_ENUM)136); //B0(16-19) B1(16-19) B0(16-19) B1(16-19) B4(16-19) B5(16-19) B4(16-19) B5(16-19) B8(16-19) B9(16-19) B8(16-19) B9(16-19) BC(16-19) BD(16-19) BC(16-19) BD(16-19) - const __m512i rhs_mat_2367ABEF_2_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_2, (_MM_PERM_ENUM)136); //B2(16-19) B3(16-19) B2(16-19) B3(16-19) B6(16-19) B7(16-19) B6(16-19) B7(16-19) BA(16-19) BB(16-19) BA(16-19) BB(16-19) BE(16-19) BF(16-19) BE(16-19) BF(16-19) - - const __m512i rhs_mat_014589CD_3_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_3, (_MM_PERM_ENUM)136); //B0(24-27) B1(24-27) B0(24-27) B1(24-27) B4(24-27) B5(24-27) B4(24-27) B5(24-27) B8(24-27) B9(24-27) B8(24-27) B9(24-27) BC(24-27) BD(24-27) BC(24-27) BD(24-27) - const __m512i rhs_mat_2367ABEF_3_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_3, (_MM_PERM_ENUM)136); //B2(24-27) B3(24-27) B2(24-27) B3(24-27) B6(24-27) B7(24-27) B6(24-27) B7(24-27) BA(24-27) BB(24-27) BA(24-27) BB(24-27) BE(24-27) BF(24-27) BE(24-27) BF(24-27) - - // Shuffle pattern two - right side input - - const __m512i rhs_mat_014589CD_0_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_0, (_MM_PERM_ENUM)221); //B0(4-7) B1(4-7) B0(4-7) B1(4-7) B4(4-7) B5(4-7) B4(4-7) B5(4-7) B8(4-7) B9(4-7) B8(4-7) B9(4-7) BC(4-7) BD(4-7) BC(4-7) BD(4-7) - const __m512i rhs_mat_2367ABEF_0_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_0, (_MM_PERM_ENUM)221); //B2(4-7) B3(4-7) B2(4-7) B3(4-7) B6(4-7) B7(4-7) B6(4-7) B7(4-7) BA(4-7) BB(4-7) BA(4-7) BB(4-7) BE(4-7) BF(4-7) BE(4-7) BF(4-7) - - const __m512i rhs_mat_014589CD_1_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_1, (_MM_PERM_ENUM)221); //B0(12-15) B1(12-15) B0(12-15) B1(12-15) B4(12-15) B5(12-15) B4(12-15) B5(12-15) B8(12-15) B9(12-15) B8(12-15) B9(12-15) BC(12-15) BD(12-15) BC(12-15) BD(12-15) - const __m512i rhs_mat_2367ABEF_1_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_1, (_MM_PERM_ENUM)221); //B2(12-15) B3(12-15) B2(12-15) B3(12-15) B6(12-15) B7(12-15) B6(12-15) B7(12-15) BA(12-15) BB(12-15) BA(12-15) BB(12-15) BE(12-15) BF(12-15) BE(12-15) BF(12-15) - - const __m512i rhs_mat_014589CD_2_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_2, (_MM_PERM_ENUM)221); //B0(20-23) B1(20-23) B0(20-23) B1(20-23) B4(20-23) B5(20-23) B4(20-23) B5(20-23) B8(20-23) B9(20-23) B8(20-23) B9(20-23) BC(20-23) BD(20-23) BC(20-23) BD(20-23) - const __m512i rhs_mat_2367ABEF_2_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_2, (_MM_PERM_ENUM)221); //B2(20-23) B3(20-23) B2(20-23) B3(20-23) B6(20-23) B7(20-23) B6(20-23) B7(20-23) BA(20-23) BB(20-23) BA(20-23) BB(20-23) BE(20-23) BF(20-23) BE(20-23) BF(20-23) - - const __m512i rhs_mat_014589CD_3_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_3, (_MM_PERM_ENUM)221); //B0(28-31) B1(28-31) B0(28-31) B1(28-31) B4(28-31) B5(28-31) B4(28-31) B5(28-31) B8(28-31) B9(28-31) B8(28-31) B9(28-31) BC(28-31) BD(28-31) BC(28-31) BD(28-31) - const __m512i rhs_mat_2367ABEF_3_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_3, (_MM_PERM_ENUM)221); //B2(28-31) B3(28-31) B2(28-31) B3(28-31) B6(28-31) B7(28-31) B6(28-31) B7(28-31) BA(28-31) BB(28-31) BA(28-31) BB(28-31) BE(28-31) BF(28-31) BE(28-31) BF(28-31) - - // Scale values - Load the weight scale values of two block_q4_0x8 - const __m512 col_scale_f32 = GGML_F32Cx8x2_LOAD(b_ptr_0[b].d, b_ptr_1[b].d); - - // Process LHS in pairs of rows - for (int rp = 0; rp < 4; rp++) { - - // Load the four block_q4_0 quantized values interleaved with each other in chunks of eight - A0,A1,A2,A3 - // Loaded as set of 128 bit vectors and repeated and stored into a 256 bit vector before again repeating into 512 bit vector - __m256i lhs_mat_ymm_0123_0 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs))); - __m256i lhs_mat_ymm_01_0 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_0, lhs_mat_ymm_0123_0, 0); - __m256i lhs_mat_ymm_23_0 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_0, lhs_mat_ymm_0123_0, 17); - __m256i lhs_mat_ymm_0123_1 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs + 32))); - __m256i lhs_mat_ymm_01_1 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_1, lhs_mat_ymm_0123_1, 0); - __m256i lhs_mat_ymm_23_1 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_1, lhs_mat_ymm_0123_1, 17); - __m256i lhs_mat_ymm_0123_2 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs + 64))); - __m256i lhs_mat_ymm_01_2 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_2, lhs_mat_ymm_0123_2, 0); - __m256i lhs_mat_ymm_23_2 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_2, lhs_mat_ymm_0123_2, 17); - __m256i lhs_mat_ymm_0123_3 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs + 96))); - __m256i lhs_mat_ymm_01_3 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_3, lhs_mat_ymm_0123_3, 0); - __m256i lhs_mat_ymm_23_3 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_3, lhs_mat_ymm_0123_3, 17); - - __m512i lhs_mat_01_0 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_0), lhs_mat_ymm_01_0, 1); - __m512i lhs_mat_23_0 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_0), lhs_mat_ymm_23_0, 1); - __m512i lhs_mat_01_1 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_1), lhs_mat_ymm_01_1, 1); - __m512i lhs_mat_23_1 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_1), lhs_mat_ymm_23_1, 1); - __m512i lhs_mat_01_2 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_2), lhs_mat_ymm_01_2, 1); - __m512i lhs_mat_23_2 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_2), lhs_mat_ymm_23_2, 1); - __m512i lhs_mat_01_3 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_3), lhs_mat_ymm_01_3, 1); - __m512i lhs_mat_23_3 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_3), lhs_mat_ymm_23_3, 1); - - // Shuffle pattern one - left side input - - const __m512i lhs_mat_01_0_sp1 = _mm512_shuffle_epi32(lhs_mat_01_0, (_MM_PERM_ENUM)160); //A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3) - const __m512i lhs_mat_23_0_sp1 = _mm512_shuffle_epi32(lhs_mat_23_0, (_MM_PERM_ENUM)160); //A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3) - - const __m512i lhs_mat_01_1_sp1 = _mm512_shuffle_epi32(lhs_mat_01_1, (_MM_PERM_ENUM)160); //A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11) - const __m512i lhs_mat_23_1_sp1 = _mm512_shuffle_epi32(lhs_mat_23_1, (_MM_PERM_ENUM)160); //A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11) - - const __m512i lhs_mat_01_2_sp1 = _mm512_shuffle_epi32(lhs_mat_01_2, (_MM_PERM_ENUM)160); //A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19) - const __m512i lhs_mat_23_2_sp1 = _mm512_shuffle_epi32(lhs_mat_23_2, (_MM_PERM_ENUM)160); //A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19) - - const __m512i lhs_mat_01_3_sp1 = _mm512_shuffle_epi32(lhs_mat_01_3, (_MM_PERM_ENUM)160); //A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27) - const __m512i lhs_mat_23_3_sp1 = _mm512_shuffle_epi32(lhs_mat_23_3, (_MM_PERM_ENUM)160); //A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27) - - // Shuffle pattern two - left side input - - const __m512i lhs_mat_01_0_sp2 = _mm512_shuffle_epi32(lhs_mat_01_0, (_MM_PERM_ENUM)245); //A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7) - const __m512i lhs_mat_23_0_sp2 = _mm512_shuffle_epi32(lhs_mat_23_0, (_MM_PERM_ENUM)245); //A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7) - - const __m512i lhs_mat_01_1_sp2 = _mm512_shuffle_epi32(lhs_mat_01_1, (_MM_PERM_ENUM)245); //A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15) - const __m512i lhs_mat_23_1_sp2 = _mm512_shuffle_epi32(lhs_mat_23_1, (_MM_PERM_ENUM)245); //A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15) - - const __m512i lhs_mat_01_2_sp2 = _mm512_shuffle_epi32(lhs_mat_01_2, (_MM_PERM_ENUM)245); //A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23) - const __m512i lhs_mat_23_2_sp2 = _mm512_shuffle_epi32(lhs_mat_23_2, (_MM_PERM_ENUM)245); //A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23) - - const __m512i lhs_mat_01_3_sp2 = _mm512_shuffle_epi32(lhs_mat_01_3, (_MM_PERM_ENUM)245); //A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31) - const __m512i lhs_mat_23_3_sp2 = _mm512_shuffle_epi32(lhs_mat_23_3, (_MM_PERM_ENUM)245); //A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31) - - // The values arranged in shuffle patterns are operated with dot product operation within 32 bit lane i.e corresponding bytes and multiplied and added into 32 bit integers within 32 bit lane - // Resembles MMLAs into 2x2 matrices in ARM Version - const __m512i zero = _mm512_setzero_epi32(); - __m512i iacc_mat_00_sp1 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_01_3_sp1, rhs_mat_014589CD_3_sp1), lhs_mat_01_2_sp1, rhs_mat_014589CD_2_sp1), lhs_mat_01_1_sp1, rhs_mat_014589CD_1_sp1), lhs_mat_01_0_sp1, rhs_mat_014589CD_0_sp1); - __m512i iacc_mat_01_sp1 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_01_3_sp1, rhs_mat_2367ABEF_3_sp1), lhs_mat_01_2_sp1, rhs_mat_2367ABEF_2_sp1), lhs_mat_01_1_sp1, rhs_mat_2367ABEF_1_sp1), lhs_mat_01_0_sp1, rhs_mat_2367ABEF_0_sp1); - __m512i iacc_mat_10_sp1 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_23_3_sp1, rhs_mat_014589CD_3_sp1), lhs_mat_23_2_sp1, rhs_mat_014589CD_2_sp1), lhs_mat_23_1_sp1, rhs_mat_014589CD_1_sp1), lhs_mat_23_0_sp1, rhs_mat_014589CD_0_sp1); - __m512i iacc_mat_11_sp1 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_23_3_sp1, rhs_mat_2367ABEF_3_sp1), lhs_mat_23_2_sp1, rhs_mat_2367ABEF_2_sp1), lhs_mat_23_1_sp1, rhs_mat_2367ABEF_1_sp1), lhs_mat_23_0_sp1, rhs_mat_2367ABEF_0_sp1); - __m512i iacc_mat_00_sp2 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_01_3_sp2, rhs_mat_014589CD_3_sp2), lhs_mat_01_2_sp2, rhs_mat_014589CD_2_sp2), lhs_mat_01_1_sp2, rhs_mat_014589CD_1_sp2), lhs_mat_01_0_sp2, rhs_mat_014589CD_0_sp2); - __m512i iacc_mat_01_sp2 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_01_3_sp2, rhs_mat_2367ABEF_3_sp2), lhs_mat_01_2_sp2, rhs_mat_2367ABEF_2_sp2), lhs_mat_01_1_sp2, rhs_mat_2367ABEF_1_sp2), lhs_mat_01_0_sp2, rhs_mat_2367ABEF_0_sp2); - __m512i iacc_mat_10_sp2 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_23_3_sp2, rhs_mat_014589CD_3_sp2), lhs_mat_23_2_sp2, rhs_mat_014589CD_2_sp2), lhs_mat_23_1_sp2, rhs_mat_014589CD_1_sp2), lhs_mat_23_0_sp2, rhs_mat_014589CD_0_sp2); - __m512i iacc_mat_11_sp2 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_23_3_sp2, rhs_mat_2367ABEF_3_sp2), lhs_mat_23_2_sp2, rhs_mat_2367ABEF_2_sp2), lhs_mat_23_1_sp2, rhs_mat_2367ABEF_1_sp2), lhs_mat_23_0_sp2, rhs_mat_2367ABEF_0_sp2); - - // Output of both shuffle patterns are added in order to sum dot product outputs of all 32 values in block - __m512i iacc_mat_00 = _mm512_add_epi32(iacc_mat_00_sp1, iacc_mat_00_sp2); - __m512i iacc_mat_01 = _mm512_add_epi32(iacc_mat_01_sp1, iacc_mat_01_sp2); - __m512i iacc_mat_10 = _mm512_add_epi32(iacc_mat_10_sp1, iacc_mat_10_sp2); - __m512i iacc_mat_11 = _mm512_add_epi32(iacc_mat_11_sp1, iacc_mat_11_sp2); - - - // Straighten out to make 4 row vectors - __m512i iacc_row_0 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_00, _mm512_shuffle_epi32(iacc_mat_01, (_MM_PERM_ENUM)78)); - __m512i iacc_row_1 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_00, (_MM_PERM_ENUM)78), iacc_mat_01); - __m512i iacc_row_2 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_10, _mm512_shuffle_epi32(iacc_mat_11, (_MM_PERM_ENUM)78)); - __m512i iacc_row_3 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_10, (_MM_PERM_ENUM)78), iacc_mat_11); - - // Load the scale(d) values for all the 4 Q8_0 blocks and repeat it across lanes - const __m128i row_scale_f16 = _mm_shuffle_epi32(_mm_maskload_epi32((int const*)(a_ptrs[rp][b].d), loadMask), 68); - const __m512 row_scale_f32 = GGML_F32Cx16_REPEAT_LOAD(row_scale_f16); - - // Multiply with appropiate scales and accumulate - acc_rows[rp * 4] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_0), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 0)), acc_rows[rp * 4]); - acc_rows[rp * 4 + 1] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_1), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 85)), acc_rows[rp * 4 + 1]); - acc_rows[rp * 4 + 2] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_2), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_rows[rp * 4 + 2]); - acc_rows[rp * 4 + 3] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_3), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 255)), acc_rows[rp * 4 + 3]); - } - } - - // Store the accumulated values - for (int i = 0; i < 16; i++) { - _mm512_storeu_ps((float *)(s + ((y * 4 + i) * bs + x * 8)), acc_rows[i]); - } + acc_row = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc), _mm256_mul_ps(col_scale_f32, row_scale_f32), acc_row); } + + // Accumulated output values permuted so as to be stored in appropriate order post accumulation + acc_row = _mm256_permutevar8x32_ps(acc_row, finalpermutemask); + _mm256_storeu_ps(s + (y * nr + x * 8), acc_row); } - // Take a block_q8_0x4 structures at each pass of the loop and perform dot product operation - for (; y < nr / 4; y ++) { + } +} - const block_q8_0x4 * a_ptr = a_ptr_start + (y * nb); +// GEMM for 8x blocks of 32 4-bit quants with a single scale factor per block +template +static void gemm_q4_b32_8x8_q8_0_lut_avx(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc, __m256i signextendlut) { + static_assert( + std::is_same_v || + std::is_same_v, + "Unsupported block type"); - // Take group of two block_q4_0x8 structures at each pass of the loop and perform dot product operation - for (int64_t x = 0; x < anc / 8; x += 2) { + const int qk = QK8_0; + const int nb = n / qk; - const block_q4_0x8 * b_ptr_0 = b_ptr_start + ((x) * b_nb); - const block_q4_0x8 * b_ptr_1 = b_ptr_start + ((x + 1) * b_nb); + const block_tx8 * b_ptr_start = (const block_tx8 *)vx; + const block_q8_0x4 * a_ptr_start = (const block_q8_0x4 *)vy; - // Master FP accumulators - __m512 acc_rows[4]; - for (int i = 0; i < 4; i++) { - acc_rows[i] = _mm512_setzero_ps(); - } + int64_t b_nb = n / 32; + int64_t y = 0; + // Mask to mask out nibbles from packed bytes + const __m256i m4b = _mm256_set1_epi8(0x0F); + const __m128i loadMask = _mm_blend_epi32(_mm_setzero_si128(), _mm_set1_epi32(0xFFFFFFFF), 3); + // Permute mask used for easier vector processing at later stages + __m256i requiredOrder = _mm256_set_epi32(3, 2, 1, 0, 7, 6, 5, 4); + int64_t xstart = 0; + int anr = nr - nr%16; // Used to align nr with boundary of 16 +#ifdef __AVX512F__ + int anc = nc - nc%16; // Used to align nc with boundary of 16 + // Mask to mask out nibbles from packed bytes expanded to 512 bit length + const __m512i m4bexpanded = _mm512_set1_epi8(0x0F); + // Lookup table to convert signed nibbles to signed bytes expanded to 512 bit length + __m512i signextendlutexpanded = _mm512_inserti32x8(_mm512_castsi256_si512(signextendlut), signextendlut, 1); - for (int64_t b = 0; b < nb; b++) { - // Load the sixteen block_q4_0 quantized values interleaved with each other in chunks of eight - B0,B1 ....BE,BF - const __m256i rhs_raw_mat_0123_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs)); - const __m256i rhs_raw_mat_4567_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs + 32)); - const __m256i rhs_raw_mat_0123_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs + 64)); - const __m256i rhs_raw_mat_4567_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs + 96)); + // Take group of four block_q8_0x4 structures at each pass of the loop and perform dot product operation + for (; y < anr / 4; y += 4) { - const __m256i rhs_raw_mat_89AB_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs)); - const __m256i rhs_raw_mat_CDEF_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs + 32)); - const __m256i rhs_raw_mat_89AB_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs + 64)); - const __m256i rhs_raw_mat_CDEF_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs + 96)); + const block_q8_0x4 * a_ptrs[4]; - // Save the values in the following vectors in the formats B0B1B4B5, B2B3B6B7 for further processing and storing of valuess - const __m256i rhs_raw_mat_0145_0 = _mm256_blend_epi32(rhs_raw_mat_0123_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_0, requiredOrder), 240); - const __m256i rhs_raw_mat_2367_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_0, requiredOrder), rhs_raw_mat_4567_0, 240); - const __m256i rhs_raw_mat_0145_1 = _mm256_blend_epi32(rhs_raw_mat_0123_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_1, requiredOrder), 240); - const __m256i rhs_raw_mat_2367_1 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_1, requiredOrder), rhs_raw_mat_4567_1, 240); + a_ptrs[0] = a_ptr_start + (y * nb); + for (int i = 0; i < 3; ++i) { + a_ptrs[i + 1] = a_ptrs[i] + nb; + } - const __m256i rhs_raw_mat_89CD_0 = _mm256_blend_epi32(rhs_raw_mat_89AB_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_CDEF_0, requiredOrder), 240); - const __m256i rhs_raw_mat_ABEF_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_89AB_0, requiredOrder), rhs_raw_mat_CDEF_0, 240); - const __m256i rhs_raw_mat_89CD_1 = _mm256_blend_epi32(rhs_raw_mat_89AB_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_CDEF_1, requiredOrder), 240); - const __m256i rhs_raw_mat_ABEF_1 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_89AB_1, requiredOrder), rhs_raw_mat_CDEF_1, 240); + // Take group of two block_tx8 structures at each pass of the loop and perform dot product operation + for (int64_t x = 0; x < anc / 8; x += 2) { - const __m512i rhs_raw_mat_014589CD_0 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_0), rhs_raw_mat_89CD_0, 1); - const __m512i rhs_raw_mat_2367ABEF_0 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_0), rhs_raw_mat_ABEF_0, 1); - const __m512i rhs_raw_mat_014589CD_1 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_1), rhs_raw_mat_89CD_1, 1); - const __m512i rhs_raw_mat_2367ABEF_1 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_1), rhs_raw_mat_ABEF_1, 1); + const block_tx8 * b_ptr_0 = b_ptr_start + ((x) * b_nb); + const block_tx8 * b_ptr_1 = b_ptr_start + ((x + 1) * b_nb); + + // Master FP accumulators + __m512 acc_rows[16]; + for (int i = 0; i < 16; i++) { + acc_rows[i] = _mm512_setzero_ps(); + } + + for (int64_t b = 0; b < nb; b++) { + // Load the sixteen blocks of quantized values interleaved with each other in chunks of eight - B0,B1 ....BE,BF + const __m256i rhs_raw_mat_0123_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs)); + const __m256i rhs_raw_mat_4567_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs + 32)); + const __m256i rhs_raw_mat_0123_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs + 64)); + const __m256i rhs_raw_mat_4567_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs + 96)); + + const __m256i rhs_raw_mat_89AB_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs)); + const __m256i rhs_raw_mat_CDEF_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs + 32)); + const __m256i rhs_raw_mat_89AB_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs + 64)); + const __m256i rhs_raw_mat_CDEF_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs + 96)); + + // Save the values in the following vectors in the formats B0B1B4B5B8B9BCBD, B2B3B6B7BABBBEBF for further processing and storing of values + const __m256i rhs_raw_mat_0145_0 = _mm256_blend_epi32(rhs_raw_mat_0123_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_0, requiredOrder), 240); + const __m256i rhs_raw_mat_2367_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_0, requiredOrder), rhs_raw_mat_4567_0, 240); + const __m256i rhs_raw_mat_0145_1 = _mm256_blend_epi32(rhs_raw_mat_0123_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_1, requiredOrder), 240); + const __m256i rhs_raw_mat_2367_1 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_1, requiredOrder), rhs_raw_mat_4567_1, 240); + + const __m256i rhs_raw_mat_89CD_0 = _mm256_blend_epi32(rhs_raw_mat_89AB_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_CDEF_0, requiredOrder), 240); + const __m256i rhs_raw_mat_ABEF_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_89AB_0, requiredOrder), rhs_raw_mat_CDEF_0, 240); + const __m256i rhs_raw_mat_89CD_1 = _mm256_blend_epi32(rhs_raw_mat_89AB_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_CDEF_1, requiredOrder), 240); + const __m256i rhs_raw_mat_ABEF_1 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_89AB_1, requiredOrder), rhs_raw_mat_CDEF_1, 240); + + const __m512i rhs_raw_mat_014589CD_0 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_0), rhs_raw_mat_89CD_0, 1); + const __m512i rhs_raw_mat_2367ABEF_0 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_0), rhs_raw_mat_ABEF_0, 1); + const __m512i rhs_raw_mat_014589CD_1 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_1), rhs_raw_mat_89CD_1, 1); + const __m512i rhs_raw_mat_2367ABEF_1 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_1), rhs_raw_mat_ABEF_1, 1); - // 4-bit -> 8-bit - Sign is maintained - const __m512i rhs_mat_014589CD_0 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_014589CD_0, m4bexpanded)); //B0(0-7) B1(0-7) B4(0-7) B5(0-7) B8(0-7) B9(0-7) BC(0-7) BD(0-7) - const __m512i rhs_mat_2367ABEF_0 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_2367ABEF_0, m4bexpanded)); //B2(0-7) B3(0-7) B6(0-7) B7(0-7) BA(0-7) BB(0-7) BE(0-7) BF(0-7) + // 4-bit -> 8-bit - Sign is maintained + const __m512i rhs_mat_014589CD_0 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_014589CD_0, m4bexpanded)); //B0(0-7) B1(0-7) B4(0-7) B5(0-7) B8(0-7) B9(0-7) BC(0-7) BD(0-7) + const __m512i rhs_mat_2367ABEF_0 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_2367ABEF_0, m4bexpanded)); //B2(0-7) B3(0-7) B6(0-7) B7(0-7) BA(0-7) BB(0-7) BE(0-7) BF(0-7) - const __m512i rhs_mat_014589CD_1 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_014589CD_1, m4bexpanded)); //B0(8-15) B1(8-15) B4(8-15) B5(8-15) B8(8-15) B9(8-15) BC(8-15) BD(8-15) - const __m512i rhs_mat_2367ABEF_1 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_2367ABEF_1, m4bexpanded)); //B2(8-15) B3(8-15) B6(8-15) B7(8-15) BA(8-15) BB(8-15) BE(8-15) BF(8-15) + const __m512i rhs_mat_014589CD_1 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_014589CD_1, m4bexpanded)); //B0(8-15) B1(8-15) B4(8-15) B5(8-15) B8(8-15) B9(8-15) BC(8-15) BD(8-15) + const __m512i rhs_mat_2367ABEF_1 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_2367ABEF_1, m4bexpanded)); //B2(8-15) B3(8-15) B6(8-15) B7(8-15) BA(8-15) BB(8-15) BE(8-15) BF(8-15) - const __m512i rhs_mat_014589CD_2 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_0, 4), m4bexpanded)); //B0(16-23) B1(16-23) B4(16-23) B5(16-23) B8(16-23) B9(16-23) BC(16-23) BD(16-23) - const __m512i rhs_mat_2367ABEF_2 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_0, 4), m4bexpanded)); //B2(16-23) B3(16-23) B6(16-23) B7(16-23) BA(16-23) BB(16-23) BE(16-23) BF(16-23) + const __m512i rhs_mat_014589CD_2 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_0, 4), m4bexpanded)); //B0(16-23) B1(16-23) B4(16-23) B5(16-23) B8(16-23) B9(16-23) BC(16-23) BD(16-23) + const __m512i rhs_mat_2367ABEF_2 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_0, 4), m4bexpanded)); //B2(16-23) B3(16-23) B6(16-23) B7(16-23) BA(16-23) BB(16-23) BE(16-23) BF(16-23) - const __m512i rhs_mat_014589CD_3 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_1, 4), m4bexpanded)); //B0(24-31) B1(24-31) B4(24-31) B5(24-31) B8(24-31) B9(24-31) BC(24-31) BD(24-31) - const __m512i rhs_mat_2367ABEF_3 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_1, 4), m4bexpanded)); //B2(24-31) B3(24-31) B6(24-31) B7(24-31) BA(24-31) BB(24-31) BE(24-31) BF(24-31) + const __m512i rhs_mat_014589CD_3 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_1, 4), m4bexpanded)); //B0(24-31) B1(24-31) B4(24-31) B5(24-31) B8(24-31) B9(24-31) BC(24-31) BD(24-31) + const __m512i rhs_mat_2367ABEF_3 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_1, 4), m4bexpanded)); //B2(24-31) B3(24-31) B6(24-31) B7(24-31) BA(24-31) BB(24-31) BE(24-31) BF(24-31) - // Shuffle pattern one - right side input - const __m512i rhs_mat_014589CD_0_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_0, (_MM_PERM_ENUM)136); //B0(0-3) B1(0-3) B0(0-3) B1(0-3) B4(0-3) B5(0-3) B4(0-3) B5(0-3) B8(0-3) B9(0-3) B8(0-3) B9(0-3) BC(0-3) BD(0-3) BC(0-3) BD(0-3) - const __m512i rhs_mat_2367ABEF_0_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_0, (_MM_PERM_ENUM)136); //B2(0-3) B3(0-3) B2(0-3) B3(0-3) B6(0-3) B7(0-3) B6(0-3) B7(0-3) BA(0-3) BB(0-3) BA(0-3) BB(0-3) BE(0-3) BF(0-3) BE(0-3) BF(0-3) + // Shuffle pattern one - right side input + const __m512i rhs_mat_014589CD_0_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_0, (_MM_PERM_ENUM)136); //B0(0-3) B1(0-3) B0(0-3) B1(0-3) B4(0-3) B5(0-3) B4(0-3) B5(0-3) B8(0-3) B9(0-3) B8(0-3) B9(0-3) BC(0-3) BD(0-3) BC(0-3) BD(0-3) + const __m512i rhs_mat_2367ABEF_0_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_0, (_MM_PERM_ENUM)136); //B2(0-3) B3(0-3) B2(0-3) B3(0-3) B6(0-3) B7(0-3) B6(0-3) B7(0-3) BA(0-3) BB(0-3) BA(0-3) BB(0-3) BE(0-3) BF(0-3) BE(0-3) BF(0-3) - const __m512i rhs_mat_014589CD_1_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_1, (_MM_PERM_ENUM)136); //B0(8-11) B1(8-11) B0(8-11) B1(8-11) B4(8-11) B5(8-11) B4(8-11) B5(8-11) B8(8-11) B9(8-11) B8(8-11) B9(8-11) BC(8-11) BD(8-11) BC(8-11) BD(8-11) - const __m512i rhs_mat_2367ABEF_1_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_1, (_MM_PERM_ENUM)136); //B2(8-11) B3(8-11) B2(8-11) B3(8-11) B6(8-11) B7(8-11) B6(8-11) B7(8-11) BA(8-11) BB(8-11) BA(8-11) BB(8-11) BE(8-11) BF(8-11) BE(8-11) BF(8-11) + const __m512i rhs_mat_014589CD_1_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_1, (_MM_PERM_ENUM)136); //B0(8-11) B1(8-11) B0(8-11) B1(8-11) B4(8-11) B5(8-11) B4(8-11) B5(8-11) B8(8-11) B9(8-11) B8(8-11) B9(8-11) BC(8-11) BD(8-11) BC(8-11) BD(8-11) + const __m512i rhs_mat_2367ABEF_1_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_1, (_MM_PERM_ENUM)136); //B2(8-11) B3(8-11) B2(8-11) B3(8-11) B6(8-11) B7(8-11) B6(8-11) B7(8-11) BA(8-11) BB(8-11) BA(8-11) BB(8-11) BE(8-11) BF(8-11) BE(8-11) BF(8-11) - const __m512i rhs_mat_014589CD_2_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_2, (_MM_PERM_ENUM)136); //B0(16-19) B1(16-19) B0(16-19) B1(16-19) B4(16-19) B5(16-19) B4(16-19) B5(16-19) B8(16-19) B9(16-19) B8(16-19) B9(16-19) BC(16-19) BD(16-19) BC(16-19) BD(16-19) - const __m512i rhs_mat_2367ABEF_2_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_2, (_MM_PERM_ENUM)136); //B2(16-19) B3(16-19) B2(16-19) B3(16-19) B6(16-19) B7(16-19) B6(16-19) B7(16-19) BA(16-19) BB(16-19) BA(16-19) BB(16-19) BE(16-19) BF(16-19) BE(16-19) BF(16-19) + const __m512i rhs_mat_014589CD_2_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_2, (_MM_PERM_ENUM)136); //B0(16-19) B1(16-19) B0(16-19) B1(16-19) B4(16-19) B5(16-19) B4(16-19) B5(16-19) B8(16-19) B9(16-19) B8(16-19) B9(16-19) BC(16-19) BD(16-19) BC(16-19) BD(16-19) + const __m512i rhs_mat_2367ABEF_2_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_2, (_MM_PERM_ENUM)136); //B2(16-19) B3(16-19) B2(16-19) B3(16-19) B6(16-19) B7(16-19) B6(16-19) B7(16-19) BA(16-19) BB(16-19) BA(16-19) BB(16-19) BE(16-19) BF(16-19) BE(16-19) BF(16-19) - const __m512i rhs_mat_014589CD_3_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_3, (_MM_PERM_ENUM)136); //B0(24-27) B1(24-27) B0(24-27) B1(24-27) B4(24-27) B5(24-27) B4(24-27) B5(24-27) B8(24-27) B9(24-27) B8(24-27) B9(24-27) BC(24-27) BD(24-27) BC(24-27) BD(24-27) - const __m512i rhs_mat_2367ABEF_3_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_3, (_MM_PERM_ENUM)136); //B2(24-27) B3(24-27) B2(24-27) B3(24-27) B6(24-27) B7(24-27) B6(24-27) B7(24-27) BA(24-27) BB(24-27) BA(24-27) BB(24-27) BE(24-27) BF(24-27) BE(24-27) BF(24-27) + const __m512i rhs_mat_014589CD_3_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_3, (_MM_PERM_ENUM)136); //B0(24-27) B1(24-27) B0(24-27) B1(24-27) B4(24-27) B5(24-27) B4(24-27) B5(24-27) B8(24-27) B9(24-27) B8(24-27) B9(24-27) BC(24-27) BD(24-27) BC(24-27) BD(24-27) + const __m512i rhs_mat_2367ABEF_3_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_3, (_MM_PERM_ENUM)136); //B2(24-27) B3(24-27) B2(24-27) B3(24-27) B6(24-27) B7(24-27) B6(24-27) B7(24-27) BA(24-27) BB(24-27) BA(24-27) BB(24-27) BE(24-27) BF(24-27) BE(24-27) BF(24-27) - // Shuffle pattern two - right side input + // Shuffle pattern two - right side input - const __m512i rhs_mat_014589CD_0_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_0, (_MM_PERM_ENUM)221); //B0(4-7) B1(4-7) B0(4-7) B1(4-7) B4(4-7) B5(4-7) B4(4-7) B5(4-7) B8(4-7) B9(4-7) B8(4-7) B9(4-7) BC(4-7) BD(4-7) BC(4-7) BD(4-7) - const __m512i rhs_mat_2367ABEF_0_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_0, (_MM_PERM_ENUM)221); //B2(4-7) B3(4-7) B2(4-7) B3(4-7) B6(4-7) B7(4-7) B6(4-7) B7(4-7) BA(4-7) BB(4-7) BA(4-7) BB(4-7) BE(4-7) BF(4-7) BE(4-7) BF(4-7) + const __m512i rhs_mat_014589CD_0_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_0, (_MM_PERM_ENUM)221); //B0(4-7) B1(4-7) B0(4-7) B1(4-7) B4(4-7) B5(4-7) B4(4-7) B5(4-7) B8(4-7) B9(4-7) B8(4-7) B9(4-7) BC(4-7) BD(4-7) BC(4-7) BD(4-7) + const __m512i rhs_mat_2367ABEF_0_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_0, (_MM_PERM_ENUM)221); //B2(4-7) B3(4-7) B2(4-7) B3(4-7) B6(4-7) B7(4-7) B6(4-7) B7(4-7) BA(4-7) BB(4-7) BA(4-7) BB(4-7) BE(4-7) BF(4-7) BE(4-7) BF(4-7) - const __m512i rhs_mat_014589CD_1_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_1, (_MM_PERM_ENUM)221); //B0(12-15) B1(12-15) B0(12-15) B1(12-15) B4(12-15) B5(12-15) B4(12-15) B5(12-15) B8(12-15) B9(12-15) B8(12-15) B9(12-15) BC(12-15) BD(12-15) BC(12-15) BD(12-15) - const __m512i rhs_mat_2367ABEF_1_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_1, (_MM_PERM_ENUM)221); //B2(12-15) B3(12-15) B2(12-15) B3(12-15) B6(12-15) B7(12-15) B6(12-15) B7(12-15) BA(12-15) BB(12-15) BA(12-15) BB(12-15) BE(12-15) BF(12-15) BE(12-15) BF(12-15) + const __m512i rhs_mat_014589CD_1_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_1, (_MM_PERM_ENUM)221); //B0(12-15) B1(12-15) B0(12-15) B1(12-15) B4(12-15) B5(12-15) B4(12-15) B5(12-15) B8(12-15) B9(12-15) B8(12-15) B9(12-15) BC(12-15) BD(12-15) BC(12-15) BD(12-15) + const __m512i rhs_mat_2367ABEF_1_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_1, (_MM_PERM_ENUM)221); //B2(12-15) B3(12-15) B2(12-15) B3(12-15) B6(12-15) B7(12-15) B6(12-15) B7(12-15) BA(12-15) BB(12-15) BA(12-15) BB(12-15) BE(12-15) BF(12-15) BE(12-15) BF(12-15) - const __m512i rhs_mat_014589CD_2_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_2, (_MM_PERM_ENUM)221); //B0(20-23) B1(20-23) B0(20-23) B1(20-23) B4(20-23) B5(20-23) B4(20-23) B5(20-23) B8(20-23) B9(20-23) B8(20-23) B9(20-23) BC(20-23) BD(20-23) BC(20-23) BD(20-23) - const __m512i rhs_mat_2367ABEF_2_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_2, (_MM_PERM_ENUM)221); //B2(20-23) B3(20-23) B2(20-23) B3(20-23) B6(20-23) B7(20-23) B6(20-23) B7(20-23) BA(20-23) BB(20-23) BA(20-23) BB(20-23) BE(20-23) BF(20-23) BE(20-23) BF(20-23) + const __m512i rhs_mat_014589CD_2_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_2, (_MM_PERM_ENUM)221); //B0(20-23) B1(20-23) B0(20-23) B1(20-23) B4(20-23) B5(20-23) B4(20-23) B5(20-23) B8(20-23) B9(20-23) B8(20-23) B9(20-23) BC(20-23) BD(20-23) BC(20-23) BD(20-23) + const __m512i rhs_mat_2367ABEF_2_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_2, (_MM_PERM_ENUM)221); //B2(20-23) B3(20-23) B2(20-23) B3(20-23) B6(20-23) B7(20-23) B6(20-23) B7(20-23) BA(20-23) BB(20-23) BA(20-23) BB(20-23) BE(20-23) BF(20-23) BE(20-23) BF(20-23) - const __m512i rhs_mat_014589CD_3_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_3, (_MM_PERM_ENUM)221); //B0(28-31) B1(28-31) B0(28-31) B1(28-31) B4(28-31) B5(28-31) B4(28-31) B5(28-31) B8(28-31) B9(28-31) B8(28-31) B9(28-31) BC(28-31) BD(28-31) BC(28-31) BD(28-31) - const __m512i rhs_mat_2367ABEF_3_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_3, (_MM_PERM_ENUM)221); //B2(28-31) B3(28-31) B2(28-31) B3(28-31) B6(28-31) B7(28-31) B6(28-31) B7(28-31) BA(28-31) BB(28-31) BA(28-31) BB(28-31) BE(28-31) BF(28-31) BE(28-31) BF(28-31) + const __m512i rhs_mat_014589CD_3_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_3, (_MM_PERM_ENUM)221); //B0(28-31) B1(28-31) B0(28-31) B1(28-31) B4(28-31) B5(28-31) B4(28-31) B5(28-31) B8(28-31) B9(28-31) B8(28-31) B9(28-31) BC(28-31) BD(28-31) BC(28-31) BD(28-31) + const __m512i rhs_mat_2367ABEF_3_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_3, (_MM_PERM_ENUM)221); //B2(28-31) B3(28-31) B2(28-31) B3(28-31) B6(28-31) B7(28-31) B6(28-31) B7(28-31) BA(28-31) BB(28-31) BA(28-31) BB(28-31) BE(28-31) BF(28-31) BE(28-31) BF(28-31) + // Scale values - Load the weight scale values of two block_tx8 + __m512 col_scale_f32; + if constexpr ( + std::is_same_v || + std::is_same_v) { + col_scale_f32 = GGML_F32Cx8x2_LOAD(b_ptr_0[b].d, b_ptr_1[b].d); + } - // Scale values - Load the weight scale values of two block_q4_0x8 - const __m512 col_scale_f32 = GGML_F32Cx8x2_LOAD(b_ptr_0[b].d, b_ptr_1[b].d); + // Process LHS in pairs of rows + for (int rp = 0; rp < 4; rp++) { - // Load the four block_q4_0 quantized values interleaved with each other in chunks of eight - A0,A1,A2,A3 + // Load the four blocks of quantized values interleaved with each other in chunks of eight - A0,A1,A2,A3 // Loaded as set of 128 bit vectors and repeated and stored into a 256 bit vector before again repeating into 512 bit vector - __m256i lhs_mat_ymm_0123_0 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs))); + __m256i lhs_mat_ymm_0123_0 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs))); __m256i lhs_mat_ymm_01_0 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_0, lhs_mat_ymm_0123_0, 0); __m256i lhs_mat_ymm_23_0 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_0, lhs_mat_ymm_0123_0, 17); - __m256i lhs_mat_ymm_0123_1 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs + 32))); + __m256i lhs_mat_ymm_0123_1 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs + 32))); __m256i lhs_mat_ymm_01_1 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_1, lhs_mat_ymm_0123_1, 0); __m256i lhs_mat_ymm_23_1 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_1, lhs_mat_ymm_0123_1, 17); - __m256i lhs_mat_ymm_0123_2 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs + 64))); + __m256i lhs_mat_ymm_0123_2 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs + 64))); __m256i lhs_mat_ymm_01_2 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_2, lhs_mat_ymm_0123_2, 0); __m256i lhs_mat_ymm_23_2 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_2, lhs_mat_ymm_0123_2, 17); - __m256i lhs_mat_ymm_0123_3 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs + 96))); + __m256i lhs_mat_ymm_0123_3 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs + 96))); __m256i lhs_mat_ymm_01_3 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_3, lhs_mat_ymm_0123_3, 0); __m256i lhs_mat_ymm_23_3 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_3, lhs_mat_ymm_0123_3, 17); @@ -1250,350 +832,814 @@ void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo __m512i iacc_row_3 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_10, (_MM_PERM_ENUM)78), iacc_mat_11); // Load the scale(d) values for all the 4 Q8_0 blocks and repeat it across lanes - const __m128i row_scale_f16 = _mm_shuffle_epi32(_mm_maskload_epi32((int const*)(a_ptr[b].d), loadMask), 68); + const __m128i row_scale_f16 = _mm_shuffle_epi32(_mm_maskload_epi32((int const*)(a_ptrs[rp][b].d), loadMask), 68); const __m512 row_scale_f32 = GGML_F32Cx16_REPEAT_LOAD(row_scale_f16); // Multiply with appropiate scales and accumulate - acc_rows[0] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_0), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 0)), acc_rows[0]); - acc_rows[1] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_1), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 85)), acc_rows[1]); - acc_rows[2] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_2), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_rows[2]); - acc_rows[3] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_3), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 255)), acc_rows[3]); + acc_rows[rp * 4] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_0), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 0)), acc_rows[rp * 4]); + acc_rows[rp * 4 + 1] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_1), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 85)), acc_rows[rp * 4 + 1]); + acc_rows[rp * 4 + 2] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_2), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_rows[rp * 4 + 2]); + acc_rows[rp * 4 + 3] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_3), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 255)), acc_rows[rp * 4 + 3]); } + } + + // Store the accumulated values + for (int i = 0; i < 16; i++) { + _mm512_storeu_ps((float *)(s + ((y * 4 + i) * bs + x * 8)), acc_rows[i]); + } + } + } + + // Take a block_q8_0x4 structures at each pass of the loop and perform dot product operation + for (; y < nr / 4; y ++) { + const block_q8_0x4 * a_ptr = a_ptr_start + (y * nb); + + // Take group of two block_tx8 structures at each pass of the loop and perform dot product operation + for (int64_t x = 0; x < anc / 8; x += 2) { + + const block_tx8 * b_ptr_0 = b_ptr_start + ((x) * b_nb); + const block_tx8 * b_ptr_1 = b_ptr_start + ((x + 1) * b_nb); + + // Master FP accumulators + __m512 acc_rows[4]; + for (int i = 0; i < 4; i++) { + acc_rows[i] = _mm512_setzero_ps(); + } + + for (int64_t b = 0; b < nb; b++) { + // Load the sixteen blocks of quantized values interleaved with each other in chunks of eight - B0,B1 ....BE,BF + const __m256i rhs_raw_mat_0123_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs)); + const __m256i rhs_raw_mat_4567_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs + 32)); + const __m256i rhs_raw_mat_0123_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs + 64)); + const __m256i rhs_raw_mat_4567_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs + 96)); + + const __m256i rhs_raw_mat_89AB_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs)); + const __m256i rhs_raw_mat_CDEF_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs + 32)); + const __m256i rhs_raw_mat_89AB_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs + 64)); + const __m256i rhs_raw_mat_CDEF_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs + 96)); + + // Save the values in the following vectors in the formats B0B1B4B5, B2B3B6B7 for further processing and storing of valuess + const __m256i rhs_raw_mat_0145_0 = _mm256_blend_epi32(rhs_raw_mat_0123_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_0, requiredOrder), 240); + const __m256i rhs_raw_mat_2367_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_0, requiredOrder), rhs_raw_mat_4567_0, 240); + const __m256i rhs_raw_mat_0145_1 = _mm256_blend_epi32(rhs_raw_mat_0123_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_1, requiredOrder), 240); + const __m256i rhs_raw_mat_2367_1 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_1, requiredOrder), rhs_raw_mat_4567_1, 240); + + const __m256i rhs_raw_mat_89CD_0 = _mm256_blend_epi32(rhs_raw_mat_89AB_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_CDEF_0, requiredOrder), 240); + const __m256i rhs_raw_mat_ABEF_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_89AB_0, requiredOrder), rhs_raw_mat_CDEF_0, 240); + const __m256i rhs_raw_mat_89CD_1 = _mm256_blend_epi32(rhs_raw_mat_89AB_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_CDEF_1, requiredOrder), 240); + const __m256i rhs_raw_mat_ABEF_1 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_89AB_1, requiredOrder), rhs_raw_mat_CDEF_1, 240); + + const __m512i rhs_raw_mat_014589CD_0 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_0), rhs_raw_mat_89CD_0, 1); + const __m512i rhs_raw_mat_2367ABEF_0 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_0), rhs_raw_mat_ABEF_0, 1); + const __m512i rhs_raw_mat_014589CD_1 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_1), rhs_raw_mat_89CD_1, 1); + const __m512i rhs_raw_mat_2367ABEF_1 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_1), rhs_raw_mat_ABEF_1, 1); + + // 4-bit -> 8-bit - Sign is maintained + const __m512i rhs_mat_014589CD_0 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_014589CD_0, m4bexpanded)); //B0(0-7) B1(0-7) B4(0-7) B5(0-7) B8(0-7) B9(0-7) BC(0-7) BD(0-7) + const __m512i rhs_mat_2367ABEF_0 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_2367ABEF_0, m4bexpanded)); //B2(0-7) B3(0-7) B6(0-7) B7(0-7) BA(0-7) BB(0-7) BE(0-7) BF(0-7) + + const __m512i rhs_mat_014589CD_1 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_014589CD_1, m4bexpanded)); //B0(8-15) B1(8-15) B4(8-15) B5(8-15) B8(8-15) B9(8-15) BC(8-15) BD(8-15) + const __m512i rhs_mat_2367ABEF_1 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_2367ABEF_1, m4bexpanded)); //B2(8-15) B3(8-15) B6(8-15) B7(8-15) BA(8-15) BB(8-15) BE(8-15) BF(8-15) + + const __m512i rhs_mat_014589CD_2 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_0, 4), m4bexpanded)); //B0(16-23) B1(16-23) B4(16-23) B5(16-23) B8(16-23) B9(16-23) BC(16-23) BD(16-23) + const __m512i rhs_mat_2367ABEF_2 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_0, 4), m4bexpanded)); //B2(16-23) B3(16-23) B6(16-23) B7(16-23) BA(16-23) BB(16-23) BE(16-23) BF(16-23) + + const __m512i rhs_mat_014589CD_3 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_1, 4), m4bexpanded)); //B0(24-31) B1(24-31) B4(24-31) B5(24-31) B8(24-31) B9(24-31) BC(24-31) BD(24-31) + const __m512i rhs_mat_2367ABEF_3 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_1, 4), m4bexpanded)); //B2(24-31) B3(24-31) B6(24-31) B7(24-31) BA(24-31) BB(24-31) BE(24-31) BF(24-31) + + // Shuffle pattern one - right side input + const __m512i rhs_mat_014589CD_0_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_0, (_MM_PERM_ENUM)136); //B0(0-3) B1(0-3) B0(0-3) B1(0-3) B4(0-3) B5(0-3) B4(0-3) B5(0-3) B8(0-3) B9(0-3) B8(0-3) B9(0-3) BC(0-3) BD(0-3) BC(0-3) BD(0-3) + const __m512i rhs_mat_2367ABEF_0_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_0, (_MM_PERM_ENUM)136); //B2(0-3) B3(0-3) B2(0-3) B3(0-3) B6(0-3) B7(0-3) B6(0-3) B7(0-3) BA(0-3) BB(0-3) BA(0-3) BB(0-3) BE(0-3) BF(0-3) BE(0-3) BF(0-3) + + const __m512i rhs_mat_014589CD_1_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_1, (_MM_PERM_ENUM)136); //B0(8-11) B1(8-11) B0(8-11) B1(8-11) B4(8-11) B5(8-11) B4(8-11) B5(8-11) B8(8-11) B9(8-11) B8(8-11) B9(8-11) BC(8-11) BD(8-11) BC(8-11) BD(8-11) + const __m512i rhs_mat_2367ABEF_1_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_1, (_MM_PERM_ENUM)136); //B2(8-11) B3(8-11) B2(8-11) B3(8-11) B6(8-11) B7(8-11) B6(8-11) B7(8-11) BA(8-11) BB(8-11) BA(8-11) BB(8-11) BE(8-11) BF(8-11) BE(8-11) BF(8-11) + + const __m512i rhs_mat_014589CD_2_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_2, (_MM_PERM_ENUM)136); //B0(16-19) B1(16-19) B0(16-19) B1(16-19) B4(16-19) B5(16-19) B4(16-19) B5(16-19) B8(16-19) B9(16-19) B8(16-19) B9(16-19) BC(16-19) BD(16-19) BC(16-19) BD(16-19) + const __m512i rhs_mat_2367ABEF_2_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_2, (_MM_PERM_ENUM)136); //B2(16-19) B3(16-19) B2(16-19) B3(16-19) B6(16-19) B7(16-19) B6(16-19) B7(16-19) BA(16-19) BB(16-19) BA(16-19) BB(16-19) BE(16-19) BF(16-19) BE(16-19) BF(16-19) + + const __m512i rhs_mat_014589CD_3_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_3, (_MM_PERM_ENUM)136); //B0(24-27) B1(24-27) B0(24-27) B1(24-27) B4(24-27) B5(24-27) B4(24-27) B5(24-27) B8(24-27) B9(24-27) B8(24-27) B9(24-27) BC(24-27) BD(24-27) BC(24-27) BD(24-27) + const __m512i rhs_mat_2367ABEF_3_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_3, (_MM_PERM_ENUM)136); //B2(24-27) B3(24-27) B2(24-27) B3(24-27) B6(24-27) B7(24-27) B6(24-27) B7(24-27) BA(24-27) BB(24-27) BA(24-27) BB(24-27) BE(24-27) BF(24-27) BE(24-27) BF(24-27) + + // Shuffle pattern two - right side input + + const __m512i rhs_mat_014589CD_0_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_0, (_MM_PERM_ENUM)221); //B0(4-7) B1(4-7) B0(4-7) B1(4-7) B4(4-7) B5(4-7) B4(4-7) B5(4-7) B8(4-7) B9(4-7) B8(4-7) B9(4-7) BC(4-7) BD(4-7) BC(4-7) BD(4-7) + const __m512i rhs_mat_2367ABEF_0_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_0, (_MM_PERM_ENUM)221); //B2(4-7) B3(4-7) B2(4-7) B3(4-7) B6(4-7) B7(4-7) B6(4-7) B7(4-7) BA(4-7) BB(4-7) BA(4-7) BB(4-7) BE(4-7) BF(4-7) BE(4-7) BF(4-7) - // Store the accumulated values - for (int i = 0; i < 4; i++) { - _mm512_storeu_ps((float *)(s + ((y * 4 + i) * bs + x * 8)), acc_rows[i]); + const __m512i rhs_mat_014589CD_1_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_1, (_MM_PERM_ENUM)221); //B0(12-15) B1(12-15) B0(12-15) B1(12-15) B4(12-15) B5(12-15) B4(12-15) B5(12-15) B8(12-15) B9(12-15) B8(12-15) B9(12-15) BC(12-15) BD(12-15) BC(12-15) BD(12-15) + const __m512i rhs_mat_2367ABEF_1_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_1, (_MM_PERM_ENUM)221); //B2(12-15) B3(12-15) B2(12-15) B3(12-15) B6(12-15) B7(12-15) B6(12-15) B7(12-15) BA(12-15) BB(12-15) BA(12-15) BB(12-15) BE(12-15) BF(12-15) BE(12-15) BF(12-15) + + const __m512i rhs_mat_014589CD_2_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_2, (_MM_PERM_ENUM)221); //B0(20-23) B1(20-23) B0(20-23) B1(20-23) B4(20-23) B5(20-23) B4(20-23) B5(20-23) B8(20-23) B9(20-23) B8(20-23) B9(20-23) BC(20-23) BD(20-23) BC(20-23) BD(20-23) + const __m512i rhs_mat_2367ABEF_2_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_2, (_MM_PERM_ENUM)221); //B2(20-23) B3(20-23) B2(20-23) B3(20-23) B6(20-23) B7(20-23) B6(20-23) B7(20-23) BA(20-23) BB(20-23) BA(20-23) BB(20-23) BE(20-23) BF(20-23) BE(20-23) BF(20-23) + + const __m512i rhs_mat_014589CD_3_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_3, (_MM_PERM_ENUM)221); //B0(28-31) B1(28-31) B0(28-31) B1(28-31) B4(28-31) B5(28-31) B4(28-31) B5(28-31) B8(28-31) B9(28-31) B8(28-31) B9(28-31) BC(28-31) BD(28-31) BC(28-31) BD(28-31) + const __m512i rhs_mat_2367ABEF_3_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_3, (_MM_PERM_ENUM)221); //B2(28-31) B3(28-31) B2(28-31) B3(28-31) B6(28-31) B7(28-31) B6(28-31) B7(28-31) BA(28-31) BB(28-31) BA(28-31) BB(28-31) BE(28-31) BF(28-31) BE(28-31) BF(28-31) + + + // Scale values - Load the weight scale values of two block_tx8 + __m512 col_scale_f32; + if constexpr ( + std::is_same_v || + std::is_same_v) { + col_scale_f32 = GGML_F32Cx8x2_LOAD(b_ptr_0[b].d, b_ptr_1[b].d); } + + // Load the four blocks of quantized values interleaved with each other in chunks of eight - A0,A1,A2,A3 + // Loaded as set of 128 bit vectors and repeated and stored into a 256 bit vector before again repeating into 512 bit vector + __m256i lhs_mat_ymm_0123_0 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs))); + __m256i lhs_mat_ymm_01_0 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_0, lhs_mat_ymm_0123_0, 0); + __m256i lhs_mat_ymm_23_0 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_0, lhs_mat_ymm_0123_0, 17); + __m256i lhs_mat_ymm_0123_1 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs + 32))); + __m256i lhs_mat_ymm_01_1 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_1, lhs_mat_ymm_0123_1, 0); + __m256i lhs_mat_ymm_23_1 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_1, lhs_mat_ymm_0123_1, 17); + __m256i lhs_mat_ymm_0123_2 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs + 64))); + __m256i lhs_mat_ymm_01_2 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_2, lhs_mat_ymm_0123_2, 0); + __m256i lhs_mat_ymm_23_2 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_2, lhs_mat_ymm_0123_2, 17); + __m256i lhs_mat_ymm_0123_3 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs + 96))); + __m256i lhs_mat_ymm_01_3 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_3, lhs_mat_ymm_0123_3, 0); + __m256i lhs_mat_ymm_23_3 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_3, lhs_mat_ymm_0123_3, 17); + + __m512i lhs_mat_01_0 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_0), lhs_mat_ymm_01_0, 1); + __m512i lhs_mat_23_0 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_0), lhs_mat_ymm_23_0, 1); + __m512i lhs_mat_01_1 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_1), lhs_mat_ymm_01_1, 1); + __m512i lhs_mat_23_1 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_1), lhs_mat_ymm_23_1, 1); + __m512i lhs_mat_01_2 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_2), lhs_mat_ymm_01_2, 1); + __m512i lhs_mat_23_2 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_2), lhs_mat_ymm_23_2, 1); + __m512i lhs_mat_01_3 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_3), lhs_mat_ymm_01_3, 1); + __m512i lhs_mat_23_3 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_3), lhs_mat_ymm_23_3, 1); + + // Shuffle pattern one - left side input + + const __m512i lhs_mat_01_0_sp1 = _mm512_shuffle_epi32(lhs_mat_01_0, (_MM_PERM_ENUM)160); //A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3) + const __m512i lhs_mat_23_0_sp1 = _mm512_shuffle_epi32(lhs_mat_23_0, (_MM_PERM_ENUM)160); //A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3) + + const __m512i lhs_mat_01_1_sp1 = _mm512_shuffle_epi32(lhs_mat_01_1, (_MM_PERM_ENUM)160); //A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11) + const __m512i lhs_mat_23_1_sp1 = _mm512_shuffle_epi32(lhs_mat_23_1, (_MM_PERM_ENUM)160); //A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11) + + const __m512i lhs_mat_01_2_sp1 = _mm512_shuffle_epi32(lhs_mat_01_2, (_MM_PERM_ENUM)160); //A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19) + const __m512i lhs_mat_23_2_sp1 = _mm512_shuffle_epi32(lhs_mat_23_2, (_MM_PERM_ENUM)160); //A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19) + + const __m512i lhs_mat_01_3_sp1 = _mm512_shuffle_epi32(lhs_mat_01_3, (_MM_PERM_ENUM)160); //A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27) + const __m512i lhs_mat_23_3_sp1 = _mm512_shuffle_epi32(lhs_mat_23_3, (_MM_PERM_ENUM)160); //A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27) + + // Shuffle pattern two - left side input + + const __m512i lhs_mat_01_0_sp2 = _mm512_shuffle_epi32(lhs_mat_01_0, (_MM_PERM_ENUM)245); //A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7) + const __m512i lhs_mat_23_0_sp2 = _mm512_shuffle_epi32(lhs_mat_23_0, (_MM_PERM_ENUM)245); //A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7) + + const __m512i lhs_mat_01_1_sp2 = _mm512_shuffle_epi32(lhs_mat_01_1, (_MM_PERM_ENUM)245); //A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15) + const __m512i lhs_mat_23_1_sp2 = _mm512_shuffle_epi32(lhs_mat_23_1, (_MM_PERM_ENUM)245); //A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15) + + const __m512i lhs_mat_01_2_sp2 = _mm512_shuffle_epi32(lhs_mat_01_2, (_MM_PERM_ENUM)245); //A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23) + const __m512i lhs_mat_23_2_sp2 = _mm512_shuffle_epi32(lhs_mat_23_2, (_MM_PERM_ENUM)245); //A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23) + + const __m512i lhs_mat_01_3_sp2 = _mm512_shuffle_epi32(lhs_mat_01_3, (_MM_PERM_ENUM)245); //A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31) + const __m512i lhs_mat_23_3_sp2 = _mm512_shuffle_epi32(lhs_mat_23_3, (_MM_PERM_ENUM)245); //A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31) + + // The values arranged in shuffle patterns are operated with dot product operation within 32 bit lane i.e corresponding bytes and multiplied and added into 32 bit integers within 32 bit lane + // Resembles MMLAs into 2x2 matrices in ARM Version + const __m512i zero = _mm512_setzero_epi32(); + __m512i iacc_mat_00_sp1 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_01_3_sp1, rhs_mat_014589CD_3_sp1), lhs_mat_01_2_sp1, rhs_mat_014589CD_2_sp1), lhs_mat_01_1_sp1, rhs_mat_014589CD_1_sp1), lhs_mat_01_0_sp1, rhs_mat_014589CD_0_sp1); + __m512i iacc_mat_01_sp1 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_01_3_sp1, rhs_mat_2367ABEF_3_sp1), lhs_mat_01_2_sp1, rhs_mat_2367ABEF_2_sp1), lhs_mat_01_1_sp1, rhs_mat_2367ABEF_1_sp1), lhs_mat_01_0_sp1, rhs_mat_2367ABEF_0_sp1); + __m512i iacc_mat_10_sp1 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_23_3_sp1, rhs_mat_014589CD_3_sp1), lhs_mat_23_2_sp1, rhs_mat_014589CD_2_sp1), lhs_mat_23_1_sp1, rhs_mat_014589CD_1_sp1), lhs_mat_23_0_sp1, rhs_mat_014589CD_0_sp1); + __m512i iacc_mat_11_sp1 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_23_3_sp1, rhs_mat_2367ABEF_3_sp1), lhs_mat_23_2_sp1, rhs_mat_2367ABEF_2_sp1), lhs_mat_23_1_sp1, rhs_mat_2367ABEF_1_sp1), lhs_mat_23_0_sp1, rhs_mat_2367ABEF_0_sp1); + __m512i iacc_mat_00_sp2 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_01_3_sp2, rhs_mat_014589CD_3_sp2), lhs_mat_01_2_sp2, rhs_mat_014589CD_2_sp2), lhs_mat_01_1_sp2, rhs_mat_014589CD_1_sp2), lhs_mat_01_0_sp2, rhs_mat_014589CD_0_sp2); + __m512i iacc_mat_01_sp2 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_01_3_sp2, rhs_mat_2367ABEF_3_sp2), lhs_mat_01_2_sp2, rhs_mat_2367ABEF_2_sp2), lhs_mat_01_1_sp2, rhs_mat_2367ABEF_1_sp2), lhs_mat_01_0_sp2, rhs_mat_2367ABEF_0_sp2); + __m512i iacc_mat_10_sp2 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_23_3_sp2, rhs_mat_014589CD_3_sp2), lhs_mat_23_2_sp2, rhs_mat_014589CD_2_sp2), lhs_mat_23_1_sp2, rhs_mat_014589CD_1_sp2), lhs_mat_23_0_sp2, rhs_mat_014589CD_0_sp2); + __m512i iacc_mat_11_sp2 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_23_3_sp2, rhs_mat_2367ABEF_3_sp2), lhs_mat_23_2_sp2, rhs_mat_2367ABEF_2_sp2), lhs_mat_23_1_sp2, rhs_mat_2367ABEF_1_sp2), lhs_mat_23_0_sp2, rhs_mat_2367ABEF_0_sp2); + + // Output of both shuffle patterns are added in order to sum dot product outputs of all 32 values in block + __m512i iacc_mat_00 = _mm512_add_epi32(iacc_mat_00_sp1, iacc_mat_00_sp2); + __m512i iacc_mat_01 = _mm512_add_epi32(iacc_mat_01_sp1, iacc_mat_01_sp2); + __m512i iacc_mat_10 = _mm512_add_epi32(iacc_mat_10_sp1, iacc_mat_10_sp2); + __m512i iacc_mat_11 = _mm512_add_epi32(iacc_mat_11_sp1, iacc_mat_11_sp2); + + + // Straighten out to make 4 row vectors + __m512i iacc_row_0 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_00, _mm512_shuffle_epi32(iacc_mat_01, (_MM_PERM_ENUM)78)); + __m512i iacc_row_1 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_00, (_MM_PERM_ENUM)78), iacc_mat_01); + __m512i iacc_row_2 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_10, _mm512_shuffle_epi32(iacc_mat_11, (_MM_PERM_ENUM)78)); + __m512i iacc_row_3 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_10, (_MM_PERM_ENUM)78), iacc_mat_11); + + // Load the scale(d) values for all the 4 Q8_0 blocks and repeat it across lanes + const __m128i row_scale_f16 = _mm_shuffle_epi32(_mm_maskload_epi32((int const*)(a_ptr[b].d), loadMask), 68); + const __m512 row_scale_f32 = GGML_F32Cx16_REPEAT_LOAD(row_scale_f16); + + // Multiply with appropiate scales and accumulate + acc_rows[0] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_0), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 0)), acc_rows[0]); + acc_rows[1] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_1), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 85)), acc_rows[1]); + acc_rows[2] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_2), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_rows[2]); + acc_rows[3] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_3), _mm512_mul_ps(col_scale_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 255)), acc_rows[3]); + } + + // Store the accumulated values + for (int i = 0; i < 4; i++) { + _mm512_storeu_ps((float *)(s + ((y * 4 + i) * bs + x * 8)), acc_rows[i]); } } - if (anc != nc) { - xstart = anc/8; - y = 0; + } + if (anc != nc) { + xstart = anc/8; + y = 0; + } +#endif // __AVX512F__ + + // Take group of four block_q8_0x4 structures at each pass of the loop and perform dot product operation + + for (; y < anr / 4; y += 4) { + const block_q8_0x4 * a_ptrs[4]; + + a_ptrs[0] = a_ptr_start + (y * nb); + for (int i = 0; i < 3; ++i) { + a_ptrs[i + 1] = a_ptrs[i] + nb; } - #endif // __AVX512F__ - // Take group of four block_q8_0x4 structures at each pass of the loop and perform dot product operation + // Take group of eight block_tx8 structures at each pass of the loop and perform dot product operation + for (int64_t x = xstart; x < nc / 8; x++) { - for (; y < anr / 4; y += 4) { - const block_q8_0x4 * a_ptrs[4]; + const block_tx8 * b_ptr = b_ptr_start + (x * b_nb); - a_ptrs[0] = a_ptr_start + (y * nb); - for (int i = 0; i < 3; ++i) { - a_ptrs[i + 1] = a_ptrs[i] + nb; + // Master FP accumulators + __m256 acc_rows[16]; + for (int i = 0; i < 16; i++) { + acc_rows[i] = _mm256_setzero_ps(); } - // Take group of eight block_q4_0x8 structures at each pass of the loop and perform dot product operation - for (int64_t x = xstart; x < nc / 8; x++) { + for (int64_t b = 0; b < nb; b++) { + // Load the eight blocks of quantized values interleaved with each other in chunks of eight - B0,B1 ....B6,B7 + const __m256i rhs_raw_mat_0123_0 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs)); + const __m256i rhs_raw_mat_4567_0 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 32)); + const __m256i rhs_raw_mat_0123_1 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 64)); + const __m256i rhs_raw_mat_4567_1 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 96)); + + // Save the values in the following vectors in the formats B0B1B4B5, B2B3B6B7 for further processing and storing of values + const __m256i rhs_raw_mat_0145_0 = _mm256_blend_epi32(rhs_raw_mat_0123_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_0, requiredOrder), 240); + const __m256i rhs_raw_mat_2367_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_0, requiredOrder), rhs_raw_mat_4567_0, 240); + const __m256i rhs_raw_mat_0145_1 = _mm256_blend_epi32(rhs_raw_mat_0123_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_1, requiredOrder), 240); + const __m256i rhs_raw_mat_2367_1 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_1, requiredOrder), rhs_raw_mat_4567_1, 240); + + // 4-bit -> 8-bit - Sign is maintained + const __m256i rhs_mat_0145_0 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_0145_0, m4b)); //B0(0-7) B1(0-7) B4(0-7) B5(0-7) + const __m256i rhs_mat_2367_0 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_2367_0, m4b)); //B2(0-7) B3(0-7) B6(0-7) B7(0-7) + + const __m256i rhs_mat_0145_1 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_0145_1, m4b)); //B0(8-15) B1(8-15) B4(8-15) B5(8-15) + const __m256i rhs_mat_2367_1 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_2367_1, m4b)); //B2(8-15) B3(8-15) B6(8-15) B7(8-15) + + const __m256i rhs_mat_0145_2 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_0, 4), m4b)); //B0(16-23) B1(16-23) B4(16-23) B5(16-23) + const __m256i rhs_mat_2367_2 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_0, 4), m4b)); //B2(16-23) B3(16-23) B6(16-23) B7(16-23) + + const __m256i rhs_mat_0145_3 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_1, 4), m4b)); //B0(24-31) B1(24-31) B4(24-31) B5(24-31) + const __m256i rhs_mat_2367_3 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_1, 4), m4b)); //B2(24-31) B3(24-31) B6(24-31) B7(24-31) + + // Shuffle pattern one - right side input + const __m256i rhs_mat_0145_0_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_0, 136); //B0(0-3) B1(0-3) B0(0-3) B1(0-3) B4(0-3) B5(0-3) B4(0-3) B5(0-3) + const __m256i rhs_mat_2367_0_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_0, 136); //B2(0-3) B3(0-3) B2(0-3) B3(0-3) B6(0-3) B7(0-3) B6(0-3) B7(0-3) + + const __m256i rhs_mat_0145_1_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_1, 136); //B0(8-11) B1(8-11) B0(8-11) B1(8-11) B4(8-11) B5(8-11) B4(8-11) B5(8-11) + const __m256i rhs_mat_2367_1_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_1, 136); //B2(8-11) B3(8-11) B2(8-11) B3(8-11) B6(8-11) B7(8-11) B6(8-11) B7(8-11) + + const __m256i rhs_mat_0145_2_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_2, 136); //B0(16-19) B1(16-19) B0(16-19) B1(16-19) B4(16-19) B5(16-19) B4(16-19) B5(16-19) + const __m256i rhs_mat_2367_2_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_2, 136); //B2(16-19) B3(16-19) B2(16-19) B3(16-19) B6(16-19) B7(16-19) B6(16-19) B7(16-19) + + const __m256i rhs_mat_0145_3_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_3, 136); //B0(24-27) B1(24-27) B0(24-27) B1(24-27) B4(24-27) B5(24-27) B4(24-27) B5(24-27) + const __m256i rhs_mat_2367_3_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_3, 136); //B2(24-27) B3(24-27) B2(24-27) B3(24-27) B6(24-27) B7(24-27) B6(24-27) B7(24-27) + + // Shuffle pattern two - right side input + + const __m256i rhs_mat_0145_0_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_0, 221); //B0(4-7) B1(4-7) B0(4-7) B1(4-7) B4(4-7) B5(4-7) B4(4-7) B5(4-7) + const __m256i rhs_mat_2367_0_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_0, 221); //B2(4-7) B3(4-7) B2(4-7) B3(4-7) B6(4-7) B7(4-7) B6(4-7) B7(4-7) + + const __m256i rhs_mat_0145_1_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_1, 221); //B0(12-15) B1(12-15) B0(12-15) B1(12-15) B4(12-15) B5(12-15) B4(12-15) B5(12-15) + const __m256i rhs_mat_2367_1_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_1, 221); //B2(12-15) B3(12-15) B2(12-15) B3(12-15) B6(12-15) B7(12-15) B6(12-15) B7(12-15) + + const __m256i rhs_mat_0145_2_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_2, 221); //B0(20-23) B1(20-23) B0(20-23) B1(20-23) B4(20-23) B5(20-23) B4(20-23) B5(20-23) + const __m256i rhs_mat_2367_2_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_2, 221); //B2(20-23) B3(20-23) B2(20-23) B3(20-23) B6(20-23) B7(20-23) B6(20-23) B7(20-23) + + const __m256i rhs_mat_0145_3_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_3, 221); //B0(28-31) B1(28-31) B0(28-31) B1(28-31) B4(28-31) B5(28-31) B4(28-31) B5(28-31) + const __m256i rhs_mat_2367_3_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_3, 221); //B2(28-31) B3(28-31) B2(28-31) B3(28-31) B6(28-31) B7(28-31) B6(28-31) B7(28-31) + + // Scale values - Load the wight scale values of block_tx8 + __m256 col_scale_f32; + if constexpr ( + std::is_same_v || + std::is_same_v) { + col_scale_f32 = GGML_F32Cx8_LOAD(b_ptr[b].d); + } + + // Process LHS in groups of four + for (int rp = 0; rp < 4; rp++) { + // Load the four blocks of quantized values interleaved with each other in chunks of eight - A0,A1,A2,A3 + // Loaded as set of 128 bit vectors and repeated into a 256 bit vector + __m256i lhs_mat_0123_0 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs))); + __m256i lhs_mat_01_0 = _mm256_permute2f128_si256(lhs_mat_0123_0, lhs_mat_0123_0, 0); + __m256i lhs_mat_23_0 = _mm256_permute2f128_si256(lhs_mat_0123_0, lhs_mat_0123_0, 17); + __m256i lhs_mat_0123_1 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs + 32))); + __m256i lhs_mat_01_1 = _mm256_permute2f128_si256(lhs_mat_0123_1, lhs_mat_0123_1, 0); + __m256i lhs_mat_23_1 = _mm256_permute2f128_si256(lhs_mat_0123_1, lhs_mat_0123_1, 17); + __m256i lhs_mat_0123_2 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs + 64))); + __m256i lhs_mat_01_2 = _mm256_permute2f128_si256(lhs_mat_0123_2, lhs_mat_0123_2, 0); + __m256i lhs_mat_23_2 = _mm256_permute2f128_si256(lhs_mat_0123_2, lhs_mat_0123_2, 17); + __m256i lhs_mat_0123_3 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs + 96))); + __m256i lhs_mat_01_3 = _mm256_permute2f128_si256(lhs_mat_0123_3, lhs_mat_0123_3, 0); + __m256i lhs_mat_23_3 = _mm256_permute2f128_si256(lhs_mat_0123_3, lhs_mat_0123_3, 17); + + // Shuffle pattern one - left side input + const __m256i lhs_mat_01_0_sp1 = _mm256_shuffle_epi32(lhs_mat_01_0, 160); //A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3) + const __m256i lhs_mat_23_0_sp1 = _mm256_shuffle_epi32(lhs_mat_23_0, 160); //A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3) + + const __m256i lhs_mat_01_1_sp1 = _mm256_shuffle_epi32(lhs_mat_01_1, 160); //A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11) + const __m256i lhs_mat_23_1_sp1 = _mm256_shuffle_epi32(lhs_mat_23_1, 160); //A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11) + + const __m256i lhs_mat_01_2_sp1 = _mm256_shuffle_epi32(lhs_mat_01_2, 160); //A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19) + const __m256i lhs_mat_23_2_sp1 = _mm256_shuffle_epi32(lhs_mat_23_2, 160); //A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19) + + const __m256i lhs_mat_01_3_sp1 = _mm256_shuffle_epi32(lhs_mat_01_3, 160); //A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27) + const __m256i lhs_mat_23_3_sp1 = _mm256_shuffle_epi32(lhs_mat_23_3, 160); //A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27) + + // Shuffle pattern two - left side input + const __m256i lhs_mat_01_0_sp2 = _mm256_shuffle_epi32(lhs_mat_01_0, 245); //A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7) + const __m256i lhs_mat_23_0_sp2 = _mm256_shuffle_epi32(lhs_mat_23_0, 245); //A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7) + + const __m256i lhs_mat_01_1_sp2 = _mm256_shuffle_epi32(lhs_mat_01_1, 245); //A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15) + const __m256i lhs_mat_23_1_sp2 = _mm256_shuffle_epi32(lhs_mat_23_1, 245); //A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15) + + const __m256i lhs_mat_01_2_sp2 = _mm256_shuffle_epi32(lhs_mat_01_2, 245); //A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23) + const __m256i lhs_mat_23_2_sp2 = _mm256_shuffle_epi32(lhs_mat_23_2, 245); //A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23) + + const __m256i lhs_mat_01_3_sp2 = _mm256_shuffle_epi32(lhs_mat_01_3, 245); //A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31) + const __m256i lhs_mat_23_3_sp2 = _mm256_shuffle_epi32(lhs_mat_23_3, 245); //A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31) + + // The values arranged in shuffle patterns are operated with dot product operation within 32 bit lane i.e corresponding bytes and multiplied and added into 32 bit integers within 32 bit lane + // Resembles MMLAs into 2x2 matrices in ARM Version + const __m256i zero = _mm256_setzero_si256(); + __m256i iacc_mat_00_sp1 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_01_3_sp1, rhs_mat_0145_3_sp1), lhs_mat_01_2_sp1, rhs_mat_0145_2_sp1), lhs_mat_01_1_sp1, rhs_mat_0145_1_sp1), lhs_mat_01_0_sp1, rhs_mat_0145_0_sp1); + __m256i iacc_mat_01_sp1 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_01_3_sp1, rhs_mat_2367_3_sp1), lhs_mat_01_2_sp1, rhs_mat_2367_2_sp1), lhs_mat_01_1_sp1, rhs_mat_2367_1_sp1), lhs_mat_01_0_sp1, rhs_mat_2367_0_sp1); + __m256i iacc_mat_10_sp1 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_23_3_sp1, rhs_mat_0145_3_sp1), lhs_mat_23_2_sp1, rhs_mat_0145_2_sp1), lhs_mat_23_1_sp1, rhs_mat_0145_1_sp1), lhs_mat_23_0_sp1, rhs_mat_0145_0_sp1); + __m256i iacc_mat_11_sp1 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_23_3_sp1, rhs_mat_2367_3_sp1), lhs_mat_23_2_sp1, rhs_mat_2367_2_sp1), lhs_mat_23_1_sp1, rhs_mat_2367_1_sp1), lhs_mat_23_0_sp1, rhs_mat_2367_0_sp1); + __m256i iacc_mat_00_sp2 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_01_3_sp2, rhs_mat_0145_3_sp2), lhs_mat_01_2_sp2, rhs_mat_0145_2_sp2), lhs_mat_01_1_sp2, rhs_mat_0145_1_sp2), lhs_mat_01_0_sp2, rhs_mat_0145_0_sp2); + __m256i iacc_mat_01_sp2 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_01_3_sp2, rhs_mat_2367_3_sp2), lhs_mat_01_2_sp2, rhs_mat_2367_2_sp2), lhs_mat_01_1_sp2, rhs_mat_2367_1_sp2), lhs_mat_01_0_sp2, rhs_mat_2367_0_sp2); + __m256i iacc_mat_10_sp2 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_23_3_sp2, rhs_mat_0145_3_sp2), lhs_mat_23_2_sp2, rhs_mat_0145_2_sp2), lhs_mat_23_1_sp2, rhs_mat_0145_1_sp2), lhs_mat_23_0_sp2, rhs_mat_0145_0_sp2); + __m256i iacc_mat_11_sp2 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_23_3_sp2, rhs_mat_2367_3_sp2), lhs_mat_23_2_sp2, rhs_mat_2367_2_sp2), lhs_mat_23_1_sp2, rhs_mat_2367_1_sp2), lhs_mat_23_0_sp2, rhs_mat_2367_0_sp2); + + // Output of both shuffle patterns are added in order to sum dot product outputs of all 32 values in block + __m256i iacc_mat_00 = _mm256_add_epi32(iacc_mat_00_sp1, iacc_mat_00_sp2); + __m256i iacc_mat_01 = _mm256_add_epi32(iacc_mat_01_sp1, iacc_mat_01_sp2); + __m256i iacc_mat_10 = _mm256_add_epi32(iacc_mat_10_sp1, iacc_mat_10_sp2); + __m256i iacc_mat_11 = _mm256_add_epi32(iacc_mat_11_sp1, iacc_mat_11_sp2); + + // Straighten out to make 4 row vectors + __m256i iacc_row_0 = _mm256_blend_epi32(iacc_mat_00, _mm256_shuffle_epi32(iacc_mat_01, 78), 204); + __m256i iacc_row_1 = _mm256_blend_epi32(_mm256_shuffle_epi32(iacc_mat_00, 78), iacc_mat_01, 204); + __m256i iacc_row_2 = _mm256_blend_epi32(iacc_mat_10, _mm256_shuffle_epi32(iacc_mat_11, 78), 204); + __m256i iacc_row_3 = _mm256_blend_epi32(_mm256_shuffle_epi32(iacc_mat_10, 78), iacc_mat_11, 204); - const block_q4_0x8 * b_ptr = b_ptr_start + (x * b_nb); + // Load the scale(d) values for all the 4 Q8_0 blocks and repeat it across lanes + const __m256 row_scale_f32 = GGML_F32Cx8_REPEAT_LOAD(a_ptrs[rp][b].d, loadMask); - // Master FP accumulators - __m256 acc_rows[16]; - for (int i = 0; i < 16; i++) { - acc_rows[i] = _mm256_setzero_ps(); + // Multiply with appropiate scales and accumulate + acc_rows[rp * 4] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_0), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 0)), acc_rows[rp * 4]); + acc_rows[rp * 4 + 1] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_1), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 85)), acc_rows[rp * 4 + 1]); + acc_rows[rp * 4 + 2] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_2), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_rows[rp * 4 + 2]); + acc_rows[rp * 4 + 3] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_3), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 255)), acc_rows[rp * 4 + 3]); } + } - for (int64_t b = 0; b < nb; b++) { - // Load the eight block_q4_0 quantized values interleaved with each other in chunks of eight - B0,B1 ....B6,B7 - const __m256i rhs_raw_mat_0123_0 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs)); - const __m256i rhs_raw_mat_4567_0 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 32)); - const __m256i rhs_raw_mat_0123_1 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 64)); - const __m256i rhs_raw_mat_4567_1 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 96)); + // Store the accumulated values + for (int i = 0; i < 16; i++) { + _mm256_storeu_ps((float *)(s + ((y * 4 + i) * bs + x * 8)), acc_rows[i]); + } + } + } - // Save the values in the following vectors in the formats B0B1B4B5, B2B3B6B7 for further processing and storing of values - const __m256i rhs_raw_mat_0145_0 = _mm256_blend_epi32(rhs_raw_mat_0123_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_0, requiredOrder), 240); - const __m256i rhs_raw_mat_2367_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_0, requiredOrder), rhs_raw_mat_4567_0, 240); - const __m256i rhs_raw_mat_0145_1 = _mm256_blend_epi32(rhs_raw_mat_0123_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_1, requiredOrder), 240); - const __m256i rhs_raw_mat_2367_1 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_1, requiredOrder), rhs_raw_mat_4567_1, 240); + // Take a block_q8_0x4 structures at each pass of the loop and perform dot product operation + for (; y < nr / 4; y ++) { + const block_q8_0x4 * a_ptr = a_ptr_start + (y * nb); - // 4-bit -> 8-bit - Sign is maintained - const __m256i rhs_mat_0145_0 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_0145_0, m4b)); //B0(0-7) B1(0-7) B4(0-7) B5(0-7) - const __m256i rhs_mat_2367_0 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_2367_0, m4b)); //B2(0-7) B3(0-7) B6(0-7) B7(0-7) + // Load the eight blocks of quantized values interleaved with each other in chunks of eight - B0,B1 ....B6,B7 + for (int64_t x = xstart; x < nc / 8; x++) { + const block_tx8 * b_ptr = b_ptr_start + (x * b_nb); - const __m256i rhs_mat_0145_1 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_0145_1, m4b)); //B0(8-15) B1(8-15) B4(8-15) B5(8-15) - const __m256i rhs_mat_2367_1 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_2367_1, m4b)); //B2(8-15) B3(8-15) B6(8-15) B7(8-15) + // Master FP accumulators + __m256 acc_rows[4]; + for (int i = 0; i < 4; i++) { + acc_rows[i] = _mm256_setzero_ps(); + } - const __m256i rhs_mat_0145_2 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_0, 4), m4b)); //B0(16-23) B1(16-23) B4(16-23) B5(16-23) - const __m256i rhs_mat_2367_2 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_0, 4), m4b)); //B2(16-23) B3(16-23) B6(16-23) B7(16-23) + for (int64_t b = 0; b < nb; b++) { + // Load the eight block_q8_0 quantized values interleaved with each other in chunks of eight - B0,B1 ....B6,B7 + const __m256i rhs_raw_mat_0123_0 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs)); + const __m256i rhs_raw_mat_4567_0 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 32)); + const __m256i rhs_raw_mat_0123_1 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 64)); + const __m256i rhs_raw_mat_4567_1 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 96)); + + // Save the values in the following vectors in the formats B0B1B4B5, B2B3B6B7 for further processing and storing of valuess + const __m256i rhs_raw_mat_0145_0 = _mm256_blend_epi32(rhs_raw_mat_0123_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_0, requiredOrder), 240); + const __m256i rhs_raw_mat_2367_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_0, requiredOrder), rhs_raw_mat_4567_0, 240); + const __m256i rhs_raw_mat_0145_1 = _mm256_blend_epi32(rhs_raw_mat_0123_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_1, requiredOrder), 240); + const __m256i rhs_raw_mat_2367_1 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_1, requiredOrder), rhs_raw_mat_4567_1, 240); - const __m256i rhs_mat_0145_3 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_1, 4), m4b)); //B0(24-31) B1(24-31) B4(24-31) B5(24-31) - const __m256i rhs_mat_2367_3 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_1, 4), m4b)); //B2(24-31) B3(24-31) B6(24-31) B7(24-31) + // 4-bit -> 8-bit - Sign is maintained + const __m256i rhs_mat_0145_0 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_0145_0, m4b)); //B0(0-7) B1(0-7) B4(0-7) B5(0-7) + const __m256i rhs_mat_2367_0 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_2367_0, m4b)); //B2(0-7) B3(0-7) B6(0-7) B7(0-7) - // Shuffle pattern one - right side input - const __m256i rhs_mat_0145_0_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_0, 136); //B0(0-3) B1(0-3) B0(0-3) B1(0-3) B4(0-3) B5(0-3) B4(0-3) B5(0-3) - const __m256i rhs_mat_2367_0_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_0, 136); //B2(0-3) B3(0-3) B2(0-3) B3(0-3) B6(0-3) B7(0-3) B6(0-3) B7(0-3) + const __m256i rhs_mat_0145_1 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_0145_1, m4b)); //B0(8-15) B1(8-15) B4(8-15) B5(8-15) + const __m256i rhs_mat_2367_1 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_2367_1, m4b)); //B2(8-15) B3(8-15) B6(8-15) B7(8-15) - const __m256i rhs_mat_0145_1_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_1, 136); //B0(8-11) B1(8-11) B0(8-11) B1(8-11) B4(8-11) B5(8-11) B4(8-11) B5(8-11) - const __m256i rhs_mat_2367_1_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_1, 136); //B2(8-11) B3(8-11) B2(8-11) B3(8-11) B6(8-11) B7(8-11) B6(8-11) B7(8-11) + const __m256i rhs_mat_0145_2 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_0, 4), m4b)); //B0(16-23) B1(16-23) B4(16-23) B5(16-23) + const __m256i rhs_mat_2367_2 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_0, 4), m4b)); //B2(16-23) B3(16-23) B6(16-23) B7(16-23) - const __m256i rhs_mat_0145_2_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_2, 136); //B0(16-19) B1(16-19) B0(16-19) B1(16-19) B4(16-19) B5(16-19) B4(16-19) B5(16-19) - const __m256i rhs_mat_2367_2_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_2, 136); //B2(16-19) B3(16-19) B2(16-19) B3(16-19) B6(16-19) B7(16-19) B6(16-19) B7(16-19) + const __m256i rhs_mat_0145_3 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_1, 4), m4b)); //B0(24-31) B1(24-31) B4(24-31) B5(24-31) + const __m256i rhs_mat_2367_3 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_1, 4), m4b)); //B2(24-31) B3(24-31) B6(24-31) B7(24-31) - const __m256i rhs_mat_0145_3_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_3, 136); //B0(24-27) B1(24-27) B0(24-27) B1(24-27) B4(24-27) B5(24-27) B4(24-27) B5(24-27) - const __m256i rhs_mat_2367_3_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_3, 136); //B2(24-27) B3(24-27) B2(24-27) B3(24-27) B6(24-27) B7(24-27) B6(24-27) B7(24-27) + // Shuffle pattern one - right side input + const __m256i rhs_mat_0145_0_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_0, 136); //B0(0-3) B1(0-3) B0(0-3) B1(0-3) B4(0-3) B5(0-3) B4(0-3) B5(0-3) + const __m256i rhs_mat_2367_0_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_0, 136); //B2(0-3) B3(0-3) B2(0-3) B3(0-3) B6(0-3) B7(0-3) B6(0-3) B7(0-3) - // Shuffle pattern two - right side input + const __m256i rhs_mat_0145_1_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_1, 136); //B0(8-11) B1(8-11) B0(8-11) B1(8-11) B4(8-11) B5(8-11) B4(8-11) B5(8-11) + const __m256i rhs_mat_2367_1_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_1, 136); //B2(8-11) B3(8-11) B2(8-11) B3(8-11) B6(8-11) B7(8-11) B6(8-11) B7(8-11) - const __m256i rhs_mat_0145_0_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_0, 221); //B0(4-7) B1(4-7) B0(4-7) B1(4-7) B4(4-7) B5(4-7) B4(4-7) B5(4-7) - const __m256i rhs_mat_2367_0_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_0, 221); //B2(4-7) B3(4-7) B2(4-7) B3(4-7) B6(4-7) B7(4-7) B6(4-7) B7(4-7) + const __m256i rhs_mat_0145_2_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_2, 136); //B0(16-19) B1(16-19) B0(16-19) B1(16-19) B4(16-19) B5(16-19) B4(16-19) B5(16-19) + const __m256i rhs_mat_2367_2_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_2, 136); //B2(16-19) B3(16-19) B2(16-19) B3(16-19) B6(16-19) B7(16-19) B6(16-19) B7(16-19) - const __m256i rhs_mat_0145_1_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_1, 221); //B0(12-15) B1(12-15) B0(12-15) B1(12-15) B4(12-15) B5(12-15) B4(12-15) B5(12-15) - const __m256i rhs_mat_2367_1_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_1, 221); //B2(12-15) B3(12-15) B2(12-15) B3(12-15) B6(12-15) B7(12-15) B6(12-15) B7(12-15) + const __m256i rhs_mat_0145_3_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_3, 136); //B0(24-27) B1(24-27) B0(24-27) B1(24-27) B4(24-27) B5(24-27) B4(24-27) B5(24-27) + const __m256i rhs_mat_2367_3_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_3, 136); //B2(24-27) B3(24-27) B2(24-27) B3(24-27) B6(24-27) B7(24-27) B6(24-27) B7(24-27) - const __m256i rhs_mat_0145_2_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_2, 221); //B0(20-23) B1(20-23) B0(20-23) B1(20-23) B4(20-23) B5(20-23) B4(20-23) B5(20-23) - const __m256i rhs_mat_2367_2_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_2, 221); //B2(20-23) B3(20-23) B2(20-23) B3(20-23) B6(20-23) B7(20-23) B6(20-23) B7(20-23) + // Shuffle pattern two - right side input - const __m256i rhs_mat_0145_3_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_3, 221); //B0(28-31) B1(28-31) B0(28-31) B1(28-31) B4(28-31) B5(28-31) B4(28-31) B5(28-31) - const __m256i rhs_mat_2367_3_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_3, 221); //B2(28-31) B3(28-31) B2(28-31) B3(28-31) B6(28-31) B7(28-31) B6(28-31) B7(28-31) + const __m256i rhs_mat_0145_0_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_0, 221); //B0(4-7) B1(4-7) B0(4-7) B1(4-7) B4(4-7) B5(4-7) B4(4-7) B5(4-7) + const __m256i rhs_mat_2367_0_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_0, 221); //B2(4-7) B3(4-7) B2(4-7) B3(4-7) B6(4-7) B7(4-7) B6(4-7) B7(4-7) - // Scale values - Load the wight scale values of block_q4_0x8 - const __m256 col_scale_f32 = GGML_F32Cx8_LOAD(b_ptr[b].d); + const __m256i rhs_mat_0145_1_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_1, 221); //B0(12-15) B1(12-15) B0(12-15) B1(12-15) B4(12-15) B5(12-15) B4(12-15) B5(12-15) + const __m256i rhs_mat_2367_1_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_1, 221); //B2(12-15) B3(12-15) B2(12-15) B3(12-15) B6(12-15) B7(12-15) B6(12-15) B7(12-15) - // Process LHS in groups of four - for (int rp = 0; rp < 4; rp++) { - // Load the four block_q4_0 quantized values interleaved with each other in chunks of eight - A0,A1,A2,A3 - // Loaded as set of 128 bit vectors and repeated into a 256 bit vector - __m256i lhs_mat_0123_0 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs))); - __m256i lhs_mat_01_0 = _mm256_permute2f128_si256(lhs_mat_0123_0, lhs_mat_0123_0, 0); - __m256i lhs_mat_23_0 = _mm256_permute2f128_si256(lhs_mat_0123_0, lhs_mat_0123_0, 17); - __m256i lhs_mat_0123_1 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs + 32))); - __m256i lhs_mat_01_1 = _mm256_permute2f128_si256(lhs_mat_0123_1, lhs_mat_0123_1, 0); - __m256i lhs_mat_23_1 = _mm256_permute2f128_si256(lhs_mat_0123_1, lhs_mat_0123_1, 17); - __m256i lhs_mat_0123_2 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs + 64))); - __m256i lhs_mat_01_2 = _mm256_permute2f128_si256(lhs_mat_0123_2, lhs_mat_0123_2, 0); - __m256i lhs_mat_23_2 = _mm256_permute2f128_si256(lhs_mat_0123_2, lhs_mat_0123_2, 17); - __m256i lhs_mat_0123_3 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs + 96))); - __m256i lhs_mat_01_3 = _mm256_permute2f128_si256(lhs_mat_0123_3, lhs_mat_0123_3, 0); - __m256i lhs_mat_23_3 = _mm256_permute2f128_si256(lhs_mat_0123_3, lhs_mat_0123_3, 17); + const __m256i rhs_mat_0145_2_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_2, 221); //B0(20-23) B1(20-23) B0(20-23) B1(20-23) B4(20-23) B5(20-23) B4(20-23) B5(20-23) + const __m256i rhs_mat_2367_2_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_2, 221); //B2(20-23) B3(20-23) B2(20-23) B3(20-23) B6(20-23) B7(20-23) B6(20-23) B7(20-23) + + const __m256i rhs_mat_0145_3_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_3, 221); //B0(28-31) B1(28-31) B0(28-31) B1(28-31) B4(28-31) B5(28-31) B4(28-31) B5(28-31) + const __m256i rhs_mat_2367_3_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_3, 221); //B2(28-31) B3(28-31) B2(28-31) B3(28-31) B6(28-31) B7(28-31) B6(28-31) B7(28-31) + + // Scale values - Load the wight scale values of block_tx8 + __m256 col_scale_f32; + if constexpr ( + std::is_same_v || + std::is_same_v) { + col_scale_f32 = GGML_F32Cx8_LOAD(b_ptr[b].d); + } + + // Load the four blocks of quantized values interleaved with each other in chunks of eight - A0,A1,A2,A3 + // Loaded as set of 128 bit vectors and repeated into a 256 bit vector + __m256i lhs_mat_0123_0 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs))); + __m256i lhs_mat_01_0 = _mm256_permute2f128_si256(lhs_mat_0123_0, lhs_mat_0123_0, 0); + __m256i lhs_mat_23_0 = _mm256_permute2f128_si256(lhs_mat_0123_0, lhs_mat_0123_0, 17); + __m256i lhs_mat_0123_1 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs + 32))); + __m256i lhs_mat_01_1 = _mm256_permute2f128_si256(lhs_mat_0123_1, lhs_mat_0123_1, 0); + __m256i lhs_mat_23_1 = _mm256_permute2f128_si256(lhs_mat_0123_1, lhs_mat_0123_1, 17); + __m256i lhs_mat_0123_2 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs + 64))); + __m256i lhs_mat_01_2 = _mm256_permute2f128_si256(lhs_mat_0123_2, lhs_mat_0123_2, 0); + __m256i lhs_mat_23_2 = _mm256_permute2f128_si256(lhs_mat_0123_2, lhs_mat_0123_2, 17); + __m256i lhs_mat_0123_3 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs + 96))); + __m256i lhs_mat_01_3 = _mm256_permute2f128_si256(lhs_mat_0123_3, lhs_mat_0123_3, 0); + __m256i lhs_mat_23_3 = _mm256_permute2f128_si256(lhs_mat_0123_3, lhs_mat_0123_3, 17); + + // Shuffle pattern one - left side input + + const __m256i lhs_mat_01_0_sp1 = _mm256_shuffle_epi32(lhs_mat_01_0, 160); //A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3) + const __m256i lhs_mat_23_0_sp1 = _mm256_shuffle_epi32(lhs_mat_23_0, 160); //A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3) + + const __m256i lhs_mat_01_1_sp1 = _mm256_shuffle_epi32(lhs_mat_01_1, 160); //A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11) + const __m256i lhs_mat_23_1_sp1 = _mm256_shuffle_epi32(lhs_mat_23_1, 160); //A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11) + + const __m256i lhs_mat_01_2_sp1 = _mm256_shuffle_epi32(lhs_mat_01_2, 160); //A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19) + const __m256i lhs_mat_23_2_sp1 = _mm256_shuffle_epi32(lhs_mat_23_2, 160); //A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19) + + const __m256i lhs_mat_01_3_sp1 = _mm256_shuffle_epi32(lhs_mat_01_3, 160); //A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27) + const __m256i lhs_mat_23_3_sp1 = _mm256_shuffle_epi32(lhs_mat_23_3, 160); //A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27) + + // Shuffle pattern two - left side input + + const __m256i lhs_mat_01_0_sp2 = _mm256_shuffle_epi32(lhs_mat_01_0, 245); //A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7) + const __m256i lhs_mat_23_0_sp2 = _mm256_shuffle_epi32(lhs_mat_23_0, 245); //A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7) + + const __m256i lhs_mat_01_1_sp2 = _mm256_shuffle_epi32(lhs_mat_01_1, 245); //A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15) + const __m256i lhs_mat_23_1_sp2 = _mm256_shuffle_epi32(lhs_mat_23_1, 245); //A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15) + + const __m256i lhs_mat_01_2_sp2 = _mm256_shuffle_epi32(lhs_mat_01_2, 245); //A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23) + const __m256i lhs_mat_23_2_sp2 = _mm256_shuffle_epi32(lhs_mat_23_2, 245); //A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23) + + const __m256i lhs_mat_01_3_sp2 = _mm256_shuffle_epi32(lhs_mat_01_3, 245); //A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31) + const __m256i lhs_mat_23_3_sp2 = _mm256_shuffle_epi32(lhs_mat_23_3, 245); //A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31) + + // The values arranged in shuffle patterns are operated with dot product operation within 32 bit lane i.e corresponding bytes and multiplied and added into 32 bit integers within 32 bit lane + // Resembles MMLAs into 2x2 matrices in ARM Version + const __m256i zero = _mm256_setzero_si256(); + __m256i iacc_mat_00_sp1 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_01_3_sp1, rhs_mat_0145_3_sp1), lhs_mat_01_2_sp1, rhs_mat_0145_2_sp1), lhs_mat_01_1_sp1, rhs_mat_0145_1_sp1), lhs_mat_01_0_sp1, rhs_mat_0145_0_sp1); + __m256i iacc_mat_01_sp1 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_01_3_sp1, rhs_mat_2367_3_sp1), lhs_mat_01_2_sp1, rhs_mat_2367_2_sp1), lhs_mat_01_1_sp1, rhs_mat_2367_1_sp1), lhs_mat_01_0_sp1, rhs_mat_2367_0_sp1); + __m256i iacc_mat_10_sp1 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_23_3_sp1, rhs_mat_0145_3_sp1), lhs_mat_23_2_sp1, rhs_mat_0145_2_sp1), lhs_mat_23_1_sp1, rhs_mat_0145_1_sp1), lhs_mat_23_0_sp1, rhs_mat_0145_0_sp1); + __m256i iacc_mat_11_sp1 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_23_3_sp1, rhs_mat_2367_3_sp1), lhs_mat_23_2_sp1, rhs_mat_2367_2_sp1), lhs_mat_23_1_sp1, rhs_mat_2367_1_sp1), lhs_mat_23_0_sp1, rhs_mat_2367_0_sp1); + __m256i iacc_mat_00_sp2 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_01_3_sp2, rhs_mat_0145_3_sp2), lhs_mat_01_2_sp2, rhs_mat_0145_2_sp2), lhs_mat_01_1_sp2, rhs_mat_0145_1_sp2), lhs_mat_01_0_sp2, rhs_mat_0145_0_sp2); + __m256i iacc_mat_01_sp2 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_01_3_sp2, rhs_mat_2367_3_sp2), lhs_mat_01_2_sp2, rhs_mat_2367_2_sp2), lhs_mat_01_1_sp2, rhs_mat_2367_1_sp2), lhs_mat_01_0_sp2, rhs_mat_2367_0_sp2); + __m256i iacc_mat_10_sp2 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_23_3_sp2, rhs_mat_0145_3_sp2), lhs_mat_23_2_sp2, rhs_mat_0145_2_sp2), lhs_mat_23_1_sp2, rhs_mat_0145_1_sp2), lhs_mat_23_0_sp2, rhs_mat_0145_0_sp2); + __m256i iacc_mat_11_sp2 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_23_3_sp2, rhs_mat_2367_3_sp2), lhs_mat_23_2_sp2, rhs_mat_2367_2_sp2), lhs_mat_23_1_sp2, rhs_mat_2367_1_sp2), lhs_mat_23_0_sp2, rhs_mat_2367_0_sp2); + + // Output of both shuffle patterns are added in order to sum dot product outputs of all 32 values in block + __m256i iacc_mat_00 = _mm256_add_epi32(iacc_mat_00_sp1, iacc_mat_00_sp2); + __m256i iacc_mat_01 = _mm256_add_epi32(iacc_mat_01_sp1, iacc_mat_01_sp2); + __m256i iacc_mat_10 = _mm256_add_epi32(iacc_mat_10_sp1, iacc_mat_10_sp2); + __m256i iacc_mat_11 = _mm256_add_epi32(iacc_mat_11_sp1, iacc_mat_11_sp2); + + + // Straighten out to make 4 row vectors + __m256i iacc_row_0 = _mm256_blend_epi32(iacc_mat_00, _mm256_shuffle_epi32(iacc_mat_01, 78), 204); + __m256i iacc_row_1 = _mm256_blend_epi32(_mm256_shuffle_epi32(iacc_mat_00, 78), iacc_mat_01, 204); + __m256i iacc_row_2 = _mm256_blend_epi32(iacc_mat_10, _mm256_shuffle_epi32(iacc_mat_11, 78), 204); + __m256i iacc_row_3 = _mm256_blend_epi32(_mm256_shuffle_epi32(iacc_mat_10, 78), iacc_mat_11, 204); + + // Load the scale(d) values for all the 4 Q8_0 blocks and repeat it across lanes + const __m256 row_scale_f32 = GGML_F32Cx8_REPEAT_LOAD(a_ptr[b].d, loadMask); + + // Multiply with appropiate scales and accumulate + acc_rows[0] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_0), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 0)), acc_rows[0]); + acc_rows[1] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_1), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 85)), acc_rows[1]); + acc_rows[2] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_2), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_rows[2]); + acc_rows[3] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_3), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 255)), acc_rows[3]); + } + + // Store the accumulated values + for (int i = 0; i < 4; i++) { + _mm256_storeu_ps((float *)(s + ((y * 4 + i) * bs + x * 8)), acc_rows[i]); + } + } + } +} + +#endif // defined(__AVX2__) || defined(__AVX512F__) + +void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) { +#if defined(__AVX2__) || defined(__AVX512F__) + { + // Lookup table to convert signed nibbles to signed bytes + __m256i signextendlut = _mm256_castsi128_si256(_mm_set_epi8(-1, -2, -3, -4, -5, -6, -7, -8, 7, 6, 5, 4, 3, 2, 1, 0)); + signextendlut = _mm256_permute2f128_si256(signextendlut, signextendlut, 0); + + gemv_q4_b32_8x8_q8_0_lut_avx(n, s, bs, vx, vy, nr, nc, signextendlut); + + return; + } +#endif + + ggml_gemv_q4_0_8x8_q8_0_generic(n, s, bs, vx, vy, nr, nc); +} + +void ggml_gemv_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) { + const int qk = QK_K; + const int nb = n / qk; + const int ncols_interleaved = 8; + const int blocklen = 8; + static const uint32_t kmask1 = 0x3f3f3f3f; + static const uint32_t kmask2 = 0x0f0f0f0f; + static const uint32_t kmask3 = 0x03030303; + + assert (n % qk == 0); + assert (nc % ncols_interleaved == 0); + + UNUSED(s); + UNUSED(bs); + UNUSED(vx); + UNUSED(vy); + UNUSED(nr); + UNUSED(nc); + UNUSED(nb); + UNUSED(ncols_interleaved); + UNUSED(blocklen); - // Shuffle pattern one - left side input - const __m256i lhs_mat_01_0_sp1 = _mm256_shuffle_epi32(lhs_mat_01_0, 160); //A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3) - const __m256i lhs_mat_23_0_sp1 = _mm256_shuffle_epi32(lhs_mat_23_0, 160); //A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3) +#if defined(__AVX2__) + // Lookup table to convert signed nibbles to signed bytes + __m256i signextendlut = _mm256_castsi128_si256(_mm_set_epi8(-1, -2, -3, -4, -5, -6, -7, -8, 7, 6, 5, 4, 3, 2, 1, 0)); + signextendlut = _mm256_permute2f128_si256(signextendlut, signextendlut, 0); + // Shuffle masks to rearrange delta and scale values to multiply with appropriate scales + __m128i deltamask = _mm_set_epi8(15, 14, 7, 6, 13, 12, 5, 4, 11, 10, 3, 2, 9, 8, 1, 0); + __m128i scalemask = _mm_set_epi8(7, 7, 3, 3, 6, 6, 2, 2, 5, 5, 1, 1, 4, 4, 0, 0); + // Permute mask used for easier vector processing at later stages + __m256i finalpermutemask = _mm256_set_epi32(7, 5, 3, 1, 6, 4, 2, 0); - const __m256i lhs_mat_01_1_sp1 = _mm256_shuffle_epi32(lhs_mat_01_1, 160); //A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11) - const __m256i lhs_mat_23_1_sp1 = _mm256_shuffle_epi32(lhs_mat_23_1, 160); //A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11) + // Mask to extract nibbles from bytes + const __m256i m4b = _mm256_set1_epi8(0x0F); - const __m256i lhs_mat_01_2_sp1 = _mm256_shuffle_epi32(lhs_mat_01_2, 160); //A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19) - const __m256i lhs_mat_23_2_sp1 = _mm256_shuffle_epi32(lhs_mat_23_2, 160); //A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19) + int64_t b_nb = n / QK_K; - const __m256i lhs_mat_01_3_sp1 = _mm256_shuffle_epi32(lhs_mat_01_3, 160); //A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27) - const __m256i lhs_mat_23_3_sp1 = _mm256_shuffle_epi32(lhs_mat_23_3, 160); //A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27) + const block_q4_Kx8 * b_ptr_start = (const block_q4_Kx8 *)vx; + const block_q8_K * a_ptr_start = (const block_q8_K *)vy; - // Shuffle pattern two - left side input - const __m256i lhs_mat_01_0_sp2 = _mm256_shuffle_epi32(lhs_mat_01_0, 245); //A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7) - const __m256i lhs_mat_23_0_sp2 = _mm256_shuffle_epi32(lhs_mat_23_0, 245); //A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7) + // Process Q8_K blocks one by one + for (int64_t y = 0; y < nr; y++) { - const __m256i lhs_mat_01_1_sp2 = _mm256_shuffle_epi32(lhs_mat_01_1, 245); //A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15) - const __m256i lhs_mat_23_1_sp2 = _mm256_shuffle_epi32(lhs_mat_23_1, 245); //A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15) + // Pointers to LHS blocks of block_q8_K format + const block_q8_K * a_ptr = a_ptr_start + (y * nb); - const __m256i lhs_mat_01_2_sp2 = _mm256_shuffle_epi32(lhs_mat_01_2, 245); //A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23) - const __m256i lhs_mat_23_2_sp2 = _mm256_shuffle_epi32(lhs_mat_23_2, 245); //A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23) + // Take group of eight interleaved block_q4_K structures at each pass of the loop and perform dot product operation + for (int64_t x = 0; x < nc / 8; x++) { - const __m256i lhs_mat_01_3_sp2 = _mm256_shuffle_epi32(lhs_mat_01_3, 245); //A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31) - const __m256i lhs_mat_23_3_sp2 = _mm256_shuffle_epi32(lhs_mat_23_3, 245); //A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31) + // Pointers to RHS blocks + const block_q4_Kx8 * b_ptr = b_ptr_start + (x * b_nb); - // The values arranged in shuffle patterns are operated with dot product operation within 32 bit lane i.e corresponding bytes and multiplied and added into 32 bit integers within 32 bit lane - // Resembles MMLAs into 2x2 matrices in ARM Version - const __m256i zero = _mm256_setzero_si256(); - __m256i iacc_mat_00_sp1 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_01_3_sp1, rhs_mat_0145_3_sp1), lhs_mat_01_2_sp1, rhs_mat_0145_2_sp1), lhs_mat_01_1_sp1, rhs_mat_0145_1_sp1), lhs_mat_01_0_sp1, rhs_mat_0145_0_sp1); - __m256i iacc_mat_01_sp1 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_01_3_sp1, rhs_mat_2367_3_sp1), lhs_mat_01_2_sp1, rhs_mat_2367_2_sp1), lhs_mat_01_1_sp1, rhs_mat_2367_1_sp1), lhs_mat_01_0_sp1, rhs_mat_2367_0_sp1); - __m256i iacc_mat_10_sp1 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_23_3_sp1, rhs_mat_0145_3_sp1), lhs_mat_23_2_sp1, rhs_mat_0145_2_sp1), lhs_mat_23_1_sp1, rhs_mat_0145_1_sp1), lhs_mat_23_0_sp1, rhs_mat_0145_0_sp1); - __m256i iacc_mat_11_sp1 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_23_3_sp1, rhs_mat_2367_3_sp1), lhs_mat_23_2_sp1, rhs_mat_2367_2_sp1), lhs_mat_23_1_sp1, rhs_mat_2367_1_sp1), lhs_mat_23_0_sp1, rhs_mat_2367_0_sp1); - __m256i iacc_mat_00_sp2 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_01_3_sp2, rhs_mat_0145_3_sp2), lhs_mat_01_2_sp2, rhs_mat_0145_2_sp2), lhs_mat_01_1_sp2, rhs_mat_0145_1_sp2), lhs_mat_01_0_sp2, rhs_mat_0145_0_sp2); - __m256i iacc_mat_01_sp2 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_01_3_sp2, rhs_mat_2367_3_sp2), lhs_mat_01_2_sp2, rhs_mat_2367_2_sp2), lhs_mat_01_1_sp2, rhs_mat_2367_1_sp2), lhs_mat_01_0_sp2, rhs_mat_2367_0_sp2); - __m256i iacc_mat_10_sp2 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_23_3_sp2, rhs_mat_0145_3_sp2), lhs_mat_23_2_sp2, rhs_mat_0145_2_sp2), lhs_mat_23_1_sp2, rhs_mat_0145_1_sp2), lhs_mat_23_0_sp2, rhs_mat_0145_0_sp2); - __m256i iacc_mat_11_sp2 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_23_3_sp2, rhs_mat_2367_3_sp2), lhs_mat_23_2_sp2, rhs_mat_2367_2_sp2), lhs_mat_23_1_sp2, rhs_mat_2367_1_sp2), lhs_mat_23_0_sp2, rhs_mat_2367_0_sp2); + // Master FP accumulators + __m256 acc_row = _mm256_setzero_ps(); + __m256 acc_min_rows = _mm256_setzero_ps(); - // Output of both shuffle patterns are added in order to sum dot product outputs of all 32 values in block - __m256i iacc_mat_00 = _mm256_add_epi32(iacc_mat_00_sp1, iacc_mat_00_sp2); - __m256i iacc_mat_01 = _mm256_add_epi32(iacc_mat_01_sp1, iacc_mat_01_sp2); - __m256i iacc_mat_10 = _mm256_add_epi32(iacc_mat_10_sp1, iacc_mat_10_sp2); - __m256i iacc_mat_11 = _mm256_add_epi32(iacc_mat_11_sp1, iacc_mat_11_sp2); + for (int64_t b = 0; b < nb; b++) { - // Straighten out to make 4 row vectors - __m256i iacc_row_0 = _mm256_blend_epi32(iacc_mat_00, _mm256_shuffle_epi32(iacc_mat_01, 78), 204); - __m256i iacc_row_1 = _mm256_blend_epi32(_mm256_shuffle_epi32(iacc_mat_00, 78), iacc_mat_01, 204); - __m256i iacc_row_2 = _mm256_blend_epi32(iacc_mat_10, _mm256_shuffle_epi32(iacc_mat_11, 78), 204); - __m256i iacc_row_3 = _mm256_blend_epi32(_mm256_shuffle_epi32(iacc_mat_10, 78), iacc_mat_11, 204); + // Load and convert to FP32 scale from block_q8_K + const __m256 row_scale_f32 = _mm256_set1_ps((a_ptr[b].d)); - // Load the scale(d) values for all the 4 Q8_0 blocks and repeat it across lanes - const __m256 row_scale_f32 = GGML_F32Cx8_REPEAT_LOAD(a_ptrs[rp][b].d, loadMask); + // Load the scale values for the 8 blocks interleaved in block_q4_Kx8 + // col_scale_f32 rearranged so as to multiply with appropriate quants + const __m256 col_scale_f32 = GGML_F32Cx8_REARRANGE_LOAD(b_ptr[b].d, deltamask); + const __m256 col_dmin_f32 = GGML_F32Cx8_LOAD(b_ptr[b].dmin); - // Multiply with appropiate scales and accumulate - acc_rows[rp * 4] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_0), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 0)), acc_rows[rp * 4]); - acc_rows[rp * 4 + 1] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_1), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 85)), acc_rows[rp * 4 + 1]); - acc_rows[rp * 4 + 2] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_2), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_rows[rp * 4 + 2]); - acc_rows[rp * 4 + 3] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_3), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 255)), acc_rows[rp * 4 + 3]); - } - } + __m256i iacc_b = _mm256_setzero_si256(); + __m256i iacc_min_b = _mm256_setzero_si256(); - // Store the accumulated values - for (int i = 0; i < 16; i++) { - _mm256_storeu_ps((float *)(s + ((y * 4 + i) * bs + x * 8)), acc_rows[i]); - } - } - } + const __m256i q8sums = _mm256_loadu_si256((const __m256i * )(a_ptr[b].bsums)); + __m256i q8s = _mm256_castsi128_si256(_mm_hadd_epi16(_mm256_castsi256_si128(q8sums), _mm256_extracti128_si256(q8sums, 1))); + q8s = _mm256_permute2f128_si256(q8s, q8s, 0); - // Take a block_q8_0x4 structures at each pass of the loop and perform dot product operation - for (; y < nr / 4; y ++) { + // Processes two sub blocks from each Q4_K in each iteration + for (int sb = 0; sb < QK_K / 64; sb++) { - const block_q8_0x4 * a_ptr = a_ptr_start + (y * nb); + // Load the eight block_q4_K for two sub blocks quantized values interleaved with each other in chunks of eight - B0,B1 ....B6,B7 + const __m256i rhs_raw_vec_0123_0 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + sb * 256)); + const __m256i rhs_raw_vec_4567_0 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 32 + sb * 256)); + const __m256i rhs_raw_vec_0123_1 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 64 + sb * 256)); + const __m256i rhs_raw_vec_4567_1 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 96 + sb * 256)); + const __m256i rhs_raw_vec_0123_2 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 128 + sb * 256)); + const __m256i rhs_raw_vec_4567_2 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 160 + sb * 256)); + const __m256i rhs_raw_vec_0123_3 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 192 + sb * 256)); + const __m256i rhs_raw_vec_4567_3 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 224 + sb * 256)); - // Load the eight block_q4_0 quantized values interleaved with each other in chunks of eight - B0,B1 ....B6,B7 - for (int64_t x = xstart; x < nc / 8; x++) { + // 4-bit -> 8-bit + // Values of the first sub block of eight block_q4_K structures for the sb loop + const __m256i rhs_vec_0123_00 = _mm256_and_si256(rhs_raw_vec_0123_0, m4b); + const __m256i rhs_vec_4567_00 = _mm256_and_si256(rhs_raw_vec_4567_0, m4b); + const __m256i rhs_vec_0123_01 = _mm256_and_si256(rhs_raw_vec_0123_1, m4b); + const __m256i rhs_vec_4567_01 = _mm256_and_si256(rhs_raw_vec_4567_1, m4b); + const __m256i rhs_vec_0123_02 = _mm256_and_si256(rhs_raw_vec_0123_2, m4b); + const __m256i rhs_vec_4567_02 = _mm256_and_si256(rhs_raw_vec_4567_2, m4b); + const __m256i rhs_vec_0123_03 = _mm256_and_si256(rhs_raw_vec_0123_3, m4b); + const __m256i rhs_vec_4567_03 = _mm256_and_si256(rhs_raw_vec_4567_3, m4b); - const block_q4_0x8 * b_ptr = b_ptr_start + (x * b_nb); + // Values of the second sub block of eight block_q4_K structures when sb = 1 + const __m256i rhs_vec_0123_10 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_0, 4), m4b); + const __m256i rhs_vec_4567_10 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_0, 4), m4b); + const __m256i rhs_vec_0123_11 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_1, 4), m4b); + const __m256i rhs_vec_4567_11 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_1, 4), m4b); + const __m256i rhs_vec_0123_12 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_2, 4), m4b); + const __m256i rhs_vec_4567_12 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_2, 4), m4b); + const __m256i rhs_vec_0123_13 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_3, 4), m4b); + const __m256i rhs_vec_4567_13 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_3, 4), m4b); - // Master FP accumulators - __m256 acc_rows[4]; - for (int i = 0; i < 4; i++) { - acc_rows[i] = _mm256_setzero_ps(); - } + uint32_t utmp_0[4], utmp_1[4]; - for (int64_t b = 0; b < nb; b++) { - // Load the eight block_q8_0 quantized values interleaved with each other in chunks of eight - B0,B1 ....B6,B7 - const __m256i rhs_raw_mat_0123_0 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs)); - const __m256i rhs_raw_mat_4567_0 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 32)); - const __m256i rhs_raw_mat_0123_1 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 64)); - const __m256i rhs_raw_mat_4567_1 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 96)); + // Scales and Mins of corresponding sub blocks from different Q8_K structures are stored together + // The below block is for eg to extract first sub block's scales and mins from different Q4_K structures for the sb loop + memcpy(utmp_0, b_ptr[b].scales + 24 * sb, 12); + utmp_0[3] = ((utmp_0[2] >> 4) & kmask2) | (((utmp_0[1] >> 6) & kmask3) << 4); + const uint32_t uaux_0 = utmp_0[1] & kmask1; + utmp_0[1] = (utmp_0[2] & kmask2) | (((utmp_0[0] >> 6) & kmask3) << 4); + utmp_0[2] = uaux_0; + utmp_0[0] &= kmask1; - // Save the values in the following vectors in the formats B0B1B4B5, B2B3B6B7 for further processing and storing of valuess - const __m256i rhs_raw_mat_0145_0 = _mm256_blend_epi32(rhs_raw_mat_0123_0, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_0, requiredOrder), 240); - const __m256i rhs_raw_mat_2367_0 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_0, requiredOrder), rhs_raw_mat_4567_0, 240); - const __m256i rhs_raw_mat_0145_1 = _mm256_blend_epi32(rhs_raw_mat_0123_1, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_1, requiredOrder), 240); - const __m256i rhs_raw_mat_2367_1 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_1, requiredOrder), rhs_raw_mat_4567_1, 240); + // The below block is for eg to extract second sub block's scales and mins from different Q4_K structures for the sb loop + memcpy(utmp_1, b_ptr[b].scales + 12 + sb * 24, 12); + utmp_1[3] = ((utmp_1[2] >> 4) & kmask2) | (((utmp_1[1] >> 6) & kmask3) << 4); + const uint32_t uaux_1 = utmp_1[1] & kmask1; + utmp_1[1] = (utmp_1[2] & kmask2) | (((utmp_1[0] >> 6) & kmask3) << 4); + utmp_1[2] = uaux_1; + utmp_1[0] &= kmask1; - // 4-bit -> 8-bit - Sign is maintained - const __m256i rhs_mat_0145_0 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_0145_0, m4b)); //B0(0-7) B1(0-7) B4(0-7) B5(0-7) - const __m256i rhs_mat_2367_0 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_2367_0, m4b)); //B2(0-7) B3(0-7) B6(0-7) B7(0-7) + // Scales of first sub block in the sb loop + const __m128i mins_and_scales_0 = _mm_set_epi32(utmp_0[3], utmp_0[2], utmp_0[1], utmp_0[0]); + __m128i scales_rearrange_0 = _mm_shuffle_epi8(mins_and_scales_0, scalemask); + __m256i scales_0 = _mm256_cvtepu8_epi16(scales_rearrange_0); - const __m256i rhs_mat_0145_1 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_0145_1, m4b)); //B0(8-15) B1(8-15) B4(8-15) B5(8-15) - const __m256i rhs_mat_2367_1 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_2367_1, m4b)); //B2(8-15) B3(8-15) B6(8-15) B7(8-15) + // Scales of second sub block in the sb loop + __m128i mins_and_scales_1 = _mm_set_epi32(utmp_1[3], utmp_1[2], utmp_1[1], utmp_1[0]); + __m128i scales_rearrange_1 = _mm_shuffle_epi8(mins_and_scales_1, scalemask); + __m256i scales_1 = _mm256_cvtepu8_epi16(scales_rearrange_1); - const __m256i rhs_mat_0145_2 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_0, 4), m4b)); //B0(16-23) B1(16-23) B4(16-23) B5(16-23) - const __m256i rhs_mat_2367_2 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_0, 4), m4b)); //B2(16-23) B3(16-23) B6(16-23) B7(16-23) + // Mins of first and second sub block of Q4_K block are arranged side by side + __m256i mins_01 = _mm256_cvtepu8_epi16(_mm_unpacklo_epi8(_mm_shuffle_epi32(mins_and_scales_0, 78), _mm_shuffle_epi32(mins_and_scales_1, 78))); - const __m256i rhs_mat_0145_3 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_1, 4), m4b)); //B0(24-31) B1(24-31) B4(24-31) B5(24-31) - const __m256i rhs_mat_2367_3 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_1, 4), m4b)); //B2(24-31) B3(24-31) B6(24-31) B7(24-31) + // Load the two sub block values corresponding to sb in block_q8_K in batches of 16 bytes and replicate the same across 256 bit vector + __m256i lhs_vec_00 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + sb * 64))); + __m256i lhs_vec_01 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 16 + sb * 64))); + __m256i lhs_vec_10 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 32 + sb * 64))); + __m256i lhs_vec_11 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 48 + sb * 64))); - // Shuffle pattern one - right side input - const __m256i rhs_mat_0145_0_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_0, 136); //B0(0-3) B1(0-3) B0(0-3) B1(0-3) B4(0-3) B5(0-3) B4(0-3) B5(0-3) - const __m256i rhs_mat_2367_0_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_0, 136); //B2(0-3) B3(0-3) B2(0-3) B3(0-3) B6(0-3) B7(0-3) B6(0-3) B7(0-3) + lhs_vec_00 = _mm256_permute2f128_si256(lhs_vec_00, lhs_vec_00, 0); + lhs_vec_01 = _mm256_permute2f128_si256(lhs_vec_01, lhs_vec_01, 0); + lhs_vec_10 = _mm256_permute2f128_si256(lhs_vec_10, lhs_vec_10, 0); + lhs_vec_11 = _mm256_permute2f128_si256(lhs_vec_11, lhs_vec_11, 0); - const __m256i rhs_mat_0145_1_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_1, 136); //B0(8-11) B1(8-11) B0(8-11) B1(8-11) B4(8-11) B5(8-11) B4(8-11) B5(8-11) - const __m256i rhs_mat_2367_1_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_1, 136); //B2(8-11) B3(8-11) B2(8-11) B3(8-11) B6(8-11) B7(8-11) B6(8-11) B7(8-11) + // Dot product done within 32 bit lanes and accumulated in the same vector + // First done for first sub block and thenn for second sub block in each sb + // B0(0-3) B4(0-3) B1(0-3) B5(0-3) B2(0-3) B6(0-3) B3(0-3) B7(0-3) with A0(0-3) + // B0(4-7) B4(4-7) B1(4-7) B5(4-7) B2(4-7) B6(4-7) B3(4-7) B7(4-7) with A0(4-7) + // ........................................................................... + // B0(28-31) B4(28-31) B1(28-31) B5(28-31) B2(28-31) B6(28-31) B3(28-31) B7(28-31) with A0(28-31) - const __m256i rhs_mat_0145_2_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_2, 136); //B0(16-19) B1(16-19) B0(16-19) B1(16-19) B4(16-19) B5(16-19) B4(16-19) B5(16-19) - const __m256i rhs_mat_2367_2_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_2, 136); //B2(16-19) B3(16-19) B2(16-19) B3(16-19) B6(16-19) B7(16-19) B6(16-19) B7(16-19) - const __m256i rhs_mat_0145_3_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_3, 136); //B0(24-27) B1(24-27) B0(24-27) B1(24-27) B4(24-27) B5(24-27) B4(24-27) B5(24-27) - const __m256i rhs_mat_2367_3_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_3, 136); //B2(24-27) B3(24-27) B2(24-27) B3(24-27) B6(24-27) B7(24-27) B6(24-27) B7(24-27) + __m256i iacc_0 = _mm256_setzero_si256(); + __m256i iacc_1 = _mm256_setzero_si256(); - // Shuffle pattern two - right side input + iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_00 ,_mm256_shuffle_epi32(rhs_vec_4567_00, 177), 170), _mm256_shuffle_epi32(lhs_vec_00, 0))); + iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_00, 177) ,rhs_vec_4567_00, 170), _mm256_shuffle_epi32(lhs_vec_00, 85))); - const __m256i rhs_mat_0145_0_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_0, 221); //B0(4-7) B1(4-7) B0(4-7) B1(4-7) B4(4-7) B5(4-7) B4(4-7) B5(4-7) - const __m256i rhs_mat_2367_0_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_0, 221); //B2(4-7) B3(4-7) B2(4-7) B3(4-7) B6(4-7) B7(4-7) B6(4-7) B7(4-7) + iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_01 ,_mm256_shuffle_epi32(rhs_vec_4567_01, 177), 170), _mm256_shuffle_epi32(lhs_vec_00, 170))); + iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_01, 177) ,rhs_vec_4567_01, 170), _mm256_shuffle_epi32(lhs_vec_00, 255))); - const __m256i rhs_mat_0145_1_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_1, 221); //B0(12-15) B1(12-15) B0(12-15) B1(12-15) B4(12-15) B5(12-15) B4(12-15) B5(12-15) - const __m256i rhs_mat_2367_1_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_1, 221); //B2(12-15) B3(12-15) B2(12-15) B3(12-15) B6(12-15) B7(12-15) B6(12-15) B7(12-15) + iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_02 ,_mm256_shuffle_epi32(rhs_vec_4567_02, 177), 170), _mm256_shuffle_epi32(lhs_vec_01, 0))); + iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_02, 177) ,rhs_vec_4567_02, 170), _mm256_shuffle_epi32(lhs_vec_01, 85))); - const __m256i rhs_mat_0145_2_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_2, 221); //B0(20-23) B1(20-23) B0(20-23) B1(20-23) B4(20-23) B5(20-23) B4(20-23) B5(20-23) - const __m256i rhs_mat_2367_2_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_2, 221); //B2(20-23) B3(20-23) B2(20-23) B3(20-23) B6(20-23) B7(20-23) B6(20-23) B7(20-23) + iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_03 ,_mm256_shuffle_epi32(rhs_vec_4567_03, 177), 170), _mm256_shuffle_epi32(lhs_vec_01, 170))); + iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_03, 177) ,rhs_vec_4567_03, 170), _mm256_shuffle_epi32(lhs_vec_01, 255))); - const __m256i rhs_mat_0145_3_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_3, 221); //B0(28-31) B1(28-31) B0(28-31) B1(28-31) B4(28-31) B5(28-31) B4(28-31) B5(28-31) - const __m256i rhs_mat_2367_3_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_3, 221); //B2(28-31) B3(28-31) B2(28-31) B3(28-31) B6(28-31) B7(28-31) B6(28-31) B7(28-31) + iacc_0 = _mm256_madd_epi16(iacc_0, scales_0); - // Scale values - Load the wight scale values of block_q4_0x8 - const __m256 col_scale_f32 = GGML_F32Cx8_LOAD(b_ptr[b].d); + iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_10 ,_mm256_shuffle_epi32(rhs_vec_4567_10, 177), 170), _mm256_shuffle_epi32(lhs_vec_10, 0))); + iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_10, 177) ,rhs_vec_4567_10, 170), _mm256_shuffle_epi32(lhs_vec_10, 85))); - // Load the four block_q4_0 quantized values interleaved with each other in chunks of eight - A0,A1,A2,A3 - // Loaded as set of 128 bit vectors and repeated into a 256 bit vector - __m256i lhs_mat_0123_0 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs))); - __m256i lhs_mat_01_0 = _mm256_permute2f128_si256(lhs_mat_0123_0, lhs_mat_0123_0, 0); - __m256i lhs_mat_23_0 = _mm256_permute2f128_si256(lhs_mat_0123_0, lhs_mat_0123_0, 17); - __m256i lhs_mat_0123_1 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs + 32))); - __m256i lhs_mat_01_1 = _mm256_permute2f128_si256(lhs_mat_0123_1, lhs_mat_0123_1, 0); - __m256i lhs_mat_23_1 = _mm256_permute2f128_si256(lhs_mat_0123_1, lhs_mat_0123_1, 17); - __m256i lhs_mat_0123_2 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs + 64))); - __m256i lhs_mat_01_2 = _mm256_permute2f128_si256(lhs_mat_0123_2, lhs_mat_0123_2, 0); - __m256i lhs_mat_23_2 = _mm256_permute2f128_si256(lhs_mat_0123_2, lhs_mat_0123_2, 17); - __m256i lhs_mat_0123_3 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs + 96))); - __m256i lhs_mat_01_3 = _mm256_permute2f128_si256(lhs_mat_0123_3, lhs_mat_0123_3, 0); - __m256i lhs_mat_23_3 = _mm256_permute2f128_si256(lhs_mat_0123_3, lhs_mat_0123_3, 17); + iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_11 ,_mm256_shuffle_epi32(rhs_vec_4567_11, 177), 170), _mm256_shuffle_epi32(lhs_vec_10, 170))); + iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_11, 177) ,rhs_vec_4567_11, 170), _mm256_shuffle_epi32(lhs_vec_10, 255))); - // Shuffle pattern one - left side input + iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_12 ,_mm256_shuffle_epi32(rhs_vec_4567_12, 177), 170), _mm256_shuffle_epi32(lhs_vec_11, 0))); + iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_12, 177) ,rhs_vec_4567_12, 170), _mm256_shuffle_epi32(lhs_vec_11, 85))); - const __m256i lhs_mat_01_0_sp1 = _mm256_shuffle_epi32(lhs_mat_01_0, 160); //A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3) - const __m256i lhs_mat_23_0_sp1 = _mm256_shuffle_epi32(lhs_mat_23_0, 160); //A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3) + iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_13 ,_mm256_shuffle_epi32(rhs_vec_4567_13, 177), 170), _mm256_shuffle_epi32(lhs_vec_11, 170))); + iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_13, 177) ,rhs_vec_4567_13, 170), _mm256_shuffle_epi32(lhs_vec_11, 255))); - const __m256i lhs_mat_01_1_sp1 = _mm256_shuffle_epi32(lhs_mat_01_1, 160); //A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11) - const __m256i lhs_mat_23_1_sp1 = _mm256_shuffle_epi32(lhs_mat_23_1, 160); //A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11) + iacc_1 = _mm256_madd_epi16(iacc_1, scales_1); - const __m256i lhs_mat_01_2_sp1 = _mm256_shuffle_epi32(lhs_mat_01_2, 160); //A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19) - const __m256i lhs_mat_23_2_sp1 = _mm256_shuffle_epi32(lhs_mat_23_2, 160); //A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19) + // Accumulate the iacc value for one sb + __m256i iacc_sb = _mm256_add_epi32(iacc_0, iacc_1); - const __m256i lhs_mat_01_3_sp1 = _mm256_shuffle_epi32(lhs_mat_01_3, 160); //A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27) - const __m256i lhs_mat_23_3_sp1 = _mm256_shuffle_epi32(lhs_mat_23_3, 160); //A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27) + // Broadcast the bsums of the two sub blocks of the iteration of Q8_K across the vector + // Multiply-Add with corresponding mins of Q4_Kx8 with bsums + __m256i q8s_sb = _mm256_shuffle_epi32(q8s, 0); + __m256i iacc_min_sb = _mm256_madd_epi16(q8s_sb, mins_01); + q8s = _mm256_bsrli_epi128(q8s, 4); - // Shuffle pattern two - left side input + // Accumulate for the complete block + iacc_b = _mm256_add_epi32(iacc_b, iacc_sb); + iacc_min_b = _mm256_add_epi32(iacc_min_b, iacc_min_sb); + } - const __m256i lhs_mat_01_0_sp2 = _mm256_shuffle_epi32(lhs_mat_01_0, 245); //A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7) - const __m256i lhs_mat_23_0_sp2 = _mm256_shuffle_epi32(lhs_mat_23_0, 245); //A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7) + // Multiply-Add with scale values for the complete super block + acc_row = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_b), _mm256_mul_ps(col_scale_f32, row_scale_f32), acc_row); + acc_min_rows = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_min_b), _mm256_mul_ps(col_dmin_f32, row_scale_f32), acc_min_rows); - const __m256i lhs_mat_01_1_sp2 = _mm256_shuffle_epi32(lhs_mat_01_1, 245); //A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15) - const __m256i lhs_mat_23_1_sp2 = _mm256_shuffle_epi32(lhs_mat_23_1, 245); //A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15) + } - const __m256i lhs_mat_01_2_sp2 = _mm256_shuffle_epi32(lhs_mat_01_2, 245); //A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23) - const __m256i lhs_mat_23_2_sp2 = _mm256_shuffle_epi32(lhs_mat_23_2, 245); //A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23) + // Accumulated output values permuted so as to be stored in appropriate order post accumulation + acc_row = _mm256_permutevar8x32_ps(acc_row, finalpermutemask); + _mm256_storeu_ps(s + (y * nr + x * 8), _mm256_sub_ps(acc_row, acc_min_rows)); + } + } - const __m256i lhs_mat_01_3_sp2 = _mm256_shuffle_epi32(lhs_mat_01_3, 245); //A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31) - const __m256i lhs_mat_23_3_sp2 = _mm256_shuffle_epi32(lhs_mat_23_3, 245); //A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31) +#else + UNUSED(kmask1); + UNUSED(kmask2); + UNUSED(kmask3); + ggml_gemv_q4_K_8x8_q8_K_generic(n, s, bs, vx, vy, nr, nc); +#endif +} - // The values arranged in shuffle patterns are operated with dot product operation within 32 bit lane i.e corresponding bytes and multiplied and added into 32 bit integers within 32 bit lane - // Resembles MMLAs into 2x2 matrices in ARM Version - const __m256i zero = _mm256_setzero_si256(); - __m256i iacc_mat_00_sp1 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_01_3_sp1, rhs_mat_0145_3_sp1), lhs_mat_01_2_sp1, rhs_mat_0145_2_sp1), lhs_mat_01_1_sp1, rhs_mat_0145_1_sp1), lhs_mat_01_0_sp1, rhs_mat_0145_0_sp1); - __m256i iacc_mat_01_sp1 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_01_3_sp1, rhs_mat_2367_3_sp1), lhs_mat_01_2_sp1, rhs_mat_2367_2_sp1), lhs_mat_01_1_sp1, rhs_mat_2367_1_sp1), lhs_mat_01_0_sp1, rhs_mat_2367_0_sp1); - __m256i iacc_mat_10_sp1 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_23_3_sp1, rhs_mat_0145_3_sp1), lhs_mat_23_2_sp1, rhs_mat_0145_2_sp1), lhs_mat_23_1_sp1, rhs_mat_0145_1_sp1), lhs_mat_23_0_sp1, rhs_mat_0145_0_sp1); - __m256i iacc_mat_11_sp1 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_23_3_sp1, rhs_mat_2367_3_sp1), lhs_mat_23_2_sp1, rhs_mat_2367_2_sp1), lhs_mat_23_1_sp1, rhs_mat_2367_1_sp1), lhs_mat_23_0_sp1, rhs_mat_2367_0_sp1); - __m256i iacc_mat_00_sp2 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_01_3_sp2, rhs_mat_0145_3_sp2), lhs_mat_01_2_sp2, rhs_mat_0145_2_sp2), lhs_mat_01_1_sp2, rhs_mat_0145_1_sp2), lhs_mat_01_0_sp2, rhs_mat_0145_0_sp2); - __m256i iacc_mat_01_sp2 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_01_3_sp2, rhs_mat_2367_3_sp2), lhs_mat_01_2_sp2, rhs_mat_2367_2_sp2), lhs_mat_01_1_sp2, rhs_mat_2367_1_sp2), lhs_mat_01_0_sp2, rhs_mat_2367_0_sp2); - __m256i iacc_mat_10_sp2 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_23_3_sp2, rhs_mat_0145_3_sp2), lhs_mat_23_2_sp2, rhs_mat_0145_2_sp2), lhs_mat_23_1_sp2, rhs_mat_0145_1_sp2), lhs_mat_23_0_sp2, rhs_mat_0145_0_sp2); - __m256i iacc_mat_11_sp2 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_23_3_sp2, rhs_mat_2367_3_sp2), lhs_mat_23_2_sp2, rhs_mat_2367_2_sp2), lhs_mat_23_1_sp2, rhs_mat_2367_1_sp2), lhs_mat_23_0_sp2, rhs_mat_2367_0_sp2); +void ggml_gemv_iq4_nl_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) { +#if defined(__AVX2__) + { + __m256i signextendlut = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i*)kvalues_iq4nl)); + signextendlut = _mm256_permute2f128_si256(signextendlut, signextendlut, 0); - // Output of both shuffle patterns are added in order to sum dot product outputs of all 32 values in block - __m256i iacc_mat_00 = _mm256_add_epi32(iacc_mat_00_sp1, iacc_mat_00_sp2); - __m256i iacc_mat_01 = _mm256_add_epi32(iacc_mat_01_sp1, iacc_mat_01_sp2); - __m256i iacc_mat_10 = _mm256_add_epi32(iacc_mat_10_sp1, iacc_mat_10_sp2); - __m256i iacc_mat_11 = _mm256_add_epi32(iacc_mat_11_sp1, iacc_mat_11_sp2); + gemv_q4_b32_8x8_q8_0_lut_avx(n, s, bs, vx, vy, nr, nc, signextendlut); + return; + } +#endif - // Straighten out to make 4 row vectors - __m256i iacc_row_0 = _mm256_blend_epi32(iacc_mat_00, _mm256_shuffle_epi32(iacc_mat_01, 78), 204); - __m256i iacc_row_1 = _mm256_blend_epi32(_mm256_shuffle_epi32(iacc_mat_00, 78), iacc_mat_01, 204); - __m256i iacc_row_2 = _mm256_blend_epi32(iacc_mat_10, _mm256_shuffle_epi32(iacc_mat_11, 78), 204); - __m256i iacc_row_3 = _mm256_blend_epi32(_mm256_shuffle_epi32(iacc_mat_10, 78), iacc_mat_11, 204); + ggml_gemv_iq4_nl_8x8_q8_0_generic(n, s, bs, vx, vy, nr, nc); +} - // Load the scale(d) values for all the 4 Q8_0 blocks and repeat it across lanes - const __m256 row_scale_f32 = GGML_F32Cx8_REPEAT_LOAD(a_ptr[b].d, loadMask); +void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) { +#if defined(__AVX2__) || defined(__AVX512F__) + { + // Lookup table to convert signed nibbles to signed bytes + __m256i signextendlut = _mm256_castsi128_si256(_mm_set_epi8(-1, -2, -3, -4, -5, -6, -7, -8, 7, 6, 5, 4, 3, 2, 1, 0)); + signextendlut = _mm256_permute2f128_si256(signextendlut, signextendlut, 0); - // Multiply with appropiate scales and accumulate - acc_rows[0] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_0), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 0)), acc_rows[0]); - acc_rows[1] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_1), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 85)), acc_rows[1]); - acc_rows[2] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_2), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_rows[2]); - acc_rows[3] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_3), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 255)), acc_rows[3]); - } + gemm_q4_b32_8x8_q8_0_lut_avx(n, s, bs, vx, vy, nr, nc, signextendlut); - // Store the accumulated values - for (int i = 0; i < 4; i++) { - _mm256_storeu_ps((float *)(s + ((y * 4 + i) * bs + x * 8)), acc_rows[i]); - } - } - } return; } +#endif // defined(__AVX2__) || defined(__AVX512F__) -#endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__) ggml_gemm_q4_0_8x8_q8_0_generic(n, s, bs, vx, vy, nr, nc); } @@ -3050,3 +3096,18 @@ void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo ggml_gemm_q4_K_8x8_q8_K_generic(n, s, bs, vx, vy, nr, nc); #endif } + +void ggml_gemm_iq4_nl_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) { +#if defined(__AVX2__) || defined(__AVX512F__) + { + __m256i signextendlut = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i*)kvalues_iq4nl)); + signextendlut = _mm256_permute2f128_si256(signextendlut, signextendlut, 0); + + gemm_q4_b32_8x8_q8_0_lut_avx(n, s, bs, vx, vy, nr, nc, signextendlut); + + return; + } +#endif // defined(__AVX2__) || defined(__AVX512F__) + + ggml_gemm_iq4_nl_8x8_q8_0_generic(n, s, bs, vx, vy, nr, nc); +} diff --git a/ggml/src/ggml-cpu/repack.cpp b/ggml/src/ggml-cpu/repack.cpp index 74c1c029b946b..26132786f4678 100644 --- a/ggml/src/ggml-cpu/repack.cpp +++ b/ggml/src/ggml-cpu/repack.cpp @@ -206,8 +206,9 @@ void ggml_gemv_q4_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const int ncols_interleaved = 4; const int blocklen = 4; - assert (n % qk == 0); - assert (nc % ncols_interleaved == 0); + assert(nr == 1); + assert(n % qk == 0); + assert(nc % ncols_interleaved == 0); UNUSED(s); UNUSED(bs); @@ -307,30 +308,28 @@ void ggml_gemv_q4_0_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, UNUSED(ncols_interleaved); UNUSED(blocklen); - { - float sumf[8]; - int sumi; + float sumf[8]; + int sumi; - const block_q8_0 * a_ptr = (const block_q8_0 *) vy; - for (int x = 0; x < nc / ncols_interleaved; x++) { - const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb); + const block_q8_0 * a_ptr = (const block_q8_0 *) vy; + for (int x = 0; x < nc / ncols_interleaved; x++) { + const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb); - for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0; - for (int l = 0; l < nb; l++) { - for (int k = 0; k < (qk / (2 * blocklen)); k++) { - for (int j = 0; j < ncols_interleaved; j++) { - sumi = 0; - for (int i = 0; i < blocklen; ++i) { - const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4); - const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0); - sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4; - } - sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d); + for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0; + for (int l = 0; l < nb; l++) { + for (int k = 0; k < (qk / (2 * blocklen)); k++) { + for (int j = 0; j < ncols_interleaved; j++) { + sumi = 0; + for (int i = 0; i < blocklen; ++i) { + const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4); + const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0); + sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4; } + sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d); } } - for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j]; } + for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j]; } } @@ -418,43 +417,73 @@ void ggml_gemv_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs const int ncols_interleaved = 4; const int blocklen = 4; - assert (n % qk == 0); - assert (nc % ncols_interleaved == 0); + assert(nr == 1); + assert(n % qk == 0); + assert(nc % ncols_interleaved == 0); - UNUSED(s); UNUSED(bs); - UNUSED(vx); - UNUSED(vy); UNUSED(nr); - UNUSED(nc); - UNUSED(nb); - UNUSED(ncols_interleaved); - UNUSED(blocklen); - { - float sumf[4]; - int sumi; + float sumf[4]; + int sumi; - const block_q8_0 * a_ptr = (const block_q8_0 *) vy; - for (int x = 0; x < nc / ncols_interleaved; x++) { - const block_iq4_nlx4 * b_ptr = (const block_iq4_nlx4 *) vx + (x * nb); + const block_q8_0 * a_ptr = (const block_q8_0 *) vy; + for (int x = 0; x < nc / ncols_interleaved; x++) { + const block_iq4_nlx4 * b_ptr = (const block_iq4_nlx4 *) vx + (x * nb); - for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0; - for (int l = 0; l < nb; l++) { - for (int k = 0; k < (qk / (2 * blocklen)); k++) { - for (int j = 0; j < ncols_interleaved; j++) { - sumi = 0; - for (int i = 0; i < blocklen; ++i) { - const int v0 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0x0F]; - const int v1 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4]; - sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])); - } - sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d); + for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0; + for (int l = 0; l < nb; l++) { + for (int k = 0; k < (qk / (2 * blocklen)); k++) { + for (int j = 0; j < ncols_interleaved; j++) { + sumi = 0; + for (int i = 0; i < blocklen; ++i) { + const int v0 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0x0F]; + const int v1 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4]; + sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])); } + sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d); } } - for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j]; } + for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j]; + } +} + +void ggml_gemv_iq4_nl_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) { + const int qk = QK8_0; + const int nb = n / qk; + const int ncols_interleaved = 8; + const int blocklen = 8; + + assert(nr == 1); + assert(n % qk == 0); + assert(nc % ncols_interleaved == 0); + + UNUSED(bs); + UNUSED(nr); + + float sumf[8]; + int sumi; + + const block_q8_0 * a_ptr = (const block_q8_0 *) vy; + for (int x = 0; x < nc / ncols_interleaved; x++) { + const block_iq4_nlx8 * b_ptr = (const block_iq4_nlx8 *) vx + (x * nb); + + for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0; + for (int l = 0; l < nb; l++) { + for (int k = 0; k < (qk / (2 * blocklen)); k++) { + for (int j = 0; j < ncols_interleaved; j++) { + sumi = 0; + for (int i = 0; i < blocklen; ++i) { + const int v0 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0x0F]; + const int v1 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4]; + sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])); + } + sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d); + } + } + } + for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j]; } } @@ -767,6 +796,50 @@ void ggml_gemm_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs } } +void ggml_gemm_iq4_nl_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) { + const int qk = QK8_0; + const int nb = n / qk; + const int ncols_interleaved = 8; + const int blocklen = 8; + + assert(n % qk == 0); + assert(nr % 4 == 0); + assert(nc % ncols_interleaved == 0); + + float sumf[4][8]; + int sumi; + + for (int y = 0; y < nr / 4; y++) { + const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb); + for (int x = 0; x < nc / ncols_interleaved; x++) { + const block_iq4_nlx8 * b_ptr = (const block_iq4_nlx8 *) vx + (x * nb); + for (int m = 0; m < 4; m++) { + for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0; + } + for (int l = 0; l < nb; l++) { + for (int k = 0; k < (qk / (2 * blocklen)); k++) { + for (int m = 0; m < 4; m++) { + for (int j = 0; j < ncols_interleaved; j++) { + sumi = 0; + for (int i = 0; i < blocklen; ++i) { + const int v0 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0x0F]; + const int v1 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4]; + sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) + + (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])); + } + sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]); + } + } + } + } + for (int m = 0; m < 4; m++) { + for (int j = 0; j < ncols_interleaved; j++) + s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j]; + } + } + } +} + } // extern "C" static block_q4_0x4 make_block_q4_0x4(block_q4_0 * in, unsigned int blck_size_interleave) { @@ -1043,15 +1116,16 @@ static block_iq4_nlx4 make_block_iq4_nlx4(block_iq4_nl * in, unsigned int blck_s static int repack_iq4_nl_to_iq4_nl_4_bl(struct ggml_tensor * t, int interleave_block, const void * GGML_RESTRICT data, size_t data_size) { GGML_ASSERT(t->type == GGML_TYPE_IQ4_NL); - //GGML_ASSERT(interleave_block == 4 || interleave_block == 8); GGML_ASSERT(interleave_block == 4); - block_iq4_nlx4 * dst = (block_iq4_nlx4 *)t->data; - const block_iq4_nl * src = (const block_iq4_nl *)data; + const block_iq4_nl * src = (const block_iq4_nl *)data; + block_iq4_nlx4 * dst = ( block_iq4_nlx4 *)t->data; + block_iq4_nl dst_tmp[4]; + int nrow = ggml_nrows(t); int nrows_interleaved = 4; - int nblocks = t->ne[0] / QK4_0; + int nblocks = t->ne[0] / QK4_NL; GGML_ASSERT(data_size == nrow * nblocks * sizeof(block_iq4_nl)); @@ -1073,6 +1147,63 @@ static int repack_iq4_nl_to_iq4_nl_4_bl(struct ggml_tensor * t, int interleave_b GGML_UNUSED(data_size); } +static block_iq4_nlx8 make_block_iq4_nlx8(block_iq4_nl * in, unsigned int blck_size_interleave) { + block_iq4_nlx8 out; + + for (int i = 0; i < 8; i++) { + out.d[i] = in[i].d; + } + + const int end = QK4_NL * 4 / blck_size_interleave; + + if (blck_size_interleave == 8) { + for (int i = 0; i < end; ++i) { + int src_id = i % 8; + int src_offset = (i / 8) * blck_size_interleave; + int dst_offset = i * blck_size_interleave; + + memcpy(&out.qs[dst_offset], &in[src_id].qs[src_offset], sizeof(uint64_t)); + } + } else { + GGML_ASSERT(false); + } + + return out; +} + +static int repack_iq4_nl_to_iq4_nl_8_bl(struct ggml_tensor * t, int interleave_block, const void * GGML_RESTRICT data, size_t data_size) { + GGML_ASSERT(t->type == GGML_TYPE_IQ4_NL); + GGML_ASSERT(interleave_block == 8); + + const block_iq4_nl * src = (const block_iq4_nl *)data; + block_iq4_nlx8 * dst = ( block_iq4_nlx8 *)t->data; + + block_iq4_nl dst_tmp[8]; + + int nrow = ggml_nrows(t); + int nrows_interleaved = 8; + int nblocks = t->ne[0] / QK4_NL; + + GGML_ASSERT(data_size == nrow * nblocks * sizeof(block_iq4_nl)); + + if (t->ne[1] % nrows_interleaved != 0) { + return -1; + } + + for (int b = 0; b < nrow; b += nrows_interleaved) { + for (int64_t x = 0; x < nblocks; x++) { + for (int i = 0; i < nrows_interleaved; i++) { + dst_tmp[i] = src[x + i * nblocks]; + } + *dst++ = make_block_iq4_nlx8(dst_tmp, interleave_block); + } + src += nrows_interleaved * nblocks; + } + return 0; + + GGML_UNUSED(data_size); +} + namespace ggml::cpu::repack { // repack template @@ -1104,6 +1235,10 @@ template <> int repack(struct ggml_tensor * t, const void * // return repack_iq4_nl_to_iq4_nl_4_bl(t, 8, data, data_size); //} +template <> int repack(struct ggml_tensor * t, const void * data, size_t data_size) { + return repack_iq4_nl_to_iq4_nl_8_bl(t, 8, data, data_size); +} + // gemv template void gemv(int, float *, size_t, const void *, const void *, int, int); @@ -1128,6 +1263,10 @@ template <> void gemv(int n, float * s, size ggml_gemv_iq4_nl_4x4_q8_0(n, s, bs, vx, vy, nr, nc); } +template <> void gemv(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { + ggml_gemv_iq4_nl_8x8_q8_0(n, s, bs, vx, vy, nr, nc); +} + // gemm template void gemm(int, float *, size_t, const void *, const void *, int, int); @@ -1152,6 +1291,10 @@ template <> void gemm(int n, float * s, size ggml_gemm_iq4_nl_4x4_q8_0(n, s, bs, vx, vy, nr, nc); } +template <> void gemm(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { + ggml_gemm_iq4_nl_8x8_q8_0(n, s, bs, vx, vy, nr, nc); +} + class tensor_traits_base : public ggml::cpu::tensor_traits { public: virtual int repack(struct ggml_tensor * t, const void * data, size_t data_size) = 0; @@ -1423,6 +1566,7 @@ static const ggml::cpu::tensor_traits * ggml_repack_get_optimal_repack_type(cons // instance for IQ4 static const ggml::cpu::repack::tensor_traits iq4_nl_4x4_q8_0; + static const ggml::cpu::repack::tensor_traits iq4_nl_8x8_q8_0; if (cur->type == GGML_TYPE_Q4_0) { if (ggml_cpu_has_avx2() || (ggml_cpu_has_sve() && ggml_cpu_has_matmul_int8() && ggml_cpu_get_sve_cnt() == QK8_0)) { @@ -1447,6 +1591,11 @@ static const ggml::cpu::tensor_traits * ggml_repack_get_optimal_repack_type(cons } } } else if (cur->type == GGML_TYPE_IQ4_NL) { + if (ggml_cpu_has_avx2()) { + if (cur->ne[1] % 8 == 0) { + return &iq4_nl_8x8_q8_0; + } + } if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) { if (cur->ne[1] % 4 == 0) { return &iq4_nl_4x4_q8_0; diff --git a/ggml/src/ggml-cpu/repack.h b/ggml/src/ggml-cpu/repack.h index 4421e5f8e7046..0a3457308b5a0 100644 --- a/ggml/src/ggml-cpu/repack.h +++ b/ggml/src/ggml-cpu/repack.h @@ -60,6 +60,13 @@ struct block_iq4_nlx4 { static_assert(sizeof(block_iq4_nlx4) == 4 * sizeof(ggml_half) + QK4_NL * 2, "wrong iq4_nlx4 block size/padding"); +struct block_iq4_nlx8 { + ggml_half d[8]; // deltas for 8 iq4_nl blocks + uint8_t qs[QK4_NL * 4]; // nibbles / quants for 8 iq4_nl blocks +}; + +static_assert(sizeof(block_iq4_nlx8) == 8 * sizeof(ggml_half) + QK4_NL * 4, "wrong iq4_nlx8 block size/padding"); + #if defined(__cplusplus) extern "C" { #endif @@ -72,11 +79,13 @@ void ggml_gemv_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); void ggml_gemv_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); void ggml_gemv_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); +void ggml_gemv_iq4_nl_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); void ggml_gemm_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); void ggml_gemm_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); void ggml_gemm_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); +void ggml_gemm_iq4_nl_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); // Native implementations void ggml_quantize_mat_q8_0_4x4_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k); @@ -87,11 +96,13 @@ void ggml_gemv_q4_0_4x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, void ggml_gemv_q4_0_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); void ggml_gemv_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); void ggml_gemv_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); +void ggml_gemv_iq4_nl_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); void ggml_gemm_q4_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); void ggml_gemm_q4_0_4x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); void ggml_gemm_q4_0_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); void ggml_gemm_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); void ggml_gemm_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); +void ggml_gemm_iq4_nl_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc); #if defined(__cplusplus) } // extern "C"