diff --git a/convert_hf_to_gguf.py b/convert_hf_to_gguf.py index 2bf97475f78..89522dee8b8 100755 --- a/convert_hf_to_gguf.py +++ b/convert_hf_to_gguf.py @@ -4422,6 +4422,10 @@ def set_vocab(self): self._set_vocab_gpt2() def set_gguf_parameters(self): + + # note: deepseek2 using MLA converts into MQA (ie: GQA with 1 group) + self.hparams["num_key_value_heads"] = 1 + super().set_gguf_parameters() hparams = self.hparams @@ -4430,8 +4434,13 @@ def set_gguf_parameters(self): if "q_lora_rank" in hparams and hparams["q_lora_rank"] is not None: self.gguf_writer.add_q_lora_rank(hparams["q_lora_rank"]) self.gguf_writer.add_kv_lora_rank(hparams["kv_lora_rank"]) - self.gguf_writer.add_key_length(hparams["qk_nope_head_dim"] + hparams["qk_rope_head_dim"]) - self.gguf_writer.add_value_length(hparams["v_head_dim"]) + + # note: deepseek2 using MLA converts into MQA with larger heads, then decompresses to MHA + self.gguf_writer.add_key_length(hparams["kv_lora_rank"] + hparams["qk_rope_head_dim"]) + self.gguf_writer.add_value_length(hparams["kv_lora_rank"]) + self.gguf_writer.add_key_length_mla(hparams["qk_nope_head_dim"] + hparams["qk_rope_head_dim"]) + self.gguf_writer.add_value_length_mla(hparams["v_head_dim"]) + self.gguf_writer.add_expert_feed_forward_length(hparams["moe_intermediate_size"]) self.gguf_writer.add_expert_count(hparams["n_routed_experts"]) self.gguf_writer.add_expert_shared_count(hparams["n_shared_experts"]) @@ -4500,6 +4509,26 @@ def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iter else: return [] + # note: MLA with the absorption optimization, needs these two split and k_b_proj transposed + if name.endswith("kv_b_proj.weight"): + name_kb = name.replace("kv_b_proj", "k_b_proj") + name_vb = name.replace("kv_b_proj", "v_b_proj") + + n_head_kv = self.hparams["num_key_value_heads"] + v_head_dim = self.hparams["v_head_dim"] + qk_nope_head_dim = self.hparams["qk_nope_head_dim"] + + assert data_torch.shape[0] == n_head_kv * (v_head_dim + qk_nope_head_dim) + + kv_b = data_torch.view(n_head_kv, v_head_dim + qk_nope_head_dim, data_torch.shape[-1]) + k_b, v_b = torch.split(kv_b, [qk_nope_head_dim, v_head_dim], dim=1) + k_b = k_b.transpose(1, 2) + + return [ + (self.map_tensor_name(name_kb), k_b), + (self.map_tensor_name(name_vb), v_b) + ] + return [(self.map_tensor_name(name), data_torch)] def prepare_tensors(self): diff --git a/ggml/src/ggml-cpu/ggml-cpu-aarch64.cpp b/ggml/src/ggml-cpu/ggml-cpu-aarch64.cpp index ced09c05c2c..175cba329b7 100644 --- a/ggml/src/ggml-cpu/ggml-cpu-aarch64.cpp +++ b/ggml/src/ggml-cpu/ggml-cpu-aarch64.cpp @@ -4044,7 +4044,7 @@ static void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, c UNUSED(ncols_interleaved); UNUSED(blocklen); -#if defined(__AVX2__) +#if defined(__AVX2__) || defined(__AVX512F__) const block_q4_Kx8 * b_ptr_start = (const block_q4_Kx8 * ) vx; const block_q8_Kx4 * a_ptr_start = (const block_q8_Kx4 * ) vy; int64_t b_nb = n / QK_K; @@ -4054,8 +4054,748 @@ static void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, c const __m256i m4b = _mm256_set1_epi8(0x0F); // 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); + //Take group of four block_q8_Kx4 structures at each pass of the loop and perform dot product operation + for (; y < anr / 4; y += 4) { + + const block_q8_Kx4 * 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 eight block_q4_kx8 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_Kx8 * b_ptr_0 = b_ptr_start + ((x) * b_nb); + const block_q4_Kx8 * 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(); + } + + __m512 acc_min_rows[16]; + for (int i = 0; i < 16; i++) { + acc_min_rows[i] = _mm512_setzero_ps(); + } + + // For super block + for (int64_t b = 0; b < nb; b++) { + // Scale values - Load the sixteen scale values from two block_q4_kx8 structures + const __m512 col_scale_f32 = GGML_F32Cx8x2_LOAD(b_ptr_0[b].d, b_ptr_1[b].d); + + // dmin values - Load the sixteen dmin values from two block_q4_kx8 structures + const __m512 col_dmin_f32 = GGML_F32Cx8x2_LOAD(b_ptr_0[b].dmin, b_ptr_1[b].dmin); + + // Loop to iterate over the eight sub blocks of a super block - two sub blocks are processed per iteration + for (int sb = 0; sb < QK_K / 64; sb++) { + + const __m256i rhs_raw_mat_0123_0 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + sb * 256)); + const __m256i rhs_raw_mat_4567_0 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 32 + sb * 256)); + const __m256i rhs_raw_mat_0123_1 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 64 + sb * 256)); + const __m256i rhs_raw_mat_4567_1 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 96 + sb * 256)); + const __m256i rhs_raw_mat_0123_2 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 128 + sb * 256)); + const __m256i rhs_raw_mat_4567_2 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 160 + sb * 256)); + const __m256i rhs_raw_mat_0123_3 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 192 + sb * 256)); + const __m256i rhs_raw_mat_4567_3 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 224 + sb * 256)); + + const __m256i rhs_raw_mat_89AB_0 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + sb * 256)); + const __m256i rhs_raw_mat_CDEF_0 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 32 + sb * 256)); + const __m256i rhs_raw_mat_89AB_1 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 64 + sb * 256)); + const __m256i rhs_raw_mat_CDEF_1 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 96 + sb * 256)); + const __m256i rhs_raw_mat_89AB_2 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 128 + sb * 256)); + const __m256i rhs_raw_mat_CDEF_2 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 160 + sb * 256)); + const __m256i rhs_raw_mat_89AB_3 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 192 + sb * 256)); + const __m256i rhs_raw_mat_CDEF_3 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 224 + sb * 256)); + + 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_0145_2 = _mm256_blend_epi32(rhs_raw_mat_0123_2, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_2, requiredOrder), 240); + const __m256i rhs_raw_mat_2367_2 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_2, requiredOrder), rhs_raw_mat_4567_2, 240); + const __m256i rhs_raw_mat_0145_3 = _mm256_blend_epi32(rhs_raw_mat_0123_3, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_3, requiredOrder), 240); + const __m256i rhs_raw_mat_2367_3 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_3, requiredOrder), rhs_raw_mat_4567_3, 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 __m256i rhs_raw_mat_89CD_2 = _mm256_blend_epi32(rhs_raw_mat_89AB_2, _mm256_permutevar8x32_epi32(rhs_raw_mat_CDEF_2, requiredOrder), 240); + const __m256i rhs_raw_mat_ABEF_2 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_89AB_2, requiredOrder), rhs_raw_mat_CDEF_2, 240); + const __m256i rhs_raw_mat_89CD_3 = _mm256_blend_epi32(rhs_raw_mat_89AB_3, _mm256_permutevar8x32_epi32(rhs_raw_mat_CDEF_3, requiredOrder), 240); + const __m256i rhs_raw_mat_ABEF_3 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_89AB_3, requiredOrder), rhs_raw_mat_CDEF_3, 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); + + const __m512i rhs_raw_mat_014589CD_2 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_2), rhs_raw_mat_89CD_2, 1); + const __m512i rhs_raw_mat_2367ABEF_2 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_2), rhs_raw_mat_ABEF_2, 1); + const __m512i rhs_raw_mat_014589CD_3 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_3), rhs_raw_mat_89CD_3, 1); + const __m512i rhs_raw_mat_2367ABEF_3 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_3), rhs_raw_mat_ABEF_3, 1); + + //4-bit -> 8-bit + const __m512i rhs_mat_014589CD_00 = _mm512_and_si512(rhs_raw_mat_014589CD_0, m4bexpanded); //B00(0-7) B01(0-7) B04(0-7) B05(0-7) B08(0-7) B09(0-7) B0C(0-7) B0D(0-7) + const __m512i rhs_mat_2367ABEF_00 = _mm512_and_si512(rhs_raw_mat_2367ABEF_0, m4bexpanded); //B02(0-7) B03(0-7) B06(0-7) B07(0-7) B0A(0-7) B0B(0-7) B0E(0-7) B0F(0-7) + const __m512i rhs_mat_014589CD_01 = _mm512_and_si512(rhs_raw_mat_014589CD_1, m4bexpanded); //B00(8-15) B01(8-15) B04(8-15) B05(8-15) B08(8-15) B09(8-15) B0C(8-15) B0D(8-15) + const __m512i rhs_mat_2367ABEF_01 = _mm512_and_si512(rhs_raw_mat_2367ABEF_1, m4bexpanded); //B02(8-15) B03(8-15) B06(8-15) B07(8-15) B0A(8-15) B0B(8-15) B0E(8-15) B0F(8-15) + + const __m512i rhs_mat_014589CD_02 = _mm512_and_si512(rhs_raw_mat_014589CD_2, m4bexpanded); //B00(16-23) B01(16-23) B04(16-23) B05(16-23) B08(16-23) B09(16-23) B0C(16-23) B0D(16-23) + const __m512i rhs_mat_2367ABEF_02 = _mm512_and_si512(rhs_raw_mat_2367ABEF_2, m4bexpanded); //B02(16-23) B03(16-23) B06(16-23) B07(16-23) B0A(16-23) B0B(16-23) B0E(16-23) B0F(16-23) + const __m512i rhs_mat_014589CD_03 = _mm512_and_si512(rhs_raw_mat_014589CD_3, m4bexpanded); //B00(24-31) B01(24-31) B04(24-31) B05(24-31) B08(24-31) B09(24-31) B0C(24-31) B0D(24-31) + const __m512i rhs_mat_2367ABEF_03 = _mm512_and_si512(rhs_raw_mat_2367ABEF_3, m4bexpanded); //B02(24-31) B03(24-31) B06(24-31) B07(24-31) B0A(24-31) B0B(24-31) B0E(24-31) B0F(24-31) + + const __m512i rhs_mat_014589CD_10 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_0, 4), m4bexpanded); //B10(0-7) B11(0-7) B14(0-7) B15(0-7) B18(0-7) B19(0-7) B1C(0-7) B1D(0-7) + const __m512i rhs_mat_2367ABEF_10 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_0, 4), m4bexpanded); //B12(0-7) B13(0-7) B16(0-7) B17(0-7) B1A(0-7) B1B(0-7) B1E(0-7) B1F(0-7) + const __m512i rhs_mat_014589CD_11 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_1, 4), m4bexpanded); //B10(8-15) B11(8-15) B14(8-15) B15(8-15) B18(8-15) B19(8-15) B1C(8-15) B1D(8-15) + const __m512i rhs_mat_2367ABEF_11 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_1, 4), m4bexpanded); //B12(8-15) B13(8-15) B16(8-15) B17(8-15) B1A(8-15) B1B(8-15) B1E(8-15) B1F(8-15) + + const __m512i rhs_mat_014589CD_12 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_2, 4), m4bexpanded); //B10(16-23) B11(16-23) B14(16-23) B15(16-23) B18(16-23) B19(16-23) B1C(16-23) B1D(16-23) + const __m512i rhs_mat_2367ABEF_12 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_2, 4), m4bexpanded); //B12(16-23) B13(16-23) B16(16-23) B17(16-23) B1A(16-23) B1B(16-23) B1E(16-23) B1F(16-23) + const __m512i rhs_mat_014589CD_13 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_3, 4), m4bexpanded); //B10(24-31) B11(24-31) B14(24-31) B15(24-31) B18(24-31) B19(24-31) B1C(24-31) B1D(24-31) + const __m512i rhs_mat_2367ABEF_13 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_3, 4), m4bexpanded); //B12(24-31) B13(24-31) B16(24-31) B17(24-31) B1A(24-31) B1B(24-31) B1E(24-31) B1F(24-31) + + // Shuffle pattern one - right side input + const __m512i rhs_mat_014589CD_00_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_00, (_MM_PERM_ENUM)136); //B00(0-3) B01(0-3) B00(0-3) B01(0-3) B04(0-3) B05(0-3) B04(0-3) B05(0-3) B08(0-3) B09(0-3) B08(0-3) B09(0-3) B0C(0-3) B0D(0-3) B0C(0-3) B0D(0-3) + const __m512i rhs_mat_2367ABEF_00_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_00, (_MM_PERM_ENUM)136); //B02(0-3) B03(0-3) B02(0-3) B03(0-3) B06(0-3) B07(0-3) B06(0-3) B07(0-3) B0A(0-3) B0B(0-3) B0A(0-3) B0B(0-3) B0E(0-3) B0F(0-3) B0E(0-3) B0F(0-3) + const __m512i rhs_mat_014589CD_01_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_01, (_MM_PERM_ENUM)136); //B00(8-11) B01(8-11) B00(8-11) B01(8-11) B04(8-11) B05(8-11) B04(8-11) B05(8-11) B08(8-11) B09(8-11) B08(8-11) B09(8-11) B0C(8-11) B0D(8-11) B0C(8-11) B0D(8-11) + const __m512i rhs_mat_2367ABEF_01_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_01, (_MM_PERM_ENUM)136); //B02(8-11) B03(8-11) B02(8-11) B03(8-11) B06(8-11) B07(8-11) B06(8-11) B07(8-11) B0A(8-11) B0B(8-11) B0A(8-11) B0B(8-11) B0E(8-11) B0F(8-11) B0E(8-11) B0F(8-11) + const __m512i rhs_mat_014589CD_02_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_02, (_MM_PERM_ENUM)136); //B00(16-19) B01(16-19) B00(16-19) B01(16-19) B04(16-19) B05(16-19) B04(16-19) B05(16-19) B08(16-19) B09(16-19) B08(16-19) B09(16-19) B0C(16-19) B0D(16-19) B0C(16-19) B0D(16-19) + const __m512i rhs_mat_2367ABEF_02_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_02, (_MM_PERM_ENUM)136); //B02(16-19) B03(16-19) B02(16-19) B03(16-19) B06(16-19) B07(16-19) B06(16-19) B07(16-19) B0A(16-19) B0B(16-19) B0A(16-19) B0B(16-19) B0E(16-19) B0F(16-19) B0E(16-19) B0F(16-19) + const __m512i rhs_mat_014589CD_03_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_03, (_MM_PERM_ENUM)136); //B00(24-27) B01(24-27) B00(24-27) B01(24-27) B04(24-27) B05(24-27) B04(24-27) B05(24-27) B08(24-27) B09(24-27) B08(24-27) B09(24-27) B0C(24-27) B0D(24-27) B0C(24-27) B0D(24-27) + const __m512i rhs_mat_2367ABEF_03_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_03, (_MM_PERM_ENUM)136); //B02(24-27) B03(24-27) B02(24-27) B03(24-27) B06(24-27) B07(24-27) B06(24-27) B07(24-27) B0A(24-27) B0B(24-27) B0A(24-27) B0B(24-27) B0E(24-27) B0F(24-27) B0E(24-27) B0F(24-27) + + const __m512i rhs_mat_014589CD_10_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_10, (_MM_PERM_ENUM)136); //B10(0-3) B11(0-3) B10(0-3) B11(0-3) B14(0-3) B15(0-3) B14(0-3) B15(0-3) B18(0-3) B19(0-3) B18(0-3) B19(0-3) B1C(0-3) B1D(0-3) B1C(0-3) B1D(0-3) + const __m512i rhs_mat_2367ABEF_10_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_10, (_MM_PERM_ENUM)136); //B12(0-3) B13(0-3) B12(0-3) B13(0-3) B16(0-3) B17(0-3) B16(0-3) B17(0-3) B1A(0-3) B1B(0-3) B1A(0-3) B1B(0-3) B1E(0-3) B1F(0-3) B1E(0-3) B1F(0-3) + const __m512i rhs_mat_014589CD_11_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_11, (_MM_PERM_ENUM)136); //B10(8-11) B11(8-11) B10(8-11) B11(8-11) B14(8-11) B15(8-11) B14(8-11) B15(8-11) B18(8-11) B19(8-11) B18(8-11) B19(8-11) B1C(8-11) B1D(8-11) B1C(8-11) B1D(8-11) + const __m512i rhs_mat_2367ABEF_11_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_11, (_MM_PERM_ENUM)136); //B12(8-11) B13(8-11) B12(8-11) B13(8-11) B16(8-11) B17(8-11) B16(8-11) B17(8-11) B1A(8-11) B1B(8-11) B1A(8-11) B1B(8-11) B1E(8-11) B1F(8-11) B1E(8-11) B1F(8-11) + const __m512i rhs_mat_014589CD_12_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_12, (_MM_PERM_ENUM)136); //B10(16-19) B11(16-19) B10(16-19) B11(16-19) B14(16-19) B15(16-19) B14(16-19) B15(16-19) B18(16-19) B19(16-19) B18(16-19) B19(16-19) B1C(16-19) B1D(16-19) B1C(16-19) B1D(16-19) + const __m512i rhs_mat_2367ABEF_12_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_12, (_MM_PERM_ENUM)136); //B12(16-19) B13(16-19) B12(16-19) B13(16-19) B16(16-19) B17(16-19) B16(16-19) B17(16-19) B1A(16-19) B1B(16-19) B1A(16-19) B1B(16-19) B1E(16-19) B1F(16-19) B1E(16-19) B1F(16-19) + const __m512i rhs_mat_014589CD_13_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_13, (_MM_PERM_ENUM)136); //B10(24-27) B11(24-27) B10(24-27) B11(24-27) B14(24-27) B15(24-27) B14(24-27) B15(24-27) B18(24-27) B19(24-27) B18(24-27) B19(24-27) B1C(24-27) B1D(24-27) B1C(24-27) B1D(24-27) + const __m512i rhs_mat_2367ABEF_13_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_13, (_MM_PERM_ENUM)136); //B12(24-27) B13(24-27) B12(24-27) B13(24-27) B16(24-27) B17(24-27) B16(24-27) B17(24-27) B1A(24-27) B1B(24-27) B1A(24-27) B1B(24-27) B1E(24-27) B1F(24-27) B1E(24-27) B1F(24-27) + + // Shuffle pattern two - right side input + const __m512i rhs_mat_014589CD_00_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_00, (_MM_PERM_ENUM)221); //B00(4-7) B01(4-7) B00(4-7) B01(4-7) B04(4-7) B05(4-7) B04(4-7) B05(4-7) B08(4-7) B09(4-7) B08(4-7) B09(4-7) B0C(4-7) B0D(4-7) B0C(4-7) B0D(4-7) + const __m512i rhs_mat_2367ABEF_00_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_00, (_MM_PERM_ENUM)221); //B02(4-7) B03(4-7) B02(4-7) B03(4-7) B06(4-7) B07(4-7) B06(4-7) B07(4-7) B0A(4-7) B0B(4-7) B0A(4-7) B0B(4-7) B0E(4-7) B0F(4-7) B0E(4-7) B0F(4-7) + const __m512i rhs_mat_014589CD_01_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_01, (_MM_PERM_ENUM)221); //B00(12-15) B01(12-15) B00(12-15) B01(12-15) B04(12-15) B05(12-15) B04(12-15) B05(12-15) B08(12-15) B09(12-15) B08(12-15) B09(12-15) B0C(12-15) B0D(12-15) B0C(12-15) B0D(12-15) + const __m512i rhs_mat_2367ABEF_01_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_01, (_MM_PERM_ENUM)221); //B02(12-15) B03(12-15) B02(12-15) B03(12-15) B06(12-15) B07(12-15) B06(12-15) B07(12-15) B0A(12-15) B0B(12-15) B0A(12-15) B0B(12-15) B0E(12-15) B0F(12-15) B0E(12-15) B0F(12-15) + const __m512i rhs_mat_014589CD_02_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_02, (_MM_PERM_ENUM)221); //B00(20-23) B01(20-23) B00(20-23) B01(20-23) B04(20-23) B05(20-23) B04(20-23) B05(20-23) B08(20-23) B09(20-23) B08(20-23) B09(20-23) B0C(20-23) B0D(20-23) B0C(20-23) B0D(20-23) + const __m512i rhs_mat_2367ABEF_02_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_02, (_MM_PERM_ENUM)221); //B02(20-23) B03(20-23) B02(20-23) B03(20-23) B06(20-23) B07(20-23) B06(20-23) B07(20-23) B0A(20-23) B0B(20-23) B0A(20-23) B0B(20-23) B0E(20-23) B0F(20-23) B0E(20-23) B0F(20-23) + const __m512i rhs_mat_014589CD_03_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_03, (_MM_PERM_ENUM)221); //B00(28-31) B01(28-31) B00(28-31) B01(28-31) B04(28-31) B05(28-31) B04(28-31) B05(28-31) B08(28-31) B09(28-31) B08(28-31) B09(28-31) B0C(28-31) B0D(28-31) B0C(28-31) 0BD(28-31) + const __m512i rhs_mat_2367ABEF_03_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_03, (_MM_PERM_ENUM)221); //B02(28-31) B03(28-31) B02(28-31) B03(28-31) B06(28-31) B07(28-31) B06(28-31) B07(28-31) B0A(28-31) B0B(28-31) B0A(28-31) B0B(28-31) B0E(28-31) B0F(28-31) B0E(28-31) B0F(28-31) + + const __m512i rhs_mat_014589CD_10_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_10, (_MM_PERM_ENUM)221); //B10(4-7) B11(4-7) B10(4-7) B11(4-7) B14(4-7) B15(4-7) B14(4-7) B15(4-7) B18(4-7) B19(4-7) B18(4-7) B19(4-7) B1C(4-7) B1D(4-7) B1C(4-7) B1D(4-7) + const __m512i rhs_mat_2367ABEF_10_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_10, (_MM_PERM_ENUM)221); //B12(4-7) B13(4-7) B12(4-7) B13(4-7) B16(4-7) B17(4-7) B16(4-7) B17(4-7) B1A(4-7) B1B(4-7) B1A(4-7) B1B(4-7) B1E(4-7) B1F(4-7) B1E(4-7) B1F(4-7) + const __m512i rhs_mat_014589CD_11_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_11, (_MM_PERM_ENUM)221); //B10(12-15) B11(12-15) B10(12-15) B11(12-15) B14(12-15) B15(12-15) B14(12-15) B15(12-15) B18(12-15) B19(12-15) B18(12-15) B19(12-15) B1C(12-15) B1D(12-15) B1C(12-15) B1D(12-15) + const __m512i rhs_mat_2367ABEF_11_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_11, (_MM_PERM_ENUM)221); //B12(12-15) B13(12-15) B12(12-15) B13(12-15) B16(12-15) B17(12-15) B16(12-15) B17(12-15) B1A(12-15) B1B(12-15) B1A(12-15) B1B(12-15) B1E(12-15) B1F(12-15) B1E(12-15) B1F(12-15) + const __m512i rhs_mat_014589CD_12_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_12, (_MM_PERM_ENUM)221); //B10(20-23) B11(20-23) B10(20-23) B11(20-23) B14(20-23) B15(20-23) B14(20-23) B15(20-23) B18(20-23) B19(20-23) B18(20-23) B19(20-23) B1C(20-23) B1D(20-23) B1C(20-23) B1D(20-23) + const __m512i rhs_mat_2367ABEF_12_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_12, (_MM_PERM_ENUM)221); //B12(20-23) B13(20-23) B12(20-23) B13(20-23) B16(20-23) B17(20-23) B16(20-23) B17(20-23) B1A(20-23) B1B(20-23) B1A(20-23) B1B(20-23) B1E(20-23) B1F(20-23) B1E(20-23) B1F(20-23) + const __m512i rhs_mat_014589CD_13_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_13, (_MM_PERM_ENUM)221); //B10(28-31) B11(28-31) B10(28-31) B11(28-31) B14(28-31) B15(28-31) B14(28-31) B15(28-31) B18(28-31) B19(28-31) B18(28-31) B19(28-31) B1C(28-31) B1D(28-31) B1C(28-31) B1D(28-31) + const __m512i rhs_mat_2367ABEF_13_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_13, (_MM_PERM_ENUM)221); //B12(28-31) B13(28-31) B12(28-31) B13(28-31) B16(28-31) B17(28-31) B16(28-31) B17(28-31) B1A(28-31) B1B(28-31) B1A(28-31) B1B(28-31) B1E(28-31) B1F(28-31) B1E(28-31) B1F(28-31) + + uint32_t utmp_00[4], utmp_01[4], utmp_10[4], utmp_11[4]; + + // Scales and Mins of corresponding sub blocks from different Q4_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_00, b_ptr_0[b].scales + 24 * sb, 12); + utmp_00[3] = ((utmp_00[2] >> 4) & kmask2) | (((utmp_00[1] >> 6) & kmask3) << 4); + const uint32_t uaux_00 = utmp_00[1] & kmask1; + utmp_00[1] = (utmp_00[2] & kmask2) | (((utmp_00[0] >> 6) & kmask3) << 4); + utmp_00[2] = uaux_00; + utmp_00[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_01, b_ptr_0[b].scales + 12 + sb * 24, 12); + utmp_01[3] = ((utmp_01[2] >> 4) & kmask2) | (((utmp_01[1] >> 6) & kmask3) << 4); + const uint32_t uaux_01 = utmp_01[1] & kmask1; + utmp_01[1] = (utmp_01[2] & kmask2) | (((utmp_01[0] >> 6) & kmask3) << 4); + utmp_01[2] = uaux_01; + utmp_01[0] &= kmask1; + + memcpy(utmp_10, b_ptr_1[b].scales + sb * 24, 12); + utmp_10[3] = ((utmp_10[2] >> 4) & kmask2) | (((utmp_10[1] >> 6) & kmask3) << 4); + const uint32_t uaux_10 = utmp_10[1] & kmask1; + utmp_10[1] = (utmp_10[2] & kmask2) | (((utmp_10[0] >> 6) & kmask3) << 4); + utmp_10[2] = uaux_10; + utmp_10[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_11, b_ptr_1[b].scales + 12 + sb * 24, 12); + utmp_11[3] = ((utmp_11[2] >> 4) & kmask2) | (((utmp_11[1] >> 6) & kmask3) << 4); + const uint32_t uaux_11 = utmp_11[1] & kmask1; + utmp_11[1] = (utmp_11[2] & kmask2) | (((utmp_11[0] >> 6) & kmask3) << 4); + utmp_11[2] = uaux_11; + utmp_11[0] &= kmask1; + + // Scales of first sub block in the sb loop + const __m256i mins_and_scales_0 = _mm256_set_epi32(utmp_10[3], utmp_10[2], utmp_10[1], utmp_10[0], utmp_00[3], utmp_00[2], utmp_00[1], utmp_00[0]); + const __m512i scales_0 = _mm512_cvtepu8_epi16(_mm256_unpacklo_epi8(mins_and_scales_0, mins_and_scales_0)); + + // Scales of second sub block in the sb loop + const __m256i mins_and_scales_1 = _mm256_set_epi32(utmp_11[3], utmp_11[2], utmp_11[1], utmp_11[0], utmp_01[3], utmp_01[2], utmp_01[1], utmp_01[0]); + const __m512i scales_1 = _mm512_cvtepu8_epi16(_mm256_unpacklo_epi8(mins_and_scales_1, mins_and_scales_1)); + + // Mins of first and second sub block of Q4_K block are arranged side by side + const __m512i mins_01 = _mm512_cvtepu8_epi16(_mm256_unpacklo_epi8(_mm256_shuffle_epi32(mins_and_scales_0, 78), _mm256_shuffle_epi32(mins_and_scales_1, 78))); + + const __m512i scale_014589CD_0 = _mm512_shuffle_epi32(scales_0, (_MM_PERM_ENUM)68); + const __m512i scale_2367ABEF_0 = _mm512_shuffle_epi32(scales_0, (_MM_PERM_ENUM)238); + + const __m512i scale_014589CD_1 = _mm512_shuffle_epi32(scales_1, (_MM_PERM_ENUM)68); + const __m512i scale_2367ABEF_1 = _mm512_shuffle_epi32(scales_1, (_MM_PERM_ENUM)238); + + for (int rp = 0; rp < 4; rp++) { + + // Load the four block_q8_k quantized values interleaved with each other in chunks of eight bytes - 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_00 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 256 * sb))); + __m256i lhs_mat_ymm_01_00 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_00, lhs_mat_ymm_0123_00, 0); + __m256i lhs_mat_ymm_23_00 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_00, lhs_mat_ymm_0123_00, 17); + __m256i lhs_mat_ymm_0123_01 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 32 + 256 * sb))); + __m256i lhs_mat_ymm_01_01 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_01, lhs_mat_ymm_0123_01, 0); + __m256i lhs_mat_ymm_23_01 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_01, lhs_mat_ymm_0123_01, 17); + __m256i lhs_mat_ymm_0123_02 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 64 + 256 * sb))); + __m256i lhs_mat_ymm_01_02 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_02, lhs_mat_ymm_0123_02, 0); + __m256i lhs_mat_ymm_23_02 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_02, lhs_mat_ymm_0123_02, 17); + __m256i lhs_mat_ymm_0123_03 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 96 + 256 * sb))); + __m256i lhs_mat_ymm_01_03 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_03, lhs_mat_ymm_0123_03, 0); + __m256i lhs_mat_ymm_23_03 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_03, lhs_mat_ymm_0123_03, 17); + __m256i lhs_mat_ymm_0123_10 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 128 + 256 * sb))); + __m256i lhs_mat_ymm_01_10 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_10, lhs_mat_ymm_0123_10, 0); + __m256i lhs_mat_ymm_23_10 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_10, lhs_mat_ymm_0123_10, 17); + __m256i lhs_mat_ymm_0123_11 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 160 + 256 * sb))); + __m256i lhs_mat_ymm_01_11 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_11, lhs_mat_ymm_0123_11, 0); + __m256i lhs_mat_ymm_23_11 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_11, lhs_mat_ymm_0123_11, 17); + __m256i lhs_mat_ymm_0123_12 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 192 + 256 * sb))); + __m256i lhs_mat_ymm_01_12 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_12, lhs_mat_ymm_0123_12, 0); + __m256i lhs_mat_ymm_23_12 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_12, lhs_mat_ymm_0123_12, 17); + __m256i lhs_mat_ymm_0123_13 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 224 + 256 * sb))); + __m256i lhs_mat_ymm_01_13 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_13, lhs_mat_ymm_0123_13, 0); + __m256i lhs_mat_ymm_23_13 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_13, lhs_mat_ymm_0123_13, 17); + + __m512i lhs_mat_01_00 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_00), lhs_mat_ymm_01_00, 1); + __m512i lhs_mat_23_00 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_00), lhs_mat_ymm_23_00, 1); + __m512i lhs_mat_01_01 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_01), lhs_mat_ymm_01_01, 1); + __m512i lhs_mat_23_01 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_01), lhs_mat_ymm_23_01, 1); + __m512i lhs_mat_01_02 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_02), lhs_mat_ymm_01_02, 1); + __m512i lhs_mat_23_02 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_02), lhs_mat_ymm_23_02, 1); + __m512i lhs_mat_01_03 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_03), lhs_mat_ymm_01_03, 1); + __m512i lhs_mat_23_03 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_03), lhs_mat_ymm_23_03, 1); + + __m512i lhs_mat_01_10 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_10), lhs_mat_ymm_01_10, 1); + __m512i lhs_mat_23_10 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_10), lhs_mat_ymm_23_10, 1); + __m512i lhs_mat_01_11 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_11), lhs_mat_ymm_01_11, 1); + __m512i lhs_mat_23_11 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_11), lhs_mat_ymm_23_11, 1); + __m512i lhs_mat_01_12 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_12), lhs_mat_ymm_01_12, 1); + __m512i lhs_mat_23_12 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_12), lhs_mat_ymm_23_12, 1); + __m512i lhs_mat_01_13 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_13), lhs_mat_ymm_01_13, 1); + __m512i lhs_mat_23_13 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_13), lhs_mat_ymm_23_13, 1); + + // Bsums are loaded - four bsums are loaded (for two sub blocks) for the different Q8_K blocks + __m256i lhs_bsums_0123_01 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].bsums + 16 * sb))); + __m256i lhs_bsums_hsum_ymm_0123_01 = _mm256_castsi128_si256(_mm_hadd_epi16(_mm256_castsi256_si128(lhs_bsums_0123_01), _mm256_extractf128_si256(lhs_bsums_0123_01, 1))); + lhs_bsums_hsum_ymm_0123_01 = _mm256_permute2x128_si256(lhs_bsums_hsum_ymm_0123_01, lhs_bsums_hsum_ymm_0123_01, 0); + __m512i lhs_bsums_hsum_0123_01 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_bsums_hsum_ymm_0123_01), lhs_bsums_hsum_ymm_0123_01, 1); + + // Shuffle pattern one - left side input + const __m512i lhs_mat_01_00_sp1 = _mm512_shuffle_epi32(lhs_mat_01_00, (_MM_PERM_ENUM)160); //A00(0-3) A00(0-3) A01(0-3) A01(0-3) A00(0-3) A00(0-3) A01(0-3) A01(0-3) A00(0-3) A00(0-3) A01(0-3) A01(0-3) A00(0-3) A00(0-3) A01(0-3) A01(0-3) + const __m512i lhs_mat_23_00_sp1 = _mm512_shuffle_epi32(lhs_mat_23_00, (_MM_PERM_ENUM)160); //A02(0-3) A02(0-3) A03(0-3) A03(0-3) A02(0-3) A02(0-3) A03(0-3) A03(0-3) A02(0-3) A02(0-3) A03(0-3) A03(0-3) A02(0-3) A02(0-3) A03(0-3) A03(0-3) + const __m512i lhs_mat_01_01_sp1 = _mm512_shuffle_epi32(lhs_mat_01_01, (_MM_PERM_ENUM)160); //A00(8-11) A00(8-11) A01(8-11) A01(8-11) A00(8-11) A00(8-11) A01(8-11) A01(8-11) A00(8-11) A00(8-11) A01(8-11) A01(8-11) A00(8-11) A00(8-11) A01(8-11) A01(8-11) + const __m512i lhs_mat_23_01_sp1 = _mm512_shuffle_epi32(lhs_mat_23_01, (_MM_PERM_ENUM)160); //A02(8-11) A02(8-11) A03(8-11) A03(8-11) A02(8-11) A02(8-11) A03(8-11) A03(8-11) A02(8-11) A02(8-11) A03(8-11) A03(8-11) A02(8-11) A02(8-11) A03(8-11) A03(8-11) + const __m512i lhs_mat_01_02_sp1 = _mm512_shuffle_epi32(lhs_mat_01_02, (_MM_PERM_ENUM)160); //A00(16-19) A00(16-19) A01(16-19) A01(16-19) A00(16-19) A00(16-19) A01(16-19) A01(16-19) A00(16-19) A00(16-19) A01(16-19) A01(16-19) A00(16-19) A00(16-19) A01(16-19) A01(16-19) + const __m512i lhs_mat_23_02_sp1 = _mm512_shuffle_epi32(lhs_mat_23_02, (_MM_PERM_ENUM)160); //A02(16-19) A02(16-19) A03(16-19) A03(16-19) A02(16-19) A02(16-19) A03(16-19) A03(16-19) A02(16-19) A02(16-19) A03(16-19) A03(16-19) A02(16-19) A02(16-19) A03(16-19) A03(16-19) + const __m512i lhs_mat_01_03_sp1 = _mm512_shuffle_epi32(lhs_mat_01_03, (_MM_PERM_ENUM)160); //A00(24-27) A00(24-27) A01(24-27) A01(24-27) A00(24-27) A00(24-27) A01(24-27) A01(24-27) A00(24-27) A00(24-27) A01(24-27) A01(24-27) A00(24-27) A00(24-27) A01(24-27) A01(24-27) + const __m512i lhs_mat_23_03_sp1 = _mm512_shuffle_epi32(lhs_mat_23_03, (_MM_PERM_ENUM)160); //A02(24-27) A02(24-27) A03(24-27) A03(24-27) A02(24-27) A02(24-27) A03(24-27) A03(24-27) A02(24-27) A02(24-27) A03(24-27) A03(24-27) A02(24-27) A02(24-27) A03(24-27) A03(24-27) + + const __m512i lhs_mat_01_10_sp1 = _mm512_shuffle_epi32(lhs_mat_01_10, (_MM_PERM_ENUM)160); //A10(0-3) A10(0-3) A11(0-3) A11(0-3) A10(0-3) A10(0-3) A11(0-3) A11(0-3) A10(0-3) A10(0-3) A11(0-3) A11(0-3) A10(0-3) A10(0-3) A11(0-3) A11(0-3) + const __m512i lhs_mat_23_10_sp1 = _mm512_shuffle_epi32(lhs_mat_23_10, (_MM_PERM_ENUM)160); //A12(0-3) A12(0-3) A13(0-3) A13(0-3) A12(0-3) A12(0-3) A13(0-3) A13(0-3) A12(0-3) A12(0-3) A13(0-3) A13(0-3) A12(0-3) A12(0-3) A13(0-3) A13(0-3) + const __m512i lhs_mat_01_11_sp1 = _mm512_shuffle_epi32(lhs_mat_01_11, (_MM_PERM_ENUM)160); //A10(8-11) A10(8-11) A11(8-11) A11(8-11) A10(8-11) A10(8-11) A11(8-11) A11(8-11) A10(8-11) A10(8-11) A11(8-11) A11(8-11) A10(8-11) A10(8-11) A11(8-11) A11(8-11) + const __m512i lhs_mat_23_11_sp1 = _mm512_shuffle_epi32(lhs_mat_23_11, (_MM_PERM_ENUM)160); //A12(8-11) A12(8-11) A13(8-11) A13(8-11) A12(8-11) A12(8-11) A13(8-11) A13(8-11) A12(8-11) A12(8-11) A13(8-11) A13(8-11) A12(8-11) A12(8-11) A13(8-11) A13(8-11) + const __m512i lhs_mat_01_12_sp1 = _mm512_shuffle_epi32(lhs_mat_01_12, (_MM_PERM_ENUM)160); //A10(16-19) A10(16-19) A11(16-19) A11(16-19) A10(16-19) A10(16-19) A11(16-19) A11(16-19) A10(16-19) A10(16-19) A11(16-19) A11(16-19) A10(16-19) A10(16-19) A11(16-19) A11(16-19) + const __m512i lhs_mat_23_12_sp1 = _mm512_shuffle_epi32(lhs_mat_23_12, (_MM_PERM_ENUM)160); //A12(16-19) A12(16-19) A13(16-19) A13(16-19) A12(16-19) A12(16-19) A13(16-19) A13(16-19) A12(16-19) A12(16-19) A13(16-19) A13(16-19) A12(16-19) A12(16-19) A13(16-19) A13(16-19) + const __m512i lhs_mat_01_13_sp1 = _mm512_shuffle_epi32(lhs_mat_01_13, (_MM_PERM_ENUM)160); //A10(24-27) A10(24-27) A11(24-27) A11(24-27) A10(24-27) A10(24-27) A11(24-27) A11(24-27) A10(24-27) A10(24-27) A11(24-27) A11(24-27) A10(24-27) A10(24-27) A11(24-27) A11(24-27) + const __m512i lhs_mat_23_13_sp1 = _mm512_shuffle_epi32(lhs_mat_23_13, (_MM_PERM_ENUM)160); //A12(24-27) A12(24-27) A13(24-27) A13(24-27) A12(24-27) A12(24-27) A13(24-27) A13(24-27) A12(24-27) A12(24-27) A13(24-27) A13(24-27) A12(24-27) A12(24-27) A13(24-27) A13(24-27) + + const __m512i lhs_mat_01_00_sp2 = _mm512_shuffle_epi32(lhs_mat_01_00, (_MM_PERM_ENUM)245); //A00(4-7) A00(4-7) A01(4-7) A01(4-7) A00(4-7) A00(4-7) A01(4-7) A01(4-7) A00(4-7) A00(4-7) A01(4-7) A01(4-7) A00(4-7) A00(4-7) A01(4-7) A01(4-7) + const __m512i lhs_mat_23_00_sp2 = _mm512_shuffle_epi32(lhs_mat_23_00, (_MM_PERM_ENUM)245); //A02(4-7) A02(4-7) A03(4-7) A03(4-7) A02(4-7) A02(4-7) A03(4-7) A03(4-7) A02(4-7) A02(4-7) A03(4-7) A03(4-7) A02(4-7) A02(4-7) A03(4-7) A03(4-7) + const __m512i lhs_mat_01_01_sp2 = _mm512_shuffle_epi32(lhs_mat_01_01, (_MM_PERM_ENUM)245); //A00(12-15) A00(12-15) A01(12-15) A01(12-15) A00(12-15) A00(12-15) A01(12-15) A01(12-15) A00(12-15) A00(12-15) A01(12-15) A01(12-15) A00(12-15) A00(12-15) A01(12-15) A01(12-15) + const __m512i lhs_mat_23_01_sp2 = _mm512_shuffle_epi32(lhs_mat_23_01, (_MM_PERM_ENUM)245); //A02(12-15) A02(12-15) A03(12-15) A03(12-15) A02(12-15) A02(12-15) A03(12-15) A03(12-15) A02(12-15) A02(12-15) A03(12-15) A03(12-15) A02(12-15) A02(12-15) A03(12-15) A03(12-15) + const __m512i lhs_mat_01_02_sp2 = _mm512_shuffle_epi32(lhs_mat_01_02, (_MM_PERM_ENUM)245); //A00(20-23) A00(20-23) A01(20-23) A01(20-23) A00(20-23) A00(20-23) A01(20-23) A01(20-23) A00(20-23) A00(20-23) A01(20-23) A01(20-23) A00(20-23) A00(20-23) A01(20-23) A01(20-23) + const __m512i lhs_mat_23_02_sp2 = _mm512_shuffle_epi32(lhs_mat_23_02, (_MM_PERM_ENUM)245); //A02(20-23) A02(20-23) A03(20-23) A03(20-23) A02(20-23) A02(20-23) A03(20-23) A03(20-23) A02(20-23) A02(20-23) A03(20-23) A03(20-23) A02(20-23) A02(20-23) A03(20-23) A03(20-23) + const __m512i lhs_mat_01_03_sp2 = _mm512_shuffle_epi32(lhs_mat_01_03, (_MM_PERM_ENUM)245); //A00(28-31) A00(28-31) A01(28-31) A01(28-31) A00(28-31) A00(28-31) A01(28-31) A01(28-31) A00(28-31) A00(28-31) A01(28-31) A01(28-31) A00(28-31) A00(28-31) A01(28-31) A01(28-31) + const __m512i lhs_mat_23_03_sp2 = _mm512_shuffle_epi32(lhs_mat_23_03, (_MM_PERM_ENUM)245); //A02(28-31) A02(28-31) A03(28-31) A03(28-31) A02(28-31) A02(28-31) A03(28-31) A03(28-31) A02(28-31) A02(28-31) A03(28-31) A03(28-31) A02(28-31) A02(28-31) A03(28-31) A03(28-31) + + const __m512i lhs_mat_01_10_sp2 = _mm512_shuffle_epi32(lhs_mat_01_10, (_MM_PERM_ENUM)245); //A10(4-7) A10(4-7) A11(4-7) A11(4-7) A10(4-7) A10(4-7) A11(4-7) A11(4-7) A10(4-7) A10(4-7) A11(4-7) A11(4-7) A10(4-7) A10(4-7) A11(4-7) A11(4-7) + const __m512i lhs_mat_23_10_sp2 = _mm512_shuffle_epi32(lhs_mat_23_10, (_MM_PERM_ENUM)245); //A12(4-7) A12(4-7) A13(4-7) A13(4-7) A12(4-7) A12(4-7) A13(4-7) A13(4-7) A12(4-7) A12(4-7) A13(4-7) A13(4-7) A12(4-7) A12(4-7) A13(4-7) A13(4-7) + const __m512i lhs_mat_01_11_sp2 = _mm512_shuffle_epi32(lhs_mat_01_11, (_MM_PERM_ENUM)245); //A10(12-15) A10(12-15) A11(12-15) A11(12-15) A10(12-15) A10(12-15) A11(12-15) A11(12-15) A10(12-15) A10(12-15) A11(12-15) A11(12-15) A10(12-15) A10(12-15) A11(12-15) A11(12-15) + const __m512i lhs_mat_23_11_sp2 = _mm512_shuffle_epi32(lhs_mat_23_11, (_MM_PERM_ENUM)245); //A12(12-15) A12(12-15) A13(12-15) A13(12-15) A12(12-15) A12(12-15) A13(12-15) A13(12-15) A12(12-15) A12(12-15) A13(12-15) A13(12-15) A12(12-15) A12(12-15) A13(12-15) A13(12-15) + const __m512i lhs_mat_01_12_sp2 = _mm512_shuffle_epi32(lhs_mat_01_12, (_MM_PERM_ENUM)245); //A10(20-23) A10(20-23) A11(20-23) A11(20-23) A10(20-23) A10(20-23) A11(20-23) A11(20-23) A10(20-23) A10(20-23) A11(20-23) A11(20-23) A10(20-23) A10(20-23) A11(20-23) A11(20-23) + const __m512i lhs_mat_23_12_sp2 = _mm512_shuffle_epi32(lhs_mat_23_12, (_MM_PERM_ENUM)245); //A12(20-23) A12(20-23) A13(20-23) A13(20-23) A12(20-23) A12(20-23) A13(20-23) A13(20-23) A12(20-23) A12(20-23) A13(20-23) A13(20-23) A12(20-23) A12(20-23) A13(20-23) A13(20-23) + const __m512i lhs_mat_01_13_sp2 = _mm512_shuffle_epi32(lhs_mat_01_13, (_MM_PERM_ENUM)245); //A10(28-31) A10(28-31) A11(28-31) A11(28-31) A10(28-31) A10(28-31) A11(28-31) A11(28-31) A10(28-31) A10(28-31) A11(28-31) A11(28-31) A10(28-31) A10(28-31) A11(28-31) A11(28-31) + const __m512i lhs_mat_23_13_sp2 = _mm512_shuffle_epi32(lhs_mat_23_13, (_MM_PERM_ENUM)245); //A12(28-31) A12(28-31) A13(28-31) A13(28-31) A12(28-31) A12(28-31) A13(28-31) A13(28-31) A12(28-31) A12(28-31) A13(28-31) A13(28-31) A12(28-31) A12(28-31) A13(28-31) A13(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 + __m512i iacc_mat_00_0_sp1 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_03_sp1, lhs_mat_01_03_sp1), _mm512_maddubs_epi16(rhs_mat_014589CD_02_sp1, lhs_mat_01_02_sp1)), _mm512_maddubs_epi16(rhs_mat_014589CD_01_sp1, lhs_mat_01_01_sp1)), _mm512_maddubs_epi16(rhs_mat_014589CD_00_sp1, lhs_mat_01_00_sp1)); + __m512i iacc_mat_01_0_sp1 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_03_sp1, lhs_mat_01_03_sp1), _mm512_maddubs_epi16(rhs_mat_2367ABEF_02_sp1, lhs_mat_01_02_sp1)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_01_sp1, lhs_mat_01_01_sp1)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_00_sp1, lhs_mat_01_00_sp1)); + __m512i iacc_mat_10_0_sp1 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_03_sp1, lhs_mat_23_03_sp1), _mm512_maddubs_epi16(rhs_mat_014589CD_02_sp1, lhs_mat_23_02_sp1)), _mm512_maddubs_epi16(rhs_mat_014589CD_01_sp1, lhs_mat_23_01_sp1)), _mm512_maddubs_epi16(rhs_mat_014589CD_00_sp1, lhs_mat_23_00_sp1)); + __m512i iacc_mat_11_0_sp1 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_03_sp1, lhs_mat_23_03_sp1), _mm512_maddubs_epi16(rhs_mat_2367ABEF_02_sp1, lhs_mat_23_02_sp1)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_01_sp1, lhs_mat_23_01_sp1)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_00_sp1, lhs_mat_23_00_sp1)); + __m512i iacc_mat_00_1_sp1 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_13_sp1, lhs_mat_01_13_sp1), _mm512_maddubs_epi16(rhs_mat_014589CD_12_sp1, lhs_mat_01_12_sp1)), _mm512_maddubs_epi16(rhs_mat_014589CD_11_sp1, lhs_mat_01_11_sp1)), _mm512_maddubs_epi16(rhs_mat_014589CD_10_sp1, lhs_mat_01_10_sp1)); + __m512i iacc_mat_01_1_sp1 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_13_sp1, lhs_mat_01_13_sp1), _mm512_maddubs_epi16(rhs_mat_2367ABEF_12_sp1, lhs_mat_01_12_sp1)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_11_sp1, lhs_mat_01_11_sp1)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_10_sp1, lhs_mat_01_10_sp1)); + __m512i iacc_mat_10_1_sp1 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_13_sp1, lhs_mat_23_13_sp1), _mm512_maddubs_epi16(rhs_mat_014589CD_12_sp1, lhs_mat_23_12_sp1)), _mm512_maddubs_epi16(rhs_mat_014589CD_11_sp1, lhs_mat_23_11_sp1)), _mm512_maddubs_epi16(rhs_mat_014589CD_10_sp1, lhs_mat_23_10_sp1)); + __m512i iacc_mat_11_1_sp1 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_13_sp1, lhs_mat_23_13_sp1), _mm512_maddubs_epi16(rhs_mat_2367ABEF_12_sp1, lhs_mat_23_12_sp1)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_11_sp1, lhs_mat_23_11_sp1)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_10_sp1, lhs_mat_23_10_sp1)); + + __m512i iacc_mat_00_0_sp2 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_03_sp2, lhs_mat_01_03_sp2), _mm512_maddubs_epi16(rhs_mat_014589CD_02_sp2, lhs_mat_01_02_sp2)), _mm512_maddubs_epi16(rhs_mat_014589CD_01_sp2, lhs_mat_01_01_sp2)), _mm512_maddubs_epi16(rhs_mat_014589CD_00_sp2, lhs_mat_01_00_sp2)); + __m512i iacc_mat_01_0_sp2 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_03_sp2, lhs_mat_01_03_sp2), _mm512_maddubs_epi16(rhs_mat_2367ABEF_02_sp2, lhs_mat_01_02_sp2)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_01_sp2, lhs_mat_01_01_sp2)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_00_sp2, lhs_mat_01_00_sp2)); + __m512i iacc_mat_10_0_sp2 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_03_sp2, lhs_mat_23_03_sp2), _mm512_maddubs_epi16(rhs_mat_014589CD_02_sp2, lhs_mat_23_02_sp2)), _mm512_maddubs_epi16(rhs_mat_014589CD_01_sp2, lhs_mat_23_01_sp2)), _mm512_maddubs_epi16(rhs_mat_014589CD_00_sp2, lhs_mat_23_00_sp2)); + __m512i iacc_mat_11_0_sp2 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_03_sp2, lhs_mat_23_03_sp2), _mm512_maddubs_epi16(rhs_mat_2367ABEF_02_sp2, lhs_mat_23_02_sp2)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_01_sp2, lhs_mat_23_01_sp2)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_00_sp2, lhs_mat_23_00_sp2)); + __m512i iacc_mat_00_1_sp2 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_13_sp2, lhs_mat_01_13_sp2), _mm512_maddubs_epi16(rhs_mat_014589CD_12_sp2, lhs_mat_01_12_sp2)), _mm512_maddubs_epi16(rhs_mat_014589CD_11_sp2, lhs_mat_01_11_sp2)), _mm512_maddubs_epi16(rhs_mat_014589CD_10_sp2, lhs_mat_01_10_sp2)); + __m512i iacc_mat_01_1_sp2 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_13_sp2, lhs_mat_01_13_sp2), _mm512_maddubs_epi16(rhs_mat_2367ABEF_12_sp2, lhs_mat_01_12_sp2)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_11_sp2, lhs_mat_01_11_sp2)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_10_sp2, lhs_mat_01_10_sp2)); + __m512i iacc_mat_10_1_sp2 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_13_sp2, lhs_mat_23_13_sp2), _mm512_maddubs_epi16(rhs_mat_014589CD_12_sp2, lhs_mat_23_12_sp2)), _mm512_maddubs_epi16(rhs_mat_014589CD_11_sp2, lhs_mat_23_11_sp2)), _mm512_maddubs_epi16(rhs_mat_014589CD_10_sp2, lhs_mat_23_10_sp2)); + __m512i iacc_mat_11_1_sp2 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_13_sp2, lhs_mat_23_13_sp2), _mm512_maddubs_epi16(rhs_mat_2367ABEF_12_sp2, lhs_mat_23_12_sp2)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_11_sp2, lhs_mat_23_11_sp2)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_10_sp2, lhs_mat_23_10_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_0 = _mm512_add_epi16(iacc_mat_00_0_sp1, iacc_mat_00_0_sp2); + __m512i iacc_mat_01_0 = _mm512_add_epi16(iacc_mat_01_0_sp1, iacc_mat_01_0_sp2); + __m512i iacc_mat_10_0 = _mm512_add_epi16(iacc_mat_10_0_sp1, iacc_mat_10_0_sp2); + __m512i iacc_mat_11_0 = _mm512_add_epi16(iacc_mat_11_0_sp1, iacc_mat_11_0_sp2); + + __m512i iacc_mat_00_1 = _mm512_add_epi16(iacc_mat_00_1_sp1, iacc_mat_00_1_sp2); + __m512i iacc_mat_01_1 = _mm512_add_epi16(iacc_mat_01_1_sp1, iacc_mat_01_1_sp2); + __m512i iacc_mat_10_1 = _mm512_add_epi16(iacc_mat_10_1_sp1, iacc_mat_10_1_sp2); + __m512i iacc_mat_11_1 = _mm512_add_epi16(iacc_mat_11_1_sp1, iacc_mat_11_1_sp2); + + iacc_mat_00_0 = _mm512_madd_epi16(iacc_mat_00_0, scale_014589CD_0); + iacc_mat_01_0 = _mm512_madd_epi16(iacc_mat_01_0, scale_2367ABEF_0); + iacc_mat_10_0 = _mm512_madd_epi16(iacc_mat_10_0, scale_014589CD_0); + iacc_mat_11_0 = _mm512_madd_epi16(iacc_mat_11_0, scale_2367ABEF_0); + + iacc_mat_00_1 = _mm512_madd_epi16(iacc_mat_00_1, scale_014589CD_1); + iacc_mat_01_1 = _mm512_madd_epi16(iacc_mat_01_1, scale_2367ABEF_1); + iacc_mat_10_1 = _mm512_madd_epi16(iacc_mat_10_1, scale_014589CD_1); + iacc_mat_11_1 = _mm512_madd_epi16(iacc_mat_11_1, scale_2367ABEF_1); + + // Straighten out to make 4 row vectors (4 for each sub block which are accumulated together in the next step) + __m512i iacc_row_0_0 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_00_0, _mm512_shuffle_epi32(iacc_mat_01_0, (_MM_PERM_ENUM)78)); + __m512i iacc_row_1_0 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_00_0, (_MM_PERM_ENUM)78), iacc_mat_01_0); + __m512i iacc_row_2_0 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_10_0, _mm512_shuffle_epi32(iacc_mat_11_0, (_MM_PERM_ENUM)78)); + __m512i iacc_row_3_0 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_10_0, (_MM_PERM_ENUM)78), iacc_mat_11_0); + __m512i iacc_row_0_1 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_00_1, _mm512_shuffle_epi32(iacc_mat_01_1, (_MM_PERM_ENUM)78)); + __m512i iacc_row_1_1 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_00_1, (_MM_PERM_ENUM)78), iacc_mat_01_1); + __m512i iacc_row_2_1 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_10_1, _mm512_shuffle_epi32(iacc_mat_11_1, (_MM_PERM_ENUM)78)); + __m512i iacc_row_3_1 = _mm512_mask_blend_epi32(0xCCCC,_mm512_shuffle_epi32(iacc_mat_10_1, (_MM_PERM_ENUM)78), iacc_mat_11_1); + + __m512i iacc_row_0 = _mm512_add_epi32(iacc_row_0_0, iacc_row_0_1); + __m512i iacc_row_1 = _mm512_add_epi32(iacc_row_1_0, iacc_row_1_1); + __m512i iacc_row_2 = _mm512_add_epi32(iacc_row_2_0, iacc_row_2_1); + __m512i iacc_row_3 = _mm512_add_epi32(iacc_row_3_0, iacc_row_3_1); + + // Load the scale(d) values for all the 4 Q8_k blocks and repeat it across lanes + const __m128 row_scale_f32_sse = _mm_load_ps(a_ptrs[rp][b].d); + const __m256 row_scale_f32_ymm = _mm256_set_m128(row_scale_f32_sse, row_scale_f32_sse); + const __m512 row_scale_f32 = _mm512_insertf32x8(_mm512_castps256_ps512(row_scale_f32_ymm), row_scale_f32_ymm, 1); + + // Multiply with appropiate scales and accumulate (for both d and dmin) below + 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]); + + __m512i iacc_row_min_0 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_hsum_0123_01, (_MM_PERM_ENUM)0), mins_01); + __m512i iacc_row_min_1 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_hsum_0123_01, (_MM_PERM_ENUM)85), mins_01); + __m512i iacc_row_min_2 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_hsum_0123_01, (_MM_PERM_ENUM)170), mins_01); + __m512i iacc_row_min_3 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_hsum_0123_01, (_MM_PERM_ENUM)255), mins_01); + + acc_min_rows[rp * 4] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_min_0), _mm512_mul_ps(col_dmin_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 0)), acc_min_rows[rp * 4]); + acc_min_rows[rp * 4 + 1] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_min_1), _mm512_mul_ps(col_dmin_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 85)), acc_min_rows[rp * 4 + 1]); + acc_min_rows[rp * 4 + 2] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_min_2), _mm512_mul_ps(col_dmin_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_min_rows[rp * 4 + 2]); + acc_min_rows[rp * 4 + 3] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_min_3), _mm512_mul_ps(col_dmin_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 255)), acc_min_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)), _mm512_sub_ps(acc_rows[i], acc_min_rows[i])); + } + } + } + + for (; y < nr / 4; y++) { + + const block_q8_Kx4 * a_ptr = a_ptr_start + (y * nb); + + // Take group of eight block_q4_kx8 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_Kx8 * b_ptr_0 = b_ptr_start + ((x) * b_nb); + const block_q4_Kx8 * 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(); + } + + __m512 acc_min_rows[4]; + for (int i = 0; i < 4; i++) { + acc_min_rows[i] = _mm512_setzero_ps(); + } + + // For super block + for (int64_t b = 0; b < nb; b++) { + // Scale values - Load the sixteen scale values from two block_q4_kx8 structures + const __m512 col_scale_f32 = GGML_F32Cx8x2_LOAD(b_ptr_0[b].d, b_ptr_1[b].d); + + // dmin values - Load the sixteen dmin values from two block_q4_kx8 structures + const __m512 col_dmin_f32 = GGML_F32Cx8x2_LOAD(b_ptr_0[b].dmin, b_ptr_1[b].dmin); + + // Loop to iterate over the eight sub blocks of a super block - two sub blocks are processed per iteration + for (int sb = 0; sb < QK_K / 64; sb++) { + + const __m256i rhs_raw_mat_0123_0 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + sb * 256)); + const __m256i rhs_raw_mat_4567_0 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 32 + sb * 256)); + const __m256i rhs_raw_mat_0123_1 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 64 + sb * 256)); + const __m256i rhs_raw_mat_4567_1 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 96 + sb * 256)); + const __m256i rhs_raw_mat_0123_2 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 128 + sb * 256)); + const __m256i rhs_raw_mat_4567_2 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 160 + sb * 256)); + const __m256i rhs_raw_mat_0123_3 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 192 + sb * 256)); + const __m256i rhs_raw_mat_4567_3 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 224 + sb * 256)); + + const __m256i rhs_raw_mat_89AB_0 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + sb * 256)); + const __m256i rhs_raw_mat_CDEF_0 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 32 + sb * 256)); + const __m256i rhs_raw_mat_89AB_1 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 64 + sb * 256)); + const __m256i rhs_raw_mat_CDEF_1 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 96 + sb * 256)); + const __m256i rhs_raw_mat_89AB_2 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 128 + sb * 256)); + const __m256i rhs_raw_mat_CDEF_2 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 160 + sb * 256)); + const __m256i rhs_raw_mat_89AB_3 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 192 + sb * 256)); + const __m256i rhs_raw_mat_CDEF_3 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 224 + sb * 256)); + + 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_0145_2 = _mm256_blend_epi32(rhs_raw_mat_0123_2, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_2, requiredOrder), 240); + const __m256i rhs_raw_mat_2367_2 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_2, requiredOrder), rhs_raw_mat_4567_2, 240); + const __m256i rhs_raw_mat_0145_3 = _mm256_blend_epi32(rhs_raw_mat_0123_3, _mm256_permutevar8x32_epi32(rhs_raw_mat_4567_3, requiredOrder), 240); + const __m256i rhs_raw_mat_2367_3 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_0123_3, requiredOrder), rhs_raw_mat_4567_3, 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 __m256i rhs_raw_mat_89CD_2 = _mm256_blend_epi32(rhs_raw_mat_89AB_2, _mm256_permutevar8x32_epi32(rhs_raw_mat_CDEF_2, requiredOrder), 240); + const __m256i rhs_raw_mat_ABEF_2 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_89AB_2, requiredOrder), rhs_raw_mat_CDEF_2, 240); + const __m256i rhs_raw_mat_89CD_3 = _mm256_blend_epi32(rhs_raw_mat_89AB_3, _mm256_permutevar8x32_epi32(rhs_raw_mat_CDEF_3, requiredOrder), 240); + const __m256i rhs_raw_mat_ABEF_3 = _mm256_blend_epi32(_mm256_permutevar8x32_epi32(rhs_raw_mat_89AB_3, requiredOrder), rhs_raw_mat_CDEF_3, 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); + + const __m512i rhs_raw_mat_014589CD_2 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_2), rhs_raw_mat_89CD_2, 1); + const __m512i rhs_raw_mat_2367ABEF_2 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_2), rhs_raw_mat_ABEF_2, 1); + const __m512i rhs_raw_mat_014589CD_3 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_3), rhs_raw_mat_89CD_3, 1); + const __m512i rhs_raw_mat_2367ABEF_3 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_3), rhs_raw_mat_ABEF_3, 1); + + //4-bit -> 8-bit + const __m512i rhs_mat_014589CD_00 = _mm512_and_si512(rhs_raw_mat_014589CD_0, m4bexpanded); //B00(0-7) B01(0-7) B04(0-7) B05(0-7) B08(0-7) B09(0-7) B0C(0-7) B0D(0-7) + const __m512i rhs_mat_2367ABEF_00 = _mm512_and_si512(rhs_raw_mat_2367ABEF_0, m4bexpanded); //B02(0-7) B03(0-7) B06(0-7) B07(0-7) B0A(0-7) B0B(0-7) B0E(0-7) B0F(0-7) + const __m512i rhs_mat_014589CD_01 = _mm512_and_si512(rhs_raw_mat_014589CD_1, m4bexpanded); //B00(8-15) B01(8-15) B04(8-15) B05(8-15) B08(8-15) B09(8-15) B0C(8-15) B0D(8-15) + const __m512i rhs_mat_2367ABEF_01 = _mm512_and_si512(rhs_raw_mat_2367ABEF_1, m4bexpanded); //B02(8-15) B03(8-15) B06(8-15) B07(8-15) B0A(8-15) B0B(8-15) B0E(8-15) B0F(8-15) + + const __m512i rhs_mat_014589CD_02 = _mm512_and_si512(rhs_raw_mat_014589CD_2, m4bexpanded); //B00(16-23) B01(16-23) B04(16-23) B05(16-23) B08(16-23) B09(16-23) B0C(16-23) B0D(16-23) + const __m512i rhs_mat_2367ABEF_02 = _mm512_and_si512(rhs_raw_mat_2367ABEF_2, m4bexpanded); //B02(16-23) B03(16-23) B06(16-23) B07(16-23) B0A(16-23) B0B(16-23) B0E(16-23) B0F(16-23) + const __m512i rhs_mat_014589CD_03 = _mm512_and_si512(rhs_raw_mat_014589CD_3, m4bexpanded); //B00(24-31) B01(24-31) B04(24-31) B05(24-31) B08(24-31) B09(24-31) B0C(24-31) B0D(24-31) + const __m512i rhs_mat_2367ABEF_03 = _mm512_and_si512(rhs_raw_mat_2367ABEF_3, m4bexpanded); //B02(24-31) B03(24-31) B06(24-31) B07(24-31) B0A(24-31) B0B(24-31) B0E(24-31) B0F(24-31) + + const __m512i rhs_mat_014589CD_10 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_0, 4), m4bexpanded); //B10(0-7) B11(0-7) B14(0-7) B15(0-7) B18(0-7) B19(0-7) B1C(0-7) B1D(0-7) + const __m512i rhs_mat_2367ABEF_10 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_0, 4), m4bexpanded); //B12(0-7) B13(0-7) B16(0-7) B17(0-7) B1A(0-7) B1B(0-7) B1E(0-7) B1F(0-7) + const __m512i rhs_mat_014589CD_11 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_1, 4), m4bexpanded); //B10(8-15) B11(8-15) B14(8-15) B15(8-15) B18(8-15) B19(8-15) B1C(8-15) B1D(8-15) + const __m512i rhs_mat_2367ABEF_11 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_1, 4), m4bexpanded); //B12(8-15) B13(8-15) B16(8-15) B17(8-15) B1A(8-15) B1B(8-15) B1E(8-15) B1F(8-15) + + const __m512i rhs_mat_014589CD_12 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_2, 4), m4bexpanded); //B10(16-23) B11(16-23) B14(16-23) B15(16-23) B18(16-23) B19(16-23) B1C(16-23) B1D(16-23) + const __m512i rhs_mat_2367ABEF_12 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_2, 4), m4bexpanded); //B12(16-23) B13(16-23) B16(16-23) B17(16-23) B1A(16-23) B1B(16-23) B1E(16-23) B1F(16-23) + const __m512i rhs_mat_014589CD_13 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_3, 4), m4bexpanded); //B10(24-31) B11(24-31) B14(24-31) B15(24-31) B18(24-31) B19(24-31) B1C(24-31) B1D(24-31) + const __m512i rhs_mat_2367ABEF_13 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_3, 4), m4bexpanded); //B12(24-31) B13(24-31) B16(24-31) B17(24-31) B1A(24-31) B1B(24-31) B1E(24-31) B1F(24-31) + + // Shuffle pattern one - right side input + const __m512i rhs_mat_014589CD_00_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_00, (_MM_PERM_ENUM)136); //B00(0-3) B01(0-3) B00(0-3) B01(0-3) B04(0-3) B05(0-3) B04(0-3) B05(0-3) B08(0-3) B09(0-3) B08(0-3) B09(0-3) B0C(0-3) B0D(0-3) B0C(0-3) B0D(0-3) + const __m512i rhs_mat_2367ABEF_00_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_00, (_MM_PERM_ENUM)136); //B02(0-3) B03(0-3) B02(0-3) B03(0-3) B06(0-3) B07(0-3) B06(0-3) B07(0-3) B0A(0-3) B0B(0-3) B0A(0-3) B0B(0-3) B0E(0-3) B0F(0-3) B0E(0-3) B0F(0-3) + const __m512i rhs_mat_014589CD_01_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_01, (_MM_PERM_ENUM)136); //B00(8-11) B01(8-11) B00(8-11) B01(8-11) B04(8-11) B05(8-11) B04(8-11) B05(8-11) B08(8-11) B09(8-11) B08(8-11) B09(8-11) B0C(8-11) B0D(8-11) B0C(8-11) B0D(8-11) + const __m512i rhs_mat_2367ABEF_01_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_01, (_MM_PERM_ENUM)136); //B02(8-11) B03(8-11) B02(8-11) B03(8-11) B06(8-11) B07(8-11) B06(8-11) B07(8-11) B0A(8-11) B0B(8-11) B0A(8-11) B0B(8-11) B0E(8-11) B0F(8-11) B0E(8-11) B0F(8-11) + const __m512i rhs_mat_014589CD_02_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_02, (_MM_PERM_ENUM)136); //B00(16-19) B01(16-19) B00(16-19) B01(16-19) B04(16-19) B05(16-19) B04(16-19) B05(16-19) B08(16-19) B09(16-19) B08(16-19) B09(16-19) B0C(16-19) B0D(16-19) B0C(16-19) B0D(16-19) + const __m512i rhs_mat_2367ABEF_02_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_02, (_MM_PERM_ENUM)136); //B02(16-19) B03(16-19) B02(16-19) B03(16-19) B06(16-19) B07(16-19) B06(16-19) B07(16-19) B0A(16-19) B0B(16-19) B0A(16-19) B0B(16-19) B0E(16-19) B0F(16-19) B0E(16-19) B0F(16-19) + const __m512i rhs_mat_014589CD_03_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_03, (_MM_PERM_ENUM)136); //B00(24-27) B01(24-27) B00(24-27) B01(24-27) B04(24-27) B05(24-27) B04(24-27) B05(24-27) B08(24-27) B09(24-27) B08(24-27) B09(24-27) B0C(24-27) B0D(24-27) B0C(24-27) B0D(24-27) + const __m512i rhs_mat_2367ABEF_03_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_03, (_MM_PERM_ENUM)136); //B02(24-27) B03(24-27) B02(24-27) B03(24-27) B06(24-27) B07(24-27) B06(24-27) B07(24-27) B0A(24-27) B0B(24-27) B0A(24-27) B0B(24-27) B0E(24-27) B0F(24-27) B0E(24-27) B0F(24-27) + + const __m512i rhs_mat_014589CD_10_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_10, (_MM_PERM_ENUM)136); //B10(0-3) B11(0-3) B10(0-3) B11(0-3) B14(0-3) B15(0-3) B14(0-3) B15(0-3) B18(0-3) B19(0-3) B18(0-3) B19(0-3) B1C(0-3) B1D(0-3) B1C(0-3) B1D(0-3) + const __m512i rhs_mat_2367ABEF_10_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_10, (_MM_PERM_ENUM)136); //B12(0-3) B13(0-3) B12(0-3) B13(0-3) B16(0-3) B17(0-3) B16(0-3) B17(0-3) B1A(0-3) B1B(0-3) B1A(0-3) B1B(0-3) B1E(0-3) B1F(0-3) B1E(0-3) B1F(0-3) + const __m512i rhs_mat_014589CD_11_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_11, (_MM_PERM_ENUM)136); //B10(8-11) B11(8-11) B10(8-11) B11(8-11) B14(8-11) B15(8-11) B14(8-11) B15(8-11) B18(8-11) B19(8-11) B18(8-11) B19(8-11) B1C(8-11) B1D(8-11) B1C(8-11) B1D(8-11) + const __m512i rhs_mat_2367ABEF_11_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_11, (_MM_PERM_ENUM)136); //B12(8-11) B13(8-11) B12(8-11) B13(8-11) B16(8-11) B17(8-11) B16(8-11) B17(8-11) B1A(8-11) B1B(8-11) B1A(8-11) B1B(8-11) B1E(8-11) B1F(8-11) B1E(8-11) B1F(8-11) + const __m512i rhs_mat_014589CD_12_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_12, (_MM_PERM_ENUM)136); //B10(16-19) B11(16-19) B10(16-19) B11(16-19) B14(16-19) B15(16-19) B14(16-19) B15(16-19) B18(16-19) B19(16-19) B18(16-19) B19(16-19) B1C(16-19) B1D(16-19) B1C(16-19) B1D(16-19) + const __m512i rhs_mat_2367ABEF_12_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_12, (_MM_PERM_ENUM)136); //B12(16-19) B13(16-19) B12(16-19) B13(16-19) B16(16-19) B17(16-19) B16(16-19) B17(16-19) B1A(16-19) B1B(16-19) B1A(16-19) B1B(16-19) B1E(16-19) B1F(16-19) B1E(16-19) B1F(16-19) + const __m512i rhs_mat_014589CD_13_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_13, (_MM_PERM_ENUM)136); //B10(24-27) B11(24-27) B10(24-27) B11(24-27) B14(24-27) B15(24-27) B14(24-27) B15(24-27) B18(24-27) B19(24-27) B18(24-27) B19(24-27) B1C(24-27) B1D(24-27) B1C(24-27) B1D(24-27) + const __m512i rhs_mat_2367ABEF_13_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_13, (_MM_PERM_ENUM)136); //B12(24-27) B13(24-27) B12(24-27) B13(24-27) B16(24-27) B17(24-27) B16(24-27) B17(24-27) B1A(24-27) B1B(24-27) B1A(24-27) B1B(24-27) B1E(24-27) B1F(24-27) B1E(24-27) B1F(24-27) + + // Shuffle pattern two - right side input + const __m512i rhs_mat_014589CD_00_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_00, (_MM_PERM_ENUM)221); //B00(4-7) B01(4-7) B00(4-7) B01(4-7) B04(4-7) B05(4-7) B04(4-7) B05(4-7) B08(4-7) B09(4-7) B08(4-7) B09(4-7) B0C(4-7) B0D(4-7) B0C(4-7) B0D(4-7) + const __m512i rhs_mat_2367ABEF_00_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_00, (_MM_PERM_ENUM)221); //B02(4-7) B03(4-7) B02(4-7) B03(4-7) B06(4-7) B07(4-7) B06(4-7) B07(4-7) B0A(4-7) B0B(4-7) B0A(4-7) B0B(4-7) B0E(4-7) B0F(4-7) B0E(4-7) B0F(4-7) + const __m512i rhs_mat_014589CD_01_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_01, (_MM_PERM_ENUM)221); //B00(12-15) B01(12-15) B00(12-15) B01(12-15) B04(12-15) B05(12-15) B04(12-15) B05(12-15) B08(12-15) B09(12-15) B08(12-15) B09(12-15) B0C(12-15) B0D(12-15) B0C(12-15) B0D(12-15) + const __m512i rhs_mat_2367ABEF_01_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_01, (_MM_PERM_ENUM)221); //B02(12-15) B03(12-15) B02(12-15) B03(12-15) B06(12-15) B07(12-15) B06(12-15) B07(12-15) B0A(12-15) B0B(12-15) B0A(12-15) B0B(12-15) B0E(12-15) B0F(12-15) B0E(12-15) B0F(12-15) + const __m512i rhs_mat_014589CD_02_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_02, (_MM_PERM_ENUM)221); //B00(20-23) B01(20-23) B00(20-23) B01(20-23) B04(20-23) B05(20-23) B04(20-23) B05(20-23) B08(20-23) B09(20-23) B08(20-23) B09(20-23) B0C(20-23) B0D(20-23) B0C(20-23) B0D(20-23) + const __m512i rhs_mat_2367ABEF_02_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_02, (_MM_PERM_ENUM)221); //B02(20-23) B03(20-23) B02(20-23) B03(20-23) B06(20-23) B07(20-23) B06(20-23) B07(20-23) B0A(20-23) B0B(20-23) B0A(20-23) B0B(20-23) B0E(20-23) B0F(20-23) B0E(20-23) B0F(20-23) + const __m512i rhs_mat_014589CD_03_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_03, (_MM_PERM_ENUM)221); //B00(28-31) B01(28-31) B00(28-31) B01(28-31) B04(28-31) B05(28-31) B04(28-31) B05(28-31) B08(28-31) B09(28-31) B08(28-31) B09(28-31) B0C(28-31) B0D(28-31) B0C(28-31) 0BD(28-31) + const __m512i rhs_mat_2367ABEF_03_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_03, (_MM_PERM_ENUM)221); //B02(28-31) B03(28-31) B02(28-31) B03(28-31) B06(28-31) B07(28-31) B06(28-31) B07(28-31) B0A(28-31) B0B(28-31) B0A(28-31) B0B(28-31) B0E(28-31) B0F(28-31) B0E(28-31) B0F(28-31) + + const __m512i rhs_mat_014589CD_10_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_10, (_MM_PERM_ENUM)221); //B10(4-7) B11(4-7) B10(4-7) B11(4-7) B14(4-7) B15(4-7) B14(4-7) B15(4-7) B18(4-7) B19(4-7) B18(4-7) B19(4-7) B1C(4-7) B1D(4-7) B1C(4-7) B1D(4-7) + const __m512i rhs_mat_2367ABEF_10_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_10, (_MM_PERM_ENUM)221); //B12(4-7) B13(4-7) B12(4-7) B13(4-7) B16(4-7) B17(4-7) B16(4-7) B17(4-7) B1A(4-7) B1B(4-7) B1A(4-7) B1B(4-7) B1E(4-7) B1F(4-7) B1E(4-7) B1F(4-7) + const __m512i rhs_mat_014589CD_11_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_11, (_MM_PERM_ENUM)221); //B10(12-15) B11(12-15) B10(12-15) B11(12-15) B14(12-15) B15(12-15) B14(12-15) B15(12-15) B18(12-15) B19(12-15) B18(12-15) B19(12-15) B1C(12-15) B1D(12-15) B1C(12-15) B1D(12-15) + const __m512i rhs_mat_2367ABEF_11_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_11, (_MM_PERM_ENUM)221); //B12(12-15) B13(12-15) B12(12-15) B13(12-15) B16(12-15) B17(12-15) B16(12-15) B17(12-15) B1A(12-15) B1B(12-15) B1A(12-15) B1B(12-15) B1E(12-15) B1F(12-15) B1E(12-15) B1F(12-15) + const __m512i rhs_mat_014589CD_12_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_12, (_MM_PERM_ENUM)221); //B10(20-23) B11(20-23) B10(20-23) B11(20-23) B14(20-23) B15(20-23) B14(20-23) B15(20-23) B18(20-23) B19(20-23) B18(20-23) B19(20-23) B1C(20-23) B1D(20-23) B1C(20-23) B1D(20-23) + const __m512i rhs_mat_2367ABEF_12_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_12, (_MM_PERM_ENUM)221); //B12(20-23) B13(20-23) B12(20-23) B13(20-23) B16(20-23) B17(20-23) B16(20-23) B17(20-23) B1A(20-23) B1B(20-23) B1A(20-23) B1B(20-23) B1E(20-23) B1F(20-23) B1E(20-23) B1F(20-23) + const __m512i rhs_mat_014589CD_13_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_13, (_MM_PERM_ENUM)221); //B10(28-31) B11(28-31) B10(28-31) B11(28-31) B14(28-31) B15(28-31) B14(28-31) B15(28-31) B18(28-31) B19(28-31) B18(28-31) B19(28-31) B1C(28-31) B1D(28-31) B1C(28-31) B1D(28-31) + const __m512i rhs_mat_2367ABEF_13_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_13, (_MM_PERM_ENUM)221); //B12(28-31) B13(28-31) B12(28-31) B13(28-31) B16(28-31) B17(28-31) B16(28-31) B17(28-31) B1A(28-31) B1B(28-31) B1A(28-31) B1B(28-31) B1E(28-31) B1F(28-31) B1E(28-31) B1F(28-31) + + uint32_t utmp_00[4], utmp_01[4], utmp_10[4], utmp_11[4]; + + // Scales and Mins of corresponding sub blocks from different Q4_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_00, b_ptr_0[b].scales + 24 * sb, 12); + utmp_00[3] = ((utmp_00[2] >> 4) & kmask2) | (((utmp_00[1] >> 6) & kmask3) << 4); + const uint32_t uaux_00 = utmp_00[1] & kmask1; + utmp_00[1] = (utmp_00[2] & kmask2) | (((utmp_00[0] >> 6) & kmask3) << 4); + utmp_00[2] = uaux_00; + utmp_00[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_01, b_ptr_0[b].scales + 12 + sb * 24, 12); + utmp_01[3] = ((utmp_01[2] >> 4) & kmask2) | (((utmp_01[1] >> 6) & kmask3) << 4); + const uint32_t uaux_01 = utmp_01[1] & kmask1; + utmp_01[1] = (utmp_01[2] & kmask2) | (((utmp_01[0] >> 6) & kmask3) << 4); + utmp_01[2] = uaux_01; + utmp_01[0] &= kmask1; + + // 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_10, b_ptr_1[b].scales + sb * 24, 12); + utmp_10[3] = ((utmp_10[2] >> 4) & kmask2) | (((utmp_10[1] >> 6) & kmask3) << 4); + const uint32_t uaux_10 = utmp_10[1] & kmask1; + utmp_10[1] = (utmp_10[2] & kmask2) | (((utmp_10[0] >> 6) & kmask3) << 4); + utmp_10[2] = uaux_10; + utmp_10[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_11, b_ptr_1[b].scales + 12 + sb * 24, 12); + utmp_11[3] = ((utmp_11[2] >> 4) & kmask2) | (((utmp_11[1] >> 6) & kmask3) << 4); + const uint32_t uaux_11 = utmp_11[1] & kmask1; + utmp_11[1] = (utmp_11[2] & kmask2) | (((utmp_11[0] >> 6) & kmask3) << 4); + utmp_11[2] = uaux_11; + utmp_11[0] &= kmask1; + + // Scales of first sub block in the sb loop + const __m256i mins_and_scales_0 = _mm256_set_epi32(utmp_10[3], utmp_10[2], utmp_10[1], utmp_10[0], utmp_00[3], utmp_00[2], utmp_00[1], utmp_00[0]); + const __m512i scales_0 = _mm512_cvtepu8_epi16(_mm256_unpacklo_epi8(mins_and_scales_0, mins_and_scales_0)); + + // Scales of second sub block in the sb loop + const __m256i mins_and_scales_1 = _mm256_set_epi32(utmp_11[3], utmp_11[2], utmp_11[1], utmp_11[0], utmp_01[3], utmp_01[2], utmp_01[1], utmp_01[0]); + const __m512i scales_1 = _mm512_cvtepu8_epi16(_mm256_unpacklo_epi8(mins_and_scales_1, mins_and_scales_1)); + + // Mins of first and second sub block of Q4_K block are arranged side by side + const __m512i mins_01 = _mm512_cvtepu8_epi16(_mm256_unpacklo_epi8(_mm256_shuffle_epi32(mins_and_scales_0, 78), _mm256_shuffle_epi32(mins_and_scales_1, 78))); + + const __m512i scale_014589CD_0 = _mm512_shuffle_epi32(scales_0, (_MM_PERM_ENUM)68); + const __m512i scale_2367ABEF_0 = _mm512_shuffle_epi32(scales_0, (_MM_PERM_ENUM)238); + + const __m512i scale_014589CD_1 = _mm512_shuffle_epi32(scales_1, (_MM_PERM_ENUM)68); + const __m512i scale_2367ABEF_1 = _mm512_shuffle_epi32(scales_1, (_MM_PERM_ENUM)238); + + // Load the four block_q8_k quantized values interleaved with each other in chunks of eight bytes - A0,A1,A2,A3 + // Loaded as set of 128 bit vectors and repeated into a 256 bit vector + __m256i lhs_mat_ymm_0123_00 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 256 * sb))); + __m256i lhs_mat_ymm_01_00 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_00, lhs_mat_ymm_0123_00, 0); + __m256i lhs_mat_ymm_23_00 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_00, lhs_mat_ymm_0123_00, 17); + __m256i lhs_mat_ymm_0123_01 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 32 + 256 * sb))); + __m256i lhs_mat_ymm_01_01 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_01, lhs_mat_ymm_0123_01, 0); + __m256i lhs_mat_ymm_23_01 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_01, lhs_mat_ymm_0123_01, 17); + __m256i lhs_mat_ymm_0123_02 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 64 + 256 * sb))); + __m256i lhs_mat_ymm_01_02 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_02, lhs_mat_ymm_0123_02, 0); + __m256i lhs_mat_ymm_23_02 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_02, lhs_mat_ymm_0123_02, 17); + __m256i lhs_mat_ymm_0123_03 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 96 + 256 * sb))); + __m256i lhs_mat_ymm_01_03 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_03, lhs_mat_ymm_0123_03, 0); + __m256i lhs_mat_ymm_23_03 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_03, lhs_mat_ymm_0123_03, 17); + __m256i lhs_mat_ymm_0123_10 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 128 + 256 * sb))); + __m256i lhs_mat_ymm_01_10 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_10, lhs_mat_ymm_0123_10, 0); + __m256i lhs_mat_ymm_23_10 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_10, lhs_mat_ymm_0123_10, 17); + __m256i lhs_mat_ymm_0123_11 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 160 + 256 * sb))); + __m256i lhs_mat_ymm_01_11 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_11, lhs_mat_ymm_0123_11, 0); + __m256i lhs_mat_ymm_23_11 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_11, lhs_mat_ymm_0123_11, 17); + __m256i lhs_mat_ymm_0123_12 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 192 + 256 * sb))); + __m256i lhs_mat_ymm_01_12 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_12, lhs_mat_ymm_0123_12, 0); + __m256i lhs_mat_ymm_23_12 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_12, lhs_mat_ymm_0123_12, 17); + __m256i lhs_mat_ymm_0123_13 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 224 + 256 * sb))); + __m256i lhs_mat_ymm_01_13 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_13, lhs_mat_ymm_0123_13, 0); + __m256i lhs_mat_ymm_23_13 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_13, lhs_mat_ymm_0123_13, 17); + + //Loaded as set of 128 bit vectors and repeated and stored into a 256 bit vector before again repeating into a 512 bit vector + __m512i lhs_mat_01_00 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_00), lhs_mat_ymm_01_00, 1); + __m512i lhs_mat_23_00 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_00), lhs_mat_ymm_23_00, 1); + __m512i lhs_mat_01_01 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_01), lhs_mat_ymm_01_01, 1); + __m512i lhs_mat_23_01 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_01), lhs_mat_ymm_23_01, 1); + __m512i lhs_mat_01_02 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_02), lhs_mat_ymm_01_02, 1); + __m512i lhs_mat_23_02 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_02), lhs_mat_ymm_23_02, 1); + __m512i lhs_mat_01_03 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_03), lhs_mat_ymm_01_03, 1); + __m512i lhs_mat_23_03 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_03), lhs_mat_ymm_23_03, 1); + + __m512i lhs_mat_01_10 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_10), lhs_mat_ymm_01_10, 1); + __m512i lhs_mat_23_10 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_10), lhs_mat_ymm_23_10, 1); + __m512i lhs_mat_01_11 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_11), lhs_mat_ymm_01_11, 1); + __m512i lhs_mat_23_11 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_11), lhs_mat_ymm_23_11, 1); + __m512i lhs_mat_01_12 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_12), lhs_mat_ymm_01_12, 1); + __m512i lhs_mat_23_12 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_12), lhs_mat_ymm_23_12, 1); + __m512i lhs_mat_01_13 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_13), lhs_mat_ymm_01_13, 1); + __m512i lhs_mat_23_13 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_13), lhs_mat_ymm_23_13, 1); + + // Bsums are loaded - four bsums are loaded (for two sub blocks) for the different Q8_K blocks + __m256i lhs_bsums_0123_01 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].bsums + 16 * sb))); + __m256i lhs_bsums_hsum_ymm_0123_01 = _mm256_castsi128_si256(_mm_hadd_epi16(_mm256_castsi256_si128(lhs_bsums_0123_01), _mm256_extractf128_si256(lhs_bsums_0123_01, 1))); + lhs_bsums_hsum_ymm_0123_01 = _mm256_permute2x128_si256(lhs_bsums_hsum_ymm_0123_01, lhs_bsums_hsum_ymm_0123_01, 0); + __m512i lhs_bsums_hsum_0123_01 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_bsums_hsum_ymm_0123_01), lhs_bsums_hsum_ymm_0123_01, 1); + + // Shuffle pattern one - left side input + const __m512i lhs_mat_01_00_sp1 = _mm512_shuffle_epi32(lhs_mat_01_00, (_MM_PERM_ENUM)160); //A00(0-3) A00(0-3) A01(0-3) A01(0-3) A00(0-3) A00(0-3) A01(0-3) A01(0-3) A00(0-3) A00(0-3) A01(0-3) A01(0-3) A00(0-3) A00(0-3) A01(0-3) A01(0-3) + const __m512i lhs_mat_23_00_sp1 = _mm512_shuffle_epi32(lhs_mat_23_00, (_MM_PERM_ENUM)160); //A02(0-3) A02(0-3) A03(0-3) A03(0-3) A02(0-3) A02(0-3) A03(0-3) A03(0-3) A02(0-3) A02(0-3) A03(0-3) A03(0-3) A02(0-3) A02(0-3) A03(0-3) A03(0-3) + const __m512i lhs_mat_01_01_sp1 = _mm512_shuffle_epi32(lhs_mat_01_01, (_MM_PERM_ENUM)160); //A00(8-11) A00(8-11) A01(8-11) A01(8-11) A00(8-11) A00(8-11) A01(8-11) A01(8-11) A00(8-11) A00(8-11) A01(8-11) A01(8-11) A00(8-11) A00(8-11) A01(8-11) A01(8-11) + const __m512i lhs_mat_23_01_sp1 = _mm512_shuffle_epi32(lhs_mat_23_01, (_MM_PERM_ENUM)160); //A02(8-11) A02(8-11) A03(8-11) A03(8-11) A02(8-11) A02(8-11) A03(8-11) A03(8-11) A02(8-11) A02(8-11) A03(8-11) A03(8-11) A02(8-11) A02(8-11) A03(8-11) A03(8-11) + const __m512i lhs_mat_01_02_sp1 = _mm512_shuffle_epi32(lhs_mat_01_02, (_MM_PERM_ENUM)160); //A00(16-19) A00(16-19) A01(16-19) A01(16-19) A00(16-19) A00(16-19) A01(16-19) A01(16-19) A00(16-19) A00(16-19) A01(16-19) A01(16-19) A00(16-19) A00(16-19) A01(16-19) A01(16-19) + const __m512i lhs_mat_23_02_sp1 = _mm512_shuffle_epi32(lhs_mat_23_02, (_MM_PERM_ENUM)160); //A02(16-19) A02(16-19) A03(16-19) A03(16-19) A02(16-19) A02(16-19) A03(16-19) A03(16-19) A02(16-19) A02(16-19) A03(16-19) A03(16-19) A02(16-19) A02(16-19) A03(16-19) A03(16-19) + const __m512i lhs_mat_01_03_sp1 = _mm512_shuffle_epi32(lhs_mat_01_03, (_MM_PERM_ENUM)160); //A00(24-27) A00(24-27) A01(24-27) A01(24-27) A00(24-27) A00(24-27) A01(24-27) A01(24-27) A00(24-27) A00(24-27) A01(24-27) A01(24-27) A00(24-27) A00(24-27) A01(24-27) A01(24-27) + const __m512i lhs_mat_23_03_sp1 = _mm512_shuffle_epi32(lhs_mat_23_03, (_MM_PERM_ENUM)160); //A02(24-27) A02(24-27) A03(24-27) A03(24-27) A02(24-27) A02(24-27) A03(24-27) A03(24-27) A02(24-27) A02(24-27) A03(24-27) A03(24-27) A02(24-27) A02(24-27) A03(24-27) A03(24-27) + + const __m512i lhs_mat_01_10_sp1 = _mm512_shuffle_epi32(lhs_mat_01_10, (_MM_PERM_ENUM)160); //A10(0-3) A10(0-3) A11(0-3) A11(0-3) A10(0-3) A10(0-3) A11(0-3) A11(0-3) A10(0-3) A10(0-3) A11(0-3) A11(0-3) A10(0-3) A10(0-3) A11(0-3) A11(0-3) + const __m512i lhs_mat_23_10_sp1 = _mm512_shuffle_epi32(lhs_mat_23_10, (_MM_PERM_ENUM)160); //A12(0-3) A12(0-3) A13(0-3) A13(0-3) A12(0-3) A12(0-3) A13(0-3) A13(0-3) A12(0-3) A12(0-3) A13(0-3) A13(0-3) A12(0-3) A12(0-3) A13(0-3) A13(0-3) + const __m512i lhs_mat_01_11_sp1 = _mm512_shuffle_epi32(lhs_mat_01_11, (_MM_PERM_ENUM)160); //A10(8-11) A10(8-11) A11(8-11) A11(8-11) A10(8-11) A10(8-11) A11(8-11) A11(8-11) A10(8-11) A10(8-11) A11(8-11) A11(8-11) A10(8-11) A10(8-11) A11(8-11) A11(8-11) + const __m512i lhs_mat_23_11_sp1 = _mm512_shuffle_epi32(lhs_mat_23_11, (_MM_PERM_ENUM)160); //A12(8-11) A12(8-11) A13(8-11) A13(8-11) A12(8-11) A12(8-11) A13(8-11) A13(8-11) A12(8-11) A12(8-11) A13(8-11) A13(8-11) A12(8-11) A12(8-11) A13(8-11) A13(8-11) + const __m512i lhs_mat_01_12_sp1 = _mm512_shuffle_epi32(lhs_mat_01_12, (_MM_PERM_ENUM)160); //A10(16-19) A10(16-19) A11(16-19) A11(16-19) A10(16-19) A10(16-19) A11(16-19) A11(16-19) A10(16-19) A10(16-19) A11(16-19) A11(16-19) A10(16-19) A10(16-19) A11(16-19) A11(16-19) + const __m512i lhs_mat_23_12_sp1 = _mm512_shuffle_epi32(lhs_mat_23_12, (_MM_PERM_ENUM)160); //A12(16-19) A12(16-19) A13(16-19) A13(16-19) A12(16-19) A12(16-19) A13(16-19) A13(16-19) A12(16-19) A12(16-19) A13(16-19) A13(16-19) A12(16-19) A12(16-19) A13(16-19) A13(16-19) + const __m512i lhs_mat_01_13_sp1 = _mm512_shuffle_epi32(lhs_mat_01_13, (_MM_PERM_ENUM)160); //A10(24-27) A10(24-27) A11(24-27) A11(24-27) A10(24-27) A10(24-27) A11(24-27) A11(24-27) A10(24-27) A10(24-27) A11(24-27) A11(24-27) A10(24-27) A10(24-27) A11(24-27) A11(24-27) + const __m512i lhs_mat_23_13_sp1 = _mm512_shuffle_epi32(lhs_mat_23_13, (_MM_PERM_ENUM)160); //A12(24-27) A12(24-27) A13(24-27) A13(24-27) A12(24-27) A12(24-27) A13(24-27) A13(24-27) A12(24-27) A12(24-27) A13(24-27) A13(24-27) A12(24-27) A12(24-27) A13(24-27) A13(24-27) + + const __m512i lhs_mat_01_00_sp2 = _mm512_shuffle_epi32(lhs_mat_01_00, (_MM_PERM_ENUM)245); //A00(4-7) A00(4-7) A01(4-7) A01(4-7) A00(4-7) A00(4-7) A01(4-7) A01(4-7) A00(4-7) A00(4-7) A01(4-7) A01(4-7) A00(4-7) A00(4-7) A01(4-7) A01(4-7) + const __m512i lhs_mat_23_00_sp2 = _mm512_shuffle_epi32(lhs_mat_23_00, (_MM_PERM_ENUM)245); //A02(4-7) A02(4-7) A03(4-7) A03(4-7) A02(4-7) A02(4-7) A03(4-7) A03(4-7) A02(4-7) A02(4-7) A03(4-7) A03(4-7) A02(4-7) A02(4-7) A03(4-7) A03(4-7) + const __m512i lhs_mat_01_01_sp2 = _mm512_shuffle_epi32(lhs_mat_01_01, (_MM_PERM_ENUM)245); //A00(12-15) A00(12-15) A01(12-15) A01(12-15) A00(12-15) A00(12-15) A01(12-15) A01(12-15) A00(12-15) A00(12-15) A01(12-15) A01(12-15) A00(12-15) A00(12-15) A01(12-15) A01(12-15) + const __m512i lhs_mat_23_01_sp2 = _mm512_shuffle_epi32(lhs_mat_23_01, (_MM_PERM_ENUM)245); //A02(12-15) A02(12-15) A03(12-15) A03(12-15) A02(12-15) A02(12-15) A03(12-15) A03(12-15) A02(12-15) A02(12-15) A03(12-15) A03(12-15) A02(12-15) A02(12-15) A03(12-15) A03(12-15) + const __m512i lhs_mat_01_02_sp2 = _mm512_shuffle_epi32(lhs_mat_01_02, (_MM_PERM_ENUM)245); //A00(20-23) A00(20-23) A01(20-23) A01(20-23) A00(20-23) A00(20-23) A01(20-23) A01(20-23) A00(20-23) A00(20-23) A01(20-23) A01(20-23) A00(20-23) A00(20-23) A01(20-23) A01(20-23) + const __m512i lhs_mat_23_02_sp2 = _mm512_shuffle_epi32(lhs_mat_23_02, (_MM_PERM_ENUM)245); //A02(20-23) A02(20-23) A03(20-23) A03(20-23) A02(20-23) A02(20-23) A03(20-23) A03(20-23) A02(20-23) A02(20-23) A03(20-23) A03(20-23) A02(20-23) A02(20-23) A03(20-23) A03(20-23) + const __m512i lhs_mat_01_03_sp2 = _mm512_shuffle_epi32(lhs_mat_01_03, (_MM_PERM_ENUM)245); //A00(28-31) A00(28-31) A01(28-31) A01(28-31) A00(28-31) A00(28-31) A01(28-31) A01(28-31) A00(28-31) A00(28-31) A01(28-31) A01(28-31) A00(28-31) A00(28-31) A01(28-31) A01(28-31) + const __m512i lhs_mat_23_03_sp2 = _mm512_shuffle_epi32(lhs_mat_23_03, (_MM_PERM_ENUM)245); //A02(28-31) A02(28-31) A03(28-31) A03(28-31) A02(28-31) A02(28-31) A03(28-31) A03(28-31) A02(28-31) A02(28-31) A03(28-31) A03(28-31) A02(28-31) A02(28-31) A03(28-31) A03(28-31) + + const __m512i lhs_mat_01_10_sp2 = _mm512_shuffle_epi32(lhs_mat_01_10, (_MM_PERM_ENUM)245); //A10(4-7) A10(4-7) A11(4-7) A11(4-7) A10(4-7) A10(4-7) A11(4-7) A11(4-7) A10(4-7) A10(4-7) A11(4-7) A11(4-7) A10(4-7) A10(4-7) A11(4-7) A11(4-7) + const __m512i lhs_mat_23_10_sp2 = _mm512_shuffle_epi32(lhs_mat_23_10, (_MM_PERM_ENUM)245); //A12(4-7) A12(4-7) A13(4-7) A13(4-7) A12(4-7) A12(4-7) A13(4-7) A13(4-7) A12(4-7) A12(4-7) A13(4-7) A13(4-7) A12(4-7) A12(4-7) A13(4-7) A13(4-7) + const __m512i lhs_mat_01_11_sp2 = _mm512_shuffle_epi32(lhs_mat_01_11, (_MM_PERM_ENUM)245); //A10(12-15) A10(12-15) A11(12-15) A11(12-15) A10(12-15) A10(12-15) A11(12-15) A11(12-15) A10(12-15) A10(12-15) A11(12-15) A11(12-15) A10(12-15) A10(12-15) A11(12-15) A11(12-15) + const __m512i lhs_mat_23_11_sp2 = _mm512_shuffle_epi32(lhs_mat_23_11, (_MM_PERM_ENUM)245); //A12(12-15) A12(12-15) A13(12-15) A13(12-15) A12(12-15) A12(12-15) A13(12-15) A13(12-15) A12(12-15) A12(12-15) A13(12-15) A13(12-15) A12(12-15) A12(12-15) A13(12-15) A13(12-15) + const __m512i lhs_mat_01_12_sp2 = _mm512_shuffle_epi32(lhs_mat_01_12, (_MM_PERM_ENUM)245); //A10(20-23) A10(20-23) A11(20-23) A11(20-23) A10(20-23) A10(20-23) A11(20-23) A11(20-23) A10(20-23) A10(20-23) A11(20-23) A11(20-23) A10(20-23) A10(20-23) A11(20-23) A11(20-23) + const __m512i lhs_mat_23_12_sp2 = _mm512_shuffle_epi32(lhs_mat_23_12, (_MM_PERM_ENUM)245); //A12(20-23) A12(20-23) A13(20-23) A13(20-23) A12(20-23) A12(20-23) A13(20-23) A13(20-23) A12(20-23) A12(20-23) A13(20-23) A13(20-23) A12(20-23) A12(20-23) A13(20-23) A13(20-23) + const __m512i lhs_mat_01_13_sp2 = _mm512_shuffle_epi32(lhs_mat_01_13, (_MM_PERM_ENUM)245); //A10(28-31) A10(28-31) A11(28-31) A11(28-31) A10(28-31) A10(28-31) A11(28-31) A11(28-31) A10(28-31) A10(28-31) A11(28-31) A11(28-31) A10(28-31) A10(28-31) A11(28-31) A11(28-31) + const __m512i lhs_mat_23_13_sp2 = _mm512_shuffle_epi32(lhs_mat_23_13, (_MM_PERM_ENUM)245); //A12(28-31) A12(28-31) A13(28-31) A13(28-31) A12(28-31) A12(28-31) A13(28-31) A13(28-31) A12(28-31) A12(28-31) A13(28-31) A13(28-31) A12(28-31) A12(28-31) A13(28-31) A13(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 + __m512i iacc_mat_00_0_sp1 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_03_sp1, lhs_mat_01_03_sp1), _mm512_maddubs_epi16(rhs_mat_014589CD_02_sp1, lhs_mat_01_02_sp1)), _mm512_maddubs_epi16(rhs_mat_014589CD_01_sp1, lhs_mat_01_01_sp1)), _mm512_maddubs_epi16(rhs_mat_014589CD_00_sp1, lhs_mat_01_00_sp1)); + __m512i iacc_mat_01_0_sp1 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_03_sp1, lhs_mat_01_03_sp1), _mm512_maddubs_epi16(rhs_mat_2367ABEF_02_sp1, lhs_mat_01_02_sp1)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_01_sp1, lhs_mat_01_01_sp1)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_00_sp1, lhs_mat_01_00_sp1)); + __m512i iacc_mat_10_0_sp1 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_03_sp1, lhs_mat_23_03_sp1), _mm512_maddubs_epi16(rhs_mat_014589CD_02_sp1, lhs_mat_23_02_sp1)), _mm512_maddubs_epi16(rhs_mat_014589CD_01_sp1, lhs_mat_23_01_sp1)), _mm512_maddubs_epi16(rhs_mat_014589CD_00_sp1, lhs_mat_23_00_sp1)); + __m512i iacc_mat_11_0_sp1 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_03_sp1, lhs_mat_23_03_sp1), _mm512_maddubs_epi16(rhs_mat_2367ABEF_02_sp1, lhs_mat_23_02_sp1)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_01_sp1, lhs_mat_23_01_sp1)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_00_sp1, lhs_mat_23_00_sp1)); + __m512i iacc_mat_00_1_sp1 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_13_sp1, lhs_mat_01_13_sp1), _mm512_maddubs_epi16(rhs_mat_014589CD_12_sp1, lhs_mat_01_12_sp1)), _mm512_maddubs_epi16(rhs_mat_014589CD_11_sp1, lhs_mat_01_11_sp1)), _mm512_maddubs_epi16(rhs_mat_014589CD_10_sp1, lhs_mat_01_10_sp1)); + __m512i iacc_mat_01_1_sp1 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_13_sp1, lhs_mat_01_13_sp1), _mm512_maddubs_epi16(rhs_mat_2367ABEF_12_sp1, lhs_mat_01_12_sp1)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_11_sp1, lhs_mat_01_11_sp1)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_10_sp1, lhs_mat_01_10_sp1)); + __m512i iacc_mat_10_1_sp1 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_13_sp1, lhs_mat_23_13_sp1), _mm512_maddubs_epi16(rhs_mat_014589CD_12_sp1, lhs_mat_23_12_sp1)), _mm512_maddubs_epi16(rhs_mat_014589CD_11_sp1, lhs_mat_23_11_sp1)), _mm512_maddubs_epi16(rhs_mat_014589CD_10_sp1, lhs_mat_23_10_sp1)); + __m512i iacc_mat_11_1_sp1 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_13_sp1, lhs_mat_23_13_sp1), _mm512_maddubs_epi16(rhs_mat_2367ABEF_12_sp1, lhs_mat_23_12_sp1)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_11_sp1, lhs_mat_23_11_sp1)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_10_sp1, lhs_mat_23_10_sp1)); + + __m512i iacc_mat_00_0_sp2 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_03_sp2, lhs_mat_01_03_sp2), _mm512_maddubs_epi16(rhs_mat_014589CD_02_sp2, lhs_mat_01_02_sp2)), _mm512_maddubs_epi16(rhs_mat_014589CD_01_sp2, lhs_mat_01_01_sp2)), _mm512_maddubs_epi16(rhs_mat_014589CD_00_sp2, lhs_mat_01_00_sp2)); + __m512i iacc_mat_01_0_sp2 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_03_sp2, lhs_mat_01_03_sp2), _mm512_maddubs_epi16(rhs_mat_2367ABEF_02_sp2, lhs_mat_01_02_sp2)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_01_sp2, lhs_mat_01_01_sp2)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_00_sp2, lhs_mat_01_00_sp2)); + __m512i iacc_mat_10_0_sp2 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_03_sp2, lhs_mat_23_03_sp2), _mm512_maddubs_epi16(rhs_mat_014589CD_02_sp2, lhs_mat_23_02_sp2)), _mm512_maddubs_epi16(rhs_mat_014589CD_01_sp2, lhs_mat_23_01_sp2)), _mm512_maddubs_epi16(rhs_mat_014589CD_00_sp2, lhs_mat_23_00_sp2)); + __m512i iacc_mat_11_0_sp2 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_03_sp2, lhs_mat_23_03_sp2), _mm512_maddubs_epi16(rhs_mat_2367ABEF_02_sp2, lhs_mat_23_02_sp2)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_01_sp2, lhs_mat_23_01_sp2)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_00_sp2, lhs_mat_23_00_sp2)); + __m512i iacc_mat_00_1_sp2 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_13_sp2, lhs_mat_01_13_sp2), _mm512_maddubs_epi16(rhs_mat_014589CD_12_sp2, lhs_mat_01_12_sp2)), _mm512_maddubs_epi16(rhs_mat_014589CD_11_sp2, lhs_mat_01_11_sp2)), _mm512_maddubs_epi16(rhs_mat_014589CD_10_sp2, lhs_mat_01_10_sp2)); + __m512i iacc_mat_01_1_sp2 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_13_sp2, lhs_mat_01_13_sp2), _mm512_maddubs_epi16(rhs_mat_2367ABEF_12_sp2, lhs_mat_01_12_sp2)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_11_sp2, lhs_mat_01_11_sp2)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_10_sp2, lhs_mat_01_10_sp2)); + __m512i iacc_mat_10_1_sp2 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_13_sp2, lhs_mat_23_13_sp2), _mm512_maddubs_epi16(rhs_mat_014589CD_12_sp2, lhs_mat_23_12_sp2)), _mm512_maddubs_epi16(rhs_mat_014589CD_11_sp2, lhs_mat_23_11_sp2)), _mm512_maddubs_epi16(rhs_mat_014589CD_10_sp2, lhs_mat_23_10_sp2)); + __m512i iacc_mat_11_1_sp2 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_13_sp2, lhs_mat_23_13_sp2), _mm512_maddubs_epi16(rhs_mat_2367ABEF_12_sp2, lhs_mat_23_12_sp2)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_11_sp2, lhs_mat_23_11_sp2)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_10_sp2, lhs_mat_23_10_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_0 = _mm512_add_epi16(iacc_mat_00_0_sp1, iacc_mat_00_0_sp2); + __m512i iacc_mat_01_0 = _mm512_add_epi16(iacc_mat_01_0_sp1, iacc_mat_01_0_sp2); + __m512i iacc_mat_10_0 = _mm512_add_epi16(iacc_mat_10_0_sp1, iacc_mat_10_0_sp2); + __m512i iacc_mat_11_0 = _mm512_add_epi16(iacc_mat_11_0_sp1, iacc_mat_11_0_sp2); + + __m512i iacc_mat_00_1 = _mm512_add_epi16(iacc_mat_00_1_sp1, iacc_mat_00_1_sp2); + __m512i iacc_mat_01_1 = _mm512_add_epi16(iacc_mat_01_1_sp1, iacc_mat_01_1_sp2); + __m512i iacc_mat_10_1 = _mm512_add_epi16(iacc_mat_10_1_sp1, iacc_mat_10_1_sp2); + __m512i iacc_mat_11_1 = _mm512_add_epi16(iacc_mat_11_1_sp1, iacc_mat_11_1_sp2); + + iacc_mat_00_0 = _mm512_madd_epi16(iacc_mat_00_0, scale_014589CD_0); + iacc_mat_01_0 = _mm512_madd_epi16(iacc_mat_01_0, scale_2367ABEF_0); + iacc_mat_10_0 = _mm512_madd_epi16(iacc_mat_10_0, scale_014589CD_0); + iacc_mat_11_0 = _mm512_madd_epi16(iacc_mat_11_0, scale_2367ABEF_0); + + iacc_mat_00_1 = _mm512_madd_epi16(iacc_mat_00_1, scale_014589CD_1); + iacc_mat_01_1 = _mm512_madd_epi16(iacc_mat_01_1, scale_2367ABEF_1); + iacc_mat_10_1 = _mm512_madd_epi16(iacc_mat_10_1, scale_014589CD_1); + iacc_mat_11_1 = _mm512_madd_epi16(iacc_mat_11_1, scale_2367ABEF_1); + + // Straighten out to make 4 row vectors (4 for each sub block which are accumulated together in the next step) + __m512i iacc_row_0_0 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_00_0, _mm512_shuffle_epi32(iacc_mat_01_0, (_MM_PERM_ENUM)78)); + __m512i iacc_row_1_0 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_00_0, (_MM_PERM_ENUM)78), iacc_mat_01_0); + __m512i iacc_row_2_0 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_10_0, _mm512_shuffle_epi32(iacc_mat_11_0, (_MM_PERM_ENUM)78)); + __m512i iacc_row_3_0 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_10_0, (_MM_PERM_ENUM)78), iacc_mat_11_0); + __m512i iacc_row_0_1 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_00_1, _mm512_shuffle_epi32(iacc_mat_01_1, (_MM_PERM_ENUM)78)); + __m512i iacc_row_1_1 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_00_1, (_MM_PERM_ENUM)78), iacc_mat_01_1); + __m512i iacc_row_2_1 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_10_1, _mm512_shuffle_epi32(iacc_mat_11_1, (_MM_PERM_ENUM)78)); + __m512i iacc_row_3_1 = _mm512_mask_blend_epi32(0xCCCC,_mm512_shuffle_epi32(iacc_mat_10_1, (_MM_PERM_ENUM)78), iacc_mat_11_1); + + __m512i iacc_row_0 = _mm512_add_epi32(iacc_row_0_0, iacc_row_0_1); + __m512i iacc_row_1 = _mm512_add_epi32(iacc_row_1_0, iacc_row_1_1); + __m512i iacc_row_2 = _mm512_add_epi32(iacc_row_2_0, iacc_row_2_1); + __m512i iacc_row_3 = _mm512_add_epi32(iacc_row_3_0, iacc_row_3_1); + + // Load the scale(d) values for all the 4 Q8_k blocks and repeat it across lanes + const __m128 row_scale_f32_sse = _mm_load_ps(a_ptr[b].d); + const __m256 row_scale_f32_ymm = _mm256_set_m128(row_scale_f32_sse, row_scale_f32_sse); + const __m512 row_scale_f32 = _mm512_insertf32x8(_mm512_castps256_ps512(row_scale_f32_ymm), row_scale_f32_ymm, 1); + + // Multiply with appropiate scales and accumulate (for both d and dmin) below + 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]); + + __m512i iacc_row_min_0 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_hsum_0123_01, (_MM_PERM_ENUM)0), mins_01); + __m512i iacc_row_min_1 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_hsum_0123_01, (_MM_PERM_ENUM)85), mins_01); + __m512i iacc_row_min_2 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_hsum_0123_01, (_MM_PERM_ENUM)170), mins_01); + __m512i iacc_row_min_3 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_hsum_0123_01, (_MM_PERM_ENUM)255), mins_01); + + acc_min_rows[0] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_min_0), _mm512_mul_ps(col_dmin_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 0)), acc_min_rows[0]); + acc_min_rows[1] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_min_1), _mm512_mul_ps(col_dmin_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 85)), acc_min_rows[1]); + acc_min_rows[2] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_min_2), _mm512_mul_ps(col_dmin_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_min_rows[2]); + acc_min_rows[3] = _mm512_fmadd_ps(_mm512_cvtepi32_ps(iacc_row_min_3), _mm512_mul_ps(col_dmin_f32, _mm512_shuffle_ps(row_scale_f32, row_scale_f32, 255)), acc_min_rows[3]); + } + } + // Store accumlated values + for (int i = 0; i < 4; i++) { + _mm512_storeu_ps((float * )(s + ((y * 4 + i) * bs + x * 8)), _mm512_sub_ps(acc_rows[i], acc_min_rows[i])); + } + } + } + if (anc != nc) { + xstart = anc/8; + y = 0; + } +#endif //AVX512F + // Take group of four block_q8_Kx4 structures at each pass of the loop and perform dot product operation for (; y < anr / 4; y += 4) { @@ -4067,7 +4807,7 @@ static void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, c } // Take group of eight block_q4_kx8 structures at each pass of the loop and perform dot product operation - for (int64_t x = 0; x < nc / 8; x++) { + for (int64_t x = xstart; x < nc / 8; x++) { const block_q4_Kx8 * b_ptr = b_ptr_start + (x * b_nb); @@ -4401,7 +5141,7 @@ static void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, c const block_q8_Kx4 * a_ptr = a_ptr_start + (y * nb); - for (int64_t x = 0; x < nc / 8; x++) { + for (int64_t x = xstart; x < nc / 8; x++) { const block_q4_Kx8 * b_ptr = b_ptr_start + (x * b_nb); diff --git a/ggml/src/ggml-sycl/ggml-sycl.cpp b/ggml/src/ggml-sycl/ggml-sycl.cpp index 09800d3403f..4d2fda0bfa6 100644 --- a/ggml/src/ggml-sycl/ggml-sycl.cpp +++ b/ggml/src/ggml-sycl/ggml-sycl.cpp @@ -4009,10 +4009,8 @@ static bool ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, const g case GGML_OP_ROPE: { const int mode = ((const int32_t *) op->op_params)[2]; - if (mode & GGML_ROPE_TYPE_MROPE) { - return false; - } - if (mode & GGML_ROPE_TYPE_VISION) { + // mode is not used as a bitmask in practice, the various rope type modes are independent implementations + if (mode == GGML_ROPE_TYPE_MROPE) { return false; } return ggml_is_contiguous(op->src[0]); diff --git a/ggml/src/ggml-sycl/rope.cpp b/ggml/src/ggml-sycl/rope.cpp index bbcb356e979..80e050f2414 100644 --- a/ggml/src/ggml-sycl/rope.cpp +++ b/ggml/src/ggml-sycl/rope.cpp @@ -1,9 +1,15 @@ #include "rope.hpp" +#include "ggml-sycl/common.hpp" +#include "ggml.h" struct rope_corr_dims { float v[2]; }; +struct mrope_sections { + int v[4]; +}; + static float rope_yarn_ramp(const float low, const float high, const int i0) { const float y = (i0 / 2 - low) / sycl::max(0.001f, high - low); return 1.0f - sycl::min(1.0f, sycl::max(0.0f, y)); @@ -114,6 +120,48 @@ static void rope_neox( dst[i + n_dims/2] = x0*sin_theta + x1*cos_theta; } +template +static void rope_vision(const T * x, T * dst, const int ne0, const int ne1, const int ne2, const size_t s1, + const size_t s2, const int n_dims, const int32_t * pos, const float freq_scale, + const float ext_factor, const float attn_factor, const rope_corr_dims corr_dims, + const float theta_scale, const float * freq_factors, const mrope_sections sections, + const sycl::nd_item<3> & item_ct1) { + // get index pos + const int i0 = 2 * (item_ct1.get_group(1) * item_ct1.get_local_range(1) + item_ct1.get_local_id(1)); + if (i0 >= ne0) { + return; + } + const int row_dst = (item_ct1.get_group(2) * item_ct1.get_local_range(2)) + item_ct1.get_local_id(2); + const int row_x = row_dst % ne1; + const int channel_x = row_dst / ne1; + const int idst = (row_dst * ne0) + (i0 / 2); + const size_t ix = ((size_t) channel_x * s2) + ((size_t) row_x * s1) + (i0 / 2); + + const int sect_dims = sections.v[0] + sections.v[1]; + const int sector = (i0 / 2) % sect_dims; + + float theta_base = 0.0f; + if (sector < sections.v[0]) { + const int p = sector; + theta_base = pos[channel_x] * sycl::pow(theta_scale, (float) p); + } else { + // Simplified from CUDA backend code: if (sector >= sections.v[0] && sector < sec_w) which is just sector >= sections.v[0] + const int p = sector - sections.v[0]; + theta_base = pos[channel_x + ne2] * sycl::pow(theta_scale, (float) p); + } + + const float freq_factor = has_ff ? freq_factors[i0 / 2] : 1.0f; + float cos_theta; + float sin_theta; + rope_yarn(theta_base / freq_factor, freq_scale, corr_dims, i0, ext_factor, attn_factor, &cos_theta, &sin_theta); + const float x0 = x[ix + 0]; + const float x1 = x[ix + n_dims]; + + // store results in dst + dst[idst + 0] = x0 * cos_theta - x1 * sin_theta; + dst[idst + n_dims] = x0 * sin_theta + x1 * cos_theta; +} + template static void rope_norm_sycl( const T *x, T *dst, int ne0, int n_dims, int nr, const int32_t *pos, float freq_scale, int p_delta_rows, @@ -192,21 +240,58 @@ static void rope_neox_sycl( } } +// rope vision +template +static void rope_vision_sycl(const T * x, T * dst, const int ne0, const int ne1, const int ne2, const size_t s1, + const size_t s2, const int n_dims, const int nr, const int32_t * pos, + const float freq_scale, const float freq_base, const float ext_factor, + const float attn_factor, const rope_corr_dims corr_dims, const float * freq_factors, + const mrope_sections sections, queue_ptr stream) { + GGML_ASSERT(ne0 % 2 == 0); + const sycl::range<3> block_dims(1, SYCL_ROPE_BLOCK_SIZE, 1); + const int n_blocks_y = (ne0 + 2 * SYCL_ROPE_BLOCK_SIZE - 1) / (2 * SYCL_ROPE_BLOCK_SIZE); + const sycl::range<3> grid_dims(1, n_blocks_y, nr); + const sycl::nd_range<3> nd_range(grid_dims * block_dims, block_dims); + + const float theta_scale = std::pow(freq_base, -2.0f / n_dims); + // Add FP16 capability check if T could be sycl::half + if constexpr (std::is_same_v) { + dpct::has_capability_or_fail(stream->get_device(), { sycl::aspect::fp16 }); + } + // launch kernel + if (freq_factors == nullptr) { + stream->parallel_for(nd_range, [=](sycl::nd_item<3> item_ct1) { + rope_vision(x, dst, ne0, ne1, ne2, s1, s2, n_dims, pos, freq_scale, ext_factor, attn_factor, + corr_dims, theta_scale, freq_factors, sections, item_ct1); + }); + } else { + stream->parallel_for(nd_range, [=](sycl::nd_item<3> item_ct1) { + rope_vision(x, dst, ne0, ne1, ne2, s1, s2, n_dims, pos, freq_scale, ext_factor, attn_factor, + corr_dims, theta_scale, freq_factors, sections, item_ct1); + }); + } +} + void ggml_sycl_op_rope(ggml_backend_sycl_context & ctx, ggml_tensor *dst) { GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32 || dst->src[0]->type == GGML_TYPE_F16); GGML_ASSERT( dst->type == GGML_TYPE_F32 || dst->type == GGML_TYPE_F16); GGML_ASSERT(dst->src[0]->type == dst->type); - - const int64_t ne00 = dst->src[0]->ne[0]; - const int64_t ne01 = dst->src[0]->ne[1]; + const int64_t ne00 = dst->src[0]->ne[0]; // head dims + const int64_t ne01 = dst->src[0]->ne[1]; // num heads + const int64_t ne02 = dst->src[0]->ne[2]; // num heads const int64_t nr = ggml_nrows(dst->src[0]); + const size_t s01 = dst->src[0]->nb[1] / ggml_type_size(dst->src[0]->type); + const size_t s02 = dst->src[0]->nb[2] / ggml_type_size(dst->src[0]->type); + + //const int n_past = ((int32_t *) dst->op_params)[0]; const int n_dims = ((int32_t *) dst->op_params)[1]; const int mode = ((int32_t *) dst->op_params)[2]; //const int n_ctx = ((int32_t *) dst->op_params)[3]; const int n_ctx_orig = ((int32_t *) dst->op_params)[4]; + mrope_sections sections; // RoPE alteration for extended context float freq_base; @@ -222,8 +307,10 @@ void ggml_sycl_op_rope(ggml_backend_sycl_context & ctx, ggml_tensor *dst) { memcpy(&attn_factor, (int32_t *) dst->op_params + 8, sizeof(float)); memcpy(&beta_fast, (int32_t *) dst->op_params + 9, sizeof(float)); memcpy(&beta_slow, (int32_t *) dst->op_params + 10, sizeof(float)); + memcpy(§ions.v, (int32_t *) dst->op_params + 11, sizeof(int)*4); const bool is_neox = mode & GGML_ROPE_TYPE_NEOX; + const bool is_vision = mode == GGML_ROPE_TYPE_VISION; const int32_t * pos = (const int32_t *) dst->src[1]->data; @@ -240,6 +327,7 @@ void ggml_sycl_op_rope(ggml_backend_sycl_context & ctx, ggml_tensor *dst) { // compute if (is_neox) { + GGML_SYCL_DEBUG("%s: neox path\n", __func__); if (dst->src[0]->type == GGML_TYPE_F32) { rope_neox_sycl( (const float *)dst->src[0]->data, (float *)dst->data, ne00, n_dims, nr, pos, freq_scale, ne01, freq_base, ext_factor, @@ -253,7 +341,19 @@ void ggml_sycl_op_rope(ggml_backend_sycl_context & ctx, ggml_tensor *dst) { } else { GGML_ABORT("fatal error"); } + } else if (is_vision) { + GGML_SYCL_DEBUG("%s: vision path\n", __func__); + if (dst->src[0]->type == GGML_TYPE_F16) { + rope_vision_sycl((const sycl::half *)dst->src[0]->data, (sycl::half *)dst->data, ne00, ne01, ne02, s01, s02, n_dims, nr, pos, freq_scale, + freq_base, ext_factor, attn_factor, corr_dims, freq_factors, sections, main_stream); + } else if (dst->src[0]->type == GGML_TYPE_F32) { + rope_vision_sycl((const float *) dst->src[0]->data, (float *)dst->data, ne00, ne01, ne02, s01, s02, n_dims, nr, pos, freq_scale, + freq_base, ext_factor, attn_factor, corr_dims, freq_factors, sections, main_stream); + } else { + GGML_ABORT("Fatal error: Tensor type unsupported!"); + } } else { + GGML_SYCL_DEBUG("%s: norm path\n", __func__); if (dst->src[0]->type == GGML_TYPE_F32) { rope_norm_sycl( (const float *)dst->src[0]->data, (float *)dst->data, ne00, n_dims, nr, pos, freq_scale, ne01, freq_base, ext_factor, diff --git a/gguf-py/gguf/constants.py b/gguf-py/gguf/constants.py index 162070e6e19..8fcde2626aa 100644 --- a/gguf-py/gguf/constants.py +++ b/gguf-py/gguf/constants.py @@ -139,6 +139,8 @@ class Attention: REL_BUCKETS_COUNT = "{arch}.attention.relative_buckets_count" SLIDING_WINDOW = "{arch}.attention.sliding_window" SCALE = "{arch}.attention.scale" + KEY_LENGTH_MLA = "{arch}.attention.key_length_mla" + VALUE_LENGTH_MLA = "{arch}.attention.value_length_mla" class Rope: DIMENSION_COUNT = "{arch}.rope.dimension_count" @@ -382,6 +384,8 @@ class MODEL_TENSOR(IntEnum): ATTN_Q_B = auto() ATTN_KV_A_MQA = auto() ATTN_KV_B = auto() + ATTN_K_B = auto() + ATTN_V_B = auto() ATTN_Q_A_NORM = auto() ATTN_KV_A_NORM = auto() FFN_SUB_NORM = auto() @@ -590,6 +594,8 @@ class MODEL_TENSOR(IntEnum): MODEL_TENSOR.ATTN_Q_B: "blk.{bid}.attn_q_b", MODEL_TENSOR.ATTN_KV_A_MQA: "blk.{bid}.attn_kv_a_mqa", MODEL_TENSOR.ATTN_KV_B: "blk.{bid}.attn_kv_b", + MODEL_TENSOR.ATTN_K_B: "blk.{bid}.attn_k_b", + MODEL_TENSOR.ATTN_V_B: "blk.{bid}.attn_v_b", MODEL_TENSOR.ATTN_Q_A_NORM: "blk.{bid}.attn_q_a_norm", MODEL_TENSOR.ATTN_KV_A_NORM: "blk.{bid}.attn_kv_a_norm", MODEL_TENSOR.ATTN_SUB_NORM: "blk.{bid}.attn_sub_norm", @@ -1517,6 +1523,8 @@ class MODEL_TENSOR(IntEnum): MODEL_TENSOR.ATTN_Q_B, MODEL_TENSOR.ATTN_KV_A_MQA, MODEL_TENSOR.ATTN_KV_B, + MODEL_TENSOR.ATTN_K_B, + MODEL_TENSOR.ATTN_V_B, MODEL_TENSOR.ATTN_Q_A_NORM, MODEL_TENSOR.ATTN_KV_A_NORM, MODEL_TENSOR.ATTN_OUT, diff --git a/gguf-py/gguf/gguf_writer.py b/gguf-py/gguf/gguf_writer.py index 485550aad6d..aef03db1577 100644 --- a/gguf-py/gguf/gguf_writer.py +++ b/gguf-py/gguf/gguf_writer.py @@ -689,6 +689,12 @@ def add_key_length(self, length: int) -> None: def add_value_length(self, length: int) -> None: self.add_uint32(Keys.Attention.VALUE_LENGTH.format(arch=self.arch), length) + def add_key_length_mla(self, length: int) -> None: + self.add_uint32(Keys.Attention.KEY_LENGTH_MLA.format(arch=self.arch), length) + + def add_value_length_mla(self, length: int) -> None: + self.add_uint32(Keys.Attention.VALUE_LENGTH_MLA.format(arch=self.arch), length) + def add_max_alibi_bias(self, bias: float) -> None: self.add_float32(Keys.Attention.MAX_ALIBI_BIAS.format(arch=self.arch), bias) diff --git a/gguf-py/gguf/tensor_mapping.py b/gguf-py/gguf/tensor_mapping.py index 35154e9b5da..0bc75cf513a 100644 --- a/gguf-py/gguf/tensor_mapping.py +++ b/gguf-py/gguf/tensor_mapping.py @@ -677,6 +677,14 @@ class TensorNameMap: "model.layers.{bid}.self_attn.kv_b_proj", # deepseek2 ), + MODEL_TENSOR.ATTN_K_B: ( + "model.layers.{bid}.self_attn.k_b_proj", # deepseek2 + ), + + MODEL_TENSOR.ATTN_V_B: ( + "model.layers.{bid}.self_attn.v_b_proj", # deepseek2 + ), + MODEL_TENSOR.ATTN_Q_A_NORM: ( "model.layers.{bid}.self_attn.q_a_layernorm", # deepseek2 ), diff --git a/src/llama-arch.cpp b/src/llama-arch.cpp index a6fddc7fd2e..62e1480bb58 100644 --- a/src/llama-arch.cpp +++ b/src/llama-arch.cpp @@ -140,6 +140,8 @@ static const std::map LLM_KV_NAMES = { { LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT, "%s.attention.relative_buckets_count" }, { LLM_KV_ATTENTION_SLIDING_WINDOW, "%s.attention.sliding_window" }, { LLM_KV_ATTENTION_SCALE, "%s.attention.scale" }, + { LLM_KV_ATTENTION_KEY_LENGTH_MLA, "%s.attention.key_length_mla" }, + { LLM_KV_ATTENTION_VALUE_LENGTH_MLA, "%s.attention.value_length_mla" }, { LLM_KV_ROPE_DIMENSION_COUNT, "%s.rope.dimension_count" }, { LLM_KV_ROPE_DIMENSION_SECTIONS, "%s.rope.dimension_sections" }, @@ -1103,6 +1105,8 @@ static const std::map> LLM_TENSOR_N { LLM_TENSOR_ATTN_Q_B, "blk.%d.attn_q_b" }, { LLM_TENSOR_ATTN_KV_A_MQA, "blk.%d.attn_kv_a_mqa" }, { LLM_TENSOR_ATTN_KV_B, "blk.%d.attn_kv_b" }, + { LLM_TENSOR_ATTN_K_B, "blk.%d.attn_k_b" }, + { LLM_TENSOR_ATTN_V_B, "blk.%d.attn_v_b" }, { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, { LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" }, @@ -1563,23 +1567,8 @@ static const std::map LLM_TENSOR_INFOS = { {LLM_TENSOR_ATTN_Q_B, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, {LLM_TENSOR_ATTN_KV_A_MQA, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, {LLM_TENSOR_ATTN_KV_B, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, - {LLM_TENSOR_DEC_ATTN_Q, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, - {LLM_TENSOR_DEC_ATTN_K, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, - {LLM_TENSOR_ATTN_Q, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, - {LLM_TENSOR_ATTN_K, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, - {LLM_TENSOR_ATTN_V, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, - {LLM_TENSOR_ATTN_QKV, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, - {LLM_TENSOR_ATTN_OUT, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, - {LLM_TENSOR_FFN_GATE, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, - {LLM_TENSOR_FFN_DOWN, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, - {LLM_TENSOR_FFN_UP, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, - {LLM_TENSOR_FFN_DOWN_SHEXP, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, - {LLM_TENSOR_FFN_GATE_SHEXP, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, - {LLM_TENSOR_FFN_UP_SHEXP, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, - {LLM_TENSOR_ATTN_Q_A, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, - {LLM_TENSOR_ATTN_Q_B, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, - {LLM_TENSOR_ATTN_KV_A_MQA, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, - {LLM_TENSOR_ATTN_KV_B, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, + {LLM_TENSOR_ATTN_K_B, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, + {LLM_TENSOR_ATTN_V_B, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, {LLM_TENSOR_DEC_ATTN_Q, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, {LLM_TENSOR_DEC_ATTN_K, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, {LLM_TENSOR_DEC_ATTN_V, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, diff --git a/src/llama-arch.h b/src/llama-arch.h index 2c2099b3c38..98ca00a1bd0 100644 --- a/src/llama-arch.h +++ b/src/llama-arch.h @@ -144,6 +144,8 @@ enum llm_kv { LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT, LLM_KV_ATTENTION_SLIDING_WINDOW, LLM_KV_ATTENTION_SCALE, + LLM_KV_ATTENTION_KEY_LENGTH_MLA, + LLM_KV_ATTENTION_VALUE_LENGTH_MLA, LLM_KV_ROPE_DIMENSION_COUNT, LLM_KV_ROPE_DIMENSION_SECTIONS, @@ -306,6 +308,8 @@ enum llm_tensor { LLM_TENSOR_ATTN_Q_B, LLM_TENSOR_ATTN_KV_A_MQA, LLM_TENSOR_ATTN_KV_B, + LLM_TENSOR_ATTN_K_B, + LLM_TENSOR_ATTN_V_B, LLM_TENSOR_ATTN_Q_A_NORM, LLM_TENSOR_ATTN_KV_A_NORM, LLM_TENSOR_ATTN_SUB_NORM, diff --git a/src/llama-context.cpp b/src/llama-context.cpp index 4735e98ea04..d3ef1cbdeb6 100644 --- a/src/llama-context.cpp +++ b/src/llama-context.cpp @@ -10,6 +10,7 @@ #include #include #include +#include // // llama_context @@ -473,7 +474,6 @@ ggml_tensor * llama_context::build_rope_shift( const auto & n_ctx_orig = cparams.n_ctx_orig_yarn; const auto & yarn_ext_factor = cparams.yarn_ext_factor; - const auto & yarn_attn_factor = cparams.yarn_attn_factor; const auto & yarn_beta_fast = cparams.yarn_beta_fast; const auto & yarn_beta_slow = cparams.yarn_beta_slow; @@ -482,6 +482,10 @@ ggml_tensor * llama_context::build_rope_shift( const auto & n_rot = hparams.n_rot; const auto & rope_type = hparams.rope_type; + // See llm_build_deepseek2() for why attn_factor has to be scaled for YaRN RoPE to work correctly. + // See https://github.com/ggerganov/llama.cpp/discussions/7416 for detailed explanation. + const float yarn_attn_factor_scaled = model.arch == LLM_ARCH_DEEPSEEK2 ? 1.0f / (1.0f + 0.1f * logf(1.0f / freq_scale)) : cparams.yarn_attn_factor; + ggml_tensor * tmp; if (ggml_is_quantized(cur->type)) { @@ -500,14 +504,14 @@ ggml_tensor * llama_context::build_rope_shift( tmp = ggml_rope_ext_inplace(ctx0, tmp, shift, factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - yarn_ext_factor, yarn_attn_factor, yarn_beta_fast, yarn_beta_slow); + yarn_ext_factor, yarn_attn_factor_scaled, yarn_beta_fast, yarn_beta_slow); tmp = ggml_cpy(ctx0, tmp, cur); } else { // we rotate only the first n_rot dimensions tmp = ggml_rope_ext_inplace(ctx0, cur, shift, factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - yarn_ext_factor, yarn_attn_factor, yarn_beta_fast, yarn_beta_slow); + yarn_ext_factor, yarn_attn_factor_scaled, yarn_beta_fast, yarn_beta_slow); } return tmp; @@ -2274,6 +2278,11 @@ llama_context * llama_init_from_model( params.flash_attn = false; } + if (params.flash_attn && model->arch == LLM_ARCH_DEEPSEEK2) { + LLAMA_LOG_WARN("%s: flash_attn is not compatible with Deepseek2 - forcing off\n", __func__); + params.flash_attn = false; + } + if (ggml_is_quantized(params.type_v) && !params.flash_attn) { LLAMA_LOG_ERROR("%s: V cache quantization requires flash_attn\n", __func__); return nullptr; diff --git a/src/llama-graph.cpp b/src/llama-graph.cpp index cd955d63bc3..5d0222b9810 100644 --- a/src/llama-graph.cpp +++ b/src/llama-graph.cpp @@ -1188,6 +1188,7 @@ ggml_tensor * llm_graph_context::build_attn_mha( ggml_tensor * v, ggml_tensor * kq_b, ggml_tensor * kq_mask, + ggml_tensor * v_mla, bool v_trans, float kq_scale) const { //const int64_t n_embd_k_gqa = hparams.n_embd_k_gqa(il); @@ -1199,7 +1200,8 @@ ggml_tensor * llm_graph_context::build_attn_mha( //const auto & n_embd_head_k = hparams.n_embd_head_k; //const auto & n_embd_head_v = hparams.n_embd_head_v; - const auto n_embd_head_v = v_trans ? v->ne[1] : v->ne[0]; + // note: for MLA with the absorption optimization, the final embedding size will be changed via v_mla + const auto n_embd_head_v = v_mla == nullptr ? v_trans ? v->ne[1] : v->ne[0] : v_mla->ne[1]; const auto n_tokens = q->ne[1]; const auto n_head = q->ne[2]; @@ -1267,6 +1269,11 @@ ggml_tensor * llm_graph_context::build_attn_mha( ggml_tensor * kqv = ggml_mul_mat(ctx0, v, kq); + // for MLA with the absorption optimization, we need to "decompress" from MQA back to MHA + if (v_mla) { + kqv = ggml_mul_mat(ctx0, v_mla, kqv); + } + ggml_tensor * kqv_merged = ggml_permute(ctx0, kqv, 0, 2, 1, 3); cur = ggml_cont_2d(ctx0, kqv_merged, n_embd_head_v*n_head, n_tokens); @@ -1304,6 +1311,7 @@ ggml_tensor * llm_graph_context::build_attn( ggml_tensor * k_cur, ggml_tensor * v_cur, ggml_tensor * kq_b, + ggml_tensor * v_mla, float kq_scale, int il) const { GGML_UNUSED(n_tokens); @@ -1325,7 +1333,7 @@ ggml_tensor * llm_graph_context::build_attn( ggml_tensor * v = ggml_permute(ctx0, v_cur, 0, 2, 1, 3); //cb(k, "v", il); - ggml_tensor * cur = build_attn_mha(gf, q, k, v, kq_b, kq_mask, false, kq_scale); + ggml_tensor * cur = build_attn_mha(gf, q, k, v, kq_b, kq_mask, v_mla, false, kq_scale); cb(cur, "kqv_out", il); @@ -1379,6 +1387,7 @@ ggml_tensor * llm_graph_context::build_attn( ggml_tensor * k_cur, ggml_tensor * v_cur, ggml_tensor * kq_b, + ggml_tensor * v_mla, float kq_scale, int il) const { // these nodes are added to the graph together so that they are not reordered @@ -1464,7 +1473,7 @@ ggml_tensor * llm_graph_context::build_attn( ggml_element_size(kv_self->v_l[il])*n_ctx*n_embd_head_v, 0); - ggml_tensor * cur = build_attn_mha(gf, q, k, v, kq_b, kq_mask, v_trans, kq_scale); + ggml_tensor * cur = build_attn_mha(gf, q, k, v, kq_b, kq_mask, v_mla, v_trans, kq_scale); cb(cur, "kqv_out", il); if (wo) { @@ -1504,6 +1513,7 @@ ggml_tensor * llm_graph_context::build_attn( ggml_tensor * k_cur, ggml_tensor * v_cur, ggml_tensor * kq_b, + ggml_tensor * v_mla, float kq_scale, int il) const { // these nodes are added to the graph together so that they are not reordered @@ -1523,7 +1533,7 @@ ggml_tensor * llm_graph_context::build_attn( ggml_tensor * v = ggml_permute(ctx0, v_cur, 0, 2, 1, 3); //cb(k, "v", il); - ggml_tensor * cur = build_attn_mha(gf, q, k, v, kq_b, kq_mask, false, kq_scale); + ggml_tensor * cur = build_attn_mha(gf, q, k, v, kq_b, kq_mask, v_mla, false, kq_scale); cb(cur, "kqv_out", il); @@ -1692,4 +1702,3 @@ void llm_graph_context::build_pooling( ggml_build_forward_expand(gf, cur); } - diff --git a/src/llama-graph.h b/src/llama-graph.h index 5b6618f9e55..d192dc14957 100644 --- a/src/llama-graph.h +++ b/src/llama-graph.h @@ -505,11 +505,12 @@ struct llm_graph_context { ggml_tensor * build_attn_mha( ggml_cgraph * gf, - ggml_tensor * q, // [n_embd_head_q, n_tokens, n_head_q] - ggml_tensor * k, // [n_embd_head_k, n_tokens, n_head_k] - ggml_tensor * v, // [n_embd_head_v, n_tokens, n_head_v] (v_trans == false) + ggml_tensor * q, // [n_embd_head_q, n_tokens, n_head_q] + ggml_tensor * k, // [n_embd_head_k, n_tokens, n_head_k] + ggml_tensor * v, // [n_embd_head_v, n_tokens, n_head_v] (v_trans == false) ggml_tensor * kq_b, ggml_tensor * kq_mask, + ggml_tensor * v_mla, // [n_embd_head_v_mla, n_embd_head_v, n_head_v] bool v_trans, float kq_scale) const; @@ -524,6 +525,7 @@ struct llm_graph_context { ggml_tensor * k_cur, // [n_embd_head_k, n_head_k, n_tokens] ggml_tensor * v_cur, // [n_embd_head_v, n_head_v, n_tokens] ggml_tensor * kq_b, + ggml_tensor * v_mla, // [n_embd_head_v_mla, n_embd_head_v, n_head_v] float kq_scale, int il) const; @@ -538,6 +540,7 @@ struct llm_graph_context { ggml_tensor * k_cur, // [n_embd_head_k, n_head_k, n_tokens] ggml_tensor * v_cur, // [n_embd_head_v, n_head_v, n_tokens] ggml_tensor * kq_b, + ggml_tensor * v_mla, // [n_embd_head_v_mla, n_embd_head_v, n_head_v] float kq_scale, int il) const; @@ -552,6 +555,7 @@ struct llm_graph_context { ggml_tensor * k_cur, // [n_embd_head_k, n_head_k, n_tokens] ggml_tensor * v_cur, // [n_embd_head_v, n_head_v, n_tokens] ggml_tensor * kq_b, + ggml_tensor * v_mla, // [n_embd_head_v_mla, n_embd_head_v, n_head_v] float kq_scale, int il) const; diff --git a/src/llama-hparams.h b/src/llama-hparams.h index 4e0b57190a3..80fcd65df0d 100644 --- a/src/llama-hparams.h +++ b/src/llama-hparams.h @@ -43,6 +43,10 @@ struct llama_hparams { uint32_t n_expert_used = 0; uint32_t n_rel_attn_bkts = 0; + // note: deepseek2 using MLA converts into MQA with larger heads, then decompresses to MHA + uint32_t n_embd_head_k_mla = 0; + uint32_t n_embd_head_v_mla = 0; + // for WavTokenizer struct llama_hparams_posnet posnet; struct llama_hparams_convnext convnext; diff --git a/src/llama-kv-cache.cpp b/src/llama-kv-cache.cpp index dbf5f1187d9..7c9d46d8119 100644 --- a/src/llama-kv-cache.cpp +++ b/src/llama-kv-cache.cpp @@ -27,7 +27,7 @@ bool llama_kv_cache_unified::init( recurrent = llama_model_is_recurrent(&model); v_trans = !recurrent && !cparams.flash_attn; - can_shift = !recurrent && model.arch != LLM_ARCH_DEEPSEEK2; // not supported due to MLA + can_shift = !recurrent; LLAMA_LOG_INFO("%s: kv_size = %d, offload = %d, type_k = '%s', type_v = '%s', n_layer = %d, can_shift = %d\n", __func__, kv_size, offload, ggml_type_name(type_k), ggml_type_name(type_v), n_layer, can_shift); diff --git a/src/llama-model.cpp b/src/llama-model.cpp index b74dd72cfbf..248c61748ea 100644 --- a/src/llama-model.cpp +++ b/src/llama-model.cpp @@ -1156,6 +1156,8 @@ void llama_model::load_hparams(llama_model_loader & ml) { ml.get_key(LLM_KV_ATTENTION_Q_LORA_RANK, hparams.n_lora_q); } ml.get_key(LLM_KV_ATTENTION_KV_LORA_RANK, hparams.n_lora_kv); + ml.get_key(LLM_KV_ATTENTION_KEY_LENGTH_MLA, hparams.n_embd_head_k_mla, false); + ml.get_key(LLM_KV_ATTENTION_VALUE_LENGTH_MLA, hparams.n_embd_head_v_mla, false); ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp); ml.get_key(LLM_KV_EXPERT_SHARED_COUNT, hparams.n_expert_shared); ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE, hparams.expert_weights_scale); @@ -3205,8 +3207,14 @@ bool llama_model::load_tensors(llama_model_loader & ml) { { const bool is_lite = (hparams.n_layer == 27); + const bool is_mla = (hparams.n_embd_head_k_mla != 0 && hparams.n_embd_head_v_mla != 0); + + // note: these are the actual head sizes you get when treating as MHA or after "decompression" using wv_b for MLA + const int64_t n_embd_head_k_mla = is_mla ? hparams.n_embd_head_k_mla : hparams.n_embd_head_k; + const int64_t n_embd_head_v_mla = is_mla ? hparams.n_embd_head_v_mla : hparams.n_embd_head_v; + const int64_t n_embd_head_qk_rope = hparams.n_rot; - const int64_t n_embd_head_qk_nope = hparams.n_embd_head_k - hparams.n_rot; + const int64_t n_embd_head_qk_nope = n_embd_head_k_mla - n_embd_head_qk_rope; const int64_t q_lora_rank = hparams.n_lora_q; const int64_t kv_lora_rank = hparams.n_lora_kv; @@ -3232,14 +3240,22 @@ bool llama_model::load_tensors(llama_model_loader & ml) { if (!is_lite) { layer.wq_a = create_tensor(tn(LLM_TENSOR_ATTN_Q_A, "weight", i), {n_embd, q_lora_rank}, 0); - layer.wq_b = create_tensor(tn(LLM_TENSOR_ATTN_Q_B, "weight", i), {q_lora_rank, n_head * n_embd_head_k}, 0); + layer.wq_b = create_tensor(tn(LLM_TENSOR_ATTN_Q_B, "weight", i), {q_lora_rank, n_head * n_embd_head_k_mla}, 0); } else { - layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd_k_gqa}, 0); + layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_head * n_embd_head_k_mla}, 0); } - layer.wkv_a_mqa = create_tensor(tn(LLM_TENSOR_ATTN_KV_A_MQA, "weight", i), {n_embd, kv_lora_rank + (n_embd_head_qk_rope)}, 0); - layer.wkv_b = create_tensor(tn(LLM_TENSOR_ATTN_KV_B, "weight", i), {kv_lora_rank, n_head * (n_embd_head_qk_nope + n_embd_head_v)}, 0); - layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), { n_head * ( n_embd_head_v), n_embd}, 0); + layer.wkv_a_mqa = create_tensor(tn(LLM_TENSOR_ATTN_KV_A_MQA, "weight", i), {n_embd, kv_lora_rank + n_embd_head_qk_rope}, 0); + + // note: only old legacy GGUF files will have the unsplit wkv_b tensor in + if (is_mla) { + layer.wk_b = create_tensor(tn(LLM_TENSOR_ATTN_K_B, "weight", i), {n_embd_head_qk_nope, kv_lora_rank, n_head}, 0); + layer.wv_b = create_tensor(tn(LLM_TENSOR_ATTN_V_B, "weight", i), {kv_lora_rank, n_embd_head_v_mla, n_head}, 0); + } else { + layer.wkv_b = create_tensor(tn(LLM_TENSOR_ATTN_KV_B, "weight", i), {kv_lora_rank, n_head * (n_embd_head_qk_nope + n_embd_head_v_mla)}, 0); + } + + layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_head * n_embd_head_v_mla, n_embd}, 0); layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0); @@ -4290,6 +4306,8 @@ void llama_model::print_info() const { LLAMA_LOG_INFO("%s: n_layer_dense_lead = %d\n", __func__, hparams.n_layer_dense_lead); LLAMA_LOG_INFO("%s: n_lora_q = %d\n", __func__, hparams.n_lora_q); LLAMA_LOG_INFO("%s: n_lora_kv = %d\n", __func__, hparams.n_lora_kv); + LLAMA_LOG_INFO("%s: n_embd_head_k_mla = %d\n", __func__, hparams.n_embd_head_k_mla); + LLAMA_LOG_INFO("%s: n_embd_head_v_mla = %d\n", __func__, hparams.n_embd_head_v_mla); LLAMA_LOG_INFO("%s: n_ff_exp = %d\n", __func__, hparams.n_ff_exp); LLAMA_LOG_INFO("%s: n_expert_shared = %d\n", __func__, hparams.n_expert_shared); LLAMA_LOG_INFO("%s: expert_weights_scale = %.1f\n", __func__, hparams.expert_weights_scale); @@ -4496,7 +4514,7 @@ struct llm_build_llama : public llm_graph_context { cur = build_attn(inp_attn, gf, model.layers[il].wo, model.layers[il].bo, - Qcur, Kcur, Vcur, nullptr, kq_scale, il); + Qcur, Kcur, Vcur, nullptr, nullptr, kq_scale, il); cb(cur, "attn_out", il); } @@ -4709,7 +4727,7 @@ struct llm_build_deci : public llm_graph_context { cur = build_attn(inp_attn, gf, model.layers[il].wo, model.layers[il].bo, - Qcur, Kcur, Vcur, nullptr, kq_scale, il); + Qcur, Kcur, Vcur, nullptr, nullptr, kq_scale, il); } if (il == n_layer - 1) { @@ -4851,7 +4869,7 @@ struct llm_build_baichuan : public llm_graph_context { cur = build_attn(inp_attn, gf, model.layers[il].wo, NULL, - Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } if (il == n_layer - 1) { @@ -4966,7 +4984,7 @@ struct llm_build_xverse : public llm_graph_context { cur = build_attn(inp_attn, gf, model.layers[il].wo, NULL, - Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } if (il == n_layer - 1) { @@ -5091,7 +5109,7 @@ struct llm_build_falcon : public llm_graph_context { cur = build_attn(inp_attn, gf, model.layers[il].wo, NULL, - Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } if (il == n_layer - 1) { @@ -5221,7 +5239,7 @@ struct llm_build_grok : public llm_graph_context { cur = build_attn(inp_attn, gf, model.layers[il].wo, model.layers[il].bo, - Qcur, Kcur, Vcur, nullptr, 1.0f, il); + Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f, il); } if (il == n_layer - 1) { @@ -5372,7 +5390,7 @@ struct llm_build_dbrx : public llm_graph_context { cur = build_attn(inp_attn, gf, model.layers[il].wo, NULL, - Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } if (il == n_layer - 1) { @@ -5486,7 +5504,7 @@ struct llm_build_starcoder : public llm_graph_context { cur = build_attn(inp_attn, gf, model.layers[il].wo, model.layers[il].bo, - Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } if (il == n_layer - 1) { @@ -5585,7 +5603,7 @@ struct llm_build_refact : public llm_graph_context { cur = build_attn(inp_attn, gf, model.layers[il].wo, NULL, - Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } if (il == n_layer - 1) { @@ -5739,7 +5757,7 @@ struct llm_build_bert : public llm_graph_context { cur = build_attn(inp_attn, gf, model.layers[il].wo, model.layers[il].bo, - Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); cb(cur, "kqv_out", il); if (il == n_layer - 1 && pooling_type == LLAMA_POOLING_TYPE_NONE) { @@ -5856,7 +5874,7 @@ struct llm_build_bloom : public llm_graph_context { cur = build_attn(inp_attn, gf, model.layers[il].wo, model.layers[il].bo, - Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } if (il == n_layer - 1) { @@ -5997,7 +6015,7 @@ struct llm_build_mpt : public llm_graph_context { cur = build_attn(inp_attn, gf, model.layers[il].wo, model.layers[il].bo, - Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } if (il == n_layer - 1) { @@ -6143,7 +6161,7 @@ struct llm_build_stablelm : public llm_graph_context { cur = build_attn(inp_attn, gf, model.layers[il].wo, NULL, - Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } if (il == n_layer - 1) { @@ -6266,7 +6284,7 @@ struct llm_build_qwen : public llm_graph_context { cur = build_attn(inp_attn, gf, model.layers[il].wo, NULL, - Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } if (il == n_layer - 1) { @@ -6386,7 +6404,7 @@ struct llm_build_qwen2 : public llm_graph_context { cur = build_attn(inp_attn, gf, model.layers[il].wo, model.layers[il].bo, - Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } if (il == n_layer - 1) { @@ -6507,7 +6525,7 @@ struct llm_build_qwen2vl : public llm_graph_context { cur = build_attn(inp_attn, gf, model.layers[il].wo, model.layers[il].bo, - Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } if (il == n_layer - 1) { @@ -6634,7 +6652,7 @@ struct llm_build_qwen2moe : public llm_graph_context { cur = build_attn(inp_attn, gf, model.layers[il].wo, model.layers[il].bo, - Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } if (il == n_layer - 1) { @@ -6787,7 +6805,7 @@ struct llm_build_qwen3 : public llm_graph_context { cur = build_attn(inp_attn, gf, model.layers[il].wo, model.layers[il].bo, - Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } if (il == n_layer - 1) { @@ -6908,7 +6926,7 @@ struct llm_build_qwen3moe : public llm_graph_context { cur = build_attn(inp_attn, gf, model.layers[il].wo, model.layers[il].bo, - Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } if (il == n_layer - 1) { @@ -7048,7 +7066,7 @@ struct llm_build_phi2 : public llm_graph_context { cur = build_attn(inp_attn, gf, model.layers[il].wo, model.layers[il].bo, - Qcur, Kcur, Vcur, nullptr, 1.0f, il); + Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f, il); } if (il == n_layer - 1) { @@ -7177,7 +7195,7 @@ struct llm_build_phi3 : public llm_graph_context { cur = build_attn(inp_attn, gf, model.layers[il].wo, model.layers[il].bo, - Qcur, Kcur, Vcur, nullptr, 1.0f, il); + Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f, il); } if (il == n_layer - 1) { @@ -7312,7 +7330,7 @@ struct llm_build_plamo : public llm_graph_context { cur = build_attn(inp_attn, gf, model.layers[il].wo, NULL, - Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } ggml_tensor * sa_out = cur; @@ -7419,7 +7437,7 @@ struct llm_build_gpt2 : public llm_graph_context { cur = build_attn(inp_attn, gf, model.layers[il].wo, model.layers[il].bo, - Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } if (il == n_layer - 1) { @@ -7535,7 +7553,7 @@ struct llm_build_codeshell : public llm_graph_context { cur = build_attn(inp_attn, gf, model.layers[il].wo, model.layers[il].bo, - Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } if (il == n_layer - 1) { @@ -7664,7 +7682,7 @@ struct llm_build_orion : public llm_graph_context { cur = build_attn(inp_attn, gf, model.layers[il].wo, NULL, - Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } if (il == n_layer - 1) { @@ -7791,7 +7809,7 @@ struct llm_build_internlm2 : public llm_graph_context { cur = build_attn(inp_attn, gf, model.layers[il].wo, model.layers[il].bo, - Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } if (il == n_layer - 1) { @@ -7988,7 +8006,7 @@ struct llm_build_minicpm3 : public llm_graph_context { cur = build_attn(inp_attn, gf, model.layers[il].wo, NULL, - q_states, k_states, v_states, nullptr, kq_scale, il); + q_states, k_states, v_states, nullptr, nullptr, kq_scale, il); } if (il == n_layer - 1) { @@ -8118,7 +8136,7 @@ struct llm_build_gemma : public llm_graph_context { cur = build_attn(inp_attn, gf, model.layers[il].wo, NULL, - Qcur, Kcur, Vcur, nullptr, 1.0f, il); + Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f, il); } if (il == n_layer - 1) { @@ -8240,7 +8258,7 @@ struct llm_build_gemma2 : public llm_graph_context { cur = build_attn(inp_attn, gf, model.layers[il].wo, NULL, - Qcur, Kcur, Vcur, nullptr, 1.0f, il); + Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f, il); } cur = build_norm(cur, @@ -8381,7 +8399,7 @@ struct llm_build_gemma3 : public llm_graph_context { cur = build_attn(inp_attn, gf, model.layers[il].wo, NULL, - Qcur, Kcur, Vcur, nullptr, hparams.f_attention_scale, il); + Qcur, Kcur, Vcur, nullptr, nullptr, hparams.f_attention_scale, il); } cur = build_norm(cur, @@ -8521,7 +8539,7 @@ struct llm_build_starcoder2 : public llm_graph_context { cur = build_attn(inp_attn, gf, model.layers[il].wo, model.layers[il].bo, - Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } if (il == n_layer - 1) { @@ -8856,7 +8874,7 @@ struct llm_build_command_r : public llm_graph_context { cur = build_attn(inp_attn, gf, model.layers[il].wo, model.layers[il].bo, - Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } if (il == n_layer - 1) { @@ -8991,7 +9009,7 @@ struct llm_build_cohere2 : public llm_graph_context { cur = build_attn(inp_attn, gf, model.layers[il].wo, model.layers[il].bo, - Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } if (il == n_layer - 1) { @@ -9122,7 +9140,7 @@ struct llm_build_olmo : public llm_graph_context { cur = build_attn(inp_attn, gf, model.layers[il].wo, nullptr, - Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } if (il == n_layer - 1) { @@ -9242,7 +9260,7 @@ struct llm_build_olmo2 : public llm_graph_context { cur = build_attn(inp_attn, gf, model.layers[il].wo, NULL, - Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } cur = build_norm(cur, @@ -9375,7 +9393,7 @@ struct llm_build_olmoe : public llm_graph_context { cur = build_attn(inp_attn, gf, model.layers[il].wo, NULL, - Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } if (il == n_layer - 1) { @@ -9508,7 +9526,7 @@ struct llm_build_openelm : public llm_graph_context { cur = build_attn(inp_attn, gf, model.layers[il].wo, NULL, - Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } if (il == n_layer - 1) { @@ -9622,7 +9640,7 @@ struct llm_build_gptneox : public llm_graph_context { cur = build_attn(inp_attn, gf, model.layers[il].wo, model.layers[il].bo, - Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } if (il == n_layer - 1) { @@ -9772,7 +9790,7 @@ struct llm_build_arctic : public llm_graph_context { cur = build_attn(inp_attn, gf, model.layers[il].wo, NULL, - Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } if (il == n_layer - 1) { @@ -9927,7 +9945,7 @@ struct llm_build_deepseek : public llm_graph_context { cur = build_attn(inp_attn, gf, model.layers[il].wo, model.layers[il].bo, - Qcur, Kcur, Vcur, nullptr, kq_scale, il); + Qcur, Kcur, Vcur, nullptr, nullptr, kq_scale, il); } if (il == n_layer - 1) { @@ -10017,16 +10035,23 @@ struct llm_build_deepseek2 : public llm_graph_context { llm_build_deepseek2(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) { bool is_lite = (hparams.n_layer == 27); + const bool is_mla = (hparams.n_embd_head_k_mla != 0 && hparams.n_embd_head_v_mla != 0); + + // note: these are the actual head sizes you get when treating as MHA or after "decompression" using wv_b for MLA + const int64_t n_embd_head_k = is_mla ? hparams.n_embd_head_k_mla : hparams.n_embd_head_k; + const int64_t n_embd_head_v = is_mla ? hparams.n_embd_head_v_mla : hparams.n_embd_head_v; + + const int64_t n_embd_head_qk_rope = hparams.n_rot; + const int64_t n_embd_head_qk_nope = n_embd_head_k - n_embd_head_qk_rope; + + const uint32_t kv_lora_rank = hparams.n_lora_kv; + // We have to pre-scale kq_scale and attn_factor to make the YaRN RoPE work correctly. // See https://github.com/ggerganov/llama.cpp/discussions/7416 for detailed explanation. const float mscale = attn_factor * (1.0f + hparams.rope_yarn_log_mul * logf(1.0f / freq_scale)); - const float kq_scale = 1.0f*mscale*mscale/sqrtf(float(hparams.n_embd_head_k)); + const float kq_scale = 1.0f*mscale*mscale/sqrtf(float(n_embd_head_k)); const float attn_factor_scaled = 1.0f / (1.0f + 0.1f * logf(1.0f / freq_scale)); - const uint32_t n_embd_head_qk_rope = hparams.n_rot; - const uint32_t n_embd_head_qk_nope = hparams.n_embd_head_k - hparams.n_rot; - const uint32_t kv_lora_rank = hparams.n_lora_kv; - ggml_tensor * cur; ggml_tensor * inpL; @@ -10051,16 +10076,14 @@ struct llm_build_deepseek2 : public llm_graph_context { { ggml_tensor * q = NULL; if (!is_lite) { - // {n_embd, q_lora_rank} * {n_embd, n_tokens} -> {q_lora_rank, n_tokens} q = ggml_mul_mat(ctx0, model.layers[il].wq_a, cur); cb(q, "q", il); q = build_norm(q, - model.layers[il].attn_q_a_norm, NULL, + model.layers[il].attn_q_a_norm, nullptr, LLM_NORM_RMS, il); cb(q, "q", il); - // {q_lora_rank, n_head * hparams.n_embd_head_k} * {q_lora_rank, n_tokens} -> {n_head * hparams.n_embd_head_k, n_tokens} q = ggml_mul_mat(ctx0, model.layers[il].wq_b, q); cb(q, "q", il); } else { @@ -10068,96 +10091,125 @@ struct llm_build_deepseek2 : public llm_graph_context { cb(q, "q", il); } - // split into {n_head * n_embd_head_qk_nope, n_tokens} - ggml_tensor * q_nope = ggml_view_3d(ctx0, q, n_embd_head_qk_nope, n_head, n_tokens, - ggml_row_size(q->type, hparams.n_embd_head_k), - ggml_row_size(q->type, hparams.n_embd_head_k * n_head), + // split into {n_embd_head_qk_nope, n_head, n_tokens} + ggml_tensor * q_nope = ggml_view_3d(ctx0, q, + n_embd_head_qk_nope, n_head, n_tokens, + ggml_row_size(q->type, n_embd_head_k), + ggml_row_size(q->type, n_embd_head_k) * n_head, 0); cb(q_nope, "q_nope", il); - // and {n_head * n_embd_head_qk_rope, n_tokens} - ggml_tensor * q_pe = ggml_view_3d(ctx0, q, n_embd_head_qk_rope, n_head, n_tokens, - ggml_row_size(q->type, hparams.n_embd_head_k), - ggml_row_size(q->type, hparams.n_embd_head_k * n_head), + // and {n_embd_head_qk_rope, n_head, n_tokens} + ggml_tensor * q_pe = ggml_view_3d(ctx0, q, + n_embd_head_qk_rope, n_head, n_tokens, + ggml_row_size(q->type, n_embd_head_k), + ggml_row_size(q->type, n_embd_head_k) * n_head, ggml_row_size(q->type, n_embd_head_qk_nope)); cb(q_pe, "q_pe", il); - // {n_embd, kv_lora_rank + n_embd_head_qk_rope} * {n_embd, n_tokens} -> {kv_lora_rank + n_embd_head_qk_rope, n_tokens} - ggml_tensor * kv_pe_compresseed = ggml_mul_mat(ctx0, model.layers[il].wkv_a_mqa, cur); - cb(kv_pe_compresseed, "kv_pe_compresseed", il); + ggml_tensor * kv_cmpr_pe = ggml_mul_mat(ctx0, model.layers[il].wkv_a_mqa, cur); + cb(kv_cmpr_pe, "kv_cmpr_pe", il); // split into {kv_lora_rank, n_tokens} - ggml_tensor * kv_compressed = ggml_view_2d(ctx0, kv_pe_compresseed, kv_lora_rank, n_tokens, - kv_pe_compresseed->nb[1], + ggml_tensor * kv_cmpr = ggml_view_2d(ctx0, kv_cmpr_pe, + kv_lora_rank, n_tokens, + ggml_row_size(kv_cmpr_pe->type, kv_lora_rank + n_embd_head_qk_rope), 0); - cb(kv_compressed, "kv_compressed", il); + cb(kv_cmpr, "kv_cmpr", il); + + // and {n_embd_head_qk_rope, 1, n_tokens} + ggml_tensor * k_pe = ggml_view_3d(ctx0, kv_cmpr_pe, + n_embd_head_qk_rope, 1, n_tokens, + ggml_row_size(kv_cmpr_pe->type, kv_lora_rank + n_embd_head_qk_rope), + ggml_row_size(kv_cmpr_pe->type, kv_lora_rank + n_embd_head_qk_rope), + ggml_row_size(kv_cmpr_pe->type, kv_lora_rank)); + cb(k_pe, "k_pe", il); - // and {n_embd_head_qk_rope, n_tokens} - ggml_tensor * k_pe = ggml_view_3d(ctx0, kv_pe_compresseed, n_embd_head_qk_rope, 1, n_tokens, - kv_pe_compresseed->nb[1], - kv_pe_compresseed->nb[1], - ggml_row_size(kv_pe_compresseed->type, kv_lora_rank)); + q_pe = ggml_rope_ext(ctx0, q_pe, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor_scaled, beta_fast, beta_slow + ); + cb(q_pe, "q_pe", il); + + k_pe = ggml_rope_ext(ctx0, k_pe, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor_scaled, beta_fast, beta_slow + ); cb(k_pe, "k_pe", il); - // TODO: the CUDA backend used to not support non-cont. (RMS) norm, investigate removing ggml_cont - kv_compressed = ggml_cont(ctx0, kv_compressed); - kv_compressed = build_norm(kv_compressed, - model.layers[il].attn_kv_a_norm, NULL, + kv_cmpr = build_norm(kv_cmpr, + model.layers[il].attn_kv_a_norm, nullptr, LLM_NORM_RMS, il); - cb(kv_compressed, "kv_compressed", il); + cb(kv_cmpr, "kv_cmpr", il); - // {kv_lora_rank, n_head * (n_embd_head_qk_nope + n_embd_head_v)} * {kv_lora_rank, n_tokens} -> {n_head * (n_embd_head_qk_nope + n_embd_head_v), n_tokens} - ggml_tensor * kv = ggml_mul_mat(ctx0, model.layers[il].wkv_b, kv_compressed); - cb(kv, "kv", il); + if (is_mla) { + // {n_embd_head_qk_nope, n_tokens, n_head} + q_nope = ggml_permute(ctx0, q_nope, 0, 2, 1, 3); + cb(q_nope, "q_nope_perm", il); - // split into {n_head * n_embd_head_qk_nope, n_tokens} - ggml_tensor * k_nope = ggml_view_3d(ctx0, kv, n_embd_head_qk_nope, n_head, n_tokens, - ggml_row_size(kv->type, n_embd_head_qk_nope + hparams.n_embd_head_v), - ggml_row_size(kv->type, n_head * (n_embd_head_qk_nope + hparams.n_embd_head_v)), - 0); - cb(k_nope, "k_nope", il); + // {n_embd_head_qk_nope, kv_lora_rank, n_head} x {n_embd_head_qk_nope, n_tokens, n_head} + ggml_tensor * q_nope_absorbed = ggml_mul_mat(ctx0, model.layers[il].wk_b, q_nope); + cb(q_nope_absorbed, "q_nope_absorbed", il); - // and {n_head * n_embd_head_v, n_tokens} - ggml_tensor * v_states = ggml_view_3d(ctx0, kv, hparams.n_embd_head_v, n_head, n_tokens, - ggml_row_size(kv->type, (n_embd_head_qk_nope + hparams.n_embd_head_v)), - ggml_row_size(kv->type, (n_embd_head_qk_nope + hparams.n_embd_head_v)*n_head), - ggml_row_size(kv->type, (n_embd_head_qk_nope))); - cb(v_states, "v_states", il); + // {kv_lora_rank, n_head, n_tokens} + q_nope_absorbed = ggml_permute(ctx0, q_nope_absorbed, 0, 2, 1, 3); + cb(q_nope_absorbed, "q_nope_absorbed_perm", il); - v_states = ggml_cont(ctx0, v_states); - cb(v_states, "v_states", il); + // {n_embd_head_qk_rope + kv_lora_rank, n_head, n_tokens} + // note: rope must go first for in-place context shifting in build_rope_shift() + ggml_tensor * Qcur = ggml_concat(ctx0, q_pe, q_nope_absorbed, 0); + cb(Qcur, "Qcur", il); - v_states = ggml_view_2d(ctx0, v_states, hparams.n_embd_head_v * n_head, n_tokens, - ggml_row_size(kv->type, hparams.n_embd_head_v * n_head), - 0); - cb(v_states, "v_states", il); + kv_cmpr = ggml_reshape_3d(ctx0, kv_cmpr, kv_lora_rank, 1, n_tokens); + cb(kv_cmpr, "kv_cmpr_reshape", il); - q_pe = ggml_cont(ctx0, q_pe); // TODO: the CUDA backend used to not support non-cont. RoPE, investigate removing this - q_pe = ggml_rope_ext( - ctx0, q_pe, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor_scaled, beta_fast, beta_slow - ); - cb(q_pe, "q_pe", il); + // {n_embd_head_qk_rope + kv_lora_rank, 1, n_tokens} + ggml_tensor * Kcur = ggml_concat(ctx0, k_pe, kv_cmpr, 0); + cb(Kcur, "Kcur", il); - // shared RoPE key - k_pe = ggml_cont(ctx0, k_pe); // TODO: the CUDA backend used to not support non-cont. RoPE, investigate removing this - k_pe = ggml_rope_ext( - ctx0, k_pe, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor_scaled, beta_fast, beta_slow - ); - cb(k_pe, "k_pe", il); + // {kv_lora_rank, 1, n_tokens} + ggml_tensor * Vcur = kv_cmpr; + cb(Vcur, "Vcur", il); - ggml_tensor * q_states = ggml_concat(ctx0, q_nope, q_pe, 0); - cb(q_states, "q_states", il); + // note: MLA with the absorption optimzation converts into MQA (ie: GQA with 1 group) + cur = build_attn(inp_attn, gf, + model.layers[il].wo, NULL, + Qcur, Kcur, Vcur, nullptr, model.layers[il].wv_b, kq_scale, il); + } else { + ggml_tensor * kv = ggml_mul_mat(ctx0, model.layers[il].wkv_b, kv_cmpr); + cb(kv, "kv", il); + + // split into {n_embd_head_qk_nope, n_head, n_tokens} + ggml_tensor * k_nope = ggml_view_3d(ctx0, kv, + n_embd_head_qk_nope, n_head, n_tokens, + ggml_row_size(kv->type, n_embd_head_qk_nope + n_embd_head_v), + ggml_row_size(kv->type, n_embd_head_qk_nope + n_embd_head_v) * n_head, + 0); + cb(k_nope, "k_nope_view", il); - ggml_tensor * k_states = ggml_concat(ctx0, k_nope, ggml_repeat(ctx0, k_pe, q_pe), 0); - cb(k_states, "k_states", il); + // and {n_embd_head_v, n_head, n_tokens} + ggml_tensor * Vcur = ggml_view_3d(ctx0, kv, + n_embd_head_v, n_head, n_tokens, + ggml_row_size(kv->type, n_embd_head_qk_nope + n_embd_head_v), + ggml_row_size(kv->type, n_embd_head_qk_nope + n_embd_head_v) * n_head, + ggml_row_size(kv->type, n_embd_head_qk_nope)); + cb(Vcur, "Vcur_view", il); - cur = build_attn(inp_attn, gf, - model.layers[il].wo, NULL, - q_states, k_states, v_states, nullptr, kq_scale, il); + Vcur = ggml_cont(ctx0, Vcur); + cb(Vcur, "Vcur_cont", il); + + // note: rope must go first for in-place context shifting in build_rope_shift() + ggml_tensor * Qcur = ggml_concat(ctx0, q_pe, q_nope, 0); + cb(Qcur, "Qcur", il); + + ggml_tensor * Kcur = ggml_concat(ctx0, ggml_repeat(ctx0, k_pe, q_pe), k_nope, 0); + cb(Kcur, "Kcur", il); + + // note: MLA without the absorption optimization converts into MHA (ie: GQA with full n_head groups) + cur = build_attn(inp_attn, gf, + model.layers[il].wo, NULL, + Qcur, Kcur, Vcur, nullptr, nullptr, kq_scale, il); + } } if (il == n_layer - 1) { @@ -10323,7 +10375,7 @@ struct llm_build_bitnet : public llm_graph_context { cur = build_attn(inp_attn, gf, NULL, NULL, - Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); cur = build_norm(cur, model.layers[il].attn_sub_norm, NULL, @@ -10446,7 +10498,7 @@ struct llm_build_t5_enc : public llm_graph_context { cur = build_attn(inp_attn, gf, model.layers[il].wo_enc, nullptr, - Qcur, Kcur, Vcur, kq_b, 1.0f, il); + Qcur, Kcur, Vcur, kq_b, nullptr, 1.0f, il); cb(cur, "kqv_out", il); } @@ -10552,7 +10604,7 @@ struct llm_build_t5_dec : public llm_graph_context { cur = build_attn(inp_attn_self, gf, model.layers[il].wo, model.layers[il].bo, - Qcur, Kcur, Vcur, kq_b, 1.0f, il); + Qcur, Kcur, Vcur, kq_b, nullptr, 1.0f, il); cb(cur, "kqv_out", il); } @@ -10584,7 +10636,7 @@ struct llm_build_t5_dec : public llm_graph_context { cur = build_attn(inp_attn_cross, gf, model.layers[il].wo_cross, nullptr, - Qcur, Kcur, Vcur, nullptr, 1.0f, il); + Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f, il); cb(cur, "kqv_out", il); //ggml_tensor * q = ggml_permute(ctx0, Qcur, 0, 2, 1, 3); @@ -10717,7 +10769,7 @@ struct llm_build_jais : public llm_graph_context { cur = build_attn(inp_attn, gf, model.layers[il].wo, model.layers[il].bo, - Qcur, Kcur, Vcur, nullptr, 1.0f/float(n_embd_head), il); + Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/float(n_embd_head), il); } if (il == n_layer - 1) { @@ -10849,7 +10901,7 @@ struct llm_build_chatglm : public llm_graph_context { cur = build_attn(inp_attn, gf, model.layers[il].wo, NULL, - Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } if (il == n_layer - 1) { @@ -10982,7 +11034,7 @@ struct llm_build_glm4 : public llm_graph_context { cur = build_attn(inp_attn, gf, model.layers[il].wo, NULL, - Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } if (il == n_layer - 1) { @@ -11126,7 +11178,7 @@ struct llm_build_nemotron : public llm_graph_context { cur = build_attn(inp_attn, gf, model.layers[il].wo, model.layers[il].bo, - Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } if (il == n_layer - 1) { @@ -11257,7 +11309,7 @@ struct llm_build_exaone : public llm_graph_context { cur = build_attn(inp_attn, gf, model.layers[il].wo, model.layers[il].bo, - Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); } if (il == n_layer - 1) { @@ -12159,7 +12211,7 @@ struct llm_build_chameleon : public llm_graph_context { cur = build_attn(inp_attn, gf, model.layers[il].wo, nullptr, - Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); if (hparams.swin_norm) { cur = build_norm(cur, @@ -12515,7 +12567,7 @@ struct llm_build_plm : public llm_graph_context { cur = build_attn(inp_attn, gf, model.layers[il].wo, NULL, - q_states, k_states, v_states, nullptr, kq_scale, il); + q_states, k_states, v_states, nullptr, nullptr, kq_scale, il); } if (il == n_layer - 1) { @@ -12638,7 +12690,7 @@ struct llm_build_bailingmoe : public llm_graph_context { cur = build_attn(inp_attn, gf, model.layers[il].wo, model.layers[il].bo, - Qcur, Kcur, Vcur, nullptr, 1.0f/sqrtf(float(n_rot)), il); + Qcur, Kcur, Vcur, nullptr, nullptr, 1.0f/sqrtf(float(n_rot)), il); } if (il == n_layer - 1) { diff --git a/src/llama-model.h b/src/llama-model.h index 0f18dac1673..fd82d106ccd 100644 --- a/src/llama-model.h +++ b/src/llama-model.h @@ -171,6 +171,8 @@ struct llama_layer { struct ggml_tensor * wq_b = nullptr; struct ggml_tensor * wkv_a_mqa = nullptr; struct ggml_tensor * wkv_b = nullptr; + struct ggml_tensor * wk_b = nullptr; + struct ggml_tensor * wv_b = nullptr; struct ggml_tensor * wq_cross = nullptr; struct ggml_tensor * wk_cross = nullptr; struct ggml_tensor * wv_cross = nullptr;