diff --git a/ggml/src/ggml-cpu/arch-fallback.h b/ggml/src/ggml-cpu/arch-fallback.h
index 10e5342516a9c..f02cfe8fa5c06 100644
--- a/ggml/src/ggml-cpu/arch-fallback.h
+++ b/ggml/src/ggml-cpu/arch-fallback.h
@@ -37,17 +37,21 @@
#define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
#define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0
#define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
+#define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
#define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
#define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
#define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
#define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
#define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
+#define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
#define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
#elif defined(__aarch64__) || defined(__arm__) || defined(_M_ARM) || defined(_M_ARM64)
// repack.cpp
#define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8
#define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
+#define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
#define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
+#define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
#elif defined(__x86_64__) || defined(__i386__) || defined(_M_IX86) || defined(_M_X64)
// repack.cpp
#define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
@@ -72,11 +76,13 @@
#define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
#define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0
#define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
+#define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
#define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
#define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
#define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
#define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
#define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
+#define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
#define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
#elif defined(__loongarch64)
// quants.c
@@ -92,11 +98,13 @@
#define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
#define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0
#define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
+#define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
#define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
#define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
#define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
#define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
#define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
+#define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
#define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
#elif defined(__riscv)
// quants.c
@@ -119,10 +127,12 @@
#define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
#define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
#define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
+#define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
#define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
#define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
#define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
#define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
+#define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
#define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
#elif defined(__s390x__)
// quants.c
@@ -147,11 +157,13 @@
#define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
#define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0
#define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
+#define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
#define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
#define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
#define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
#define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
#define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
+#define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
#define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
#elif defined(__wasm__)
// quants.c
@@ -175,10 +187,12 @@
#define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
#define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0
#define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
+#define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
#define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
#define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
#define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
#define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
#define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
+#define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
#define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
#endif
diff --git a/ggml/src/ggml-cpu/arch/x86/repack.cpp b/ggml/src/ggml-cpu/arch/x86/repack.cpp
index c00c1e541cb44..429a6b33af1da 100644
--- a/ggml/src/ggml-cpu/arch/x86/repack.cpp
+++ b/ggml/src/ggml-cpu/arch/x86/repack.cpp
@@ -990,6 +990,372 @@ void ggml_gemv_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
#endif
}
+void ggml_gemv_q2_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+ const int qk = QK_K;
+ const int nb = n / qk;
+ const int ncols_interleaved = 8;
+ const int blocklen = 8;
+
+ assert (n % qk == 0);
+ assert (nc % ncols_interleaved == 0);
+
+ UNUSED(s);
+ UNUSED(bs);
+ UNUSED(vx);
+ UNUSED(vy);
+ UNUSED(nr);
+ UNUSED(nc);
+ UNUSED(nb);
+ UNUSED(ncols_interleaved);
+ UNUSED(blocklen);
+
+#if defined(__AVX2__)
+ // Lookup table to convert signed nibbles to signed bytes
+ __m256i signextendlut = _mm256_castsi128_si256(_mm_set_epi8(-1, -2, -3, -4, -5, -6, -7, -8, 7, 6, 5, 4, 3, 2, 1, 0));
+ signextendlut = _mm256_permute2f128_si256(signextendlut, signextendlut, 0);
+ // Shuffle masks to rearrange delta values to multiply with appropriate scales
+ __m128i deltamask = _mm_set_epi8(15, 14, 7, 6, 13, 12, 5, 4, 11, 10, 3, 2, 9, 8, 1, 0);
+ // Permute mask used for easier vector processing at later stages
+ __m256i finalpermutemask = _mm256_set_epi32(7, 5, 3, 1, 6, 4, 2, 0);
+
+ const __m256i m3b = _mm256_set1_epi8(3);
+ const __m128i m4b_sse = _mm_set1_epi8(0xF);
+
+ //Mask to get appropriate scales
+ __m128i scalemask1 = _mm_set_epi8(14,14,6,6,12,12,4,4,10,10,2,2,8,8,0,0);
+ __m128i scalemask2 = _mm_set_epi8(15,15,7,7,13,13,5,5,11,11,3,3,9,9,1,1);
+
+ int64_t b_nb = n / QK_K;
+
+ const block_q2_Kx8 * b_ptr_start = (const block_q2_Kx8 *)vx;
+ const block_q8_K * a_ptr_start = (const block_q8_K *)vy;
+
+ // Process Q8_K blocks one by one
+ for (int64_t y = 0; y < nr; y++) {
+
+ // Pointers to LHS blocks of block_q8_K format
+ const block_q8_K * a_ptr = a_ptr_start + (y * nb);
+
+ // Take group of eight interleaved block_q2_K structures at each pass of the loop and perform dot product operation
+ for(int64_t x = 0; x < nc / 8; x++) {
+
+ // Pointers to RHS blocks
+ const block_q2_Kx8 * b_ptr = b_ptr_start + (x * b_nb);
+
+ // Master FP accumulators
+ __m256 acc_row = _mm256_setzero_ps();
+ __m256 acc_min_rows = _mm256_setzero_ps();
+
+ for (int64_t b = 0; b < nb; b++) {
+
+ // Load and convert to FP32 delta from block_q8_K
+ const __m256 row_scale_f32 = _mm256_set1_ps((a_ptr[b].d));
+
+ // Load the delta values for the 8 blocks interleaved in block_q2_Kx8
+ // col_scale_f32 rearranged so as to multiply with appropriate quants
+ const __m256 col_scale_f32 = GGML_F32Cx8_REARRANGE_LOAD(b_ptr[b].d, deltamask);
+ const __m256 col_dmin_f32 = GGML_F32Cx8_LOAD(b_ptr[b].dmin);
+
+ __m256i iacc_b = _mm256_setzero_si256();
+ __m256i iacc_min_b = _mm256_setzero_si256();
+
+ // Processes eight sub blocks from each Q2_K in each iteration
+ for(int sb = 0; sb < QK_K / 128; sb++) {
+
+ // Load the eight block_q2_K for eight sub blocks quantized values interleaved with each other in chunks of eight - B0,B1 ....B6,B7
+ const __m256i rhs_raw_vec_0123_0 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + sb * 256));
+ const __m256i rhs_raw_vec_4567_0 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 32 + sb * 256));
+ const __m256i rhs_raw_vec_0123_1 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 64 + sb * 256));
+ const __m256i rhs_raw_vec_4567_1 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 96 + sb * 256));
+ const __m256i rhs_raw_vec_0123_2 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 128 + sb * 256));
+ const __m256i rhs_raw_vec_4567_2 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 160 + sb * 256));
+ const __m256i rhs_raw_vec_0123_3 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 192 + sb * 256));
+ const __m256i rhs_raw_vec_4567_3 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 224 + sb * 256));
+
+ // 2-bit -> 8-bit
+ // Values of the 0th,2nd,4th,6th sub blocks of eight block_q2_K structures for the sb loop
+ const __m256i rhs_vec_0123_00 = _mm256_and_si256(rhs_raw_vec_0123_0, m3b); //B00(0-7) B01(0-7) B02(0-7) B03(0-7)
+ const __m256i rhs_vec_0123_20 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_0, 2), m3b); //B20(0-7) B21(0-7) B22(0-7) B23(0-7)
+ const __m256i rhs_vec_0123_40 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_0, 4), m3b); //B40(0-7) B41(0-7) B42(0-7) B43(0-7)
+ const __m256i rhs_vec_0123_60 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_0, 6), m3b); //B60(0-7) B61(0-7) B62(0-7) B63(0-7)
+
+ const __m256i rhs_vec_4567_00 = _mm256_and_si256(rhs_raw_vec_4567_0, m3b); //B04(0-7) B05(0-7) B06(0-7) B07(0-7)
+ const __m256i rhs_vec_4567_20 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_0, 2), m3b); //B24(0-7) B25(0-7) B26(0-7) B27(0-7)
+ const __m256i rhs_vec_4567_40 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_0, 4), m3b); //B44(0-7) B45(0-7) B46(0-7) B47(0-7)
+ const __m256i rhs_vec_4567_60 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_0, 6), m3b); //B64(0-7) B65(0-7) B66(0-7) B67(0-7)
+
+ const __m256i rhs_vec_0123_01 = _mm256_and_si256(rhs_raw_vec_0123_1, m3b); //B00(8-15) B01(8-15) B02(8-15) B03(8-15)
+ const __m256i rhs_vec_0123_21 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_1, 2), m3b); //B20(8-15) B21(8-15) B22(8-15) B23(8-15)
+ const __m256i rhs_vec_0123_41 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_1, 4), m3b); //B40(8-15) B41(8-15) B42(8-15) B43(8-15)
+ const __m256i rhs_vec_0123_61 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_1, 6), m3b); //B60(8-15) B61(8-15) B62(8-15) B63(8-15)
+
+ const __m256i rhs_vec_4567_01 = _mm256_and_si256(rhs_raw_vec_4567_1, m3b); //B04(8-15) B05(8-15) B06(8-15) B07(8-15)
+ const __m256i rhs_vec_4567_21 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_1, 2), m3b); //B24(8-15) B25(8-15) B26(8-15) B27(8-15)
+ const __m256i rhs_vec_4567_41 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_1, 4), m3b); //B44(8-15) B45(8-15) B46(8-15) B47(8-15)
+ const __m256i rhs_vec_4567_61 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_1, 6), m3b); //B64(8-15) B65(8-15) B66(8-15) B67(8-15)
+
+ // Values of the 1st,3rd,5th,7th sub blocks of eight block_q2_K structures for the sb loop
+ const __m256i rhs_vec_0123_10 = _mm256_and_si256(rhs_raw_vec_0123_2, m3b); //B10(0-7) B11(0-7) B12(0-7) B13(0-7)
+ const __m256i rhs_vec_0123_30 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_2, 2), m3b); //B30(0-7) B31(0-7) B32(0-7) B33(0-7)
+ const __m256i rhs_vec_0123_50 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_2, 4), m3b); //B50(0-7) B51(0-7) B52(0-7) B53(0-7)
+ const __m256i rhs_vec_0123_70 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_2, 6), m3b); //B70(0-7) B71(0-7) B72(0-7) B73(0-7)
+
+ const __m256i rhs_vec_4567_10 = _mm256_and_si256(rhs_raw_vec_4567_2, m3b); //B14(0-7) B15(0-7) B16(0-7) B17(0-7)
+ const __m256i rhs_vec_4567_30 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_2, 2), m3b); //B34(0-7) B35(0-7) B36(0-7) B37(0-7)
+ const __m256i rhs_vec_4567_50 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_2, 4), m3b); //B54(0-7) B55(0-7) B56(0-7) B57(0-7)
+ const __m256i rhs_vec_4567_70 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_2, 6), m3b); //B74(0-7) B75(0-7) B76(0-7) B77(0-7)
+
+ const __m256i rhs_vec_0123_11 = _mm256_and_si256(rhs_raw_vec_0123_3, m3b); //B10(8-15) B11(8-15) B12(8-15) B13(8-15)
+ const __m256i rhs_vec_0123_31 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_3, 2), m3b); //B30(8-15) B31(8-15) B32(8-15) B33(8-15)
+ const __m256i rhs_vec_0123_51 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_3, 4), m3b); //B50(8-15) B51(8-15) B52(8-15) B53(8-15)
+ const __m256i rhs_vec_0123_71 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_3, 6), m3b); //B70(8-15) B71(8-15) B72(8-15) B73(8-15)
+
+ const __m256i rhs_vec_4567_11 = _mm256_and_si256(rhs_raw_vec_4567_3, m3b); //B14(8-15) B15(8-15) B16(8-15) B17(8-15)
+ const __m256i rhs_vec_4567_31 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_3, 2), m3b); //B34(8-15) B35(8-15) B36(8-15) B37(8-15)
+ const __m256i rhs_vec_4567_51 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_3, 4), m3b); //B54(8-15) B55(8-15) B56(8-15) B57(8-15)
+ const __m256i rhs_vec_4567_71 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_3, 6), m3b); //B74(8-15) B75(8-15) B76(8-15) B77(8-15)
+
+ //Scales and Mins of corresponding sub blocks from different Q2_K structures are stored together
+ //s00 m00 s01 m01 s10 m10 s11 m11 s20 m20 s21 m21 s30 m30 s31 m31 s40 m40 s41 m41 s50 m50 s51 m51 s60 m60 s61 m61 s70 m70 s71 m71
+
+ const __m128i mins_and_scales_01 = _mm_loadu_si128((const __m128i *)(b_ptr[b].scales + sb * 64));
+ const __m128i mins_and_scales_23 = _mm_loadu_si128((const __m128i *)(b_ptr[b].scales + 16 + sb * 64));
+ const __m128i mins_and_scales_45 = _mm_loadu_si128((const __m128i *)(b_ptr[b].scales + 32 + sb * 64));
+ const __m128i mins_and_scales_67 = _mm_loadu_si128((const __m128i *)(b_ptr[b].scales + 48 + sb * 64));
+
+ // Extract scales which is lower half from mins_and_scales
+ const __m128i scales_01 = _mm_and_si128(mins_and_scales_01, m4b_sse);
+ const __m128i scales_23 = _mm_and_si128(mins_and_scales_23, m4b_sse);
+ const __m128i scales_45 = _mm_and_si128(mins_and_scales_45, m4b_sse);
+ const __m128i scales_67 = _mm_and_si128(mins_and_scales_67, m4b_sse);
+
+ // Extract mins which is upper half from mins_and_scales
+ const __m256i mins_01 = _mm256_cvtepu8_epi16(_mm_and_si128(_mm_srli_epi16(mins_and_scales_01, 4), m4b_sse));
+ const __m256i mins_23 = _mm256_cvtepu8_epi16(_mm_and_si128(_mm_srli_epi16(mins_and_scales_23, 4), m4b_sse));
+ const __m256i mins_45 = _mm256_cvtepu8_epi16(_mm_and_si128(_mm_srli_epi16(mins_and_scales_45, 4), m4b_sse));
+ const __m256i mins_67 = _mm256_cvtepu8_epi16(_mm_and_si128(_mm_srli_epi16(mins_and_scales_67, 4), m4b_sse));
+
+ // Scales of sub blocks in the sb loop
+ // Scales of the 0th sub block from each super block
+ __m128i scales_rearrange_0 = _mm_shuffle_epi8(scales_01, scalemask1);
+ __m256i scales_0 = _mm256_cvtepu8_epi16(scales_rearrange_0);
+
+ // Scales of the 1st sub block from each super block
+ __m128i scales_rearrange_1 = _mm_shuffle_epi8(scales_01, scalemask2);
+ __m256i scales_1 = _mm256_cvtepu8_epi16(scales_rearrange_1);
+
+ // Scales of the 2nd sub block from each super block
+ __m128i scales_rearrange_2 = _mm_shuffle_epi8(scales_23, scalemask1);
+ __m256i scales_2 = _mm256_cvtepu8_epi16(scales_rearrange_2);
+
+ // Scales of the 3rd sub block from each super block
+ __m128i scales_rearrange_3 = _mm_shuffle_epi8(scales_23, scalemask2);
+ __m256i scales_3 = _mm256_cvtepu8_epi16(scales_rearrange_3);
+
+ // Scales of the 4th sub block from each super block
+ __m128i scales_rearrange_4 = _mm_shuffle_epi8(scales_45, scalemask1);
+ __m256i scales_4 = _mm256_cvtepu8_epi16(scales_rearrange_4);
+
+ // Scales of the 5th sub block from each super block
+ __m128i scales_rearrange_5 = _mm_shuffle_epi8(scales_45, scalemask2);
+ __m256i scales_5 = _mm256_cvtepu8_epi16(scales_rearrange_5);
+
+ // Scales of the 6th sub block from each super block
+ __m128i scales_rearrange_6 = _mm_shuffle_epi8(scales_67, scalemask1);
+ __m256i scales_6 = _mm256_cvtepu8_epi16(scales_rearrange_6);
+
+ // Scales of the 7th sub block from each super block
+ __m128i scales_rearrange_7 = _mm_shuffle_epi8(scales_67, scalemask2);
+ __m256i scales_7 = _mm256_cvtepu8_epi16(scales_rearrange_7);
+
+ // Load the sub block values corresponding to sb in block_q8_K in batches of 16 bytes and replicate the same across 256 bit vector
+ __m256i lhs_vec_0 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + sb * 128)));
+ __m256i lhs_vec_1 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 16 + sb * 128)));
+ __m256i lhs_vec_2 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 32 + sb * 128)));
+ __m256i lhs_vec_3 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 48 + sb * 128)));
+ __m256i lhs_vec_4 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 64 + sb * 128)));
+ __m256i lhs_vec_5 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 80 + sb * 128)));
+ __m256i lhs_vec_6 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 96 + sb * 128)));
+ __m256i lhs_vec_7 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 112 + sb * 128)));
+
+ lhs_vec_0 = _mm256_permute2f128_si256(lhs_vec_0, lhs_vec_0, 0);
+ lhs_vec_1 = _mm256_permute2f128_si256(lhs_vec_1, lhs_vec_1, 0);
+ lhs_vec_2 = _mm256_permute2f128_si256(lhs_vec_2, lhs_vec_2, 0);
+ lhs_vec_3 = _mm256_permute2f128_si256(lhs_vec_3, lhs_vec_3, 0);
+ lhs_vec_4 = _mm256_permute2f128_si256(lhs_vec_4, lhs_vec_4, 0);
+ lhs_vec_5 = _mm256_permute2f128_si256(lhs_vec_5, lhs_vec_5, 0);
+ lhs_vec_6 = _mm256_permute2f128_si256(lhs_vec_6, lhs_vec_6, 0);
+ lhs_vec_7 = _mm256_permute2f128_si256(lhs_vec_7, lhs_vec_7, 0);
+
+ __m256i iacc_0 = _mm256_setzero_si256();
+ __m256i iacc_1 = _mm256_setzero_si256();
+ __m256i iacc_2 = _mm256_setzero_si256();
+ __m256i iacc_3 = _mm256_setzero_si256();
+ __m256i iacc_4 = _mm256_setzero_si256();
+ __m256i iacc_5 = _mm256_setzero_si256();
+ __m256i iacc_6 = _mm256_setzero_si256();
+ __m256i iacc_7 = _mm256_setzero_si256();
+
+ // Dot product done within 32 bit lanes and accumulated in the same vector
+ // First done for 0th sub block and then for seven (1st - 7th) other sub blocks processed for each sb (sb < QK_K/128 loop) // B0(0-3) B4(0-3) B1(0-3) B5(0-3) B2(0-3) B6(0-3) B3(0-3) B7(0-3) with A0(0-3)
+ // B0(4-7) B4(4-7) B1(4-7) B5(4-7) B2(4-7) B6(4-7) B3(4-7) B7(4-7) with A0(4-7)
+ // B0(8-11) B4(8-11) B1(8-11) B5(8-11) B2(8-11) B6(8-11) B3(8-11) B7(8-11) with A0(8-11)
+ // B0(12-15) B4(12-15) B1(12-15) B5(12-15) B2(12-15) B6(12-15) B3(12-15) B7(12-15) with A0(12-15)
+
+ iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_00 ,_mm256_shuffle_epi32(rhs_vec_4567_00, 177), 170), _mm256_shuffle_epi32(lhs_vec_0, 0)));
+ iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_00, 177) ,rhs_vec_4567_00, 170), _mm256_shuffle_epi32(lhs_vec_0, 85)));
+
+ iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_01 ,_mm256_shuffle_epi32(rhs_vec_4567_01, 177), 170), _mm256_shuffle_epi32(lhs_vec_0, 170)));
+ iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_01, 177) ,rhs_vec_4567_01, 170), _mm256_shuffle_epi32(lhs_vec_0, 255)));
+
+ iacc_0 = _mm256_madd_epi16(iacc_0, scales_0);
+
+ iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_10 ,_mm256_shuffle_epi32(rhs_vec_4567_10, 177), 170), _mm256_shuffle_epi32(lhs_vec_1, 0)));
+ iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_10, 177) ,rhs_vec_4567_10, 170), _mm256_shuffle_epi32(lhs_vec_1, 85)));
+
+ iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_11 ,_mm256_shuffle_epi32(rhs_vec_4567_11, 177), 170), _mm256_shuffle_epi32(lhs_vec_1, 170)));
+ iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_11, 177) ,rhs_vec_4567_11, 170), _mm256_shuffle_epi32(lhs_vec_1, 255)));
+
+ iacc_1 = _mm256_madd_epi16(iacc_1, scales_1);
+
+ iacc_2 = _mm256_add_epi16(iacc_2, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_20 ,_mm256_shuffle_epi32(rhs_vec_4567_20, 177), 170), _mm256_shuffle_epi32(lhs_vec_2, 0)));
+ iacc_2 = _mm256_add_epi16(iacc_2, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_20, 177) ,rhs_vec_4567_20, 170), _mm256_shuffle_epi32(lhs_vec_2, 85)));
+
+ iacc_2 = _mm256_add_epi16(iacc_2, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_21 ,_mm256_shuffle_epi32(rhs_vec_4567_21, 177), 170), _mm256_shuffle_epi32(lhs_vec_2, 170)));
+ iacc_2 = _mm256_add_epi16(iacc_2, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_21, 177) ,rhs_vec_4567_21, 170), _mm256_shuffle_epi32(lhs_vec_2, 255)));
+
+ iacc_2 = _mm256_madd_epi16(iacc_2, scales_2);
+
+ iacc_3 = _mm256_add_epi16(iacc_3, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_30 ,_mm256_shuffle_epi32(rhs_vec_4567_30, 177), 170), _mm256_shuffle_epi32(lhs_vec_3, 0)));
+ iacc_3 = _mm256_add_epi16(iacc_3, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_30, 177) ,rhs_vec_4567_30, 170), _mm256_shuffle_epi32(lhs_vec_3, 85)));
+
+ iacc_3 = _mm256_add_epi16(iacc_3, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_31 ,_mm256_shuffle_epi32(rhs_vec_4567_31, 177), 170), _mm256_shuffle_epi32(lhs_vec_3, 170)));
+ iacc_3 = _mm256_add_epi16(iacc_3, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_31, 177) ,rhs_vec_4567_31, 170), _mm256_shuffle_epi32(lhs_vec_3, 255)));
+
+ iacc_3 = _mm256_madd_epi16(iacc_3, scales_3);
+
+ iacc_4 = _mm256_add_epi16(iacc_4, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_40 ,_mm256_shuffle_epi32(rhs_vec_4567_40, 177), 170), _mm256_shuffle_epi32(lhs_vec_4, 0)));
+ iacc_4 = _mm256_add_epi16(iacc_4, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_40, 177) ,rhs_vec_4567_40, 170), _mm256_shuffle_epi32(lhs_vec_4, 85)));
+
+ iacc_4 = _mm256_add_epi16(iacc_4, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_41 ,_mm256_shuffle_epi32(rhs_vec_4567_41, 177), 170), _mm256_shuffle_epi32(lhs_vec_4, 170)));
+ iacc_4 = _mm256_add_epi16(iacc_4, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_41, 177) ,rhs_vec_4567_41, 170), _mm256_shuffle_epi32(lhs_vec_4, 255)));
+
+ iacc_4 = _mm256_madd_epi16(iacc_4, scales_4);
+
+ iacc_5 = _mm256_add_epi16(iacc_5, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_50 ,_mm256_shuffle_epi32(rhs_vec_4567_50, 177), 170), _mm256_shuffle_epi32(lhs_vec_5, 0)));
+ iacc_5 = _mm256_add_epi16(iacc_5, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_50, 177) ,rhs_vec_4567_50, 170), _mm256_shuffle_epi32(lhs_vec_5, 85)));
+
+ iacc_5 = _mm256_add_epi16(iacc_5, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_51 ,_mm256_shuffle_epi32(rhs_vec_4567_51, 177), 170), _mm256_shuffle_epi32(lhs_vec_5, 170)));
+ iacc_5 = _mm256_add_epi16(iacc_5, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_51, 177) ,rhs_vec_4567_51, 170), _mm256_shuffle_epi32(lhs_vec_5, 255)));
+
+ iacc_5 = _mm256_madd_epi16(iacc_5, scales_5);
+
+ iacc_6 = _mm256_add_epi16(iacc_6, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_60 ,_mm256_shuffle_epi32(rhs_vec_4567_60, 177), 170), _mm256_shuffle_epi32(lhs_vec_6, 0)));
+ iacc_6 = _mm256_add_epi16(iacc_6, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_60, 177) ,rhs_vec_4567_60, 170), _mm256_shuffle_epi32(lhs_vec_6, 85)));
+
+ iacc_6 = _mm256_add_epi16(iacc_6, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_61 ,_mm256_shuffle_epi32(rhs_vec_4567_61, 177), 170), _mm256_shuffle_epi32(lhs_vec_6, 170)));
+ iacc_6 = _mm256_add_epi16(iacc_6, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_61, 177) ,rhs_vec_4567_61, 170), _mm256_shuffle_epi32(lhs_vec_6, 255)));
+
+ iacc_6 = _mm256_madd_epi16(iacc_6, scales_6);
+
+ iacc_7 = _mm256_add_epi16(iacc_7, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_70 ,_mm256_shuffle_epi32(rhs_vec_4567_70, 177), 170), _mm256_shuffle_epi32(lhs_vec_7, 0)));
+ iacc_7 = _mm256_add_epi16(iacc_7, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_70, 177) ,rhs_vec_4567_70, 170), _mm256_shuffle_epi32(lhs_vec_7, 85)));
+
+ iacc_7 = _mm256_add_epi16(iacc_7, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_71 ,_mm256_shuffle_epi32(rhs_vec_4567_71, 177), 170), _mm256_shuffle_epi32(lhs_vec_7, 170)));
+ iacc_7 = _mm256_add_epi16(iacc_7, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_71, 177) ,rhs_vec_4567_71, 170), _mm256_shuffle_epi32(lhs_vec_7, 255)));
+
+ iacc_7 = _mm256_madd_epi16(iacc_7, scales_7);
+
+ // Accumulate the iacc value for one sb
+ __m256i iacc_sb = _mm256_add_epi32(_mm256_add_epi32(_mm256_add_epi32(iacc_0, iacc_1), _mm256_add_epi32(iacc_2, iacc_3)), _mm256_add_epi32(_mm256_add_epi32(iacc_4, iacc_5), _mm256_add_epi32(iacc_6, iacc_7)));
+
+ __m128i q8sums = _mm_loadu_si128((const __m128i *)(a_ptr[b].bsums + sb * 8));
+ __m256i q8s = _mm256_castsi128_si256(q8sums);
+ q8s= _mm256_permute2f128_si256(q8s, q8s, 0);
+
+ // Broadcast the bsums of the two corresponding subblocks of q8_k
+ // Multiply-Add with corresponding mins of Q2_Kx8 with bsums
+ __m256i iacc_min_sb_01 = _mm256_madd_epi16(_mm256_shuffle_epi32(q8s, 0), mins_01);
+ __m256i iacc_min_sb_23 = _mm256_madd_epi16(_mm256_shuffle_epi32(q8s, 85), mins_23);
+ __m256i iacc_min_sb_45 = _mm256_madd_epi16(_mm256_shuffle_epi32(q8s, 170), mins_45);
+ __m256i iacc_min_sb_67 = _mm256_madd_epi16(_mm256_shuffle_epi32(q8s, 255), mins_67);
+
+ __m256i iacc_min_sb = _mm256_add_epi32(_mm256_add_epi32(iacc_min_sb_01, iacc_min_sb_23), _mm256_add_epi32(iacc_min_sb_45,iacc_min_sb_67));
+
+ // Accumulate for the complete block
+ iacc_b = _mm256_add_epi32(iacc_b, iacc_sb);
+ iacc_min_b = _mm256_add_epi32(iacc_min_b, iacc_min_sb);
+ }
+
+ //Multiply-Add with scale values for complete super block
+ acc_row = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_b), _mm256_mul_ps(col_scale_f32, row_scale_f32), acc_row);
+ acc_min_rows = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_min_b), _mm256_mul_ps(col_dmin_f32, row_scale_f32), acc_min_rows);
+ }
+ // Accumulated output values permuted so as to be stored in appropriate order post accumulation
+ acc_row = _mm256_permutevar8x32_ps(acc_row, finalpermutemask);
+ _mm256_storeu_ps(s + (y * nr + x * 8), _mm256_sub_ps(acc_row, acc_min_rows));
+ }
+ }
+#else
+
+ float sumf[8];
+ float sum_minf[8];
+ int sumi1,sumi2,sumi3,sumi4;
+ int sumi;
+
+ const block_q8_K * a_ptr = (const block_q8_K *)vy;
+ for(int x = 0; x < nc / ncols_interleaved; x++) {
+ const block_q2_Kx8 * b_ptr = (const block_q2_Kx8 *) vx + (x * nb);
+ for (int j = 0; j < ncols_interleaved; j++) {
+ sumf[j] = 0.0;
+ sum_minf[j] = 0.0;
+ }
+ for (int l = 0; l < nb; l++) {
+ for (int k = 0; k < (qk / (4 * blocklen)); k++) {
+ uint8_t *scales_0 = (uint8_t*) b_ptr[l].scales + (k / 4) * 64 ;
+ uint8_t *scales_1 = (uint8_t*) b_ptr[l].scales + (k / 4) * 64 + 16;
+ uint8_t *scales_2 = (uint8_t*) b_ptr[l].scales + (k / 4) * 64 + 32;
+ uint8_t *scales_3 = (uint8_t*) b_ptr[l].scales + (k / 4) * 64 + 48;
+ for (int j = 0; j < ncols_interleaved; j++) {
+ sumi1 = 0;
+ sumi2 = 0;
+ sumi3 = 0;
+ sumi4 = 0;
+ sumi = 0;
+ int offset = ((k / 2) % 2) + j * 2;
+ for (int i = 0; i < blocklen; ++i){
+ const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 3);
+ const int v1 = (int8_t) ((b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 2 ) & 3);
+ const int v2 = (int8_t) ((b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4 ) & 3);
+ const int v3 = (int8_t) ((b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 6 ) & 3);
+ sumi1 = (v0 * a_ptr[l].qs[(k >> 2) * 128 + (k % 4) * blocklen + i]);
+ sumi2 = (v1 * a_ptr[l].qs[(k >> 2) * 128 + (k % 4) * blocklen + i + 32]);
+ sumi3 = (v2 * a_ptr[l].qs[(k >> 2) * 128 + (k % 4) * blocklen + i + 64]);
+ sumi4 = (v3 * a_ptr[l].qs[(k >> 2) * 128 + (k % 4) * blocklen + i + 96]);
+
+ sumi1 = sumi1 * (scales_0[offset] & 0xF);
+ sumi2 = sumi2 * (scales_1[offset] & 0xF);
+ sumi3 = sumi3 * (scales_2[offset] & 0xF);
+ sumi4 = sumi4 * (scales_3[offset] & 0xF);
+ sumi += sumi1 + sumi2 + sumi3 + sumi4;
+ }
+ sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d;
+ }
+ }
+ for(int sb = 0; sb < 8; sb++) {
+ uint8_t *mins = (uint8_t*) b_ptr[l].scales + sb * 16;
+ for(int j = 0; j < ncols_interleaved; j++){
+ sum_minf[j] += ((mins[j * 2] >> 4) * a_ptr[l].bsums[sb * 2] + (mins[(j * 2)+ 1] >> 4) * a_ptr[l].bsums[sb * 2 + 1]) * GGML_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d;
+ }
+ }
+ }
+ for (int j = 0; j < ncols_interleaved; j++) {
+ s[x * ncols_interleaved + j] = sumf[j] - sum_minf[j];
+ }
+ }
+#endif
+}
+
void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
const int qk = QK8_0;
const int nb = n / qk;
@@ -3283,3 +3649,2952 @@ void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
}
#endif
}
+
+void ggml_gemm_q2_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+ const int qk = QK_K;
+ const int nb = n / qk;
+ const int ncols_interleaved = 8;
+ const int blocklen = 8;
+
+ assert (n % qk == 0);
+ assert (nr % 4 == 0);
+ assert (nc % ncols_interleaved == 0);
+
+ UNUSED(s);
+ UNUSED(bs);
+ UNUSED(vx);
+ UNUSED(vy);
+ UNUSED(nr);
+ UNUSED(nc);
+ UNUSED(nb);
+ UNUSED(ncols_interleaved);
+ UNUSED(blocklen);
+
+#if defined(__AVX2__) || defined(__AVX512F__)
+ const block_q2_Kx8 * b_ptr_start = (const block_q2_Kx8 * ) vx;
+ const block_q8_Kx4 * a_ptr_start = (const block_q8_Kx4 * ) vy;
+ int64_t b_nb = n / QK_K;
+ int64_t y = 0;
+
+ // Permute mask used for easier vector processing at later stages
+ __m256i requiredOrder = _mm256_set_epi32(3, 2, 1, 0, 7, 6, 5, 4);
+ int64_t xstart = 0;
+ int anr = nr - nr % 16; // Used to align nr with boundary of 16
+
+ // Mask to convert 2 bit and 4 bit values into a bytes
+ const __m256i m3b = _mm256_set1_epi8(3);
+ const __m128i m4b_sse = _mm_set1_epi8(0xF);
+
+ //Mask to get appropriate scales
+ __m128i scalesmask1_sse = _mm_set_epi8(14,14,12,12,10,10,8,8,6,6,4,4,2,2,0,0);
+ __m128i scalesmask2_sse = _mm_set_epi8(15,15,13,13,11,11,9,9,7,7,5,5,3,3,1,1);
+
+ __m256i scalesmask1 = _mm256_castsi128_si256(scalesmask1_sse);
+ scalesmask1 = _mm256_permute2f128_si256(scalesmask1, scalesmask1, 0);
+ __m256i scalesmask2 = _mm256_castsi128_si256(scalesmask2_sse);
+ scalesmask2 = _mm256_permute2f128_si256(scalesmask2, scalesmask2, 0);
+
+#ifdef __AVX512F__
+
+ int anc = nc - nc % 16; // Used to align nc with boundary of 16
+
+ // Mask to mask out nibbles from packed bytes
+ const __m256i m4b = _mm256_set1_epi8(0x0F);
+ // Mask to mask out nibbles from packed bytes expanded to 512 bit length
+ const __m512i m3bexpanded = _mm512_set1_epi8(3);
+ //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_q2_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_q2_Kx8 * b_ptr_0 = b_ptr_start + ((x) * b_nb);
+ const block_q2_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++) {
+ // Delta values - Load the sixteen scale values from two block_q2_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_q2_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 sixteen sub blocks of a super block - eight sub blocks are processed per iteration
+ for (int sb = 0; sb < QK_K / 128; sb++) {
+
+ // Load the eight block_q4_k for eight sub blocks quantized values interleaved with each other in chunks of eight bytes - B0,B1 ....B6,B7
+ 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);
+
+ //2-bit -> 8-bit
+ const __m512i rhs_mat_014589CD_00 = _mm512_and_si512(rhs_raw_mat_014589CD_0,m3bexpanded); //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,m3bexpanded); //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,m3bexpanded); //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,m3bexpanded); //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_10 = _mm512_and_si512(rhs_raw_mat_014589CD_2,m3bexpanded); //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(rhs_raw_mat_2367ABEF_2,m3bexpanded); //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(rhs_raw_mat_014589CD_3,m3bexpanded); //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(rhs_raw_mat_2367ABEF_3,m3bexpanded); //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_20 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_0, 2), m3bexpanded); //B20(0-7) B21(0-7) B24(0-7) B25(0-7) B28(0-7) B29(0-7) B2C(0-7) B2D(0-7)
+ const __m512i rhs_mat_2367ABEF_20 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_0, 2), m3bexpanded); //B22(0-7) B23(0-7) B26(0-7) B27(0-7) B2A(0-7) B2B(0-7) B2E(0-7) B2F(0-7)
+
+ const __m512i rhs_mat_014589CD_21 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_1, 2), m3bexpanded); //B20(8-15) B21(8-15) B24(8-15) B25(8-15) B28(8-15) B29(8-15) B2C(8-15) B2D(8-15)
+ const __m512i rhs_mat_2367ABEF_21 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_1, 2), m3bexpanded); //B22(8-15) B23(8-15) B26(8-15) B27(8-15) B2A(8-15) B2B(8-15) B2E(8-15) B2F(8-15)
+
+ const __m512i rhs_mat_014589CD_30 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_2, 2), m3bexpanded); //B30(0-7) B31(0-7) B34(0-7) B35(0-7) B38(0-7) B39(0-7) B3C(0-7) B3D(0-7)
+ const __m512i rhs_mat_2367ABEF_30 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_2, 2), m3bexpanded); //B32(0-7) B33(0-7) B36(0-7) B37(0-7) B3A(0-7) B3B(0-7) B3E(0-7) B3F(0-7)
+
+ const __m512i rhs_mat_014589CD_31 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_3, 2), m3bexpanded); //B30(8-15) B31(8-15) B34(8-15) B35(8-15) B38(8-15) B39(8-15) B3C(8-15) B3D(8-15)
+ const __m512i rhs_mat_2367ABEF_31 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_3, 2), m3bexpanded); //B32(8-15) B33(8-15) B36(8-15) B37(8-15) B3A(8-15) B3B(8-15) B3E(8-15) B3F(8-15)
+
+ const __m512i rhs_mat_014589CD_40 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_0, 4), m3bexpanded); //B40(0-7) B41(0-7) B44(0-7) B45(0-7) B48(0-7) B49(0-7) B4C(0-7) B4D(0-7)
+ const __m512i rhs_mat_2367ABEF_40 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_0, 4), m3bexpanded); //B42(0-7) B43(0-7) B46(0-7) B47(0-7) B4A(0-7) B4B(0-7) B4E(0-7) B4F(0-7)
+
+ const __m512i rhs_mat_014589CD_41 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_1, 4), m3bexpanded); //B40(8-15) B41(8-15) B44(8-15) B45(8-15) B48(8-15) B49(8-15) B4C(8-15) B4D(8-15)
+ const __m512i rhs_mat_2367ABEF_41 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_1, 4), m3bexpanded); //B42(8-15) B43(8-15) B46(8-15) B47(8-15) B4A(8-15) B4B(8-15) B4E(8-15) B4F(8-15)
+
+ const __m512i rhs_mat_014589CD_50 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_2, 4), m3bexpanded); //B50(0-7) B51(0-7) B54(0-7) B55(0-7) B58(0-7) B59(0-7) B5C(0-7) B5D(0-7)
+ const __m512i rhs_mat_2367ABEF_50 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_2, 4), m3bexpanded); //B52(0-7) B53(0-7) B56(0-7) B57(0-7) B5A(0-7) B5B(0-7) B5E(0-7) B5F(0-7)
+
+ const __m512i rhs_mat_014589CD_51 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_3, 4), m3bexpanded); //B50(8-15) B51(8-15) B54(8-15) B55(8-15) B58(8-15) B59(8-15) B5C(8-15) B5D(8-15)
+ const __m512i rhs_mat_2367ABEF_51 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_3, 4), m3bexpanded); //B52(8-15) B53(8-15) B56(8-15) B57(8-15) B5A(8-15) B5B(8-15) B5E(8-15) B5F(8-15)
+
+ const __m512i rhs_mat_014589CD_60 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_0, 6), m3bexpanded); //B60(0-7) B61(0-7) B64(0-7) B65(0-7) B68(0-7) B69(0-7) B6C(0-7) B6D(0-7)
+ const __m512i rhs_mat_2367ABEF_60 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_0, 6), m3bexpanded); //B62(0-7) B63(0-7) B66(0-7) B67(0-7) B6A(0-7) B6B(0-7) B6E(0-7) B6F(0-7)
+
+ const __m512i rhs_mat_014589CD_61 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_1, 6), m3bexpanded); //B60(8-15) B61(8-15) B64(8-15) B65(8-15) B68(8-15) B69(8-15) B6C(8-15) B6D(8-15)
+ const __m512i rhs_mat_2367ABEF_61 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_1, 6), m3bexpanded); //B62(8-15) B63(8-15) B66(8-15) B67(8-15) B6A(8-15) B6B(8-15) B6E(8-15) B6F(8-15)
+
+ const __m512i rhs_mat_014589CD_70 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_2, 6), m3bexpanded); //B70(0-7) B71(0-7) B74(0-7) B75(0-7) B78(0-7) B79(0-7) B7C(0-7) B7D(0-7)
+ const __m512i rhs_mat_2367ABEF_70 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_2, 6), m3bexpanded); //B72(0-7) B73(0-7) B76(0-7) B77(0-7) B7A(0-7) B7B(0-7) B7E(0-7) B7F(0-7)
+
+ const __m512i rhs_mat_014589CD_71 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_3, 6), m3bexpanded); //B70(8-15) B71(8-15) B74(8-15) B75(8-15) B78(8-15) B79(8-15) B7C(8-15) B7D(8-15)
+ const __m512i rhs_mat_2367ABEF_71 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_3, 6), m3bexpanded); //B72(8-15) B73(8-15) B76(8-15) B77(8-15) B7A(8-15) B7B(8-15) B7E(8-15) B7F(8-15)
+
+ 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_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_20_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_20, (_MM_PERM_ENUM)136); //B20(0-3) B21(0-3) B20(0-3) B21(0-3) B24(0-3) B25(0-3) B24(0-3) B25(0-3) B28(0-3) B29(0-3) B28(0-3) B29(0-3) B2C(0-3) B2D(0-3) B2C(0-3) B2D(0-3)
+ const __m512i rhs_mat_2367ABEF_20_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_20, (_MM_PERM_ENUM)136); //B22(0-3) B23(0-3) B22(0-3) B23(0-3) B26(0-3) B27(0-3) B26(0-3) B27(0-3) B2A(0-3) B2B(0-3) B2A(0-3) B2B(0-3) B2E(0-3) B2F(0-3) B2E(0-3) B2F(0-3)
+
+ const __m512i rhs_mat_014589CD_21_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_21, (_MM_PERM_ENUM)136); //B20(8-11) B21(8-11) B20(8-11) B21(8-11) B24(8-11) B25(8-11) B24(8-11) B25(8-11) B28(8-11) B29(8-11) B28(8-11) B29(8-11) B2C(8-11) B2D(8-11) B2C(8-11) B2D(8-11)
+ const __m512i rhs_mat_2367ABEF_21_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_21, (_MM_PERM_ENUM)136); //B22(8-11) B23(8-11) B22(8-11) B23(8-11) B26(8-11) B27(8-11) B26(8-11) B27(8-11) B2A(8-11) B2B(8-11) B2A(8-11) B2B(8-11) B2E(8-11) B2F(8-11) B2E(8-11) B2F(8-11)
+
+ const __m512i rhs_mat_014589CD_30_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_30, (_MM_PERM_ENUM)136); ///B30(0-3) B31(0-3) B30(0-3) B31(0-3) B34(0-3) B35(0-3) B34(0-3) B35(0-3) B38(0-3) B39(0-3) B38(0-3) B39(0-3) B3C(0-3) B3D(0-3) B3C(0-3) B3D(0-3)
+ const __m512i rhs_mat_2367ABEF_30_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_30, (_MM_PERM_ENUM)136); //B32(0-3) B33(0-3) B32(0-3) B33(0-3) B36(0-3) B37(0-3) B36(0-3) B37(0-3) B3A(0-3) B3B(0-3) B3A(0-3) B3B(0-3) B3E(0-3) B3F(0-3) B3E(0-3) B3F(0-3)
+
+ const __m512i rhs_mat_014589CD_31_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_31, (_MM_PERM_ENUM)136); //B30(8-11) B31(8-11) B30(8-11) B31(8-11) B34(8-11) B35(8-11) B34(8-11) B35(8-11) B38(8-11) B39(8-11) B38(8-11) B39(8-11) B3C(8-11) B3D(8-11) B3C(8-11) B3D(8-11)
+ const __m512i rhs_mat_2367ABEF_31_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_31, (_MM_PERM_ENUM)136); //B32(8-11) B33(8-11) B32(8-11) B33(8-11) B36(8-11) B37(8-11) B36(8-11) B37(8-11) B3A(8-11) B3B(8-11) B3A(8-11) B3B(8-11) B3E(8-11) B3F(8-11) B3E(8-11) B3F(8-11)
+
+ const __m512i rhs_mat_014589CD_40_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_40, (_MM_PERM_ENUM)136); //B40(0-3) B41(0-3) B40(0-3) B41(0-3) B44(0-3) B45(0-3) B44(0-3) B45(0-3) B48(0-3) B49(0-3) B48(0-3) B49(0-3) B4C(0-3) B4D(0-3) B4C(0-3) B4D(0-3)
+ const __m512i rhs_mat_2367ABEF_40_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_40, (_MM_PERM_ENUM)136); //B42(0-3) B43(0-3) B42(0-3) B43(0-3) B46(0-3) B47(0-3) B46(0-3) B47(0-3) B4A(0-3) B4B(0-3) B4A(0-3) B4B(0-3) B4E(0-3) B4F(0-3) B4E(0-3) B4F(0-3)
+
+ const __m512i rhs_mat_014589CD_41_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_41, (_MM_PERM_ENUM)136); //B40(8-11) B41(8-11) B40(8-11) B41(8-11) B44(8-11) B45(8-11) B44(8-11) B45(8-11) B48(8-11) B49(8-11) B48(8-11) B49(8-11) B4C(8-11) B4D(8-11) B4C(8-11) B4D(8-11)
+ const __m512i rhs_mat_2367ABEF_41_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_41, (_MM_PERM_ENUM)136); //B42(8-11) B43(8-11) B42(8-11) B43(8-11) B46(8-11) B47(8-11) B46(8-11) B47(8-11) B4A(8-11) B4B(8-11) B4A(8-11) B4B(8-11) B4E(8-11) B4F(8-11) B4E(8-11) B4F(8-11)
+
+ const __m512i rhs_mat_014589CD_50_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_50, (_MM_PERM_ENUM)136); //B50(0-3) B51(0-3) B50(0-3) B51(0-3) B54(0-3) B55(0-3) B54(0-3) B55(0-3) B58(0-3) B59(0-3) B58(0-3) B59(0-3) B5C(0-3) B5D(0-3) B5C(0-3) B5D(0-3)
+ const __m512i rhs_mat_2367ABEF_50_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_50, (_MM_PERM_ENUM)136); //B52(0-3) B53(0-3) B52(0-3) B53(0-3) B56(0-3) B57(0-3) B56(0-3) B57(0-3) B5A(0-3) B5B(0-3) B5A(0-3) B5B(0-3) B5E(0-3) B5F(0-3) B5E(0-3) B5F(0-3)
+
+ const __m512i rhs_mat_014589CD_51_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_51, (_MM_PERM_ENUM)136); //B50(8-11) B51(8-11) B50(8-11) B51(8-11) B54(8-11) B55(8-11) B54(8-11) B55(8-11) B58(8-11) B59(8-11) B58(8-11) B59(8-11) B5C(8-11) B5D(8-11) B5C(8-11) B5D(8-11)
+ const __m512i rhs_mat_2367ABEF_51_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_51, (_MM_PERM_ENUM)136); //B52(8-11) B53(8-11) B52(8-11) B53(8-11) B56(8-11) B57(8-11) B56(8-11) B57(8-11) B5A(8-11) B5B(8-11) B5A(8-11) B5B(8-11) B5E(8-11) B5F(8-11) B5E(8-11) B5F(8-11)
+
+ const __m512i rhs_mat_014589CD_60_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_60, (_MM_PERM_ENUM)136); //B60(0-3) B61(0-3) B60(0-3) B61(0-3) B64(0-3) B65(0-3) B64(0-3) B65(0-3) B68(0-3) B69(0-3) B68(0-3) B69(0-3) B6C(0-3) B6D(0-3) B6C(0-3) B6D(0-3)
+ const __m512i rhs_mat_2367ABEF_60_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_60, (_MM_PERM_ENUM)136); //B62(0-3) B63(0-3) B62(0-3) B63(0-3) B66(0-3) B67(0-3) B66(0-3) B67(0-3) B6A(0-3) B6B(0-3) B6A(0-3) B6B(0-3) B6E(0-3) B6F(0-3) B6E(0-3) B6F(0-3)
+
+ const __m512i rhs_mat_014589CD_61_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_61, (_MM_PERM_ENUM)136); //B60(8-11) B61(8-11) B60(8-11) B61(8-11) B64(8-11) B65(8-11) B64(8-11) B65(8-11) B68(8-11) B69(8-11) B68(8-11) B69(8-11) B6C(8-11) B6D(8-11) B6C(8-11) B6D(8-11)
+ const __m512i rhs_mat_2367ABEF_61_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_61, (_MM_PERM_ENUM)136); //B62(8-11) B63(8-11) B62(8-11) B63(8-11) B66(8-11) B67(8-11) B66(8-11) B67(8-11) B6A(8-11) B6B(8-11) B6A(8-11) B6B(8-11) B6E(8-11) B6F(8-11) B6E(8-11) B6F(8-11)
+
+ const __m512i rhs_mat_014589CD_70_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_70, (_MM_PERM_ENUM)136); //B70(0-3) B71(0-3) B70(0-3) B71(0-3) B74(0-3) B75(0-3) B74(0-3) B75(0-3) B78(0-3) B79(0-3) B78(0-3) B79(0-3) B7C(0-3) B7D(0-3) B7C(0-3) B7D(0-3)
+ const __m512i rhs_mat_2367ABEF_70_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_70, (_MM_PERM_ENUM)136); //B72(0-3) B73(0-3) B72(0-3) B73(0-3) B76(0-3) B77(0-3) B76(0-3) B77(0-3) B7A(0-3) B7B(0-3) B7A(0-3) B7B(0-3) B7E(0-3) B7F(0-3) B7E(0-3) B7F(0-3)
+
+ const __m512i rhs_mat_014589CD_71_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_71, (_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_71_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_71, (_MM_PERM_ENUM)136); //B72(8-11) B73(8-11) B72(8-11) B73(8-11) B76(8-11) B77(8-11) B76(8-11) B77(8-11) B7A(8-11) B7B(8-11) B7A(8-11) B7B(8-11) B7E(8-11) B7F(8-11) B7E(8-11) B7F(8-11)
+
+ 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_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_20_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_20, (_MM_PERM_ENUM)221); //B20(4-7) B21(4-7) B20(4-7) B21(4-7) B24(4-7) B25(4-7) B24(4-7) B25(4-7) B28(4-7) B29(4-7) B28(4-7) B29(4-7) B2C(4-7) B2D(4-7) B2C(4-7) B2D(4-7)
+ const __m512i rhs_mat_2367ABEF_20_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_20, (_MM_PERM_ENUM)221); //B22(4-7) B23(4-7) B22(4-7) B23(4-7) B26(4-7) B27(4-7) B26(4-7) B27(4-7) B2A(4-7) B2B(4-7) B2A(4-7) B2B(4-7) B2E(4-7) B2F(4-7) B2E(4-7) B2F(4-7)
+
+ const __m512i rhs_mat_014589CD_21_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_21, (_MM_PERM_ENUM)221); //B20(12-15) B21(12-15) B20(12-15) B21(12-15) B24(12-15) B25(12-15) B24(12-15) B25(12-15) B28(12-15) B29(12-15) B28(12-15) B29(12-15) B2C(12-15) B2D(12-15) B2C(12-15) B2D(12-15)
+ const __m512i rhs_mat_2367ABEF_21_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_21, (_MM_PERM_ENUM)221); //B22(12-15) B23(12-15) B22(12-15) B23(12-15) B26(12-15) B27(12-15) B26(12-15) B27(12-15) B2A(12-15) B2B(12-15) B2A(12-15) B2B(12-15) B2E(12-15) B2F(12-15) B2E(12-15) B2F(12-15)
+
+ const __m512i rhs_mat_014589CD_30_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_30, (_MM_PERM_ENUM)221); //B30(4-7) B31(4-7) B30(4-7) B31(4-7) B34(4-7) B35(4-7) B34(4-7) B35(4-7) B38(4-7) B39(4-7) B38(4-7) B39(4-7) B3C(4-7) B3D(4-7) B3C(4-7) B3D(4-7)
+ const __m512i rhs_mat_2367ABEF_30_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_30, (_MM_PERM_ENUM)221); //B32(4-7) B33(4-7) B32(4-7) B33(4-7) B36(4-7) B37(4-7) B36(4-7) B37(4-7) B3A(4-7) B3B(4-7) B3A(4-7) B3B(4-7) B3E(4-7) B3F(4-7) B3E(4-7) B3F(4-7)
+
+ const __m512i rhs_mat_014589CD_31_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_31, (_MM_PERM_ENUM)221); //B30(12-15) B31(12-15) B30(12-15) B31(12-15) B34(12-15) B35(12-15) B34(12-15) B35(12-15) B38(12-15) B39(12-15) B38(12-15) B39(12-15) B3C(12-15) B3D(12-15) B3C(12-15) B3D(12-15)
+ const __m512i rhs_mat_2367ABEF_31_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_31, (_MM_PERM_ENUM)221); //B32(12-15) B33(12-15) B32(12-15) B33(12-15) B36(12-15) B37(12-15) B36(12-15) B37(12-15) B3A(12-15) B3B(12-15) B3A(12-15) B3B(12-15) B3E(12-15) B3F(12-15) B3E(12-15) B3F(12-15)
+
+ const __m512i rhs_mat_014589CD_40_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_40, (_MM_PERM_ENUM)221); //B40(4-7) B41(4-7) B40(4-7) B41(4-7) B44(4-7) B45(4-7) B44(4-7) B45(4-7) B48(4-7) B49(4-7) B48(4-7) B49(4-7) B4C(4-7) B4D(4-7) B4C(4-7) B4D(4-7)
+ const __m512i rhs_mat_2367ABEF_40_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_40, (_MM_PERM_ENUM)221); //B42(4-7) B43(4-7) B42(4-7) B43(4-7) B46(4-7) B47(4-7) B46(4-7) B47(4-7) B4A(4-7) B4B(4-7) B4A(4-7) B4B(4-7) B4E(4-7) B4F(4-7) B4E(4-7) B4F(4-7)
+
+ const __m512i rhs_mat_014589CD_41_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_41, (_MM_PERM_ENUM)221); //B40(12-15) B41(12-15) B40(12-15) B41(12-15) B44(12-15) B45(12-15) B44(12-15) B45(12-15) B48(12-15) B49(12-15) B48(12-15) B49(12-15) B4C(12-15) B4D(12-15) B4C(12-15) B4D(12-15)
+ const __m512i rhs_mat_2367ABEF_41_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_41, (_MM_PERM_ENUM)221); //B42(12-15) B43(12-15) B42(12-15) B43(12-15) B46(12-15) B47(12-15) B46(12-15) B47(12-15) B4A(12-15) B4B(12-15) B4A(12-15) B4B(12-15) B4E(12-15) B4F(12-15) B4E(12-15) B4F(12-15)
+
+ const __m512i rhs_mat_014589CD_50_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_50, (_MM_PERM_ENUM)221); //B50(4-7) B51(4-7) B50(4-7) B51(4-7) B54(4-7) B55(4-7) B54(4-7) B55(4-7) B58(4-7) B59(4-7) B58(4-7) B59(4-7) B5C(4-7) B5D(4-7) B5C(4-7) B5D(4-7)
+ const __m512i rhs_mat_2367ABEF_50_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_50, (_MM_PERM_ENUM)221); //B52(4-7) B53(4-7) B52(4-7) B53(4-7) B56(4-7) B57(4-7) B56(4-7) B57(4-7) B5A(4-7) B5B(4-7) B5A(4-7) B5B(4-7) B5E(4-7) B5F(4-7) B5E(4-7) B5F(4-7)
+
+ const __m512i rhs_mat_014589CD_51_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_51, (_MM_PERM_ENUM)221); //B50(12-15) B51(12-15) B50(12-15) B51(12-15) B54(12-15) B55(12-15) B54(12-15) B55(12-15) B58(12-15) B59(12-15) B58(12-15) B59(12-15) B5C(12-15) B5D(12-15) B5C(12-15) B5D(12-15)
+ const __m512i rhs_mat_2367ABEF_51_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_51, (_MM_PERM_ENUM)221); //B52(12-15) B53(12-15) B52(12-15) B53(12-15) B56(12-15) B57(12-15) B56(12-15) B57(12-15) B5A(12-15) B5B(12-15) B5A(12-15) B5B(12-15) B5E(12-15) B5F(12-15) B5E(12-15) B5F(12-15)
+
+ const __m512i rhs_mat_014589CD_60_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_60, (_MM_PERM_ENUM)221); //B60(4-7) B61(4-7) B60(4-7) B61(4-7) B64(4-7) B65(4-7) B64(4-7) B65(4-7) B68(4-7) B69(4-7) B68(4-7) B69(4-7) B6C(4-7) B6D(4-7) B6C(4-7) B6D(4-7)
+ const __m512i rhs_mat_2367ABEF_60_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_60, (_MM_PERM_ENUM)221); //B62(4-7) B63(4-7) B62(4-7) B63(4-7) B66(4-7) B67(4-7) B66(4-7) B67(4-7) B6A(4-7) B6B(4-7) B6A(4-7) B6B(4-7) B6E(4-7) B6F(4-7) B6E(4-7) B6F(4-7)
+
+ const __m512i rhs_mat_014589CD_61_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_61, (_MM_PERM_ENUM)221); //B60(12-15) B61(12-15) B60(12-15) B61(12-15) B64(12-15) B65(12-15) B64(12-15) B65(12-15) B68(12-15) B69(12-15) B68(12-15) B69(12-15) B6C(12-15) B6D(12-15) B6C(12-15) B6D(12-15)
+ const __m512i rhs_mat_2367ABEF_61_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_61, (_MM_PERM_ENUM)221); //B62(12-15) B63(12-15) B62(12-15) B63(12-15) B66(12-15) B67(12-15) B66(12-15) B67(12-15) B6A(12-15) B6B(12-15) B6A(12-15) B6B(12-15) B6E(12-15) B6F(12-15) B6E(12-15) B6F(12-15)
+
+ const __m512i rhs_mat_014589CD_70_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_70, (_MM_PERM_ENUM)221); //B70(4-7) B71(4-7) B70(4-7) B71(4-7) B74(4-7) B75(4-7) B74(4-7) B75(4-7) B78(4-7) B79(4-7) B78(4-7) B79(4-7) B7C(4-7) B7D(4-7) B7C(4-7) B7D(4-7)
+ const __m512i rhs_mat_2367ABEF_70_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_70, (_MM_PERM_ENUM)221); //B72(4-7) B73(4-7) B72(4-7) B73(4-7) B76(4-7) B77(4-7) B76(4-7) B77(4-7) B7A(4-7) B7B(4-7) B7A(4-7) B7B(4-7) B7E(4-7) B7F(4-7) B7E(4-7) B7F(4-7)
+
+ const __m512i rhs_mat_014589CD_71_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_71, (_MM_PERM_ENUM)221); //B70(12-15) B71(12-15) B70(12-15) B71(12-15) B74(12-15) B75(12-15) B74(12-15) B75(12-15) B78(12-15) B79(12-15) B78(12-15) B79(12-15) B7C(12-15) B7D(12-15) B7C(12-15) B7D(12-15)
+ const __m512i rhs_mat_2367ABEF_71_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_71, (_MM_PERM_ENUM)221); //B72(12-15) B73(12-15) B72(12-15) B73(12-15) B76(12-15) B77(12-15) B76(12-15) B77(12-15) B7A(12-15) B7B(12-15) B7A(12-15) B7B(12-15) B7E(12-15) B7F(12-15) B7E(12-15) B7F(12-15)
+
+ //notation:superblock subblock
+ //s00 m00 s01 m01 s10 m10 s11 m11 s20 m20 s21 m21 s30 m30 s31 m31 s40 m40 s41 m41 s50 m50 s51 m51 s60 m60 s61 m61 s70 m70 s71 m71
+
+ const __m128i mins_and_scales_01_0 = _mm_loadu_si128((const __m128i *)(b_ptr_0[b].scales + sb * 64));
+ const __m128i mins_and_scales_23_0 = _mm_loadu_si128((const __m128i *)(b_ptr_0[b].scales + 16 + sb * 64));
+ const __m128i mins_and_scales_45_0 = _mm_loadu_si128((const __m128i *)(b_ptr_0[b].scales + 32 + sb * 64));
+ const __m128i mins_and_scales_67_0 = _mm_loadu_si128((const __m128i *)(b_ptr_0[b].scales + 48 + sb * 64));
+
+ const __m128i mins_and_scales_01_1 = _mm_loadu_si128((const __m128i *)(b_ptr_1[b].scales + sb * 64));
+ const __m128i mins_and_scales_23_1 = _mm_loadu_si128((const __m128i *)(b_ptr_1[b].scales + 16 + sb * 64));
+ const __m128i mins_and_scales_45_1 = _mm_loadu_si128((const __m128i *)(b_ptr_1[b].scales + 32 + sb * 64));
+ const __m128i mins_and_scales_67_1 = _mm_loadu_si128((const __m128i *)(b_ptr_1[b].scales + 48 + sb * 64));
+
+ // Combine mins and scales for sub-blocks: 0-1, 2-3, 4-5, 6-7 in the sb loop
+ const __m256i mins_and_scales_01 = _mm256_insertf128_si256(_mm256_castsi128_si256(mins_and_scales_01_0), mins_and_scales_01_1, 1);
+ const __m256i mins_and_scales_23 = _mm256_insertf128_si256(_mm256_castsi128_si256(mins_and_scales_23_0), mins_and_scales_23_1, 1);
+ const __m256i mins_and_scales_45 = _mm256_insertf128_si256(_mm256_castsi128_si256(mins_and_scales_45_0), mins_and_scales_45_1, 1);
+ const __m256i mins_and_scales_67 = _mm256_insertf128_si256(_mm256_castsi128_si256(mins_and_scales_67_0), mins_and_scales_67_1, 1);
+
+ // Extract scales which is lower half from mins_and_scales
+ const __m256i scales_01 = _mm256_and_si256(mins_and_scales_01, m4b);
+ const __m256i scales_23 = _mm256_and_si256(mins_and_scales_23, m4b);
+ const __m256i scales_45 = _mm256_and_si256(mins_and_scales_45, m4b);
+ const __m256i scales_67 = _mm256_and_si256(mins_and_scales_67, m4b);
+
+ // Extract mins which is upper half from mins_and_scales
+ const __m512i mins_01 = _mm512_cvtepu8_epi16(_mm256_and_si256(_mm256_srli_epi16(mins_and_scales_01, 4), m4b));
+ const __m512i mins_23 = _mm512_cvtepu8_epi16(_mm256_and_si256(_mm256_srli_epi16(mins_and_scales_23, 4), m4b));
+ const __m512i mins_45 = _mm512_cvtepu8_epi16(_mm256_and_si256(_mm256_srli_epi16(mins_and_scales_45, 4), m4b));
+ const __m512i mins_67 = _mm512_cvtepu8_epi16(_mm256_and_si256(_mm256_srli_epi16(mins_and_scales_67, 4), m4b));
+
+ const __m512i scales_0 = _mm512_cvtepu8_epi16(_mm256_shuffle_epi8(scales_01,scalesmask1));
+ const __m512i scales_1 = _mm512_cvtepu8_epi16(_mm256_shuffle_epi8(scales_01,scalesmask2));
+ const __m512i scales_2 = _mm512_cvtepu8_epi16(_mm256_shuffle_epi8(scales_23,scalesmask1));
+ const __m512i scales_3 = _mm512_cvtepu8_epi16(_mm256_shuffle_epi8(scales_23,scalesmask2));
+ const __m512i scales_4 = _mm512_cvtepu8_epi16(_mm256_shuffle_epi8(scales_45,scalesmask1));
+ const __m512i scales_5 = _mm512_cvtepu8_epi16(_mm256_shuffle_epi8(scales_45,scalesmask2));
+ const __m512i scales_6 = _mm512_cvtepu8_epi16(_mm256_shuffle_epi8(scales_67,scalesmask1));
+ const __m512i scales_7 = _mm512_cvtepu8_epi16(_mm256_shuffle_epi8(scales_67,scalesmask2));
+
+ 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);
+
+ const __m512i scale_014589CD_2 = _mm512_shuffle_epi32(scales_2, (_MM_PERM_ENUM)68);
+ const __m512i scale_2367ABEF_2 = _mm512_shuffle_epi32(scales_2, (_MM_PERM_ENUM)238);
+
+ const __m512i scale_014589CD_3 = _mm512_shuffle_epi32(scales_3, (_MM_PERM_ENUM)68);
+ const __m512i scale_2367ABEF_3 = _mm512_shuffle_epi32(scales_3, (_MM_PERM_ENUM)238);
+
+ const __m512i scale_014589CD_4 = _mm512_shuffle_epi32(scales_4, (_MM_PERM_ENUM)68);
+ const __m512i scale_2367ABEF_4 = _mm512_shuffle_epi32(scales_4, (_MM_PERM_ENUM)238);
+
+ const __m512i scale_014589CD_5 = _mm512_shuffle_epi32(scales_5, (_MM_PERM_ENUM)68);
+ const __m512i scale_2367ABEF_5 = _mm512_shuffle_epi32(scales_5, (_MM_PERM_ENUM)238);
+
+ const __m512i scale_014589CD_6 = _mm512_shuffle_epi32(scales_6, (_MM_PERM_ENUM)68);
+ const __m512i scale_2367ABEF_6 = _mm512_shuffle_epi32(scales_6, (_MM_PERM_ENUM)238);
+
+ const __m512i scale_014589CD_7 = _mm512_shuffle_epi32(scales_7, (_MM_PERM_ENUM)68);
+ const __m512i scale_2367ABEF_7 = _mm512_shuffle_epi32(scales_7, (_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 + 512 * 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 + 512 * 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_10 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 64 + 512 * 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 + 96 + 512 * 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_20 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 128 + 512 * sb)));
+ __m256i lhs_mat_ymm_01_20 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_20, lhs_mat_ymm_0123_20, 0);
+ __m256i lhs_mat_ymm_23_20 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_20, lhs_mat_ymm_0123_20, 17);
+ __m256i lhs_mat_ymm_0123_21 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 160 + 512 * sb)));
+ __m256i lhs_mat_ymm_01_21 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_21, lhs_mat_ymm_0123_21, 0);
+ __m256i lhs_mat_ymm_23_21 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_21, lhs_mat_ymm_0123_21, 17);
+ __m256i lhs_mat_ymm_0123_30 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 192 + 512 * sb)));
+ __m256i lhs_mat_ymm_01_30 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_30, lhs_mat_ymm_0123_30, 0);
+ __m256i lhs_mat_ymm_23_30 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_30, lhs_mat_ymm_0123_30, 17);
+ __m256i lhs_mat_ymm_0123_31 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 224 + 512 * sb)));
+ __m256i lhs_mat_ymm_01_31 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_31, lhs_mat_ymm_0123_31, 0);
+ __m256i lhs_mat_ymm_23_31 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_31, lhs_mat_ymm_0123_31, 17);
+
+ __m256i lhs_mat_ymm_0123_40 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 256 + 512 * sb)));
+ __m256i lhs_mat_ymm_01_40 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_40, lhs_mat_ymm_0123_40, 0);
+ __m256i lhs_mat_ymm_23_40 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_40, lhs_mat_ymm_0123_40, 17);
+ __m256i lhs_mat_ymm_0123_41 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 288 + 512 * sb)));
+ __m256i lhs_mat_ymm_01_41 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_41, lhs_mat_ymm_0123_41, 0);
+ __m256i lhs_mat_ymm_23_41 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_41, lhs_mat_ymm_0123_41, 17);
+ __m256i lhs_mat_ymm_0123_50 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 320 + 512 * sb)));
+ __m256i lhs_mat_ymm_01_50 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_50, lhs_mat_ymm_0123_50, 0);
+ __m256i lhs_mat_ymm_23_50 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_50, lhs_mat_ymm_0123_50, 17);
+ __m256i lhs_mat_ymm_0123_51 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 352 + 512 * sb)));
+ __m256i lhs_mat_ymm_01_51 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_51, lhs_mat_ymm_0123_51, 0);
+ __m256i lhs_mat_ymm_23_51 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_51, lhs_mat_ymm_0123_51, 17);
+ __m256i lhs_mat_ymm_0123_60 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 384 + 512 * sb)));
+ __m256i lhs_mat_ymm_01_60 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_60, lhs_mat_ymm_0123_60, 0);
+ __m256i lhs_mat_ymm_23_60 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_60, lhs_mat_ymm_0123_60, 17);
+ __m256i lhs_mat_ymm_0123_61 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 416 + 512 * sb)));
+ __m256i lhs_mat_ymm_01_61 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_61, lhs_mat_ymm_0123_61, 0);
+ __m256i lhs_mat_ymm_23_61 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_61, lhs_mat_ymm_0123_61, 17);
+ __m256i lhs_mat_ymm_0123_70 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 448 + 512 * sb)));
+ __m256i lhs_mat_ymm_01_70 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_70, lhs_mat_ymm_0123_70, 0);
+ __m256i lhs_mat_ymm_23_70 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_70, lhs_mat_ymm_0123_70, 17);
+ __m256i lhs_mat_ymm_0123_71 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 480 + 512 * sb)));
+ __m256i lhs_mat_ymm_01_71 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_71, lhs_mat_ymm_0123_71, 0);
+ __m256i lhs_mat_ymm_23_71 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_71, lhs_mat_ymm_0123_71, 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_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_20 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_20), lhs_mat_ymm_01_20, 1);
+ __m512i lhs_mat_23_20 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_20), lhs_mat_ymm_23_20, 1);
+ __m512i lhs_mat_01_21 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_21), lhs_mat_ymm_01_21, 1);
+ __m512i lhs_mat_23_21 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_21), lhs_mat_ymm_23_21, 1);
+
+ __m512i lhs_mat_01_30 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_30), lhs_mat_ymm_01_30, 1);
+ __m512i lhs_mat_23_30 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_30), lhs_mat_ymm_23_30, 1);
+ __m512i lhs_mat_01_31 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_31), lhs_mat_ymm_01_31, 1);
+ __m512i lhs_mat_23_31 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_31), lhs_mat_ymm_23_31, 1);
+
+ __m512i lhs_mat_01_40 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_40), lhs_mat_ymm_01_40, 1);
+ __m512i lhs_mat_23_40 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_40), lhs_mat_ymm_23_40, 1);
+ __m512i lhs_mat_01_41 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_41), lhs_mat_ymm_01_41, 1);
+ __m512i lhs_mat_23_41 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_41), lhs_mat_ymm_23_41, 1);
+
+ __m512i lhs_mat_01_50 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_50), lhs_mat_ymm_01_50, 1);
+ __m512i lhs_mat_23_50 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_50), lhs_mat_ymm_23_50, 1);
+ __m512i lhs_mat_01_51 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_51), lhs_mat_ymm_01_51, 1);
+ __m512i lhs_mat_23_51 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_51), lhs_mat_ymm_23_51, 1);
+
+ __m512i lhs_mat_01_60 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_60), lhs_mat_ymm_01_60, 1);
+ __m512i lhs_mat_23_60 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_60), lhs_mat_ymm_23_60, 1);
+ __m512i lhs_mat_01_61 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_61), lhs_mat_ymm_01_61, 1);
+ __m512i lhs_mat_23_61 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_61), lhs_mat_ymm_23_61, 1);
+
+ __m512i lhs_mat_01_70 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_70), lhs_mat_ymm_01_70, 1);
+ __m512i lhs_mat_23_70 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_70), lhs_mat_ymm_23_70, 1);
+ __m512i lhs_mat_01_71 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_71), lhs_mat_ymm_01_71, 1);
+ __m512i lhs_mat_23_71 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_71), lhs_mat_ymm_23_71, 1);
+
+ // Bsums are loaded for the different Q8_K blocks
+ __m128i lhs_raw_bsums_01_0123 = _mm_loadu_si128((const __m128i *)((a_ptrs[rp][b].bsums + 32 * sb)));
+ __m128i lhs_raw_bsums_23_0123 = _mm_loadu_si128((const __m128i *)(a_ptrs[rp][b].bsums + 8 + 32 * sb));
+ __m128i lhs_raw_bsums_01_4567 = _mm_loadu_si128((const __m128i *)((a_ptrs[rp][b].bsums + 16 + 32 * sb)));
+ __m128i lhs_raw_bsums_23_4567 = _mm_loadu_si128((const __m128i *)(a_ptrs[rp][b].bsums + 24 + 32 * sb));
+
+ __m256i lhs_bsums_ymm_01_0123 = _mm256_inserti128_si256(_mm256_castsi128_si256(lhs_raw_bsums_01_0123), lhs_raw_bsums_01_0123, 1);
+ __m512i lhs_bsums_01_0123 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_bsums_ymm_01_0123), lhs_bsums_ymm_01_0123, 1);
+ __m256i lhs_bsums_ymm_23_0123 = _mm256_inserti128_si256(_mm256_castsi128_si256(lhs_raw_bsums_23_0123), lhs_raw_bsums_23_0123, 1);
+ __m512i lhs_bsums_23_0123 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_bsums_ymm_23_0123), lhs_bsums_ymm_23_0123, 1); __m256i lhs_bsums_ymm_01_4567 = _mm256_inserti128_si256(_mm256_castsi128_si256(lhs_raw_bsums_01_4567), lhs_raw_bsums_01_4567, 1);
+ __m512i lhs_bsums_01_4567 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_bsums_ymm_01_4567), lhs_bsums_ymm_01_4567, 1);
+ __m256i lhs_bsums_ymm_23_4567 = _mm256_inserti128_si256(_mm256_castsi128_si256(lhs_raw_bsums_23_4567), lhs_raw_bsums_23_4567, 1);
+ __m512i lhs_bsums_23_4567 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_bsums_ymm_23_4567), lhs_bsums_ymm_23_4567, 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_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_20_sp1 = _mm512_shuffle_epi32(lhs_mat_01_20, (_MM_PERM_ENUM)160); //A20(0-3) A20(0-3) A21(0-3) A21(0-3) A20(0-3) A20(0-3) A21(0-3) A21(0-3) A20(0-3) A20(0-3) A21(0-3) A21(0-3) A20(0-3) A20(0-3) A21(0-3) A21(0-3)
+ const __m512i lhs_mat_23_20_sp1 = _mm512_shuffle_epi32(lhs_mat_23_20, (_MM_PERM_ENUM)160); //A22(0-3) A22(0-3) A23(0-3) A23(0-3) A22(0-3) A22(0-3) A23(0-3) A23(0-3) A22(0-3) A22(0-3) A23(0-3) A23(0-3) A22(0-3) A22(0-3) A23(0-3) A23(0-3)
+
+ const __m512i lhs_mat_01_21_sp1 = _mm512_shuffle_epi32(lhs_mat_01_21, (_MM_PERM_ENUM)160); //A20(8-11) A20(8-11) A21(8-11) A21(8-11) A20(8-11) A20(8-11) A21(8-11) A21(8-11) A20(8-11) A20(8-11) A21(8-11) A21(8-11) A20(8-11) A20(8-11) A21(8-11) A21(8-11)
+ const __m512i lhs_mat_23_21_sp1 = _mm512_shuffle_epi32(lhs_mat_23_21, (_MM_PERM_ENUM)160); //A22(8-11) A22(8-11) A23(8-11) A23(8-11) A22(8-11) A22(8-11) A23(8-11) A23(8-11) A22(8-11) A22(8-11) A23(8-11) A23(8-11) A22(8-11) A22(8-11) A23(8-11) A23(8-11)
+
+ const __m512i lhs_mat_01_30_sp1 = _mm512_shuffle_epi32(lhs_mat_01_30, (_MM_PERM_ENUM)160); //A30(0-3) A30(0-3) A31(0-3) A31(0-3) A30(0-3) A30(0-3) A31(0-3) A31(0-3) A30(0-3) A30(0-3) A31(0-3) A31(0-3) A30(0-3) A30(0-3) A31(0-3) A31(0-3)
+ const __m512i lhs_mat_23_30_sp1 = _mm512_shuffle_epi32(lhs_mat_23_30, (_MM_PERM_ENUM)160); //A32(0-3) A32(0-3) A33(0-3) A33(0-3) A32(0-3) A32(0-3) A33(0-3) A33(0-3) A32(0-3) A32(0-3) A33(0-3) A33(0-3) A32(0-3) A32(0-3) A33(0-3) A33(0-3)
+
+ const __m512i lhs_mat_01_31_sp1 = _mm512_shuffle_epi32(lhs_mat_01_31, (_MM_PERM_ENUM)160); //A30(8-11) A30(8-11) A31(8-11) A31(8-11) A30(8-11) A30(8-11) A31(8-11) A31(8-11) A30(8-11) A30(8-11) A31(8-11) A31(8-11) A30(8-11) A30(8-11) A31(8-11) A31(8-11)
+ const __m512i lhs_mat_23_31_sp1 = _mm512_shuffle_epi32(lhs_mat_23_31, (_MM_PERM_ENUM)160); //A32(8-11) A32(8-11) A33(8-11) A33(8-11) A32(8-11) A32(8-11) A33(8-11) A33(8-11) A32(8-11) A32(8-11) A33(8-11) A33(8-11) A32(8-11) A32(8-11) A33(8-11) A33(8-11)
+
+ const __m512i lhs_mat_01_40_sp1 = _mm512_shuffle_epi32(lhs_mat_01_40, (_MM_PERM_ENUM)160); //A40(0-3) A40(0-3) A41(0-3) A41(0-3) A40(0-3) A40(0-3) A41(0-3) A41(0-3) A40(0-3) A40(0-3) A41(0-3) A41(0-3) A40(0-3) A40(0-3) A41(0-3) A41(0-3)
+ const __m512i lhs_mat_23_40_sp1 = _mm512_shuffle_epi32(lhs_mat_23_40, (_MM_PERM_ENUM)160); //A42(0-3) A42(0-3) A43(0-3) A43(0-3) A42(0-3) A42(0-3) A43(0-3) A43(0-3) A42(0-3) A42(0-3) A43(0-3) A43(0-3) A42(0-3) A42(0-3) A43(0-3) A43(0-3)
+
+ const __m512i lhs_mat_01_41_sp1 = _mm512_shuffle_epi32(lhs_mat_01_41, (_MM_PERM_ENUM)160); //A40(8-11) A40(8-11) A41(8-11) A41(8-11) A40(8-11) A40(8-11) A41(8-11) A41(8-11) A40(8-11) A40(8-11) A41(8-11) A41(8-11) A40(8-11) A40(8-11) A41(8-11) A41(8-11)
+ const __m512i lhs_mat_23_41_sp1 = _mm512_shuffle_epi32(lhs_mat_23_41, (_MM_PERM_ENUM)160); //A42(8-11) A42(8-11) A43(8-11) A43(8-11) A42(8-11) A42(8-11) A43(8-11) A43(8-11) A42(8-11) A42(8-11) A43(8-11) A43(8-11) A42(8-11) A42(8-11) A43(8-11) A43(8-11)
+
+ const __m512i lhs_mat_01_50_sp1 = _mm512_shuffle_epi32(lhs_mat_01_50, (_MM_PERM_ENUM)160); //A50(0-3) A50(0-3) A51(0-3) A51(0-3) A50(0-3) A50(0-3) A51(0-3) A51(0-3) A50(0-3) A50(0-3) A51(0-3) A51(0-3) A50(0-3) A50(0-3) A51(0-3) A51(0-3)
+ const __m512i lhs_mat_23_50_sp1 = _mm512_shuffle_epi32(lhs_mat_23_50, (_MM_PERM_ENUM)160); //A52(0-3) A52(0-3) A53(0-3) A53(0-3) A52(0-3) A52(0-3) A53(0-3) A53(0-3) A52(0-3) A52(0-3) A53(0-3) A53(0-3) A52(0-3) A52(0-3) A53(0-3) A53(0-3)
+
+ const __m512i lhs_mat_01_51_sp1 = _mm512_shuffle_epi32(lhs_mat_01_51, (_MM_PERM_ENUM)160); //A50(8-11) A50(8-11) A51(8-11) A51(8-11) A50(8-11) A50(8-11) A51(8-11) A51(8-11) A50(8-11) A50(8-11) A51(8-11) A51(8-11) A50(8-11) A50(8-11) A51(8-11) A51(8-11)
+ const __m512i lhs_mat_23_51_sp1 = _mm512_shuffle_epi32(lhs_mat_23_51, (_MM_PERM_ENUM)160); //A52(8-11) A52(8-11) A53(8-11) A53(8-11) A52(8-11) A52(8-11) A53(8-11) A53(8-11) A52(8-11) A52(8-11) A53(8-11) A53(8-11) A52(8-11) A52(8-11) A53(8-11) A53(8-11)
+
+ const __m512i lhs_mat_01_60_sp1 = _mm512_shuffle_epi32(lhs_mat_01_60, (_MM_PERM_ENUM)160); //A60(0-3) A60(0-3) A61(0-3) A61(0-3) A60(0-3) A60(0-3) A61(0-3) A61(0-3) A60(0-3) A60(0-3) A61(0-3) A61(0-3) A60(0-3) A60(0-3) A61(0-3) A61(0-3)
+ const __m512i lhs_mat_23_60_sp1 = _mm512_shuffle_epi32(lhs_mat_23_60, (_MM_PERM_ENUM)160); //A62(0-3) A62(0-3) A63(0-3) A63(0-3) A62(0-3) A62(0-3) A63(0-3) A63(0-3) A62(0-3) A62(0-3) A63(0-3) A63(0-3) A62(0-3) A62(0-3) A63(0-3) A63(0-3)
+
+ const __m512i lhs_mat_01_61_sp1 = _mm512_shuffle_epi32(lhs_mat_01_61, (_MM_PERM_ENUM)160); //A60(8-11) A60(8-11) A61(8-11) A61(8-11) A60(8-11) A60(8-11) A61(8-11) A61(8-11) A60(8-11) A60(8-11) A61(8-11) A61(8-11) A60(8-11) A60(8-11) A61(8-11) A61(8-11)
+ const __m512i lhs_mat_23_61_sp1 = _mm512_shuffle_epi32(lhs_mat_23_61, (_MM_PERM_ENUM)160); //A62(8-11) A62(8-11) A63(8-11) A63(8-11) A62(8-11) A62(8-11) A63(8-11) A63(8-11) A62(8-11) A62(8-11) A63(8-11) A63(8-11) A62(8-11) A62(8-11) A63(8-11) A63(8-11)
+
+ const __m512i lhs_mat_01_70_sp1 = _mm512_shuffle_epi32(lhs_mat_01_70, (_MM_PERM_ENUM)160); //A70(0-3) A70(0-3) A71(0-3) A71(0-3) A70(0-3) A70(0-3) A71(0-3) A71(0-3) A70(0-3) A70(0-3) A71(0-3) A71(0-3) A70(0-3) A70(0-3) A71(0-3) A71(0-3)
+ const __m512i lhs_mat_23_70_sp1 = _mm512_shuffle_epi32(lhs_mat_23_70, (_MM_PERM_ENUM)160); //A72(0-3) A72(0-3) A73(0-3) A73(0-3) A72(0-3) A72(0-3) A73(0-3) A73(0-3) A72(0-3) A72(0-3) A73(0-3) A73(0-3) A72(0-3) A72(0-3) A73(0-3) A73(0-3)
+
+ const __m512i lhs_mat_01_71_sp1 = _mm512_shuffle_epi32(lhs_mat_01_71, (_MM_PERM_ENUM)160); //A70(8-11) A70(8-11) A71(8-11) A71(8-11) A70(8-11) A70(8-11) A71(8-11) A71(8-11) A70(8-11) A70(8-11) A71(8-11) A71(8-11) A70(8-11) A70(8-11) A71(8-11) A71(8-11)
+ const __m512i lhs_mat_23_71_sp1 = _mm512_shuffle_epi32(lhs_mat_23_71, (_MM_PERM_ENUM)160); //A72(8-11) A72(8-11) A73(8-11) A73(8-11) A72(8-11) A72(8-11) A73(8-11) A73(8-11) A72(8-11) A72(8-11) A73(8-11) A73(8-11) A72(8-11) A72(8-11) A73(8-11) A73(8-11)
+
+ 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_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_20_sp2 = _mm512_shuffle_epi32(lhs_mat_01_20, (_MM_PERM_ENUM)245); //A20(4-7) A20(4-7) A21(4-7) A21(4-7) A20(4-7) A20(4-7) A21(4-7) A21(4-7) A20(4-7) A20(4-7) A21(4-7) A21(4-7) A20(4-7) A20(4-7) A21(4-7) A21(4-7)
+ const __m512i lhs_mat_23_20_sp2 = _mm512_shuffle_epi32(lhs_mat_23_20, (_MM_PERM_ENUM)245); //A22(4-7) A22(4-7) A23(4-7) A23(4-7) A22(4-7) A22(4-7) A23(4-7) A23(4-7) A22(4-7) A22(4-7) A23(4-7) A23(4-7) A22(4-7) A22(4-7) A23(4-7) A23(4-7)
+
+ const __m512i lhs_mat_01_21_sp2 = _mm512_shuffle_epi32(lhs_mat_01_21, (_MM_PERM_ENUM)245); //A20(12-15) A20(12-15) A21(12-15) A21(12-15) A20(12-15) A20(12-15) A21(12-15) A21(12-15) A20(12-15) A20(12-15) A21(12-15) A21(12-15) A20(12-15) A20(12-15) A21(12-15) A21(12-15)
+ const __m512i lhs_mat_23_21_sp2 = _mm512_shuffle_epi32(lhs_mat_23_21, (_MM_PERM_ENUM)245); //A22(12-15) A22(12-15) A23(12-15) A23(12-15) A22(12-15) A22(12-15) A23(12-15) A23(12-15) A22(12-15) A22(12-15) A23(12-15) A23(12-15) A22(12-15) A22(12-15) A23(12-15) A23(12-15)
+
+ const __m512i lhs_mat_01_30_sp2 = _mm512_shuffle_epi32(lhs_mat_01_30, (_MM_PERM_ENUM)245); //A30(4-7) A30(4-7) A31(4-7) A31(4-7) A30(4-7) A30(4-7) A31(4-7) A31(4-7) A30(4-7) A30(4-7) A31(4-7) A31(4-7) A30(4-7) A30(4-7) A31(4-7) A31(4-7)
+ const __m512i lhs_mat_23_30_sp2 = _mm512_shuffle_epi32(lhs_mat_23_30, (_MM_PERM_ENUM)245); //A32(4-7) A32(4-7) A33(4-7) A33(4-7) A32(4-7) A32(4-7) A33(4-7) A33(4-7) A32(4-7) A32(4-7) A33(4-7) A33(4-7) A32(4-7) A32(4-7) A33(4-7) A33(4-7)
+
+ const __m512i lhs_mat_01_31_sp2 = _mm512_shuffle_epi32(lhs_mat_01_31, (_MM_PERM_ENUM)245); //A30(12-15) A30(12-15) A31(12-15) A31(12-15) A30(12-15) A30(12-15) A31(12-15) A31(12-15) A30(12-15) A30(12-15) A31(12-15) A31(12-15) A30(12-15) A30(12-15) A31(12-15) A31(12-15)
+ const __m512i lhs_mat_23_31_sp2 = _mm512_shuffle_epi32(lhs_mat_23_31, (_MM_PERM_ENUM)245); //A32(12-15) A32(12-15) A33(12-15) A33(12-15) A32(12-15) A32(12-15) A33(12-15) A33(12-15) A32(12-15) A32(12-15) A33(12-15) A33(12-15) A32(12-15) A32(12-15) A33(12-15) A33(12-15)
+
+ const __m512i lhs_mat_01_40_sp2 = _mm512_shuffle_epi32(lhs_mat_01_40, (_MM_PERM_ENUM)245); //A40(4-7) A40(4-7) A41(4-7) A41(4-7) A40(4-7) A40(4-7) A41(4-7) A41(4-7) A40(4-7) A40(4-7) A41(4-7) A41(4-7) A40(4-7) A40(4-7) A41(4-7) A41(4-7)
+ const __m512i lhs_mat_23_40_sp2 = _mm512_shuffle_epi32(lhs_mat_23_40, (_MM_PERM_ENUM)245); //A42(4-7) A42(4-7) A43(4-7) A43(4-7) A42(4-7) A42(4-7) A43(4-7) A43(4-7) A42(4-7) A42(4-7) A43(4-7) A43(4-7) A42(4-7) A42(4-7) A43(4-7) A43(4-7)
+
+ const __m512i lhs_mat_01_41_sp2 = _mm512_shuffle_epi32(lhs_mat_01_41, (_MM_PERM_ENUM)245); //A40(12-15) A40(12-15) A41(12-15) A41(12-15) A40(12-15) A40(12-15) A41(12-15) A41(12-15) A40(12-15) A40(12-15) A41(12-15) A41(12-15) A40(12-15) A40(12-15) A41(12-15) A41(12-15)
+ const __m512i lhs_mat_23_41_sp2 = _mm512_shuffle_epi32(lhs_mat_23_41, (_MM_PERM_ENUM)245); //A42(12-15) A42(12-15) A43(12-15) A43(12-15) A42(12-15) A42(12-15) A43(12-15) A43(12-15) A42(12-15) A42(12-15) A43(12-15) A43(12-15) A42(12-15) A42(12-15) A43(12-15) A43(12-15)
+
+ const __m512i lhs_mat_01_50_sp2 = _mm512_shuffle_epi32(lhs_mat_01_50, (_MM_PERM_ENUM)245); //A50(4-7) A50(4-7) A51(4-7) A51(4-7) A50(4-7) A50(4-7) A51(4-7) A51(4-7) A50(4-7) A50(4-7) A51(4-7) A51(4-7) A50(4-7) A50(4-7) A51(4-7) A51(4-7)
+ const __m512i lhs_mat_23_50_sp2 = _mm512_shuffle_epi32(lhs_mat_23_50, (_MM_PERM_ENUM)245); //A52(4-7) A52(4-7) A53(4-7) A53(4-7) A52(4-7) A52(4-7) A53(4-7) A53(4-7) A52(4-7) A52(4-7) A53(4-7) A53(4-7) A52(4-7) A52(4-7) A53(4-7) A53(4-7)
+
+ const __m512i lhs_mat_01_51_sp2 = _mm512_shuffle_epi32(lhs_mat_01_51, (_MM_PERM_ENUM)245); //A50(12-15) A50(12-15) A51(12-15) A51(12-15) A50(12-15) A50(12-15) A51(12-15) A51(12-15) A50(12-15) A50(12-15) A51(12-15) A51(12-15) A50(12-15) A50(12-15) A51(12-15) A51(12-15)
+ const __m512i lhs_mat_23_51_sp2 = _mm512_shuffle_epi32(lhs_mat_23_51, (_MM_PERM_ENUM)245); //A52(12-15) A52(12-15) A53(12-15) A53(12-15) A52(12-15) A52(12-15) A53(12-15) A53(12-15) A52(12-15) A52(12-15) A53(12-15) A53(12-15) A52(12-15) A52(12-15) A53(12-15) A53(12-15)
+
+ const __m512i lhs_mat_01_60_sp2 = _mm512_shuffle_epi32(lhs_mat_01_60, (_MM_PERM_ENUM)245); //A60(4-7) A60(4-7) A61(4-7) A61(4-7) A60(4-7) A60(4-7) A61(4-7) A61(4-7) A60(4-7) A60(4-7) A61(4-7) A61(4-7) A60(4-7) A60(4-7) A61(4-7) A61(4-7)
+ const __m512i lhs_mat_23_60_sp2 = _mm512_shuffle_epi32(lhs_mat_23_60, (_MM_PERM_ENUM)245); //A62(4-7) A62(4-7) A63(4-7) A63(4-7) A62(4-7) A62(4-7) A63(4-7) A63(4-7) A62(4-7) A62(4-7) A63(4-7) A63(4-7) A62(4-7) A62(4-7) A63(4-7) A63(4-7)
+
+ const __m512i lhs_mat_01_61_sp2 = _mm512_shuffle_epi32(lhs_mat_01_61, (_MM_PERM_ENUM)245); //A60(12-15) A60(12-15) A61(12-15) A61(12-15) A60(12-15) A60(12-15) A61(12-15) A61(12-15) A60(12-15) A60(12-15) A61(12-15) A61(12-15) A60(12-15) A60(12-15) A61(12-15) A61(12-15)
+ const __m512i lhs_mat_23_61_sp2 = _mm512_shuffle_epi32(lhs_mat_23_61, (_MM_PERM_ENUM)245); //A62(12-15) A62(12-15) A63(12-15) A63(12-15) A62(12-15) A62(12-15) A63(12-15) A63(12-15) A62(12-15) A62(12-15) A63(12-15) A63(12-15) A62(12-15) A62(12-15) A63(12-15) A63(12-15)
+
+ const __m512i lhs_mat_01_70_sp2 = _mm512_shuffle_epi32(lhs_mat_01_70, (_MM_PERM_ENUM)245); //A70(4-7) A70(4-7) A71(4-7) A71(4-7) A70(4-7) A70(4-7) A71(4-7) A71(4-7) A70(4-7) A70(4-7) A71(4-7) A71(4-7) A70(4-7) A70(4-7) A71(4-7) A71(4-7)
+ const __m512i lhs_mat_23_70_sp2 = _mm512_shuffle_epi32(lhs_mat_23_70, (_MM_PERM_ENUM)245); //A72(4-7) A72(4-7) A73(4-7) A73(4-7) A72(4-7) A72(4-7) A73(4-7) A73(4-7) A72(4-7) A72(4-7) A73(4-7) A73(4-7) A72(4-7) A72(4-7) A73(4-7) A73(4-7)
+
+ const __m512i lhs_mat_01_71_sp2 = _mm512_shuffle_epi32(lhs_mat_01_71, (_MM_PERM_ENUM)245); //A70(12-15) A70(12-15) A71(12-15) A71(12-15) A70(12-15) A70(12-15) A71(12-15) A71(12-15) A70(12-15) A70(12-15) A71(12-15) A71(12-15) A70(12-15) A70(12-15) A71(12-15) A71(12-15)
+ const __m512i lhs_mat_23_71_sp2 = _mm512_shuffle_epi32(lhs_mat_23_71, (_MM_PERM_ENUM)245); //A72(12-15) A72(12-15) A73(12-15) A73(12-15) A72(12-15) A72(12-15) A73(12-15) A73(12-15) A72(12-15) A72(12-15) A73(12-15) A73(12-15) A72(12-15) A72(12-15) A73(12-15) A73(12-15)
+
+ // 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_maddubs_epi16(rhs_mat_014589CD_00_sp1, lhs_mat_01_00_sp1),_mm512_maddubs_epi16(rhs_mat_014589CD_01_sp1, lhs_mat_01_01_sp1));
+ __m512i iacc_mat_01_0_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_00_sp1, lhs_mat_01_00_sp1),_mm512_maddubs_epi16(rhs_mat_2367ABEF_01_sp1, lhs_mat_01_01_sp1));
+
+ __m512i iacc_mat_10_0_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_00_sp1, lhs_mat_23_00_sp1),_mm512_maddubs_epi16(rhs_mat_014589CD_01_sp1, lhs_mat_23_01_sp1));
+ __m512i iacc_mat_11_0_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_00_sp1, lhs_mat_23_00_sp1),_mm512_maddubs_epi16(rhs_mat_2367ABEF_01_sp1, lhs_mat_23_01_sp1));
+
+ __m512i iacc_mat_00_1_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_10_sp1, lhs_mat_01_10_sp1),_mm512_maddubs_epi16(rhs_mat_014589CD_11_sp1, lhs_mat_01_11_sp1));
+ __m512i iacc_mat_01_1_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_10_sp1, lhs_mat_01_10_sp1),_mm512_maddubs_epi16(rhs_mat_2367ABEF_11_sp1, lhs_mat_01_11_sp1));
+
+ __m512i iacc_mat_10_1_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_10_sp1, lhs_mat_23_10_sp1),_mm512_maddubs_epi16(rhs_mat_014589CD_11_sp1, lhs_mat_23_11_sp1));
+ __m512i iacc_mat_11_1_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_10_sp1, lhs_mat_23_10_sp1),_mm512_maddubs_epi16(rhs_mat_2367ABEF_11_sp1, lhs_mat_23_11_sp1));
+
+ __m512i iacc_mat_00_2_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_20_sp1, lhs_mat_01_20_sp1),_mm512_maddubs_epi16(rhs_mat_014589CD_21_sp1, lhs_mat_01_21_sp1));
+ __m512i iacc_mat_01_2_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_20_sp1, lhs_mat_01_20_sp1),_mm512_maddubs_epi16(rhs_mat_2367ABEF_21_sp1, lhs_mat_01_21_sp1));
+
+ __m512i iacc_mat_10_2_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_20_sp1, lhs_mat_23_20_sp1),_mm512_maddubs_epi16(rhs_mat_014589CD_21_sp1, lhs_mat_23_21_sp1));
+ __m512i iacc_mat_11_2_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_20_sp1, lhs_mat_23_20_sp1),_mm512_maddubs_epi16(rhs_mat_2367ABEF_21_sp1, lhs_mat_23_21_sp1));
+
+ __m512i iacc_mat_00_3_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_30_sp1, lhs_mat_01_30_sp1),_mm512_maddubs_epi16(rhs_mat_014589CD_31_sp1, lhs_mat_01_31_sp1));
+ __m512i iacc_mat_01_3_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_30_sp1, lhs_mat_01_30_sp1),_mm512_maddubs_epi16(rhs_mat_2367ABEF_31_sp1, lhs_mat_01_31_sp1));
+
+ __m512i iacc_mat_10_3_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_30_sp1, lhs_mat_23_30_sp1),_mm512_maddubs_epi16(rhs_mat_014589CD_31_sp1, lhs_mat_23_31_sp1));
+ __m512i iacc_mat_11_3_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_30_sp1, lhs_mat_23_30_sp1),_mm512_maddubs_epi16(rhs_mat_2367ABEF_31_sp1, lhs_mat_23_31_sp1));
+
+ __m512i iacc_mat_00_4_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_40_sp1, lhs_mat_01_40_sp1),_mm512_maddubs_epi16(rhs_mat_014589CD_41_sp1, lhs_mat_01_41_sp1));
+ __m512i iacc_mat_01_4_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_40_sp1, lhs_mat_01_40_sp1),_mm512_maddubs_epi16(rhs_mat_2367ABEF_41_sp1, lhs_mat_01_41_sp1));
+
+ __m512i iacc_mat_10_4_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_40_sp1, lhs_mat_23_40_sp1),_mm512_maddubs_epi16(rhs_mat_014589CD_41_sp1, lhs_mat_23_41_sp1));
+ __m512i iacc_mat_11_4_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_40_sp1, lhs_mat_23_40_sp1),_mm512_maddubs_epi16(rhs_mat_2367ABEF_41_sp1, lhs_mat_23_41_sp1));
+
+ __m512i iacc_mat_00_5_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_50_sp1, lhs_mat_01_50_sp1),_mm512_maddubs_epi16(rhs_mat_014589CD_51_sp1, lhs_mat_01_51_sp1));
+ __m512i iacc_mat_01_5_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_50_sp1, lhs_mat_01_50_sp1),_mm512_maddubs_epi16(rhs_mat_2367ABEF_51_sp1, lhs_mat_01_51_sp1));
+
+ __m512i iacc_mat_10_5_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_50_sp1, lhs_mat_23_50_sp1),_mm512_maddubs_epi16(rhs_mat_014589CD_51_sp1, lhs_mat_23_51_sp1));
+ __m512i iacc_mat_11_5_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_50_sp1, lhs_mat_23_50_sp1),_mm512_maddubs_epi16(rhs_mat_2367ABEF_51_sp1, lhs_mat_23_51_sp1));
+
+ __m512i iacc_mat_00_6_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_60_sp1, lhs_mat_01_60_sp1),_mm512_maddubs_epi16(rhs_mat_014589CD_61_sp1, lhs_mat_01_61_sp1));
+ __m512i iacc_mat_01_6_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_60_sp1, lhs_mat_01_60_sp1),_mm512_maddubs_epi16(rhs_mat_2367ABEF_61_sp1, lhs_mat_01_61_sp1));
+
+ __m512i iacc_mat_10_6_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_60_sp1, lhs_mat_23_60_sp1),_mm512_maddubs_epi16(rhs_mat_014589CD_61_sp1, lhs_mat_23_61_sp1));
+ __m512i iacc_mat_11_6_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_60_sp1, lhs_mat_23_60_sp1),_mm512_maddubs_epi16(rhs_mat_2367ABEF_61_sp1, lhs_mat_23_61_sp1));
+
+ __m512i iacc_mat_00_7_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_70_sp1, lhs_mat_01_70_sp1),_mm512_maddubs_epi16(rhs_mat_014589CD_71_sp1, lhs_mat_01_71_sp1));
+ __m512i iacc_mat_01_7_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_70_sp1, lhs_mat_01_70_sp1),_mm512_maddubs_epi16(rhs_mat_2367ABEF_71_sp1, lhs_mat_01_71_sp1));
+
+ __m512i iacc_mat_10_7_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_70_sp1, lhs_mat_23_70_sp1),_mm512_maddubs_epi16(rhs_mat_014589CD_71_sp1, lhs_mat_23_71_sp1));
+ __m512i iacc_mat_11_7_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_70_sp1, lhs_mat_23_70_sp1),_mm512_maddubs_epi16(rhs_mat_2367ABEF_71_sp1, lhs_mat_23_71_sp1));
+
+
+ __m512i iacc_mat_00_0_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_00_sp2, lhs_mat_01_00_sp2),_mm512_maddubs_epi16(rhs_mat_014589CD_01_sp2, lhs_mat_01_01_sp2));
+ __m512i iacc_mat_01_0_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_00_sp2, lhs_mat_01_00_sp2),_mm512_maddubs_epi16(rhs_mat_2367ABEF_01_sp2, lhs_mat_01_01_sp2));
+
+ __m512i iacc_mat_10_0_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_00_sp2, lhs_mat_23_00_sp2),_mm512_maddubs_epi16(rhs_mat_014589CD_01_sp2, lhs_mat_23_01_sp2));
+ __m512i iacc_mat_11_0_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_00_sp2, lhs_mat_23_00_sp2),_mm512_maddubs_epi16(rhs_mat_2367ABEF_01_sp2, lhs_mat_23_01_sp2));
+
+ __m512i iacc_mat_00_1_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_10_sp2, lhs_mat_01_10_sp2),_mm512_maddubs_epi16(rhs_mat_014589CD_11_sp2, lhs_mat_01_11_sp2));
+ __m512i iacc_mat_01_1_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_10_sp2, lhs_mat_01_10_sp2),_mm512_maddubs_epi16(rhs_mat_2367ABEF_11_sp2, lhs_mat_01_11_sp2));
+
+ __m512i iacc_mat_10_1_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_10_sp2, lhs_mat_23_10_sp2),_mm512_maddubs_epi16(rhs_mat_014589CD_11_sp2, lhs_mat_23_11_sp2));
+ __m512i iacc_mat_11_1_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_10_sp2, lhs_mat_23_10_sp2),_mm512_maddubs_epi16(rhs_mat_2367ABEF_11_sp2, lhs_mat_23_11_sp2));
+
+ __m512i iacc_mat_00_2_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_20_sp2, lhs_mat_01_20_sp2),_mm512_maddubs_epi16(rhs_mat_014589CD_21_sp2, lhs_mat_01_21_sp2));
+ __m512i iacc_mat_01_2_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_20_sp2, lhs_mat_01_20_sp2),_mm512_maddubs_epi16(rhs_mat_2367ABEF_21_sp2, lhs_mat_01_21_sp2));
+
+ __m512i iacc_mat_10_2_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_20_sp2, lhs_mat_23_20_sp2),_mm512_maddubs_epi16(rhs_mat_014589CD_21_sp2, lhs_mat_23_21_sp2));
+ __m512i iacc_mat_11_2_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_20_sp2, lhs_mat_23_20_sp2),_mm512_maddubs_epi16(rhs_mat_2367ABEF_21_sp2, lhs_mat_23_21_sp2));
+
+ __m512i iacc_mat_00_3_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_30_sp2, lhs_mat_01_30_sp2),_mm512_maddubs_epi16(rhs_mat_014589CD_31_sp2, lhs_mat_01_31_sp2));
+ __m512i iacc_mat_01_3_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_30_sp2, lhs_mat_01_30_sp2),_mm512_maddubs_epi16(rhs_mat_2367ABEF_31_sp2, lhs_mat_01_31_sp2));
+
+ __m512i iacc_mat_10_3_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_30_sp2, lhs_mat_23_30_sp2),_mm512_maddubs_epi16(rhs_mat_014589CD_31_sp2, lhs_mat_23_31_sp2));
+ __m512i iacc_mat_11_3_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_30_sp2, lhs_mat_23_30_sp2),_mm512_maddubs_epi16(rhs_mat_2367ABEF_31_sp2, lhs_mat_23_31_sp2));
+
+ __m512i iacc_mat_00_4_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_40_sp2, lhs_mat_01_40_sp2),_mm512_maddubs_epi16(rhs_mat_014589CD_41_sp2, lhs_mat_01_41_sp2));
+ __m512i iacc_mat_01_4_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_40_sp2, lhs_mat_01_40_sp2),_mm512_maddubs_epi16(rhs_mat_2367ABEF_41_sp2, lhs_mat_01_41_sp2));
+
+ __m512i iacc_mat_10_4_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_40_sp2, lhs_mat_23_40_sp2),_mm512_maddubs_epi16(rhs_mat_014589CD_41_sp2, lhs_mat_23_41_sp2));
+ __m512i iacc_mat_11_4_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_40_sp2, lhs_mat_23_40_sp2),_mm512_maddubs_epi16(rhs_mat_2367ABEF_41_sp2, lhs_mat_23_41_sp2));
+
+ __m512i iacc_mat_00_5_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_50_sp2, lhs_mat_01_50_sp2),_mm512_maddubs_epi16(rhs_mat_014589CD_51_sp2, lhs_mat_01_51_sp2));
+ __m512i iacc_mat_01_5_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_50_sp2, lhs_mat_01_50_sp2),_mm512_maddubs_epi16(rhs_mat_2367ABEF_51_sp2, lhs_mat_01_51_sp2));
+
+ __m512i iacc_mat_10_5_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_50_sp2, lhs_mat_23_50_sp2),_mm512_maddubs_epi16(rhs_mat_014589CD_51_sp2, lhs_mat_23_51_sp2));
+ __m512i iacc_mat_11_5_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_50_sp2, lhs_mat_23_50_sp2),_mm512_maddubs_epi16(rhs_mat_2367ABEF_51_sp2, lhs_mat_23_51_sp2));
+
+ __m512i iacc_mat_00_6_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_60_sp2, lhs_mat_01_60_sp2),_mm512_maddubs_epi16(rhs_mat_014589CD_61_sp2, lhs_mat_01_61_sp2));
+ __m512i iacc_mat_01_6_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_60_sp2, lhs_mat_01_60_sp2),_mm512_maddubs_epi16(rhs_mat_2367ABEF_61_sp2, lhs_mat_01_61_sp2));
+
+ __m512i iacc_mat_10_6_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_60_sp2, lhs_mat_23_60_sp2),_mm512_maddubs_epi16(rhs_mat_014589CD_61_sp2, lhs_mat_23_61_sp2));
+ __m512i iacc_mat_11_6_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_60_sp2, lhs_mat_23_60_sp2),_mm512_maddubs_epi16(rhs_mat_2367ABEF_61_sp2, lhs_mat_23_61_sp2));
+
+ __m512i iacc_mat_00_7_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_70_sp2, lhs_mat_01_70_sp2),_mm512_maddubs_epi16(rhs_mat_014589CD_71_sp2, lhs_mat_01_71_sp2));
+ __m512i iacc_mat_01_7_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_70_sp2, lhs_mat_01_70_sp2),_mm512_maddubs_epi16(rhs_mat_2367ABEF_71_sp2, lhs_mat_01_71_sp2));
+
+ __m512i iacc_mat_10_7_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_70_sp2, lhs_mat_23_70_sp2),_mm512_maddubs_epi16(rhs_mat_014589CD_71_sp2, lhs_mat_23_71_sp2));
+ __m512i iacc_mat_11_7_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_70_sp2, lhs_mat_23_70_sp2),_mm512_maddubs_epi16(rhs_mat_2367ABEF_71_sp2, lhs_mat_23_71_sp2));
+
+ // Combine results from both shuffle patterns for each output 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);
+
+ __m512i iacc_mat_00_2 = _mm512_add_epi16(iacc_mat_00_2_sp1, iacc_mat_00_2_sp2);
+ __m512i iacc_mat_01_2 = _mm512_add_epi16(iacc_mat_01_2_sp1, iacc_mat_01_2_sp2);
+ __m512i iacc_mat_10_2 = _mm512_add_epi16(iacc_mat_10_2_sp1, iacc_mat_10_2_sp2);
+ __m512i iacc_mat_11_2 = _mm512_add_epi16(iacc_mat_11_2_sp1, iacc_mat_11_2_sp2);
+
+ __m512i iacc_mat_00_3 = _mm512_add_epi16(iacc_mat_00_3_sp1, iacc_mat_00_3_sp2);
+ __m512i iacc_mat_01_3 = _mm512_add_epi16(iacc_mat_01_3_sp1, iacc_mat_01_3_sp2);
+ __m512i iacc_mat_10_3 = _mm512_add_epi16(iacc_mat_10_3_sp1, iacc_mat_10_3_sp2);
+ __m512i iacc_mat_11_3 = _mm512_add_epi16(iacc_mat_11_3_sp1, iacc_mat_11_3_sp2);
+
+ __m512i iacc_mat_00_4 = _mm512_add_epi16(iacc_mat_00_4_sp1, iacc_mat_00_4_sp2);
+ __m512i iacc_mat_01_4 = _mm512_add_epi16(iacc_mat_01_4_sp1, iacc_mat_01_4_sp2);
+ __m512i iacc_mat_10_4 = _mm512_add_epi16(iacc_mat_10_4_sp1, iacc_mat_10_4_sp2);
+ __m512i iacc_mat_11_4 = _mm512_add_epi16(iacc_mat_11_4_sp1, iacc_mat_11_4_sp2);
+
+ __m512i iacc_mat_00_5 = _mm512_add_epi16(iacc_mat_00_5_sp1, iacc_mat_00_5_sp2);
+ __m512i iacc_mat_01_5 = _mm512_add_epi16(iacc_mat_01_5_sp1, iacc_mat_01_5_sp2);
+ __m512i iacc_mat_10_5 = _mm512_add_epi16(iacc_mat_10_5_sp1, iacc_mat_10_5_sp2);
+ __m512i iacc_mat_11_5 = _mm512_add_epi16(iacc_mat_11_5_sp1, iacc_mat_11_5_sp2);
+
+ __m512i iacc_mat_00_6 = _mm512_add_epi16(iacc_mat_00_6_sp1, iacc_mat_00_6_sp2);
+ __m512i iacc_mat_01_6 = _mm512_add_epi16(iacc_mat_01_6_sp1, iacc_mat_01_6_sp2);
+ __m512i iacc_mat_10_6 = _mm512_add_epi16(iacc_mat_10_6_sp1, iacc_mat_10_6_sp2);
+ __m512i iacc_mat_11_6 = _mm512_add_epi16(iacc_mat_11_6_sp1, iacc_mat_11_6_sp2);
+
+ __m512i iacc_mat_00_7 = _mm512_add_epi16(iacc_mat_00_7_sp1, iacc_mat_00_7_sp2);
+ __m512i iacc_mat_01_7 = _mm512_add_epi16(iacc_mat_01_7_sp1, iacc_mat_01_7_sp2);
+ __m512i iacc_mat_10_7 = _mm512_add_epi16(iacc_mat_10_7_sp1, iacc_mat_10_7_sp2);
+ __m512i iacc_mat_11_7 = _mm512_add_epi16(iacc_mat_11_7_sp1, iacc_mat_11_7_sp2);
+
+ // Output of both shuffle patterns are added in order to sum dot product outputs of all 32 values in block
+ 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);
+
+ iacc_mat_00_2 = _mm512_madd_epi16(iacc_mat_00_2, scale_014589CD_2);
+ iacc_mat_01_2 = _mm512_madd_epi16(iacc_mat_01_2, scale_2367ABEF_2);
+ iacc_mat_10_2 = _mm512_madd_epi16(iacc_mat_10_2, scale_014589CD_2);
+ iacc_mat_11_2 = _mm512_madd_epi16(iacc_mat_11_2, scale_2367ABEF_2);
+
+ iacc_mat_00_3 = _mm512_madd_epi16(iacc_mat_00_3, scale_014589CD_3);
+ iacc_mat_01_3 = _mm512_madd_epi16(iacc_mat_01_3, scale_2367ABEF_3);
+ iacc_mat_10_3 = _mm512_madd_epi16(iacc_mat_10_3, scale_014589CD_3);
+ iacc_mat_11_3 = _mm512_madd_epi16(iacc_mat_11_3, scale_2367ABEF_3);
+
+ iacc_mat_00_4 = _mm512_madd_epi16(iacc_mat_00_4, scale_014589CD_4);
+ iacc_mat_01_4 = _mm512_madd_epi16(iacc_mat_01_4, scale_2367ABEF_4);
+ iacc_mat_10_4 = _mm512_madd_epi16(iacc_mat_10_4, scale_014589CD_4);
+ iacc_mat_11_4 = _mm512_madd_epi16(iacc_mat_11_4, scale_2367ABEF_4);
+
+ iacc_mat_00_5 = _mm512_madd_epi16(iacc_mat_00_5, scale_014589CD_5);
+ iacc_mat_01_5 = _mm512_madd_epi16(iacc_mat_01_5, scale_2367ABEF_5);
+ iacc_mat_10_5 = _mm512_madd_epi16(iacc_mat_10_5, scale_014589CD_5);
+ iacc_mat_11_5 = _mm512_madd_epi16(iacc_mat_11_5, scale_2367ABEF_5);
+
+ iacc_mat_00_6 = _mm512_madd_epi16(iacc_mat_00_6, scale_014589CD_6);
+ iacc_mat_01_6 = _mm512_madd_epi16(iacc_mat_01_6, scale_2367ABEF_6);
+ iacc_mat_10_6 = _mm512_madd_epi16(iacc_mat_10_6, scale_014589CD_6);
+ iacc_mat_11_6 = _mm512_madd_epi16(iacc_mat_11_6, scale_2367ABEF_6);
+
+ iacc_mat_00_7 = _mm512_madd_epi16(iacc_mat_00_7, scale_014589CD_7);
+ iacc_mat_01_7 = _mm512_madd_epi16(iacc_mat_01_7, scale_2367ABEF_7);
+ iacc_mat_10_7 = _mm512_madd_epi16(iacc_mat_10_7, scale_014589CD_7);
+ iacc_mat_11_7 = _mm512_madd_epi16(iacc_mat_11_7, scale_2367ABEF_7);
+
+ __m512i iacc_mat_00 = _mm512_add_epi32(_mm512_add_epi32(_mm512_add_epi32(iacc_mat_00_0, iacc_mat_00_1), _mm512_add_epi32(iacc_mat_00_2, iacc_mat_00_3)), _mm512_add_epi32(_mm512_add_epi32(iacc_mat_00_4, iacc_mat_00_5), _mm512_add_epi32(iacc_mat_00_6, iacc_mat_00_7)));
+ __m512i iacc_mat_01 = _mm512_add_epi32(_mm512_add_epi32(_mm512_add_epi32(iacc_mat_01_0, iacc_mat_01_1), _mm512_add_epi32(iacc_mat_01_2, iacc_mat_01_3)), _mm512_add_epi32(_mm512_add_epi32(iacc_mat_01_4, iacc_mat_01_5), _mm512_add_epi32(iacc_mat_01_6, iacc_mat_01_7)));
+ __m512i iacc_mat_10 = _mm512_add_epi32(_mm512_add_epi32(_mm512_add_epi32(iacc_mat_10_0, iacc_mat_10_1), _mm512_add_epi32(iacc_mat_10_2, iacc_mat_10_3)), _mm512_add_epi32(_mm512_add_epi32(iacc_mat_10_4, iacc_mat_10_5), _mm512_add_epi32(iacc_mat_10_6, iacc_mat_10_7)));
+ __m512i iacc_mat_11 = _mm512_add_epi32(_mm512_add_epi32(_mm512_add_epi32(iacc_mat_11_0, iacc_mat_11_1), _mm512_add_epi32(iacc_mat_11_2, iacc_mat_11_3)), _mm512_add_epi32(_mm512_add_epi32(iacc_mat_11_4, iacc_mat_11_5), _mm512_add_epi32(iacc_mat_11_6, iacc_mat_11_7)));
+
+ // Straighten out to make 4 row vectors
+ __m512i iacc_row_0 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_00, _mm512_shuffle_epi32(iacc_mat_01, (_MM_PERM_ENUM)78));
+ __m512i iacc_row_1 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_00, (_MM_PERM_ENUM)78), iacc_mat_01);
+ __m512i iacc_row_2 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_10, _mm512_shuffle_epi32(iacc_mat_11, (_MM_PERM_ENUM)78));
+ __m512i iacc_row_3 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_10, (_MM_PERM_ENUM)78), iacc_mat_11);
+
+ // Load the scale(d) values for all the 4 Q8_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]);
+
+ // Take two bsums from two Q8_Ks at a time and multiply with corresponding mins values from each Q2_K
+ __m512i iacc_row_min_0_01 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_01_0123, (_MM_PERM_ENUM)0), mins_01);
+ __m512i iacc_row_min_1_01 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_01_0123, (_MM_PERM_ENUM)170), mins_01);
+ __m512i iacc_row_min_2_01 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_23_0123, (_MM_PERM_ENUM)0), mins_01);
+ __m512i iacc_row_min_3_01 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_23_0123, (_MM_PERM_ENUM)170), mins_01);
+
+ __m512i iacc_row_min_0_23 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_01_0123, (_MM_PERM_ENUM)85), mins_23);
+ __m512i iacc_row_min_1_23 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_01_0123, (_MM_PERM_ENUM)255), mins_23);
+ __m512i iacc_row_min_2_23 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_23_0123, (_MM_PERM_ENUM)85), mins_23);
+ __m512i iacc_row_min_3_23 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_23_0123, (_MM_PERM_ENUM)255), mins_23);
+
+ __m512i iacc_row_min_0_45 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_01_4567, (_MM_PERM_ENUM)0), mins_45);
+ __m512i iacc_row_min_1_45 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_01_4567, (_MM_PERM_ENUM)170), mins_45);
+ __m512i iacc_row_min_2_45 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_23_4567, (_MM_PERM_ENUM)0), mins_45);
+ __m512i iacc_row_min_3_45 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_23_4567, (_MM_PERM_ENUM)170), mins_45);
+
+ __m512i iacc_row_min_0_67 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_01_4567, (_MM_PERM_ENUM)85), mins_67);
+ __m512i iacc_row_min_1_67 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_01_4567, (_MM_PERM_ENUM)255), mins_67);
+ __m512i iacc_row_min_2_67 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_23_4567, (_MM_PERM_ENUM)85), mins_67);
+ __m512i iacc_row_min_3_67 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_23_4567, (_MM_PERM_ENUM)255), mins_67);
+
+ __m512i iacc_row_min_0 = _mm512_add_epi32(_mm512_add_epi32(iacc_row_min_0_01, iacc_row_min_0_23), _mm512_add_epi32(iacc_row_min_0_45,iacc_row_min_0_67));
+ __m512i iacc_row_min_1 = _mm512_add_epi32(_mm512_add_epi32(iacc_row_min_1_01, iacc_row_min_1_23), _mm512_add_epi32(iacc_row_min_1_45,iacc_row_min_1_67));
+ __m512i iacc_row_min_2 = _mm512_add_epi32(_mm512_add_epi32(iacc_row_min_2_01, iacc_row_min_2_23), _mm512_add_epi32(iacc_row_min_2_45,iacc_row_min_2_67));
+ __m512i iacc_row_min_3 = _mm512_add_epi32(_mm512_add_epi32(iacc_row_min_3_01, iacc_row_min_3_23), _mm512_add_epi32(iacc_row_min_3_45,iacc_row_min_3_67));
+
+ 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_q2_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_q2_Kx8 * b_ptr_0 = b_ptr_start + ((x) * b_nb);
+ const block_q2_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++) {
+ // Delta 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 sixteen sub blocks of a super block - eight sub blocks are processed per iteration
+ for (int sb = 0; sb < QK_K / 128; sb++) {
+
+ // Load the eight block_q4_k for eight sub blocks quantized values interleaved with each other in chunks of eight bytes - B0,B1 ....B6,B7
+ 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);
+
+ //2-bit -> 8-bit
+ const __m512i rhs_mat_014589CD_00 = _mm512_and_si512(rhs_raw_mat_014589CD_0,m3bexpanded); //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,m3bexpanded); //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,m3bexpanded); //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,m3bexpanded); //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_10 = _mm512_and_si512(rhs_raw_mat_014589CD_2,m3bexpanded); //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(rhs_raw_mat_2367ABEF_2,m3bexpanded); //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(rhs_raw_mat_014589CD_3,m3bexpanded); //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(rhs_raw_mat_2367ABEF_3,m3bexpanded); //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_20 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_0, 2), m3bexpanded); //B20(0-7) B21(0-7) B24(0-7) B25(0-7) B28(0-7) B29(0-7) B2C(0-7) B2D(0-7)
+ const __m512i rhs_mat_2367ABEF_20 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_0, 2), m3bexpanded); //B22(0-7) B23(0-7) B26(0-7) B27(0-7) B2A(0-7) B2B(0-7) B2E(0-7) B2F(0-7)
+
+ const __m512i rhs_mat_014589CD_21 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_1, 2), m3bexpanded); //B20(8-15) B21(8-15) B24(8-15) B25(8-15) B28(8-15) B29(8-15) B2C(8-15) B2D(8-15)
+ const __m512i rhs_mat_2367ABEF_21 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_1, 2), m3bexpanded); //B22(8-15) B23(8-15) B26(8-15) B27(8-15) B2A(8-15) B2B(8-15) B2E(8-15) B2F(8-15)
+
+ const __m512i rhs_mat_014589CD_30 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_2, 2), m3bexpanded); //B30(0-7) B31(0-7) B34(0-7) B35(0-7) B38(0-7) B39(0-7) B3C(0-7) B3D(0-7)
+ const __m512i rhs_mat_2367ABEF_30 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_2, 2), m3bexpanded); //B32(0-7) B33(0-7) B36(0-7) B37(0-7) B3A(0-7) B3B(0-7) B3E(0-7) B3F(0-7)
+
+ const __m512i rhs_mat_014589CD_31 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_3, 2), m3bexpanded); //B30(8-15) B31(8-15) B34(8-15) B35(8-15) B38(8-15) B39(8-15) B3C(8-15) B3D(8-15)
+ const __m512i rhs_mat_2367ABEF_31 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_3, 2), m3bexpanded); //B32(8-15) B33(8-15) B36(8-15) B37(8-15) B3A(8-15) B3B(8-15) B3E(8-15) B3F(8-15)
+
+ const __m512i rhs_mat_014589CD_40 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_0, 4), m3bexpanded); //B40(0-7) B41(0-7) B44(0-7) B45(0-7) B48(0-7) B49(0-7) B4C(0-7) B4D(0-7)
+ const __m512i rhs_mat_2367ABEF_40 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_0, 4), m3bexpanded); //B42(0-7) B43(0-7) B46(0-7) B47(0-7) B4A(0-7) B4B(0-7) B4E(0-7) B4F(0-7)
+
+ const __m512i rhs_mat_014589CD_41 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_1, 4), m3bexpanded); //B40(8-15) B41(8-15) B44(8-15) B45(8-15) B48(8-15) B49(8-15) B4C(8-15) B4D(8-15)
+ const __m512i rhs_mat_2367ABEF_41 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_1, 4), m3bexpanded); //B42(8-15) B43(8-15) B46(8-15) B47(8-15) B4A(8-15) B4B(8-15) B4E(8-15) B4F(8-15)
+
+ const __m512i rhs_mat_014589CD_50 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_2, 4), m3bexpanded); //B50(0-7) B51(0-7) B54(0-7) B55(0-7) B58(0-7) B59(0-7) B5C(0-7) B5D(0-7)
+ const __m512i rhs_mat_2367ABEF_50 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_2, 4), m3bexpanded); //B52(0-7) B53(0-7) B56(0-7) B57(0-7) B5A(0-7) B5B(0-7) B5E(0-7) B5F(0-7)
+
+ const __m512i rhs_mat_014589CD_51 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_3, 4), m3bexpanded); //B50(8-15) B51(8-15) B54(8-15) B55(8-15) B58(8-15) B59(8-15) B5C(8-15) B5D(8-15)
+ const __m512i rhs_mat_2367ABEF_51 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_3, 4), m3bexpanded); //B52(8-15) B53(8-15) B56(8-15) B57(8-15) B5A(8-15) B5B(8-15) B5E(8-15) B5F(8-15)
+
+ const __m512i rhs_mat_014589CD_60 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_0, 6), m3bexpanded); //B60(0-7) B61(0-7) B64(0-7) B65(0-7) B68(0-7) B69(0-7) B6C(0-7) B6D(0-7)
+ const __m512i rhs_mat_2367ABEF_60 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_0, 6), m3bexpanded); //B62(0-7) B63(0-7) B66(0-7) B67(0-7) B6A(0-7) B6B(0-7) B6E(0-7) B6F(0-7)
+
+ const __m512i rhs_mat_014589CD_61 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_1, 6), m3bexpanded); //B60(8-15) B61(8-15) B64(8-15) B65(8-15) B68(8-15) B69(8-15) B6C(8-15) B6D(8-15)
+ const __m512i rhs_mat_2367ABEF_61 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_1, 6), m3bexpanded); //B62(8-15) B63(8-15) B66(8-15) B67(8-15) B6A(8-15) B6B(8-15) B6E(8-15) B6F(8-15)
+
+ const __m512i rhs_mat_014589CD_70 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_2, 6), m3bexpanded); //B70(0-7) B71(0-7) B74(0-7) B75(0-7) B78(0-7) B79(0-7) B7C(0-7) B7D(0-7)
+ const __m512i rhs_mat_2367ABEF_70 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_2, 6), m3bexpanded); //B72(0-7) B73(0-7) B76(0-7) B77(0-7) B7A(0-7) B7B(0-7) B7E(0-7) B7F(0-7)
+
+ const __m512i rhs_mat_014589CD_71 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_3, 6), m3bexpanded); //B70(8-15) B71(8-15) B74(8-15) B75(8-15) B78(8-15) B79(8-15) B7C(8-15) B7D(8-15)
+ const __m512i rhs_mat_2367ABEF_71 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_3, 6), m3bexpanded); //B72(8-15) B73(8-15) B76(8-15) B77(8-15) B7A(8-15) B7B(8-15) B7E(8-15) B7F(8-15)
+
+ 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_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_20_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_20, (_MM_PERM_ENUM)136); //B20(0-3) B21(0-3) B20(0-3) B21(0-3) B24(0-3) B25(0-3) B24(0-3) B25(0-3) B28(0-3) B29(0-3) B28(0-3) B29(0-3) B2C(0-3) B2D(0-3) B2C(0-3) B2D(0-3)
+ const __m512i rhs_mat_2367ABEF_20_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_20, (_MM_PERM_ENUM)136); //B22(0-3) B23(0-3) B22(0-3) B23(0-3) B26(0-3) B27(0-3) B26(0-3) B27(0-3) B2A(0-3) B2B(0-3) B2A(0-3) B2B(0-3) B2E(0-3) B2F(0-3) B2E(0-3) B2F(0-3)
+
+ const __m512i rhs_mat_014589CD_21_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_21, (_MM_PERM_ENUM)136); //B20(8-11) B21(8-11) B20(8-11) B21(8-11) B24(8-11) B25(8-11) B24(8-11) B25(8-11) B28(8-11) B29(8-11) B28(8-11) B29(8-11) B2C(8-11) B2D(8-11) B2C(8-11) B2D(8-11)
+ const __m512i rhs_mat_2367ABEF_21_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_21, (_MM_PERM_ENUM)136); //B22(8-11) B23(8-11) B22(8-11) B23(8-11) B26(8-11) B27(8-11) B26(8-11) B27(8-11) B2A(8-11) B2B(8-11) B2A(8-11) B2B(8-11) B2E(8-11) B2F(8-11) B2E(8-11) B2F(8-11)
+ const __m512i rhs_mat_014589CD_30_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_30, (_MM_PERM_ENUM)136); ///B30(0-3) B31(0-3) B30(0-3) B31(0-3) B34(0-3) B35(0-3) B34(0-3) B35(0-3) B38(0-3) B39(0-3) B38(0-3) B39(0-3) B3C(0-3) B3D(0-3) B3C(0-3) B3D(0-3)
+ const __m512i rhs_mat_2367ABEF_30_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_30, (_MM_PERM_ENUM)136); //B32(0-3) B33(0-3) B32(0-3) B33(0-3) B36(0-3) B37(0-3) B36(0-3) B37(0-3) B3A(0-3) B3B(0-3) B3A(0-3) B3B(0-3) B3E(0-3) B3F(0-3) B3E(0-3) B3F(0-3)
+
+ const __m512i rhs_mat_014589CD_31_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_31, (_MM_PERM_ENUM)136); //B30(8-11) B31(8-11) B30(8-11) B31(8-11) B34(8-11) B35(8-11) B34(8-11) B35(8-11) B38(8-11) B39(8-11) B38(8-11) B39(8-11) B3C(8-11) B3D(8-11) B3C(8-11) B3D(8-11)
+ const __m512i rhs_mat_2367ABEF_31_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_31, (_MM_PERM_ENUM)136); //B32(8-11) B33(8-11) B32(8-11) B33(8-11) B36(8-11) B37(8-11) B36(8-11) B37(8-11) B3A(8-11) B3B(8-11) B3A(8-11) B3B(8-11) B3E(8-11) B3F(8-11) B3E(8-11) B3F(8-11)
+
+ const __m512i rhs_mat_014589CD_40_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_40, (_MM_PERM_ENUM)136); //B40(0-3) B41(0-3) B40(0-3) B41(0-3) B44(0-3) B45(0-3) B44(0-3) B45(0-3) B48(0-3) B49(0-3) B48(0-3) B49(0-3) B4C(0-3) B4D(0-3) B4C(0-3) B4D(0-3)
+ const __m512i rhs_mat_2367ABEF_40_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_40, (_MM_PERM_ENUM)136); //B42(0-3) B43(0-3) B42(0-3) B43(0-3) B46(0-3) B47(0-3) B46(0-3) B47(0-3) B4A(0-3) B4B(0-3) B4A(0-3) B4B(0-3) B4E(0-3) B4F(0-3) B4E(0-3) B4F(0-3)
+
+ const __m512i rhs_mat_014589CD_41_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_41, (_MM_PERM_ENUM)136); //B40(8-11) B41(8-11) B40(8-11) B41(8-11) B44(8-11) B45(8-11) B44(8-11) B45(8-11) B48(8-11) B49(8-11) B48(8-11) B49(8-11) B4C(8-11) B4D(8-11) B4C(8-11) B4D(8-11)
+ const __m512i rhs_mat_2367ABEF_41_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_41, (_MM_PERM_ENUM)136); //B42(8-11) B43(8-11) B42(8-11) B43(8-11) B46(8-11) B47(8-11) B46(8-11) B47(8-11) B4A(8-11) B4B(8-11) B4A(8-11) B4B(8-11) B4E(8-11) B4F(8-11) B4E(8-11) B4F(8-11)
+
+ const __m512i rhs_mat_014589CD_50_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_50, (_MM_PERM_ENUM)136); //B50(0-3) B51(0-3) B50(0-3) B51(0-3) B54(0-3) B55(0-3) B54(0-3) B55(0-3) B58(0-3) B59(0-3) B58(0-3) B59(0-3) B5C(0-3) B5D(0-3) B5C(0-3) B5D(0-3)
+ const __m512i rhs_mat_2367ABEF_50_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_50, (_MM_PERM_ENUM)136); //B52(0-3) B53(0-3) B52(0-3) B53(0-3) B56(0-3) B57(0-3) B56(0-3) B57(0-3) B5A(0-3) B5B(0-3) B5A(0-3) B5B(0-3) B5E(0-3) B5F(0-3) B5E(0-3) B5F(0-3)
+
+ const __m512i rhs_mat_014589CD_51_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_51, (_MM_PERM_ENUM)136); //B50(8-11) B51(8-11) B50(8-11) B51(8-11) B54(8-11) B55(8-11) B54(8-11) B55(8-11) B58(8-11) B59(8-11) B58(8-11) B59(8-11) B5C(8-11) B5D(8-11) B5C(8-11) B5D(8-11)
+ const __m512i rhs_mat_2367ABEF_51_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_51, (_MM_PERM_ENUM)136); //B52(8-11) B53(8-11) B52(8-11) B53(8-11) B56(8-11) B57(8-11) B56(8-11) B57(8-11) B5A(8-11) B5B(8-11) B5A(8-11) B5B(8-11) B5E(8-11) B5F(8-11) B5E(8-11) B5F(8-11)
+
+ const __m512i rhs_mat_014589CD_60_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_60, (_MM_PERM_ENUM)136); //B60(0-3) B61(0-3) B60(0-3) B61(0-3) B64(0-3) B65(0-3) B64(0-3) B65(0-3) B68(0-3) B69(0-3) B68(0-3) B69(0-3) B6C(0-3) B6D(0-3) B6C(0-3) B6D(0-3)
+ const __m512i rhs_mat_2367ABEF_60_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_60, (_MM_PERM_ENUM)136); //B62(0-3) B63(0-3) B62(0-3) B63(0-3) B66(0-3) B67(0-3) B66(0-3) B67(0-3) B6A(0-3) B6B(0-3) B6A(0-3) B6B(0-3) B6E(0-3) B6F(0-3) B6E(0-3) B6F(0-3)
+
+ const __m512i rhs_mat_014589CD_61_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_61, (_MM_PERM_ENUM)136); //B60(8-11) B61(8-11) B60(8-11) B61(8-11) B64(8-11) B65(8-11) B64(8-11) B65(8-11) B68(8-11) B69(8-11) B68(8-11) B69(8-11) B6C(8-11) B6D(8-11) B6C(8-11) B6D(8-11)
+ const __m512i rhs_mat_2367ABEF_61_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_61, (_MM_PERM_ENUM)136); //B62(8-11) B63(8-11) B62(8-11) B63(8-11) B66(8-11) B67(8-11) B66(8-11) B67(8-11) B6A(8-11) B6B(8-11) B6A(8-11) B6B(8-11) B6E(8-11) B6F(8-11) B6E(8-11) B6F(8-11)
+
+ const __m512i rhs_mat_014589CD_70_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_70, (_MM_PERM_ENUM)136); //B70(0-3) B71(0-3) B70(0-3) B71(0-3) B74(0-3) B75(0-3) B74(0-3) B75(0-3) B78(0-3) B79(0-3) B78(0-3) B79(0-3) B7C(0-3) B7D(0-3) B7C(0-3) B7D(0-3)
+ const __m512i rhs_mat_2367ABEF_70_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_70, (_MM_PERM_ENUM)136); //B72(0-3) B73(0-3) B72(0-3) B73(0-3) B76(0-3) B77(0-3) B76(0-3) B77(0-3) B7A(0-3) B7B(0-3) B7A(0-3) B7B(0-3) B7E(0-3) B7F(0-3) B7E(0-3) B7F(0-3)
+
+ const __m512i rhs_mat_014589CD_71_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_71, (_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_71_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_71, (_MM_PERM_ENUM)136); //B72(8-11) B73(8-11) B72(8-11) B73(8-11) B76(8-11) B77(8-11) B76(8-11) B77(8-11) B7A(8-11) B7B(8-11) B7A(8-11) B7B(8-11) B7E(8-11) B7F(8-11) B7E(8-11) B7F(8-11)
+
+ 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_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_20_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_20, (_MM_PERM_ENUM)221); //B20(4-7) B21(4-7) B20(4-7) B21(4-7) B24(4-7) B25(4-7) B24(4-7) B25(4-7) B28(4-7) B29(4-7) B28(4-7) B29(4-7) B2C(4-7) B2D(4-7) B2C(4-7) B2D(4-7)
+ const __m512i rhs_mat_2367ABEF_20_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_20, (_MM_PERM_ENUM)221); //B22(4-7) B23(4-7) B22(4-7) B23(4-7) B26(4-7) B27(4-7) B26(4-7) B27(4-7) B2A(4-7) B2B(4-7) B2A(4-7) B2B(4-7) B2E(4-7) B2F(4-7) B2E(4-7) B2F(4-7)
+
+ const __m512i rhs_mat_014589CD_21_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_21, (_MM_PERM_ENUM)221); //B20(12-15) B21(12-15) B20(12-15) B21(12-15) B24(12-15) B25(12-15) B24(12-15) B25(12-15) B28(12-15) B29(12-15) B28(12-15) B29(12-15) B2C(12-15) B2D(12-15) B2C(12-15) B2D(12-15)
+ const __m512i rhs_mat_2367ABEF_21_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_21, (_MM_PERM_ENUM)221); //B22(12-15) B23(12-15) B22(12-15) B23(12-15) B26(12-15) B27(12-15) B26(12-15) B27(12-15) B2A(12-15) B2B(12-15) B2A(12-15) B2B(12-15) B2E(12-15) B2F(12-15) B2E(12-15) B2F(12-15)
+
+ const __m512i rhs_mat_014589CD_30_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_30, (_MM_PERM_ENUM)221); //B30(4-7) B31(4-7) B30(4-7) B31(4-7) B34(4-7) B35(4-7) B34(4-7) B35(4-7) B38(4-7) B39(4-7) B38(4-7) B39(4-7) B3C(4-7) B3D(4-7) B3C(4-7) B3D(4-7)
+ const __m512i rhs_mat_2367ABEF_30_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_30, (_MM_PERM_ENUM)221); //B32(4-7) B33(4-7) B32(4-7) B33(4-7) B36(4-7) B37(4-7) B36(4-7) B37(4-7) B3A(4-7) B3B(4-7) B3A(4-7) B3B(4-7) B3E(4-7) B3F(4-7) B3E(4-7) B3F(4-7)
+
+ const __m512i rhs_mat_014589CD_31_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_31, (_MM_PERM_ENUM)221); //B30(12-15) B31(12-15) B30(12-15) B31(12-15) B34(12-15) B35(12-15) B34(12-15) B35(12-15) B38(12-15) B39(12-15) B38(12-15) B39(12-15) B3C(12-15) B3D(12-15) B3C(12-15) B3D(12-15)
+ const __m512i rhs_mat_2367ABEF_31_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_31, (_MM_PERM_ENUM)221); //B32(12-15) B33(12-15) B32(12-15) B33(12-15) B36(12-15) B37(12-15) B36(12-15) B37(12-15) B3A(12-15) B3B(12-15) B3A(12-15) B3B(12-15) B3E(12-15) B3F(12-15) B3E(12-15) B3F(12-15)
+
+ const __m512i rhs_mat_014589CD_40_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_40, (_MM_PERM_ENUM)221); //B40(4-7) B41(4-7) B40(4-7) B41(4-7) B44(4-7) B45(4-7) B44(4-7) B45(4-7) B48(4-7) B49(4-7) B48(4-7) B49(4-7) B4C(4-7) B4D(4-7) B4C(4-7) B4D(4-7)
+ const __m512i rhs_mat_2367ABEF_40_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_40, (_MM_PERM_ENUM)221); //B42(4-7) B43(4-7) B42(4-7) B43(4-7) B46(4-7) B47(4-7) B46(4-7) B47(4-7) B4A(4-7) B4B(4-7) B4A(4-7) B4B(4-7) B4E(4-7) B4F(4-7) B4E(4-7) B4F(4-7)
+
+ const __m512i rhs_mat_014589CD_41_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_41, (_MM_PERM_ENUM)221); //B40(12-15) B41(12-15) B40(12-15) B41(12-15) B44(12-15) B45(12-15) B44(12-15) B45(12-15) B48(12-15) B49(12-15) B48(12-15) B49(12-15) B4C(12-15) B4D(12-15) B4C(12-15) B4D(12-15)
+ const __m512i rhs_mat_2367ABEF_41_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_41, (_MM_PERM_ENUM)221); //B42(12-15) B43(12-15) B42(12-15) B43(12-15) B46(12-15) B47(12-15) B46(12-15) B47(12-15) B4A(12-15) B4B(12-15) B4A(12-15) B4B(12-15) B4E(12-15) B4F(12-15) B4E(12-15) B4F(12-15)
+
+ const __m512i rhs_mat_014589CD_50_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_50, (_MM_PERM_ENUM)221); //B50(4-7) B51(4-7) B50(4-7) B51(4-7) B54(4-7) B55(4-7) B54(4-7) B55(4-7) B58(4-7) B59(4-7) B58(4-7) B59(4-7) B5C(4-7) B5D(4-7) B5C(4-7) B5D(4-7)
+ const __m512i rhs_mat_2367ABEF_50_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_50, (_MM_PERM_ENUM)221); //B52(4-7) B53(4-7) B52(4-7) B53(4-7) B56(4-7) B57(4-7) B56(4-7) B57(4-7) B5A(4-7) B5B(4-7) B5A(4-7) B5B(4-7) B5E(4-7) B5F(4-7) B5E(4-7) B5F(4-7)
+
+ const __m512i rhs_mat_014589CD_51_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_51, (_MM_PERM_ENUM)221); //B50(12-15) B51(12-15) B50(12-15) B51(12-15) B54(12-15) B55(12-15) B54(12-15) B55(12-15) B58(12-15) B59(12-15) B58(12-15) B59(12-15) B5C(12-15) B5D(12-15) B5C(12-15) B5D(12-15)
+ const __m512i rhs_mat_2367ABEF_51_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_51, (_MM_PERM_ENUM)221); //B52(12-15) B53(12-15) B52(12-15) B53(12-15) B56(12-15) B57(12-15) B56(12-15) B57(12-15) B5A(12-15) B5B(12-15) B5A(12-15) B5B(12-15) B5E(12-15) B5F(12-15) B5E(12-15) B5F(12-15)
+
+ const __m512i rhs_mat_014589CD_60_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_60, (_MM_PERM_ENUM)221); //B60(4-7) B61(4-7) B60(4-7) B61(4-7) B64(4-7) B65(4-7) B64(4-7) B65(4-7) B68(4-7) B69(4-7) B68(4-7) B69(4-7) B6C(4-7) B6D(4-7) B6C(4-7) B6D(4-7)
+ const __m512i rhs_mat_2367ABEF_60_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_60, (_MM_PERM_ENUM)221); //B62(4-7) B63(4-7) B62(4-7) B63(4-7) B66(4-7) B67(4-7) B66(4-7) B67(4-7) B6A(4-7) B6B(4-7) B6A(4-7) B6B(4-7) B6E(4-7) B6F(4-7) B6E(4-7) B6F(4-7)
+
+ const __m512i rhs_mat_014589CD_61_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_61, (_MM_PERM_ENUM)221); //B60(12-15) B61(12-15) B60(12-15) B61(12-15) B64(12-15) B65(12-15) B64(12-15) B65(12-15) B68(12-15) B69(12-15) B68(12-15) B69(12-15) B6C(12-15) B6D(12-15) B6C(12-15) B6D(12-15)
+ const __m512i rhs_mat_2367ABEF_61_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_61, (_MM_PERM_ENUM)221); //B62(12-15) B63(12-15) B62(12-15) B63(12-15) B66(12-15) B67(12-15) B66(12-15) B67(12-15) B6A(12-15) B6B(12-15) B6A(12-15) B6B(12-15) B6E(12-15) B6F(12-15) B6E(12-15) B6F(12-15)
+
+ const __m512i rhs_mat_014589CD_70_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_70, (_MM_PERM_ENUM)221); //B70(4-7) B71(4-7) B70(4-7) B71(4-7) B74(4-7) B75(4-7) B74(4-7) B75(4-7) B78(4-7) B79(4-7) B78(4-7) B79(4-7) B7C(4-7) B7D(4-7) B7C(4-7) B7D(4-7)
+ const __m512i rhs_mat_2367ABEF_70_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_70, (_MM_PERM_ENUM)221); //B72(4-7) B73(4-7) B72(4-7) B73(4-7) B76(4-7) B77(4-7) B76(4-7) B77(4-7) B7A(4-7) B7B(4-7) B7A(4-7) B7B(4-7) B7E(4-7) B7F(4-7) B7E(4-7) B7F(4-7)
+
+ const __m512i rhs_mat_014589CD_71_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_71, (_MM_PERM_ENUM)221); //B70(12-15) B71(12-15) B70(12-15) B71(12-15) B74(12-15) B75(12-15) B74(12-15) B75(12-15) B78(12-15) B79(12-15) B78(12-15) B79(12-15) B7C(12-15) B7D(12-15) B7C(12-15) B7D(12-15)
+ const __m512i rhs_mat_2367ABEF_71_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_71, (_MM_PERM_ENUM)221); //B72(12-15) B73(12-15) B72(12-15) B73(12-15) B76(12-15) B77(12-15) B76(12-15) B77(12-15) B7A(12-15) B7B(12-15) B7A(12-15) B7B(12-15) B7E(12-15) B7F(12-15) B7E(12-15) B7F(12-15)
+
+ //notation:superblock subblock
+ //s00 m00 s01 m01 s10 m10 s11 m11 s20 m20 s21 m21 s30 m30 s31 m31 s40 m40 s41 m41 s50 m50 s51 m51 s60 m60 s61 m61 s70 m70 s71 m71
+
+ const __m128i mins_and_scales_01_0 = _mm_loadu_si128((const __m128i *)(b_ptr_0[b].scales + sb * 64));
+ const __m128i mins_and_scales_23_0 = _mm_loadu_si128((const __m128i *)(b_ptr_0[b].scales + 16 + sb * 64));
+ const __m128i mins_and_scales_45_0 = _mm_loadu_si128((const __m128i *)(b_ptr_0[b].scales + 32 + sb * 64));
+ const __m128i mins_and_scales_67_0 = _mm_loadu_si128((const __m128i *)(b_ptr_0[b].scales + 48 + sb * 64));
+
+ const __m128i mins_and_scales_01_1 = _mm_loadu_si128((const __m128i *)(b_ptr_1[b].scales + sb * 64));
+ const __m128i mins_and_scales_23_1 = _mm_loadu_si128((const __m128i *)(b_ptr_1[b].scales + 16 + sb * 64));
+ const __m128i mins_and_scales_45_1 = _mm_loadu_si128((const __m128i *)(b_ptr_1[b].scales + 32 + sb * 64));
+ const __m128i mins_and_scales_67_1 = _mm_loadu_si128((const __m128i *)(b_ptr_1[b].scales + 48 + sb * 64));
+
+ // Combine mins and scales for sub-blocks: 0-1, 2-3, 4-5, 6-7 in the sb loop
+ const __m256i mins_and_scales_01 = _mm256_insertf128_si256(_mm256_castsi128_si256(mins_and_scales_01_0), mins_and_scales_01_1, 1);
+ const __m256i mins_and_scales_23 = _mm256_insertf128_si256(_mm256_castsi128_si256(mins_and_scales_23_0), mins_and_scales_23_1, 1);
+ const __m256i mins_and_scales_45 = _mm256_insertf128_si256(_mm256_castsi128_si256(mins_and_scales_45_0), mins_and_scales_45_1, 1);
+ const __m256i mins_and_scales_67 = _mm256_insertf128_si256(_mm256_castsi128_si256(mins_and_scales_67_0), mins_and_scales_67_1, 1);
+
+ // Extract scales which is lower half from mins_and_scales
+ const __m256i scales_01 = _mm256_and_si256(mins_and_scales_01, m4b);
+ const __m256i scales_23 = _mm256_and_si256(mins_and_scales_23, m4b);
+ const __m256i scales_45 = _mm256_and_si256(mins_and_scales_45, m4b);
+ const __m256i scales_67 = _mm256_and_si256(mins_and_scales_67, m4b);
+
+ // Extract mins which is upper half from mins_and_scales
+ const __m512i mins_01 = _mm512_cvtepu8_epi16(_mm256_and_si256(_mm256_srli_epi16(mins_and_scales_01, 4), m4b));
+ const __m512i mins_23 = _mm512_cvtepu8_epi16(_mm256_and_si256(_mm256_srli_epi16(mins_and_scales_23, 4), m4b));
+ const __m512i mins_45 = _mm512_cvtepu8_epi16(_mm256_and_si256(_mm256_srli_epi16(mins_and_scales_45, 4), m4b));
+ const __m512i mins_67 = _mm512_cvtepu8_epi16(_mm256_and_si256(_mm256_srli_epi16(mins_and_scales_67, 4), m4b));
+
+ const __m512i scales_0 = _mm512_cvtepu8_epi16(_mm256_shuffle_epi8(scales_01, scalesmask1));
+ const __m512i scales_1 = _mm512_cvtepu8_epi16(_mm256_shuffle_epi8(scales_01, scalesmask2));
+ const __m512i scales_2 = _mm512_cvtepu8_epi16(_mm256_shuffle_epi8(scales_23, scalesmask1));
+ const __m512i scales_3 = _mm512_cvtepu8_epi16(_mm256_shuffle_epi8(scales_23, scalesmask2));
+ const __m512i scales_4 = _mm512_cvtepu8_epi16(_mm256_shuffle_epi8(scales_45, scalesmask1));
+ const __m512i scales_5 = _mm512_cvtepu8_epi16(_mm256_shuffle_epi8(scales_45, scalesmask2));
+ const __m512i scales_6 = _mm512_cvtepu8_epi16(_mm256_shuffle_epi8(scales_67, scalesmask1));
+ const __m512i scales_7 = _mm512_cvtepu8_epi16(_mm256_shuffle_epi8(scales_67, scalesmask2));
+
+ 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);
+
+ const __m512i scale_014589CD_2 = _mm512_shuffle_epi32(scales_2, (_MM_PERM_ENUM)68);
+ const __m512i scale_2367ABEF_2 = _mm512_shuffle_epi32(scales_2, (_MM_PERM_ENUM)238);
+
+ const __m512i scale_014589CD_3 = _mm512_shuffle_epi32(scales_3, (_MM_PERM_ENUM)68);
+ const __m512i scale_2367ABEF_3 = _mm512_shuffle_epi32(scales_3, (_MM_PERM_ENUM)238);
+
+ const __m512i scale_014589CD_4 = _mm512_shuffle_epi32(scales_4, (_MM_PERM_ENUM)68);
+ const __m512i scale_2367ABEF_4 = _mm512_shuffle_epi32(scales_4, (_MM_PERM_ENUM)238);
+
+ const __m512i scale_014589CD_5 = _mm512_shuffle_epi32(scales_5, (_MM_PERM_ENUM)68);
+ const __m512i scale_2367ABEF_5 = _mm512_shuffle_epi32(scales_5, (_MM_PERM_ENUM)238);
+
+ const __m512i scale_014589CD_6 = _mm512_shuffle_epi32(scales_6, (_MM_PERM_ENUM)68);
+ const __m512i scale_2367ABEF_6 = _mm512_shuffle_epi32(scales_6, (_MM_PERM_ENUM)238);
+
+ const __m512i scale_014589CD_7 = _mm512_shuffle_epi32(scales_7, (_MM_PERM_ENUM)68);
+ const __m512i scale_2367ABEF_7 = _mm512_shuffle_epi32(scales_7, (_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 + 512 * 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 + 512 * 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_10 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 64 + 512 * 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 + 96 + 512 * 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_20 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 128 + 512 * sb)));
+ __m256i lhs_mat_ymm_01_20 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_20, lhs_mat_ymm_0123_20, 0);
+ __m256i lhs_mat_ymm_23_20 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_20, lhs_mat_ymm_0123_20, 17);
+ __m256i lhs_mat_ymm_0123_21 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 160 + 512 * sb)));
+ __m256i lhs_mat_ymm_01_21 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_21, lhs_mat_ymm_0123_21, 0);
+ __m256i lhs_mat_ymm_23_21 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_21, lhs_mat_ymm_0123_21, 17);
+ __m256i lhs_mat_ymm_0123_30 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 192 + 512 * sb)));
+ __m256i lhs_mat_ymm_01_30 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_30, lhs_mat_ymm_0123_30, 0);
+ __m256i lhs_mat_ymm_23_30 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_30, lhs_mat_ymm_0123_30, 17);
+ __m256i lhs_mat_ymm_0123_31 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 224 + 512 * sb)));
+ __m256i lhs_mat_ymm_01_31 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_31, lhs_mat_ymm_0123_31, 0);
+ __m256i lhs_mat_ymm_23_31 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_31, lhs_mat_ymm_0123_31, 17);
+
+ __m256i lhs_mat_ymm_0123_40 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 256 + 512 * sb)));
+ __m256i lhs_mat_ymm_01_40 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_40, lhs_mat_ymm_0123_40, 0);
+ __m256i lhs_mat_ymm_23_40 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_40, lhs_mat_ymm_0123_40, 17);
+ __m256i lhs_mat_ymm_0123_41 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 288 + 512 * sb)));
+ __m256i lhs_mat_ymm_01_41 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_41, lhs_mat_ymm_0123_41, 0);
+ __m256i lhs_mat_ymm_23_41 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_41, lhs_mat_ymm_0123_41, 17);
+ __m256i lhs_mat_ymm_0123_50 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 320 + 512 * sb)));
+ __m256i lhs_mat_ymm_01_50 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_50, lhs_mat_ymm_0123_50, 0);
+ __m256i lhs_mat_ymm_23_50 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_50, lhs_mat_ymm_0123_50, 17);
+ __m256i lhs_mat_ymm_0123_51 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 352 + 512 * sb)));
+ __m256i lhs_mat_ymm_01_51 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_51, lhs_mat_ymm_0123_51, 0);
+ __m256i lhs_mat_ymm_23_51 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_51, lhs_mat_ymm_0123_51, 17);
+ __m256i lhs_mat_ymm_0123_60 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 384 + 512 * sb)));
+ __m256i lhs_mat_ymm_01_60 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_60, lhs_mat_ymm_0123_60, 0);
+ __m256i lhs_mat_ymm_23_60 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_60, lhs_mat_ymm_0123_60, 17);
+ __m256i lhs_mat_ymm_0123_61 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 416 + 512 * sb)));
+ __m256i lhs_mat_ymm_01_61 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_61, lhs_mat_ymm_0123_61, 0);
+ __m256i lhs_mat_ymm_23_61 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_61, lhs_mat_ymm_0123_61, 17);
+ __m256i lhs_mat_ymm_0123_70 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 448 + 512 * sb)));
+ __m256i lhs_mat_ymm_01_70 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_70, lhs_mat_ymm_0123_70, 0);
+ __m256i lhs_mat_ymm_23_70 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_70, lhs_mat_ymm_0123_70, 17);
+ __m256i lhs_mat_ymm_0123_71 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 480 + 512 * sb)));
+ __m256i lhs_mat_ymm_01_71 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_71, lhs_mat_ymm_0123_71, 0);
+ __m256i lhs_mat_ymm_23_71 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_71, lhs_mat_ymm_0123_71, 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_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_20 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_20), lhs_mat_ymm_01_20, 1);
+ __m512i lhs_mat_23_20 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_20), lhs_mat_ymm_23_20, 1);
+ __m512i lhs_mat_01_21 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_21), lhs_mat_ymm_01_21, 1);
+ __m512i lhs_mat_23_21 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_21), lhs_mat_ymm_23_21, 1);
+
+ __m512i lhs_mat_01_30 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_30), lhs_mat_ymm_01_30, 1);
+ __m512i lhs_mat_23_30 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_30), lhs_mat_ymm_23_30, 1);
+ __m512i lhs_mat_01_31 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_31), lhs_mat_ymm_01_31, 1);
+ __m512i lhs_mat_23_31 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_31), lhs_mat_ymm_23_31, 1);
+
+ __m512i lhs_mat_01_40 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_40), lhs_mat_ymm_01_40, 1);
+ __m512i lhs_mat_23_40 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_40), lhs_mat_ymm_23_40, 1);
+ __m512i lhs_mat_01_41 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_41), lhs_mat_ymm_01_41, 1);
+ __m512i lhs_mat_23_41 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_41), lhs_mat_ymm_23_41, 1);
+
+ __m512i lhs_mat_01_50 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_50), lhs_mat_ymm_01_50, 1);
+ __m512i lhs_mat_23_50 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_50), lhs_mat_ymm_23_50, 1);
+ __m512i lhs_mat_01_51 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_51), lhs_mat_ymm_01_51, 1);
+ __m512i lhs_mat_23_51 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_51), lhs_mat_ymm_23_51, 1);
+
+ __m512i lhs_mat_01_60 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_60), lhs_mat_ymm_01_60, 1);
+ __m512i lhs_mat_23_60 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_60), lhs_mat_ymm_23_60, 1);
+ __m512i lhs_mat_01_61 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_61), lhs_mat_ymm_01_61, 1);
+ __m512i lhs_mat_23_61 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_61), lhs_mat_ymm_23_61, 1);
+
+ __m512i lhs_mat_01_70 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_70), lhs_mat_ymm_01_70, 1);
+ __m512i lhs_mat_23_70 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_70), lhs_mat_ymm_23_70, 1);
+ __m512i lhs_mat_01_71 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_71), lhs_mat_ymm_01_71, 1);
+ __m512i lhs_mat_23_71 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_71), lhs_mat_ymm_23_71, 1);
+
+ // Bsums are loaded for the different Q8_K blocks
+ __m128i lhs_raw_bsums_01_0123 = _mm_loadu_si128((const __m128i *)((a_ptr[b].bsums + 32 * sb)));
+ __m128i lhs_raw_bsums_23_0123 = _mm_loadu_si128((const __m128i *)(a_ptr[b].bsums + 8 + 32 * sb));
+ __m128i lhs_raw_bsums_01_4567 = _mm_loadu_si128((const __m128i *)((a_ptr[b].bsums + 16 + 32 * sb)));
+ __m128i lhs_raw_bsums_23_4567 = _mm_loadu_si128((const __m128i *)(a_ptr[b].bsums + 24 + 32 * sb));
+
+ __m256i lhs_bsums_ymm_01_0123 = _mm256_inserti128_si256(_mm256_castsi128_si256(lhs_raw_bsums_01_0123), lhs_raw_bsums_01_0123, 1);
+ __m512i lhs_bsums_01_0123 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_bsums_ymm_01_0123), lhs_bsums_ymm_01_0123, 1);
+ __m256i lhs_bsums_ymm_23_0123 = _mm256_inserti128_si256(_mm256_castsi128_si256(lhs_raw_bsums_23_0123), lhs_raw_bsums_23_0123, 1);
+ __m512i lhs_bsums_23_0123 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_bsums_ymm_23_0123), lhs_bsums_ymm_23_0123, 1);
+ __m256i lhs_bsums_ymm_01_4567 = _mm256_inserti128_si256(_mm256_castsi128_si256(lhs_raw_bsums_01_4567), lhs_raw_bsums_01_4567, 1);
+ __m512i lhs_bsums_01_4567 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_bsums_ymm_01_4567), lhs_bsums_ymm_01_4567, 1);
+ __m256i lhs_bsums_ymm_23_4567 = _mm256_inserti128_si256(_mm256_castsi128_si256(lhs_raw_bsums_23_4567), lhs_raw_bsums_23_4567, 1);
+ __m512i lhs_bsums_23_4567 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_bsums_ymm_23_4567), lhs_bsums_ymm_23_4567, 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_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_20_sp1 = _mm512_shuffle_epi32(lhs_mat_01_20, (_MM_PERM_ENUM)160); //A20(0-3) A20(0-3) A21(0-3) A21(0-3) A20(0-3) A20(0-3) A21(0-3) A21(0-3) A20(0-3) A20(0-3) A21(0-3) A21(0-3) A20(0-3) A20(0-3) A21(0-3) A21(0-3)
+ const __m512i lhs_mat_23_20_sp1 = _mm512_shuffle_epi32(lhs_mat_23_20, (_MM_PERM_ENUM)160); //A22(0-3) A22(0-3) A23(0-3) A23(0-3) A22(0-3) A22(0-3) A23(0-3) A23(0-3) A22(0-3) A22(0-3) A23(0-3) A23(0-3) A22(0-3) A22(0-3) A23(0-3) A23(0-3)
+
+ const __m512i lhs_mat_01_21_sp1 = _mm512_shuffle_epi32(lhs_mat_01_21, (_MM_PERM_ENUM)160); //A20(8-11) A20(8-11) A21(8-11) A21(8-11) A20(8-11) A20(8-11) A21(8-11) A21(8-11) A20(8-11) A20(8-11) A21(8-11) A21(8-11) A20(8-11) A20(8-11) A21(8-11) A21(8-11)
+ const __m512i lhs_mat_23_21_sp1 = _mm512_shuffle_epi32(lhs_mat_23_21, (_MM_PERM_ENUM)160); //A22(8-11) A22(8-11) A23(8-11) A23(8-11) A22(8-11) A22(8-11) A23(8-11) A23(8-11) A22(8-11) A22(8-11) A23(8-11) A23(8-11) A22(8-11) A22(8-11) A23(8-11) A23(8-11)
+
+ const __m512i lhs_mat_01_30_sp1 = _mm512_shuffle_epi32(lhs_mat_01_30, (_MM_PERM_ENUM)160); //A30(0-3) A30(0-3) A31(0-3) A31(0-3) A30(0-3) A30(0-3) A31(0-3) A31(0-3) A30(0-3) A30(0-3) A31(0-3) A31(0-3) A30(0-3) A30(0-3) A31(0-3) A31(0-3)
+ const __m512i lhs_mat_23_30_sp1 = _mm512_shuffle_epi32(lhs_mat_23_30, (_MM_PERM_ENUM)160); //A32(0-3) A32(0-3) A33(0-3) A33(0-3) A32(0-3) A32(0-3) A33(0-3) A33(0-3) A32(0-3) A32(0-3) A33(0-3) A33(0-3) A32(0-3) A32(0-3) A33(0-3) A33(0-3)
+
+ const __m512i lhs_mat_01_31_sp1 = _mm512_shuffle_epi32(lhs_mat_01_31, (_MM_PERM_ENUM)160); //A30(8-11) A30(8-11) A31(8-11) A31(8-11) A30(8-11) A30(8-11) A31(8-11) A31(8-11) A30(8-11) A30(8-11) A31(8-11) A31(8-11) A30(8-11) A30(8-11) A31(8-11) A31(8-11)
+ const __m512i lhs_mat_23_31_sp1 = _mm512_shuffle_epi32(lhs_mat_23_31, (_MM_PERM_ENUM)160); //A32(8-11) A32(8-11) A33(8-11) A33(8-11) A32(8-11) A32(8-11) A33(8-11) A33(8-11) A32(8-11) A32(8-11) A33(8-11) A33(8-11) A32(8-11) A32(8-11) A33(8-11) A33(8-11)
+
+ const __m512i lhs_mat_01_40_sp1 = _mm512_shuffle_epi32(lhs_mat_01_40, (_MM_PERM_ENUM)160); //A40(0-3) A40(0-3) A41(0-3) A41(0-3) A40(0-3) A40(0-3) A41(0-3) A41(0-3) A40(0-3) A40(0-3) A41(0-3) A41(0-3) A40(0-3) A40(0-3) A41(0-3) A41(0-3)
+ const __m512i lhs_mat_23_40_sp1 = _mm512_shuffle_epi32(lhs_mat_23_40, (_MM_PERM_ENUM)160); //A42(0-3) A42(0-3) A43(0-3) A43(0-3) A42(0-3) A42(0-3) A43(0-3) A43(0-3) A42(0-3) A42(0-3) A43(0-3) A43(0-3) A42(0-3) A42(0-3) A43(0-3) A43(0-3)
+
+ const __m512i lhs_mat_01_41_sp1 = _mm512_shuffle_epi32(lhs_mat_01_41, (_MM_PERM_ENUM)160); //A40(8-11) A40(8-11) A41(8-11) A41(8-11) A40(8-11) A40(8-11) A41(8-11) A41(8-11) A40(8-11) A40(8-11) A41(8-11) A41(8-11) A40(8-11) A40(8-11) A41(8-11) A41(8-11)
+ const __m512i lhs_mat_23_41_sp1 = _mm512_shuffle_epi32(lhs_mat_23_41, (_MM_PERM_ENUM)160); //A42(8-11) A42(8-11) A43(8-11) A43(8-11) A42(8-11) A42(8-11) A43(8-11) A43(8-11) A42(8-11) A42(8-11) A43(8-11) A43(8-11) A42(8-11) A42(8-11) A43(8-11) A43(8-11)
+
+ const __m512i lhs_mat_01_50_sp1 = _mm512_shuffle_epi32(lhs_mat_01_50, (_MM_PERM_ENUM)160); //A50(0-3) A50(0-3) A51(0-3) A51(0-3) A50(0-3) A50(0-3) A51(0-3) A51(0-3) A50(0-3) A50(0-3) A51(0-3) A51(0-3) A50(0-3) A50(0-3) A51(0-3) A51(0-3)
+ const __m512i lhs_mat_23_50_sp1 = _mm512_shuffle_epi32(lhs_mat_23_50, (_MM_PERM_ENUM)160); //A52(0-3) A52(0-3) A53(0-3) A53(0-3) A52(0-3) A52(0-3) A53(0-3) A53(0-3) A52(0-3) A52(0-3) A53(0-3) A53(0-3) A52(0-3) A52(0-3) A53(0-3) A53(0-3)
+
+ const __m512i lhs_mat_01_51_sp1 = _mm512_shuffle_epi32(lhs_mat_01_51, (_MM_PERM_ENUM)160); //A50(8-11) A50(8-11) A51(8-11) A51(8-11) A50(8-11) A50(8-11) A51(8-11) A51(8-11) A50(8-11) A50(8-11) A51(8-11) A51(8-11) A50(8-11) A50(8-11) A51(8-11) A51(8-11)
+ const __m512i lhs_mat_23_51_sp1 = _mm512_shuffle_epi32(lhs_mat_23_51, (_MM_PERM_ENUM)160); //A52(8-11) A52(8-11) A53(8-11) A53(8-11) A52(8-11) A52(8-11) A53(8-11) A53(8-11) A52(8-11) A52(8-11) A53(8-11) A53(8-11) A52(8-11) A52(8-11) A53(8-11) A53(8-11)
+
+ const __m512i lhs_mat_01_60_sp1 = _mm512_shuffle_epi32(lhs_mat_01_60, (_MM_PERM_ENUM)160); //A60(0-3) A60(0-3) A61(0-3) A61(0-3) A60(0-3) A60(0-3) A61(0-3) A61(0-3) A60(0-3) A60(0-3) A61(0-3) A61(0-3) A60(0-3) A60(0-3) A61(0-3) A61(0-3)
+ const __m512i lhs_mat_23_60_sp1 = _mm512_shuffle_epi32(lhs_mat_23_60, (_MM_PERM_ENUM)160); //A62(0-3) A62(0-3) A63(0-3) A63(0-3) A62(0-3) A62(0-3) A63(0-3) A63(0-3) A62(0-3) A62(0-3) A63(0-3) A63(0-3) A62(0-3) A62(0-3) A63(0-3) A63(0-3)
+
+ const __m512i lhs_mat_01_61_sp1 = _mm512_shuffle_epi32(lhs_mat_01_61, (_MM_PERM_ENUM)160); //A60(8-11) A60(8-11) A61(8-11) A61(8-11) A60(8-11) A60(8-11) A61(8-11) A61(8-11) A60(8-11) A60(8-11) A61(8-11) A61(8-11) A60(8-11) A60(8-11) A61(8-11) A61(8-11)
+ const __m512i lhs_mat_23_61_sp1 = _mm512_shuffle_epi32(lhs_mat_23_61, (_MM_PERM_ENUM)160); //A62(8-11) A62(8-11) A63(8-11) A63(8-11) A62(8-11) A62(8-11) A63(8-11) A63(8-11) A62(8-11) A62(8-11) A63(8-11) A63(8-11) A62(8-11) A62(8-11) A63(8-11) A63(8-11)
+
+ const __m512i lhs_mat_01_70_sp1 = _mm512_shuffle_epi32(lhs_mat_01_70, (_MM_PERM_ENUM)160); //A70(0-3) A70(0-3) A71(0-3) A71(0-3) A70(0-3) A70(0-3) A71(0-3) A71(0-3) A70(0-3) A70(0-3) A71(0-3) A71(0-3) A70(0-3) A70(0-3) A71(0-3) A71(0-3)
+ const __m512i lhs_mat_23_70_sp1 = _mm512_shuffle_epi32(lhs_mat_23_70, (_MM_PERM_ENUM)160); //A72(0-3) A72(0-3) A73(0-3) A73(0-3) A72(0-3) A72(0-3) A73(0-3) A73(0-3) A72(0-3) A72(0-3) A73(0-3) A73(0-3) A72(0-3) A72(0-3) A73(0-3) A73(0-3)
+
+ const __m512i lhs_mat_01_71_sp1 = _mm512_shuffle_epi32(lhs_mat_01_71, (_MM_PERM_ENUM)160); //A70(8-11) A70(8-11) A71(8-11) A71(8-11) A70(8-11) A70(8-11) A71(8-11) A71(8-11) A70(8-11) A70(8-11) A71(8-11) A71(8-11) A70(8-11) A70(8-11) A71(8-11) A71(8-11)
+ const __m512i lhs_mat_23_71_sp1 = _mm512_shuffle_epi32(lhs_mat_23_71, (_MM_PERM_ENUM)160); //A72(8-11) A72(8-11) A73(8-11) A73(8-11) A72(8-11) A72(8-11) A73(8-11) A73(8-11) A72(8-11) A72(8-11) A73(8-11) A73(8-11) A72(8-11) A72(8-11) A73(8-11) A73(8-11)
+
+ 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_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_20_sp2 = _mm512_shuffle_epi32(lhs_mat_01_20, (_MM_PERM_ENUM)245); //A20(4-7) A20(4-7) A21(4-7) A21(4-7) A20(4-7) A20(4-7) A21(4-7) A21(4-7) A20(4-7) A20(4-7) A21(4-7) A21(4-7) A20(4-7) A20(4-7) A21(4-7) A21(4-7)
+ const __m512i lhs_mat_23_20_sp2 = _mm512_shuffle_epi32(lhs_mat_23_20, (_MM_PERM_ENUM)245); //A22(4-7) A22(4-7) A23(4-7) A23(4-7) A22(4-7) A22(4-7) A23(4-7) A23(4-7) A22(4-7) A22(4-7) A23(4-7) A23(4-7) A22(4-7) A22(4-7) A23(4-7) A23(4-7)
+
+ const __m512i lhs_mat_01_21_sp2 = _mm512_shuffle_epi32(lhs_mat_01_21, (_MM_PERM_ENUM)245); //A20(12-15) A20(12-15) A21(12-15) A21(12-15) A20(12-15) A20(12-15) A21(12-15) A21(12-15) A20(12-15) A20(12-15) A21(12-15) A21(12-15) A20(12-15) A20(12-15) A21(12-15) A21(12-15)
+ const __m512i lhs_mat_23_21_sp2 = _mm512_shuffle_epi32(lhs_mat_23_21, (_MM_PERM_ENUM)245); //A22(12-15) A22(12-15) A23(12-15) A23(12-15) A22(12-15) A22(12-15) A23(12-15) A23(12-15) A22(12-15) A22(12-15) A23(12-15) A23(12-15) A22(12-15) A22(12-15) A23(12-15) A23(12-15)
+
+ const __m512i lhs_mat_01_30_sp2 = _mm512_shuffle_epi32(lhs_mat_01_30, (_MM_PERM_ENUM)245); //A30(4-7) A30(4-7) A31(4-7) A31(4-7) A30(4-7) A30(4-7) A31(4-7) A31(4-7) A30(4-7) A30(4-7) A31(4-7) A31(4-7) A30(4-7) A30(4-7) A31(4-7) A31(4-7)
+ const __m512i lhs_mat_23_30_sp2 = _mm512_shuffle_epi32(lhs_mat_23_30, (_MM_PERM_ENUM)245); //A32(4-7) A32(4-7) A33(4-7) A33(4-7) A32(4-7) A32(4-7) A33(4-7) A33(4-7) A32(4-7) A32(4-7) A33(4-7) A33(4-7) A32(4-7) A32(4-7) A33(4-7) A33(4-7)
+
+ const __m512i lhs_mat_01_31_sp2 = _mm512_shuffle_epi32(lhs_mat_01_31, (_MM_PERM_ENUM)245); //A30(12-15) A30(12-15) A31(12-15) A31(12-15) A30(12-15) A30(12-15) A31(12-15) A31(12-15) A30(12-15) A30(12-15) A31(12-15) A31(12-15) A30(12-15) A30(12-15) A31(12-15) A31(12-15)
+ const __m512i lhs_mat_23_31_sp2 = _mm512_shuffle_epi32(lhs_mat_23_31, (_MM_PERM_ENUM)245); //A32(12-15) A32(12-15) A33(12-15) A33(12-15) A32(12-15) A32(12-15) A33(12-15) A33(12-15) A32(12-15) A32(12-15) A33(12-15) A33(12-15) A32(12-15) A32(12-15) A33(12-15) A33(12-15)
+
+ const __m512i lhs_mat_01_40_sp2 = _mm512_shuffle_epi32(lhs_mat_01_40, (_MM_PERM_ENUM)245); //A40(4-7) A40(4-7) A41(4-7) A41(4-7) A40(4-7) A40(4-7) A41(4-7) A41(4-7) A40(4-7) A40(4-7) A41(4-7) A41(4-7) A40(4-7) A40(4-7) A41(4-7) A41(4-7)
+ const __m512i lhs_mat_23_40_sp2 = _mm512_shuffle_epi32(lhs_mat_23_40, (_MM_PERM_ENUM)245); //A42(4-7) A42(4-7) A43(4-7) A43(4-7) A42(4-7) A42(4-7) A43(4-7) A43(4-7) A42(4-7) A42(4-7) A43(4-7) A43(4-7) A42(4-7) A42(4-7) A43(4-7) A43(4-7)
+
+ const __m512i lhs_mat_01_41_sp2 = _mm512_shuffle_epi32(lhs_mat_01_41, (_MM_PERM_ENUM)245); //A40(12-15) A40(12-15) A41(12-15) A41(12-15) A40(12-15) A40(12-15) A41(12-15) A41(12-15) A40(12-15) A40(12-15) A41(12-15) A41(12-15) A40(12-15) A40(12-15) A41(12-15) A41(12-15)
+ const __m512i lhs_mat_23_41_sp2 = _mm512_shuffle_epi32(lhs_mat_23_41, (_MM_PERM_ENUM)245); //A42(12-15) A42(12-15) A43(12-15) A43(12-15) A42(12-15) A42(12-15) A43(12-15) A43(12-15) A42(12-15) A42(12-15) A43(12-15) A43(12-15) A42(12-15) A42(12-15) A43(12-15) A43(12-15)
+
+ const __m512i lhs_mat_01_50_sp2 = _mm512_shuffle_epi32(lhs_mat_01_50, (_MM_PERM_ENUM)245); //A50(4-7) A50(4-7) A51(4-7) A51(4-7) A50(4-7) A50(4-7) A51(4-7) A51(4-7) A50(4-7) A50(4-7) A51(4-7) A51(4-7) A50(4-7) A50(4-7) A51(4-7) A51(4-7)
+ const __m512i lhs_mat_23_50_sp2 = _mm512_shuffle_epi32(lhs_mat_23_50, (_MM_PERM_ENUM)245); //A52(4-7) A52(4-7) A53(4-7) A53(4-7) A52(4-7) A52(4-7) A53(4-7) A53(4-7) A52(4-7) A52(4-7) A53(4-7) A53(4-7) A52(4-7) A52(4-7) A53(4-7) A53(4-7)
+
+ const __m512i lhs_mat_01_51_sp2 = _mm512_shuffle_epi32(lhs_mat_01_51, (_MM_PERM_ENUM)245); //A50(12-15) A50(12-15) A51(12-15) A51(12-15) A50(12-15) A50(12-15) A51(12-15) A51(12-15) A50(12-15) A50(12-15) A51(12-15) A51(12-15) A50(12-15) A50(12-15) A51(12-15) A51(12-15)
+ const __m512i lhs_mat_23_51_sp2 = _mm512_shuffle_epi32(lhs_mat_23_51, (_MM_PERM_ENUM)245); //A52(12-15) A52(12-15) A53(12-15) A53(12-15) A52(12-15) A52(12-15) A53(12-15) A53(12-15) A52(12-15) A52(12-15) A53(12-15) A53(12-15) A52(12-15) A52(12-15) A53(12-15) A53(12-15)
+
+ const __m512i lhs_mat_01_60_sp2 = _mm512_shuffle_epi32(lhs_mat_01_60, (_MM_PERM_ENUM)245); //A60(4-7) A60(4-7) A61(4-7) A61(4-7) A60(4-7) A60(4-7) A61(4-7) A61(4-7) A60(4-7) A60(4-7) A61(4-7) A61(4-7) A60(4-7) A60(4-7) A61(4-7) A61(4-7)
+ const __m512i lhs_mat_23_60_sp2 = _mm512_shuffle_epi32(lhs_mat_23_60, (_MM_PERM_ENUM)245); //A62(4-7) A62(4-7) A63(4-7) A63(4-7) A62(4-7) A62(4-7) A63(4-7) A63(4-7) A62(4-7) A62(4-7) A63(4-7) A63(4-7) A62(4-7) A62(4-7) A63(4-7) A63(4-7)
+
+ const __m512i lhs_mat_01_61_sp2 = _mm512_shuffle_epi32(lhs_mat_01_61, (_MM_PERM_ENUM)245); //A60(12-15) A60(12-15) A61(12-15) A61(12-15) A60(12-15) A60(12-15) A61(12-15) A61(12-15) A60(12-15) A60(12-15) A61(12-15) A61(12-15) A60(12-15) A60(12-15) A61(12-15) A61(12-15)
+ const __m512i lhs_mat_23_61_sp2 = _mm512_shuffle_epi32(lhs_mat_23_61, (_MM_PERM_ENUM)245); //A62(12-15) A62(12-15) A63(12-15) A63(12-15) A62(12-15) A62(12-15) A63(12-15) A63(12-15) A62(12-15) A62(12-15) A63(12-15) A63(12-15) A62(12-15) A62(12-15) A63(12-15) A63(12-15)
+
+ const __m512i lhs_mat_01_70_sp2 = _mm512_shuffle_epi32(lhs_mat_01_70, (_MM_PERM_ENUM)245); //A70(4-7) A70(4-7) A71(4-7) A71(4-7) A70(4-7) A70(4-7) A71(4-7) A71(4-7) A70(4-7) A70(4-7) A71(4-7) A71(4-7) A70(4-7) A70(4-7) A71(4-7) A71(4-7)
+ const __m512i lhs_mat_23_70_sp2 = _mm512_shuffle_epi32(lhs_mat_23_70, (_MM_PERM_ENUM)245); //A72(4-7) A72(4-7) A73(4-7) A73(4-7) A72(4-7) A72(4-7) A73(4-7) A73(4-7) A72(4-7) A72(4-7) A73(4-7) A73(4-7) A72(4-7) A72(4-7) A73(4-7) A73(4-7)
+
+ const __m512i lhs_mat_01_71_sp2 = _mm512_shuffle_epi32(lhs_mat_01_71, (_MM_PERM_ENUM)245); //A70(12-15) A70(12-15) A71(12-15) A71(12-15) A70(12-15) A70(12-15) A71(12-15) A71(12-15) A70(12-15) A70(12-15) A71(12-15) A71(12-15) A70(12-15) A70(12-15) A71(12-15) A71(12-15)
+ const __m512i lhs_mat_23_71_sp2 = _mm512_shuffle_epi32(lhs_mat_23_71, (_MM_PERM_ENUM)245); //A72(12-15) A72(12-15) A73(12-15) A73(12-15) A72(12-15) A72(12-15) A73(12-15) A73(12-15) A72(12-15) A72(12-15) A73(12-15) A73(12-15) A72(12-15) A72(12-15) A73(12-15) A73(12-15)
+
+ // 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_maddubs_epi16(rhs_mat_014589CD_00_sp1, lhs_mat_01_00_sp1),_mm512_maddubs_epi16(rhs_mat_014589CD_01_sp1, lhs_mat_01_01_sp1));
+ __m512i iacc_mat_01_0_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_00_sp1, lhs_mat_01_00_sp1),_mm512_maddubs_epi16(rhs_mat_2367ABEF_01_sp1, lhs_mat_01_01_sp1));
+
+ __m512i iacc_mat_10_0_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_00_sp1, lhs_mat_23_00_sp1),_mm512_maddubs_epi16(rhs_mat_014589CD_01_sp1, lhs_mat_23_01_sp1));
+ __m512i iacc_mat_11_0_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_00_sp1, lhs_mat_23_00_sp1),_mm512_maddubs_epi16(rhs_mat_2367ABEF_01_sp1, lhs_mat_23_01_sp1));
+
+ __m512i iacc_mat_00_1_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_10_sp1, lhs_mat_01_10_sp1),_mm512_maddubs_epi16(rhs_mat_014589CD_11_sp1, lhs_mat_01_11_sp1));
+ __m512i iacc_mat_01_1_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_10_sp1, lhs_mat_01_10_sp1),_mm512_maddubs_epi16(rhs_mat_2367ABEF_11_sp1, lhs_mat_01_11_sp1));
+
+ __m512i iacc_mat_10_1_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_10_sp1, lhs_mat_23_10_sp1),_mm512_maddubs_epi16(rhs_mat_014589CD_11_sp1, lhs_mat_23_11_sp1));
+ __m512i iacc_mat_11_1_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_10_sp1, lhs_mat_23_10_sp1),_mm512_maddubs_epi16(rhs_mat_2367ABEF_11_sp1, lhs_mat_23_11_sp1));
+
+ __m512i iacc_mat_00_2_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_20_sp1, lhs_mat_01_20_sp1),_mm512_maddubs_epi16(rhs_mat_014589CD_21_sp1, lhs_mat_01_21_sp1));
+ __m512i iacc_mat_01_2_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_20_sp1, lhs_mat_01_20_sp1),_mm512_maddubs_epi16(rhs_mat_2367ABEF_21_sp1, lhs_mat_01_21_sp1));
+
+ __m512i iacc_mat_10_2_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_20_sp1, lhs_mat_23_20_sp1),_mm512_maddubs_epi16(rhs_mat_014589CD_21_sp1, lhs_mat_23_21_sp1));
+ __m512i iacc_mat_11_2_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_20_sp1, lhs_mat_23_20_sp1),_mm512_maddubs_epi16(rhs_mat_2367ABEF_21_sp1, lhs_mat_23_21_sp1));
+
+ __m512i iacc_mat_00_3_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_30_sp1, lhs_mat_01_30_sp1),_mm512_maddubs_epi16(rhs_mat_014589CD_31_sp1, lhs_mat_01_31_sp1));
+ __m512i iacc_mat_01_3_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_30_sp1, lhs_mat_01_30_sp1),_mm512_maddubs_epi16(rhs_mat_2367ABEF_31_sp1, lhs_mat_01_31_sp1));
+
+ __m512i iacc_mat_10_3_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_30_sp1, lhs_mat_23_30_sp1),_mm512_maddubs_epi16(rhs_mat_014589CD_31_sp1, lhs_mat_23_31_sp1));
+ __m512i iacc_mat_11_3_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_30_sp1, lhs_mat_23_30_sp1),_mm512_maddubs_epi16(rhs_mat_2367ABEF_31_sp1, lhs_mat_23_31_sp1));
+
+ __m512i iacc_mat_00_4_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_40_sp1, lhs_mat_01_40_sp1),_mm512_maddubs_epi16(rhs_mat_014589CD_41_sp1, lhs_mat_01_41_sp1));
+ __m512i iacc_mat_01_4_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_40_sp1, lhs_mat_01_40_sp1),_mm512_maddubs_epi16(rhs_mat_2367ABEF_41_sp1, lhs_mat_01_41_sp1));
+
+ __m512i iacc_mat_10_4_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_40_sp1, lhs_mat_23_40_sp1),_mm512_maddubs_epi16(rhs_mat_014589CD_41_sp1, lhs_mat_23_41_sp1));
+ __m512i iacc_mat_11_4_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_40_sp1, lhs_mat_23_40_sp1),_mm512_maddubs_epi16(rhs_mat_2367ABEF_41_sp1, lhs_mat_23_41_sp1));
+
+ __m512i iacc_mat_00_5_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_50_sp1, lhs_mat_01_50_sp1),_mm512_maddubs_epi16(rhs_mat_014589CD_51_sp1, lhs_mat_01_51_sp1));
+ __m512i iacc_mat_01_5_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_50_sp1, lhs_mat_01_50_sp1),_mm512_maddubs_epi16(rhs_mat_2367ABEF_51_sp1, lhs_mat_01_51_sp1));
+
+ __m512i iacc_mat_10_5_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_50_sp1, lhs_mat_23_50_sp1),_mm512_maddubs_epi16(rhs_mat_014589CD_51_sp1, lhs_mat_23_51_sp1));
+ __m512i iacc_mat_11_5_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_50_sp1, lhs_mat_23_50_sp1),_mm512_maddubs_epi16(rhs_mat_2367ABEF_51_sp1, lhs_mat_23_51_sp1));
+
+ __m512i iacc_mat_00_6_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_60_sp1, lhs_mat_01_60_sp1),_mm512_maddubs_epi16(rhs_mat_014589CD_61_sp1, lhs_mat_01_61_sp1));
+ __m512i iacc_mat_01_6_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_60_sp1, lhs_mat_01_60_sp1),_mm512_maddubs_epi16(rhs_mat_2367ABEF_61_sp1, lhs_mat_01_61_sp1));
+
+ __m512i iacc_mat_10_6_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_60_sp1, lhs_mat_23_60_sp1),_mm512_maddubs_epi16(rhs_mat_014589CD_61_sp1, lhs_mat_23_61_sp1));
+ __m512i iacc_mat_11_6_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_60_sp1, lhs_mat_23_60_sp1),_mm512_maddubs_epi16(rhs_mat_2367ABEF_61_sp1, lhs_mat_23_61_sp1));
+
+ __m512i iacc_mat_00_7_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_70_sp1, lhs_mat_01_70_sp1),_mm512_maddubs_epi16(rhs_mat_014589CD_71_sp1, lhs_mat_01_71_sp1));
+ __m512i iacc_mat_01_7_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_70_sp1, lhs_mat_01_70_sp1),_mm512_maddubs_epi16(rhs_mat_2367ABEF_71_sp1, lhs_mat_01_71_sp1));
+
+ __m512i iacc_mat_10_7_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_70_sp1, lhs_mat_23_70_sp1),_mm512_maddubs_epi16(rhs_mat_014589CD_71_sp1, lhs_mat_23_71_sp1));
+ __m512i iacc_mat_11_7_sp1 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_70_sp1, lhs_mat_23_70_sp1),_mm512_maddubs_epi16(rhs_mat_2367ABEF_71_sp1, lhs_mat_23_71_sp1));
+
+
+ __m512i iacc_mat_00_0_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_00_sp2, lhs_mat_01_00_sp2),_mm512_maddubs_epi16(rhs_mat_014589CD_01_sp2, lhs_mat_01_01_sp2));
+ __m512i iacc_mat_01_0_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_00_sp2, lhs_mat_01_00_sp2),_mm512_maddubs_epi16(rhs_mat_2367ABEF_01_sp2, lhs_mat_01_01_sp2));
+
+ __m512i iacc_mat_10_0_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_00_sp2, lhs_mat_23_00_sp2),_mm512_maddubs_epi16(rhs_mat_014589CD_01_sp2, lhs_mat_23_01_sp2));
+ __m512i iacc_mat_11_0_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_00_sp2, lhs_mat_23_00_sp2),_mm512_maddubs_epi16(rhs_mat_2367ABEF_01_sp2, lhs_mat_23_01_sp2));
+
+ __m512i iacc_mat_00_1_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_10_sp2, lhs_mat_01_10_sp2),_mm512_maddubs_epi16(rhs_mat_014589CD_11_sp2, lhs_mat_01_11_sp2));
+ __m512i iacc_mat_01_1_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_10_sp2, lhs_mat_01_10_sp2),_mm512_maddubs_epi16(rhs_mat_2367ABEF_11_sp2, lhs_mat_01_11_sp2));
+
+ __m512i iacc_mat_10_1_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_10_sp2, lhs_mat_23_10_sp2),_mm512_maddubs_epi16(rhs_mat_014589CD_11_sp2, lhs_mat_23_11_sp2));
+ __m512i iacc_mat_11_1_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_10_sp2, lhs_mat_23_10_sp2),_mm512_maddubs_epi16(rhs_mat_2367ABEF_11_sp2, lhs_mat_23_11_sp2));
+
+ __m512i iacc_mat_00_2_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_20_sp2, lhs_mat_01_20_sp2),_mm512_maddubs_epi16(rhs_mat_014589CD_21_sp2, lhs_mat_01_21_sp2));
+ __m512i iacc_mat_01_2_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_20_sp2, lhs_mat_01_20_sp2),_mm512_maddubs_epi16(rhs_mat_2367ABEF_21_sp2, lhs_mat_01_21_sp2));
+
+ __m512i iacc_mat_10_2_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_20_sp2, lhs_mat_23_20_sp2),_mm512_maddubs_epi16(rhs_mat_014589CD_21_sp2, lhs_mat_23_21_sp2));
+ __m512i iacc_mat_11_2_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_20_sp2, lhs_mat_23_20_sp2),_mm512_maddubs_epi16(rhs_mat_2367ABEF_21_sp2, lhs_mat_23_21_sp2));
+
+ __m512i iacc_mat_00_3_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_30_sp2, lhs_mat_01_30_sp2),_mm512_maddubs_epi16(rhs_mat_014589CD_31_sp2, lhs_mat_01_31_sp2));
+ __m512i iacc_mat_01_3_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_30_sp2, lhs_mat_01_30_sp2),_mm512_maddubs_epi16(rhs_mat_2367ABEF_31_sp2, lhs_mat_01_31_sp2));
+
+ __m512i iacc_mat_10_3_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_30_sp2, lhs_mat_23_30_sp2),_mm512_maddubs_epi16(rhs_mat_014589CD_31_sp2, lhs_mat_23_31_sp2));
+ __m512i iacc_mat_11_3_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_30_sp2, lhs_mat_23_30_sp2),_mm512_maddubs_epi16(rhs_mat_2367ABEF_31_sp2, lhs_mat_23_31_sp2));
+
+ __m512i iacc_mat_00_4_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_40_sp2, lhs_mat_01_40_sp2),_mm512_maddubs_epi16(rhs_mat_014589CD_41_sp2, lhs_mat_01_41_sp2));
+ __m512i iacc_mat_01_4_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_40_sp2, lhs_mat_01_40_sp2),_mm512_maddubs_epi16(rhs_mat_2367ABEF_41_sp2, lhs_mat_01_41_sp2));
+
+ __m512i iacc_mat_10_4_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_40_sp2, lhs_mat_23_40_sp2),_mm512_maddubs_epi16(rhs_mat_014589CD_41_sp2, lhs_mat_23_41_sp2));
+ __m512i iacc_mat_11_4_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_40_sp2, lhs_mat_23_40_sp2),_mm512_maddubs_epi16(rhs_mat_2367ABEF_41_sp2, lhs_mat_23_41_sp2));
+
+ __m512i iacc_mat_00_5_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_50_sp2, lhs_mat_01_50_sp2),_mm512_maddubs_epi16(rhs_mat_014589CD_51_sp2, lhs_mat_01_51_sp2));
+ __m512i iacc_mat_01_5_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_50_sp2, lhs_mat_01_50_sp2),_mm512_maddubs_epi16(rhs_mat_2367ABEF_51_sp2, lhs_mat_01_51_sp2));
+
+ __m512i iacc_mat_10_5_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_50_sp2, lhs_mat_23_50_sp2),_mm512_maddubs_epi16(rhs_mat_014589CD_51_sp2, lhs_mat_23_51_sp2));
+ __m512i iacc_mat_11_5_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_50_sp2, lhs_mat_23_50_sp2),_mm512_maddubs_epi16(rhs_mat_2367ABEF_51_sp2, lhs_mat_23_51_sp2));
+
+ __m512i iacc_mat_00_6_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_60_sp2, lhs_mat_01_60_sp2),_mm512_maddubs_epi16(rhs_mat_014589CD_61_sp2, lhs_mat_01_61_sp2));
+ __m512i iacc_mat_01_6_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_60_sp2, lhs_mat_01_60_sp2),_mm512_maddubs_epi16(rhs_mat_2367ABEF_61_sp2, lhs_mat_01_61_sp2));
+
+ __m512i iacc_mat_10_6_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_60_sp2, lhs_mat_23_60_sp2),_mm512_maddubs_epi16(rhs_mat_014589CD_61_sp2, lhs_mat_23_61_sp2));
+ __m512i iacc_mat_11_6_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_60_sp2, lhs_mat_23_60_sp2),_mm512_maddubs_epi16(rhs_mat_2367ABEF_61_sp2, lhs_mat_23_61_sp2));
+
+ __m512i iacc_mat_00_7_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_70_sp2, lhs_mat_01_70_sp2),_mm512_maddubs_epi16(rhs_mat_014589CD_71_sp2, lhs_mat_01_71_sp2));
+ __m512i iacc_mat_01_7_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_70_sp2, lhs_mat_01_70_sp2),_mm512_maddubs_epi16(rhs_mat_2367ABEF_71_sp2, lhs_mat_01_71_sp2));
+
+ __m512i iacc_mat_10_7_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_70_sp2, lhs_mat_23_70_sp2),_mm512_maddubs_epi16(rhs_mat_014589CD_71_sp2, lhs_mat_23_71_sp2));
+ __m512i iacc_mat_11_7_sp2 = _mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_70_sp2, lhs_mat_23_70_sp2),_mm512_maddubs_epi16(rhs_mat_2367ABEF_71_sp2, lhs_mat_23_71_sp2));
+
+ // Combine results from both shuffle patterns for each output 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);
+
+ __m512i iacc_mat_00_2 = _mm512_add_epi16(iacc_mat_00_2_sp1, iacc_mat_00_2_sp2);
+ __m512i iacc_mat_01_2 = _mm512_add_epi16(iacc_mat_01_2_sp1, iacc_mat_01_2_sp2);
+ __m512i iacc_mat_10_2 = _mm512_add_epi16(iacc_mat_10_2_sp1, iacc_mat_10_2_sp2);
+ __m512i iacc_mat_11_2 = _mm512_add_epi16(iacc_mat_11_2_sp1, iacc_mat_11_2_sp2);
+
+ __m512i iacc_mat_00_3 = _mm512_add_epi16(iacc_mat_00_3_sp1, iacc_mat_00_3_sp2);
+ __m512i iacc_mat_01_3 = _mm512_add_epi16(iacc_mat_01_3_sp1, iacc_mat_01_3_sp2);
+ __m512i iacc_mat_10_3 = _mm512_add_epi16(iacc_mat_10_3_sp1, iacc_mat_10_3_sp2);
+ __m512i iacc_mat_11_3 = _mm512_add_epi16(iacc_mat_11_3_sp1, iacc_mat_11_3_sp2);
+
+ __m512i iacc_mat_00_4 = _mm512_add_epi16(iacc_mat_00_4_sp1, iacc_mat_00_4_sp2);
+ __m512i iacc_mat_01_4 = _mm512_add_epi16(iacc_mat_01_4_sp1, iacc_mat_01_4_sp2);
+ __m512i iacc_mat_10_4 = _mm512_add_epi16(iacc_mat_10_4_sp1, iacc_mat_10_4_sp2);
+ __m512i iacc_mat_11_4 = _mm512_add_epi16(iacc_mat_11_4_sp1, iacc_mat_11_4_sp2);
+
+ __m512i iacc_mat_00_5 = _mm512_add_epi16(iacc_mat_00_5_sp1, iacc_mat_00_5_sp2);
+ __m512i iacc_mat_01_5 = _mm512_add_epi16(iacc_mat_01_5_sp1, iacc_mat_01_5_sp2);
+ __m512i iacc_mat_10_5 = _mm512_add_epi16(iacc_mat_10_5_sp1, iacc_mat_10_5_sp2);
+ __m512i iacc_mat_11_5 = _mm512_add_epi16(iacc_mat_11_5_sp1, iacc_mat_11_5_sp2);
+
+ __m512i iacc_mat_00_6 = _mm512_add_epi16(iacc_mat_00_6_sp1, iacc_mat_00_6_sp2);
+ __m512i iacc_mat_01_6 = _mm512_add_epi16(iacc_mat_01_6_sp1, iacc_mat_01_6_sp2);
+ __m512i iacc_mat_10_6 = _mm512_add_epi16(iacc_mat_10_6_sp1, iacc_mat_10_6_sp2);
+ __m512i iacc_mat_11_6 = _mm512_add_epi16(iacc_mat_11_6_sp1, iacc_mat_11_6_sp2);
+
+ __m512i iacc_mat_00_7 = _mm512_add_epi16(iacc_mat_00_7_sp1, iacc_mat_00_7_sp2);
+ __m512i iacc_mat_01_7 = _mm512_add_epi16(iacc_mat_01_7_sp1, iacc_mat_01_7_sp2);
+ __m512i iacc_mat_10_7 = _mm512_add_epi16(iacc_mat_10_7_sp1, iacc_mat_10_7_sp2);
+ __m512i iacc_mat_11_7 = _mm512_add_epi16(iacc_mat_11_7_sp1, iacc_mat_11_7_sp2);
+
+ // Output of both shuffle patterns are added in order to sum dot product outputs of all 32 values in block
+ 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);
+
+ iacc_mat_00_2 = _mm512_madd_epi16(iacc_mat_00_2, scale_014589CD_2);
+ iacc_mat_01_2 = _mm512_madd_epi16(iacc_mat_01_2, scale_2367ABEF_2);
+ iacc_mat_10_2 = _mm512_madd_epi16(iacc_mat_10_2, scale_014589CD_2);
+ iacc_mat_11_2 = _mm512_madd_epi16(iacc_mat_11_2, scale_2367ABEF_2);
+
+ iacc_mat_00_3 = _mm512_madd_epi16(iacc_mat_00_3, scale_014589CD_3);
+ iacc_mat_01_3 = _mm512_madd_epi16(iacc_mat_01_3, scale_2367ABEF_3);
+ iacc_mat_10_3 = _mm512_madd_epi16(iacc_mat_10_3, scale_014589CD_3);
+ iacc_mat_11_3 = _mm512_madd_epi16(iacc_mat_11_3, scale_2367ABEF_3);
+
+ iacc_mat_00_4 = _mm512_madd_epi16(iacc_mat_00_4, scale_014589CD_4);
+ iacc_mat_01_4 = _mm512_madd_epi16(iacc_mat_01_4, scale_2367ABEF_4);
+ iacc_mat_10_4 = _mm512_madd_epi16(iacc_mat_10_4, scale_014589CD_4);
+ iacc_mat_11_4 = _mm512_madd_epi16(iacc_mat_11_4, scale_2367ABEF_4);
+
+ iacc_mat_00_5 = _mm512_madd_epi16(iacc_mat_00_5, scale_014589CD_5);
+ iacc_mat_01_5 = _mm512_madd_epi16(iacc_mat_01_5, scale_2367ABEF_5);
+ iacc_mat_10_5 = _mm512_madd_epi16(iacc_mat_10_5, scale_014589CD_5);
+ iacc_mat_11_5 = _mm512_madd_epi16(iacc_mat_11_5, scale_2367ABEF_5);
+
+ iacc_mat_00_6 = _mm512_madd_epi16(iacc_mat_00_6, scale_014589CD_6);
+ iacc_mat_01_6 = _mm512_madd_epi16(iacc_mat_01_6, scale_2367ABEF_6);
+ iacc_mat_10_6 = _mm512_madd_epi16(iacc_mat_10_6, scale_014589CD_6);
+ iacc_mat_11_6 = _mm512_madd_epi16(iacc_mat_11_6, scale_2367ABEF_6);
+
+ iacc_mat_00_7 = _mm512_madd_epi16(iacc_mat_00_7, scale_014589CD_7);
+ iacc_mat_01_7 = _mm512_madd_epi16(iacc_mat_01_7, scale_2367ABEF_7);
+ iacc_mat_10_7 = _mm512_madd_epi16(iacc_mat_10_7, scale_014589CD_7);
+ iacc_mat_11_7 = _mm512_madd_epi16(iacc_mat_11_7, scale_2367ABEF_7);
+
+ __m512i iacc_mat_00 = _mm512_add_epi32(_mm512_add_epi32(_mm512_add_epi32(iacc_mat_00_0, iacc_mat_00_1), _mm512_add_epi32(iacc_mat_00_2, iacc_mat_00_3)), _mm512_add_epi32(_mm512_add_epi32(iacc_mat_00_4, iacc_mat_00_5), _mm512_add_epi32(iacc_mat_00_6, iacc_mat_00_7)));
+ __m512i iacc_mat_01 = _mm512_add_epi32(_mm512_add_epi32(_mm512_add_epi32(iacc_mat_01_0, iacc_mat_01_1), _mm512_add_epi32(iacc_mat_01_2, iacc_mat_01_3)), _mm512_add_epi32(_mm512_add_epi32(iacc_mat_01_4, iacc_mat_01_5), _mm512_add_epi32(iacc_mat_01_6, iacc_mat_01_7)));
+ __m512i iacc_mat_10 = _mm512_add_epi32(_mm512_add_epi32(_mm512_add_epi32(iacc_mat_10_0, iacc_mat_10_1), _mm512_add_epi32(iacc_mat_10_2, iacc_mat_10_3)), _mm512_add_epi32(_mm512_add_epi32(iacc_mat_10_4, iacc_mat_10_5), _mm512_add_epi32(iacc_mat_10_6, iacc_mat_10_7)));
+ __m512i iacc_mat_11 = _mm512_add_epi32(_mm512_add_epi32(_mm512_add_epi32(iacc_mat_11_0, iacc_mat_11_1), _mm512_add_epi32(iacc_mat_11_2, iacc_mat_11_3)), _mm512_add_epi32(_mm512_add_epi32(iacc_mat_11_4, iacc_mat_11_5), _mm512_add_epi32(iacc_mat_11_6, iacc_mat_11_7)));
+
+ // Straighten out to make 4 row vectors
+ __m512i iacc_row_0 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_00, _mm512_shuffle_epi32(iacc_mat_01, (_MM_PERM_ENUM)78));
+ __m512i iacc_row_1 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_00, (_MM_PERM_ENUM)78), iacc_mat_01);
+ __m512i iacc_row_2 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_10, _mm512_shuffle_epi32(iacc_mat_11, (_MM_PERM_ENUM)78));
+ __m512i iacc_row_3 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_10, (_MM_PERM_ENUM)78), iacc_mat_11);
+
+ // Load the scale(d) values for all the 4 Q8_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]);
+
+ // Take two bsums from two Q8_Ks at a time and multiply with corresponding mins values from each Q2_K
+ __m512i iacc_row_min_0_01 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_01_0123, (_MM_PERM_ENUM)0), mins_01);
+ __m512i iacc_row_min_1_01 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_01_0123, (_MM_PERM_ENUM)170), mins_01);
+ __m512i iacc_row_min_2_01 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_23_0123, (_MM_PERM_ENUM)0), mins_01);
+ __m512i iacc_row_min_3_01 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_23_0123, (_MM_PERM_ENUM)170), mins_01);
+
+ __m512i iacc_row_min_0_23 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_01_0123, (_MM_PERM_ENUM)85), mins_23);
+ __m512i iacc_row_min_1_23 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_01_0123, (_MM_PERM_ENUM)255), mins_23);
+ __m512i iacc_row_min_2_23 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_23_0123, (_MM_PERM_ENUM)85), mins_23);
+ __m512i iacc_row_min_3_23 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_23_0123, (_MM_PERM_ENUM)255), mins_23);
+
+ __m512i iacc_row_min_0_45 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_01_4567, (_MM_PERM_ENUM)0), mins_45);
+ __m512i iacc_row_min_1_45 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_01_4567, (_MM_PERM_ENUM)170), mins_45);
+ __m512i iacc_row_min_2_45 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_23_4567, (_MM_PERM_ENUM)0), mins_45);
+ __m512i iacc_row_min_3_45 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_23_4567, (_MM_PERM_ENUM)170), mins_45);
+
+ __m512i iacc_row_min_0_67 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_01_4567, (_MM_PERM_ENUM)85), mins_67);
+ __m512i iacc_row_min_1_67 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_01_4567, (_MM_PERM_ENUM)255), mins_67);
+ __m512i iacc_row_min_2_67 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_23_4567, (_MM_PERM_ENUM)85), mins_67);
+ __m512i iacc_row_min_3_67 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_23_4567, (_MM_PERM_ENUM)255), mins_67);
+
+ __m512i iacc_row_min_0 = _mm512_add_epi32(_mm512_add_epi32(iacc_row_min_0_01, iacc_row_min_0_23), _mm512_add_epi32(iacc_row_min_0_45,iacc_row_min_0_67));
+ __m512i iacc_row_min_1 = _mm512_add_epi32(_mm512_add_epi32(iacc_row_min_1_01, iacc_row_min_1_23), _mm512_add_epi32(iacc_row_min_1_45,iacc_row_min_1_67));
+ __m512i iacc_row_min_2 = _mm512_add_epi32(_mm512_add_epi32(iacc_row_min_2_01, iacc_row_min_2_23), _mm512_add_epi32(iacc_row_min_2_45,iacc_row_min_2_67));
+ __m512i iacc_row_min_3 = _mm512_add_epi32(_mm512_add_epi32(iacc_row_min_3_01, iacc_row_min_3_23), _mm512_add_epi32(iacc_row_min_3_45,iacc_row_min_3_67));
+
+ 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) {
+
+ 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_q2_kx8 structures at each pass of the loop and perform dot product operation
+ for (int64_t x = xstart; x < nc / 8; x++) {
+
+ const block_q2_Kx8 * b_ptr = b_ptr_start + (x * b_nb);
+
+ // Master FP accumulators
+ __m256 acc_rows[16];
+ for (int i = 0; i < 16; i++) {
+ acc_rows[i] = _mm256_setzero_ps();
+ }
+
+ __m256 acc_min_rows[16];
+ for (int i = 0; i < 16; i++) {
+ acc_min_rows[i] = _mm256_setzero_ps();
+ }
+
+ // For super block
+ for (int64_t b = 0; b < nb; b++) {
+ // Delta values - Load the eight scale values of block_q2_kx8
+ const __m256 col_scale_f32 = GGML_F32Cx8_LOAD(b_ptr[b].d);
+
+ // dmin values - Load the eight dmin values of block_q2_kx8
+ const __m256 col_dmin_f32 = GGML_F32Cx8_LOAD(b_ptr[b].dmin);
+
+ // Loop to iterate over the sixteen sub blocks of a super block - eight sub blocks are processed per iteration
+ for (int sb = 0; sb < QK_K / 128; sb++) {
+
+ // Load the eight block_q4_K for eight sub blocks quantized values interleaved with each other in chunks of eight bytes - B0,B1 ....B6,B7
+ const __m256i rhs_raw_mat_0123_0 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + sb * 256));
+ const __m256i rhs_raw_mat_4567_0 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 32 + sb * 256));
+ const __m256i rhs_raw_mat_0123_1 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 64 + sb * 256));
+ const __m256i rhs_raw_mat_4567_1 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 96 + sb * 256));
+ const __m256i rhs_raw_mat_0123_2 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 128 + sb * 256));
+ const __m256i rhs_raw_mat_4567_2 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 160 + sb * 256));
+ const __m256i rhs_raw_mat_0123_3 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 192 + sb * 256));
+ const __m256i rhs_raw_mat_4567_3 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 224 + sb * 256));
+
+ // Save the values in the following vectors in the formats B0B1B4B5, B2B3B6B7 for further processing and storing of values
+ //superblock sub block which part of sub block
+ 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);
+
+ // 2-bit -> 8-bit
+ // First sub block of the eight sub blocks processed in the iteration
+ const __m256i rhs_mat_0145_00 = _mm256_and_si256(rhs_raw_mat_0145_0, m3b); //B00(0-7) B01(0-7) B04(0-7) B05(0-7)
+ const __m256i rhs_mat_2367_00 = _mm256_and_si256(rhs_raw_mat_2367_0, m3b); //B02(0-7) B03(0-7) B06(0-7) B07(0-7)
+
+ const __m256i rhs_mat_0145_01 = _mm256_and_si256(rhs_raw_mat_0145_1, m3b); //B00(8-15) B01(8-15) B04(8-15) B05(8-15)
+ const __m256i rhs_mat_2367_01 = _mm256_and_si256(rhs_raw_mat_2367_1, m3b); //B02(8-15) B03(8-15) B06(8-15) B07(8-15)
+
+ // Second sub block of the eight sub blocks processed in the iteration
+ const __m256i rhs_mat_0145_10 = _mm256_and_si256(rhs_raw_mat_0145_2, m3b); //B10(0-7) B11(0-7) B14(0-7) B15(0-7)
+ const __m256i rhs_mat_2367_10 = _mm256_and_si256(rhs_raw_mat_2367_2, m3b); //B12(0-7) B13(0-7) B16(0-7) B17(0-7)
+
+ const __m256i rhs_mat_0145_11 = _mm256_and_si256(rhs_raw_mat_0145_3, m3b); //B10(8-15) B11(8-15) B14(8-15) B15(8-15)
+ const __m256i rhs_mat_2367_11 = _mm256_and_si256(rhs_raw_mat_2367_3, m3b); //B12(8-15) B13(8-15) B16(8-15) B17(8-15)
+
+ // Third sub block of the eight sub blocks processed in the iteration
+ const __m256i rhs_mat_0145_20 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_0, 2), m3b); //B20(0-7) B21(0-7) B24(0-7) B25(0-7)
+ const __m256i rhs_mat_2367_20 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_0, 2), m3b); //B22(0-7) B23(0-7) B26(0-7) B27(0-7)
+
+ const __m256i rhs_mat_0145_21 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_1, 2), m3b); //B20(8-15) B21(8-15) B24(8-15) B25(8-15)
+ const __m256i rhs_mat_2367_21 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_1, 2), m3b); //B22(8-15) B23(8-15) B26(8-15) B27(8-15)
+
+ // Fourth sub block of the eight sub blocks processed in the iteration
+ const __m256i rhs_mat_0145_30 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_2, 2), m3b); //B30(0-7) B31(0-7) B34(0-7) B35(0-7)
+ const __m256i rhs_mat_2367_30 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_2, 2), m3b); //B32(0-7) B33(0-7) B36(0-7) B37(0-7)
+
+ const __m256i rhs_mat_0145_31 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_3, 2), m3b); //B30(8-15) B31(8-15) B34(8-15) B35(8-15)
+ const __m256i rhs_mat_2367_31 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_3, 2), m3b); //B32(8-15) B33(8-15) B36(8-15) B37(8-15)
+
+ // Fifth sub block of the eight sub blocks processed in the iteration
+ const __m256i rhs_mat_0145_40 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_0, 4), m3b); //B40(0-7) B41(0-7) B44(0-7) B45(0-7)
+ const __m256i rhs_mat_2367_40 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_0, 4), m3b); //B42(0-7) B43(0-7) B46(0-7) B47(0-7)
+
+ const __m256i rhs_mat_0145_41 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_1, 4), m3b); //B40(8-15) B41(8-15) B44(8-15) B45(8-15)
+ const __m256i rhs_mat_2367_41 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_1, 4), m3b); //B42(8-15) B43(8-15) B46(8-15) B47(8-15)
+
+ // Sixth sub block of the eight sub blocks processed in the iteration
+ const __m256i rhs_mat_0145_50 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_2, 4), m3b); //B50(0-7) B51(0-7) B54(0-7) B55(0-7)
+ const __m256i rhs_mat_2367_50 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_2, 4), m3b); //B52(0-7) B53(0-7) B56(0-7) B57(0-7)
+
+ const __m256i rhs_mat_0145_51 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_3, 4), m3b); //B50(8-15) B51(8-15) B54(8-15) B55(8-15)
+ const __m256i rhs_mat_2367_51 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_3, 4), m3b); //B52(8-15) B53(8-15) B56(8-15) B57(8-15)
+
+ // Seventh sub block of the eight sub blocks processed in the iteration
+ const __m256i rhs_mat_0145_60 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_0, 6), m3b); //B60(0-7) B61(0-7) B64(0-7) B65(0-7)
+ const __m256i rhs_mat_2367_60 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_0, 6), m3b); //B62(0-7) B63(0-7) B66(0-7) B67(0-7)
+
+ const __m256i rhs_mat_0145_61 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_1, 6), m3b); //B60(8-15) B61(8-15) B64(8-15) B65(8-15)
+ const __m256i rhs_mat_2367_61 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_1, 6), m3b); //B62(8-15) B63(8-15) B66(8-15) B67(8-15)
+
+ // Eighth sub block of the eight sub blocks processed in the iteration
+ const __m256i rhs_mat_0145_70 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_2, 6), m3b); //B70(0-7) B71(0-7) B74(0-7) B75(0-7)
+ const __m256i rhs_mat_2367_70 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_2, 6), m3b); //B72(0-7) B73(0-7) B76(0-7) B77(0-7)
+
+ const __m256i rhs_mat_0145_71 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_3, 6), m3b); //B70(8-15) B71(8-15) B74(8-15) B75(8-15)
+ const __m256i rhs_mat_2367_71 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_3, 6), m3b); //B72(8-15) B73(8-15) B76(8-15) B77(8-15)
+
+ // Shuffle pattern one - right side input
+ const __m256i rhs_mat_0145_00_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_00, 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)
+ const __m256i rhs_mat_2367_00_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_00, 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)
+
+ const __m256i rhs_mat_0145_01_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_01, 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)
+ const __m256i rhs_mat_2367_01_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_01, 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)
+
+ const __m256i rhs_mat_0145_10_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_10, 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)
+ const __m256i rhs_mat_2367_10_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_10, 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)
+
+ const __m256i rhs_mat_0145_11_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_11, 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)
+ const __m256i rhs_mat_2367_11_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_11, 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)
+
+ const __m256i rhs_mat_0145_20_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_20, 136); //B20(0-3) B21(0-3) B20(0-3) B21(0-3) B24(0-3) B25(0-3) B24(0-3) B25(0-3)
+ const __m256i rhs_mat_2367_20_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_20, 136); //B22(0-3) B23(0-3) B22(0-3) B23(0-3) B26(0-3) B27(0-3) B26(0-3) B27(0-3)
+
+ const __m256i rhs_mat_0145_21_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_21, 136); //B20(8-11) B21(8-11) B20(8-11) B21(8-11) B24(8-11) B25(8-11) B24(8-11) B25(8-11)
+ const __m256i rhs_mat_2367_21_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_21, 136); //B22(8-11) B23(8-11) B22(8-11) B23(8-11) B26(8-11) B27(8-11) B26(8-11) B27(8-11)
+
+ const __m256i rhs_mat_0145_30_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_30, 136); //B30(0-3) B31(0-3) B30(0-3) B31(0-3) B34(0-3) B35(0-3) B34(0-3) B35(0-3)
+ const __m256i rhs_mat_2367_30_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_30, 136); //B32(0-3) B33(0-3) B32(0-3) B33(0-3) B36(0-3) B37(0-3) B36(0-3) B37(0-3)
+
+ const __m256i rhs_mat_0145_31_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_31, 136); //B30(8-11) B31(8-11) B30(8-11) B31(8-11) B34(8-11) B35(8-11) B34(8-11) B35(8-11
+ const __m256i rhs_mat_2367_31_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_31, 136); //B32(8-11) B33(8-11) B32(8-11) B33(8-11) B36(8-11) B37(8-11) B36(8-11) B37(8-11)
+
+ const __m256i rhs_mat_0145_40_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_40, 136); //B40(0-3) B41(0-3) B40(0-3) B41(0-3) B44(0-3) B45(0-3) B44(0-3) B45(0-3)
+ const __m256i rhs_mat_2367_40_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_40, 136); //B42(0-3) B43(0-3) B42(0-3) B43(0-3) B46(0-3) B47(0-3) B46(0-3) B47(0-3)
+
+ const __m256i rhs_mat_0145_41_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_41, 136); //B40(8-11) B41(8-11) B40(8-11) B41(8-11) B44(8-11) B45(8-11) B44(8-11) B45(8-11)
+ const __m256i rhs_mat_2367_41_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_41, 136); //B42(8-11) B43(8-11) B42(8-11) B43(8-11) B46(8-11) B47(8-11) B46(8-11) B47(8-11)
+
+ const __m256i rhs_mat_0145_50_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_50, 136); //B50(0-3) B51(0-3) B50(0-3) B51(0-3) B54(0-3) B55(0-3) B54(0-3) B55(0-3)
+ const __m256i rhs_mat_2367_50_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_50, 136); //B52(0-3) B53(0-3) B52(0-3) B53(0-3) B56(0-3) B57(0-3) B56(0-3) B57(0-3)
+
+ const __m256i rhs_mat_0145_51_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_51, 136); //B50(8-11) B51(8-11) B50(8-11) B51(8-11) B54(8-11) B55(8-11) B54(8-11) B55(8-11)
+ const __m256i rhs_mat_2367_51_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_51, 136); //B52(8-11) B53(8-11) B52(8-11) B53(8-11) B56(8-11) B57(8-11) B56(8-11) B57(8-11)
+
+ const __m256i rhs_mat_0145_60_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_60, 136); //B60(0-3) B61(0-3) B60(0-3) B61(0-3) B64(0-3) B65(0-3) B64(0-3) B65(0-3)
+ const __m256i rhs_mat_2367_60_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_60, 136); //B62(0-3) B63(0-3) B62(0-3) B63(0-3) B66(0-3) B67(0-3) B66(0-3) B67(0-3)
+
+ const __m256i rhs_mat_0145_61_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_61, 136); //B60(8-11) B61(8-11) B60(8-11) B61(8-11) B64(8-11) B65(8-11) B64(8-11) B65(8-11)
+ const __m256i rhs_mat_2367_61_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_61, 136); //B62(8-11) B63(8-11) B62(8-11) B63(8-11) B66(8-11) B67(8-11) B66(8-11) B67(8-11)
+
+ const __m256i rhs_mat_0145_70_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_70, 136); //B70(0-3) B71(0-3) B70(0-3) B71(0-3) B74(0-3) B75(0-3) B74(0-3) B75(0-3)
+ const __m256i rhs_mat_2367_70_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_70, 136); //B72(0-3) B73(0-3) B72(0-3) B73(0-3) B76(0-3) B77(0-3) B76(0-3) B77(0-3)
+
+ const __m256i rhs_mat_0145_71_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_71, 136); //B70(8-11) B71(8-11) B70(8-11) B71(8-11) B74(8-11) B75(8-11) B74(8-11) B75(8-11)
+ const __m256i rhs_mat_2367_71_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_71, 136); //B72(8-11) B73(8-11) B72(8-11) B73(8-11) B76(8-11) B77(8-11) B76(8-11) B77(8-11)
+
+
+ // Shuffle pattern two - right side input
+ const __m256i rhs_mat_0145_00_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_00, 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)
+ const __m256i rhs_mat_2367_00_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_00, 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)
+
+ const __m256i rhs_mat_0145_01_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_01, 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)
+ const __m256i rhs_mat_2367_01_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_01, 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)
+
+ const __m256i rhs_mat_0145_10_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_10, 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)
+ const __m256i rhs_mat_2367_10_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_10, 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)
+
+ const __m256i rhs_mat_0145_11_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_11, 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)
+ const __m256i rhs_mat_2367_11_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_11, 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)
+
+ const __m256i rhs_mat_0145_20_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_20, 221); //B20(4-7) B21(4-7) B20(4-7) B21(4-7) B24(4-7) B25(4-7) B24(4-7) B25(4-7)
+ const __m256i rhs_mat_2367_20_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_20, 221); //B22(4-7) B23(4-7) B22(4-7) B23(4-7) B26(4-7) B27(4-7) B26(4-7) B27(4-7)
+
+ const __m256i rhs_mat_0145_21_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_21, 221); //B20(12-15) B21(12-15) B20(12-15) B21(12-15) B24(12-15) B25(12-15) B24(12-15) B25(12-15)
+ const __m256i rhs_mat_2367_21_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_21, 221); //B22(12-15) B23(12-15) B22(12-15) B23(12-15) B26(12-15) B27(12-15) B26(12-15) B27(12-15)
+
+ const __m256i rhs_mat_0145_30_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_30, 221); //B30(4-7) B31(4-7) B30(4-7) B31(4-7) B34(4-7) B35(4-7) B34(4-7) B35(4-7)
+ const __m256i rhs_mat_2367_30_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_30, 221); //B32(4-7) B33(4-7) B32(4-7) B33(4-7) B36(4-7) B37(4-7) B36(4-7) B37(4-7)
+
+ const __m256i rhs_mat_0145_31_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_31, 221); //B30(12-15) B31(12-15) B30(12-15) B31(12-15) B34(12-15) B35(12-15) B34(12-15) B35(12-15)
+ const __m256i rhs_mat_2367_31_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_31, 221); //B32(12-15) B33(12-15) B32(12-15) B33(12-15) B36(12-15) B37(12-15) B36(12-15) B37(12-15)
+
+ const __m256i rhs_mat_0145_40_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_40, 221); //B40(4-7) B41(4-7) B40(4-7) B41(4-7) B44(4-7) B45(4-7) B44(4-7) B45(4-7)
+ const __m256i rhs_mat_2367_40_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_40, 221); //B42(4-7) B43(4-7) B42(4-7) B43(4-7) B46(4-7) B47(4-7) B46(4-7) B47(4-7)
+
+ const __m256i rhs_mat_0145_41_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_41, 221); //B40(12-15) B41(12-15) B40(12-15) B41(12-15) B44(12-15) B45(12-15) B44(12-15) B45(12-15)
+ const __m256i rhs_mat_2367_41_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_41, 221); //B42(12-15) B43(12-15) B42(12-15) B43(12-15) B46(12-15) B47(12-15) B46(12-15) B47(12-15)
+
+ const __m256i rhs_mat_0145_50_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_50, 221); //B50(4-7) B51(4-7) B50(4-7) B51(4-7) B54(4-7) B55(4-7) B54(4-7) B55(4-7)
+ const __m256i rhs_mat_2367_50_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_50, 221); //B52(4-7) B53(4-7) B52(4-7) B53(4-7) B56(4-7) B57(4-7) B56(4-7) B57(4-7)
+
+ const __m256i rhs_mat_0145_51_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_51, 221); //B50(12-15) B51(12-15) B50(12-15) B51(12-15) B54(12-15) B55(12-15) B54(12-15) B55(12-15)
+ const __m256i rhs_mat_2367_51_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_51, 221); //B52(12-15) B53(12-15) B52(12-15) B53(12-15) B56(12-15) B57(12-15) B56(12-15) B57(12-15)
+
+ const __m256i rhs_mat_0145_60_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_60, 221); //B60(4-7) B61(4-7) B60(4-7) B61(4-7) B64(4-7) B65(4-7) B64(4-7) B65(4-7)
+ const __m256i rhs_mat_2367_60_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_60, 221); //B62(4-7) B63(4-7) B62(4-7) B63(4-7) B66(4-7) B67(4-7) B66(4-7) B67(4-7)
+
+ const __m256i rhs_mat_0145_61_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_61, 221); //B60(12-15) B61(12-15) B60(12-15) B61(12-15) B64(12-15) B65(12-15) B64(12-15) B65(12-15)
+ const __m256i rhs_mat_2367_61_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_61, 221); //B62(12-15) B63(12-15) B62(12-15) B63(12-15) B66(12-15) B67(12-15) B66(12-15) B67(12-15)
+
+ const __m256i rhs_mat_0145_70_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_70, 221); //B70(4-7) B71(4-7) B70(4-7) B71(4-7) B74(4-7) B75(4-7) B74(4-7) B75(4-7)
+ const __m256i rhs_mat_2367_70_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_70, 221); //B72(4-7) B73(4-7) B72(4-7) B73(4-7) B76(4-7) B77(4-7) B76(4-7) B77(4-7)
+
+ const __m256i rhs_mat_0145_71_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_71, 221); //B70(12-15) B71(12-15) B70(12-15) B71(12-15) B74(12-15) B75(12-15) B74(12-15) B75(12-15)
+ const __m256i rhs_mat_2367_71_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_71, 221); //B72(12-15) B73(12-15) B72(12-15) B73(12-15) B76(12-15) B77(12-15) B76(12-15) B77(12-15)
+
+ //Scales and Mins of corresponding sub blocks from different Q2_K structures are stored together
+ //s00 m00 s01 m01 s10 m10 s11 m11 s20 m20 s21 m21 s30 m30 s31 m31 s40 m40 s41 m41 s50 m50 s51 m51 s60 m60 s61 m61 s70 m70 s71 m71
+
+ // Combine mins and scales for sub-blocks: 0-1, 2-3, 4-5, 6-7 in the sb loop
+ const __m128i mins_and_scales_01 = _mm_loadu_si128((const __m128i *)(b_ptr[b].scales + sb * 64));
+ const __m128i mins_and_scales_23 = _mm_loadu_si128((const __m128i *)(b_ptr[b].scales + 16 + sb * 64));
+ const __m128i mins_and_scales_45 = _mm_loadu_si128((const __m128i *)(b_ptr[b].scales + 32 + sb * 64));
+ const __m128i mins_and_scales_67 = _mm_loadu_si128((const __m128i *)(b_ptr[b].scales + 48 + sb * 64));
+
+ // Extract scales which is lower half from mins_and_scales
+ const __m128i scales_01 = _mm_and_si128(mins_and_scales_01, m4b_sse);
+ const __m128i scales_23 = _mm_and_si128(mins_and_scales_23, m4b_sse);
+ const __m128i scales_45 = _mm_and_si128(mins_and_scales_45, m4b_sse);
+ const __m128i scales_67 = _mm_and_si128(mins_and_scales_67, m4b_sse);
+
+ // Extract mins which is upper half from mins_and_scales
+ const __m256i mins_01 = _mm256_cvtepu8_epi16(_mm_and_si128(_mm_srli_epi16(mins_and_scales_01, 4), m4b_sse));
+ const __m256i mins_23 = _mm256_cvtepu8_epi16(_mm_and_si128(_mm_srli_epi16(mins_and_scales_23, 4), m4b_sse));
+ const __m256i mins_45 = _mm256_cvtepu8_epi16(_mm_and_si128(_mm_srli_epi16(mins_and_scales_45, 4), m4b_sse));
+ const __m256i mins_67 = _mm256_cvtepu8_epi16(_mm_and_si128(_mm_srli_epi16(mins_and_scales_67, 4), m4b_sse));
+
+ const __m256i scales_0 = _mm256_cvtepu8_epi16(_mm_shuffle_epi8(scales_01, scalesmask1_sse));
+ const __m256i scales_1 = _mm256_cvtepu8_epi16(_mm_shuffle_epi8(scales_01, scalesmask2_sse));
+
+ const __m256i scales_2 = _mm256_cvtepu8_epi16(_mm_shuffle_epi8(scales_23, scalesmask1_sse));
+ const __m256i scales_3 = _mm256_cvtepu8_epi16(_mm_shuffle_epi8(scales_23, scalesmask2_sse));
+
+ const __m256i scales_4 = _mm256_cvtepu8_epi16(_mm_shuffle_epi8(scales_45, scalesmask1_sse));
+ const __m256i scales_5 = _mm256_cvtepu8_epi16(_mm_shuffle_epi8(scales_45, scalesmask2_sse));
+
+ const __m256i scales_6 = _mm256_cvtepu8_epi16(_mm_shuffle_epi8(scales_67, scalesmask1_sse));
+ const __m256i scales_7 = _mm256_cvtepu8_epi16(_mm_shuffle_epi8(scales_67, scalesmask2_sse));
+
+ const __m256i scale_0145_0 = _mm256_shuffle_epi32(scales_0, 68);
+ const __m256i scale_2367_0 = _mm256_shuffle_epi32(scales_0, 238);
+
+ const __m256i scale_0145_1 = _mm256_shuffle_epi32(scales_1, 68);
+ const __m256i scale_2367_1 = _mm256_shuffle_epi32(scales_1, 238);
+
+ const __m256i scale_0145_2 = _mm256_shuffle_epi32(scales_2, 68);
+ const __m256i scale_2367_2 = _mm256_shuffle_epi32(scales_2, 238);
+
+ const __m256i scale_0145_3 = _mm256_shuffle_epi32(scales_3, 68);
+ const __m256i scale_2367_3 = _mm256_shuffle_epi32(scales_3, 238);
+
+ const __m256i scale_0145_4 = _mm256_shuffle_epi32(scales_4, 68);
+ const __m256i scale_2367_4 = _mm256_shuffle_epi32(scales_4, 238);
+
+ const __m256i scale_0145_5 = _mm256_shuffle_epi32(scales_5, 68);
+ const __m256i scale_2367_5 = _mm256_shuffle_epi32(scales_5, 238);
+
+ const __m256i scale_0145_6 = _mm256_shuffle_epi32(scales_6, 68);
+ const __m256i scale_2367_6 = _mm256_shuffle_epi32(scales_6, 238);
+
+ const __m256i scale_0145_7 = _mm256_shuffle_epi32(scales_7, 68);
+ const __m256i scale_2367_7 = _mm256_shuffle_epi32(scales_7, 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 into a 256 bit vector
+ __m256i lhs_mat_0123_00 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 512 * sb)));
+ __m256i lhs_mat_01_00 = _mm256_permute2f128_si256(lhs_mat_0123_00, lhs_mat_0123_00, 0);
+ __m256i lhs_mat_23_00 = _mm256_permute2f128_si256(lhs_mat_0123_00, lhs_mat_0123_00, 17);
+ __m256i lhs_mat_0123_01 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 32 + 512 * sb)));
+ __m256i lhs_mat_01_01 = _mm256_permute2f128_si256(lhs_mat_0123_01, lhs_mat_0123_01, 0);
+ __m256i lhs_mat_23_01 = _mm256_permute2f128_si256(lhs_mat_0123_01, lhs_mat_0123_01, 17);
+ __m256i lhs_mat_0123_10 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 64 + 512 * sb)));
+ __m256i lhs_mat_01_10 = _mm256_permute2f128_si256(lhs_mat_0123_10, lhs_mat_0123_10, 0);
+ __m256i lhs_mat_23_10 = _mm256_permute2f128_si256(lhs_mat_0123_10, lhs_mat_0123_10, 17);
+ __m256i lhs_mat_0123_11 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 96 + 512 * sb)));
+ __m256i lhs_mat_01_11 = _mm256_permute2f128_si256(lhs_mat_0123_11, lhs_mat_0123_11, 0);
+ __m256i lhs_mat_23_11 = _mm256_permute2f128_si256(lhs_mat_0123_11, lhs_mat_0123_11, 17);
+ __m256i lhs_mat_0123_20 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 128 + 512 * sb)));
+ __m256i lhs_mat_01_20 = _mm256_permute2f128_si256(lhs_mat_0123_20, lhs_mat_0123_20, 0);
+ __m256i lhs_mat_23_20 = _mm256_permute2f128_si256(lhs_mat_0123_20, lhs_mat_0123_20, 17);
+ __m256i lhs_mat_0123_21 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 160 + 512 * sb)));
+ __m256i lhs_mat_01_21 = _mm256_permute2f128_si256(lhs_mat_0123_21, lhs_mat_0123_21, 0);
+ __m256i lhs_mat_23_21 = _mm256_permute2f128_si256(lhs_mat_0123_21, lhs_mat_0123_21, 17);
+ __m256i lhs_mat_0123_30 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 192 + 512 * sb)));
+ __m256i lhs_mat_01_30 = _mm256_permute2f128_si256(lhs_mat_0123_30, lhs_mat_0123_30, 0);
+ __m256i lhs_mat_23_30 = _mm256_permute2f128_si256(lhs_mat_0123_30, lhs_mat_0123_30, 17);
+ __m256i lhs_mat_0123_31 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 224 + 512 * sb)));
+ __m256i lhs_mat_01_31 = _mm256_permute2f128_si256(lhs_mat_0123_31, lhs_mat_0123_31, 0);
+ __m256i lhs_mat_23_31 = _mm256_permute2f128_si256(lhs_mat_0123_31, lhs_mat_0123_31, 17);
+
+ __m256i lhs_mat_0123_40 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 256 + 512 * sb)));
+ __m256i lhs_mat_01_40 = _mm256_permute2f128_si256(lhs_mat_0123_40, lhs_mat_0123_40, 0);
+ __m256i lhs_mat_23_40 = _mm256_permute2f128_si256(lhs_mat_0123_40, lhs_mat_0123_40, 17);
+ __m256i lhs_mat_0123_41 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 288 + 512 * sb)));
+ __m256i lhs_mat_01_41 = _mm256_permute2f128_si256(lhs_mat_0123_41, lhs_mat_0123_41, 0);
+ __m256i lhs_mat_23_41 = _mm256_permute2f128_si256(lhs_mat_0123_41, lhs_mat_0123_41, 17);
+ __m256i lhs_mat_0123_50 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 320 + 512 * sb)));
+ __m256i lhs_mat_01_50 = _mm256_permute2f128_si256(lhs_mat_0123_50, lhs_mat_0123_50, 0);
+ __m256i lhs_mat_23_50 = _mm256_permute2f128_si256(lhs_mat_0123_50, lhs_mat_0123_50, 17);
+ __m256i lhs_mat_0123_51 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 352 + 512 * sb)));
+ __m256i lhs_mat_01_51 = _mm256_permute2f128_si256(lhs_mat_0123_51, lhs_mat_0123_51, 0);
+ __m256i lhs_mat_23_51 = _mm256_permute2f128_si256(lhs_mat_0123_51, lhs_mat_0123_51, 17);
+ __m256i lhs_mat_0123_60 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 384 + 512 * sb)));
+ __m256i lhs_mat_01_60 = _mm256_permute2f128_si256(lhs_mat_0123_60, lhs_mat_0123_60, 0);
+ __m256i lhs_mat_23_60 = _mm256_permute2f128_si256(lhs_mat_0123_60, lhs_mat_0123_60, 17);
+ __m256i lhs_mat_0123_61 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 416 + 512 * sb)));
+ __m256i lhs_mat_01_61 = _mm256_permute2f128_si256(lhs_mat_0123_61, lhs_mat_0123_61, 0);
+ __m256i lhs_mat_23_61 = _mm256_permute2f128_si256(lhs_mat_0123_61, lhs_mat_0123_61, 17);
+ __m256i lhs_mat_0123_70 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 448 + 512 * sb)));
+ __m256i lhs_mat_01_70 = _mm256_permute2f128_si256(lhs_mat_0123_70, lhs_mat_0123_70, 0);
+ __m256i lhs_mat_23_70 = _mm256_permute2f128_si256(lhs_mat_0123_70, lhs_mat_0123_70, 17);
+ __m256i lhs_mat_0123_71 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 480 + 512 * sb)));
+ __m256i lhs_mat_01_71 = _mm256_permute2f128_si256(lhs_mat_0123_71, lhs_mat_0123_71, 0);
+ __m256i lhs_mat_23_71 = _mm256_permute2f128_si256(lhs_mat_0123_71, lhs_mat_0123_71, 17);
+
+ // Bsums are loaded for the different Q8_K blocks
+ __m128i lhs_raw_bsums_01_0123 = _mm_loadu_si128((const __m128i *)((a_ptrs[rp][b].bsums + 32 * sb)));
+ __m128i lhs_raw_bsums_23_0123 = _mm_loadu_si128((const __m128i *)(a_ptrs[rp][b].bsums + 8 + 32 * sb));
+ __m128i lhs_raw_bsums_01_4567 = _mm_loadu_si128((const __m128i *)((a_ptrs[rp][b].bsums + 16 + 32 * sb)));
+ __m128i lhs_raw_bsums_23_4567 = _mm_loadu_si128((const __m128i *)(a_ptrs[rp][b].bsums + 24 + 32 * sb));
+
+ // Shuffle pattern one - left side input
+ const __m256i lhs_mat_01_00_sp1 = _mm256_shuffle_epi32(lhs_mat_01_00, 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)
+ const __m256i lhs_mat_23_00_sp1 = _mm256_shuffle_epi32(lhs_mat_23_00, 160); //A02(0-3) A03(0-3) A02(0-3) A03(0-3) A02(0-3) A03(0-3) A02(0-3) A03(0-3)
+
+ const __m256i lhs_mat_01_01_sp1 = _mm256_shuffle_epi32(lhs_mat_01_01, 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)
+ const __m256i lhs_mat_23_01_sp1 = _mm256_shuffle_epi32(lhs_mat_23_01, 160); //A02(8-11) A03(8-11) A02(8-11) A03(8-11) A02(8-11) A03(8-11) A02(8-11) A03(8-11)
+
+ const __m256i lhs_mat_01_10_sp1 = _mm256_shuffle_epi32(lhs_mat_01_10, 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)
+ const __m256i lhs_mat_23_10_sp1 = _mm256_shuffle_epi32(lhs_mat_23_10, 160); //A12(0-3) A13(0-3) A12(0-3) A13(0-3) A12(0-3) A13(0-3) A12(0-3) A13(0-3)
+
+ const __m256i lhs_mat_01_11_sp1 = _mm256_shuffle_epi32(lhs_mat_01_11, 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)
+ const __m256i lhs_mat_23_11_sp1 = _mm256_shuffle_epi32(lhs_mat_23_11, 160); //A12(8-11) A13(8-11) A12(8-11) A13(8-11) A12(8-11) A13(8-11) A12(8-11) A13(8-11)
+
+ const __m256i lhs_mat_01_20_sp1 = _mm256_shuffle_epi32(lhs_mat_01_20, 160); //A20(0-3) A20(0-3) A21(0-3) A21(0-3) A20(0-3) A20(0-3) A21(0-3) A21(0-3)
+ const __m256i lhs_mat_23_20_sp1 = _mm256_shuffle_epi32(lhs_mat_23_20, 160); //A22(0-3) A23(0-3) A22(0-3) A23(0-3) A22(0-3) A23(0-3) A22(0-3) A23(0-3)
+
+ const __m256i lhs_mat_01_21_sp1 = _mm256_shuffle_epi32(lhs_mat_01_21, 160); //A20(8-11) A20(8-11) A21(8-11) A21(8-11) A20(8-11) A20(8-11) A21(8-11) A21(8-11)
+ const __m256i lhs_mat_23_21_sp1 = _mm256_shuffle_epi32(lhs_mat_23_21, 160); //A22(8-11) A23(8-11) A22(8-11) A23(8-11) A22(8-11) A23(8-11) A22(8-11) A23(8-11)
+
+ const __m256i lhs_mat_01_30_sp1 = _mm256_shuffle_epi32(lhs_mat_01_30, 160); //A30(0-3) A30(0-3) A31(0-3) A31(0-3) A30(0-3) A30(0-3) A31(0-3) A31(0-3)
+ const __m256i lhs_mat_23_30_sp1 = _mm256_shuffle_epi32(lhs_mat_23_30, 160); //A32(0-3) A33(0-3) A32(0-3) A33(0-3) A32(0-3) A33(0-3) A32(0-3) A33(0-3)
+
+ const __m256i lhs_mat_01_31_sp1 = _mm256_shuffle_epi32(lhs_mat_01_31, 160); //A30(8-11) A30(8-11) A31(8-11) A31(8-11) A30(8-11) A30(8-11) A31(8-11) A31(8-11)
+ const __m256i lhs_mat_23_31_sp1 = _mm256_shuffle_epi32(lhs_mat_23_31, 160); //A32(8-11) A33(8-11) A32(8-11) A33(8-11) A32(8-11) A33(8-11) A32(8-11) A33(8-11)
+
+ const __m256i lhs_mat_01_40_sp1 = _mm256_shuffle_epi32(lhs_mat_01_40, 160); //A40(0-3) A40(0-3) A41(0-3) A41(0-3) A40(0-3) A40(0-3) A41(0-3) A41(0-3)
+ const __m256i lhs_mat_23_40_sp1 = _mm256_shuffle_epi32(lhs_mat_23_40, 160); //A42(0-3) A43(0-3) A42(0-3) A43(0-3) A42(0-3) A43(0-3) A42(0-3) A43(0-3)
+
+ const __m256i lhs_mat_01_41_sp1 = _mm256_shuffle_epi32(lhs_mat_01_41, 160); //A40(8-11) A40(8-11) A41(8-11) A41(8-11) A40(8-11) A40(8-11) A41(8-11) A41(8-11)
+ const __m256i lhs_mat_23_41_sp1 = _mm256_shuffle_epi32(lhs_mat_23_41, 160); //A42(8-11) A43(8-11) A42(8-11) A43(8-11) A42(8-11) A43(8-11) A42(8-11) A43(8-11)
+
+ const __m256i lhs_mat_01_50_sp1 = _mm256_shuffle_epi32(lhs_mat_01_50, 160); //A50(0-3) A50(0-3) A51(0-3) A51(0-3) A50(0-3) A50(0-3) A51(0-3) A51(0-3)
+ const __m256i lhs_mat_23_50_sp1 = _mm256_shuffle_epi32(lhs_mat_23_50, 160); //A52(0-3) A53(0-3) A52(0-3) A53(0-3) A52(0-3) A53(0-3) A52(0-3) A53(0-3)
+
+ const __m256i lhs_mat_01_51_sp1 = _mm256_shuffle_epi32(lhs_mat_01_51, 160); //A50(8-11) A50(8-11) A51(8-11) A51(8-11) A50(8-11) A50(8-11) A51(8-11) A51(8-11)
+ const __m256i lhs_mat_23_51_sp1 = _mm256_shuffle_epi32(lhs_mat_23_51, 160); //A52(8-11) A53(8-11) A52(8-11) A53(8-11) A52(8-11) A53(8-11) A52(8-11) A53(8-11)
+
+ const __m256i lhs_mat_01_60_sp1 = _mm256_shuffle_epi32(lhs_mat_01_60, 160); //A60(0-3) A60(0-3) A61(0-3) A61(0-3) A60(0-3) A60(0-3) A61(0-3) A61(0-3)
+ const __m256i lhs_mat_23_60_sp1 = _mm256_shuffle_epi32(lhs_mat_23_60, 160); //A62(0-3) A63(0-3) A62(0-3) A63(0-3) A62(0-3) A63(0-3) A62(0-3) A63(0-3)
+
+ const __m256i lhs_mat_01_61_sp1 = _mm256_shuffle_epi32(lhs_mat_01_61, 160); //A60(8-11) A60(8-11) A61(8-11) A61(8-11) A60(8-11) A60(8-11) A61(8-11) A61(8-11)
+ const __m256i lhs_mat_23_61_sp1 = _mm256_shuffle_epi32(lhs_mat_23_61, 160); //A62(8-11) A63(8-11) A62(8-11) A63(8-11) A62(8-11) A63(8-11) A62(8-11) A63(8-11)
+
+ const __m256i lhs_mat_01_70_sp1 = _mm256_shuffle_epi32(lhs_mat_01_70, 160); //A70(0-3) A70(0-3) A71(0-3) A71(0-3) A70(0-3) A70(0-3) A71(0-3) A71(0-3)
+ const __m256i lhs_mat_23_70_sp1 = _mm256_shuffle_epi32(lhs_mat_23_70, 160); //A72(0-3) A73(0-3) A72(0-3) A73(0-3) A72(0-3) A73(0-3) A72(0-3) A73(0-3)
+
+ const __m256i lhs_mat_01_71_sp1 = _mm256_shuffle_epi32(lhs_mat_01_71, 160); //A70(8-11) A70(8-11) A71(8-11) A71(8-11) A70(8-11) A70(8-11) A71(8-11) A71(8-11)
+ const __m256i lhs_mat_23_71_sp1 = _mm256_shuffle_epi32(lhs_mat_23_71, 160); //A72(8-11) A73(8-11) A72(8-11) A73(8-11) A72(8-11) A73(8-11) A72(8-11) A73(8-11)
+
+ // Shuffle pattern two- left side input
+ const __m256i lhs_mat_01_00_sp2 = _mm256_shuffle_epi32(lhs_mat_01_00, 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)
+ const __m256i lhs_mat_23_00_sp2 = _mm256_shuffle_epi32(lhs_mat_23_00, 245); //A02(4-7) A03(4-7) A02(4-7) A03(4-7) A02(4-7) A03(4-7) A02(4-7) A03(4-7)
+
+ const __m256i lhs_mat_01_01_sp2 = _mm256_shuffle_epi32(lhs_mat_01_01, 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)
+ const __m256i lhs_mat_23_01_sp2 = _mm256_shuffle_epi32(lhs_mat_23_01, 245); //A02(12-15) A03(12-15) A02(12-15) A03(12-15) A02(12-15) A03(12-15) A02(12-15) A03(12-15)
+
+ const __m256i lhs_mat_01_10_sp2 = _mm256_shuffle_epi32(lhs_mat_01_10, 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)
+ const __m256i lhs_mat_23_10_sp2 = _mm256_shuffle_epi32(lhs_mat_23_10, 245); //A12(4-7) A13(4-7) A12(4-7) A13(4-7) A12(4-7) A13(4-7) A12(4-7) A13(4-7)
+
+ const __m256i lhs_mat_01_11_sp2 = _mm256_shuffle_epi32(lhs_mat_01_11, 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)
+ const __m256i lhs_mat_23_11_sp2 = _mm256_shuffle_epi32(lhs_mat_23_11, 245); //A12(12-15) A13(12-15) A12(12-15) A13(12-15) A12(12-15) A13(12-15) A12(12-15) A13(12-15)
+
+ const __m256i lhs_mat_01_20_sp2 = _mm256_shuffle_epi32(lhs_mat_01_20, 245); //A20(4-7) A20(4-7) A21(4-7) A21(4-7) A20(4-7) A20(4-7) A21(4-7) A21(4-7)
+ const __m256i lhs_mat_23_20_sp2 = _mm256_shuffle_epi32(lhs_mat_23_20, 245); //A22(4-7) A23(4-7) A22(4-7) A23(4-7) A22(4-7) A23(4-7) A22(4-7) A23(4-7)
+
+ const __m256i lhs_mat_01_21_sp2 = _mm256_shuffle_epi32(lhs_mat_01_21, 245); //A20(12-15) A20(12-15) A21(12-15) A21(12-15) A20(12-15) A20(12-15) A21(12-15) A21(12-15)
+ const __m256i lhs_mat_23_21_sp2 = _mm256_shuffle_epi32(lhs_mat_23_21, 245); //A22(12-15) A23(12-15) A22(12-15) A23(12-15) A22(12-15) A23(12-15) A22(12-15) A23(12-15)
+
+ const __m256i lhs_mat_01_30_sp2 = _mm256_shuffle_epi32(lhs_mat_01_30, 245); //A30(4-7) A30(4-7) A31(4-7) A31(4-7) A30(4-7) A30(4-7) A31(4-7) A31(4-7)
+ const __m256i lhs_mat_23_30_sp2 = _mm256_shuffle_epi32(lhs_mat_23_30, 245); //A32(4-7) A33(4-7) A32(4-7) A33(4-7) A32(4-7) A33(4-7) A32(4-7) A33(4-7)
+
+ const __m256i lhs_mat_01_31_sp2 = _mm256_shuffle_epi32(lhs_mat_01_31, 245); //A30(12-15) A30(12-15) A31(12-15) A31(12-15) A30(12-15) A30(12-15) A31(12-15) A31(12-15)
+ const __m256i lhs_mat_23_31_sp2 = _mm256_shuffle_epi32(lhs_mat_23_31, 245); //A32(12-15) A33(12-15) A32(12-15) A33(12-15) A32(12-15) A33(12-15) A32(12-15) A33(12-15)
+
+ const __m256i lhs_mat_01_40_sp2 = _mm256_shuffle_epi32(lhs_mat_01_40, 245); //A40(4-7) A40(4-7) A41(4-7) A41(4-7) A40(4-7) A40(4-7) A41(4-7) A41(4-7)
+ const __m256i lhs_mat_23_40_sp2 = _mm256_shuffle_epi32(lhs_mat_23_40, 245); //A42(4-7) A43(4-7) A42(4-7) A43(4-7) A42(4-7) A43(4-7) A42(4-7) A43(4-7)
+
+ const __m256i lhs_mat_01_41_sp2 = _mm256_shuffle_epi32(lhs_mat_01_41, 245); //A40(12-15) A40(12-15) A41(12-15) A41(12-15) A40(12-15) A40(12-15) A41(12-15) A41(12-15)
+ const __m256i lhs_mat_23_41_sp2 = _mm256_shuffle_epi32(lhs_mat_23_41, 245); //A42(12-15) A43(12-15) A42(12-15) A43(12-15) A42(12-15) A43(12-15) A42(12-15) A43(12-15)
+
+ const __m256i lhs_mat_01_50_sp2 = _mm256_shuffle_epi32(lhs_mat_01_50, 245); //A50(4-7) A50(4-7) A51(4-7) A51(4-7) A50(4-7) A50(4-7) A51(4-7) A51(4-7)
+ const __m256i lhs_mat_23_50_sp2 = _mm256_shuffle_epi32(lhs_mat_23_50, 245); //A52(4-7) A53(4-7) A52(4-7) A53(4-7) A52(4-7) A53(4-7) A52(4-7) A53(4-7)
+
+ const __m256i lhs_mat_01_51_sp2 = _mm256_shuffle_epi32(lhs_mat_01_51, 245); //A50(12-15) A50(12-15) A51(12-15) A51(12-15) A50(12-15) A50(12-15) A51(12-15) A51(12-15)
+ const __m256i lhs_mat_23_51_sp2 = _mm256_shuffle_epi32(lhs_mat_23_51, 245); //A52(12-15) A53(12-15) A52(12-15) A53(12-15) A52(12-15) A53(12-15) A52(12-15) A53(12-15)
+
+ const __m256i lhs_mat_01_60_sp2 = _mm256_shuffle_epi32(lhs_mat_01_60, 245); //A60(4-7) A60(4-7) A61(4-7) A61(4-7) A60(4-7) A60(4-7) A61(4-7) A61(4-7)
+ const __m256i lhs_mat_23_60_sp2 = _mm256_shuffle_epi32(lhs_mat_23_60, 245); //A62(4-7) A63(4-7) A62(4-7) A63(4-7) A62(4-7) A63(4-7) A62(4-7) A63(4-7)
+
+ const __m256i lhs_mat_01_61_sp2 = _mm256_shuffle_epi32(lhs_mat_01_61, 245); //A60(12-15) A60(12-15) A61(12-15) A61(12-15) A60(12-15) A60(12-15) A61(12-15) A61(12-15)
+ const __m256i lhs_mat_23_61_sp2 = _mm256_shuffle_epi32(lhs_mat_23_61, 245); //A62(12-15) A63(12-15) A62(12-15) A63(12-15) A62(12-15) A63(12-15) A62(12-15) A63(12-15)
+
+ const __m256i lhs_mat_01_70_sp2 = _mm256_shuffle_epi32(lhs_mat_01_70, 245); //A70(4-7) A70(4-7) A71(4-7) A71(4-7) A70(4-7) A70(4-7) A71(4-7) A71(4-7)
+ const __m256i lhs_mat_23_70_sp2 = _mm256_shuffle_epi32(lhs_mat_23_70, 245); //A72(4-7) A73(4-7) A72(4-7) A73(4-7) A72(4-7) A73(4-7) A72(4-7) A73(4-7)
+
+ const __m256i lhs_mat_01_71_sp2 = _mm256_shuffle_epi32(lhs_mat_01_71, 245); //A70(12-15) A70(12-15) A71(12-15) A71(12-15) A70(12-15) A70(12-15) A71(12-15) A71(12-15)
+ const __m256i lhs_mat_23_71_sp2 = _mm256_shuffle_epi32(lhs_mat_23_71, 245); //A72(12-15) A73(12-15) A72(12-15) A73(12-15) A72(12-15) A73(12-15) A72(12-15) A73(12-15)
+
+ // 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
+ __m256i iacc_mat_00_0_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_00_sp1, lhs_mat_01_00_sp1),_mm256_maddubs_epi16(rhs_mat_0145_01_sp1, lhs_mat_01_01_sp1));
+ __m256i iacc_mat_01_0_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_00_sp1, lhs_mat_01_00_sp1),_mm256_maddubs_epi16(rhs_mat_2367_01_sp1, lhs_mat_01_01_sp1));
+
+ __m256i iacc_mat_10_0_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_00_sp1, lhs_mat_23_00_sp1),_mm256_maddubs_epi16(rhs_mat_0145_01_sp1, lhs_mat_23_01_sp1));
+ __m256i iacc_mat_11_0_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_00_sp1, lhs_mat_23_00_sp1),_mm256_maddubs_epi16(rhs_mat_2367_01_sp1, lhs_mat_23_01_sp1));
+
+ __m256i iacc_mat_00_1_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_10_sp1, lhs_mat_01_10_sp1),_mm256_maddubs_epi16(rhs_mat_0145_11_sp1, lhs_mat_01_11_sp1));
+ __m256i iacc_mat_01_1_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_10_sp1, lhs_mat_01_10_sp1),_mm256_maddubs_epi16(rhs_mat_2367_11_sp1, lhs_mat_01_11_sp1));
+
+ __m256i iacc_mat_10_1_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_10_sp1, lhs_mat_23_10_sp1),_mm256_maddubs_epi16(rhs_mat_0145_11_sp1, lhs_mat_23_11_sp1));
+ __m256i iacc_mat_11_1_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_10_sp1, lhs_mat_23_10_sp1),_mm256_maddubs_epi16(rhs_mat_2367_11_sp1, lhs_mat_23_11_sp1));
+
+ __m256i iacc_mat_00_2_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_20_sp1, lhs_mat_01_20_sp1),_mm256_maddubs_epi16(rhs_mat_0145_21_sp1, lhs_mat_01_21_sp1));
+ __m256i iacc_mat_01_2_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_20_sp1, lhs_mat_01_20_sp1),_mm256_maddubs_epi16(rhs_mat_2367_21_sp1, lhs_mat_01_21_sp1));
+
+ __m256i iacc_mat_10_2_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_20_sp1, lhs_mat_23_20_sp1),_mm256_maddubs_epi16(rhs_mat_0145_21_sp1, lhs_mat_23_21_sp1));
+ __m256i iacc_mat_11_2_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_20_sp1, lhs_mat_23_20_sp1),_mm256_maddubs_epi16(rhs_mat_2367_21_sp1, lhs_mat_23_21_sp1));
+
+ __m256i iacc_mat_00_3_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_30_sp1, lhs_mat_01_30_sp1),_mm256_maddubs_epi16(rhs_mat_0145_31_sp1, lhs_mat_01_31_sp1));
+ __m256i iacc_mat_01_3_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_30_sp1, lhs_mat_01_30_sp1),_mm256_maddubs_epi16(rhs_mat_2367_31_sp1, lhs_mat_01_31_sp1));
+
+ __m256i iacc_mat_10_3_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_30_sp1, lhs_mat_23_30_sp1),_mm256_maddubs_epi16(rhs_mat_0145_31_sp1, lhs_mat_23_31_sp1));
+ __m256i iacc_mat_11_3_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_30_sp1, lhs_mat_23_30_sp1),_mm256_maddubs_epi16(rhs_mat_2367_31_sp1, lhs_mat_23_31_sp1));
+
+ __m256i iacc_mat_00_4_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_40_sp1, lhs_mat_01_40_sp1),_mm256_maddubs_epi16(rhs_mat_0145_41_sp1, lhs_mat_01_41_sp1));
+ __m256i iacc_mat_01_4_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_40_sp1, lhs_mat_01_40_sp1),_mm256_maddubs_epi16(rhs_mat_2367_41_sp1, lhs_mat_01_41_sp1));
+
+ __m256i iacc_mat_10_4_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_40_sp1, lhs_mat_23_40_sp1),_mm256_maddubs_epi16(rhs_mat_0145_41_sp1, lhs_mat_23_41_sp1));
+ __m256i iacc_mat_11_4_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_40_sp1, lhs_mat_23_40_sp1),_mm256_maddubs_epi16(rhs_mat_2367_41_sp1, lhs_mat_23_41_sp1));
+
+ __m256i iacc_mat_00_5_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_50_sp1, lhs_mat_01_50_sp1),_mm256_maddubs_epi16(rhs_mat_0145_51_sp1, lhs_mat_01_51_sp1));
+ __m256i iacc_mat_01_5_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_50_sp1, lhs_mat_01_50_sp1),_mm256_maddubs_epi16(rhs_mat_2367_51_sp1, lhs_mat_01_51_sp1));
+
+ __m256i iacc_mat_10_5_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_50_sp1, lhs_mat_23_50_sp1),_mm256_maddubs_epi16(rhs_mat_0145_51_sp1, lhs_mat_23_51_sp1));
+ __m256i iacc_mat_11_5_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_50_sp1, lhs_mat_23_50_sp1),_mm256_maddubs_epi16(rhs_mat_2367_51_sp1, lhs_mat_23_51_sp1));
+
+ __m256i iacc_mat_00_6_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_60_sp1, lhs_mat_01_60_sp1),_mm256_maddubs_epi16(rhs_mat_0145_61_sp1, lhs_mat_01_61_sp1));
+ __m256i iacc_mat_01_6_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_60_sp1, lhs_mat_01_60_sp1),_mm256_maddubs_epi16(rhs_mat_2367_61_sp1, lhs_mat_01_61_sp1));
+
+ __m256i iacc_mat_10_6_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_60_sp1, lhs_mat_23_60_sp1),_mm256_maddubs_epi16(rhs_mat_0145_61_sp1, lhs_mat_23_61_sp1));
+ __m256i iacc_mat_11_6_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_60_sp1, lhs_mat_23_60_sp1),_mm256_maddubs_epi16(rhs_mat_2367_61_sp1, lhs_mat_23_61_sp1));
+
+ __m256i iacc_mat_00_7_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_70_sp1, lhs_mat_01_70_sp1),_mm256_maddubs_epi16(rhs_mat_0145_71_sp1, lhs_mat_01_71_sp1));
+ __m256i iacc_mat_01_7_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_70_sp1, lhs_mat_01_70_sp1),_mm256_maddubs_epi16(rhs_mat_2367_71_sp1, lhs_mat_01_71_sp1));
+
+ __m256i iacc_mat_10_7_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_70_sp1, lhs_mat_23_70_sp1),_mm256_maddubs_epi16(rhs_mat_0145_71_sp1, lhs_mat_23_71_sp1));
+ __m256i iacc_mat_11_7_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_70_sp1, lhs_mat_23_70_sp1),_mm256_maddubs_epi16(rhs_mat_2367_71_sp1, lhs_mat_23_71_sp1));
+
+
+ __m256i iacc_mat_00_0_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_00_sp2, lhs_mat_01_00_sp2),_mm256_maddubs_epi16(rhs_mat_0145_01_sp2, lhs_mat_01_01_sp2));
+ __m256i iacc_mat_01_0_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_00_sp2, lhs_mat_01_00_sp2),_mm256_maddubs_epi16(rhs_mat_2367_01_sp2, lhs_mat_01_01_sp2));
+
+ __m256i iacc_mat_10_0_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_00_sp2, lhs_mat_23_00_sp2),_mm256_maddubs_epi16(rhs_mat_0145_01_sp2, lhs_mat_23_01_sp2));
+ __m256i iacc_mat_11_0_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_00_sp2, lhs_mat_23_00_sp2),_mm256_maddubs_epi16(rhs_mat_2367_01_sp2, lhs_mat_23_01_sp2));
+
+ __m256i iacc_mat_00_1_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_10_sp2, lhs_mat_01_10_sp2),_mm256_maddubs_epi16(rhs_mat_0145_11_sp2, lhs_mat_01_11_sp2));
+ __m256i iacc_mat_01_1_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_10_sp2, lhs_mat_01_10_sp2),_mm256_maddubs_epi16(rhs_mat_2367_11_sp2, lhs_mat_01_11_sp2));
+
+ __m256i iacc_mat_10_1_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_10_sp2, lhs_mat_23_10_sp2),_mm256_maddubs_epi16(rhs_mat_0145_11_sp2, lhs_mat_23_11_sp2));
+ __m256i iacc_mat_11_1_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_10_sp2, lhs_mat_23_10_sp2),_mm256_maddubs_epi16(rhs_mat_2367_11_sp2, lhs_mat_23_11_sp2));
+
+ __m256i iacc_mat_00_2_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_20_sp2, lhs_mat_01_20_sp2),_mm256_maddubs_epi16(rhs_mat_0145_21_sp2, lhs_mat_01_21_sp2));
+ __m256i iacc_mat_01_2_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_20_sp2, lhs_mat_01_20_sp2),_mm256_maddubs_epi16(rhs_mat_2367_21_sp2, lhs_mat_01_21_sp2));
+
+ __m256i iacc_mat_10_2_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_20_sp2, lhs_mat_23_20_sp2),_mm256_maddubs_epi16(rhs_mat_0145_21_sp2, lhs_mat_23_21_sp2));
+ __m256i iacc_mat_11_2_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_20_sp2, lhs_mat_23_20_sp2),_mm256_maddubs_epi16(rhs_mat_2367_21_sp2, lhs_mat_23_21_sp2));
+
+ __m256i iacc_mat_00_3_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_30_sp2, lhs_mat_01_30_sp2),_mm256_maddubs_epi16(rhs_mat_0145_31_sp2, lhs_mat_01_31_sp2));
+ __m256i iacc_mat_01_3_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_30_sp2, lhs_mat_01_30_sp2),_mm256_maddubs_epi16(rhs_mat_2367_31_sp2, lhs_mat_01_31_sp2));
+
+ __m256i iacc_mat_10_3_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_30_sp2, lhs_mat_23_30_sp2),_mm256_maddubs_epi16(rhs_mat_0145_31_sp2, lhs_mat_23_31_sp2));
+ __m256i iacc_mat_11_3_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_30_sp2, lhs_mat_23_30_sp2),_mm256_maddubs_epi16(rhs_mat_2367_31_sp2, lhs_mat_23_31_sp2));
+
+ __m256i iacc_mat_00_4_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_40_sp2, lhs_mat_01_40_sp2),_mm256_maddubs_epi16(rhs_mat_0145_41_sp2, lhs_mat_01_41_sp2));
+ __m256i iacc_mat_01_4_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_40_sp2, lhs_mat_01_40_sp2),_mm256_maddubs_epi16(rhs_mat_2367_41_sp2, lhs_mat_01_41_sp2));
+
+ __m256i iacc_mat_10_4_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_40_sp2, lhs_mat_23_40_sp2),_mm256_maddubs_epi16(rhs_mat_0145_41_sp2, lhs_mat_23_41_sp2));
+ __m256i iacc_mat_11_4_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_40_sp2, lhs_mat_23_40_sp2),_mm256_maddubs_epi16(rhs_mat_2367_41_sp2, lhs_mat_23_41_sp2));
+
+ __m256i iacc_mat_00_5_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_50_sp2, lhs_mat_01_50_sp2),_mm256_maddubs_epi16(rhs_mat_0145_51_sp2, lhs_mat_01_51_sp2));
+ __m256i iacc_mat_01_5_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_50_sp2, lhs_mat_01_50_sp2),_mm256_maddubs_epi16(rhs_mat_2367_51_sp2, lhs_mat_01_51_sp2));
+
+ __m256i iacc_mat_10_5_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_50_sp2, lhs_mat_23_50_sp2),_mm256_maddubs_epi16(rhs_mat_0145_51_sp2, lhs_mat_23_51_sp2));
+ __m256i iacc_mat_11_5_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_50_sp2, lhs_mat_23_50_sp2),_mm256_maddubs_epi16(rhs_mat_2367_51_sp2, lhs_mat_23_51_sp2));
+
+ __m256i iacc_mat_00_6_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_60_sp2, lhs_mat_01_60_sp2),_mm256_maddubs_epi16(rhs_mat_0145_61_sp2, lhs_mat_01_61_sp2));
+ __m256i iacc_mat_01_6_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_60_sp2, lhs_mat_01_60_sp2),_mm256_maddubs_epi16(rhs_mat_2367_61_sp2, lhs_mat_01_61_sp2));
+
+ __m256i iacc_mat_10_6_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_60_sp2, lhs_mat_23_60_sp2),_mm256_maddubs_epi16(rhs_mat_0145_61_sp2, lhs_mat_23_61_sp2));
+ __m256i iacc_mat_11_6_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_60_sp2, lhs_mat_23_60_sp2),_mm256_maddubs_epi16(rhs_mat_2367_61_sp2, lhs_mat_23_61_sp2));
+
+ __m256i iacc_mat_00_7_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_70_sp2, lhs_mat_01_70_sp2),_mm256_maddubs_epi16(rhs_mat_0145_71_sp2, lhs_mat_01_71_sp2));
+ __m256i iacc_mat_01_7_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_70_sp2, lhs_mat_01_70_sp2),_mm256_maddubs_epi16(rhs_mat_2367_71_sp2, lhs_mat_01_71_sp2));
+
+ __m256i iacc_mat_10_7_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_70_sp2, lhs_mat_23_70_sp2),_mm256_maddubs_epi16(rhs_mat_0145_71_sp2, lhs_mat_23_71_sp2));
+ __m256i iacc_mat_11_7_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_70_sp2, lhs_mat_23_70_sp2),_mm256_maddubs_epi16(rhs_mat_2367_71_sp2, lhs_mat_23_71_sp2));
+
+ // Combine results from both shuffle patterns for each output block
+ __m256i iacc_mat_00_0 = _mm256_add_epi16(iacc_mat_00_0_sp1, iacc_mat_00_0_sp2);
+ __m256i iacc_mat_01_0 = _mm256_add_epi16(iacc_mat_01_0_sp1, iacc_mat_01_0_sp2);
+ __m256i iacc_mat_10_0 = _mm256_add_epi16(iacc_mat_10_0_sp1, iacc_mat_10_0_sp2);
+ __m256i iacc_mat_11_0 = _mm256_add_epi16(iacc_mat_11_0_sp1, iacc_mat_11_0_sp2);
+
+ __m256i iacc_mat_00_1 = _mm256_add_epi16(iacc_mat_00_1_sp1, iacc_mat_00_1_sp2);
+ __m256i iacc_mat_01_1 = _mm256_add_epi16(iacc_mat_01_1_sp1, iacc_mat_01_1_sp2);
+ __m256i iacc_mat_10_1 = _mm256_add_epi16(iacc_mat_10_1_sp1, iacc_mat_10_1_sp2);
+ __m256i iacc_mat_11_1 = _mm256_add_epi16(iacc_mat_11_1_sp1, iacc_mat_11_1_sp2);
+
+ __m256i iacc_mat_00_2 = _mm256_add_epi16(iacc_mat_00_2_sp1, iacc_mat_00_2_sp2);
+ __m256i iacc_mat_01_2 = _mm256_add_epi16(iacc_mat_01_2_sp1, iacc_mat_01_2_sp2);
+ __m256i iacc_mat_10_2 = _mm256_add_epi16(iacc_mat_10_2_sp1, iacc_mat_10_2_sp2);
+ __m256i iacc_mat_11_2 = _mm256_add_epi16(iacc_mat_11_2_sp1, iacc_mat_11_2_sp2);
+
+ __m256i iacc_mat_00_3 = _mm256_add_epi16(iacc_mat_00_3_sp1, iacc_mat_00_3_sp2);
+ __m256i iacc_mat_01_3 = _mm256_add_epi16(iacc_mat_01_3_sp1, iacc_mat_01_3_sp2);
+ __m256i iacc_mat_10_3 = _mm256_add_epi16(iacc_mat_10_3_sp1, iacc_mat_10_3_sp2);
+ __m256i iacc_mat_11_3 = _mm256_add_epi16(iacc_mat_11_3_sp1, iacc_mat_11_3_sp2);
+
+ __m256i iacc_mat_00_4 = _mm256_add_epi16(iacc_mat_00_4_sp1, iacc_mat_00_4_sp2);
+ __m256i iacc_mat_01_4 = _mm256_add_epi16(iacc_mat_01_4_sp1, iacc_mat_01_4_sp2);
+ __m256i iacc_mat_10_4 = _mm256_add_epi16(iacc_mat_10_4_sp1, iacc_mat_10_4_sp2);
+ __m256i iacc_mat_11_4 = _mm256_add_epi16(iacc_mat_11_4_sp1, iacc_mat_11_4_sp2);
+
+ __m256i iacc_mat_00_5 = _mm256_add_epi16(iacc_mat_00_5_sp1, iacc_mat_00_5_sp2);
+ __m256i iacc_mat_01_5 = _mm256_add_epi16(iacc_mat_01_5_sp1, iacc_mat_01_5_sp2);
+ __m256i iacc_mat_10_5 = _mm256_add_epi16(iacc_mat_10_5_sp1, iacc_mat_10_5_sp2);
+ __m256i iacc_mat_11_5 = _mm256_add_epi16(iacc_mat_11_5_sp1, iacc_mat_11_5_sp2);
+
+ __m256i iacc_mat_00_6 = _mm256_add_epi16(iacc_mat_00_6_sp1, iacc_mat_00_6_sp2);
+ __m256i iacc_mat_01_6 = _mm256_add_epi16(iacc_mat_01_6_sp1, iacc_mat_01_6_sp2);
+ __m256i iacc_mat_10_6 = _mm256_add_epi16(iacc_mat_10_6_sp1, iacc_mat_10_6_sp2);
+ __m256i iacc_mat_11_6 = _mm256_add_epi16(iacc_mat_11_6_sp1, iacc_mat_11_6_sp2);
+
+ __m256i iacc_mat_00_7 = _mm256_add_epi16(iacc_mat_00_7_sp1, iacc_mat_00_7_sp2);
+ __m256i iacc_mat_01_7 = _mm256_add_epi16(iacc_mat_01_7_sp1, iacc_mat_01_7_sp2);
+ __m256i iacc_mat_10_7 = _mm256_add_epi16(iacc_mat_10_7_sp1, iacc_mat_10_7_sp2);
+ __m256i iacc_mat_11_7 = _mm256_add_epi16(iacc_mat_11_7_sp1, iacc_mat_11_7_sp2);
+
+ // Output of both shuffle patterns are added in order to sum dot product outputs of all 32 values in block
+ iacc_mat_00_0 = _mm256_madd_epi16(iacc_mat_00_0, scale_0145_0);
+ iacc_mat_01_0 = _mm256_madd_epi16(iacc_mat_01_0, scale_2367_0);
+ iacc_mat_10_0 = _mm256_madd_epi16(iacc_mat_10_0, scale_0145_0);
+ iacc_mat_11_0 = _mm256_madd_epi16(iacc_mat_11_0, scale_2367_0);
+
+ iacc_mat_00_1 = _mm256_madd_epi16(iacc_mat_00_1, scale_0145_1);
+ iacc_mat_01_1 = _mm256_madd_epi16(iacc_mat_01_1, scale_2367_1);
+ iacc_mat_10_1 = _mm256_madd_epi16(iacc_mat_10_1, scale_0145_1);
+ iacc_mat_11_1 = _mm256_madd_epi16(iacc_mat_11_1, scale_2367_1);
+
+ iacc_mat_00_2 = _mm256_madd_epi16(iacc_mat_00_2, scale_0145_2);
+ iacc_mat_01_2 = _mm256_madd_epi16(iacc_mat_01_2, scale_2367_2);
+ iacc_mat_10_2 = _mm256_madd_epi16(iacc_mat_10_2, scale_0145_2);
+ iacc_mat_11_2 = _mm256_madd_epi16(iacc_mat_11_2, scale_2367_2);
+
+ iacc_mat_00_3 = _mm256_madd_epi16(iacc_mat_00_3, scale_0145_3);
+ iacc_mat_01_3 = _mm256_madd_epi16(iacc_mat_01_3, scale_2367_3);
+ iacc_mat_10_3 = _mm256_madd_epi16(iacc_mat_10_3, scale_0145_3);
+ iacc_mat_11_3 = _mm256_madd_epi16(iacc_mat_11_3, scale_2367_3);
+
+ iacc_mat_00_4 = _mm256_madd_epi16(iacc_mat_00_4, scale_0145_4);
+ iacc_mat_01_4 = _mm256_madd_epi16(iacc_mat_01_4, scale_2367_4);
+ iacc_mat_10_4 = _mm256_madd_epi16(iacc_mat_10_4, scale_0145_4);
+ iacc_mat_11_4 = _mm256_madd_epi16(iacc_mat_11_4, scale_2367_4);
+
+ iacc_mat_00_5 = _mm256_madd_epi16(iacc_mat_00_5, scale_0145_5);
+ iacc_mat_01_5 = _mm256_madd_epi16(iacc_mat_01_5, scale_2367_5);
+ iacc_mat_10_5 = _mm256_madd_epi16(iacc_mat_10_5, scale_0145_5);
+ iacc_mat_11_5 = _mm256_madd_epi16(iacc_mat_11_5, scale_2367_5);
+
+ iacc_mat_00_6 = _mm256_madd_epi16(iacc_mat_00_6, scale_0145_6);
+ iacc_mat_01_6 = _mm256_madd_epi16(iacc_mat_01_6, scale_2367_6);
+ iacc_mat_10_6 = _mm256_madd_epi16(iacc_mat_10_6, scale_0145_6);
+ iacc_mat_11_6 = _mm256_madd_epi16(iacc_mat_11_6, scale_2367_6);
+
+ iacc_mat_00_7 = _mm256_madd_epi16(iacc_mat_00_7, scale_0145_7);
+ iacc_mat_01_7 = _mm256_madd_epi16(iacc_mat_01_7, scale_2367_7);
+ iacc_mat_10_7 = _mm256_madd_epi16(iacc_mat_10_7, scale_0145_7);
+ iacc_mat_11_7 = _mm256_madd_epi16(iacc_mat_11_7, scale_2367_7);
+
+ __m256i iacc_mat_00 = _mm256_add_epi32(_mm256_add_epi32(_mm256_add_epi32(iacc_mat_00_0, iacc_mat_00_1), _mm256_add_epi32(iacc_mat_00_2, iacc_mat_00_3)), _mm256_add_epi32(_mm256_add_epi32(iacc_mat_00_4, iacc_mat_00_5), _mm256_add_epi32(iacc_mat_00_6, iacc_mat_00_7)));
+ __m256i iacc_mat_01 = _mm256_add_epi32(_mm256_add_epi32(_mm256_add_epi32(iacc_mat_01_0, iacc_mat_01_1), _mm256_add_epi32(iacc_mat_01_2, iacc_mat_01_3)), _mm256_add_epi32(_mm256_add_epi32(iacc_mat_01_4, iacc_mat_01_5), _mm256_add_epi32(iacc_mat_01_6, iacc_mat_01_7)));
+ __m256i iacc_mat_10 = _mm256_add_epi32(_mm256_add_epi32(_mm256_add_epi32(iacc_mat_10_0, iacc_mat_10_1), _mm256_add_epi32(iacc_mat_10_2, iacc_mat_10_3)), _mm256_add_epi32(_mm256_add_epi32(iacc_mat_10_4, iacc_mat_10_5), _mm256_add_epi32(iacc_mat_10_6, iacc_mat_10_7)));
+ __m256i iacc_mat_11 = _mm256_add_epi32(_mm256_add_epi32(_mm256_add_epi32(iacc_mat_11_0, iacc_mat_11_1), _mm256_add_epi32(iacc_mat_11_2, iacc_mat_11_3)), _mm256_add_epi32(_mm256_add_epi32(iacc_mat_11_4, iacc_mat_11_5), _mm256_add_epi32(iacc_mat_11_6, iacc_mat_11_7)));
+
+ // Straighten out to make 4 row vectors
+ __m256i iacc_row_0 = _mm256_blend_epi32(iacc_mat_00, _mm256_shuffle_epi32(iacc_mat_01, 78), 204);
+ __m256i iacc_row_1 = _mm256_blend_epi32(_mm256_shuffle_epi32(iacc_mat_00, 78), iacc_mat_01, 204);
+ __m256i iacc_row_2 = _mm256_blend_epi32(iacc_mat_10, _mm256_shuffle_epi32(iacc_mat_11, 78), 204);
+ __m256i iacc_row_3 = _mm256_blend_epi32(_mm256_shuffle_epi32(iacc_mat_10, 78), iacc_mat_11, 204);
+
+ // Load the scale(d) values for all the 4 Q8_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 = _mm256_set_m128(row_scale_f32_sse, row_scale_f32_sse);
+
+ // Multiply with appropiate scales and accumulate (for both d and dmin) below
+ acc_rows[rp * 4] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_0), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 0)), acc_rows[rp * 4]);
+ acc_rows[rp * 4 + 1] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_1), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 85)), acc_rows[rp * 4 + 1]);
+ acc_rows[rp * 4 + 2] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_2), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_rows[rp * 4 + 2]);
+ acc_rows[rp * 4 + 3] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_3), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 255)), acc_rows[rp * 4 + 3]);
+
+ __m256i lhs_bsums_01_0123 = _mm256_inserti128_si256(_mm256_castsi128_si256(lhs_raw_bsums_01_0123), lhs_raw_bsums_01_0123, 1);
+ __m256i lhs_bsums_23_0123 = _mm256_inserti128_si256(_mm256_castsi128_si256(lhs_raw_bsums_23_0123), lhs_raw_bsums_23_0123, 1);
+ __m256i lhs_bsums_01_4567 = _mm256_inserti128_si256(_mm256_castsi128_si256(lhs_raw_bsums_01_4567), lhs_raw_bsums_01_4567, 1);
+ __m256i lhs_bsums_23_4567 = _mm256_inserti128_si256(_mm256_castsi128_si256(lhs_raw_bsums_23_4567), lhs_raw_bsums_23_4567, 1);
+
+ // Take two bsums from two Q8_Ks at a time and multiply with corresponding mins values from each Q2_K
+ __m256i iacc_row_min_0_01 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_01_0123, 0), mins_01);
+ __m256i iacc_row_min_1_01 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_01_0123, 170), mins_01);
+ __m256i iacc_row_min_2_01 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_23_0123, 0), mins_01);
+ __m256i iacc_row_min_3_01 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_23_0123, 170), mins_01);
+
+ __m256i iacc_row_min_0_23 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_01_0123, 85), mins_23);
+ __m256i iacc_row_min_1_23 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_01_0123, 255), mins_23);
+ __m256i iacc_row_min_2_23 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_23_0123, 85), mins_23);
+ __m256i iacc_row_min_3_23 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_23_0123, 255), mins_23);
+
+ __m256i iacc_row_min_0_45 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_01_4567, 0), mins_45);
+ __m256i iacc_row_min_1_45 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_01_4567, 170), mins_45);
+ __m256i iacc_row_min_2_45 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_23_4567, 0), mins_45);
+ __m256i iacc_row_min_3_45 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_23_4567, 170), mins_45);
+
+ __m256i iacc_row_min_0_67 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_01_4567, 85), mins_67);
+ __m256i iacc_row_min_1_67 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_01_4567, 255), mins_67);
+ __m256i iacc_row_min_2_67 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_23_4567, 85), mins_67);
+ __m256i iacc_row_min_3_67 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_23_4567, 255), mins_67);
+
+ __m256i iacc_row_min_0 = _mm256_add_epi32(_mm256_add_epi32(iacc_row_min_0_01, iacc_row_min_0_23), _mm256_add_epi32(iacc_row_min_0_45,iacc_row_min_0_67));
+ __m256i iacc_row_min_1 = _mm256_add_epi32(_mm256_add_epi32(iacc_row_min_1_01, iacc_row_min_1_23), _mm256_add_epi32(iacc_row_min_1_45,iacc_row_min_1_67));
+ __m256i iacc_row_min_2 = _mm256_add_epi32(_mm256_add_epi32(iacc_row_min_2_01, iacc_row_min_2_23), _mm256_add_epi32(iacc_row_min_2_45,iacc_row_min_2_67));
+ __m256i iacc_row_min_3 = _mm256_add_epi32(_mm256_add_epi32(iacc_row_min_3_01, iacc_row_min_3_23), _mm256_add_epi32(iacc_row_min_3_45,iacc_row_min_3_67));
+
+ acc_min_rows[rp * 4] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_min_0), _mm256_mul_ps(col_dmin_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 0)), acc_min_rows[rp * 4]);
+ acc_min_rows[rp * 4 + 1] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_min_1), _mm256_mul_ps(col_dmin_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 85)), acc_min_rows[rp * 4 + 1]);
+ acc_min_rows[rp * 4 + 2] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_min_2), _mm256_mul_ps(col_dmin_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_min_rows[rp * 4 + 2]);
+ acc_min_rows[rp * 4 + 3] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_min_3), _mm256_mul_ps(col_dmin_f32, _mm256_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++) {
+ _mm256_storeu_ps((float * )(s + ((y * 4 + i) * bs + x * 8)), _mm256_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_q2_kx8 structures at each pass of the loop and perform dot product operation
+ for (int64_t x = xstart; x < nc / 8; x++) {
+
+ const block_q2_Kx8 * b_ptr = b_ptr_start + (x * b_nb);
+
+ // Master FP accumulators
+ __m256 acc_rows[4];
+ for (int i = 0; i < 4; i++) {
+ acc_rows[i] = _mm256_setzero_ps();
+ }
+
+ __m256 acc_min_rows[4];
+ for (int i = 0; i < 4; i++) {
+ acc_min_rows[i] = _mm256_setzero_ps();
+ }
+
+ for (int64_t b = 0; b < nb; b++) {
+ // Delta values - Load the eight scale values of block_q2_kx8
+ const __m256 col_scale_f32 = GGML_F32Cx8_LOAD(b_ptr[b].d);
+
+ // dmin values - Load the eight dmin values of block_q2_kx8
+ const __m256 col_dmin_f32 = GGML_F32Cx8_LOAD(b_ptr[b].dmin);
+
+ // Loop to iterate over the sixteen sub blocks of a super block - eight sub blocks are processed per iteration
+ for (int sb = 0; sb < QK_K / 128; sb++) {
+
+ // Load the eight block_q4_k for eight sub blocks quantized values interleaved with each other in chunks of eight bytes - B0,B1 ....B6,B7
+ const __m256i rhs_raw_mat_0123_0 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + sb * 256));
+ const __m256i rhs_raw_mat_4567_0 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 32 + sb * 256));
+ const __m256i rhs_raw_mat_0123_1 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 64 + sb * 256));
+ const __m256i rhs_raw_mat_4567_1 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 96 + sb * 256));
+ const __m256i rhs_raw_mat_0123_2 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 128 + sb * 256));
+ const __m256i rhs_raw_mat_4567_2 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 160 + sb * 256));
+ const __m256i rhs_raw_mat_0123_3 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 192 + sb * 256));
+ const __m256i rhs_raw_mat_4567_3 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 224 + sb * 256));
+
+ // Save the values in the following vectors in the formats B0B1B4B5, B2B3B6B7 for further processing and storing of values
+ //superblock sub block which part of sub block
+ 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);
+
+ // 2-bit -> 8-bit
+ // First sub block of the eight sub blocks processed in the iteration
+ const __m256i rhs_mat_0145_00 = _mm256_and_si256(rhs_raw_mat_0145_0, m3b); //B00(0-7) B01(0-7) B04(0-7) B05(0-7)
+ const __m256i rhs_mat_2367_00 = _mm256_and_si256(rhs_raw_mat_2367_0, m3b); //B02(0-7) B03(0-7) B06(0-7) B07(0-7)
+
+ const __m256i rhs_mat_0145_01 = _mm256_and_si256(rhs_raw_mat_0145_1, m3b); //B00(8-15) B01(8-15) B04(8-15) B05(8-15)
+ const __m256i rhs_mat_2367_01 = _mm256_and_si256(rhs_raw_mat_2367_1, m3b); //B02(8-15) B03(8-15) B06(8-15) B07(8-15)
+
+ // Second sub block of the eight sub blocks processed in the iteration
+ const __m256i rhs_mat_0145_10 = _mm256_and_si256(rhs_raw_mat_0145_2, m3b); //B10(0-7) B11(0-7) B14(0-7) B15(0-7)
+ const __m256i rhs_mat_2367_10 = _mm256_and_si256(rhs_raw_mat_2367_2, m3b); //B12(0-7) B13(0-7) B16(0-7) B17(0-7)
+
+ const __m256i rhs_mat_0145_11 = _mm256_and_si256(rhs_raw_mat_0145_3, m3b); //B10(8-15) B11(8-15) B14(8-15) B15(8-15)
+ const __m256i rhs_mat_2367_11 = _mm256_and_si256(rhs_raw_mat_2367_3, m3b); //B12(8-15) B13(8-15) B16(8-15) B17(8-15)
+
+ // Third sub block of the eight sub blocks processed in the iteration
+ const __m256i rhs_mat_0145_20 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_0, 2), m3b); //B20(0-7) B21(0-7) B24(0-7) B25(0-7)
+ const __m256i rhs_mat_2367_20 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_0, 2), m3b); //B22(0-7) B23(0-7) B26(0-7) B27(0-7)
+
+ const __m256i rhs_mat_0145_21 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_1, 2), m3b); //B20(8-15) B21(8-15) B24(8-15) B25(8-15)
+ const __m256i rhs_mat_2367_21 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_1, 2), m3b); //B22(8-15) B23(8-15) B26(8-15) B27(8-15)
+
+ // Fourth sub block of the eight sub blocks processed in the iteration
+ const __m256i rhs_mat_0145_30 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_2, 2), m3b); //B30(0-7) B31(0-7) B34(0-7) B35(0-7)
+ const __m256i rhs_mat_2367_30 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_2, 2), m3b); //B32(0-7) B33(0-7) B36(0-7) B37(0-7)
+
+ const __m256i rhs_mat_0145_31 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_3, 2), m3b); //B30(8-15) B31(8-15) B34(8-15) B35(8-15)
+ const __m256i rhs_mat_2367_31 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_3, 2), m3b); //B32(8-15) B33(8-15) B36(8-15) B37(8-15)
+
+ // Fifth sub block of the eight sub blocks processed in the iteration
+ const __m256i rhs_mat_0145_40 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_0, 4), m3b); //B40(0-7) B41(0-7) B44(0-7) B45(0-7)
+ const __m256i rhs_mat_2367_40 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_0, 4), m3b); //B42(0-7) B43(0-7) B46(0-7) B47(0-7)
+
+ const __m256i rhs_mat_0145_41 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_1, 4), m3b); //B40(8-15) B41(8-15) B44(8-15) B45(8-15)
+ const __m256i rhs_mat_2367_41 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_1, 4), m3b); //B42(8-15) B43(8-15) B46(8-15) B47(8-15)
+
+ // Sixth sub block of the eight sub blocks processed in the iteration
+ const __m256i rhs_mat_0145_50 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_2, 4), m3b); //B50(0-7) B51(0-7) B54(0-7) B55(0-7)
+ const __m256i rhs_mat_2367_50 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_2, 4), m3b); //B52(0-7) B53(0-7) B56(0-7) B57(0-7)
+
+ const __m256i rhs_mat_0145_51 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_3, 4), m3b); //B50(8-15) B51(8-15) B54(8-15) B55(8-15)
+ const __m256i rhs_mat_2367_51 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_3, 4), m3b); //B52(8-15) B53(8-15) B56(8-15) B57(8-15)
+
+ // Seventh sub block of the eight sub blocks processed in the iteration
+ const __m256i rhs_mat_0145_60 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_0, 6), m3b); //B60(0-7) B61(0-7) B64(0-7) B65(0-7)
+ const __m256i rhs_mat_2367_60 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_0, 6), m3b); //B62(0-7) B63(0-7) B66(0-7) B67(0-7)
+
+ const __m256i rhs_mat_0145_61 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_1, 6), m3b); //B60(8-15) B61(8-15) B64(8-15) B65(8-15)
+ const __m256i rhs_mat_2367_61 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_1, 6), m3b); //B62(8-15) B63(8-15) B66(8-15) B67(8-15)
+
+ // Eighth sub block of the eight sub blocks processed in the iteration
+ const __m256i rhs_mat_0145_70 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_2, 6), m3b); //B70(0-7) B71(0-7) B74(0-7) B75(0-7)
+ const __m256i rhs_mat_2367_70 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_2, 6), m3b); //B72(0-7) B73(0-7) B76(0-7) B77(0-7)
+
+ const __m256i rhs_mat_0145_71 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_3, 6), m3b); //B70(8-15) B71(8-15) B74(8-15) B75(8-15)
+ const __m256i rhs_mat_2367_71 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_3, 6), m3b); //B72(8-15) B73(8-15) B76(8-15) B77(8-15)
+
+ // Shuffle pattern one - right side input
+ const __m256i rhs_mat_0145_00_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_00, 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)
+ const __m256i rhs_mat_2367_00_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_00, 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)
+
+ const __m256i rhs_mat_0145_01_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_01, 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)
+ const __m256i rhs_mat_2367_01_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_01, 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)
+
+ const __m256i rhs_mat_0145_10_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_10, 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)
+ const __m256i rhs_mat_2367_10_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_10, 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)
+
+ const __m256i rhs_mat_0145_11_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_11, 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)
+ const __m256i rhs_mat_2367_11_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_11, 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)
+
+ const __m256i rhs_mat_0145_20_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_20, 136); //B20(0-3) B21(0-3) B20(0-3) B21(0-3) B24(0-3) B25(0-3) B24(0-3) B25(0-3)
+ const __m256i rhs_mat_2367_20_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_20, 136); //B22(0-3) B23(0-3) B22(0-3) B23(0-3) B26(0-3) B27(0-3) B26(0-3) B27(0-3)
+
+ const __m256i rhs_mat_0145_21_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_21, 136); //B20(8-11) B21(8-11) B20(8-11) B21(8-11) B24(8-11) B25(8-11) B24(8-11) B25(8-11)
+ const __m256i rhs_mat_2367_21_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_21, 136); //B22(8-11) B23(8-11) B22(8-11) B23(8-11) B26(8-11) B27(8-11) B26(8-11) B27(8-11)
+
+ const __m256i rhs_mat_0145_30_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_30, 136); //B30(0-3) B31(0-3) B30(0-3) B31(0-3) B34(0-3) B35(0-3) B34(0-3) B35(0-3)
+ const __m256i rhs_mat_2367_30_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_30, 136); //B32(0-3) B33(0-3) B32(0-3) B33(0-3) B36(0-3) B37(0-3) B36(0-3) B37(0-3)
+
+ const __m256i rhs_mat_0145_31_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_31, 136); //B30(8-11) B31(8-11) B30(8-11) B31(8-11) B34(8-11) B35(8-11) B34(8-11) B35(8-11
+ const __m256i rhs_mat_2367_31_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_31, 136); //B32(8-11) B33(8-11) B32(8-11) B33(8-11) B36(8-11) B37(8-11) B36(8-11) B37(8-11)
+
+ const __m256i rhs_mat_0145_40_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_40, 136); //B40(0-3) B41(0-3) B40(0-3) B41(0-3) B44(0-3) B45(0-3) B44(0-3) B45(0-3)
+ const __m256i rhs_mat_2367_40_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_40, 136); //B42(0-3) B43(0-3) B42(0-3) B43(0-3) B46(0-3) B47(0-3) B46(0-3) B47(0-3)
+
+ const __m256i rhs_mat_0145_41_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_41, 136); //B40(8-11) B41(8-11) B40(8-11) B41(8-11) B44(8-11) B45(8-11) B44(8-11) B45(8-11)
+ const __m256i rhs_mat_2367_41_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_41, 136); //B42(8-11) B43(8-11) B42(8-11) B43(8-11) B46(8-11) B47(8-11) B46(8-11) B47(8-11)
+
+ const __m256i rhs_mat_0145_50_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_50, 136); //B50(0-3) B51(0-3) B50(0-3) B51(0-3) B54(0-3) B55(0-3) B54(0-3) B55(0-3)
+ const __m256i rhs_mat_2367_50_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_50, 136); //B52(0-3) B53(0-3) B52(0-3) B53(0-3) B56(0-3) B57(0-3) B56(0-3) B57(0-3)
+
+ const __m256i rhs_mat_0145_51_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_51, 136); //B50(8-11) B51(8-11) B50(8-11) B51(8-11) B54(8-11) B55(8-11) B54(8-11) B55(8-11)
+ const __m256i rhs_mat_2367_51_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_51, 136); //B52(8-11) B53(8-11) B52(8-11) B53(8-11) B56(8-11) B57(8-11) B56(8-11) B57(8-11)
+
+ const __m256i rhs_mat_0145_60_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_60, 136); //B60(0-3) B61(0-3) B60(0-3) B61(0-3) B64(0-3) B65(0-3) B64(0-3) B65(0-3)
+ const __m256i rhs_mat_2367_60_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_60, 136); //B62(0-3) B63(0-3) B62(0-3) B63(0-3) B66(0-3) B67(0-3) B66(0-3) B67(0-3)
+
+ const __m256i rhs_mat_0145_61_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_61, 136); //B60(8-11) B61(8-11) B60(8-11) B61(8-11) B64(8-11) B65(8-11) B64(8-11) B65(8-11)
+ const __m256i rhs_mat_2367_61_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_61, 136); //B62(8-11) B63(8-11) B62(8-11) B63(8-11) B66(8-11) B67(8-11) B66(8-11) B67(8-11)
+
+ const __m256i rhs_mat_0145_70_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_70, 136); //B70(0-3) B71(0-3) B70(0-3) B71(0-3) B74(0-3) B75(0-3) B74(0-3) B75(0-3)
+ const __m256i rhs_mat_2367_70_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_70, 136); //B72(0-3) B73(0-3) B72(0-3) B73(0-3) B76(0-3) B77(0-3) B76(0-3) B77(0-3)
+
+ const __m256i rhs_mat_0145_71_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_71, 136); //B70(8-11) B71(8-11) B70(8-11) B71(8-11) B74(8-11) B75(8-11) B74(8-11) B75(8-11)
+ const __m256i rhs_mat_2367_71_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_71, 136); //B72(8-11) B73(8-11) B72(8-11) B73(8-11) B76(8-11) B77(8-11) B76(8-11) B77(8-11)
+
+
+ // Shuffle pattern two - right side input
+ const __m256i rhs_mat_0145_00_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_00, 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)
+ const __m256i rhs_mat_2367_00_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_00, 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)
+
+ const __m256i rhs_mat_0145_01_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_01, 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)
+ const __m256i rhs_mat_2367_01_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_01, 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)
+
+ const __m256i rhs_mat_0145_10_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_10, 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)
+ const __m256i rhs_mat_2367_10_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_10, 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)
+
+ const __m256i rhs_mat_0145_11_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_11, 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)
+ const __m256i rhs_mat_2367_11_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_11, 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)
+
+ const __m256i rhs_mat_0145_20_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_20, 221); //B20(4-7) B21(4-7) B20(4-7) B21(4-7) B24(4-7) B25(4-7) B24(4-7) B25(4-7)
+ const __m256i rhs_mat_2367_20_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_20, 221); //B22(4-7) B23(4-7) B22(4-7) B23(4-7) B26(4-7) B27(4-7) B26(4-7) B27(4-7)
+
+ const __m256i rhs_mat_0145_21_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_21, 221); //B20(12-15) B21(12-15) B20(12-15) B21(12-15) B24(12-15) B25(12-15) B24(12-15) B25(12-15)
+ const __m256i rhs_mat_2367_21_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_21, 221); //B22(12-15) B23(12-15) B22(12-15) B23(12-15) B26(12-15) B27(12-15) B26(12-15) B27(12-15)
+
+ const __m256i rhs_mat_0145_30_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_30, 221); //B30(4-7) B31(4-7) B30(4-7) B31(4-7) B34(4-7) B35(4-7) B34(4-7) B35(4-7)
+ const __m256i rhs_mat_2367_30_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_30, 221); //B32(4-7) B33(4-7) B32(4-7) B33(4-7) B36(4-7) B37(4-7) B36(4-7) B37(4-7)
+
+ const __m256i rhs_mat_0145_31_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_31, 221); //B30(12-15) B31(12-15) B30(12-15) B31(12-15) B34(12-15) B35(12-15) B34(12-15) B35(12-15)
+ const __m256i rhs_mat_2367_31_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_31, 221); //B32(12-15) B33(12-15) B32(12-15) B33(12-15) B36(12-15) B37(12-15) B36(12-15) B37(12-15)
+
+ const __m256i rhs_mat_0145_40_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_40, 221); //B40(4-7) B41(4-7) B40(4-7) B41(4-7) B44(4-7) B45(4-7) B44(4-7) B45(4-7)
+ const __m256i rhs_mat_2367_40_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_40, 221); //B42(4-7) B43(4-7) B42(4-7) B43(4-7) B46(4-7) B47(4-7) B46(4-7) B47(4-7)
+
+ const __m256i rhs_mat_0145_41_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_41, 221); //B40(12-15) B41(12-15) B40(12-15) B41(12-15) B44(12-15) B45(12-15) B44(12-15) B45(12-15)
+ const __m256i rhs_mat_2367_41_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_41, 221); //B42(12-15) B43(12-15) B42(12-15) B43(12-15) B46(12-15) B47(12-15) B46(12-15) B47(12-15)
+
+ const __m256i rhs_mat_0145_50_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_50, 221); //B50(4-7) B51(4-7) B50(4-7) B51(4-7) B54(4-7) B55(4-7) B54(4-7) B55(4-7)
+ const __m256i rhs_mat_2367_50_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_50, 221); //B52(4-7) B53(4-7) B52(4-7) B53(4-7) B56(4-7) B57(4-7) B56(4-7) B57(4-7)
+
+ const __m256i rhs_mat_0145_51_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_51, 221); //B50(12-15) B51(12-15) B50(12-15) B51(12-15) B54(12-15) B55(12-15) B54(12-15) B55(12-15)
+ const __m256i rhs_mat_2367_51_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_51, 221); //B52(12-15) B53(12-15) B52(12-15) B53(12-15) B56(12-15) B57(12-15) B56(12-15) B57(12-15)
+
+ const __m256i rhs_mat_0145_60_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_60, 221); //B60(4-7) B61(4-7) B60(4-7) B61(4-7) B64(4-7) B65(4-7) B64(4-7) B65(4-7)
+ const __m256i rhs_mat_2367_60_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_60, 221); //B62(4-7) B63(4-7) B62(4-7) B63(4-7) B66(4-7) B67(4-7) B66(4-7) B67(4-7)
+
+ const __m256i rhs_mat_0145_61_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_61, 221); //B60(12-15) B61(12-15) B60(12-15) B61(12-15) B64(12-15) B65(12-15) B64(12-15) B65(12-15)
+ const __m256i rhs_mat_2367_61_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_61, 221); //B62(12-15) B63(12-15) B62(12-15) B63(12-15) B66(12-15) B67(12-15) B66(12-15) B67(12-15)
+
+ const __m256i rhs_mat_0145_70_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_70, 221); //B70(4-7) B71(4-7) B70(4-7) B71(4-7) B74(4-7) B75(4-7) B74(4-7) B75(4-7)
+ const __m256i rhs_mat_2367_70_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_70, 221); //B72(4-7) B73(4-7) B72(4-7) B73(4-7) B76(4-7) B77(4-7) B76(4-7) B77(4-7)
+
+ const __m256i rhs_mat_0145_71_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_71, 221); //B70(12-15) B71(12-15) B70(12-15) B71(12-15) B74(12-15) B75(12-15) B74(12-15) B75(12-15)
+ const __m256i rhs_mat_2367_71_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_71, 221); //B72(12-15) B73(12-15) B72(12-15) B73(12-15) B76(12-15) B77(12-15) B76(12-15) B77(12-15)
+
+
+ //Scales and Mins of corresponding sub blocks from different Q2_K structures are stored together
+ //s00 m00 s01 m01 s10 m10 s11 m11 s20 m20 s21 m21 s30 m30 s31 m31 s40 m40 s41 m41 s50 m50 s51 m51 s60 m60 s61 m61 s70 m70 s71 m71
+
+ // Combine mins and scales for sub-blocks: 0-1, 2-3, 4-5, 6-7 in the sb loop
+ const __m128i mins_and_scales_01 = _mm_loadu_si128((const __m128i *)(b_ptr[b].scales + sb * 64));
+ const __m128i mins_and_scales_23 = _mm_loadu_si128((const __m128i *)(b_ptr[b].scales + 16 + sb * 64));
+ const __m128i mins_and_scales_45 = _mm_loadu_si128((const __m128i *)(b_ptr[b].scales + 32 + sb * 64));
+ const __m128i mins_and_scales_67 = _mm_loadu_si128((const __m128i *)(b_ptr[b].scales + 48 + sb * 64));
+
+ // Extract scales which is lower half from mins_and_scales
+ const __m128i scales_01 = _mm_and_si128(mins_and_scales_01, m4b_sse);
+ const __m128i scales_23 = _mm_and_si128(mins_and_scales_23, m4b_sse);
+ const __m128i scales_45 = _mm_and_si128(mins_and_scales_45, m4b_sse);
+ const __m128i scales_67 = _mm_and_si128(mins_and_scales_67, m4b_sse);
+
+ // Extract mins which is upper half from mins_and_scales
+ const __m256i mins_01 = _mm256_cvtepu8_epi16(_mm_and_si128(_mm_srli_epi16(mins_and_scales_01, 4), m4b_sse));
+ const __m256i mins_23 = _mm256_cvtepu8_epi16(_mm_and_si128(_mm_srli_epi16(mins_and_scales_23, 4), m4b_sse));
+ const __m256i mins_45 = _mm256_cvtepu8_epi16(_mm_and_si128(_mm_srli_epi16(mins_and_scales_45, 4), m4b_sse));
+ const __m256i mins_67 = _mm256_cvtepu8_epi16(_mm_and_si128(_mm_srli_epi16(mins_and_scales_67, 4), m4b_sse));
+
+ const __m256i scales_0 = _mm256_cvtepu8_epi16(_mm_shuffle_epi8(scales_01, scalesmask1_sse));
+ const __m256i scales_1 = _mm256_cvtepu8_epi16(_mm_shuffle_epi8(scales_01, scalesmask2_sse));
+
+ const __m256i scales_2 = _mm256_cvtepu8_epi16(_mm_shuffle_epi8(scales_23, scalesmask1_sse));
+ const __m256i scales_3 = _mm256_cvtepu8_epi16(_mm_shuffle_epi8(scales_23, scalesmask2_sse));
+
+ const __m256i scales_4 = _mm256_cvtepu8_epi16(_mm_shuffle_epi8(scales_45, scalesmask1_sse));
+ const __m256i scales_5 = _mm256_cvtepu8_epi16(_mm_shuffle_epi8(scales_45, scalesmask2_sse));
+
+ const __m256i scales_6 = _mm256_cvtepu8_epi16(_mm_shuffle_epi8(scales_67, scalesmask1_sse));
+ const __m256i scales_7 = _mm256_cvtepu8_epi16(_mm_shuffle_epi8(scales_67, scalesmask2_sse));
+
+ const __m256i scale_0145_0 = _mm256_shuffle_epi32(scales_0, 68);
+ const __m256i scale_2367_0 = _mm256_shuffle_epi32(scales_0, 238);
+
+ const __m256i scale_0145_1 = _mm256_shuffle_epi32(scales_1, 68);
+ const __m256i scale_2367_1 = _mm256_shuffle_epi32(scales_1, 238);
+
+ const __m256i scale_0145_2 = _mm256_shuffle_epi32(scales_2, 68);
+ const __m256i scale_2367_2 = _mm256_shuffle_epi32(scales_2, 238);
+
+ const __m256i scale_0145_3 = _mm256_shuffle_epi32(scales_3, 68);
+ const __m256i scale_2367_3 = _mm256_shuffle_epi32(scales_3, 238);
+
+ const __m256i scale_0145_4 = _mm256_shuffle_epi32(scales_4, 68);
+ const __m256i scale_2367_4 = _mm256_shuffle_epi32(scales_4, 238);
+
+ const __m256i scale_0145_5 = _mm256_shuffle_epi32(scales_5, 68);
+ const __m256i scale_2367_5 = _mm256_shuffle_epi32(scales_5, 238);
+
+ const __m256i scale_0145_6 = _mm256_shuffle_epi32(scales_6, 68);
+ const __m256i scale_2367_6 = _mm256_shuffle_epi32(scales_6, 238);
+
+ const __m256i scale_0145_7 = _mm256_shuffle_epi32(scales_7, 68);
+ const __m256i scale_2367_7 = _mm256_shuffle_epi32(scales_7, 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_0123_00 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 512 * sb)));
+ __m256i lhs_mat_01_00 = _mm256_permute2f128_si256(lhs_mat_0123_00, lhs_mat_0123_00, 0);
+ __m256i lhs_mat_23_00 = _mm256_permute2f128_si256(lhs_mat_0123_00, lhs_mat_0123_00, 17);
+ __m256i lhs_mat_0123_01 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 32 + 512 * sb)));
+ __m256i lhs_mat_01_01 = _mm256_permute2f128_si256(lhs_mat_0123_01, lhs_mat_0123_01, 0);
+ __m256i lhs_mat_23_01 = _mm256_permute2f128_si256(lhs_mat_0123_01, lhs_mat_0123_01, 17);
+ __m256i lhs_mat_0123_10 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 64 + 512 * sb)));
+ __m256i lhs_mat_01_10 = _mm256_permute2f128_si256(lhs_mat_0123_10, lhs_mat_0123_10, 0);
+ __m256i lhs_mat_23_10 = _mm256_permute2f128_si256(lhs_mat_0123_10, lhs_mat_0123_10, 17);
+ __m256i lhs_mat_0123_11 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 96 + 512 * sb)));
+ __m256i lhs_mat_01_11 = _mm256_permute2f128_si256(lhs_mat_0123_11, lhs_mat_0123_11, 0);
+ __m256i lhs_mat_23_11 = _mm256_permute2f128_si256(lhs_mat_0123_11, lhs_mat_0123_11, 17);
+ __m256i lhs_mat_0123_20 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 128 + 512 * sb)));
+ __m256i lhs_mat_01_20 = _mm256_permute2f128_si256(lhs_mat_0123_20, lhs_mat_0123_20, 0);
+ __m256i lhs_mat_23_20 = _mm256_permute2f128_si256(lhs_mat_0123_20, lhs_mat_0123_20, 17);
+ __m256i lhs_mat_0123_21 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 160 + 512 * sb)));
+ __m256i lhs_mat_01_21 = _mm256_permute2f128_si256(lhs_mat_0123_21, lhs_mat_0123_21, 0);
+ __m256i lhs_mat_23_21 = _mm256_permute2f128_si256(lhs_mat_0123_21, lhs_mat_0123_21, 17);
+ __m256i lhs_mat_0123_30 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 192 + 512 * sb)));
+ __m256i lhs_mat_01_30 = _mm256_permute2f128_si256(lhs_mat_0123_30, lhs_mat_0123_30, 0);
+ __m256i lhs_mat_23_30 = _mm256_permute2f128_si256(lhs_mat_0123_30, lhs_mat_0123_30, 17);
+ __m256i lhs_mat_0123_31 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 224 + 512 * sb)));
+ __m256i lhs_mat_01_31 = _mm256_permute2f128_si256(lhs_mat_0123_31, lhs_mat_0123_31, 0);
+ __m256i lhs_mat_23_31 = _mm256_permute2f128_si256(lhs_mat_0123_31, lhs_mat_0123_31, 17);
+
+ __m256i lhs_mat_0123_40 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 256 + 512 * sb)));
+ __m256i lhs_mat_01_40 = _mm256_permute2f128_si256(lhs_mat_0123_40, lhs_mat_0123_40, 0);
+ __m256i lhs_mat_23_40 = _mm256_permute2f128_si256(lhs_mat_0123_40, lhs_mat_0123_40, 17);
+ __m256i lhs_mat_0123_41 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 288 + 512 * sb)));
+ __m256i lhs_mat_01_41 = _mm256_permute2f128_si256(lhs_mat_0123_41, lhs_mat_0123_41, 0);
+ __m256i lhs_mat_23_41 = _mm256_permute2f128_si256(lhs_mat_0123_41, lhs_mat_0123_41, 17);
+ __m256i lhs_mat_0123_50 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 320 + 512 * sb)));
+ __m256i lhs_mat_01_50 = _mm256_permute2f128_si256(lhs_mat_0123_50, lhs_mat_0123_50, 0);
+ __m256i lhs_mat_23_50 = _mm256_permute2f128_si256(lhs_mat_0123_50, lhs_mat_0123_50, 17);
+ __m256i lhs_mat_0123_51 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 352 + 512 * sb)));
+ __m256i lhs_mat_01_51 = _mm256_permute2f128_si256(lhs_mat_0123_51, lhs_mat_0123_51, 0);
+ __m256i lhs_mat_23_51 = _mm256_permute2f128_si256(lhs_mat_0123_51, lhs_mat_0123_51, 17);
+ __m256i lhs_mat_0123_60 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 384 + 512 * sb)));
+ __m256i lhs_mat_01_60 = _mm256_permute2f128_si256(lhs_mat_0123_60, lhs_mat_0123_60, 0);
+ __m256i lhs_mat_23_60 = _mm256_permute2f128_si256(lhs_mat_0123_60, lhs_mat_0123_60, 17);
+ __m256i lhs_mat_0123_61 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 416 + 512 * sb)));
+ __m256i lhs_mat_01_61 = _mm256_permute2f128_si256(lhs_mat_0123_61, lhs_mat_0123_61, 0);
+ __m256i lhs_mat_23_61 = _mm256_permute2f128_si256(lhs_mat_0123_61, lhs_mat_0123_61, 17);
+ __m256i lhs_mat_0123_70 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 448 + 512 * sb)));
+ __m256i lhs_mat_01_70 = _mm256_permute2f128_si256(lhs_mat_0123_70, lhs_mat_0123_70, 0);
+ __m256i lhs_mat_23_70 = _mm256_permute2f128_si256(lhs_mat_0123_70, lhs_mat_0123_70, 17);
+ __m256i lhs_mat_0123_71 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 480 + 512 * sb)));
+ __m256i lhs_mat_01_71 = _mm256_permute2f128_si256(lhs_mat_0123_71, lhs_mat_0123_71, 0);
+ __m256i lhs_mat_23_71 = _mm256_permute2f128_si256(lhs_mat_0123_71, lhs_mat_0123_71, 17);
+
+ // Bsums are loaded for the different Q8_K blocks
+ __m128i lhs_raw_bsums_01_0123 = _mm_loadu_si128((const __m128i *)((a_ptr[b].bsums + 32 * sb)));
+ __m128i lhs_raw_bsums_23_0123 = _mm_loadu_si128((const __m128i *)(a_ptr[b].bsums + 8 + 32 * sb));
+ __m128i lhs_raw_bsums_01_4567 = _mm_loadu_si128((const __m128i *)((a_ptr[b].bsums + 16 + 32 * sb)));
+ __m128i lhs_raw_bsums_23_4567 = _mm_loadu_si128((const __m128i *)(a_ptr[b].bsums + 24 + 32 * sb));
+
+ // Shuffle pattern one - left side input
+ const __m256i lhs_mat_01_00_sp1 = _mm256_shuffle_epi32(lhs_mat_01_00, 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)
+ const __m256i lhs_mat_23_00_sp1 = _mm256_shuffle_epi32(lhs_mat_23_00, 160); //A02(0-3) A03(0-3) A02(0-3) A03(0-3) A02(0-3) A03(0-3) A02(0-3) A03(0-3)
+
+ const __m256i lhs_mat_01_01_sp1 = _mm256_shuffle_epi32(lhs_mat_01_01, 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)
+ const __m256i lhs_mat_23_01_sp1 = _mm256_shuffle_epi32(lhs_mat_23_01, 160); //A02(8-11) A03(8-11) A02(8-11) A03(8-11) A02(8-11) A03(8-11) A02(8-11) A03(8-11)
+
+ const __m256i lhs_mat_01_10_sp1 = _mm256_shuffle_epi32(lhs_mat_01_10, 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)
+ const __m256i lhs_mat_23_10_sp1 = _mm256_shuffle_epi32(lhs_mat_23_10, 160); //A12(0-3) A13(0-3) A12(0-3) A13(0-3) A12(0-3) A13(0-3) A12(0-3) A13(0-3)
+
+ const __m256i lhs_mat_01_11_sp1 = _mm256_shuffle_epi32(lhs_mat_01_11, 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)
+ const __m256i lhs_mat_23_11_sp1 = _mm256_shuffle_epi32(lhs_mat_23_11, 160); //A12(8-11) A13(8-11) A12(8-11) A13(8-11) A12(8-11) A13(8-11) A12(8-11) A13(8-11)
+
+ const __m256i lhs_mat_01_20_sp1 = _mm256_shuffle_epi32(lhs_mat_01_20, 160); //A20(0-3) A20(0-3) A21(0-3) A21(0-3) A20(0-3) A20(0-3) A21(0-3) A21(0-3)
+ const __m256i lhs_mat_23_20_sp1 = _mm256_shuffle_epi32(lhs_mat_23_20, 160); //A22(0-3) A23(0-3) A22(0-3) A23(0-3) A22(0-3) A23(0-3) A22(0-3) A23(0-3)
+
+ const __m256i lhs_mat_01_21_sp1 = _mm256_shuffle_epi32(lhs_mat_01_21, 160); //A20(8-11) A20(8-11) A21(8-11) A21(8-11) A20(8-11) A20(8-11) A21(8-11) A21(8-11)
+ const __m256i lhs_mat_23_21_sp1 = _mm256_shuffle_epi32(lhs_mat_23_21, 160); //A22(8-11) A23(8-11) A22(8-11) A23(8-11) A22(8-11) A23(8-11) A22(8-11) A23(8-11)
+
+ const __m256i lhs_mat_01_30_sp1 = _mm256_shuffle_epi32(lhs_mat_01_30, 160); //A30(0-3) A30(0-3) A31(0-3) A31(0-3) A30(0-3) A30(0-3) A31(0-3) A31(0-3)
+ const __m256i lhs_mat_23_30_sp1 = _mm256_shuffle_epi32(lhs_mat_23_30, 160); //A32(0-3) A33(0-3) A32(0-3) A33(0-3) A32(0-3) A33(0-3) A32(0-3) A33(0-3)
+
+ const __m256i lhs_mat_01_31_sp1 = _mm256_shuffle_epi32(lhs_mat_01_31, 160); //A30(8-11) A30(8-11) A31(8-11) A31(8-11) A30(8-11) A30(8-11) A31(8-11) A31(8-11)
+ const __m256i lhs_mat_23_31_sp1 = _mm256_shuffle_epi32(lhs_mat_23_31, 160); //A32(8-11) A33(8-11) A32(8-11) A33(8-11) A32(8-11) A33(8-11) A32(8-11) A33(8-11)
+
+ const __m256i lhs_mat_01_40_sp1 = _mm256_shuffle_epi32(lhs_mat_01_40, 160); //A40(0-3) A40(0-3) A41(0-3) A41(0-3) A40(0-3) A40(0-3) A41(0-3) A41(0-3)
+ const __m256i lhs_mat_23_40_sp1 = _mm256_shuffle_epi32(lhs_mat_23_40, 160); //A42(0-3) A43(0-3) A42(0-3) A43(0-3) A42(0-3) A43(0-3) A42(0-3) A43(0-3)
+
+ const __m256i lhs_mat_01_41_sp1 = _mm256_shuffle_epi32(lhs_mat_01_41, 160); //A40(8-11) A40(8-11) A41(8-11) A41(8-11) A40(8-11) A40(8-11) A41(8-11) A41(8-11)
+ const __m256i lhs_mat_23_41_sp1 = _mm256_shuffle_epi32(lhs_mat_23_41, 160); //A42(8-11) A43(8-11) A42(8-11) A43(8-11) A42(8-11) A43(8-11) A42(8-11) A43(8-11)
+
+ const __m256i lhs_mat_01_50_sp1 = _mm256_shuffle_epi32(lhs_mat_01_50, 160); //A50(0-3) A50(0-3) A51(0-3) A51(0-3) A50(0-3) A50(0-3) A51(0-3) A51(0-3)
+ const __m256i lhs_mat_23_50_sp1 = _mm256_shuffle_epi32(lhs_mat_23_50, 160); //A52(0-3) A53(0-3) A52(0-3) A53(0-3) A52(0-3) A53(0-3) A52(0-3) A53(0-3)
+
+ const __m256i lhs_mat_01_51_sp1 = _mm256_shuffle_epi32(lhs_mat_01_51, 160); //A50(8-11) A50(8-11) A51(8-11) A51(8-11) A50(8-11) A50(8-11) A51(8-11) A51(8-11)
+ const __m256i lhs_mat_23_51_sp1 = _mm256_shuffle_epi32(lhs_mat_23_51, 160); //A52(8-11) A53(8-11) A52(8-11) A53(8-11) A52(8-11) A53(8-11) A52(8-11) A53(8-11)
+
+ const __m256i lhs_mat_01_60_sp1 = _mm256_shuffle_epi32(lhs_mat_01_60, 160); //A60(0-3) A60(0-3) A61(0-3) A61(0-3) A60(0-3) A60(0-3) A61(0-3) A61(0-3)
+ const __m256i lhs_mat_23_60_sp1 = _mm256_shuffle_epi32(lhs_mat_23_60, 160); //A62(0-3) A63(0-3) A62(0-3) A63(0-3) A62(0-3) A63(0-3) A62(0-3) A63(0-3)
+
+ const __m256i lhs_mat_01_61_sp1 = _mm256_shuffle_epi32(lhs_mat_01_61, 160); //A60(8-11) A60(8-11) A61(8-11) A61(8-11) A60(8-11) A60(8-11) A61(8-11) A61(8-11)
+ const __m256i lhs_mat_23_61_sp1 = _mm256_shuffle_epi32(lhs_mat_23_61, 160); //A62(8-11) A63(8-11) A62(8-11) A63(8-11) A62(8-11) A63(8-11) A62(8-11) A63(8-11)
+
+ const __m256i lhs_mat_01_70_sp1 = _mm256_shuffle_epi32(lhs_mat_01_70, 160); //A70(0-3) A70(0-3) A71(0-3) A71(0-3) A70(0-3) A70(0-3) A71(0-3) A71(0-3)
+ const __m256i lhs_mat_23_70_sp1 = _mm256_shuffle_epi32(lhs_mat_23_70, 160); //A72(0-3) A73(0-3) A72(0-3) A73(0-3) A72(0-3) A73(0-3) A72(0-3) A73(0-3)
+
+ const __m256i lhs_mat_01_71_sp1 = _mm256_shuffle_epi32(lhs_mat_01_71, 160); //A70(8-11) A70(8-11) A71(8-11) A71(8-11) A70(8-11) A70(8-11) A71(8-11) A71(8-11)
+ const __m256i lhs_mat_23_71_sp1 = _mm256_shuffle_epi32(lhs_mat_23_71, 160); //A72(8-11) A73(8-11) A72(8-11) A73(8-11) A72(8-11) A73(8-11) A72(8-11) A73(8-11)
+
+ // Shuffle pattern two- left side input
+ const __m256i lhs_mat_01_00_sp2 = _mm256_shuffle_epi32(lhs_mat_01_00, 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)
+ const __m256i lhs_mat_23_00_sp2 = _mm256_shuffle_epi32(lhs_mat_23_00, 245); //A02(4-7) A03(4-7) A02(4-7) A03(4-7) A02(4-7) A03(4-7) A02(4-7) A03(4-7)
+
+ const __m256i lhs_mat_01_01_sp2 = _mm256_shuffle_epi32(lhs_mat_01_01, 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)
+ const __m256i lhs_mat_23_01_sp2 = _mm256_shuffle_epi32(lhs_mat_23_01, 245); //A02(12-15) A03(12-15) A02(12-15) A03(12-15) A02(12-15) A03(12-15) A02(12-15) A03(12-15)
+
+ const __m256i lhs_mat_01_10_sp2 = _mm256_shuffle_epi32(lhs_mat_01_10, 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)
+ const __m256i lhs_mat_23_10_sp2 = _mm256_shuffle_epi32(lhs_mat_23_10, 245); //A12(4-7) A13(4-7) A12(4-7) A13(4-7) A12(4-7) A13(4-7) A12(4-7) A13(4-7)
+
+ const __m256i lhs_mat_01_11_sp2 = _mm256_shuffle_epi32(lhs_mat_01_11, 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)
+ const __m256i lhs_mat_23_11_sp2 = _mm256_shuffle_epi32(lhs_mat_23_11, 245); //A12(12-15) A13(12-15) A12(12-15) A13(12-15) A12(12-15) A13(12-15) A12(12-15) A13(12-15)
+
+ const __m256i lhs_mat_01_20_sp2 = _mm256_shuffle_epi32(lhs_mat_01_20, 245); //A20(4-7) A20(4-7) A21(4-7) A21(4-7) A20(4-7) A20(4-7) A21(4-7) A21(4-7)
+ const __m256i lhs_mat_23_20_sp2 = _mm256_shuffle_epi32(lhs_mat_23_20, 245); //A22(4-7) A23(4-7) A22(4-7) A23(4-7) A22(4-7) A23(4-7) A22(4-7) A23(4-7)
+
+ const __m256i lhs_mat_01_21_sp2 = _mm256_shuffle_epi32(lhs_mat_01_21, 245); //A20(12-15) A20(12-15) A21(12-15) A21(12-15) A20(12-15) A20(12-15) A21(12-15) A21(12-15)
+ const __m256i lhs_mat_23_21_sp2 = _mm256_shuffle_epi32(lhs_mat_23_21, 245); //A22(12-15) A23(12-15) A22(12-15) A23(12-15) A22(12-15) A23(12-15) A22(12-15) A23(12-15)
+
+ const __m256i lhs_mat_01_30_sp2 = _mm256_shuffle_epi32(lhs_mat_01_30, 245); //A30(4-7) A30(4-7) A31(4-7) A31(4-7) A30(4-7) A30(4-7) A31(4-7) A31(4-7)
+ const __m256i lhs_mat_23_30_sp2 = _mm256_shuffle_epi32(lhs_mat_23_30, 245); //A32(4-7) A33(4-7) A32(4-7) A33(4-7) A32(4-7) A33(4-7) A32(4-7) A33(4-7)
+
+ const __m256i lhs_mat_01_31_sp2 = _mm256_shuffle_epi32(lhs_mat_01_31, 245); //A30(12-15) A30(12-15) A31(12-15) A31(12-15) A30(12-15) A30(12-15) A31(12-15) A31(12-15)
+ const __m256i lhs_mat_23_31_sp2 = _mm256_shuffle_epi32(lhs_mat_23_31, 245); //A32(12-15) A33(12-15) A32(12-15) A33(12-15) A32(12-15) A33(12-15) A32(12-15) A33(12-15)
+
+ const __m256i lhs_mat_01_40_sp2 = _mm256_shuffle_epi32(lhs_mat_01_40, 245); //A40(4-7) A40(4-7) A41(4-7) A41(4-7) A40(4-7) A40(4-7) A41(4-7) A41(4-7)
+ const __m256i lhs_mat_23_40_sp2 = _mm256_shuffle_epi32(lhs_mat_23_40, 245); //A42(4-7) A43(4-7) A42(4-7) A43(4-7) A42(4-7) A43(4-7) A42(4-7) A43(4-7)
+
+ const __m256i lhs_mat_01_41_sp2 = _mm256_shuffle_epi32(lhs_mat_01_41, 245); //A40(12-15) A40(12-15) A41(12-15) A41(12-15) A40(12-15) A40(12-15) A41(12-15) A41(12-15)
+ const __m256i lhs_mat_23_41_sp2 = _mm256_shuffle_epi32(lhs_mat_23_41, 245); //A42(12-15) A43(12-15) A42(12-15) A43(12-15) A42(12-15) A43(12-15) A42(12-15) A43(12-15)
+
+ const __m256i lhs_mat_01_50_sp2 = _mm256_shuffle_epi32(lhs_mat_01_50, 245); //A50(4-7) A50(4-7) A51(4-7) A51(4-7) A50(4-7) A50(4-7) A51(4-7) A51(4-7)
+ const __m256i lhs_mat_23_50_sp2 = _mm256_shuffle_epi32(lhs_mat_23_50, 245); //A52(4-7) A53(4-7) A52(4-7) A53(4-7) A52(4-7) A53(4-7) A52(4-7) A53(4-7)
+
+ const __m256i lhs_mat_01_51_sp2 = _mm256_shuffle_epi32(lhs_mat_01_51, 245); //A50(12-15) A50(12-15) A51(12-15) A51(12-15) A50(12-15) A50(12-15) A51(12-15) A51(12-15)
+ const __m256i lhs_mat_23_51_sp2 = _mm256_shuffle_epi32(lhs_mat_23_51, 245); //A52(12-15) A53(12-15) A52(12-15) A53(12-15) A52(12-15) A53(12-15) A52(12-15) A53(12-15)
+
+ const __m256i lhs_mat_01_60_sp2 = _mm256_shuffle_epi32(lhs_mat_01_60, 245); //A60(4-7) A60(4-7) A61(4-7) A61(4-7) A60(4-7) A60(4-7) A61(4-7) A61(4-7)
+ const __m256i lhs_mat_23_60_sp2 = _mm256_shuffle_epi32(lhs_mat_23_60, 245); //A62(4-7) A63(4-7) A62(4-7) A63(4-7) A62(4-7) A63(4-7) A62(4-7) A63(4-7)
+
+ const __m256i lhs_mat_01_61_sp2 = _mm256_shuffle_epi32(lhs_mat_01_61, 245); //A60(12-15) A60(12-15) A61(12-15) A61(12-15) A60(12-15) A60(12-15) A61(12-15) A61(12-15)
+ const __m256i lhs_mat_23_61_sp2 = _mm256_shuffle_epi32(lhs_mat_23_61, 245); //A62(12-15) A63(12-15) A62(12-15) A63(12-15) A62(12-15) A63(12-15) A62(12-15) A63(12-15)
+
+ const __m256i lhs_mat_01_70_sp2 = _mm256_shuffle_epi32(lhs_mat_01_70, 245); //A70(4-7) A70(4-7) A71(4-7) A71(4-7) A70(4-7) A70(4-7) A71(4-7) A71(4-7)
+ const __m256i lhs_mat_23_70_sp2 = _mm256_shuffle_epi32(lhs_mat_23_70, 245); //A72(4-7) A73(4-7) A72(4-7) A73(4-7) A72(4-7) A73(4-7) A72(4-7) A73(4-7)
+
+ const __m256i lhs_mat_01_71_sp2 = _mm256_shuffle_epi32(lhs_mat_01_71, 245); //A70(12-15) A70(12-15) A71(12-15) A71(12-15) A70(12-15) A70(12-15) A71(12-15) A71(12-15)
+ const __m256i lhs_mat_23_71_sp2 = _mm256_shuffle_epi32(lhs_mat_23_71, 245); //A72(12-15) A73(12-15) A72(12-15) A73(12-15) A72(12-15) A73(12-15) A72(12-15) A73(12-15)
+
+ // 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
+ __m256i iacc_mat_00_0_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_00_sp1, lhs_mat_01_00_sp1),_mm256_maddubs_epi16(rhs_mat_0145_01_sp1, lhs_mat_01_01_sp1));
+ __m256i iacc_mat_01_0_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_00_sp1, lhs_mat_01_00_sp1),_mm256_maddubs_epi16(rhs_mat_2367_01_sp1, lhs_mat_01_01_sp1));
+
+ __m256i iacc_mat_10_0_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_00_sp1, lhs_mat_23_00_sp1),_mm256_maddubs_epi16(rhs_mat_0145_01_sp1, lhs_mat_23_01_sp1));
+ __m256i iacc_mat_11_0_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_00_sp1, lhs_mat_23_00_sp1),_mm256_maddubs_epi16(rhs_mat_2367_01_sp1, lhs_mat_23_01_sp1));
+
+ __m256i iacc_mat_00_1_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_10_sp1, lhs_mat_01_10_sp1),_mm256_maddubs_epi16(rhs_mat_0145_11_sp1, lhs_mat_01_11_sp1));
+ __m256i iacc_mat_01_1_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_10_sp1, lhs_mat_01_10_sp1),_mm256_maddubs_epi16(rhs_mat_2367_11_sp1, lhs_mat_01_11_sp1));
+
+ __m256i iacc_mat_10_1_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_10_sp1, lhs_mat_23_10_sp1),_mm256_maddubs_epi16(rhs_mat_0145_11_sp1, lhs_mat_23_11_sp1));
+ __m256i iacc_mat_11_1_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_10_sp1, lhs_mat_23_10_sp1),_mm256_maddubs_epi16(rhs_mat_2367_11_sp1, lhs_mat_23_11_sp1));
+
+ __m256i iacc_mat_00_2_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_20_sp1, lhs_mat_01_20_sp1),_mm256_maddubs_epi16(rhs_mat_0145_21_sp1, lhs_mat_01_21_sp1));
+ __m256i iacc_mat_01_2_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_20_sp1, lhs_mat_01_20_sp1),_mm256_maddubs_epi16(rhs_mat_2367_21_sp1, lhs_mat_01_21_sp1));
+
+ __m256i iacc_mat_10_2_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_20_sp1, lhs_mat_23_20_sp1),_mm256_maddubs_epi16(rhs_mat_0145_21_sp1, lhs_mat_23_21_sp1));
+ __m256i iacc_mat_11_2_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_20_sp1, lhs_mat_23_20_sp1),_mm256_maddubs_epi16(rhs_mat_2367_21_sp1, lhs_mat_23_21_sp1));
+
+ __m256i iacc_mat_00_3_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_30_sp1, lhs_mat_01_30_sp1),_mm256_maddubs_epi16(rhs_mat_0145_31_sp1, lhs_mat_01_31_sp1));
+ __m256i iacc_mat_01_3_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_30_sp1, lhs_mat_01_30_sp1),_mm256_maddubs_epi16(rhs_mat_2367_31_sp1, lhs_mat_01_31_sp1));
+
+ __m256i iacc_mat_10_3_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_30_sp1, lhs_mat_23_30_sp1),_mm256_maddubs_epi16(rhs_mat_0145_31_sp1, lhs_mat_23_31_sp1));
+ __m256i iacc_mat_11_3_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_30_sp1, lhs_mat_23_30_sp1),_mm256_maddubs_epi16(rhs_mat_2367_31_sp1, lhs_mat_23_31_sp1));
+
+ __m256i iacc_mat_00_4_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_40_sp1, lhs_mat_01_40_sp1),_mm256_maddubs_epi16(rhs_mat_0145_41_sp1, lhs_mat_01_41_sp1));
+ __m256i iacc_mat_01_4_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_40_sp1, lhs_mat_01_40_sp1),_mm256_maddubs_epi16(rhs_mat_2367_41_sp1, lhs_mat_01_41_sp1));
+
+ __m256i iacc_mat_10_4_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_40_sp1, lhs_mat_23_40_sp1),_mm256_maddubs_epi16(rhs_mat_0145_41_sp1, lhs_mat_23_41_sp1));
+ __m256i iacc_mat_11_4_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_40_sp1, lhs_mat_23_40_sp1),_mm256_maddubs_epi16(rhs_mat_2367_41_sp1, lhs_mat_23_41_sp1));
+
+ __m256i iacc_mat_00_5_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_50_sp1, lhs_mat_01_50_sp1),_mm256_maddubs_epi16(rhs_mat_0145_51_sp1, lhs_mat_01_51_sp1));
+ __m256i iacc_mat_01_5_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_50_sp1, lhs_mat_01_50_sp1),_mm256_maddubs_epi16(rhs_mat_2367_51_sp1, lhs_mat_01_51_sp1));
+
+ __m256i iacc_mat_10_5_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_50_sp1, lhs_mat_23_50_sp1),_mm256_maddubs_epi16(rhs_mat_0145_51_sp1, lhs_mat_23_51_sp1));
+ __m256i iacc_mat_11_5_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_50_sp1, lhs_mat_23_50_sp1),_mm256_maddubs_epi16(rhs_mat_2367_51_sp1, lhs_mat_23_51_sp1));
+
+ __m256i iacc_mat_00_6_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_60_sp1, lhs_mat_01_60_sp1),_mm256_maddubs_epi16(rhs_mat_0145_61_sp1, lhs_mat_01_61_sp1));
+ __m256i iacc_mat_01_6_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_60_sp1, lhs_mat_01_60_sp1),_mm256_maddubs_epi16(rhs_mat_2367_61_sp1, lhs_mat_01_61_sp1));
+
+ __m256i iacc_mat_10_6_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_60_sp1, lhs_mat_23_60_sp1),_mm256_maddubs_epi16(rhs_mat_0145_61_sp1, lhs_mat_23_61_sp1));
+ __m256i iacc_mat_11_6_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_60_sp1, lhs_mat_23_60_sp1),_mm256_maddubs_epi16(rhs_mat_2367_61_sp1, lhs_mat_23_61_sp1));
+
+ __m256i iacc_mat_00_7_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_70_sp1, lhs_mat_01_70_sp1),_mm256_maddubs_epi16(rhs_mat_0145_71_sp1, lhs_mat_01_71_sp1));
+ __m256i iacc_mat_01_7_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_70_sp1, lhs_mat_01_70_sp1),_mm256_maddubs_epi16(rhs_mat_2367_71_sp1, lhs_mat_01_71_sp1));
+
+ __m256i iacc_mat_10_7_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_70_sp1, lhs_mat_23_70_sp1),_mm256_maddubs_epi16(rhs_mat_0145_71_sp1, lhs_mat_23_71_sp1));
+ __m256i iacc_mat_11_7_sp1 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_70_sp1, lhs_mat_23_70_sp1),_mm256_maddubs_epi16(rhs_mat_2367_71_sp1, lhs_mat_23_71_sp1));
+
+
+ __m256i iacc_mat_00_0_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_00_sp2, lhs_mat_01_00_sp2),_mm256_maddubs_epi16(rhs_mat_0145_01_sp2, lhs_mat_01_01_sp2));
+ __m256i iacc_mat_01_0_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_00_sp2, lhs_mat_01_00_sp2),_mm256_maddubs_epi16(rhs_mat_2367_01_sp2, lhs_mat_01_01_sp2));
+
+ __m256i iacc_mat_10_0_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_00_sp2, lhs_mat_23_00_sp2),_mm256_maddubs_epi16(rhs_mat_0145_01_sp2, lhs_mat_23_01_sp2));
+ __m256i iacc_mat_11_0_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_00_sp2, lhs_mat_23_00_sp2),_mm256_maddubs_epi16(rhs_mat_2367_01_sp2, lhs_mat_23_01_sp2));
+
+ __m256i iacc_mat_00_1_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_10_sp2, lhs_mat_01_10_sp2),_mm256_maddubs_epi16(rhs_mat_0145_11_sp2, lhs_mat_01_11_sp2));
+ __m256i iacc_mat_01_1_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_10_sp2, lhs_mat_01_10_sp2),_mm256_maddubs_epi16(rhs_mat_2367_11_sp2, lhs_mat_01_11_sp2));
+
+ __m256i iacc_mat_10_1_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_10_sp2, lhs_mat_23_10_sp2),_mm256_maddubs_epi16(rhs_mat_0145_11_sp2, lhs_mat_23_11_sp2));
+ __m256i iacc_mat_11_1_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_10_sp2, lhs_mat_23_10_sp2),_mm256_maddubs_epi16(rhs_mat_2367_11_sp2, lhs_mat_23_11_sp2));
+
+ __m256i iacc_mat_00_2_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_20_sp2, lhs_mat_01_20_sp2),_mm256_maddubs_epi16(rhs_mat_0145_21_sp2, lhs_mat_01_21_sp2));
+ __m256i iacc_mat_01_2_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_20_sp2, lhs_mat_01_20_sp2),_mm256_maddubs_epi16(rhs_mat_2367_21_sp2, lhs_mat_01_21_sp2));
+
+ __m256i iacc_mat_10_2_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_20_sp2, lhs_mat_23_20_sp2),_mm256_maddubs_epi16(rhs_mat_0145_21_sp2, lhs_mat_23_21_sp2));
+ __m256i iacc_mat_11_2_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_20_sp2, lhs_mat_23_20_sp2),_mm256_maddubs_epi16(rhs_mat_2367_21_sp2, lhs_mat_23_21_sp2));
+
+ __m256i iacc_mat_00_3_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_30_sp2, lhs_mat_01_30_sp2),_mm256_maddubs_epi16(rhs_mat_0145_31_sp2, lhs_mat_01_31_sp2));
+ __m256i iacc_mat_01_3_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_30_sp2, lhs_mat_01_30_sp2),_mm256_maddubs_epi16(rhs_mat_2367_31_sp2, lhs_mat_01_31_sp2));
+
+ __m256i iacc_mat_10_3_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_30_sp2, lhs_mat_23_30_sp2),_mm256_maddubs_epi16(rhs_mat_0145_31_sp2, lhs_mat_23_31_sp2));
+ __m256i iacc_mat_11_3_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_30_sp2, lhs_mat_23_30_sp2),_mm256_maddubs_epi16(rhs_mat_2367_31_sp2, lhs_mat_23_31_sp2));
+
+ __m256i iacc_mat_00_4_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_40_sp2, lhs_mat_01_40_sp2),_mm256_maddubs_epi16(rhs_mat_0145_41_sp2, lhs_mat_01_41_sp2));
+ __m256i iacc_mat_01_4_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_40_sp2, lhs_mat_01_40_sp2),_mm256_maddubs_epi16(rhs_mat_2367_41_sp2, lhs_mat_01_41_sp2));
+
+ __m256i iacc_mat_10_4_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_40_sp2, lhs_mat_23_40_sp2),_mm256_maddubs_epi16(rhs_mat_0145_41_sp2, lhs_mat_23_41_sp2));
+ __m256i iacc_mat_11_4_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_40_sp2, lhs_mat_23_40_sp2),_mm256_maddubs_epi16(rhs_mat_2367_41_sp2, lhs_mat_23_41_sp2));
+
+ __m256i iacc_mat_00_5_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_50_sp2, lhs_mat_01_50_sp2),_mm256_maddubs_epi16(rhs_mat_0145_51_sp2, lhs_mat_01_51_sp2));
+ __m256i iacc_mat_01_5_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_50_sp2, lhs_mat_01_50_sp2),_mm256_maddubs_epi16(rhs_mat_2367_51_sp2, lhs_mat_01_51_sp2));
+
+ __m256i iacc_mat_10_5_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_50_sp2, lhs_mat_23_50_sp2),_mm256_maddubs_epi16(rhs_mat_0145_51_sp2, lhs_mat_23_51_sp2));
+ __m256i iacc_mat_11_5_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_50_sp2, lhs_mat_23_50_sp2),_mm256_maddubs_epi16(rhs_mat_2367_51_sp2, lhs_mat_23_51_sp2));
+
+ __m256i iacc_mat_00_6_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_60_sp2, lhs_mat_01_60_sp2),_mm256_maddubs_epi16(rhs_mat_0145_61_sp2, lhs_mat_01_61_sp2));
+ __m256i iacc_mat_01_6_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_60_sp2, lhs_mat_01_60_sp2),_mm256_maddubs_epi16(rhs_mat_2367_61_sp2, lhs_mat_01_61_sp2));
+
+ __m256i iacc_mat_10_6_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_60_sp2, lhs_mat_23_60_sp2),_mm256_maddubs_epi16(rhs_mat_0145_61_sp2, lhs_mat_23_61_sp2));
+ __m256i iacc_mat_11_6_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_60_sp2, lhs_mat_23_60_sp2),_mm256_maddubs_epi16(rhs_mat_2367_61_sp2, lhs_mat_23_61_sp2));
+
+ __m256i iacc_mat_00_7_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_70_sp2, lhs_mat_01_70_sp2),_mm256_maddubs_epi16(rhs_mat_0145_71_sp2, lhs_mat_01_71_sp2));
+ __m256i iacc_mat_01_7_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_70_sp2, lhs_mat_01_70_sp2),_mm256_maddubs_epi16(rhs_mat_2367_71_sp2, lhs_mat_01_71_sp2));
+
+ __m256i iacc_mat_10_7_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_70_sp2, lhs_mat_23_70_sp2),_mm256_maddubs_epi16(rhs_mat_0145_71_sp2, lhs_mat_23_71_sp2));
+ __m256i iacc_mat_11_7_sp2 = _mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_70_sp2, lhs_mat_23_70_sp2),_mm256_maddubs_epi16(rhs_mat_2367_71_sp2, lhs_mat_23_71_sp2));
+
+ // Combine results from both shuffle patterns for each output block.
+ __m256i iacc_mat_00_0 = _mm256_add_epi16(iacc_mat_00_0_sp1, iacc_mat_00_0_sp2);
+ __m256i iacc_mat_01_0 = _mm256_add_epi16(iacc_mat_01_0_sp1, iacc_mat_01_0_sp2);
+ __m256i iacc_mat_10_0 = _mm256_add_epi16(iacc_mat_10_0_sp1, iacc_mat_10_0_sp2);
+ __m256i iacc_mat_11_0 = _mm256_add_epi16(iacc_mat_11_0_sp1, iacc_mat_11_0_sp2);
+
+ __m256i iacc_mat_00_1 = _mm256_add_epi16(iacc_mat_00_1_sp1, iacc_mat_00_1_sp2);
+ __m256i iacc_mat_01_1 = _mm256_add_epi16(iacc_mat_01_1_sp1, iacc_mat_01_1_sp2);
+ __m256i iacc_mat_10_1 = _mm256_add_epi16(iacc_mat_10_1_sp1, iacc_mat_10_1_sp2);
+ __m256i iacc_mat_11_1 = _mm256_add_epi16(iacc_mat_11_1_sp1, iacc_mat_11_1_sp2);
+
+ __m256i iacc_mat_00_2 = _mm256_add_epi16(iacc_mat_00_2_sp1, iacc_mat_00_2_sp2);
+ __m256i iacc_mat_01_2 = _mm256_add_epi16(iacc_mat_01_2_sp1, iacc_mat_01_2_sp2);
+ __m256i iacc_mat_10_2 = _mm256_add_epi16(iacc_mat_10_2_sp1, iacc_mat_10_2_sp2);
+ __m256i iacc_mat_11_2 = _mm256_add_epi16(iacc_mat_11_2_sp1, iacc_mat_11_2_sp2);
+
+ __m256i iacc_mat_00_3 = _mm256_add_epi16(iacc_mat_00_3_sp1, iacc_mat_00_3_sp2);
+ __m256i iacc_mat_01_3 = _mm256_add_epi16(iacc_mat_01_3_sp1, iacc_mat_01_3_sp2);
+ __m256i iacc_mat_10_3 = _mm256_add_epi16(iacc_mat_10_3_sp1, iacc_mat_10_3_sp2);
+ __m256i iacc_mat_11_3 = _mm256_add_epi16(iacc_mat_11_3_sp1, iacc_mat_11_3_sp2);
+
+ __m256i iacc_mat_00_4 = _mm256_add_epi16(iacc_mat_00_4_sp1, iacc_mat_00_4_sp2);
+ __m256i iacc_mat_01_4 = _mm256_add_epi16(iacc_mat_01_4_sp1, iacc_mat_01_4_sp2);
+ __m256i iacc_mat_10_4 = _mm256_add_epi16(iacc_mat_10_4_sp1, iacc_mat_10_4_sp2);
+ __m256i iacc_mat_11_4 = _mm256_add_epi16(iacc_mat_11_4_sp1, iacc_mat_11_4_sp2);
+
+ __m256i iacc_mat_00_5 = _mm256_add_epi16(iacc_mat_00_5_sp1, iacc_mat_00_5_sp2);
+ __m256i iacc_mat_01_5 = _mm256_add_epi16(iacc_mat_01_5_sp1, iacc_mat_01_5_sp2);
+ __m256i iacc_mat_10_5 = _mm256_add_epi16(iacc_mat_10_5_sp1, iacc_mat_10_5_sp2);
+ __m256i iacc_mat_11_5 = _mm256_add_epi16(iacc_mat_11_5_sp1, iacc_mat_11_5_sp2);
+
+ __m256i iacc_mat_00_6 = _mm256_add_epi16(iacc_mat_00_6_sp1, iacc_mat_00_6_sp2);
+ __m256i iacc_mat_01_6 = _mm256_add_epi16(iacc_mat_01_6_sp1, iacc_mat_01_6_sp2);
+ __m256i iacc_mat_10_6 = _mm256_add_epi16(iacc_mat_10_6_sp1, iacc_mat_10_6_sp2);
+ __m256i iacc_mat_11_6 = _mm256_add_epi16(iacc_mat_11_6_sp1, iacc_mat_11_6_sp2);
+
+ __m256i iacc_mat_00_7 = _mm256_add_epi16(iacc_mat_00_7_sp1, iacc_mat_00_7_sp2);
+ __m256i iacc_mat_01_7 = _mm256_add_epi16(iacc_mat_01_7_sp1, iacc_mat_01_7_sp2);
+ __m256i iacc_mat_10_7 = _mm256_add_epi16(iacc_mat_10_7_sp1, iacc_mat_10_7_sp2);
+ __m256i iacc_mat_11_7 = _mm256_add_epi16(iacc_mat_11_7_sp1, iacc_mat_11_7_sp2);
+
+ // Output of both shuffle patterns are added in order to sum dot product outputs of all 32 values in block
+ iacc_mat_00_0 = _mm256_madd_epi16(iacc_mat_00_0, scale_0145_0);
+ iacc_mat_01_0 = _mm256_madd_epi16(iacc_mat_01_0, scale_2367_0);
+ iacc_mat_10_0 = _mm256_madd_epi16(iacc_mat_10_0, scale_0145_0);
+ iacc_mat_11_0 = _mm256_madd_epi16(iacc_mat_11_0, scale_2367_0);
+
+ iacc_mat_00_1 = _mm256_madd_epi16(iacc_mat_00_1, scale_0145_1);
+ iacc_mat_01_1 = _mm256_madd_epi16(iacc_mat_01_1, scale_2367_1);
+ iacc_mat_10_1 = _mm256_madd_epi16(iacc_mat_10_1, scale_0145_1);
+ iacc_mat_11_1 = _mm256_madd_epi16(iacc_mat_11_1, scale_2367_1);
+
+ iacc_mat_00_2 = _mm256_madd_epi16(iacc_mat_00_2, scale_0145_2);
+ iacc_mat_01_2 = _mm256_madd_epi16(iacc_mat_01_2, scale_2367_2);
+ iacc_mat_10_2 = _mm256_madd_epi16(iacc_mat_10_2, scale_0145_2);
+ iacc_mat_11_2 = _mm256_madd_epi16(iacc_mat_11_2, scale_2367_2);
+
+ iacc_mat_00_3 = _mm256_madd_epi16(iacc_mat_00_3, scale_0145_3);
+ iacc_mat_01_3 = _mm256_madd_epi16(iacc_mat_01_3, scale_2367_3);
+ iacc_mat_10_3 = _mm256_madd_epi16(iacc_mat_10_3, scale_0145_3);
+ iacc_mat_11_3 = _mm256_madd_epi16(iacc_mat_11_3, scale_2367_3);
+
+ iacc_mat_00_4 = _mm256_madd_epi16(iacc_mat_00_4, scale_0145_4);
+ iacc_mat_01_4 = _mm256_madd_epi16(iacc_mat_01_4, scale_2367_4);
+ iacc_mat_10_4 = _mm256_madd_epi16(iacc_mat_10_4, scale_0145_4);
+ iacc_mat_11_4 = _mm256_madd_epi16(iacc_mat_11_4, scale_2367_4);
+
+ iacc_mat_00_5 = _mm256_madd_epi16(iacc_mat_00_5, scale_0145_5);
+ iacc_mat_01_5 = _mm256_madd_epi16(iacc_mat_01_5, scale_2367_5);
+ iacc_mat_10_5 = _mm256_madd_epi16(iacc_mat_10_5, scale_0145_5);
+ iacc_mat_11_5 = _mm256_madd_epi16(iacc_mat_11_5, scale_2367_5);
+
+ iacc_mat_00_6 = _mm256_madd_epi16(iacc_mat_00_6, scale_0145_6);
+ iacc_mat_01_6 = _mm256_madd_epi16(iacc_mat_01_6, scale_2367_6);
+ iacc_mat_10_6 = _mm256_madd_epi16(iacc_mat_10_6, scale_0145_6);
+ iacc_mat_11_6 = _mm256_madd_epi16(iacc_mat_11_6, scale_2367_6);
+
+ iacc_mat_00_7 = _mm256_madd_epi16(iacc_mat_00_7, scale_0145_7);
+ iacc_mat_01_7 = _mm256_madd_epi16(iacc_mat_01_7, scale_2367_7);
+ iacc_mat_10_7 = _mm256_madd_epi16(iacc_mat_10_7, scale_0145_7);
+ iacc_mat_11_7 = _mm256_madd_epi16(iacc_mat_11_7, scale_2367_7);
+
+ __m256i iacc_mat_00 = _mm256_add_epi32(_mm256_add_epi32(_mm256_add_epi32(iacc_mat_00_0, iacc_mat_00_1), _mm256_add_epi32(iacc_mat_00_2, iacc_mat_00_3)), _mm256_add_epi32(_mm256_add_epi32(iacc_mat_00_4, iacc_mat_00_5), _mm256_add_epi32(iacc_mat_00_6, iacc_mat_00_7)));
+ __m256i iacc_mat_01 = _mm256_add_epi32(_mm256_add_epi32(_mm256_add_epi32(iacc_mat_01_0, iacc_mat_01_1), _mm256_add_epi32(iacc_mat_01_2, iacc_mat_01_3)), _mm256_add_epi32(_mm256_add_epi32(iacc_mat_01_4, iacc_mat_01_5), _mm256_add_epi32(iacc_mat_01_6, iacc_mat_01_7)));
+ __m256i iacc_mat_10 = _mm256_add_epi32(_mm256_add_epi32(_mm256_add_epi32(iacc_mat_10_0, iacc_mat_10_1), _mm256_add_epi32(iacc_mat_10_2, iacc_mat_10_3)), _mm256_add_epi32(_mm256_add_epi32(iacc_mat_10_4, iacc_mat_10_5), _mm256_add_epi32(iacc_mat_10_6, iacc_mat_10_7)));
+ __m256i iacc_mat_11 = _mm256_add_epi32(_mm256_add_epi32(_mm256_add_epi32(iacc_mat_11_0, iacc_mat_11_1), _mm256_add_epi32(iacc_mat_11_2, iacc_mat_11_3)), _mm256_add_epi32(_mm256_add_epi32(iacc_mat_11_4, iacc_mat_11_5), _mm256_add_epi32(iacc_mat_11_6, iacc_mat_11_7)));
+
+ // Straighten out to make 4 row vectors
+ __m256i iacc_row_0 = _mm256_blend_epi32(iacc_mat_00, _mm256_shuffle_epi32(iacc_mat_01, 78), 204);
+ __m256i iacc_row_1 = _mm256_blend_epi32(_mm256_shuffle_epi32(iacc_mat_00, 78), iacc_mat_01, 204);
+ __m256i iacc_row_2 = _mm256_blend_epi32(iacc_mat_10, _mm256_shuffle_epi32(iacc_mat_11, 78), 204);
+ __m256i iacc_row_3 = _mm256_blend_epi32(_mm256_shuffle_epi32(iacc_mat_10, 78), iacc_mat_11, 204);
+
+ // Load the scale(d) values for all the 4 Q8_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 = _mm256_set_m128(row_scale_f32_sse, row_scale_f32_sse);
+
+ // Multiply with appropiate scales and accumulate (for both d and dmin) below
+ acc_rows[0] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_0), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 0)), acc_rows[0]);
+ acc_rows[1] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_1), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 85)), acc_rows[1]);
+ acc_rows[2] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_2), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_rows[2]);
+ acc_rows[3] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_3), _mm256_mul_ps(col_scale_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 255)), acc_rows[3]);
+
+ __m256i lhs_bsums_01_0123 = _mm256_inserti128_si256(_mm256_castsi128_si256(lhs_raw_bsums_01_0123), lhs_raw_bsums_01_0123, 1);
+ __m256i lhs_bsums_23_0123 = _mm256_inserti128_si256(_mm256_castsi128_si256(lhs_raw_bsums_23_0123), lhs_raw_bsums_23_0123, 1);
+ __m256i lhs_bsums_01_4567 = _mm256_inserti128_si256(_mm256_castsi128_si256(lhs_raw_bsums_01_4567), lhs_raw_bsums_01_4567, 1);
+ __m256i lhs_bsums_23_4567 = _mm256_inserti128_si256(_mm256_castsi128_si256(lhs_raw_bsums_23_4567), lhs_raw_bsums_23_4567, 1);
+
+ // Take two bsums from two Q8_Ks at a time and multiply with corresponding mins values from each Q2_K
+ __m256i iacc_row_min_0_01 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_01_0123, 0), mins_01);
+ __m256i iacc_row_min_1_01 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_01_0123, 170), mins_01);
+ __m256i iacc_row_min_2_01 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_23_0123, 0), mins_01);
+ __m256i iacc_row_min_3_01 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_23_0123, 170), mins_01);
+
+ __m256i iacc_row_min_0_23 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_01_0123, 85), mins_23);
+ __m256i iacc_row_min_1_23 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_01_0123, 255), mins_23);
+ __m256i iacc_row_min_2_23 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_23_0123, 85), mins_23);
+ __m256i iacc_row_min_3_23 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_23_0123, 255), mins_23);
+
+ __m256i iacc_row_min_0_45 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_01_4567, 0), mins_45);
+ __m256i iacc_row_min_1_45 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_01_4567, 170), mins_45);
+ __m256i iacc_row_min_2_45 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_23_4567, 0), mins_45);
+ __m256i iacc_row_min_3_45 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_23_4567, 170), mins_45);
+
+ __m256i iacc_row_min_0_67 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_01_4567, 85), mins_67);
+ __m256i iacc_row_min_1_67 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_01_4567, 255), mins_67);
+ __m256i iacc_row_min_2_67 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_23_4567, 85), mins_67);
+ __m256i iacc_row_min_3_67 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_23_4567, 255), mins_67);
+
+ __m256i iacc_row_min_0 = _mm256_add_epi32(_mm256_add_epi32(iacc_row_min_0_01, iacc_row_min_0_23), _mm256_add_epi32(iacc_row_min_0_45,iacc_row_min_0_67));
+ __m256i iacc_row_min_1 = _mm256_add_epi32(_mm256_add_epi32(iacc_row_min_1_01, iacc_row_min_1_23), _mm256_add_epi32(iacc_row_min_1_45,iacc_row_min_1_67));
+ __m256i iacc_row_min_2 = _mm256_add_epi32(_mm256_add_epi32(iacc_row_min_2_01, iacc_row_min_2_23), _mm256_add_epi32(iacc_row_min_2_45,iacc_row_min_2_67));
+ __m256i iacc_row_min_3 = _mm256_add_epi32(_mm256_add_epi32(iacc_row_min_3_01, iacc_row_min_3_23), _mm256_add_epi32(iacc_row_min_3_45,iacc_row_min_3_67));
+
+ acc_min_rows[0] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_min_0), _mm256_mul_ps(col_dmin_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 0)), acc_min_rows[0]);
+ acc_min_rows[1] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_min_1), _mm256_mul_ps(col_dmin_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 85)), acc_min_rows[1]);
+ acc_min_rows[2] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_min_2), _mm256_mul_ps(col_dmin_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 170)), acc_min_rows[2]);
+ acc_min_rows[3] = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_row_min_3), _mm256_mul_ps(col_dmin_f32, _mm256_shuffle_ps(row_scale_f32, row_scale_f32, 255)), acc_min_rows[3]);
+ }
+ }
+ // Store the accumulated values
+ for (int i = 0; i < 4; i++) {
+ _mm256_storeu_ps((float * )(s + ((y * 4 + i) * bs + x * 8)), _mm256_sub_ps(acc_rows[i], acc_min_rows[i]));
+ }
+ }
+ }
+#else
+
+ float sumf[4][8];
+ float sum_minf[4][8];
+ int sumi1, sumi2, sumi3, sumi4;
+ int sumi;
+
+ for (int y = 0; y < nr / 4; y++) {
+ const block_q8_Kx4 * a_ptr = (const block_q8_Kx4 *) vy + (y * nb);
+ for (int x = 0; x < nc / ncols_interleaved; x++) {
+ const block_q2_Kx8 * b_ptr = (const block_q2_Kx8 *) vx + (x * nb);
+ for (int m = 0; m < 4; m++) {
+ for (int j = 0; j < ncols_interleaved; j++) {
+ sumf[m][j] = 0.0;
+ sum_minf[m][j] = 0.0;
+ }
+ }
+ for (int l = 0; l < nb; l++) {
+ for (int k = 0; k < (qk / (4 * blocklen)); k++) {
+
+ uint8_t *scales_0 = (uint8_t*) b_ptr[l].scales + (k / 4) * 64 ;
+ uint8_t *scales_1 = (uint8_t*) b_ptr[l].scales + (k / 4) * 64 + 16;
+ uint8_t *scales_2 = (uint8_t*) b_ptr[l].scales + (k / 4) * 64 + 32;
+ uint8_t *scales_3 = (uint8_t*) b_ptr[l].scales + (k / 4) * 64 + 48;
+ for (int m = 0; m < 4; m++) {
+ for (int j = 0; j < ncols_interleaved; j++) {
+ sumi1 = 0;
+ sumi2 = 0;
+ sumi3 = 0;
+ sumi4 = 0;
+ sumi = 0;
+ int offset = ((k / 2) % 2) + j * 2;
+ for (int i = 0; i < blocklen; ++i){
+ const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0x03);
+ const int v1 = (int8_t) ((b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 2 ) & 0x03);
+ const int v2 = (int8_t) ((b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4 ) & 0x03);
+ const int v3 = (int8_t) ((b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 6 ) & 0x03);
+ sumi1 = (v0 * a_ptr[l].qs[(k >> 2) * 512 + (k % 4) * 4 * blocklen + m * blocklen + i]);
+ sumi2 = (v1 * a_ptr[l].qs[(k >> 2) * 512 + (k % 4) * 4 * blocklen + m * blocklen + i + 128]);
+ sumi3 = (v2 * a_ptr[l].qs[(k >> 2) * 512 + (k % 4) * 4 * blocklen + m * blocklen + i + 256]);
+ sumi4 = (v3 * a_ptr[l].qs[(k >> 2) * 512 + (k % 4) * 4 * blocklen + m * blocklen + i + 384]);
+ sumi1 = sumi1 * (scales_0[offset] & 0xF);
+ sumi2 = sumi2 * (scales_1[offset] & 0xF);
+ sumi3 = sumi3 * (scales_2[offset] & 0xF);
+ sumi4 = sumi4 * (scales_3[offset] & 0xF);
+ sumi += sumi1 + sumi2 + sumi3 + sumi4;
+ }
+ sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d[m];
+ }
+ }
+ }
+ for(int sb = 0; sb < 8; sb++) {
+ uint8_t *mins = (uint8_t*) b_ptr[l].scales + sb * 16;
+ for(int m = 0; m < 4; m++) {
+ const int16_t *bsums = a_ptr[l].bsums + (sb * 8) + (m * 4) - ((sb % 2) * 6);
+ for(int j = 0; j < ncols_interleaved; j++) {
+ int mins_prod = ((mins[j * 2] >> 4) * bsums[0] + (mins[(j * 2)+ 1] >> 4) * bsums[1]);
+ sum_minf[m][j] += (mins_prod) * GGML_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d[m];
+ }
+ }
+ }
+ }
+
+ for (int m = 0; m < 4; m++) {
+ for (int j = 0; j < ncols_interleaved; j++) {
+ s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j] - sum_minf[m][j];
+ }
+ }
+ }
+ }
+
+#endif
+}
diff --git a/ggml/src/ggml-cpu/repack.cpp b/ggml/src/ggml-cpu/repack.cpp
index 72ee93a5abc7c..c2cca6b3a582a 100644
--- a/ggml/src/ggml-cpu/repack.cpp
+++ b/ggml/src/ggml-cpu/repack.cpp
@@ -413,6 +413,82 @@ void ggml_gemv_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs,
}
}
+void ggml_gemv_q2_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+ const int qk = QK_K;
+ const int nb = n / qk;
+ const int ncols_interleaved = 8;
+ const int blocklen = 8;
+
+ assert (n % qk == 0);
+ assert (nc % ncols_interleaved == 0);
+
+ UNUSED(s);
+ UNUSED(bs);
+ UNUSED(vx);
+ UNUSED(vy);
+ UNUSED(nr);
+ UNUSED(nc);
+ UNUSED(nb);
+ UNUSED(ncols_interleaved);
+ UNUSED(blocklen);
+
+ float sumf[8];
+ float sum_minf[8];
+ int sumi1,sumi2,sumi3,sumi4;
+ int sumi;
+
+ const block_q8_K * a_ptr = (const block_q8_K *)vy;
+ for(int x = 0; x < nc / ncols_interleaved; x++) {
+ const block_q2_Kx8 * b_ptr = (const block_q2_Kx8 *) vx + (x * nb);
+ for (int j = 0; j < ncols_interleaved; j++) {
+ sumf[j] = 0.0;
+ sum_minf[j] = 0.0;
+ }
+ for (int l = 0; l < nb; l++) {
+ for (int k = 0; k < (qk / (4 * blocklen)); k++) {
+ const uint8_t *scales_0 = b_ptr[l].scales + (k / 4) * 64 ;
+ const uint8_t *scales_1 = b_ptr[l].scales + (k / 4) * 64 + 16;
+ const uint8_t *scales_2 = b_ptr[l].scales + (k / 4) * 64 + 32;
+ const uint8_t *scales_3 = b_ptr[l].scales + (k / 4) * 64 + 48;
+ for (int j = 0; j < ncols_interleaved; j++) {
+ sumi1 = 0;
+ sumi2 = 0;
+ sumi3 = 0;
+ sumi4 = 0;
+ sumi = 0;
+ int offset = ((k / 2) % 2) + j * 2;
+ for (int i = 0; i < blocklen; ++i){
+ const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 3);
+ const int v1 = (int8_t) ((b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 2 ) & 3);
+ const int v2 = (int8_t) ((b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4 ) & 3);
+ const int v3 = (int8_t) ((b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 6 ) & 3);
+ sumi1 = (v0 * a_ptr[l].qs[(k >> 2) * 128 + (k % 4) * blocklen + i]);
+ sumi2 = (v1 * a_ptr[l].qs[(k >> 2) * 128 + (k % 4) * blocklen + i + 32]);
+ sumi3 = (v2 * a_ptr[l].qs[(k >> 2) * 128 + (k % 4) * blocklen + i + 64]);
+ sumi4 = (v3 * a_ptr[l].qs[(k >> 2) * 128 + (k % 4) * blocklen + i + 96]);
+
+ sumi1 = sumi1 * (scales_0[offset] & 0xF);
+ sumi2 = sumi2 * (scales_1[offset] & 0xF);
+ sumi3 = sumi3 * (scales_2[offset] & 0xF);
+ sumi4 = sumi4 * (scales_3[offset] & 0xF);
+ sumi += sumi1 + sumi2 + sumi3 + sumi4;
+ }
+ sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d;
+ }
+ }
+ for(int sb = 0; sb < 8; sb++) {
+ const uint8_t *mins = b_ptr[l].scales + sb * 16;
+ for(int j = 0; j < ncols_interleaved; j++){
+ sum_minf[j] += ((mins[j * 2] >> 4) * a_ptr[l].bsums[sb * 2] + (mins[(j * 2)+ 1] >> 4) * a_ptr[l].bsums[sb * 2 + 1]) * GGML_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d;
+ }
+ }
+ }
+ for (int j = 0; j < ncols_interleaved; j++) {
+ s[x * ncols_interleaved + j] = sumf[j] - sum_minf[j];
+ }
+ }
+}
+
void ggml_gemv_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
const int qk = QK8_0;
const int nb = n / qk;
@@ -712,6 +788,97 @@ void ggml_gemm_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs,
}
}
+void ggml_gemm_q2_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+ const int qk = QK_K;
+ const int nb = n / qk;
+ const int ncols_interleaved = 8;
+ const int blocklen = 8;
+
+ assert (n % qk == 0);
+ assert (nr % 4 == 0);
+ assert (nc % ncols_interleaved == 0);
+
+ UNUSED(s);
+ UNUSED(bs);
+ UNUSED(vx);
+ UNUSED(vy);
+ UNUSED(nr);
+ UNUSED(nc);
+ UNUSED(nb);
+ UNUSED(ncols_interleaved);
+ UNUSED(blocklen);
+
+ float sumf[4][8];
+ float sum_minf[4][8];
+ int sumi1, sumi2, sumi3, sumi4;
+ int sumi;
+
+ for (int y = 0; y < nr / 4; y++) {
+ const block_q8_Kx4 * a_ptr = (const block_q8_Kx4 *) vy + (y * nb);
+ for (int x = 0; x < nc / ncols_interleaved; x++) {
+ const block_q2_Kx8 * b_ptr = (const block_q2_Kx8 *) vx + (x * nb);
+ for (int m = 0; m < 4; m++) {
+ for (int j = 0; j < ncols_interleaved; j++) {
+ sumf[m][j] = 0.0;
+ sum_minf[m][j] = 0.0;
+ }
+ }
+ for (int l = 0; l < nb; l++) {
+ for (int k = 0; k < (qk / (4 * blocklen)); k++) {
+
+ const uint8_t *scales_0 = b_ptr[l].scales + (k / 4) * 64 ;
+ const uint8_t *scales_1 = b_ptr[l].scales + (k / 4) * 64 + 16;
+ const uint8_t *scales_2 = b_ptr[l].scales + (k / 4) * 64 + 32;
+ const uint8_t *scales_3 = b_ptr[l].scales + (k / 4) * 64 + 48;
+ for (int m = 0; m < 4; m++) {
+ for (int j = 0; j < ncols_interleaved; j++) {
+ sumi1 = 0;
+ sumi2 = 0;
+ sumi3 = 0;
+ sumi4 = 0;
+ sumi = 0;
+ int offset = ((k / 2) % 2) + j * 2;
+ for (int i = 0; i < blocklen; ++i){
+ const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 3);
+ const int v1 = (int8_t) ((b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 2 ) & 3);
+ const int v2 = (int8_t) ((b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4 ) & 3);
+ const int v3 = (int8_t) ((b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 6 ) & 3);
+ sumi1 = (v0 * a_ptr[l].qs[(k >> 2) * 512 + (k % 4) * 4 * blocklen + m * blocklen + i]);
+ sumi2 = (v1 * a_ptr[l].qs[(k >> 2) * 512 + (k % 4) * 4 * blocklen + m * blocklen + i + 128]);
+ sumi3 = (v2 * a_ptr[l].qs[(k >> 2) * 512 + (k % 4) * 4 * blocklen + m * blocklen + i + 256]);
+ sumi4 = (v3 * a_ptr[l].qs[(k >> 2) * 512 + (k % 4) * 4 * blocklen + m * blocklen + i + 384]);
+ sumi1 = sumi1 * (scales_0[offset] & 0xF);
+ sumi2 = sumi2 * (scales_1[offset] & 0xF);
+ sumi3 = sumi3 * (scales_2[offset] & 0xF);
+ sumi4 = sumi4 * (scales_3[offset] & 0xF);
+ sumi += sumi1 + sumi2 + sumi3 + sumi4;
+ }
+ sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d[m];
+ }
+ }
+ }
+ for(int sb = 0; sb < 8; sb++) {
+ const uint8_t *mins = b_ptr[l].scales + sb * 16;
+ for(int m = 0; m < 4; m++) {
+ const int16_t *bsums = a_ptr[l].bsums + (sb * 8) + (m * 4) - ((sb % 2) * 6);
+ for(int j = 0; j < ncols_interleaved; j++) {
+ int mins_prod = ((mins[j * 2] >> 4) * bsums[0] + (mins[(j * 2)+ 1] >> 4) * bsums[1]);
+ sum_minf[m][j] += (mins_prod) * GGML_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d[m];
+ }
+ }
+ }
+ }
+
+ for (int m = 0; m < 4; m++) {
+ for (int j = 0; j < ncols_interleaved; j++) {
+ s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j] - sum_minf[m][j];
+ }
+ }
+ }
+ }
+}
+
+
void ggml_gemm_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
const int qk = QK8_0;
const int nb = n / qk;
@@ -915,6 +1082,50 @@ static block_q4_Kx8 make_block_q4_Kx8(block_q4_K * in, unsigned int blck_size_in
return out;
}
+static block_q2_Kx8 make_block_q2_Kx8(block_q2_K * in, unsigned int blck_size_interleave) {
+ block_q2_Kx8 out;
+
+ // Delta(scale) and dmin values of the eight Q2_K structures are copied onto the output interleaved structure
+ for (int i = 0; i < 8; i++) {
+ out.d[i] = in[i].GGML_COMMON_AGGR_U.GGML_COMMON_AGGR_S.d;
+ }
+
+ for (int i = 0; i < 8; i++) {
+ out.dmin[i] = in[i].GGML_COMMON_AGGR_U.GGML_COMMON_AGGR_S.dmin;
+ }
+
+ const int end = QK_K * 2 / blck_size_interleave;
+
+ // Interleave Q4_K quants by taking 8 bytes at a time
+ for (int i = 0; i < end; ++i) {
+ int src_id = i % 8;
+ int src_offset = (i / 8) * blck_size_interleave;
+ int dst_offset = i * blck_size_interleave;
+
+ uint64_t elems;
+ memcpy(&elems, &in[src_id].qs[src_offset], sizeof(uint64_t));
+ memcpy(&out.qs[dst_offset], &elems, sizeof(uint64_t));
+ }
+
+ // The below logic is designed so as to unapck and rearrange scales and mins values in Q2_K
+ // Currently the Q2_K structure has 16 scales and 16 mins packed in 16 bytes ( 4 bits for each value)
+ // The output Q2_Kx8 structure has 128 bytes for storing scales and mins
+ // Every 16 byte is packed such that it contains scales and mins for corresponding sub blocks from Q2_K structure
+ // For eg - First 16 bytes contains 16 scales and 16 mins - each of first and second sub blocks from different Q2_K structures
+
+ for(int i = 0; i < 128; i++){
+
+ // Index for selecting which q2k super block
+ int src1 = (i % 16) / 2;
+ // Index for selecting scale
+ int src2 = ((i / 16) * 2) + (i % 2);
+
+ out.scales[i] = in[src1].scales[src2];
+ }
+ return out;
+
+}
+
static int repack_q4_0_to_q4_0_4_bl(struct ggml_tensor * t, int interleave_block, const void * GGML_RESTRICT data, size_t data_size) {
GGML_ASSERT(t->type == GGML_TYPE_Q4_0);
GGML_ASSERT(interleave_block == 4 || interleave_block == 8);
@@ -976,6 +1187,37 @@ static int repack_q4_K_to_q4_K_8_bl(struct ggml_tensor * t, int interleave_block
GGML_UNUSED(data_size);
}
+static int repack_q2_K_to_q2_K_8_bl(struct ggml_tensor * t, int interleave_block, const void * GGML_RESTRICT data, size_t data_size) {
+ GGML_ASSERT(t->type == GGML_TYPE_Q2_K);
+ GGML_ASSERT(interleave_block == 8);
+ constexpr int nrows_interleaved = 8;
+
+ block_q2_Kx8 * dst = (block_q2_Kx8*)t->data;
+ const block_q2_K * src = (const block_q2_K*) data;
+ block_q2_K dst_tmp[8];
+ int nrow = ggml_nrows(t);
+ int nblocks = t->ne[0] / QK_K;
+
+ GGML_ASSERT(data_size == nrow * nblocks * sizeof(block_q2_K));
+
+ if (t->ne[1] % nrows_interleaved != 0 || t->ne[0] % 8 != 0) {
+ return -1;
+ }
+
+ for (int b = 0; b < nrow; b += nrows_interleaved) {
+ for (int64_t x = 0; x < nblocks; x++) {
+ for (int i = 0; i < nrows_interleaved; i++ ) {
+ dst_tmp[i] = src[x + i * nblocks];
+ }
+ *dst++ = make_block_q2_Kx8(dst_tmp, interleave_block);
+ }
+ src += nrows_interleaved * nblocks;
+ }
+ return 0;
+
+ GGML_UNUSED(data_size);
+}
+
static int repack_q4_0_to_q4_0_8_bl(struct ggml_tensor * t, int interleave_block, const void * GGML_RESTRICT data, size_t data_size) {
GGML_ASSERT(t->type == GGML_TYPE_Q4_0);
GGML_ASSERT(interleave_block == 8);
@@ -1096,6 +1338,10 @@ template <> int repack<block_q4_K, 8, 8>(struct ggml_tensor * t, const void * da
return repack_q4_K_to_q4_K_8_bl(t, 8, data, data_size);
}
+template <> int repack<block_q2_K, 8, 8>(struct ggml_tensor * t, const void * data, size_t data_size) {
+ return repack_q2_K_to_q2_K_8_bl(t, 8, data, data_size);
+}
+
template <> int repack<block_iq4_nl, 4, 4>(struct ggml_tensor * t, const void * data, size_t data_size) {
return repack_iq4_nl_to_iq4_nl_4_bl(t, 4, data, data_size);
}
@@ -1125,6 +1371,10 @@ template <> void gemv<block_q4_K, 8, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t
ggml_gemv_q4_K_8x8_q8_K(n, s, bs, vx, vy, nr, nc);
}
+template <> void gemv<block_q2_K, 8, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
+ ggml_gemv_q2_K_8x8_q8_K(n, s, bs, vx, vy, nr, nc);
+}
+
template <> void gemv<block_iq4_nl, 4, 4, GGML_TYPE_Q8_0>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
ggml_gemv_iq4_nl_4x4_q8_0(n, s, bs, vx, vy, nr, nc);
}
@@ -1149,6 +1399,10 @@ template <> void gemm<block_q4_K, 8, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t
ggml_gemm_q4_K_8x8_q8_K(n, s, bs, vx, vy, nr, nc);
}
+template <> void gemm<block_q2_K, 8, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
+ ggml_gemm_q2_K_8x8_q8_K(n, s, bs, vx, vy, nr, nc);
+}
+
template <> void gemm<block_iq4_nl, 4, 4, GGML_TYPE_Q8_0>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
ggml_gemm_iq4_nl_4x4_q8_0(n, s, bs, vx, vy, nr, nc);
}
@@ -1422,6 +1676,9 @@ static const ggml::cpu::tensor_traits * ggml_repack_get_optimal_repack_type(cons
static const ggml::cpu::repack::tensor_traits<block_q4_0, 8, 8, GGML_TYPE_Q8_0> q4_0_8x8_q8_0;
static const ggml::cpu::repack::tensor_traits<block_q4_K, 8, 8, GGML_TYPE_Q8_K> q4_K_8x8_q8_K;
+ // instance for Q2
+ static const ggml::cpu::repack::tensor_traits<block_q2_K, 8, 8, GGML_TYPE_Q8_K> q2_K_8x8_q8_K;
+
// instance for IQ4
static const ggml::cpu::repack::tensor_traits<block_iq4_nl, 4, 4, GGML_TYPE_Q8_0> iq4_nl_4x4_q8_0;
@@ -1447,6 +1704,12 @@ static const ggml::cpu::tensor_traits * ggml_repack_get_optimal_repack_type(cons
return &q4_K_8x8_q8_K;
}
}
+ } else if (cur->type == GGML_TYPE_Q2_K) {
+ if (ggml_cpu_has_avx512()) {
+ if (cur->ne[1] % 8 == 0) {
+ return &q2_K_8x8_q8_K;
+ }
+ }
} else if (cur->type == GGML_TYPE_IQ4_NL) {
if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) {
if (cur->ne[1] % 4 == 0) {
diff --git a/ggml/src/ggml-cpu/repack.h b/ggml/src/ggml-cpu/repack.h
index 4421e5f8e7046..cd322e74389d6 100644
--- a/ggml/src/ggml-cpu/repack.h
+++ b/ggml/src/ggml-cpu/repack.h
@@ -44,7 +44,14 @@ struct block_q4_Kx8 {
};
static_assert(sizeof(block_q4_Kx8) == sizeof(ggml_half) * 16 + K_SCALE_SIZE * 8 + QK_K * 4, "wrong q4_K block size/padding");
+struct block_q2_Kx8 {
+ ggml_half d[8]; // super-block scale for quantized scales
+ ggml_half dmin[8]; // super-block scale for quantized mins
+ uint8_t scales[128]; // scales and mins, quantized with 4 bits
+ uint8_t qs[512]; // 2--bit quants
+};
+static_assert(sizeof(block_q2_Kx8) == sizeof(ggml_half) * 16 + QK_K/2 + QK_K * 2, "wrong q2_K block size/padding");
struct block_q8_Kx4 {
float d[4]; // delta
int8_t qs[QK_K * 4]; // quants
@@ -71,11 +78,13 @@ void ggml_gemv_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
void ggml_gemv_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemv_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q2_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemv_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemm_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemm_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q2_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemm_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
// Native implementations
@@ -86,11 +95,13 @@ void ggml_gemv_q4_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs,
void ggml_gemv_q4_0_4x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemv_q4_0_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemv_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q2_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemv_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemm_q4_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemm_q4_0_4x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemm_q4_0_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemm_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q2_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
void ggml_gemm_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
#if defined(__cplusplus)