hey! this is a static git repo viewer. commit diffs, file history, older commits and other revisions won't load without javascript.

you are still able to view source files on branch/tag tips, but the experience is degraded.

char-fork/llama.cpp

git clone https://git.t4t.associates/char-fork/llama.cpp

Shaw Nguyenggml-cpu/x86: fix unused changemask warning in repack (#20692)78d550b542026-03-18

1#define GGML_COMMON_IMPL_CPP
2#define GGML_COMMON_DECL_CPP
3#include "ggml-common.h"
4#include "ggml-backend-impl.h"
5
6#include "ggml-impl.h"
7#include "ggml-cpu.h"
8#include "ggml-cpu-impl.h"
9#include "simd-mappings.h"
10#include "traits.h"
11
12#include <cmath>
13#include <cstring>
14#include <cassert>
15#include <cstdlib> // for qsort
16#include <cstdio>  // for GGML_ASSERT
17
18#define GGML_CPU_CLANG_WORKAROUND
19#include "../../repack.h"
20
21#if defined(__GNUC__)
22#pragma GCC diagnostic ignored "-Woverlength-strings"
23#endif
24
25#define UNUSED GGML_UNUSED
26
27#if defined(__AVX__)
28#if defined(__F16C__)
29#if defined(__AVX512F__)
30#define GGML_F32Cx8x2_LOAD(x, y)     _mm512_cvtph_ps(_mm256_set_m128i(_mm_loadu_si128((const __m128i *)(y)), _mm_loadu_si128((const __m128i *)(x))))
31#define GGML_F32Cx16_REPEAT_LOAD(x)  _mm512_cvtph_ps(_mm256_set_m128i(x, x))
32#endif
33// the  _mm256_cvt intrinsics require F16C
34#define GGML_F32Cx8_LOAD(x)     _mm256_cvtph_ps(_mm_loadu_si128((const __m128i *)(x)))
35#define GGML_F32Cx8_REPEAT_LOAD(x, loadMask)     _mm256_cvtph_ps(_mm_shuffle_epi32(_mm_maskload_epi32((int const*)(x), loadMask), 68))
36#define GGML_F32Cx8_REARRANGE_LOAD(x, arrangeMask)     _mm256_cvtph_ps(_mm_shuffle_epi8(_mm_loadu_si128((const __m128i *) x), arrangeMask))
37#else
38#if defined(__AVX512F__)
39static inline __m512 __avx512_f32cx8x2_load(ggml_fp16_t *x, ggml_fp16_t *y) {
40    float tmp[16];
41
42    for (int i = 0; i < 8; i++) {
43        tmp[i] = GGML_CPU_FP16_TO_FP32(x[i]);
44    }
45
46    for (int i = 0; i < 8; i++) {
47        tmp[i + 8] = GGML_CPU_FP16_TO_FP32(y[i]);
48    }
49
50    return _mm512_loadu_ps(tmp);
51}
52static inline __m512 __avx512_repeat_f32cx16_load(__m128i x) {
53    float tmp[16];
54    uint16_t tmphalf[8];
55    _mm_storeu_si128((__m128i*)tmphalf, x);
56
57    for (int i = 0; i < 4; i++) {
58        tmp[i] = GGML_CPU_FP16_TO_FP32(tmphalf[i]);
59        tmp[i + 4] = GGML_CPU_FP16_TO_FP32(tmphalf[i]);
60        tmp[i + 8] = GGML_CPU_FP16_TO_FP32(tmphalf[i]);
61        tmp[i + 12] = GGML_CPU_FP16_TO_FP32(tmphalf[i]);
62    }
63
64    return _mm512_loadu_ps(tmp);
65}
66#endif
67static inline __m256 __avx_f32cx8_load(ggml_fp16_t *x) {
68    float tmp[8];
69
70    for (int i = 0; i < 8; i++) {
71        tmp[i] = GGML_CPU_FP16_TO_FP32(x[i]);
72    }
73
74    return _mm256_loadu_ps(tmp);
75}
76static inline __m256 __avx_repeat_f32cx8_load(ggml_fp16_t *x) {
77    float tmp[8];
78
79    for (int i = 0; i < 4; i++) {
80        tmp[i] = GGML_CPU_FP16_TO_FP32(x[i]);
81        tmp[i + 4] = GGML_CPU_FP16_TO_FP32(x[i]);
82    }
83
84    return _mm256_loadu_ps(tmp);
85}
86static inline __m256 __avx_rearranged_f32cx8_load(ggml_fp16_t *x, __m128i arrangeMask) {
87    uint16_t tmphalf[8];
88    float tmp[8];
89
90    _mm_storeu_si128((__m128i*)tmphalf, _mm_shuffle_epi8(_mm_loadu_si128((const __m128i *) x), arrangeMask));
91    for (int i = 0; i < 8; i++) {
92        tmp[i] = GGML_CPU_FP16_TO_FP32(tmphalf[i]);
93    }
94
95    return _mm256_loadu_ps(tmp);
96}
97
98#define GGML_F32Cx8_LOAD(x)     __avx_f32cx8_load(x)
99#define GGML_F32Cx8_REPEAT_LOAD(x, loadMask)     __avx_repeat_f32cx8_load(x)
100#define GGML_F32Cx8_REARRANGE_LOAD(x, arrangeMask)     __avx_rearranged_f32cx8_load(x, arrangeMask)
101#if defined(__AVX512F__)
102#define GGML_F32Cx8x2_LOAD(x, y)     __avx512_f32cx8x2_load(x, y)
103#define GGML_F32Cx16_REPEAT_LOAD(x)  __avx512_repeat_f32cx16_load(x)
104#endif
105#endif
106#endif
107
108static inline int nearest_int(float fval) {
109    assert(fabsf(fval) <= 4194303.f);
110    float val = fval + 12582912.f;
111    int i; memcpy(&i, &val, sizeof(int));
112    return (i & 0x007fffff) - 0x00400000;
113}
114
115#if defined(__AVX2__) || defined(__AVX512F__)
116#if defined(__AVX512F__)
117// add int16_t pairwise and return as 512 bit int vector, then add the accumulator
118static inline __m512i sum_i16_pairs_acc_int32x16(const __m512i acc, const __m512i x) {
119    const __m512i ones = _mm512_set1_epi16(1);
120    return _mm512_add_epi32(acc, _mm512_madd_epi16(ones, x));
121}
122
123static inline __m512i mul_sum_us8_pairs_acc_int32x16(const __m512i acc, const __m512i ax, const __m512i sy) {
124#if defined(__AVX512VNNI__)
125    return _mm512_dpbusd_epi32(acc, ax, sy);
126#else
127    // Perform multiplication and create 16-bit values
128    const __m512i dot = _mm512_maddubs_epi16(ax, sy);
129    return sum_i16_pairs_acc_int32x16(acc, dot);
130#endif
131}
132
133// multiply int8_t, add results pairwise twice and return as 512 bit int vector，then add the accumulator
134static inline __m512i mul_sum_i8_pairs_acc_int32x16(const __m512i acc, const __m512i x, const __m512i y) {
135    const __m512i zero = _mm512_setzero_si512();
136    // Get absolute values of x vectors
137    const __m512i ax = _mm512_abs_epi8(x);
138    // Sign the values of the y vectors
139    __mmask64 blt0 = _mm512_movepi8_mask(x);
140    const __m512i sy = _mm512_mask_sub_epi8(y, blt0, zero, y);
141    return mul_sum_us8_pairs_acc_int32x16(acc, ax, sy);
142}
143#endif
144
145// add int16_t pairwise and return as 256 bit int vector, then add the accumulator
146static inline __m256i sum_i16_pairs_acc_int32x8(const __m256i acc, const __m256i x) {
147    const __m256i ones = _mm256_set1_epi16(1);
148    return _mm256_add_epi32(acc, _mm256_madd_epi16(ones, x));
149}
150
151static inline __m256i mul_sum_us8_pairs_acc_int32x8(const __m256i acc, const __m256i ax, const __m256i sy) {
152#if defined(__AVX512VNNI__) && defined(__AVX512VL__)
153    return _mm256_dpbusd_epi32(acc, ax, sy);
154#elif defined(__AVXVNNI__)
155    return _mm256_dpbusd_avx_epi32(acc, ax, sy);
156#else
157    // Perform multiplication and create 16-bit values
158    const __m256i dot = _mm256_maddubs_epi16(ax, sy);
159    return sum_i16_pairs_acc_int32x8(acc, dot);
160#endif
161}
162
163// Integer variant of the function defined in ggml-quants.c
164// multiply int8_t, add results pairwise twice and return as 256 bit int vector, then add the accumulator
165static inline __m256i mul_sum_i8_pairs_acc_int32x8(const __m256i acc, const __m256i x, const __m256i y) {
166#if defined(__AVXVNNIINT8__)
167    return _mm256_dpbssd_epi32(acc, x, y);
168#else
169    // Get absolute values of x vectors
170    const __m256i ax = _mm256_sign_epi8(x, x);
171    // Sign the values of the y vectors
172    const __m256i sy = _mm256_sign_epi8(y, x);
173    return mul_sum_us8_pairs_acc_int32x8(acc, ax, sy);
174#endif
175}
176#endif
177
178void ggml_quantize_mat_q8_0_4x8(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) {
179    assert(QK8_0 == 32);
180    assert(k % QK8_0 == 0);
181    const int nb = k / QK8_0;
182
183    block_q8_0x4 * GGML_RESTRICT y = (block_q8_0x4 *) vy;
184
185#if defined(__AVX2__) || defined(__AVX__)
186    float id[4];
187    __m256 srcv[4][4];
188    __m256 idvec[4];
189
190    for (int i = 0; i < nb; i++) {
191        for (int row_iter = 0; row_iter < 4; row_iter++) {
192            // Load elements into 4 AVX vectors
193            __m256 v0 = _mm256_loadu_ps( x + row_iter * k + i * 32 );
194            __m256 v1 = _mm256_loadu_ps( x + row_iter * k + i * 32 + 8 );
195            __m256 v2 = _mm256_loadu_ps( x + row_iter * k + i * 32 + 16 );
196            __m256 v3 = _mm256_loadu_ps( x + row_iter * k + i * 32 + 24 );
197
198            // Compute max(abs(e)) for the block
199            const __m256 signBit = _mm256_set1_ps( -0.0f );
200            __m256 maxAbs = _mm256_andnot_ps( signBit, v0 );
201            maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v1 ) );
202            maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v2 ) );
203            maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v3 ) );
204
205            __m128 max4 = _mm_max_ps( _mm256_extractf128_ps( maxAbs, 1 ), _mm256_castps256_ps128( maxAbs ) );
206            max4 = _mm_max_ps( max4, _mm_movehl_ps( max4, max4 ) );
207            max4 = _mm_max_ss( max4, _mm_movehdup_ps( max4 ) );
208            const float maxScalar = _mm_cvtss_f32( max4 );
209
210            // Divided by 127.f to mirror results in quantize_row_q8_0
211            const float d = maxScalar  / 127.f;
212            id[row_iter] = ( maxScalar != 0.0f ) ? 127.f / maxScalar : 0.0f; //d ? 1.0f / d : 0.0f;
213
214            // Store the scale for the individual block
215            y[i].d[row_iter] = GGML_CPU_FP32_TO_FP16(d);
216
217            // Store the values in blocks of eight values - Aim is to use these later for block interleaving
218            srcv[row_iter][0] = v0;
219            srcv[row_iter][1] = v1;
220            srcv[row_iter][2] = v2;
221            srcv[row_iter][3] = v3;
222            idvec[row_iter] = _mm256_set1_ps(id[row_iter]);
223        }
224
225        // The loop iterates four times - The aim is to get 4 corresponding chunks of eight bytes from the original weight blocks that are interleaved
226        for (int j = 0; j < 4; j++) {
227            // Apply the multiplier
228            __m256 v0 = _mm256_mul_ps(srcv[0][j], idvec[0]);
229            __m256 v1 = _mm256_mul_ps(srcv[1][j], idvec[1]);
230            __m256 v2 = _mm256_mul_ps(srcv[2][j], idvec[2]);
231            __m256 v3 = _mm256_mul_ps(srcv[3][j], idvec[3]);
232
233            // Round to nearest integer
234            v0 = _mm256_round_ps( v0, _MM_ROUND_NEAREST );
235            v1 = _mm256_round_ps( v1, _MM_ROUND_NEAREST );
236            v2 = _mm256_round_ps( v2, _MM_ROUND_NEAREST );
237            v3 = _mm256_round_ps( v3, _MM_ROUND_NEAREST );
238
239            // Convert floats to integers
240            __m256i i0 = _mm256_cvtps_epi32( v0 );
241            __m256i i1 = _mm256_cvtps_epi32( v1 );
242            __m256i i2 = _mm256_cvtps_epi32( v2 );
243            __m256i i3 = _mm256_cvtps_epi32( v3 );
244
245#if defined(__AVX2__)
246            // Convert int32 to int16
247            i0 = _mm256_packs_epi32( i0, i1 );
248            i2 = _mm256_packs_epi32( i2, i3 );
249            // Convert int16 to int8
250            i0 = _mm256_packs_epi16( i0, i2 );
251
252            //  Permute and store the quantized weights in the required order after the pack instruction
253            const __m256i perm = _mm256_setr_epi32( 0, 4, 1, 5, 2, 6, 3, 7 );
254            i0 = _mm256_permutevar8x32_epi32( i0, perm );
255
256            _mm256_storeu_si256((__m256i *)(y[i].qs + 32 * j), i0);
257#else
258            // Since we don't have in AVX some necessary functions,
259            // we split the registers in half and call AVX2 analogs from SSE
260            __m128i ni0 = _mm256_castsi256_si128( i0 );
261            __m128i ni1 = _mm256_extractf128_si256( i0, 1);
262            __m128i ni2 = _mm256_castsi256_si128( i1 );
263            __m128i ni3 = _mm256_extractf128_si256( i1, 1);
264            __m128i ni4 = _mm256_castsi256_si128( i2 );
265            __m128i ni5 = _mm256_extractf128_si256( i2, 1);
266            __m128i ni6 = _mm256_castsi256_si128( i3 );
267            __m128i ni7 = _mm256_extractf128_si256( i3, 1);
268
269            // Convert int32 to int16
270            ni0 = _mm_packs_epi32( ni0, ni1 );
271            ni2 = _mm_packs_epi32( ni2, ni3 );
272            ni4 = _mm_packs_epi32( ni4, ni5 );
273            ni6 = _mm_packs_epi32( ni6, ni7 );
274            // Convert int16 to int8
275            ni0 = _mm_packs_epi16( ni0, ni2 );
276            ni4 = _mm_packs_epi16( ni4, ni6 );
277            _mm_storeu_si128((__m128i *)(y[i].qs + 32 * j), ni0);
278            _mm_storeu_si128((__m128i *)(y[i].qs + 32 * j + 16), ni4);
279#endif
280        }
281    }
282
283#else
284    UNUSED(nb);
285    UNUSED(y);
286    ggml_quantize_mat_q8_0_4x8_generic(x, vy, k);
287#endif
288}
289
290void ggml_quantize_mat_q8_K_4x8(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) {
291    assert(QK_K == 256);
292    assert(k % QK_K == 0);
293    const int nb = k / QK_K;
294
295    block_q8_Kx4 * GGML_RESTRICT y = (block_q8_Kx4 *) vy;
296
297#if defined(__AVX2__)
298    float iscale[4];
299    __m256 srcv[4][32];
300    __m256 iscale_vec[4];
301
302    for (int i = 0; i < nb; i++) {
303        for (int row_iter = 0; row_iter < 4; row_iter++) {
304            // Load elements into 4 AVX vectors
305            __m256 v0 = _mm256_loadu_ps( x + row_iter * k + i * 256 );
306            __m256 v1 = _mm256_loadu_ps( x + row_iter * k + i * 256 + 8 );
307            __m256 v2 = _mm256_loadu_ps( x + row_iter * k + i * 256 + 16 );
308            __m256 v3 = _mm256_loadu_ps( x + row_iter * k + i * 256 + 24 );
309
310            // Compute max(abs(e)) for the block
311            const __m256 signBit = _mm256_set1_ps( -0.0f );
312            __m256 abs0 = _mm256_andnot_ps( signBit, v0 );
313            __m256 abs1 = _mm256_andnot_ps( signBit, v1 );
314            __m256 abs2 = _mm256_andnot_ps( signBit, v2 );
315            __m256 abs3 = _mm256_andnot_ps( signBit, v3 );
316
317            __m256 maxAbs = _mm256_max_ps( abs0, abs1 );
318            maxAbs = _mm256_max_ps( maxAbs, abs2 );
319            maxAbs = _mm256_max_ps( maxAbs, abs3 );
320
321            __m256 mask0 = _mm256_cmp_ps( maxAbs, v0, _CMP_EQ_OQ );
322            __m256 mask1 = _mm256_cmp_ps( maxAbs, v1, _CMP_EQ_OQ );
323            __m256 mask2 = _mm256_cmp_ps( maxAbs, v2, _CMP_EQ_OQ );
324            __m256 mask3 = _mm256_cmp_ps( maxAbs, v3, _CMP_EQ_OQ );
325
326            __m256 maskAbs = _mm256_or_ps(_mm256_or_ps(mask0, mask1),_mm256_or_ps(mask2, mask3));
327
328            srcv[row_iter][0] = v0;
329            srcv[row_iter][1] = v1;
330            srcv[row_iter][2] = v2;
331            srcv[row_iter][3] = v3;
332
333            for (int sb = 1; sb < 8; sb++) {
334                // Temporarily stores absolute quant values
335                __m256 tempAbs = maxAbs;
336
337                // Load elements into 4 AVX vectors
338                __m256 v0 = _mm256_loadu_ps( x + row_iter * k + i * 256 + sb * 32);
339                __m256 v1 = _mm256_loadu_ps( x + row_iter * k + i * 256 + sb * 32 + 8 );
340                __m256 v2 = _mm256_loadu_ps( x + row_iter * k + i * 256 + sb * 32 + 16 );
341                __m256 v3 = _mm256_loadu_ps( x + row_iter * k + i * 256 + sb * 32 + 24 );
342
343                // Compute max(abs(e)) for the block
344                __m256 abs0 = _mm256_andnot_ps( signBit, v0 );
345                __m256 abs1 = _mm256_andnot_ps( signBit, v1 );
346                __m256 abs2 = _mm256_andnot_ps( signBit, v2 );
347                __m256 abs3 = _mm256_andnot_ps( signBit, v3 );
348
349                maxAbs = _mm256_max_ps( maxAbs, abs0 );
350                maxAbs = _mm256_max_ps( maxAbs, abs1 );
351                maxAbs = _mm256_max_ps( maxAbs, abs2 );
352                maxAbs = _mm256_max_ps( maxAbs, abs3 );
353
354                __m256 mask_prev = _mm256_cmp_ps( tempAbs, maxAbs, _CMP_EQ_OQ );
355                maskAbs = _mm256_and_ps( maskAbs, mask_prev );
356
357                mask0 = _mm256_cmp_ps( maxAbs, v0, _CMP_EQ_OQ );
358                mask1 = _mm256_cmp_ps( maxAbs, v1, _CMP_EQ_OQ );
359                mask2 = _mm256_cmp_ps( maxAbs, v2, _CMP_EQ_OQ );
360                mask3 = _mm256_cmp_ps( maxAbs, v3, _CMP_EQ_OQ );
361
362                __m256 mask_curr = _mm256_or_ps(_mm256_or_ps(mask0, mask1),_mm256_or_ps(mask2, mask3));
363                maskAbs =  _mm256_or_ps(maskAbs, mask_curr);
364
365                srcv[row_iter][sb * 4] = v0;
366                srcv[row_iter][sb * 4 + 1] = v1;
367                srcv[row_iter][sb * 4 + 2] = v2;
368                srcv[row_iter][sb * 4 + 3] = v3;
369            }
370
371            __m128 max4 = _mm_max_ps( _mm256_extractf128_ps( maxAbs, 1 ), _mm256_castps256_ps128( maxAbs ) );
372            max4 = _mm_max_ps( max4, _mm_movehl_ps( max4, max4 ) );
373            max4 = _mm_max_ss( max4, _mm_movehdup_ps( max4 ) );
374            const float maxScalar = _mm_cvtss_f32( max4 );
375
376            __m256 maxScalarVec = _mm256_set1_ps(maxScalar);
377
378            __m256 mask_next = _mm256_cmp_ps( maxScalarVec, maxAbs, _CMP_EQ_OQ );
379            __m256 finalMask = _mm256_and_ps(maskAbs, mask_next);
380
381            const int mask = _mm256_movemask_ps(finalMask);
382            iscale[row_iter] = ( maxScalar != 0.0f ) ? 127.f / maxScalar : 0.0f;
383
384            if(mask) {
385                iscale[row_iter] = ( maxScalar != 0.0f ) ? -127.f / maxScalar: 0.0f;
386            }
387
388            y[i].d[row_iter] = maxScalar ? 1/iscale[row_iter] : 0;
389            iscale_vec[row_iter] = _mm256_set1_ps(iscale[row_iter]);
390        }
391
392        __m256i quants_interleaved[32];
393        for (int j = 0; j < 32; j++) {
394            // Apply the multiplier
395            __m256 v0 = _mm256_mul_ps(srcv[0][j], iscale_vec[0]);
396            __m256 v1 = _mm256_mul_ps(srcv[1][j], iscale_vec[1]);
397            __m256 v2 = _mm256_mul_ps(srcv[2][j], iscale_vec[2]);
398            __m256 v3 = _mm256_mul_ps(srcv[3][j], iscale_vec[3]);
399
400            // Round to nearest integer
401            v0 = _mm256_round_ps( v0, _MM_ROUND_NEAREST );
402            v1 = _mm256_round_ps( v1, _MM_ROUND_NEAREST );
403            v2 = _mm256_round_ps( v2, _MM_ROUND_NEAREST );
404            v3 = _mm256_round_ps( v3, _MM_ROUND_NEAREST );
405
406            // Convert floats to integers
407            __m256i i0 = _mm256_cvtps_epi32( v0 );
408            __m256i i1 = _mm256_cvtps_epi32( v1 );
409            __m256i i2 = _mm256_cvtps_epi32( v2 );
410            __m256i i3 = _mm256_cvtps_epi32( v3 );
411
412            // Convert int32 to int16
413            i0 = _mm256_packs_epi32( i0, i1 );
414            i2 = _mm256_packs_epi32( i2, i3 );
415            // Convert int16 to int8
416            i0 = _mm256_packs_epi16( i0, i2 );
417
418            //  Permute and store the quantized weights in the required order after the pack instruction
419            const __m256i perm = _mm256_setr_epi32( 0, 4, 1, 5, 2, 6, 3, 7 );
420            i0 = _mm256_permutevar8x32_epi32( i0, perm );
421
422            _mm256_storeu_si256((__m256i *)(y[i].qs + 32 * j), i0);
423            quants_interleaved[j] = i0;
424        }
425
426        // Masks to shuffle the quants of corresponding sub blocks for rearranging quants for vectorized bsums computation
427        __m256i shuffle_mask_sb2 = _mm256_castsi128_si256(_mm_setr_epi8(0, 1, 0, 1, 4, 5, 6, 7, 8, 9, 8, 9, 12, 13, 14, 15));
428        shuffle_mask_sb2 = _mm256_permute2f128_si256(shuffle_mask_sb2, shuffle_mask_sb2, 0);
429        __m256i shuffle_mask_sb3 = _mm256_castsi128_si256(_mm_setr_epi8(0, 1, 2, 3, 0, 1, 6, 7, 8, 9, 10, 11, 8, 9, 14, 15));
430        shuffle_mask_sb3 = _mm256_permute2f128_si256(shuffle_mask_sb3, shuffle_mask_sb3, 0);
431        __m256i shuffle_mask_sb4 = _mm256_castsi128_si256(_mm_setr_epi8(0, 1, 2, 3, 4, 5, 0, 1, 8, 9, 10, 11, 12, 13, 8, 9));
432        shuffle_mask_sb4 = _mm256_permute2f128_si256(shuffle_mask_sb4, shuffle_mask_sb4, 0);
433
434        for (int k = 0; k < 4; k++) {
435            // Quants from four different sub blocks are taken
436            __m256i q0 = quants_interleaved[k * 8 + 0];
437            __m256i q1 = quants_interleaved[k * 8 + 1];
438            __m256i q2 = quants_interleaved[k * 8 + 2];
439            __m256i q3 = quants_interleaved[k * 8 + 3];
440            __m256i q4 = quants_interleaved[k * 8 + 4];
441            __m256i q5 = quants_interleaved[k * 8 + 5];
442            __m256i q6 = quants_interleaved[k * 8 + 6];
443            __m256i q7 = quants_interleaved[k * 8 + 7];
444
445
446            // The below code block has the first half of different sub blocks shuffled and blended so as to process 2 values from each sub block at a time
447            __m256i sb2_h1_shuffled = _mm256_shuffle_epi8(q2, shuffle_mask_sb2);
448            __m256i sb_h1_interleaved = _mm256_blend_epi16(q0, sb2_h1_shuffled, 34);
449            __m256i sb3_h1_shuffled = _mm256_shuffle_epi8(q4, shuffle_mask_sb3);
450            sb_h1_interleaved = _mm256_blend_epi16(sb_h1_interleaved, sb3_h1_shuffled, 68);
451            __m256i sb4_h1_shuffled = _mm256_shuffle_epi8(q6, shuffle_mask_sb4);
452            sb_h1_interleaved = _mm256_blend_epi16(sb_h1_interleaved, sb4_h1_shuffled, 136);
453
454            __m256i one = _mm256_set1_epi8(1);
455            __m256i bsums_r1 = _mm256_maddubs_epi16(one, sb_h1_interleaved);
456
457            for (int l = 0; l < 3; l++) {
458                // Quants value shifted to process next two values from each sub block
459                q0 = _mm256_srli_epi64(q0, 16);
460                q2 = _mm256_srli_epi64(q2, 16);
461                q4 = _mm256_srli_epi64(q4, 16);
462                q6 = _mm256_srli_epi64(q6, 16);
463
464                sb2_h1_shuffled = _mm256_shuffle_epi8(q2, shuffle_mask_sb2);
465                sb_h1_interleaved = _mm256_blend_epi16(q0, sb2_h1_shuffled, 34);
466                sb3_h1_shuffled = _mm256_shuffle_epi8(q4, shuffle_mask_sb3);
467                sb_h1_interleaved = _mm256_blend_epi16(sb_h1_interleaved, sb3_h1_shuffled, 68);
468                sb4_h1_shuffled = _mm256_shuffle_epi8(q6, shuffle_mask_sb4);
469                sb_h1_interleaved = _mm256_blend_epi16(sb_h1_interleaved, sb4_h1_shuffled, 136);
470
471                bsums_r1 = _mm256_add_epi16(bsums_r1, _mm256_maddubs_epi16(one, sb_h1_interleaved));
472            }
473
474            // The below code block has the second half of different sub blocks shuffled and blended so as to process 2 values from each sub block at a time
475            __m256i sb2_h2_shuffled = _mm256_shuffle_epi8(q3, shuffle_mask_sb2);
476            __m256i sb_h2_interleaved = _mm256_blend_epi16(q1, sb2_h2_shuffled, 34);
477            __m256i sb3_h2_shuffled = _mm256_shuffle_epi8(q5, shuffle_mask_sb3);
478            sb_h2_interleaved = _mm256_blend_epi16(sb_h2_interleaved, sb3_h2_shuffled, 68);
479            __m256i sb4_h2_shuffled = _mm256_shuffle_epi8(q7, shuffle_mask_sb4);
480            sb_h2_interleaved = _mm256_blend_epi16(sb_h2_interleaved, sb4_h2_shuffled, 136);
481
482            __m256i bsums_r2 = _mm256_maddubs_epi16(one, sb_h2_interleaved);
483
484            for (int l = 0; l < 3; l++) {
485                // Quants value shifted to process next two values from each sub block
486                q1 = _mm256_srli_epi64(q1, 16);
487                q3 = _mm256_srli_epi64(q3, 16);
488                q5 = _mm256_srli_epi64(q5, 16);
489                q7 = _mm256_srli_epi64(q7, 16);
490
491                sb2_h2_shuffled = _mm256_shuffle_epi8(q3, shuffle_mask_sb2);
492                sb_h2_interleaved = _mm256_blend_epi16(q1, sb2_h2_shuffled, 34);
493                sb3_h2_shuffled = _mm256_shuffle_epi8(q5, shuffle_mask_sb3);
494                sb_h2_interleaved = _mm256_blend_epi16(sb_h2_interleaved, sb3_h2_shuffled, 68);
495                sb4_h2_shuffled = _mm256_shuffle_epi8(q7, shuffle_mask_sb4);
496                sb_h2_interleaved = _mm256_blend_epi16(sb_h2_interleaved, sb4_h2_shuffled, 136);
497
498                bsums_r2 = _mm256_add_epi16(bsums_r2, _mm256_maddubs_epi16(one, sb_h2_interleaved));
499            }
500
501            // Overall bsums in interleaved fashion computed by adding results of both halves
502            __m256i bsums_r = _mm256_add_epi16(bsums_r1, bsums_r2);
503            _mm256_storeu_si256((__m256i *)(y[i].bsums + 16 * k), bsums_r);
504        }
505    }
506
507#else
508    UNUSED(nb);
509    UNUSED(y);
510    ggml_quantize_mat_q8_K_4x8_generic(x, vy, k);
511#endif
512}
513
514//
515// GEMV/GEMM templates
516//
517
518#if defined(__AVX2__) || defined(__AVX512F__)
519
520// GEMV for 8x blocks of 32 4-bit quants with a single scale factor per block
521template<typename block_tx8>
522static void gemv_q4_b32_8x8_q8_0_lut_avx(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc, __m256i signextendlut) {
523    static_assert(
524            std::is_same_v<block_tx8, block_q4_0x8> ||
525            std::is_same_v<block_tx8, block_iq4_nlx8> ||
526            std::is_same_v<block_tx8, block_mxfp4x8>,
527            "Unsupported block type");
528
529    const int qk = QK8_0;
530    const int nb = n / qk;
531
532    UNUSED(bs);
533
534    __m256i finalpermutemask = _mm256_set_epi32(7, 5, 3, 1, 6, 4, 2, 0);
535
536    // Permute mask used for easier vector processing at later stages
537    const __m256i m4b = _mm256_set1_epi8(0x0F);
538
539    int64_t b_nb = n / 32;
540
541    const block_tx8  * b_ptr_start = (const block_tx8  *)vx;
542    const block_q8_0 * a_ptr_start = (const block_q8_0 *)vy;
543
544    // Process Q8_0 blocks one by one
545    for (int64_t y = 0; y < nr; y++) {
546
547        // Pointers to LHS blocks of block_q8_0 format
548        const block_q8_0 * a_ptr = a_ptr_start + (y * nb);
549
550        // Take group of eight blocks at each pass of the loop and perform dot product operation
551        for (int64_t x = 0; x < nc / 8; x++) {
552
553            // Pointers to RHS blocks
554            const block_tx8 * b_ptr = b_ptr_start + (x * b_nb);
555
556            // Master FP accumulator
557            __m256 acc_row = _mm256_setzero_ps();
558
559            for (int64_t b = 0; b < nb; b++) {
560                // Load 8 blocks of 32 interleaved as 8 bytes (B0 - B7)
561                const __m256i rhs_raw_vec_0123_0 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs));
562                const __m256i rhs_raw_vec_4567_0 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs) + 1);
563                const __m256i rhs_raw_vec_0123_1 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs) + 2);
564                const __m256i rhs_raw_vec_4567_1 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs) + 3);
565
566                // 4-bit -> 8-bit - Sign is maintained
567                const __m256i rhs_vec_0123_0 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_vec_0123_0, m4b)); // B0(0-7) B1(0-7) B2(0-7) B3(0-7)
568                const __m256i rhs_vec_4567_0 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_vec_4567_0, m4b)); // B4(0-7) B5(0-7) B6(0-7) B7(0-7)
569                const __m256i rhs_vec_0123_1 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_vec_0123_1, m4b)); // B0(8-15) B1(8-15) B2(8-15) B3(8-15)
570                const __m256i rhs_vec_4567_1 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_vec_4567_1, m4b)); // B0(8-15) B1(8-15) B2(8-15) B3(8-15)
571
572                const __m256i rhs_vec_0123_2 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_0, 4), m4b)); // B0(16-23) B1(16-23) B2(16-23) B3(16-23)
573                const __m256i rhs_vec_4567_2 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_0, 4), m4b)); // B4(16-23) B5(16-23) B6(16-23) B7(16-23)
574                const __m256i rhs_vec_0123_3 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_1, 4), m4b)); // B0(24-31) B1(24-31) B2(24-31) B3(24-31)
575                const __m256i rhs_vec_4567_3 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_1, 4), m4b)); // B4(24-31) B5(24-31) B6(24-31) B7(24-31)
576
577                // Load the scale values for the 8 blocks interleaved in block_tx8
578                __m256 col_scale_f32;
579                if constexpr (
580                        std::is_same_v<block_tx8, block_q4_0x8> ||
581                        std::is_same_v<block_tx8, block_iq4_nlx8>) {
582                    const __m128i changemask = _mm_set_epi8(15, 14, 7, 6, 13, 12, 5, 4, 11, 10, 3, 2, 9, 8, 1, 0);
583                    col_scale_f32 = GGML_F32Cx8_REARRANGE_LOAD(b_ptr[b].d, changemask);
584                } else if constexpr (std::is_same_v<block_tx8, block_mxfp4x8>) {
585                    // Load 8 E8M0 exponents and convert to float via LUT
586                    // Rearranged to match changemask order: 0,4,1,5,2,6,3,7
587                    col_scale_f32 = _mm256_set_ps(
588                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[b].e[7]),
589                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[b].e[3]),
590                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[b].e[6]),
591                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[b].e[2]),
592                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[b].e[5]),
593                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[b].e[1]),
594                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[b].e[4]),
595                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[b].e[0]));
596                }
597
598                // Load and convert to FP32 scale from block_q8_0
599                const __m256 row_scale_f32 = _mm256_set1_ps(GGML_CPU_FP16_TO_FP32(a_ptr[b].d));
600
601                // Load the block values in block_q8_0 in batches of 16 bytes and replicate the same across 256 bit vector
602                __m256i lhs_vec_0 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)a_ptr[b].qs));
603                __m256i lhs_vec_1 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 16)));
604
605                lhs_vec_0 = _mm256_permute2f128_si256(lhs_vec_0, lhs_vec_0, 0); // A0 (0-15) A0(0-15)
606                lhs_vec_1 = _mm256_permute2f128_si256(lhs_vec_1, lhs_vec_1, 0); // A0 (16-31) A0(16-31))
607
608                __m256i iacc = _mm256_setzero_si256();
609
610                // Dot product done within 32 bit lanes and accumulated in the same vector
611                // 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)
612                // 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)
613                // ...........................................................................
614                // B0(28-31) B4(28-31) B1(28-31) B5(28-31) B2(28-31) B6(28-31) B3(28-31) B7(28-31) with A0(28-31)
615
616                iacc = mul_sum_i8_pairs_acc_int32x8(iacc, _mm256_blend_epi32(rhs_vec_0123_0 ,_mm256_shuffle_epi32(rhs_vec_4567_0, 177), 170), _mm256_shuffle_epi32(lhs_vec_0, 0));
617                iacc = mul_sum_i8_pairs_acc_int32x8(iacc, _mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_0, 177) ,rhs_vec_4567_0, 170), _mm256_shuffle_epi32(lhs_vec_0, 85));
618
619                iacc = mul_sum_i8_pairs_acc_int32x8(iacc, _mm256_blend_epi32(rhs_vec_0123_1 ,_mm256_shuffle_epi32(rhs_vec_4567_1, 177), 170), _mm256_shuffle_epi32(lhs_vec_0, 170));
620                iacc = mul_sum_i8_pairs_acc_int32x8(iacc, _mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_1, 177) ,rhs_vec_4567_1, 170), _mm256_shuffle_epi32(lhs_vec_0, 255));
621
622                iacc = mul_sum_i8_pairs_acc_int32x8(iacc, _mm256_blend_epi32(rhs_vec_0123_2 ,_mm256_shuffle_epi32(rhs_vec_4567_2, 177), 170), _mm256_shuffle_epi32(lhs_vec_1, 0));
623                iacc = mul_sum_i8_pairs_acc_int32x8(iacc, _mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_2, 177) ,rhs_vec_4567_2, 170), _mm256_shuffle_epi32(lhs_vec_1, 85));
624
625                iacc = mul_sum_i8_pairs_acc_int32x8(iacc, _mm256_blend_epi32(rhs_vec_0123_3 ,_mm256_shuffle_epi32(rhs_vec_4567_3, 177), 170), _mm256_shuffle_epi32(lhs_vec_1, 170));
626                iacc = mul_sum_i8_pairs_acc_int32x8(iacc, _mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_3, 177) ,rhs_vec_4567_3, 170), _mm256_shuffle_epi32(lhs_vec_1, 255));
627
628                // Accumulated values multiplied with appropriate scales
629                acc_row = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc), _mm256_mul_ps(col_scale_f32, row_scale_f32), acc_row);
630            }
631
632            // Accumulated output values permuted so as to be stored in appropriate order post accumulation
633            acc_row = _mm256_permutevar8x32_ps(acc_row, finalpermutemask);
634            _mm256_storeu_ps(s + (y * nr + x * 8), acc_row);
635        }
636    }
637}
638
639// GEMM for 8x blocks of 32 4-bit quants with a single scale factor per block
640template<typename block_tx8>
641static void gemm_q4_b32_8x8_q8_0_lut_avx(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc, __m256i signextendlut) {
642    static_assert(
643            std::is_same_v<block_tx8, block_q4_0x8> ||
644            std::is_same_v<block_tx8, block_iq4_nlx8> ||
645            std::is_same_v<block_tx8, block_mxfp4x8>,
646            "Unsupported block type");
647
648    const int qk = QK8_0;
649    const int nb = n / qk;
650
651    const block_tx8    * b_ptr_start = (const block_tx8    *)vx;
652    const block_q8_0x4 * a_ptr_start = (const block_q8_0x4 *)vy;
653
654    int64_t b_nb = n / 32;
655    int64_t y = 0;
656    // Mask to mask out nibbles from packed bytes
657    const __m256i m4b = _mm256_set1_epi8(0x0F);
658    const __m128i loadMask = _mm_blend_epi32(_mm_setzero_si128(), _mm_set1_epi32(0xFFFFFFFF), 3);
659    // Permute mask used for easier vector processing at later stages
660    __m256i requiredOrder = _mm256_set_epi32(3, 2, 1, 0, 7, 6, 5, 4);
661    int64_t xstart = 0;
662    int anr = nr - nr%16; // Used to align nr with boundary of 16
663#if defined(__AVX512BW__) && defined(__AVX512DQ__)
664    int anc = nc - nc%16; // Used to align nc with boundary of 16
665                          // Mask to mask out nibbles from packed bytes expanded to 512 bit length
666    const __m512i m4bexpanded = _mm512_set1_epi8(0x0F);
667    // Lookup table to convert signed nibbles to signed bytes expanded to 512 bit length
668    __m512i signextendlutexpanded = _mm512_inserti32x8(_mm512_castsi256_si512(signextendlut), signextendlut, 1);
669
670    // Take group of four block_q8_0x4 structures at each pass of the loop and perform dot product operation
671    for (; y < anr / 4; y += 4) {
672
673        const block_q8_0x4 * a_ptrs[4];
674
675        a_ptrs[0] = a_ptr_start + (y * nb);
676        for (int i = 0; i < 3; ++i) {
677            a_ptrs[i + 1] = a_ptrs[i] + nb;
678        }
679
680        // Take group of two block_tx8 structures at each pass of the loop and perform dot product operation
681        for (int64_t x = 0; x < anc / 8; x += 2) {
682
683            const block_tx8 * b_ptr_0 = b_ptr_start + ((x)     * b_nb);
684            const block_tx8 * b_ptr_1 = b_ptr_start + ((x + 1) * b_nb);
685
686            // Master FP accumulators
687            __m512 acc_rows[16];
688            for (int i = 0; i < 16; i++) {
689                acc_rows[i] = _mm512_setzero_ps();
690            }
691
692            for (int64_t b = 0; b < nb; b++) {
693                // Load the sixteen blocks of quantized values interleaved with each other in chunks of eight - B0,B1 ....BE,BF
694                const __m256i rhs_raw_mat_0123_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs));
695                const __m256i rhs_raw_mat_4567_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs + 32));
696                const __m256i rhs_raw_mat_0123_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs + 64));
697                const __m256i rhs_raw_mat_4567_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs + 96));
698
699                const __m256i rhs_raw_mat_89AB_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs));
700                const __m256i rhs_raw_mat_CDEF_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs + 32));
701                const __m256i rhs_raw_mat_89AB_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs + 64));
702                const __m256i rhs_raw_mat_CDEF_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs + 96));
703
704                // Save the values in the following vectors in the formats B0B1B4B5B8B9BCBD, B2B3B6B7BABBBEBF for further processing and storing of values
705                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);
706                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);
707                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);
708                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);
709
710                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);
711                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);
712                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);
713                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);
714
715                const __m512i rhs_raw_mat_014589CD_0 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_0), rhs_raw_mat_89CD_0, 1);
716                const __m512i rhs_raw_mat_2367ABEF_0 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_0), rhs_raw_mat_ABEF_0, 1);
717                const __m512i rhs_raw_mat_014589CD_1 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_1), rhs_raw_mat_89CD_1, 1);
718                const __m512i rhs_raw_mat_2367ABEF_1 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_1), rhs_raw_mat_ABEF_1, 1);
719
720                // 4-bit -> 8-bit - Sign is maintained
721                const __m512i rhs_mat_014589CD_0 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_014589CD_0, m4bexpanded)); //B0(0-7) B1(0-7) B4(0-7) B5(0-7) B8(0-7) B9(0-7) BC(0-7) BD(0-7)
722                const __m512i rhs_mat_2367ABEF_0 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_2367ABEF_0, m4bexpanded)); //B2(0-7) B3(0-7) B6(0-7) B7(0-7) BA(0-7) BB(0-7) BE(0-7) BF(0-7)
723
724                const __m512i rhs_mat_014589CD_1 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_014589CD_1, m4bexpanded)); //B0(8-15) B1(8-15) B4(8-15) B5(8-15) B8(8-15) B9(8-15) BC(8-15) BD(8-15)
725                const __m512i rhs_mat_2367ABEF_1 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_2367ABEF_1, m4bexpanded)); //B2(8-15) B3(8-15) B6(8-15) B7(8-15) BA(8-15) BB(8-15) BE(8-15) BF(8-15)
726
727                const __m512i rhs_mat_014589CD_2 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_0, 4), m4bexpanded)); //B0(16-23) B1(16-23) B4(16-23) B5(16-23) B8(16-23) B9(16-23) BC(16-23) BD(16-23)
728                const __m512i rhs_mat_2367ABEF_2 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_0, 4), m4bexpanded)); //B2(16-23) B3(16-23) B6(16-23) B7(16-23) BA(16-23) BB(16-23) BE(16-23) BF(16-23)
729
730                const __m512i rhs_mat_014589CD_3 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_1, 4), m4bexpanded)); //B0(24-31) B1(24-31) B4(24-31) B5(24-31) B8(24-31) B9(24-31) BC(24-31) BD(24-31)
731                const __m512i rhs_mat_2367ABEF_3 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_1, 4), m4bexpanded)); //B2(24-31) B3(24-31) B6(24-31) B7(24-31) BA(24-31) BB(24-31) BE(24-31) BF(24-31)
732
733                // Shuffle pattern one - right side input
734                const __m512i rhs_mat_014589CD_0_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_0, (_MM_PERM_ENUM)136); //B0(0-3) B1(0-3) B0(0-3) B1(0-3) B4(0-3) B5(0-3) B4(0-3) B5(0-3) B8(0-3) B9(0-3) B8(0-3) B9(0-3) BC(0-3) BD(0-3) BC(0-3) BD(0-3)
735                const __m512i rhs_mat_2367ABEF_0_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_0, (_MM_PERM_ENUM)136); //B2(0-3) B3(0-3) B2(0-3) B3(0-3) B6(0-3) B7(0-3) B6(0-3) B7(0-3) BA(0-3) BB(0-3) BA(0-3) BB(0-3) BE(0-3) BF(0-3) BE(0-3) BF(0-3)
736
737                const __m512i rhs_mat_014589CD_1_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_1, (_MM_PERM_ENUM)136); //B0(8-11) B1(8-11) B0(8-11) B1(8-11) B4(8-11) B5(8-11) B4(8-11) B5(8-11) B8(8-11) B9(8-11) B8(8-11) B9(8-11) BC(8-11) BD(8-11) BC(8-11) BD(8-11)
738                const __m512i rhs_mat_2367ABEF_1_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_1, (_MM_PERM_ENUM)136); //B2(8-11) B3(8-11) B2(8-11) B3(8-11) B6(8-11) B7(8-11) B6(8-11) B7(8-11) BA(8-11) BB(8-11) BA(8-11) BB(8-11) BE(8-11) BF(8-11) BE(8-11) BF(8-11)
739
740                const __m512i rhs_mat_014589CD_2_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_2, (_MM_PERM_ENUM)136); //B0(16-19) B1(16-19) B0(16-19) B1(16-19) B4(16-19) B5(16-19) B4(16-19) B5(16-19) B8(16-19) B9(16-19) B8(16-19) B9(16-19) BC(16-19) BD(16-19) BC(16-19) BD(16-19)
741                const __m512i rhs_mat_2367ABEF_2_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_2, (_MM_PERM_ENUM)136); //B2(16-19) B3(16-19) B2(16-19) B3(16-19) B6(16-19) B7(16-19) B6(16-19) B7(16-19) BA(16-19) BB(16-19) BA(16-19) BB(16-19) BE(16-19) BF(16-19) BE(16-19) BF(16-19)
742
743                const __m512i rhs_mat_014589CD_3_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_3, (_MM_PERM_ENUM)136); //B0(24-27) B1(24-27) B0(24-27) B1(24-27) B4(24-27) B5(24-27) B4(24-27) B5(24-27) B8(24-27) B9(24-27) B8(24-27) B9(24-27) BC(24-27) BD(24-27) BC(24-27) BD(24-27)
744                const __m512i rhs_mat_2367ABEF_3_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_3, (_MM_PERM_ENUM)136); //B2(24-27) B3(24-27) B2(24-27) B3(24-27) B6(24-27) B7(24-27) B6(24-27) B7(24-27) BA(24-27) BB(24-27) BA(24-27) BB(24-27) BE(24-27) BF(24-27) BE(24-27) BF(24-27)
745
746                // Shuffle pattern two - right side input
747
748                const __m512i rhs_mat_014589CD_0_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_0, (_MM_PERM_ENUM)221); //B0(4-7) B1(4-7) B0(4-7) B1(4-7) B4(4-7) B5(4-7) B4(4-7) B5(4-7) B8(4-7) B9(4-7) B8(4-7) B9(4-7) BC(4-7) BD(4-7) BC(4-7) BD(4-7)
749                const __m512i rhs_mat_2367ABEF_0_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_0, (_MM_PERM_ENUM)221); //B2(4-7) B3(4-7) B2(4-7) B3(4-7) B6(4-7) B7(4-7) B6(4-7) B7(4-7) BA(4-7) BB(4-7) BA(4-7) BB(4-7) BE(4-7) BF(4-7) BE(4-7) BF(4-7)
750
751                const __m512i rhs_mat_014589CD_1_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_1, (_MM_PERM_ENUM)221); //B0(12-15) B1(12-15) B0(12-15) B1(12-15) B4(12-15) B5(12-15) B4(12-15) B5(12-15) B8(12-15) B9(12-15) B8(12-15) B9(12-15) BC(12-15) BD(12-15) BC(12-15) BD(12-15)
752                const __m512i rhs_mat_2367ABEF_1_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_1, (_MM_PERM_ENUM)221); //B2(12-15) B3(12-15) B2(12-15) B3(12-15) B6(12-15) B7(12-15) B6(12-15) B7(12-15) BA(12-15) BB(12-15) BA(12-15) BB(12-15) BE(12-15) BF(12-15) BE(12-15) BF(12-15)
753
754                const __m512i rhs_mat_014589CD_2_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_2, (_MM_PERM_ENUM)221); //B0(20-23) B1(20-23) B0(20-23) B1(20-23) B4(20-23) B5(20-23) B4(20-23) B5(20-23) B8(20-23) B9(20-23) B8(20-23) B9(20-23) BC(20-23) BD(20-23) BC(20-23) BD(20-23)
755                const __m512i rhs_mat_2367ABEF_2_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_2, (_MM_PERM_ENUM)221); //B2(20-23) B3(20-23) B2(20-23) B3(20-23) B6(20-23) B7(20-23) B6(20-23) B7(20-23) BA(20-23) BB(20-23) BA(20-23) BB(20-23) BE(20-23) BF(20-23) BE(20-23) BF(20-23)
756
757                const __m512i rhs_mat_014589CD_3_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_3, (_MM_PERM_ENUM)221); //B0(28-31) B1(28-31) B0(28-31) B1(28-31) B4(28-31) B5(28-31) B4(28-31) B5(28-31) B8(28-31) B9(28-31) B8(28-31) B9(28-31) BC(28-31) BD(28-31) BC(28-31) BD(28-31)
758                const __m512i rhs_mat_2367ABEF_3_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_3, (_MM_PERM_ENUM)221); //B2(28-31) B3(28-31) B2(28-31) B3(28-31) B6(28-31) B7(28-31) B6(28-31) B7(28-31) BA(28-31) BB(28-31) BA(28-31) BB(28-31) BE(28-31) BF(28-31) BE(28-31) BF(28-31)
759
760                // Scale values - Load the weight scale values of two block_tx8
761                __m512 col_scale_f32;
762                if constexpr (
763                        std::is_same_v<block_tx8, block_q4_0x8> ||
764                        std::is_same_v<block_tx8, block_iq4_nlx8>) {
765                    col_scale_f32 = GGML_F32Cx8x2_LOAD(b_ptr_0[b].d, b_ptr_1[b].d);
766                } else if constexpr (std::is_same_v<block_tx8, block_mxfp4x8>) {
767                    //TODO: simd-ify
768                    col_scale_f32 = _mm512_set_ps(
769                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_1[b].e[7]),
770                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_1[b].e[6]),
771                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_1[b].e[5]),
772                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_1[b].e[4]),
773                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_1[b].e[3]),
774                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_1[b].e[2]),
775                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_1[b].e[1]),
776                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_1[b].e[0]),
777                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_0[b].e[7]),
778                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_0[b].e[6]),
779                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_0[b].e[5]),
780                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_0[b].e[4]),
781                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_0[b].e[3]),
782                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_0[b].e[2]),
783                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_0[b].e[1]),
784                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_0[b].e[0]));
785                }
786
787                // Process LHS in pairs of rows
788                for (int rp = 0; rp < 4; rp++) {
789
790                    // Load the four blocks of quantized values interleaved with each other in chunks of eight - A0,A1,A2,A3
791                    // Loaded as set of 128 bit vectors and repeated and stored into a 256 bit vector before again repeating into 512 bit vector
792                    __m256i lhs_mat_ymm_0123_0 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs)));
793                    __m256i lhs_mat_ymm_01_0 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_0, lhs_mat_ymm_0123_0, 0);
794                    __m256i lhs_mat_ymm_23_0 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_0, lhs_mat_ymm_0123_0, 17);
795                    __m256i lhs_mat_ymm_0123_1 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs + 32)));
796                    __m256i lhs_mat_ymm_01_1 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_1, lhs_mat_ymm_0123_1, 0);
797                    __m256i lhs_mat_ymm_23_1 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_1, lhs_mat_ymm_0123_1, 17);
798                    __m256i lhs_mat_ymm_0123_2 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs + 64)));
799                    __m256i lhs_mat_ymm_01_2 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_2, lhs_mat_ymm_0123_2, 0);
800                    __m256i lhs_mat_ymm_23_2 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_2, lhs_mat_ymm_0123_2, 17);
801                    __m256i lhs_mat_ymm_0123_3 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs + 96)));
802                    __m256i lhs_mat_ymm_01_3 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_3, lhs_mat_ymm_0123_3, 0);
803                    __m256i lhs_mat_ymm_23_3 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_3, lhs_mat_ymm_0123_3, 17);
804
805                    __m512i lhs_mat_01_0 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_0), lhs_mat_ymm_01_0, 1);
806                    __m512i lhs_mat_23_0 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_0), lhs_mat_ymm_23_0, 1);
807                    __m512i lhs_mat_01_1 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_1), lhs_mat_ymm_01_1, 1);
808                    __m512i lhs_mat_23_1 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_1), lhs_mat_ymm_23_1, 1);
809                    __m512i lhs_mat_01_2 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_2), lhs_mat_ymm_01_2, 1);
810                    __m512i lhs_mat_23_2 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_2), lhs_mat_ymm_23_2, 1);
811                    __m512i lhs_mat_01_3 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_3), lhs_mat_ymm_01_3, 1);
812                    __m512i lhs_mat_23_3 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_3), lhs_mat_ymm_23_3, 1);
813
814                    // Shuffle pattern one - left side input
815
816                    const __m512i lhs_mat_01_0_sp1 = _mm512_shuffle_epi32(lhs_mat_01_0, (_MM_PERM_ENUM)160);  //A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3)
817                    const __m512i lhs_mat_23_0_sp1 = _mm512_shuffle_epi32(lhs_mat_23_0, (_MM_PERM_ENUM)160);  //A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3)
818
819                    const __m512i lhs_mat_01_1_sp1 = _mm512_shuffle_epi32(lhs_mat_01_1, (_MM_PERM_ENUM)160);  //A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11)
820                    const __m512i lhs_mat_23_1_sp1 = _mm512_shuffle_epi32(lhs_mat_23_1, (_MM_PERM_ENUM)160);  //A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11)
821
822                    const __m512i lhs_mat_01_2_sp1 = _mm512_shuffle_epi32(lhs_mat_01_2, (_MM_PERM_ENUM)160);  //A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19)
823                    const __m512i lhs_mat_23_2_sp1 = _mm512_shuffle_epi32(lhs_mat_23_2, (_MM_PERM_ENUM)160);  //A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19)
824
825                    const __m512i lhs_mat_01_3_sp1 = _mm512_shuffle_epi32(lhs_mat_01_3, (_MM_PERM_ENUM)160);  //A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27)
826                    const __m512i lhs_mat_23_3_sp1 = _mm512_shuffle_epi32(lhs_mat_23_3, (_MM_PERM_ENUM)160);  //A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27)
827
828                    // Shuffle pattern two - left side input
829
830                    const __m512i lhs_mat_01_0_sp2 = _mm512_shuffle_epi32(lhs_mat_01_0, (_MM_PERM_ENUM)245);  //A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7)
831                    const __m512i lhs_mat_23_0_sp2 = _mm512_shuffle_epi32(lhs_mat_23_0, (_MM_PERM_ENUM)245);  //A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7)
832
833                    const __m512i lhs_mat_01_1_sp2 = _mm512_shuffle_epi32(lhs_mat_01_1, (_MM_PERM_ENUM)245);  //A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15)
834                    const __m512i lhs_mat_23_1_sp2 = _mm512_shuffle_epi32(lhs_mat_23_1, (_MM_PERM_ENUM)245);  //A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15)
835
836                    const __m512i lhs_mat_01_2_sp2 = _mm512_shuffle_epi32(lhs_mat_01_2, (_MM_PERM_ENUM)245);  //A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23)
837                    const __m512i lhs_mat_23_2_sp2 = _mm512_shuffle_epi32(lhs_mat_23_2, (_MM_PERM_ENUM)245);  //A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23)
838
839                    const __m512i lhs_mat_01_3_sp2 = _mm512_shuffle_epi32(lhs_mat_01_3, (_MM_PERM_ENUM)245);  //A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31)
840                    const __m512i lhs_mat_23_3_sp2 = _mm512_shuffle_epi32(lhs_mat_23_3, (_MM_PERM_ENUM)245);  //A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31)
841
842                    // 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
843                    // Resembles MMLAs into 2x2 matrices in ARM Version
844                    const __m512i zero = _mm512_setzero_epi32();
845                    __m512i iacc_mat_00_sp1 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_01_3_sp1, rhs_mat_014589CD_3_sp1), lhs_mat_01_2_sp1, rhs_mat_014589CD_2_sp1), lhs_mat_01_1_sp1, rhs_mat_014589CD_1_sp1), lhs_mat_01_0_sp1, rhs_mat_014589CD_0_sp1);
846                    __m512i iacc_mat_01_sp1 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_01_3_sp1, rhs_mat_2367ABEF_3_sp1), lhs_mat_01_2_sp1, rhs_mat_2367ABEF_2_sp1), lhs_mat_01_1_sp1, rhs_mat_2367ABEF_1_sp1), lhs_mat_01_0_sp1, rhs_mat_2367ABEF_0_sp1);
847                    __m512i iacc_mat_10_sp1 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_23_3_sp1, rhs_mat_014589CD_3_sp1), lhs_mat_23_2_sp1, rhs_mat_014589CD_2_sp1), lhs_mat_23_1_sp1, rhs_mat_014589CD_1_sp1), lhs_mat_23_0_sp1, rhs_mat_014589CD_0_sp1);
848                    __m512i iacc_mat_11_sp1 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_23_3_sp1, rhs_mat_2367ABEF_3_sp1), lhs_mat_23_2_sp1, rhs_mat_2367ABEF_2_sp1), lhs_mat_23_1_sp1, rhs_mat_2367ABEF_1_sp1), lhs_mat_23_0_sp1, rhs_mat_2367ABEF_0_sp1);
849                    __m512i iacc_mat_00_sp2 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_01_3_sp2, rhs_mat_014589CD_3_sp2), lhs_mat_01_2_sp2, rhs_mat_014589CD_2_sp2), lhs_mat_01_1_sp2, rhs_mat_014589CD_1_sp2), lhs_mat_01_0_sp2, rhs_mat_014589CD_0_sp2);
850                    __m512i iacc_mat_01_sp2 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_01_3_sp2, rhs_mat_2367ABEF_3_sp2), lhs_mat_01_2_sp2, rhs_mat_2367ABEF_2_sp2), lhs_mat_01_1_sp2, rhs_mat_2367ABEF_1_sp2), lhs_mat_01_0_sp2, rhs_mat_2367ABEF_0_sp2);
851                    __m512i iacc_mat_10_sp2 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_23_3_sp2, rhs_mat_014589CD_3_sp2), lhs_mat_23_2_sp2, rhs_mat_014589CD_2_sp2), lhs_mat_23_1_sp2, rhs_mat_014589CD_1_sp2), lhs_mat_23_0_sp2, rhs_mat_014589CD_0_sp2);
852                    __m512i iacc_mat_11_sp2 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_23_3_sp2, rhs_mat_2367ABEF_3_sp2), lhs_mat_23_2_sp2, rhs_mat_2367ABEF_2_sp2), lhs_mat_23_1_sp2, rhs_mat_2367ABEF_1_sp2), lhs_mat_23_0_sp2, rhs_mat_2367ABEF_0_sp2);
853
854                    // Output of both shuffle patterns are added in order to sum dot product outputs of all 32 values in block
855                    __m512i iacc_mat_00 = _mm512_add_epi32(iacc_mat_00_sp1, iacc_mat_00_sp2);
856                    __m512i iacc_mat_01 = _mm512_add_epi32(iacc_mat_01_sp1, iacc_mat_01_sp2);
857                    __m512i iacc_mat_10 = _mm512_add_epi32(iacc_mat_10_sp1, iacc_mat_10_sp2);
858                    __m512i iacc_mat_11 = _mm512_add_epi32(iacc_mat_11_sp1, iacc_mat_11_sp2);
859
860
861                    // Straighten out to make 4 row vectors
862                    __m512i iacc_row_0 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_00, _mm512_shuffle_epi32(iacc_mat_01, (_MM_PERM_ENUM)78));
863                    __m512i iacc_row_1 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_00, (_MM_PERM_ENUM)78), iacc_mat_01);
864                    __m512i iacc_row_2 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_10, _mm512_shuffle_epi32(iacc_mat_11, (_MM_PERM_ENUM)78));
865                    __m512i iacc_row_3 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_10, (_MM_PERM_ENUM)78), iacc_mat_11);
866
867                    // Load the scale(d) values for all the 4 Q8_0 blocks and repeat it across lanes
868                    const __m128i row_scale_f16 = _mm_shuffle_epi32(_mm_maskload_epi32((int const*)(a_ptrs[rp][b].d), loadMask), 68);
869                    const __m512 row_scale_f32 = GGML_F32Cx16_REPEAT_LOAD(row_scale_f16);
870
871                    // Multiply with appropriate scales and accumulate
872                    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]);
873                    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]);
874                    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]);
875                    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]);
876                }
877            }
878
879            // Store the accumulated values
880            for (int i = 0; i < 16; i++) {
881                _mm512_storeu_ps((float *)(s + ((y * 4 + i) * bs + x * 8)), acc_rows[i]);
882            }
883        }
884    }
885
886    // Take a block_q8_0x4 structures at each pass of the loop and perform dot product operation
887    for (; y < nr / 4; y ++) {
888        const block_q8_0x4 * a_ptr = a_ptr_start + (y * nb);
889
890        // Take group of two block_tx8 structures at each pass of the loop and perform dot product operation
891        for (int64_t x = 0; x < anc / 8; x += 2) {
892
893            const block_tx8 * b_ptr_0 = b_ptr_start + ((x)     * b_nb);
894            const block_tx8 * b_ptr_1 = b_ptr_start + ((x + 1) * b_nb);
895
896            // Master FP accumulators
897            __m512 acc_rows[4];
898            for (int i = 0; i < 4; i++) {
899                acc_rows[i] = _mm512_setzero_ps();
900            }
901
902            for (int64_t b = 0; b < nb; b++) {
903                // Load the sixteen blocks of quantized values interleaved with each other in chunks of eight - B0,B1 ....BE,BF
904                const __m256i rhs_raw_mat_0123_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs));
905                const __m256i rhs_raw_mat_4567_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs + 32));
906                const __m256i rhs_raw_mat_0123_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs + 64));
907                const __m256i rhs_raw_mat_4567_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_0[b].qs + 96));
908
909                const __m256i rhs_raw_mat_89AB_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs));
910                const __m256i rhs_raw_mat_CDEF_0 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs + 32));
911                const __m256i rhs_raw_mat_89AB_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs + 64));
912                const __m256i rhs_raw_mat_CDEF_1 = _mm256_loadu_si256((const __m256i *)(b_ptr_1[b].qs + 96));
913
914                // Save the values in the following vectors in the formats B0B1B4B5, B2B3B6B7 for further processing and storing of values
915                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);
916                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);
917                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);
918                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);
919
920                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);
921                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);
922                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);
923                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);
924
925                const __m512i rhs_raw_mat_014589CD_0 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_0), rhs_raw_mat_89CD_0, 1);
926                const __m512i rhs_raw_mat_2367ABEF_0 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_0), rhs_raw_mat_ABEF_0, 1);
927                const __m512i rhs_raw_mat_014589CD_1 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_1), rhs_raw_mat_89CD_1, 1);
928                const __m512i rhs_raw_mat_2367ABEF_1 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_1), rhs_raw_mat_ABEF_1, 1);
929
930                // 4-bit -> 8-bit - Sign is maintained
931                const __m512i rhs_mat_014589CD_0 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_014589CD_0, m4bexpanded)); //B0(0-7) B1(0-7) B4(0-7) B5(0-7) B8(0-7) B9(0-7) BC(0-7) BD(0-7)
932                const __m512i rhs_mat_2367ABEF_0 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_2367ABEF_0, m4bexpanded)); //B2(0-7) B3(0-7) B6(0-7) B7(0-7) BA(0-7) BB(0-7) BE(0-7) BF(0-7)
933
934                const __m512i rhs_mat_014589CD_1 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_014589CD_1, m4bexpanded)); //B0(8-15) B1(8-15) B4(8-15) B5(8-15) B8(8-15) B9(8-15) BC(8-15) BD(8-15)
935                const __m512i rhs_mat_2367ABEF_1 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(rhs_raw_mat_2367ABEF_1, m4bexpanded)); //B2(8-15) B3(8-15) B6(8-15) B7(8-15) BA(8-15) BB(8-15) BE(8-15) BF(8-15)
936
937                const __m512i rhs_mat_014589CD_2 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_0, 4), m4bexpanded)); //B0(16-23) B1(16-23) B4(16-23) B5(16-23) B8(16-23) B9(16-23) BC(16-23) BD(16-23)
938                const __m512i rhs_mat_2367ABEF_2 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_0, 4), m4bexpanded)); //B2(16-23) B3(16-23) B6(16-23) B7(16-23) BA(16-23) BB(16-23) BE(16-23) BF(16-23)
939
940                const __m512i rhs_mat_014589CD_3 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_1, 4), m4bexpanded)); //B0(24-31) B1(24-31) B4(24-31) B5(24-31) B8(24-31) B9(24-31) BC(24-31) BD(24-31)
941                const __m512i rhs_mat_2367ABEF_3 = _mm512_shuffle_epi8(signextendlutexpanded, _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_1, 4), m4bexpanded)); //B2(24-31) B3(24-31) B6(24-31) B7(24-31) BA(24-31) BB(24-31) BE(24-31) BF(24-31)
942
943                // Shuffle pattern one - right side input
944                const __m512i rhs_mat_014589CD_0_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_0, (_MM_PERM_ENUM)136); //B0(0-3) B1(0-3) B0(0-3) B1(0-3) B4(0-3) B5(0-3) B4(0-3) B5(0-3) B8(0-3) B9(0-3) B8(0-3) B9(0-3) BC(0-3) BD(0-3) BC(0-3) BD(0-3)
945                const __m512i rhs_mat_2367ABEF_0_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_0, (_MM_PERM_ENUM)136); //B2(0-3) B3(0-3) B2(0-3) B3(0-3) B6(0-3) B7(0-3) B6(0-3) B7(0-3) BA(0-3) BB(0-3) BA(0-3) BB(0-3) BE(0-3) BF(0-3) BE(0-3) BF(0-3)
946
947                const __m512i rhs_mat_014589CD_1_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_1, (_MM_PERM_ENUM)136); //B0(8-11) B1(8-11) B0(8-11) B1(8-11) B4(8-11) B5(8-11) B4(8-11) B5(8-11) B8(8-11) B9(8-11) B8(8-11) B9(8-11) BC(8-11) BD(8-11) BC(8-11) BD(8-11)
948                const __m512i rhs_mat_2367ABEF_1_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_1, (_MM_PERM_ENUM)136); //B2(8-11) B3(8-11) B2(8-11) B3(8-11) B6(8-11) B7(8-11) B6(8-11) B7(8-11) BA(8-11) BB(8-11) BA(8-11) BB(8-11) BE(8-11) BF(8-11) BE(8-11) BF(8-11)
949
950                const __m512i rhs_mat_014589CD_2_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_2, (_MM_PERM_ENUM)136); //B0(16-19) B1(16-19) B0(16-19) B1(16-19) B4(16-19) B5(16-19) B4(16-19) B5(16-19) B8(16-19) B9(16-19) B8(16-19) B9(16-19) BC(16-19) BD(16-19) BC(16-19) BD(16-19)
951                const __m512i rhs_mat_2367ABEF_2_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_2, (_MM_PERM_ENUM)136); //B2(16-19) B3(16-19) B2(16-19) B3(16-19) B6(16-19) B7(16-19) B6(16-19) B7(16-19) BA(16-19) BB(16-19) BA(16-19) BB(16-19) BE(16-19) BF(16-19) BE(16-19) BF(16-19)
952
953                const __m512i rhs_mat_014589CD_3_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_3, (_MM_PERM_ENUM)136); //B0(24-27) B1(24-27) B0(24-27) B1(24-27) B4(24-27) B5(24-27) B4(24-27) B5(24-27) B8(24-27) B9(24-27) B8(24-27) B9(24-27) BC(24-27) BD(24-27) BC(24-27) BD(24-27)
954                const __m512i rhs_mat_2367ABEF_3_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_3, (_MM_PERM_ENUM)136); //B2(24-27) B3(24-27) B2(24-27) B3(24-27) B6(24-27) B7(24-27) B6(24-27) B7(24-27) BA(24-27) BB(24-27) BA(24-27) BB(24-27) BE(24-27) BF(24-27) BE(24-27) BF(24-27)
955
956                // Shuffle pattern two - right side input
957
958                const __m512i rhs_mat_014589CD_0_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_0, (_MM_PERM_ENUM)221); //B0(4-7) B1(4-7) B0(4-7) B1(4-7) B4(4-7) B5(4-7) B4(4-7) B5(4-7) B8(4-7) B9(4-7) B8(4-7) B9(4-7) BC(4-7) BD(4-7) BC(4-7) BD(4-7)
959                const __m512i rhs_mat_2367ABEF_0_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_0, (_MM_PERM_ENUM)221); //B2(4-7) B3(4-7) B2(4-7) B3(4-7) B6(4-7) B7(4-7) B6(4-7) B7(4-7) BA(4-7) BB(4-7) BA(4-7) BB(4-7) BE(4-7) BF(4-7) BE(4-7) BF(4-7)
960
961                const __m512i rhs_mat_014589CD_1_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_1, (_MM_PERM_ENUM)221); //B0(12-15) B1(12-15) B0(12-15) B1(12-15) B4(12-15) B5(12-15) B4(12-15) B5(12-15) B8(12-15) B9(12-15) B8(12-15) B9(12-15) BC(12-15) BD(12-15) BC(12-15) BD(12-15)
962                const __m512i rhs_mat_2367ABEF_1_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_1, (_MM_PERM_ENUM)221); //B2(12-15) B3(12-15) B2(12-15) B3(12-15) B6(12-15) B7(12-15) B6(12-15) B7(12-15) BA(12-15) BB(12-15) BA(12-15) BB(12-15) BE(12-15) BF(12-15) BE(12-15) BF(12-15)
963
964                const __m512i rhs_mat_014589CD_2_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_2, (_MM_PERM_ENUM)221); //B0(20-23) B1(20-23) B0(20-23) B1(20-23) B4(20-23) B5(20-23) B4(20-23) B5(20-23) B8(20-23) B9(20-23) B8(20-23) B9(20-23) BC(20-23) BD(20-23) BC(20-23) BD(20-23)
965                const __m512i rhs_mat_2367ABEF_2_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_2, (_MM_PERM_ENUM)221); //B2(20-23) B3(20-23) B2(20-23) B3(20-23) B6(20-23) B7(20-23) B6(20-23) B7(20-23) BA(20-23) BB(20-23) BA(20-23) BB(20-23) BE(20-23) BF(20-23) BE(20-23) BF(20-23)
966
967                const __m512i rhs_mat_014589CD_3_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_3, (_MM_PERM_ENUM)221); //B0(28-31) B1(28-31) B0(28-31) B1(28-31) B4(28-31) B5(28-31) B4(28-31) B5(28-31) B8(28-31) B9(28-31) B8(28-31) B9(28-31) BC(28-31) BD(28-31) BC(28-31) BD(28-31)
968                const __m512i rhs_mat_2367ABEF_3_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_3, (_MM_PERM_ENUM)221); //B2(28-31) B3(28-31) B2(28-31) B3(28-31) B6(28-31) B7(28-31) B6(28-31) B7(28-31) BA(28-31) BB(28-31) BA(28-31) BB(28-31) BE(28-31) BF(28-31) BE(28-31) BF(28-31)
969
970
971                // Scale values - Load the weight scale values of two block_tx8
972                __m512 col_scale_f32;
973                if constexpr (
974                        std::is_same_v<block_tx8, block_q4_0x8> ||
975                        std::is_same_v<block_tx8, block_iq4_nlx8>) {
976                    col_scale_f32 = GGML_F32Cx8x2_LOAD(b_ptr_0[b].d, b_ptr_1[b].d);
977                } else if constexpr (std::is_same_v<block_tx8, block_mxfp4x8>) {
978                    //TODO: simd-ify
979                    col_scale_f32 = _mm512_set_ps(
980                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_1[b].e[7]),
981                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_1[b].e[6]),
982                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_1[b].e[5]),
983                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_1[b].e[4]),
984                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_1[b].e[3]),
985                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_1[b].e[2]),
986                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_1[b].e[1]),
987                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_1[b].e[0]),
988                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_0[b].e[7]),
989                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_0[b].e[6]),
990                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_0[b].e[5]),
991                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_0[b].e[4]),
992                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_0[b].e[3]),
993                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_0[b].e[2]),
994                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_0[b].e[1]),
995                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr_0[b].e[0]));
996                }
997
998                // Load the four blocks of quantized values interleaved with each other in chunks of eight - A0,A1,A2,A3
999                // Loaded as set of 128 bit vectors and repeated and stored into a 256 bit vector before again repeating into 512 bit vector
1000                __m256i lhs_mat_ymm_0123_0 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs)));
1001                __m256i lhs_mat_ymm_01_0 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_0, lhs_mat_ymm_0123_0, 0);
1002                __m256i lhs_mat_ymm_23_0 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_0, lhs_mat_ymm_0123_0, 17);
1003                __m256i lhs_mat_ymm_0123_1 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs + 32)));
1004                __m256i lhs_mat_ymm_01_1 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_1, lhs_mat_ymm_0123_1, 0);
1005                __m256i lhs_mat_ymm_23_1 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_1, lhs_mat_ymm_0123_1, 17);
1006                __m256i lhs_mat_ymm_0123_2 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs + 64)));
1007                __m256i lhs_mat_ymm_01_2 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_2, lhs_mat_ymm_0123_2, 0);
1008                __m256i lhs_mat_ymm_23_2 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_2, lhs_mat_ymm_0123_2, 17);
1009                __m256i lhs_mat_ymm_0123_3 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs + 96)));
1010                __m256i lhs_mat_ymm_01_3 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_3, lhs_mat_ymm_0123_3, 0);
1011                __m256i lhs_mat_ymm_23_3 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_3, lhs_mat_ymm_0123_3, 17);
1012
1013                __m512i lhs_mat_01_0 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_0), lhs_mat_ymm_01_0, 1);
1014                __m512i lhs_mat_23_0 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_0), lhs_mat_ymm_23_0, 1);
1015                __m512i lhs_mat_01_1 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_1), lhs_mat_ymm_01_1, 1);
1016                __m512i lhs_mat_23_1 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_1), lhs_mat_ymm_23_1, 1);
1017                __m512i lhs_mat_01_2 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_2), lhs_mat_ymm_01_2, 1);
1018                __m512i lhs_mat_23_2 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_2), lhs_mat_ymm_23_2, 1);
1019                __m512i lhs_mat_01_3 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_3), lhs_mat_ymm_01_3, 1);
1020                __m512i lhs_mat_23_3 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_3), lhs_mat_ymm_23_3, 1);
1021
1022                // Shuffle pattern one - left side input
1023
1024                const __m512i lhs_mat_01_0_sp1 = _mm512_shuffle_epi32(lhs_mat_01_0, (_MM_PERM_ENUM)160);  //A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3)
1025                const __m512i lhs_mat_23_0_sp1 = _mm512_shuffle_epi32(lhs_mat_23_0, (_MM_PERM_ENUM)160);  //A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3)
1026
1027                const __m512i lhs_mat_01_1_sp1 = _mm512_shuffle_epi32(lhs_mat_01_1, (_MM_PERM_ENUM)160);  //A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11)
1028                const __m512i lhs_mat_23_1_sp1 = _mm512_shuffle_epi32(lhs_mat_23_1, (_MM_PERM_ENUM)160);  //A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11)
1029
1030                const __m512i lhs_mat_01_2_sp1 = _mm512_shuffle_epi32(lhs_mat_01_2, (_MM_PERM_ENUM)160);  //A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19)
1031                const __m512i lhs_mat_23_2_sp1 = _mm512_shuffle_epi32(lhs_mat_23_2, (_MM_PERM_ENUM)160);  //A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19)
1032
1033                const __m512i lhs_mat_01_3_sp1 = _mm512_shuffle_epi32(lhs_mat_01_3, (_MM_PERM_ENUM)160);  //A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27)
1034                const __m512i lhs_mat_23_3_sp1 = _mm512_shuffle_epi32(lhs_mat_23_3, (_MM_PERM_ENUM)160);  //A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27)
1035
1036                // Shuffle pattern two - left side input
1037
1038                const __m512i lhs_mat_01_0_sp2 = _mm512_shuffle_epi32(lhs_mat_01_0, (_MM_PERM_ENUM)245);  //A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7)
1039                const __m512i lhs_mat_23_0_sp2 = _mm512_shuffle_epi32(lhs_mat_23_0, (_MM_PERM_ENUM)245);  //A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7)
1040
1041                const __m512i lhs_mat_01_1_sp2 = _mm512_shuffle_epi32(lhs_mat_01_1, (_MM_PERM_ENUM)245);  //A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15)
1042                const __m512i lhs_mat_23_1_sp2 = _mm512_shuffle_epi32(lhs_mat_23_1, (_MM_PERM_ENUM)245);  //A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15)
1043
1044                const __m512i lhs_mat_01_2_sp2 = _mm512_shuffle_epi32(lhs_mat_01_2, (_MM_PERM_ENUM)245);  //A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23)
1045                const __m512i lhs_mat_23_2_sp2 = _mm512_shuffle_epi32(lhs_mat_23_2, (_MM_PERM_ENUM)245);  //A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23)
1046
1047                const __m512i lhs_mat_01_3_sp2 = _mm512_shuffle_epi32(lhs_mat_01_3, (_MM_PERM_ENUM)245);  //A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31)
1048                const __m512i lhs_mat_23_3_sp2 = _mm512_shuffle_epi32(lhs_mat_23_3, (_MM_PERM_ENUM)245);  //A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31)
1049
1050                // 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
1051                // Resembles MMLAs into 2x2 matrices in ARM Version
1052                const __m512i zero = _mm512_setzero_epi32();
1053                __m512i iacc_mat_00_sp1 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_01_3_sp1, rhs_mat_014589CD_3_sp1), lhs_mat_01_2_sp1, rhs_mat_014589CD_2_sp1), lhs_mat_01_1_sp1, rhs_mat_014589CD_1_sp1), lhs_mat_01_0_sp1, rhs_mat_014589CD_0_sp1);
1054                __m512i iacc_mat_01_sp1 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_01_3_sp1, rhs_mat_2367ABEF_3_sp1), lhs_mat_01_2_sp1, rhs_mat_2367ABEF_2_sp1), lhs_mat_01_1_sp1, rhs_mat_2367ABEF_1_sp1), lhs_mat_01_0_sp1, rhs_mat_2367ABEF_0_sp1);
1055                __m512i iacc_mat_10_sp1 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_23_3_sp1, rhs_mat_014589CD_3_sp1), lhs_mat_23_2_sp1, rhs_mat_014589CD_2_sp1), lhs_mat_23_1_sp1, rhs_mat_014589CD_1_sp1), lhs_mat_23_0_sp1, rhs_mat_014589CD_0_sp1);
1056                __m512i iacc_mat_11_sp1 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_23_3_sp1, rhs_mat_2367ABEF_3_sp1), lhs_mat_23_2_sp1, rhs_mat_2367ABEF_2_sp1), lhs_mat_23_1_sp1, rhs_mat_2367ABEF_1_sp1), lhs_mat_23_0_sp1, rhs_mat_2367ABEF_0_sp1);
1057                __m512i iacc_mat_00_sp2 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_01_3_sp2, rhs_mat_014589CD_3_sp2), lhs_mat_01_2_sp2, rhs_mat_014589CD_2_sp2), lhs_mat_01_1_sp2, rhs_mat_014589CD_1_sp2), lhs_mat_01_0_sp2, rhs_mat_014589CD_0_sp2);
1058                __m512i iacc_mat_01_sp2 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_01_3_sp2, rhs_mat_2367ABEF_3_sp2), lhs_mat_01_2_sp2, rhs_mat_2367ABEF_2_sp2), lhs_mat_01_1_sp2, rhs_mat_2367ABEF_1_sp2), lhs_mat_01_0_sp2, rhs_mat_2367ABEF_0_sp2);
1059                __m512i iacc_mat_10_sp2 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_23_3_sp2, rhs_mat_014589CD_3_sp2), lhs_mat_23_2_sp2, rhs_mat_014589CD_2_sp2), lhs_mat_23_1_sp2, rhs_mat_014589CD_1_sp2), lhs_mat_23_0_sp2, rhs_mat_014589CD_0_sp2);
1060                __m512i iacc_mat_11_sp2 = mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(mul_sum_i8_pairs_acc_int32x16(zero, lhs_mat_23_3_sp2, rhs_mat_2367ABEF_3_sp2), lhs_mat_23_2_sp2, rhs_mat_2367ABEF_2_sp2), lhs_mat_23_1_sp2, rhs_mat_2367ABEF_1_sp2), lhs_mat_23_0_sp2, rhs_mat_2367ABEF_0_sp2);
1061
1062                // Output of both shuffle patterns are added in order to sum dot product outputs of all 32 values in block
1063                __m512i iacc_mat_00 = _mm512_add_epi32(iacc_mat_00_sp1, iacc_mat_00_sp2);
1064                __m512i iacc_mat_01 = _mm512_add_epi32(iacc_mat_01_sp1, iacc_mat_01_sp2);
1065                __m512i iacc_mat_10 = _mm512_add_epi32(iacc_mat_10_sp1, iacc_mat_10_sp2);
1066                __m512i iacc_mat_11 = _mm512_add_epi32(iacc_mat_11_sp1, iacc_mat_11_sp2);
1067
1068
1069                // Straighten out to make 4 row vectors
1070                __m512i iacc_row_0 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_00, _mm512_shuffle_epi32(iacc_mat_01, (_MM_PERM_ENUM)78));
1071                __m512i iacc_row_1 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_00, (_MM_PERM_ENUM)78), iacc_mat_01);
1072                __m512i iacc_row_2 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_10, _mm512_shuffle_epi32(iacc_mat_11, (_MM_PERM_ENUM)78));
1073                __m512i iacc_row_3 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_10, (_MM_PERM_ENUM)78), iacc_mat_11);
1074
1075                // Load the scale(d) values for all the 4 Q8_0 blocks and repeat it across lanes
1076                const __m128i row_scale_f16 = _mm_shuffle_epi32(_mm_maskload_epi32((int const*)(a_ptr[b].d), loadMask), 68);
1077                const __m512 row_scale_f32 = GGML_F32Cx16_REPEAT_LOAD(row_scale_f16);
1078
1079                // Multiply with appropriate scales and accumulate
1080                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]);
1081                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]);
1082                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]);
1083                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]);
1084            }
1085
1086            // Store the accumulated values
1087            for (int i = 0; i < 4; i++) {
1088                _mm512_storeu_ps((float *)(s + ((y * 4 + i) * bs + x * 8)), acc_rows[i]);
1089            }
1090        }
1091    }
1092    if (anc != nc) {
1093        xstart = anc/8;
1094        y = 0;
1095    }
1096#endif // __AVX512BW__ && __AVX512DQ__
1097
1098    // Take group of four block_q8_0x4 structures at each pass of the loop and perform dot product operation
1099
1100    for (; y < anr / 4; y += 4) {
1101        const block_q8_0x4 * a_ptrs[4];
1102
1103        a_ptrs[0] = a_ptr_start + (y * nb);
1104        for (int i = 0; i < 3; ++i) {
1105            a_ptrs[i + 1] = a_ptrs[i] + nb;
1106        }
1107
1108        // Take group of eight block_tx8 structures at each pass of the loop and perform dot product operation
1109        for (int64_t x = xstart; x < nc / 8; x++) {
1110
1111            const block_tx8 * b_ptr = b_ptr_start + (x * b_nb);
1112
1113            // Master FP accumulators
1114            __m256 acc_rows[16];
1115            for (int i = 0; i < 16; i++) {
1116                acc_rows[i] = _mm256_setzero_ps();
1117            }
1118
1119            for (int64_t b = 0; b < nb; b++) {
1120                // Load the eight blocks of quantized values interleaved with each other in chunks of eight - B0,B1 ....B6,B7
1121                const __m256i rhs_raw_mat_0123_0 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs));
1122                const __m256i rhs_raw_mat_4567_0 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 32));
1123                const __m256i rhs_raw_mat_0123_1 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 64));
1124                const __m256i rhs_raw_mat_4567_1 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 96));
1125
1126                // Save the values in the following vectors in the formats B0B1B4B5, B2B3B6B7 for further processing and storing of values
1127                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);
1128                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);
1129                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);
1130                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);
1131
1132                // 4-bit -> 8-bit - Sign is maintained
1133                const __m256i rhs_mat_0145_0 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_0145_0, m4b)); //B0(0-7) B1(0-7) B4(0-7) B5(0-7)
1134                const __m256i rhs_mat_2367_0 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_2367_0, m4b)); //B2(0-7) B3(0-7) B6(0-7) B7(0-7)
1135
1136                const __m256i rhs_mat_0145_1 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_0145_1, m4b)); //B0(8-15) B1(8-15) B4(8-15) B5(8-15)
1137                const __m256i rhs_mat_2367_1 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_2367_1, m4b)); //B2(8-15) B3(8-15) B6(8-15) B7(8-15)
1138
1139                const __m256i rhs_mat_0145_2 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_0, 4), m4b)); //B0(16-23) B1(16-23) B4(16-23) B5(16-23)
1140                const __m256i rhs_mat_2367_2 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_0, 4), m4b)); //B2(16-23) B3(16-23) B6(16-23) B7(16-23)
1141
1142                const __m256i rhs_mat_0145_3 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_1, 4), m4b)); //B0(24-31) B1(24-31) B4(24-31) B5(24-31)
1143                const __m256i rhs_mat_2367_3 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_1, 4), m4b)); //B2(24-31) B3(24-31) B6(24-31) B7(24-31)
1144
1145                // Shuffle pattern one - right side input
1146                const __m256i rhs_mat_0145_0_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_0, 136);  //B0(0-3) B1(0-3) B0(0-3) B1(0-3) B4(0-3) B5(0-3) B4(0-3) B5(0-3)
1147                const __m256i rhs_mat_2367_0_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_0, 136);  //B2(0-3) B3(0-3) B2(0-3) B3(0-3) B6(0-3) B7(0-3) B6(0-3) B7(0-3)
1148
1149                const __m256i rhs_mat_0145_1_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_1, 136);  //B0(8-11) B1(8-11) B0(8-11) B1(8-11) B4(8-11) B5(8-11) B4(8-11) B5(8-11)
1150                const __m256i rhs_mat_2367_1_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_1, 136);  //B2(8-11) B3(8-11) B2(8-11) B3(8-11) B6(8-11) B7(8-11) B6(8-11) B7(8-11)
1151
1152                const __m256i rhs_mat_0145_2_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_2, 136);  //B0(16-19) B1(16-19) B0(16-19) B1(16-19) B4(16-19) B5(16-19) B4(16-19) B5(16-19)
1153                const __m256i rhs_mat_2367_2_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_2, 136);  //B2(16-19) B3(16-19) B2(16-19) B3(16-19) B6(16-19) B7(16-19) B6(16-19) B7(16-19)
1154
1155                const __m256i rhs_mat_0145_3_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_3, 136);  //B0(24-27) B1(24-27) B0(24-27) B1(24-27) B4(24-27) B5(24-27) B4(24-27) B5(24-27)
1156                const __m256i rhs_mat_2367_3_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_3, 136);  //B2(24-27) B3(24-27) B2(24-27) B3(24-27) B6(24-27) B7(24-27) B6(24-27) B7(24-27)
1157
1158                // Shuffle pattern two - right side input
1159
1160                const __m256i rhs_mat_0145_0_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_0, 221);  //B0(4-7) B1(4-7) B0(4-7) B1(4-7) B4(4-7) B5(4-7) B4(4-7) B5(4-7)
1161                const __m256i rhs_mat_2367_0_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_0, 221);  //B2(4-7) B3(4-7) B2(4-7) B3(4-7) B6(4-7) B7(4-7) B6(4-7) B7(4-7)
1162
1163                const __m256i rhs_mat_0145_1_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_1, 221);  //B0(12-15) B1(12-15) B0(12-15) B1(12-15) B4(12-15) B5(12-15) B4(12-15) B5(12-15)
1164                const __m256i rhs_mat_2367_1_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_1, 221);  //B2(12-15) B3(12-15) B2(12-15) B3(12-15) B6(12-15) B7(12-15) B6(12-15) B7(12-15)
1165
1166                const __m256i rhs_mat_0145_2_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_2, 221);  //B0(20-23) B1(20-23) B0(20-23) B1(20-23) B4(20-23) B5(20-23) B4(20-23) B5(20-23)
1167                const __m256i rhs_mat_2367_2_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_2, 221);  //B2(20-23) B3(20-23) B2(20-23) B3(20-23) B6(20-23) B7(20-23) B6(20-23) B7(20-23)
1168
1169                const __m256i rhs_mat_0145_3_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_3, 221);  //B0(28-31) B1(28-31) B0(28-31) B1(28-31) B4(28-31) B5(28-31) B4(28-31) B5(28-31)
1170                const __m256i rhs_mat_2367_3_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_3, 221);  //B2(28-31) B3(28-31) B2(28-31) B3(28-31) B6(28-31) B7(28-31) B6(28-31) B7(28-31)
1171
1172                // Scale values - Load the wight scale values of block_tx8
1173                __m256 col_scale_f32;
1174                if constexpr (
1175                        std::is_same_v<block_tx8, block_q4_0x8> ||
1176                        std::is_same_v<block_tx8, block_iq4_nlx8>) {
1177                    col_scale_f32 = GGML_F32Cx8_LOAD(b_ptr[b].d);
1178                } else if constexpr (std::is_same_v<block_tx8, block_mxfp4x8>) {
1179                    col_scale_f32 = _mm256_set_ps(
1180                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[b].e[7]),
1181                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[b].e[6]),
1182                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[b].e[5]),
1183                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[b].e[4]),
1184                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[b].e[3]),
1185                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[b].e[2]),
1186                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[b].e[1]),
1187                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[b].e[0]));
1188                }
1189
1190                // Process LHS in groups of four
1191                for (int rp = 0; rp < 4; rp++) {
1192                    // Load the four blocks of quantized values interleaved with each other in chunks of eight - A0,A1,A2,A3
1193                    // Loaded as set of 128 bit vectors and repeated into a 256 bit vector
1194                    __m256i lhs_mat_0123_0 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs)));
1195                    __m256i lhs_mat_01_0 = _mm256_permute2f128_si256(lhs_mat_0123_0, lhs_mat_0123_0, 0);
1196                    __m256i lhs_mat_23_0 = _mm256_permute2f128_si256(lhs_mat_0123_0, lhs_mat_0123_0, 17);
1197                    __m256i lhs_mat_0123_1 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs + 32)));
1198                    __m256i lhs_mat_01_1 = _mm256_permute2f128_si256(lhs_mat_0123_1, lhs_mat_0123_1, 0);
1199                    __m256i lhs_mat_23_1 = _mm256_permute2f128_si256(lhs_mat_0123_1, lhs_mat_0123_1, 17);
1200                    __m256i lhs_mat_0123_2 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs + 64)));
1201                    __m256i lhs_mat_01_2 = _mm256_permute2f128_si256(lhs_mat_0123_2, lhs_mat_0123_2, 0);
1202                    __m256i lhs_mat_23_2 = _mm256_permute2f128_si256(lhs_mat_0123_2, lhs_mat_0123_2, 17);
1203                    __m256i lhs_mat_0123_3 = _mm256_loadu_si256((const __m256i *)((a_ptrs[rp][b].qs + 96)));
1204                    __m256i lhs_mat_01_3 = _mm256_permute2f128_si256(lhs_mat_0123_3, lhs_mat_0123_3, 0);
1205                    __m256i lhs_mat_23_3 = _mm256_permute2f128_si256(lhs_mat_0123_3, lhs_mat_0123_3, 17);
1206
1207                    // Shuffle pattern one - left side input
1208                    const __m256i lhs_mat_01_0_sp1 = _mm256_shuffle_epi32(lhs_mat_01_0, 160);  //A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3)
1209                    const __m256i lhs_mat_23_0_sp1 = _mm256_shuffle_epi32(lhs_mat_23_0, 160);  //A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3)
1210
1211                    const __m256i lhs_mat_01_1_sp1 = _mm256_shuffle_epi32(lhs_mat_01_1, 160);  //A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11)
1212                    const __m256i lhs_mat_23_1_sp1 = _mm256_shuffle_epi32(lhs_mat_23_1, 160);  //A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11)
1213
1214                    const __m256i lhs_mat_01_2_sp1 = _mm256_shuffle_epi32(lhs_mat_01_2, 160);  //A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19)
1215                    const __m256i lhs_mat_23_2_sp1 = _mm256_shuffle_epi32(lhs_mat_23_2, 160);  //A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19)
1216
1217                    const __m256i lhs_mat_01_3_sp1 = _mm256_shuffle_epi32(lhs_mat_01_3, 160);  //A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27)
1218                    const __m256i lhs_mat_23_3_sp1 = _mm256_shuffle_epi32(lhs_mat_23_3, 160);  //A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27)
1219
1220                    // Shuffle pattern two - left side input
1221                    const __m256i lhs_mat_01_0_sp2 = _mm256_shuffle_epi32(lhs_mat_01_0, 245);  //A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7)
1222                    const __m256i lhs_mat_23_0_sp2 = _mm256_shuffle_epi32(lhs_mat_23_0, 245);  //A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7)
1223
1224                    const __m256i lhs_mat_01_1_sp2 = _mm256_shuffle_epi32(lhs_mat_01_1, 245);  //A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15)
1225                    const __m256i lhs_mat_23_1_sp2 = _mm256_shuffle_epi32(lhs_mat_23_1, 245);  //A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15)
1226
1227                    const __m256i lhs_mat_01_2_sp2 = _mm256_shuffle_epi32(lhs_mat_01_2, 245);  //A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23)
1228                    const __m256i lhs_mat_23_2_sp2 = _mm256_shuffle_epi32(lhs_mat_23_2, 245);  //A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23)
1229
1230                    const __m256i lhs_mat_01_3_sp2 = _mm256_shuffle_epi32(lhs_mat_01_3, 245);  //A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31)
1231                    const __m256i lhs_mat_23_3_sp2 = _mm256_shuffle_epi32(lhs_mat_23_3, 245);  //A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31)
1232
1233                    // 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
1234                    // Resembles MMLAs into 2x2 matrices in ARM Version
1235                    const __m256i zero = _mm256_setzero_si256();
1236                    __m256i iacc_mat_00_sp1 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_01_3_sp1, rhs_mat_0145_3_sp1), lhs_mat_01_2_sp1, rhs_mat_0145_2_sp1), lhs_mat_01_1_sp1, rhs_mat_0145_1_sp1), lhs_mat_01_0_sp1, rhs_mat_0145_0_sp1);
1237                    __m256i iacc_mat_01_sp1 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_01_3_sp1, rhs_mat_2367_3_sp1), lhs_mat_01_2_sp1, rhs_mat_2367_2_sp1), lhs_mat_01_1_sp1, rhs_mat_2367_1_sp1), lhs_mat_01_0_sp1, rhs_mat_2367_0_sp1);
1238                    __m256i iacc_mat_10_sp1 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_23_3_sp1, rhs_mat_0145_3_sp1), lhs_mat_23_2_sp1, rhs_mat_0145_2_sp1), lhs_mat_23_1_sp1, rhs_mat_0145_1_sp1), lhs_mat_23_0_sp1, rhs_mat_0145_0_sp1);
1239                    __m256i iacc_mat_11_sp1 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_23_3_sp1, rhs_mat_2367_3_sp1), lhs_mat_23_2_sp1, rhs_mat_2367_2_sp1), lhs_mat_23_1_sp1, rhs_mat_2367_1_sp1), lhs_mat_23_0_sp1, rhs_mat_2367_0_sp1);
1240                    __m256i iacc_mat_00_sp2 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_01_3_sp2, rhs_mat_0145_3_sp2), lhs_mat_01_2_sp2, rhs_mat_0145_2_sp2), lhs_mat_01_1_sp2, rhs_mat_0145_1_sp2), lhs_mat_01_0_sp2, rhs_mat_0145_0_sp2);
1241                    __m256i iacc_mat_01_sp2 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_01_3_sp2, rhs_mat_2367_3_sp2), lhs_mat_01_2_sp2, rhs_mat_2367_2_sp2), lhs_mat_01_1_sp2, rhs_mat_2367_1_sp2), lhs_mat_01_0_sp2, rhs_mat_2367_0_sp2);
1242                    __m256i iacc_mat_10_sp2 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_23_3_sp2, rhs_mat_0145_3_sp2), lhs_mat_23_2_sp2, rhs_mat_0145_2_sp2), lhs_mat_23_1_sp2, rhs_mat_0145_1_sp2), lhs_mat_23_0_sp2, rhs_mat_0145_0_sp2);
1243                    __m256i iacc_mat_11_sp2 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_23_3_sp2, rhs_mat_2367_3_sp2), lhs_mat_23_2_sp2, rhs_mat_2367_2_sp2), lhs_mat_23_1_sp2, rhs_mat_2367_1_sp2), lhs_mat_23_0_sp2, rhs_mat_2367_0_sp2);
1244
1245                    // Output of both shuffle patterns are added in order to sum dot product outputs of all 32 values in block
1246                    __m256i iacc_mat_00 = _mm256_add_epi32(iacc_mat_00_sp1, iacc_mat_00_sp2);
1247                    __m256i iacc_mat_01 = _mm256_add_epi32(iacc_mat_01_sp1, iacc_mat_01_sp2);
1248                    __m256i iacc_mat_10 = _mm256_add_epi32(iacc_mat_10_sp1, iacc_mat_10_sp2);
1249                    __m256i iacc_mat_11 = _mm256_add_epi32(iacc_mat_11_sp1, iacc_mat_11_sp2);
1250
1251                    // Straighten out to make 4 row vectors
1252                    __m256i iacc_row_0 = _mm256_blend_epi32(iacc_mat_00, _mm256_shuffle_epi32(iacc_mat_01, 78), 204);
1253                    __m256i iacc_row_1 = _mm256_blend_epi32(_mm256_shuffle_epi32(iacc_mat_00, 78), iacc_mat_01, 204);
1254                    __m256i iacc_row_2 = _mm256_blend_epi32(iacc_mat_10, _mm256_shuffle_epi32(iacc_mat_11, 78), 204);
1255                    __m256i iacc_row_3 = _mm256_blend_epi32(_mm256_shuffle_epi32(iacc_mat_10, 78), iacc_mat_11, 204);
1256
1257                    // Load the scale(d) values for all the 4 Q8_0 blocks and repeat it across lanes
1258                    const __m256 row_scale_f32 = GGML_F32Cx8_REPEAT_LOAD(a_ptrs[rp][b].d, loadMask);
1259
1260                    // Multiply with appropriate scales and accumulate
1261                    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]);
1262                    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]);
1263                    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]);
1264                    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]);
1265                }
1266            }
1267
1268            // Store the accumulated values
1269            for (int i = 0; i < 16; i++) {
1270                _mm256_storeu_ps((float *)(s + ((y * 4 + i) * bs + x * 8)), acc_rows[i]);
1271            }
1272        }
1273    }
1274
1275    // Take a block_q8_0x4 structures at each pass of the loop and perform dot product operation
1276    for (; y < nr / 4; y ++) {
1277        const block_q8_0x4 * a_ptr = a_ptr_start + (y * nb);
1278
1279        // Load the eight blocks of quantized values interleaved with each other in chunks of eight - B0,B1 ....B6,B7
1280        for (int64_t x = xstart; x < nc / 8; x++) {
1281            const block_tx8 * b_ptr = b_ptr_start + (x * b_nb);
1282
1283            // Master FP accumulators
1284            __m256 acc_rows[4];
1285            for (int i = 0; i < 4; i++) {
1286                acc_rows[i] = _mm256_setzero_ps();
1287            }
1288
1289            for (int64_t b = 0; b < nb; b++) {
1290                // Load the eight block_q8_0 quantized values interleaved with each other in chunks of eight - B0,B1 ....B6,B7
1291                const __m256i rhs_raw_mat_0123_0 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs));
1292                const __m256i rhs_raw_mat_4567_0 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 32));
1293                const __m256i rhs_raw_mat_0123_1 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 64));
1294                const __m256i rhs_raw_mat_4567_1 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 96));
1295
1296                // Save the values in the following vectors in the formats B0B1B4B5, B2B3B6B7 for further processing and storing of values
1297                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);
1298                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);
1299                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);
1300                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);
1301
1302                // 4-bit -> 8-bit - Sign is maintained
1303                const __m256i rhs_mat_0145_0 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_0145_0, m4b));  //B0(0-7) B1(0-7) B4(0-7) B5(0-7)
1304                const __m256i rhs_mat_2367_0 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_2367_0, m4b));  //B2(0-7) B3(0-7) B6(0-7) B7(0-7)
1305
1306                const __m256i rhs_mat_0145_1 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_0145_1, m4b));  //B0(8-15) B1(8-15) B4(8-15) B5(8-15)
1307                const __m256i rhs_mat_2367_1 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(rhs_raw_mat_2367_1, m4b));  //B2(8-15) B3(8-15) B6(8-15) B7(8-15)
1308
1309                const __m256i rhs_mat_0145_2 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_0, 4), m4b));  //B0(16-23) B1(16-23) B4(16-23) B5(16-23)
1310                const __m256i rhs_mat_2367_2 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_0, 4), m4b));  //B2(16-23) B3(16-23) B6(16-23) B7(16-23)
1311
1312                const __m256i rhs_mat_0145_3 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_1, 4), m4b));  //B0(24-31) B1(24-31) B4(24-31) B5(24-31)
1313                const __m256i rhs_mat_2367_3 = _mm256_shuffle_epi8(signextendlut, _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_1, 4), m4b));  //B2(24-31) B3(24-31) B6(24-31) B7(24-31)
1314
1315                // Shuffle pattern one - right side input
1316                const __m256i rhs_mat_0145_0_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_0, 136);  //B0(0-3) B1(0-3) B0(0-3) B1(0-3) B4(0-3) B5(0-3) B4(0-3) B5(0-3)
1317                const __m256i rhs_mat_2367_0_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_0, 136);  //B2(0-3) B3(0-3) B2(0-3) B3(0-3) B6(0-3) B7(0-3) B6(0-3) B7(0-3)
1318
1319                const __m256i rhs_mat_0145_1_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_1, 136);  //B0(8-11) B1(8-11) B0(8-11) B1(8-11) B4(8-11) B5(8-11) B4(8-11) B5(8-11)
1320                const __m256i rhs_mat_2367_1_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_1, 136);  //B2(8-11) B3(8-11) B2(8-11) B3(8-11) B6(8-11) B7(8-11) B6(8-11) B7(8-11)
1321
1322                const __m256i rhs_mat_0145_2_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_2, 136);  //B0(16-19) B1(16-19) B0(16-19) B1(16-19) B4(16-19) B5(16-19) B4(16-19) B5(16-19)
1323                const __m256i rhs_mat_2367_2_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_2, 136);  //B2(16-19) B3(16-19) B2(16-19) B3(16-19) B6(16-19) B7(16-19) B6(16-19) B7(16-19)
1324
1325                const __m256i rhs_mat_0145_3_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_3, 136);  //B0(24-27) B1(24-27) B0(24-27) B1(24-27) B4(24-27) B5(24-27) B4(24-27) B5(24-27)
1326                const __m256i rhs_mat_2367_3_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_3, 136);  //B2(24-27) B3(24-27) B2(24-27) B3(24-27) B6(24-27) B7(24-27) B6(24-27) B7(24-27)
1327
1328                // Shuffle pattern two - right side input
1329
1330                const __m256i rhs_mat_0145_0_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_0, 221);  //B0(4-7) B1(4-7) B0(4-7) B1(4-7) B4(4-7) B5(4-7) B4(4-7) B5(4-7)
1331                const __m256i rhs_mat_2367_0_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_0, 221);  //B2(4-7) B3(4-7) B2(4-7) B3(4-7) B6(4-7) B7(4-7) B6(4-7) B7(4-7)
1332
1333                const __m256i rhs_mat_0145_1_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_1, 221);  //B0(12-15) B1(12-15) B0(12-15) B1(12-15) B4(12-15) B5(12-15) B4(12-15) B5(12-15)
1334                const __m256i rhs_mat_2367_1_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_1, 221);  //B2(12-15) B3(12-15) B2(12-15) B3(12-15) B6(12-15) B7(12-15) B6(12-15) B7(12-15)
1335
1336                const __m256i rhs_mat_0145_2_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_2, 221);  //B0(20-23) B1(20-23) B0(20-23) B1(20-23) B4(20-23) B5(20-23) B4(20-23) B5(20-23)
1337                const __m256i rhs_mat_2367_2_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_2, 221);  //B2(20-23) B3(20-23) B2(20-23) B3(20-23) B6(20-23) B7(20-23) B6(20-23) B7(20-23)
1338
1339                const __m256i rhs_mat_0145_3_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_3, 221);  //B0(28-31) B1(28-31) B0(28-31) B1(28-31) B4(28-31) B5(28-31) B4(28-31) B5(28-31)
1340                const __m256i rhs_mat_2367_3_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_3, 221);  //B2(28-31) B3(28-31) B2(28-31) B3(28-31) B6(28-31) B7(28-31) B6(28-31) B7(28-31)
1341
1342                // Scale values - Load the wight scale values of block_tx8
1343                __m256 col_scale_f32;
1344                if constexpr (
1345                        std::is_same_v<block_tx8, block_q4_0x8> ||
1346                        std::is_same_v<block_tx8, block_iq4_nlx8>) {
1347                    col_scale_f32 = GGML_F32Cx8_LOAD(b_ptr[b].d);
1348                } else if constexpr (std::is_same_v<block_tx8, block_mxfp4x8>) {
1349                    col_scale_f32 = _mm256_set_ps(
1350                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[b].e[7]),
1351                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[b].e[6]),
1352                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[b].e[5]),
1353                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[b].e[4]),
1354                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[b].e[3]),
1355                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[b].e[2]),
1356                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[b].e[1]),
1357                        GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[b].e[0]));
1358                }
1359
1360                // Load the four blocks of quantized values interleaved with each other in chunks of eight - A0,A1,A2,A3
1361                // Loaded as set of 128 bit vectors and repeated into a 256 bit vector
1362                __m256i lhs_mat_0123_0 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs)));
1363                __m256i lhs_mat_01_0 = _mm256_permute2f128_si256(lhs_mat_0123_0, lhs_mat_0123_0, 0);
1364                __m256i lhs_mat_23_0 = _mm256_permute2f128_si256(lhs_mat_0123_0, lhs_mat_0123_0, 17);
1365                __m256i lhs_mat_0123_1 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs + 32)));
1366                __m256i lhs_mat_01_1 = _mm256_permute2f128_si256(lhs_mat_0123_1, lhs_mat_0123_1, 0);
1367                __m256i lhs_mat_23_1 = _mm256_permute2f128_si256(lhs_mat_0123_1, lhs_mat_0123_1, 17);
1368                __m256i lhs_mat_0123_2 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs + 64)));
1369                __m256i lhs_mat_01_2 = _mm256_permute2f128_si256(lhs_mat_0123_2, lhs_mat_0123_2, 0);
1370                __m256i lhs_mat_23_2 = _mm256_permute2f128_si256(lhs_mat_0123_2, lhs_mat_0123_2, 17);
1371                __m256i lhs_mat_0123_3 = _mm256_loadu_si256((const __m256i *)((a_ptr[b].qs + 96)));
1372                __m256i lhs_mat_01_3 = _mm256_permute2f128_si256(lhs_mat_0123_3, lhs_mat_0123_3, 0);
1373                __m256i lhs_mat_23_3 = _mm256_permute2f128_si256(lhs_mat_0123_3, lhs_mat_0123_3, 17);
1374
1375                // Shuffle pattern one - left side input
1376
1377                const __m256i lhs_mat_01_0_sp1 = _mm256_shuffle_epi32(lhs_mat_01_0, 160);  //A0(0-3) A0(0-3) A1(0-3) A1(0-3) A0(0-3) A0(0-3) A1(0-3) A1(0-3)
1378                const __m256i lhs_mat_23_0_sp1 = _mm256_shuffle_epi32(lhs_mat_23_0, 160);  //A2(0-3) A2(0-3) A3(0-3) A3(0-3) A2(0-3) A2(0-3) A3(0-3) A3(0-3)
1379
1380                const __m256i lhs_mat_01_1_sp1 = _mm256_shuffle_epi32(lhs_mat_01_1, 160);  //A0(8-11) A0(8-11) A1(8-11) A1(8-11) A0(8-11) A0(8-11) A1(8-11) A1(8-11)
1381                const __m256i lhs_mat_23_1_sp1 = _mm256_shuffle_epi32(lhs_mat_23_1, 160);  //A2(8-11) A2(8-11) A3(8-11) A3(8-11) A2(8-11) A2(8-11) A3(8-11) A3(8-11)
1382
1383                const __m256i lhs_mat_01_2_sp1 = _mm256_shuffle_epi32(lhs_mat_01_2, 160);  //A0(16-19) A0(16-19) A1(16-19) A1(16-19) A0(16-19) A0(16-19) A1(16-19) A1(16-19)
1384                const __m256i lhs_mat_23_2_sp1 = _mm256_shuffle_epi32(lhs_mat_23_2, 160);  //A2(16-19) A2(16-19) A3(16-19) A3(16-19) A2(16-19) A2(16-19) A3(16-19) A3(16-19)
1385
1386                const __m256i lhs_mat_01_3_sp1 = _mm256_shuffle_epi32(lhs_mat_01_3, 160);  //A0(24-27) A0(24-27) A1(24-27) A1(24-27) A0(24-27) A0(24-27) A1(24-27) A1(24-27)
1387                const __m256i lhs_mat_23_3_sp1 = _mm256_shuffle_epi32(lhs_mat_23_3, 160);  //A2(24-27) A2(24-27) A3(24-27) A3(24-27) A2(24-27) A2(24-27) A3(24-27) A3(24-27)
1388
1389                // Shuffle pattern two - left side input
1390
1391                const __m256i lhs_mat_01_0_sp2 = _mm256_shuffle_epi32(lhs_mat_01_0, 245);  //A0(4-7) A0(4-7) A1(4-7) A1(4-7) A0(4-7) A0(4-7) A1(4-7) A1(4-7)
1392                const __m256i lhs_mat_23_0_sp2 = _mm256_shuffle_epi32(lhs_mat_23_0, 245);  //A2(4-7) A2(4-7) A3(4-7) A3(4-7) A2(4-7) A2(4-7) A3(4-7) A3(4-7)
1393
1394                const __m256i lhs_mat_01_1_sp2 = _mm256_shuffle_epi32(lhs_mat_01_1, 245);  //A0(12-15) A0(12-15) A1(12-15) A1(12-15) A0(12-15) A0(12-15) A1(12-15) A1(12-15)
1395                const __m256i lhs_mat_23_1_sp2 = _mm256_shuffle_epi32(lhs_mat_23_1, 245);  //A2(12-15) A2(12-15) A3(12-15) A3(12-15) A2(12-15) A2(12-15) A3(12-15) A3(12-15)
1396
1397                const __m256i lhs_mat_01_2_sp2 = _mm256_shuffle_epi32(lhs_mat_01_2, 245);  //A0(20-23) A0(20-23) A1(20-23) A1(20-23) A0(20-23) A0(20-23) A1(20-23) A1(20-23)
1398                const __m256i lhs_mat_23_2_sp2 = _mm256_shuffle_epi32(lhs_mat_23_2, 245);  //A2(20-23) A2(20-23) A3(20-23) A3(20-23) A2(20-23) A2(20-23) A3(20-23) A3(20-23)
1399
1400                const __m256i lhs_mat_01_3_sp2 = _mm256_shuffle_epi32(lhs_mat_01_3, 245);  //A0(28-31) A0(28-31) A1(28-31) A1(28-31) A0(28-31) A0(28-31) A1(28-31) A1(28-31)
1401                const __m256i lhs_mat_23_3_sp2 = _mm256_shuffle_epi32(lhs_mat_23_3, 245);  //A2(28-31) A2(28-31) A3(28-31) A3(28-31) A2(28-31) A2(28-31) A3(28-31) A3(28-31)
1402
1403                // 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
1404                // Resembles MMLAs into 2x2 matrices in ARM Version
1405                const __m256i zero = _mm256_setzero_si256();
1406                __m256i iacc_mat_00_sp1 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_01_3_sp1, rhs_mat_0145_3_sp1), lhs_mat_01_2_sp1, rhs_mat_0145_2_sp1), lhs_mat_01_1_sp1, rhs_mat_0145_1_sp1), lhs_mat_01_0_sp1, rhs_mat_0145_0_sp1);
1407                __m256i iacc_mat_01_sp1 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_01_3_sp1, rhs_mat_2367_3_sp1), lhs_mat_01_2_sp1, rhs_mat_2367_2_sp1), lhs_mat_01_1_sp1, rhs_mat_2367_1_sp1), lhs_mat_01_0_sp1, rhs_mat_2367_0_sp1);
1408                __m256i iacc_mat_10_sp1 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_23_3_sp1, rhs_mat_0145_3_sp1), lhs_mat_23_2_sp1, rhs_mat_0145_2_sp1), lhs_mat_23_1_sp1, rhs_mat_0145_1_sp1), lhs_mat_23_0_sp1, rhs_mat_0145_0_sp1);
1409                __m256i iacc_mat_11_sp1 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_23_3_sp1, rhs_mat_2367_3_sp1), lhs_mat_23_2_sp1, rhs_mat_2367_2_sp1), lhs_mat_23_1_sp1, rhs_mat_2367_1_sp1), lhs_mat_23_0_sp1, rhs_mat_2367_0_sp1);
1410                __m256i iacc_mat_00_sp2 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_01_3_sp2, rhs_mat_0145_3_sp2), lhs_mat_01_2_sp2, rhs_mat_0145_2_sp2), lhs_mat_01_1_sp2, rhs_mat_0145_1_sp2), lhs_mat_01_0_sp2, rhs_mat_0145_0_sp2);
1411                __m256i iacc_mat_01_sp2 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_01_3_sp2, rhs_mat_2367_3_sp2), lhs_mat_01_2_sp2, rhs_mat_2367_2_sp2), lhs_mat_01_1_sp2, rhs_mat_2367_1_sp2), lhs_mat_01_0_sp2, rhs_mat_2367_0_sp2);
1412                __m256i iacc_mat_10_sp2 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_23_3_sp2, rhs_mat_0145_3_sp2), lhs_mat_23_2_sp2, rhs_mat_0145_2_sp2), lhs_mat_23_1_sp2, rhs_mat_0145_1_sp2), lhs_mat_23_0_sp2, rhs_mat_0145_0_sp2);
1413                __m256i iacc_mat_11_sp2 = mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(mul_sum_i8_pairs_acc_int32x8(zero, lhs_mat_23_3_sp2, rhs_mat_2367_3_sp2), lhs_mat_23_2_sp2, rhs_mat_2367_2_sp2), lhs_mat_23_1_sp2, rhs_mat_2367_1_sp2), lhs_mat_23_0_sp2, rhs_mat_2367_0_sp2);
1414
1415                // Output of both shuffle patterns are added in order to sum dot product outputs of all 32 values in block
1416                __m256i iacc_mat_00 = _mm256_add_epi32(iacc_mat_00_sp1, iacc_mat_00_sp2);
1417                __m256i iacc_mat_01 = _mm256_add_epi32(iacc_mat_01_sp1, iacc_mat_01_sp2);
1418                __m256i iacc_mat_10 = _mm256_add_epi32(iacc_mat_10_sp1, iacc_mat_10_sp2);
1419                __m256i iacc_mat_11 = _mm256_add_epi32(iacc_mat_11_sp1, iacc_mat_11_sp2);
1420
1421
1422                // Straighten out to make 4 row vectors
1423                __m256i iacc_row_0 = _mm256_blend_epi32(iacc_mat_00, _mm256_shuffle_epi32(iacc_mat_01, 78), 204);
1424                __m256i iacc_row_1 = _mm256_blend_epi32(_mm256_shuffle_epi32(iacc_mat_00, 78), iacc_mat_01, 204);
1425                __m256i iacc_row_2 = _mm256_blend_epi32(iacc_mat_10, _mm256_shuffle_epi32(iacc_mat_11, 78), 204);
1426                __m256i iacc_row_3 = _mm256_blend_epi32(_mm256_shuffle_epi32(iacc_mat_10, 78), iacc_mat_11, 204);
1427
1428                // Load the scale(d) values for all the 4 Q8_0 blocks and repeat it across lanes
1429                const __m256 row_scale_f32 = GGML_F32Cx8_REPEAT_LOAD(a_ptr[b].d, loadMask);
1430
1431                // Multiply with appropriate scales and accumulate
1432                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]);
1433                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]);
1434                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]);
1435                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]);
1436            }
1437
1438            // Store the accumulated values
1439            for (int i = 0; i < 4; i++) {
1440                _mm256_storeu_ps((float *)(s + ((y * 4 + i) * bs + x * 8)), acc_rows[i]);
1441            }
1442        }
1443    }
1444}
1445
1446#endif // defined(__AVX2__) || defined(__AVX512F__)
1447
1448void 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) {
1449#if defined(__AVX2__) || defined(__AVX512F__)
1450    {
1451        // Lookup table to convert signed nibbles to signed bytes
1452        __m256i signextendlut = _mm256_castsi128_si256(_mm_set_epi8(-1, -2, -3, -4, -5, -6, -7, -8, 7, 6, 5, 4, 3, 2, 1, 0));
1453        signextendlut = _mm256_permute2f128_si256(signextendlut, signextendlut, 0);
1454
1455        gemv_q4_b32_8x8_q8_0_lut_avx<block_q4_0x8>(n, s, bs, vx, vy, nr, nc, signextendlut);
1456
1457        return;
1458    }
1459#endif
1460
1461    ggml_gemv_q4_0_8x8_q8_0_generic(n, s, bs, vx, vy, nr, nc);
1462}
1463
1464void 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) {
1465    const int qk = QK_K;
1466    const int nb = n / qk;
1467    const int ncols_interleaved = 8;
1468    const int blocklen = 8;
1469    static const uint32_t kmask1 = 0x3f3f3f3f;
1470    static const uint32_t kmask2 = 0x0f0f0f0f;
1471    static const uint32_t kmask3 = 0x03030303;
1472
1473    assert (n % qk == 0);
1474    assert (nc % ncols_interleaved == 0);
1475
1476    UNUSED(s);
1477    UNUSED(bs);
1478    UNUSED(vx);
1479    UNUSED(vy);
1480    UNUSED(nr);
1481    UNUSED(nc);
1482    UNUSED(nb);
1483    UNUSED(ncols_interleaved);
1484    UNUSED(blocklen);
1485
1486#if defined(__AVX2__)
1487    // Lookup table to convert signed nibbles to signed bytes
1488    __m256i signextendlut = _mm256_castsi128_si256(_mm_set_epi8(-1, -2, -3, -4, -5, -6, -7, -8, 7, 6, 5, 4, 3, 2, 1, 0));
1489    signextendlut = _mm256_permute2f128_si256(signextendlut, signextendlut, 0);
1490    // Shuffle masks to rearrange delta and scale values to multiply with appropriate scales
1491    __m128i deltamask = _mm_set_epi8(15, 14, 7, 6, 13, 12, 5, 4, 11, 10, 3, 2, 9, 8, 1, 0);
1492    __m128i scalemask = _mm_set_epi8(7, 7, 3, 3, 6, 6, 2, 2, 5, 5, 1, 1, 4, 4, 0, 0);
1493    // Permute mask used for easier vector processing at later stages
1494    __m256i finalpermutemask = _mm256_set_epi32(7, 5, 3, 1, 6, 4, 2, 0);
1495
1496    // Mask to extract nibbles from bytes
1497    const __m256i m4b = _mm256_set1_epi8(0x0F);
1498
1499    int64_t b_nb = n / QK_K;
1500
1501    const block_q4_Kx8 * b_ptr_start = (const block_q4_Kx8 *)vx;
1502    const block_q8_K * a_ptr_start = (const block_q8_K *)vy;
1503
1504    // Process Q8_K blocks one by one
1505    for (int64_t y = 0; y < nr; y++) {
1506
1507        // Pointers to LHS blocks of block_q8_K format
1508        const block_q8_K * a_ptr = a_ptr_start + (y * nb);
1509
1510        // Take group of eight interleaved block_q4_K structures at each pass of the loop and perform dot product operation
1511        for (int64_t x = 0; x < nc / 8; x++) {
1512
1513            // Pointers to RHS blocks
1514            const block_q4_Kx8 * b_ptr = b_ptr_start + (x * b_nb);
1515
1516            // Master FP accumulators
1517            __m256 acc_row = _mm256_setzero_ps();
1518            __m256 acc_min_rows = _mm256_setzero_ps();
1519
1520            for (int64_t b = 0; b < nb; b++) {
1521
1522                // Load and convert to FP32 scale from block_q8_K
1523                const __m256 row_scale_f32 = _mm256_set1_ps((a_ptr[b].d));
1524
1525                // Load the scale values for the 8 blocks interleaved in block_q4_Kx8
1526                // col_scale_f32 rearranged so as to multiply with appropriate quants
1527                const __m256 col_scale_f32 = GGML_F32Cx8_REARRANGE_LOAD(b_ptr[b].d, deltamask);
1528                const __m256 col_dmin_f32 = GGML_F32Cx8_LOAD(b_ptr[b].dmin);
1529
1530                __m256i iacc_b = _mm256_setzero_si256();
1531                __m256i iacc_min_b = _mm256_setzero_si256();
1532
1533                const __m256i q8sums = _mm256_loadu_si256((const __m256i * )(a_ptr[b].bsums));
1534                __m256i q8s = _mm256_castsi128_si256(_mm_hadd_epi16(_mm256_castsi256_si128(q8sums), _mm256_extracti128_si256(q8sums, 1)));
1535                q8s = _mm256_permute2f128_si256(q8s, q8s, 0);
1536
1537                // Processes two sub blocks from each Q4_K in each iteration
1538                for (int sb = 0; sb < QK_K / 64; sb++) {
1539
1540                    // Load the eight block_q4_K for two sub blocks quantized values interleaved with each other in chunks of eight - B0,B1 ....B6,B7
1541                    const __m256i rhs_raw_vec_0123_0 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + sb * 256));
1542                    const __m256i rhs_raw_vec_4567_0 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 32 + sb * 256));
1543                    const __m256i rhs_raw_vec_0123_1 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 64 + sb * 256));
1544                    const __m256i rhs_raw_vec_4567_1 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 96 + sb * 256));
1545                    const __m256i rhs_raw_vec_0123_2 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 128 + sb * 256));
1546                    const __m256i rhs_raw_vec_4567_2 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 160 + sb * 256));
1547                    const __m256i rhs_raw_vec_0123_3 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 192 + sb * 256));
1548                    const __m256i rhs_raw_vec_4567_3 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 224 + sb * 256));
1549
1550                    // 4-bit -> 8-bit
1551                    // Values of the first sub block of eight block_q4_K structures for the sb loop
1552                    const __m256i rhs_vec_0123_00 = _mm256_and_si256(rhs_raw_vec_0123_0, m4b);
1553                    const __m256i rhs_vec_4567_00 = _mm256_and_si256(rhs_raw_vec_4567_0, m4b);
1554                    const __m256i rhs_vec_0123_01 = _mm256_and_si256(rhs_raw_vec_0123_1, m4b);
1555                    const __m256i rhs_vec_4567_01 = _mm256_and_si256(rhs_raw_vec_4567_1, m4b);
1556                    const __m256i rhs_vec_0123_02 = _mm256_and_si256(rhs_raw_vec_0123_2, m4b);
1557                    const __m256i rhs_vec_4567_02 = _mm256_and_si256(rhs_raw_vec_4567_2, m4b);
1558                    const __m256i rhs_vec_0123_03 = _mm256_and_si256(rhs_raw_vec_0123_3, m4b);
1559                    const __m256i rhs_vec_4567_03 = _mm256_and_si256(rhs_raw_vec_4567_3, m4b);
1560
1561                    // Values of the second sub block of eight block_q4_K structures when sb = 1
1562                    const __m256i rhs_vec_0123_10 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_0, 4), m4b);
1563                    const __m256i rhs_vec_4567_10 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_0, 4), m4b);
1564                    const __m256i rhs_vec_0123_11 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_1, 4), m4b);
1565                    const __m256i rhs_vec_4567_11 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_1, 4), m4b);
1566                    const __m256i rhs_vec_0123_12 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_2, 4), m4b);
1567                    const __m256i rhs_vec_4567_12 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_2, 4), m4b);
1568                    const __m256i rhs_vec_0123_13 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_0123_3, 4), m4b);
1569                    const __m256i rhs_vec_4567_13 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_vec_4567_3, 4), m4b);
1570
1571                    uint32_t utmp_0[4], utmp_1[4];
1572
1573                    // Scales and Mins of corresponding sub blocks from different Q8_K structures are stored together
1574                    // The below block is for eg to extract first sub block's scales and mins from different Q4_K structures for the sb loop
1575                    memcpy(utmp_0, b_ptr[b].scales + 24 * sb, 12);
1576                    utmp_0[3] = ((utmp_0[2] >> 4) & kmask2) | (((utmp_0[1] >> 6) & kmask3) << 4);
1577                    const uint32_t uaux_0 = utmp_0[1] & kmask1;
1578                    utmp_0[1] = (utmp_0[2] & kmask2) | (((utmp_0[0] >> 6) & kmask3) << 4);
1579                    utmp_0[2] = uaux_0;
1580                    utmp_0[0] &= kmask1;
1581
1582                    // The below block is for eg to extract second sub block's scales and mins from different Q4_K structures for the sb loop
1583                    memcpy(utmp_1, b_ptr[b].scales + 12 + sb * 24, 12);
1584                    utmp_1[3] = ((utmp_1[2] >> 4) & kmask2) | (((utmp_1[1] >> 6) & kmask3) << 4);
1585                    const uint32_t uaux_1 = utmp_1[1] & kmask1;
1586                    utmp_1[1] = (utmp_1[2] & kmask2) | (((utmp_1[0] >> 6) & kmask3) << 4);
1587                    utmp_1[2] = uaux_1;
1588                    utmp_1[0] &= kmask1;
1589
1590                    // Scales of first sub block in the sb loop
1591                    const __m128i mins_and_scales_0 = _mm_set_epi32(utmp_0[3], utmp_0[2], utmp_0[1], utmp_0[0]);
1592                    __m128i scales_rearrange_0 = _mm_shuffle_epi8(mins_and_scales_0, scalemask);
1593                    __m256i scales_0 = _mm256_cvtepu8_epi16(scales_rearrange_0);
1594
1595                    // Scales of second sub block in the sb loop
1596                    __m128i mins_and_scales_1 = _mm_set_epi32(utmp_1[3], utmp_1[2], utmp_1[1], utmp_1[0]);
1597                    __m128i scales_rearrange_1 = _mm_shuffle_epi8(mins_and_scales_1, scalemask);
1598                    __m256i scales_1 = _mm256_cvtepu8_epi16(scales_rearrange_1);
1599
1600                    // Mins of first and second sub block of Q4_K block are arranged side by side
1601                    __m256i mins_01 = _mm256_cvtepu8_epi16(_mm_unpacklo_epi8(_mm_shuffle_epi32(mins_and_scales_0, 78), _mm_shuffle_epi32(mins_and_scales_1, 78)));
1602
1603                    // Load the two sub block values corresponding to sb in block_q8_K in batches of 16 bytes and replicate the same across 256 bit vector
1604                    __m256i lhs_vec_00 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + sb * 64)));
1605                    __m256i lhs_vec_01 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 16 + sb * 64)));
1606                    __m256i lhs_vec_10 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 32 + sb * 64)));
1607                    __m256i lhs_vec_11 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 48 + sb * 64)));
1608
1609                    lhs_vec_00 = _mm256_permute2f128_si256(lhs_vec_00, lhs_vec_00, 0);
1610                    lhs_vec_01 = _mm256_permute2f128_si256(lhs_vec_01, lhs_vec_01, 0);
1611                    lhs_vec_10 = _mm256_permute2f128_si256(lhs_vec_10, lhs_vec_10, 0);
1612                    lhs_vec_11 = _mm256_permute2f128_si256(lhs_vec_11, lhs_vec_11, 0);
1613
1614                    // Dot product done within 32 bit lanes and accumulated in the same vector
1615                    // First done for first sub block and then for second sub block in each sb
1616                    // 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)
1617                    // 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)
1618                    // ...........................................................................
1619                    // B0(28-31) B4(28-31) B1(28-31) B5(28-31) B2(28-31) B6(28-31) B3(28-31) B7(28-31) with A0(28-31)
1620
1621
1622                    __m256i iacc_0 = _mm256_setzero_si256();
1623                    __m256i iacc_1 = _mm256_setzero_si256();
1624
1625                    iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_00 ,_mm256_shuffle_epi32(rhs_vec_4567_00, 177), 170), _mm256_shuffle_epi32(lhs_vec_00, 0)));
1626                    iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_00, 177) ,rhs_vec_4567_00, 170), _mm256_shuffle_epi32(lhs_vec_00, 85)));
1627
1628                    iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_01 ,_mm256_shuffle_epi32(rhs_vec_4567_01, 177), 170), _mm256_shuffle_epi32(lhs_vec_00, 170)));
1629                    iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_01, 177) ,rhs_vec_4567_01, 170), _mm256_shuffle_epi32(lhs_vec_00, 255)));
1630
1631                    iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_02 ,_mm256_shuffle_epi32(rhs_vec_4567_02, 177), 170), _mm256_shuffle_epi32(lhs_vec_01, 0)));
1632                    iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_02, 177) ,rhs_vec_4567_02, 170), _mm256_shuffle_epi32(lhs_vec_01, 85)));
1633
1634                    iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_03 ,_mm256_shuffle_epi32(rhs_vec_4567_03, 177), 170), _mm256_shuffle_epi32(lhs_vec_01, 170)));
1635                    iacc_0 = _mm256_add_epi16(iacc_0, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_03, 177) ,rhs_vec_4567_03, 170), _mm256_shuffle_epi32(lhs_vec_01, 255)));
1636
1637                    iacc_0 = _mm256_madd_epi16(iacc_0, scales_0);
1638
1639                    iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_10 ,_mm256_shuffle_epi32(rhs_vec_4567_10, 177), 170), _mm256_shuffle_epi32(lhs_vec_10, 0)));
1640                    iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_10, 177) ,rhs_vec_4567_10, 170), _mm256_shuffle_epi32(lhs_vec_10, 85)));
1641
1642                    iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_11 ,_mm256_shuffle_epi32(rhs_vec_4567_11, 177), 170), _mm256_shuffle_epi32(lhs_vec_10, 170)));
1643                    iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_11, 177) ,rhs_vec_4567_11, 170), _mm256_shuffle_epi32(lhs_vec_10, 255)));
1644
1645                    iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_12 ,_mm256_shuffle_epi32(rhs_vec_4567_12, 177), 170), _mm256_shuffle_epi32(lhs_vec_11, 0)));
1646                    iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_12, 177) ,rhs_vec_4567_12, 170), _mm256_shuffle_epi32(lhs_vec_11, 85)));
1647
1648                    iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(rhs_vec_0123_13 ,_mm256_shuffle_epi32(rhs_vec_4567_13, 177), 170), _mm256_shuffle_epi32(lhs_vec_11, 170)));
1649                    iacc_1 = _mm256_add_epi16(iacc_1, _mm256_maddubs_epi16(_mm256_blend_epi32(_mm256_shuffle_epi32(rhs_vec_0123_13, 177) ,rhs_vec_4567_13, 170), _mm256_shuffle_epi32(lhs_vec_11, 255)));
1650
1651                    iacc_1 = _mm256_madd_epi16(iacc_1, scales_1);
1652
1653                    // Accumulate the iacc value for one sb
1654                    __m256i iacc_sb = _mm256_add_epi32(iacc_0, iacc_1);
1655
1656                    // Broadcast the bsums of the two sub blocks  of the iteration of Q8_K across the vector
1657                    // Multiply-Add with corresponding mins of Q4_Kx8 with bsums
1658                    __m256i q8s_sb = _mm256_shuffle_epi32(q8s, 0);
1659                    __m256i iacc_min_sb = _mm256_madd_epi16(q8s_sb, mins_01);
1660                    q8s = _mm256_bsrli_epi128(q8s, 4);
1661
1662                    // Accumulate for the complete block
1663                    iacc_b = _mm256_add_epi32(iacc_b, iacc_sb);
1664                    iacc_min_b = _mm256_add_epi32(iacc_min_b, iacc_min_sb);
1665                }
1666
1667                // Multiply-Add with scale values for the complete super block
1668                acc_row = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_b), _mm256_mul_ps(col_scale_f32, row_scale_f32), acc_row);
1669                acc_min_rows = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_min_b), _mm256_mul_ps(col_dmin_f32, row_scale_f32), acc_min_rows);
1670
1671            }
1672
1673            // Accumulated output values permuted so as to be stored in appropriate order post accumulation
1674            acc_row = _mm256_permutevar8x32_ps(acc_row, finalpermutemask);
1675            _mm256_storeu_ps(s + (y * nr + x * 8), _mm256_sub_ps(acc_row, acc_min_rows));
1676        }
1677    }
1678
1679#else
1680    UNUSED(kmask1);
1681    UNUSED(kmask2);
1682    UNUSED(kmask3);
1683    ggml_gemv_q4_K_8x8_q8_K_generic(n, s, bs, vx, vy, nr, nc);
1684#endif
1685}
1686
1687void ggml_gemv_iq4_nl_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
1688#if defined(__AVX2__)
1689    __m256i signextendlut = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i*)kvalues_iq4nl));
1690    signextendlut = _mm256_permute2f128_si256(signextendlut, signextendlut, 0);
1691
1692    gemv_q4_b32_8x8_q8_0_lut_avx<block_iq4_nlx8>(n, s, bs, vx, vy, nr, nc, signextendlut);
1693
1694    return;
1695#endif
1696
1697    ggml_gemv_iq4_nl_8x8_q8_0_generic(n, s, bs, vx, vy, nr, nc);
1698}
1699
1700void ggml_gemv_mxfp4_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) {
1701#if defined(__AVX2__)
1702    __m256i signextendlut = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i*)kvalues_mxfp4));
1703    signextendlut = _mm256_permute2f128_si256(signextendlut, signextendlut, 0);
1704
1705    gemv_q4_b32_8x8_q8_0_lut_avx<block_mxfp4x8>(n, s, bs, vx, vy, nr, nc, signextendlut);
1706
1707    return;
1708#endif
1709
1710    ggml_gemv_mxfp4_8x8_q8_0_generic(n, s, bs, vx, vy, nr, nc);
1711}
1712
1713void 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) {
1714    const int qk = QK_K;
1715    const int nb = n / qk;
1716    const int ncols_interleaved = 8;
1717    const int blocklen = 8;
1718
1719    assert (n % qk == 0);
1720    assert (nc % ncols_interleaved == 0);
1721
1722    UNUSED(s);
1723    UNUSED(bs);
1724    UNUSED(vx);
1725    UNUSED(vy);
1726    UNUSED(nr);
1727    UNUSED(nc);
1728    UNUSED(nb);
1729    UNUSED(ncols_interleaved);
1730    UNUSED(blocklen);
1731
1732#if defined(__AVX2__)
1733    // Lookup table to convert signed nibbles to signed bytes
1734    __m256i signextendlut = _mm256_castsi128_si256(_mm_set_epi8(-1, -2, -3, -4, -5, -6, -7, -8, 7, 6, 5, 4, 3, 2, 1, 0));
1735    signextendlut = _mm256_permute2f128_si256(signextendlut, signextendlut, 0);
1736    // Shuffle masks to rearrange delta values to multiply with appropriate scales
1737    __m128i deltamask = _mm_set_epi8(15, 14, 7, 6, 13, 12, 5, 4, 11, 10, 3, 2, 9, 8, 1, 0);
1738    // Permute mask used for easier vector processing at later stages
1739    __m256i finalpermutemask = _mm256_set_epi32(7, 5, 3, 1, 6, 4, 2, 0);
1740
1741    const __m256i m3b = _mm256_set1_epi8(3);
1742    const __m128i m4b_sse = _mm_set1_epi8(0xF);
1743
1744    //Mask to get appropriate scales
1745    __m128i scalemask1 = _mm_set_epi8(14,14,6,6,12,12,4,4,10,10,2,2,8,8,0,0);
1746    __m128i scalemask2 = _mm_set_epi8(15,15,7,7,13,13,5,5,11,11,3,3,9,9,1,1);
1747
1748    int64_t b_nb = n / QK_K;
1749
1750    const block_q2_Kx8 * b_ptr_start = (const block_q2_Kx8 *)vx;
1751    const block_q8_K * a_ptr_start = (const block_q8_K *)vy;
1752
1753    // Process Q8_K blocks one by one
1754    for (int64_t y = 0; y < nr; y++) {
1755
1756        // Pointers to LHS blocks of block_q8_K format
1757        const block_q8_K * a_ptr = a_ptr_start + (y * nb);
1758
1759        // Take group of eight interleaved block_q2_K structures at each pass of the loop and perform dot product operation
1760        for(int64_t x = 0; x < nc / 8; x++) {
1761
1762            // Pointers to RHS blocks
1763            const block_q2_Kx8 * b_ptr = b_ptr_start + (x * b_nb);
1764
1765            // Master FP accumulators
1766            __m256 acc_row = _mm256_setzero_ps();
1767            __m256 acc_min_rows = _mm256_setzero_ps();
1768
1769            for (int64_t b = 0; b < nb; b++) {
1770
1771                // Load and convert to FP32 delta from block_q8_K
1772                const __m256 row_scale_f32 = _mm256_set1_ps((a_ptr[b].d));
1773
1774                // Load the delta values for the 8 blocks interleaved in block_q2_Kx8
1775                // col_scale_f32 rearranged so as to multiply with appropriate quants
1776                const __m256 col_scale_f32 = GGML_F32Cx8_REARRANGE_LOAD(b_ptr[b].d, deltamask);
1777                const __m256 col_dmin_f32 = GGML_F32Cx8_LOAD(b_ptr[b].dmin);
1778
1779                __m256i iacc_b = _mm256_setzero_si256();
1780                __m256i iacc_min_b = _mm256_setzero_si256();
1781
1782                // Processes eight sub blocks from each Q2_K in each iteration
1783                for(int sb = 0; sb < QK_K / 128; sb++) {
1784
1785                    // Load the eight block_q2_K for eight sub blocks quantized values interleaved with each other in chunks of eight - B0,B1 ....B6,B7
1786                    const __m256i rhs_raw_vec_0123_0 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + sb * 256));
1787                    const __m256i rhs_raw_vec_4567_0 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 32 + sb * 256));
1788                    const __m256i rhs_raw_vec_0123_1 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 64 + sb * 256));
1789                    const __m256i rhs_raw_vec_4567_1 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 96 + sb * 256));
1790                    const __m256i rhs_raw_vec_0123_2 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 128 + sb * 256));
1791                    const __m256i rhs_raw_vec_4567_2 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 160 + sb * 256));
1792                    const __m256i rhs_raw_vec_0123_3 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 192 + sb * 256));
1793                    const __m256i rhs_raw_vec_4567_3 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 224 + sb * 256));
1794
1795                    // 2-bit -> 8-bit
1796                    // Values of the 0th,2nd,4th,6th sub blocks of eight block_q2_K structures for the sb loop
1797                    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)
1798                    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)
1799                    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)
1800                    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)
1801
1802                    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)
1803                    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)
1804                    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)
1805                    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)
1806
1807                    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)
1808                    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)
1809                    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)
1810                    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)
1811
1812                    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)
1813                    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)
1814                    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)
1815                    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)
1816
1817                    // Values of the 1st,3rd,5th,7th sub blocks of eight block_q2_K structures for the sb loop
1818                    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)
1819                    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)
1820                    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)
1821                    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)
1822
1823                    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)
1824                    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)
1825                    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)
1826                    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)
1827
1828                    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)
1829                    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)
1830                    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)
1831                    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)
1832
1833                    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)
1834                    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)
1835                    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)
1836                    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)
1837
1838                    //Scales and Mins of corresponding sub blocks from different Q2_K structures are stored together
1839                    //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
1840
1841                    const __m128i mins_and_scales_01 = _mm_loadu_si128((const __m128i *)(b_ptr[b].scales + sb * 64));
1842                    const __m128i mins_and_scales_23 = _mm_loadu_si128((const __m128i *)(b_ptr[b].scales + 16 + sb * 64));
1843                    const __m128i mins_and_scales_45 = _mm_loadu_si128((const __m128i *)(b_ptr[b].scales + 32 + sb * 64));
1844                    const __m128i mins_and_scales_67 = _mm_loadu_si128((const __m128i *)(b_ptr[b].scales + 48 + sb * 64));
1845
1846                    // Extract scales which is lower half from mins_and_scales
1847                    const __m128i scales_01 = _mm_and_si128(mins_and_scales_01, m4b_sse);
1848                    const __m128i scales_23 = _mm_and_si128(mins_and_scales_23, m4b_sse);
1849                    const __m128i scales_45 = _mm_and_si128(mins_and_scales_45, m4b_sse);
1850                    const __m128i scales_67 = _mm_and_si128(mins_and_scales_67, m4b_sse);
1851
1852                    // Extract mins which is upper half from mins_and_scales
1853                    const __m256i mins_01 = _mm256_cvtepu8_epi16(_mm_and_si128(_mm_srli_epi16(mins_and_scales_01, 4), m4b_sse));
1854                    const __m256i mins_23 = _mm256_cvtepu8_epi16(_mm_and_si128(_mm_srli_epi16(mins_and_scales_23, 4), m4b_sse));
1855                    const __m256i mins_45 = _mm256_cvtepu8_epi16(_mm_and_si128(_mm_srli_epi16(mins_and_scales_45, 4), m4b_sse));
1856                    const __m256i mins_67 = _mm256_cvtepu8_epi16(_mm_and_si128(_mm_srli_epi16(mins_and_scales_67, 4), m4b_sse));
1857
1858                    // Scales of sub blocks in the sb loop
1859                    // Scales of the 0th sub block from each super block
1860                    __m128i scales_rearrange_0 = _mm_shuffle_epi8(scales_01, scalemask1);
1861                    __m256i scales_0 = _mm256_cvtepu8_epi16(scales_rearrange_0);
1862
1863                    // Scales of the 1st sub block from each super block
1864                    __m128i scales_rearrange_1 = _mm_shuffle_epi8(scales_01, scalemask2);
1865                    __m256i scales_1 = _mm256_cvtepu8_epi16(scales_rearrange_1);
1866
1867                    // Scales of the 2nd sub block from each super block
1868                    __m128i scales_rearrange_2 = _mm_shuffle_epi8(scales_23, scalemask1);
1869                    __m256i scales_2 = _mm256_cvtepu8_epi16(scales_rearrange_2);
1870
1871                    // Scales of the 3rd sub block from each super block
1872                    __m128i scales_rearrange_3 = _mm_shuffle_epi8(scales_23, scalemask2);
1873                    __m256i scales_3 = _mm256_cvtepu8_epi16(scales_rearrange_3);
1874
1875                    // Scales of the 4th sub block from each super block
1876                    __m128i scales_rearrange_4 = _mm_shuffle_epi8(scales_45, scalemask1);
1877                    __m256i scales_4 = _mm256_cvtepu8_epi16(scales_rearrange_4);
1878
1879                    // Scales of the 5th sub block from each super block
1880                    __m128i scales_rearrange_5 = _mm_shuffle_epi8(scales_45, scalemask2);
1881                    __m256i scales_5 = _mm256_cvtepu8_epi16(scales_rearrange_5);
1882
1883                    // Scales of the 6th sub block from each super block
1884                    __m128i scales_rearrange_6 = _mm_shuffle_epi8(scales_67, scalemask1);
1885                    __m256i scales_6 = _mm256_cvtepu8_epi16(scales_rearrange_6);
1886
1887                    // Scales of the 7th sub block from each super block
1888                    __m128i scales_rearrange_7 = _mm_shuffle_epi8(scales_67, scalemask2);
1889                    __m256i scales_7 = _mm256_cvtepu8_epi16(scales_rearrange_7);
1890
1891                    // 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
1892                    __m256i lhs_vec_0 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + sb * 128)));
1893                    __m256i lhs_vec_1 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 16 + sb * 128)));
1894                    __m256i lhs_vec_2 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 32 + sb * 128)));
1895                    __m256i lhs_vec_3 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 48 + sb * 128)));
1896                    __m256i lhs_vec_4 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 64 + sb * 128)));
1897                    __m256i lhs_vec_5 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 80 + sb * 128)));
1898                    __m256i lhs_vec_6 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 96 + sb * 128)));
1899                    __m256i lhs_vec_7 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)(a_ptr[b].qs + 112 + sb * 128)));
1900
1901                    lhs_vec_0 = _mm256_permute2f128_si256(lhs_vec_0, lhs_vec_0, 0);
1902                    lhs_vec_1 = _mm256_permute2f128_si256(lhs_vec_1, lhs_vec_1, 0);
1903                    lhs_vec_2 = _mm256_permute2f128_si256(lhs_vec_2, lhs_vec_2, 0);
1904                    lhs_vec_3 = _mm256_permute2f128_si256(lhs_vec_3, lhs_vec_3, 0);
1905                    lhs_vec_4 = _mm256_permute2f128_si256(lhs_vec_4, lhs_vec_4, 0);
1906                    lhs_vec_5 = _mm256_permute2f128_si256(lhs_vec_5, lhs_vec_5, 0);
1907                    lhs_vec_6 = _mm256_permute2f128_si256(lhs_vec_6, lhs_vec_6, 0);
1908                    lhs_vec_7 = _mm256_permute2f128_si256(lhs_vec_7, lhs_vec_7, 0);
1909
1910                    __m256i iacc_0 = _mm256_setzero_si256();
1911                    __m256i iacc_1 = _mm256_setzero_si256();
1912                    __m256i iacc_2 = _mm256_setzero_si256();
1913                    __m256i iacc_3 = _mm256_setzero_si256();
1914                    __m256i iacc_4 = _mm256_setzero_si256();
1915                    __m256i iacc_5 = _mm256_setzero_si256();
1916                    __m256i iacc_6 = _mm256_setzero_si256();
1917                    __m256i iacc_7 = _mm256_setzero_si256();
1918
1919                    // Dot product done within 32 bit lanes and accumulated in the same vector
1920                    // 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)
1921                    // 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)
1922                    // 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)
1923                    // 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)
1924
1925                    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)));
1926                    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)));
1927
1928                    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)));
1929                    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)));
1930
1931                    iacc_0 = _mm256_madd_epi16(iacc_0, scales_0);
1932
1933                    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)));
1934                    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)));
1935
1936                    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)));
1937                    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)));
1938
1939                    iacc_1 = _mm256_madd_epi16(iacc_1, scales_1);
1940
1941                    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)));
1942                    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)));
1943
1944                    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)));
1945                    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)));
1946
1947                    iacc_2 = _mm256_madd_epi16(iacc_2, scales_2);
1948
1949                    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)));
1950                    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)));
1951
1952                    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)));
1953                    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)));
1954
1955                    iacc_3 = _mm256_madd_epi16(iacc_3, scales_3);
1956
1957                    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)));
1958                    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)));
1959
1960                    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)));
1961                    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)));
1962
1963                    iacc_4 = _mm256_madd_epi16(iacc_4, scales_4);
1964
1965                    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)));
1966                    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)));
1967
1968                    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)));
1969                    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)));
1970
1971                    iacc_5 = _mm256_madd_epi16(iacc_5, scales_5);
1972
1973                    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)));
1974                    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)));
1975
1976                    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)));
1977                    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)));
1978
1979                    iacc_6 = _mm256_madd_epi16(iacc_6, scales_6);
1980
1981                    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)));
1982                    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)));
1983
1984                    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)));
1985                    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)));
1986
1987                    iacc_7 = _mm256_madd_epi16(iacc_7, scales_7);
1988
1989                    // Accumulate the iacc value for one sb
1990                    __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)));
1991
1992                    __m128i q8sums = _mm_loadu_si128((const __m128i *)(a_ptr[b].bsums + sb * 8));
1993                    __m256i q8s = _mm256_castsi128_si256(q8sums);
1994                    q8s= _mm256_permute2f128_si256(q8s, q8s, 0);
1995
1996                    // Broadcast the bsums of the two corresponding subblocks of q8_k
1997                    // Multiply-Add with corresponding mins of Q2_Kx8 with bsums
1998                    __m256i iacc_min_sb_01 = _mm256_madd_epi16(_mm256_shuffle_epi32(q8s, 0), mins_01);
1999                    __m256i iacc_min_sb_23 = _mm256_madd_epi16(_mm256_shuffle_epi32(q8s, 85), mins_23);
2000                    __m256i iacc_min_sb_45 = _mm256_madd_epi16(_mm256_shuffle_epi32(q8s, 170), mins_45);
2001                    __m256i iacc_min_sb_67 = _mm256_madd_epi16(_mm256_shuffle_epi32(q8s, 255), mins_67);
2002
2003                    __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));
2004
2005                    // Accumulate for the complete block
2006                    iacc_b = _mm256_add_epi32(iacc_b, iacc_sb);
2007                    iacc_min_b = _mm256_add_epi32(iacc_min_b, iacc_min_sb);
2008                }
2009
2010                //Multiply-Add with scale values for complete super block
2011                acc_row = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_b), _mm256_mul_ps(col_scale_f32, row_scale_f32), acc_row);
2012                acc_min_rows = _mm256_fmadd_ps(_mm256_cvtepi32_ps(iacc_min_b), _mm256_mul_ps(col_dmin_f32, row_scale_f32), acc_min_rows);
2013            }
2014            // Accumulated output values permuted so as to be stored in appropriate order post accumulation
2015            acc_row = _mm256_permutevar8x32_ps(acc_row, finalpermutemask);
2016            _mm256_storeu_ps(s + (y * nr + x * 8), _mm256_sub_ps(acc_row, acc_min_rows));
2017        }
2018    }
2019#else
2020
2021    ggml_gemv_q2_K_8x8_q8_K_generic(n, s, bs, vx, vy, nr, nc);
2022
2023#endif
2024}
2025
2026void 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) {
2027#if defined(__AVX2__) || defined(__AVX512F__)
2028    {
2029        // Lookup table to convert signed nibbles to signed bytes
2030        __m256i signextendlut = _mm256_castsi128_si256(_mm_set_epi8(-1, -2, -3, -4, -5, -6, -7, -8, 7, 6, 5, 4, 3, 2, 1, 0));
2031        signextendlut = _mm256_permute2f128_si256(signextendlut, signextendlut, 0);
2032
2033        gemm_q4_b32_8x8_q8_0_lut_avx<block_q4_0x8>(n, s, bs, vx, vy, nr, nc, signextendlut);
2034
2035        return;
2036    }
2037#endif // defined(__AVX2__) || defined(__AVX512F__)
2038
2039    ggml_gemm_q4_0_8x8_q8_0_generic(n, s, bs, vx, vy, nr, nc);
2040}
2041
2042void 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) {
2043    const int qk = QK_K;
2044    const int nb = n / qk;
2045    const int ncols_interleaved = 8;
2046    const int blocklen = 8;
2047    static const uint32_t kmask1 = 0x3f3f3f3f;
2048    static const uint32_t kmask2 = 0x0f0f0f0f;
2049    static const uint32_t kmask3 = 0x03030303;
2050
2051    assert (n % qk == 0);
2052    assert (nr % 4 == 0);
2053    assert (nc % ncols_interleaved == 0);
2054
2055    UNUSED(s);
2056    UNUSED(bs);
2057    UNUSED(vx);
2058    UNUSED(vy);
2059    UNUSED(nr);
2060    UNUSED(nc);
2061    UNUSED(nb);
2062    UNUSED(ncols_interleaved);
2063    UNUSED(blocklen);
2064
2065#if defined(__AVX2__) || defined(__AVX512F__)
2066    const block_q4_Kx8 * b_ptr_start = (const block_q4_Kx8 * ) vx;
2067    const block_q8_Kx4 * a_ptr_start = (const block_q8_Kx4 * ) vy;
2068    int64_t b_nb = n / QK_K;
2069    int64_t y = 0;
2070
2071    // Mask to mask out nibbles from packed bytes
2072    const __m256i m4b = _mm256_set1_epi8(0x0F);
2073    // Permute mask used for easier vector processing at later stages
2074    __m256i requiredOrder = _mm256_set_epi32(3, 2, 1, 0, 7, 6, 5, 4);
2075    int64_t xstart = 0;
2076    int anr = nr - nr % 16;; // Used to align nr with boundary of 16
2077#if defined(__AVX512BW__) && defined(__AVX512DQ__)
2078    int anc = nc - nc % 16; // Used to align nc with boundary of 16
2079    // Mask to mask out nibbles from packed bytes expanded to 512 bit length
2080    const __m512i m4bexpanded = _mm512_set1_epi8(0x0F);
2081    //Take group of four block_q8_Kx4 structures at each pass of the loop and perform dot product operation
2082    for (; y < anr / 4; y += 4) {
2083
2084        const block_q8_Kx4 * a_ptrs[4];
2085
2086        a_ptrs[0] = a_ptr_start + (y * nb);
2087        for (int i = 0; i < 3; ++i) {
2088            a_ptrs[i + 1] = a_ptrs[i] + nb;
2089        }
2090
2091        // Take group of eight block_q4_kx8 structures at each pass of the loop and perform dot product operation
2092        for (int64_t x = 0; x < anc / 8; x += 2) {
2093
2094            const block_q4_Kx8 * b_ptr_0 = b_ptr_start + ((x) * b_nb);
2095            const block_q4_Kx8 * b_ptr_1 = b_ptr_start + ((x + 1) * b_nb);
2096
2097            // Master FP accumulators
2098            __m512 acc_rows[16];
2099            for (int i = 0; i < 16; i++) {
2100                acc_rows[i] = _mm512_setzero_ps();
2101            }
2102
2103            __m512 acc_min_rows[16];
2104            for (int i = 0; i < 16; i++) {
2105                acc_min_rows[i] = _mm512_setzero_ps();
2106            }
2107
2108            // For super block
2109            for (int64_t b = 0; b < nb; b++) {
2110                // Scale values - Load the sixteen scale values from two block_q4_kx8 structures
2111                const __m512 col_scale_f32 = GGML_F32Cx8x2_LOAD(b_ptr_0[b].d, b_ptr_1[b].d);
2112
2113                // dmin values - Load the sixteen dmin values from two block_q4_kx8 structures
2114                const __m512 col_dmin_f32 = GGML_F32Cx8x2_LOAD(b_ptr_0[b].dmin, b_ptr_1[b].dmin);
2115
2116                // Loop to iterate over the eight sub blocks of a super block - two sub blocks are processed per iteration
2117                for (int sb = 0; sb < QK_K / 64; sb++) {
2118
2119                    const __m256i rhs_raw_mat_0123_0 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + sb * 256));
2120                    const __m256i rhs_raw_mat_4567_0 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 32 + sb * 256));
2121                    const __m256i rhs_raw_mat_0123_1 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 64 + sb * 256));
2122                    const __m256i rhs_raw_mat_4567_1 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 96 + sb * 256));
2123                    const __m256i rhs_raw_mat_0123_2 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 128 + sb * 256));
2124                    const __m256i rhs_raw_mat_4567_2 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 160 + sb * 256));
2125                    const __m256i rhs_raw_mat_0123_3 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 192 + sb * 256));
2126                    const __m256i rhs_raw_mat_4567_3 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 224 + sb * 256));
2127
2128                    const __m256i rhs_raw_mat_89AB_0 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + sb * 256));
2129                    const __m256i rhs_raw_mat_CDEF_0 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 32 + sb * 256));
2130                    const __m256i rhs_raw_mat_89AB_1 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 64 + sb * 256));
2131                    const __m256i rhs_raw_mat_CDEF_1 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 96 + sb * 256));
2132                    const __m256i rhs_raw_mat_89AB_2 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 128 + sb * 256));
2133                    const __m256i rhs_raw_mat_CDEF_2 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 160 + sb * 256));
2134                    const __m256i rhs_raw_mat_89AB_3 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 192 + sb * 256));
2135                    const __m256i rhs_raw_mat_CDEF_3 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 224 + sb * 256));
2136
2137                    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);
2138                    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);
2139                    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);
2140                    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);
2141                    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);
2142                    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);
2143                    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);
2144                    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);
2145
2146                    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);
2147                    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);
2148                    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);
2149                    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);
2150                    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);
2151                    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);
2152                    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);
2153                    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);
2154
2155                    const __m512i rhs_raw_mat_014589CD_0 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_0), rhs_raw_mat_89CD_0, 1);
2156                    const __m512i rhs_raw_mat_2367ABEF_0 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_0), rhs_raw_mat_ABEF_0, 1);
2157                    const __m512i rhs_raw_mat_014589CD_1 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_1), rhs_raw_mat_89CD_1, 1);
2158                    const __m512i rhs_raw_mat_2367ABEF_1 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_1), rhs_raw_mat_ABEF_1, 1);
2159
2160                    const __m512i rhs_raw_mat_014589CD_2 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_2), rhs_raw_mat_89CD_2, 1);
2161                    const __m512i rhs_raw_mat_2367ABEF_2 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_2), rhs_raw_mat_ABEF_2, 1);
2162                    const __m512i rhs_raw_mat_014589CD_3 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_3), rhs_raw_mat_89CD_3, 1);
2163                    const __m512i rhs_raw_mat_2367ABEF_3 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_3), rhs_raw_mat_ABEF_3, 1);
2164
2165                    //4-bit -> 8-bit
2166                    const __m512i rhs_mat_014589CD_00 = _mm512_and_si512(rhs_raw_mat_014589CD_0, m4bexpanded); //B00(0-7) B01(0-7) B04(0-7) B05(0-7) B08(0-7) B09(0-7) B0C(0-7) B0D(0-7)
2167                    const __m512i rhs_mat_2367ABEF_00 = _mm512_and_si512(rhs_raw_mat_2367ABEF_0, m4bexpanded); //B02(0-7) B03(0-7) B06(0-7) B07(0-7) B0A(0-7) B0B(0-7) B0E(0-7) B0F(0-7)
2168                    const __m512i rhs_mat_014589CD_01 = _mm512_and_si512(rhs_raw_mat_014589CD_1, m4bexpanded); //B00(8-15) B01(8-15) B04(8-15) B05(8-15) B08(8-15) B09(8-15) B0C(8-15) B0D(8-15)
2169                    const __m512i rhs_mat_2367ABEF_01 = _mm512_and_si512(rhs_raw_mat_2367ABEF_1, m4bexpanded); //B02(8-15) B03(8-15) B06(8-15) B07(8-15) B0A(8-15) B0B(8-15) B0E(8-15) B0F(8-15)
2170
2171                    const __m512i rhs_mat_014589CD_02 = _mm512_and_si512(rhs_raw_mat_014589CD_2, m4bexpanded); //B00(16-23) B01(16-23) B04(16-23) B05(16-23) B08(16-23) B09(16-23) B0C(16-23) B0D(16-23)
2172                    const __m512i rhs_mat_2367ABEF_02 = _mm512_and_si512(rhs_raw_mat_2367ABEF_2, m4bexpanded); //B02(16-23) B03(16-23) B06(16-23) B07(16-23) B0A(16-23) B0B(16-23) B0E(16-23) B0F(16-23)
2173                    const __m512i rhs_mat_014589CD_03 = _mm512_and_si512(rhs_raw_mat_014589CD_3, m4bexpanded); //B00(24-31) B01(24-31) B04(24-31) B05(24-31) B08(24-31) B09(24-31) B0C(24-31) B0D(24-31)
2174                    const __m512i rhs_mat_2367ABEF_03 = _mm512_and_si512(rhs_raw_mat_2367ABEF_3, m4bexpanded); //B02(24-31) B03(24-31) B06(24-31) B07(24-31) B0A(24-31) B0B(24-31) B0E(24-31) B0F(24-31)
2175
2176                    const __m512i rhs_mat_014589CD_10 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_0, 4), m4bexpanded); //B10(0-7) B11(0-7) B14(0-7) B15(0-7) B18(0-7) B19(0-7) B1C(0-7) B1D(0-7)
2177                    const __m512i rhs_mat_2367ABEF_10 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_0, 4), m4bexpanded); //B12(0-7) B13(0-7) B16(0-7) B17(0-7) B1A(0-7) B1B(0-7) B1E(0-7) B1F(0-7)
2178                    const __m512i rhs_mat_014589CD_11 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_1, 4), m4bexpanded); //B10(8-15) B11(8-15) B14(8-15) B15(8-15) B18(8-15) B19(8-15) B1C(8-15) B1D(8-15)
2179                    const __m512i rhs_mat_2367ABEF_11 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_1, 4), m4bexpanded); //B12(8-15) B13(8-15) B16(8-15) B17(8-15) B1A(8-15) B1B(8-15) B1E(8-15) B1F(8-15)
2180
2181                    const __m512i rhs_mat_014589CD_12 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_2, 4), m4bexpanded); //B10(16-23) B11(16-23) B14(16-23) B15(16-23) B18(16-23) B19(16-23) B1C(16-23) B1D(16-23)
2182                    const __m512i rhs_mat_2367ABEF_12 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_2, 4), m4bexpanded); //B12(16-23) B13(16-23) B16(16-23) B17(16-23) B1A(16-23) B1B(16-23) B1E(16-23) B1F(16-23)
2183                    const __m512i rhs_mat_014589CD_13 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_3, 4), m4bexpanded); //B10(24-31) B11(24-31) B14(24-31) B15(24-31) B18(24-31) B19(24-31) B1C(24-31) B1D(24-31)
2184                    const __m512i rhs_mat_2367ABEF_13 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_3, 4), m4bexpanded); //B12(24-31) B13(24-31) B16(24-31) B17(24-31) B1A(24-31) B1B(24-31) B1E(24-31) B1F(24-31)
2185
2186                    // Shuffle pattern one - right side input
2187                    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)
2188                    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)
2189                    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)
2190                    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)
2191                    const __m512i rhs_mat_014589CD_02_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_02, (_MM_PERM_ENUM)136); //B00(16-19) B01(16-19) B00(16-19) B01(16-19) B04(16-19) B05(16-19) B04(16-19) B05(16-19) B08(16-19) B09(16-19) B08(16-19) B09(16-19) B0C(16-19) B0D(16-19) B0C(16-19) B0D(16-19)
2192                    const __m512i rhs_mat_2367ABEF_02_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_02, (_MM_PERM_ENUM)136); //B02(16-19) B03(16-19) B02(16-19) B03(16-19) B06(16-19) B07(16-19) B06(16-19) B07(16-19) B0A(16-19) B0B(16-19) B0A(16-19) B0B(16-19) B0E(16-19) B0F(16-19) B0E(16-19) B0F(16-19)
2193                    const __m512i rhs_mat_014589CD_03_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_03, (_MM_PERM_ENUM)136); //B00(24-27) B01(24-27) B00(24-27) B01(24-27) B04(24-27) B05(24-27) B04(24-27) B05(24-27) B08(24-27) B09(24-27) B08(24-27) B09(24-27) B0C(24-27) B0D(24-27) B0C(24-27) B0D(24-27)
2194                    const __m512i rhs_mat_2367ABEF_03_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_03, (_MM_PERM_ENUM)136); //B02(24-27) B03(24-27) B02(24-27) B03(24-27) B06(24-27) B07(24-27) B06(24-27) B07(24-27) B0A(24-27) B0B(24-27) B0A(24-27) B0B(24-27) B0E(24-27) B0F(24-27) B0E(24-27) B0F(24-27)
2195
2196                    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)
2197                    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)
2198                    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)
2199                    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)
2200                    const __m512i rhs_mat_014589CD_12_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_12, (_MM_PERM_ENUM)136); //B10(16-19) B11(16-19) B10(16-19) B11(16-19) B14(16-19) B15(16-19) B14(16-19) B15(16-19) B18(16-19) B19(16-19) B18(16-19) B19(16-19) B1C(16-19) B1D(16-19) B1C(16-19) B1D(16-19)
2201                    const __m512i rhs_mat_2367ABEF_12_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_12, (_MM_PERM_ENUM)136); //B12(16-19) B13(16-19) B12(16-19) B13(16-19) B16(16-19) B17(16-19) B16(16-19) B17(16-19) B1A(16-19) B1B(16-19) B1A(16-19) B1B(16-19) B1E(16-19) B1F(16-19) B1E(16-19) B1F(16-19)
2202                    const __m512i rhs_mat_014589CD_13_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_13, (_MM_PERM_ENUM)136); //B10(24-27) B11(24-27) B10(24-27) B11(24-27) B14(24-27) B15(24-27) B14(24-27) B15(24-27) B18(24-27) B19(24-27) B18(24-27) B19(24-27) B1C(24-27) B1D(24-27) B1C(24-27) B1D(24-27)
2203                    const __m512i rhs_mat_2367ABEF_13_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_13, (_MM_PERM_ENUM)136); //B12(24-27) B13(24-27) B12(24-27) B13(24-27) B16(24-27) B17(24-27) B16(24-27) B17(24-27) B1A(24-27) B1B(24-27) B1A(24-27) B1B(24-27) B1E(24-27) B1F(24-27) B1E(24-27) B1F(24-27)
2204
2205                    // Shuffle pattern two - right side input
2206                    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)
2207                    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)
2208                    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)
2209                    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)
2210                    const __m512i rhs_mat_014589CD_02_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_02, (_MM_PERM_ENUM)221); //B00(20-23) B01(20-23) B00(20-23) B01(20-23) B04(20-23) B05(20-23) B04(20-23) B05(20-23) B08(20-23) B09(20-23) B08(20-23) B09(20-23) B0C(20-23) B0D(20-23) B0C(20-23) B0D(20-23)
2211                    const __m512i rhs_mat_2367ABEF_02_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_02, (_MM_PERM_ENUM)221); //B02(20-23) B03(20-23) B02(20-23) B03(20-23) B06(20-23) B07(20-23) B06(20-23) B07(20-23) B0A(20-23) B0B(20-23) B0A(20-23) B0B(20-23) B0E(20-23) B0F(20-23) B0E(20-23) B0F(20-23)
2212                    const __m512i rhs_mat_014589CD_03_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_03, (_MM_PERM_ENUM)221); //B00(28-31) B01(28-31) B00(28-31) B01(28-31) B04(28-31) B05(28-31) B04(28-31) B05(28-31) B08(28-31) B09(28-31) B08(28-31) B09(28-31) B0C(28-31) B0D(28-31) B0C(28-31) 0BD(28-31)
2213                    const __m512i rhs_mat_2367ABEF_03_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_03, (_MM_PERM_ENUM)221); //B02(28-31) B03(28-31) B02(28-31) B03(28-31) B06(28-31) B07(28-31) B06(28-31) B07(28-31) B0A(28-31) B0B(28-31) B0A(28-31) B0B(28-31) B0E(28-31) B0F(28-31) B0E(28-31) B0F(28-31)
2214
2215                    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)
2216                    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)
2217                    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)
2218                    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)
2219                    const __m512i rhs_mat_014589CD_12_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_12, (_MM_PERM_ENUM)221); //B10(20-23) B11(20-23) B10(20-23) B11(20-23) B14(20-23) B15(20-23) B14(20-23) B15(20-23) B18(20-23) B19(20-23) B18(20-23) B19(20-23) B1C(20-23) B1D(20-23) B1C(20-23) B1D(20-23)
2220                    const __m512i rhs_mat_2367ABEF_12_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_12, (_MM_PERM_ENUM)221); //B12(20-23) B13(20-23) B12(20-23) B13(20-23) B16(20-23) B17(20-23) B16(20-23) B17(20-23) B1A(20-23) B1B(20-23) B1A(20-23) B1B(20-23) B1E(20-23) B1F(20-23) B1E(20-23) B1F(20-23)
2221                    const __m512i rhs_mat_014589CD_13_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_13, (_MM_PERM_ENUM)221); //B10(28-31) B11(28-31) B10(28-31) B11(28-31) B14(28-31) B15(28-31) B14(28-31) B15(28-31) B18(28-31) B19(28-31) B18(28-31) B19(28-31) B1C(28-31) B1D(28-31) B1C(28-31) B1D(28-31)
2222                    const __m512i rhs_mat_2367ABEF_13_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_13, (_MM_PERM_ENUM)221); //B12(28-31) B13(28-31) B12(28-31) B13(28-31) B16(28-31) B17(28-31) B16(28-31) B17(28-31) B1A(28-31) B1B(28-31) B1A(28-31) B1B(28-31) B1E(28-31) B1F(28-31) B1E(28-31) B1F(28-31)
2223
2224                    uint32_t utmp_00[4], utmp_01[4], utmp_10[4], utmp_11[4];
2225
2226                    // Scales and Mins of corresponding sub blocks from different Q4_K structures are stored together
2227                    // The below block is for eg to extract first sub block's scales and mins from different Q4_K structures for the sb loop
2228                    memcpy(utmp_00, b_ptr_0[b].scales + 24 * sb, 12);
2229                    utmp_00[3] = ((utmp_00[2] >> 4) & kmask2) | (((utmp_00[1] >> 6) & kmask3) << 4);
2230                    const uint32_t uaux_00 = utmp_00[1] & kmask1;
2231                    utmp_00[1] = (utmp_00[2] & kmask2) | (((utmp_00[0] >> 6) & kmask3) << 4);
2232                    utmp_00[2] = uaux_00;
2233                    utmp_00[0] &= kmask1;
2234
2235                    // The below block is for eg to extract second sub block's scales and mins from different Q4_K structures for the sb loop
2236                    memcpy(utmp_01, b_ptr_0[b].scales + 12 + sb * 24, 12);
2237                    utmp_01[3] = ((utmp_01[2] >> 4) & kmask2) | (((utmp_01[1] >> 6) & kmask3) << 4);
2238                    const uint32_t uaux_01 = utmp_01[1] & kmask1;
2239                    utmp_01[1] = (utmp_01[2] & kmask2) | (((utmp_01[0] >> 6) & kmask3) << 4);
2240                    utmp_01[2] = uaux_01;
2241                    utmp_01[0] &= kmask1;
2242
2243                    memcpy(utmp_10, b_ptr_1[b].scales + sb * 24, 12);
2244                    utmp_10[3] = ((utmp_10[2] >> 4) & kmask2) | (((utmp_10[1] >> 6) & kmask3) << 4);
2245                    const uint32_t uaux_10 = utmp_10[1] & kmask1;
2246                    utmp_10[1] = (utmp_10[2] & kmask2) | (((utmp_10[0] >> 6) & kmask3) << 4);
2247                    utmp_10[2] = uaux_10;
2248                    utmp_10[0] &= kmask1;
2249
2250                    // The below block is for eg to extract second sub block's scales and mins from different Q4_K structures for the sb loop
2251                    memcpy(utmp_11, b_ptr_1[b].scales + 12 + sb * 24, 12);
2252                    utmp_11[3] = ((utmp_11[2] >> 4) & kmask2) | (((utmp_11[1] >> 6) & kmask3) << 4);
2253                    const uint32_t uaux_11 = utmp_11[1] & kmask1;
2254                    utmp_11[1] = (utmp_11[2] & kmask2) | (((utmp_11[0] >> 6) & kmask3) << 4);
2255                    utmp_11[2] = uaux_11;
2256                    utmp_11[0] &= kmask1;
2257
2258                    // Scales of first sub block in the sb loop
2259                    const __m256i mins_and_scales_0 = _mm256_set_epi32(utmp_10[3], utmp_10[2], utmp_10[1], utmp_10[0], utmp_00[3], utmp_00[2], utmp_00[1], utmp_00[0]);
2260                    const __m512i scales_0 = _mm512_cvtepu8_epi16(_mm256_unpacklo_epi8(mins_and_scales_0, mins_and_scales_0));
2261
2262                    // Scales of second sub block in the sb loop
2263                    const __m256i mins_and_scales_1 = _mm256_set_epi32(utmp_11[3], utmp_11[2], utmp_11[1], utmp_11[0], utmp_01[3], utmp_01[2], utmp_01[1], utmp_01[0]);
2264                    const __m512i scales_1 = _mm512_cvtepu8_epi16(_mm256_unpacklo_epi8(mins_and_scales_1, mins_and_scales_1));
2265
2266                    // Mins of first and second sub block of Q4_K block are arranged side by side
2267                    const __m512i mins_01 = _mm512_cvtepu8_epi16(_mm256_unpacklo_epi8(_mm256_shuffle_epi32(mins_and_scales_0, 78), _mm256_shuffle_epi32(mins_and_scales_1, 78)));
2268
2269                    const __m512i scale_014589CD_0 = _mm512_shuffle_epi32(scales_0, (_MM_PERM_ENUM)68);
2270                    const __m512i scale_2367ABEF_0 = _mm512_shuffle_epi32(scales_0, (_MM_PERM_ENUM)238);
2271
2272                    const __m512i scale_014589CD_1 = _mm512_shuffle_epi32(scales_1, (_MM_PERM_ENUM)68);
2273                    const __m512i scale_2367ABEF_1 = _mm512_shuffle_epi32(scales_1, (_MM_PERM_ENUM)238);
2274
2275                    for (int rp = 0; rp < 4; rp++) {
2276
2277                        // Load the four block_q8_k quantized values interleaved with each other in chunks of eight bytes - A0,A1,A2,A3
2278                        // Loaded as set of 128 bit vectors and repeated and stored into a 256 bit vector before again repeating into 512 bit vector
2279                        __m256i lhs_mat_ymm_0123_00 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 256 * sb)));
2280                        __m256i lhs_mat_ymm_01_00 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_00, lhs_mat_ymm_0123_00, 0);
2281                        __m256i lhs_mat_ymm_23_00 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_00, lhs_mat_ymm_0123_00, 17);
2282                        __m256i lhs_mat_ymm_0123_01 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 32 + 256 * sb)));
2283                        __m256i lhs_mat_ymm_01_01 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_01, lhs_mat_ymm_0123_01, 0);
2284                        __m256i lhs_mat_ymm_23_01 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_01, lhs_mat_ymm_0123_01, 17);
2285                        __m256i lhs_mat_ymm_0123_02 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 64 + 256 * sb)));
2286                        __m256i lhs_mat_ymm_01_02 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_02, lhs_mat_ymm_0123_02, 0);
2287                        __m256i lhs_mat_ymm_23_02 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_02, lhs_mat_ymm_0123_02, 17);
2288                        __m256i lhs_mat_ymm_0123_03 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 96 + 256 * sb)));
2289                        __m256i lhs_mat_ymm_01_03 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_03, lhs_mat_ymm_0123_03, 0);
2290                        __m256i lhs_mat_ymm_23_03 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_03, lhs_mat_ymm_0123_03, 17);
2291                        __m256i lhs_mat_ymm_0123_10 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 128 + 256 * sb)));
2292                        __m256i lhs_mat_ymm_01_10 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_10, lhs_mat_ymm_0123_10, 0);
2293                        __m256i lhs_mat_ymm_23_10 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_10, lhs_mat_ymm_0123_10, 17);
2294                        __m256i lhs_mat_ymm_0123_11 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 160 + 256 * sb)));
2295                        __m256i lhs_mat_ymm_01_11 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_11, lhs_mat_ymm_0123_11, 0);
2296                        __m256i lhs_mat_ymm_23_11 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_11, lhs_mat_ymm_0123_11, 17);
2297                        __m256i lhs_mat_ymm_0123_12 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 192 + 256 * sb)));
2298                        __m256i lhs_mat_ymm_01_12 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_12, lhs_mat_ymm_0123_12, 0);
2299                        __m256i lhs_mat_ymm_23_12 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_12, lhs_mat_ymm_0123_12, 17);
2300                        __m256i lhs_mat_ymm_0123_13 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 224 + 256 * sb)));
2301                        __m256i lhs_mat_ymm_01_13 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_13, lhs_mat_ymm_0123_13, 0);
2302                        __m256i lhs_mat_ymm_23_13 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_13, lhs_mat_ymm_0123_13, 17);
2303
2304                        __m512i lhs_mat_01_00 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_00), lhs_mat_ymm_01_00, 1);
2305                        __m512i lhs_mat_23_00 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_00), lhs_mat_ymm_23_00, 1);
2306                        __m512i lhs_mat_01_01 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_01), lhs_mat_ymm_01_01, 1);
2307                        __m512i lhs_mat_23_01 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_01), lhs_mat_ymm_23_01, 1);
2308                        __m512i lhs_mat_01_02 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_02), lhs_mat_ymm_01_02, 1);
2309                        __m512i lhs_mat_23_02 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_02), lhs_mat_ymm_23_02, 1);
2310                        __m512i lhs_mat_01_03 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_03), lhs_mat_ymm_01_03, 1);
2311                        __m512i lhs_mat_23_03 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_03), lhs_mat_ymm_23_03, 1);
2312
2313                        __m512i lhs_mat_01_10 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_10), lhs_mat_ymm_01_10, 1);
2314                        __m512i lhs_mat_23_10 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_10), lhs_mat_ymm_23_10, 1);
2315                        __m512i lhs_mat_01_11 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_11), lhs_mat_ymm_01_11, 1);
2316                        __m512i lhs_mat_23_11 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_11), lhs_mat_ymm_23_11, 1);
2317                        __m512i lhs_mat_01_12 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_12), lhs_mat_ymm_01_12, 1);
2318                        __m512i lhs_mat_23_12 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_12), lhs_mat_ymm_23_12, 1);
2319                        __m512i lhs_mat_01_13 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_13), lhs_mat_ymm_01_13, 1);
2320                        __m512i lhs_mat_23_13 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_13), lhs_mat_ymm_23_13, 1);
2321
2322                        // Bsums are loaded - four bsums are loaded (for two sub blocks) for the different Q8_K blocks
2323                        __m256i lhs_bsums_0123_01 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].bsums + 16 * sb)));
2324                        __m256i lhs_bsums_hsum_ymm_0123_01 = _mm256_castsi128_si256(_mm_hadd_epi16(_mm256_castsi256_si128(lhs_bsums_0123_01), _mm256_extractf128_si256(lhs_bsums_0123_01, 1)));
2325                        lhs_bsums_hsum_ymm_0123_01 = _mm256_permute2x128_si256(lhs_bsums_hsum_ymm_0123_01, lhs_bsums_hsum_ymm_0123_01, 0);
2326                        __m512i lhs_bsums_hsum_0123_01 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_bsums_hsum_ymm_0123_01), lhs_bsums_hsum_ymm_0123_01, 1);
2327
2328                        // Shuffle pattern one - left side input
2329                        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)
2330                        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)
2331                        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)
2332                        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)
2333                        const __m512i lhs_mat_01_02_sp1 = _mm512_shuffle_epi32(lhs_mat_01_02, (_MM_PERM_ENUM)160); //A00(16-19) A00(16-19) A01(16-19) A01(16-19) A00(16-19) A00(16-19) A01(16-19) A01(16-19) A00(16-19) A00(16-19) A01(16-19) A01(16-19) A00(16-19) A00(16-19) A01(16-19) A01(16-19)
2334                        const __m512i lhs_mat_23_02_sp1 = _mm512_shuffle_epi32(lhs_mat_23_02, (_MM_PERM_ENUM)160); //A02(16-19) A02(16-19) A03(16-19) A03(16-19) A02(16-19) A02(16-19) A03(16-19) A03(16-19) A02(16-19) A02(16-19) A03(16-19) A03(16-19) A02(16-19) A02(16-19) A03(16-19) A03(16-19)
2335                        const __m512i lhs_mat_01_03_sp1 = _mm512_shuffle_epi32(lhs_mat_01_03, (_MM_PERM_ENUM)160); //A00(24-27) A00(24-27) A01(24-27) A01(24-27) A00(24-27) A00(24-27) A01(24-27) A01(24-27) A00(24-27) A00(24-27) A01(24-27) A01(24-27) A00(24-27) A00(24-27) A01(24-27) A01(24-27)
2336                        const __m512i lhs_mat_23_03_sp1 = _mm512_shuffle_epi32(lhs_mat_23_03, (_MM_PERM_ENUM)160); //A02(24-27) A02(24-27) A03(24-27) A03(24-27) A02(24-27) A02(24-27) A03(24-27) A03(24-27) A02(24-27) A02(24-27) A03(24-27) A03(24-27) A02(24-27) A02(24-27) A03(24-27) A03(24-27)
2337
2338                        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)
2339                        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)
2340                        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)
2341                        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)
2342                        const __m512i lhs_mat_01_12_sp1 = _mm512_shuffle_epi32(lhs_mat_01_12, (_MM_PERM_ENUM)160); //A10(16-19) A10(16-19) A11(16-19) A11(16-19) A10(16-19) A10(16-19) A11(16-19) A11(16-19) A10(16-19) A10(16-19) A11(16-19) A11(16-19) A10(16-19) A10(16-19) A11(16-19) A11(16-19)
2343                        const __m512i lhs_mat_23_12_sp1 = _mm512_shuffle_epi32(lhs_mat_23_12, (_MM_PERM_ENUM)160); //A12(16-19) A12(16-19) A13(16-19) A13(16-19) A12(16-19) A12(16-19) A13(16-19) A13(16-19) A12(16-19) A12(16-19) A13(16-19) A13(16-19) A12(16-19) A12(16-19) A13(16-19) A13(16-19)
2344                        const __m512i lhs_mat_01_13_sp1 = _mm512_shuffle_epi32(lhs_mat_01_13, (_MM_PERM_ENUM)160); //A10(24-27) A10(24-27) A11(24-27) A11(24-27) A10(24-27) A10(24-27) A11(24-27) A11(24-27) A10(24-27) A10(24-27) A11(24-27) A11(24-27) A10(24-27) A10(24-27) A11(24-27) A11(24-27)
2345                        const __m512i lhs_mat_23_13_sp1 = _mm512_shuffle_epi32(lhs_mat_23_13, (_MM_PERM_ENUM)160); //A12(24-27) A12(24-27) A13(24-27) A13(24-27) A12(24-27) A12(24-27) A13(24-27) A13(24-27) A12(24-27) A12(24-27) A13(24-27) A13(24-27) A12(24-27) A12(24-27) A13(24-27) A13(24-27)
2346
2347                        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)
2348                        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)
2349                        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)
2350                        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)
2351                        const __m512i lhs_mat_01_02_sp2 = _mm512_shuffle_epi32(lhs_mat_01_02, (_MM_PERM_ENUM)245); //A00(20-23) A00(20-23) A01(20-23) A01(20-23) A00(20-23) A00(20-23) A01(20-23) A01(20-23) A00(20-23) A00(20-23) A01(20-23) A01(20-23) A00(20-23) A00(20-23) A01(20-23) A01(20-23)
2352                        const __m512i lhs_mat_23_02_sp2 = _mm512_shuffle_epi32(lhs_mat_23_02, (_MM_PERM_ENUM)245); //A02(20-23) A02(20-23) A03(20-23) A03(20-23) A02(20-23) A02(20-23) A03(20-23) A03(20-23) A02(20-23) A02(20-23) A03(20-23) A03(20-23) A02(20-23) A02(20-23) A03(20-23) A03(20-23)
2353                        const __m512i lhs_mat_01_03_sp2 = _mm512_shuffle_epi32(lhs_mat_01_03, (_MM_PERM_ENUM)245); //A00(28-31) A00(28-31) A01(28-31) A01(28-31) A00(28-31) A00(28-31) A01(28-31) A01(28-31) A00(28-31) A00(28-31) A01(28-31) A01(28-31) A00(28-31) A00(28-31) A01(28-31) A01(28-31)
2354                        const __m512i lhs_mat_23_03_sp2 = _mm512_shuffle_epi32(lhs_mat_23_03, (_MM_PERM_ENUM)245); //A02(28-31) A02(28-31) A03(28-31) A03(28-31) A02(28-31) A02(28-31) A03(28-31) A03(28-31) A02(28-31) A02(28-31) A03(28-31) A03(28-31) A02(28-31) A02(28-31) A03(28-31) A03(28-31)
2355
2356                        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)
2357                        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)
2358                        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)
2359                        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)
2360                        const __m512i lhs_mat_01_12_sp2 = _mm512_shuffle_epi32(lhs_mat_01_12, (_MM_PERM_ENUM)245); //A10(20-23) A10(20-23) A11(20-23) A11(20-23) A10(20-23) A10(20-23) A11(20-23) A11(20-23) A10(20-23) A10(20-23) A11(20-23) A11(20-23) A10(20-23) A10(20-23) A11(20-23) A11(20-23)
2361                        const __m512i lhs_mat_23_12_sp2 = _mm512_shuffle_epi32(lhs_mat_23_12, (_MM_PERM_ENUM)245); //A12(20-23) A12(20-23) A13(20-23) A13(20-23) A12(20-23) A12(20-23) A13(20-23) A13(20-23) A12(20-23) A12(20-23) A13(20-23) A13(20-23) A12(20-23) A12(20-23) A13(20-23) A13(20-23)
2362                        const __m512i lhs_mat_01_13_sp2 = _mm512_shuffle_epi32(lhs_mat_01_13, (_MM_PERM_ENUM)245); //A10(28-31) A10(28-31) A11(28-31) A11(28-31) A10(28-31) A10(28-31) A11(28-31) A11(28-31) A10(28-31) A10(28-31) A11(28-31) A11(28-31) A10(28-31) A10(28-31) A11(28-31) A11(28-31)
2363                        const __m512i lhs_mat_23_13_sp2 = _mm512_shuffle_epi32(lhs_mat_23_13, (_MM_PERM_ENUM)245); //A12(28-31) A12(28-31) A13(28-31) A13(28-31) A12(28-31) A12(28-31) A13(28-31) A13(28-31) A12(28-31) A12(28-31) A13(28-31) A13(28-31) A12(28-31) A12(28-31) A13(28-31) A13(28-31)
2364
2365                        // 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
2366                        __m512i iacc_mat_00_0_sp1 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_03_sp1, lhs_mat_01_03_sp1), _mm512_maddubs_epi16(rhs_mat_014589CD_02_sp1, lhs_mat_01_02_sp1)), _mm512_maddubs_epi16(rhs_mat_014589CD_01_sp1, lhs_mat_01_01_sp1)), _mm512_maddubs_epi16(rhs_mat_014589CD_00_sp1, lhs_mat_01_00_sp1));
2367                        __m512i iacc_mat_01_0_sp1 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_03_sp1, lhs_mat_01_03_sp1), _mm512_maddubs_epi16(rhs_mat_2367ABEF_02_sp1, lhs_mat_01_02_sp1)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_01_sp1, lhs_mat_01_01_sp1)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_00_sp1, lhs_mat_01_00_sp1));
2368                        __m512i iacc_mat_10_0_sp1 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_03_sp1, lhs_mat_23_03_sp1), _mm512_maddubs_epi16(rhs_mat_014589CD_02_sp1, lhs_mat_23_02_sp1)), _mm512_maddubs_epi16(rhs_mat_014589CD_01_sp1, lhs_mat_23_01_sp1)), _mm512_maddubs_epi16(rhs_mat_014589CD_00_sp1, lhs_mat_23_00_sp1));
2369                        __m512i iacc_mat_11_0_sp1 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_03_sp1, lhs_mat_23_03_sp1), _mm512_maddubs_epi16(rhs_mat_2367ABEF_02_sp1, lhs_mat_23_02_sp1)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_01_sp1, lhs_mat_23_01_sp1)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_00_sp1, lhs_mat_23_00_sp1));
2370                        __m512i iacc_mat_00_1_sp1 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_13_sp1, lhs_mat_01_13_sp1), _mm512_maddubs_epi16(rhs_mat_014589CD_12_sp1, lhs_mat_01_12_sp1)), _mm512_maddubs_epi16(rhs_mat_014589CD_11_sp1, lhs_mat_01_11_sp1)), _mm512_maddubs_epi16(rhs_mat_014589CD_10_sp1, lhs_mat_01_10_sp1));
2371                        __m512i iacc_mat_01_1_sp1 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_13_sp1, lhs_mat_01_13_sp1), _mm512_maddubs_epi16(rhs_mat_2367ABEF_12_sp1, lhs_mat_01_12_sp1)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_11_sp1, lhs_mat_01_11_sp1)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_10_sp1, lhs_mat_01_10_sp1));
2372                        __m512i iacc_mat_10_1_sp1 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_13_sp1, lhs_mat_23_13_sp1), _mm512_maddubs_epi16(rhs_mat_014589CD_12_sp1, lhs_mat_23_12_sp1)), _mm512_maddubs_epi16(rhs_mat_014589CD_11_sp1, lhs_mat_23_11_sp1)), _mm512_maddubs_epi16(rhs_mat_014589CD_10_sp1, lhs_mat_23_10_sp1));
2373                        __m512i iacc_mat_11_1_sp1 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_13_sp1, lhs_mat_23_13_sp1), _mm512_maddubs_epi16(rhs_mat_2367ABEF_12_sp1, lhs_mat_23_12_sp1)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_11_sp1, lhs_mat_23_11_sp1)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_10_sp1, lhs_mat_23_10_sp1));
2374
2375                        __m512i iacc_mat_00_0_sp2 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_03_sp2, lhs_mat_01_03_sp2), _mm512_maddubs_epi16(rhs_mat_014589CD_02_sp2, lhs_mat_01_02_sp2)), _mm512_maddubs_epi16(rhs_mat_014589CD_01_sp2, lhs_mat_01_01_sp2)), _mm512_maddubs_epi16(rhs_mat_014589CD_00_sp2, lhs_mat_01_00_sp2));
2376                        __m512i iacc_mat_01_0_sp2 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_03_sp2, lhs_mat_01_03_sp2), _mm512_maddubs_epi16(rhs_mat_2367ABEF_02_sp2, lhs_mat_01_02_sp2)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_01_sp2, lhs_mat_01_01_sp2)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_00_sp2, lhs_mat_01_00_sp2));
2377                        __m512i iacc_mat_10_0_sp2 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_03_sp2, lhs_mat_23_03_sp2), _mm512_maddubs_epi16(rhs_mat_014589CD_02_sp2, lhs_mat_23_02_sp2)), _mm512_maddubs_epi16(rhs_mat_014589CD_01_sp2, lhs_mat_23_01_sp2)), _mm512_maddubs_epi16(rhs_mat_014589CD_00_sp2, lhs_mat_23_00_sp2));
2378                        __m512i iacc_mat_11_0_sp2 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_03_sp2, lhs_mat_23_03_sp2), _mm512_maddubs_epi16(rhs_mat_2367ABEF_02_sp2, lhs_mat_23_02_sp2)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_01_sp2, lhs_mat_23_01_sp2)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_00_sp2, lhs_mat_23_00_sp2));
2379                        __m512i iacc_mat_00_1_sp2 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_13_sp2, lhs_mat_01_13_sp2), _mm512_maddubs_epi16(rhs_mat_014589CD_12_sp2, lhs_mat_01_12_sp2)), _mm512_maddubs_epi16(rhs_mat_014589CD_11_sp2, lhs_mat_01_11_sp2)), _mm512_maddubs_epi16(rhs_mat_014589CD_10_sp2, lhs_mat_01_10_sp2));
2380                        __m512i iacc_mat_01_1_sp2 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_13_sp2, lhs_mat_01_13_sp2), _mm512_maddubs_epi16(rhs_mat_2367ABEF_12_sp2, lhs_mat_01_12_sp2)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_11_sp2, lhs_mat_01_11_sp2)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_10_sp2, lhs_mat_01_10_sp2));
2381                        __m512i iacc_mat_10_1_sp2 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_13_sp2, lhs_mat_23_13_sp2), _mm512_maddubs_epi16(rhs_mat_014589CD_12_sp2, lhs_mat_23_12_sp2)), _mm512_maddubs_epi16(rhs_mat_014589CD_11_sp2, lhs_mat_23_11_sp2)), _mm512_maddubs_epi16(rhs_mat_014589CD_10_sp2, lhs_mat_23_10_sp2));
2382                        __m512i iacc_mat_11_1_sp2 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_13_sp2, lhs_mat_23_13_sp2), _mm512_maddubs_epi16(rhs_mat_2367ABEF_12_sp2, lhs_mat_23_12_sp2)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_11_sp2, lhs_mat_23_11_sp2)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_10_sp2, lhs_mat_23_10_sp2));
2383
2384                        // Output of both shuffle patterns are added in order to sum dot product outputs of all 32 values in block
2385                        __m512i iacc_mat_00_0 = _mm512_add_epi16(iacc_mat_00_0_sp1, iacc_mat_00_0_sp2);
2386                        __m512i iacc_mat_01_0 = _mm512_add_epi16(iacc_mat_01_0_sp1, iacc_mat_01_0_sp2);
2387                        __m512i iacc_mat_10_0 = _mm512_add_epi16(iacc_mat_10_0_sp1, iacc_mat_10_0_sp2);
2388                        __m512i iacc_mat_11_0 = _mm512_add_epi16(iacc_mat_11_0_sp1, iacc_mat_11_0_sp2);
2389
2390                        __m512i iacc_mat_00_1 = _mm512_add_epi16(iacc_mat_00_1_sp1, iacc_mat_00_1_sp2);
2391                        __m512i iacc_mat_01_1 = _mm512_add_epi16(iacc_mat_01_1_sp1, iacc_mat_01_1_sp2);
2392                        __m512i iacc_mat_10_1 = _mm512_add_epi16(iacc_mat_10_1_sp1, iacc_mat_10_1_sp2);
2393                        __m512i iacc_mat_11_1 = _mm512_add_epi16(iacc_mat_11_1_sp1, iacc_mat_11_1_sp2);
2394
2395                        iacc_mat_00_0 = _mm512_madd_epi16(iacc_mat_00_0, scale_014589CD_0);
2396                        iacc_mat_01_0 = _mm512_madd_epi16(iacc_mat_01_0, scale_2367ABEF_0);
2397                        iacc_mat_10_0 = _mm512_madd_epi16(iacc_mat_10_0, scale_014589CD_0);
2398                        iacc_mat_11_0 = _mm512_madd_epi16(iacc_mat_11_0, scale_2367ABEF_0);
2399
2400                        iacc_mat_00_1 = _mm512_madd_epi16(iacc_mat_00_1, scale_014589CD_1);
2401                        iacc_mat_01_1 = _mm512_madd_epi16(iacc_mat_01_1, scale_2367ABEF_1);
2402                        iacc_mat_10_1 = _mm512_madd_epi16(iacc_mat_10_1, scale_014589CD_1);
2403                        iacc_mat_11_1 = _mm512_madd_epi16(iacc_mat_11_1, scale_2367ABEF_1);
2404
2405                        // Straighten out to make 4 row vectors (4 for each sub block which are accumulated together in the next step)
2406                        __m512i iacc_row_0_0 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_00_0, _mm512_shuffle_epi32(iacc_mat_01_0, (_MM_PERM_ENUM)78));
2407                        __m512i iacc_row_1_0 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_00_0, (_MM_PERM_ENUM)78), iacc_mat_01_0);
2408                        __m512i iacc_row_2_0 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_10_0, _mm512_shuffle_epi32(iacc_mat_11_0, (_MM_PERM_ENUM)78));
2409                        __m512i iacc_row_3_0 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_10_0, (_MM_PERM_ENUM)78), iacc_mat_11_0);
2410                        __m512i iacc_row_0_1 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_00_1, _mm512_shuffle_epi32(iacc_mat_01_1, (_MM_PERM_ENUM)78));
2411                        __m512i iacc_row_1_1 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_00_1, (_MM_PERM_ENUM)78), iacc_mat_01_1);
2412                        __m512i iacc_row_2_1 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_10_1, _mm512_shuffle_epi32(iacc_mat_11_1, (_MM_PERM_ENUM)78));
2413                        __m512i iacc_row_3_1 = _mm512_mask_blend_epi32(0xCCCC,_mm512_shuffle_epi32(iacc_mat_10_1, (_MM_PERM_ENUM)78), iacc_mat_11_1);
2414
2415                        __m512i iacc_row_0 = _mm512_add_epi32(iacc_row_0_0, iacc_row_0_1);
2416                        __m512i iacc_row_1 = _mm512_add_epi32(iacc_row_1_0, iacc_row_1_1);
2417                        __m512i iacc_row_2 = _mm512_add_epi32(iacc_row_2_0, iacc_row_2_1);
2418                        __m512i iacc_row_3 = _mm512_add_epi32(iacc_row_3_0, iacc_row_3_1);
2419
2420                        // Load the scale(d) values for all the 4 Q8_k blocks and repeat it across lanes
2421                        const __m128 row_scale_f32_sse = _mm_load_ps(a_ptrs[rp][b].d);
2422                        const __m256 row_scale_f32_ymm = _mm256_set_m128(row_scale_f32_sse, row_scale_f32_sse);
2423                        const __m512 row_scale_f32 = _mm512_insertf32x8(_mm512_castps256_ps512(row_scale_f32_ymm), row_scale_f32_ymm, 1);
2424
2425                        // Multiply with appropriate scales and accumulate (for both d and dmin) below
2426                        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]);
2427                        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]);
2428                        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]);
2429                        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]);
2430
2431                        __m512i iacc_row_min_0 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_hsum_0123_01, (_MM_PERM_ENUM)0), mins_01);
2432                        __m512i iacc_row_min_1 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_hsum_0123_01, (_MM_PERM_ENUM)85), mins_01);
2433                        __m512i iacc_row_min_2 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_hsum_0123_01, (_MM_PERM_ENUM)170), mins_01);
2434                        __m512i iacc_row_min_3 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_hsum_0123_01, (_MM_PERM_ENUM)255), mins_01);
2435
2436                        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]);
2437                        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]);
2438                        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]);
2439                        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]);
2440                    }
2441                }
2442            }
2443            // Store the accumulated values
2444            for (int i = 0; i < 16; i++) {
2445                _mm512_storeu_ps((float * )(s + ((y * 4 + i) * bs + x * 8)), _mm512_sub_ps(acc_rows[i], acc_min_rows[i]));
2446            }
2447        }
2448    }
2449
2450    for (; y < nr / 4; y++) {
2451
2452        const block_q8_Kx4 * a_ptr = a_ptr_start + (y * nb);
2453
2454        // Take group of eight block_q4_kx8 structures at each pass of the loop and perform dot product operation
2455        for (int64_t x = 0; x < anc / 8; x += 2) {
2456
2457            const block_q4_Kx8 * b_ptr_0 = b_ptr_start + ((x) * b_nb);
2458            const block_q4_Kx8 * b_ptr_1 = b_ptr_start + ((x + 1) * b_nb);
2459
2460            // Master FP accumulators
2461            __m512 acc_rows[4];
2462            for (int i = 0; i < 4; i++) {
2463                acc_rows[i] = _mm512_setzero_ps();
2464            }
2465
2466            __m512 acc_min_rows[4];
2467            for (int i = 0; i < 4; i++) {
2468                acc_min_rows[i] = _mm512_setzero_ps();
2469            }
2470
2471            // For super block
2472            for (int64_t b = 0; b < nb; b++) {
2473                // Scale values - Load the sixteen scale values from two block_q4_kx8 structures
2474                const __m512 col_scale_f32 = GGML_F32Cx8x2_LOAD(b_ptr_0[b].d, b_ptr_1[b].d);
2475
2476                // dmin values - Load the sixteen dmin values from two block_q4_kx8 structures
2477                const __m512 col_dmin_f32 = GGML_F32Cx8x2_LOAD(b_ptr_0[b].dmin, b_ptr_1[b].dmin);
2478
2479                // Loop to iterate over the eight sub blocks of a super block - two sub blocks are processed per iteration
2480                for (int sb = 0; sb < QK_K / 64; sb++) {
2481
2482                    const __m256i rhs_raw_mat_0123_0 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + sb * 256));
2483                    const __m256i rhs_raw_mat_4567_0 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 32 + sb * 256));
2484                    const __m256i rhs_raw_mat_0123_1 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 64 + sb * 256));
2485                    const __m256i rhs_raw_mat_4567_1 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 96 + sb * 256));
2486                    const __m256i rhs_raw_mat_0123_2 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 128 + sb * 256));
2487                    const __m256i rhs_raw_mat_4567_2 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 160 + sb * 256));
2488                    const __m256i rhs_raw_mat_0123_3 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 192 + sb * 256));
2489                    const __m256i rhs_raw_mat_4567_3 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 224 + sb * 256));
2490
2491                    const __m256i rhs_raw_mat_89AB_0 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + sb * 256));
2492                    const __m256i rhs_raw_mat_CDEF_0 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 32 + sb * 256));
2493                    const __m256i rhs_raw_mat_89AB_1 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 64 + sb * 256));
2494                    const __m256i rhs_raw_mat_CDEF_1 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 96 + sb * 256));
2495                    const __m256i rhs_raw_mat_89AB_2 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 128 + sb * 256));
2496                    const __m256i rhs_raw_mat_CDEF_2 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 160 + sb * 256));
2497                    const __m256i rhs_raw_mat_89AB_3 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 192 + sb * 256));
2498                    const __m256i rhs_raw_mat_CDEF_3 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 224 + sb * 256));
2499
2500                    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);
2501                    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);
2502                    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);
2503                    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);
2504                    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);
2505                    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);
2506                    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);
2507                    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);
2508
2509                    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);
2510                    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);
2511                    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);
2512                    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);
2513                    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);
2514                    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);
2515                    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);
2516                    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);
2517
2518                    const __m512i rhs_raw_mat_014589CD_0 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_0), rhs_raw_mat_89CD_0, 1);
2519                    const __m512i rhs_raw_mat_2367ABEF_0 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_0), rhs_raw_mat_ABEF_0, 1);
2520                    const __m512i rhs_raw_mat_014589CD_1 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_1), rhs_raw_mat_89CD_1, 1);
2521                    const __m512i rhs_raw_mat_2367ABEF_1 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_1), rhs_raw_mat_ABEF_1, 1);
2522
2523                    const __m512i rhs_raw_mat_014589CD_2 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_2), rhs_raw_mat_89CD_2, 1);
2524                    const __m512i rhs_raw_mat_2367ABEF_2 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_2), rhs_raw_mat_ABEF_2, 1);
2525                    const __m512i rhs_raw_mat_014589CD_3 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_3), rhs_raw_mat_89CD_3, 1);
2526                    const __m512i rhs_raw_mat_2367ABEF_3 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_3), rhs_raw_mat_ABEF_3, 1);
2527
2528                    //4-bit -> 8-bit
2529                    const __m512i rhs_mat_014589CD_00 = _mm512_and_si512(rhs_raw_mat_014589CD_0, m4bexpanded); //B00(0-7) B01(0-7) B04(0-7) B05(0-7) B08(0-7) B09(0-7) B0C(0-7) B0D(0-7)
2530                    const __m512i rhs_mat_2367ABEF_00 = _mm512_and_si512(rhs_raw_mat_2367ABEF_0, m4bexpanded); //B02(0-7) B03(0-7) B06(0-7) B07(0-7) B0A(0-7) B0B(0-7) B0E(0-7) B0F(0-7)
2531                    const __m512i rhs_mat_014589CD_01 = _mm512_and_si512(rhs_raw_mat_014589CD_1, m4bexpanded); //B00(8-15) B01(8-15) B04(8-15) B05(8-15) B08(8-15) B09(8-15) B0C(8-15) B0D(8-15)
2532                    const __m512i rhs_mat_2367ABEF_01 = _mm512_and_si512(rhs_raw_mat_2367ABEF_1, m4bexpanded); //B02(8-15) B03(8-15) B06(8-15) B07(8-15) B0A(8-15) B0B(8-15) B0E(8-15) B0F(8-15)
2533
2534                    const __m512i rhs_mat_014589CD_02 = _mm512_and_si512(rhs_raw_mat_014589CD_2, m4bexpanded); //B00(16-23) B01(16-23) B04(16-23) B05(16-23) B08(16-23) B09(16-23) B0C(16-23) B0D(16-23)
2535                    const __m512i rhs_mat_2367ABEF_02 = _mm512_and_si512(rhs_raw_mat_2367ABEF_2, m4bexpanded); //B02(16-23) B03(16-23) B06(16-23) B07(16-23) B0A(16-23) B0B(16-23) B0E(16-23) B0F(16-23)
2536                    const __m512i rhs_mat_014589CD_03 = _mm512_and_si512(rhs_raw_mat_014589CD_3, m4bexpanded); //B00(24-31) B01(24-31) B04(24-31) B05(24-31) B08(24-31) B09(24-31) B0C(24-31) B0D(24-31)
2537                    const __m512i rhs_mat_2367ABEF_03 = _mm512_and_si512(rhs_raw_mat_2367ABEF_3, m4bexpanded); //B02(24-31) B03(24-31) B06(24-31) B07(24-31) B0A(24-31) B0B(24-31) B0E(24-31) B0F(24-31)
2538
2539                    const __m512i rhs_mat_014589CD_10 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_0, 4), m4bexpanded); //B10(0-7) B11(0-7) B14(0-7) B15(0-7) B18(0-7) B19(0-7) B1C(0-7) B1D(0-7)
2540                    const __m512i rhs_mat_2367ABEF_10 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_0, 4), m4bexpanded); //B12(0-7) B13(0-7) B16(0-7) B17(0-7) B1A(0-7) B1B(0-7) B1E(0-7) B1F(0-7)
2541                    const __m512i rhs_mat_014589CD_11 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_1, 4), m4bexpanded); //B10(8-15) B11(8-15) B14(8-15) B15(8-15) B18(8-15) B19(8-15) B1C(8-15) B1D(8-15)
2542                    const __m512i rhs_mat_2367ABEF_11 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_1, 4), m4bexpanded); //B12(8-15) B13(8-15) B16(8-15) B17(8-15) B1A(8-15) B1B(8-15) B1E(8-15) B1F(8-15)
2543
2544                    const __m512i rhs_mat_014589CD_12 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_2, 4), m4bexpanded); //B10(16-23) B11(16-23) B14(16-23) B15(16-23) B18(16-23) B19(16-23) B1C(16-23) B1D(16-23)
2545                    const __m512i rhs_mat_2367ABEF_12 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_2, 4), m4bexpanded); //B12(16-23) B13(16-23) B16(16-23) B17(16-23) B1A(16-23) B1B(16-23) B1E(16-23) B1F(16-23)
2546                    const __m512i rhs_mat_014589CD_13 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_014589CD_3, 4), m4bexpanded); //B10(24-31) B11(24-31) B14(24-31) B15(24-31) B18(24-31) B19(24-31) B1C(24-31) B1D(24-31)
2547                    const __m512i rhs_mat_2367ABEF_13 = _mm512_and_si512(_mm512_srli_epi16(rhs_raw_mat_2367ABEF_3, 4), m4bexpanded); //B12(24-31) B13(24-31) B16(24-31) B17(24-31) B1A(24-31) B1B(24-31) B1E(24-31) B1F(24-31)
2548
2549                    // Shuffle pattern one - right side input
2550                    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)
2551                    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)
2552                    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)
2553                    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)
2554                    const __m512i rhs_mat_014589CD_02_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_02, (_MM_PERM_ENUM)136); //B00(16-19) B01(16-19) B00(16-19) B01(16-19) B04(16-19) B05(16-19) B04(16-19) B05(16-19) B08(16-19) B09(16-19) B08(16-19) B09(16-19) B0C(16-19) B0D(16-19) B0C(16-19) B0D(16-19)
2555                    const __m512i rhs_mat_2367ABEF_02_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_02, (_MM_PERM_ENUM)136); //B02(16-19) B03(16-19) B02(16-19) B03(16-19) B06(16-19) B07(16-19) B06(16-19) B07(16-19) B0A(16-19) B0B(16-19) B0A(16-19) B0B(16-19) B0E(16-19) B0F(16-19) B0E(16-19) B0F(16-19)
2556                    const __m512i rhs_mat_014589CD_03_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_03, (_MM_PERM_ENUM)136); //B00(24-27) B01(24-27) B00(24-27) B01(24-27) B04(24-27) B05(24-27) B04(24-27) B05(24-27) B08(24-27) B09(24-27) B08(24-27) B09(24-27) B0C(24-27) B0D(24-27) B0C(24-27) B0D(24-27)
2557                    const __m512i rhs_mat_2367ABEF_03_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_03, (_MM_PERM_ENUM)136); //B02(24-27) B03(24-27) B02(24-27) B03(24-27) B06(24-27) B07(24-27) B06(24-27) B07(24-27) B0A(24-27) B0B(24-27) B0A(24-27) B0B(24-27) B0E(24-27) B0F(24-27) B0E(24-27) B0F(24-27)
2558
2559                    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)
2560                    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)
2561                    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)
2562                    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)
2563                    const __m512i rhs_mat_014589CD_12_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_12, (_MM_PERM_ENUM)136); //B10(16-19) B11(16-19) B10(16-19) B11(16-19) B14(16-19) B15(16-19) B14(16-19) B15(16-19) B18(16-19) B19(16-19) B18(16-19) B19(16-19) B1C(16-19) B1D(16-19) B1C(16-19) B1D(16-19)
2564                    const __m512i rhs_mat_2367ABEF_12_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_12, (_MM_PERM_ENUM)136); //B12(16-19) B13(16-19) B12(16-19) B13(16-19) B16(16-19) B17(16-19) B16(16-19) B17(16-19) B1A(16-19) B1B(16-19) B1A(16-19) B1B(16-19) B1E(16-19) B1F(16-19) B1E(16-19) B1F(16-19)
2565                    const __m512i rhs_mat_014589CD_13_sp1 = _mm512_shuffle_epi32(rhs_mat_014589CD_13, (_MM_PERM_ENUM)136); //B10(24-27) B11(24-27) B10(24-27) B11(24-27) B14(24-27) B15(24-27) B14(24-27) B15(24-27) B18(24-27) B19(24-27) B18(24-27) B19(24-27) B1C(24-27) B1D(24-27) B1C(24-27) B1D(24-27)
2566                    const __m512i rhs_mat_2367ABEF_13_sp1 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_13, (_MM_PERM_ENUM)136); //B12(24-27) B13(24-27) B12(24-27) B13(24-27) B16(24-27) B17(24-27) B16(24-27) B17(24-27) B1A(24-27) B1B(24-27) B1A(24-27) B1B(24-27) B1E(24-27) B1F(24-27) B1E(24-27) B1F(24-27)
2567
2568                    // Shuffle pattern two - right side input
2569                    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)
2570                    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)
2571                    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)
2572                    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)
2573                    const __m512i rhs_mat_014589CD_02_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_02, (_MM_PERM_ENUM)221); //B00(20-23) B01(20-23) B00(20-23) B01(20-23) B04(20-23) B05(20-23) B04(20-23) B05(20-23) B08(20-23) B09(20-23) B08(20-23) B09(20-23) B0C(20-23) B0D(20-23) B0C(20-23) B0D(20-23)
2574                    const __m512i rhs_mat_2367ABEF_02_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_02, (_MM_PERM_ENUM)221); //B02(20-23) B03(20-23) B02(20-23) B03(20-23) B06(20-23) B07(20-23) B06(20-23) B07(20-23) B0A(20-23) B0B(20-23) B0A(20-23) B0B(20-23) B0E(20-23) B0F(20-23) B0E(20-23) B0F(20-23)
2575                    const __m512i rhs_mat_014589CD_03_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_03, (_MM_PERM_ENUM)221); //B00(28-31) B01(28-31) B00(28-31) B01(28-31) B04(28-31) B05(28-31) B04(28-31) B05(28-31) B08(28-31) B09(28-31) B08(28-31) B09(28-31) B0C(28-31) B0D(28-31) B0C(28-31) 0BD(28-31)
2576                    const __m512i rhs_mat_2367ABEF_03_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_03, (_MM_PERM_ENUM)221); //B02(28-31) B03(28-31) B02(28-31) B03(28-31) B06(28-31) B07(28-31) B06(28-31) B07(28-31) B0A(28-31) B0B(28-31) B0A(28-31) B0B(28-31) B0E(28-31) B0F(28-31) B0E(28-31) B0F(28-31)
2577
2578                    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)
2579                    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)
2580                    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)
2581                    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)
2582                    const __m512i rhs_mat_014589CD_12_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_12, (_MM_PERM_ENUM)221); //B10(20-23) B11(20-23) B10(20-23) B11(20-23) B14(20-23) B15(20-23) B14(20-23) B15(20-23) B18(20-23) B19(20-23) B18(20-23) B19(20-23) B1C(20-23) B1D(20-23) B1C(20-23) B1D(20-23)
2583                    const __m512i rhs_mat_2367ABEF_12_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_12, (_MM_PERM_ENUM)221); //B12(20-23) B13(20-23) B12(20-23) B13(20-23) B16(20-23) B17(20-23) B16(20-23) B17(20-23) B1A(20-23) B1B(20-23) B1A(20-23) B1B(20-23) B1E(20-23) B1F(20-23) B1E(20-23) B1F(20-23)
2584                    const __m512i rhs_mat_014589CD_13_sp2 = _mm512_shuffle_epi32(rhs_mat_014589CD_13, (_MM_PERM_ENUM)221); //B10(28-31) B11(28-31) B10(28-31) B11(28-31) B14(28-31) B15(28-31) B14(28-31) B15(28-31) B18(28-31) B19(28-31) B18(28-31) B19(28-31) B1C(28-31) B1D(28-31) B1C(28-31) B1D(28-31)
2585                    const __m512i rhs_mat_2367ABEF_13_sp2 = _mm512_shuffle_epi32(rhs_mat_2367ABEF_13, (_MM_PERM_ENUM)221); //B12(28-31) B13(28-31) B12(28-31) B13(28-31) B16(28-31) B17(28-31) B16(28-31) B17(28-31) B1A(28-31) B1B(28-31) B1A(28-31) B1B(28-31) B1E(28-31) B1F(28-31) B1E(28-31) B1F(28-31)
2586
2587                    uint32_t utmp_00[4], utmp_01[4], utmp_10[4], utmp_11[4];
2588
2589                    // Scales and Mins of corresponding sub blocks from different Q4_K structures are stored together
2590                    // The below block is for eg to extract first sub block's scales and mins from different Q4_K structures for the sb loop
2591                    memcpy(utmp_00, b_ptr_0[b].scales + 24 * sb, 12);
2592                    utmp_00[3] = ((utmp_00[2] >> 4) & kmask2) | (((utmp_00[1] >> 6) & kmask3) << 4);
2593                    const uint32_t uaux_00 = utmp_00[1] & kmask1;
2594                    utmp_00[1] = (utmp_00[2] & kmask2) | (((utmp_00[0] >> 6) & kmask3) << 4);
2595                    utmp_00[2] = uaux_00;
2596                    utmp_00[0] &= kmask1;
2597
2598                    // The below block is for eg to extract second sub block's scales and mins from different Q4_K structures for the sb loop
2599                    memcpy(utmp_01, b_ptr_0[b].scales + 12 + sb * 24, 12);
2600                    utmp_01[3] = ((utmp_01[2] >> 4) & kmask2) | (((utmp_01[1] >> 6) & kmask3) << 4);
2601                    const uint32_t uaux_01 = utmp_01[1] & kmask1;
2602                    utmp_01[1] = (utmp_01[2] & kmask2) | (((utmp_01[0] >> 6) & kmask3) << 4);
2603                    utmp_01[2] = uaux_01;
2604                    utmp_01[0] &= kmask1;
2605
2606                    // The below block is for eg to extract first sub block's scales and mins from different Q4_K structures for the sb loop
2607                    memcpy(utmp_10, b_ptr_1[b].scales + sb * 24, 12);
2608                    utmp_10[3] = ((utmp_10[2] >> 4) & kmask2) | (((utmp_10[1] >> 6) & kmask3) << 4);
2609                    const uint32_t uaux_10 = utmp_10[1] & kmask1;
2610                    utmp_10[1] = (utmp_10[2] & kmask2) | (((utmp_10[0] >> 6) & kmask3) << 4);
2611                    utmp_10[2] = uaux_10;
2612                    utmp_10[0] &= kmask1;
2613
2614                    // The below block is for eg to extract second sub block's scales and mins from different Q4_K structures for the sb loop
2615                    memcpy(utmp_11, b_ptr_1[b].scales + 12 + sb * 24, 12);
2616                    utmp_11[3] = ((utmp_11[2] >> 4) & kmask2) | (((utmp_11[1] >> 6) & kmask3) << 4);
2617                    const uint32_t uaux_11 = utmp_11[1] & kmask1;
2618                    utmp_11[1] = (utmp_11[2] & kmask2) | (((utmp_11[0] >> 6) & kmask3) << 4);
2619                    utmp_11[2] = uaux_11;
2620                    utmp_11[0] &= kmask1;
2621
2622                    // Scales of first sub block in the sb loop
2623                    const __m256i mins_and_scales_0 = _mm256_set_epi32(utmp_10[3], utmp_10[2], utmp_10[1], utmp_10[0], utmp_00[3], utmp_00[2], utmp_00[1], utmp_00[0]);
2624                    const __m512i scales_0 = _mm512_cvtepu8_epi16(_mm256_unpacklo_epi8(mins_and_scales_0, mins_and_scales_0));
2625
2626                    // Scales of second sub block in the sb loop
2627                    const __m256i mins_and_scales_1 = _mm256_set_epi32(utmp_11[3], utmp_11[2], utmp_11[1], utmp_11[0], utmp_01[3], utmp_01[2], utmp_01[1], utmp_01[0]);
2628                    const __m512i scales_1 = _mm512_cvtepu8_epi16(_mm256_unpacklo_epi8(mins_and_scales_1, mins_and_scales_1));
2629
2630                    // Mins of first and second sub block of Q4_K block are arranged side by side
2631                    const __m512i mins_01 = _mm512_cvtepu8_epi16(_mm256_unpacklo_epi8(_mm256_shuffle_epi32(mins_and_scales_0, 78), _mm256_shuffle_epi32(mins_and_scales_1, 78)));
2632
2633                    const __m512i scale_014589CD_0 = _mm512_shuffle_epi32(scales_0, (_MM_PERM_ENUM)68);
2634                    const __m512i scale_2367ABEF_0 = _mm512_shuffle_epi32(scales_0, (_MM_PERM_ENUM)238);
2635
2636                    const __m512i scale_014589CD_1 = _mm512_shuffle_epi32(scales_1, (_MM_PERM_ENUM)68);
2637                    const __m512i scale_2367ABEF_1 = _mm512_shuffle_epi32(scales_1, (_MM_PERM_ENUM)238);
2638
2639                    // Load the four block_q8_k quantized values interleaved with each other in chunks of eight bytes - A0,A1,A2,A3
2640                    // Loaded as set of 128 bit vectors and repeated into a 256 bit vector
2641                    __m256i lhs_mat_ymm_0123_00 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 256 * sb)));
2642                    __m256i lhs_mat_ymm_01_00 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_00, lhs_mat_ymm_0123_00, 0);
2643                    __m256i lhs_mat_ymm_23_00 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_00, lhs_mat_ymm_0123_00, 17);
2644                    __m256i lhs_mat_ymm_0123_01 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 32 + 256 * sb)));
2645                    __m256i lhs_mat_ymm_01_01 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_01, lhs_mat_ymm_0123_01, 0);
2646                    __m256i lhs_mat_ymm_23_01 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_01, lhs_mat_ymm_0123_01, 17);
2647                    __m256i lhs_mat_ymm_0123_02 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 64 + 256 * sb)));
2648                    __m256i lhs_mat_ymm_01_02 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_02, lhs_mat_ymm_0123_02, 0);
2649                    __m256i lhs_mat_ymm_23_02 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_02, lhs_mat_ymm_0123_02, 17);
2650                    __m256i lhs_mat_ymm_0123_03 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 96 + 256 * sb)));
2651                    __m256i lhs_mat_ymm_01_03 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_03, lhs_mat_ymm_0123_03, 0);
2652                    __m256i lhs_mat_ymm_23_03 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_03, lhs_mat_ymm_0123_03, 17);
2653                    __m256i lhs_mat_ymm_0123_10 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 128 + 256 * sb)));
2654                    __m256i lhs_mat_ymm_01_10 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_10, lhs_mat_ymm_0123_10, 0);
2655                    __m256i lhs_mat_ymm_23_10 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_10, lhs_mat_ymm_0123_10, 17);
2656                    __m256i lhs_mat_ymm_0123_11 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 160 + 256 * sb)));
2657                    __m256i lhs_mat_ymm_01_11 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_11, lhs_mat_ymm_0123_11, 0);
2658                    __m256i lhs_mat_ymm_23_11 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_11, lhs_mat_ymm_0123_11, 17);
2659                    __m256i lhs_mat_ymm_0123_12 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 192 + 256 * sb)));
2660                    __m256i lhs_mat_ymm_01_12 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_12, lhs_mat_ymm_0123_12, 0);
2661                    __m256i lhs_mat_ymm_23_12 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_12, lhs_mat_ymm_0123_12, 17);
2662                    __m256i lhs_mat_ymm_0123_13 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 224 + 256 * sb)));
2663                    __m256i lhs_mat_ymm_01_13 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_13, lhs_mat_ymm_0123_13, 0);
2664                    __m256i lhs_mat_ymm_23_13 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_13, lhs_mat_ymm_0123_13, 17);
2665
2666                    //Loaded as set of 128 bit vectors and repeated and stored into a 256 bit vector before again repeating into a 512 bit vector
2667                    __m512i lhs_mat_01_00 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_00), lhs_mat_ymm_01_00, 1);
2668                    __m512i lhs_mat_23_00 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_00), lhs_mat_ymm_23_00, 1);
2669                    __m512i lhs_mat_01_01 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_01), lhs_mat_ymm_01_01, 1);
2670                    __m512i lhs_mat_23_01 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_01), lhs_mat_ymm_23_01, 1);
2671                    __m512i lhs_mat_01_02 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_02), lhs_mat_ymm_01_02, 1);
2672                    __m512i lhs_mat_23_02 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_02), lhs_mat_ymm_23_02, 1);
2673                    __m512i lhs_mat_01_03 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_03), lhs_mat_ymm_01_03, 1);
2674                    __m512i lhs_mat_23_03 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_03), lhs_mat_ymm_23_03, 1);
2675
2676                    __m512i lhs_mat_01_10 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_10), lhs_mat_ymm_01_10, 1);
2677                    __m512i lhs_mat_23_10 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_10), lhs_mat_ymm_23_10, 1);
2678                    __m512i lhs_mat_01_11 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_11), lhs_mat_ymm_01_11, 1);
2679                    __m512i lhs_mat_23_11 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_11), lhs_mat_ymm_23_11, 1);
2680                    __m512i lhs_mat_01_12 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_12), lhs_mat_ymm_01_12, 1);
2681                    __m512i lhs_mat_23_12 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_12), lhs_mat_ymm_23_12, 1);
2682                    __m512i lhs_mat_01_13 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_13), lhs_mat_ymm_01_13, 1);
2683                    __m512i lhs_mat_23_13 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_13), lhs_mat_ymm_23_13, 1);
2684
2685                    // Bsums are loaded - four bsums are loaded (for two sub blocks) for the different Q8_K blocks
2686                    __m256i lhs_bsums_0123_01 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].bsums + 16 * sb)));
2687                    __m256i lhs_bsums_hsum_ymm_0123_01 = _mm256_castsi128_si256(_mm_hadd_epi16(_mm256_castsi256_si128(lhs_bsums_0123_01), _mm256_extractf128_si256(lhs_bsums_0123_01, 1)));
2688                    lhs_bsums_hsum_ymm_0123_01 = _mm256_permute2x128_si256(lhs_bsums_hsum_ymm_0123_01, lhs_bsums_hsum_ymm_0123_01, 0);
2689                    __m512i lhs_bsums_hsum_0123_01 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_bsums_hsum_ymm_0123_01), lhs_bsums_hsum_ymm_0123_01, 1);
2690
2691                    // Shuffle pattern one - left side input
2692                    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)
2693                    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)
2694                    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)
2695                    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)
2696                    const __m512i lhs_mat_01_02_sp1 = _mm512_shuffle_epi32(lhs_mat_01_02, (_MM_PERM_ENUM)160); //A00(16-19) A00(16-19) A01(16-19) A01(16-19) A00(16-19) A00(16-19) A01(16-19) A01(16-19) A00(16-19) A00(16-19) A01(16-19) A01(16-19) A00(16-19) A00(16-19) A01(16-19) A01(16-19)
2697                    const __m512i lhs_mat_23_02_sp1 = _mm512_shuffle_epi32(lhs_mat_23_02, (_MM_PERM_ENUM)160); //A02(16-19) A02(16-19) A03(16-19) A03(16-19) A02(16-19) A02(16-19) A03(16-19) A03(16-19) A02(16-19) A02(16-19) A03(16-19) A03(16-19) A02(16-19) A02(16-19) A03(16-19) A03(16-19)
2698                    const __m512i lhs_mat_01_03_sp1 = _mm512_shuffle_epi32(lhs_mat_01_03, (_MM_PERM_ENUM)160); //A00(24-27) A00(24-27) A01(24-27) A01(24-27) A00(24-27) A00(24-27) A01(24-27) A01(24-27) A00(24-27) A00(24-27) A01(24-27) A01(24-27) A00(24-27) A00(24-27) A01(24-27) A01(24-27)
2699                    const __m512i lhs_mat_23_03_sp1 = _mm512_shuffle_epi32(lhs_mat_23_03, (_MM_PERM_ENUM)160); //A02(24-27) A02(24-27) A03(24-27) A03(24-27) A02(24-27) A02(24-27) A03(24-27) A03(24-27) A02(24-27) A02(24-27) A03(24-27) A03(24-27) A02(24-27) A02(24-27) A03(24-27) A03(24-27)
2700
2701                    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)
2702                    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)
2703                    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)
2704                    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)
2705                    const __m512i lhs_mat_01_12_sp1 = _mm512_shuffle_epi32(lhs_mat_01_12, (_MM_PERM_ENUM)160); //A10(16-19) A10(16-19) A11(16-19) A11(16-19) A10(16-19) A10(16-19) A11(16-19) A11(16-19) A10(16-19) A10(16-19) A11(16-19) A11(16-19) A10(16-19) A10(16-19) A11(16-19) A11(16-19)
2706                    const __m512i lhs_mat_23_12_sp1 = _mm512_shuffle_epi32(lhs_mat_23_12, (_MM_PERM_ENUM)160); //A12(16-19) A12(16-19) A13(16-19) A13(16-19) A12(16-19) A12(16-19) A13(16-19) A13(16-19) A12(16-19) A12(16-19) A13(16-19) A13(16-19) A12(16-19) A12(16-19) A13(16-19) A13(16-19)
2707                    const __m512i lhs_mat_01_13_sp1 = _mm512_shuffle_epi32(lhs_mat_01_13, (_MM_PERM_ENUM)160); //A10(24-27) A10(24-27) A11(24-27) A11(24-27) A10(24-27) A10(24-27) A11(24-27) A11(24-27) A10(24-27) A10(24-27) A11(24-27) A11(24-27) A10(24-27) A10(24-27) A11(24-27) A11(24-27)
2708                    const __m512i lhs_mat_23_13_sp1 = _mm512_shuffle_epi32(lhs_mat_23_13, (_MM_PERM_ENUM)160); //A12(24-27) A12(24-27) A13(24-27) A13(24-27) A12(24-27) A12(24-27) A13(24-27) A13(24-27) A12(24-27) A12(24-27) A13(24-27) A13(24-27) A12(24-27) A12(24-27) A13(24-27) A13(24-27)
2709
2710                    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)
2711                    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)
2712                    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)
2713                    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)
2714                    const __m512i lhs_mat_01_02_sp2 = _mm512_shuffle_epi32(lhs_mat_01_02, (_MM_PERM_ENUM)245); //A00(20-23) A00(20-23) A01(20-23) A01(20-23) A00(20-23) A00(20-23) A01(20-23) A01(20-23) A00(20-23) A00(20-23) A01(20-23) A01(20-23) A00(20-23) A00(20-23) A01(20-23) A01(20-23)
2715                    const __m512i lhs_mat_23_02_sp2 = _mm512_shuffle_epi32(lhs_mat_23_02, (_MM_PERM_ENUM)245); //A02(20-23) A02(20-23) A03(20-23) A03(20-23) A02(20-23) A02(20-23) A03(20-23) A03(20-23) A02(20-23) A02(20-23) A03(20-23) A03(20-23) A02(20-23) A02(20-23) A03(20-23) A03(20-23)
2716                    const __m512i lhs_mat_01_03_sp2 = _mm512_shuffle_epi32(lhs_mat_01_03, (_MM_PERM_ENUM)245); //A00(28-31) A00(28-31) A01(28-31) A01(28-31) A00(28-31) A00(28-31) A01(28-31) A01(28-31) A00(28-31) A00(28-31) A01(28-31) A01(28-31) A00(28-31) A00(28-31) A01(28-31) A01(28-31)
2717                    const __m512i lhs_mat_23_03_sp2 = _mm512_shuffle_epi32(lhs_mat_23_03, (_MM_PERM_ENUM)245); //A02(28-31) A02(28-31) A03(28-31) A03(28-31) A02(28-31) A02(28-31) A03(28-31) A03(28-31) A02(28-31) A02(28-31) A03(28-31) A03(28-31) A02(28-31) A02(28-31) A03(28-31) A03(28-31)
2718
2719                    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)
2720                    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)
2721                    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)
2722                    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)
2723                    const __m512i lhs_mat_01_12_sp2 = _mm512_shuffle_epi32(lhs_mat_01_12, (_MM_PERM_ENUM)245); //A10(20-23) A10(20-23) A11(20-23) A11(20-23) A10(20-23) A10(20-23) A11(20-23) A11(20-23) A10(20-23) A10(20-23) A11(20-23) A11(20-23) A10(20-23) A10(20-23) A11(20-23) A11(20-23)
2724                    const __m512i lhs_mat_23_12_sp2 = _mm512_shuffle_epi32(lhs_mat_23_12, (_MM_PERM_ENUM)245); //A12(20-23) A12(20-23) A13(20-23) A13(20-23) A12(20-23) A12(20-23) A13(20-23) A13(20-23) A12(20-23) A12(20-23) A13(20-23) A13(20-23) A12(20-23) A12(20-23) A13(20-23) A13(20-23)
2725                    const __m512i lhs_mat_01_13_sp2 = _mm512_shuffle_epi32(lhs_mat_01_13, (_MM_PERM_ENUM)245); //A10(28-31) A10(28-31) A11(28-31) A11(28-31) A10(28-31) A10(28-31) A11(28-31) A11(28-31) A10(28-31) A10(28-31) A11(28-31) A11(28-31) A10(28-31) A10(28-31) A11(28-31) A11(28-31)
2726                    const __m512i lhs_mat_23_13_sp2 = _mm512_shuffle_epi32(lhs_mat_23_13, (_MM_PERM_ENUM)245); //A12(28-31) A12(28-31) A13(28-31) A13(28-31) A12(28-31) A12(28-31) A13(28-31) A13(28-31) A12(28-31) A12(28-31) A13(28-31) A13(28-31) A12(28-31) A12(28-31) A13(28-31) A13(28-31)
2727
2728                    // 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
2729                    __m512i iacc_mat_00_0_sp1 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_03_sp1, lhs_mat_01_03_sp1), _mm512_maddubs_epi16(rhs_mat_014589CD_02_sp1, lhs_mat_01_02_sp1)), _mm512_maddubs_epi16(rhs_mat_014589CD_01_sp1, lhs_mat_01_01_sp1)), _mm512_maddubs_epi16(rhs_mat_014589CD_00_sp1, lhs_mat_01_00_sp1));
2730                    __m512i iacc_mat_01_0_sp1 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_03_sp1, lhs_mat_01_03_sp1), _mm512_maddubs_epi16(rhs_mat_2367ABEF_02_sp1, lhs_mat_01_02_sp1)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_01_sp1, lhs_mat_01_01_sp1)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_00_sp1, lhs_mat_01_00_sp1));
2731                    __m512i iacc_mat_10_0_sp1 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_03_sp1, lhs_mat_23_03_sp1), _mm512_maddubs_epi16(rhs_mat_014589CD_02_sp1, lhs_mat_23_02_sp1)), _mm512_maddubs_epi16(rhs_mat_014589CD_01_sp1, lhs_mat_23_01_sp1)), _mm512_maddubs_epi16(rhs_mat_014589CD_00_sp1, lhs_mat_23_00_sp1));
2732                    __m512i iacc_mat_11_0_sp1 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_03_sp1, lhs_mat_23_03_sp1), _mm512_maddubs_epi16(rhs_mat_2367ABEF_02_sp1, lhs_mat_23_02_sp1)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_01_sp1, lhs_mat_23_01_sp1)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_00_sp1, lhs_mat_23_00_sp1));
2733                    __m512i iacc_mat_00_1_sp1 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_13_sp1, lhs_mat_01_13_sp1), _mm512_maddubs_epi16(rhs_mat_014589CD_12_sp1, lhs_mat_01_12_sp1)), _mm512_maddubs_epi16(rhs_mat_014589CD_11_sp1, lhs_mat_01_11_sp1)), _mm512_maddubs_epi16(rhs_mat_014589CD_10_sp1, lhs_mat_01_10_sp1));
2734                    __m512i iacc_mat_01_1_sp1 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_13_sp1, lhs_mat_01_13_sp1), _mm512_maddubs_epi16(rhs_mat_2367ABEF_12_sp1, lhs_mat_01_12_sp1)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_11_sp1, lhs_mat_01_11_sp1)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_10_sp1, lhs_mat_01_10_sp1));
2735                    __m512i iacc_mat_10_1_sp1 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_13_sp1, lhs_mat_23_13_sp1), _mm512_maddubs_epi16(rhs_mat_014589CD_12_sp1, lhs_mat_23_12_sp1)), _mm512_maddubs_epi16(rhs_mat_014589CD_11_sp1, lhs_mat_23_11_sp1)), _mm512_maddubs_epi16(rhs_mat_014589CD_10_sp1, lhs_mat_23_10_sp1));
2736                    __m512i iacc_mat_11_1_sp1 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_13_sp1, lhs_mat_23_13_sp1), _mm512_maddubs_epi16(rhs_mat_2367ABEF_12_sp1, lhs_mat_23_12_sp1)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_11_sp1, lhs_mat_23_11_sp1)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_10_sp1, lhs_mat_23_10_sp1));
2737
2738                    __m512i iacc_mat_00_0_sp2 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_03_sp2, lhs_mat_01_03_sp2), _mm512_maddubs_epi16(rhs_mat_014589CD_02_sp2, lhs_mat_01_02_sp2)), _mm512_maddubs_epi16(rhs_mat_014589CD_01_sp2, lhs_mat_01_01_sp2)), _mm512_maddubs_epi16(rhs_mat_014589CD_00_sp2, lhs_mat_01_00_sp2));
2739                    __m512i iacc_mat_01_0_sp2 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_03_sp2, lhs_mat_01_03_sp2), _mm512_maddubs_epi16(rhs_mat_2367ABEF_02_sp2, lhs_mat_01_02_sp2)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_01_sp2, lhs_mat_01_01_sp2)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_00_sp2, lhs_mat_01_00_sp2));
2740                    __m512i iacc_mat_10_0_sp2 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_03_sp2, lhs_mat_23_03_sp2), _mm512_maddubs_epi16(rhs_mat_014589CD_02_sp2, lhs_mat_23_02_sp2)), _mm512_maddubs_epi16(rhs_mat_014589CD_01_sp2, lhs_mat_23_01_sp2)), _mm512_maddubs_epi16(rhs_mat_014589CD_00_sp2, lhs_mat_23_00_sp2));
2741                    __m512i iacc_mat_11_0_sp2 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_03_sp2, lhs_mat_23_03_sp2), _mm512_maddubs_epi16(rhs_mat_2367ABEF_02_sp2, lhs_mat_23_02_sp2)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_01_sp2, lhs_mat_23_01_sp2)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_00_sp2, lhs_mat_23_00_sp2));
2742                    __m512i iacc_mat_00_1_sp2 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_13_sp2, lhs_mat_01_13_sp2), _mm512_maddubs_epi16(rhs_mat_014589CD_12_sp2, lhs_mat_01_12_sp2)), _mm512_maddubs_epi16(rhs_mat_014589CD_11_sp2, lhs_mat_01_11_sp2)), _mm512_maddubs_epi16(rhs_mat_014589CD_10_sp2, lhs_mat_01_10_sp2));
2743                    __m512i iacc_mat_01_1_sp2 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_13_sp2, lhs_mat_01_13_sp2), _mm512_maddubs_epi16(rhs_mat_2367ABEF_12_sp2, lhs_mat_01_12_sp2)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_11_sp2, lhs_mat_01_11_sp2)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_10_sp2, lhs_mat_01_10_sp2));
2744                    __m512i iacc_mat_10_1_sp2 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_014589CD_13_sp2, lhs_mat_23_13_sp2), _mm512_maddubs_epi16(rhs_mat_014589CD_12_sp2, lhs_mat_23_12_sp2)), _mm512_maddubs_epi16(rhs_mat_014589CD_11_sp2, lhs_mat_23_11_sp2)), _mm512_maddubs_epi16(rhs_mat_014589CD_10_sp2, lhs_mat_23_10_sp2));
2745                    __m512i iacc_mat_11_1_sp2 = _mm512_add_epi16(_mm512_add_epi16(_mm512_add_epi16(_mm512_maddubs_epi16(rhs_mat_2367ABEF_13_sp2, lhs_mat_23_13_sp2), _mm512_maddubs_epi16(rhs_mat_2367ABEF_12_sp2, lhs_mat_23_12_sp2)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_11_sp2, lhs_mat_23_11_sp2)), _mm512_maddubs_epi16(rhs_mat_2367ABEF_10_sp2, lhs_mat_23_10_sp2));
2746
2747                    // Output of both shuffle patterns are added in order to sum dot product outputs of all 32 values in block
2748                    __m512i iacc_mat_00_0 = _mm512_add_epi16(iacc_mat_00_0_sp1, iacc_mat_00_0_sp2);
2749                    __m512i iacc_mat_01_0 = _mm512_add_epi16(iacc_mat_01_0_sp1, iacc_mat_01_0_sp2);
2750                    __m512i iacc_mat_10_0 = _mm512_add_epi16(iacc_mat_10_0_sp1, iacc_mat_10_0_sp2);
2751                    __m512i iacc_mat_11_0 = _mm512_add_epi16(iacc_mat_11_0_sp1, iacc_mat_11_0_sp2);
2752
2753                    __m512i iacc_mat_00_1 = _mm512_add_epi16(iacc_mat_00_1_sp1, iacc_mat_00_1_sp2);
2754                    __m512i iacc_mat_01_1 = _mm512_add_epi16(iacc_mat_01_1_sp1, iacc_mat_01_1_sp2);
2755                    __m512i iacc_mat_10_1 = _mm512_add_epi16(iacc_mat_10_1_sp1, iacc_mat_10_1_sp2);
2756                    __m512i iacc_mat_11_1 = _mm512_add_epi16(iacc_mat_11_1_sp1, iacc_mat_11_1_sp2);
2757
2758                    iacc_mat_00_0 = _mm512_madd_epi16(iacc_mat_00_0, scale_014589CD_0);
2759                    iacc_mat_01_0 = _mm512_madd_epi16(iacc_mat_01_0, scale_2367ABEF_0);
2760                    iacc_mat_10_0 = _mm512_madd_epi16(iacc_mat_10_0, scale_014589CD_0);
2761                    iacc_mat_11_0 = _mm512_madd_epi16(iacc_mat_11_0, scale_2367ABEF_0);
2762
2763                    iacc_mat_00_1 = _mm512_madd_epi16(iacc_mat_00_1, scale_014589CD_1);
2764                    iacc_mat_01_1 = _mm512_madd_epi16(iacc_mat_01_1, scale_2367ABEF_1);
2765                    iacc_mat_10_1 = _mm512_madd_epi16(iacc_mat_10_1, scale_014589CD_1);
2766                    iacc_mat_11_1 = _mm512_madd_epi16(iacc_mat_11_1, scale_2367ABEF_1);
2767
2768                    // Straighten out to make 4 row vectors (4 for each sub block which are accumulated together in the next step)
2769                    __m512i iacc_row_0_0 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_00_0, _mm512_shuffle_epi32(iacc_mat_01_0, (_MM_PERM_ENUM)78));
2770                    __m512i iacc_row_1_0 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_00_0, (_MM_PERM_ENUM)78), iacc_mat_01_0);
2771                    __m512i iacc_row_2_0 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_10_0, _mm512_shuffle_epi32(iacc_mat_11_0, (_MM_PERM_ENUM)78));
2772                    __m512i iacc_row_3_0 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_10_0, (_MM_PERM_ENUM)78), iacc_mat_11_0);
2773                    __m512i iacc_row_0_1 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_00_1, _mm512_shuffle_epi32(iacc_mat_01_1, (_MM_PERM_ENUM)78));
2774                    __m512i iacc_row_1_1 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_00_1, (_MM_PERM_ENUM)78), iacc_mat_01_1);
2775                    __m512i iacc_row_2_1 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_10_1, _mm512_shuffle_epi32(iacc_mat_11_1, (_MM_PERM_ENUM)78));
2776                    __m512i iacc_row_3_1 = _mm512_mask_blend_epi32(0xCCCC,_mm512_shuffle_epi32(iacc_mat_10_1, (_MM_PERM_ENUM)78), iacc_mat_11_1);
2777
2778                    __m512i iacc_row_0 = _mm512_add_epi32(iacc_row_0_0, iacc_row_0_1);
2779                    __m512i iacc_row_1 = _mm512_add_epi32(iacc_row_1_0, iacc_row_1_1);
2780                    __m512i iacc_row_2 = _mm512_add_epi32(iacc_row_2_0, iacc_row_2_1);
2781                    __m512i iacc_row_3 = _mm512_add_epi32(iacc_row_3_0, iacc_row_3_1);
2782
2783                    // Load the scale(d) values for all the 4 Q8_k blocks and repeat it across lanes
2784                    const __m128 row_scale_f32_sse = _mm_load_ps(a_ptr[b].d);
2785                    const __m256 row_scale_f32_ymm = _mm256_set_m128(row_scale_f32_sse, row_scale_f32_sse);
2786                    const __m512 row_scale_f32 = _mm512_insertf32x8(_mm512_castps256_ps512(row_scale_f32_ymm), row_scale_f32_ymm, 1);
2787
2788                    // Multiply with appropriate scales and accumulate (for both d and dmin) below
2789                    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]);
2790                    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]);
2791                    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]);
2792                    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]);
2793
2794                    __m512i iacc_row_min_0 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_hsum_0123_01, (_MM_PERM_ENUM)0), mins_01);
2795                    __m512i iacc_row_min_1 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_hsum_0123_01, (_MM_PERM_ENUM)85), mins_01);
2796                    __m512i iacc_row_min_2 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_hsum_0123_01, (_MM_PERM_ENUM)170), mins_01);
2797                    __m512i iacc_row_min_3 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_hsum_0123_01, (_MM_PERM_ENUM)255), mins_01);
2798
2799                    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]);
2800                    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]);
2801                    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]);
2802                    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]);
2803                }
2804            }
2805            // Store accumulated values
2806            for (int i = 0; i < 4; i++) {
2807                _mm512_storeu_ps((float * )(s + ((y * 4 + i) * bs + x * 8)), _mm512_sub_ps(acc_rows[i], acc_min_rows[i]));
2808            }
2809        }
2810    }
2811    if (anc != nc) {
2812        xstart = anc/8;
2813        y = 0;
2814    }
2815#endif // __AVX512BW__ && __AVX512DQ__
2816
2817    // Take group of four block_q8_Kx4 structures at each pass of the loop and perform dot product operation
2818    for (; y < anr / 4; y += 4) {
2819
2820        const block_q8_Kx4 * a_ptrs[4];
2821
2822        a_ptrs[0] = a_ptr_start + (y * nb);
2823        for (int i = 0; i < 3; ++i) {
2824            a_ptrs[i + 1] = a_ptrs[i] + nb;
2825        }
2826
2827        // Take group of eight block_q4_kx8 structures at each pass of the loop and perform dot product operation
2828        for (int64_t x = xstart; x < nc / 8; x++) {
2829
2830            const block_q4_Kx8 * b_ptr = b_ptr_start + (x * b_nb);
2831
2832            // Master FP accumulators
2833            __m256 acc_rows[16];
2834            for (int i = 0; i < 16; i++) {
2835                acc_rows[i] = _mm256_setzero_ps();
2836            }
2837
2838            __m256 acc_min_rows[16];
2839            for (int i = 0; i < 16; i++) {
2840                acc_min_rows[i] = _mm256_setzero_ps();
2841            }
2842
2843            // For super block
2844            for (int64_t b = 0; b < nb; b++) {
2845
2846                // Scale values - Load the eight scale values of block_q4_kx8
2847                const __m256 col_scale_f32 = GGML_F32Cx8_LOAD(b_ptr[b].d);
2848
2849                // dmin values - Load the eight dmin values of block_q4_kx8
2850                const __m256 col_dmin_f32 = GGML_F32Cx8_LOAD(b_ptr[b].dmin);
2851
2852                // Loop to iterate over the eight sub blocks of a super block - two sub blocks are processed per iteration
2853                for (int sb = 0; sb < QK_K / 64; sb++) {
2854
2855                    // Load the eight block_q4_K for two sub blocks quantized values interleaved with each other in chunks of eight bytes - B0,B1 ....B6,B7
2856                    const __m256i rhs_raw_mat_0123_0 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + sb * 256));
2857                    const __m256i rhs_raw_mat_4567_0 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 32 + sb * 256));
2858                    const __m256i rhs_raw_mat_0123_1 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 64 + sb * 256));
2859                    const __m256i rhs_raw_mat_4567_1 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 96 + sb * 256));
2860                    const __m256i rhs_raw_mat_0123_2 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 128 + sb * 256));
2861                    const __m256i rhs_raw_mat_4567_2 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 160 + sb * 256));
2862                    const __m256i rhs_raw_mat_0123_3 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 192 + sb * 256));
2863                    const __m256i rhs_raw_mat_4567_3 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 224 + sb * 256));
2864
2865                    // Save the values in the following vectors in the formats B0B1B4B5, B2B3B6B7 for further processing and storing of values
2866                    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);
2867                    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);
2868                    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);
2869                    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);
2870                    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);
2871                    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);
2872                    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);
2873                    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);
2874
2875                    // 4-bit -> 8-bit
2876                    // First sub block of the two sub blocks processed in the iteration
2877                    const __m256i rhs_mat_0145_00 = _mm256_and_si256(rhs_raw_mat_0145_0, m4b); //B00(0-7) B01(0-7) B04(0-7) B05(0-7)
2878                    const __m256i rhs_mat_2367_00 = _mm256_and_si256(rhs_raw_mat_2367_0, m4b); //B02(0-7) B03(0-7) B06(0-7) B07(0-7)
2879
2880                    const __m256i rhs_mat_0145_01 = _mm256_and_si256(rhs_raw_mat_0145_1, m4b); //B00(8-15) B01(8-15) B04(8-15) B05(8-15)
2881                    const __m256i rhs_mat_2367_01 = _mm256_and_si256(rhs_raw_mat_2367_1, m4b); //B02(8-15) B03(8-15) B06(8-15) B07(8-15)
2882
2883                    const __m256i rhs_mat_0145_02 = _mm256_and_si256(rhs_raw_mat_0145_2, m4b); //B00(16-23) B01(16-23) B04(16-23) B05(16-23)
2884                    const __m256i rhs_mat_2367_02 = _mm256_and_si256(rhs_raw_mat_2367_2, m4b); //B02(16-23) B03(16-23) B06(16-23) B07(16-23)
2885
2886                    const __m256i rhs_mat_0145_03 = _mm256_and_si256(rhs_raw_mat_0145_3, m4b); //B00(24-31) B01(24-31) B04(24-31) B05(24-31)
2887                    const __m256i rhs_mat_2367_03 = _mm256_and_si256(rhs_raw_mat_2367_3, m4b); //B02(24-31) B03(24-31) B06(24-31) B07(24-31)
2888
2889                    // Second sub block of the two sub blocks processed in the iteration
2890                    const __m256i rhs_mat_0145_10 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_0, 4), m4b); //B10(0-7) B11(0-7) B14(0-7) B15(0-7)
2891                    const __m256i rhs_mat_2367_10 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_0, 4), m4b); //B12(0-7) B13(0-7) B16(0-7) B17(0-7)
2892
2893                    const __m256i rhs_mat_0145_11 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_1, 4), m4b); //B10(8-15) B11(8-15) B14(8-15) B15(8-15)
2894                    const __m256i rhs_mat_2367_11 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_1, 4), m4b); //B12(8-15) B13(8-15) B16(8-15) B17(8-15)
2895
2896                    const __m256i rhs_mat_0145_12 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_2, 4), m4b); //B10(16-23) B11(16-23) B14(16-23) B15(16-23)
2897                    const __m256i rhs_mat_2367_12 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_2, 4), m4b); //B12(16-23) B13(16-23) B16(16-23) B17(16-23)
2898
2899                    const __m256i rhs_mat_0145_13 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_3, 4), m4b); //B10(24-31) B11(24-31) B14(24-31) B15(24-31)
2900                    const __m256i rhs_mat_2367_13 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_3, 4), m4b); //B12(24-31) B13(24-31) B16(24-31) B17(24-31)
2901
2902                    // Shuffle pattern one - right side input
2903                    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)
2904                    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)
2905
2906                    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)
2907                    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)
2908
2909                    const __m256i rhs_mat_0145_02_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_02, 136); //B00(16-19) B01(16-19) B00(16-19) B01(16-19) B04(16-19) B05(16-19) B04(16-19) B05(16-19)
2910                    const __m256i rhs_mat_2367_02_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_02, 136); //B02(16-19) B03(16-19) B02(16-19) B03(16-19) B06(16-19) B07(16-19) B06(16-19) B07(16-19)
2911
2912                    const __m256i rhs_mat_0145_03_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_03, 136); //B00(24-27) B01(24-27) B00(24-27) B01(24-27) B04(24-27) B05(24-27) B04(24-27) B05(24-27)
2913                    const __m256i rhs_mat_2367_03_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_03, 136); //B02(24-27) B03(24-27) B02(24-27) B03(24-27) B06(24-27) B07(24-27) B06(24-27) B07(24-27)
2914
2915                    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)
2916                    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)
2917
2918                    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)
2919                    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)
2920
2921                    const __m256i rhs_mat_0145_12_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_12, 136); //B10(16-19) B11(16-19) B10(16-19) B11(16-19) B14(16-19) B15(16-19) B14(16-19) B15(16-19)
2922                    const __m256i rhs_mat_2367_12_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_12, 136); //B12(16-19) B13(16-19) B12(16-19) B13(16-19) B16(16-19) B17(16-19) B16(16-19) B17(16-19)
2923
2924                    const __m256i rhs_mat_0145_13_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_13, 136); //B10(24-27) B11(24-27) B10(24-27) B11(24-27) B14(24-27) B15(24-27) B14(24-27) B15(24-27)
2925                    const __m256i rhs_mat_2367_13_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_13, 136); //B12(24-27) B13(24-27) B12(24-27) B13(24-27) B16(24-27) B17(24-27) B16(24-27) B17(24-27)
2926
2927
2928                    // Shuffle pattern two - right side input
2929                    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)
2930                    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)
2931
2932                    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)
2933                    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)
2934
2935                    const __m256i rhs_mat_0145_02_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_02, 221); //B00(20-23) B01(20-23) B00(20-23) B01(20-23) B04(20-23) B05(20-23) B04(20-23) B05(20-23)
2936                    const __m256i rhs_mat_2367_02_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_02, 221); //B02(20-23) B03(20-23) B02(20-23) B03(20-23) B06(20-23) B07(20-23) B06(20-23) B07(20-23)
2937
2938                    const __m256i rhs_mat_0145_03_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_03, 221); //B00(28-31) B01(28-31) B00(28-31) B01(28-31) B04(28-31) B05(28-31) B04(28-31) B05(28-31)
2939                    const __m256i rhs_mat_2367_03_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_03, 221); //B02(28-31) B03(28-31) B02(28-31) B03(28-31) B06(28-31) B07(28-31) B06(28-31) B07(28-31)
2940
2941                    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)
2942                    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)
2943
2944                    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)
2945                    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)
2946
2947                    const __m256i rhs_mat_0145_12_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_12, 221); //B10(20-23) B11(20-23) B10(20-23) B11(20-23) B14(20-23) B15(20-23) B14(20-23) B15(20-23)
2948                    const __m256i rhs_mat_2367_12_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_12, 221); //B12(20-23) B13(20-23) B12(20-23) B13(20-23) B16(20-23) B17(20-23) B16(20-23) B17(20-23)
2949
2950                    const __m256i rhs_mat_0145_13_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_13, 221); //B10(28-31) B11(28-31) B10(28-31) B11(28-31) B14(28-31) B15(28-31) B14(28-31) B15(28-31)
2951                    const __m256i rhs_mat_2367_13_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_13, 221); //B12(28-31) B13(28-31) B12(28-31) B13(28-31) B16(28-31) B17(28-31) B16(28-31) B17(28-31)
2952
2953                    uint32_t utmp_0[4], utmp_1[4];
2954
2955                    // Scales and Mins of corresponding sub blocks from different Q4_K structures are stored together
2956                    // The below block is for eg to extract first sub block's scales and mins from different Q4_K structures for the sb loop
2957                    memcpy(utmp_0, b_ptr[b].scales + 24 * sb, 12);
2958                    utmp_0[3] = ((utmp_0[2] >> 4) & kmask2) | (((utmp_0[1] >> 6) & kmask3) << 4);
2959                    const uint32_t uaux_0 = utmp_0[1] & kmask1;
2960                    utmp_0[1] = (utmp_0[2] & kmask2) | (((utmp_0[0] >> 6) & kmask3) << 4);
2961                    utmp_0[2] = uaux_0;
2962                    utmp_0[0] &= kmask1;
2963
2964                    // The below block is for eg to extract second sub block's scales and mins from different Q4_K structures for the sb loop
2965                    memcpy(utmp_1, b_ptr[b].scales + 12 + sb * 24, 12);
2966                    utmp_1[3] = ((utmp_1[2] >> 4) & kmask2) | (((utmp_1[1] >> 6) & kmask3) << 4);
2967                    const uint32_t uaux_1 = utmp_1[1] & kmask1;
2968                    utmp_1[1] = (utmp_1[2] & kmask2) | (((utmp_1[0] >> 6) & kmask3) << 4);
2969                    utmp_1[2] = uaux_1;
2970                    utmp_1[0] &= kmask1;
2971
2972                    // Scales of first sub block in the sb loop
2973                    const __m128i mins_and_scales_0 = _mm_set_epi32(utmp_0[3], utmp_0[2], utmp_0[1], utmp_0[0]);
2974                    const __m256i scales_0 = _mm256_cvtepu8_epi16(_mm_unpacklo_epi8(mins_and_scales_0, mins_and_scales_0));
2975
2976                    // Scales of second sub block in the sb loop
2977                    const __m128i mins_and_scales_1 = _mm_set_epi32(utmp_1[3], utmp_1[2], utmp_1[1], utmp_1[0]);
2978                    const __m256i scales_1 = _mm256_cvtepu8_epi16(_mm_unpacklo_epi8(mins_and_scales_1, mins_and_scales_1));
2979
2980                    // Mins of first and second sub block of Q4_K block are arranged side by side
2981                    const __m256i mins_01 = _mm256_cvtepu8_epi16(_mm_unpacklo_epi8(_mm_shuffle_epi32(mins_and_scales_0, 78), _mm_shuffle_epi32(mins_and_scales_1, 78)));
2982
2983                    const __m256i scale_0145_0 = _mm256_shuffle_epi32(scales_0, 68);
2984                    const __m256i scale_2367_0 = _mm256_shuffle_epi32(scales_0, 238);
2985
2986                    const __m256i scale_0145_1 = _mm256_shuffle_epi32(scales_1, 68);
2987                    const __m256i scale_2367_1 = _mm256_shuffle_epi32(scales_1, 238);
2988
2989                    for (int rp = 0; rp < 4; rp++) {
2990
2991                        // Load the four block_q8_k quantized values interleaved with each other in chunks of eight bytes - A0,A1,A2,A3
2992                        // Loaded as set of 128 bit vectors and repeated into a 256 bit vector
2993                        __m256i lhs_mat_0123_00 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 256 * sb)));
2994                        __m256i lhs_mat_01_00 = _mm256_permute2f128_si256(lhs_mat_0123_00, lhs_mat_0123_00, 0);
2995                        __m256i lhs_mat_23_00 = _mm256_permute2f128_si256(lhs_mat_0123_00, lhs_mat_0123_00, 17);
2996                        __m256i lhs_mat_0123_01 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 32 + 256 * sb)));
2997                        __m256i lhs_mat_01_01 = _mm256_permute2f128_si256(lhs_mat_0123_01, lhs_mat_0123_01, 0);
2998                        __m256i lhs_mat_23_01 = _mm256_permute2f128_si256(lhs_mat_0123_01, lhs_mat_0123_01, 17);
2999                        __m256i lhs_mat_0123_02 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 64 + 256 * sb)));
3000                        __m256i lhs_mat_01_02 = _mm256_permute2f128_si256(lhs_mat_0123_02, lhs_mat_0123_02, 0);
3001                        __m256i lhs_mat_23_02 = _mm256_permute2f128_si256(lhs_mat_0123_02, lhs_mat_0123_02, 17);
3002                        __m256i lhs_mat_0123_03 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 96 + 256 * sb)));
3003                        __m256i lhs_mat_01_03 = _mm256_permute2f128_si256(lhs_mat_0123_03, lhs_mat_0123_03, 0);
3004                        __m256i lhs_mat_23_03 = _mm256_permute2f128_si256(lhs_mat_0123_03, lhs_mat_0123_03, 17);
3005                        __m256i lhs_mat_0123_10 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 128 + 256 * sb)));
3006                        __m256i lhs_mat_01_10 = _mm256_permute2f128_si256(lhs_mat_0123_10, lhs_mat_0123_10, 0);
3007                        __m256i lhs_mat_23_10 = _mm256_permute2f128_si256(lhs_mat_0123_10, lhs_mat_0123_10, 17);
3008                        __m256i lhs_mat_0123_11 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 160 + 256 * sb)));
3009                        __m256i lhs_mat_01_11 = _mm256_permute2f128_si256(lhs_mat_0123_11, lhs_mat_0123_11, 0);
3010                        __m256i lhs_mat_23_11 = _mm256_permute2f128_si256(lhs_mat_0123_11, lhs_mat_0123_11, 17);
3011                        __m256i lhs_mat_0123_12 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 192 + 256 * sb)));
3012                        __m256i lhs_mat_01_12 = _mm256_permute2f128_si256(lhs_mat_0123_12, lhs_mat_0123_12, 0);
3013                        __m256i lhs_mat_23_12 = _mm256_permute2f128_si256(lhs_mat_0123_12, lhs_mat_0123_12, 17);
3014                        __m256i lhs_mat_0123_13 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 224 + 256 * sb)));
3015                        __m256i lhs_mat_01_13 = _mm256_permute2f128_si256(lhs_mat_0123_13, lhs_mat_0123_13, 0);
3016                        __m256i lhs_mat_23_13 = _mm256_permute2f128_si256(lhs_mat_0123_13, lhs_mat_0123_13, 17);
3017
3018                        // Bsums are loaded - four bsums are loaded (for two sub blocks) for the different Q8_K blocks
3019                        __m256i lhs_bsums_0123_01 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].bsums + 16 * sb)));
3020                        __m256i lhs_bsums_hsum_0123_01 = _mm256_castsi128_si256(_mm_hadd_epi16(_mm256_castsi256_si128(lhs_bsums_0123_01), _mm256_extractf128_si256(lhs_bsums_0123_01, 1)));
3021                        lhs_bsums_hsum_0123_01 = _mm256_permute2x128_si256(lhs_bsums_hsum_0123_01, lhs_bsums_hsum_0123_01, 0);
3022
3023                        // Shuffle pattern one - left side input
3024                        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)
3025                        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)
3026
3027                        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)
3028                        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)
3029
3030                        const __m256i lhs_mat_01_02_sp1 = _mm256_shuffle_epi32(lhs_mat_01_02, 160);  //A00(16-19) A00(16-19) A01(16-19) A01(16-19) A00(16-19) A00(16-19) A01(16-19) A01(16-19)
3031                        const __m256i lhs_mat_23_02_sp1 = _mm256_shuffle_epi32(lhs_mat_23_02, 160);  //A02(16-19) A03(16-19) A02(16-19) A03(16-19) A02(16-19) A03(16-19) A02(16-19) A03(16-19)
3032
3033                        const __m256i lhs_mat_01_03_sp1 = _mm256_shuffle_epi32(lhs_mat_01_03, 160);  //A00(24-27) A00(24-27) A01(24-27) A01(24-27) A00(24-27) A00(24-27) A01(24-27) A01(24-27)
3034                        const __m256i lhs_mat_23_03_sp1 = _mm256_shuffle_epi32(lhs_mat_23_03, 160); //A02(24-27) A03(24-27) A02(24-27) A03(24-27) A02(24-27) A03(24-27) A02(24-27) A03(24-27)
3035
3036                        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)
3037                        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)
3038
3039                        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)
3040                        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)
3041
3042                        const __m256i lhs_mat_01_12_sp1 = _mm256_shuffle_epi32(lhs_mat_01_12, 160);  //A10(16-19) A10(16-19) A11(16-19) A11(16-19) A10(16-19) A10(16-19) A11(16-19) A11(16-19)
3043                        const __m256i lhs_mat_23_12_sp1 = _mm256_shuffle_epi32(lhs_mat_23_12, 160);  //A12(16-19) A13(16-19) A12(16-19) A13(16-19) A12(16-19) A13(16-19) A12(16-19) A13(16-19)
3044
3045                        const __m256i lhs_mat_01_13_sp1 = _mm256_shuffle_epi32(lhs_mat_01_13, 160); //A10(24-27) A10(24-27) A11(24-27) A11(24-27) A10(24-27) A10(24-27) A11(24-27) A11(24-27)
3046                        const __m256i lhs_mat_23_13_sp1 = _mm256_shuffle_epi32(lhs_mat_23_13, 160); //A12(24-27) A13(24-27) A12(24-27) A13(24-27) A12(24-27) A13(24-27) A12(24-27) A13(24-27)
3047
3048                        // Shuffle pattern two- left side input
3049                        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)
3050                        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)
3051
3052                        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)
3053                        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)
3054
3055                        const __m256i lhs_mat_01_02_sp2 = _mm256_shuffle_epi32(lhs_mat_01_02, 245); //A00(20-23) A00(20-23) A01(20-23) A01(20-23) A00(20-23) A00(20-23) A01(20-23) A01(20-23)
3056                        const __m256i lhs_mat_23_02_sp2 = _mm256_shuffle_epi32(lhs_mat_23_02, 245); //A02(20-23) A03(20-23) A02(20-23) A03(20-23) A02(20-23) A03(20-23) A02(20-23) A03(20-23)
3057
3058                        const __m256i lhs_mat_01_03_sp2 = _mm256_shuffle_epi32(lhs_mat_01_03, 245); //A00(28-31) A00(28-31) A01(28-31) A01(28-31) A00(28-31) A00(28-31) A01(28-31) A01(28-31)
3059                        const __m256i lhs_mat_23_03_sp2 = _mm256_shuffle_epi32(lhs_mat_23_03, 245); //A02(28-31) A03(28-31) A02(28-31) A03(28-31) A02(28-31) A03(28-31) A02(28-31) A03(28-31)
3060
3061                        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)
3062                        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)
3063
3064                        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)
3065                        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)
3066
3067                        const __m256i lhs_mat_01_12_sp2 = _mm256_shuffle_epi32(lhs_mat_01_12, 245); //A10(20-23) A10(20-23) A11(20-23) A11(20-23) A10(20-23) A10(20-23) A11(20-23) A11(20-23)
3068                        const __m256i lhs_mat_23_12_sp2 = _mm256_shuffle_epi32(lhs_mat_23_12, 245); //A12(20-23) A13(20-23) A12(20-23) A13(20-23) A12(20-23) A13(20-23) A12(20-23) A13(20-23)
3069
3070                        const __m256i lhs_mat_01_13_sp2 = _mm256_shuffle_epi32(lhs_mat_01_13, 245); //A10(28-31) A10(28-31) A11(28-31) A11(28-31) A10(28-31) A10(28-31) A11(28-31) A11(28-31)
3071                        const __m256i lhs_mat_23_13_sp2 = _mm256_shuffle_epi32(lhs_mat_23_13, 245); //A12(28-31) A13(28-31) A12(28-31) A13(28-31) A12(28-31) A13(28-31) A12(28-31) A13(28-31)
3072
3073                        // 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
3074                        __m256i iacc_mat_00_0_sp1 = _mm256_add_epi16(_mm256_add_epi16(_mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_03_sp1, lhs_mat_01_03_sp1), _mm256_maddubs_epi16(rhs_mat_0145_02_sp1, lhs_mat_01_02_sp1)), _mm256_maddubs_epi16(rhs_mat_0145_01_sp1, lhs_mat_01_01_sp1)), _mm256_maddubs_epi16(rhs_mat_0145_00_sp1, lhs_mat_01_00_sp1));
3075                        __m256i iacc_mat_01_0_sp1 = _mm256_add_epi16(_mm256_add_epi16(_mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_03_sp1, lhs_mat_01_03_sp1), _mm256_maddubs_epi16(rhs_mat_2367_02_sp1, lhs_mat_01_02_sp1)), _mm256_maddubs_epi16(rhs_mat_2367_01_sp1, lhs_mat_01_01_sp1)), _mm256_maddubs_epi16(rhs_mat_2367_00_sp1, lhs_mat_01_00_sp1));
3076                        __m256i iacc_mat_10_0_sp1 = _mm256_add_epi16(_mm256_add_epi16(_mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_03_sp1, lhs_mat_23_03_sp1), _mm256_maddubs_epi16(rhs_mat_0145_02_sp1, lhs_mat_23_02_sp1)), _mm256_maddubs_epi16(rhs_mat_0145_01_sp1, lhs_mat_23_01_sp1)), _mm256_maddubs_epi16(rhs_mat_0145_00_sp1, lhs_mat_23_00_sp1));
3077                        __m256i iacc_mat_11_0_sp1 = _mm256_add_epi16(_mm256_add_epi16(_mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_03_sp1, lhs_mat_23_03_sp1), _mm256_maddubs_epi16(rhs_mat_2367_02_sp1, lhs_mat_23_02_sp1)), _mm256_maddubs_epi16(rhs_mat_2367_01_sp1, lhs_mat_23_01_sp1)), _mm256_maddubs_epi16(rhs_mat_2367_00_sp1, lhs_mat_23_00_sp1));
3078                        __m256i iacc_mat_00_1_sp1 = _mm256_add_epi16(_mm256_add_epi16(_mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_13_sp1, lhs_mat_01_13_sp1), _mm256_maddubs_epi16(rhs_mat_0145_12_sp1, lhs_mat_01_12_sp1)), _mm256_maddubs_epi16(rhs_mat_0145_11_sp1, lhs_mat_01_11_sp1)), _mm256_maddubs_epi16(rhs_mat_0145_10_sp1, lhs_mat_01_10_sp1));
3079                        __m256i iacc_mat_01_1_sp1 = _mm256_add_epi16(_mm256_add_epi16(_mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_13_sp1, lhs_mat_01_13_sp1), _mm256_maddubs_epi16(rhs_mat_2367_12_sp1, lhs_mat_01_12_sp1)), _mm256_maddubs_epi16(rhs_mat_2367_11_sp1, lhs_mat_01_11_sp1)), _mm256_maddubs_epi16(rhs_mat_2367_10_sp1, lhs_mat_01_10_sp1));
3080                        __m256i iacc_mat_10_1_sp1 = _mm256_add_epi16(_mm256_add_epi16(_mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_13_sp1, lhs_mat_23_13_sp1), _mm256_maddubs_epi16(rhs_mat_0145_12_sp1, lhs_mat_23_12_sp1)), _mm256_maddubs_epi16(rhs_mat_0145_11_sp1, lhs_mat_23_11_sp1)), _mm256_maddubs_epi16(rhs_mat_0145_10_sp1, lhs_mat_23_10_sp1));
3081                        __m256i iacc_mat_11_1_sp1 = _mm256_add_epi16(_mm256_add_epi16(_mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_13_sp1, lhs_mat_23_13_sp1), _mm256_maddubs_epi16(rhs_mat_2367_12_sp1, lhs_mat_23_12_sp1)), _mm256_maddubs_epi16(rhs_mat_2367_11_sp1, lhs_mat_23_11_sp1)), _mm256_maddubs_epi16(rhs_mat_2367_10_sp1, lhs_mat_23_10_sp1));
3082
3083                        __m256i iacc_mat_00_0_sp2 = _mm256_add_epi16(_mm256_add_epi16(_mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_03_sp2, lhs_mat_01_03_sp2), _mm256_maddubs_epi16(rhs_mat_0145_02_sp2, lhs_mat_01_02_sp2)), _mm256_maddubs_epi16(rhs_mat_0145_01_sp2, lhs_mat_01_01_sp2)), _mm256_maddubs_epi16(rhs_mat_0145_00_sp2, lhs_mat_01_00_sp2));
3084                        __m256i iacc_mat_01_0_sp2 = _mm256_add_epi16(_mm256_add_epi16(_mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_03_sp2, lhs_mat_01_03_sp2), _mm256_maddubs_epi16(rhs_mat_2367_02_sp2, lhs_mat_01_02_sp2)), _mm256_maddubs_epi16(rhs_mat_2367_01_sp2, lhs_mat_01_01_sp2)), _mm256_maddubs_epi16(rhs_mat_2367_00_sp2, lhs_mat_01_00_sp2));
3085                        __m256i iacc_mat_10_0_sp2 = _mm256_add_epi16(_mm256_add_epi16(_mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_03_sp2, lhs_mat_23_03_sp2), _mm256_maddubs_epi16(rhs_mat_0145_02_sp2, lhs_mat_23_02_sp2)), _mm256_maddubs_epi16(rhs_mat_0145_01_sp2, lhs_mat_23_01_sp2)), _mm256_maddubs_epi16(rhs_mat_0145_00_sp2, lhs_mat_23_00_sp2));
3086                        __m256i iacc_mat_11_0_sp2 = _mm256_add_epi16(_mm256_add_epi16(_mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_03_sp2, lhs_mat_23_03_sp2), _mm256_maddubs_epi16(rhs_mat_2367_02_sp2, lhs_mat_23_02_sp2)), _mm256_maddubs_epi16(rhs_mat_2367_01_sp2, lhs_mat_23_01_sp2)), _mm256_maddubs_epi16(rhs_mat_2367_00_sp2, lhs_mat_23_00_sp2));
3087                        __m256i iacc_mat_00_1_sp2 = _mm256_add_epi16(_mm256_add_epi16(_mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_13_sp2, lhs_mat_01_13_sp2), _mm256_maddubs_epi16(rhs_mat_0145_12_sp2, lhs_mat_01_12_sp2)), _mm256_maddubs_epi16(rhs_mat_0145_11_sp2, lhs_mat_01_11_sp2)), _mm256_maddubs_epi16(rhs_mat_0145_10_sp2, lhs_mat_01_10_sp2));
3088                        __m256i iacc_mat_01_1_sp2 = _mm256_add_epi16(_mm256_add_epi16(_mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_13_sp2, lhs_mat_01_13_sp2), _mm256_maddubs_epi16(rhs_mat_2367_12_sp2, lhs_mat_01_12_sp2)), _mm256_maddubs_epi16(rhs_mat_2367_11_sp2, lhs_mat_01_11_sp2)), _mm256_maddubs_epi16(rhs_mat_2367_10_sp2, lhs_mat_01_10_sp2));
3089                        __m256i iacc_mat_10_1_sp2 = _mm256_add_epi16(_mm256_add_epi16(_mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_13_sp2, lhs_mat_23_13_sp2), _mm256_maddubs_epi16(rhs_mat_0145_12_sp2, lhs_mat_23_12_sp2)), _mm256_maddubs_epi16(rhs_mat_0145_11_sp2, lhs_mat_23_11_sp2)), _mm256_maddubs_epi16(rhs_mat_0145_10_sp2, lhs_mat_23_10_sp2));
3090                        __m256i iacc_mat_11_1_sp2 = _mm256_add_epi16(_mm256_add_epi16(_mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_13_sp2, lhs_mat_23_13_sp2), _mm256_maddubs_epi16(rhs_mat_2367_12_sp2, lhs_mat_23_12_sp2)), _mm256_maddubs_epi16(rhs_mat_2367_11_sp2, lhs_mat_23_11_sp2)), _mm256_maddubs_epi16(rhs_mat_2367_10_sp2, lhs_mat_23_10_sp2));
3091
3092                        // Output of both shuffle patterns are added in order to sum dot product outputs of all 32 values in block
3093                        __m256i iacc_mat_00_0 = _mm256_add_epi16(iacc_mat_00_0_sp1, iacc_mat_00_0_sp2);
3094                        __m256i iacc_mat_01_0 = _mm256_add_epi16(iacc_mat_01_0_sp1, iacc_mat_01_0_sp2);
3095                        __m256i iacc_mat_10_0 = _mm256_add_epi16(iacc_mat_10_0_sp1, iacc_mat_10_0_sp2);
3096                        __m256i iacc_mat_11_0 = _mm256_add_epi16(iacc_mat_11_0_sp1, iacc_mat_11_0_sp2);
3097
3098                        __m256i iacc_mat_00_1 = _mm256_add_epi16(iacc_mat_00_1_sp1, iacc_mat_00_1_sp2);
3099                        __m256i iacc_mat_01_1 = _mm256_add_epi16(iacc_mat_01_1_sp1, iacc_mat_01_1_sp2);
3100                        __m256i iacc_mat_10_1 = _mm256_add_epi16(iacc_mat_10_1_sp1, iacc_mat_10_1_sp2);
3101                        __m256i iacc_mat_11_1 = _mm256_add_epi16(iacc_mat_11_1_sp1, iacc_mat_11_1_sp2);
3102
3103                        // Output of both shuffle patterns are added in order to sum dot product outputs of all 32 values in block
3104                        iacc_mat_00_0 = _mm256_madd_epi16(iacc_mat_00_0, scale_0145_0);
3105                        iacc_mat_01_0 = _mm256_madd_epi16(iacc_mat_01_0, scale_2367_0);
3106                        iacc_mat_10_0 = _mm256_madd_epi16(iacc_mat_10_0, scale_0145_0);
3107                        iacc_mat_11_0 = _mm256_madd_epi16(iacc_mat_11_0, scale_2367_0);
3108
3109                        iacc_mat_00_1 = _mm256_madd_epi16(iacc_mat_00_1, scale_0145_1);
3110                        iacc_mat_01_1 = _mm256_madd_epi16(iacc_mat_01_1, scale_2367_1);
3111                        iacc_mat_10_1 = _mm256_madd_epi16(iacc_mat_10_1, scale_0145_1);
3112                        iacc_mat_11_1 = _mm256_madd_epi16(iacc_mat_11_1, scale_2367_1);
3113
3114                        // Straighten out to make 4 row vectors (4 for each sub block which are accumulated together in the next step)
3115                        __m256i iacc_row_0_0 = _mm256_blend_epi32(iacc_mat_00_0, _mm256_shuffle_epi32(iacc_mat_01_0, 78), 204);
3116                        __m256i iacc_row_1_0 = _mm256_blend_epi32(_mm256_shuffle_epi32(iacc_mat_00_0, 78), iacc_mat_01_0, 204);
3117                        __m256i iacc_row_2_0 = _mm256_blend_epi32(iacc_mat_10_0, _mm256_shuffle_epi32(iacc_mat_11_0, 78), 204);
3118                        __m256i iacc_row_3_0 = _mm256_blend_epi32(_mm256_shuffle_epi32(iacc_mat_10_0, 78), iacc_mat_11_0, 204);
3119                        __m256i iacc_row_0_1 = _mm256_blend_epi32(iacc_mat_00_1, _mm256_shuffle_epi32(iacc_mat_01_1, 78), 204);
3120                        __m256i iacc_row_1_1 = _mm256_blend_epi32(_mm256_shuffle_epi32(iacc_mat_00_1, 78), iacc_mat_01_1, 204);
3121                        __m256i iacc_row_2_1 = _mm256_blend_epi32(iacc_mat_10_1, _mm256_shuffle_epi32(iacc_mat_11_1, 78), 204);
3122                        __m256i iacc_row_3_1 = _mm256_blend_epi32(_mm256_shuffle_epi32(iacc_mat_10_1, 78), iacc_mat_11_1, 204);
3123
3124                        __m256i iacc_row_0 = _mm256_add_epi32(iacc_row_0_0, iacc_row_0_1);
3125                        __m256i iacc_row_1 = _mm256_add_epi32(iacc_row_1_0, iacc_row_1_1);
3126                        __m256i iacc_row_2 = _mm256_add_epi32(iacc_row_2_0, iacc_row_2_1);
3127                        __m256i iacc_row_3 = _mm256_add_epi32(iacc_row_3_0, iacc_row_3_1);
3128
3129                        // Load the scale(d) values for all the 4 Q8_k blocks and repeat it across lanes
3130                        const __m128 row_scale_f32_sse = _mm_load_ps(a_ptrs[rp][b].d);
3131                        const __m256 row_scale_f32 = _mm256_set_m128(row_scale_f32_sse, row_scale_f32_sse);//GGML_F32Cx8_REPEAT_LOAD(a_ptrs[rp][b].d, loadMask);
3132
3133                        // Multiply with appropriate scales and accumulate (for both d and dmin) below
3134                        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]);
3135                        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]);
3136                        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]);
3137                        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]);
3138
3139                        __m256i iacc_row_min_0 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_hsum_0123_01, 0), mins_01);
3140                        __m256i iacc_row_min_1 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_hsum_0123_01, 85), mins_01);
3141                        __m256i iacc_row_min_2 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_hsum_0123_01, 170), mins_01);
3142                        __m256i iacc_row_min_3 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_hsum_0123_01, 255), mins_01);
3143
3144                        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]);
3145                        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]);
3146                        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]);
3147                        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]);
3148
3149                    }
3150                }
3151            }
3152            // Store the accumulated values
3153            for (int i = 0; i < 16; i++) {
3154                _mm256_storeu_ps((float * )(s + ((y * 4 + i) * bs + x * 8)), _mm256_sub_ps(acc_rows[i], acc_min_rows[i]));
3155            }
3156        }
3157    }
3158    for (; y < nr / 4; y++) {
3159
3160        const block_q8_Kx4 * a_ptr = a_ptr_start + (y * nb);
3161
3162        for (int64_t x = xstart; x < nc / 8; x++) {
3163
3164            const block_q4_Kx8 * b_ptr = b_ptr_start + (x * b_nb);
3165
3166            // Master FP accumulators
3167            __m256 acc_rows[4];
3168            for (int i = 0; i < 4; i++) {
3169                acc_rows[i] = _mm256_setzero_ps();
3170            }
3171
3172            __m256 acc_min_rows[4];
3173            for (int i = 0; i < 4; i++) {
3174                acc_min_rows[i] = _mm256_setzero_ps();
3175            }
3176
3177            for (int64_t b = 0; b < nb; b++) {
3178
3179                // Scale values - Load the eight scale values of block_q4_Kx8
3180                const __m256 col_scale_f32 = GGML_F32Cx8_LOAD(b_ptr[b].d);
3181
3182                // dmin values - Load the eight dmin values of block_q4_Kx8
3183                const __m256 col_dmin_f32 = GGML_F32Cx8_LOAD(b_ptr[b].dmin);
3184
3185                // Loop to iterate over the eight sub blocks of a super block - two sub blocks are processed per iteration
3186                for (int sb = 0; sb < QK_K / 64; sb++) {
3187
3188                    // Load the eight block_q4_k for two sub blocks quantized values interleaved with each other in chunks of eight bytes - B0,B1 ....B6,B7
3189                    const __m256i rhs_raw_mat_0123_0 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + sb * 256));
3190                    const __m256i rhs_raw_mat_4567_0 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 32 + sb * 256));
3191                    const __m256i rhs_raw_mat_0123_1 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 64 + sb * 256));
3192                    const __m256i rhs_raw_mat_4567_1 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 96 + sb * 256));
3193                    const __m256i rhs_raw_mat_0123_2 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 128 + sb * 256));
3194                    const __m256i rhs_raw_mat_4567_2 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 160 + sb * 256));
3195                    const __m256i rhs_raw_mat_0123_3 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 192 + sb * 256));
3196                    const __m256i rhs_raw_mat_4567_3 = _mm256_loadu_si256((const __m256i * )(b_ptr[b].qs + 224 + sb * 256));
3197
3198                    // Save the values in the following vectors in the formats B0B1B4B5, B2B3B6B7 for further processing and storing of values
3199                    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);
3200                    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);
3201                    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);
3202                    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);
3203                    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);
3204                    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);
3205                    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);
3206                    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);
3207
3208                    // 4-bit -> 8-bit
3209                    // First sub block of the two sub blocks processed in the iteration
3210                    const __m256i rhs_mat_0145_00 = _mm256_and_si256(rhs_raw_mat_0145_0, m4b); //B00(0-7) B01(0-7) B04(0-7) B05(0-7)
3211                    const __m256i rhs_mat_2367_00 = _mm256_and_si256(rhs_raw_mat_2367_0, m4b); //B02(0-7) B03(0-7) B06(0-7) B07(0-7)
3212
3213                    const __m256i rhs_mat_0145_01 = _mm256_and_si256(rhs_raw_mat_0145_1, m4b); //B00(8-15) B01(8-15) B04(8-15) B05(8-15)
3214                    const __m256i rhs_mat_2367_01 = _mm256_and_si256(rhs_raw_mat_2367_1, m4b); //B02(8-15) B03(8-15) B06(8-15) B07(8-15)
3215
3216                    const __m256i rhs_mat_0145_02 = _mm256_and_si256(rhs_raw_mat_0145_2, m4b); //B00(16-23) B01(16-23) B04(16-23) B05(16-23)
3217                    const __m256i rhs_mat_2367_02 = _mm256_and_si256(rhs_raw_mat_2367_2, m4b); //B02(16-23) B03(16-23) B06(16-23) B07(16-23)
3218
3219                    const __m256i rhs_mat_0145_03 = _mm256_and_si256(rhs_raw_mat_0145_3, m4b); //B00(24-31) B01(24-31) B04(24-31) B05(24-31)
3220                    const __m256i rhs_mat_2367_03 = _mm256_and_si256(rhs_raw_mat_2367_3, m4b); //B02(24-31) B03(24-31) B06(24-31) B07(24-31)
3221
3222                    // Second sub block of the two sub blocks processed in the iteration
3223                    const __m256i rhs_mat_0145_10 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_0, 4), m4b); //B10(0-7) B11(0-7) B14(0-7) B15(0-7)
3224                    const __m256i rhs_mat_2367_10 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_0, 4), m4b); //B12(0-7) B13(0-7) B16(0-7) B17(0-7)
3225
3226                    const __m256i rhs_mat_0145_11 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_1, 4), m4b); //B10(8-15) B11(8-15) B14(8-15) B15(8-15)
3227                    const __m256i rhs_mat_2367_11 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_1, 4), m4b); //B12(8-15) B13(8-15) B16(8-15) B17(8-15)
3228
3229                    const __m256i rhs_mat_0145_12 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_2, 4), m4b); //B10(16-23) B11(16-23) B14(16-23) B15(16-23)
3230                    const __m256i rhs_mat_2367_12 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_2, 4), m4b); //B12(16-23) B13(16-23) B16(16-23) B17(16-23)
3231
3232                    const __m256i rhs_mat_0145_13 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_0145_3, 4), m4b); //B10(24-31) B11(24-31) B14(24-31) B15(24-31)
3233                    const __m256i rhs_mat_2367_13 = _mm256_and_si256(_mm256_srli_epi16(rhs_raw_mat_2367_3, 4), m4b); //B12(24-31) B13(24-31) B16(24-31) B17(24-31)
3234
3235                    // Shuffle pattern one - right side input
3236                    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)
3237                    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)
3238
3239                    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)
3240                    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)
3241
3242                    const __m256i rhs_mat_0145_02_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_02, 136); //B00(16-19) B01(16-19) B00(16-19) B01(16-19) B04(16-19) B05(16-19) B04(16-19) B05(16-19)
3243                    const __m256i rhs_mat_2367_02_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_02, 136); //B02(16-19) B03(16-19) B02(16-19) B03(16-19) B06(16-19) B07(16-19) B06(16-19) B07(16-19)
3244
3245                    const __m256i rhs_mat_0145_03_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_03, 136); //B00(24-27) B01(24-27) B00(24-27) B01(24-27) B04(24-27) B05(24-27) B04(24-27) B05(24-27)
3246                    const __m256i rhs_mat_2367_03_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_03, 136); //B02(24-27) B03(24-27) B02(24-27) B03(24-27) B06(24-27) B07(24-27) B06(24-27) B07(24-27)
3247
3248                    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)
3249                    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)
3250
3251                    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)
3252                    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)
3253
3254                    const __m256i rhs_mat_0145_12_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_12, 136); //B10(16-19) B11(16-19) B10(16-19) B11(16-19) B14(16-19) B15(16-19) B14(16-19) B15(16-19)
3255                    const __m256i rhs_mat_2367_12_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_12, 136); //B12(16-19) B13(16-19) B12(16-19) B13(16-19) B16(16-19) B17(16-19) B16(16-19) B17(16-19)
3256
3257                    const __m256i rhs_mat_0145_13_sp1 = _mm256_shuffle_epi32(rhs_mat_0145_13, 136); //B10(24-27) B11(24-27) B10(24-27) B11(24-27) B14(24-27) B15(24-27) B14(24-27) B15(24-27)
3258                    const __m256i rhs_mat_2367_13_sp1 = _mm256_shuffle_epi32(rhs_mat_2367_13, 136); //B12(24-27) B13(24-27) B12(24-27) B13(24-27) B16(24-27) B17(24-27) B16(24-27) B17(24-27)
3259
3260                    // Shuffle pattern two - right side input
3261                    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)
3262                    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)
3263
3264                    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)
3265                    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)
3266
3267                    const __m256i rhs_mat_0145_02_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_02, 221); //B00(20-23) B01(20-23) B00(20-23) B01(20-23) B04(20-23) B05(20-23) B04(20-23) B05(20-23)
3268                    const __m256i rhs_mat_2367_02_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_02, 221); //B02(20-23) B03(20-23) B02(20-23) B03(20-23) B06(20-23) B07(20-23) B06(20-23) B07(20-23)
3269
3270                    const __m256i rhs_mat_0145_03_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_03, 221); //B00(28-31) B01(28-31) B00(28-31) B01(28-31) B04(28-31) B05(28-31) B04(28-31) B05(28-31)
3271                    const __m256i rhs_mat_2367_03_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_03, 221); //B02(28-31) B03(28-31) B02(28-31) B03(28-31) B06(28-31) B07(28-31) B06(28-31) B07(28-31)
3272
3273                    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)
3274                    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)
3275
3276                    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)
3277                    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)
3278
3279                    const __m256i rhs_mat_0145_12_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_12, 221); //B10(20-23) B11(20-23) B10(20-23) B11(20-23) B14(20-23) B15(20-23) B14(20-23) B15(20-23)
3280                    const __m256i rhs_mat_2367_12_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_12, 221); //B12(20-23) B13(20-23) B12(20-23) B13(20-23) B16(20-23) B17(20-23) B16(20-23) B17(20-23)
3281
3282                    const __m256i rhs_mat_0145_13_sp2 = _mm256_shuffle_epi32(rhs_mat_0145_13, 221); //B10(28-31) B11(28-31) B10(28-31) B11(28-31) B14(28-31) B15(28-31) B14(28-31) B15(28-31)
3283                    const __m256i rhs_mat_2367_13_sp2 = _mm256_shuffle_epi32(rhs_mat_2367_13, 221); //B12(28-31) B13(28-31) B12(28-31) B13(28-31) B16(28-31) B17(28-31) B16(28-31) B17(28-31)
3284
3285                    uint32_t utmp_0[4], utmp_1[4];
3286
3287                    // Scales and Mins of corresponding sub blocks from different Q4_K structures are stored together
3288                    // The below block is for eg to extract first sub block's scales and mins from different Q4_K structures for the sb loop
3289                    memcpy(utmp_0, b_ptr[b].scales + 24 * sb, 12);
3290                    utmp_0[3] = ((utmp_0[2] >> 4) & kmask2) | (((utmp_0[1] >> 6) & kmask3) << 4);
3291                    const uint32_t uaux_0 = utmp_0[1] & kmask1;
3292                    utmp_0[1] = (utmp_0[2] & kmask2) | (((utmp_0[0] >> 6) & kmask3) << 4);
3293                    utmp_0[2] = uaux_0;
3294                    utmp_0[0] &= kmask1;
3295
3296                    // The below block is for eg to extract second sub block's scales and mins from different Q4_K structures when sb = 1
3297                    memcpy(utmp_1, b_ptr[b].scales + 12 + sb * 24, 12);
3298                    utmp_1[3] = ((utmp_1[2] >> 4) & kmask2) | (((utmp_1[1] >> 6) & kmask3) << 4);
3299                    const uint32_t uaux_1 = utmp_1[1] & kmask1;
3300                    utmp_1[1] = (utmp_1[2] & kmask2) | (((utmp_1[0] >> 6) & kmask3) << 4);
3301                    utmp_1[2] = uaux_1;
3302                    utmp_1[0] &= kmask1;
3303
3304                    // Scales of first sub block in the sb loop
3305                    const __m128i mins_and_scales_0 = _mm_set_epi32(utmp_0[3], utmp_0[2], utmp_0[1], utmp_0[0]);
3306                    const __m256i scales_0 = _mm256_cvtepu8_epi16(_mm_unpacklo_epi8(mins_and_scales_0, mins_and_scales_0));
3307
3308                    // Scales of second sub block in the sb loop
3309                    const __m128i mins_and_scales_1 = _mm_set_epi32(utmp_1[3], utmp_1[2], utmp_1[1], utmp_1[0]);
3310                    const __m256i scales_1 = _mm256_cvtepu8_epi16(_mm_unpacklo_epi8(mins_and_scales_1, mins_and_scales_1));
3311
3312                    // Mins of first and second sub block of Q4_K block are arranged side by side
3313                    const __m256i mins_01 = _mm256_cvtepu8_epi16(_mm_unpacklo_epi8(_mm_shuffle_epi32(mins_and_scales_0, 78), _mm_shuffle_epi32(mins_and_scales_1, 78)));
3314
3315                    const __m256i scale_0145_0 = _mm256_shuffle_epi32(scales_0, 68);
3316                    const __m256i scale_2367_0 = _mm256_shuffle_epi32(scales_0, 238);
3317
3318                    const __m256i scale_0145_1 = _mm256_shuffle_epi32(scales_1, 68);
3319                    const __m256i scale_2367_1 = _mm256_shuffle_epi32(scales_1, 238);
3320
3321                    // Load the four block_q8_k quantized values interleaved with each other in chunks of eight bytes - A0,A1,A2,A3
3322                    // Loaded as set of 128 bit vectors and repeated into a 256 bit vector
3323                    __m256i lhs_mat_0123_00 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 256 * sb)));
3324                    __m256i lhs_mat_01_00 = _mm256_permute2f128_si256(lhs_mat_0123_00, lhs_mat_0123_00, 0);
3325                    __m256i lhs_mat_23_00 = _mm256_permute2f128_si256(lhs_mat_0123_00, lhs_mat_0123_00, 17);
3326                    __m256i lhs_mat_0123_01 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 32 + 256 * sb)));
3327                    __m256i lhs_mat_01_01 = _mm256_permute2f128_si256(lhs_mat_0123_01, lhs_mat_0123_01, 0);
3328                    __m256i lhs_mat_23_01 = _mm256_permute2f128_si256(lhs_mat_0123_01, lhs_mat_0123_01, 17);
3329                    __m256i lhs_mat_0123_02 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 64 + 256 * sb)));
3330                    __m256i lhs_mat_01_02 = _mm256_permute2f128_si256(lhs_mat_0123_02, lhs_mat_0123_02, 0);
3331                    __m256i lhs_mat_23_02 = _mm256_permute2f128_si256(lhs_mat_0123_02, lhs_mat_0123_02, 17);
3332                    __m256i lhs_mat_0123_03 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 96 + 256 * sb)));
3333                    __m256i lhs_mat_01_03 = _mm256_permute2f128_si256(lhs_mat_0123_03, lhs_mat_0123_03, 0);
3334                    __m256i lhs_mat_23_03 = _mm256_permute2f128_si256(lhs_mat_0123_03, lhs_mat_0123_03, 17);
3335                    __m256i lhs_mat_0123_10 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 128 + 256 * sb)));
3336                    __m256i lhs_mat_01_10 = _mm256_permute2f128_si256(lhs_mat_0123_10, lhs_mat_0123_10, 0);
3337                    __m256i lhs_mat_23_10 = _mm256_permute2f128_si256(lhs_mat_0123_10, lhs_mat_0123_10, 17);
3338                    __m256i lhs_mat_0123_11 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 160 + 256 * sb)));
3339                    __m256i lhs_mat_01_11 = _mm256_permute2f128_si256(lhs_mat_0123_11, lhs_mat_0123_11, 0);
3340                    __m256i lhs_mat_23_11 = _mm256_permute2f128_si256(lhs_mat_0123_11, lhs_mat_0123_11, 17);
3341                    __m256i lhs_mat_0123_12 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 192 + 256 * sb)));
3342                    __m256i lhs_mat_01_12 = _mm256_permute2f128_si256(lhs_mat_0123_12, lhs_mat_0123_12, 0);
3343                    __m256i lhs_mat_23_12 = _mm256_permute2f128_si256(lhs_mat_0123_12, lhs_mat_0123_12, 17);
3344                    __m256i lhs_mat_0123_13 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 224 + 256 * sb)));
3345                    __m256i lhs_mat_01_13 = _mm256_permute2f128_si256(lhs_mat_0123_13, lhs_mat_0123_13, 0);
3346                    __m256i lhs_mat_23_13 = _mm256_permute2f128_si256(lhs_mat_0123_13, lhs_mat_0123_13, 17);
3347
3348                    // Bsums are loaded - four bsums are loaded (for two sub blocks) for the different Q8_K blocks
3349                    __m256i lhs_bsums_0123_01 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].bsums + 16 * sb)));
3350                    __m256i lhs_bsums_hsum_0123_01 = _mm256_castsi128_si256(_mm_hadd_epi16(_mm256_castsi256_si128(lhs_bsums_0123_01), _mm256_extractf128_si256(lhs_bsums_0123_01, 1)));
3351                    lhs_bsums_hsum_0123_01 = _mm256_permute2x128_si256(lhs_bsums_hsum_0123_01, lhs_bsums_hsum_0123_01, 0);
3352
3353                    // Shuffle pattern one - left side input
3354                    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)
3355                    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)
3356
3357                    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)
3358                    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)
3359
3360                    const __m256i lhs_mat_01_02_sp1 = _mm256_shuffle_epi32(lhs_mat_01_02, 160);  //A00(16-19) A00(16-19) A01(16-19) A01(16-19) A00(16-19) A00(16-19) A01(16-19) A01(16-19)
3361                    const __m256i lhs_mat_23_02_sp1 = _mm256_shuffle_epi32(lhs_mat_23_02, 160);  //A02(16-19) A03(16-19) A02(16-19) A03(16-19) A02(16-19) A03(16-19) A02(16-19) A03(16-19)
3362
3363                    const __m256i lhs_mat_01_03_sp1 = _mm256_shuffle_epi32(lhs_mat_01_03, 160);  //A00(24-27) A00(24-27) A01(24-27) A01(24-27) A00(24-27) A00(24-27) A01(24-27) A01(24-27)
3364                    const __m256i lhs_mat_23_03_sp1 = _mm256_shuffle_epi32(lhs_mat_23_03, 160); //A02(24-27) A03(24-27) A02(24-27) A03(24-27) A02(24-27) A03(24-27) A02(24-27) A03(24-27)
3365
3366                    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)
3367                    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)
3368
3369                    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)
3370                    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)
3371
3372                    const __m256i lhs_mat_01_12_sp1 = _mm256_shuffle_epi32(lhs_mat_01_12, 160);  //A10(16-19) A10(16-19) A11(16-19) A11(16-19) A10(16-19) A10(16-19) A11(16-19) A11(16-19)
3373                    const __m256i lhs_mat_23_12_sp1 = _mm256_shuffle_epi32(lhs_mat_23_12, 160);  //A12(16-19) A13(16-19) A12(16-19) A13(16-19) A12(16-19) A13(16-19) A12(16-19) A13(16-19)
3374
3375                    const __m256i lhs_mat_01_13_sp1 = _mm256_shuffle_epi32(lhs_mat_01_13, 160); //A10(24-27) A10(24-27) A11(24-27) A11(24-27) A10(24-27) A10(24-27) A11(24-27) A11(24-27)
3376                    const __m256i lhs_mat_23_13_sp1 = _mm256_shuffle_epi32(lhs_mat_23_13, 160); //A12(24-27) A13(24-27) A12(24-27) A13(24-27) A12(24-27) A13(24-27) A12(24-27) A13(24-27)
3377
3378                    // Shuffle pattern two- left side input
3379                    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)
3380                    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)
3381
3382                    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)
3383                    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)
3384
3385                    const __m256i lhs_mat_01_02_sp2 = _mm256_shuffle_epi32(lhs_mat_01_02, 245); //A00(20-23) A00(20-23) A01(20-23) A01(20-23) A00(20-23) A00(20-23) A01(20-23) A01(20-23)
3386                    const __m256i lhs_mat_23_02_sp2 = _mm256_shuffle_epi32(lhs_mat_23_02, 245); //A02(20-23) A03(20-23) A02(20-23) A03(20-23) A02(20-23) A03(20-23) A02(20-23) A03(20-23)
3387
3388                    const __m256i lhs_mat_01_03_sp2 = _mm256_shuffle_epi32(lhs_mat_01_03, 245); //A00(28-31) A00(28-31) A01(28-31) A01(28-31) A00(28-31) A00(28-31) A01(28-31) A01(28-31)
3389                    const __m256i lhs_mat_23_03_sp2 = _mm256_shuffle_epi32(lhs_mat_23_03, 245); //A02(28-31) A03(28-31) A02(28-31) A03(28-31) A02(28-31) A03(28-31) A02(28-31) A03(28-31)
3390
3391                    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)
3392                    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)
3393
3394                    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)
3395                    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)
3396
3397                    const __m256i lhs_mat_01_12_sp2 = _mm256_shuffle_epi32(lhs_mat_01_12, 245); //A10(20-23) A10(20-23) A11(20-23) A11(20-23) A10(20-23) A10(20-23) A11(20-23) A11(20-23)
3398                    const __m256i lhs_mat_23_12_sp2 = _mm256_shuffle_epi32(lhs_mat_23_12, 245); //A12(20-23) A13(20-23) A12(20-23) A13(20-23) A12(20-23) A13(20-23) A12(20-23) A13(20-23)
3399
3400                    const __m256i lhs_mat_01_13_sp2 = _mm256_shuffle_epi32(lhs_mat_01_13, 245); //A10(28-31) A10(28-31) A11(28-31) A11(28-31) A10(28-31) A10(28-31) A11(28-31) A11(28-31)
3401                    const __m256i lhs_mat_23_13_sp2 = _mm256_shuffle_epi32(lhs_mat_23_13, 245); //A12(28-31) A13(28-31) A12(28-31) A13(28-31) A12(28-31) A13(28-31) A12(28-31) A13(28-31)
3402
3403                    // 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
3404                    __m256i iacc_mat_00_0_sp1 = _mm256_add_epi16(_mm256_add_epi16(_mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_03_sp1, lhs_mat_01_03_sp1), _mm256_maddubs_epi16(rhs_mat_0145_02_sp1, lhs_mat_01_02_sp1)), _mm256_maddubs_epi16(rhs_mat_0145_01_sp1, lhs_mat_01_01_sp1)), _mm256_maddubs_epi16(rhs_mat_0145_00_sp1, lhs_mat_01_00_sp1));
3405                    __m256i iacc_mat_01_0_sp1 = _mm256_add_epi16(_mm256_add_epi16(_mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_03_sp1, lhs_mat_01_03_sp1), _mm256_maddubs_epi16(rhs_mat_2367_02_sp1, lhs_mat_01_02_sp1)), _mm256_maddubs_epi16(rhs_mat_2367_01_sp1, lhs_mat_01_01_sp1)), _mm256_maddubs_epi16(rhs_mat_2367_00_sp1, lhs_mat_01_00_sp1));
3406                    __m256i iacc_mat_10_0_sp1 = _mm256_add_epi16(_mm256_add_epi16(_mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_03_sp1, lhs_mat_23_03_sp1), _mm256_maddubs_epi16(rhs_mat_0145_02_sp1, lhs_mat_23_02_sp1)), _mm256_maddubs_epi16(rhs_mat_0145_01_sp1, lhs_mat_23_01_sp1)), _mm256_maddubs_epi16(rhs_mat_0145_00_sp1, lhs_mat_23_00_sp1));
3407                    __m256i iacc_mat_11_0_sp1 = _mm256_add_epi16(_mm256_add_epi16(_mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_03_sp1, lhs_mat_23_03_sp1), _mm256_maddubs_epi16(rhs_mat_2367_02_sp1, lhs_mat_23_02_sp1)), _mm256_maddubs_epi16(rhs_mat_2367_01_sp1, lhs_mat_23_01_sp1)), _mm256_maddubs_epi16(rhs_mat_2367_00_sp1, lhs_mat_23_00_sp1));
3408                    __m256i iacc_mat_00_1_sp1 = _mm256_add_epi16(_mm256_add_epi16(_mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_13_sp1, lhs_mat_01_13_sp1), _mm256_maddubs_epi16(rhs_mat_0145_12_sp1, lhs_mat_01_12_sp1)), _mm256_maddubs_epi16(rhs_mat_0145_11_sp1, lhs_mat_01_11_sp1)), _mm256_maddubs_epi16(rhs_mat_0145_10_sp1, lhs_mat_01_10_sp1));
3409                    __m256i iacc_mat_01_1_sp1 = _mm256_add_epi16(_mm256_add_epi16(_mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_13_sp1, lhs_mat_01_13_sp1), _mm256_maddubs_epi16(rhs_mat_2367_12_sp1, lhs_mat_01_12_sp1)), _mm256_maddubs_epi16(rhs_mat_2367_11_sp1, lhs_mat_01_11_sp1)), _mm256_maddubs_epi16(rhs_mat_2367_10_sp1, lhs_mat_01_10_sp1));
3410                    __m256i iacc_mat_10_1_sp1 = _mm256_add_epi16(_mm256_add_epi16(_mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_13_sp1, lhs_mat_23_13_sp1), _mm256_maddubs_epi16(rhs_mat_0145_12_sp1, lhs_mat_23_12_sp1)), _mm256_maddubs_epi16(rhs_mat_0145_11_sp1, lhs_mat_23_11_sp1)), _mm256_maddubs_epi16(rhs_mat_0145_10_sp1, lhs_mat_23_10_sp1));
3411                    __m256i iacc_mat_11_1_sp1 = _mm256_add_epi16(_mm256_add_epi16(_mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_13_sp1, lhs_mat_23_13_sp1), _mm256_maddubs_epi16(rhs_mat_2367_12_sp1, lhs_mat_23_12_sp1)), _mm256_maddubs_epi16(rhs_mat_2367_11_sp1, lhs_mat_23_11_sp1)), _mm256_maddubs_epi16(rhs_mat_2367_10_sp1, lhs_mat_23_10_sp1));
3412
3413                    __m256i iacc_mat_00_0_sp2 = _mm256_add_epi16(_mm256_add_epi16(_mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_03_sp2, lhs_mat_01_03_sp2), _mm256_maddubs_epi16(rhs_mat_0145_02_sp2, lhs_mat_01_02_sp2)), _mm256_maddubs_epi16(rhs_mat_0145_01_sp2, lhs_mat_01_01_sp2)), _mm256_maddubs_epi16(rhs_mat_0145_00_sp2, lhs_mat_01_00_sp2));
3414                    __m256i iacc_mat_01_0_sp2 = _mm256_add_epi16(_mm256_add_epi16(_mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_03_sp2, lhs_mat_01_03_sp2), _mm256_maddubs_epi16(rhs_mat_2367_02_sp2, lhs_mat_01_02_sp2)), _mm256_maddubs_epi16(rhs_mat_2367_01_sp2, lhs_mat_01_01_sp2)), _mm256_maddubs_epi16(rhs_mat_2367_00_sp2, lhs_mat_01_00_sp2));
3415                    __m256i iacc_mat_10_0_sp2 = _mm256_add_epi16(_mm256_add_epi16(_mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_03_sp2, lhs_mat_23_03_sp2), _mm256_maddubs_epi16(rhs_mat_0145_02_sp2, lhs_mat_23_02_sp2)), _mm256_maddubs_epi16(rhs_mat_0145_01_sp2, lhs_mat_23_01_sp2)), _mm256_maddubs_epi16(rhs_mat_0145_00_sp2, lhs_mat_23_00_sp2));
3416                    __m256i iacc_mat_11_0_sp2 = _mm256_add_epi16(_mm256_add_epi16(_mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_03_sp2, lhs_mat_23_03_sp2), _mm256_maddubs_epi16(rhs_mat_2367_02_sp2, lhs_mat_23_02_sp2)), _mm256_maddubs_epi16(rhs_mat_2367_01_sp2, lhs_mat_23_01_sp2)), _mm256_maddubs_epi16(rhs_mat_2367_00_sp2, lhs_mat_23_00_sp2));
3417                    __m256i iacc_mat_00_1_sp2 = _mm256_add_epi16(_mm256_add_epi16(_mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_13_sp2, lhs_mat_01_13_sp2), _mm256_maddubs_epi16(rhs_mat_0145_12_sp2, lhs_mat_01_12_sp2)), _mm256_maddubs_epi16(rhs_mat_0145_11_sp2, lhs_mat_01_11_sp2)), _mm256_maddubs_epi16(rhs_mat_0145_10_sp2, lhs_mat_01_10_sp2));
3418                    __m256i iacc_mat_01_1_sp2 = _mm256_add_epi16(_mm256_add_epi16(_mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_13_sp2, lhs_mat_01_13_sp2), _mm256_maddubs_epi16(rhs_mat_2367_12_sp2, lhs_mat_01_12_sp2)), _mm256_maddubs_epi16(rhs_mat_2367_11_sp2, lhs_mat_01_11_sp2)), _mm256_maddubs_epi16(rhs_mat_2367_10_sp2, lhs_mat_01_10_sp2));
3419                    __m256i iacc_mat_10_1_sp2 = _mm256_add_epi16(_mm256_add_epi16(_mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_0145_13_sp2, lhs_mat_23_13_sp2), _mm256_maddubs_epi16(rhs_mat_0145_12_sp2, lhs_mat_23_12_sp2)), _mm256_maddubs_epi16(rhs_mat_0145_11_sp2, lhs_mat_23_11_sp2)), _mm256_maddubs_epi16(rhs_mat_0145_10_sp2, lhs_mat_23_10_sp2));
3420                    __m256i iacc_mat_11_1_sp2 = _mm256_add_epi16(_mm256_add_epi16(_mm256_add_epi16(_mm256_maddubs_epi16(rhs_mat_2367_13_sp2, lhs_mat_23_13_sp2), _mm256_maddubs_epi16(rhs_mat_2367_12_sp2, lhs_mat_23_12_sp2)), _mm256_maddubs_epi16(rhs_mat_2367_11_sp2, lhs_mat_23_11_sp2)), _mm256_maddubs_epi16(rhs_mat_2367_10_sp2, lhs_mat_23_10_sp2));
3421
3422                    // Output of both shuffle patterns are added in order to sum dot product outputs of all 32 values in block
3423                    __m256i iacc_mat_00_0 = _mm256_add_epi16(iacc_mat_00_0_sp1, iacc_mat_00_0_sp2);
3424                    __m256i iacc_mat_01_0 = _mm256_add_epi16(iacc_mat_01_0_sp1, iacc_mat_01_0_sp2);
3425                    __m256i iacc_mat_10_0 = _mm256_add_epi16(iacc_mat_10_0_sp1, iacc_mat_10_0_sp2);
3426                    __m256i iacc_mat_11_0 = _mm256_add_epi16(iacc_mat_11_0_sp1, iacc_mat_11_0_sp2);
3427
3428                    __m256i iacc_mat_00_1 = _mm256_add_epi16(iacc_mat_00_1_sp1, iacc_mat_00_1_sp2);
3429                    __m256i iacc_mat_01_1 = _mm256_add_epi16(iacc_mat_01_1_sp1, iacc_mat_01_1_sp2);
3430                    __m256i iacc_mat_10_1 = _mm256_add_epi16(iacc_mat_10_1_sp1, iacc_mat_10_1_sp2);
3431                    __m256i iacc_mat_11_1 = _mm256_add_epi16(iacc_mat_11_1_sp1, iacc_mat_11_1_sp2);
3432
3433                    // Output of both shuffle patterns are added in order to sum dot product outputs of all 32 values in block
3434                    iacc_mat_00_0 = _mm256_madd_epi16(iacc_mat_00_0, scale_0145_0);
3435                    iacc_mat_01_0 = _mm256_madd_epi16(iacc_mat_01_0, scale_2367_0);
3436                    iacc_mat_10_0 = _mm256_madd_epi16(iacc_mat_10_0, scale_0145_0);
3437                    iacc_mat_11_0 = _mm256_madd_epi16(iacc_mat_11_0, scale_2367_0);
3438
3439                    iacc_mat_00_1 = _mm256_madd_epi16(iacc_mat_00_1, scale_0145_1);
3440                    iacc_mat_01_1 = _mm256_madd_epi16(iacc_mat_01_1, scale_2367_1);
3441                    iacc_mat_10_1 = _mm256_madd_epi16(iacc_mat_10_1, scale_0145_1);
3442                    iacc_mat_11_1 = _mm256_madd_epi16(iacc_mat_11_1, scale_2367_1);
3443
3444                    // Straighten out to make 4 row vectors (4 for each sub block which are accumulated together in the next step)
3445                    __m256i iacc_row_0_0 = _mm256_blend_epi32(iacc_mat_00_0, _mm256_shuffle_epi32(iacc_mat_01_0, 78), 204);
3446                    __m256i iacc_row_1_0 = _mm256_blend_epi32(_mm256_shuffle_epi32(iacc_mat_00_0, 78), iacc_mat_01_0, 204);
3447                    __m256i iacc_row_2_0 = _mm256_blend_epi32(iacc_mat_10_0, _mm256_shuffle_epi32(iacc_mat_11_0, 78), 204);
3448                    __m256i iacc_row_3_0 = _mm256_blend_epi32(_mm256_shuffle_epi32(iacc_mat_10_0, 78), iacc_mat_11_0, 204);
3449                    __m256i iacc_row_0_1 = _mm256_blend_epi32(iacc_mat_00_1, _mm256_shuffle_epi32(iacc_mat_01_1, 78), 204);
3450                    __m256i iacc_row_1_1 = _mm256_blend_epi32(_mm256_shuffle_epi32(iacc_mat_00_1, 78), iacc_mat_01_1, 204);
3451                    __m256i iacc_row_2_1 = _mm256_blend_epi32(iacc_mat_10_1, _mm256_shuffle_epi32(iacc_mat_11_1, 78), 204);
3452                    __m256i iacc_row_3_1 = _mm256_blend_epi32(_mm256_shuffle_epi32(iacc_mat_10_1, 78), iacc_mat_11_1, 204);
3453
3454                    __m256i iacc_row_0 = _mm256_add_epi32(iacc_row_0_0, iacc_row_0_1);
3455                    __m256i iacc_row_1 = _mm256_add_epi32(iacc_row_1_0, iacc_row_1_1);
3456                    __m256i iacc_row_2 = _mm256_add_epi32(iacc_row_2_0, iacc_row_2_1);
3457                    __m256i iacc_row_3 = _mm256_add_epi32(iacc_row_3_0, iacc_row_3_1);
3458
3459                    // Load the scale(d) values for all the 4 Q8_k blocks and repeat it across lanes
3460                    const __m128 row_scale_f32_sse = _mm_load_ps(a_ptr[b].d);
3461                    const __m256 row_scale_f32 = _mm256_set_m128(row_scale_f32_sse, row_scale_f32_sse); //GGML_F32Cx8_REPEAT_LOAD(a_ptrs[rp][b].d, loadMask);
3462
3463                    // Multiply with appropriate scales and accumulate (for both d and dmin) below
3464                    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]);
3465                    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]);
3466                    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]);
3467                    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]);
3468
3469                    __m256i iacc_row_min_0 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_hsum_0123_01, 0), mins_01);
3470                    __m256i iacc_row_min_1 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_hsum_0123_01, 85), mins_01);
3471                    __m256i iacc_row_min_2 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_hsum_0123_01, 170), mins_01);
3472                    __m256i iacc_row_min_3 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_hsum_0123_01, 255), mins_01);
3473
3474                    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]);
3475                    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]);
3476                    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]);
3477                    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]);
3478                }
3479            }
3480
3481            // Store the accumulated values
3482            for (int i = 0; i < 4; i++) {
3483                _mm256_storeu_ps((float * )(s + ((y * 4 + i) * bs + x * 8)), _mm256_sub_ps(acc_rows[i], acc_min_rows[i]));
3484            }
3485        }
3486    }
3487
3488#else
3489    UNUSED(kmask1);
3490    UNUSED(kmask2);
3491    UNUSED(kmask3);
3492    ggml_gemm_q4_K_8x8_q8_K_generic(n, s, bs, vx, vy, nr, nc);
3493#endif
3494}
3495
3496void ggml_gemm_iq4_nl_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
3497#if defined(__AVX2__) || defined(__AVX512F__)
3498    {
3499        __m256i signextendlut = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i*)kvalues_iq4nl));
3500        signextendlut = _mm256_permute2f128_si256(signextendlut, signextendlut, 0);
3501
3502        gemm_q4_b32_8x8_q8_0_lut_avx<block_iq4_nlx8>(n, s, bs, vx, vy, nr, nc, signextendlut);
3503
3504        return;
3505    }
3506#endif // defined(__AVX2__) || defined(__AVX512F__)
3507
3508    ggml_gemm_iq4_nl_4x4_q8_0(n, s, bs, vx, vy, nr, nc);
3509}
3510
3511void ggml_gemm_mxfp4_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) {
3512#if defined(__AVX2__) || defined(__AVX512F__)
3513    {
3514        __m256i signextendlut = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i*)kvalues_mxfp4));
3515        signextendlut = _mm256_permute2f128_si256(signextendlut, signextendlut, 0);
3516
3517        gemm_q4_b32_8x8_q8_0_lut_avx<block_mxfp4x8>(n, s, bs, vx, vy, nr, nc, signextendlut);
3518
3519        return;
3520    }
3521#endif // defined(__AVX2__) || defined(__AVX512F__)
3522
3523    ggml_gemm_mxfp4_8x8_q8_0_generic(n, s, bs, vx, vy, nr, nc);
3524}
3525
3526void 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) {
3527    const int qk = QK_K;
3528    const int nb = n / qk;
3529    const int ncols_interleaved = 8;
3530    const int blocklen = 8;
3531
3532    assert (n % qk == 0);
3533    assert (nr % 4 == 0);
3534    assert (nc % ncols_interleaved == 0);
3535
3536    UNUSED(s);
3537    UNUSED(bs);
3538    UNUSED(vx);
3539    UNUSED(vy);
3540    UNUSED(nr);
3541    UNUSED(nc);
3542    UNUSED(nb);
3543    UNUSED(ncols_interleaved);
3544    UNUSED(blocklen);
3545
3546#if defined(__AVX2__) || defined(__AVX512F__)
3547    const block_q2_Kx8 * b_ptr_start = (const block_q2_Kx8 * ) vx;
3548    const block_q8_Kx4 * a_ptr_start = (const block_q8_Kx4 * ) vy;
3549    int64_t b_nb = n / QK_K;
3550    int64_t y = 0;
3551
3552    // Permute mask used for easier vector processing at later stages
3553    __m256i requiredOrder = _mm256_set_epi32(3, 2, 1, 0, 7, 6, 5, 4);
3554    int64_t xstart = 0;
3555    int anr = nr - nr % 16; // Used to align nr with boundary of 16
3556
3557    // Mask to convert 2 bit and 4 bit values into a bytes
3558    const __m256i m3b = _mm256_set1_epi8(3);
3559    const __m128i m4b_sse = _mm_set1_epi8(0xF);
3560
3561    //Mask to get appropriate scales
3562    __m128i scalesmask1_sse = _mm_set_epi8(14,14,12,12,10,10,8,8,6,6,4,4,2,2,0,0);
3563    __m128i scalesmask2_sse = _mm_set_epi8(15,15,13,13,11,11,9,9,7,7,5,5,3,3,1,1);
3564
3565    __m256i scalesmask1 = _mm256_castsi128_si256(scalesmask1_sse);
3566    scalesmask1 = _mm256_permute2f128_si256(scalesmask1, scalesmask1, 0);
3567    __m256i scalesmask2 = _mm256_castsi128_si256(scalesmask2_sse);
3568    scalesmask2 = _mm256_permute2f128_si256(scalesmask2, scalesmask2, 0);
3569
3570#if defined(__AVX512BW__) && defined(__AVX512DQ__)
3571
3572    int anc = nc - nc % 16; // Used to align nc with boundary of 16
3573
3574    // Mask to mask out nibbles from packed bytes
3575    const __m256i m4b = _mm256_set1_epi8(0x0F);
3576    // Mask to mask out nibbles from packed bytes expanded to 512 bit length
3577    const __m512i m3bexpanded = _mm512_set1_epi8(3);
3578    //Take group of four block_q8_Kx4 structures at each pass of the loop and perform dot product operation
3579    for (; y < anr / 4; y += 4) {
3580
3581        const block_q8_Kx4 * a_ptrs[4];
3582
3583        a_ptrs[0] = a_ptr_start + (y * nb);
3584        for (int i = 0; i < 3; ++i) {
3585            a_ptrs[i + 1] = a_ptrs[i] + nb;
3586        }
3587
3588        // Take group of eight block_q2_kx8 structures at each pass of the loop and perform dot product operation
3589        for (int64_t x = 0; x < anc / 8; x += 2) {
3590
3591            const block_q2_Kx8 * b_ptr_0 = b_ptr_start + ((x) * b_nb);
3592            const block_q2_Kx8 * b_ptr_1 = b_ptr_start + ((x + 1) * b_nb);
3593
3594            // Master FP accumulators
3595            __m512 acc_rows[16];
3596            for (int i = 0; i < 16; i++) {
3597                acc_rows[i] = _mm512_setzero_ps();
3598            }
3599
3600            __m512 acc_min_rows[16];
3601            for (int i = 0; i < 16; i++) {
3602                acc_min_rows[i] = _mm512_setzero_ps();
3603            }
3604            // For super block
3605            for (int64_t b = 0; b < nb; b++) {
3606                // Delta values - Load the sixteen scale values from two block_q2_kx8 structures
3607                const __m512 col_scale_f32 = GGML_F32Cx8x2_LOAD(b_ptr_0[b].d, b_ptr_1[b].d);
3608
3609                // dmin values - Load the sixteen dmin values from two block_q2_kx8 structures
3610                const __m512 col_dmin_f32 = GGML_F32Cx8x2_LOAD(b_ptr_0[b].dmin, b_ptr_1[b].dmin);
3611
3612                // Loop to iterate over the sixteen sub blocks of a super block - eight sub blocks are processed per iteration
3613                for (int sb = 0; sb < QK_K / 128; sb++) {
3614
3615                    // Load the eight block_q2_k for eight sub blocks quantized values interleaved with each other in chunks of eight bytes - B0,B1 ....B6,B7
3616                    const __m256i rhs_raw_mat_0123_0 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + sb * 256));
3617                    const __m256i rhs_raw_mat_4567_0 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 32 + sb * 256));
3618                    const __m256i rhs_raw_mat_0123_1 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 64 + sb * 256));
3619                    const __m256i rhs_raw_mat_4567_1 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 96 + sb * 256));
3620                    const __m256i rhs_raw_mat_0123_2 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 128 + sb * 256));
3621                    const __m256i rhs_raw_mat_4567_2 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 160 + sb * 256));
3622                    const __m256i rhs_raw_mat_0123_3 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 192 + sb * 256));
3623                    const __m256i rhs_raw_mat_4567_3 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 224 + sb * 256));
3624
3625                    const __m256i rhs_raw_mat_89AB_0 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + sb * 256));
3626                    const __m256i rhs_raw_mat_CDEF_0 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 32 + sb * 256));
3627                    const __m256i rhs_raw_mat_89AB_1 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 64 + sb * 256));
3628                    const __m256i rhs_raw_mat_CDEF_1 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 96 + sb * 256));
3629                    const __m256i rhs_raw_mat_89AB_2 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 128 + sb * 256));
3630                    const __m256i rhs_raw_mat_CDEF_2 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 160 + sb * 256));
3631                    const __m256i rhs_raw_mat_89AB_3 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 192 + sb * 256));
3632                    const __m256i rhs_raw_mat_CDEF_3 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 224 + sb * 256));
3633
3634                    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);
3635                    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);
3636                    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);
3637                    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);
3638                    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);
3639                    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);
3640                    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);
3641                    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);
3642
3643                    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);
3644                    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);
3645                    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);
3646                    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);
3647                    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);
3648                    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);
3649                    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);
3650                    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);
3651
3652                    const __m512i rhs_raw_mat_014589CD_0 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_0), rhs_raw_mat_89CD_0, 1);
3653                    const __m512i rhs_raw_mat_2367ABEF_0 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_0), rhs_raw_mat_ABEF_0, 1);
3654                    const __m512i rhs_raw_mat_014589CD_1 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_1), rhs_raw_mat_89CD_1, 1);
3655                    const __m512i rhs_raw_mat_2367ABEF_1 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_1), rhs_raw_mat_ABEF_1, 1);
3656
3657                    const __m512i rhs_raw_mat_014589CD_2 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_2), rhs_raw_mat_89CD_2, 1);
3658                    const __m512i rhs_raw_mat_2367ABEF_2 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_2), rhs_raw_mat_ABEF_2, 1);
3659                    const __m512i rhs_raw_mat_014589CD_3 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_3), rhs_raw_mat_89CD_3, 1);
3660                    const __m512i rhs_raw_mat_2367ABEF_3 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_3), rhs_raw_mat_ABEF_3, 1);
3661
3662                    //2-bit -> 8-bit
3663                    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)
3664                    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)
3665                    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)
3666                    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)
3667                    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)
3668                    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)
3669                    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)
3670                    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)
3671
3672                    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)
3673                    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)
3674
3675                    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)
3676                    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)
3677
3678                    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)
3679                    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)
3680
3681                    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)
3682                    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)
3683
3684                    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)
3685                    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)
3686
3687                    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)
3688                    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)
3689
3690                    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)
3691                    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)
3692
3693                    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)
3694                    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)
3695
3696                    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)
3697                    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)
3698
3699                    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)
3700                    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)
3701
3702                    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)
3703                    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)
3704
3705                    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)
3706                    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)
3707
3708                    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)
3709                    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)
3710
3711                    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)
3712                    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)
3713
3714                    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)
3715                    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)
3716
3717                    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)
3718                    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)
3719
3720                    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)
3721                    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)
3722
3723                    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)
3724                    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)
3725
3726                    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)
3727                    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)
3728
3729                    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)
3730                    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)
3731
3732                    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)
3733                    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)
3734
3735                    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)
3736                    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)
3737
3738                    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)
3739                    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)
3740
3741                    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)
3742                    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)
3743
3744                    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)
3745                    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)
3746
3747                    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)
3748                    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)
3749
3750                    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)
3751                    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)
3752
3753                    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)
3754                    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)
3755
3756                    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)
3757                    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)
3758
3759                    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)
3760                    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)
3761
3762                    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)
3763                    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)
3764
3765                    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)
3766                    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)
3767
3768                    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)
3769                    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)
3770
3771                    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)
3772                    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)
3773
3774                    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)
3775                    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)
3776
3777                    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)
3778                    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)
3779
3780                    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)
3781                    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)
3782
3783                    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)
3784                    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)
3785
3786                    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)
3787                    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)
3788
3789                    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)
3790                    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)
3791
3792                    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)
3793                    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)
3794
3795                    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)
3796                    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)
3797
3798                    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)
3799                    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)
3800
3801                    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)
3802                    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)
3803
3804                    //notation:superblock subblock
3805                    //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
3806
3807                    const __m128i mins_and_scales_01_0 = _mm_loadu_si128((const __m128i *)(b_ptr_0[b].scales + sb * 64));
3808                    const __m128i mins_and_scales_23_0 = _mm_loadu_si128((const __m128i *)(b_ptr_0[b].scales + 16 + sb * 64));
3809                    const __m128i mins_and_scales_45_0 = _mm_loadu_si128((const __m128i *)(b_ptr_0[b].scales + 32 + sb * 64));
3810                    const __m128i mins_and_scales_67_0 = _mm_loadu_si128((const __m128i *)(b_ptr_0[b].scales + 48 + sb * 64));
3811
3812                    const __m128i mins_and_scales_01_1 = _mm_loadu_si128((const __m128i *)(b_ptr_1[b].scales + sb * 64));
3813                    const __m128i mins_and_scales_23_1 = _mm_loadu_si128((const __m128i *)(b_ptr_1[b].scales + 16 + sb * 64));
3814                    const __m128i mins_and_scales_45_1 = _mm_loadu_si128((const __m128i *)(b_ptr_1[b].scales + 32 + sb * 64));
3815                    const __m128i mins_and_scales_67_1 = _mm_loadu_si128((const __m128i *)(b_ptr_1[b].scales + 48 + sb * 64));
3816
3817                    // Combine mins and scales for sub-blocks: 0-1, 2-3, 4-5, 6-7 in the sb loop
3818                    const __m256i mins_and_scales_01 = _mm256_insertf128_si256(_mm256_castsi128_si256(mins_and_scales_01_0), mins_and_scales_01_1, 1);
3819                    const __m256i mins_and_scales_23 = _mm256_insertf128_si256(_mm256_castsi128_si256(mins_and_scales_23_0), mins_and_scales_23_1, 1);
3820                    const __m256i mins_and_scales_45 = _mm256_insertf128_si256(_mm256_castsi128_si256(mins_and_scales_45_0), mins_and_scales_45_1, 1);
3821                    const __m256i mins_and_scales_67 = _mm256_insertf128_si256(_mm256_castsi128_si256(mins_and_scales_67_0), mins_and_scales_67_1, 1);
3822
3823                    // Extract scales which is lower half from mins_and_scales
3824                    const __m256i scales_01 = _mm256_and_si256(mins_and_scales_01, m4b);
3825                    const __m256i scales_23 = _mm256_and_si256(mins_and_scales_23, m4b);
3826                    const __m256i scales_45 = _mm256_and_si256(mins_and_scales_45, m4b);
3827                    const __m256i scales_67 = _mm256_and_si256(mins_and_scales_67, m4b);
3828
3829                    // Extract mins which is upper half from mins_and_scales
3830                    const __m512i mins_01 = _mm512_cvtepu8_epi16(_mm256_and_si256(_mm256_srli_epi16(mins_and_scales_01, 4), m4b));
3831                    const __m512i mins_23 = _mm512_cvtepu8_epi16(_mm256_and_si256(_mm256_srli_epi16(mins_and_scales_23, 4), m4b));
3832                    const __m512i mins_45 = _mm512_cvtepu8_epi16(_mm256_and_si256(_mm256_srli_epi16(mins_and_scales_45, 4), m4b));
3833                    const __m512i mins_67 = _mm512_cvtepu8_epi16(_mm256_and_si256(_mm256_srli_epi16(mins_and_scales_67, 4), m4b));
3834
3835                    const __m512i scales_0 = _mm512_cvtepu8_epi16(_mm256_shuffle_epi8(scales_01,scalesmask1));
3836                    const __m512i scales_1 = _mm512_cvtepu8_epi16(_mm256_shuffle_epi8(scales_01,scalesmask2));
3837                    const __m512i scales_2 = _mm512_cvtepu8_epi16(_mm256_shuffle_epi8(scales_23,scalesmask1));
3838                    const __m512i scales_3 = _mm512_cvtepu8_epi16(_mm256_shuffle_epi8(scales_23,scalesmask2));
3839                    const __m512i scales_4 = _mm512_cvtepu8_epi16(_mm256_shuffle_epi8(scales_45,scalesmask1));
3840                    const __m512i scales_5 = _mm512_cvtepu8_epi16(_mm256_shuffle_epi8(scales_45,scalesmask2));
3841                    const __m512i scales_6 = _mm512_cvtepu8_epi16(_mm256_shuffle_epi8(scales_67,scalesmask1));
3842                    const __m512i scales_7 = _mm512_cvtepu8_epi16(_mm256_shuffle_epi8(scales_67,scalesmask2));
3843
3844                    const __m512i scale_014589CD_0 = _mm512_shuffle_epi32(scales_0, (_MM_PERM_ENUM)68);
3845                    const __m512i scale_2367ABEF_0 = _mm512_shuffle_epi32(scales_0, (_MM_PERM_ENUM)238);
3846
3847                    const __m512i scale_014589CD_1 = _mm512_shuffle_epi32(scales_1, (_MM_PERM_ENUM)68);
3848                    const __m512i scale_2367ABEF_1 = _mm512_shuffle_epi32(scales_1, (_MM_PERM_ENUM)238);
3849
3850                    const __m512i scale_014589CD_2 = _mm512_shuffle_epi32(scales_2, (_MM_PERM_ENUM)68);
3851                    const __m512i scale_2367ABEF_2 = _mm512_shuffle_epi32(scales_2, (_MM_PERM_ENUM)238);
3852
3853                    const __m512i scale_014589CD_3 = _mm512_shuffle_epi32(scales_3, (_MM_PERM_ENUM)68);
3854                    const __m512i scale_2367ABEF_3 = _mm512_shuffle_epi32(scales_3, (_MM_PERM_ENUM)238);
3855
3856                    const __m512i scale_014589CD_4 = _mm512_shuffle_epi32(scales_4, (_MM_PERM_ENUM)68);
3857                    const __m512i scale_2367ABEF_4 = _mm512_shuffle_epi32(scales_4, (_MM_PERM_ENUM)238);
3858
3859                    const __m512i scale_014589CD_5 = _mm512_shuffle_epi32(scales_5, (_MM_PERM_ENUM)68);
3860                    const __m512i scale_2367ABEF_5 = _mm512_shuffle_epi32(scales_5, (_MM_PERM_ENUM)238);
3861
3862                    const __m512i scale_014589CD_6 = _mm512_shuffle_epi32(scales_6, (_MM_PERM_ENUM)68);
3863                    const __m512i scale_2367ABEF_6 = _mm512_shuffle_epi32(scales_6, (_MM_PERM_ENUM)238);
3864
3865                    const __m512i scale_014589CD_7 = _mm512_shuffle_epi32(scales_7, (_MM_PERM_ENUM)68);
3866                    const __m512i scale_2367ABEF_7 = _mm512_shuffle_epi32(scales_7, (_MM_PERM_ENUM)238);
3867
3868
3869                    for (int rp = 0; rp < 4; rp++) {
3870
3871                        // Load the four block_q8_k quantized values interleaved with each other in chunks of eight bytes - A0,A1,A2,A3
3872                        // Loaded as set of 128 bit vectors and repeated and stored into a 256 bit vector before again repeating into 512 bit vector
3873                        __m256i lhs_mat_ymm_0123_00 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 512 * sb)));
3874                        __m256i lhs_mat_ymm_01_00 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_00, lhs_mat_ymm_0123_00, 0);
3875                        __m256i lhs_mat_ymm_23_00 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_00, lhs_mat_ymm_0123_00, 17);
3876                        __m256i lhs_mat_ymm_0123_01 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 32 + 512 * sb)));
3877                        __m256i lhs_mat_ymm_01_01 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_01, lhs_mat_ymm_0123_01, 0);
3878                        __m256i lhs_mat_ymm_23_01 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_01, lhs_mat_ymm_0123_01, 17);
3879                        __m256i lhs_mat_ymm_0123_10 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 64 + 512 * sb)));
3880                        __m256i lhs_mat_ymm_01_10 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_10, lhs_mat_ymm_0123_10, 0);
3881                        __m256i lhs_mat_ymm_23_10 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_10, lhs_mat_ymm_0123_10, 17);
3882                        __m256i lhs_mat_ymm_0123_11 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 96 + 512 * sb)));
3883                        __m256i lhs_mat_ymm_01_11 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_11, lhs_mat_ymm_0123_11, 0);
3884                        __m256i lhs_mat_ymm_23_11 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_11, lhs_mat_ymm_0123_11, 17);
3885                        __m256i lhs_mat_ymm_0123_20 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 128 + 512 * sb)));
3886                        __m256i lhs_mat_ymm_01_20 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_20, lhs_mat_ymm_0123_20, 0);
3887                        __m256i lhs_mat_ymm_23_20 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_20, lhs_mat_ymm_0123_20, 17);
3888                        __m256i lhs_mat_ymm_0123_21 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 160 + 512 * sb)));
3889                        __m256i lhs_mat_ymm_01_21 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_21, lhs_mat_ymm_0123_21, 0);
3890                        __m256i lhs_mat_ymm_23_21 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_21, lhs_mat_ymm_0123_21, 17);
3891                        __m256i lhs_mat_ymm_0123_30 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 192 + 512 * sb)));
3892                        __m256i lhs_mat_ymm_01_30 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_30, lhs_mat_ymm_0123_30, 0);
3893                        __m256i lhs_mat_ymm_23_30 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_30, lhs_mat_ymm_0123_30, 17);
3894                        __m256i lhs_mat_ymm_0123_31 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 224 + 512 * sb)));
3895                        __m256i lhs_mat_ymm_01_31 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_31, lhs_mat_ymm_0123_31, 0);
3896                        __m256i lhs_mat_ymm_23_31 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_31, lhs_mat_ymm_0123_31, 17);
3897
3898                        __m256i lhs_mat_ymm_0123_40 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 256 + 512 * sb)));
3899                        __m256i lhs_mat_ymm_01_40 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_40, lhs_mat_ymm_0123_40, 0);
3900                        __m256i lhs_mat_ymm_23_40 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_40, lhs_mat_ymm_0123_40, 17);
3901                        __m256i lhs_mat_ymm_0123_41 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 288 + 512 * sb)));
3902                        __m256i lhs_mat_ymm_01_41 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_41, lhs_mat_ymm_0123_41, 0);
3903                        __m256i lhs_mat_ymm_23_41 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_41, lhs_mat_ymm_0123_41, 17);
3904                        __m256i lhs_mat_ymm_0123_50 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 320 + 512 * sb)));
3905                        __m256i lhs_mat_ymm_01_50 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_50, lhs_mat_ymm_0123_50, 0);
3906                        __m256i lhs_mat_ymm_23_50 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_50, lhs_mat_ymm_0123_50, 17);
3907                        __m256i lhs_mat_ymm_0123_51 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 352 + 512 * sb)));
3908                        __m256i lhs_mat_ymm_01_51 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_51, lhs_mat_ymm_0123_51, 0);
3909                        __m256i lhs_mat_ymm_23_51 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_51, lhs_mat_ymm_0123_51, 17);
3910                        __m256i lhs_mat_ymm_0123_60 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 384 + 512 * sb)));
3911                        __m256i lhs_mat_ymm_01_60 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_60, lhs_mat_ymm_0123_60, 0);
3912                        __m256i lhs_mat_ymm_23_60 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_60, lhs_mat_ymm_0123_60, 17);
3913                        __m256i lhs_mat_ymm_0123_61 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 416 + 512 * sb)));
3914                        __m256i lhs_mat_ymm_01_61 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_61, lhs_mat_ymm_0123_61, 0);
3915                        __m256i lhs_mat_ymm_23_61 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_61, lhs_mat_ymm_0123_61, 17);
3916                        __m256i lhs_mat_ymm_0123_70 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 448 + 512 * sb)));
3917                        __m256i lhs_mat_ymm_01_70 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_70, lhs_mat_ymm_0123_70, 0);
3918                        __m256i lhs_mat_ymm_23_70 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_70, lhs_mat_ymm_0123_70, 17);
3919                        __m256i lhs_mat_ymm_0123_71 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 480 + 512 * sb)));
3920                        __m256i lhs_mat_ymm_01_71 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_71, lhs_mat_ymm_0123_71, 0);
3921                        __m256i lhs_mat_ymm_23_71 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_71, lhs_mat_ymm_0123_71, 17);
3922
3923
3924                        __m512i lhs_mat_01_00 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_00), lhs_mat_ymm_01_00, 1);
3925                        __m512i lhs_mat_23_00 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_00), lhs_mat_ymm_23_00, 1);
3926                        __m512i lhs_mat_01_01 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_01), lhs_mat_ymm_01_01, 1);
3927                        __m512i lhs_mat_23_01 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_01), lhs_mat_ymm_23_01, 1);
3928
3929                        __m512i lhs_mat_01_10 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_10), lhs_mat_ymm_01_10, 1);
3930                        __m512i lhs_mat_23_10 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_10), lhs_mat_ymm_23_10, 1);
3931                        __m512i lhs_mat_01_11 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_11), lhs_mat_ymm_01_11, 1);
3932                        __m512i lhs_mat_23_11 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_11), lhs_mat_ymm_23_11, 1);
3933
3934                        __m512i lhs_mat_01_20 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_20), lhs_mat_ymm_01_20, 1);
3935                        __m512i lhs_mat_23_20 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_20), lhs_mat_ymm_23_20, 1);
3936                        __m512i lhs_mat_01_21 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_21), lhs_mat_ymm_01_21, 1);
3937                        __m512i lhs_mat_23_21 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_21), lhs_mat_ymm_23_21, 1);
3938
3939                        __m512i lhs_mat_01_30 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_30), lhs_mat_ymm_01_30, 1);
3940                        __m512i lhs_mat_23_30 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_30), lhs_mat_ymm_23_30, 1);
3941                        __m512i lhs_mat_01_31 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_31), lhs_mat_ymm_01_31, 1);
3942                        __m512i lhs_mat_23_31 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_31), lhs_mat_ymm_23_31, 1);
3943
3944                        __m512i lhs_mat_01_40 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_40), lhs_mat_ymm_01_40, 1);
3945                        __m512i lhs_mat_23_40 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_40), lhs_mat_ymm_23_40, 1);
3946                        __m512i lhs_mat_01_41 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_41), lhs_mat_ymm_01_41, 1);
3947                        __m512i lhs_mat_23_41 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_41), lhs_mat_ymm_23_41, 1);
3948
3949                        __m512i lhs_mat_01_50 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_50), lhs_mat_ymm_01_50, 1);
3950                        __m512i lhs_mat_23_50 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_50), lhs_mat_ymm_23_50, 1);
3951                        __m512i lhs_mat_01_51 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_51), lhs_mat_ymm_01_51, 1);
3952                        __m512i lhs_mat_23_51 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_51), lhs_mat_ymm_23_51, 1);
3953
3954                        __m512i lhs_mat_01_60 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_60), lhs_mat_ymm_01_60, 1);
3955                        __m512i lhs_mat_23_60 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_60), lhs_mat_ymm_23_60, 1);
3956                        __m512i lhs_mat_01_61 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_61), lhs_mat_ymm_01_61, 1);
3957                        __m512i lhs_mat_23_61 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_61), lhs_mat_ymm_23_61, 1);
3958
3959                        __m512i lhs_mat_01_70 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_70), lhs_mat_ymm_01_70, 1);
3960                        __m512i lhs_mat_23_70 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_70), lhs_mat_ymm_23_70, 1);
3961                        __m512i lhs_mat_01_71 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_71), lhs_mat_ymm_01_71, 1);
3962                        __m512i lhs_mat_23_71 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_71), lhs_mat_ymm_23_71, 1);
3963
3964                        // Bsums are loaded for the different Q8_K blocks
3965                        __m128i lhs_raw_bsums_01_0123 = _mm_loadu_si128((const __m128i *)((a_ptrs[rp][b].bsums + 32 * sb)));
3966                        __m128i lhs_raw_bsums_23_0123 = _mm_loadu_si128((const __m128i *)(a_ptrs[rp][b].bsums + 8 + 32 * sb));
3967                        __m128i lhs_raw_bsums_01_4567 = _mm_loadu_si128((const __m128i *)((a_ptrs[rp][b].bsums + 16 + 32 * sb)));
3968                        __m128i lhs_raw_bsums_23_4567 = _mm_loadu_si128((const __m128i *)(a_ptrs[rp][b].bsums + 24 + 32 * sb));
3969
3970                        __m256i lhs_bsums_ymm_01_0123 = _mm256_inserti128_si256(_mm256_castsi128_si256(lhs_raw_bsums_01_0123), lhs_raw_bsums_01_0123, 1);
3971                        __m512i lhs_bsums_01_0123 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_bsums_ymm_01_0123), lhs_bsums_ymm_01_0123, 1);
3972                        __m256i lhs_bsums_ymm_23_0123 = _mm256_inserti128_si256(_mm256_castsi128_si256(lhs_raw_bsums_23_0123), lhs_raw_bsums_23_0123, 1);
3973                        __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);
3974                        __m512i lhs_bsums_01_4567 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_bsums_ymm_01_4567), lhs_bsums_ymm_01_4567, 1);
3975                        __m256i lhs_bsums_ymm_23_4567 = _mm256_inserti128_si256(_mm256_castsi128_si256(lhs_raw_bsums_23_4567), lhs_raw_bsums_23_4567, 1);
3976                        __m512i lhs_bsums_23_4567 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_bsums_ymm_23_4567), lhs_bsums_ymm_23_4567, 1);
3977
3978                        // Shuffle pattern one - left side input
3979                        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)
3980                        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)
3981
3982                        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)
3983                        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)
3984
3985                        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)
3986                        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)
3987
3988                        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)
3989                        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)
3990
3991                        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)
3992                        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)
3993
3994                        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)
3995                        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)
3996
3997                        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)
3998                        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)
3999
4000                        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)
4001                        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)
4002
4003                        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)
4004                        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)
4005
4006                        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)
4007                        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)
4008
4009                        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)
4010                        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)
4011
4012                        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)
4013                        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)
4014
4015                        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)
4016                        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)
4017
4018                        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)
4019                        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)
4020
4021                        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)
4022                        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)
4023
4024                        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)
4025                        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)
4026
4027                        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)
4028                        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)
4029
4030                        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)
4031                        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)
4032
4033                        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)
4034                        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)
4035
4036                        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)
4037                        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)
4038
4039                        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)
4040                        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)
4041
4042                        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)
4043                        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)
4044
4045                        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)
4046                        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)
4047
4048                        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)
4049                        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)
4050
4051                        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)
4052                        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)
4053
4054                        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)
4055                        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)
4056
4057                        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)
4058                        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)
4059
4060                        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)
4061                        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)
4062
4063                        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)
4064                        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)
4065
4066                        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)
4067                        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)
4068
4069                        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)
4070                        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)
4071
4072                        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)
4073                        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)
4074
4075                        // 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
4076                        __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));
4077                        __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));
4078
4079                        __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));
4080                        __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));
4081
4082                        __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));
4083                        __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));
4084
4085                        __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));
4086                        __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));
4087
4088                        __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));
4089                        __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));
4090
4091                        __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));
4092                        __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));
4093
4094                        __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));
4095                        __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));
4096
4097                        __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));
4098                        __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));
4099
4100                        __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));
4101                        __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));
4102
4103                        __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));
4104                        __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));
4105
4106                        __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));
4107                        __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));
4108
4109                        __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));
4110                        __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));
4111
4112                        __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));
4113                        __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));
4114
4115                        __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));
4116                        __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));
4117
4118                        __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));
4119                        __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));
4120
4121                        __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));
4122                        __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));
4123
4124
4125                        __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));
4126                        __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));
4127
4128                        __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));
4129                        __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));
4130
4131                        __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));
4132                        __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));
4133
4134                        __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));
4135                        __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));
4136
4137                        __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));
4138                        __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));
4139
4140                        __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));
4141                        __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));
4142
4143                        __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));
4144                        __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));
4145
4146                        __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));
4147                        __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));
4148
4149                        __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));
4150                        __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));
4151
4152                        __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));
4153                        __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));
4154
4155                        __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));
4156                        __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));
4157
4158                        __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));
4159                        __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));
4160
4161                        __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));
4162                        __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));
4163
4164                        __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));
4165                        __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));
4166
4167                        __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));
4168                        __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));
4169
4170                        __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));
4171                        __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));
4172
4173                        // Combine results from both shuffle patterns for each output block
4174                        __m512i iacc_mat_00_0 = _mm512_add_epi16(iacc_mat_00_0_sp1, iacc_mat_00_0_sp2);
4175                        __m512i iacc_mat_01_0 = _mm512_add_epi16(iacc_mat_01_0_sp1, iacc_mat_01_0_sp2);
4176                        __m512i iacc_mat_10_0 = _mm512_add_epi16(iacc_mat_10_0_sp1, iacc_mat_10_0_sp2);
4177                        __m512i iacc_mat_11_0 = _mm512_add_epi16(iacc_mat_11_0_sp1, iacc_mat_11_0_sp2);
4178
4179                        __m512i iacc_mat_00_1 = _mm512_add_epi16(iacc_mat_00_1_sp1, iacc_mat_00_1_sp2);
4180                        __m512i iacc_mat_01_1 = _mm512_add_epi16(iacc_mat_01_1_sp1, iacc_mat_01_1_sp2);
4181                        __m512i iacc_mat_10_1 = _mm512_add_epi16(iacc_mat_10_1_sp1, iacc_mat_10_1_sp2);
4182                        __m512i iacc_mat_11_1 = _mm512_add_epi16(iacc_mat_11_1_sp1, iacc_mat_11_1_sp2);
4183
4184                        __m512i iacc_mat_00_2 = _mm512_add_epi16(iacc_mat_00_2_sp1, iacc_mat_00_2_sp2);
4185                        __m512i iacc_mat_01_2 = _mm512_add_epi16(iacc_mat_01_2_sp1, iacc_mat_01_2_sp2);
4186                        __m512i iacc_mat_10_2 = _mm512_add_epi16(iacc_mat_10_2_sp1, iacc_mat_10_2_sp2);
4187                        __m512i iacc_mat_11_2 = _mm512_add_epi16(iacc_mat_11_2_sp1, iacc_mat_11_2_sp2);
4188
4189                        __m512i iacc_mat_00_3 = _mm512_add_epi16(iacc_mat_00_3_sp1, iacc_mat_00_3_sp2);
4190                        __m512i iacc_mat_01_3 = _mm512_add_epi16(iacc_mat_01_3_sp1, iacc_mat_01_3_sp2);
4191                        __m512i iacc_mat_10_3 = _mm512_add_epi16(iacc_mat_10_3_sp1, iacc_mat_10_3_sp2);
4192                        __m512i iacc_mat_11_3 = _mm512_add_epi16(iacc_mat_11_3_sp1, iacc_mat_11_3_sp2);
4193
4194                        __m512i iacc_mat_00_4 = _mm512_add_epi16(iacc_mat_00_4_sp1, iacc_mat_00_4_sp2);
4195                        __m512i iacc_mat_01_4 = _mm512_add_epi16(iacc_mat_01_4_sp1, iacc_mat_01_4_sp2);
4196                        __m512i iacc_mat_10_4 = _mm512_add_epi16(iacc_mat_10_4_sp1, iacc_mat_10_4_sp2);
4197                        __m512i iacc_mat_11_4 = _mm512_add_epi16(iacc_mat_11_4_sp1, iacc_mat_11_4_sp2);
4198
4199                        __m512i iacc_mat_00_5 = _mm512_add_epi16(iacc_mat_00_5_sp1, iacc_mat_00_5_sp2);
4200                        __m512i iacc_mat_01_5 = _mm512_add_epi16(iacc_mat_01_5_sp1, iacc_mat_01_5_sp2);
4201                        __m512i iacc_mat_10_5 = _mm512_add_epi16(iacc_mat_10_5_sp1, iacc_mat_10_5_sp2);
4202                        __m512i iacc_mat_11_5 = _mm512_add_epi16(iacc_mat_11_5_sp1, iacc_mat_11_5_sp2);
4203
4204                        __m512i iacc_mat_00_6 = _mm512_add_epi16(iacc_mat_00_6_sp1, iacc_mat_00_6_sp2);
4205                        __m512i iacc_mat_01_6 = _mm512_add_epi16(iacc_mat_01_6_sp1, iacc_mat_01_6_sp2);
4206                        __m512i iacc_mat_10_6 = _mm512_add_epi16(iacc_mat_10_6_sp1, iacc_mat_10_6_sp2);
4207                        __m512i iacc_mat_11_6 = _mm512_add_epi16(iacc_mat_11_6_sp1, iacc_mat_11_6_sp2);
4208
4209                        __m512i iacc_mat_00_7 = _mm512_add_epi16(iacc_mat_00_7_sp1, iacc_mat_00_7_sp2);
4210                        __m512i iacc_mat_01_7 = _mm512_add_epi16(iacc_mat_01_7_sp1, iacc_mat_01_7_sp2);
4211                        __m512i iacc_mat_10_7 = _mm512_add_epi16(iacc_mat_10_7_sp1, iacc_mat_10_7_sp2);
4212                        __m512i iacc_mat_11_7 = _mm512_add_epi16(iacc_mat_11_7_sp1, iacc_mat_11_7_sp2);
4213
4214                        // Output of both shuffle patterns are added in order to sum dot product outputs of all 32 values in block
4215                        iacc_mat_00_0 = _mm512_madd_epi16(iacc_mat_00_0, scale_014589CD_0);
4216                        iacc_mat_01_0 = _mm512_madd_epi16(iacc_mat_01_0, scale_2367ABEF_0);
4217                        iacc_mat_10_0 = _mm512_madd_epi16(iacc_mat_10_0, scale_014589CD_0);
4218                        iacc_mat_11_0 = _mm512_madd_epi16(iacc_mat_11_0, scale_2367ABEF_0);
4219
4220                        iacc_mat_00_1 = _mm512_madd_epi16(iacc_mat_00_1, scale_014589CD_1);
4221                        iacc_mat_01_1 = _mm512_madd_epi16(iacc_mat_01_1, scale_2367ABEF_1);
4222                        iacc_mat_10_1 = _mm512_madd_epi16(iacc_mat_10_1, scale_014589CD_1);
4223                        iacc_mat_11_1 = _mm512_madd_epi16(iacc_mat_11_1, scale_2367ABEF_1);
4224
4225                        iacc_mat_00_2 = _mm512_madd_epi16(iacc_mat_00_2, scale_014589CD_2);
4226                        iacc_mat_01_2 = _mm512_madd_epi16(iacc_mat_01_2, scale_2367ABEF_2);
4227                        iacc_mat_10_2 = _mm512_madd_epi16(iacc_mat_10_2, scale_014589CD_2);
4228                        iacc_mat_11_2 = _mm512_madd_epi16(iacc_mat_11_2, scale_2367ABEF_2);
4229
4230                        iacc_mat_00_3 = _mm512_madd_epi16(iacc_mat_00_3, scale_014589CD_3);
4231                        iacc_mat_01_3 = _mm512_madd_epi16(iacc_mat_01_3, scale_2367ABEF_3);
4232                        iacc_mat_10_3 = _mm512_madd_epi16(iacc_mat_10_3, scale_014589CD_3);
4233                        iacc_mat_11_3 = _mm512_madd_epi16(iacc_mat_11_3, scale_2367ABEF_3);
4234
4235                        iacc_mat_00_4 = _mm512_madd_epi16(iacc_mat_00_4, scale_014589CD_4);
4236                        iacc_mat_01_4 = _mm512_madd_epi16(iacc_mat_01_4, scale_2367ABEF_4);
4237                        iacc_mat_10_4 = _mm512_madd_epi16(iacc_mat_10_4, scale_014589CD_4);
4238                        iacc_mat_11_4 = _mm512_madd_epi16(iacc_mat_11_4, scale_2367ABEF_4);
4239
4240                        iacc_mat_00_5 = _mm512_madd_epi16(iacc_mat_00_5, scale_014589CD_5);
4241                        iacc_mat_01_5 = _mm512_madd_epi16(iacc_mat_01_5, scale_2367ABEF_5);
4242                        iacc_mat_10_5 = _mm512_madd_epi16(iacc_mat_10_5, scale_014589CD_5);
4243                        iacc_mat_11_5 = _mm512_madd_epi16(iacc_mat_11_5, scale_2367ABEF_5);
4244
4245                        iacc_mat_00_6 = _mm512_madd_epi16(iacc_mat_00_6, scale_014589CD_6);
4246                        iacc_mat_01_6 = _mm512_madd_epi16(iacc_mat_01_6, scale_2367ABEF_6);
4247                        iacc_mat_10_6 = _mm512_madd_epi16(iacc_mat_10_6, scale_014589CD_6);
4248                        iacc_mat_11_6 = _mm512_madd_epi16(iacc_mat_11_6, scale_2367ABEF_6);
4249
4250                        iacc_mat_00_7 = _mm512_madd_epi16(iacc_mat_00_7, scale_014589CD_7);
4251                        iacc_mat_01_7 = _mm512_madd_epi16(iacc_mat_01_7, scale_2367ABEF_7);
4252                        iacc_mat_10_7 = _mm512_madd_epi16(iacc_mat_10_7, scale_014589CD_7);
4253                        iacc_mat_11_7 = _mm512_madd_epi16(iacc_mat_11_7, scale_2367ABEF_7);
4254
4255                        __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)));
4256                        __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)));
4257                        __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)));
4258                        __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)));
4259
4260                        // Straighten out to make 4 row vectors
4261                        __m512i iacc_row_0 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_00, _mm512_shuffle_epi32(iacc_mat_01, (_MM_PERM_ENUM)78));
4262                        __m512i iacc_row_1 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_00, (_MM_PERM_ENUM)78), iacc_mat_01);
4263                        __m512i iacc_row_2 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_10, _mm512_shuffle_epi32(iacc_mat_11, (_MM_PERM_ENUM)78));
4264                        __m512i iacc_row_3 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_10, (_MM_PERM_ENUM)78), iacc_mat_11);
4265
4266                        // Load the scale(d) values for all the 4 Q8_k blocks and repeat it across lanes
4267                        const __m128 row_scale_f32_sse = _mm_load_ps(a_ptrs[rp][b].d);
4268                        const __m256 row_scale_f32_ymm = _mm256_set_m128(row_scale_f32_sse, row_scale_f32_sse);
4269                        const __m512 row_scale_f32 = _mm512_insertf32x8(_mm512_castps256_ps512(row_scale_f32_ymm), row_scale_f32_ymm, 1);
4270
4271                        // Multiply with appropriate scales and accumulate (for both d and dmin) below
4272                        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]);
4273                        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]);
4274                        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]);
4275                        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]);
4276
4277                        // Take two bsums from two Q8_Ks at a time and multiply with corresponding mins values from each Q2_K
4278                        __m512i iacc_row_min_0_01 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_01_0123, (_MM_PERM_ENUM)0), mins_01);
4279                        __m512i iacc_row_min_1_01 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_01_0123, (_MM_PERM_ENUM)170), mins_01);
4280                        __m512i iacc_row_min_2_01 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_23_0123, (_MM_PERM_ENUM)0), mins_01);
4281                        __m512i iacc_row_min_3_01 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_23_0123, (_MM_PERM_ENUM)170), mins_01);
4282
4283                        __m512i iacc_row_min_0_23 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_01_0123, (_MM_PERM_ENUM)85), mins_23);
4284                        __m512i iacc_row_min_1_23 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_01_0123, (_MM_PERM_ENUM)255), mins_23);
4285                        __m512i iacc_row_min_2_23 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_23_0123, (_MM_PERM_ENUM)85), mins_23);
4286                        __m512i iacc_row_min_3_23 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_23_0123, (_MM_PERM_ENUM)255), mins_23);
4287
4288                        __m512i iacc_row_min_0_45 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_01_4567, (_MM_PERM_ENUM)0), mins_45);
4289                        __m512i iacc_row_min_1_45 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_01_4567, (_MM_PERM_ENUM)170), mins_45);
4290                        __m512i iacc_row_min_2_45 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_23_4567, (_MM_PERM_ENUM)0), mins_45);
4291                        __m512i iacc_row_min_3_45 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_23_4567, (_MM_PERM_ENUM)170), mins_45);
4292
4293                        __m512i iacc_row_min_0_67 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_01_4567, (_MM_PERM_ENUM)85), mins_67);
4294                        __m512i iacc_row_min_1_67 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_01_4567, (_MM_PERM_ENUM)255), mins_67);
4295                        __m512i iacc_row_min_2_67 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_23_4567, (_MM_PERM_ENUM)85), mins_67);
4296                        __m512i iacc_row_min_3_67 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_23_4567, (_MM_PERM_ENUM)255), mins_67);
4297
4298                        __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));
4299                        __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));
4300                        __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));
4301                        __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));
4302
4303                        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]);
4304                        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]);
4305                        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]);
4306                        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]);
4307                    }
4308                }
4309            }
4310            // Store the accumulated values
4311            for (int i = 0; i < 16; i++) {
4312                _mm512_storeu_ps((float * )(s + ((y * 4 + i) * bs + x * 8)), _mm512_sub_ps(acc_rows[i], acc_min_rows[i]));
4313            }
4314        }
4315    }
4316
4317    for (; y < nr / 4; y ++) {
4318
4319        const block_q8_Kx4 * a_ptr = a_ptr_start + (y * nb);
4320
4321        // Take group of eight block_q2_kx8 structures at each pass of the loop and perform dot product operation
4322        for (int64_t x = 0; x < anc / 8; x += 2) {
4323
4324            const block_q2_Kx8 * b_ptr_0 = b_ptr_start + ((x) * b_nb);
4325            const block_q2_Kx8 * b_ptr_1 = b_ptr_start + ((x + 1) * b_nb);
4326
4327            // Master FP accumulators
4328            __m512 acc_rows[4];
4329            for (int i = 0; i < 4; i++) {
4330                acc_rows[i] = _mm512_setzero_ps();
4331            }
4332
4333            __m512 acc_min_rows[4];
4334            for (int i = 0; i < 4; i++) {
4335                acc_min_rows[i] = _mm512_setzero_ps();
4336            }
4337            // For super block
4338            for (int64_t b = 0; b < nb; b++) {
4339                // Delta values - Load the sixteen scale values from two block_q2_kx8 structures
4340                const __m512 col_scale_f32 = GGML_F32Cx8x2_LOAD(b_ptr_0[b].d, b_ptr_1[b].d);
4341
4342                // dmin values - Load the sixteen dmin values from two block_q2_kx8 structures
4343                const __m512 col_dmin_f32 = GGML_F32Cx8x2_LOAD(b_ptr_0[b].dmin, b_ptr_1[b].dmin);
4344
4345                // Loop to iterate over the sixteen sub blocks of a super block - eight sub blocks are processed per iteration
4346                for (int sb = 0; sb < QK_K / 128; sb++) {
4347
4348                    // Load the eight block_q2_k for eight sub blocks quantized values interleaved with each other in chunks of eight bytes - B0,B1 ....B6,B7
4349                    const __m256i rhs_raw_mat_0123_0 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + sb * 256));
4350                    const __m256i rhs_raw_mat_4567_0 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 32 + sb * 256));
4351                    const __m256i rhs_raw_mat_0123_1 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 64 + sb * 256));
4352                    const __m256i rhs_raw_mat_4567_1 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 96 + sb * 256));
4353                    const __m256i rhs_raw_mat_0123_2 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 128 + sb * 256));
4354                    const __m256i rhs_raw_mat_4567_2 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 160 + sb * 256));
4355                    const __m256i rhs_raw_mat_0123_3 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 192 + sb * 256));
4356                    const __m256i rhs_raw_mat_4567_3 = _mm256_loadu_si256((const __m256i * )(b_ptr_0[b].qs + 224 + sb * 256));
4357
4358                    const __m256i rhs_raw_mat_89AB_0 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + sb * 256));
4359                    const __m256i rhs_raw_mat_CDEF_0 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 32 + sb * 256));
4360                    const __m256i rhs_raw_mat_89AB_1 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 64 + sb * 256));
4361                    const __m256i rhs_raw_mat_CDEF_1 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 96 + sb * 256));
4362                    const __m256i rhs_raw_mat_89AB_2 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 128 + sb * 256));
4363                    const __m256i rhs_raw_mat_CDEF_2 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 160 + sb * 256));
4364                    const __m256i rhs_raw_mat_89AB_3 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 192 + sb * 256));
4365                    const __m256i rhs_raw_mat_CDEF_3 = _mm256_loadu_si256((const __m256i * )(b_ptr_1[b].qs + 224 + sb * 256));
4366
4367                    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);
4368                    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);
4369                    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);
4370                    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);
4371                    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);
4372                    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);
4373                    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);
4374                    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);
4375
4376                    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);
4377                    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);
4378                    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);
4379                    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);
4380                    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);
4381                    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);
4382                    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);
4383                    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);
4384
4385                    const __m512i rhs_raw_mat_014589CD_0 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_0), rhs_raw_mat_89CD_0, 1);
4386                    const __m512i rhs_raw_mat_2367ABEF_0 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_0), rhs_raw_mat_ABEF_0, 1);
4387                    const __m512i rhs_raw_mat_014589CD_1 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_1), rhs_raw_mat_89CD_1, 1);
4388                    const __m512i rhs_raw_mat_2367ABEF_1 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_1), rhs_raw_mat_ABEF_1, 1);
4389
4390                    const __m512i rhs_raw_mat_014589CD_2 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_2), rhs_raw_mat_89CD_2, 1);
4391                    const __m512i rhs_raw_mat_2367ABEF_2 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_2), rhs_raw_mat_ABEF_2, 1);
4392                    const __m512i rhs_raw_mat_014589CD_3 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_0145_3), rhs_raw_mat_89CD_3, 1);
4393                    const __m512i rhs_raw_mat_2367ABEF_3 = _mm512_inserti32x8(_mm512_castsi256_si512(rhs_raw_mat_2367_3), rhs_raw_mat_ABEF_3, 1);
4394
4395                    //2-bit -> 8-bit
4396                    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)
4397                    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)
4398                    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)
4399                    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)
4400                    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)
4401                    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)
4402                    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)
4403                    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)
4404
4405                    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)
4406                    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)
4407
4408                    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)
4409                    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)
4410
4411                    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)
4412                    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)
4413
4414                    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)
4415                    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)
4416
4417                    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)
4418                    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)
4419
4420                    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)
4421                    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)
4422
4423                    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)
4424                    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)
4425
4426                    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)
4427                    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)
4428
4429                    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)
4430                    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)
4431
4432                    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)
4433                    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)
4434
4435                    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)
4436                    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)
4437
4438                    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)
4439                    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)
4440
4441                    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)
4442                    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)
4443
4444                    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)
4445                    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)
4446
4447                    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)
4448                    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)
4449
4450                    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)
4451                    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)
4452
4453                    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)
4454                    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)
4455
4456                    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)
4457                    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)
4458                    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)
4459                    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)
4460
4461                    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)
4462                    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)
4463
4464                    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)
4465                    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)
4466
4467                    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)
4468                    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)
4469
4470                    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)
4471                    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)
4472
4473                    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)
4474                    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)
4475
4476                    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)
4477                    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)
4478
4479                    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)
4480                    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)
4481
4482                    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)
4483                    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)
4484
4485                    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)
4486                    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)
4487
4488                    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)
4489                    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)
4490
4491                    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)
4492                    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)
4493
4494                    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)
4495                    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)
4496
4497                    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)
4498                    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)
4499
4500                    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)
4501                    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)
4502
4503                    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)
4504                    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)
4505
4506                    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)
4507                    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)
4508
4509                    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)
4510                    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)
4511
4512                    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)
4513                    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)
4514
4515                    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)
4516                    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)
4517
4518                    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)
4519                    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)
4520
4521                    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)
4522                    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)
4523
4524                    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)
4525                    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)
4526
4527                    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)
4528                    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)
4529
4530                    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)
4531                    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)
4532
4533                    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)
4534                    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)
4535
4536                    //notation:superblock subblock
4537                    //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
4538
4539                    const __m128i mins_and_scales_01_0 = _mm_loadu_si128((const __m128i *)(b_ptr_0[b].scales + sb * 64));
4540                    const __m128i mins_and_scales_23_0 = _mm_loadu_si128((const __m128i *)(b_ptr_0[b].scales + 16 + sb * 64));
4541                    const __m128i mins_and_scales_45_0 = _mm_loadu_si128((const __m128i *)(b_ptr_0[b].scales + 32 + sb * 64));
4542                    const __m128i mins_and_scales_67_0 = _mm_loadu_si128((const __m128i *)(b_ptr_0[b].scales + 48 + sb * 64));
4543
4544                    const __m128i mins_and_scales_01_1 = _mm_loadu_si128((const __m128i *)(b_ptr_1[b].scales + sb * 64));
4545                    const __m128i mins_and_scales_23_1 = _mm_loadu_si128((const __m128i *)(b_ptr_1[b].scales + 16 + sb * 64));
4546                    const __m128i mins_and_scales_45_1 = _mm_loadu_si128((const __m128i *)(b_ptr_1[b].scales + 32 + sb * 64));
4547                    const __m128i mins_and_scales_67_1 = _mm_loadu_si128((const __m128i *)(b_ptr_1[b].scales + 48 + sb * 64));
4548
4549                    // Combine mins and scales for sub-blocks: 0-1, 2-3, 4-5, 6-7 in the sb loop
4550                    const __m256i mins_and_scales_01 = _mm256_insertf128_si256(_mm256_castsi128_si256(mins_and_scales_01_0), mins_and_scales_01_1, 1);
4551                    const __m256i mins_and_scales_23 = _mm256_insertf128_si256(_mm256_castsi128_si256(mins_and_scales_23_0), mins_and_scales_23_1, 1);
4552                    const __m256i mins_and_scales_45 = _mm256_insertf128_si256(_mm256_castsi128_si256(mins_and_scales_45_0), mins_and_scales_45_1, 1);
4553                    const __m256i mins_and_scales_67 = _mm256_insertf128_si256(_mm256_castsi128_si256(mins_and_scales_67_0), mins_and_scales_67_1, 1);
4554
4555                    // Extract scales which is lower half from mins_and_scales
4556                    const __m256i scales_01 = _mm256_and_si256(mins_and_scales_01, m4b);
4557                    const __m256i scales_23 = _mm256_and_si256(mins_and_scales_23, m4b);
4558                    const __m256i scales_45 = _mm256_and_si256(mins_and_scales_45, m4b);
4559                    const __m256i scales_67 = _mm256_and_si256(mins_and_scales_67, m4b);
4560
4561                    // Extract mins which is upper half from mins_and_scales
4562                    const __m512i mins_01 = _mm512_cvtepu8_epi16(_mm256_and_si256(_mm256_srli_epi16(mins_and_scales_01, 4), m4b));
4563                    const __m512i mins_23 = _mm512_cvtepu8_epi16(_mm256_and_si256(_mm256_srli_epi16(mins_and_scales_23, 4), m4b));
4564                    const __m512i mins_45 = _mm512_cvtepu8_epi16(_mm256_and_si256(_mm256_srli_epi16(mins_and_scales_45, 4), m4b));
4565                    const __m512i mins_67 = _mm512_cvtepu8_epi16(_mm256_and_si256(_mm256_srli_epi16(mins_and_scales_67, 4), m4b));
4566
4567                    const __m512i scales_0 = _mm512_cvtepu8_epi16(_mm256_shuffle_epi8(scales_01, scalesmask1));
4568                    const __m512i scales_1 = _mm512_cvtepu8_epi16(_mm256_shuffle_epi8(scales_01, scalesmask2));
4569                    const __m512i scales_2 = _mm512_cvtepu8_epi16(_mm256_shuffle_epi8(scales_23, scalesmask1));
4570                    const __m512i scales_3 = _mm512_cvtepu8_epi16(_mm256_shuffle_epi8(scales_23, scalesmask2));
4571                    const __m512i scales_4 = _mm512_cvtepu8_epi16(_mm256_shuffle_epi8(scales_45, scalesmask1));
4572                    const __m512i scales_5 = _mm512_cvtepu8_epi16(_mm256_shuffle_epi8(scales_45, scalesmask2));
4573                    const __m512i scales_6 = _mm512_cvtepu8_epi16(_mm256_shuffle_epi8(scales_67, scalesmask1));
4574                    const __m512i scales_7 = _mm512_cvtepu8_epi16(_mm256_shuffle_epi8(scales_67, scalesmask2));
4575
4576                    const __m512i scale_014589CD_0 = _mm512_shuffle_epi32(scales_0, (_MM_PERM_ENUM)68);
4577                    const __m512i scale_2367ABEF_0 = _mm512_shuffle_epi32(scales_0, (_MM_PERM_ENUM)238);
4578
4579                    const __m512i scale_014589CD_1 = _mm512_shuffle_epi32(scales_1, (_MM_PERM_ENUM)68);
4580                    const __m512i scale_2367ABEF_1 = _mm512_shuffle_epi32(scales_1, (_MM_PERM_ENUM)238);
4581
4582                    const __m512i scale_014589CD_2 = _mm512_shuffle_epi32(scales_2, (_MM_PERM_ENUM)68);
4583                    const __m512i scale_2367ABEF_2 = _mm512_shuffle_epi32(scales_2, (_MM_PERM_ENUM)238);
4584
4585                    const __m512i scale_014589CD_3 = _mm512_shuffle_epi32(scales_3, (_MM_PERM_ENUM)68);
4586                    const __m512i scale_2367ABEF_3 = _mm512_shuffle_epi32(scales_3, (_MM_PERM_ENUM)238);
4587
4588                    const __m512i scale_014589CD_4 = _mm512_shuffle_epi32(scales_4, (_MM_PERM_ENUM)68);
4589                    const __m512i scale_2367ABEF_4 = _mm512_shuffle_epi32(scales_4, (_MM_PERM_ENUM)238);
4590
4591                    const __m512i scale_014589CD_5 = _mm512_shuffle_epi32(scales_5, (_MM_PERM_ENUM)68);
4592                    const __m512i scale_2367ABEF_5 = _mm512_shuffle_epi32(scales_5, (_MM_PERM_ENUM)238);
4593
4594                    const __m512i scale_014589CD_6 = _mm512_shuffle_epi32(scales_6, (_MM_PERM_ENUM)68);
4595                    const __m512i scale_2367ABEF_6 = _mm512_shuffle_epi32(scales_6, (_MM_PERM_ENUM)238);
4596
4597                    const __m512i scale_014589CD_7 = _mm512_shuffle_epi32(scales_7, (_MM_PERM_ENUM)68);
4598                    const __m512i scale_2367ABEF_7 = _mm512_shuffle_epi32(scales_7, (_MM_PERM_ENUM)238);
4599
4600                    // Load the four block_q8_k quantized values interleaved with each other in chunks of eight bytes - A0,A1,A2,A3
4601                    // Loaded as set of 128 bit vectors and repeated into a 256 bit vector
4602                    __m256i lhs_mat_ymm_0123_00 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 512 * sb)));
4603                    __m256i lhs_mat_ymm_01_00 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_00, lhs_mat_ymm_0123_00, 0);
4604                    __m256i lhs_mat_ymm_23_00 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_00, lhs_mat_ymm_0123_00, 17);
4605                    __m256i lhs_mat_ymm_0123_01 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 32 + 512 * sb)));
4606                    __m256i lhs_mat_ymm_01_01 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_01, lhs_mat_ymm_0123_01, 0);
4607                    __m256i lhs_mat_ymm_23_01 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_01, lhs_mat_ymm_0123_01, 17);
4608                    __m256i lhs_mat_ymm_0123_10 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 64 + 512 * sb)));
4609                    __m256i lhs_mat_ymm_01_10 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_10, lhs_mat_ymm_0123_10, 0);
4610                    __m256i lhs_mat_ymm_23_10 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_10, lhs_mat_ymm_0123_10, 17);
4611                    __m256i lhs_mat_ymm_0123_11 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 96 + 512 * sb)));
4612                    __m256i lhs_mat_ymm_01_11 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_11, lhs_mat_ymm_0123_11, 0);
4613                    __m256i lhs_mat_ymm_23_11 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_11, lhs_mat_ymm_0123_11, 17);
4614                    __m256i lhs_mat_ymm_0123_20 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 128 + 512 * sb)));
4615                    __m256i lhs_mat_ymm_01_20 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_20, lhs_mat_ymm_0123_20, 0);
4616                    __m256i lhs_mat_ymm_23_20 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_20, lhs_mat_ymm_0123_20, 17);
4617                    __m256i lhs_mat_ymm_0123_21 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 160 + 512 * sb)));
4618                    __m256i lhs_mat_ymm_01_21 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_21, lhs_mat_ymm_0123_21, 0);
4619                    __m256i lhs_mat_ymm_23_21 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_21, lhs_mat_ymm_0123_21, 17);
4620                    __m256i lhs_mat_ymm_0123_30 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 192 + 512 * sb)));
4621                    __m256i lhs_mat_ymm_01_30 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_30, lhs_mat_ymm_0123_30, 0);
4622                    __m256i lhs_mat_ymm_23_30 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_30, lhs_mat_ymm_0123_30, 17);
4623                    __m256i lhs_mat_ymm_0123_31 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 224 + 512 * sb)));
4624                    __m256i lhs_mat_ymm_01_31 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_31, lhs_mat_ymm_0123_31, 0);
4625                    __m256i lhs_mat_ymm_23_31 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_31, lhs_mat_ymm_0123_31, 17);
4626
4627                    __m256i lhs_mat_ymm_0123_40 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 256 + 512 * sb)));
4628                    __m256i lhs_mat_ymm_01_40 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_40, lhs_mat_ymm_0123_40, 0);
4629                    __m256i lhs_mat_ymm_23_40 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_40, lhs_mat_ymm_0123_40, 17);
4630                    __m256i lhs_mat_ymm_0123_41 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 288 + 512 * sb)));
4631                    __m256i lhs_mat_ymm_01_41 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_41, lhs_mat_ymm_0123_41, 0);
4632                    __m256i lhs_mat_ymm_23_41 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_41, lhs_mat_ymm_0123_41, 17);
4633                    __m256i lhs_mat_ymm_0123_50 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 320 + 512 * sb)));
4634                    __m256i lhs_mat_ymm_01_50 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_50, lhs_mat_ymm_0123_50, 0);
4635                    __m256i lhs_mat_ymm_23_50 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_50, lhs_mat_ymm_0123_50, 17);
4636                    __m256i lhs_mat_ymm_0123_51 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 352 + 512 * sb)));
4637                    __m256i lhs_mat_ymm_01_51 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_51, lhs_mat_ymm_0123_51, 0);
4638                    __m256i lhs_mat_ymm_23_51 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_51, lhs_mat_ymm_0123_51, 17);
4639                    __m256i lhs_mat_ymm_0123_60 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 384 + 512 * sb)));
4640                    __m256i lhs_mat_ymm_01_60 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_60, lhs_mat_ymm_0123_60, 0);
4641                    __m256i lhs_mat_ymm_23_60 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_60, lhs_mat_ymm_0123_60, 17);
4642                    __m256i lhs_mat_ymm_0123_61 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 416 + 512 * sb)));
4643                    __m256i lhs_mat_ymm_01_61 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_61, lhs_mat_ymm_0123_61, 0);
4644                    __m256i lhs_mat_ymm_23_61 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_61, lhs_mat_ymm_0123_61, 17);
4645                    __m256i lhs_mat_ymm_0123_70 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 448 + 512 * sb)));
4646                    __m256i lhs_mat_ymm_01_70 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_70, lhs_mat_ymm_0123_70, 0);
4647                    __m256i lhs_mat_ymm_23_70 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_70, lhs_mat_ymm_0123_70, 17);
4648                    __m256i lhs_mat_ymm_0123_71 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 480 + 512 * sb)));
4649                    __m256i lhs_mat_ymm_01_71 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_71, lhs_mat_ymm_0123_71, 0);
4650                    __m256i lhs_mat_ymm_23_71 = _mm256_permute2f128_si256(lhs_mat_ymm_0123_71, lhs_mat_ymm_0123_71, 17);
4651
4652                    __m512i lhs_mat_01_00 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_00), lhs_mat_ymm_01_00, 1);
4653                    __m512i lhs_mat_23_00 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_00), lhs_mat_ymm_23_00, 1);
4654                    __m512i lhs_mat_01_01 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_01), lhs_mat_ymm_01_01, 1);
4655                    __m512i lhs_mat_23_01 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_01), lhs_mat_ymm_23_01, 1);
4656
4657                    __m512i lhs_mat_01_10 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_10), lhs_mat_ymm_01_10, 1);
4658                    __m512i lhs_mat_23_10 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_10), lhs_mat_ymm_23_10, 1);
4659                    __m512i lhs_mat_01_11 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_11), lhs_mat_ymm_01_11, 1);
4660                    __m512i lhs_mat_23_11 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_11), lhs_mat_ymm_23_11, 1);
4661
4662                    __m512i lhs_mat_01_20 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_20), lhs_mat_ymm_01_20, 1);
4663                    __m512i lhs_mat_23_20 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_20), lhs_mat_ymm_23_20, 1);
4664                    __m512i lhs_mat_01_21 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_21), lhs_mat_ymm_01_21, 1);
4665                    __m512i lhs_mat_23_21 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_21), lhs_mat_ymm_23_21, 1);
4666
4667                    __m512i lhs_mat_01_30 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_30), lhs_mat_ymm_01_30, 1);
4668                    __m512i lhs_mat_23_30 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_30), lhs_mat_ymm_23_30, 1);
4669                    __m512i lhs_mat_01_31 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_31), lhs_mat_ymm_01_31, 1);
4670                    __m512i lhs_mat_23_31 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_31), lhs_mat_ymm_23_31, 1);
4671
4672                    __m512i lhs_mat_01_40 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_40), lhs_mat_ymm_01_40, 1);
4673                    __m512i lhs_mat_23_40 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_40), lhs_mat_ymm_23_40, 1);
4674                    __m512i lhs_mat_01_41 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_41), lhs_mat_ymm_01_41, 1);
4675                    __m512i lhs_mat_23_41 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_41), lhs_mat_ymm_23_41, 1);
4676
4677                    __m512i lhs_mat_01_50 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_50), lhs_mat_ymm_01_50, 1);
4678                    __m512i lhs_mat_23_50 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_50), lhs_mat_ymm_23_50, 1);
4679                    __m512i lhs_mat_01_51 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_51), lhs_mat_ymm_01_51, 1);
4680                    __m512i lhs_mat_23_51 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_51), lhs_mat_ymm_23_51, 1);
4681
4682                    __m512i lhs_mat_01_60 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_60), lhs_mat_ymm_01_60, 1);
4683                    __m512i lhs_mat_23_60 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_60), lhs_mat_ymm_23_60, 1);
4684                    __m512i lhs_mat_01_61 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_61), lhs_mat_ymm_01_61, 1);
4685                    __m512i lhs_mat_23_61 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_61), lhs_mat_ymm_23_61, 1);
4686
4687                    __m512i lhs_mat_01_70 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_70), lhs_mat_ymm_01_70, 1);
4688                    __m512i lhs_mat_23_70 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_70), lhs_mat_ymm_23_70, 1);
4689                    __m512i lhs_mat_01_71 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_01_71), lhs_mat_ymm_01_71, 1);
4690                    __m512i lhs_mat_23_71 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_mat_ymm_23_71), lhs_mat_ymm_23_71, 1);
4691
4692                    // Bsums are loaded for the different Q8_K blocks
4693                    __m128i lhs_raw_bsums_01_0123 = _mm_loadu_si128((const __m128i *)((a_ptr[b].bsums + 32 * sb)));
4694                    __m128i lhs_raw_bsums_23_0123 = _mm_loadu_si128((const __m128i *)(a_ptr[b].bsums + 8 + 32 * sb));
4695                    __m128i lhs_raw_bsums_01_4567 = _mm_loadu_si128((const __m128i *)((a_ptr[b].bsums + 16 + 32 * sb)));
4696                    __m128i lhs_raw_bsums_23_4567 = _mm_loadu_si128((const __m128i *)(a_ptr[b].bsums + 24 + 32 * sb));
4697
4698                    __m256i lhs_bsums_ymm_01_0123 = _mm256_inserti128_si256(_mm256_castsi128_si256(lhs_raw_bsums_01_0123), lhs_raw_bsums_01_0123, 1);
4699                    __m512i lhs_bsums_01_0123 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_bsums_ymm_01_0123), lhs_bsums_ymm_01_0123, 1);
4700                    __m256i lhs_bsums_ymm_23_0123 = _mm256_inserti128_si256(_mm256_castsi128_si256(lhs_raw_bsums_23_0123), lhs_raw_bsums_23_0123, 1);
4701                    __m512i lhs_bsums_23_0123 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_bsums_ymm_23_0123), lhs_bsums_ymm_23_0123, 1);
4702                    __m256i lhs_bsums_ymm_01_4567 = _mm256_inserti128_si256(_mm256_castsi128_si256(lhs_raw_bsums_01_4567), lhs_raw_bsums_01_4567, 1);
4703                    __m512i lhs_bsums_01_4567 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_bsums_ymm_01_4567), lhs_bsums_ymm_01_4567, 1);
4704                    __m256i lhs_bsums_ymm_23_4567 = _mm256_inserti128_si256(_mm256_castsi128_si256(lhs_raw_bsums_23_4567), lhs_raw_bsums_23_4567, 1);
4705                    __m512i lhs_bsums_23_4567 = _mm512_inserti32x8(_mm512_castsi256_si512(lhs_bsums_ymm_23_4567), lhs_bsums_ymm_23_4567, 1);
4706
4707                    // Shuffle pattern one - left side input
4708                    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)
4709                    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)
4710
4711                    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)
4712                    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)
4713
4714                    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)
4715                    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)
4716
4717                    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)
4718                    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)
4719
4720                    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)
4721                    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)
4722
4723                    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)
4724                    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)
4725
4726                    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)
4727                    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)
4728
4729                    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)
4730                    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)
4731
4732                    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)
4733                    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)
4734
4735                    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)
4736                    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)
4737
4738                    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)
4739                    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)
4740
4741                    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)
4742                    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)
4743
4744                    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)
4745                    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)
4746
4747                    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)
4748                    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)
4749
4750                    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)
4751                    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)
4752
4753                    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)
4754                    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)
4755
4756                    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)
4757                    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)
4758
4759                    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)
4760                    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)
4761
4762                    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)
4763                    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)
4764
4765                    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)
4766                    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)
4767
4768                    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)
4769                    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)
4770
4771                    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)
4772                    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)
4773
4774                    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)
4775                    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)
4776
4777                    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)
4778                    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)
4779
4780                    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)
4781                    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)
4782
4783                    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)
4784                    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)
4785
4786                    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)
4787                    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)
4788
4789                    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)
4790                    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)
4791
4792                    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)
4793                    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)
4794
4795                    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)
4796                    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)
4797
4798                    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)
4799                    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)
4800
4801                    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)
4802                    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)
4803
4804                    // 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
4805                    __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));
4806                    __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));
4807
4808                    __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));
4809                    __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));
4810
4811                    __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));
4812                    __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));
4813
4814                    __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));
4815                    __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));
4816
4817                    __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));
4818                    __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));
4819
4820                    __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));
4821                    __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));
4822
4823                    __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));
4824                    __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));
4825
4826                    __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));
4827                    __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));
4828
4829                    __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));
4830                    __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));
4831
4832                    __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));
4833                    __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));
4834
4835                    __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));
4836                    __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));
4837
4838                    __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));
4839                    __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));
4840
4841                    __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));
4842                    __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));
4843
4844                    __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));
4845                    __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));
4846
4847                    __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));
4848                    __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));
4849
4850                    __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));
4851                    __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));
4852
4853
4854                    __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));
4855                    __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));
4856
4857                    __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));
4858                    __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));
4859
4860                    __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));
4861                    __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));
4862
4863                    __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));
4864                    __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));
4865
4866                    __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));
4867                    __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));
4868
4869                    __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));
4870                    __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));
4871
4872                    __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));
4873                    __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));
4874
4875                    __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));
4876                    __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));
4877
4878                    __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));
4879                    __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));
4880
4881                    __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));
4882                    __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));
4883
4884                    __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));
4885                    __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));
4886
4887                    __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));
4888                    __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));
4889
4890                    __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));
4891                    __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));
4892
4893                    __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));
4894                    __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));
4895
4896                    __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));
4897                    __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));
4898
4899                    __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));
4900                    __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));
4901
4902                    // Combine results from both shuffle patterns for each output block
4903                    __m512i iacc_mat_00_0 = _mm512_add_epi16(iacc_mat_00_0_sp1, iacc_mat_00_0_sp2);
4904                    __m512i iacc_mat_01_0 = _mm512_add_epi16(iacc_mat_01_0_sp1, iacc_mat_01_0_sp2);
4905                    __m512i iacc_mat_10_0 = _mm512_add_epi16(iacc_mat_10_0_sp1, iacc_mat_10_0_sp2);
4906                    __m512i iacc_mat_11_0 = _mm512_add_epi16(iacc_mat_11_0_sp1, iacc_mat_11_0_sp2);
4907
4908                    __m512i iacc_mat_00_1 = _mm512_add_epi16(iacc_mat_00_1_sp1, iacc_mat_00_1_sp2);
4909                    __m512i iacc_mat_01_1 = _mm512_add_epi16(iacc_mat_01_1_sp1, iacc_mat_01_1_sp2);
4910                    __m512i iacc_mat_10_1 = _mm512_add_epi16(iacc_mat_10_1_sp1, iacc_mat_10_1_sp2);
4911                    __m512i iacc_mat_11_1 = _mm512_add_epi16(iacc_mat_11_1_sp1, iacc_mat_11_1_sp2);
4912
4913                    __m512i iacc_mat_00_2 = _mm512_add_epi16(iacc_mat_00_2_sp1, iacc_mat_00_2_sp2);
4914                    __m512i iacc_mat_01_2 = _mm512_add_epi16(iacc_mat_01_2_sp1, iacc_mat_01_2_sp2);
4915                    __m512i iacc_mat_10_2 = _mm512_add_epi16(iacc_mat_10_2_sp1, iacc_mat_10_2_sp2);
4916                    __m512i iacc_mat_11_2 = _mm512_add_epi16(iacc_mat_11_2_sp1, iacc_mat_11_2_sp2);
4917
4918                    __m512i iacc_mat_00_3 = _mm512_add_epi16(iacc_mat_00_3_sp1, iacc_mat_00_3_sp2);
4919                    __m512i iacc_mat_01_3 = _mm512_add_epi16(iacc_mat_01_3_sp1, iacc_mat_01_3_sp2);
4920                    __m512i iacc_mat_10_3 = _mm512_add_epi16(iacc_mat_10_3_sp1, iacc_mat_10_3_sp2);
4921                    __m512i iacc_mat_11_3 = _mm512_add_epi16(iacc_mat_11_3_sp1, iacc_mat_11_3_sp2);
4922
4923                    __m512i iacc_mat_00_4 = _mm512_add_epi16(iacc_mat_00_4_sp1, iacc_mat_00_4_sp2);
4924                    __m512i iacc_mat_01_4 = _mm512_add_epi16(iacc_mat_01_4_sp1, iacc_mat_01_4_sp2);
4925                    __m512i iacc_mat_10_4 = _mm512_add_epi16(iacc_mat_10_4_sp1, iacc_mat_10_4_sp2);
4926                    __m512i iacc_mat_11_4 = _mm512_add_epi16(iacc_mat_11_4_sp1, iacc_mat_11_4_sp2);
4927
4928                    __m512i iacc_mat_00_5 = _mm512_add_epi16(iacc_mat_00_5_sp1, iacc_mat_00_5_sp2);
4929                    __m512i iacc_mat_01_5 = _mm512_add_epi16(iacc_mat_01_5_sp1, iacc_mat_01_5_sp2);
4930                    __m512i iacc_mat_10_5 = _mm512_add_epi16(iacc_mat_10_5_sp1, iacc_mat_10_5_sp2);
4931                    __m512i iacc_mat_11_5 = _mm512_add_epi16(iacc_mat_11_5_sp1, iacc_mat_11_5_sp2);
4932
4933                    __m512i iacc_mat_00_6 = _mm512_add_epi16(iacc_mat_00_6_sp1, iacc_mat_00_6_sp2);
4934                    __m512i iacc_mat_01_6 = _mm512_add_epi16(iacc_mat_01_6_sp1, iacc_mat_01_6_sp2);
4935                    __m512i iacc_mat_10_6 = _mm512_add_epi16(iacc_mat_10_6_sp1, iacc_mat_10_6_sp2);
4936                    __m512i iacc_mat_11_6 = _mm512_add_epi16(iacc_mat_11_6_sp1, iacc_mat_11_6_sp2);
4937
4938                    __m512i iacc_mat_00_7 = _mm512_add_epi16(iacc_mat_00_7_sp1, iacc_mat_00_7_sp2);
4939                    __m512i iacc_mat_01_7 = _mm512_add_epi16(iacc_mat_01_7_sp1, iacc_mat_01_7_sp2);
4940                    __m512i iacc_mat_10_7 = _mm512_add_epi16(iacc_mat_10_7_sp1, iacc_mat_10_7_sp2);
4941                    __m512i iacc_mat_11_7 = _mm512_add_epi16(iacc_mat_11_7_sp1, iacc_mat_11_7_sp2);
4942
4943                    // Output of both shuffle patterns are added in order to sum dot product outputs of all 32 values in block
4944                    iacc_mat_00_0 = _mm512_madd_epi16(iacc_mat_00_0, scale_014589CD_0);
4945                    iacc_mat_01_0 = _mm512_madd_epi16(iacc_mat_01_0, scale_2367ABEF_0);
4946                    iacc_mat_10_0 = _mm512_madd_epi16(iacc_mat_10_0, scale_014589CD_0);
4947                    iacc_mat_11_0 = _mm512_madd_epi16(iacc_mat_11_0, scale_2367ABEF_0);
4948
4949                    iacc_mat_00_1 = _mm512_madd_epi16(iacc_mat_00_1, scale_014589CD_1);
4950                    iacc_mat_01_1 = _mm512_madd_epi16(iacc_mat_01_1, scale_2367ABEF_1);
4951                    iacc_mat_10_1 = _mm512_madd_epi16(iacc_mat_10_1, scale_014589CD_1);
4952                    iacc_mat_11_1 = _mm512_madd_epi16(iacc_mat_11_1, scale_2367ABEF_1);
4953
4954                    iacc_mat_00_2 = _mm512_madd_epi16(iacc_mat_00_2, scale_014589CD_2);
4955                    iacc_mat_01_2 = _mm512_madd_epi16(iacc_mat_01_2, scale_2367ABEF_2);
4956                    iacc_mat_10_2 = _mm512_madd_epi16(iacc_mat_10_2, scale_014589CD_2);
4957                    iacc_mat_11_2 = _mm512_madd_epi16(iacc_mat_11_2, scale_2367ABEF_2);
4958
4959                    iacc_mat_00_3 = _mm512_madd_epi16(iacc_mat_00_3, scale_014589CD_3);
4960                    iacc_mat_01_3 = _mm512_madd_epi16(iacc_mat_01_3, scale_2367ABEF_3);
4961                    iacc_mat_10_3 = _mm512_madd_epi16(iacc_mat_10_3, scale_014589CD_3);
4962                    iacc_mat_11_3 = _mm512_madd_epi16(iacc_mat_11_3, scale_2367ABEF_3);
4963
4964                    iacc_mat_00_4 = _mm512_madd_epi16(iacc_mat_00_4, scale_014589CD_4);
4965                    iacc_mat_01_4 = _mm512_madd_epi16(iacc_mat_01_4, scale_2367ABEF_4);
4966                    iacc_mat_10_4 = _mm512_madd_epi16(iacc_mat_10_4, scale_014589CD_4);
4967                    iacc_mat_11_4 = _mm512_madd_epi16(iacc_mat_11_4, scale_2367ABEF_4);
4968
4969                    iacc_mat_00_5 = _mm512_madd_epi16(iacc_mat_00_5, scale_014589CD_5);
4970                    iacc_mat_01_5 = _mm512_madd_epi16(iacc_mat_01_5, scale_2367ABEF_5);
4971                    iacc_mat_10_5 = _mm512_madd_epi16(iacc_mat_10_5, scale_014589CD_5);
4972                    iacc_mat_11_5 = _mm512_madd_epi16(iacc_mat_11_5, scale_2367ABEF_5);
4973
4974                    iacc_mat_00_6 = _mm512_madd_epi16(iacc_mat_00_6, scale_014589CD_6);
4975                    iacc_mat_01_6 = _mm512_madd_epi16(iacc_mat_01_6, scale_2367ABEF_6);
4976                    iacc_mat_10_6 = _mm512_madd_epi16(iacc_mat_10_6, scale_014589CD_6);
4977                    iacc_mat_11_6 = _mm512_madd_epi16(iacc_mat_11_6, scale_2367ABEF_6);
4978
4979                    iacc_mat_00_7 = _mm512_madd_epi16(iacc_mat_00_7, scale_014589CD_7);
4980                    iacc_mat_01_7 = _mm512_madd_epi16(iacc_mat_01_7, scale_2367ABEF_7);
4981                    iacc_mat_10_7 = _mm512_madd_epi16(iacc_mat_10_7, scale_014589CD_7);
4982                    iacc_mat_11_7 = _mm512_madd_epi16(iacc_mat_11_7, scale_2367ABEF_7);
4983
4984                    __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)));
4985                    __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)));
4986                    __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)));
4987                    __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)));
4988
4989                    // Straighten out to make 4 row vectors
4990                    __m512i iacc_row_0 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_00, _mm512_shuffle_epi32(iacc_mat_01, (_MM_PERM_ENUM)78));
4991                    __m512i iacc_row_1 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_00, (_MM_PERM_ENUM)78), iacc_mat_01);
4992                    __m512i iacc_row_2 = _mm512_mask_blend_epi32(0xCCCC, iacc_mat_10, _mm512_shuffle_epi32(iacc_mat_11, (_MM_PERM_ENUM)78));
4993                    __m512i iacc_row_3 = _mm512_mask_blend_epi32(0xCCCC, _mm512_shuffle_epi32(iacc_mat_10, (_MM_PERM_ENUM)78), iacc_mat_11);
4994
4995                    // Load the scale(d) values for all the 4 Q8_k blocks and repeat it across lanes
4996                    const __m128 row_scale_f32_sse = _mm_load_ps(a_ptr[b].d);
4997                    const __m256 row_scale_f32_ymm = _mm256_set_m128(row_scale_f32_sse, row_scale_f32_sse);
4998                    const __m512 row_scale_f32 = _mm512_insertf32x8(_mm512_castps256_ps512(row_scale_f32_ymm), row_scale_f32_ymm, 1);
4999
5000                    // Multiply with appropiate scales and accumulate (for both d and dmin) below
5001                    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]);
5002                    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]);
5003                    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]);
5004                    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]);
5005
5006                    // Take two bsums from two Q8_Ks at a time and multiply with corresponding mins values from each Q2_K
5007                    __m512i iacc_row_min_0_01 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_01_0123, (_MM_PERM_ENUM)0), mins_01);
5008                    __m512i iacc_row_min_1_01 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_01_0123, (_MM_PERM_ENUM)170), mins_01);
5009                    __m512i iacc_row_min_2_01 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_23_0123, (_MM_PERM_ENUM)0), mins_01);
5010                    __m512i iacc_row_min_3_01 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_23_0123, (_MM_PERM_ENUM)170), mins_01);
5011
5012                    __m512i iacc_row_min_0_23 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_01_0123, (_MM_PERM_ENUM)85), mins_23);
5013                    __m512i iacc_row_min_1_23 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_01_0123, (_MM_PERM_ENUM)255), mins_23);
5014                    __m512i iacc_row_min_2_23 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_23_0123, (_MM_PERM_ENUM)85), mins_23);
5015                    __m512i iacc_row_min_3_23 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_23_0123, (_MM_PERM_ENUM)255), mins_23);
5016
5017                    __m512i iacc_row_min_0_45 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_01_4567, (_MM_PERM_ENUM)0), mins_45);
5018                    __m512i iacc_row_min_1_45 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_01_4567, (_MM_PERM_ENUM)170), mins_45);
5019                    __m512i iacc_row_min_2_45 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_23_4567, (_MM_PERM_ENUM)0), mins_45);
5020                    __m512i iacc_row_min_3_45 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_23_4567, (_MM_PERM_ENUM)170), mins_45);
5021
5022                    __m512i iacc_row_min_0_67 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_01_4567, (_MM_PERM_ENUM)85), mins_67);
5023                    __m512i iacc_row_min_1_67 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_01_4567, (_MM_PERM_ENUM)255), mins_67);
5024                    __m512i iacc_row_min_2_67 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_23_4567, (_MM_PERM_ENUM)85), mins_67);
5025                    __m512i iacc_row_min_3_67 = _mm512_madd_epi16(_mm512_shuffle_epi32(lhs_bsums_23_4567, (_MM_PERM_ENUM)255), mins_67);
5026
5027                    __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));
5028                    __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));
5029                    __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));
5030                    __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));
5031
5032                    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]);
5033                    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]);
5034                    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]);
5035                    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]);
5036                }
5037            }
5038            // Store accumulated values
5039            for (int i = 0; i < 4; i++) {
5040                _mm512_storeu_ps((float * )(s + ((y * 4 + i) * bs + x * 8)), _mm512_sub_ps(acc_rows[i], acc_min_rows[i]));
5041            }
5042        }
5043    }
5044
5045    if (anc != nc) {
5046        xstart = anc/8;
5047        y = 0;
5048    }
5049
5050#endif // __AVX512BW__ && __AVX512DQ__
5051
5052    // Take group of four block_q8_Kx4 structures at each pass of the loop and perform dot product operation
5053    for (; y < anr / 4; y += 4) {
5054
5055        const block_q8_Kx4 * a_ptrs[4];
5056
5057        a_ptrs[0] = a_ptr_start + (y * nb);
5058        for (int i = 0; i < 3; ++i) {
5059            a_ptrs[i + 1] = a_ptrs[i] + nb;
5060        }
5061
5062        // Take group of eight block_q2_kx8 structures at each pass of the loop and perform dot product operation
5063        for (int64_t x = xstart; x < nc / 8; x++) {
5064
5065            const block_q2_Kx8 * b_ptr = b_ptr_start + (x * b_nb);
5066
5067            // Master FP accumulators
5068            __m256 acc_rows[16];
5069            for (int i = 0; i < 16; i++) {
5070                acc_rows[i] = _mm256_setzero_ps();
5071            }
5072
5073            __m256 acc_min_rows[16];
5074            for (int i = 0; i < 16; i++) {
5075                acc_min_rows[i] = _mm256_setzero_ps();
5076            }
5077
5078            // For super block
5079            for (int64_t b = 0; b < nb; b++) {
5080                // Delta values - Load the eight scale values of block_q2_kx8
5081                const __m256 col_scale_f32 = GGML_F32Cx8_LOAD(b_ptr[b].d);
5082
5083                // dmin values - Load the eight dmin values of block_q2_kx8
5084                const __m256 col_dmin_f32 = GGML_F32Cx8_LOAD(b_ptr[b].dmin);
5085
5086                // Loop to iterate over the sixteen sub blocks of a super block - eight sub blocks are processed per iteration
5087                for (int sb = 0; sb < QK_K / 128; sb++) {
5088
5089                    // Load the eight block_q2_K for eight sub blocks quantized values interleaved with each other in chunks of eight bytes - B0,B1 ....B6,B7
5090                    const __m256i rhs_raw_mat_0123_0 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + sb * 256));
5091                    const __m256i rhs_raw_mat_4567_0 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 32 + sb * 256));
5092                    const __m256i rhs_raw_mat_0123_1 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 64 + sb * 256));
5093                    const __m256i rhs_raw_mat_4567_1 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 96 + sb * 256));
5094                    const __m256i rhs_raw_mat_0123_2 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 128 + sb * 256));
5095                    const __m256i rhs_raw_mat_4567_2 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 160 + sb * 256));
5096                    const __m256i rhs_raw_mat_0123_3 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 192 + sb * 256));
5097                    const __m256i rhs_raw_mat_4567_3 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 224 + sb * 256));
5098
5099                    // Save the values in the following vectors in the formats B0B1B4B5, B2B3B6B7 for further processing and storing of values
5100                    //superblock    sub block   which part of sub block
5101                    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);
5102                    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);
5103
5104                    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);
5105                    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);
5106
5107                    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);
5108                    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);
5109
5110                    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);
5111                    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);
5112
5113                    // 2-bit -> 8-bit
5114                    // First sub block of the eight sub blocks processed in the iteration
5115                    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)
5116                    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)
5117
5118                    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)
5119                    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)
5120
5121                    // Second sub block of the eight sub blocks processed in the iteration
5122                    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)
5123                    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)
5124
5125                    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)
5126                    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)
5127
5128                    // Third sub block of the eight sub blocks processed in the iteration
5129                    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)
5130                    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)
5131
5132                    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)
5133                    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)
5134
5135                    // Fourth sub block of the eight sub blocks processed in the iteration
5136                    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)
5137                    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)
5138
5139                    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)
5140                    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)
5141
5142                    // Fifth sub block of the eight sub blocks processed in the iteration
5143                    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)
5144                    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)
5145
5146                    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)
5147                    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)
5148
5149                    // Sixth sub block of the eight sub blocks processed in the iteration
5150                    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)
5151                    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)
5152
5153                    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)
5154                    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)
5155
5156                    // Seventh sub block of the eight sub blocks processed in the iteration
5157                    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)
5158                    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)
5159
5160                    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)
5161                    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)
5162
5163                    // Eighth sub block of the eight sub blocks processed in the iteration
5164                    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)
5165                    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)
5166
5167                    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)
5168                    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)
5169
5170                    // Shuffle pattern one - right side input
5171                    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)
5172                    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)
5173
5174                    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)
5175                    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)
5176
5177                    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)
5178                    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)
5179
5180                    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)
5181                    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)
5182
5183                    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)
5184                    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)
5185
5186                    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)
5187                    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)
5188
5189                    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)
5190                    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)
5191
5192                    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
5193                    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)
5194
5195                    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)
5196                    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)
5197
5198                    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)
5199                    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)
5200
5201                    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)
5202                    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)
5203
5204                    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)
5205                    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)
5206
5207                    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)
5208                    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)
5209
5210                    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)
5211                    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)
5212
5213                    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)
5214                    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)
5215
5216                    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)
5217                    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)
5218
5219
5220                    // Shuffle pattern two - right side input
5221                    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)
5222                    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)
5223
5224                    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)
5225                    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)
5226
5227                    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)
5228                    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)
5229
5230                    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)
5231                    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)
5232
5233                    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)
5234                    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)
5235
5236                    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)
5237                    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)
5238
5239                    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)
5240                    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)
5241
5242                    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)
5243                    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)
5244
5245                    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)
5246                    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)
5247
5248                    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)
5249                    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)
5250
5251                    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)
5252                    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)
5253
5254                    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)
5255                    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)
5256
5257                    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)
5258                    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)
5259
5260                    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)
5261                    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)
5262
5263                    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)
5264                    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)
5265
5266                    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)
5267                    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)
5268
5269                    //Scales and Mins of corresponding sub blocks from different Q2_K structures are stored together
5270                    //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
5271
5272                    // Combine mins and scales for sub-blocks: 0-1, 2-3, 4-5, 6-7 in the sb loop
5273                    const __m128i mins_and_scales_01 = _mm_loadu_si128((const __m128i *)(b_ptr[b].scales + sb * 64));
5274                    const __m128i mins_and_scales_23 = _mm_loadu_si128((const __m128i *)(b_ptr[b].scales + 16 + sb * 64));
5275                    const __m128i mins_and_scales_45 = _mm_loadu_si128((const __m128i *)(b_ptr[b].scales + 32 + sb * 64));
5276                    const __m128i mins_and_scales_67 = _mm_loadu_si128((const __m128i *)(b_ptr[b].scales + 48 + sb * 64));
5277
5278                    // Extract scales which is lower half from mins_and_scales
5279                    const __m128i scales_01 = _mm_and_si128(mins_and_scales_01, m4b_sse);
5280                    const __m128i scales_23 = _mm_and_si128(mins_and_scales_23, m4b_sse);
5281                    const __m128i scales_45 = _mm_and_si128(mins_and_scales_45, m4b_sse);
5282                    const __m128i scales_67 = _mm_and_si128(mins_and_scales_67, m4b_sse);
5283
5284                    // Extract mins which is upper half from mins_and_scales
5285                    const __m256i mins_01 = _mm256_cvtepu8_epi16(_mm_and_si128(_mm_srli_epi16(mins_and_scales_01, 4), m4b_sse));
5286                    const __m256i mins_23 = _mm256_cvtepu8_epi16(_mm_and_si128(_mm_srli_epi16(mins_and_scales_23, 4), m4b_sse));
5287                    const __m256i mins_45 = _mm256_cvtepu8_epi16(_mm_and_si128(_mm_srli_epi16(mins_and_scales_45, 4), m4b_sse));
5288                    const __m256i mins_67 = _mm256_cvtepu8_epi16(_mm_and_si128(_mm_srli_epi16(mins_and_scales_67, 4), m4b_sse));
5289
5290                    const __m256i scales_0 = _mm256_cvtepu8_epi16(_mm_shuffle_epi8(scales_01, scalesmask1_sse));
5291                    const __m256i scales_1 = _mm256_cvtepu8_epi16(_mm_shuffle_epi8(scales_01, scalesmask2_sse));
5292
5293                    const __m256i scales_2 = _mm256_cvtepu8_epi16(_mm_shuffle_epi8(scales_23, scalesmask1_sse));
5294                    const __m256i scales_3 = _mm256_cvtepu8_epi16(_mm_shuffle_epi8(scales_23, scalesmask2_sse));
5295
5296                    const __m256i scales_4 = _mm256_cvtepu8_epi16(_mm_shuffle_epi8(scales_45, scalesmask1_sse));
5297                    const __m256i scales_5 = _mm256_cvtepu8_epi16(_mm_shuffle_epi8(scales_45, scalesmask2_sse));
5298
5299                    const __m256i scales_6 = _mm256_cvtepu8_epi16(_mm_shuffle_epi8(scales_67, scalesmask1_sse));
5300                    const __m256i scales_7 = _mm256_cvtepu8_epi16(_mm_shuffle_epi8(scales_67, scalesmask2_sse));
5301
5302                    const __m256i scale_0145_0 = _mm256_shuffle_epi32(scales_0, 68);
5303                    const __m256i scale_2367_0 = _mm256_shuffle_epi32(scales_0, 238);
5304
5305                    const __m256i scale_0145_1 = _mm256_shuffle_epi32(scales_1, 68);
5306                    const __m256i scale_2367_1 = _mm256_shuffle_epi32(scales_1, 238);
5307
5308                    const __m256i scale_0145_2 = _mm256_shuffle_epi32(scales_2, 68);
5309                    const __m256i scale_2367_2 = _mm256_shuffle_epi32(scales_2, 238);
5310
5311                    const __m256i scale_0145_3 = _mm256_shuffle_epi32(scales_3, 68);
5312                    const __m256i scale_2367_3 = _mm256_shuffle_epi32(scales_3, 238);
5313
5314                    const __m256i scale_0145_4 = _mm256_shuffle_epi32(scales_4, 68);
5315                    const __m256i scale_2367_4 = _mm256_shuffle_epi32(scales_4, 238);
5316
5317                    const __m256i scale_0145_5 = _mm256_shuffle_epi32(scales_5, 68);
5318                    const __m256i scale_2367_5 = _mm256_shuffle_epi32(scales_5, 238);
5319
5320                    const __m256i scale_0145_6 = _mm256_shuffle_epi32(scales_6, 68);
5321                    const __m256i scale_2367_6 = _mm256_shuffle_epi32(scales_6, 238);
5322
5323                    const __m256i scale_0145_7 = _mm256_shuffle_epi32(scales_7, 68);
5324                    const __m256i scale_2367_7 = _mm256_shuffle_epi32(scales_7, 238);
5325
5326
5327                    for (int rp = 0; rp < 4; rp++) {
5328
5329                        // Load the four block_q8_k quantized values interleaved with each other in chunks of eight bytes - A0,A1,A2,A3
5330                        // Loaded as set of 128 bit vectors and repeated into a 256 bit vector
5331                        __m256i lhs_mat_0123_00 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 512 * sb)));
5332                        __m256i lhs_mat_01_00 = _mm256_permute2f128_si256(lhs_mat_0123_00, lhs_mat_0123_00, 0);
5333                        __m256i lhs_mat_23_00 = _mm256_permute2f128_si256(lhs_mat_0123_00, lhs_mat_0123_00, 17);
5334                        __m256i lhs_mat_0123_01 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 32 + 512 * sb)));
5335                        __m256i lhs_mat_01_01 = _mm256_permute2f128_si256(lhs_mat_0123_01, lhs_mat_0123_01, 0);
5336                        __m256i lhs_mat_23_01 = _mm256_permute2f128_si256(lhs_mat_0123_01, lhs_mat_0123_01, 17);
5337                        __m256i lhs_mat_0123_10 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 64 + 512 * sb)));
5338                        __m256i lhs_mat_01_10 = _mm256_permute2f128_si256(lhs_mat_0123_10, lhs_mat_0123_10, 0);
5339                        __m256i lhs_mat_23_10 = _mm256_permute2f128_si256(lhs_mat_0123_10, lhs_mat_0123_10, 17);
5340                        __m256i lhs_mat_0123_11 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 96 + 512 * sb)));
5341                        __m256i lhs_mat_01_11 = _mm256_permute2f128_si256(lhs_mat_0123_11, lhs_mat_0123_11, 0);
5342                        __m256i lhs_mat_23_11 = _mm256_permute2f128_si256(lhs_mat_0123_11, lhs_mat_0123_11, 17);
5343                        __m256i lhs_mat_0123_20 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 128 + 512 * sb)));
5344                        __m256i lhs_mat_01_20 = _mm256_permute2f128_si256(lhs_mat_0123_20, lhs_mat_0123_20, 0);
5345                        __m256i lhs_mat_23_20 = _mm256_permute2f128_si256(lhs_mat_0123_20, lhs_mat_0123_20, 17);
5346                        __m256i lhs_mat_0123_21 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 160 + 512 * sb)));
5347                        __m256i lhs_mat_01_21 = _mm256_permute2f128_si256(lhs_mat_0123_21, lhs_mat_0123_21, 0);
5348                        __m256i lhs_mat_23_21 = _mm256_permute2f128_si256(lhs_mat_0123_21, lhs_mat_0123_21, 17);
5349                        __m256i lhs_mat_0123_30 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 192 + 512 * sb)));
5350                        __m256i lhs_mat_01_30 = _mm256_permute2f128_si256(lhs_mat_0123_30, lhs_mat_0123_30, 0);
5351                        __m256i lhs_mat_23_30 = _mm256_permute2f128_si256(lhs_mat_0123_30, lhs_mat_0123_30, 17);
5352                        __m256i lhs_mat_0123_31 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 224 + 512 * sb)));
5353                        __m256i lhs_mat_01_31 = _mm256_permute2f128_si256(lhs_mat_0123_31, lhs_mat_0123_31, 0);
5354                        __m256i lhs_mat_23_31 = _mm256_permute2f128_si256(lhs_mat_0123_31, lhs_mat_0123_31, 17);
5355
5356                        __m256i lhs_mat_0123_40 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 256 + 512 * sb)));
5357                        __m256i lhs_mat_01_40 = _mm256_permute2f128_si256(lhs_mat_0123_40, lhs_mat_0123_40, 0);
5358                        __m256i lhs_mat_23_40 = _mm256_permute2f128_si256(lhs_mat_0123_40, lhs_mat_0123_40, 17);
5359                        __m256i lhs_mat_0123_41 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 288 + 512 * sb)));
5360                        __m256i lhs_mat_01_41 = _mm256_permute2f128_si256(lhs_mat_0123_41, lhs_mat_0123_41, 0);
5361                        __m256i lhs_mat_23_41 = _mm256_permute2f128_si256(lhs_mat_0123_41, lhs_mat_0123_41, 17);
5362                        __m256i lhs_mat_0123_50 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 320 + 512 * sb)));
5363                        __m256i lhs_mat_01_50 = _mm256_permute2f128_si256(lhs_mat_0123_50, lhs_mat_0123_50, 0);
5364                        __m256i lhs_mat_23_50 = _mm256_permute2f128_si256(lhs_mat_0123_50, lhs_mat_0123_50, 17);
5365                        __m256i lhs_mat_0123_51 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 352 + 512 * sb)));
5366                        __m256i lhs_mat_01_51 = _mm256_permute2f128_si256(lhs_mat_0123_51, lhs_mat_0123_51, 0);
5367                        __m256i lhs_mat_23_51 = _mm256_permute2f128_si256(lhs_mat_0123_51, lhs_mat_0123_51, 17);
5368                        __m256i lhs_mat_0123_60 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 384 + 512 * sb)));
5369                        __m256i lhs_mat_01_60 = _mm256_permute2f128_si256(lhs_mat_0123_60, lhs_mat_0123_60, 0);
5370                        __m256i lhs_mat_23_60 = _mm256_permute2f128_si256(lhs_mat_0123_60, lhs_mat_0123_60, 17);
5371                        __m256i lhs_mat_0123_61 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 416 + 512 * sb)));
5372                        __m256i lhs_mat_01_61 = _mm256_permute2f128_si256(lhs_mat_0123_61, lhs_mat_0123_61, 0);
5373                        __m256i lhs_mat_23_61 = _mm256_permute2f128_si256(lhs_mat_0123_61, lhs_mat_0123_61, 17);
5374                        __m256i lhs_mat_0123_70 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 448 + 512 * sb)));
5375                        __m256i lhs_mat_01_70 = _mm256_permute2f128_si256(lhs_mat_0123_70, lhs_mat_0123_70, 0);
5376                        __m256i lhs_mat_23_70 = _mm256_permute2f128_si256(lhs_mat_0123_70, lhs_mat_0123_70, 17);
5377                        __m256i lhs_mat_0123_71 = _mm256_loadu_si256((const __m256i * )((a_ptrs[rp][b].qs + 480 + 512 * sb)));
5378                        __m256i lhs_mat_01_71 = _mm256_permute2f128_si256(lhs_mat_0123_71, lhs_mat_0123_71, 0);
5379                        __m256i lhs_mat_23_71 = _mm256_permute2f128_si256(lhs_mat_0123_71, lhs_mat_0123_71, 17);
5380
5381                        // Bsums are loaded for the different Q8_K blocks
5382                        __m128i lhs_raw_bsums_01_0123 = _mm_loadu_si128((const __m128i *)((a_ptrs[rp][b].bsums + 32 * sb)));
5383                        __m128i lhs_raw_bsums_23_0123 = _mm_loadu_si128((const __m128i *)(a_ptrs[rp][b].bsums + 8 + 32 * sb));
5384                        __m128i lhs_raw_bsums_01_4567 = _mm_loadu_si128((const __m128i *)((a_ptrs[rp][b].bsums + 16 + 32 * sb)));
5385                        __m128i lhs_raw_bsums_23_4567 = _mm_loadu_si128((const __m128i *)(a_ptrs[rp][b].bsums + 24 + 32 * sb));
5386
5387                        // Shuffle pattern one - left side input
5388                        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)
5389                        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)
5390
5391                        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)
5392                        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)
5393
5394                        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)
5395                        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)
5396
5397                        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)
5398                        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)
5399
5400                        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)
5401                        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)
5402
5403                        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)
5404                        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)
5405
5406                        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)
5407                        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)
5408
5409                        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)
5410                        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)
5411
5412                        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)
5413                        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)
5414
5415                        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)
5416                        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)
5417
5418                        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)
5419                        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)
5420
5421                        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)
5422                        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)
5423
5424                        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)
5425                        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)
5426
5427                        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)
5428                        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)
5429
5430                        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)
5431                        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)
5432
5433                        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)
5434                        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)
5435
5436                        // Shuffle pattern two- left side input
5437                        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)
5438                        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)
5439
5440                        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)
5441                        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)
5442
5443                        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)
5444                        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)
5445
5446                        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)
5447                        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)
5448
5449                        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)
5450                        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)
5451
5452                        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)
5453                        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)
5454
5455                        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)
5456                        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)
5457
5458                        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)
5459                        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)
5460
5461                        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)
5462                        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)
5463
5464                        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)
5465                        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)
5466
5467                        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)
5468                        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)
5469
5470                        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)
5471                        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)
5472
5473                        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)
5474                        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)
5475
5476                        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)
5477                        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)
5478
5479                        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)
5480                        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)
5481
5482                        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)
5483                        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)
5484
5485                        // 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
5486                        __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));
5487                        __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));
5488
5489                        __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));
5490                        __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));
5491
5492                        __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));
5493                        __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));
5494
5495                        __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));
5496                        __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));
5497
5498                        __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));
5499                        __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));
5500
5501                        __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));
5502                        __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));
5503
5504                        __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));
5505                        __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));
5506
5507                        __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));
5508                        __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));
5509
5510                        __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));
5511                        __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));
5512
5513                        __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));
5514                        __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));
5515
5516                        __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));
5517                        __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));
5518
5519                        __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));
5520                        __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));
5521
5522                        __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));
5523                        __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));
5524
5525                        __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));
5526                        __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));
5527
5528                        __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));
5529                        __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));
5530
5531                        __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));
5532                        __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));
5533
5534
5535                        __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));
5536                        __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));
5537
5538                        __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));
5539                        __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));
5540
5541                        __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));
5542                        __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));
5543
5544                        __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));
5545                        __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));
5546
5547                        __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));
5548                        __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));
5549
5550                        __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));
5551                        __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));
5552
5553                        __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));
5554                        __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));
5555
5556                        __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));
5557                        __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));
5558
5559                        __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));
5560                        __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));
5561
5562                        __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));
5563                        __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));
5564
5565                        __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));
5566                        __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));
5567
5568                        __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));
5569                        __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));
5570
5571                        __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));
5572                        __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));
5573
5574                        __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));
5575                        __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));
5576
5577                        __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));
5578                        __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));
5579
5580                        __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));
5581                        __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));
5582
5583                        // Combine results from both shuffle patterns for each output block
5584                        __m256i iacc_mat_00_0 = _mm256_add_epi16(iacc_mat_00_0_sp1, iacc_mat_00_0_sp2);
5585                        __m256i iacc_mat_01_0 = _mm256_add_epi16(iacc_mat_01_0_sp1, iacc_mat_01_0_sp2);
5586                        __m256i iacc_mat_10_0 = _mm256_add_epi16(iacc_mat_10_0_sp1, iacc_mat_10_0_sp2);
5587                        __m256i iacc_mat_11_0 = _mm256_add_epi16(iacc_mat_11_0_sp1, iacc_mat_11_0_sp2);
5588
5589                        __m256i iacc_mat_00_1 = _mm256_add_epi16(iacc_mat_00_1_sp1, iacc_mat_00_1_sp2);
5590                        __m256i iacc_mat_01_1 = _mm256_add_epi16(iacc_mat_01_1_sp1, iacc_mat_01_1_sp2);
5591                        __m256i iacc_mat_10_1 = _mm256_add_epi16(iacc_mat_10_1_sp1, iacc_mat_10_1_sp2);
5592                        __m256i iacc_mat_11_1 = _mm256_add_epi16(iacc_mat_11_1_sp1, iacc_mat_11_1_sp2);
5593
5594                        __m256i iacc_mat_00_2 = _mm256_add_epi16(iacc_mat_00_2_sp1, iacc_mat_00_2_sp2);
5595                        __m256i iacc_mat_01_2 = _mm256_add_epi16(iacc_mat_01_2_sp1, iacc_mat_01_2_sp2);
5596                        __m256i iacc_mat_10_2 = _mm256_add_epi16(iacc_mat_10_2_sp1, iacc_mat_10_2_sp2);
5597                        __m256i iacc_mat_11_2 = _mm256_add_epi16(iacc_mat_11_2_sp1, iacc_mat_11_2_sp2);
5598
5599                        __m256i iacc_mat_00_3 = _mm256_add_epi16(iacc_mat_00_3_sp1, iacc_mat_00_3_sp2);
5600                        __m256i iacc_mat_01_3 = _mm256_add_epi16(iacc_mat_01_3_sp1, iacc_mat_01_3_sp2);
5601                        __m256i iacc_mat_10_3 = _mm256_add_epi16(iacc_mat_10_3_sp1, iacc_mat_10_3_sp2);
5602                        __m256i iacc_mat_11_3 = _mm256_add_epi16(iacc_mat_11_3_sp1, iacc_mat_11_3_sp2);
5603
5604                        __m256i iacc_mat_00_4 = _mm256_add_epi16(iacc_mat_00_4_sp1, iacc_mat_00_4_sp2);
5605                        __m256i iacc_mat_01_4 = _mm256_add_epi16(iacc_mat_01_4_sp1, iacc_mat_01_4_sp2);
5606                        __m256i iacc_mat_10_4 = _mm256_add_epi16(iacc_mat_10_4_sp1, iacc_mat_10_4_sp2);
5607                        __m256i iacc_mat_11_4 = _mm256_add_epi16(iacc_mat_11_4_sp1, iacc_mat_11_4_sp2);
5608
5609                        __m256i iacc_mat_00_5 = _mm256_add_epi16(iacc_mat_00_5_sp1, iacc_mat_00_5_sp2);
5610                        __m256i iacc_mat_01_5 = _mm256_add_epi16(iacc_mat_01_5_sp1, iacc_mat_01_5_sp2);
5611                        __m256i iacc_mat_10_5 = _mm256_add_epi16(iacc_mat_10_5_sp1, iacc_mat_10_5_sp2);
5612                        __m256i iacc_mat_11_5 = _mm256_add_epi16(iacc_mat_11_5_sp1, iacc_mat_11_5_sp2);
5613
5614                        __m256i iacc_mat_00_6 = _mm256_add_epi16(iacc_mat_00_6_sp1, iacc_mat_00_6_sp2);
5615                        __m256i iacc_mat_01_6 = _mm256_add_epi16(iacc_mat_01_6_sp1, iacc_mat_01_6_sp2);
5616                        __m256i iacc_mat_10_6 = _mm256_add_epi16(iacc_mat_10_6_sp1, iacc_mat_10_6_sp2);
5617                        __m256i iacc_mat_11_6 = _mm256_add_epi16(iacc_mat_11_6_sp1, iacc_mat_11_6_sp2);
5618
5619                        __m256i iacc_mat_00_7 = _mm256_add_epi16(iacc_mat_00_7_sp1, iacc_mat_00_7_sp2);
5620                        __m256i iacc_mat_01_7 = _mm256_add_epi16(iacc_mat_01_7_sp1, iacc_mat_01_7_sp2);
5621                        __m256i iacc_mat_10_7 = _mm256_add_epi16(iacc_mat_10_7_sp1, iacc_mat_10_7_sp2);
5622                        __m256i iacc_mat_11_7 = _mm256_add_epi16(iacc_mat_11_7_sp1, iacc_mat_11_7_sp2);
5623
5624                        // Output of both shuffle patterns are added in order to sum dot product outputs of all 32 values in block
5625                        iacc_mat_00_0 = _mm256_madd_epi16(iacc_mat_00_0, scale_0145_0);
5626                        iacc_mat_01_0 = _mm256_madd_epi16(iacc_mat_01_0, scale_2367_0);
5627                        iacc_mat_10_0 = _mm256_madd_epi16(iacc_mat_10_0, scale_0145_0);
5628                        iacc_mat_11_0 = _mm256_madd_epi16(iacc_mat_11_0, scale_2367_0);
5629
5630                        iacc_mat_00_1 = _mm256_madd_epi16(iacc_mat_00_1, scale_0145_1);
5631                        iacc_mat_01_1 = _mm256_madd_epi16(iacc_mat_01_1, scale_2367_1);
5632                        iacc_mat_10_1 = _mm256_madd_epi16(iacc_mat_10_1, scale_0145_1);
5633                        iacc_mat_11_1 = _mm256_madd_epi16(iacc_mat_11_1, scale_2367_1);
5634
5635                        iacc_mat_00_2 = _mm256_madd_epi16(iacc_mat_00_2, scale_0145_2);
5636                        iacc_mat_01_2 = _mm256_madd_epi16(iacc_mat_01_2, scale_2367_2);
5637                        iacc_mat_10_2 = _mm256_madd_epi16(iacc_mat_10_2, scale_0145_2);
5638                        iacc_mat_11_2 = _mm256_madd_epi16(iacc_mat_11_2, scale_2367_2);
5639
5640                        iacc_mat_00_3 = _mm256_madd_epi16(iacc_mat_00_3, scale_0145_3);
5641                        iacc_mat_01_3 = _mm256_madd_epi16(iacc_mat_01_3, scale_2367_3);
5642                        iacc_mat_10_3 = _mm256_madd_epi16(iacc_mat_10_3, scale_0145_3);
5643                        iacc_mat_11_3 = _mm256_madd_epi16(iacc_mat_11_3, scale_2367_3);
5644
5645                        iacc_mat_00_4 = _mm256_madd_epi16(iacc_mat_00_4, scale_0145_4);
5646                        iacc_mat_01_4 = _mm256_madd_epi16(iacc_mat_01_4, scale_2367_4);
5647                        iacc_mat_10_4 = _mm256_madd_epi16(iacc_mat_10_4, scale_0145_4);
5648                        iacc_mat_11_4 = _mm256_madd_epi16(iacc_mat_11_4, scale_2367_4);
5649
5650                        iacc_mat_00_5 = _mm256_madd_epi16(iacc_mat_00_5, scale_0145_5);
5651                        iacc_mat_01_5 = _mm256_madd_epi16(iacc_mat_01_5, scale_2367_5);
5652                        iacc_mat_10_5 = _mm256_madd_epi16(iacc_mat_10_5, scale_0145_5);
5653                        iacc_mat_11_5 = _mm256_madd_epi16(iacc_mat_11_5, scale_2367_5);
5654
5655                        iacc_mat_00_6 = _mm256_madd_epi16(iacc_mat_00_6, scale_0145_6);
5656                        iacc_mat_01_6 = _mm256_madd_epi16(iacc_mat_01_6, scale_2367_6);
5657                        iacc_mat_10_6 = _mm256_madd_epi16(iacc_mat_10_6, scale_0145_6);
5658                        iacc_mat_11_6 = _mm256_madd_epi16(iacc_mat_11_6, scale_2367_6);
5659
5660                        iacc_mat_00_7 = _mm256_madd_epi16(iacc_mat_00_7, scale_0145_7);
5661                        iacc_mat_01_7 = _mm256_madd_epi16(iacc_mat_01_7, scale_2367_7);
5662                        iacc_mat_10_7 = _mm256_madd_epi16(iacc_mat_10_7, scale_0145_7);
5663                        iacc_mat_11_7 = _mm256_madd_epi16(iacc_mat_11_7, scale_2367_7);
5664
5665                        __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)));
5666                        __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)));
5667                        __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)));
5668                        __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)));
5669
5670                        // Straighten out to make 4 row vectors
5671                        __m256i iacc_row_0 = _mm256_blend_epi32(iacc_mat_00, _mm256_shuffle_epi32(iacc_mat_01, 78), 204);
5672                        __m256i iacc_row_1 = _mm256_blend_epi32(_mm256_shuffle_epi32(iacc_mat_00, 78), iacc_mat_01, 204);
5673                        __m256i iacc_row_2 = _mm256_blend_epi32(iacc_mat_10, _mm256_shuffle_epi32(iacc_mat_11, 78), 204);
5674                        __m256i iacc_row_3 = _mm256_blend_epi32(_mm256_shuffle_epi32(iacc_mat_10, 78), iacc_mat_11, 204);
5675
5676                        // Load the scale(d) values for all the 4 Q8_k blocks and repeat it across lanes
5677                        const __m128 row_scale_f32_sse = _mm_load_ps(a_ptrs[rp][b].d);
5678                        const __m256 row_scale_f32 = _mm256_set_m128(row_scale_f32_sse, row_scale_f32_sse);
5679
5680                        // Multiply with appropriate scales and accumulate (for both d and dmin) below
5681                        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]);
5682                        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]);
5683                        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]);
5684                        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]);
5685
5686                        __m256i lhs_bsums_01_0123 = _mm256_inserti128_si256(_mm256_castsi128_si256(lhs_raw_bsums_01_0123), lhs_raw_bsums_01_0123, 1);
5687                        __m256i lhs_bsums_23_0123 = _mm256_inserti128_si256(_mm256_castsi128_si256(lhs_raw_bsums_23_0123), lhs_raw_bsums_23_0123, 1);
5688                        __m256i lhs_bsums_01_4567 = _mm256_inserti128_si256(_mm256_castsi128_si256(lhs_raw_bsums_01_4567), lhs_raw_bsums_01_4567, 1);
5689                        __m256i lhs_bsums_23_4567 = _mm256_inserti128_si256(_mm256_castsi128_si256(lhs_raw_bsums_23_4567), lhs_raw_bsums_23_4567, 1);
5690
5691                       // Take two bsums from two Q8_Ks at a time and multiply with corresponding mins values from each Q2_K
5692                        __m256i iacc_row_min_0_01 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_01_0123, 0), mins_01);
5693                        __m256i iacc_row_min_1_01 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_01_0123, 170), mins_01);
5694                        __m256i iacc_row_min_2_01 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_23_0123, 0), mins_01);
5695                        __m256i iacc_row_min_3_01 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_23_0123, 170), mins_01);
5696
5697                        __m256i iacc_row_min_0_23 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_01_0123, 85), mins_23);
5698                        __m256i iacc_row_min_1_23 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_01_0123, 255), mins_23);
5699                        __m256i iacc_row_min_2_23 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_23_0123, 85), mins_23);
5700                        __m256i iacc_row_min_3_23 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_23_0123, 255), mins_23);
5701
5702                        __m256i iacc_row_min_0_45 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_01_4567, 0), mins_45);
5703                        __m256i iacc_row_min_1_45 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_01_4567, 170), mins_45);
5704                        __m256i iacc_row_min_2_45 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_23_4567, 0), mins_45);
5705                        __m256i iacc_row_min_3_45 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_23_4567, 170), mins_45);
5706
5707                        __m256i iacc_row_min_0_67 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_01_4567, 85), mins_67);
5708                        __m256i iacc_row_min_1_67 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_01_4567, 255), mins_67);
5709                        __m256i iacc_row_min_2_67 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_23_4567, 85), mins_67);
5710                        __m256i iacc_row_min_3_67 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_23_4567, 255), mins_67);
5711
5712                        __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));
5713                        __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));
5714                        __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));
5715                        __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));
5716
5717                        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]);
5718                        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]);
5719                        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]);
5720                        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]);
5721
5722                    }
5723                }
5724            }
5725            // Store the accumulated values
5726            for (int i = 0; i < 16; i++) {
5727                _mm256_storeu_ps((float * )(s + ((y * 4 + i) * bs + x * 8)), _mm256_sub_ps(acc_rows[i], acc_min_rows[i]));
5728
5729            }
5730        }
5731    }
5732
5733    for (; y < nr / 4; y ++) {
5734
5735        const block_q8_Kx4 * a_ptr = a_ptr_start + (y * nb);
5736
5737        // Take group of eight block_q2_kx8 structures at each pass of the loop and perform dot product operation
5738        for (int64_t x = xstart; x < nc / 8; x++) {
5739
5740            const block_q2_Kx8 * b_ptr = b_ptr_start + (x * b_nb);
5741
5742            // Master FP accumulators
5743            __m256 acc_rows[4];
5744            for (int i = 0; i < 4; i++) {
5745                acc_rows[i] = _mm256_setzero_ps();
5746            }
5747
5748            __m256 acc_min_rows[4];
5749            for (int i = 0; i < 4; i++) {
5750                acc_min_rows[i] = _mm256_setzero_ps();
5751            }
5752
5753            for (int64_t b = 0; b < nb; b++) {
5754                // Delta values - Load the eight scale values of block_q2_kx8
5755                const __m256 col_scale_f32 = GGML_F32Cx8_LOAD(b_ptr[b].d);
5756
5757                // dmin values - Load the eight dmin values of block_q2_kx8
5758                const __m256 col_dmin_f32 = GGML_F32Cx8_LOAD(b_ptr[b].dmin);
5759
5760                // Loop to iterate over the sixteen sub blocks of a super block - eight sub blocks are processed per iteration
5761                for (int sb = 0; sb < QK_K / 128; sb++) {
5762
5763                    // Load the eight block_q2_k for eight sub blocks quantized values interleaved with each other in chunks of eight bytes - B0,B1 ....B6,B7
5764                    const __m256i rhs_raw_mat_0123_0 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + sb * 256));
5765                    const __m256i rhs_raw_mat_4567_0 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 32 + sb * 256));
5766                    const __m256i rhs_raw_mat_0123_1 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 64 + sb * 256));
5767                    const __m256i rhs_raw_mat_4567_1 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 96 + sb * 256));
5768                    const __m256i rhs_raw_mat_0123_2 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 128 + sb * 256));
5769                    const __m256i rhs_raw_mat_4567_2 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 160 + sb * 256));
5770                    const __m256i rhs_raw_mat_0123_3 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 192 + sb * 256));
5771                    const __m256i rhs_raw_mat_4567_3 = _mm256_loadu_si256((const __m256i *)(b_ptr[b].qs + 224 + sb * 256));
5772
5773                    // Save the values in the following vectors in the formats B0B1B4B5, B2B3B6B7 for further processing and storing of values
5774                    //superblock    sub block   which part of sub block
5775                    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);
5776                    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);
5777
5778                    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);
5779                    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);
5780
5781                    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);
5782                    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);
5783
5784                    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);
5785                    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);
5786
5787                    // 2-bit -> 8-bit
5788                    // First sub block of the eight sub blocks processed in the iteration
5789                    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)
5790                    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)
5791
5792                    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)
5793                    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)
5794
5795                    // Second sub block of the eight sub blocks processed in the iteration
5796                    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)
5797                    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)
5798
5799                    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)
5800                    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)
5801
5802                    // Third sub block of the eight sub blocks processed in the iteration
5803                    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)
5804                    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)
5805
5806                    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)
5807                    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)
5808
5809                    // Fourth sub block of the eight sub blocks processed in the iteration
5810                    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)
5811                    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)
5812
5813                    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)
5814                    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)
5815
5816                    // Fifth sub block of the eight sub blocks processed in the iteration
5817                    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)
5818                    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)
5819
5820                    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)
5821                    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)
5822
5823                    // Sixth sub block of the eight sub blocks processed in the iteration
5824                    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)
5825                    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)
5826
5827                    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)
5828                    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)
5829
5830                    // Seventh sub block of the eight sub blocks processed in the iteration
5831                    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)
5832                    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)
5833
5834                    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)
5835                    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)
5836
5837                    // Eighth sub block of the eight sub blocks processed in the iteration
5838                    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)
5839                    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)
5840
5841                    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)
5842                    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)
5843
5844                    // Shuffle pattern one - right side input
5845                    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)
5846                    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)
5847
5848                    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)
5849                    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)
5850
5851                    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)
5852                    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)
5853
5854                    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)
5855                    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)
5856
5857                    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)
5858                    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)
5859
5860                    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)
5861                    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)
5862
5863                    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)
5864                    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)
5865
5866                    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
5867                    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)
5868
5869                    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)
5870                    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)
5871
5872                    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)
5873                    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)
5874
5875                    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)
5876                    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)
5877
5878                    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)
5879                    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)
5880
5881                    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)
5882                    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)
5883
5884                    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)
5885                    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)
5886
5887                    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)
5888                    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)
5889
5890                    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)
5891                    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)
5892
5893
5894                    // Shuffle pattern two - right side input
5895                    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)
5896                    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)
5897
5898                    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)
5899                    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)
5900
5901                    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)
5902                    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)
5903
5904                    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)
5905                    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)
5906
5907                    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)
5908                    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)
5909
5910                    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)
5911                    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)
5912
5913                    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)
5914                    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)
5915
5916                    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)
5917                    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)
5918
5919                    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)
5920                    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)
5921
5922                    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)
5923                    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)
5924
5925                    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)
5926                    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)
5927
5928                    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)
5929                    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)
5930
5931                    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)
5932                    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)
5933
5934                    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)
5935                    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)
5936
5937                    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)
5938                    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)
5939
5940                    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)
5941                    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)
5942
5943
5944                    //Scales and Mins of corresponding sub blocks from different Q2_K structures are stored together
5945                    //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
5946
5947                    // Combine mins and scales for sub-blocks: 0-1, 2-3, 4-5, 6-7 in the sb loop
5948                    const __m128i mins_and_scales_01 = _mm_loadu_si128((const __m128i *)(b_ptr[b].scales + sb * 64));
5949                    const __m128i mins_and_scales_23 = _mm_loadu_si128((const __m128i *)(b_ptr[b].scales + 16 + sb * 64));
5950                    const __m128i mins_and_scales_45 = _mm_loadu_si128((const __m128i *)(b_ptr[b].scales + 32 + sb * 64));
5951                    const __m128i mins_and_scales_67 = _mm_loadu_si128((const __m128i *)(b_ptr[b].scales + 48 + sb * 64));
5952
5953                    // Extract scales which is lower half from mins_and_scales
5954                    const __m128i scales_01 = _mm_and_si128(mins_and_scales_01, m4b_sse);
5955                    const __m128i scales_23 = _mm_and_si128(mins_and_scales_23, m4b_sse);
5956                    const __m128i scales_45 = _mm_and_si128(mins_and_scales_45, m4b_sse);
5957                    const __m128i scales_67 = _mm_and_si128(mins_and_scales_67, m4b_sse);
5958
5959                    // Extract mins which is upper half from mins_and_scales
5960                    const __m256i mins_01 = _mm256_cvtepu8_epi16(_mm_and_si128(_mm_srli_epi16(mins_and_scales_01, 4), m4b_sse));
5961                    const __m256i mins_23 = _mm256_cvtepu8_epi16(_mm_and_si128(_mm_srli_epi16(mins_and_scales_23, 4), m4b_sse));
5962                    const __m256i mins_45 = _mm256_cvtepu8_epi16(_mm_and_si128(_mm_srli_epi16(mins_and_scales_45, 4), m4b_sse));
5963                    const __m256i mins_67 = _mm256_cvtepu8_epi16(_mm_and_si128(_mm_srli_epi16(mins_and_scales_67, 4), m4b_sse));
5964
5965                    const __m256i scales_0 = _mm256_cvtepu8_epi16(_mm_shuffle_epi8(scales_01, scalesmask1_sse));
5966                    const __m256i scales_1 = _mm256_cvtepu8_epi16(_mm_shuffle_epi8(scales_01, scalesmask2_sse));
5967
5968                    const __m256i scales_2 = _mm256_cvtepu8_epi16(_mm_shuffle_epi8(scales_23, scalesmask1_sse));
5969                    const __m256i scales_3 = _mm256_cvtepu8_epi16(_mm_shuffle_epi8(scales_23, scalesmask2_sse));
5970
5971                    const __m256i scales_4 = _mm256_cvtepu8_epi16(_mm_shuffle_epi8(scales_45, scalesmask1_sse));
5972                    const __m256i scales_5 = _mm256_cvtepu8_epi16(_mm_shuffle_epi8(scales_45, scalesmask2_sse));
5973
5974                    const __m256i scales_6 = _mm256_cvtepu8_epi16(_mm_shuffle_epi8(scales_67, scalesmask1_sse));
5975                    const __m256i scales_7 = _mm256_cvtepu8_epi16(_mm_shuffle_epi8(scales_67, scalesmask2_sse));
5976
5977                    const __m256i scale_0145_0 = _mm256_shuffle_epi32(scales_0, 68);
5978                    const __m256i scale_2367_0 = _mm256_shuffle_epi32(scales_0, 238);
5979
5980                    const __m256i scale_0145_1 = _mm256_shuffle_epi32(scales_1, 68);
5981                    const __m256i scale_2367_1 = _mm256_shuffle_epi32(scales_1, 238);
5982
5983                    const __m256i scale_0145_2 = _mm256_shuffle_epi32(scales_2, 68);
5984                    const __m256i scale_2367_2 = _mm256_shuffle_epi32(scales_2, 238);
5985
5986                    const __m256i scale_0145_3 = _mm256_shuffle_epi32(scales_3, 68);
5987                    const __m256i scale_2367_3 = _mm256_shuffle_epi32(scales_3, 238);
5988
5989                    const __m256i scale_0145_4 = _mm256_shuffle_epi32(scales_4, 68);
5990                    const __m256i scale_2367_4 = _mm256_shuffle_epi32(scales_4, 238);
5991
5992                    const __m256i scale_0145_5 = _mm256_shuffle_epi32(scales_5, 68);
5993                    const __m256i scale_2367_5 = _mm256_shuffle_epi32(scales_5, 238);
5994
5995                    const __m256i scale_0145_6 = _mm256_shuffle_epi32(scales_6, 68);
5996                    const __m256i scale_2367_6 = _mm256_shuffle_epi32(scales_6, 238);
5997
5998                    const __m256i scale_0145_7 = _mm256_shuffle_epi32(scales_7, 68);
5999                    const __m256i scale_2367_7 = _mm256_shuffle_epi32(scales_7, 238);
6000
6001                    // Load the four block_q8_k quantized values interleaved with each other in chunks of eight bytes - A0,A1,A2,A3
6002                    // Loaded as set of 128 bit vectors and repeated into a 256 bit vector
6003                    __m256i lhs_mat_0123_00 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 512 * sb)));
6004                    __m256i lhs_mat_01_00 = _mm256_permute2f128_si256(lhs_mat_0123_00, lhs_mat_0123_00, 0);
6005                    __m256i lhs_mat_23_00 = _mm256_permute2f128_si256(lhs_mat_0123_00, lhs_mat_0123_00, 17);
6006                    __m256i lhs_mat_0123_01 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 32 + 512 * sb)));
6007                    __m256i lhs_mat_01_01 = _mm256_permute2f128_si256(lhs_mat_0123_01, lhs_mat_0123_01, 0);
6008                    __m256i lhs_mat_23_01 = _mm256_permute2f128_si256(lhs_mat_0123_01, lhs_mat_0123_01, 17);
6009                    __m256i lhs_mat_0123_10 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 64 + 512 * sb)));
6010                    __m256i lhs_mat_01_10 = _mm256_permute2f128_si256(lhs_mat_0123_10, lhs_mat_0123_10, 0);
6011                    __m256i lhs_mat_23_10 = _mm256_permute2f128_si256(lhs_mat_0123_10, lhs_mat_0123_10, 17);
6012                    __m256i lhs_mat_0123_11 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 96 + 512 * sb)));
6013                    __m256i lhs_mat_01_11 = _mm256_permute2f128_si256(lhs_mat_0123_11, lhs_mat_0123_11, 0);
6014                    __m256i lhs_mat_23_11 = _mm256_permute2f128_si256(lhs_mat_0123_11, lhs_mat_0123_11, 17);
6015                    __m256i lhs_mat_0123_20 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 128 + 512 * sb)));
6016                    __m256i lhs_mat_01_20 = _mm256_permute2f128_si256(lhs_mat_0123_20, lhs_mat_0123_20, 0);
6017                    __m256i lhs_mat_23_20 = _mm256_permute2f128_si256(lhs_mat_0123_20, lhs_mat_0123_20, 17);
6018                    __m256i lhs_mat_0123_21 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 160 + 512 * sb)));
6019                    __m256i lhs_mat_01_21 = _mm256_permute2f128_si256(lhs_mat_0123_21, lhs_mat_0123_21, 0);
6020                    __m256i lhs_mat_23_21 = _mm256_permute2f128_si256(lhs_mat_0123_21, lhs_mat_0123_21, 17);
6021                    __m256i lhs_mat_0123_30 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 192 + 512 * sb)));
6022                    __m256i lhs_mat_01_30 = _mm256_permute2f128_si256(lhs_mat_0123_30, lhs_mat_0123_30, 0);
6023                    __m256i lhs_mat_23_30 = _mm256_permute2f128_si256(lhs_mat_0123_30, lhs_mat_0123_30, 17);
6024                    __m256i lhs_mat_0123_31 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 224 + 512 * sb)));
6025                    __m256i lhs_mat_01_31 = _mm256_permute2f128_si256(lhs_mat_0123_31, lhs_mat_0123_31, 0);
6026                    __m256i lhs_mat_23_31 = _mm256_permute2f128_si256(lhs_mat_0123_31, lhs_mat_0123_31, 17);
6027
6028                    __m256i lhs_mat_0123_40 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 256 + 512 * sb)));
6029                    __m256i lhs_mat_01_40 = _mm256_permute2f128_si256(lhs_mat_0123_40, lhs_mat_0123_40, 0);
6030                    __m256i lhs_mat_23_40 = _mm256_permute2f128_si256(lhs_mat_0123_40, lhs_mat_0123_40, 17);
6031                    __m256i lhs_mat_0123_41 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 288 + 512 * sb)));
6032                    __m256i lhs_mat_01_41 = _mm256_permute2f128_si256(lhs_mat_0123_41, lhs_mat_0123_41, 0);
6033                    __m256i lhs_mat_23_41 = _mm256_permute2f128_si256(lhs_mat_0123_41, lhs_mat_0123_41, 17);
6034                    __m256i lhs_mat_0123_50 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 320 + 512 * sb)));
6035                    __m256i lhs_mat_01_50 = _mm256_permute2f128_si256(lhs_mat_0123_50, lhs_mat_0123_50, 0);
6036                    __m256i lhs_mat_23_50 = _mm256_permute2f128_si256(lhs_mat_0123_50, lhs_mat_0123_50, 17);
6037                    __m256i lhs_mat_0123_51 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 352 + 512 * sb)));
6038                    __m256i lhs_mat_01_51 = _mm256_permute2f128_si256(lhs_mat_0123_51, lhs_mat_0123_51, 0);
6039                    __m256i lhs_mat_23_51 = _mm256_permute2f128_si256(lhs_mat_0123_51, lhs_mat_0123_51, 17);
6040                    __m256i lhs_mat_0123_60 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 384 + 512 * sb)));
6041                    __m256i lhs_mat_01_60 = _mm256_permute2f128_si256(lhs_mat_0123_60, lhs_mat_0123_60, 0);
6042                    __m256i lhs_mat_23_60 = _mm256_permute2f128_si256(lhs_mat_0123_60, lhs_mat_0123_60, 17);
6043                    __m256i lhs_mat_0123_61 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 416 + 512 * sb)));
6044                    __m256i lhs_mat_01_61 = _mm256_permute2f128_si256(lhs_mat_0123_61, lhs_mat_0123_61, 0);
6045                    __m256i lhs_mat_23_61 = _mm256_permute2f128_si256(lhs_mat_0123_61, lhs_mat_0123_61, 17);
6046                    __m256i lhs_mat_0123_70 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 448 + 512 * sb)));
6047                    __m256i lhs_mat_01_70 = _mm256_permute2f128_si256(lhs_mat_0123_70, lhs_mat_0123_70, 0);
6048                    __m256i lhs_mat_23_70 = _mm256_permute2f128_si256(lhs_mat_0123_70, lhs_mat_0123_70, 17);
6049                    __m256i lhs_mat_0123_71 = _mm256_loadu_si256((const __m256i * )((a_ptr[b].qs + 480 + 512 * sb)));
6050                    __m256i lhs_mat_01_71 = _mm256_permute2f128_si256(lhs_mat_0123_71, lhs_mat_0123_71, 0);
6051                    __m256i lhs_mat_23_71 = _mm256_permute2f128_si256(lhs_mat_0123_71, lhs_mat_0123_71, 17);
6052
6053                    // Bsums are loaded for the different Q8_K blocks
6054                    __m128i lhs_raw_bsums_01_0123 = _mm_loadu_si128((const __m128i *)((a_ptr[b].bsums + 32 * sb)));
6055                    __m128i lhs_raw_bsums_23_0123 = _mm_loadu_si128((const __m128i *)(a_ptr[b].bsums + 8 + 32 * sb));
6056                    __m128i lhs_raw_bsums_01_4567 = _mm_loadu_si128((const __m128i *)((a_ptr[b].bsums + 16 + 32 * sb)));
6057                    __m128i lhs_raw_bsums_23_4567 = _mm_loadu_si128((const __m128i *)(a_ptr[b].bsums + 24 + 32 * sb));
6058
6059                    // Shuffle pattern one - left side input
6060                    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)
6061                    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)
6062
6063                    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)
6064                    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)
6065
6066                    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)
6067                    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)
6068
6069                    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)
6070                    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)
6071
6072                    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)
6073                    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)
6074
6075                    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)
6076                    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)
6077
6078                    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)
6079                    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)
6080
6081                    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)
6082                    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)
6083
6084                    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)
6085                    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)
6086
6087                    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)
6088                    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)
6089
6090                    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)
6091                    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)
6092
6093                    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)
6094                    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)
6095
6096                    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)
6097                    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)
6098
6099                    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)
6100                    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)
6101
6102                    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)
6103                    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)
6104
6105                    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)
6106                    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)
6107
6108                    // Shuffle pattern two- left side input
6109                    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)
6110                    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)
6111
6112                    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)
6113                    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)
6114
6115                    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)
6116                    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)
6117
6118                    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)
6119                    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)
6120
6121                    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)
6122                    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)
6123
6124                    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)
6125                    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)
6126
6127                    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)
6128                    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)
6129
6130                    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)
6131                    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)
6132
6133                    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)
6134                    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)
6135
6136                    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)
6137                    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)
6138
6139                    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)
6140                    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)
6141
6142                    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)
6143                    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)
6144
6145                    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)
6146                    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)
6147
6148                    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)
6149                    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)
6150
6151                    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)
6152                    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)
6153
6154                    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)
6155                    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)
6156
6157                    // 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
6158                    __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));
6159                    __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));
6160
6161                    __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));
6162                    __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));
6163
6164                    __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));
6165                    __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));
6166
6167                    __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));
6168                    __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));
6169
6170                    __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));
6171                    __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));
6172
6173                    __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));
6174                    __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));
6175
6176                    __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));
6177                    __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));
6178
6179                    __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));
6180                    __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));
6181
6182                    __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));
6183                    __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));
6184
6185                    __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));
6186                    __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));
6187
6188                    __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));
6189                    __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));
6190
6191                    __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));
6192                    __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));
6193
6194                    __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));
6195                    __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));
6196
6197                    __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));
6198                    __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));
6199
6200                    __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));
6201                    __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));
6202
6203                    __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));
6204                    __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));
6205
6206
6207                    __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));
6208                    __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));
6209
6210                    __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));
6211                    __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));
6212
6213                    __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));
6214                    __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));
6215
6216                    __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));
6217                    __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));
6218
6219                    __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));
6220                    __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));
6221
6222                    __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));
6223                    __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));
6224
6225                    __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));
6226                    __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));
6227
6228                    __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));
6229                    __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));
6230
6231                    __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));
6232                    __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));
6233
6234                    __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));
6235                    __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));
6236
6237                    __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));
6238                    __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));
6239
6240                    __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));
6241                    __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));
6242
6243                    __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));
6244                    __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));
6245
6246                    __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));
6247                    __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));
6248
6249                    __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));
6250                    __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));
6251
6252                    __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));
6253                    __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));
6254
6255                    // Combine results from both shuffle patterns for each output block.
6256                    __m256i iacc_mat_00_0 = _mm256_add_epi16(iacc_mat_00_0_sp1, iacc_mat_00_0_sp2);
6257                    __m256i iacc_mat_01_0 = _mm256_add_epi16(iacc_mat_01_0_sp1, iacc_mat_01_0_sp2);
6258                    __m256i iacc_mat_10_0 = _mm256_add_epi16(iacc_mat_10_0_sp1, iacc_mat_10_0_sp2);
6259                    __m256i iacc_mat_11_0 = _mm256_add_epi16(iacc_mat_11_0_sp1, iacc_mat_11_0_sp2);
6260
6261                    __m256i iacc_mat_00_1 = _mm256_add_epi16(iacc_mat_00_1_sp1, iacc_mat_00_1_sp2);
6262                    __m256i iacc_mat_01_1 = _mm256_add_epi16(iacc_mat_01_1_sp1, iacc_mat_01_1_sp2);
6263                    __m256i iacc_mat_10_1 = _mm256_add_epi16(iacc_mat_10_1_sp1, iacc_mat_10_1_sp2);
6264                    __m256i iacc_mat_11_1 = _mm256_add_epi16(iacc_mat_11_1_sp1, iacc_mat_11_1_sp2);
6265
6266                    __m256i iacc_mat_00_2 = _mm256_add_epi16(iacc_mat_00_2_sp1, iacc_mat_00_2_sp2);
6267                    __m256i iacc_mat_01_2 = _mm256_add_epi16(iacc_mat_01_2_sp1, iacc_mat_01_2_sp2);
6268                    __m256i iacc_mat_10_2 = _mm256_add_epi16(iacc_mat_10_2_sp1, iacc_mat_10_2_sp2);
6269                    __m256i iacc_mat_11_2 = _mm256_add_epi16(iacc_mat_11_2_sp1, iacc_mat_11_2_sp2);
6270
6271                    __m256i iacc_mat_00_3 = _mm256_add_epi16(iacc_mat_00_3_sp1, iacc_mat_00_3_sp2);
6272                    __m256i iacc_mat_01_3 = _mm256_add_epi16(iacc_mat_01_3_sp1, iacc_mat_01_3_sp2);
6273                    __m256i iacc_mat_10_3 = _mm256_add_epi16(iacc_mat_10_3_sp1, iacc_mat_10_3_sp2);
6274                    __m256i iacc_mat_11_3 = _mm256_add_epi16(iacc_mat_11_3_sp1, iacc_mat_11_3_sp2);
6275
6276                    __m256i iacc_mat_00_4 = _mm256_add_epi16(iacc_mat_00_4_sp1, iacc_mat_00_4_sp2);
6277                    __m256i iacc_mat_01_4 = _mm256_add_epi16(iacc_mat_01_4_sp1, iacc_mat_01_4_sp2);
6278                    __m256i iacc_mat_10_4 = _mm256_add_epi16(iacc_mat_10_4_sp1, iacc_mat_10_4_sp2);
6279                    __m256i iacc_mat_11_4 = _mm256_add_epi16(iacc_mat_11_4_sp1, iacc_mat_11_4_sp2);
6280
6281                    __m256i iacc_mat_00_5 = _mm256_add_epi16(iacc_mat_00_5_sp1, iacc_mat_00_5_sp2);
6282                    __m256i iacc_mat_01_5 = _mm256_add_epi16(iacc_mat_01_5_sp1, iacc_mat_01_5_sp2);
6283                    __m256i iacc_mat_10_5 = _mm256_add_epi16(iacc_mat_10_5_sp1, iacc_mat_10_5_sp2);
6284                    __m256i iacc_mat_11_5 = _mm256_add_epi16(iacc_mat_11_5_sp1, iacc_mat_11_5_sp2);
6285
6286                    __m256i iacc_mat_00_6 = _mm256_add_epi16(iacc_mat_00_6_sp1, iacc_mat_00_6_sp2);
6287                    __m256i iacc_mat_01_6 = _mm256_add_epi16(iacc_mat_01_6_sp1, iacc_mat_01_6_sp2);
6288                    __m256i iacc_mat_10_6 = _mm256_add_epi16(iacc_mat_10_6_sp1, iacc_mat_10_6_sp2);
6289                    __m256i iacc_mat_11_6 = _mm256_add_epi16(iacc_mat_11_6_sp1, iacc_mat_11_6_sp2);
6290
6291                    __m256i iacc_mat_00_7 = _mm256_add_epi16(iacc_mat_00_7_sp1, iacc_mat_00_7_sp2);
6292                    __m256i iacc_mat_01_7 = _mm256_add_epi16(iacc_mat_01_7_sp1, iacc_mat_01_7_sp2);
6293                    __m256i iacc_mat_10_7 = _mm256_add_epi16(iacc_mat_10_7_sp1, iacc_mat_10_7_sp2);
6294                    __m256i iacc_mat_11_7 = _mm256_add_epi16(iacc_mat_11_7_sp1, iacc_mat_11_7_sp2);
6295
6296                    // Output of both shuffle patterns are added in order to sum dot product outputs of all 32 values in block
6297                    iacc_mat_00_0 = _mm256_madd_epi16(iacc_mat_00_0, scale_0145_0);
6298                    iacc_mat_01_0 = _mm256_madd_epi16(iacc_mat_01_0, scale_2367_0);
6299                    iacc_mat_10_0 = _mm256_madd_epi16(iacc_mat_10_0, scale_0145_0);
6300                    iacc_mat_11_0 = _mm256_madd_epi16(iacc_mat_11_0, scale_2367_0);
6301
6302                    iacc_mat_00_1 = _mm256_madd_epi16(iacc_mat_00_1, scale_0145_1);
6303                    iacc_mat_01_1 = _mm256_madd_epi16(iacc_mat_01_1, scale_2367_1);
6304                    iacc_mat_10_1 = _mm256_madd_epi16(iacc_mat_10_1, scale_0145_1);
6305                    iacc_mat_11_1 = _mm256_madd_epi16(iacc_mat_11_1, scale_2367_1);
6306
6307                    iacc_mat_00_2 = _mm256_madd_epi16(iacc_mat_00_2, scale_0145_2);
6308                    iacc_mat_01_2 = _mm256_madd_epi16(iacc_mat_01_2, scale_2367_2);
6309                    iacc_mat_10_2 = _mm256_madd_epi16(iacc_mat_10_2, scale_0145_2);
6310                    iacc_mat_11_2 = _mm256_madd_epi16(iacc_mat_11_2, scale_2367_2);
6311
6312                    iacc_mat_00_3 = _mm256_madd_epi16(iacc_mat_00_3, scale_0145_3);
6313                    iacc_mat_01_3 = _mm256_madd_epi16(iacc_mat_01_3, scale_2367_3);
6314                    iacc_mat_10_3 = _mm256_madd_epi16(iacc_mat_10_3, scale_0145_3);
6315                    iacc_mat_11_3 = _mm256_madd_epi16(iacc_mat_11_3, scale_2367_3);
6316
6317                    iacc_mat_00_4 = _mm256_madd_epi16(iacc_mat_00_4, scale_0145_4);
6318                    iacc_mat_01_4 = _mm256_madd_epi16(iacc_mat_01_4, scale_2367_4);
6319                    iacc_mat_10_4 = _mm256_madd_epi16(iacc_mat_10_4, scale_0145_4);
6320                    iacc_mat_11_4 = _mm256_madd_epi16(iacc_mat_11_4, scale_2367_4);
6321
6322                    iacc_mat_00_5 = _mm256_madd_epi16(iacc_mat_00_5, scale_0145_5);
6323                    iacc_mat_01_5 = _mm256_madd_epi16(iacc_mat_01_5, scale_2367_5);
6324                    iacc_mat_10_5 = _mm256_madd_epi16(iacc_mat_10_5, scale_0145_5);
6325                    iacc_mat_11_5 = _mm256_madd_epi16(iacc_mat_11_5, scale_2367_5);
6326
6327                    iacc_mat_00_6 = _mm256_madd_epi16(iacc_mat_00_6, scale_0145_6);
6328                    iacc_mat_01_6 = _mm256_madd_epi16(iacc_mat_01_6, scale_2367_6);
6329                    iacc_mat_10_6 = _mm256_madd_epi16(iacc_mat_10_6, scale_0145_6);
6330                    iacc_mat_11_6 = _mm256_madd_epi16(iacc_mat_11_6, scale_2367_6);
6331
6332                    iacc_mat_00_7 = _mm256_madd_epi16(iacc_mat_00_7, scale_0145_7);
6333                    iacc_mat_01_7 = _mm256_madd_epi16(iacc_mat_01_7, scale_2367_7);
6334                    iacc_mat_10_7 = _mm256_madd_epi16(iacc_mat_10_7, scale_0145_7);
6335                    iacc_mat_11_7 = _mm256_madd_epi16(iacc_mat_11_7, scale_2367_7);
6336
6337                    __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)));
6338                    __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)));
6339                    __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)));
6340                    __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)));
6341
6342                    // Straighten out to make 4 row vectors
6343                    __m256i iacc_row_0 = _mm256_blend_epi32(iacc_mat_00, _mm256_shuffle_epi32(iacc_mat_01, 78), 204);
6344                    __m256i iacc_row_1 = _mm256_blend_epi32(_mm256_shuffle_epi32(iacc_mat_00, 78), iacc_mat_01, 204);
6345                    __m256i iacc_row_2 = _mm256_blend_epi32(iacc_mat_10, _mm256_shuffle_epi32(iacc_mat_11, 78), 204);
6346                    __m256i iacc_row_3 = _mm256_blend_epi32(_mm256_shuffle_epi32(iacc_mat_10, 78), iacc_mat_11, 204);
6347
6348                    // Load the scale(d) values for all the 4 Q8_k blocks and repeat it across lanes
6349                    const __m128 row_scale_f32_sse = _mm_load_ps(a_ptr[b].d);
6350                    const __m256 row_scale_f32 = _mm256_set_m128(row_scale_f32_sse, row_scale_f32_sse);
6351
6352                    // Multiply with appropriate scales and accumulate (for both d and dmin) below
6353                    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]);
6354                    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]);
6355                    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]);
6356                    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]);
6357
6358                    __m256i lhs_bsums_01_0123 = _mm256_inserti128_si256(_mm256_castsi128_si256(lhs_raw_bsums_01_0123), lhs_raw_bsums_01_0123, 1);
6359                    __m256i lhs_bsums_23_0123 = _mm256_inserti128_si256(_mm256_castsi128_si256(lhs_raw_bsums_23_0123), lhs_raw_bsums_23_0123, 1);
6360                    __m256i lhs_bsums_01_4567 = _mm256_inserti128_si256(_mm256_castsi128_si256(lhs_raw_bsums_01_4567), lhs_raw_bsums_01_4567, 1);
6361                    __m256i lhs_bsums_23_4567 = _mm256_inserti128_si256(_mm256_castsi128_si256(lhs_raw_bsums_23_4567), lhs_raw_bsums_23_4567, 1);
6362
6363                    // Take two bsums from two Q8_Ks at a time and multiply with corresponding mins values from each Q2_K
6364                    __m256i iacc_row_min_0_01 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_01_0123, 0), mins_01);
6365                    __m256i iacc_row_min_1_01 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_01_0123, 170), mins_01);
6366                    __m256i iacc_row_min_2_01 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_23_0123, 0), mins_01);
6367                    __m256i iacc_row_min_3_01 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_23_0123, 170), mins_01);
6368
6369                    __m256i iacc_row_min_0_23 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_01_0123, 85), mins_23);
6370                    __m256i iacc_row_min_1_23 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_01_0123, 255), mins_23);
6371                    __m256i iacc_row_min_2_23 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_23_0123, 85), mins_23);
6372                    __m256i iacc_row_min_3_23 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_23_0123, 255), mins_23);
6373
6374                    __m256i iacc_row_min_0_45 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_01_4567, 0), mins_45);
6375                    __m256i iacc_row_min_1_45 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_01_4567, 170), mins_45);
6376                    __m256i iacc_row_min_2_45 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_23_4567, 0), mins_45);
6377                    __m256i iacc_row_min_3_45 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_23_4567, 170), mins_45);
6378
6379                    __m256i iacc_row_min_0_67 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_01_4567, 85), mins_67);
6380                    __m256i iacc_row_min_1_67 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_01_4567, 255), mins_67);
6381                    __m256i iacc_row_min_2_67 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_23_4567, 85), mins_67);
6382                    __m256i iacc_row_min_3_67 = _mm256_madd_epi16(_mm256_shuffle_epi32(lhs_bsums_23_4567, 255), mins_67);
6383
6384                    __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));
6385                    __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));
6386                    __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));
6387                    __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));
6388
6389                    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]);
6390                    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]);
6391                    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]);
6392                    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]);
6393                }
6394            }
6395            // Store the accumulated values
6396            for (int i = 0; i < 4; i++) {
6397                _mm256_storeu_ps((float * )(s + ((y * 4 + i) * bs + x * 8)), _mm256_sub_ps(acc_rows[i], acc_min_rows[i]));
6398            }
6399        }
6400    }
6401#else
6402
6403    ggml_gemm_q2_K_8x8_q8_K_generic(n, s, bs, vx, vy, nr, nc);
6404
6405
6406#endif
6407}