File size: 35,842 Bytes
dffe123
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
// This file is part of OpenCV project.
// It is subject to the license terms in the LICENSE file found in the top-level directory
// of this distribution and at http://opencv.org/license.html.

// This file is modified from the ficus (https://github.com/vpisarev/ficus/blob/master/lib/NN/OpConv_Winograd.fx).
// Here is the original license:
/*
    This file is a part of ficus language project.
    See ficus/LICENSE for the licensing terms
*/

#include "../../precomp.hpp"
#include "convolution.hpp"

#include "conv_winograd_f63.simd.hpp"
#include "layers/cpu_kernels/conv_winograd_f63.simd_declarations.hpp" // defines CV_CPU_DISPATCH_MODES_ALL=AVX2,...,BASELINE based on CMakeLists.txt content

namespace cv { namespace dnn {

#if CV_NEON || CV_SIMD128 || CV_TRY_AVX2
enum { VEC_ALIGN = 32, DFT_TYPE = CV_32F }; // Memory alignment.

void winofunc_accum_F32(const float* inwptr, const float* wptr, float* outbuf, int Cg, int iblock,
                            const int winoIblock, const int winoKblock, const int winoAtomF32, const int winoNatomF32);

/*Input transform*/
void winofunc_BtXB_8x8_F32(const float* inptr, int inpstep,
                          float* outptr, int Cg, const int winoIblock, const int winoAtomF32);

/*Output transform*/
void winofunc_AtXA_8x8_F32(const float* inptr, int inpstep, float* bpptr, int bpstep, float* outptr, int outstep,
                          float bias, float minval, float maxval, bool ifMinMaxAct);

int runWinograd63(InputArray _input, InputArray _fusedAddMat, OutputArray _output, const Ptr<FastConv>& conv,
                  int ntasks, float minval, float maxval, ActivationLayer* activ, bool ifMinMaxAct)
{
    Mat input = _input.getMat();
    Mat output = _output.getMat();
    Mat fusedAddMat = _fusedAddMat.getMat();

    MatShape inputShape = shape(input);
    MatShape outputShape = shape(output);
    CV_Assert(inputShape.size() == 4 && outputShape.size() == 4);

    int N = inputShape[0], C = inputShape[1], Hi = inputShape[2], Wi = inputShape[3];  // [N, C, H, W]
    int K = conv->K;
    int H0 = outputShape[2], W0 = outputShape[3];

    int pad_top = conv->pad_top;
    int pad_left = conv->pad_left;

    int ngroups = conv->ngroups, Cg = C/ngroups, Kg = K/ngroups;

    const int CONV_WINO_KBLOCK = 4;
#if (CV_NEON && CV_NEON_AARCH64)
    const int CONV_WINO_IBLOCK = 6;
#elif  CV_TRY_AVX || CV_TRY_AVX2
    const int CONV_WINO_IBLOCK = (conv->useAVX || conv->useAVX2) ? 6 : 3;
#else
    const int CONV_WINO_IBLOCK = 3;
#endif

#if CV_TRY_AVX || CV_TRY_AVX2
    const int CONV_WINO_ATOM_F32 = (conv->useAVX || conv->useAVX2) ? 8 : 4;
#else
    const int CONV_WINO_ATOM_F32 = 4;
#endif
    const int CONV_WINO_NATOMS_F32 = CONV_WINO_AREA / CONV_WINO_ATOM_F32; // for AVX2, it is 8, otherwise, it's 16.

    int CONV_WINO_ATOM = CONV_WINO_ATOM_F32;
    int CONV_WINO_NATOMS = CONV_WINO_NATOMS_F32;

#ifdef CONV_ARM_FP16
    // FP 16
    const int CONV_WINO_ATOM_F16 = CONV_WINO_ATOM_F32 * 2;
    const int CONV_WINO_NATOMS_F16 = CONV_WINO_AREA / CONV_WINO_ATOM_F16;
#endif

    int esz = sizeof(float );

#ifdef CONV_ARM_FP16
    const bool useFP16 = conv->useFP16;
    if (useFP16)
    {
        // works at FP 16.
        CONV_WINO_ATOM = CONV_WINO_ATOM_F16;
        CONV_WINO_NATOMS = CONV_WINO_NATOMS_F16;
        esz = sizeof(__fp16);
    }
#endif

    int Kg_nblocks = (Kg + CONV_WINO_KBLOCK - 1)/CONV_WINO_KBLOCK;
    const size_t inp_planesize = (size_t)Hi*Wi;
    const size_t out_planesize = (size_t)H0*W0;

    int blocks_per_row = (W0+CONV_WINO_STEP-1)/CONV_WINO_STEP;
    int blocks_per_plane = ((H0+CONV_WINO_STEP-1)/CONV_WINO_STEP)*blocks_per_row;
    int blocks_per_plane_aligned = ((blocks_per_plane +
                                     CONV_WINO_IBLOCK-1)/CONV_WINO_IBLOCK)*CONV_WINO_IBLOCK;

    size_t totalbufsize = (size_t)N*C*blocks_per_plane_aligned*CONV_WINO_AREA;

    AutoBuffer<char> _buf;
    _buf.allocate((totalbufsize + VEC_ALIGN) * esz);
    char* wbuf_all = alignPtr(_buf.data(), VEC_ALIGN * esz);

    float* inp = input.ptr<float>();
    float* out = output.ptr<float>();

    float* fusedAddPtr = fusedAddMat.empty() ? nullptr : fusedAddMat.ptr<float>();

    // Phase 1. compute forward Winograd transforms for all input blocks,
    // all input planes, all samples in the batch.
    // [TODO]: maybe, if there are too many input channels, it makes sense to
    // transform only part of input channels at once and then compute the partial
    // accumulated sums (i.e. update the output buffers several times,
    // rather than compute them in one pass).
    parallel_for_(Range(0, ntasks), [&](const Range& r0) {
    for (int task_id = r0.start; task_id < r0.end; task_id++)
    {
        int nc0 = (N*C)*task_id/ntasks;
        int nc1 = (N*C)*(task_id+1)/ntasks;
        for(; nc0 < nc1; nc0++)
        {
            int n = nc0 / C;
            int c = nc0 - n*C;
            int g = c / Cg;
            c -= g*Cg;

            for (int block_id = 0; block_id < blocks_per_plane; block_id += CONV_WINO_IBLOCK)
            {
                for (int db = 0; db < CONV_WINO_IBLOCK; db++)
                {
                    size_t inwofs = ((n*ngroups + g)*blocks_per_plane_aligned +
                                     block_id)*Cg*CONV_WINO_AREA +
                                    (c*CONV_WINO_IBLOCK + db) * CONV_WINO_ATOM;
                    char* inwptr = wbuf_all + inwofs * esz;

                    if (block_id + db < blocks_per_plane)
                    {
                        int y0 = (block_id + db) / blocks_per_row;
                        int x0 = (block_id + db) - y0 * blocks_per_row;
                        y0 = y0*CONV_WINO_STEP - pad_top;
                        x0 = x0*CONV_WINO_STEP - pad_left;
                        bool partial = y0 < 0 || y0 + CONV_WINO_SIZE > Hi ||
                                       x0 < 0 || x0 + CONV_WINO_SIZE > Wi;
                        int dx1 = 0, dx2 = CONV_WINO_SIZE, dy1 = 0, dy2 = CONV_WINO_SIZE;
                        int inpstep = Wi;

                        float inpbuf[CONV_WINO_AREA];
                        float* inptr0 = (float*)inp + nc0*inp_planesize + y0*Wi + x0;
                        float* inptr = inptr0;

                        if (partial)
                        {
                            memset(inpbuf, 0, sizeof(inpbuf));
                            dy1 = -y0 > 0 ? -y0 : 0;
                            dy2 = Hi - y0 < CONV_WINO_SIZE ? Hi - y0 : CONV_WINO_SIZE;

                            if (dy2 < dy1) {dy2 = dy1 = 0;}
                            dx1 = -x0 > 0 ? -x0 : 0;
                            dx2 = Wi - x0 < CONV_WINO_SIZE ? Wi - x0 : CONV_WINO_SIZE;

                            if (dx2 < dx1) {dx2 = dx1 = 0;}
                            inptr0 -= y0*Wi + x0;

                            if (dx1 < dx2 && dy1 < dy2)
                            {
                                for(int dy = dy1; dy < dy2; dy++)
                                    memcpy(&inpbuf[dy*CONV_WINO_SIZE + dx1],
                                           inptr0 + (y0+dy)*Wi + (x0+dx1),
                                           (dx2-dx1)*sizeof(inpbuf[0]));
                            }

                            inptr = inpbuf;
                            inpstep = CONV_WINO_SIZE;
                        }

#if CV_TRY_AVX2
                        if (conv->useAVX2)
                            opt_AVX2::winofunc_BtXB_8x8_F32(inptr, inpstep, (float *)inwptr, Cg, CONV_WINO_IBLOCK, CONV_WINO_ATOM);
                        else
#endif
#if CV_TRY_AVX
                        if (conv->useAVX)
                            opt_AVX::winofunc_BtXB_8x8_F32(inptr, inpstep, (float *)inwptr, Cg, CONV_WINO_IBLOCK, CONV_WINO_ATOM);
                        else
#endif
#if CV_NEON && CV_NEON_AARCH64
                        if (conv->useNEON)
                        {
#ifdef CONV_ARM_FP16
                            if (useFP16)
                            {
                                opt_NEON_FP16::winofunc_BtXB_8x8_F16(inptr, inpstep, inwptr, Cg, CONV_WINO_IBLOCK,
                                                                CONV_WINO_ATOM);
                            }
                            else
#endif
                            opt_NEON::winofunc_BtXB_8x8_F32(inptr, inpstep, (float *)inwptr, Cg, CONV_WINO_IBLOCK,
                                                            CONV_WINO_ATOM);
                        }
                        else
#endif
                        winofunc_BtXB_8x8_F32(inptr, inpstep, (float *)inwptr, Cg, CONV_WINO_IBLOCK, CONV_WINO_ATOM);

                    }
                    else
                    {
                        for (int i = 0; i < CONV_WINO_NATOMS; i++, inwptr += CONV_WINO_IBLOCK * CONV_WINO_ATOM * esz)
                            memset(inwptr, 0, CONV_WINO_ATOM * esz);
                    }
                }
            }
        }
    }});

    // Phase 2. compute elemwise-weighted sums of transformed blocks,
    // apply inverse Winograd transforms to the sums,
    // add bias, apply activation function if any and store the results.
    char* wptr0 = nullptr;
#ifdef CONV_ARM_FP16
    if (useFP16)
    {
        CV_Assert(!conv->weightsWinoBuf_FP16.empty());
        wptr0 = (char *)conv->getWeightsWinoFP16();
    }
    else
#endif
    {
        CV_Assert(!conv->weightsWinoBuf.empty());
        wptr0 = (char *)conv->getWeightsWino();
    }

    parallel_for_(Range(0, ntasks), [&](const Range& r0) {
    for (int task_id = r0.start; task_id < r0.end; task_id++)
    {
        size_t out_wbuf_size = CONV_WINO_AREA * CONV_WINO_KBLOCK * CONV_WINO_IBLOCK;
        size_t outbuf_size = CONV_WINO_AREA;

        // For saving the accumulation output.
        AutoBuffer<char> out_wbuf_;
        out_wbuf_.allocate((out_wbuf_size + VEC_ALIGN) * esz);
        char* out_wbuf = alignPtr(out_wbuf_.data(), VEC_ALIGN * esz);
        memset(out_wbuf, 0, out_wbuf_size * esz);

        // For saving the fuse_Add data.
        AutoBuffer<float> outbuf_;
        outbuf_.allocate(outbuf_size + VEC_ALIGN);
        float* outbuf = alignPtr(outbuf_.data(), VEC_ALIGN);
        memset(outbuf, 0, outbuf_size * sizeof(outbuf[0]));

        int ngk0 = (int)(((int64_t)N*Kg_nblocks*ngroups)*task_id/ntasks);
        int ngk1 = (int)(((int64_t)N*Kg_nblocks*ngroups)*(task_id+1)/ntasks);

        for(; ngk0 < ngk1; ngk0++)
        {
            int n = ngk0 / (Kg_nblocks*ngroups);
            int gk0 = ngk0 % (Kg_nblocks*ngroups);
            int g = gk0 / Kg_nblocks;
            int k0 = (gk0 % Kg_nblocks)*CONV_WINO_KBLOCK;
            int k1 = k0 + CONV_WINO_KBLOCK <= Kg ? k0 + CONV_WINO_KBLOCK : Kg;

            for (int block_id0 = 0; block_id0 < blocks_per_plane; block_id0 += CONV_WINO_IBLOCK)
            {
                int block_id1 = block_id0 + CONV_WINO_IBLOCK;
                block_id1 = block_id1 < blocks_per_plane ? block_id1 : blocks_per_plane;
                size_t inwofs = ((n*ngroups + g)*blocks_per_plane_aligned + block_id0)*Cg*CONV_WINO_AREA;
                size_t wofs = (g*Kg_nblocks*CONV_WINO_KBLOCK + k0)*Cg*CONV_WINO_AREA;

                char* inwptr = wbuf_all + inwofs * esz;
                char* wptr = wptr0 + wofs * esz;

#if CV_TRY_AVX2
                if (conv->useAVX2)
                    opt_AVX2::winofunc_accum_F32((float *)inwptr, (float *)wptr, (float *)out_wbuf, Cg, block_id1 - block_id0, CONV_WINO_IBLOCK,
                                       CONV_WINO_KBLOCK, CONV_WINO_ATOM, CONV_WINO_NATOMS);
                else
#endif
#if CV_TRY_AVX
                if (conv->useAVX)
                    opt_AVX::winofunc_accum_F32((float *)inwptr, (float *)wptr, (float *)out_wbuf, Cg, block_id1 - block_id0, CONV_WINO_IBLOCK,
                                       CONV_WINO_KBLOCK, CONV_WINO_ATOM, CONV_WINO_NATOMS);
                else
#endif
#if CV_NEON && CV_NEON_AARCH64
                if (conv->useNEON)
                {
#ifdef CONV_ARM_FP16
                    if (useFP16)
                    {
                        opt_NEON_FP16::winofunc_accum_F16(inwptr, wptr, out_wbuf, Cg, block_id1 - block_id0, CONV_WINO_IBLOCK,
                                                     CONV_WINO_KBLOCK, CONV_WINO_ATOM, CONV_WINO_NATOMS);
                    }
                    else
#endif
                    opt_NEON::winofunc_accum_F32((float *)inwptr, (float *)wptr, (float *)out_wbuf, Cg, block_id1 - block_id0, CONV_WINO_IBLOCK,
                                                 CONV_WINO_KBLOCK, CONV_WINO_ATOM, CONV_WINO_NATOMS);
                }
                else
#endif
                winofunc_accum_F32((float *)inwptr, (float *)wptr, (float *)out_wbuf, Cg, block_id1 - block_id0, CONV_WINO_IBLOCK,
                                       CONV_WINO_KBLOCK, CONV_WINO_ATOM, CONV_WINO_NATOMS);

                for (int k = k0; k < k1; k++)
                {
                    float biasv = conv->biasBuf[g*Kg + k];
                    for (int block_id = block_id0; block_id < block_id1; block_id++)
                    {
                        int y0 = block_id / blocks_per_row;
                        int x0 = block_id - y0 * blocks_per_row;
                        y0 = y0*CONV_WINO_STEP;
                        x0 = x0*CONV_WINO_STEP;
                        int dy1 = H0 - y0;
                        if (dy1 > CONV_WINO_STEP) dy1 = CONV_WINO_STEP;
                        int dx1 = W0 - x0;
                        if (dx1 > CONV_WINO_STEP) dx1 = CONV_WINO_STEP;
                        assert(dx1 > 0 && dy1 > 0);
                        bool partial = activ || dy1 < CONV_WINO_STEP || dx1 < CONV_WINO_STEP;
                        size_t outofs = (n*K + g*Kg + k)*out_planesize + y0*W0 + x0;
                        int outstep = W0;

                        float* outptr0 = (float*)out + outofs;
                        float* pbptr0 = fusedAddPtr ? fusedAddPtr + outofs : nullptr;
                        float *outptr = outptr0, *bpptr = pbptr0;

                        if (partial)
                        {
                            outptr = outbuf;
                            outstep = CONV_WINO_SIZE;
                            if (pbptr0)
                            {
                                bpptr = outbuf;
                                for (int y = 0; y < dy1; y++)
                                    memcpy(outbuf + y*CONV_WINO_SIZE, pbptr0 + y*W0,
                                           dx1*sizeof(pbptr0[0]));
                            }
                        }
#if CV_TRY_AVX2
                        if (conv->useAVX2)
                            opt_AVX2::winofunc_AtXA_8x8_F32((float *)out_wbuf + ((k - k0)*CONV_WINO_IBLOCK + (block_id - block_id0))*CONV_WINO_AREA, CONV_WINO_SIZE,
                                                                bpptr, outstep, outptr, outstep, biasv, minval, maxval, ifMinMaxAct);
                        else
#endif
#if CV_TRY_AVX
                        if (conv->useAVX)
                            opt_AVX::winofunc_AtXA_8x8_F32((float *)out_wbuf + ((k - k0)*CONV_WINO_IBLOCK + (block_id - block_id0))*CONV_WINO_AREA, CONV_WINO_SIZE,
                                                                bpptr, outstep, outptr, outstep, biasv, minval, maxval, ifMinMaxAct);
                        else
#endif
#if CV_NEON && CV_NEON_AARCH64
                        // NEON optimization is only for ARMv8 device, and for ARMv7 device, we use the Universal intrinsics.
                        if (conv->useNEON)
                        {
#ifdef CONV_ARM_FP16
                            if (useFP16)
                            {
                                opt_NEON_FP16::winofunc_AtXA_8x8_F16(out_wbuf + ((k - k0)*CONV_WINO_IBLOCK + (block_id - block_id0))*CONV_WINO_AREA * esz, CONV_WINO_SIZE,
                                                                bpptr, outstep, outptr, outstep, biasv, minval, maxval, ifMinMaxAct);
                            }
                            else
#endif
                            opt_NEON::winofunc_AtXA_8x8_F32((float *)out_wbuf + ((k - k0)*CONV_WINO_IBLOCK + (block_id - block_id0))*CONV_WINO_AREA, CONV_WINO_SIZE,
                                                            bpptr, outstep, outptr, outstep, biasv, minval, maxval, ifMinMaxAct);
                        }
                        else
#endif
                        winofunc_AtXA_8x8_F32((float *)out_wbuf + ((k - k0)*CONV_WINO_IBLOCK + (block_id - block_id0))*CONV_WINO_AREA, CONV_WINO_SIZE,
                                                  bpptr, outstep, outptr, outstep, biasv, minval, maxval, ifMinMaxAct);

                        if (partial)
                        {
                            if (activ)
                                activ->forwardSlice(outptr, outptr, CONV_WINO_SIZE*CONV_WINO_STEP, 0, g*Kg + k, g*Kg + k + 1);
                            for (int y = 0; y < dy1; y++)
                                memcpy(outptr0 + y*W0, outptr + y*CONV_WINO_SIZE, dx1*sizeof(outptr0[0]));
                        }
                    }
                }
            }
        }
    }});
    return 1;
}

/****************************************************************************************\
                                    SIMD for winograd function
\****************************************************************************************/

#if CV_SIMD128

void winofunc_accum_F32(const float* inwptr, const float* wptr, float* outbuf, int Cg, int iblock,
                            const int winoIblock, const int winoKblock, const int winoAtomF32, const int winoNatomF32)
{
#if 1
    CV_Assert(winoIblock == 3 && winoKblock == 4 && winoAtomF32 == 4);
    for (int atom_id = 0; atom_id < winoNatomF32; atom_id++,
            outbuf += winoAtomF32)
    {
        v_float32x4 s00 = v_setzero_f32(), s01 = s00, s02 = s00;
        v_float32x4 s10 = v_setzero_f32(), s11 = s00, s12 = s00;
        v_float32x4 s20 = v_setzero_f32(), s21 = s00, s22 = s00;
        v_float32x4 s30 = v_setzero_f32(), s31 = s00, s32 = s00;

        for (int c = 0; c < Cg; c++, inwptr += winoIblock*winoAtomF32,
                                     wptr += winoKblock*winoAtomF32)
        {
            v_float32x4 x0, x1, x2;
            x0 = v_load(inwptr);
            x1 = v_load(inwptr + 4);
            x2 = v_load(inwptr + 8);

            v_float32x4 w0 = v_load(wptr);
            s00 = v_fma(w0, x0, s00);
            s01 = v_fma(w0, x1, s01);
            s02 = v_fma(w0, x2, s02);

            w0 = v_load(wptr + 4);
            s10 = v_fma(w0, x0, s10);
            s11 = v_fma(w0, x1, s11);
            s12 = v_fma(w0, x2, s12);

            w0 = v_load(wptr + 8);
            s20 = v_fma(w0, x0, s20);
            s21 = v_fma(w0, x1, s21);
            s22 = v_fma(w0, x2, s22);

            w0 = v_load(wptr + 12);
            s30 = v_fma(w0, x0, s30);
            s31 = v_fma(w0, x1, s31);
            s32 = v_fma(w0, x2, s32);
        }

        v_store(outbuf, s00);
        v_store(outbuf + 1*64, s01);
        v_store(outbuf + 2*64, s02);
        v_store(outbuf + 3*64, s10);
        v_store(outbuf + 4*64, s11);
        v_store(outbuf + 5*64, s12);
        v_store(outbuf + 6*64, s20);
        v_store(outbuf + 7*64, s21);
        v_store(outbuf + 8*64, s22);
        v_store(outbuf + 9*64, s30);
        v_store(outbuf + 10*64, s31);
        v_store(outbuf + 11*64, s32);
    }
#else
    // Naive C++ code, the code should never be run here.
    for (int atom_id = 0; atom_id < winoNatomF32;
                atom_id++, outbuf += winoAtomF32)
    {
        float sumbuf[winoIblock*winoKblock*winoAtomF32];
        memset(sumbuf, 0, sizeof(sumbuf));
        for (int c = 0; c < Cg; c++, inwptr += winoIblock*winoAtomF32,
                                     wptr += winoKblock*winoAtomF32)
        {
            for (int i = 0; i < winoKblock; i++)
            {
                for (int j = 0; j < winoIblock; j++)
                {
                    int i_ = i*winoAtomF32;
                    int j_ = j*winoAtomF32;
                    int ij_ = i_*winoIblock + j_;
                    float s0 = inwptr[j_ + 0]*wptr[i_ + 0];
                    float s1 = inwptr[j_ + 1]*wptr[i_ + 1];
                    float s2 = inwptr[j_ + 2]*wptr[i_ + 2];
                    float s3 = inwptr[j_ + 3]*wptr[i_ + 3];
                    sumbuf[ij_ + 0] += s0;
                    sumbuf[ij_ + 1] += s1;
                    sumbuf[ij_ + 2] += s2;
                    sumbuf[ij_ + 3] += s3;
                }
            }
        }
        for (int ij = 0; ij < winoKblock*winoIblock; ij++)
        {
            int ij_ = ij*winoAtomF32;
            int ij_out = ij*CONV_WINO_AREA;
            outbuf[ij_out + 0] = sumbuf[ij_ + 0];
            outbuf[ij_out + 1] = sumbuf[ij_ + 1];
            outbuf[ij_out + 2] = sumbuf[ij_ + 2];
            outbuf[ij_out + 3] = sumbuf[ij_ + 3];
        }
    }
#endif
}

/*Input transform*/
void winofunc_BtXB_8x8_F32(const float* inptr, int inpstep,
                          float* outptr, int Cg, const int winoIblock, const int winoAtomF32)
{
    CV_Assert(winoIblock == 3 && winoAtomF32 == 4);
    v_float32x4 x00 = v_load(inptr), x01 = v_load(inptr + 4);
    v_float32x4 x10 = v_load(inptr + inpstep), x11 = v_load(inptr + inpstep + 4);
    v_float32x4 x20 = v_load(inptr + inpstep*2), x21 = v_load(inptr + inpstep*2 + 4);
    v_float32x4 x30 = v_load(inptr + inpstep*3), x31 = v_load(inptr + inpstep*3 + 4);
    v_float32x4 x40 = v_load(inptr + inpstep*4), x41 = v_load(inptr + inpstep*4 + 4);
    v_float32x4 x50 = v_load(inptr + inpstep*5), x51 = v_load(inptr + inpstep*5 + 4);
    v_float32x4 x60 = v_load(inptr + inpstep*6), x61 = v_load(inptr + inpstep*6 + 4);
    v_float32x4 x70 = v_load(inptr + inpstep*7), x71 = v_load(inptr + inpstep*7 + 4);

    v_float32x4 z00, z01, z10, z11, z20, z21, z30, z31, z40, z41, z50, z51, z60, z61, z70, z71;

    {
        /* Y[0] = [1.f, 0.f, -5.25f, 0.f, 5.25f, 0.f, -1.f, 0.f]*X */
        /* Y[7] = [0.f, -1.f, 0.f, 5.25f, 0.f, -5.25f, 0.f, 1.f]*X */
        v_float32x4 q5_25 = v_setall_f32(5.25f), t00, t01, t10, t11;
        t00 = v_sub(x40, x20);
        t01 = v_sub(x41, x21);
        t10 = v_sub(x30, x50);
        t11 = v_sub(x31, x51);
        v_float32x4 y00 = v_fma(t00, q5_25, v_sub(x00, x60));
        v_float32x4 y01 = v_fma(t01, q5_25, v_sub(x01, x61));
        v_float32x4 y70 = v_fma(t10, q5_25, v_sub(x70, x10));
        v_float32x4 y71 = v_fma(t11, q5_25, v_sub(x71, x11));

        /* Y[1] = [0.f, 1.f, 1.f, -4.25f, -4.25f, 1.f, 1.f, 0.f]*X */
        /* Y[2] = [0.f, -1.f, 1.f, 4.25f, -4.25f, -1.f, 1.f, 0.f]*X */
        v_float32x4 qm4_25 = v_setall_f32(-4.25f);
        t00 = v_fma(x30, qm4_25, v_add(x10, x50));
        t01 = v_fma(x31, qm4_25, v_add(x11, x51));
        t10 = v_fma(x40, qm4_25, v_add(x20, x60));
        t11 = v_fma(x41, qm4_25, v_add(x21, x61));

        v_float32x4 y10 = v_add(t00, t10), y11 = v_add(t01, t11);
        v_float32x4 y20 = v_sub(t10, t00), y21 = v_sub(t11, t01);

        /* Y[3] = [0.f, 0.5f, 0.25f, -2.5f, -1.25f, 2.f, 1.f, 0.f]*X */
        /* Y[4] = [0.f, -0.5f, 0.25f, 2.5f, -1.25f, -2.f, 1.f, 0.f]*X */
        v_float32x4 q0_5 = v_setall_f32(0.5f), q0_25 = v_setall_f32(0.25f);
        v_float32x4 qm2_5 = v_setall_f32(-2.5f), qm1_25 = v_setall_f32(-1.25f);
        t00 = v_fma(x10, q0_5, v_add(x50, x50));
        t01 = v_fma(x11, q0_5, v_add(x51, x51));
        t10 = v_fma(x20, q0_25, x60);
        t11 = v_fma(x21, q0_25, x61);
        t00 = v_fma(x30, qm2_5, t00);
        t01 = v_fma(x31, qm2_5, t01);
        t10 = v_fma(x40, qm1_25, t10);
        t11 = v_fma(x41, qm1_25, t11);

        v_float32x4 y30 = v_add(t00, t10), y31 = v_add(t01, t11);
        v_float32x4 y40 = v_sub(t10, t00), y41 = v_sub(t11, t01);

        /* Y[5] = [0.f, 2.f, 4.f, -2.5f, -5.f, 0.5f, 1.f, 0.f]*X */
        /* Y[6] = [0.f, -2.f, 4.f, 2.5f, -5.f, -0.5f, 1.f, 0.f]*X */
        v_float32x4 q4 = v_setall_f32(4.f), qm5 = v_setall_f32(-5.f);
        t00 = v_fma(x50, q0_5, v_add(x10, x10));
        t01 = v_fma(x51, q0_5, v_add(x11, x11));
        t10 = v_fma(x20, q4   , x60);
        t11 = v_fma(x21, q4   , x61);
        t00 = v_fma(x30, qm2_5, t00);
        t01 = v_fma(x31, qm2_5, t01);
        t10 = v_fma(x40, qm5  , t10);
        t11 = v_fma(x41, qm5  , t11);

        v_float32x4 y50 = v_add(t00, t10), y51 = v_add(t01, t11);
        v_float32x4 y60 = v_sub(t10, t00), y61 = v_sub(t11, t01);

        /* transpose 8x8 matrix with v_transpose4x4 */

        v_float32x4 y000, y100, y200, y300, y010, y110, y210, y310, y400, y500, y600, y700, y410, y510, y610, y710;
        v_transpose4x4(y00, y10, y20, y30, y000, y100, y200, y300);
        v_transpose4x4(y01, y11, y21, y31, y010, y110, y210, y310);
        v_transpose4x4(y40, y50, y60, y70, y400, y500, y600, y700);
        v_transpose4x4(y41, y51, y61, y71, y410, y510, y610, y710);

        /* Z[0] = [1.f, 0.f, -5.25f, 0.f, 5.25f, 0.f, -1.f, 0.f]*Y */
        /* Z[7] = [0.f, -1.f, 0.f, 5.25f, 0.f, -5.25f, 0.f, 1.f]*Y */
        t00 = v_sub(y010, y200);
        t01 = v_sub(y410, y600);
        t10 = v_sub(y300, y110);
        t11 = v_sub(y700, y510);
        z00 = v_fma(t00, q5_25, v_sub(y000, y210));
        z01 = v_fma(t01, q5_25, v_sub(y400, y610));
        z70 = v_fma(t10, q5_25, v_sub(y310, y100));
        z71 = v_fma(t11, q5_25, v_sub(y710, y500));

        /* Z[1] = [0.f, 1.f, 1.f, -4.25f, -4.25f, 1.f, 1.f, 0.f]*Y */
        /* Z[2] = [0.f, -1.f, 1.f, 4.25f, -4.25f, -1.f, 1.f, 0.f]*Y */
        t00 = v_fma(y300, qm4_25, v_add(y100, y110));
        t01 = v_fma(y700, qm4_25, v_add(y500, y510));
        t10 = v_fma(y010, qm4_25, v_add(y200, y210));
        t11 = v_fma(y410, qm4_25, v_add(y600, y610));

        z10 = v_add(t00, t10); z11 = v_add(t01, t11);
        z20 = v_sub(t10, t00); z21 = v_sub(t11, t01);

        /* Z[3] = [0.f, 0.5f, 0.25f, -2.5f, -1.25f, 2.f, 1.f, 0.f]*Y */
        /* Z[4] = [0.f, -0.5f, 0.25f, 2.5f, -1.25f, -2.f, 1.f, 0.f]*Y */
        t00 = v_fma(y100, q0_5, v_add(y110, y110));
        t01 = v_fma(y500, q0_5, v_add(y510, y510));
        t10 = v_fma(y200, q0_25, y210);
        t11 = v_fma(y600, q0_25, y610);
        t00 = v_fma(y300, qm2_5, t00);
        t01 = v_fma(y700, qm2_5, t01);
        t10 = v_fma(y010, qm1_25, t10);
        t11 = v_fma(y410, qm1_25, t11);

        z30 = v_add(t00, t10); z31 = v_add(t01, t11);
        z40 = v_sub(t10, t00); z41 = v_sub(t11, t01);

        /* Z[5] = [0.f, 2.f, 4.f, -2.5f, -5.f, 0.5f, 1.f, 0.f]*Y */
        /* Z[6] = [0.f, -2.f, 4.f, 2.5f, -5.f, -0.5f, 1.f, 0.f]*Y */
        t00 = v_fma(y110, q0_5, v_add(y100, y100));
        t01 = v_fma(y510, q0_5, v_add(y500, y500));
        t10 = v_fma(y200, q4, y210);
        t11 = v_fma(y600, q4, y610);
        t00 = v_fma(y300, qm2_5, t00);
        t01 = v_fma(y700, qm2_5, t01);
        t10 = v_fma(y010, qm5, t10);
        t11 = v_fma(y410, qm5, t11);

        z50 = v_add(t00, t10); z51 = v_add(t01, t11);
        z60 = v_sub(t10, t00); z61 = v_sub(t11, t01);
    }

    const int outstep = winoIblock*winoAtomF32*Cg;

    v_store(outptr, z00);
    v_store(outptr + outstep, z01);
    v_store(outptr + outstep*2, z10);
    v_store(outptr + outstep*3, z11);
    v_store(outptr + outstep*4, z20);
    v_store(outptr + outstep*5, z21);
    v_store(outptr + outstep*6, z30);
    v_store(outptr + outstep*7, z31);
    v_store(outptr + outstep*8, z40);
    v_store(outptr + outstep*9, z41);
    v_store(outptr + outstep*10, z50);
    v_store(outptr + outstep*11, z51);
    v_store(outptr + outstep*12, z60);
    v_store(outptr + outstep*13, z61);
    v_store(outptr + outstep*14, z70);
    v_store(outptr + outstep*15, z71);
}

/*Output transform*/
/*  Inverse Winograd 8x8 transform:
    out = (A'*inp*A)', where
    inp is input 8x8 FP32 matrix,
    A' is
    [1.f, 1.f, 1.f, 1.f, 1.f, 1.f, 1.f, 0.f,
     0.f, 1.f, -1.f, 2.f, -2.f, 0.5f, -0.5f, 0.f,
     0.f, 1.f, 1.f, 4.f, 4.f, 0.25f, 0.25f, 0.f,
     0.f, 1.f, -1.f, 8.f, -8.f, 0.125f, -0.125f, 0.f,
     0.f, 1.f, 1.f, 16.f, 16.f, 1.f/16, 1.f/16, 0.f,
     0.f, 1.f, -1.f, 32.f, -32.f, 1.f/32, -1.f/32, 1.f]

    inp is pre-loaded into xij registers,
    out will be stored in zij, where (0<=i<=7 for x, 0<=i<=5 for z), 0<=j<=1.

    After the inverse transform is done, we add bias,
    optionally add results from the earlier tensors (by-pass),
    optionally apply activation function and then
    store the final results.

    That is, after both forward and then inverse transformation,
    we get non-transposed result.
    Of course, for the correct work of Winograd-based convolution,
    the Winograd-transformed weights should also be transposed.
    init_conv() (see OpConv.fx) takes care of that.
*/
void winofunc_AtXA_8x8_F32(const float* inptr, int inpstep,
                          float* bpptr, int bpstep, float* outptr, int outstep,
                          float bias, float minval, float maxval, bool ifMinMaxAct)
{
    v_float32x4 x00 = v_load(inptr), x01 = v_load(inptr + 4);
    v_float32x4 x10 = v_load(inptr + inpstep), x11 = v_load(inptr + inpstep + 4);
    v_float32x4 x20 = v_load(inptr + inpstep*2), x21 = v_load(inptr + inpstep*2 + 4);
    v_float32x4 x30 = v_load(inptr + inpstep*3), x31 = v_load(inptr + inpstep*3 + 4);
    v_float32x4 x40 = v_load(inptr + inpstep*4), x41 = v_load(inptr + inpstep*4 + 4);
    v_float32x4 x50 = v_load(inptr + inpstep*5), x51 = v_load(inptr + inpstep*5 + 4);
    v_float32x4 x60 = v_load(inptr + inpstep*6), x61 = v_load(inptr + inpstep*6 + 4);
    v_float32x4 x70 = v_load(inptr + inpstep*7), x71 = v_load(inptr + inpstep*7 + 4);
    v_float32x4 z00, z01, z10, z11, z20, z21, z30, z31, z40, z41, z50, z51;

    {
        v_float32x4 s12_0, s12_1, s34_0, s34_1, s56_0, s56_1;
        s12_0 = v_add(x10, x20); s12_1 = v_add(x11, x21);
        s34_0 = v_add(x30, x40); s34_1 = v_add(x31, x41);
        s56_0 = v_add(x50, x60); s56_1 = v_add(x51, x61);

        v_float32x4 y00 = v_add(v_add(v_add(x00, s12_0), s34_0), s56_0);
        v_float32x4 y01 = v_add(v_add(v_add(x01, s12_1), s34_1), s56_1);

        v_float32x4 a0 = v_setall_f32(0.25f), a1 = v_setall_f32(4.0f);
        v_float32x4 y20 = v_fma(s56_0, a0, v_fma(s34_0, a1, s12_0));
        v_float32x4 y21 = v_fma(s56_1, a0 ,v_fma(s34_1, a1, s12_1) );

        a0 = v_setall_f32(1.f/16), a1 = v_setall_f32(16.0f);
        v_float32x4 y40 = v_fma(s56_0, a0, v_fma(s34_0, a1, s12_0));
        v_float32x4 y41 = v_fma(s56_1, a0, v_fma(s34_1, a1, s12_1));

        s12_0 = v_sub(x10, x20); s12_1 = v_sub(x11, x21);
        s34_0 = v_sub(x30, x40); s34_1 = v_sub(x31, x41);
        s56_0 = v_sub(x50, x60); s56_1 = v_sub(x51, x61);

        a0 = v_setall_f32(1.f/32), a1 = v_setall_f32(32.f);
        v_float32x4 y50 = v_fma(s56_0, a0, v_fma(s34_0, a1, v_add(x70, s12_0)));
        v_float32x4 y51 = v_fma(s56_1, a0, v_fma(s34_1, a1, v_add(x71, s12_1)));

        a0 = v_setall_f32(0.5f), a1 = v_setall_f32(2.f);
        v_float32x4 y10 = v_fma(s56_0, a0, v_fma(s34_0, a1, s12_0));
        v_float32x4 y11 = v_fma(s56_1, a0, v_fma(s34_1, a1, s12_1));

        a0 = v_setall_f32(0.125f), a1 = v_setall_f32(8.f);
        v_float32x4 y30 = v_fma(s56_0, a0, v_fma(s34_0, a1, s12_0));
        v_float32x4 y31 = v_fma(s56_1, a0, v_fma(s34_1, a1, s12_1));

        v_float32x4 y60 = v_setall_f32(0.f), y61 = y60, y70 = y60, y71 = y60;

        /* transpose 8x8 matrix with v_transpose4x4 */

        v_float32x4 y000, y100, y200, y300, y010, y110, y210, y310, y400, y500, y600, y700, y410, y510, y610, y710;
        v_transpose4x4(y00, y10, y20, y30, y000, y100, y200, y300);
        v_transpose4x4(y01, y11, y21, y31, y010, y110, y210, y310);
        v_transpose4x4(y40, y50, y60, y70, y400, y500, y600, y700);
        v_transpose4x4(y41, y51, y61, y71, y410, y510, y610, y710);

        s12_0 = v_add(y100, y200); s12_1 = v_add(y500, y600);
        s34_0 = v_add(y300, y010); s34_1 = v_add(y700, y410);
        s56_0 = v_add(y110, y210); s56_1 = v_add(y510, y610);

        z00 = v_add(v_add(v_add(y000, s12_0), s34_0), s56_0);
        z01 = v_add(v_add(v_add(y400, s12_1), s34_1), s56_1);

        a0 = v_setall_f32(0.25f), a1 = v_setall_f32(4.0f);
        z20 = v_fma(s56_0, a0, v_fma(s34_0, a1, s12_0));
        z21 = v_fma(s56_1, a0, v_fma(s34_1, a1, s12_1));

        a0 = v_setall_f32(1.f/16), a1 = v_setall_f32(16.0f);
        z40 = v_fma(s56_0, a0, v_fma(s34_0, a1, s12_0));
        z41 = v_fma(s56_1, a0, v_fma(s34_1, a1, s12_1));

        s12_0 = v_sub(y100, y200); s12_1 = v_sub(y500, y600);
        s34_0 = v_sub(y300, y010); s34_1 = v_sub(y700, y410);
        s56_0 = v_sub(y110, y210); s56_1 = v_sub(y510, y610);

        a0 = v_setall_f32(1.f/32), a1 = v_setall_f32(32.0f);
        z50 = v_fma(s56_0, a0, v_fma(s34_0, a1, v_add(y310, s12_0)));
        z51 = v_fma(s56_1, a0, v_fma(s34_1, a1, v_add(y710, s12_1)));
        a0 = v_setall_f32(0.5f), a1 = v_setall_f32(2.0f);
        z10 = v_fma(s56_0, a0, v_fma(s34_0, a1, s12_0));
        z11 = v_fma(s56_1, a0, v_fma(s34_1, a1, s12_1));

        a0 = v_setall_f32(0.125f), a1 = v_setall_f32(8.0f);
        z30 = v_fma(s56_0, a0, v_fma(s34_0, a1, s12_0));
        z31 = v_fma(s56_1, a0, v_fma(s34_1, a1, s12_1));

        v_float32x4 vbias = v_setall_f32(bias);
        z00 = v_add(z00, vbias);
        z01 = v_add(z01, vbias);
        z10 = v_add(z10, vbias);
        z11 = v_add(z11, vbias);
        z20 = v_add(z20, vbias);
        z21 = v_add(z21, vbias);
        z30 = v_add(z30, vbias);
        z31 = v_add(z31, vbias);
        z40 = v_add(z40, vbias);
        z41 = v_add(z41, vbias);
        z50 = v_add(z50, vbias);
        z51 = v_add(z51, vbias);
    }

    if (bpptr)
    {
        z00 = v_add(z00, v_load(bpptr));
        z01 = v_add(z01, v_load_low(bpptr + 4));
        z10 = v_add(z10, v_load(bpptr + bpstep));
        z11 = v_add(z11, v_load_low(bpptr + bpstep + 4));
        z20 = v_add(z20, v_load(bpptr + bpstep * 2));
        z21 = v_add(z21, v_load_low(bpptr + bpstep * 2 + 4));
        z30 = v_add(z30, v_load(bpptr + bpstep * 3));
        z31 = v_add(z31, v_load_low(bpptr + bpstep * 3 + 4));
        z40 = v_add(z40, v_load(bpptr + bpstep * 4));
        z41 = v_add(z41, v_load_low(bpptr + bpstep * 4 + 4));
        z50 = v_add(z50, v_load(bpptr + bpstep * 5));
        z51 = v_add(z51, v_load_low(bpptr + bpstep * 5 + 4));
    }

    if (ifMinMaxAct)
    {
        v_float32x4 vmax = v_setall_f32(maxval);
        v_float32x4 vmin = v_setall_f32(minval);

        z00 = v_min(v_max(z00, vmin), vmax);
        z01 = v_min(v_max(z01, vmin), vmax);
        z10 = v_min(v_max(z10, vmin), vmax);
        z11 = v_min(v_max(z11, vmin), vmax);
        z20 = v_min(v_max(z20, vmin), vmax);
        z21 = v_min(v_max(z21, vmin), vmax);
        z30 = v_min(v_max(z30, vmin), vmax);
        z31 = v_min(v_max(z31, vmin), vmax);
        z40 = v_min(v_max(z40, vmin), vmax);
        z41 = v_min(v_max(z41, vmin), vmax);
        z50 = v_min(v_max(z50, vmin), vmax);
        z51 = v_min(v_max(z51, vmin), vmax);
    }

    v_store(outptr, z00);
    v_store_low(outptr + 4, z01);
    v_store(outptr + outstep, z10);
    v_store_low(outptr + outstep + 4, z11);
    v_store(outptr + outstep*2, z20);
    v_store_low(outptr + outstep*2 + 4, z21);
    v_store(outptr + outstep*3, z30);
    v_store_low(outptr + outstep*3 + 4, z31);
    v_store(outptr + outstep*4, z40);
    v_store_low(outptr + outstep*4 + 4, z41);
    v_store(outptr + outstep*5, z50);
    v_store_low(outptr + outstep*5 + 4, z51);
}
#endif

#else
int runWinograd63(InputArray _input, InputArray _fusedAddMat, OutputArray _output, const Ptr<FastConv>& conv,
                  int ntasks, float minval, float maxval, ActivationLayer* activ, bool ifMinMaxAct)
{
    return 0;
}
#endif

}} // namespace cv::dnn