File: conv2d_grad_op_avx.h

package info (click to toggle)
tiny-dnn 1.0.0a3%2Bds-6
  • links: PTS, VCS
  • area: main
  • in suites: sid
  • size: 4,784 kB
  • sloc: cpp: 16,471; ansic: 11,829; lisp: 3,682; python: 3,422; makefile: 208
file content (509 lines) | stat: -rw-r--r-- 25,237 bytes parent folder | download | duplicates (3)
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
/*
    Copyright (c) 2016, Taiga Nomi, Edgar Riba
    All rights reserved.

    Redistribution and use in source and binary forms, with or without
    modification, are permitted provided that the following conditions are met:
    * Redistributions of source code must retain the above copyright
    notice, this list of conditions and the following disclaimer.
    * Redistributions in binary form must reproduce the above copyright
    notice, this list of conditions and the following disclaimer in the
    documentation and/or other materials provided with the distribution.
    * Neither the name of the <organization> nor the
    names of its contributors may be used to endorse or promote products
    derived from this software without specific prior written permission.

    THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY
    EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
    WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
    DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY
    DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
    (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
    LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
    ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
    (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
    SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
*/
#pragma once

#include <vector>
#include "tiny_dnn/core/params/conv_params.h"
#include "tiny_dnn/core/kernels/conv2d_op_internal.h"

#ifdef CNN_USE_AVX
#include "tiny_dnn/core/kernels/avx_kernel_common.h"
#endif

namespace tiny_dnn {
namespace kernels {

#ifdef CNN_USE_AVX

// float ver
template <typename Allocator>
void avx_conv2d_5x5_back_kernel_one(const core::conv_params& params,
                                    const std::vector<float, Allocator>& prev_out,
                                    const std::vector<float, Allocator>& W,
                                    std::vector<float, Allocator>&       dW,
                                    std::vector<float, Allocator>&       db,
                                    std::vector<float, Allocator>&       curr_delta,
                                    std::vector<float, Allocator>*       prev_delta) {
    auto& in        = params.in;
    auto& out       = params.out;
    auto& in_padded = params.in_padded;
    auto& tbl       = params.tbl;
    auto  w_stride  = params.w_stride;
    const size_t in_padded_area = in_padded.area();
    float* pdelta_dst_org = &(*prev_delta)[0];
    const size_t  h_stride2 = params.h_stride * in_padded.width_;
    static const __m256i imask = _mm256_setr_epi32(-1, -1, -1, -1, -1, 0, 0, 0);
    static const __m256 mask = _mm256_castsi256_ps(_mm256_setr_epi32(-1, -1, -1, -1, -1, 0, 0, 0));
    // propagate delta to previous layer
    if (w_stride == 1 && out.width_ >= 4) {
        const serial_size_t nblocks = out.width_ / 4;
        for (serial_size_t inc = 0; inc < in.depth_; ++inc, pdelta_dst_org += in_padded_area) {
            for (serial_size_t outc = 0; outc < out.depth_; outc++) {
                if (!tbl.is_connected(outc, inc)) continue;
                const float* pw = &W[25 * (in.depth_ * outc + inc)];
                const float* pdelta_src = &curr_delta[out.get_index(0, 0, outc)];
                float* pdelta_dst = pdelta_dst_org;
                __m256 w0a = _mm256_and_ps(_mm256_loadu_ps(pw+0), mask);
                __m256 w1a = _mm256_and_ps(_mm256_loadu_ps(pw+5), mask);
                __m256 w2a = _mm256_and_ps(_mm256_loadu_ps(pw+10), mask);
                __m256 w3a = _mm256_and_ps(_mm256_loadu_ps(pw+15), mask);
                __m256 w4a = _mm256_and_ps(_mm256_loadu_ps(pw+20), mask);
                __m256 w0b = leftShift<4>(w0a);
                __m256 w1b = leftShift<4>(w1a);
                __m256 w2b = leftShift<4>(w2a);
                __m256 w3b = leftShift<4>(w3a);
                __m256 w4b = leftShift<4>(w4a);
                __m256 w0c = leftShift<8>(w0a);
                __m256 w1c = leftShift<8>(w1a);
                __m256 w2c = leftShift<8>(w2a);
                __m256 w3c = leftShift<8>(w3a);
                __m256 w4c = leftShift<8>(w4a);
                __m256 w0d = leftShift<12>(w0a);
                __m256 w1d = leftShift<12>(w1a);
                __m256 w2d = leftShift<12>(w2a);
                __m256 w3d = leftShift<12>(w3a);
                __m256 w4d = leftShift<12>(w4a);
                for (serial_size_t y = 0; y < out.height_; y++) {
                    const float* pdelta_src2 = pdelta_src;
                    float* delta_dst0 = pdelta_dst;
                    float* delta_dst1 = &pdelta_dst[in_padded.width_ * 1];
                    float* delta_dst2 = &pdelta_dst[in_padded.width_ * 2];
                    float* delta_dst3 = &pdelta_dst[in_padded.width_ * 3];
                    float* delta_dst4 = &pdelta_dst[in_padded.width_ * 4];
                    for (serial_size_t n = 0; n < nblocks; ++n) {
                        __m256 delta_src = _mm256_broadcast_ps((const __m128*)pdelta_src2);
                        __m256 dst0 = _mm256_loadu_ps(delta_dst0 + 4 * n);
                        __m256 dst1 = _mm256_loadu_ps(delta_dst1 + 4 * n);
                        __m256 dst2 = _mm256_loadu_ps(delta_dst2 + 4 * n);
                        __m256 dst3 = _mm256_loadu_ps(delta_dst3 + 4 * n);
                        __m256 dst4 = _mm256_loadu_ps(delta_dst4 + 4 * n);
                        __m256 delta_src0 = _mm256_permute_ps(delta_src, _MM_SHUFFLE(0, 0, 0, 0));
                        __m256 delta_src1 = _mm256_permute_ps(delta_src, _MM_SHUFFLE(1, 1, 1, 1));
                        __m256 delta_src2 = _mm256_permute_ps(delta_src, _MM_SHUFFLE(2, 2, 2, 2));
                        __m256 delta_src3 = _mm256_permute_ps(delta_src, _MM_SHUFFLE(3, 3, 3, 3));
                        dst0 = madd256_ps(w0a, delta_src0, dst0);
                        dst1 = madd256_ps(w1a, delta_src0, dst1);
                        dst2 = madd256_ps(w2a, delta_src0, dst2);
                        dst3 = madd256_ps(w3a, delta_src0, dst3);
                        dst4 = madd256_ps(w4a, delta_src0, dst4);
                        dst0 = madd256_ps(w0b, delta_src1, dst0);
                        dst1 = madd256_ps(w1b, delta_src1, dst1);
                        dst2 = madd256_ps(w2b, delta_src1, dst2);
                        dst3 = madd256_ps(w3b, delta_src1, dst3);
                        dst4 = madd256_ps(w4b, delta_src1, dst4);
                        dst0 = madd256_ps(w0c, delta_src2, dst0);
                        dst1 = madd256_ps(w1c, delta_src2, dst1);
                        dst2 = madd256_ps(w2c, delta_src2, dst2);
                        dst3 = madd256_ps(w3c, delta_src2, dst3);
                        dst4 = madd256_ps(w4c, delta_src2, dst4);
                        dst0 = madd256_ps(w0d, delta_src3, dst0);
                        _mm256_storeu_ps(delta_dst0 + 4 * n, dst0);
                        dst1 = madd256_ps(w1d, delta_src3, dst1);
                        _mm256_storeu_ps(delta_dst1 + 4 * n, dst1);
                        dst2 = madd256_ps(w2d, delta_src3, dst2);
                        _mm256_storeu_ps(delta_dst2 + 4 * n, dst2);
                        dst3 = madd256_ps(w3d, delta_src3, dst3);
                        _mm256_storeu_ps(delta_dst3 + 4 * n, dst3);
                        dst4 = madd256_ps(w4d, delta_src3, dst4);
                        _mm256_storeu_ps(delta_dst4 + 4 * n, dst4);
                        pdelta_src2 += 4;
                    }
                    for (serial_size_t x = nblocks * 4; x < out.width_; x++) {
                        __m256 delta_src = _mm256_broadcast_ss(pdelta_src + x);
                        __m256 dst0 = _mm256_loadu_ps(delta_dst0 + x);
                        __m256 dst1 = _mm256_loadu_ps(delta_dst1 + x);
                        __m256 dst2 = _mm256_loadu_ps(delta_dst2 + x);
                        __m256 dst3 = _mm256_loadu_ps(delta_dst3 + x);
                        __m256 dst4 = _mm256_loadu_ps(delta_dst4 + x);
                        dst0 = madd256_ps(w0a, delta_src, dst0);
                        dst1 = madd256_ps(w1a, delta_src, dst1);
                        dst2 = madd256_ps(w2a, delta_src, dst2);
                        dst3 = madd256_ps(w3a, delta_src, dst3);
                        dst4 = madd256_ps(w4a, delta_src, dst4);
                        _mm256_storeu_ps(delta_dst0 + x, dst0);
                        _mm256_storeu_ps(delta_dst1 + x, dst1);
                        _mm256_storeu_ps(delta_dst2 + x, dst2);
                        _mm256_storeu_ps(delta_dst3 + x, dst3);
                        _mm256_storeu_ps(delta_dst4 + x, dst4);
                    }
                    pdelta_src += out.width_;
                    pdelta_dst += h_stride2;
                }
            }
        }
    } else if (out.height_ == 1 && out.width_ == 1) {
        for (serial_size_t inc = 0; inc < in.depth_; ++inc, pdelta_dst_org += in_padded_area) {
            float* delta_dst0 = pdelta_dst_org;
            float* delta_dst1 = &pdelta_dst_org[in_padded.width_ * 1];
            float* delta_dst2 = &pdelta_dst_org[in_padded.width_ * 2];
            float* delta_dst3 = &pdelta_dst_org[in_padded.width_ * 3];
            float* delta_dst4 = &pdelta_dst_org[in_padded.width_ * 4];
            __m256 dst0 = _mm256_loadu_ps(delta_dst0);
            __m256 dst1 = _mm256_loadu_ps(delta_dst1);
            __m256 dst2 = _mm256_loadu_ps(delta_dst2);
            __m256 dst3 = _mm256_loadu_ps(delta_dst3);
            __m256 dst4 = _mm256_maskload_ps(delta_dst4, imask);

            // *FROM
            // ---0 0000
            // ---1 1111
            // ---2 2222
            // ---3 3333
            // ---4 4444
            //
            // *TO
            // 1110 0000
            // 3222 2211
            // 4444 3333
            // ---- ---4
            __m256 sum0 = _mm256_blend_ps(
                dst0,
                leftShift<20>(dst1),
                0xE0 /* 0b11100000 */
            );
            __m256 sum1 = _mm256_blend_ps(
                leftShift<28>(dst3),
                _mm256_blend_ps(leftShift<8>(dst2), rightShift<12>(dst1), 0x03 /* 0b00000011 */),
                0x7F /* 0b01111111 */
            );
            __m256 sum2 = _mm256_blend_ps(
                leftShift<16>(dst4),
                rightShift<4>(dst3),
                0x0F /* 0b00001111 */
            );
            __m128 sum3 = _mm256_extractf128_ps(dst4, 1);

            size_t widx = 25 * inc;
            size_t wstep = 25 * in.depth_;

            if (tbl.is_empty()) {
                for (serial_size_t outc = 0; outc < out.depth_; outc++, widx+=wstep) {
                    __m256 delta_src = _mm256_broadcast_ss(&curr_delta[outc]);
                    const float* pw = (const float*)&W[widx];
                    __m256 w0 = _mm256_loadu_ps(pw+0);
                    __m256 w1 = _mm256_loadu_ps(pw + 8);
                    __m256 w2 = _mm256_loadu_ps(pw + 16);
                    __m128 w3 = _mm_load_ss(pw + 24);
                    sum0 = madd256_ps(w0, delta_src, sum0);
                    sum1 = madd256_ps(w1, delta_src, sum1);
                    sum2 = madd256_ps(w2, delta_src, sum2);
                    sum3 = madd128_ss(w3, _mm256_castps256_ps128(delta_src), sum3);
                }
            }
            else {
                for (serial_size_t outc = 0; outc < out.depth_; outc++, widx += wstep) {
                    if (!tbl.is_connected(outc, inc)) {
                        continue;
                    }
                    __m256 delta_src = _mm256_broadcast_ss(&curr_delta[outc]);
                    const float* pw = (const float*)&W[widx];
                    __m256 w0 = _mm256_loadu_ps(pw + 0);
                    __m256 w1 = _mm256_loadu_ps(pw + 8);
                    __m256 w2 = _mm256_loadu_ps(pw + 16);
                    __m128 w3 = _mm_load_ss(pw + 24);
                    sum0 = madd256_ps(w0, delta_src, sum0);
                    sum1 = madd256_ps(w1, delta_src, sum1);
                    sum2 = madd256_ps(w2, delta_src, sum2);
                    sum3 = madd128_ss(w3, _mm256_castps256_ps128(delta_src), sum3);
                }
            }

            // *FROM
            // 1110 0000
            // 3222 2211
            // 4444 3333
            // ---- ---4
            //
            // *TO
            // ---0 0000
            // ---1 1111
            // ---2 2222
            // ---3 3333
            // ---4 4444
            dst0 = _mm256_blend_ps(
                dst0,
                sum0,
                0x1F /* 0b00011111 */
            );
            dst1 = _mm256_blend_ps(
                dst1,
                _mm256_or_ps(
                    rightShift<20>(sum0),
                    leftShift<12>(sum1)
                ),
                0x1F /* 0b00011111 */
            );
            dst2 = _mm256_blend_ps(
                dst2,
                rightShift<8>(sum1),
                0x1F /* 0b00011111 */
            );
            dst3 = _mm256_blend_ps(
                dst3,
                _mm256_or_ps(
                    rightShift<28>(sum1),
                    leftShift<4>(sum2)
                ),
                0x1F /* 0b00011111 */
            );
            dst4 = _mm256_blend_ps(
                dst4,
                _mm256_set_m128(
                    sum3,
                    _mm256_extractf128_ps(sum2, 1)
                ),
                0x1F /* 0b00011111 */
            );

            _mm256_storeu_ps(delta_dst0, dst0);
            _mm256_storeu_ps(delta_dst1, dst1);
            _mm256_storeu_ps(delta_dst2, dst2);
            _mm256_storeu_ps(delta_dst3, dst3);
            _mm256_maskstore_ps(delta_dst4, imask, dst4);
        } // for
    } else {
        for (serial_size_t inc = 0; inc < in.depth_; ++inc, pdelta_dst_org += in_padded_area) {
            for (serial_size_t outc = 0; outc < out.depth_; outc++) {
                if (!tbl.is_connected(outc, inc)) continue;

                const float* pw = &W[25 * (in.depth_ * outc + inc)];
                const float* pdelta_src = &curr_delta[out.get_index(0, 0, outc)];
                float* pdelta_dst = pdelta_dst_org;
                __m256 w0a = _mm256_maskload_ps(pw+0, imask);
                __m256 w1a = _mm256_maskload_ps(pw+5, imask);
                __m256 w2a = _mm256_maskload_ps(pw+10, imask);
                __m256 w3a = _mm256_maskload_ps(pw+15, imask);
                __m256 w4a = _mm256_maskload_ps(pw+20, imask);
                for (serial_size_t y = 0; y < out.height_; y++) {
                    float* delta_dst0 = pdelta_dst;
                    float* delta_dst1 = &pdelta_dst[in_padded.width_ * 1];
                    float* delta_dst2 = &pdelta_dst[in_padded.width_ * 2];
                    float* delta_dst3 = &pdelta_dst[in_padded.width_ * 3];
                    float* delta_dst4 = &pdelta_dst[in_padded.width_ * 4];
                    for (serial_size_t x = 0; x < out.width_; x++) {
                        __m256 delta_src = _mm256_broadcast_ss(pdelta_src + x);
                        __m256 dst0 = _mm256_loadu_ps(delta_dst0);
                        __m256 dst1 = _mm256_loadu_ps(delta_dst1);
                        __m256 dst2 = _mm256_loadu_ps(delta_dst2);
                        __m256 dst3 = _mm256_loadu_ps(delta_dst3);
                        __m256 dst4 = _mm256_maskload_ps(delta_dst4, imask);
                        dst0 = madd256_ps(w0a, delta_src, dst0);
                        dst1 = madd256_ps(w1a, delta_src, dst1);
                        dst2 = madd256_ps(w2a, delta_src, dst2);
                        dst3 = madd256_ps(w3a, delta_src, dst3);
                        dst4 = madd256_ps(w4a, delta_src, dst4);
                        _mm256_storeu_ps(delta_dst0, dst0);
                        _mm256_storeu_ps(delta_dst1, dst1);
                        _mm256_storeu_ps(delta_dst2, dst2);
                        _mm256_storeu_ps(delta_dst3, dst3);
                        _mm256_maskstore_ps(delta_dst4, imask, dst4);
                        delta_dst0 += w_stride;
                        delta_dst1 += w_stride;
                        delta_dst2 += w_stride;
                        delta_dst3 += w_stride;
                        delta_dst4 += w_stride;
                    } // for x
                    pdelta_src += out.width_;
                    pdelta_dst += h_stride2;
                } // for y
            } // for outc
        } // for inc
    }

    // accumulate dw
    if (out.width_ == 1 && out.height_ == 1) {
        const float* pprev_out = &prev_out[0];
        for (serial_size_t inc = 0; inc < in.depth_; ++inc, pprev_out += in_padded_area) {
            VECTORIZE_ALIGN(32) float floats[28];
            size_t in_padded_width = in_padded.width_;
            _mm256_store_ps(&floats[0], _mm256_loadu_ps(pprev_out + in_padded_width * 0));
            _mm256_storeu_ps(&floats[5], _mm256_loadu_ps(pprev_out + in_padded_width * 1));
            _mm256_storeu_ps(&floats[10], _mm256_loadu_ps(pprev_out + in_padded_width * 2));
            _mm256_storeu_ps(&floats[15], _mm256_loadu_ps(pprev_out + in_padded_width * 3));
            _mm256_storeu_ps(&floats[20], _mm256_maskload_ps(pprev_out + in_padded_width * 4, imask));
            __m256 prevos0 = _mm256_load_ps(&floats[0]);
            __m256 prevos1 = _mm256_load_ps(&floats[8]);
            __m256 prevos2 = _mm256_load_ps(&floats[16]);
            __m128 prevos3 = _mm_load_ss(&floats[24]);
            serial_size_t widx = 25 * inc;
            serial_size_t widx_delta = 25 * in.depth_;
            float* pdW = &dW[widx];
            for (serial_size_t outc = 0; outc < out.depth_; outc++, pdW += widx_delta) {
                if (!tbl.is_connected(outc, inc)) {
                    continue;
                }
                __m256 delta = _mm256_broadcast_ss(&curr_delta[outc]);
                __m256 w0 = _mm256_loadu_ps(pdW+0);
                __m256 w1 = _mm256_loadu_ps(pdW+8);
                __m256 w2 = _mm256_loadu_ps(pdW + 16);
                __m128 w3 = _mm_load_ss(pdW + 24);
                w0 = madd256_ps(prevos0, delta, w0);
                w1 = madd256_ps(prevos1, delta, w1);
                w2 = madd256_ps(prevos2, delta, w2);
                w3 = madd128_ss(prevos3, _mm256_castps256_ps128(delta), w3);
                _mm256_storeu_ps(pdW + 0, w0);
                _mm256_storeu_ps(pdW + 8, w1);
                _mm256_storeu_ps(pdW+16, w2);
                _mm_store_ss(pdW+24, w3);
            }
        }
    } else {
        // prepare load-mask beforehand
        const size_t nblocks = out.width_ >> 3;
        static const int32_t masks[] = {
            -1, -1, -1, -1,
            -1, -1, -1, -1,
            0, 0, 0, 0,
            0, 0, 0, 0,
        };
        const size_t remainder = out.width_ & 7;
        __m256i mask = _mm256_loadu_si256((const __m256i*)(masks + 8 - remainder));
        auto& weight = params.weight;
        for (serial_size_t inc = 0; inc < in.depth_; ++inc) {
            for (serial_size_t outc = 0; outc < out.depth_; outc++) {

                if (!tbl.is_connected(outc, inc)) continue;
                const float* delta = &curr_delta[out.get_index(0, 0, outc)];

                serial_size_t widx = weight.get_index(0, 0, in.depth_ * outc + inc);
                for (serial_size_t wy = 0; wy < 5 /* weight.height_ */; wy++) {
                    for (serial_size_t wx = 0; wx < 5 /* weight.width_ */; wx++) {
                        const float* prevo = &prev_out[in_padded.get_index(wx, wy, inc)];

                        if (w_stride > 1) {
                            float_t dst = float_t(0);

                            for (serial_size_t y = 0; y < params.out.height_; y++) {
                                serial_size_t prevo_idx = y * params.in_padded.width_ * params.h_stride;
                                serial_size_t delta_idx = y * params.out.width_;

                                for (serial_size_t x = 0; x < params.out.width_; x++) {
                                    dst += prevo[prevo_idx + x * params.w_stride] * delta[delta_idx + x];
                                }
                            }
                            dW[widx] += dst;
                        }
                        else {
                            __m128 prev_sum = _mm_load_ss(&dW[widx]);
                            __m256 sum0 = _mm256_setzero_ps();
                            __m256 sum1 = _mm256_setzero_ps();
                            for (serial_size_t y = 0; y < out.height_; y++) {
                                // vectorize::dot
                                const float* pa = prevo + y * in_padded.width_ * params.h_stride;
                                const float* pb = delta + y * out.width_;
                                for (size_t i = 0; i < nblocks; ++i) {
                                    __m256 a = _mm256_loadu_ps(pa + 8 * i);
                                    __m256 b = _mm256_loadu_ps(pb + 8 * i);
                                    sum0 = madd256_ps(a, b, sum0);
                                }
                                if (remainder) {
                                    __m256 a = _mm256_maskload_ps(pa + 8 * nblocks, mask);
                                    __m256 b = _mm256_maskload_ps(pb + 8 * nblocks, mask);
                                    sum1 = madd256_ps(a, b, sum1);
                                }
                            }
                            sum1 = _mm256_and_ps(sum1, _mm256_castsi256_ps(mask));
                            __m256 sum = _mm256_add_ps(sum0, sum1);
                            _mm_store_ss(&dW[widx], _mm_add_ps(prev_sum, hsum256_ps(sum)));
                        }
                        ++widx;
                    }
                }
            }
        }
    }

    // accumulate db
    if (params.has_bias) {
        //fvec_t& db = *in_grad[2];
        
        if (out.width_ == 1 && out.height_ == 1) {
            for (serial_size_t outc = 0; outc < out.depth_; outc++) {
                db[outc] += curr_delta[outc];
            }
        } else {
            for (serial_size_t outc = 0; outc < out.depth_; outc++) {
                const float *delta = &curr_delta[out.get_index(0, 0, outc)];
                db[outc] += std::accumulate(delta, delta + out.width_ * out.height_, float(0));
            }
        }
    }
} // avx_conv2d_5x5_back_kernel float ver

// double ver
template <typename Allocator>
void avx_conv2d_5x5_back_kernel(const core::conv_params& params,
                                const std::vector<std::vector<double, Allocator>>& prev_out,
                                const std::vector<double, Allocator>& W,
                                std::vector<std::vector<double, Allocator>>&       dW,
                                std::vector<std::vector<double, Allocator>>&       db,
                                std::vector<std::vector<double, Allocator>>&       curr_delta,
                                std::vector<std::vector<double, Allocator>>&       prev_delta) {
    // backward-pass fallbacks to tiny-backend at float_t == double
    conv2d_op_internal(prev_out, W, dW, db, curr_delta, prev_delta, params, true);
}

// float ver
template <typename Allocator>
void avx_conv2d_5x5_back_kernel(const core::conv_params& params,
                                const std::vector<std::vector<float, Allocator>>& prev_out,
                                const std::vector<float, Allocator>& W,
                                std::vector<std::vector<float, Allocator>>&       dW,
                                std::vector<std::vector<float, Allocator>>&       db,
                                std::vector<std::vector<float, Allocator>>&       curr_delta,
                                std::vector<std::vector<float, Allocator>>&       prev_delta) {
    for_i(prev_out.size(), [&](int sample) {
        avx_conv2d_5x5_back_kernel_one(params, prev_out[sample], W, dW[sample], db[sample],
            curr_delta[sample], &prev_delta[sample]);
    });
} 


#endif // CNN_USE_AVX

inline void
conv2d_grad_op_avx(const tensor_t&        prev_out,
                   const vec_t&                  W,
                   tensor_t&                    dW,
                   tensor_t&                    db,
                   tensor_t&            curr_delta,
                   tensor_t&            prev_delta,
                   const core::conv_params& params,
                   const bool    layer_parallelize) {
#ifdef CNN_USE_AVX
    if (params.weight.height_ == 5 && params.weight.width_ == 5) {
        avx_conv2d_5x5_back_kernel(params, prev_out, W, dW, db, curr_delta, prev_delta);
        return;
    }
#endif

    conv2d_op_internal(prev_out, W, dW, db, curr_delta,
                       prev_delta, params, layer_parallelize);
}

}  // namespace kernels
}  // namespace tiny_dnn