1 // Tencent is pleased to support the open source community by making ncnn available.
2 //
3 // Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
4 //
5 // Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
6 // in compliance with the License. You may obtain a copy of the License at
7 //
8 // https://opensource.org/licenses/BSD-3-Clause
9 //
10 // Unless required by applicable law or agreed to in writing, software distributed
11 // under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
12 // CONDITIONS OF ANY KIND, either express or implied. See the License for the
13 // specific language governing permissions and limitations under the License.
14
15 #include "deconvolution_arm.h"
16
17 #include "layer_type.h"
18
19 #if __ARM_NEON
20 #include <arm_neon.h>
21 #endif // __ARM_NEON
22
23 #include "arm_activation.h"
24
25 namespace ncnn {
26
27 #include "deconvolution_3x3.h"
28 #include "deconvolution_4x4.h"
29
30 #if __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
31 #include "deconvolution_4x4_fp16s.h"
32 #endif
33
Deconvolution_arm()34 Deconvolution_arm::Deconvolution_arm()
35 {
36 #if __ARM_NEON
37 support_packing = true;
38 #if __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
39 support_fp16_storage = true;
40 #endif
41 #endif // __ARM_NEON
42
43 support_bf16_storage = true;
44
45 activation = 0;
46 }
47
create_pipeline(const Option & opt)48 int Deconvolution_arm::create_pipeline(const Option& opt)
49 {
50 if (activation_type == 1)
51 {
52 activation = ncnn::create_layer(ncnn::LayerType::ReLU);
53
54 ncnn::ParamDict pd;
55 activation->load_param(pd);
56 }
57 else if (activation_type == 2)
58 {
59 activation = ncnn::create_layer(ncnn::LayerType::ReLU);
60
61 ncnn::ParamDict pd;
62 pd.set(0, activation_params[0]); // slope
63 activation->load_param(pd);
64 }
65 else if (activation_type == 3)
66 {
67 activation = ncnn::create_layer(ncnn::LayerType::Clip);
68
69 ncnn::ParamDict pd;
70 pd.set(0, activation_params[0]); // min
71 pd.set(1, activation_params[1]); // max
72 activation->load_param(pd);
73 }
74 else if (activation_type == 4)
75 {
76 activation = ncnn::create_layer(ncnn::LayerType::Sigmoid);
77
78 ncnn::ParamDict pd;
79 activation->load_param(pd);
80 }
81
82 if (activation)
83 {
84 activation->create_pipeline(opt);
85 }
86
87 #if __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
88 if (opt.use_fp16_storage)
89 {
90 return create_pipeline_fp16s(opt);
91 }
92 #endif
93
94 if (opt.use_bf16_storage)
95 {
96 return create_pipeline_bf16s(opt);
97 }
98
99 const int maxk = kernel_w * kernel_h;
100 int num_input = weight_data_size / maxk / num_output;
101
102 Mat weight_data_transposed(weight_data.w);
103 {
104 float* pt = weight_data_transposed;
105 const float* p = weight_data;
106
107 for (int i = 0; i < num_input * num_output; i++)
108 {
109 for (int k = 0; k < maxk; k++)
110 {
111 pt[maxk - 1 - k] = p[k];
112 }
113
114 p += maxk;
115 pt += maxk;
116 }
117 }
118
119 int elempack = (support_packing && opt.use_packing_layout && num_input % 4 == 0) ? 4 : 1;
120 int out_elempack = (support_packing && opt.use_packing_layout && num_output % 4 == 0) ? 4 : 1;
121
122 #if __ARM_NEON
123 // pack4
124 if (elempack == 4 && out_elempack == 4)
125 {
126 // src = kw-kh-inch-outch
127 // dst = 4b-4a-kw-kh-inch/4a-outch/4b
128 {
129 Mat weight_data_r2 = weight_data_transposed.reshape(maxk, num_input, num_output);
130
131 weight_data_pack4.create(maxk, num_input / 4, num_output / 4, (size_t)4 * 16, 16);
132
133 for (int q = 0; q + 3 < num_output; q += 4)
134 {
135 const Mat k0 = weight_data_r2.channel(q);
136 const Mat k1 = weight_data_r2.channel(q + 1);
137 const Mat k2 = weight_data_r2.channel(q + 2);
138 const Mat k3 = weight_data_r2.channel(q + 3);
139
140 Mat g0 = weight_data_pack4.channel(q / 4);
141
142 for (int p = 0; p + 3 < num_input; p += 4)
143 {
144 const float* k00 = k0.row(p);
145 const float* k01 = k0.row(p + 1);
146 const float* k02 = k0.row(p + 2);
147 const float* k03 = k0.row(p + 3);
148
149 const float* k10 = k1.row(p);
150 const float* k11 = k1.row(p + 1);
151 const float* k12 = k1.row(p + 2);
152 const float* k13 = k1.row(p + 3);
153
154 const float* k20 = k2.row(p);
155 const float* k21 = k2.row(p + 1);
156 const float* k22 = k2.row(p + 2);
157 const float* k23 = k2.row(p + 3);
158
159 const float* k30 = k3.row(p);
160 const float* k31 = k3.row(p + 1);
161 const float* k32 = k3.row(p + 2);
162 const float* k33 = k3.row(p + 3);
163
164 float* g00 = g0.row(p / 4);
165
166 for (int k = 0; k < maxk; k++)
167 {
168 g00[0] = k00[k];
169 g00[1] = k10[k];
170 g00[2] = k20[k];
171 g00[3] = k30[k];
172
173 g00[4] = k01[k];
174 g00[5] = k11[k];
175 g00[6] = k21[k];
176 g00[7] = k31[k];
177
178 g00[8] = k02[k];
179 g00[9] = k12[k];
180 g00[10] = k22[k];
181 g00[11] = k32[k];
182
183 g00[12] = k03[k];
184 g00[13] = k13[k];
185 g00[14] = k23[k];
186 g00[15] = k33[k];
187
188 g00 += 16;
189 }
190 }
191 }
192 }
193 }
194
195 // pack1to4
196 if (elempack == 1 && out_elempack == 4)
197 {
198 // src = kw-kh-inch-outch
199 // dst = 4b-kw-kh-inch-outch/4b
200 {
201 Mat weight_data_r2 = weight_data_transposed.reshape(maxk, num_input, num_output);
202
203 weight_data_pack1to4.create(maxk, num_input, num_output / 4, (size_t)4 * 4, 4);
204
205 for (int q = 0; q + 3 < num_output; q += 4)
206 {
207 const Mat k0 = weight_data_r2.channel(q);
208 const Mat k1 = weight_data_r2.channel(q + 1);
209 const Mat k2 = weight_data_r2.channel(q + 2);
210 const Mat k3 = weight_data_r2.channel(q + 3);
211
212 Mat g0 = weight_data_pack1to4.channel(q / 4);
213
214 for (int p = 0; p < num_input; p++)
215 {
216 const float* k00 = k0.row(p);
217 const float* k10 = k1.row(p);
218 const float* k20 = k2.row(p);
219 const float* k30 = k3.row(p);
220
221 float* g00 = g0.row(p);
222
223 for (int k = 0; k < maxk; k++)
224 {
225 g00[0] = k00[k];
226 g00[1] = k10[k];
227 g00[2] = k20[k];
228 g00[3] = k30[k];
229
230 g00 += 4;
231 }
232 }
233 }
234 }
235 }
236
237 // pack4to1
238 if (elempack == 4 && out_elempack == 1)
239 {
240 // src = kw-kh-inch-outch
241 // dst = 4a-kw-kh-inch/4a-outch
242 {
243 Mat weight_data_r2 = weight_data_transposed.reshape(maxk, num_input, num_output);
244
245 weight_data_pack4to1.create(maxk, num_input / 4, num_output, (size_t)4 * 4, 4);
246
247 for (int q = 0; q < num_output; q++)
248 {
249 const Mat k0 = weight_data_r2.channel(q);
250 Mat g0 = weight_data_pack4to1.channel(q);
251
252 for (int p = 0; p + 3 < num_input; p += 4)
253 {
254 const float* k00 = k0.row(p);
255 const float* k01 = k0.row(p + 1);
256 const float* k02 = k0.row(p + 2);
257 const float* k03 = k0.row(p + 3);
258
259 float* g00 = g0.row(p / 4);
260
261 for (int k = 0; k < maxk; k++)
262 {
263 g00[0] = k00[k];
264 g00[1] = k01[k];
265 g00[2] = k02[k];
266 g00[3] = k03[k];
267
268 g00 += 4;
269 }
270 }
271 }
272 }
273 }
274 #endif // __ARM_NEON
275
276 // pack1
277 if (elempack == 1 && out_elempack == 1)
278 {
279 weight_data_pack1 = weight_data_transposed;
280 }
281
282 return 0;
283 }
284
destroy_pipeline(const Option & opt)285 int Deconvolution_arm::destroy_pipeline(const Option& opt)
286 {
287 if (activation)
288 {
289 activation->destroy_pipeline(opt);
290 delete activation;
291 activation = 0;
292 }
293
294 return 0;
295 }
296
forward(const Mat & bottom_blob,Mat & top_blob,const Option & opt) const297 int Deconvolution_arm::forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const
298 {
299 int elembits = bottom_blob.elembits();
300
301 #if __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
302 if (opt.use_fp16_storage && elembits == 16)
303 {
304 if (opt.use_fp16_arithmetic)
305 return forward_fp16sa(bottom_blob, top_blob, opt);
306 else
307 return forward_fp16s(bottom_blob, top_blob, opt);
308 }
309 #endif
310
311 if (opt.use_bf16_storage && elembits == 16)
312 return forward_bf16s(bottom_blob, top_blob, opt);
313
314 // deconvolv with NxN kernel
315 // value = value + bias
316
317 int w = bottom_blob.w;
318 int h = bottom_blob.h;
319 int channels = bottom_blob.c;
320 size_t elemsize = bottom_blob.elemsize;
321 int elempack = bottom_blob.elempack;
322
323 // NCNN_LOGE("Deconvolution input %d x %d pad = %d %d ksize=%d %d stride=%d %d", w, h, pad_w, pad_h, kernel_w, kernel_h, stride_w, stride_h);
324
325 const int kernel_extent_w = dilation_w * (kernel_w - 1) + 1;
326 const int kernel_extent_h = dilation_h * (kernel_h - 1) + 1;
327
328 int outw = (w - 1) * stride_w + kernel_extent_w;
329 int outh = (h - 1) * stride_h + kernel_extent_h;
330 int out_elempack = (support_packing && opt.use_packing_layout && num_output % 4 == 0) ? 4 : 1;
331 size_t out_elemsize = elemsize / elempack * out_elempack;
332
333 Mat top_blob_bordered;
334 if (pad_left > 0 || pad_right > 0 || pad_top > 0 || pad_bottom > 0 || output_pad_right > 0 || output_pad_bottom > 0 || (output_w > 0 && output_h > 0))
335 {
336 top_blob_bordered.create(outw, outh, num_output / out_elempack, out_elemsize, out_elempack, opt.workspace_allocator);
337 }
338 else
339 {
340 top_blob_bordered = top_blob;
341 top_blob_bordered.create(outw, outh, num_output / out_elempack, out_elemsize, out_elempack, opt.blob_allocator);
342 }
343 if (top_blob_bordered.empty())
344 return -100;
345
346 const int maxk = kernel_w * kernel_h;
347
348 #if __ARM_NEON
349 if (elempack == 4 && out_elempack == 4)
350 {
351 // num_output
352 #pragma omp parallel for num_threads(opt.num_threads)
353 for (int p = 0; p < num_output / out_elempack; p++)
354 {
355 float* outptr = top_blob_bordered.channel(p);
356
357 for (int i = 0; i < outh; i++)
358 {
359 for (int j = 0; j < outw; j++)
360 {
361 float32x4_t _sum = vdupq_n_f32(0.f);
362
363 if (bias_term)
364 {
365 _sum = vld1q_f32(((const float*)bias_data) + p * 4);
366 }
367
368 const float* kptr = (const float*)weight_data_pack4 + maxk * channels * p * 16;
369
370 // channels
371 for (int q = 0; q < channels; q++)
372 {
373 const Mat m = bottom_blob.channel(q);
374
375 for (int y = 0; y < kernel_h; y++)
376 {
377 int sys = (i + y * dilation_h - (kernel_extent_h - 1));
378 if (sys < 0 || sys % stride_h != 0)
379 continue;
380
381 int sy = sys / stride_h;
382 if (sy >= h)
383 continue;
384
385 for (int x = 0; x < kernel_w; x++)
386 {
387 int sxs = (j + x * dilation_w - (kernel_extent_w - 1));
388 if (sxs < 0 || sxs % stride_w != 0)
389 continue;
390
391 int sx = sxs / stride_w;
392 if (sx >= w)
393 continue;
394
395 const float* sptr = m.row(sy) + sx * 4;
396
397 float32x4_t _val = vld1q_f32(sptr);
398
399 int k = y * kernel_w + x;
400
401 float32x4_t _w0 = vld1q_f32(kptr + k * 16);
402 float32x4_t _w1 = vld1q_f32(kptr + k * 16 + 4);
403 float32x4_t _w2 = vld1q_f32(kptr + k * 16 + 8);
404 float32x4_t _w3 = vld1q_f32(kptr + k * 16 + 12);
405
406 #if __aarch64__
407 _sum = vmlaq_laneq_f32(_sum, _w0, _val, 0);
408 _sum = vmlaq_laneq_f32(_sum, _w1, _val, 1);
409 _sum = vmlaq_laneq_f32(_sum, _w2, _val, 2);
410 _sum = vmlaq_laneq_f32(_sum, _w3, _val, 3);
411 #else
412 _sum = vmlaq_lane_f32(_sum, _w0, vget_low_f32(_val), 0);
413 _sum = vmlaq_lane_f32(_sum, _w1, vget_low_f32(_val), 1);
414 _sum = vmlaq_lane_f32(_sum, _w2, vget_high_f32(_val), 0);
415 _sum = vmlaq_lane_f32(_sum, _w3, vget_high_f32(_val), 1);
416 #endif
417 }
418 }
419
420 kptr += maxk * 16;
421 }
422
423 _sum = activation_ps(_sum, activation_type, activation_params);
424
425 vst1q_f32(outptr + j * 4, _sum);
426 }
427
428 outptr += outw * 4;
429 }
430 }
431 }
432
433 if (elempack == 1 && out_elempack == 4)
434 {
435 // num_output
436 #pragma omp parallel for num_threads(opt.num_threads)
437 for (int p = 0; p < num_output / out_elempack; p++)
438 {
439 float* outptr = top_blob_bordered.channel(p);
440
441 for (int i = 0; i < outh; i++)
442 {
443 for (int j = 0; j < outw; j++)
444 {
445 float32x4_t _sum = vdupq_n_f32(0.f);
446
447 if (bias_term)
448 {
449 _sum = vld1q_f32(((const float*)bias_data) + p * 4);
450 }
451
452 const float* kptr = (const float*)weight_data_pack1to4 + maxk * channels * p * 4;
453
454 // channels
455 for (int q = 0; q < channels; q++)
456 {
457 const Mat m = bottom_blob.channel(q);
458
459 for (int y = 0; y < kernel_h; y++)
460 {
461 int sys = (i + y * dilation_h - (kernel_extent_h - 1));
462 if (sys < 0 || sys % stride_h != 0)
463 continue;
464
465 int sy = sys / stride_h;
466 if (sy >= h)
467 continue;
468
469 const float* sptr = m.row(sy);
470
471 for (int x = 0; x < kernel_w; x++)
472 {
473 int sxs = (j + x * dilation_w - (kernel_extent_w - 1));
474 if (sxs < 0 || sxs % stride_w != 0)
475 continue;
476
477 int sx = sxs / stride_w;
478 if (sx >= w)
479 continue;
480
481 float32x4_t _val = vdupq_n_f32(sptr[sx]);
482
483 int k = y * kernel_w + x;
484
485 float32x4_t _w = vld1q_f32(kptr + k * 4);
486
487 _sum = vmlaq_f32(_sum, _val, _w);
488 }
489 }
490
491 kptr += maxk * 4;
492 }
493
494 _sum = activation_ps(_sum, activation_type, activation_params);
495
496 vst1q_f32(outptr + j * 4, _sum);
497 }
498
499 outptr += outw * 4;
500 }
501 }
502 }
503
504 if (elempack == 4 && out_elempack == 1)
505 {
506 // num_output
507 #pragma omp parallel for num_threads(opt.num_threads)
508 for (int p = 0; p < num_output / out_elempack; p++)
509 {
510 float* outptr = top_blob_bordered.channel(p);
511
512 for (int i = 0; i < outh; i++)
513 {
514 for (int j = 0; j < outw; j++)
515 {
516 float sum = 0.f;
517
518 if (bias_term)
519 {
520 sum = bias_data[p];
521 }
522
523 const float* kptr = (const float*)weight_data_pack4to1 + maxk * channels * p * 4;
524
525 // channels
526 for (int q = 0; q < channels; q++)
527 {
528 const Mat m = bottom_blob.channel(q);
529
530 for (int y = 0; y < kernel_h; y++)
531 {
532 int sys = (i + y * dilation_h - (kernel_extent_h - 1));
533 if (sys < 0 || sys % stride_h != 0)
534 continue;
535
536 int sy = sys / stride_h;
537 if (sy >= h)
538 continue;
539
540 for (int x = 0; x < kernel_w; x++)
541 {
542 int sxs = (j + x * dilation_w - (kernel_extent_w - 1));
543 if (sxs < 0 || sxs % stride_w != 0)
544 continue;
545
546 int sx = sxs / stride_w;
547 if (sx >= w)
548 continue;
549
550 const float* sptr = m.row(sy) + sx * 4;
551
552 float32x4_t _val = vld1q_f32(sptr);
553
554 int k = y * kernel_w + x;
555
556 float32x4_t _w = vld1q_f32(kptr + k * 4);
557
558 float32x4_t _s4 = vmulq_f32(_val, _w);
559 #if __aarch64__
560 sum += vaddvq_f32(_s4); // dot
561 #else
562 float32x2_t _ss = vadd_f32(vget_low_f32(_s4), vget_high_f32(_s4));
563 _ss = vpadd_f32(_ss, _ss);
564 sum += vget_lane_f32(_ss, 0);
565 #endif
566 }
567 }
568
569 kptr += maxk * 4;
570 }
571
572 sum = activation_ss(sum, activation_type, activation_params);
573
574 outptr[j] = sum;
575 }
576
577 outptr += outw;
578 }
579 }
580 }
581 #endif // __ARM_NEON
582
583 if (elempack == 1 && out_elempack == 1)
584 {
585 if (kernel_w == 3 && kernel_h == 3 && stride_w == 1 && stride_h == 1 && dilation_w == 1 && dilation_h == 1)
586 {
587 deconv3x3s1_neon(bottom_blob, top_blob_bordered, weight_data, bias_data, opt);
588
589 if (activation)
590 {
591 activation->forward_inplace(top_blob_bordered, opt);
592 }
593 }
594 else if (kernel_w == 3 && kernel_h == 3 && stride_w == 2 && stride_h == 2 && dilation_w == 1 && dilation_h == 1)
595 {
596 deconv3x3s2_neon(bottom_blob, top_blob_bordered, weight_data, bias_data, opt);
597
598 if (activation)
599 {
600 activation->forward_inplace(top_blob_bordered, opt);
601 }
602 }
603 else if (kernel_w == 4 && kernel_h == 4 && stride_w == 1 && stride_h == 1 && dilation_w == 1 && dilation_h == 1)
604 {
605 deconv4x4s1_neon(bottom_blob, top_blob_bordered, weight_data, bias_data, opt);
606
607 if (activation)
608 {
609 activation->forward_inplace(top_blob_bordered, opt);
610 }
611 }
612 else if (kernel_w == 4 && kernel_h == 4 && stride_w == 2 && stride_h == 2 && dilation_w == 1 && dilation_h == 1)
613 {
614 deconv4x4s2_neon(bottom_blob, top_blob_bordered, weight_data, bias_data, opt);
615
616 if (activation)
617 {
618 activation->forward_inplace(top_blob_bordered, opt);
619 }
620 }
621 else
622 {
623 // num_output
624 #pragma omp parallel for num_threads(opt.num_threads)
625 for (int p = 0; p < num_output; p++)
626 {
627 float* outptr = top_blob_bordered.channel(p);
628
629 for (int i = 0; i < outh; i++)
630 {
631 for (int j = 0; j < outw; j++)
632 {
633 float sum = 0.f;
634
635 if (bias_term)
636 {
637 sum = bias_data[p];
638 }
639
640 const float* kptr = (const float*)weight_data_pack1 + maxk * channels * p;
641
642 // channels
643 for (int q = 0; q < channels; q++)
644 {
645 const Mat m = bottom_blob.channel(q);
646
647 for (int y = 0; y < kernel_h; y++)
648 {
649 int sys = (i + y * dilation_h - (kernel_extent_h - 1));
650 if (sys < 0 || sys % stride_h != 0)
651 continue;
652
653 int sy = sys / stride_h;
654 if (sy >= h)
655 continue;
656
657 const float* sptr = m.row(sy);
658
659 for (int x = 0; x < kernel_w; x++)
660 {
661 int sxs = (j + x * dilation_w - (kernel_extent_w - 1));
662 if (sxs < 0 || sxs % stride_w != 0)
663 continue;
664
665 int sx = sxs / stride_w;
666 if (sx >= w)
667 continue;
668
669 float val = sptr[sx];
670
671 int k = y * kernel_w + x;
672
673 float w = kptr[k];
674
675 sum += val * w;
676 }
677 }
678
679 kptr += maxk;
680 }
681
682 if (activation_type == 1)
683 {
684 sum = std::max(sum, 0.f);
685 }
686 else if (activation_type == 2)
687 {
688 float slope = activation_params[0];
689 sum = sum > 0.f ? sum : sum * slope;
690 }
691 else if (activation_type == 3)
692 {
693 float min = activation_params[0];
694 float max = activation_params[1];
695 if (sum < min)
696 sum = min;
697 if (sum > max)
698 sum = max;
699 }
700 else if (activation_type == 4)
701 {
702 sum = static_cast<float>(1.f / (1.f + exp(-sum)));
703 }
704
705 outptr[j] = sum;
706 }
707
708 outptr += outw;
709 }
710 }
711 }
712 }
713
714 cut_padding(top_blob_bordered, top_blob, opt);
715 if (top_blob.empty())
716 return -100;
717
718 return 0;
719 }
720
721 #if __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
create_pipeline_fp16s(const Option & opt)722 int Deconvolution_arm::create_pipeline_fp16s(const Option& opt)
723 {
724 const int maxk = kernel_w * kernel_h;
725 const int num_input = weight_data_size / maxk / num_output;
726
727 int elempack = 1;
728 int out_elempack = 1;
729
730 if (opt.use_packing_layout)
731 {
732 elempack = opt.use_fp16_arithmetic && num_input % 8 == 0 ? 8 : num_input % 4 == 0 ? 4 : 1;
733 out_elempack = opt.use_fp16_arithmetic && num_output % 8 == 0 ? 8 : num_output % 4 == 0 ? 4 : 1;
734 }
735
736 Mat weight_data_transposed(weight_data.w);
737 {
738 float* pt = weight_data_transposed;
739 const float* p = weight_data;
740
741 for (int i = 0; i < num_input * num_output; i++)
742 {
743 for (int k = 0; k < maxk; k++)
744 {
745 pt[maxk - 1 - k] = p[k];
746 }
747
748 p += maxk;
749 pt += maxk;
750 }
751 }
752
753 // src = kw-kh-inch-outch
754 // dst = pb-pa-kw-kh-inch/pa-outch/pb
755 {
756 Mat weight_data_r2 = weight_data_transposed.reshape(maxk, num_input, num_output);
757
758 weight_data_fp16.create(maxk, num_input / elempack, num_output / out_elempack, (size_t)2u * elempack * out_elempack, elempack * out_elempack);
759
760 for (int q = 0; q + (out_elempack - 1) < num_output; q += out_elempack)
761 {
762 Mat g0 = weight_data_fp16.channel(q / out_elempack);
763
764 for (int p = 0; p + (elempack - 1) < num_input; p += elempack)
765 {
766 __fp16* g00 = g0.row<__fp16>(p / elempack);
767
768 for (int k = 0; k < maxk; k++)
769 {
770 for (int i = 0; i < elempack; i++)
771 {
772 for (int j = 0; j < out_elempack; j++)
773 {
774 const float* k00 = weight_data_r2.channel(q + j).row(p + i);
775
776 g00[0] = (__fp16)k00[k];
777
778 g00++;
779 }
780 }
781 }
782 }
783 }
784 }
785
786 if (elempack == 1 && out_elempack == 1 && opt.use_fp16_arithmetic)
787 {
788 if (kernel_w == 4 && kernel_h == 4 && stride_w == 2 && stride_h == 2 && dilation_w == 1 && dilation_h == 1)
789 {
790 ncnn::cast_float32_to_float16(weight_data, weight_data_fp16, opt);
791 }
792 }
793
794 ncnn::cast_float32_to_float16(bias_data, bias_data_fp16, opt);
795
796 return 0;
797 }
798
forward_fp16s(const Mat & bottom_blob,Mat & top_blob,const Option & opt) const799 int Deconvolution_arm::forward_fp16s(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const
800 {
801 // deconvolv with NxN kernel
802 // value = value + bias
803
804 int w = bottom_blob.w;
805 int h = bottom_blob.h;
806 int channels = bottom_blob.c;
807 size_t elemsize = bottom_blob.elemsize;
808 int elempack = bottom_blob.elempack;
809
810 // NCNN_LOGE("Deconvolution input %d x %d pad = %d %d ksize=%d %d stride=%d %d", w, h, pad_w, pad_h, kernel_w, kernel_h, stride_w, stride_h);
811
812 const int kernel_extent_w = dilation_w * (kernel_w - 1) + 1;
813 const int kernel_extent_h = dilation_h * (kernel_h - 1) + 1;
814
815 int outw = (w - 1) * stride_w + kernel_extent_w;
816 int outh = (h - 1) * stride_h + kernel_extent_h;
817 int out_elempack = opt.use_packing_layout && num_output % 4 == 0 ? 4 : 1;
818 size_t out_elemsize = elemsize / elempack * out_elempack;
819
820 Mat top_blob_bordered;
821 if (pad_left > 0 || pad_right > 0 || pad_top > 0 || pad_bottom > 0 || output_pad_right > 0 || output_pad_bottom > 0 || (output_w > 0 && output_h > 0))
822 {
823 top_blob_bordered.create(outw, outh, num_output / out_elempack, out_elemsize, out_elempack, opt.workspace_allocator);
824 }
825 else
826 {
827 top_blob_bordered = top_blob;
828 top_blob_bordered.create(outw, outh, num_output / out_elempack, out_elemsize, out_elempack, opt.blob_allocator);
829 }
830 if (top_blob_bordered.empty())
831 return -100;
832
833 const int maxk = kernel_w * kernel_h;
834
835 if (elempack == 4 && out_elempack == 4)
836 {
837 {
838 // num_output
839 #pragma omp parallel for num_threads(opt.num_threads)
840 for (int p = 0; p < num_output / out_elempack; p++)
841 {
842 __fp16* outptr = top_blob_bordered.channel(p);
843
844 for (int i = 0; i < outh; i++)
845 {
846 for (int j = 0; j < outw; j++)
847 {
848 float32x4_t _sum = vdupq_n_f32(0.f);
849
850 if (bias_term)
851 {
852 _sum = vld1q_f32(((const float*)bias_data) + p * 4);
853 }
854
855 const __fp16* kptr = weight_data_fp16.channel(p);
856
857 // channels
858 for (int q = 0; q < channels; q++)
859 {
860 const Mat m = bottom_blob.channel(q);
861
862 for (int y = 0; y < kernel_h; y++)
863 {
864 int sys = (i + y * dilation_h - (kernel_extent_h - 1));
865 if (sys < 0 || sys % stride_h != 0)
866 continue;
867
868 int sy = sys / stride_h;
869 if (sy >= h)
870 continue;
871
872 for (int x = 0; x < kernel_w; x++)
873 {
874 int sxs = (j + x * dilation_w - (kernel_extent_w - 1));
875 if (sxs < 0 || sxs % stride_w != 0)
876 continue;
877
878 int sx = sxs / stride_w;
879 if (sx >= w)
880 continue;
881
882 const __fp16* sptr = m.row<const __fp16>(sy) + sx * 4;
883
884 float32x4_t _val = vcvt_f32_f16(vld1_f16(sptr));
885
886 int k = y * kernel_w + x;
887
888 float32x4_t _w0 = vcvt_f32_f16(vld1_f16(kptr + k * 16));
889 float32x4_t _w1 = vcvt_f32_f16(vld1_f16(kptr + k * 16 + 4));
890 float32x4_t _w2 = vcvt_f32_f16(vld1_f16(kptr + k * 16 + 8));
891 float32x4_t _w3 = vcvt_f32_f16(vld1_f16(kptr + k * 16 + 12));
892
893 _sum = vfmaq_laneq_f32(_sum, _w0, _val, 0);
894 _sum = vfmaq_laneq_f32(_sum, _w1, _val, 1);
895 _sum = vfmaq_laneq_f32(_sum, _w2, _val, 2);
896 _sum = vfmaq_laneq_f32(_sum, _w3, _val, 3);
897 }
898 }
899
900 kptr += maxk * 16;
901 }
902
903 _sum = activation_ps(_sum, activation_type, activation_params);
904
905 vst1_f16(outptr + j * 4, vcvt_f16_f32(_sum));
906 }
907
908 outptr += outw * 4;
909 }
910 }
911 }
912 }
913
914 if (elempack == 1 && out_elempack == 4)
915 {
916 {
917 // num_output
918 #pragma omp parallel for num_threads(opt.num_threads)
919 for (int p = 0; p < num_output / out_elempack; p++)
920 {
921 __fp16* outptr = top_blob_bordered.channel(p);
922
923 for (int i = 0; i < outh; i++)
924 {
925 for (int j = 0; j < outw; j++)
926 {
927 float32x4_t _sum = vdupq_n_f32(0.f);
928
929 if (bias_term)
930 {
931 _sum = vld1q_f32(((const float*)bias_data) + p * 4);
932 }
933
934 const __fp16* kptr = weight_data_fp16.channel(p);
935
936 // channels
937 for (int q = 0; q < channels; q++)
938 {
939 const Mat m = bottom_blob.channel(q);
940
941 for (int y = 0; y < kernel_h; y++)
942 {
943 int sys = (i + y * dilation_h - (kernel_extent_h - 1));
944 if (sys < 0 || sys % stride_h != 0)
945 continue;
946
947 int sy = sys / stride_h;
948 if (sy >= h)
949 continue;
950
951 const __fp16* sptr = m.row<const __fp16>(sy);
952
953 for (int x = 0; x < kernel_w; x++)
954 {
955 int sxs = (j + x * dilation_w - (kernel_extent_w - 1));
956 if (sxs < 0 || sxs % stride_w != 0)
957 continue;
958
959 int sx = sxs / stride_w;
960 if (sx >= w)
961 continue;
962
963 float32x4_t _val = vdupq_n_f32((float)sptr[sx]);
964
965 int k = y * kernel_w + x;
966
967 float32x4_t _w = vcvt_f32_f16(vld1_f16(kptr + k * 4));
968
969 _sum = vfmaq_f32(_sum, _val, _w);
970 }
971 }
972
973 kptr += maxk * 4;
974 }
975
976 _sum = activation_ps(_sum, activation_type, activation_params);
977
978 vst1_f16(outptr + j * 4, vcvt_f16_f32(_sum));
979 }
980
981 outptr += outw * 4;
982 }
983 }
984 }
985 }
986
987 if (elempack == 4 && out_elempack == 1)
988 {
989 {
990 // num_output
991 #pragma omp parallel for num_threads(opt.num_threads)
992 for (int p = 0; p < num_output / out_elempack; p++)
993 {
994 __fp16* outptr = top_blob_bordered.channel(p);
995
996 for (int i = 0; i < outh; i++)
997 {
998 for (int j = 0; j < outw; j++)
999 {
1000 float sum = 0.f;
1001
1002 if (bias_term)
1003 {
1004 sum = bias_data[p];
1005 }
1006
1007 const __fp16* kptr = weight_data_fp16.channel(p);
1008
1009 // channels
1010 for (int q = 0; q < channels; q++)
1011 {
1012 const Mat m = bottom_blob.channel(q);
1013
1014 for (int y = 0; y < kernel_h; y++)
1015 {
1016 int sys = (i + y * dilation_h - (kernel_extent_h - 1));
1017 if (sys < 0 || sys % stride_h != 0)
1018 continue;
1019
1020 int sy = sys / stride_h;
1021 if (sy >= h)
1022 continue;
1023
1024 for (int x = 0; x < kernel_w; x++)
1025 {
1026 int sxs = (j + x * dilation_w - (kernel_extent_w - 1));
1027 if (sxs < 0 || sxs % stride_w != 0)
1028 continue;
1029
1030 int sx = sxs / stride_w;
1031 if (sx >= w)
1032 continue;
1033
1034 const __fp16* sptr = m.row<const __fp16>(sy) + sx * 4;
1035
1036 float32x4_t _val = vcvt_f32_f16(vld1_f16(sptr));
1037
1038 int k = y * kernel_w + x;
1039
1040 float32x4_t _w = vcvt_f32_f16(vld1_f16(kptr + k * 4));
1041
1042 float32x4_t _s4 = vmulq_f32(_val, _w);
1043
1044 sum += vaddvq_f32(_s4); // dot
1045 }
1046 }
1047
1048 kptr += maxk * 4;
1049 }
1050
1051 sum = activation_ss(sum, activation_type, activation_params);
1052
1053 outptr[j] = (__fp16)sum;
1054 }
1055
1056 outptr += outw;
1057 }
1058 }
1059 }
1060 }
1061
1062 if (elempack == 1 && out_elempack == 1)
1063 {
1064 {
1065 // num_output
1066 #pragma omp parallel for num_threads(opt.num_threads)
1067 for (int p = 0; p < num_output; p++)
1068 {
1069 __fp16* outptr = top_blob_bordered.channel(p);
1070
1071 for (int i = 0; i < outh; i++)
1072 {
1073 for (int j = 0; j < outw; j++)
1074 {
1075 float sum = 0.f;
1076
1077 if (bias_term)
1078 {
1079 sum = bias_data[p];
1080 }
1081
1082 const __fp16* kptr = weight_data_fp16.channel(p);
1083
1084 // channels
1085 for (int q = 0; q < channels; q++)
1086 {
1087 const Mat m = bottom_blob.channel(q);
1088
1089 for (int y = 0; y < kernel_h; y++)
1090 {
1091 int sys = (i + y * dilation_h - (kernel_extent_h - 1));
1092 if (sys < 0 || sys % stride_h != 0)
1093 continue;
1094
1095 int sy = sys / stride_h;
1096 if (sy >= h)
1097 continue;
1098
1099 const __fp16* sptr = m.row<const __fp16>(sy);
1100
1101 for (int x = 0; x < kernel_w; x++)
1102 {
1103 int sxs = (j + x * dilation_w - (kernel_extent_w - 1));
1104 if (sxs < 0 || sxs % stride_w != 0)
1105 continue;
1106
1107 int sx = sxs / stride_w;
1108 if (sx >= w)
1109 continue;
1110
1111 float val = (float)sptr[sx];
1112
1113 int k = y * kernel_w + x;
1114
1115 float w = (float)kptr[k];
1116
1117 sum += val * w;
1118 }
1119 }
1120
1121 kptr += maxk;
1122 }
1123
1124 sum = activation_ss(sum, activation_type, activation_params);
1125
1126 outptr[j] = (__fp16)sum;
1127 }
1128
1129 outptr += outw;
1130 }
1131 }
1132 }
1133 }
1134
1135 cut_padding(top_blob_bordered, top_blob, opt);
1136 if (top_blob.empty())
1137 return -100;
1138
1139 return 0;
1140 }
1141
forward_fp16sa(const Mat & bottom_blob,Mat & top_blob,const Option & opt) const1142 int Deconvolution_arm::forward_fp16sa(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const
1143 {
1144 // deconvolv with NxN kernel
1145 // value = value + bias
1146
1147 int w = bottom_blob.w;
1148 int h = bottom_blob.h;
1149 int channels = bottom_blob.c;
1150 size_t elemsize = bottom_blob.elemsize;
1151 int elempack = bottom_blob.elempack;
1152
1153 // NCNN_LOGE("Deconvolution input %d x %d pad = %d %d ksize=%d %d stride=%d %d", w, h, pad_w, pad_h, kernel_w, kernel_h, stride_w, stride_h);
1154
1155 const int kernel_extent_w = dilation_w * (kernel_w - 1) + 1;
1156 const int kernel_extent_h = dilation_h * (kernel_h - 1) + 1;
1157
1158 int outw = (w - 1) * stride_w + kernel_extent_w;
1159 int outh = (h - 1) * stride_h + kernel_extent_h;
1160 int out_elempack = 1;
1161 if (opt.use_packing_layout)
1162 {
1163 out_elempack = opt.use_fp16_arithmetic && num_output % 8 == 0 ? 8 : num_output % 4 == 0 ? 4 : 1;
1164 }
1165 size_t out_elemsize = elemsize / elempack * out_elempack;
1166
1167 Mat top_blob_bordered;
1168 if (pad_left > 0 || pad_right > 0 || pad_top > 0 || pad_bottom > 0 || output_pad_right > 0 || output_pad_bottom > 0 || (output_w > 0 && output_h > 0))
1169 {
1170 top_blob_bordered.create(outw, outh, num_output / out_elempack, out_elemsize, out_elempack, opt.workspace_allocator);
1171 }
1172 else
1173 {
1174 top_blob_bordered = top_blob;
1175 top_blob_bordered.create(outw, outh, num_output / out_elempack, out_elemsize, out_elempack, opt.blob_allocator);
1176 }
1177 if (top_blob_bordered.empty())
1178 return -100;
1179
1180 const int maxk = kernel_w * kernel_h;
1181
1182 if (elempack == 8 && out_elempack == 8)
1183 {
1184 {
1185 // num_output
1186 #pragma omp parallel for num_threads(opt.num_threads)
1187 for (int p = 0; p < num_output / out_elempack; p++)
1188 {
1189 __fp16* outptr = top_blob_bordered.channel(p);
1190
1191 for (int i = 0; i < outh; i++)
1192 {
1193 for (int j = 0; j < outw; j++)
1194 {
1195 float16x8_t _sum = vdupq_n_f16((__fp16)0.f);
1196
1197 if (bias_term)
1198 {
1199 _sum = vld1q_f16((const __fp16*)bias_data_fp16 + p * 8);
1200 }
1201
1202 const __fp16* kptr = weight_data_fp16.channel(p);
1203
1204 // channels
1205 for (int q = 0; q < channels; q++)
1206 {
1207 const Mat m = bottom_blob.channel(q);
1208
1209 for (int y = 0; y < kernel_h; y++)
1210 {
1211 int sys = (i + y * dilation_h - (kernel_extent_h - 1));
1212 if (sys < 0 || sys % stride_h != 0)
1213 continue;
1214
1215 int sy = sys / stride_h;
1216 if (sy >= h)
1217 continue;
1218
1219 for (int x = 0; x < kernel_w; x++)
1220 {
1221 int sxs = (j + x * dilation_w - (kernel_extent_w - 1));
1222 if (sxs < 0 || sxs % stride_w != 0)
1223 continue;
1224
1225 int sx = sxs / stride_w;
1226 if (sx >= w)
1227 continue;
1228
1229 const __fp16* sptr = m.row<const __fp16>(sy) + sx * 8;
1230
1231 float16x8_t _val = vld1q_f16(sptr);
1232
1233 int k = y * kernel_w + x;
1234
1235 float16x8_t _w0 = vld1q_f16(kptr + k * 64);
1236 float16x8_t _w1 = vld1q_f16(kptr + k * 64 + 8);
1237 float16x8_t _w2 = vld1q_f16(kptr + k * 64 + 16);
1238 float16x8_t _w3 = vld1q_f16(kptr + k * 64 + 24);
1239 float16x8_t _w4 = vld1q_f16(kptr + k * 64 + 32);
1240 float16x8_t _w5 = vld1q_f16(kptr + k * 64 + 40);
1241 float16x8_t _w6 = vld1q_f16(kptr + k * 64 + 48);
1242 float16x8_t _w7 = vld1q_f16(kptr + k * 64 + 56);
1243
1244 _sum = vfmaq_laneq_f16(_sum, _w0, _val, 0);
1245 _sum = vfmaq_laneq_f16(_sum, _w1, _val, 1);
1246 _sum = vfmaq_laneq_f16(_sum, _w2, _val, 2);
1247 _sum = vfmaq_laneq_f16(_sum, _w3, _val, 3);
1248 _sum = vfmaq_laneq_f16(_sum, _w4, _val, 4);
1249 _sum = vfmaq_laneq_f16(_sum, _w5, _val, 5);
1250 _sum = vfmaq_laneq_f16(_sum, _w6, _val, 6);
1251 _sum = vfmaq_laneq_f16(_sum, _w7, _val, 7);
1252 }
1253 }
1254
1255 kptr += maxk * 64;
1256 }
1257
1258 _sum = activation_ps(_sum, activation_type, activation_params);
1259
1260 vst1q_f16(outptr + j * 8, _sum);
1261 }
1262
1263 outptr += outw * 8;
1264 }
1265 }
1266 }
1267 }
1268
1269 if (elempack == 1 && out_elempack == 8)
1270 {
1271 {
1272 // num_output
1273 #pragma omp parallel for num_threads(opt.num_threads)
1274 for (int p = 0; p < num_output / out_elempack; p++)
1275 {
1276 __fp16* outptr = top_blob_bordered.channel(p);
1277
1278 for (int i = 0; i < outh; i++)
1279 {
1280 for (int j = 0; j < outw; j++)
1281 {
1282 float16x8_t _sum = vdupq_n_f16((__fp16)0.f);
1283
1284 if (bias_term)
1285 {
1286 _sum = vld1q_f16((const __fp16*)bias_data_fp16 + p * 8);
1287 }
1288
1289 const __fp16* kptr = weight_data_fp16.channel(p);
1290
1291 // channels
1292 for (int q = 0; q < channels; q++)
1293 {
1294 const Mat m = bottom_blob.channel(q);
1295
1296 for (int y = 0; y < kernel_h; y++)
1297 {
1298 int sys = (i + y * dilation_h - (kernel_extent_h - 1));
1299 if (sys < 0 || sys % stride_h != 0)
1300 continue;
1301
1302 int sy = sys / stride_h;
1303 if (sy >= h)
1304 continue;
1305
1306 const __fp16* sptr = m.row<const __fp16>(sy);
1307
1308 for (int x = 0; x < kernel_w; x++)
1309 {
1310 int sxs = (j + x * dilation_w - (kernel_extent_w - 1));
1311 if (sxs < 0 || sxs % stride_w != 0)
1312 continue;
1313
1314 int sx = sxs / stride_w;
1315 if (sx >= w)
1316 continue;
1317
1318 float16x8_t _val = vdupq_n_f16(sptr[sx]);
1319
1320 int k = y * kernel_w + x;
1321
1322 float16x8_t _w = vld1q_f16(kptr + k * 8);
1323
1324 _sum = vfmaq_f16(_sum, _val, _w);
1325 }
1326 }
1327
1328 kptr += maxk * 8;
1329 }
1330
1331 _sum = activation_ps(_sum, activation_type, activation_params);
1332
1333 vst1q_f16(outptr + j * 8, _sum);
1334 }
1335
1336 outptr += outw * 8;
1337 }
1338 }
1339 }
1340 }
1341
1342 if (elempack == 4 && out_elempack == 8)
1343 {
1344 {
1345 // num_output
1346 #pragma omp parallel for num_threads(opt.num_threads)
1347 for (int p = 0; p < num_output / out_elempack; p++)
1348 {
1349 __fp16* outptr = top_blob_bordered.channel(p);
1350
1351 for (int i = 0; i < outh; i++)
1352 {
1353 for (int j = 0; j < outw; j++)
1354 {
1355 float16x8_t _sum = vdupq_n_f16((__fp16)0.f);
1356
1357 if (bias_term)
1358 {
1359 _sum = vld1q_f16((const __fp16*)bias_data_fp16 + p * 8);
1360 }
1361
1362 const __fp16* kptr = weight_data_fp16.channel(p);
1363
1364 // channels
1365 for (int q = 0; q < channels; q++)
1366 {
1367 const Mat m = bottom_blob.channel(q);
1368
1369 for (int y = 0; y < kernel_h; y++)
1370 {
1371 int sys = (i + y * dilation_h - (kernel_extent_h - 1));
1372 if (sys < 0 || sys % stride_h != 0)
1373 continue;
1374
1375 int sy = sys / stride_h;
1376 if (sy >= h)
1377 continue;
1378
1379 for (int x = 0; x < kernel_w; x++)
1380 {
1381 int sxs = (j + x * dilation_w - (kernel_extent_w - 1));
1382 if (sxs < 0 || sxs % stride_w != 0)
1383 continue;
1384
1385 int sx = sxs / stride_w;
1386 if (sx >= w)
1387 continue;
1388
1389 const __fp16* sptr = m.row<const __fp16>(sy) + sx * 4;
1390
1391 float16x4_t _val = vld1_f16(sptr);
1392
1393 int k = y * kernel_w + x;
1394
1395 float16x8_t _w0 = vld1q_f16(kptr + k * 32);
1396 float16x8_t _w1 = vld1q_f16(kptr + k * 32 + 8);
1397 float16x8_t _w2 = vld1q_f16(kptr + k * 32 + 16);
1398 float16x8_t _w3 = vld1q_f16(kptr + k * 32 + 24);
1399
1400 _sum = vfmaq_lane_f16(_sum, _w0, _val, 0);
1401 _sum = vfmaq_lane_f16(_sum, _w1, _val, 1);
1402 _sum = vfmaq_lane_f16(_sum, _w2, _val, 2);
1403 _sum = vfmaq_lane_f16(_sum, _w3, _val, 3);
1404 }
1405 }
1406
1407 kptr += maxk * 32;
1408 }
1409
1410 _sum = activation_ps(_sum, activation_type, activation_params);
1411
1412 vst1q_f16(outptr + j * 8, _sum);
1413 }
1414
1415 outptr += outw * 8;
1416 }
1417 }
1418 }
1419 }
1420
1421 if (elempack == 8 && out_elempack == 1)
1422 {
1423 {
1424 // num_output
1425 #pragma omp parallel for num_threads(opt.num_threads)
1426 for (int p = 0; p < num_output / out_elempack; p++)
1427 {
1428 __fp16* outptr = top_blob_bordered.channel(p);
1429
1430 for (int i = 0; i < outh; i++)
1431 {
1432 for (int j = 0; j < outw; j++)
1433 {
1434 float sum = 0.f;
1435
1436 if (bias_term)
1437 {
1438 sum = bias_data[p];
1439 }
1440
1441 const __fp16* kptr = weight_data_fp16.channel(p);
1442
1443 // channels
1444 for (int q = 0; q < channels; q++)
1445 {
1446 const Mat m = bottom_blob.channel(q);
1447
1448 for (int y = 0; y < kernel_h; y++)
1449 {
1450 int sys = (i + y * dilation_h - (kernel_extent_h - 1));
1451 if (sys < 0 || sys % stride_h != 0)
1452 continue;
1453
1454 int sy = sys / stride_h;
1455 if (sy >= h)
1456 continue;
1457
1458 for (int x = 0; x < kernel_w; x++)
1459 {
1460 int sxs = (j + x * dilation_w - (kernel_extent_w - 1));
1461 if (sxs < 0 || sxs % stride_w != 0)
1462 continue;
1463
1464 int sx = sxs / stride_w;
1465 if (sx >= w)
1466 continue;
1467
1468 const __fp16* sptr = m.row<const __fp16>(sy) + sx * 8;
1469
1470 float16x8_t _val = vld1q_f16(sptr);
1471
1472 int k = y * kernel_w + x;
1473
1474 float16x8_t _w = vld1q_f16(kptr + k * 8);
1475
1476 float16x8_t _s8 = vmulq_f16(_val, _w);
1477
1478 float16x4_t _s4 = vadd_f16(vget_low_f16(_s8), vget_high_f16(_s8));
1479 sum += vaddvq_f32(vcvt_f32_f16(_s4)); // dot
1480 }
1481 }
1482
1483 kptr += maxk * 8;
1484 }
1485
1486 sum = activation_ss(sum, activation_type, activation_params);
1487
1488 outptr[j] = (__fp16)sum;
1489 }
1490
1491 outptr += outw;
1492 }
1493 }
1494 }
1495 }
1496
1497 if (elempack == 8 && out_elempack == 4)
1498 {
1499 {
1500 // num_output
1501 #pragma omp parallel for num_threads(opt.num_threads)
1502 for (int p = 0; p < num_output / out_elempack; p++)
1503 {
1504 __fp16* outptr = top_blob_bordered.channel(p);
1505
1506 for (int i = 0; i < outh; i++)
1507 {
1508 for (int j = 0; j < outw; j++)
1509 {
1510 float16x4_t _sum = vdup_n_f16((__fp16)0.f);
1511
1512 if (bias_term)
1513 {
1514 _sum = vld1_f16((const __fp16*)bias_data_fp16 + p * 4);
1515 }
1516
1517 const __fp16* kptr = weight_data_fp16.channel(p);
1518
1519 // channels
1520 for (int q = 0; q < channels; q++)
1521 {
1522 const Mat m = bottom_blob.channel(q);
1523
1524 for (int y = 0; y < kernel_h; y++)
1525 {
1526 int sys = (i + y * dilation_h - (kernel_extent_h - 1));
1527 if (sys < 0 || sys % stride_h != 0)
1528 continue;
1529
1530 int sy = sys / stride_h;
1531 if (sy >= h)
1532 continue;
1533
1534 for (int x = 0; x < kernel_w; x++)
1535 {
1536 int sxs = (j + x * dilation_w - (kernel_extent_w - 1));
1537 if (sxs < 0 || sxs % stride_w != 0)
1538 continue;
1539
1540 int sx = sxs / stride_w;
1541 if (sx >= w)
1542 continue;
1543
1544 const __fp16* sptr = m.row<const __fp16>(sy) + sx * 8;
1545
1546 float16x8_t _val = vld1q_f16(sptr);
1547
1548 int k = y * kernel_w + x;
1549
1550 float16x4_t _w0 = vld1_f16(kptr + k * 32);
1551 float16x4_t _w1 = vld1_f16(kptr + k * 32 + 4);
1552 float16x4_t _w2 = vld1_f16(kptr + k * 32 + 8);
1553 float16x4_t _w3 = vld1_f16(kptr + k * 32 + 12);
1554 float16x4_t _w4 = vld1_f16(kptr + k * 32 + 16);
1555 float16x4_t _w5 = vld1_f16(kptr + k * 32 + 20);
1556 float16x4_t _w6 = vld1_f16(kptr + k * 32 + 24);
1557 float16x4_t _w7 = vld1_f16(kptr + k * 32 + 28);
1558
1559 _sum = vfma_laneq_f16(_sum, _w0, _val, 0);
1560 _sum = vfma_laneq_f16(_sum, _w1, _val, 1);
1561 _sum = vfma_laneq_f16(_sum, _w2, _val, 2);
1562 _sum = vfma_laneq_f16(_sum, _w3, _val, 3);
1563 _sum = vfma_laneq_f16(_sum, _w4, _val, 4);
1564 _sum = vfma_laneq_f16(_sum, _w5, _val, 5);
1565 _sum = vfma_laneq_f16(_sum, _w6, _val, 6);
1566 _sum = vfma_laneq_f16(_sum, _w7, _val, 7);
1567 }
1568 }
1569
1570 kptr += maxk * 32;
1571 }
1572
1573 _sum = activation_ps(_sum, activation_type, activation_params);
1574
1575 vst1_f16(outptr + j * 4, _sum);
1576 }
1577
1578 outptr += outw * 4;
1579 }
1580 }
1581 }
1582 }
1583
1584 if (elempack == 4 && out_elempack == 4)
1585 {
1586 {
1587 // num_output
1588 #pragma omp parallel for num_threads(opt.num_threads)
1589 for (int p = 0; p < num_output / out_elempack; p++)
1590 {
1591 __fp16* outptr = top_blob_bordered.channel(p);
1592
1593 for (int i = 0; i < outh; i++)
1594 {
1595 for (int j = 0; j < outw; j++)
1596 {
1597 float16x4_t _sum = vdup_n_f16((__fp16)0.f);
1598
1599 if (bias_term)
1600 {
1601 _sum = vld1_f16((const __fp16*)bias_data_fp16 + p * 4);
1602 }
1603
1604 const __fp16* kptr = weight_data_fp16.channel(p);
1605
1606 // channels
1607 for (int q = 0; q < channels; q++)
1608 {
1609 const Mat m = bottom_blob.channel(q);
1610
1611 for (int y = 0; y < kernel_h; y++)
1612 {
1613 int sys = (i + y * dilation_h - (kernel_extent_h - 1));
1614 if (sys < 0 || sys % stride_h != 0)
1615 continue;
1616
1617 int sy = sys / stride_h;
1618 if (sy >= h)
1619 continue;
1620
1621 for (int x = 0; x < kernel_w; x++)
1622 {
1623 int sxs = (j + x * dilation_w - (kernel_extent_w - 1));
1624 if (sxs < 0 || sxs % stride_w != 0)
1625 continue;
1626
1627 int sx = sxs / stride_w;
1628 if (sx >= w)
1629 continue;
1630
1631 const __fp16* sptr = m.row<const __fp16>(sy) + sx * 4;
1632
1633 float16x4_t _val = vld1_f16(sptr);
1634
1635 int k = y * kernel_w + x;
1636
1637 float16x4_t _w0 = vld1_f16(kptr + k * 16);
1638 float16x4_t _w1 = vld1_f16(kptr + k * 16 + 4);
1639 float16x4_t _w2 = vld1_f16(kptr + k * 16 + 8);
1640 float16x4_t _w3 = vld1_f16(kptr + k * 16 + 12);
1641
1642 _sum = vfma_lane_f16(_sum, _w0, _val, 0);
1643 _sum = vfma_lane_f16(_sum, _w1, _val, 1);
1644 _sum = vfma_lane_f16(_sum, _w2, _val, 2);
1645 _sum = vfma_lane_f16(_sum, _w3, _val, 3);
1646 }
1647 }
1648
1649 kptr += maxk * 16;
1650 }
1651
1652 _sum = activation_ps(_sum, activation_type, activation_params);
1653
1654 vst1_f16(outptr + j * 4, _sum);
1655 }
1656
1657 outptr += outw * 4;
1658 }
1659 }
1660 }
1661 }
1662
1663 if (elempack == 1 && out_elempack == 4)
1664 {
1665 {
1666 // num_output
1667 #pragma omp parallel for num_threads(opt.num_threads)
1668 for (int p = 0; p < num_output / out_elempack; p++)
1669 {
1670 __fp16* outptr = top_blob_bordered.channel(p);
1671
1672 for (int i = 0; i < outh; i++)
1673 {
1674 for (int j = 0; j < outw; j++)
1675 {
1676 float16x4_t _sum = vdup_n_f16((__fp16)0.f);
1677
1678 if (bias_term)
1679 {
1680 _sum = vld1_f16((const __fp16*)bias_data_fp16 + p * 4);
1681 }
1682
1683 const __fp16* kptr = weight_data_fp16.channel(p);
1684
1685 // channels
1686 for (int q = 0; q < channels; q++)
1687 {
1688 const Mat m = bottom_blob.channel(q);
1689
1690 for (int y = 0; y < kernel_h; y++)
1691 {
1692 int sys = (i + y * dilation_h - (kernel_extent_h - 1));
1693 if (sys < 0 || sys % stride_h != 0)
1694 continue;
1695
1696 int sy = sys / stride_h;
1697 if (sy >= h)
1698 continue;
1699
1700 const __fp16* sptr = m.row<const __fp16>(sy);
1701
1702 for (int x = 0; x < kernel_w; x++)
1703 {
1704 int sxs = (j + x * dilation_w - (kernel_extent_w - 1));
1705 if (sxs < 0 || sxs % stride_w != 0)
1706 continue;
1707
1708 int sx = sxs / stride_w;
1709 if (sx >= w)
1710 continue;
1711
1712 float16x4_t _val = vdup_n_f16(sptr[sx]);
1713
1714 int k = y * kernel_w + x;
1715
1716 float16x4_t _w = vld1_f16(kptr + k * 4);
1717
1718 _sum = vfma_f16(_sum, _val, _w);
1719 }
1720 }
1721
1722 kptr += maxk * 4;
1723 }
1724
1725 _sum = activation_ps(_sum, activation_type, activation_params);
1726
1727 vst1_f16(outptr + j * 4, _sum);
1728 }
1729
1730 outptr += outw * 4;
1731 }
1732 }
1733 }
1734 }
1735
1736 if (elempack == 4 && out_elempack == 1)
1737 {
1738 {
1739 // num_output
1740 #pragma omp parallel for num_threads(opt.num_threads)
1741 for (int p = 0; p < num_output / out_elempack; p++)
1742 {
1743 __fp16* outptr = top_blob_bordered.channel(p);
1744
1745 for (int i = 0; i < outh; i++)
1746 {
1747 for (int j = 0; j < outw; j++)
1748 {
1749 float sum = 0.f;
1750
1751 if (bias_term)
1752 {
1753 sum = bias_data[p];
1754 }
1755
1756 const __fp16* kptr = weight_data_fp16.channel(p);
1757
1758 // channels
1759 for (int q = 0; q < channels; q++)
1760 {
1761 const Mat m = bottom_blob.channel(q);
1762
1763 for (int y = 0; y < kernel_h; y++)
1764 {
1765 int sys = (i + y * dilation_h - (kernel_extent_h - 1));
1766 if (sys < 0 || sys % stride_h != 0)
1767 continue;
1768
1769 int sy = sys / stride_h;
1770 if (sy >= h)
1771 continue;
1772
1773 for (int x = 0; x < kernel_w; x++)
1774 {
1775 int sxs = (j + x * dilation_w - (kernel_extent_w - 1));
1776 if (sxs < 0 || sxs % stride_w != 0)
1777 continue;
1778
1779 int sx = sxs / stride_w;
1780 if (sx >= w)
1781 continue;
1782
1783 const __fp16* sptr = m.row<const __fp16>(sy) + sx * 4;
1784
1785 float16x4_t _val = vld1_f16(sptr);
1786
1787 int k = y * kernel_w + x;
1788
1789 float16x4_t _w = vld1_f16(kptr + k * 4);
1790
1791 float16x4_t _s4 = vmul_f16(_val, _w);
1792
1793 sum += vaddvq_f32(vcvt_f32_f16(_s4)); // dot
1794 }
1795 }
1796
1797 kptr += maxk * 4;
1798 }
1799
1800 sum = activation_ss(sum, activation_type, activation_params);
1801
1802 outptr[j] = (__fp16)sum;
1803 }
1804
1805 outptr += outw;
1806 }
1807 }
1808 }
1809 }
1810
1811 if (elempack == 1 && out_elempack == 1)
1812 {
1813 if (kernel_w == 4 && kernel_h == 4 && stride_w == 2 && stride_h == 2 && dilation_w == 1 && dilation_h == 1)
1814 {
1815 deconv4x4s2_fp16sa_neon(bottom_blob, top_blob_bordered, weight_data_fp16, bias_data_fp16, opt);
1816
1817 if (activation)
1818 {
1819 activation->forward_inplace(top_blob_bordered, opt);
1820 }
1821 }
1822 else
1823 {
1824 // num_output
1825 #pragma omp parallel for num_threads(opt.num_threads)
1826 for (int p = 0; p < num_output; p++)
1827 {
1828 __fp16* outptr = top_blob_bordered.channel(p);
1829
1830 for (int i = 0; i < outh; i++)
1831 {
1832 for (int j = 0; j < outw; j++)
1833 {
1834 float sum = 0.f;
1835
1836 if (bias_term)
1837 {
1838 sum = bias_data[p];
1839 }
1840
1841 const __fp16* kptr = weight_data_fp16.channel(p);
1842
1843 // channels
1844 for (int q = 0; q < channels; q++)
1845 {
1846 const Mat m = bottom_blob.channel(q);
1847
1848 for (int y = 0; y < kernel_h; y++)
1849 {
1850 int sys = (i + y * dilation_h - (kernel_extent_h - 1));
1851 if (sys < 0 || sys % stride_h != 0)
1852 continue;
1853
1854 int sy = sys / stride_h;
1855 if (sy >= h)
1856 continue;
1857
1858 const __fp16* sptr = m.row<const __fp16>(sy);
1859
1860 for (int x = 0; x < kernel_w; x++)
1861 {
1862 int sxs = (j + x * dilation_w - (kernel_extent_w - 1));
1863 if (sxs < 0 || sxs % stride_w != 0)
1864 continue;
1865
1866 int sx = sxs / stride_w;
1867 if (sx >= w)
1868 continue;
1869
1870 __fp16 val = sptr[sx];
1871
1872 int k = y * kernel_w + x;
1873
1874 __fp16 w = kptr[k];
1875
1876 sum += val * w;
1877 }
1878 }
1879
1880 kptr += maxk;
1881 }
1882
1883 sum = activation_ss(sum, activation_type, activation_params);
1884
1885 outptr[j] = (__fp16)sum;
1886 }
1887
1888 outptr += outw;
1889 }
1890 }
1891 }
1892 }
1893
1894 cut_padding(top_blob_bordered, top_blob, opt);
1895 if (top_blob.empty())
1896 return -100;
1897
1898 return 0;
1899 }
1900 #endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
1901
create_pipeline_bf16s(const Option & opt)1902 int Deconvolution_arm::create_pipeline_bf16s(const Option& opt)
1903 {
1904 const int maxk = kernel_w * kernel_h;
1905 const int num_input = weight_data_size / maxk / num_output;
1906
1907 int elempack = opt.use_packing_layout && num_input % 4 == 0 ? 4 : 1;
1908 int out_elempack = opt.use_packing_layout && num_output % 4 == 0 ? 4 : 1;
1909
1910 Mat weight_data_transposed(weight_data.w);
1911 {
1912 float* pt = weight_data_transposed;
1913 const float* p = weight_data;
1914
1915 for (int i = 0; i < num_input * num_output; i++)
1916 {
1917 for (int k = 0; k < maxk; k++)
1918 {
1919 pt[maxk - 1 - k] = p[k];
1920 }
1921
1922 p += maxk;
1923 pt += maxk;
1924 }
1925 }
1926
1927 // src = kw-kh-inch-outch
1928 // dst = pb-pa-kw-kh-inch/pa-outch/pb
1929 {
1930 Mat weight_data_r2 = weight_data_transposed.reshape(maxk, num_input, num_output);
1931
1932 weight_data_bf16.create(maxk, num_input / elempack, num_output / out_elempack, (size_t)2u * elempack * out_elempack, elempack * out_elempack);
1933
1934 for (int q = 0; q + (out_elempack - 1) < num_output; q += out_elempack)
1935 {
1936 Mat g0 = weight_data_bf16.channel(q / out_elempack);
1937
1938 for (int p = 0; p + (elempack - 1) < num_input; p += elempack)
1939 {
1940 unsigned short* g00 = g0.row<unsigned short>(p / elempack);
1941
1942 for (int k = 0; k < maxk; k++)
1943 {
1944 for (int i = 0; i < elempack; i++)
1945 {
1946 for (int j = 0; j < out_elempack; j++)
1947 {
1948 const float* k00 = weight_data_r2.channel(q + j).row(p + i);
1949
1950 g00[0] = float32_to_bfloat16(k00[k]);
1951
1952 g00++;
1953 }
1954 }
1955 }
1956 }
1957 }
1958 }
1959
1960 return 0;
1961 }
1962
forward_bf16s(const Mat & bottom_blob,Mat & top_blob,const Option & opt) const1963 int Deconvolution_arm::forward_bf16s(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const
1964 {
1965 // deconvolv with NxN kernel
1966 // value = value + bias
1967
1968 int w = bottom_blob.w;
1969 int h = bottom_blob.h;
1970 int channels = bottom_blob.c;
1971 size_t elemsize = bottom_blob.elemsize;
1972 int elempack = bottom_blob.elempack;
1973
1974 // NCNN_LOGE("Deconvolution input %d x %d pad = %d %d ksize=%d %d stride=%d %d", w, h, pad_w, pad_h, kernel_w, kernel_h, stride_w, stride_h);
1975
1976 const int kernel_extent_w = dilation_w * (kernel_w - 1) + 1;
1977 const int kernel_extent_h = dilation_h * (kernel_h - 1) + 1;
1978
1979 int outw = (w - 1) * stride_w + kernel_extent_w;
1980 int outh = (h - 1) * stride_h + kernel_extent_h;
1981 int out_elempack = opt.use_packing_layout && num_output % 4 == 0 ? 4 : 1;
1982 size_t out_elemsize = elemsize / elempack * out_elempack;
1983
1984 Mat top_blob_bordered;
1985 if (pad_left > 0 || pad_right > 0 || pad_top > 0 || pad_bottom > 0 || output_pad_right > 0 || output_pad_bottom > 0 || (output_w > 0 && output_h > 0))
1986 {
1987 top_blob_bordered.create(outw, outh, num_output / out_elempack, out_elemsize, out_elempack, opt.workspace_allocator);
1988 }
1989 else
1990 {
1991 top_blob_bordered = top_blob;
1992 top_blob_bordered.create(outw, outh, num_output / out_elempack, out_elemsize, out_elempack, opt.blob_allocator);
1993 }
1994 if (top_blob_bordered.empty())
1995 return -100;
1996
1997 const int maxk = kernel_w * kernel_h;
1998
1999 #if __ARM_NEON
2000 if (elempack == 4 && out_elempack == 4)
2001 {
2002 {
2003 // num_output
2004 #pragma omp parallel for num_threads(opt.num_threads)
2005 for (int p = 0; p < num_output / out_elempack; p++)
2006 {
2007 unsigned short* outptr = top_blob_bordered.channel(p);
2008
2009 for (int i = 0; i < outh; i++)
2010 {
2011 for (int j = 0; j < outw; j++)
2012 {
2013 float32x4_t _sum = vdupq_n_f32(0.f);
2014
2015 if (bias_term)
2016 {
2017 _sum = vld1q_f32(((const float*)bias_data) + p * 4);
2018 }
2019
2020 const unsigned short* kptr = weight_data_bf16.channel(p);
2021
2022 // channels
2023 for (int q = 0; q < channels; q++)
2024 {
2025 const Mat m = bottom_blob.channel(q);
2026
2027 for (int y = 0; y < kernel_h; y++)
2028 {
2029 int sys = (i + y * dilation_h - (kernel_extent_h - 1));
2030 if (sys < 0 || sys % stride_h != 0)
2031 continue;
2032
2033 int sy = sys / stride_h;
2034 if (sy >= h)
2035 continue;
2036
2037 for (int x = 0; x < kernel_w; x++)
2038 {
2039 int sxs = (j + x * dilation_w - (kernel_extent_w - 1));
2040 if (sxs < 0 || sxs % stride_w != 0)
2041 continue;
2042
2043 int sx = sxs / stride_w;
2044 if (sx >= w)
2045 continue;
2046
2047 const unsigned short* sptr = m.row<const unsigned short>(sy) + sx * 4;
2048
2049 float32x4_t _val = vcvt_f32_bf16(vld1_u16(sptr));
2050
2051 int k = y * kernel_w + x;
2052
2053 float32x4_t _w0 = vcvt_f32_bf16(vld1_u16(kptr + k * 16));
2054 float32x4_t _w1 = vcvt_f32_bf16(vld1_u16(kptr + k * 16 + 4));
2055 float32x4_t _w2 = vcvt_f32_bf16(vld1_u16(kptr + k * 16 + 8));
2056 float32x4_t _w3 = vcvt_f32_bf16(vld1_u16(kptr + k * 16 + 12));
2057
2058 #if __aarch64__
2059 _sum = vmlaq_laneq_f32(_sum, _w0, _val, 0);
2060 _sum = vmlaq_laneq_f32(_sum, _w1, _val, 1);
2061 _sum = vmlaq_laneq_f32(_sum, _w2, _val, 2);
2062 _sum = vmlaq_laneq_f32(_sum, _w3, _val, 3);
2063 #else
2064 _sum = vmlaq_lane_f32(_sum, _w0, vget_low_f32(_val), 0);
2065 _sum = vmlaq_lane_f32(_sum, _w1, vget_low_f32(_val), 1);
2066 _sum = vmlaq_lane_f32(_sum, _w2, vget_high_f32(_val), 0);
2067 _sum = vmlaq_lane_f32(_sum, _w3, vget_high_f32(_val), 1);
2068 #endif
2069 }
2070 }
2071
2072 kptr += maxk * 16;
2073 }
2074
2075 _sum = activation_ps(_sum, activation_type, activation_params);
2076
2077 vst1_u16(outptr + j * 4, vcvt_bf16_f32(_sum));
2078 }
2079
2080 outptr += outw * 4;
2081 }
2082 }
2083 }
2084 }
2085
2086 if (elempack == 1 && out_elempack == 4)
2087 {
2088 {
2089 // num_output
2090 #pragma omp parallel for num_threads(opt.num_threads)
2091 for (int p = 0; p < num_output / out_elempack; p++)
2092 {
2093 unsigned short* outptr = top_blob_bordered.channel(p);
2094
2095 for (int i = 0; i < outh; i++)
2096 {
2097 for (int j = 0; j < outw; j++)
2098 {
2099 float32x4_t _sum = vdupq_n_f32(0.f);
2100
2101 if (bias_term)
2102 {
2103 _sum = vld1q_f32(((const float*)bias_data) + p * 4);
2104 }
2105
2106 const unsigned short* kptr = weight_data_bf16.channel(p);
2107
2108 // channels
2109 for (int q = 0; q < channels; q++)
2110 {
2111 const Mat m = bottom_blob.channel(q);
2112
2113 for (int y = 0; y < kernel_h; y++)
2114 {
2115 int sys = (i + y * dilation_h - (kernel_extent_h - 1));
2116 if (sys < 0 || sys % stride_h != 0)
2117 continue;
2118
2119 int sy = sys / stride_h;
2120 if (sy >= h)
2121 continue;
2122
2123 const unsigned short* sptr = m.row<const unsigned short>(sy);
2124
2125 for (int x = 0; x < kernel_w; x++)
2126 {
2127 int sxs = (j + x * dilation_w - (kernel_extent_w - 1));
2128 if (sxs < 0 || sxs % stride_w != 0)
2129 continue;
2130
2131 int sx = sxs / stride_w;
2132 if (sx >= w)
2133 continue;
2134
2135 float32x4_t _val = vdupq_n_f32(bfloat16_to_float32(sptr[sx]));
2136
2137 int k = y * kernel_w + x;
2138
2139 float32x4_t _w = vcvt_f32_bf16(vld1_u16(kptr + k * 4));
2140
2141 _sum = vmlaq_f32(_sum, _val, _w);
2142 }
2143 }
2144
2145 kptr += maxk * 4;
2146 }
2147
2148 _sum = activation_ps(_sum, activation_type, activation_params);
2149
2150 vst1_u16(outptr + j * 4, vcvt_bf16_f32(_sum));
2151 }
2152
2153 outptr += outw * 4;
2154 }
2155 }
2156 }
2157 }
2158
2159 if (elempack == 4 && out_elempack == 1)
2160 {
2161 {
2162 // num_output
2163 #pragma omp parallel for num_threads(opt.num_threads)
2164 for (int p = 0; p < num_output / out_elempack; p++)
2165 {
2166 unsigned short* outptr = top_blob_bordered.channel(p);
2167
2168 for (int i = 0; i < outh; i++)
2169 {
2170 for (int j = 0; j < outw; j++)
2171 {
2172 float sum = 0.f;
2173
2174 if (bias_term)
2175 {
2176 sum = bias_data[p];
2177 }
2178
2179 const unsigned short* kptr = weight_data_bf16.channel(p);
2180
2181 // channels
2182 for (int q = 0; q < channels; q++)
2183 {
2184 const Mat m = bottom_blob.channel(q);
2185
2186 for (int y = 0; y < kernel_h; y++)
2187 {
2188 int sys = (i + y * dilation_h - (kernel_extent_h - 1));
2189 if (sys < 0 || sys % stride_h != 0)
2190 continue;
2191
2192 int sy = sys / stride_h;
2193 if (sy >= h)
2194 continue;
2195
2196 for (int x = 0; x < kernel_w; x++)
2197 {
2198 int sxs = (j + x * dilation_w - (kernel_extent_w - 1));
2199 if (sxs < 0 || sxs % stride_w != 0)
2200 continue;
2201
2202 int sx = sxs / stride_w;
2203 if (sx >= w)
2204 continue;
2205
2206 const unsigned short* sptr = m.row<const unsigned short>(sy) + sx * 4;
2207
2208 float32x4_t _val = vcvt_f32_bf16(vld1_u16(sptr));
2209
2210 int k = y * kernel_w + x;
2211
2212 float32x4_t _w = vcvt_f32_bf16(vld1_u16(kptr + k * 4));
2213
2214 float32x4_t _s4 = vmulq_f32(_val, _w);
2215 #if __aarch64__
2216 sum += vaddvq_f32(_s4); // dot
2217 #else
2218 float32x2_t _ss = vadd_f32(vget_low_f32(_s4), vget_high_f32(_s4));
2219 _ss = vpadd_f32(_ss, _ss);
2220 sum += vget_lane_f32(_ss, 0);
2221 #endif
2222 }
2223 }
2224
2225 kptr += maxk * 4;
2226 }
2227
2228 sum = activation_ss(sum, activation_type, activation_params);
2229
2230 outptr[j] = float32_to_bfloat16(sum);
2231 }
2232
2233 outptr += outw;
2234 }
2235 }
2236 }
2237 }
2238 #endif // __ARM_NEON
2239
2240 if (elempack == 1 && out_elempack == 1)
2241 {
2242 {
2243 // num_output
2244 #pragma omp parallel for num_threads(opt.num_threads)
2245 for (int p = 0; p < num_output; p++)
2246 {
2247 unsigned short* outptr = top_blob_bordered.channel(p);
2248
2249 for (int i = 0; i < outh; i++)
2250 {
2251 for (int j = 0; j < outw; j++)
2252 {
2253 float sum = 0.f;
2254
2255 if (bias_term)
2256 {
2257 sum = bias_data[p];
2258 }
2259
2260 const unsigned short* kptr = weight_data_bf16.channel(p);
2261
2262 // channels
2263 for (int q = 0; q < channels; q++)
2264 {
2265 const Mat m = bottom_blob.channel(q);
2266
2267 for (int y = 0; y < kernel_h; y++)
2268 {
2269 int sys = (i + y * dilation_h - (kernel_extent_h - 1));
2270 if (sys < 0 || sys % stride_h != 0)
2271 continue;
2272
2273 int sy = sys / stride_h;
2274 if (sy >= h)
2275 continue;
2276
2277 const unsigned short* sptr = m.row<const unsigned short>(sy);
2278
2279 for (int x = 0; x < kernel_w; x++)
2280 {
2281 int sxs = (j + x * dilation_w - (kernel_extent_w - 1));
2282 if (sxs < 0 || sxs % stride_w != 0)
2283 continue;
2284
2285 int sx = sxs / stride_w;
2286 if (sx >= w)
2287 continue;
2288
2289 float val = bfloat16_to_float32(sptr[sx]);
2290
2291 int k = y * kernel_w + x;
2292
2293 float w = bfloat16_to_float32(kptr[k]);
2294
2295 sum += val * w;
2296 }
2297 }
2298
2299 kptr += maxk;
2300 }
2301
2302 if (activation_type == 1)
2303 {
2304 sum = std::max(sum, 0.f);
2305 }
2306 else if (activation_type == 2)
2307 {
2308 float slope = activation_params[0];
2309 sum = sum > 0.f ? sum : sum * slope;
2310 }
2311 else if (activation_type == 3)
2312 {
2313 float min = activation_params[0];
2314 float max = activation_params[1];
2315 if (sum < min)
2316 sum = min;
2317 if (sum > max)
2318 sum = max;
2319 }
2320 else if (activation_type == 4)
2321 {
2322 sum = static_cast<float>(1.f / (1.f + exp(-sum)));
2323 }
2324
2325 outptr[j] = float32_to_bfloat16(sum);
2326 }
2327
2328 outptr += outw;
2329 }
2330 }
2331 }
2332 }
2333
2334 cut_padding(top_blob_bordered, top_blob, opt);
2335 if (top_blob.empty())
2336 return -100;
2337
2338 return 0;
2339 }
2340
2341 } // namespace ncnn
2342