ncnn

1
// Tencent is pleased to support the open source community by making ncnn available.
2
//
3
// Copyright (C) 2021 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 "convolution_riscv.h"
16

17
#include "benchmark.h"
18
#include "cpu.h"
19
#include "layer_type.h"
20

21
#if __riscv_vector
22
#include <riscv_vector.h>
23
#endif // __riscv_vector
24

25
#include "riscv_activation.h"
26
#include "riscv_usability.h"
27

28
#include "cpu.h"
29

30
namespace ncnn {
31

32
#include "convolution_sgemm.h"
33
#include "convolution_winograd_transform.h"
34
#include "convolution_winograd_dot.h"
35
#include "convolution_1x1.h"
36
#include "convolution_3x3.h"
37

38
#if __riscv_vector
39
#include "convolution_packn.h"
40
#include "convolution_pack1ton.h"
41
#include "convolution_packnto1.h"
42

43
#include "convolution_sgemm_packn.h"
44
#include "convolution_sgemm_pack1ton.h"
45
#include "convolution_sgemm_packnto1.h"
46
#include "convolution_winograd_transform_packn.h"
47
#include "convolution_winograd_dot_packn.h"
48
#include "convolution_1x1_packn.h"
49
#include "convolution_1x1_pack1ton.h"
50
#include "convolution_1x1_packnto1.h"
51
#include "convolution_3x3_packn.h"
52
#include "convolution_3x3_pack1ton.h"
53
#include "convolution_7x7_pack1ton.h"
54

55
#if __riscv_zfh
56
#include "convolution_fp16s.h"
57
#include "convolution_packn_fp16s.h"
58
#include "convolution_pack1ton_fp16s.h"
59
#include "convolution_packnto1_fp16s.h"
60

61
#include "convolution_sgemm_fp16s.h"
62
#include "convolution_sgemm_packn_fp16s.h"
63
#include "convolution_sgemm_pack1ton_fp16s.h"
64
#include "convolution_sgemm_packnto1_fp16s.h"
65
#include "convolution_winograd_transform_packn_fp16s.h"
66
#include "convolution_winograd_dot_packn_fp16s.h"
67
#include "convolution_1x1_fp16s.h"
68
#include "convolution_1x1_packn_fp16s.h"
69
#include "convolution_1x1_pack1ton_fp16s.h"
70
#include "convolution_1x1_packnto1_fp16s.h"
71
#include "convolution_3x3_packn_fp16s.h"
72
#include "convolution_3x3_pack1ton_fp16s.h"
73
#include "convolution_7x7_pack1ton_fp16s.h"
74

75
#endif
76
#endif // __riscv_vector
77

78
Convolution_riscv::Convolution_riscv()
79
{
80
#if __riscv_vector
81
    support_packing = true;
82
#if __riscv_zfh
83
    support_fp16_storage = true;
84
#endif
85
#endif // __riscv_vector
86

87
    activation = 0;
88
}
89

90
static void convolution_transform_kernel_packed_rvv(const Mat& weight_data, Mat& weight_data_tm, int num_input, int num_output, int kernel_w, int kernel_h, int elempack, int out_elempack)
91
{
92
    const int maxk = kernel_w * kernel_h;
93

94
    // src = kw-kh-inch-outch
95
    // dst = pb-pa-kw-kh-inch/pa-outch/pb
96
    {
97
        Mat weight_data_r2 = weight_data.reshape(maxk, num_input, num_output);
98

99
        weight_data_tm.create(maxk, num_input / elempack, num_output / out_elempack, (size_t)4u * elempack * out_elempack, elempack * out_elempack);
100

101
        for (int q = 0; q + (out_elempack - 1) < num_output; q += out_elempack)
102
        {
103
            float* g00 = weight_data_tm.channel(q / out_elempack);
104

105
            for (int p = 0; p + (elempack - 1) < num_input; p += elempack)
106
            {
107
                for (int k = 0; k < maxk; k++)
108
                {
109
                    for (int i = 0; i < elempack; i++)
110
                    {
111
                        for (int j = 0; j < out_elempack; j++)
112
                        {
113
                            const float* k00 = weight_data_r2.channel(q + j).row(p + i);
114

115
                            g00[0] = k00[k];
116

117
                            g00++;
118
                        }
119
                    }
120
                }
121
            }
122
        }
123
    }
124
}
125

126
int Convolution_riscv::create_pipeline(const Option& opt)
127
{
128
    if (dynamic_weight)
129
        return 0;
130

131
    activation = create_activation_layer(activation_type, activation_params, opt);
132

133
#if NCNN_INT8
134
    if (opt.use_int8_inference && weight_data.elemsize == (size_t)1u)
135
    {
136
        // TODO implement int8
137
        return 0;
138
    }
139
#endif
140

141
#if __riscv_vector && __riscv_zfh
142
    if (opt.use_fp16_storage)
143
    {
144
        return create_pipeline_fp16s(opt);
145
    }
146
#endif
147

148
#if __riscv_vector
149
    const int packn = csrr_vlenb() / 4;
150
#endif
151

152
    const int maxk = kernel_w * kernel_h;
153
    const int num_input = weight_data_size / maxk / num_output;
154

155
    int elempack = 1;
156
    int out_elempack = 1;
157
#if __riscv_vector
158
    if (opt.use_packing_layout)
159
    {
160
        elempack = num_input % packn == 0 ? packn : 1;
161
        out_elempack = num_output % packn == 0 ? packn : 1;
162
    }
163
#endif
164

165
#if __riscv_vector
166
    // packn
167
    if (elempack == packn && out_elempack == packn)
168
    {
169
        if (opt.use_winograd_convolution && (opt.use_winograd23_convolution || opt.use_winograd43_convolution || opt.use_winograd63_convolution) && kernel_w == 3 && kernel_h == 3 && dilation_w == 1 && dilation_h == 1 && stride_w == 1 && stride_h == 1)
170
        {
171
            if ((opt.use_winograd63_convolution && num_input >= packn * 2 && num_output >= packn * 2 && num_input <= packn * 16 && num_output <= packn * 16) || (!opt.use_winograd43_convolution && !opt.use_winograd23_convolution))
172
                conv3x3s1_winograd63_transform_kernel_packn_rvv(weight_data, weight_winograd63_data, num_input, num_output, opt);
173
            else if ((opt.use_winograd43_convolution && num_input >= packn * 2 && num_output >= packn * 2) || (!opt.use_winograd63_convolution && !opt.use_winograd23_convolution))
174
                conv3x3s1_winograd43_transform_kernel_packn_rvv(weight_data, weight_winograd43_data, num_input, num_output, opt);
175
            else // if (opt.use_winograd23_convolution)
176
                conv3x3s1_winograd23_transform_kernel_packn_rvv(weight_data, weight_winograd23_data, num_input, num_output, opt);
177
        }
178
        else
179
        {
180
            convolution_transform_kernel_packed_rvv(weight_data, weight_data_tm, num_input, num_output, kernel_w, kernel_h, elempack, out_elempack);
181
        }
182
    }
183

184
    // pack1ton
185
    if (elempack == 1 && out_elempack == packn)
186
    {
187
        convolution_transform_kernel_packed_rvv(weight_data, weight_data_tm, num_input, num_output, kernel_w, kernel_h, elempack, out_elempack);
188
    }
189

190
    // packnto1
191
    if (elempack == packn && out_elempack == 1)
192
    {
193
        if (kernel_w == 1 && kernel_h == 1 && dilation_w == 1 && dilation_h == 1 && stride_w == 1 && stride_h == 1)
194
        {
195
            convolution_im2col_sgemm_transform_kernel_packnto1_rvv(weight_data, weight_data_tm, num_input, num_output, kernel_w, kernel_h);
196
        }
197
        else if (kernel_w == 1 && kernel_h == 1 && dilation_w == 1 && dilation_h == 1 && stride_w == 2 && stride_h == 2)
198
        {
199
            convolution_im2col_sgemm_transform_kernel_packnto1_rvv(weight_data, weight_data_tm, num_input, num_output, kernel_w, kernel_h);
200
        }
201
        else if (opt.use_sgemm_convolution)
202
        {
203
            convolution_im2col_sgemm_transform_kernel_packnto1_rvv(weight_data, weight_data_tm, num_input, num_output, kernel_w, kernel_h);
204
        }
205
        else
206
        {
207
            convolution_transform_kernel_packed_rvv(weight_data, weight_data_tm, num_input, num_output, kernel_w, kernel_h, elempack, out_elempack);
208
        }
209
    }
210
#endif // __riscv_vector
211

212
    // pack1
213
    if (elempack == 1 && out_elempack == 1)
214
    {
215
        if (kernel_w == 1 && kernel_h == 1 && dilation_w == 1 && dilation_h == 1 && stride_w == 1 && stride_h == 1)
216
        {
217
            convolution_im2col_sgemm_transform_kernel_rvv(weight_data, weight_data_tm, num_input, num_output, kernel_w, kernel_h);
218
        }
219
        else if (opt.use_winograd_convolution && (opt.use_winograd23_convolution || opt.use_winograd43_convolution) && kernel_w == 3 && kernel_h == 3 && dilation_w == 1 && dilation_h == 1 && stride_w == 1 && stride_h == 1)
220
        {
221
            if ((opt.use_winograd43_convolution && num_input >= 16 && num_output >= 16) || !opt.use_winograd23_convolution)
222
            {
223
                conv3x3s1_winograd43_transform_kernel_rvv(weight_data, weight_winograd43_data, num_input, num_output, opt);
224
            }
225
            else if (opt.use_winograd23_convolution)
226
            {
227
                conv3x3s1_winograd23_transform_kernel_rvv(weight_data, weight_winograd23_data, num_input, num_output, opt);
228
            }
229
        }
230
        else if (opt.use_sgemm_convolution)
231
        {
232
            convolution_im2col_sgemm_transform_kernel_rvv(weight_data, weight_data_tm, num_input, num_output, kernel_w, kernel_h);
233
        }
234
        else
235
        {
236
            weight_data_tm = weight_data;
237
        }
238
    }
239

240
    if (opt.lightmode)
241
        weight_data.release();
242

243
    return 0;
244
}
245

246
int Convolution_riscv::destroy_pipeline(const Option& opt)
247
{
248
    if (activation)
249
    {
250
        activation->destroy_pipeline(opt);
251
        delete activation;
252
        activation = 0;
253
    }
254

255
    return 0;
256
}
257

258
int Convolution_riscv::forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const
259
{
260
#if NCNN_INT8
261
    if (opt.use_int8_inference && int8_scale_term)
262
    {
263
        Mat bottom_blob_unpacked = bottom_blob;
264
        if (bottom_blob.elempack != 1)
265
        {
266
            Option opt_pack1 = opt;
267
            opt_pack1.blob_allocator = opt.workspace_allocator;
268

269
            convert_packing(bottom_blob, bottom_blob_unpacked, 1, opt_pack1);
270
        }
271

272
        Mat bottom_blob_unpacked_fp32 = bottom_blob_unpacked;
273
        if (bottom_blob_unpacked.elembits() == 16)
274
        {
275
            Option opt_pack1 = opt;
276
            opt_pack1.blob_allocator = opt.workspace_allocator;
277

278
            cast_float16_to_float32(bottom_blob_unpacked, bottom_blob_unpacked_fp32, opt_pack1);
279
        }
280

281
        Option opt_unpacked = opt;
282
        opt_unpacked.use_packing_layout = false;
283
        return Convolution::forward_int8(bottom_blob_unpacked_fp32, top_blob, opt_unpacked);
284
    }
285
#endif
286

287
    // flattened blob, implement as InnerProduct
288
    if (bottom_blob.dims == 1 && kernel_w == 1 && kernel_h == 1)
289
    {
290
        Mat bottom_blob_3d;
291
        if (bottom_blob.elemsize % 16 == 0)
292
        {
293
            bottom_blob_3d = bottom_blob;
294
            bottom_blob_3d.dims = 3;
295
            bottom_blob_3d.w = 1;
296
            bottom_blob_3d.h = 1;
297
            bottom_blob_3d.c = bottom_blob.w;
298
            bottom_blob_3d.cstep = 1;
299
        }
300
        else
301
        {
302
            bottom_blob_3d = bottom_blob.reshape(1, 1, bottom_blob.w, opt.workspace_allocator);
303
        }
304

305
        Mat top_blob_3d;
306
        int ret = forward(bottom_blob_3d, top_blob_3d, opt);
307
        if (ret != 0)
308
            return ret;
309

310
        if (top_blob_3d.elemsize % 16 == 0)
311
        {
312
            top_blob = top_blob_3d;
313
            top_blob.dims = 1;
314
            top_blob.w = top_blob_3d.c;
315
            top_blob.h = 1;
316
            top_blob.c = 1;
317
            bottom_blob_3d.cstep = top_blob_3d.c;
318
        }
319
        else
320
        {
321
            top_blob = top_blob_3d.reshape(top_blob_3d.c, opt.blob_allocator);
322
        }
323

324
        return 0;
325
    }
326

327
    int elembits = bottom_blob.elembits();
328

329
#if __riscv_vector && __riscv_zfh
330
    if (opt.use_fp16_storage && elembits == 16)
331
    {
332
        if (opt.use_fp16_arithmetic)
333
            return forward_fp16sa(bottom_blob, top_blob, opt);
334
        else
335
            return forward_fp16s(bottom_blob, top_blob, opt);
336
    }
337
#endif
338

339
#if __riscv_vector
340
    const int packn = csrr_vlenb() / 4;
341
#endif
342

343
    int w = bottom_blob.w;
344
    int h = bottom_blob.h;
345
    int channels = bottom_blob.c;
346
    size_t elemsize = bottom_blob.elemsize;
347
    int elempack = bottom_blob.elempack;
348

349
    //     NCNN_LOGE("Convolution 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);
350

351
    const int kernel_extent_w = dilation_w * (kernel_w - 1) + 1;
352
    const int kernel_extent_h = dilation_h * (kernel_h - 1) + 1;
353

354
    Mat bottom_blob_bordered;
355
    make_padding(bottom_blob, bottom_blob_bordered, opt);
356
    if (bottom_blob_bordered.empty())
357
        return -100;
358

359
    w = bottom_blob_bordered.w;
360
    h = bottom_blob_bordered.h;
361

362
    int outw = (w - kernel_extent_w) / stride_w + 1;
363
    int outh = (h - kernel_extent_h) / stride_h + 1;
364
    int out_elempack = 1;
365
#if __riscv_vector
366
    if (opt.use_packing_layout)
367
    {
368
        out_elempack = num_output % packn == 0 ? packn : 1;
369
    }
370
#endif
371
    size_t out_elemsize = elemsize / elempack * out_elempack;
372

373
    top_blob.create(outw, outh, num_output / out_elempack, out_elemsize, out_elempack, opt.blob_allocator);
374
    if (top_blob.empty())
375
        return -100;
376

377
    const int num_input = channels * elempack;
378

379
#if __riscv_vector
380
    if (elempack == packn && out_elempack == packn)
381
    {
382
        if (kernel_w == 1 && kernel_h == 1 && dilation_w == 1 && dilation_h == 1 && stride_w == 1 && stride_h == 1)
383
        {
384
            conv1x1s1_sgemm_packn_rvv(bottom_blob_bordered, top_blob, weight_data_tm, bias_data, opt);
385

386
            if (activation)
387
            {
388
                activation->forward_inplace(top_blob, opt);
389
            }
390
        }
391
        else if (kernel_w == 1 && kernel_h == 1 && dilation_w == 1 && dilation_h == 1 && stride_w == 2 && stride_h == 2)
392
        {
393
            conv1x1s2_sgemm_packn_rvv(bottom_blob_bordered, top_blob, weight_data_tm, bias_data, opt);
394

395
            if (activation)
396
            {
397
                activation->forward_inplace(top_blob, opt);
398
            }
399
        }
400
        else if (opt.use_winograd_convolution && (opt.use_winograd23_convolution || opt.use_winograd43_convolution || opt.use_winograd63_convolution) && kernel_w == 3 && kernel_h == 3 && dilation_w == 1 && dilation_h == 1 && stride_w == 1 && stride_h == 1)
401
        {
402
            if ((opt.use_winograd63_convolution && num_input >= packn * 2 && num_output >= packn * 2 && num_input <= packn * 16 && num_output <= packn * 16) || (!opt.use_winograd43_convolution && !opt.use_winograd23_convolution))
403
                conv3x3s1_winograd63_packn_rvv(bottom_blob_bordered, top_blob, weight_winograd63_data, bias_data, opt);
404
            else if ((opt.use_winograd43_convolution && num_input >= packn * 2 && num_output >= packn * 2) || (!opt.use_winograd63_convolution && !opt.use_winograd23_convolution))
405
                conv3x3s1_winograd43_packn_rvv(bottom_blob_bordered, top_blob, weight_winograd43_data, bias_data, opt);
406
            else // if (opt.use_winograd23_convolution)
407
                conv3x3s1_winograd23_packn_rvv(bottom_blob_bordered, top_blob, weight_winograd23_data, bias_data, opt);
408

409
            if (activation)
410
            {
411
                activation->forward_inplace(top_blob, opt);
412
            }
413
        }
414
        else if (opt.use_sgemm_convolution)
415
        {
416
            convolution_im2col_sgemm_packn_rvv(bottom_blob_bordered, top_blob, weight_data_tm, bias_data, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, opt);
417

418
            if (activation)
419
            {
420
                activation->forward_inplace(top_blob, opt);
421
            }
422
        }
423
        else
424
        {
425
            convolution_packn_rvv(bottom_blob_bordered, top_blob, weight_data_tm, bias_data, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, activation_type, activation_params, opt);
426
        }
427
    }
428

429
    if (elempack == 1 && out_elempack == packn)
430
    {
431
        if (kernel_w == 1 && kernel_h == 1 && dilation_w == 1 && dilation_h == 1 && stride_w == 1 && stride_h == 1)
432
        {
433
            conv1x1s1_sgemm_pack1ton_rvv(bottom_blob_bordered, top_blob, weight_data_tm, bias_data, opt);
434

435
            if (activation)
436
            {
437
                activation->forward_inplace(top_blob, opt);
438
            }
439
        }
440
        else if (kernel_w == 3 && kernel_h == 3 && dilation_w == 1 && dilation_h == 1 && stride_w == 1 && stride_h == 1)
441
        {
442
            conv3x3s1_pack1ton_rvv(bottom_blob_bordered, top_blob, weight_data_tm, bias_data, opt);
443

444
            if (activation)
445
            {
446
                activation->forward_inplace(top_blob, opt);
447
            }
448
        }
449
        else if (kernel_w == 3 && kernel_h == 3 && dilation_w == 1 && dilation_h == 1 && stride_w == 2 && stride_h == 2)
450
        {
451
            conv3x3s2_pack1ton_rvv(bottom_blob_bordered, top_blob, weight_data_tm, bias_data, opt);
452

453
            if (activation)
454
            {
455
                activation->forward_inplace(top_blob, opt);
456
            }
457
        }
458
        else if (kernel_w == 7 && kernel_h == 7 && dilation_w == 1 && dilation_h == 1 && stride_w == 2 && stride_h == 2)
459
        {
460
            conv7x7s2_pack1ton_rvv(bottom_blob_bordered, top_blob, weight_data_tm, bias_data, opt);
461

462
            if (activation)
463
            {
464
                activation->forward_inplace(top_blob, opt);
465
            }
466
        }
467
        else if (opt.use_sgemm_convolution)
468
        {
469
            convolution_im2col_sgemm_pack1ton_rvv(bottom_blob_bordered, top_blob, weight_data_tm, bias_data, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, opt);
470

471
            if (activation)
472
            {
473
                activation->forward_inplace(top_blob, opt);
474
            }
475
        }
476
        else
477
        {
478
            convolution_pack1ton_rvv(bottom_blob_bordered, top_blob, weight_data_tm, bias_data, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, activation_type, activation_params, opt);
479
        }
480
    }
481

482
    if (elempack == packn && out_elempack == 1)
483
    {
484
        if (kernel_w == 1 && kernel_h == 1 && dilation_w == 1 && dilation_h == 1 && stride_w == 1 && stride_h == 1)
485
        {
486
            conv1x1s1_sgemm_packnto1_rvv(bottom_blob_bordered, top_blob, weight_data_tm, bias_data, opt);
487

488
            if (activation)
489
            {
490
                activation->forward_inplace(top_blob, opt);
491
            }
492
        }
493
        else if (kernel_w == 1 && kernel_h == 1 && dilation_w == 1 && dilation_h == 1 && stride_w == 2 && stride_h == 2)
494
        {
495
            conv1x1s2_sgemm_packnto1_rvv(bottom_blob_bordered, top_blob, weight_data_tm, bias_data, opt);
496

497
            if (activation)
498
            {
499
                activation->forward_inplace(top_blob, opt);
500
            }
501
        }
502
        else if (opt.use_sgemm_convolution)
503
        {
504
            convolution_im2col_sgemm_packnto1_rvv(bottom_blob_bordered, top_blob, weight_data_tm, bias_data, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, opt);
505

506
            if (activation)
507
            {
508
                activation->forward_inplace(top_blob, opt);
509
            }
510
        }
511
        else
512
        {
513
            convolution_packnto1_rvv(bottom_blob_bordered, top_blob, weight_data_tm, bias_data, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, activation_type, activation_params, opt);
514
        }
515
    }
516
#endif // __riscv_vector
517

518
    if (elempack == 1 && out_elempack == 1)
519
    {
520
        if (kernel_w == 1 && kernel_h == 1 && dilation_w == 1 && dilation_h == 1 && stride_w == 1 && stride_h == 1)
521
        {
522
            conv1x1s1_sgemm_rvv(bottom_blob_bordered, top_blob, weight_data_tm, bias_data, opt);
523

524
            if (activation)
525
            {
526
                activation->forward_inplace(top_blob, opt);
527
            }
528
        }
529
        else if (opt.use_winograd_convolution && (opt.use_winograd43_convolution || opt.use_winograd23_convolution) && kernel_w == 3 && kernel_h == 3 && dilation_w == 1 && dilation_h == 1 && stride_w == 1 && stride_h == 1)
530
        {
531
            if ((opt.use_winograd43_convolution && num_input >= 16 && num_output >= 16) || !opt.use_winograd23_convolution)
532
            {
533
                conv3x3s1_winograd43_rvv(bottom_blob_bordered, top_blob, weight_winograd43_data, bias_data, opt);
534
            }
535
            else if (opt.use_winograd23_convolution)
536
            {
537
                conv3x3s1_winograd23_rvv(bottom_blob_bordered, top_blob, weight_winograd23_data, bias_data, opt);
538
            }
539

540
            if (activation)
541
            {
542
                activation->forward_inplace(top_blob, opt);
543
            }
544
        }
545
        else if (opt.use_sgemm_convolution)
546
        {
547
            convolution_im2col_sgemm_rvv(bottom_blob_bordered, top_blob, weight_data_tm, bias_data, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, opt);
548

549
            if (activation)
550
            {
551
                activation->forward_inplace(top_blob, opt);
552
            }
553
        }
554
        else
555
        {
556
            const int maxk = kernel_w * kernel_h;
557

558
            // kernel offsets
559
            std::vector<int> _space_ofs(maxk);
560
            int* space_ofs = &_space_ofs[0];
561
            {
562
                int p1 = 0;
563
                int p2 = 0;
564
                int gap = w * dilation_h - kernel_w * dilation_w;
565
                for (int i = 0; i < kernel_h; i++)
566
                {
567
                    for (int j = 0; j < kernel_w; j++)
568
                    {
569
                        space_ofs[p1] = p2;
570
                        p1++;
571
                        p2 += dilation_w;
572
                    }
573
                    p2 += gap;
574
                }
575
            }
576

577
            // num_output
578
            #pragma omp parallel for num_threads(opt.num_threads)
579
            for (int p = 0; p < num_output; p++)
580
            {
581
                float* outptr = top_blob.channel(p);
582

583
                for (int i = 0; i < outh; i++)
584
                {
585
                    for (int j = 0; j < outw; j++)
586
                    {
587
                        float sum = 0.f;
588

589
                        if (bias_term)
590
                        {
591
                            sum = bias_data[p];
592
                        }
593

594
                        const float* kptr = (const float*)weight_data_tm + maxk * channels * p;
595

596
                        // channels
597
                        for (int q = 0; q < channels; q++)
598
                        {
599
                            const Mat m = bottom_blob_bordered.channel(q);
600
                            const float* sptr = m.row(i * stride_h) + j * stride_w;
601

602
                            for (int k = 0; k < maxk; k++)
603
                            {
604
                                float val = sptr[space_ofs[k]];
605
                                float wt = kptr[k];
606
                                sum += val * wt;
607
                            }
608

609
                            kptr += maxk;
610
                        }
611

612
                        sum = activation_ss(sum, activation_type, activation_params);
613

614
                        outptr[j] = sum;
615
                    }
616

617
                    outptr += outw;
618
                }
619
            }
620
        }
621
    }
622

623
    return 0;
624
}
625

626
int Convolution_riscv::forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const
627
{
628
    const Mat& bottom_blob = bottom_blobs[0];
629
    const Mat& _weight_data = bottom_blobs[1];
630
    Mat& top_blob = top_blobs[0];
631

632
    const int _kernel_w = _weight_data.w;
633
    const int _kernel_h = _weight_data.h;
634
    const int _num_output = _weight_data.c * _weight_data.elempack;
635

636
    Mat weight_data_flattened;
637
    flatten(_weight_data, weight_data_flattened, opt);
638
    if (weight_data_flattened.empty())
639
        return -100;
640

641
#if NCNN_RVV
642
    if (opt.use_fp16_storage && cpu_support_riscv_v() && cpu_support_riscv_zfh() && weight_data_flattened.elembits() == 16)
643
    {
644
        Mat weight_data_flattened_fp32;
645
        cast_float16_to_float32(weight_data_flattened, weight_data_flattened_fp32, opt);
646
        weight_data_flattened = weight_data_flattened_fp32;
647
    }
648
#endif // NCNN_RVV
649

650
    // weight_data_flattened as pack1
651
    weight_data_flattened.w *= weight_data_flattened.elempack;
652
    weight_data_flattened.elemsize /= weight_data_flattened.elempack;
653
    weight_data_flattened.elempack = 1;
654

655
    Mat bias_data_flattened;
656
    if (bias_term)
657
    {
658
        const Mat& _bias_data = bottom_blobs[2];
659
        flatten(_bias_data, bias_data_flattened, opt);
660
        if (bias_data_flattened.empty())
661
            return -100;
662

663
#if NCNN_RVV
664
        if (opt.use_fp16_storage && cpu_support_riscv_v() && cpu_support_riscv_zfh() && bias_data_flattened.elembits() == 16)
665
        {
666
            Mat bias_data_flattened_fp32;
667
            cast_float16_to_float32(bias_data_flattened, bias_data_flattened_fp32, opt);
668
            bias_data_flattened = bias_data_flattened_fp32;
669
        }
670
#endif // NCNN_RVV
671

672
        // bias_data_flattened as pack1
673
        bias_data_flattened.w *= bias_data_flattened.elempack;
674
        bias_data_flattened.elemsize /= bias_data_flattened.elempack;
675
        bias_data_flattened.elempack = 1;
676
    }
677

678
    ncnn::Layer* op = ncnn::create_layer_cpu(ncnn::LayerType::Convolution);
679

680
    ncnn::ParamDict pd;
681
    pd.set(0, _num_output);
682
    pd.set(1, _kernel_w);
683
    pd.set(11, _kernel_h);
684
    pd.set(2, dilation_w);
685
    pd.set(12, dilation_h);
686
    pd.set(3, stride_w);
687
    pd.set(13, stride_h);
688
    pd.set(4, pad_left);
689
    pd.set(15, pad_right);
690
    pd.set(14, pad_top);
691
    pd.set(16, pad_bottom);
692
    pd.set(18, pad_value);
693
    pd.set(5, bias_term);
694
    pd.set(6, weight_data_flattened.w);
695
    pd.set(8, int8_scale_term);
696
    pd.set(9, activation_type);
697
    pd.set(10, activation_params);
698

699
    op->load_param(pd);
700

701
    ncnn::Mat weights[2];
702
    weights[0] = weight_data_flattened;
703
    weights[1] = bias_data_flattened;
704

705
    op->load_model(ncnn::ModelBinFromMatArray(weights));
706

707
    op->create_pipeline(opt);
708

709
    op->forward(bottom_blob, top_blob, opt);
710

711
    op->destroy_pipeline(opt);
712

713
    delete op;
714

715
    return 0;
716
}
717

718
#if __riscv_vector && __riscv_zfh
719
static void convolution_transform_kernel_packed_fp16s_rvv(const Mat& weight_data, Mat& weight_data_tm, int num_input, int num_output, int kernel_w, int kernel_h, int elempack, int out_elempack)
720
{
721
    const int maxk = kernel_w * kernel_h;
722

723
    // src = kw-kh-inch-outch
724
    // dst = pb-pa-kw-kh-inch/pa-outch/pb
725
    {
726
        Mat weight_data_r2 = weight_data.reshape(maxk, num_input, num_output);
727

728
        weight_data_tm.create(maxk, num_input / elempack, num_output / out_elempack, (size_t)2u * elempack * out_elempack, elempack * out_elempack);
729

730
        for (int q = 0; q + (out_elempack - 1) < num_output; q += out_elempack)
731
        {
732
            __fp16* g00 = weight_data_tm.channel(q / out_elempack);
733

734
            for (int p = 0; p + (elempack - 1) < num_input; p += elempack)
735
            {
736
                for (int k = 0; k < maxk; k++)
737
                {
738
                    for (int i = 0; i < elempack; i++)
739
                    {
740
                        for (int j = 0; j < out_elempack; j++)
741
                        {
742
                            const float* k00 = weight_data_r2.channel(q + j).row(p + i);
743

744
                            g00[0] = (__fp16)k00[k];
745

746
                            g00++;
747
                        }
748
                    }
749
                }
750
            }
751
        }
752
    }
753
}
754

755
int Convolution_riscv::create_pipeline_fp16s(const Option& opt)
756
{
757
    const int packn = csrr_vlenb() / 2;
758

759
    const int maxk = kernel_w * kernel_h;
760
    const int num_input = weight_data_size / maxk / num_output;
761

762
    int elempack = 1;
763
    int out_elempack = 1;
764

765
    if (opt.use_packing_layout)
766
    {
767
        elempack = num_input % packn == 0 ? packn : 1;
768
        out_elempack = num_output % packn == 0 ? packn : 1;
769
    }
770

771
    // packn
772
    if (elempack == packn && out_elempack == packn)
773
    {
774
        if (opt.use_winograd_convolution && (opt.use_winograd23_convolution || opt.use_winograd43_convolution || opt.use_winograd63_convolution) && opt.use_fp16_arithmetic && kernel_w == 3 && kernel_h == 3 && dilation_w == 1 && dilation_h == 1 && stride_w == 1 && stride_h == 1)
775
        {
776
            if ((opt.use_winograd63_convolution && num_input >= packn * 2 && num_output >= packn * 2 && num_input <= packn * 16 && num_output <= packn * 16) || (!opt.use_winograd43_convolution && !opt.use_winograd23_convolution))
777
                conv3x3s1_winograd63_transform_kernel_packn_fp16sa_rvv(weight_data, weight_winograd63_data, num_input, num_output, opt);
778
            else if ((opt.use_winograd43_convolution && num_input >= packn * 2 && num_output >= packn * 2) || (!opt.use_winograd63_convolution && !opt.use_winograd23_convolution))
779
                conv3x3s1_winograd43_transform_kernel_packn_fp16sa_rvv(weight_data, weight_winograd43_data, num_input, num_output, opt);
780
            else // if (opt.use_winograd23_convolution)
781
                conv3x3s1_winograd23_transform_kernel_packn_fp16sa_rvv(weight_data, weight_winograd23_data, num_input, num_output, opt);
782
        }
783
        else
784
        {
785
            convolution_transform_kernel_packed_fp16s_rvv(weight_data, weight_data_tm, num_input, num_output, kernel_w, kernel_h, elempack, out_elempack);
786
        }
787
    }
788

789
    // pack1ton
790
    if (elempack == 1 && out_elempack == packn)
791
    {
792
        convolution_transform_kernel_packed_fp16s_rvv(weight_data, weight_data_tm, num_input, num_output, kernel_w, kernel_h, elempack, out_elempack);
793
    }
794

795
    // packnto1
796
    if (elempack == packn && out_elempack == 1)
797
    {
798
        if (opt.use_fp16_arithmetic && kernel_w == 1 && kernel_h == 1 && dilation_w == 1 && dilation_h == 1 && stride_w == 1 && stride_h == 1)
799
        {
800
            convolution_im2col_sgemm_transform_kernel_packnto1_fp16sa_rvv(weight_data, weight_data_tm, num_input, num_output, kernel_w, kernel_h);
801
        }
802
        else if (opt.use_fp16_arithmetic && kernel_w == 1 && kernel_h == 1 && dilation_w == 1 && dilation_h == 1 && stride_w == 2 && stride_h == 2)
803
        {
804
            convolution_im2col_sgemm_transform_kernel_packnto1_fp16sa_rvv(weight_data, weight_data_tm, num_input, num_output, kernel_w, kernel_h);
805
        }
806
        else if (opt.use_fp16_arithmetic && opt.use_sgemm_convolution)
807
        {
808
            convolution_im2col_sgemm_transform_kernel_packnto1_fp16sa_rvv(weight_data, weight_data_tm, num_input, num_output, kernel_w, kernel_h);
809
        }
810
        else
811
        {
812
            convolution_transform_kernel_packed_fp16s_rvv(weight_data, weight_data_tm, num_input, num_output, kernel_w, kernel_h, elempack, out_elempack);
813
        }
814
    }
815

816
    // pack1
817
    if (elempack == 1 && out_elempack == 1)
818
    {
819
        if (opt.use_fp16_arithmetic && kernel_w == 1 && kernel_h == 1 && dilation_w == 1 && dilation_h == 1 && stride_w == 1 && stride_h == 1)
820
        {
821
            convolution_im2col_sgemm_transform_kernel_fp16sa_rvv(weight_data, weight_data_tm, num_input, num_output, kernel_w, kernel_h);
822
        }
823
        else if (opt.use_fp16_arithmetic && opt.use_sgemm_convolution)
824
        {
825
            convolution_im2col_sgemm_transform_kernel_fp16sa_rvv(weight_data, weight_data_tm, num_input, num_output, kernel_w, kernel_h);
826
        }
827
        else
828
        {
829
            convolution_transform_kernel_packed_fp16s_rvv(weight_data, weight_data_tm, num_input, num_output, kernel_w, kernel_h, elempack, out_elempack);
830
        }
831
    }
832

833
    if (opt.use_fp16_arithmetic)
834
    {
835
        ncnn::cast_float32_to_float16(bias_data, bias_data_fp16, opt);
836
    }
837

838
    if (opt.lightmode)
839
        weight_data.release();
840

841
    return 0;
842
}
843

844
int Convolution_riscv::forward_fp16s(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const
845
{
846
    const int packn = csrr_vlenb() / 2;
847

848
    int w = bottom_blob.w;
849
    int h = bottom_blob.h;
850
    size_t elemsize = bottom_blob.elemsize;
851
    int elempack = bottom_blob.elempack;
852

853
    //     NCNN_LOGE("Convolution forward_fp16s input %d x %d  pad = %d %d  ksize=%d %d  stride=%d %d", w, h, pad_left, pad_top, kernel_w, kernel_h, stride_w, stride_h);
854

855
    const int kernel_extent_w = dilation_w * (kernel_w - 1) + 1;
856
    const int kernel_extent_h = dilation_h * (kernel_h - 1) + 1;
857

858
    Mat bottom_blob_bordered;
859
    make_padding(bottom_blob, bottom_blob_bordered, opt);
860
    if (bottom_blob_bordered.empty())
861
        return -100;
862

863
    w = bottom_blob_bordered.w;
864
    h = bottom_blob_bordered.h;
865

866
    int outw = (w - kernel_extent_w) / stride_w + 1;
867
    int outh = (h - kernel_extent_h) / stride_h + 1;
868
    int out_elempack = (opt.use_packing_layout && num_output % packn == 0) ? packn : 1;
869
    size_t out_elemsize = elemsize / elempack * out_elempack;
870

871
    top_blob.create(outw, outh, num_output / out_elempack, out_elemsize, out_elempack, opt.blob_allocator);
872
    if (top_blob.empty())
873
        return -100;
874

875
    if (elempack == packn && out_elempack == packn)
876
    {
877
        {
878
            convolution_packn_fp16s_rvv(bottom_blob_bordered, top_blob, weight_data_tm, bias_data, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, activation_type, activation_params, opt);
879
        }
880
    }
881

882
    if (elempack == 1 && out_elempack == packn)
883
    {
884
        {
885
            convolution_pack1ton_fp16s_rvv(bottom_blob_bordered, top_blob, weight_data_tm, bias_data, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, activation_type, activation_params, opt);
886
        }
887
    }
888

889
    if (elempack == packn && out_elempack == 1)
890
    {
891
        {
892
            convolution_packnto1_fp16s_rvv(bottom_blob_bordered, top_blob, weight_data_tm, bias_data, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, activation_type, activation_params, opt);
893
        }
894
    }
895

896
    if (elempack == 1 && out_elempack == 1)
897
    {
898
        {
899
            convolution_fp16s(bottom_blob_bordered, top_blob, weight_data_tm, bias_data, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, activation_type, activation_params, opt);
900
        }
901
    }
902

903
    return 0;
904
}
905

906
int Convolution_riscv::forward_fp16sa(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const
907
{
908
    const int packn = csrr_vlenb() / 2;
909

910
    int w = bottom_blob.w;
911
    int h = bottom_blob.h;
912
    int channels = bottom_blob.c;
913
    size_t elemsize = bottom_blob.elemsize;
914
    int elempack = bottom_blob.elempack;
915

916
    // NCNN_LOGE("Convolution 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);
917

918
    const int kernel_extent_w = dilation_w * (kernel_w - 1) + 1;
919
    const int kernel_extent_h = dilation_h * (kernel_h - 1) + 1;
920

921
    Mat bottom_blob_bordered;
922
    make_padding(bottom_blob, bottom_blob_bordered, opt);
923
    if (bottom_blob_bordered.empty())
924
        return -100;
925

926
    w = bottom_blob_bordered.w;
927
    h = bottom_blob_bordered.h;
928

929
    int outw = (w - kernel_extent_w) / stride_w + 1;
930
    int outh = (h - kernel_extent_h) / stride_h + 1;
931
    int out_elempack = (opt.use_packing_layout && num_output % packn == 0) ? packn : 1;
932
    size_t out_elemsize = elemsize / elempack * out_elempack;
933

934
    top_blob.create(outw, outh, num_output / out_elempack, out_elemsize, out_elempack, opt.blob_allocator);
935
    if (top_blob.empty())
936
        return -100;
937

938
    const int num_input = channels * elempack;
939

940
    if (elempack == packn && out_elempack == packn)
941
    {
942
        if (kernel_w == 1 && kernel_h == 1 && dilation_w == 1 && dilation_h == 1 && stride_w == 1 && stride_h == 1)
943
        {
944
            conv1x1s1_sgemm_packn_fp16sa_rvv(bottom_blob_bordered, top_blob, weight_data_tm, bias_data_fp16, opt);
945

946
            if (activation)
947
            {
948
                activation->forward_inplace(top_blob, opt);
949
            }
950
        }
951
        else if (kernel_w == 1 && kernel_h == 1 && dilation_w == 1 && dilation_h == 1 && stride_w == 2 && stride_h == 2)
952
        {
953
            conv1x1s2_sgemm_packn_fp16sa_rvv(bottom_blob_bordered, top_blob, weight_data_tm, bias_data_fp16, opt);
954

955
            if (activation)
956
            {
957
                activation->forward_inplace(top_blob, opt);
958
            }
959
        }
960
        else if (opt.use_winograd_convolution && (opt.use_winograd23_convolution || opt.use_winograd43_convolution || opt.use_winograd63_convolution) && kernel_w == 3 && kernel_h == 3 && dilation_w == 1 && dilation_h == 1 && stride_w == 1 && stride_h == 1)
961
        {
962
            if ((opt.use_winograd63_convolution && num_input >= packn * 2 && num_output >= packn * 2 && num_input <= packn * 16 && num_output <= packn * 16) || (!opt.use_winograd43_convolution && !opt.use_winograd23_convolution))
963
                conv3x3s1_winograd63_packn_fp16sa_rvv(bottom_blob_bordered, top_blob, weight_winograd63_data, bias_data_fp16, opt);
964
            else if ((opt.use_winograd43_convolution && num_input >= packn * 2 && num_output >= packn * 2) || (!opt.use_winograd63_convolution && !opt.use_winograd23_convolution))
965
                conv3x3s1_winograd43_packn_fp16sa_rvv(bottom_blob_bordered, top_blob, weight_winograd43_data, bias_data_fp16, opt);
966
            else // if (opt.use_winograd23_convolution)
967
                conv3x3s1_winograd23_packn_fp16sa_rvv(bottom_blob_bordered, top_blob, weight_winograd23_data, bias_data_fp16, opt);
968

969
            if (activation)
970
            {
971
                activation->forward_inplace(top_blob, opt);
972
            }
973
        }
974
        else if (opt.use_sgemm_convolution)
975
        {
976
            convolution_im2col_sgemm_packn_fp16sa_rvv(bottom_blob_bordered, top_blob, weight_data_tm, bias_data_fp16, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, opt);
977

978
            if (activation)
979
            {
980
                activation->forward_inplace(top_blob, opt);
981
            }
982
        }
983
        else
984
        {
985
            convolution_packn_fp16sa_rvv(bottom_blob_bordered, top_blob, weight_data_tm, bias_data_fp16, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, activation_type, activation_params, opt);
986
        }
987
    }
988

989
    if (elempack == 1 && out_elempack == packn)
990
    {
991
        if (kernel_w == 1 && kernel_h == 1 && dilation_w == 1 && dilation_h == 1 && stride_w == 1 && stride_h == 1)
992
        {
993
            conv1x1s1_sgemm_pack1ton_fp16sa_rvv(bottom_blob_bordered, top_blob, weight_data_tm, bias_data_fp16, opt);
994

995
            if (activation)
996
            {
997
                activation->forward_inplace(top_blob, opt);
998
            }
999
        }
1000
        else if (kernel_w == 3 && kernel_h == 3 && dilation_w == 1 && dilation_h == 1 && stride_w == 1 && stride_h == 1)
1001
        {
1002
            conv3x3s1_pack1ton_fp16sa_rvv(bottom_blob_bordered, top_blob, weight_data_tm, bias_data_fp16, opt);
1003

1004
            if (activation)
1005
            {
1006
                activation->forward_inplace(top_blob, opt);
1007
            }
1008
        }
1009
        else if (kernel_w == 3 && kernel_h == 3 && dilation_w == 1 && dilation_h == 1 && stride_w == 2 && stride_h == 2)
1010
        {
1011
            conv3x3s2_pack1ton_fp16sa_rvv(bottom_blob_bordered, top_blob, weight_data_tm, bias_data_fp16, opt);
1012

1013
            if (activation)
1014
            {
1015
                activation->forward_inplace(top_blob, opt);
1016
            }
1017
        }
1018
        else if (kernel_w == 7 && kernel_h == 7 && dilation_w == 1 && dilation_h == 1 && stride_w == 2 && stride_h == 2)
1019
        {
1020
            conv7x7s2_pack1ton_fp16sa_rvv(bottom_blob_bordered, top_blob, weight_data_tm, bias_data_fp16, opt);
1021

1022
            if (activation)
1023
            {
1024
                activation->forward_inplace(top_blob, opt);
1025
            }
1026
        }
1027
        else if (opt.use_sgemm_convolution)
1028
        {
1029
            convolution_im2col_sgemm_pack1ton_fp16sa_rvv(bottom_blob_bordered, top_blob, weight_data_tm, bias_data_fp16, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, opt);
1030

1031
            if (activation)
1032
            {
1033
                activation->forward_inplace(top_blob, opt);
1034
            }
1035
        }
1036
        else
1037
        {
1038
            convolution_pack1ton_fp16sa_rvv(bottom_blob_bordered, top_blob, weight_data_tm, bias_data_fp16, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, activation_type, activation_params, opt);
1039
        }
1040
    }
1041

1042
    if (elempack == packn && out_elempack == 1)
1043
    {
1044
        if (kernel_w == 1 && kernel_h == 1 && dilation_w == 1 && dilation_h == 1 && stride_w == 1 && stride_h == 1)
1045
        {
1046
            conv1x1s1_sgemm_packnto1_fp16sa_rvv(bottom_blob_bordered, top_blob, weight_data_tm, bias_data_fp16, opt);
1047

1048
            if (activation)
1049
            {
1050
                activation->forward_inplace(top_blob, opt);
1051
            }
1052
        }
1053
        else if (kernel_w == 1 && kernel_h == 1 && dilation_w == 1 && dilation_h == 1 && stride_w == 2 && stride_h == 2)
1054
        {
1055
            conv1x1s2_sgemm_packnto1_fp16sa_rvv(bottom_blob_bordered, top_blob, weight_data_tm, bias_data_fp16, opt);
1056

1057
            if (activation)
1058
            {
1059
                activation->forward_inplace(top_blob, opt);
1060
            }
1061
        }
1062
        else if (opt.use_sgemm_convolution)
1063
        {
1064
            convolution_im2col_sgemm_packnto1_fp16sa_rvv(bottom_blob_bordered, top_blob, weight_data_tm, bias_data_fp16, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, opt);
1065

1066
            if (activation)
1067
            {
1068
                activation->forward_inplace(top_blob, opt);
1069
            }
1070
        }
1071
        else
1072
        {
1073
            convolution_packnto1_fp16sa_rvv(bottom_blob_bordered, top_blob, weight_data_tm, bias_data_fp16, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, activation_type, activation_params, opt);
1074
        }
1075
    }
1076

1077
    if (elempack == 1 && out_elempack == 1)
1078
    {
1079
        if (kernel_w == 1 && kernel_h == 1 && dilation_w == 1 && dilation_h == 1 && stride_w == 1 && stride_h == 1)
1080
        {
1081
            conv1x1s1_sgemm_fp16sa_rvv(bottom_blob_bordered, top_blob, weight_data_tm, bias_data_fp16, opt);
1082

1083
            if (activation)
1084
            {
1085
                activation->forward_inplace(top_blob, opt);
1086
            }
1087
        }
1088
        else if (opt.use_sgemm_convolution)
1089
        {
1090
            convolution_im2col_sgemm_fp16sa_rvv(bottom_blob_bordered, top_blob, weight_data_tm, bias_data_fp16, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, opt);
1091

1092
            if (activation)
1093
            {
1094
                activation->forward_inplace(top_blob, opt);
1095
            }
1096
        }
1097
        else
1098
        {
1099
            convolution_fp16s(bottom_blob_bordered, top_blob, weight_data_tm, bias_data, kernel_w, kernel_h, dilation_w, dilation_h, stride_w, stride_h, activation_type, activation_params, opt);
1100
        }
1101
    }
1102

1103
    return 0;
1104
}
1105
#endif // __riscv_vector && __riscv_zfh
1106

1107
} // namespace ncnn
1108