File size: 15,731 Bytes
be903e2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
// Tencent is pleased to support the open source community by making ncnn available.
//
// Copyright (C) 2017 THL A29 Limited, a Tencent company. All rights reserved.
//
// Licensed under the BSD 3-Clause License (the "License"); you may not use this file except
// in compliance with the License. You may obtain a copy of the License at
//
// https://opensource.org/licenses/BSD-3-Clause
//
// Unless required by applicable law or agreed to in writing, software distributed
// under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR
// CONDITIONS OF ANY KIND, either express or implied. See the License for the
// specific language governing permissions and limitations under the License.

#include "convolution.h"

#include "layer_type.h"

#include "fused_activation.h"

namespace ncnn {

Convolution::Convolution()
{
    one_blob_only = true;
    support_inplace = false;
}

int Convolution::load_param(const ParamDict& pd)
{
    num_output = pd.get(0, 0);
    kernel_w = pd.get(1, 0);
    kernel_h = pd.get(11, kernel_w);
    dilation_w = pd.get(2, 1);
    dilation_h = pd.get(12, dilation_w);
    stride_w = pd.get(3, 1);
    stride_h = pd.get(13, stride_w);
    pad_left = pd.get(4, 0);
    pad_right = pd.get(15, pad_left);
    pad_top = pd.get(14, pad_left);
    pad_bottom = pd.get(16, pad_top);
    pad_value = pd.get(18, 0.f);
    bias_term = pd.get(5, 0);
    weight_data_size = pd.get(6, 0);
    int8_scale_term = pd.get(8, 0);
    activation_type = pd.get(9, 0);
    activation_params = pd.get(10, Mat());

    dynamic_weight = pd.get(19, 0);

    if (dynamic_weight)
    {
        one_blob_only = false;
    }

    if (int8_scale_term)
    {
#if NCNN_INT8
        support_int8_storage = true;
#else
        NCNN_LOGE("please build ncnn with NCNN_INT8 enabled for int8 inference");
        return -1;
#endif
    }

    return 0;
}

int Convolution::load_model(const ModelBin& mb)
{
    if (dynamic_weight)
        return 0;

    weight_data = mb.load(weight_data_size, 0);
    if (weight_data.empty())
        return -100;

    if (bias_term)
    {
        bias_data = mb.load(num_output, 1);
        if (bias_data.empty())
            return -100;
    }

#if NCNN_INT8
    if (int8_scale_term)
    {
        weight_data_int8_scales = mb.load(num_output, 1);
        bottom_blob_int8_scales = mb.load(1, 1);
    }

    if (int8_scale_term > 100)
    {
        top_blob_int8_scales = mb.load(1, 1);
    }
#endif // NCNN_INT8

    return 0;
}

int Convolution::create_pipeline(const Option& opt)
{
    if (dynamic_weight)
        return 0;

#if NCNN_INT8
    // runtime quantize the weight data
    if (opt.use_int8_inference && weight_data.elemsize == (size_t)4u && int8_scale_term)
    {
        const int maxk = kernel_w * kernel_h;
        const int num_input = weight_data_size / num_output / maxk;

        Mat weight_data_r2 = weight_data.reshape(maxk, num_input, num_output);

        Mat weight_data_int8;

        Option opt_q = opt;
        opt_q.blob_allocator = weight_data.allocator;
        opt_q.use_packing_layout = false;
        quantize_to_int8(weight_data_r2, weight_data_int8, weight_data_int8_scales, opt_q);
        if (weight_data_int8.empty())
            return -100;

        weight_data = weight_data_int8.reshape(weight_data_size);
    }
#else
    (void)(opt);
#endif // NCNN_INT8

    return 0;
}

static int convolution(const Mat& bottom_blob, Mat& top_blob, const Mat& weight_data, const Mat& bias_data, int kernel_w, int kernel_h, int stride_w, int stride_h, int dilation_w, int dilation_h, int activation_type, const Mat& activation_params, const Option& opt)
{
    const int w = bottom_blob.w;
    const int inch = bottom_blob.c;

    const int outw = top_blob.w;
    const int outh = top_blob.h;
    const int outch = top_blob.c;

    const int bias_term = bias_data.empty() ? 0 : 1;

    const int maxk = kernel_w * kernel_h;

    // kernel offsets
    std::vector<int> _space_ofs(maxk);
    int* space_ofs = &_space_ofs[0];
    {
        int p1 = 0;
        int p2 = 0;
        int gap = w * dilation_h - kernel_w * dilation_w;
        for (int i = 0; i < kernel_h; i++)
        {
            for (int j = 0; j < kernel_w; j++)
            {
                space_ofs[p1] = p2;
                p1++;
                p2 += dilation_w;
            }
            p2 += gap;
        }
    }

    #pragma omp parallel for num_threads(opt.num_threads)
    for (int p = 0; p < outch; p++)
    {
        float* outptr = top_blob.channel(p);

        for (int i = 0; i < outh; i++)
        {
            for (int j = 0; j < outw; j++)
            {
                float sum = 0.f;

                if (bias_term)
                    sum = bias_data[p];

                const float* kptr = (const float*)weight_data + maxk * inch * p;

                for (int q = 0; q < inch; q++)
                {
                    const Mat m = bottom_blob.channel(q);
                    const float* sptr = m.row(i * stride_h) + j * stride_w;

                    for (int k = 0; k < maxk; k++) // 29.23
                    {
                        float val = sptr[space_ofs[k]]; // 20.72
                        float wt = kptr[k];
                        sum += val * wt; // 41.45
                    }

                    kptr += maxk;
                }

                outptr[j] = activation_ss(sum, activation_type, activation_params);
            }

            outptr += outw;
        }
    }

    return 0;
}

int Convolution::forward(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const
{
#if NCNN_INT8
    if (opt.use_int8_inference && weight_data.elemsize == (size_t)1u)
    {
        return forward_int8(bottom_blob, top_blob, opt);
    }
#endif

    // flattened blob, implement as InnerProduct
    if (bottom_blob.dims == 1 && kernel_w == 1 && kernel_h == 1)
    {
        int num_input = weight_data_size / num_output;
        if (bottom_blob.w * bottom_blob.elempack == num_input)
        {
            // call InnerProduct
            ncnn::Layer* op = ncnn::create_layer(ncnn::LayerType::InnerProduct);

            // set param
            ncnn::ParamDict pd;
            pd.set(0, num_output);
            pd.set(1, bias_term);
            pd.set(2, weight_data_size);
            pd.set(8, int8_scale_term);
            pd.set(9, activation_type);
            pd.set(10, activation_params);

            op->load_param(pd);

            // set weights
            ncnn::Mat weights[4];
            weights[0] = weight_data;
            weights[1] = bias_data;

#if NCNN_INT8
            if (int8_scale_term)
            {
                weights[2] = weight_data_int8_scales;
                weights[3] = bottom_blob_int8_scales;
            }
#endif

            op->load_model(ModelBinFromMatArray(weights));

            op->create_pipeline(opt);

            // forward
            op->forward(bottom_blob, top_blob, opt);

            op->destroy_pipeline(opt);

            delete op;

            return 0;
        }
    }

    Mat bottom_blob_bordered;
    make_padding(bottom_blob, bottom_blob_bordered, opt);
    if (bottom_blob_bordered.empty())
        return -100;

    const int w = bottom_blob_bordered.w;
    const int h = bottom_blob_bordered.h;
    const size_t elemsize = bottom_blob_bordered.elemsize;

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

    const int outw = (w - kernel_extent_w) / stride_w + 1;
    const int outh = (h - kernel_extent_h) / stride_h + 1;

    top_blob.create(outw, outh, num_output, elemsize, opt.blob_allocator);
    if (top_blob.empty())
        return -100;

    int ret = convolution(bottom_blob_bordered, top_blob, weight_data, bias_data, kernel_w, kernel_h, stride_w, stride_h, dilation_w, dilation_h, activation_type, activation_params, opt);
    if (ret != 0)
        return ret;

    return 0;
}

int Convolution::forward(const std::vector<Mat>& bottom_blobs, std::vector<Mat>& top_blobs, const Option& opt) const
{
    const Mat& bottom_blob = bottom_blobs[0];
    const Mat& _weight_data = bottom_blobs[1];
    Mat& top_blob = top_blobs[0];

    const int _kernel_w = _weight_data.w;
    const int _kernel_h = _weight_data.h;
    const int _num_output = _weight_data.c;

    Mat weight_data_flattened;
    flatten(_weight_data, weight_data_flattened, opt);
    if (weight_data_flattened.empty())
        return -100;

    Mat bias_data_flattened;
    if (bias_term)
    {
        const Mat& _bias_data = bottom_blobs[2];
        flatten(_bias_data, bias_data_flattened, opt);
        if (bias_data_flattened.empty())
            return -100;
    }

    Mat bottom_blob_bordered;
    make_padding(bottom_blob, bottom_blob_bordered, _kernel_w, _kernel_h, opt);
    if (bottom_blob_bordered.empty())
        return -100;

    const int w = bottom_blob_bordered.w;
    const int h = bottom_blob_bordered.h;
    const size_t elemsize = bottom_blob_bordered.elemsize;

    const int kernel_extent_w = dilation_w * (_kernel_w - 1) + 1;
    const int kernel_extent_h = dilation_h * (_kernel_h - 1) + 1;

    const int outw = (w - kernel_extent_w) / stride_w + 1;
    const int outh = (h - kernel_extent_h) / stride_h + 1;

    top_blob.create(outw, outh, _num_output, elemsize, opt.blob_allocator);
    if (top_blob.empty())
        return -100;

    int ret = convolution(bottom_blob_bordered, top_blob, weight_data_flattened, bias_data_flattened, _kernel_w, _kernel_h, stride_w, stride_h, dilation_w, dilation_h, activation_type, activation_params, opt);
    if (ret != 0)
        return ret;

    return 0;
}

void Convolution::make_padding(const Mat& bottom_blob, Mat& bottom_blob_bordered, const Option& opt) const
{
    make_padding(bottom_blob, bottom_blob_bordered, kernel_w, kernel_h, opt);
}

void Convolution::make_padding(const Mat& bottom_blob, Mat& bottom_blob_bordered, int _kernel_w, int _kernel_h, const Option& opt) const
{
    int w = bottom_blob.w;
    int h = bottom_blob.h;

    const int kernel_extent_w = dilation_w * (_kernel_w - 1) + 1;
    const int kernel_extent_h = dilation_h * (_kernel_h - 1) + 1;

    bottom_blob_bordered = bottom_blob;
    if (pad_left > 0 || pad_right > 0 || pad_top > 0 || pad_bottom > 0)
    {
        Option opt_b = opt;
        opt_b.blob_allocator = opt.workspace_allocator;
        copy_make_border(bottom_blob, bottom_blob_bordered, pad_top, pad_bottom, pad_left, pad_right, BORDER_CONSTANT, pad_value, opt_b);
    }
    else if (pad_left == -233 && pad_right == -233 && pad_top == -233 && pad_bottom == -233)
    {
        // tensorflow padding=SAME or onnx padding=SAME_UPPER
        int wpad = kernel_extent_w + (w - 1) / stride_w * stride_w - w;
        int hpad = kernel_extent_h + (h - 1) / stride_h * stride_h - h;
        if (wpad > 0 || hpad > 0)
        {
            Option opt_b = opt;
            opt_b.blob_allocator = opt.workspace_allocator;
            copy_make_border(bottom_blob, bottom_blob_bordered, hpad / 2, hpad - hpad / 2, wpad / 2, wpad - wpad / 2, BORDER_CONSTANT, pad_value, opt_b);
        }
    }
    else if (pad_left == -234 && pad_right == -234 && pad_top == -234 && pad_bottom == -234)
    {
        // onnx padding=SAME_LOWER
        int wpad = kernel_extent_w + (w - 1) / stride_w * stride_w - w;
        int hpad = kernel_extent_h + (h - 1) / stride_h * stride_h - h;
        if (wpad > 0 || hpad > 0)
        {
            Option opt_b = opt;
            opt_b.blob_allocator = opt.workspace_allocator;
            copy_make_border(bottom_blob, bottom_blob_bordered, hpad - hpad / 2, hpad / 2, wpad - wpad / 2, wpad / 2, BORDER_CONSTANT, pad_value, opt_b);
        }
    }
}

#if NCNN_INT8
static inline signed char float2int8(float v)
{
    int int32 = static_cast<int>(round(v));
    if (int32 > 127) return 127;
    if (int32 < -127) return -127;
    return (signed char)int32;
}

int Convolution::forward_int8(const Mat& bottom_blob, Mat& top_blob, const Option& opt) const
{
    int w = bottom_blob.w;
    int h = bottom_blob.h;
    int channels = bottom_blob.c;
    size_t elemsize = bottom_blob.elemsize;

    //     NCNN_LOGE("Convolution input %d x %d  ksize=%d %d  stride=%d %d", w, h, kernel_w, kernel_h, stride_w, stride_h);

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

    Mat bottom_blob_unbordered = bottom_blob;
    if (elemsize != 1)
    {
        Option opt_g = opt;
        opt_g.blob_allocator = opt.workspace_allocator;

        quantize_to_int8(bottom_blob, bottom_blob_unbordered, bottom_blob_int8_scales, opt_g);
    }

    Mat bottom_blob_bordered;
    make_padding(bottom_blob_unbordered, bottom_blob_bordered, opt);
    if (bottom_blob_bordered.empty())
        return -100;

    w = bottom_blob_bordered.w;
    h = bottom_blob_bordered.h;

    int outw = (w - kernel_extent_w) / stride_w + 1;
    int outh = (h - kernel_extent_h) / stride_h + 1;

    const int maxk = kernel_w * kernel_h;

    // kernel offsets
    std::vector<int> _space_ofs(maxk);
    int* space_ofs = &_space_ofs[0];
    {
        int p1 = 0;
        int p2 = 0;
        int gap = w * dilation_h - kernel_w * dilation_w;
        for (int i = 0; i < kernel_h; i++)
        {
            for (int j = 0; j < kernel_w; j++)
            {
                space_ofs[p1] = p2;
                p1++;
                p2 += dilation_w;
            }
            p2 += gap;
        }
    }

    // int8
    bool use_int8_requantize = int8_scale_term > 100;
    size_t out_elemsize = use_int8_requantize ? 1u : 4u;

    top_blob.create(outw, outh, num_output, out_elemsize, opt.blob_allocator);
    if (top_blob.empty())
        return -100;

    // num_output
    #pragma omp parallel for num_threads(opt.num_threads)
    for (int p = 0; p < num_output; p++)
    {
        signed char* outptr = top_blob.channel(p);

        for (int i = 0; i < outh; i++)
        {
            for (int j = 0; j < outw; j++)
            {
                int sum = 0;

                const signed char* kptr = (const signed char*)weight_data + maxk * channels * p;

                // channels
                for (int q = 0; q < channels; q++)
                {
                    const Mat m = bottom_blob_bordered.channel(q);
                    const signed char* sptr = m.row<signed char>(i * stride_h) + j * stride_w;

                    for (int k = 0; k < maxk; k++)
                    {
                        int val = sptr[space_ofs[k]];
                        int wt = kptr[k];
                        sum += val * wt;
                    }

                    kptr += maxk;
                }

                float scale_in;
                if (weight_data_int8_scales[p] == 0)
                    scale_in = 0;
                else
                    scale_in = 1.f / (bottom_blob_int8_scales[0] * weight_data_int8_scales[p]);

                float sumfp32 = sum * scale_in;

                if (bias_term)
                    sumfp32 += bias_data[p];

                sumfp32 = activation_ss(sumfp32, activation_type, activation_params);

                if (use_int8_requantize)
                {
                    // requantize
                    float scale_out = top_blob_int8_scales[0];
                    signed char sums8 = float2int8(sumfp32 * scale_out);
                    outptr[0] = sums8;
                    outptr += 1;
                }
                else
                {
                    // dequantize
                    ((float*)outptr)[0] = sumfp32;
                    outptr += 4;
                }
            }
        }
    }

    return 0;
}
#endif // NCNN_INT8

} // namespace ncnn