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
/*
* Copyright (c) 2018 Sergey Lavrushkin
*
* This file is part of FFmpeg.
*
* FFmpeg is free software; you can redistribute it and/or
* modify it under the terms of the GNU Lesser General Public
* License as published by the Free Software Foundation; either
* version 2.1 of the License, or (at your option) any later version.
*
* FFmpeg is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
* Lesser General Public License for more details.
*
* You should have received a copy of the GNU Lesser General Public
* License along with FFmpeg; if not, write to the Free Software
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
*/
/**
* @file
* DNN native backend implementation.
*/
#include "dnn_backend_native.h"
static DNNReturnType set_input_output_native(void *model, DNNData *input, DNNData *output)
{
ConvolutionalNetwork *network = (ConvolutionalNetwork *)model;
InputParams *input_params;
ConvolutionalParams *conv_params;
DepthToSpaceParams *depth_to_space_params;
int cur_width, cur_height, cur_channels;
int32_t layer;
if (network->layers_num <= 0 || network->layers[0].type != INPUT){
return DNN_ERROR;
}
else{
input_params = (InputParams *)network->layers[0].params;
input_params->width = cur_width = input->width;
input_params->height = cur_height = input->height;
input_params->channels = cur_channels = input->channels;
if (input->data){
av_freep(&input->data);
}
network->layers[0].output = input->data = av_malloc(cur_height * cur_width * cur_channels * sizeof(float));
if (!network->layers[0].output){
return DNN_ERROR;
}
}
for (layer = 1; layer < network->layers_num; ++layer){
switch (network->layers[layer].type){
case CONV:
conv_params = (ConvolutionalParams *)network->layers[layer].params;
if (conv_params->input_num != cur_channels){
return DNN_ERROR;
}
cur_channels = conv_params->output_num;
break;
case DEPTH_TO_SPACE:
depth_to_space_params = (DepthToSpaceParams *)network->layers[layer].params;
if (cur_channels % (depth_to_space_params->block_size * depth_to_space_params->block_size) != 0){
return DNN_ERROR;
}
cur_channels = cur_channels / (depth_to_space_params->block_size * depth_to_space_params->block_size);
cur_height *= depth_to_space_params->block_size;
cur_width *= depth_to_space_params->block_size;
break;
default:
return DNN_ERROR;
}
if (network->layers[layer].output){
av_freep(&network->layers[layer].output);
}
network->layers[layer].output = av_malloc(cur_height * cur_width * cur_channels * sizeof(float));
if (!network->layers[layer].output){
return DNN_ERROR;
}
}
output->data = network->layers[network->layers_num - 1].output;
output->height = cur_height;
output->width = cur_width;
output->channels = cur_channels;
return DNN_SUCCESS;
}
// Loads model and its parameters that are stored in a binary file with following structure:
// layers_num,layer_type,layer_parameterss,layer_type,layer_parameters...
// For CONV layer: activation_function, input_num, output_num, kernel_size, kernel, biases
// For DEPTH_TO_SPACE layer: block_size
DNNModel *ff_dnn_load_model_native(const char *model_filename)
{
DNNModel *model = NULL;
ConvolutionalNetwork *network = NULL;
AVIOContext *model_file_context;
int file_size, dnn_size, kernel_size, i;
int32_t layer;
DNNLayerType layer_type;
ConvolutionalParams *conv_params;
DepthToSpaceParams *depth_to_space_params;
model = av_malloc(sizeof(DNNModel));
if (!model){
return NULL;
}
if (avio_open(&model_file_context, model_filename, AVIO_FLAG_READ) < 0){
av_freep(&model);
return NULL;
}
file_size = avio_size(model_file_context);
network = av_malloc(sizeof(ConvolutionalNetwork));
if (!network){
avio_closep(&model_file_context);
av_freep(&model);
return NULL;
}
model->model = (void *)network;
network->layers_num = 1 + (int32_t)avio_rl32(model_file_context);
dnn_size = 4;
network->layers = av_malloc(network->layers_num * sizeof(Layer));
if (!network->layers){
av_freep(&network);
avio_closep(&model_file_context);
av_freep(&model);
return NULL;
}
for (layer = 0; layer < network->layers_num; ++layer){
network->layers[layer].output = NULL;
network->layers[layer].params = NULL;
}
network->layers[0].type = INPUT;
network->layers[0].params = av_malloc(sizeof(InputParams));
if (!network->layers[0].params){
avio_closep(&model_file_context);
ff_dnn_free_model_native(&model);
return NULL;
}
for (layer = 1; layer < network->layers_num; ++layer){
layer_type = (int32_t)avio_rl32(model_file_context);
dnn_size += 4;
switch (layer_type){
case CONV:
conv_params = av_malloc(sizeof(ConvolutionalParams));
if (!conv_params){
avio_closep(&model_file_context);
ff_dnn_free_model_native(&model);
return NULL;
}
conv_params->activation = (int32_t)avio_rl32(model_file_context);
conv_params->input_num = (int32_t)avio_rl32(model_file_context);
conv_params->output_num = (int32_t)avio_rl32(model_file_context);
conv_params->kernel_size = (int32_t)avio_rl32(model_file_context);
kernel_size = conv_params->input_num * conv_params->output_num *
conv_params->kernel_size * conv_params->kernel_size;
dnn_size += 16 + (kernel_size + conv_params->output_num << 2);
if (dnn_size > file_size || conv_params->input_num <= 0 ||
conv_params->output_num <= 0 || conv_params->kernel_size <= 0){
avio_closep(&model_file_context);
ff_dnn_free_model_native(&model);
return NULL;
}
conv_params->kernel = av_malloc(kernel_size * sizeof(float));
conv_params->biases = av_malloc(conv_params->output_num * sizeof(float));
if (!conv_params->kernel || !conv_params->biases){
avio_closep(&model_file_context);
ff_dnn_free_model_native(&model);
return NULL;
}
for (i = 0; i < kernel_size; ++i){
conv_params->kernel[i] = av_int2float(avio_rl32(model_file_context));
}
for (i = 0; i < conv_params->output_num; ++i){
conv_params->biases[i] = av_int2float(avio_rl32(model_file_context));
}
network->layers[layer].type = CONV;
network->layers[layer].params = conv_params;
break;
case DEPTH_TO_SPACE:
depth_to_space_params = av_malloc(sizeof(DepthToSpaceParams));
if (!depth_to_space_params){
avio_closep(&model_file_context);
ff_dnn_free_model_native(&model);
return NULL;
}
depth_to_space_params->block_size = (int32_t)avio_rl32(model_file_context);
dnn_size += 4;
network->layers[layer].type = DEPTH_TO_SPACE;
network->layers[layer].params = depth_to_space_params;
break;
default:
avio_closep(&model_file_context);
ff_dnn_free_model_native(&model);
return NULL;
}
}
avio_closep(&model_file_context);
if (dnn_size != file_size){
ff_dnn_free_model_native(&model);
return NULL;
}
model->set_input_output = &set_input_output_native;
return model;
}
#define CLAMP_TO_EDGE(x, w) ((x) < 0 ? 0 : ((x) >= (w) ? (w - 1) : (x)))
static void convolve(const float *input, float *output, const ConvolutionalParams *conv_params, int width, int height)
{
int y, x, n_filter, ch, kernel_y, kernel_x;
int radius = conv_params->kernel_size >> 1;
int src_linesize = width * conv_params->input_num;
int filter_linesize = conv_params->kernel_size * conv_params->input_num;
int filter_size = conv_params->kernel_size * filter_linesize;
for (y = 0; y < height; ++y){
for (x = 0; x < width; ++x){
for (n_filter = 0; n_filter < conv_params->output_num; ++n_filter){
output[n_filter] = conv_params->biases[n_filter];
for (ch = 0; ch < conv_params->input_num; ++ch){
for (kernel_y = 0; kernel_y < conv_params->kernel_size; ++kernel_y){
for (kernel_x = 0; kernel_x < conv_params->kernel_size; ++kernel_x){
output[n_filter] += input[CLAMP_TO_EDGE(y + kernel_y - radius, height) * src_linesize +
CLAMP_TO_EDGE(x + kernel_x - radius, width) * conv_params->input_num + ch] *
conv_params->kernel[n_filter * filter_size + kernel_y * filter_linesize +
kernel_x * conv_params->input_num + ch];
}
}
}
switch (conv_params->activation){
case RELU:
output[n_filter] = FFMAX(output[n_filter], 0.0);
break;
case TANH:
output[n_filter] = 2.0f / (1.0f + exp(-2.0f * output[n_filter])) - 1.0f;
break;
case SIGMOID:
output[n_filter] = 1.0f / (1.0f + exp(-output[n_filter]));
}
}
output += conv_params->output_num;
}
}
}
static void depth_to_space(const float *input, float *output, int block_size, int width, int height, int channels)
{
int y, x, by, bx, ch;
int new_channels = channels / (block_size * block_size);
int output_linesize = width * channels;
int by_linesize = output_linesize / block_size;
int x_linesize = new_channels * block_size;
for (y = 0; y < height; ++y){
for (x = 0; x < width; ++x){
for (by = 0; by < block_size; ++by){
for (bx = 0; bx < block_size; ++bx){
for (ch = 0; ch < new_channels; ++ch){
output[by * by_linesize + x * x_linesize + bx * new_channels + ch] = input[ch];
}
input += new_channels;
}
}
}
output += output_linesize;
}
}
DNNReturnType ff_dnn_execute_model_native(const DNNModel *model)
{
ConvolutionalNetwork *network = (ConvolutionalNetwork *)model->model;
int cur_width, cur_height, cur_channels;
int32_t layer;
InputParams *input_params;
ConvolutionalParams *conv_params;
DepthToSpaceParams *depth_to_space_params;
if (network->layers_num <= 0 || network->layers[0].type != INPUT || !network->layers[0].output){
return DNN_ERROR;
}
else{
input_params = (InputParams *)network->layers[0].params;
cur_width = input_params->width;
cur_height = input_params->height;
cur_channels = input_params->channels;
}
for (layer = 1; layer < network->layers_num; ++layer){
if (!network->layers[layer].output){
return DNN_ERROR;
}
switch (network->layers[layer].type){
case CONV:
conv_params = (ConvolutionalParams *)network->layers[layer].params;
convolve(network->layers[layer - 1].output, network->layers[layer].output, conv_params, cur_width, cur_height);
cur_channels = conv_params->output_num;
break;
case DEPTH_TO_SPACE:
depth_to_space_params = (DepthToSpaceParams *)network->layers[layer].params;
depth_to_space(network->layers[layer - 1].output, network->layers[layer].output,
depth_to_space_params->block_size, cur_width, cur_height, cur_channels);
cur_height *= depth_to_space_params->block_size;
cur_width *= depth_to_space_params->block_size;
cur_channels /= depth_to_space_params->block_size * depth_to_space_params->block_size;
break;
case INPUT:
return DNN_ERROR;
}
}
return DNN_SUCCESS;
}
void ff_dnn_free_model_native(DNNModel **model)
{
ConvolutionalNetwork *network;
ConvolutionalParams *conv_params;
int32_t layer;
if (*model)
{
network = (ConvolutionalNetwork *)(*model)->model;
for (layer = 0; layer < network->layers_num; ++layer){
av_freep(&network->layers[layer].output);
if (network->layers[layer].type == CONV){
conv_params = (ConvolutionalParams *)network->layers[layer].params;
av_freep(&conv_params->kernel);
av_freep(&conv_params->biases);
}
av_freep(&network->layers[layer].params);
}
av_freep(&network->layers);
av_freep(&network);
av_freep(model);
}
}