]> git.sesse.net Git - ffmpeg/blob - libavfilter/dnn/dnn_backend_native.c
dnn/native: rename struct ConvolutionalNetwork to NativeModel
[ffmpeg] / libavfilter / dnn / dnn_backend_native.c
1 /*
2  * Copyright (c) 2018 Sergey Lavrushkin
3  *
4  * This file is part of FFmpeg.
5  *
6  * FFmpeg is free software; you can redistribute it and/or
7  * modify it under the terms of the GNU Lesser General Public
8  * License as published by the Free Software Foundation; either
9  * version 2.1 of the License, or (at your option) any later version.
10  *
11  * FFmpeg is distributed in the hope that it will be useful,
12  * but WITHOUT ANY WARRANTY; without even the implied warranty of
13  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
14  * Lesser General Public License for more details.
15  *
16  * You should have received a copy of the GNU Lesser General Public
17  * License along with FFmpeg; if not, write to the Free Software
18  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
19  */
20
21 /**
22  * @file
23  * DNN native backend implementation.
24  */
25
26 #include "dnn_backend_native.h"
27 #include "libavutil/avassert.h"
28 #include "dnn_backend_native_layer_conv2d.h"
29 #include "dnn_backend_native_layers.h"
30
31 static DNNReturnType get_input_native(void *model, DNNData *input, const char *input_name)
32 {
33     NativeModel *native_model = (NativeModel *)model;
34
35     for (int i = 0; i < native_model->operands_num; ++i) {
36         DnnOperand *oprd = &native_model->operands[i];
37         if (strcmp(oprd->name, input_name) == 0) {
38             if (oprd->type != DOT_INPUT)
39                 return DNN_ERROR;
40             input->dt = oprd->data_type;
41             av_assert0(oprd->dims[0] == 1);
42             input->height = oprd->dims[1];
43             input->width = oprd->dims[2];
44             input->channels = oprd->dims[3];
45             return DNN_SUCCESS;
46         }
47     }
48
49     // do not find the input operand
50     return DNN_ERROR;
51 }
52
53 static DNNReturnType set_input_output_native(void *model, DNNData *input, const char *input_name, const char **output_names, uint32_t nb_output)
54 {
55     NativeModel *native_model = (NativeModel *)model;
56     DnnOperand *oprd = NULL;
57
58     if (native_model->layers_num <= 0 || native_model->operands_num <= 0)
59         return DNN_ERROR;
60
61     /* inputs */
62     for (int i = 0; i < native_model->operands_num; ++i) {
63         oprd = &native_model->operands[i];
64         if (strcmp(oprd->name, input_name) == 0) {
65             if (oprd->type != DOT_INPUT)
66                 return DNN_ERROR;
67             break;
68         }
69         oprd = NULL;
70     }
71
72     if (!oprd)
73         return DNN_ERROR;
74
75     oprd->dims[0] = 1;
76     oprd->dims[1] = input->height;
77     oprd->dims[2] = input->width;
78     oprd->dims[3] = input->channels;
79
80     av_freep(&oprd->data);
81     oprd->length = calculate_operand_data_length(oprd);
82     if (oprd->length <= 0)
83         return DNN_ERROR;
84     oprd->data = av_malloc(oprd->length);
85     if (!oprd->data)
86         return DNN_ERROR;
87
88     input->data = oprd->data;
89
90     /* outputs */
91     native_model->nb_output = 0;
92     av_freep(&native_model->output_indexes);
93     native_model->output_indexes = av_mallocz_array(nb_output, sizeof(*native_model->output_indexes));
94     if (!native_model->output_indexes)
95         return DNN_ERROR;
96
97     for (uint32_t i = 0; i < nb_output; ++i) {
98         const char *output_name = output_names[i];
99         for (int j = 0; j < native_model->operands_num; ++j) {
100             oprd = &native_model->operands[j];
101             if (strcmp(oprd->name, output_name) == 0) {
102                 native_model->output_indexes[native_model->nb_output++] = j;
103                 break;
104             }
105         }
106     }
107
108     if (native_model->nb_output != nb_output)
109         return DNN_ERROR;
110
111     return DNN_SUCCESS;
112 }
113
114 // Loads model and its parameters that are stored in a binary file with following structure:
115 // layers_num,layer_type,layer_parameterss,layer_type,layer_parameters...
116 // For CONV layer: activation_function, input_num, output_num, kernel_size, kernel, biases
117 // For DEPTH_TO_SPACE layer: block_size
118 DNNModel *ff_dnn_load_model_native(const char *model_filename, const char *options)
119 {
120     DNNModel *model = NULL;
121     char header_expected[] = "FFMPEGDNNNATIVE";
122     char *buf;
123     size_t size;
124     int version, header_size, major_version_expected = 1;
125     NativeModel *native_model = NULL;
126     AVIOContext *model_file_context;
127     int file_size, dnn_size, parsed_size;
128     int32_t layer;
129     DNNLayerType layer_type;
130
131     if (avio_open(&model_file_context, model_filename, AVIO_FLAG_READ) < 0){
132         return NULL;
133     }
134     file_size = avio_size(model_file_context);
135
136     model = av_mallocz(sizeof(DNNModel));
137     if (!model){
138         goto fail;
139     }
140
141     /**
142      * check file header with string and version
143      */
144     size = sizeof(header_expected);
145     buf = av_malloc(size);
146     if (!buf) {
147         goto fail;
148     }
149
150     // size - 1 to skip the ending '\0' which is not saved in file
151     avio_get_str(model_file_context, size - 1, buf, size);
152     dnn_size = size - 1;
153     if (strncmp(buf, header_expected, size) != 0) {
154         av_freep(&buf);
155         goto fail;
156     }
157     av_freep(&buf);
158
159     version = (int32_t)avio_rl32(model_file_context);
160     dnn_size += 4;
161     if (version != major_version_expected) {
162         goto fail;
163     }
164
165     // currently no need to check minor version
166     version = (int32_t)avio_rl32(model_file_context);
167     dnn_size += 4;
168     header_size = dnn_size;
169
170     native_model = av_mallocz(sizeof(NativeModel));
171     if (!native_model){
172         goto fail;
173     }
174     model->model = (void *)native_model;
175
176     avio_seek(model_file_context, file_size - 8, SEEK_SET);
177     native_model->layers_num = (int32_t)avio_rl32(model_file_context);
178     native_model->operands_num = (int32_t)avio_rl32(model_file_context);
179     dnn_size += 8;
180     avio_seek(model_file_context, header_size, SEEK_SET);
181
182     native_model->layers = av_mallocz(native_model->layers_num * sizeof(Layer));
183     if (!native_model->layers){
184         goto fail;
185     }
186
187     native_model->operands = av_mallocz(native_model->operands_num * sizeof(DnnOperand));
188     if (!native_model->operands){
189         goto fail;
190     }
191
192     for (layer = 0; layer < native_model->layers_num; ++layer){
193         layer_type = (int32_t)avio_rl32(model_file_context);
194         dnn_size += 4;
195
196         if (layer_type >= DLT_COUNT) {
197             goto fail;
198         }
199
200         native_model->layers[layer].type = layer_type;
201         parsed_size = layer_funcs[layer_type].pf_load(&native_model->layers[layer], model_file_context, file_size, native_model->operands_num);
202         if (!parsed_size) {
203             goto fail;
204         }
205         dnn_size += parsed_size;
206     }
207
208     for (int32_t i = 0; i < native_model->operands_num; ++i){
209         DnnOperand *oprd;
210         int32_t name_len;
211         int32_t operand_index = (int32_t)avio_rl32(model_file_context);
212         dnn_size += 4;
213
214         if (operand_index >= native_model->operands_num) {
215             goto fail;
216         }
217
218         oprd = &native_model->operands[operand_index];
219         name_len = (int32_t)avio_rl32(model_file_context);
220         dnn_size += 4;
221
222         avio_get_str(model_file_context, name_len, oprd->name, sizeof(oprd->name));
223         dnn_size += name_len;
224
225         oprd->type = (int32_t)avio_rl32(model_file_context);
226         dnn_size += 4;
227
228         oprd->data_type = (int32_t)avio_rl32(model_file_context);
229         dnn_size += 4;
230
231         for (int32_t dim = 0; dim < 4; ++dim) {
232             oprd->dims[dim] = (int32_t)avio_rl32(model_file_context);
233             dnn_size += 4;
234         }
235
236         oprd->isNHWC = 1;
237     }
238
239     avio_closep(&model_file_context);
240
241     if (dnn_size != file_size){
242         ff_dnn_free_model_native(&model);
243         return NULL;
244     }
245
246     model->set_input_output = &set_input_output_native;
247     model->get_input = &get_input_native;
248     model->options = options;
249
250     return model;
251
252 fail:
253     ff_dnn_free_model_native(&model);
254     avio_closep(&model_file_context);
255     return NULL;
256 }
257
258 DNNReturnType ff_dnn_execute_model_native(const DNNModel *model, DNNData *outputs, uint32_t nb_output)
259 {
260     NativeModel *native_model = (NativeModel *)model->model;
261     int32_t layer;
262     uint32_t nb = FFMIN(nb_output, native_model->nb_output);
263
264     if (native_model->layers_num <= 0 || native_model->operands_num <= 0)
265         return DNN_ERROR;
266     if (!native_model->operands[0].data)
267         return DNN_ERROR;
268
269     for (layer = 0; layer < native_model->layers_num; ++layer){
270         DNNLayerType layer_type = native_model->layers[layer].type;
271         layer_funcs[layer_type].pf_exec(native_model->operands,
272                                         native_model->layers[layer].input_operand_indexes,
273                                         native_model->layers[layer].output_operand_index,
274                                         native_model->layers[layer].params);
275     }
276
277     for (uint32_t i = 0; i < nb; ++i) {
278         DnnOperand *oprd = &native_model->operands[native_model->output_indexes[i]];
279         outputs[i].data = oprd->data;
280         outputs[i].height = oprd->dims[1];
281         outputs[i].width = oprd->dims[2];
282         outputs[i].channels = oprd->dims[3];
283         outputs[i].dt = oprd->data_type;
284     }
285
286     return DNN_SUCCESS;
287 }
288
289 int32_t calculate_operand_dims_count(const DnnOperand *oprd)
290 {
291     int32_t result = 1;
292     for (int i = 0; i < 4; ++i)
293         result *= oprd->dims[i];
294
295     return result;
296 }
297
298 int32_t calculate_operand_data_length(const DnnOperand* oprd)
299 {
300     // currently, we just support DNN_FLOAT
301     uint64_t len = sizeof(float);
302     for (int i = 0; i < 4; i++) {
303         len *= oprd->dims[i];
304         if (len > INT32_MAX)
305             return 0;
306     }
307     return len;
308 }
309
310 void ff_dnn_free_model_native(DNNModel **model)
311 {
312     NativeModel *native_model;
313     ConvolutionalParams *conv_params;
314     int32_t layer;
315
316     if (*model)
317     {
318         if ((*model)->model) {
319             native_model = (NativeModel *)(*model)->model;
320             if (native_model->layers) {
321                 for (layer = 0; layer < native_model->layers_num; ++layer){
322                     if (native_model->layers[layer].type == DLT_CONV2D){
323                         conv_params = (ConvolutionalParams *)native_model->layers[layer].params;
324                         av_freep(&conv_params->kernel);
325                         av_freep(&conv_params->biases);
326                     }
327                     av_freep(&native_model->layers[layer].params);
328                 }
329                 av_freep(&native_model->layers);
330             }
331
332             if (native_model->operands) {
333                 for (uint32_t operand = 0; operand < native_model->operands_num; ++operand)
334                     av_freep(&native_model->operands[operand].data);
335                 av_freep(&native_model->operands);
336             }
337
338             av_freep(&native_model->output_indexes);
339             av_freep(&native_model);
340         }
341         av_freep(model);
342     }
343 }