2 * Copyright (c) 2018 Sergey Lavrushkin
4 * This file is part of FFmpeg.
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.
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.
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
23 * DNN native backend implementation.
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"
31 static DNNReturnType get_input_native(void *model, DNNData *input, const char *input_name)
33 NativeModel *native_model = (NativeModel *)model;
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)
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];
49 // do not find the input operand
53 static DNNReturnType set_input_native(void *model, DNNData *input, const char *input_name)
55 NativeModel *native_model = (NativeModel *)model;
56 DnnOperand *oprd = NULL;
58 if (native_model->layers_num <= 0 || native_model->operands_num <= 0)
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)
76 oprd->dims[1] = input->height;
77 oprd->dims[2] = input->width;
78 oprd->dims[3] = input->channels;
80 av_freep(&oprd->data);
81 oprd->length = calculate_operand_data_length(oprd);
82 if (oprd->length <= 0)
84 oprd->data = av_malloc(oprd->length);
88 input->data = oprd->data;
93 // Loads model and its parameters that are stored in a binary file with following structure:
94 // layers_num,layer_type,layer_parameterss,layer_type,layer_parameters...
95 // For CONV layer: activation_function, input_num, output_num, kernel_size, kernel, biases
96 // For DEPTH_TO_SPACE layer: block_size
97 DNNModel *ff_dnn_load_model_native(const char *model_filename, const char *options)
99 DNNModel *model = NULL;
100 char header_expected[] = "FFMPEGDNNNATIVE";
103 int version, header_size, major_version_expected = 1;
104 NativeModel *native_model = NULL;
105 AVIOContext *model_file_context;
106 int file_size, dnn_size, parsed_size;
108 DNNLayerType layer_type;
110 if (avio_open(&model_file_context, model_filename, AVIO_FLAG_READ) < 0){
113 file_size = avio_size(model_file_context);
115 model = av_mallocz(sizeof(DNNModel));
121 * check file header with string and version
123 size = sizeof(header_expected);
124 buf = av_malloc(size);
129 // size - 1 to skip the ending '\0' which is not saved in file
130 avio_get_str(model_file_context, size - 1, buf, size);
132 if (strncmp(buf, header_expected, size) != 0) {
138 version = (int32_t)avio_rl32(model_file_context);
140 if (version != major_version_expected) {
144 // currently no need to check minor version
145 version = (int32_t)avio_rl32(model_file_context);
147 header_size = dnn_size;
149 native_model = av_mallocz(sizeof(NativeModel));
153 model->model = (void *)native_model;
155 avio_seek(model_file_context, file_size - 8, SEEK_SET);
156 native_model->layers_num = (int32_t)avio_rl32(model_file_context);
157 native_model->operands_num = (int32_t)avio_rl32(model_file_context);
159 avio_seek(model_file_context, header_size, SEEK_SET);
161 native_model->layers = av_mallocz(native_model->layers_num * sizeof(Layer));
162 if (!native_model->layers){
166 native_model->operands = av_mallocz(native_model->operands_num * sizeof(DnnOperand));
167 if (!native_model->operands){
171 for (layer = 0; layer < native_model->layers_num; ++layer){
172 layer_type = (int32_t)avio_rl32(model_file_context);
175 if (layer_type >= DLT_COUNT) {
179 native_model->layers[layer].type = layer_type;
180 parsed_size = layer_funcs[layer_type].pf_load(&native_model->layers[layer], model_file_context, file_size, native_model->operands_num);
184 dnn_size += parsed_size;
187 for (int32_t i = 0; i < native_model->operands_num; ++i){
190 int32_t operand_index = (int32_t)avio_rl32(model_file_context);
193 if (operand_index >= native_model->operands_num) {
197 oprd = &native_model->operands[operand_index];
198 name_len = (int32_t)avio_rl32(model_file_context);
201 avio_get_str(model_file_context, name_len, oprd->name, sizeof(oprd->name));
202 dnn_size += name_len;
204 oprd->type = (int32_t)avio_rl32(model_file_context);
207 oprd->data_type = (int32_t)avio_rl32(model_file_context);
210 for (int32_t dim = 0; dim < 4; ++dim) {
211 oprd->dims[dim] = (int32_t)avio_rl32(model_file_context);
218 avio_closep(&model_file_context);
220 if (dnn_size != file_size){
221 ff_dnn_free_model_native(&model);
225 model->set_input = &set_input_native;
226 model->get_input = &get_input_native;
227 model->options = options;
232 ff_dnn_free_model_native(&model);
233 avio_closep(&model_file_context);
237 DNNReturnType ff_dnn_execute_model_native(const DNNModel *model, DNNData *outputs, const char **output_names, uint32_t nb_output)
239 NativeModel *native_model = (NativeModel *)model->model;
242 if (native_model->layers_num <= 0 || native_model->operands_num <= 0)
244 if (!native_model->operands[0].data)
247 for (layer = 0; layer < native_model->layers_num; ++layer){
248 DNNLayerType layer_type = native_model->layers[layer].type;
249 layer_funcs[layer_type].pf_exec(native_model->operands,
250 native_model->layers[layer].input_operand_indexes,
251 native_model->layers[layer].output_operand_index,
252 native_model->layers[layer].params);
255 for (uint32_t i = 0; i < nb_output; ++i) {
256 DnnOperand *oprd = NULL;
257 const char *output_name = output_names[i];
258 for (int j = 0; j < native_model->operands_num; ++j) {
259 if (strcmp(native_model->operands[j].name, output_name) == 0) {
260 oprd = &native_model->operands[j];
268 outputs[i].data = oprd->data;
269 outputs[i].height = oprd->dims[1];
270 outputs[i].width = oprd->dims[2];
271 outputs[i].channels = oprd->dims[3];
272 outputs[i].dt = oprd->data_type;
278 int32_t calculate_operand_dims_count(const DnnOperand *oprd)
281 for (int i = 0; i < 4; ++i)
282 result *= oprd->dims[i];
287 int32_t calculate_operand_data_length(const DnnOperand* oprd)
289 // currently, we just support DNN_FLOAT
290 uint64_t len = sizeof(float);
291 for (int i = 0; i < 4; i++) {
292 len *= oprd->dims[i];
299 void ff_dnn_free_model_native(DNNModel **model)
301 NativeModel *native_model;
302 ConvolutionalParams *conv_params;
307 if ((*model)->model) {
308 native_model = (NativeModel *)(*model)->model;
309 if (native_model->layers) {
310 for (layer = 0; layer < native_model->layers_num; ++layer){
311 if (native_model->layers[layer].type == DLT_CONV2D){
312 conv_params = (ConvolutionalParams *)native_model->layers[layer].params;
313 av_freep(&conv_params->kernel);
314 av_freep(&conv_params->biases);
316 av_freep(&native_model->layers[layer].params);
318 av_freep(&native_model->layers);
321 if (native_model->operands) {
322 for (uint32_t operand = 0; operand < native_model->operands_num; ++operand)
323 av_freep(&native_model->operands[operand].data);
324 av_freep(&native_model->operands);
327 av_freep(&native_model);