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
21 #include "libavutil/avassert.h"
22 #include "libavutil/thread.h"
23 #include "libavutil/cpu.h"
24 #include "dnn_backend_native_layer_conv2d.h"
26 #define CLAMP_TO_EDGE(x, w) ((x) < 0 ? 0 : ((x) >= (w) ? (w - 1) : (x)))
28 //struct to pass parameters
29 typedef struct ThreadCommonParam{
31 const int32_t *input_operand_indexes;
32 int32_t output_operand_index;
33 const void *parameters;
38 typedef struct ThreadParam{
39 ThreadCommonParam *thread_common_param;
40 int thread_start, thread_end;
41 #if HAVE_PTHREAD_CANCEL
46 int ff_dnn_load_layer_conv2d(Layer *layer, AVIOContext *model_file_context, int file_size, int operands_num)
48 ConvolutionalParams *conv_params;
51 conv_params = av_malloc(sizeof(*conv_params));
55 conv_params->dilation = (int32_t)avio_rl32(model_file_context);
56 conv_params->padding_method = (int32_t)avio_rl32(model_file_context);
57 conv_params->activation = (int32_t)avio_rl32(model_file_context);
58 conv_params->input_num = (int32_t)avio_rl32(model_file_context);
59 conv_params->output_num = (int32_t)avio_rl32(model_file_context);
60 conv_params->kernel_size = (int32_t)avio_rl32(model_file_context);
61 conv_params->has_bias = (int32_t)avio_rl32(model_file_context);
64 kernel_size = conv_params->input_num * conv_params->output_num *
65 conv_params->kernel_size * conv_params->kernel_size;
66 dnn_size += kernel_size * 4;
67 if (conv_params->has_bias)
68 dnn_size += conv_params->output_num * 4;
70 if (dnn_size > file_size || conv_params->input_num <= 0 ||
71 conv_params->output_num <= 0 || conv_params->kernel_size <= 0){
72 av_freep(&conv_params);
76 conv_params->kernel = av_malloc_array(kernel_size, sizeof(*conv_params->kernel));
77 if (!conv_params->kernel) {
78 av_freep(&conv_params);
81 for (int i = 0; i < kernel_size; ++i) {
82 conv_params->kernel[i] = av_int2float(avio_rl32(model_file_context));
85 conv_params->biases = NULL;
86 if (conv_params->has_bias) {
87 conv_params->biases = av_malloc_array(conv_params->output_num, sizeof(*conv_params->biases));
88 if (!conv_params->biases){
89 av_freep(&conv_params->kernel);
90 av_freep(&conv_params);
93 for (int i = 0; i < conv_params->output_num; ++i){
94 conv_params->biases[i] = av_int2float(avio_rl32(model_file_context));
98 layer->params = conv_params;
100 layer->input_operand_indexes[0] = (int32_t)avio_rl32(model_file_context);
101 layer->output_operand_index = (int32_t)avio_rl32(model_file_context);
104 if (layer->input_operand_indexes[0] >= operands_num || layer->output_operand_index >= operands_num) {
111 static void * dnn_execute_layer_conv2d_thread(void *threadarg)
114 ThreadParam *thread_param = threadarg;
115 ThreadCommonParam *thread_common_param = thread_param->thread_common_param;
116 DnnOperand *operands = thread_common_param->operands;
117 int32_t input_operand_index = thread_common_param->input_operand_indexes[0];
118 int height = operands[input_operand_index].dims[1];
119 int width = operands[input_operand_index].dims[2];
120 int channel = operands[input_operand_index].dims[3];
121 const float *input = operands[input_operand_index].data;
122 const ConvolutionalParams *conv_params = thread_common_param->parameters;
124 int radius = conv_params->kernel_size >> 1;
125 int src_linesize = width * conv_params->input_num;
126 int filter_linesize = conv_params->kernel_size * conv_params->input_num;
127 int filter_size = conv_params->kernel_size * filter_linesize;
128 int pad_size = (conv_params->padding_method == VALID) ? (conv_params->kernel_size - 1) / 2 * conv_params->dilation : 0;
130 float *output = thread_common_param->output_data;
131 output += (conv_params->output_num) * (width - 2 * pad_size) * (thread_param->thread_start - pad_size);
133 av_assert0(channel == conv_params->input_num);
135 for (int y = thread_param->thread_start; y < thread_param->thread_end; ++y) {
136 for (int x = pad_size; x < width - pad_size; ++x) {
137 for (int n_filter = 0; n_filter < conv_params->output_num; ++n_filter) {
138 if (conv_params->has_bias)
139 output[n_filter] = conv_params->biases[n_filter];
141 output[n_filter] = 0.f;
143 for (int ch = 0; ch < conv_params->input_num; ++ch) {
144 for (int kernel_y = 0; kernel_y < conv_params->kernel_size; ++kernel_y) {
145 for (int kernel_x = 0; kernel_x < conv_params->kernel_size; ++kernel_x) {
147 if (conv_params->padding_method == SAME_CLAMP_TO_EDGE) {
148 int y_pos = CLAMP_TO_EDGE(y + (kernel_y - radius) * conv_params->dilation, height);
149 int x_pos = CLAMP_TO_EDGE(x + (kernel_x - radius) * conv_params->dilation, width);
150 input_pel = input[y_pos * src_linesize + x_pos * conv_params->input_num + ch];
152 int y_pos = y + (kernel_y - radius) * conv_params->dilation;
153 int x_pos = x + (kernel_x - radius) * conv_params->dilation;
154 input_pel = (x_pos < 0 || x_pos >= width || y_pos < 0 || y_pos >= height) ? 0.0 :
155 input[y_pos * src_linesize + x_pos * conv_params->input_num + ch];
159 output[n_filter] += input_pel * conv_params->kernel[n_filter * filter_size + kernel_y * filter_linesize +
160 kernel_x * conv_params->input_num + ch];
164 switch (conv_params->activation){
166 output[n_filter] = FFMAX(output[n_filter], 0.0);
169 output[n_filter] = 2.0f / (1.0f + exp(-2.0f * output[n_filter])) - 1.0f;
172 output[n_filter] = 1.0f / (1.0f + exp(-output[n_filter]));
177 output[n_filter] = FFMAX(output[n_filter], 0.0) + 0.2 * FFMIN(output[n_filter], 0.0);
180 output += conv_params->output_num;
183 return (void *)DNN_SUCCESS;
187 int ff_dnn_execute_layer_conv2d(DnnOperand *operands, const int32_t *input_operand_indexes,
188 int32_t output_operand_index, const void *parameters, NativeContext *ctx)
190 int thread_num = (ctx->options.conv2d_threads <= 0 || ctx->options.conv2d_threads > av_cpu_count())
191 ? (av_cpu_count() + 1) : (ctx->options.conv2d_threads);
192 #if HAVE_PTHREAD_CANCEL
195 ThreadParam **thread_param = av_malloc_array(thread_num, sizeof(*thread_param));
196 ThreadCommonParam thread_common_param;
197 const ConvolutionalParams *conv_params = parameters;
198 int height = operands[input_operand_indexes[0]].dims[1];
199 int width = operands[input_operand_indexes[0]].dims[2];
200 int pad_size = (conv_params->padding_method == VALID) ? (conv_params->kernel_size - 1) / 2 * conv_params->dilation : 0;
201 DnnOperand *output_operand = &operands[output_operand_index];
204 output_operand->dims[0] = operands[input_operand_indexes[0]].dims[0];
205 output_operand->dims[1] = height - pad_size * 2;
206 output_operand->dims[2] = width - pad_size * 2;
207 output_operand->dims[3] = conv_params->output_num;
208 output_operand->data_type = operands[input_operand_indexes[0]].data_type;
209 output_operand->length = ff_calculate_operand_data_length(output_operand);
210 if (output_operand->length <= 0) {
211 av_log(ctx, AV_LOG_ERROR, "The output data length overflow\n");
214 tmp = av_realloc(output_operand->data, output_operand->length);
216 av_log(ctx, AV_LOG_ERROR, "Failed to reallocate memory for output\n");
219 output_operand->data = tmp;
220 thread_common_param.output_data = output_operand->data;
221 thread_common_param.operands = operands;
222 thread_common_param.input_operand_indexes = input_operand_indexes;
223 thread_common_param.output_operand_index = output_operand_index;
224 thread_common_param.parameters = parameters;
225 thread_common_param.ctx = ctx;
227 #if HAVE_PTHREAD_CANCEL
228 thread_stride = (height - pad_size * 2) / thread_num;
230 for (int i = 0; i < thread_num; i++){
231 thread_param[i] = av_malloc(sizeof(*thread_param[0]));
232 thread_param[i]->thread_common_param = &thread_common_param;
233 thread_param[i]->thread_start = thread_stride * i + pad_size;
234 thread_param[i]->thread_end = (i == thread_num - 1) ? (height - pad_size) : (thread_param[i]->thread_start + thread_stride);
235 pthread_create(&thread_param[i]->thread, NULL, dnn_execute_layer_conv2d_thread, (void *)thread_param[i]);
238 //join threads, res gets function return
239 for (int i = 0; i < thread_num; i++){
240 pthread_join(thread_param[i]->thread, NULL);
244 for (int i = 0; i < thread_num; i++){
245 av_freep(&thread_param[i]);
248 thread_param[0] = av_malloc(sizeof(*thread_param[0]));
249 thread_param[0]->thread_common_param = &thread_common_param;
250 thread_param[0]->thread_start = pad_size;
251 thread_param[0]->thread_end = height - pad_size;
252 dnn_execute_layer_conv2d_thread((void *)thread_param[0]);
253 av_freep(&thread_param[0]);
256 av_freep(&thread_param);