git.sesse.net Git - ffmpeg/blob - libavfilter/dnn/dnn_backend_native_layer_conv2d.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 #include "libavutil/avassert.h"
  22 #include "libavutil/thread.h"
  23 #include "libavutil/cpu.h"
  24 #include "dnn_backend_native_layer_conv2d.h"
  25
  26 #define CLAMP_TO_EDGE(x, w) ((x) < 0 ? 0 : ((x) >= (w) ? (w - 1) : (x)))
  27
  28 //struct to pass parameters
  29 typedef struct thread_common_param{
  30     DnnOperand *operands;
  31     const int32_t *input_operand_indexes;
  32     int32_t output_operand_index;
  33     const void *parameters;
  34     NativeContext *ctx;
  35     float *output_data;
  36 } thread_common_param;
  37
  38 typedef struct thread_param{
  39     thread_common_param *thread_common_param;
  40     int thread_start, thread_end;
  41 } thread_param;
  42
  43 int dnn_load_layer_conv2d(Layer *layer, AVIOContext *model_file_context, int file_size, int operands_num)
  44 {
  45     ConvolutionalParams *conv_params;
  46     int kernel_size;
  47     int dnn_size = 0;
  48     conv_params = av_malloc(sizeof(*conv_params));
  49     if (!conv_params)
  50         return 0;
  51
  52     conv_params->dilation = (int32_t)avio_rl32(model_file_context);
  53     conv_params->padding_method = (int32_t)avio_rl32(model_file_context);
  54     conv_params->activation = (int32_t)avio_rl32(model_file_context);
  55     conv_params->input_num = (int32_t)avio_rl32(model_file_context);
  56     conv_params->output_num = (int32_t)avio_rl32(model_file_context);
  57     conv_params->kernel_size = (int32_t)avio_rl32(model_file_context);
  58     conv_params->has_bias = (int32_t)avio_rl32(model_file_context);
  59     dnn_size += 28;
  60
  61     kernel_size = conv_params->input_num * conv_params->output_num *
  62                       conv_params->kernel_size * conv_params->kernel_size;
  63     dnn_size += kernel_size * 4;
  64     if (conv_params->has_bias)
  65         dnn_size += conv_params->output_num * 4;
  66
  67     if (dnn_size > file_size || conv_params->input_num <= 0 ||
  68         conv_params->output_num <= 0 || conv_params->kernel_size <= 0){
  69         av_freep(&conv_params);
  70         return 0;
  71     }
  72
  73     conv_params->kernel = av_malloc(kernel_size * sizeof(float));
  74     if (!conv_params->kernel) {
  75         av_freep(&conv_params);
  76         return 0;
  77     }
  78     for (int i = 0; i < kernel_size; ++i) {
  79         conv_params->kernel[i] = av_int2float(avio_rl32(model_file_context));
  80     }
  81
  82     conv_params->biases = NULL;
  83     if (conv_params->has_bias) {
  84         conv_params->biases = av_malloc(conv_params->output_num * sizeof(float));
  85         if (!conv_params->biases){
  86             av_freep(&conv_params->kernel);
  87             av_freep(&conv_params);
  88             return 0;
  89         }
  90         for (int i = 0; i < conv_params->output_num; ++i){
  91             conv_params->biases[i] = av_int2float(avio_rl32(model_file_context));
  92         }
  93     }
  94
  95     layer->params = conv_params;
  96
  97     layer->input_operand_indexes[0] = (int32_t)avio_rl32(model_file_context);
  98     layer->output_operand_index = (int32_t)avio_rl32(model_file_context);
  99     dnn_size += 8;
 100
 101     if (layer->input_operand_indexes[0] >= operands_num || layer->output_operand_index >= operands_num) {
 102         return 0;
 103     }
 104
 105     return dnn_size;
 106 }
 107
 108 static void * dnn_execute_layer_conv2d_thread(void *threadarg)
 109 {
 110     //pass parameters
 111     thread_param *thread_param = (struct thread_param *)threadarg;
 112     thread_common_param *thread_common_param = thread_param->thread_common_param;
 113     DnnOperand *operands = thread_common_param->operands;
 114     int32_t input_operand_index = thread_common_param->input_operand_indexes[0];
 115     int height = operands[input_operand_index].dims[1];
 116     int width = operands[input_operand_index].dims[2];
 117     int channel = operands[input_operand_index].dims[3];
 118     const float *input = operands[input_operand_index].data;
 119     const ConvolutionalParams *conv_params = (const ConvolutionalParams *)(thread_common_param->parameters);
 120
 121     int radius = conv_params->kernel_size >> 1;
 122     int src_linesize = width * conv_params->input_num;
 123     int filter_linesize = conv_params->kernel_size * conv_params->input_num;
 124     int filter_size = conv_params->kernel_size * filter_linesize;
 125     int pad_size = (conv_params->padding_method == VALID) ? (conv_params->kernel_size - 1) / 2 * conv_params->dilation : 0;
 126
 127     float *output = thread_common_param->output_data;
 128     output += (conv_params->output_num) * (width - 2 * pad_size) * (thread_param->thread_start - pad_size);
 129
 130     av_assert0(channel == conv_params->input_num);
 131
 132     for (int y = thread_param->thread_start; y < thread_param->thread_end; ++y) {
 133         for (int x = pad_size; x < width - pad_size; ++x) {
 134             for (int n_filter = 0; n_filter < conv_params->output_num; ++n_filter) {
 135                 if (conv_params->has_bias)
 136                     output[n_filter] = conv_params->biases[n_filter];
 137                 else
 138                     output[n_filter] = 0.f;
 139
 140                 for (int ch = 0; ch < conv_params->input_num; ++ch) {
 141                     for (int kernel_y = 0; kernel_y < conv_params->kernel_size; ++kernel_y) {
 142                         for (int kernel_x = 0; kernel_x < conv_params->kernel_size; ++kernel_x) {
 143                             float input_pel;
 144                             if (conv_params->padding_method == SAME_CLAMP_TO_EDGE) {
 145                                 int y_pos = CLAMP_TO_EDGE(y + (kernel_y - radius) * conv_params->dilation, height);
 146                                 int x_pos = CLAMP_TO_EDGE(x + (kernel_x - radius) * conv_params->dilation, width);
 147                                 input_pel = input[y_pos * src_linesize + x_pos * conv_params->input_num + ch];
 148                             } else {
 149                                 int y_pos = y + (kernel_y - radius) * conv_params->dilation;
 150                                 int x_pos = x + (kernel_x - radius) * conv_params->dilation;
 151                                 input_pel = (x_pos < 0 || x_pos >= width || y_pos < 0 || y_pos >= height) ? 0.0 :
 152                                                    input[y_pos * src_linesize + x_pos * conv_params->input_num + ch];
 153                             }
 154
 155
 156                             output[n_filter] += input_pel * conv_params->kernel[n_filter * filter_size + kernel_y * filter_linesize +
 157                                                                                 kernel_x * conv_params->input_num + ch];
 158                         }
 159                     }
 160                 }
 161                 switch (conv_params->activation){
 162                 case RELU:
 163                     output[n_filter] = FFMAX(output[n_filter], 0.0);
 164                     break;
 165                 case TANH:
 166                     output[n_filter] = 2.0f  / (1.0f + exp(-2.0f * output[n_filter])) - 1.0f;
 167                     break;
 168                 case SIGMOID:
 169                     output[n_filter] = 1.0f / (1.0f + exp(-output[n_filter]));
 170                     break;
 171                 case NONE:
 172                     break;
 173                 case LEAKY_RELU:
 174                     output[n_filter] = FFMAX(output[n_filter], 0.0) + 0.2 * FFMIN(output[n_filter], 0.0);
 175                 }
 176             }
 177             output += conv_params->output_num;
 178         }
 179     }
 180     return (void *)DNN_SUCCESS;
 181 }
 182
 183
 184 int dnn_execute_layer_conv2d(DnnOperand *operands, const int32_t *input_operand_indexes,
 185                              int32_t output_operand_index, const void *parameters, NativeContext *ctx)
 186 {
 187     int thread_num = (ctx->options.conv2d_threads <= 0 || ctx->options.conv2d_threads > av_cpu_count())
 188         ? (av_cpu_count() + 1) : (ctx->options.conv2d_threads);
 189 #if HAVE_PTHREAD_CANCEL
 190     pthread_t *thread_id = av_malloc(thread_num * sizeof(pthread_t));
 191     int thread_stride;
 192 #endif
 193     thread_param **thread_param = av_malloc(thread_num * sizeof(*thread_param));
 194     thread_common_param thread_common_param;
 195     const ConvolutionalParams *conv_params = (const ConvolutionalParams *)(parameters);
 196     int height = operands[input_operand_indexes[0]].dims[1];
 197     int width = operands[input_operand_indexes[0]].dims[2];
 198     int pad_size = (conv_params->padding_method == VALID) ? (conv_params->kernel_size - 1) / 2 * conv_params->dilation : 0;
 199     DnnOperand *output_operand = &operands[output_operand_index];
 200
 201     output_operand->dims[0] = operands[input_operand_indexes[0]].dims[0];
 202     output_operand->dims[1] = height - pad_size * 2;
 203     output_operand->dims[2] = width - pad_size * 2;
 204     output_operand->dims[3] = conv_params->output_num;
 205     output_operand->data_type = operands[input_operand_indexes[0]].data_type;
 206     output_operand->length = calculate_operand_data_length(output_operand);
 207     if (output_operand->length <= 0) {
 208         av_log(ctx, AV_LOG_ERROR, "The output data length overflow\n");
 209         return DNN_ERROR;
 210     }
 211     output_operand->data = av_realloc(output_operand->data, output_operand->length);
 212     if (!output_operand->data) {
 213         av_log(ctx, AV_LOG_ERROR, "Failed to reallocate memory for output\n");
 214         return DNN_ERROR;
 215     }
 216     thread_common_param.output_data = output_operand->data;
 217     thread_common_param.operands = operands;
 218     thread_common_param.input_operand_indexes = input_operand_indexes;
 219     thread_common_param.output_operand_index = output_operand_index;
 220     thread_common_param.parameters = parameters;
 221     thread_common_param.ctx = ctx;
 222
 223 #if HAVE_PTHREAD_CANCEL
 224     thread_stride = (height - pad_size * 2) / thread_num;
 225     //create threads
 226     for (int i = 0; i < thread_num; i++){
 227         thread_param[i] = av_malloc(sizeof(**thread_param));
 228         thread_param[i]->thread_common_param = &thread_common_param;
 229         thread_param[i]->thread_start = thread_stride * i + pad_size;
 230         thread_param[i]->thread_end = (i == thread_num - 1) ? (height - pad_size) : (thread_param[i]->thread_start + thread_stride);
 231         pthread_create(&thread_id[i], NULL, dnn_execute_layer_conv2d_thread, (void *)thread_param[i]);
 232     }
 233
 234     //join threads, res gets function return
 235     for (int i = 0; i < thread_num; i++){
 236         pthread_join(thread_id[i], NULL);
 237     }
 238
 239     //release memory
 240     av_free(thread_id);
 241
 242     for (int i = 0; i < thread_num; i++){
 243         av_free(thread_param[i]);
 244     }
 245 #else
 246     thread_param[0] = av_malloc(sizeof(**thread_param));
 247     thread_param[0]->thread_common_param = &thread_common_param;
 248     thread_param[0]->thread_start = pad_size;
 249     thread_param[0]->thread_end = height - pad_size;
 250     dnn_execute_layer_conv2d_thread((void *)thread_param[0]);
 251     av_free(thread_param[0]);
 252 #endif
 253
 254     av_free(thread_param);
 255     return DNN_SUCCESS;
 256 }