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     int thread_num;
  37 } thread_common_param;
  38
  39 typedef struct thread_param{
  40     thread_common_param *thread_common_param;
  41     int thread_index;
  42 } thread_param;
  43
  44 int dnn_load_layer_conv2d(Layer *layer, AVIOContext *model_file_context, int file_size, int operands_num)
  45 {
  46     ConvolutionalParams *conv_params;
  47     int kernel_size;
  48     int dnn_size = 0;
  49     conv_params = av_malloc(sizeof(*conv_params));
  50     if (!conv_params)
  51         return 0;
  52
  53     conv_params->dilation = (int32_t)avio_rl32(model_file_context);
  54     conv_params->padding_method = (int32_t)avio_rl32(model_file_context);
  55     conv_params->activation = (int32_t)avio_rl32(model_file_context);
  56     conv_params->input_num = (int32_t)avio_rl32(model_file_context);
  57     conv_params->output_num = (int32_t)avio_rl32(model_file_context);
  58     conv_params->kernel_size = (int32_t)avio_rl32(model_file_context);
  59     conv_params->has_bias = (int32_t)avio_rl32(model_file_context);
  60     dnn_size += 28;
  61
  62     kernel_size = conv_params->input_num * conv_params->output_num *
  63                       conv_params->kernel_size * conv_params->kernel_size;
  64     dnn_size += kernel_size * 4;
  65     if (conv_params->has_bias)
  66         dnn_size += conv_params->output_num * 4;
  67
  68     if (dnn_size > file_size || conv_params->input_num <= 0 ||
  69         conv_params->output_num <= 0 || conv_params->kernel_size <= 0){
  70         av_freep(&conv_params);
  71         return 0;
  72     }
  73
  74     conv_params->kernel = av_malloc(kernel_size * sizeof(float));
  75     if (!conv_params->kernel) {
  76         av_freep(&conv_params);
  77         return 0;
  78     }
  79     for (int i = 0; i < kernel_size; ++i) {
  80         conv_params->kernel[i] = av_int2float(avio_rl32(model_file_context));
  81     }
  82
  83     conv_params->biases = NULL;
  84     if (conv_params->has_bias) {
  85         conv_params->biases = av_malloc(conv_params->output_num * sizeof(float));
  86         if (!conv_params->biases){
  87             av_freep(&conv_params->kernel);
  88             av_freep(&conv_params);
  89             return 0;
  90         }
  91         for (int i = 0; i < conv_params->output_num; ++i){
  92             conv_params->biases[i] = av_int2float(avio_rl32(model_file_context));
  93         }
  94     }
  95
  96     layer->params = conv_params;
  97
  98     layer->input_operand_indexes[0] = (int32_t)avio_rl32(model_file_context);
  99     layer->output_operand_index = (int32_t)avio_rl32(model_file_context);
 100     dnn_size += 8;
 101
 102     if (layer->input_operand_indexes[0] >= operands_num || layer->output_operand_index >= operands_num) {
 103         return 0;
 104     }
 105
 106     return dnn_size;
 107 }
 108
 109 static void * dnn_execute_layer_conv2d_thread(void *threadarg)
 110 {
 111     //pass parameters
 112     thread_param *thread_param = (struct thread_param *)threadarg;
 113     thread_common_param *thread_common_param = thread_param->thread_common_param;
 114     DnnOperand *operands = thread_common_param->operands;
 115     int32_t input_operand_index = thread_common_param->input_operand_indexes[0];
 116     int height = operands[input_operand_index].dims[1];
 117     int width = operands[input_operand_index].dims[2];
 118     int channel = operands[input_operand_index].dims[3];
 119     const float *input = operands[input_operand_index].data;
 120     const ConvolutionalParams *conv_params = (const ConvolutionalParams *)(thread_common_param->parameters);
 121
 122     int radius = conv_params->kernel_size >> 1;
 123     int src_linesize = width * conv_params->input_num;
 124     int filter_linesize = conv_params->kernel_size * conv_params->input_num;
 125     int filter_size = conv_params->kernel_size * filter_linesize;
 126     int pad_size = (conv_params->padding_method == VALID) ? (conv_params->kernel_size - 1) / 2 * conv_params->dilation : 0;
 127
 128     int thread_stride = (height - pad_size * 2) / thread_common_param->thread_num;
 129     int thread_start = thread_stride * thread_param->thread_index + pad_size;
 130     int thread_end = (thread_param->thread_index == thread_common_param->thread_num - 1) ? (height - pad_size) : (thread_start + thread_stride);
 131
 132     float *output = thread_common_param->output_data;
 133     output += (conv_params->output_num) * (width - 2 * pad_size) * (thread_start - pad_size);
 134
 135     av_assert0(channel == conv_params->input_num);
 136
 137     for (int y = thread_start; y < thread_end; ++y) {
 138         for (int x = pad_size; x < width - pad_size; ++x) {
 139             for (int n_filter = 0; n_filter < conv_params->output_num; ++n_filter) {
 140                 if (conv_params->has_bias)
 141                     output[n_filter] = conv_params->biases[n_filter];
 142                 else
 143                     output[n_filter] = 0.f;
 144
 145                 for (int ch = 0; ch < conv_params->input_num; ++ch) {
 146                     for (int kernel_y = 0; kernel_y < conv_params->kernel_size; ++kernel_y) {
 147                         for (int kernel_x = 0; kernel_x < conv_params->kernel_size; ++kernel_x) {
 148                             float input_pel;
 149                             if (conv_params->padding_method == SAME_CLAMP_TO_EDGE) {
 150                                 int y_pos = CLAMP_TO_EDGE(y + (kernel_y - radius) * conv_params->dilation, height);
 151                                 int x_pos = CLAMP_TO_EDGE(x + (kernel_x - radius) * conv_params->dilation, width);
 152                                 input_pel = input[y_pos * src_linesize + x_pos * conv_params->input_num + ch];
 153                             } else {
 154                                 int y_pos = y + (kernel_y - radius) * conv_params->dilation;
 155                                 int x_pos = x + (kernel_x - radius) * conv_params->dilation;
 156                                 input_pel = (x_pos < 0 || x_pos >= width || y_pos < 0 || y_pos >= height) ? 0.0 :
 157                                                    input[y_pos * src_linesize + x_pos * conv_params->input_num + ch];
 158                             }
 159
 160
 161                             output[n_filter] += input_pel * conv_params->kernel[n_filter * filter_size + kernel_y * filter_linesize +
 162                                                                                 kernel_x * conv_params->input_num + ch];
 163                         }
 164                     }
 165                 }
 166                 switch (conv_params->activation){
 167                 case RELU:
 168                     output[n_filter] = FFMAX(output[n_filter], 0.0);
 169                     break;
 170                 case TANH:
 171                     output[n_filter] = 2.0f  / (1.0f + exp(-2.0f * output[n_filter])) - 1.0f;
 172                     break;
 173                 case SIGMOID:
 174                     output[n_filter] = 1.0f / (1.0f + exp(-output[n_filter]));
 175                     break;
 176                 case NONE:
 177                     break;
 178                 case LEAKY_RELU:
 179                     output[n_filter] = FFMAX(output[n_filter], 0.0) + 0.2 * FFMIN(output[n_filter], 0.0);
 180                 }
 181             }
 182             output += conv_params->output_num;
 183         }
 184     }
 185     return (void *)DNN_SUCCESS;
 186 }
 187
 188
 189 int dnn_execute_layer_conv2d(DnnOperand *operands, const int32_t *input_operand_indexes,
 190                              int32_t output_operand_index, const void *parameters, NativeContext *ctx)
 191 {
 192     int thread_num = (ctx->options.conv2d_threads <= 0 || ctx->options.conv2d_threads > av_cpu_count())
 193         ? (av_cpu_count() + 1) : (ctx->options.conv2d_threads);
 194 #if HAVE_PTHREAD_CANCEL
 195     pthread_t *thread_id = av_malloc(thread_num * sizeof(pthread_t));
 196 #endif
 197     thread_param **thread_param = av_malloc(thread_num * sizeof(*thread_param));
 198     thread_common_param thread_common_param;
 199     const ConvolutionalParams *conv_params = (const ConvolutionalParams *)(parameters);
 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];
 202
 203     output_operand->dims[0] = operands[input_operand_indexes[0]].dims[0];
 204     output_operand->dims[1] = operands[input_operand_indexes[0]].dims[1] - pad_size * 2;
 205     output_operand->dims[2] = operands[input_operand_indexes[0]].dims[2] - pad_size * 2;
 206     output_operand->dims[3] = conv_params->output_num;
 207     output_operand->data_type = operands[input_operand_indexes[0]].data_type;
 208     output_operand->length = calculate_operand_data_length(output_operand);
 209     if (output_operand->length <= 0) {
 210         av_log(ctx, AV_LOG_ERROR, "The output data length overflow\n");
 211         return DNN_ERROR;
 212     }
 213     output_operand->data = av_realloc(output_operand->data, output_operand->length);
 214     if (!output_operand->data) {
 215         av_log(ctx, AV_LOG_ERROR, "Failed to reallocate memory for output\n");
 216         return DNN_ERROR;
 217     }
 218     thread_common_param.output_data = output_operand->data;
 219     thread_common_param.operands = operands;
 220     thread_common_param.input_operand_indexes = input_operand_indexes;
 221     thread_common_param.output_operand_index = output_operand_index;
 222     thread_common_param.parameters = parameters;
 223     thread_common_param.ctx = ctx;
 224
 225 #if HAVE_PTHREAD_CANCEL
 226     thread_common_param.thread_num = thread_num;
 227
 228     //create threads
 229     for (int i = 0; i < thread_num; i++){
 230         thread_param[i] = av_malloc(sizeof(**thread_param));
 231         thread_param[i]->thread_common_param = &thread_common_param;
 232         thread_param[i]->thread_index = i;
 233         pthread_create(&thread_id[i], NULL, dnn_execute_layer_conv2d_thread, (void *)thread_param[i]);
 234     }
 235
 236     //join threads, res gets function return
 237     for (int i = 0; i < thread_num; i++){
 238         pthread_join(thread_id[i], NULL);
 239     }
 240
 241     //release memory
 242     av_free(thread_id);
 243
 244     for (int i = 0; i < thread_num; i++){
 245         av_free(thread_param[i]);
 246     }
 247 #else
 248     thread_common_param.thread_num = 1;
 249     thread_param[0] = av_malloc(sizeof(thread_param));
 250     thread_param[0]->thread_common_param = &thread_common_param;
 251     thread_param[0]->thread_index = 0;
 252     dnn_execute_layer_conv2d_thread((void *)thread_param[0]);
 253     av_free(thread_param[0]);
 254 #endif
 255
 256     av_free(thread_param);
 257     return DNN_SUCCESS;
 258 }