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 "libavutil/avassert.h"
27 #include "dnn_backend_native_layer_avgpool.h"
29 int ff_dnn_load_layer_avg_pool(Layer *layer, AVIOContext *model_file_context, int file_size, int operands_num)
31 AvgPoolParams *avgpool_params;
33 avgpool_params = av_malloc(sizeof(*avgpool_params));
37 avgpool_params->strides = (int32_t)avio_rl32(model_file_context);
38 avgpool_params->padding_method = (int32_t)avio_rl32(model_file_context);
39 avgpool_params->kernel_size = (int32_t)avio_rl32(model_file_context);
42 if (dnn_size > file_size || avgpool_params->kernel_size <= 0 || avgpool_params->strides <=0){
43 av_freep(&avgpool_params);
47 layer->params = avgpool_params;
48 layer->input_operand_indexes[0] = (int32_t)avio_rl32(model_file_context);
49 layer->output_operand_index = (int32_t)avio_rl32(model_file_context);
52 if (layer->input_operand_indexes[0] >= operands_num || layer->output_operand_index >= operands_num) {
58 int ff_dnn_execute_layer_avg_pool(DnnOperand *operands, const int32_t *input_operand_indexes,
59 int32_t output_operand_index, const void *parameters, NativeContext *ctx)
62 int height_end, width_end, height_radius, width_radius, output_height, output_width, kernel_area;
63 int32_t input_operand_index = input_operand_indexes[0];
64 int number = operands[input_operand_index].dims[0];
65 int height = operands[input_operand_index].dims[1];
66 int width = operands[input_operand_index].dims[2];
67 int channel = operands[input_operand_index].dims[3];
68 const float *input = operands[input_operand_index].data;
69 const AvgPoolParams *avgpool_params = parameters;
71 int kernel_strides = avgpool_params->strides;
72 int src_linesize = width * channel;
73 DnnOperand *output_operand = &operands[output_operand_index];
76 * When padding_method = SAME, the tensorflow will only padding the hald number of 0 pxiels
77 * except the remainders.
78 * Eg: assuming the input height = 1080, the strides = 11, so the remainders = 1080 % 11 = 2
79 * and if ksize = 5: it will fill (5 - 2) >> 1 = 1 line before the first line of input image,
80 * and 5 - 2 - 1 = 2 lines after the last line of input image.
81 * and if ksize = 7: it will fill (7 - 2) >> 1 = 2 lines before the first line of input image,
82 * and 7 - 2 - 2 = 3 lines after the last line of input image.
84 if (avgpool_params->padding_method == SAME) {
87 height_radius = avgpool_params->kernel_size - ((height - 1) % kernel_strides + 1);
88 width_radius = avgpool_params->kernel_size - ((width - 1) % kernel_strides + 1);
89 height_radius = height_radius < 0 ? 0 : height_radius >> 1;
90 width_radius = width_radius < 0 ? 0 : width_radius >> 1;
91 output_height = ceil(height / (kernel_strides * 1.0));
92 output_width = ceil(width / (kernel_strides * 1.0));
94 av_assert0(avgpool_params->padding_method == VALID);
95 height_end = height - avgpool_params->kernel_size + 1;
96 width_end = width - avgpool_params->kernel_size + 1;
99 output_height = ceil((height - avgpool_params->kernel_size + 1) / (kernel_strides * 1.0));
100 output_width = ceil((width - avgpool_params->kernel_size + 1) / (kernel_strides * 1.0));
103 output_operand->dims[0] = number;
104 output_operand->dims[1] = output_height;
105 output_operand->dims[2] = output_width;
106 // not support pooling in channel dimension now
107 output_operand->dims[3] = channel;
108 output_operand->data_type = operands[input_operand_index].data_type;
109 output_operand->length = ff_calculate_operand_data_length(output_operand);
110 if (output_operand->length <= 0) {
111 av_log(ctx, AV_LOG_ERROR, "The output data length overflow\n");
114 output_operand->data = av_realloc(output_operand->data, output_operand->length);
115 if (!output_operand->data) {
116 av_log(ctx, AV_LOG_ERROR, "Failed to reallocate memory for output\n");
119 output = output_operand->data;
121 for (int y = 0; y < height_end; y += kernel_strides) {
122 for (int x = 0; x < width_end; x += kernel_strides) {
123 for (int n_channel = 0; n_channel < channel; ++n_channel) {
124 output[n_channel] = 0.0;
126 for (int kernel_y = 0; kernel_y < avgpool_params->kernel_size; ++kernel_y) {
127 for (int kernel_x = 0; kernel_x < avgpool_params->kernel_size; ++kernel_x) {
129 int y_pos = y + (kernel_y - height_radius);
130 int x_pos = x + (kernel_x - width_radius);
131 if (x_pos < 0 || x_pos >= width || y_pos < 0 || y_pos >= height) {
135 input_pel = input[y_pos * src_linesize + x_pos * channel + n_channel];
137 output[n_channel] += input_pel;
140 output[n_channel] /= kernel_area;