GCC Code Coverage Report
Directory: ../../../ffmpeg/ Exec Total Coverage
File: src/libavfilter/dnn/dnn_backend_native_layer_conv2d.c Lines: 78 144 54.2 %
Date: 2021-04-14 23:45:22 Branches: 39 94 41.5 %

Line Branch Exec Source
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 ThreadCommonParam{
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
} ThreadCommonParam;
37
38
typedef struct ThreadParam{
39
    ThreadCommonParam *thread_common_param;
40
    int thread_start, thread_end;
41
#if HAVE_PTHREAD_CANCEL
42
    pthread_t thread;
43
#endif
44
} ThreadParam;
45
46
int ff_dnn_load_layer_conv2d(Layer *layer, AVIOContext *model_file_context, int file_size, int operands_num)
47
{
48
    ConvolutionalParams *conv_params;
49
    int kernel_size;
50
    int dnn_size = 0;
51
    conv_params = av_malloc(sizeof(*conv_params));
52
    if (!conv_params)
53
        return 0;
54
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);
62
    dnn_size += 28;
63
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;
69
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);
73
        return 0;
74
    }
75
76
    conv_params->kernel = av_malloc_array(kernel_size, sizeof(*conv_params->kernel));
77
    if (!conv_params->kernel) {
78
        av_freep(&conv_params);
79
        return 0;
80
    }
81
    for (int i = 0; i < kernel_size; ++i) {
82
        conv_params->kernel[i] = av_int2float(avio_rl32(model_file_context));
83
    }
84
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);
91
            return 0;
92
        }
93
        for (int i = 0; i < conv_params->output_num; ++i){
94
            conv_params->biases[i] = av_int2float(avio_rl32(model_file_context));
95
        }
96
    }
97
98
    layer->params = conv_params;
99
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);
102
    dnn_size += 8;
103
104
    if (layer->input_operand_indexes[0] >= operands_num || layer->output_operand_index >= operands_num) {
105
        return 0;
106
    }
107
108
    return dnn_size;
109
}
110
111
2
static void * dnn_execute_layer_conv2d_thread(void *threadarg)
112
{
113
    //pass parameters
114
2
    ThreadParam *thread_param = threadarg;
115
2
    ThreadCommonParam *thread_common_param = thread_param->thread_common_param;
116
2
    DnnOperand *operands = thread_common_param->operands;
117
2
    int32_t input_operand_index = thread_common_param->input_operand_indexes[0];
118
2
    int height = operands[input_operand_index].dims[1];
119
2
    int width = operands[input_operand_index].dims[2];
120
2
    int channel = operands[input_operand_index].dims[3];
121
2
    const float *input = operands[input_operand_index].data;
122
2
    const ConvolutionalParams *conv_params = thread_common_param->parameters;
123
124
2
    int radius = conv_params->kernel_size >> 1;
125
2
    int src_linesize = width * conv_params->input_num;
126
2
    int filter_linesize = conv_params->kernel_size * conv_params->input_num;
127
2
    int filter_size = conv_params->kernel_size * filter_linesize;
128
2
    int pad_size = (conv_params->padding_method == VALID) ? (conv_params->kernel_size - 1) / 2 * conv_params->dilation : 0;
129
130
2
    float *output = thread_common_param->output_data;
131
2
    output += (conv_params->output_num) * (width - 2 * pad_size) * (thread_param->thread_start - pad_size);
132
133
2
    av_assert0(channel == conv_params->input_num);
134
135
10
    for (int y = thread_param->thread_start; y < thread_param->thread_end; ++y) {
136
50
        for (int x = pad_size; x < width - pad_size; ++x) {
137
126
            for (int n_filter = 0; n_filter < conv_params->output_num; ++n_filter) {
138
84
                if (conv_params->has_bias)
139
84
                    output[n_filter] = conv_params->biases[n_filter];
140
                else
141
                    output[n_filter] = 0.f;
142
143
336
                for (int ch = 0; ch < conv_params->input_num; ++ch) {
144
1008
                    for (int kernel_y = 0; kernel_y < conv_params->kernel_size; ++kernel_y) {
145
3024
                        for (int kernel_x = 0; kernel_x < conv_params->kernel_size; ++kernel_x) {
146
                            float input_pel;
147
2268
                            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];
151
                            } else {
152
2268
                                int y_pos = y + (kernel_y - radius) * conv_params->dilation;
153
2268
                                int x_pos = x + (kernel_x - radius) * conv_params->dilation;
154


3840
                                input_pel = (x_pos < 0 || x_pos >= width || y_pos < 0 || y_pos >= height) ? 0.0 :
155
1572
                                                   input[y_pos * src_linesize + x_pos * conv_params->input_num + ch];
156
                            }
157
158
159
2268
                            output[n_filter] += input_pel * conv_params->kernel[n_filter * filter_size + kernel_y * filter_linesize +
160
2268
                                                                                kernel_x * conv_params->input_num + ch];
161
                        }
162
                    }
163
                }
164

84
                switch (conv_params->activation){
165
                case RELU:
166
                    output[n_filter] = FFMAX(output[n_filter], 0.0);
167
                    break;
168
84
                case TANH:
169
84
                    output[n_filter] = 2.0f  / (1.0f + exp(-2.0f * output[n_filter])) - 1.0f;
170
84
                    break;
171
                case SIGMOID:
172
                    output[n_filter] = 1.0f / (1.0f + exp(-output[n_filter]));
173
                    break;
174
                case NONE:
175
                    break;
176
                case LEAKY_RELU:
177
                    output[n_filter] = FFMAX(output[n_filter], 0.0) + 0.2 * FFMIN(output[n_filter], 0.0);
178
                }
179
84
            }
180
42
            output += conv_params->output_num;
181
        }
182
    }
183
2
    return NULL;
184
}
185
186
187
2
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)
189
{
190
#if HAVE_PTHREAD_CANCEL
191
2
    int thread_num = (ctx->options.conv2d_threads <= 0 || ctx->options.conv2d_threads > av_cpu_count())
192
4
        ? (av_cpu_count() + 1) : (ctx->options.conv2d_threads);
193
2
    int ret = DNN_SUCCESS, thread_stride;
194
    ThreadParam *thread_param;
195
#else
196
    ThreadParam thread_param = { 0 };
197
#endif
198
    ThreadCommonParam thread_common_param;
199
2
    const ConvolutionalParams *conv_params = parameters;
200
2
    int height = operands[input_operand_indexes[0]].dims[1];
201
2
    int width = operands[input_operand_indexes[0]].dims[2];
202
2
    int pad_size = (conv_params->padding_method == VALID) ? (conv_params->kernel_size - 1) / 2 * conv_params->dilation : 0;
203
2
    DnnOperand *output_operand = &operands[output_operand_index];
204
    void *tmp;
205
206
2
    output_operand->dims[0] = operands[input_operand_indexes[0]].dims[0];
207
2
    output_operand->dims[1] = height - pad_size * 2;
208
2
    output_operand->dims[2] = width - pad_size * 2;
209
2
    output_operand->dims[3] = conv_params->output_num;
210
2
    output_operand->data_type = operands[input_operand_indexes[0]].data_type;
211
2
    output_operand->length = ff_calculate_operand_data_length(output_operand);
212
2
    if (output_operand->length <= 0) {
213
        av_log(ctx, AV_LOG_ERROR, "The output data length overflow\n");
214
        return DNN_ERROR;
215
    }
216
2
    tmp = av_realloc(output_operand->data, output_operand->length);
217
2
    if (!tmp) {
218
        av_log(ctx, AV_LOG_ERROR, "Failed to reallocate memory for output\n");
219
        return DNN_ERROR;
220
    }
221
2
    output_operand->data = tmp;
222
2
    thread_common_param.output_data = output_operand->data;
223
2
    thread_common_param.operands = operands;
224
2
    thread_common_param.input_operand_indexes = input_operand_indexes;
225
2
    thread_common_param.output_operand_index = output_operand_index;
226
2
    thread_common_param.parameters = parameters;
227
2
    thread_common_param.ctx = ctx;
228
229
#if HAVE_PTHREAD_CANCEL
230
2
    thread_param = av_malloc_array(thread_num, sizeof(*thread_param));
231
2
    if (!thread_param)
232
        return DNN_ERROR;
233
2
    thread_stride = (height - pad_size * 2) / thread_num;
234
    //create threads
235
4
    for (int i = 0; i < thread_num; i++){
236
2
        thread_param[i].thread_common_param = &thread_common_param;
237
2
        thread_param[i].thread_start = thread_stride * i + pad_size;
238
2
        thread_param[i].thread_end = (i == thread_num - 1) ? (height - pad_size) : (thread_param[i].thread_start + thread_stride);
239
2
        if (pthread_create(&thread_param[i].thread, NULL,
240
2
                           dnn_execute_layer_conv2d_thread, &thread_param[i])) {
241
            thread_num = i;
242
            ret = DNN_ERROR;
243
            break;
244
        }
245
    }
246
247
4
    for (int i = 0; i < thread_num; i++){
248
2
        pthread_join(thread_param[i].thread, NULL);
249
    }
250
251
    //release memory
252
2
    av_freep(&thread_param);
253
254
2
    return ret;
255
#else
256
    thread_param.thread_common_param = &thread_common_param;
257
    thread_param.thread_start = pad_size;
258
    thread_param.thread_end = height - pad_size;
259
    dnn_execute_layer_conv2d_thread(&thread_param);
260
261
    return DNN_SUCCESS;
262
#endif
263
}