1 |
|
|
/* |
2 |
|
|
* Copyright (c) 2019 Guo Yejun |
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 <string.h> |
22 |
|
|
#include "libavutil/avassert.h" |
23 |
|
|
#include "dnn_backend_native_layer_pad.h" |
24 |
|
|
|
25 |
|
|
int ff_dnn_load_layer_pad(Layer *layer, AVIOContext *model_file_context, int file_size, int operands_num) |
26 |
|
|
{ |
27 |
|
|
LayerPadParams *params; |
28 |
|
|
int dnn_size = 0; |
29 |
|
|
params = av_malloc(sizeof(*params)); |
30 |
|
|
if (!params) |
31 |
|
|
return 0; |
32 |
|
|
|
33 |
|
|
params->mode = (int32_t)avio_rl32(model_file_context); |
34 |
|
|
dnn_size += 4; |
35 |
|
|
for (int i = 0; i < 4; ++i) { |
36 |
|
|
params->paddings[i][0] = avio_rl32(model_file_context); |
37 |
|
|
params->paddings[i][1] = avio_rl32(model_file_context); |
38 |
|
|
dnn_size += 8; |
39 |
|
|
} |
40 |
|
|
layer->input_operand_indexes[0] = (int32_t)avio_rl32(model_file_context); |
41 |
|
|
layer->output_operand_index = (int32_t)avio_rl32(model_file_context); |
42 |
|
|
dnn_size += 8; |
43 |
|
|
layer->params = params; |
44 |
|
|
|
45 |
|
|
if (layer->input_operand_indexes[0] >= operands_num || layer->output_operand_index >= operands_num) { |
46 |
|
|
return 0; |
47 |
|
|
} |
48 |
|
|
|
49 |
|
|
return dnn_size; |
50 |
|
|
} |
51 |
|
|
|
52 |
|
30 |
static int before_get_buddy(int given, int paddings, LayerPadModeParam mode) |
53 |
|
|
{ |
54 |
✓✓ |
30 |
if (mode == LPMP_SYMMETRIC) { |
55 |
|
29 |
return (2 * paddings - 1 - given); |
56 |
✓✗ |
1 |
} else if (mode == LPMP_REFLECT) { |
57 |
|
1 |
return (2 * paddings - given); |
58 |
|
|
} else { |
59 |
|
|
av_assert0(!"should not reach here"); |
60 |
|
|
return 0; |
61 |
|
|
} |
62 |
|
|
} |
63 |
|
|
|
64 |
|
23 |
static int after_get_buddy(int given, int border, LayerPadModeParam mode) |
65 |
|
|
{ |
66 |
✓✓ |
23 |
if (mode == LPMP_SYMMETRIC) { |
67 |
|
21 |
int offset = given - border; |
68 |
|
21 |
return (border - 1 - offset); |
69 |
✓✗ |
2 |
} else if (mode == LPMP_REFLECT) { |
70 |
|
2 |
int offset = given - border; |
71 |
|
2 |
return (border - 2 - offset); |
72 |
|
|
} else { |
73 |
|
|
av_assert0(!"should not reach here"); |
74 |
|
|
return 0; |
75 |
|
|
} |
76 |
|
|
} |
77 |
|
|
|
78 |
|
3 |
int ff_dnn_execute_layer_pad(DnnOperand *operands, const int32_t *input_operand_indexes, |
79 |
|
|
int32_t output_operand_index, const void *parameters, NativeContext *ctx) |
80 |
|
|
{ |
81 |
|
|
int32_t before_paddings; |
82 |
|
|
int32_t after_paddings; |
83 |
|
|
float* output; |
84 |
|
3 |
const LayerPadParams *params = parameters; |
85 |
|
|
|
86 |
|
|
// suppose format is <N, H, W, C> |
87 |
|
3 |
int32_t input_operand_index = input_operand_indexes[0]; |
88 |
|
3 |
int number = operands[input_operand_index].dims[0]; |
89 |
|
3 |
int height = operands[input_operand_index].dims[1]; |
90 |
|
3 |
int width = operands[input_operand_index].dims[2]; |
91 |
|
3 |
int channel = operands[input_operand_index].dims[3]; |
92 |
|
3 |
const float *input = operands[input_operand_index].data; |
93 |
|
|
|
94 |
|
3 |
int new_number = number + params->paddings[0][0] + params->paddings[0][1]; |
95 |
|
3 |
int new_height = height + params->paddings[1][0] + params->paddings[1][1]; |
96 |
|
3 |
int new_width = width + params->paddings[2][0] + params->paddings[2][1]; |
97 |
|
3 |
int new_channel = channel + params->paddings[3][0] + params->paddings[3][1]; |
98 |
|
|
|
99 |
|
3 |
int c_stride = channel; |
100 |
|
3 |
int wc_stride = c_stride * width; |
101 |
|
3 |
int hwc_stride = wc_stride * height; |
102 |
|
|
|
103 |
|
3 |
int new_c_stride = new_channel; |
104 |
|
3 |
int new_wc_stride = new_c_stride * new_width; |
105 |
|
3 |
int new_hwc_stride = new_wc_stride * new_height; |
106 |
|
|
|
107 |
|
3 |
DnnOperand *output_operand = &operands[output_operand_index]; |
108 |
|
3 |
output_operand->dims[0] = new_number; |
109 |
|
3 |
output_operand->dims[1] = new_height; |
110 |
|
3 |
output_operand->dims[2] = new_width; |
111 |
|
3 |
output_operand->dims[3] = new_channel; |
112 |
|
3 |
output_operand->data_type = operands[input_operand_index].data_type; |
113 |
|
3 |
output_operand->length = ff_calculate_operand_data_length(output_operand); |
114 |
✗✓ |
3 |
if (output_operand->length <= 0) { |
115 |
|
|
av_log(ctx, AV_LOG_ERROR, "The output data length overflow\n"); |
116 |
|
|
return DNN_ERROR; |
117 |
|
|
} |
118 |
|
3 |
output_operand->data = av_realloc(output_operand->data, output_operand->length); |
119 |
✗✓ |
3 |
if (!output_operand->data) { |
120 |
|
|
av_log(ctx, AV_LOG_ERROR, "Failed to reallocate memory for output\n"); |
121 |
|
|
return DNN_ERROR; |
122 |
|
|
} |
123 |
|
3 |
output = output_operand->data; |
124 |
|
|
|
125 |
|
|
// copy the original data |
126 |
✓✓ |
8 |
for (int n = 0; n < number; n++) { |
127 |
✓✓ |
17 |
for (int h = 0; h < height; h++) { |
128 |
✓✓ |
44 |
for (int w = 0; w < width; w++) { |
129 |
|
32 |
const float *src = input + n * hwc_stride + h * wc_stride + w * c_stride; |
130 |
|
32 |
float *dst = output + (n + params->paddings[0][0]) * new_hwc_stride |
131 |
|
32 |
+ (h + params->paddings[1][0]) * new_wc_stride |
132 |
|
32 |
+ (w + params->paddings[2][0]) * new_c_stride |
133 |
|
32 |
+ params->paddings[3][0]; |
134 |
|
32 |
memcpy(dst, src, channel * sizeof(float)); |
135 |
|
|
} |
136 |
|
|
} |
137 |
|
|
} |
138 |
|
|
|
139 |
|
|
// handle the first dimension |
140 |
|
3 |
before_paddings = params->paddings[0][0]; |
141 |
|
3 |
after_paddings = params->paddings[0][1]; |
142 |
✓✓ |
4 |
for (int n = 0; n < before_paddings; n++) { |
143 |
|
1 |
float *dst = output + n * new_hwc_stride; |
144 |
✗✓ |
1 |
if (params->mode == LPMP_CONSTANT) { |
145 |
|
|
for (int i = 0; i < new_hwc_stride; i++) { |
146 |
|
|
dst[i] = params->constant_values; |
147 |
|
|
} |
148 |
|
|
} |
149 |
|
|
else { |
150 |
|
1 |
int buddy = before_get_buddy(n, before_paddings, params->mode); |
151 |
|
1 |
float *src = output + buddy * new_hwc_stride; |
152 |
|
1 |
memcpy(dst, src, new_hwc_stride * sizeof(float)); |
153 |
|
|
} |
154 |
|
|
} |
155 |
✓✓ |
5 |
for (int n = 0; n < after_paddings; n++) { |
156 |
|
2 |
int given = number + before_paddings + n; |
157 |
|
2 |
float *dst = output + given * new_hwc_stride; |
158 |
✗✓ |
2 |
if (params->mode == LPMP_CONSTANT) { |
159 |
|
|
for (int i = 0; i < new_hwc_stride; i++) { |
160 |
|
|
dst[i] = params->constant_values; |
161 |
|
|
} |
162 |
|
|
} else { |
163 |
|
2 |
int buddy = after_get_buddy(given, number + before_paddings, params->mode); |
164 |
|
2 |
float *src = output + buddy * new_hwc_stride; |
165 |
|
2 |
memcpy(dst, src, new_hwc_stride * sizeof(float)); |
166 |
|
|
} |
167 |
|
|
} |
168 |
|
|
|
169 |
|
|
// handle the second dimension |
170 |
|
3 |
before_paddings = params->paddings[1][0]; |
171 |
|
3 |
after_paddings = params->paddings[1][1]; |
172 |
✓✓ |
11 |
for (int n = 0; n < new_number; n++) { |
173 |
|
8 |
float *start = output + n * new_hwc_stride; |
174 |
✓✓ |
11 |
for (int h = 0; h < before_paddings; h++) { |
175 |
|
3 |
float *dst = start + h * new_wc_stride; |
176 |
✓✓ |
3 |
if (params->mode == LPMP_CONSTANT) { |
177 |
✓✓ |
13 |
for (int i = 0; i < new_wc_stride; i++) { |
178 |
|
12 |
dst[i] = params->constant_values; |
179 |
|
|
} |
180 |
|
|
} else { |
181 |
|
2 |
int buddy = before_get_buddy(h, before_paddings, params->mode); |
182 |
|
2 |
float *src = start + buddy * new_wc_stride; |
183 |
|
2 |
memcpy(dst, src, new_wc_stride * sizeof(float)); |
184 |
|
|
} |
185 |
|
|
} |
186 |
✓✓ |
11 |
for (int h = 0; h < after_paddings; h++) { |
187 |
|
3 |
int given = height + before_paddings + h; |
188 |
|
3 |
float *dst = start + given * new_wc_stride; |
189 |
✗✓ |
3 |
if (params->mode == LPMP_CONSTANT) { |
190 |
|
|
for (int i = 0; i < new_wc_stride; i++) { |
191 |
|
|
dst[i] = params->constant_values; |
192 |
|
|
} |
193 |
|
|
} else { |
194 |
|
3 |
int buddy = after_get_buddy(given, height + before_paddings, params->mode); |
195 |
|
3 |
float *src = start + buddy * new_wc_stride; |
196 |
|
3 |
memcpy(dst, src, new_wc_stride * sizeof(float)); |
197 |
|
|
} |
198 |
|
|
} |
199 |
|
|
} |
200 |
|
|
|
201 |
|
|
// handle the third dimension |
202 |
|
3 |
before_paddings = params->paddings[2][0]; |
203 |
|
3 |
after_paddings = params->paddings[2][1]; |
204 |
✓✓ |
11 |
for (int n = 0; n < new_number; n++) { |
205 |
✓✓ |
32 |
for (int h = 0; h < new_height; h++) { |
206 |
|
24 |
float *start = output + n * new_hwc_stride + h * new_wc_stride; |
207 |
✓✓ |
51 |
for (int w = 0; w < before_paddings; w++) { |
208 |
|
27 |
float *dst = start + w * new_c_stride; |
209 |
✗✓ |
27 |
if (params->mode == LPMP_CONSTANT) { |
210 |
|
|
for (int i = 0; i < new_c_stride; i++) { |
211 |
|
|
dst[i] = params->constant_values; |
212 |
|
|
} |
213 |
|
|
} else { |
214 |
|
27 |
int buddy = before_get_buddy(w, before_paddings, params->mode); |
215 |
|
27 |
float *src = start + buddy * new_c_stride; |
216 |
|
27 |
memcpy(dst, src, new_c_stride * sizeof(float)); |
217 |
|
|
} |
218 |
|
|
} |
219 |
✓✓ |
42 |
for (int w = 0; w < after_paddings; w++) { |
220 |
|
18 |
int given = width + before_paddings + w; |
221 |
|
18 |
float *dst = start + given * new_c_stride; |
222 |
✗✓ |
18 |
if (params->mode == LPMP_CONSTANT) { |
223 |
|
|
for (int i = 0; i < new_c_stride; i++) { |
224 |
|
|
dst[i] = params->constant_values; |
225 |
|
|
} |
226 |
|
|
} else { |
227 |
|
18 |
int buddy = after_get_buddy(given, width + before_paddings, params->mode); |
228 |
|
18 |
float *src = start + buddy * new_c_stride; |
229 |
|
18 |
memcpy(dst, src, new_c_stride * sizeof(float)); |
230 |
|
|
} |
231 |
|
|
} |
232 |
|
|
} |
233 |
|
|
} |
234 |
|
|
|
235 |
|
|
// handle the fourth dimension |
236 |
|
3 |
before_paddings = params->paddings[3][0]; |
237 |
|
3 |
after_paddings = params->paddings[3][1]; |
238 |
✓✓ |
11 |
for (int n = 0; n < new_number; n++) { |
239 |
✓✓ |
32 |
for (int h = 0; h < new_height; h++) { |
240 |
✓✓ |
135 |
for (int w = 0; w < new_width; w++) { |
241 |
|
111 |
float *start = output + n * new_hwc_stride + h * new_wc_stride + w * new_c_stride; |
242 |
✓✓ |
117 |
for (int c = 0; c < before_paddings; c++) { |
243 |
|
6 |
float *dst = start + c; |
244 |
✓✗ |
6 |
if (params->mode == LPMP_CONSTANT) { |
245 |
|
6 |
*dst = params->constant_values; |
246 |
|
|
} else { |
247 |
|
|
int buddy = before_get_buddy(c, before_paddings, params->mode); |
248 |
|
|
float *src = start + buddy; |
249 |
|
|
*dst = *src; |
250 |
|
|
} |
251 |
|
|
} |
252 |
✓✓ |
123 |
for (int c = 0; c < after_paddings; c++) { |
253 |
|
12 |
int given = channel + before_paddings + c; |
254 |
|
12 |
float *dst = start + given; |
255 |
✓✗ |
12 |
if (params->mode == LPMP_CONSTANT) { |
256 |
|
12 |
*dst = params->constant_values; |
257 |
|
|
} else { |
258 |
|
|
int buddy = after_get_buddy(given, channel + before_paddings, params->mode); |
259 |
|
|
float *src = start + buddy; |
260 |
|
|
*dst = *src; |
261 |
|
|
} |
262 |
|
|
} |
263 |
|
|
} |
264 |
|
|
} |
265 |
|
|
} |
266 |
|
|
|
267 |
|
3 |
return 0; |
268 |
|
|
} |