the reason to add this layer first is that vf_sr uses it in its tensorflow model, and the next plan is to update the python script to convert tf.pad into native model.
Showing the changes to libavfilter (3 of 3 files). Show all
Unified · Split
| @@ -1,5 +1,6 @@ | |||
| 1 | OBJS-$(CONFIG_DNN) += dnn/dnn_interface.o | 1 | OBJS-$(CONFIG_DNN) += dnn/dnn_interface.o |
| 2 | OBJS-$(CONFIG_DNN) += dnn/dnn_backend_native.o | 2 | OBJS-$(CONFIG_DNN) += dnn/dnn_backend_native.o |
| 3 | OBJS-$(CONFIG_DNN) += dnn/dnn_backend_native_layer_pad.o | ||
| 3 | 4 | ||
| 4 | DNN-OBJS-$(CONFIG_LIBTENSORFLOW) += dnn/dnn_backend_tf.o | 5 | DNN-OBJS-$(CONFIG_LIBTENSORFLOW) += dnn/dnn_backend_tf.o |
| 5 | 6 | ||
| @@ -0,0 +1,211 @@ | |||
| 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 | static int before_get_buddy(int given, int paddings, LayerPadModeParam mode) | ||
| 26 | { | ||
| 27 | if (mode == LPMP_SYMMETRIC) { | ||
| 28 | return (2 * paddings - 1 - given); | ||
| 29 | } else if (mode == LPMP_REFLECT) { | ||
| 30 | return (2 * paddings - given); | ||
| 31 | } else { | ||
| 32 | av_assert0(!"should not reach here"); | ||
| 33 | return 0; | ||
| 34 | } | ||
| 35 | } | ||
| 36 | |||
| 37 | static int after_get_buddy(int given, int border, LayerPadModeParam mode) | ||
| 38 | { | ||
| 39 | if (mode == LPMP_SYMMETRIC) { | ||
| 40 | int offset = given - border; | ||
| 41 | return (border - 1 - offset); | ||
| 42 | } else if (mode == LPMP_REFLECT) { | ||
| 43 | int offset = given - border; | ||
| 44 | return (border - 2 - offset); | ||
| 45 | } else { | ||
| 46 | av_assert0(!"should not reach here"); | ||
| 47 | return 0; | ||
| 48 | } | ||
| 49 | } | ||
| 50 | |||
| 51 | void dnn_execute_layer_pad(const float *input, float *output, const LayerPadParams *params, int number, int height, int width, int channel) | ||
| 52 | { | ||
| 53 | int32_t before_paddings; | ||
| 54 | int32_t after_paddings; | ||
| 55 | |||
| 56 | // suppose format is <N, H, W, C> | ||
| 57 | int new_number = number + params->paddings[0][0] + params->paddings[0][1]; | ||
| 58 | int new_height = height + params->paddings[1][0] + params->paddings[1][1]; | ||
| 59 | int new_width = width + params->paddings[2][0] + params->paddings[2][1]; | ||
| 60 | int new_channel = channel + params->paddings[3][0] + params->paddings[3][1]; | ||
| 61 | |||
| 62 | int c_stride = channel; | ||
| 63 | int wc_stride = c_stride * width; | ||
| 64 | int hwc_stride = wc_stride * height; | ||
| 65 | |||
| 66 | int new_c_stride = new_channel; | ||
| 67 | int new_wc_stride = new_c_stride * new_width; | ||
| 68 | int new_hwc_stride = new_wc_stride * new_height; | ||
| 69 | |||
| 70 | // copy the original data | ||
| 71 | for (int n = 0; n < number; n++) { | ||
| 72 | for (int h = 0; h < height; h++) { | ||
| 73 | for (int w = 0; w < width; w++) { | ||
| 74 | const float *src = input + n * hwc_stride + h * wc_stride + w * c_stride; | ||
| 75 | float *dst = output + (n + params->paddings[0][0]) * new_hwc_stride | ||
| 76 | + (h + params->paddings[1][0]) * new_wc_stride | ||
| 77 | + (w + params->paddings[2][0]) * new_c_stride | ||
| 78 | + params->paddings[3][0]; | ||
| 79 | memcpy(dst, src, channel * sizeof(float)); | ||
| 80 | } | ||
| 81 | } | ||
| 82 | } | ||
| 83 | |||
| 84 | // handle the first dimension | ||
| 85 | before_paddings = params->paddings[0][0]; | ||
| 86 | after_paddings = params->paddings[0][1]; | ||
| 87 | for (int n = 0; n < before_paddings; n++) { | ||
| 88 | float *dst = output + n * new_hwc_stride; | ||
| 89 | if (params->mode == LPMP_CONSTANT) { | ||
| 90 | for (int i = 0; i < new_hwc_stride; i++) { | ||
| 91 | dst[i] = params->constant_values; | ||
| 92 | } | ||
| 93 | } | ||
| 94 | else { | ||
| 95 | int buddy = before_get_buddy(n, before_paddings, params->mode); | ||
| 96 | float *src = output + buddy * new_hwc_stride; | ||
| 97 | memcpy(dst, src, new_hwc_stride * sizeof(float)); | ||
| 98 | } | ||
| 99 | } | ||
| 100 | for (int n = 0; n < after_paddings; n++) { | ||
| 101 | int given = number + before_paddings + n; | ||
| 102 | float *dst = output + given * new_hwc_stride; | ||
| 103 | if (params->mode == LPMP_CONSTANT) { | ||
| 104 | for (int i = 0; i < new_hwc_stride; i++) { | ||
| 105 | dst[i] = params->constant_values; | ||
| 106 | } | ||
| 107 | } else { | ||
| 108 | int buddy = after_get_buddy(given, number + before_paddings, params->mode); | ||
| 109 | float *src = output + buddy * new_hwc_stride; | ||
| 110 | memcpy(dst, src, new_hwc_stride * sizeof(float)); | ||
| 111 | } | ||
| 112 | } | ||
| 113 | |||
| 114 | // handle the second dimension | ||
| 115 | before_paddings = params->paddings[1][0]; | ||
| 116 | after_paddings = params->paddings[1][1]; | ||
| 117 | for (int n = 0; n < new_number; n++) { | ||
| 118 | float *start = output + n * new_hwc_stride; | ||
| 119 | for (int h = 0; h < before_paddings; h++) { | ||
| 120 | float *dst = start + h * new_wc_stride; | ||
| 121 | if (params->mode == LPMP_CONSTANT) { | ||
| 122 | for (int i = 0; i < new_wc_stride; i++) { | ||
| 123 | dst[i] = params->constant_values; | ||
| 124 | } | ||
| 125 | } else { | ||
| 126 | int buddy = before_get_buddy(h, before_paddings, params->mode); | ||
| 127 | float *src = start + buddy * new_wc_stride; | ||
| 128 | memcpy(dst, src, new_wc_stride * sizeof(float)); | ||
| 129 | } | ||
| 130 | } | ||
| 131 | for (int h = 0; h < after_paddings; h++) { | ||
| 132 | int given = height + before_paddings + h; | ||
| 133 | float *dst = start + given * new_wc_stride; | ||
| 134 | if (params->mode == LPMP_CONSTANT) { | ||
| 135 | for (int i = 0; i < new_wc_stride; i++) { | ||
| 136 | dst[i] = params->constant_values; | ||
| 137 | } | ||
| 138 | } else { | ||
| 139 | int buddy = after_get_buddy(given, height + before_paddings, params->mode); | ||
| 140 | float *src = start + buddy * new_wc_stride; | ||
| 141 | memcpy(dst, src, new_wc_stride * sizeof(float)); | ||
| 142 | } | ||
| 143 | } | ||
| 144 | } | ||
| 145 | |||
| 146 | // handle the third dimension | ||
| 147 | before_paddings = params->paddings[2][0]; | ||
| 148 | after_paddings = params->paddings[2][1]; | ||
| 149 | for (int n = 0; n < new_number; n++) { | ||
| 150 | for (int h = 0; h < new_height; h++) { | ||
| 151 | float *start = output + n * new_hwc_stride + h * new_wc_stride; | ||
| 152 | for (int w = 0; w < before_paddings; w++) { | ||
| 153 | float *dst = start + w * new_c_stride; | ||
| 154 | if (params->mode == LPMP_CONSTANT) { | ||
| 155 | for (int i = 0; i < new_c_stride; i++) { | ||
| 156 | dst[i] = params->constant_values; | ||
| 157 | } | ||
| 158 | } else { | ||
| 159 | int buddy = before_get_buddy(w, before_paddings, params->mode); | ||
| 160 | float *src = start + buddy * new_c_stride; | ||
| 161 | memcpy(dst, src, new_c_stride * sizeof(float)); | ||
| 162 | } | ||
| 163 | } | ||
| 164 | for (int w = 0; w < after_paddings; w++) { | ||
| 165 | int given = width + before_paddings + w; | ||
| 166 | float *dst = start + given * new_c_stride; | ||
| 167 | if (params->mode == LPMP_CONSTANT) { | ||
| 168 | for (int i = 0; i < new_c_stride; i++) { | ||
| 169 | dst[i] = params->constant_values; | ||
| 170 | } | ||
| 171 | } else { | ||
| 172 | int buddy = after_get_buddy(given, width + before_paddings, params->mode); | ||
| 173 | float *src = start + buddy * new_c_stride; | ||
| 174 | memcpy(dst, src, new_c_stride * sizeof(float)); | ||
| 175 | } | ||
| 176 | } | ||
| 177 | } | ||
| 178 | } | ||
| 179 | |||
| 180 | // handle the fourth dimension | ||
| 181 | before_paddings = params->paddings[3][0]; | ||
| 182 | after_paddings = params->paddings[3][1]; | ||
| 183 | for (int n = 0; n < new_number; n++) { | ||
| 184 | for (int h = 0; h < new_height; h++) { | ||
| 185 | for (int w = 0; w < new_width; w++) { | ||
| 186 | float *start = output + n * new_hwc_stride + h * new_wc_stride + w * new_c_stride; | ||
| 187 | for (int c = 0; c < before_paddings; c++) { | ||
| 188 | float *dst = start + c; | ||
| 189 | if (params->mode == LPMP_CONSTANT) { | ||
| 190 | *dst = params->constant_values; | ||
| 191 | } else { | ||
| 192 | int buddy = before_get_buddy(c, before_paddings, params->mode); | ||
| 193 | float *src = start + buddy; | ||
| 194 | *dst = *src; | ||
| 195 | } | ||
| 196 | } | ||
| 197 | for (int c = 0; c < after_paddings; c++) { | ||
| 198 | int given = channel + before_paddings + c; | ||
| 199 | float *dst = start + given; | ||
| 200 | if (params->mode == LPMP_CONSTANT) { | ||
| 201 | *dst = params->constant_values; | ||
| 202 | } else { | ||
| 203 | int buddy = after_get_buddy(given, channel + before_paddings, params->mode); | ||
| 204 | float *src = start + buddy; | ||
| 205 | *dst = *src; | ||
| 206 | } | ||
| 207 | } | ||
| 208 | } | ||
| 209 | } | ||
| 210 | } | ||
| 211 | } | ||
| @@ -0,0 +1,40 @@ | |||
| 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 | /** | ||
| 22 | * @file | ||
| 23 | * layer pad (equivalent to tf.pad) for native backend. | ||
| 24 | */ | ||
| 25 | #ifndef AVFILTER_DNN_DNN_BACKEND_NATIVE_LAYER_PAD_H | ||
| 26 | #define AVFILTER_DNN_DNN_BACKEND_NATIVE_LAYER_PAD_H | ||
| 27 | |||
| 28 | #include <stdint.h> | ||
| 29 | |||
| 30 | typedef enum {LPMP_CONSTANT, LPMP_REFLECT, LPMP_SYMMETRIC} LayerPadModeParam; | ||
| 31 | |||
| 32 | typedef struct LayerPadParams{ | ||
| 33 | int32_t paddings[4][2]; | ||
| 34 | LayerPadModeParam mode; | ||
| 35 | float constant_values; | ||
| 36 | } LayerPadParams; | ||
| 37 | |||
| 38 | void dnn_execute_layer_pad(const float *input, float *output, const LayerPadParams *params, int number, int height, int width, int channel); | ||
| 39 | |||
| 40 | #endif | ||