| File: | root/firefox-clang/third_party/aom/av1/encoder/partition_strategy.c |
| Warning: | line 2377, column 3 Null pointer passed to 1st parameter expecting 'nonnull' |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
| 1 | /* | |||
| 2 | * Copyright (c) 2019, Alliance for Open Media. All rights reserved. | |||
| 3 | * | |||
| 4 | * This source code is subject to the terms of the BSD 2 Clause License and | |||
| 5 | * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License | |||
| 6 | * was not distributed with this source code in the LICENSE file, you can | |||
| 7 | * obtain it at www.aomedia.org/license/software. If the Alliance for Open | |||
| 8 | * Media Patent License 1.0 was not distributed with this source code in the | |||
| 9 | * PATENTS file, you can obtain it at www.aomedia.org/license/patent. | |||
| 10 | */ | |||
| 11 | ||||
| 12 | #include <float.h> | |||
| 13 | ||||
| 14 | #include "config/aom_config.h" | |||
| 15 | ||||
| 16 | #include "av1/encoder/encodeframe_utils.h" | |||
| 17 | #if CONFIG_THREE_PASS0 | |||
| 18 | #include "av1/encoder/thirdpass.h" | |||
| 19 | #endif | |||
| 20 | #include "config/aom_dsp_rtcd.h" | |||
| 21 | ||||
| 22 | #include "av1/common/enums.h" | |||
| 23 | #include "av1/common/reconinter.h" | |||
| 24 | ||||
| 25 | #if !CONFIG_REALTIME_ONLY0 | |||
| 26 | #include "av1/encoder/cnn.h" | |||
| 27 | #include "av1/encoder/partition_model_weights.h" | |||
| 28 | #include "av1/encoder/partition_cnn_weights.h" | |||
| 29 | #endif | |||
| 30 | #include "av1/encoder/encoder.h" | |||
| 31 | ||||
| 32 | #include "av1/encoder/motion_search_facade.h" | |||
| 33 | #include "av1/encoder/partition_strategy.h" | |||
| 34 | #include "av1/encoder/partition_search.h" | |||
| 35 | #include "av1/encoder/rdopt.h" | |||
| 36 | ||||
| 37 | #if !CONFIG_REALTIME_ONLY0 | |||
| 38 | static inline void simple_motion_search_prune_part_features( | |||
| 39 | AV1_COMP *const cpi, MACROBLOCK *x, SIMPLE_MOTION_DATA_TREE *sms_tree, | |||
| 40 | int mi_row, int mi_col, BLOCK_SIZE bsize, float *features, | |||
| 41 | int features_to_get); | |||
| 42 | ||||
| 43 | static bool_Bool ext_ml_model_decision_before_none( | |||
| 44 | AV1_COMP *cpi, const float features_from_motion[FEATURE_SIZE_SMS_SPLIT17], | |||
| 45 | int *partition_none_allowed, int *partition_horz_allowed, | |||
| 46 | int *partition_vert_allowed, int *do_rectangular_split, | |||
| 47 | int *do_square_split); | |||
| 48 | ||||
| 49 | static bool_Bool ext_ml_model_decision_before_none_part2( | |||
| 50 | AV1_COMP *cpi, | |||
| 51 | const float features_from_motion[FEATURE_SIZE_SMS_PRUNE_PART25], | |||
| 52 | int *prune_horz, int *prune_vert); | |||
| 53 | ||||
| 54 | static bool_Bool ext_ml_model_decision_after_none( | |||
| 55 | ExtPartController *const ext_part_controller, const int is_intra_frame, | |||
| 56 | const float *const features_after_none, int *do_square_split, | |||
| 57 | int *do_rectangular_split); | |||
| 58 | ||||
| 59 | static bool_Bool ext_ml_model_decision_after_none_part2( | |||
| 60 | AV1_COMP *const cpi, const float *const features_terminate, | |||
| 61 | int *terminate_partition_search); | |||
| 62 | ||||
| 63 | static bool_Bool ext_ml_model_decision_after_split( | |||
| 64 | AV1_COMP *const cpi, const float *const features_terminate, | |||
| 65 | int *terminate_partition_search); | |||
| 66 | ||||
| 67 | static bool_Bool ext_ml_model_decision_after_split_part2( | |||
| 68 | ExtPartController *const ext_part_controller, const int is_intra_frame, | |||
| 69 | const float *const features_prune, int *prune_rect_part_horz, | |||
| 70 | int *prune_rect_part_vert); | |||
| 71 | ||||
| 72 | static bool_Bool ext_ml_model_decision_after_rect( | |||
| 73 | ExtPartController *const ext_part_controller, const int is_intra_frame, | |||
| 74 | const float *const features_after_rect, int *horza_partition_allowed, | |||
| 75 | int *horzb_partition_allowed, int *verta_partition_allowed, | |||
| 76 | int *vertb_partition_allowed); | |||
| 77 | ||||
| 78 | static bool_Bool ext_ml_model_decision_after_part_ab( | |||
| 79 | AV1_COMP *const cpi, MACROBLOCK *const x, BLOCK_SIZE bsize, int part_ctx, | |||
| 80 | int64_t best_rd, int64_t rect_part_rd[NUM_RECT_PARTS][SUB_PARTITIONS_RECT2], | |||
| 81 | int64_t split_rd[SUB_PARTITIONS_SPLIT4], int *const partition_horz4_allowed, | |||
| 82 | int *const partition_vert4_allowed, unsigned int pb_source_variance, | |||
| 83 | int mi_row, int mi_col); | |||
| 84 | ||||
| 85 | static inline int convert_bsize_to_idx(BLOCK_SIZE bsize) { | |||
| 86 | switch (bsize) { | |||
| 87 | case BLOCK_128X128: return 0; | |||
| 88 | case BLOCK_64X64: return 1; | |||
| 89 | case BLOCK_32X32: return 2; | |||
| 90 | case BLOCK_16X16: return 3; | |||
| 91 | case BLOCK_8X8: return 4; | |||
| 92 | default: assert(0 && "Invalid bsize")((void) sizeof ((0 && "Invalid bsize") ? 1 : 0), __extension__ ({ if (0 && "Invalid bsize") ; else __assert_fail ("0 && \"Invalid bsize\"" , "/root/firefox-clang/third_party/aom/av1/encoder/partition_strategy.c" , 92, __extension__ __PRETTY_FUNCTION__); })); return -1; | |||
| 93 | } | |||
| 94 | } | |||
| 95 | ||||
| 96 | static char *get_feature_file_name(int id) { | |||
| 97 | static char *feature_file_names[] = { | |||
| 98 | "feature_before_partition_none", | |||
| 99 | "feature_before_partition_none_prune_rect", | |||
| 100 | "feature_after_partition_none_prune", | |||
| 101 | "feature_after_partition_none_terminate", | |||
| 102 | "feature_after_partition_split_terminate", | |||
| 103 | "feature_after_partition_split_prune_rect", | |||
| 104 | "feature_after_partition_rect", | |||
| 105 | "feature_after_partition_ab", | |||
| 106 | }; | |||
| 107 | ||||
| 108 | return feature_file_names[id]; | |||
| 109 | } | |||
| 110 | ||||
| 111 | static void write_features_to_file(const char *const path, | |||
| 112 | const bool_Bool is_test_mode, | |||
| 113 | const float *features, | |||
| 114 | const int feature_size, const int id, | |||
| 115 | const BLOCK_SIZE bsize, const int mi_row, | |||
| 116 | const int mi_col) { | |||
| 117 | if (!WRITE_FEATURE_TO_FILE0 && !is_test_mode) return; | |||
| 118 | ||||
| 119 | char filename[256]; | |||
| 120 | snprintf(filename, sizeof(filename), "%s/%s", path, | |||
| 121 | get_feature_file_name(id)); | |||
| 122 | FILE *pfile = fopen(filename, "a"); | |||
| 123 | if (pfile == NULL((void*)0)) return; | |||
| 124 | if (!is_test_mode) { | |||
| 125 | fprintf(pfile, "%d,%d,%d,%d,%d\n", id, (int)bsize, mi_row, mi_col, | |||
| 126 | feature_size); | |||
| 127 | } | |||
| 128 | for (int i = 0; i < feature_size; ++i) { | |||
| 129 | fprintf(pfile, "%.6f", features[i]); | |||
| 130 | if (i < feature_size - 1) fprintf(pfile, ","); | |||
| 131 | } | |||
| 132 | fprintf(pfile, "\n"); | |||
| 133 | fclose(pfile); | |||
| 134 | } | |||
| 135 | ||||
| 136 | // TODO(chiyotsai@google.com): This is very much a work in progress. We still | |||
| 137 | // need to the following: | |||
| 138 | // -- add support for hdres | |||
| 139 | // -- add support for pruning rectangular partitions | |||
| 140 | // -- use reconstructed pixels instead of source pixels for padding | |||
| 141 | // -- use chroma pixels in addition to luma pixels | |||
| 142 | static void intra_mode_cnn_partition(const AV1_COMMON *const cm, MACROBLOCK *x, | |||
| 143 | int quad_tree_idx, | |||
| 144 | int intra_cnn_based_part_prune_level, | |||
| 145 | PartitionSearchState *part_state) { | |||
| 146 | assert(cm->seq_params->sb_size >= BLOCK_64X64 &&((void) sizeof ((cm->seq_params->sb_size >= BLOCK_64X64 && "Invalid sb_size for intra_cnn!") ? 1 : 0), __extension__ ({ if (cm->seq_params->sb_size >= BLOCK_64X64 && "Invalid sb_size for intra_cnn!") ; else __assert_fail ("cm->seq_params->sb_size >= BLOCK_64X64 && \"Invalid sb_size for intra_cnn!\"" , "/root/firefox-clang/third_party/aom/av1/encoder/partition_strategy.c" , 147, __extension__ __PRETTY_FUNCTION__); })) | |||
| 147 | "Invalid sb_size for intra_cnn!")((void) sizeof ((cm->seq_params->sb_size >= BLOCK_64X64 && "Invalid sb_size for intra_cnn!") ? 1 : 0), __extension__ ({ if (cm->seq_params->sb_size >= BLOCK_64X64 && "Invalid sb_size for intra_cnn!") ; else __assert_fail ("cm->seq_params->sb_size >= BLOCK_64X64 && \"Invalid sb_size for intra_cnn!\"" , "/root/firefox-clang/third_party/aom/av1/encoder/partition_strategy.c" , 147, __extension__ __PRETTY_FUNCTION__); })); | |||
| 148 | const PartitionBlkParams *blk_params = &part_state->part_blk_params; | |||
| 149 | const BLOCK_SIZE bsize = blk_params->bsize; | |||
| 150 | ||||
| 151 | const int bsize_idx = convert_bsize_to_idx(bsize); | |||
| 152 | ||||
| 153 | if (bsize == BLOCK_128X128) { | |||
| 154 | return; | |||
| 155 | } | |||
| 156 | ||||
| 157 | PartitionSearchInfo *part_info = &x->part_search_info; | |||
| 158 | ||||
| 159 | // Precompute the CNN part and cache the result in MACROBLOCK | |||
| 160 | if (bsize == BLOCK_64X64 && !part_info->cnn_output_valid) { | |||
| 161 | const CNN_CONFIG *cnn_config = &av1_intra_mode_cnn_partition_cnn_config; | |||
| 162 | ||||
| 163 | // Prepare the output | |||
| 164 | const CNN_THREAD_DATA thread_data = { .num_workers = 1, .workers = NULL((void*)0) }; | |||
| 165 | const int num_outputs = 4; | |||
| 166 | const int output_dims[4] = { 1, 2, 4, 8 }; | |||
| 167 | const int out_chs[4] = { CNN_BRANCH_0_OUT_CH20, CNN_BRANCH_1_OUT_CH4, | |||
| 168 | CNN_BRANCH_2_OUT_CH20, CNN_BRANCH_3_OUT_CH20 }; | |||
| 169 | float *output_buffer[CNN_TOT_OUT_CH(((20) + (4) + (20) + (20)))]; | |||
| 170 | ||||
| 171 | float **cur_output_buf = output_buffer; | |||
| 172 | float *curr_buf_ptr = part_info->cnn_buffer; | |||
| 173 | for (int output_idx = 0; output_idx < num_outputs; output_idx++) { | |||
| 174 | const int num_chs = out_chs[output_idx]; | |||
| 175 | const int ch_size = output_dims[output_idx] * output_dims[output_idx]; | |||
| 176 | for (int ch = 0; ch < num_chs; ch++) { | |||
| 177 | cur_output_buf[ch] = curr_buf_ptr; | |||
| 178 | curr_buf_ptr += ch_size; | |||
| 179 | } | |||
| 180 | cur_output_buf += num_chs; | |||
| 181 | } | |||
| 182 | ||||
| 183 | CNN_MULTI_OUT output = { | |||
| 184 | .num_outputs = 4, | |||
| 185 | .output_channels = out_chs, | |||
| 186 | .output_strides = output_dims, | |||
| 187 | .output_buffer = output_buffer, | |||
| 188 | }; | |||
| 189 | ||||
| 190 | // Prepare the input | |||
| 191 | const MACROBLOCKD *xd = &x->e_mbd; | |||
| 192 | const int bit_depth = xd->bd; | |||
| 193 | const int dc_q = | |||
| 194 | av1_dc_quant_QTX(x->qindex, 0, bit_depth) >> (bit_depth - 8); | |||
| 195 | part_info->log_q = log1pf((float)(dc_q * dc_q) / 256.0f); | |||
| 196 | part_info->log_q = | |||
| 197 | (part_info->log_q - av1_intra_mode_cnn_partition_mean[0]) / | |||
| 198 | av1_intra_mode_cnn_partition_std[0]; | |||
| 199 | ||||
| 200 | const int width = 65, height = 65, | |||
| 201 | stride = x->plane[AOM_PLANE_Y0].src.stride; | |||
| 202 | ||||
| 203 | if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH8) { | |||
| 204 | uint16_t *image[1] = { | |||
| 205 | CONVERT_TO_SHORTPTR(x->plane[AOM_PLANE_Y].src.buf)((uint16_t *)(((uintptr_t)(x->plane[0].src.buf)) << 1 )) - stride - 1 | |||
| 206 | }; | |||
| 207 | ||||
| 208 | if (!av1_cnn_predict_img_multi_out_highbd(image, width, height, stride, | |||
| 209 | cnn_config, &thread_data, | |||
| 210 | bit_depth, &output)) { | |||
| 211 | aom_internal_error(xd->error_info, AOM_CODEC_MEM_ERROR, | |||
| 212 | "Error allocating CNN data"); | |||
| 213 | return; | |||
| 214 | } | |||
| 215 | } else { | |||
| 216 | uint8_t *image[1] = { x->plane[AOM_PLANE_Y0].src.buf - stride - 1 }; | |||
| 217 | ||||
| 218 | if (!av1_cnn_predict_img_multi_out(image, width, height, stride, | |||
| 219 | cnn_config, &thread_data, &output)) { | |||
| 220 | aom_internal_error(xd->error_info, AOM_CODEC_MEM_ERROR, | |||
| 221 | "Error allocating CNN data"); | |||
| 222 | return; | |||
| 223 | } | |||
| 224 | } | |||
| 225 | ||||
| 226 | part_info->cnn_output_valid = 1; | |||
| 227 | } | |||
| 228 | ||||
| 229 | if (!part_info->cnn_output_valid) { | |||
| 230 | return; | |||
| 231 | } | |||
| 232 | ||||
| 233 | const NN_CONFIG *dnn_configs[5] = { | |||
| 234 | NULL((void*)0), | |||
| 235 | &av1_intra_mode_cnn_partition_branch_0_dnn_config, | |||
| 236 | &av1_intra_mode_cnn_partition_branch_1_dnn_config, | |||
| 237 | &av1_intra_mode_cnn_partition_branch_2_dnn_config, | |||
| 238 | &av1_intra_mode_cnn_partition_branch_3_dnn_config, | |||
| 239 | }; | |||
| 240 | ||||
| 241 | const NN_CONFIG *dnn_config = dnn_configs[bsize_idx]; | |||
| 242 | ||||
| 243 | float dnn_features[100]; | |||
| 244 | float logits[4] = { 0.0f }; | |||
| 245 | ||||
| 246 | const float *branch_0 = part_info->cnn_buffer; | |||
| 247 | const float *branch_1 = branch_0 + CNN_BRANCH_0_OUT_SIZE(20); | |||
| 248 | const float *branch_2 = branch_1 + CNN_BRANCH_1_OUT_SIZE((4)*2 * 2); | |||
| 249 | const float *branch_3 = branch_2 + CNN_BRANCH_2_OUT_SIZE((20)*4 * 4); | |||
| 250 | ||||
| 251 | if (bsize == BLOCK_64X64) { | |||
| 252 | int f_idx = 0; | |||
| 253 | for (int ch_idx = 0; ch_idx < CNN_BRANCH_0_OUT_CH20; ch_idx++) { | |||
| 254 | dnn_features[f_idx++] = branch_0[ch_idx]; | |||
| 255 | } | |||
| 256 | ||||
| 257 | const int spa_stride = 2 * 2; | |||
| 258 | for (int lin_idx = 0; lin_idx < spa_stride; lin_idx++) { | |||
| 259 | for (int ch_idx = 0; ch_idx < CNN_BRANCH_1_OUT_CH4; ch_idx++) { | |||
| 260 | dnn_features[f_idx++] = branch_1[lin_idx + ch_idx * spa_stride]; | |||
| 261 | } | |||
| 262 | } | |||
| 263 | dnn_features[f_idx++] = part_info->log_q; | |||
| 264 | } else if (bsize == BLOCK_32X32) { | |||
| 265 | int f_idx = 0; | |||
| 266 | for (int idx = 0; idx < CNN_BRANCH_0_OUT_CH20; idx++) { | |||
| 267 | dnn_features[f_idx++] = branch_0[idx]; | |||
| 268 | } | |||
| 269 | ||||
| 270 | const int curr_lin_idx = quad_to_linear_1[quad_tree_idx - 1]; | |||
| 271 | const int spa_stride = 2 * 2; | |||
| 272 | for (int ch_idx = 0; ch_idx < CNN_BRANCH_1_OUT_CH4; ch_idx++) { | |||
| 273 | dnn_features[f_idx++] = branch_1[curr_lin_idx + ch_idx * spa_stride]; | |||
| 274 | } | |||
| 275 | dnn_features[f_idx++] = part_info->log_q; | |||
| 276 | } else if (bsize == BLOCK_16X16) { | |||
| 277 | int f_idx = 0; | |||
| 278 | const int prev_quad_idx = (quad_tree_idx - 1) / 4; | |||
| 279 | const int prev_lin_idx = quad_to_linear_1[prev_quad_idx - 1]; | |||
| 280 | const int prev_spa_stride = 2 * 2; | |||
| 281 | for (int ch_idx = 0; ch_idx < CNN_BRANCH_1_OUT_CH4; ch_idx++) { | |||
| 282 | dnn_features[f_idx++] = branch_1[prev_lin_idx + ch_idx * prev_spa_stride]; | |||
| 283 | } | |||
| 284 | ||||
| 285 | const int curr_lin_idx = quad_to_linear_2[quad_tree_idx - 5]; | |||
| 286 | const int spa_stride = 4 * 4; | |||
| 287 | for (int ch_idx = 0; ch_idx < CNN_BRANCH_2_OUT_CH20; ch_idx++) { | |||
| 288 | dnn_features[f_idx++] = branch_2[curr_lin_idx + ch_idx * spa_stride]; | |||
| 289 | } | |||
| 290 | dnn_features[f_idx++] = part_info->log_q; | |||
| 291 | } else if (bsize == BLOCK_8X8) { | |||
| 292 | int f_idx = 0; | |||
| 293 | const int prev_quad_idx = (quad_tree_idx - 1) / 4; | |||
| 294 | const int prev_lin_idx = quad_to_linear_2[prev_quad_idx - 5]; | |||
| 295 | const int prev_spa_stride = 4 * 4; | |||
| 296 | for (int ch_idx = 0; ch_idx < CNN_BRANCH_2_OUT_CH20; ch_idx++) { | |||
| 297 | dnn_features[f_idx++] = branch_2[prev_lin_idx + ch_idx * prev_spa_stride]; | |||
| 298 | } | |||
| 299 | ||||
| 300 | const int curr_lin_idx = quad_to_linear_3[quad_tree_idx - 21]; | |||
| 301 | const int spa_stride = 8 * 8; | |||
| 302 | for (int ch_idx = 0; ch_idx < CNN_BRANCH_3_OUT_CH20; ch_idx++) { | |||
| 303 | dnn_features[f_idx++] = branch_3[curr_lin_idx + ch_idx * spa_stride]; | |||
| 304 | } | |||
| 305 | dnn_features[f_idx++] = part_info->log_q; | |||
| 306 | } else { | |||
| 307 | assert(0 && "Invalid bsize in intra_cnn partition")((void) sizeof ((0 && "Invalid bsize in intra_cnn partition" ) ? 1 : 0), __extension__ ({ if (0 && "Invalid bsize in intra_cnn partition" ) ; else __assert_fail ("0 && \"Invalid bsize in intra_cnn partition\"" , "/root/firefox-clang/third_party/aom/av1/encoder/partition_strategy.c" , 307, __extension__ __PRETTY_FUNCTION__); })); | |||
| 308 | } | |||
| 309 | ||||
| 310 | // Make decision | |||
| 311 | av1_nn_predict(dnn_features, dnn_config, 1, logits); | |||
| 312 | ||||
| 313 | const int is_720p_or_larger = AOMMIN(cm->width, cm->height)(((cm->width) < (cm->height)) ? (cm->width) : (cm ->height)) >= 720; | |||
| 314 | const int is_480p_or_larger = AOMMIN(cm->width, cm->height)(((cm->width) < (cm->height)) ? (cm->width) : (cm ->height)) >= 480; | |||
| 315 | float split_only_thresh = 100.0f, no_split_thresh = -100.0f; | |||
| 316 | if (is_720p_or_larger) { | |||
| 317 | split_only_thresh = | |||
| 318 | av1_intra_mode_cnn_partition_split_thresh_hdres[bsize_idx]; | |||
| 319 | no_split_thresh = | |||
| 320 | av1_intra_mode_cnn_partition_no_split_thresh_hdres[bsize_idx]; | |||
| 321 | } else if (is_480p_or_larger) { | |||
| 322 | split_only_thresh = | |||
| 323 | av1_intra_mode_cnn_partition_split_thresh_midres[bsize_idx]; | |||
| 324 | no_split_thresh = | |||
| 325 | av1_intra_mode_cnn_partition_no_split_thresh_midres[bsize_idx]; | |||
| 326 | } else { | |||
| 327 | split_only_thresh = | |||
| 328 | av1_intra_mode_cnn_partition_split_thresh_lowres[bsize_idx]; | |||
| 329 | no_split_thresh = | |||
| 330 | av1_intra_mode_cnn_partition_no_split_thresh_lowres[bsize_idx]; | |||
| 331 | } | |||
| 332 | ||||
| 333 | if (logits[0] > split_only_thresh) { | |||
| 334 | // As screen contents tend to choose larger partitions, do not prune | |||
| 335 | // PARTITION_NONE when intra_cnn_based_part_prune_level=1. | |||
| 336 | if (intra_cnn_based_part_prune_level != 1) { | |||
| 337 | part_state->partition_none_allowed = 0; | |||
| 338 | } | |||
| 339 | part_state->do_square_split = 1; | |||
| 340 | av1_disable_rect_partitions(part_state); | |||
| 341 | } | |||
| 342 | ||||
| 343 | if (logits[0] < no_split_thresh) { | |||
| 344 | av1_disable_square_split_partition(part_state); | |||
| 345 | } | |||
| 346 | } | |||
| 347 | ||||
| 348 | static inline int get_simple_motion_search_prune_agg(int qindex, | |||
| 349 | int prune_level, | |||
| 350 | int is_rect_part) { | |||
| 351 | assert(prune_level < TOTAL_AGG_LVLS)((void) sizeof ((prune_level < TOTAL_AGG_LVLS) ? 1 : 0), __extension__ ({ if (prune_level < TOTAL_AGG_LVLS) ; else __assert_fail ("prune_level < TOTAL_AGG_LVLS", "/root/firefox-clang/third_party/aom/av1/encoder/partition_strategy.c" , 351, __extension__ __PRETTY_FUNCTION__); })); | |||
| 352 | if (prune_level == NO_PRUNING) { | |||
| 353 | return -1; | |||
| 354 | } | |||
| 355 | ||||
| 356 | // Aggressiveness value for SIMPLE_MOTION_SEARCH_PRUNE_LEVEL except | |||
| 357 | // QIDX_BASED_AGG_LVL | |||
| 358 | const int sms_prune_agg_levels[TOTAL_SIMPLE_AGG_LVLS] = { 0, 1, 2, 3 }; | |||
| 359 | if (prune_level < TOTAL_SIMPLE_AGG_LVLS) { | |||
| 360 | return sms_prune_agg_levels[prune_level]; | |||
| 361 | } | |||
| 362 | ||||
| 363 | // Map the QIDX_BASED_AGG_LVL to corresponding aggressiveness value. | |||
| 364 | // Aggressive pruning for lower quantizers in non-boosted frames to prune | |||
| 365 | // rectangular partitions. | |||
| 366 | const int qband = is_rect_part ? (qindex <= 90 ? 1 : 0) : 0; | |||
| 367 | const int sms_prune_agg_qindex_based[2] = { 1, 2 }; | |||
| 368 | return sms_prune_agg_qindex_based[qband]; | |||
| 369 | } | |||
| 370 | ||||
| 371 | // Performs a simple_motion_search with a single reference frame and extract | |||
| 372 | // the variance of residues. Then use the features to determine whether we want | |||
| 373 | // to go straight to splitting without trying PARTITION_NONE | |||
| 374 | static void simple_motion_search_based_split(AV1_COMP *const cpi, MACROBLOCK *x, | |||
| 375 | SIMPLE_MOTION_DATA_TREE *sms_tree, | |||
| 376 | PartitionSearchState *part_state) { | |||
| 377 | const AV1_COMMON *const cm = &cpi->common; | |||
| 378 | const PartitionBlkParams *blk_params = &part_state->part_blk_params; | |||
| 379 | const int mi_row = blk_params->mi_row, mi_col = blk_params->mi_col; | |||
| 380 | const BLOCK_SIZE bsize = blk_params->bsize; | |||
| 381 | ||||
| 382 | const int bsize_idx = convert_bsize_to_idx(bsize); | |||
| 383 | const int is_720p_or_larger = AOMMIN(cm->width, cm->height)(((cm->width) < (cm->height)) ? (cm->width) : (cm ->height)) >= 720; | |||
| 384 | const int is_480p_or_larger = AOMMIN(cm->width, cm->height)(((cm->width) < (cm->height)) ? (cm->width) : (cm ->height)) >= 480; | |||
| 385 | // res_idx is 0 for res < 480p, 1 for 480p, 2 for 720p+ | |||
| 386 | const int res_idx = is_480p_or_larger + is_720p_or_larger; | |||
| 387 | ||||
| 388 | assert(bsize_idx >= 0 && bsize_idx <= 4 &&((void) sizeof ((bsize_idx >= 0 && bsize_idx <= 4 && "Invalid bsize in simple_motion_search_based_split" ) ? 1 : 0), __extension__ ({ if (bsize_idx >= 0 && bsize_idx <= 4 && "Invalid bsize in simple_motion_search_based_split" ) ; else __assert_fail ("bsize_idx >= 0 && bsize_idx <= 4 && \"Invalid bsize in simple_motion_search_based_split\"" , "/root/firefox-clang/third_party/aom/av1/encoder/partition_strategy.c" , 389, __extension__ __PRETTY_FUNCTION__); })) | |||
| 389 | "Invalid bsize in simple_motion_search_based_split")((void) sizeof ((bsize_idx >= 0 && bsize_idx <= 4 && "Invalid bsize in simple_motion_search_based_split" ) ? 1 : 0), __extension__ ({ if (bsize_idx >= 0 && bsize_idx <= 4 && "Invalid bsize in simple_motion_search_based_split" ) ; else __assert_fail ("bsize_idx >= 0 && bsize_idx <= 4 && \"Invalid bsize in simple_motion_search_based_split\"" , "/root/firefox-clang/third_party/aom/av1/encoder/partition_strategy.c" , 389, __extension__ __PRETTY_FUNCTION__); })); | |||
| 390 | ||||
| 391 | const float *ml_mean = av1_simple_motion_search_split_mean[bsize_idx]; | |||
| 392 | const float *ml_std = av1_simple_motion_search_split_std[bsize_idx]; | |||
| 393 | const NN_CONFIG *nn_config = | |||
| 394 | av1_simple_motion_search_split_nn_config[bsize_idx]; | |||
| 395 | ||||
| 396 | const int agg = get_simple_motion_search_prune_agg( | |||
| 397 | x->qindex, cpi->sf.part_sf.simple_motion_search_prune_agg, 0); | |||
| 398 | if (agg < 0) { | |||
| 399 | return; | |||
| 400 | } | |||
| 401 | ||||
| 402 | const float split_only_thresh = | |||
| 403 | av1_simple_motion_search_split_thresh[agg][res_idx][bsize_idx]; | |||
| 404 | const float no_split_thresh = | |||
| 405 | av1_simple_motion_search_no_split_thresh[agg][res_idx][bsize_idx]; | |||
| 406 | ||||
| 407 | float features[FEATURE_SIZE_SMS_SPLIT17] = { 0.0f }; | |||
| 408 | simple_motion_search_prune_part_features(cpi, x, sms_tree, mi_row, mi_col, | |||
| 409 | bsize, features, | |||
| 410 | FEATURE_SMS_SPLIT_MODEL_FLAG(1 | (1 << 1))); | |||
| 411 | ||||
| 412 | // Write features to file | |||
| 413 | write_features_to_file(cpi->oxcf.partition_info_path, | |||
| 414 | cpi->ext_part_controller.test_mode, features, | |||
| 415 | FEATURE_SIZE_SMS_SPLIT17, 0, bsize, mi_row, mi_col); | |||
| 416 | ||||
| 417 | // Note: it is intended to not normalize the features here, to keep it | |||
| 418 | // consistent for all features collected and passed to the external model. | |||
| 419 | if (ext_ml_model_decision_before_none( | |||
| 420 | cpi, features, &part_state->partition_none_allowed, | |||
| 421 | &part_state->partition_rect_allowed[HORZ], | |||
| 422 | &part_state->partition_rect_allowed[VERT], | |||
| 423 | &part_state->do_rectangular_split, &part_state->do_square_split)) { | |||
| 424 | return; | |||
| 425 | } | |||
| 426 | ||||
| 427 | for (int idx = 0; idx < FEATURE_SIZE_SMS_SPLIT17; idx++) { | |||
| 428 | features[idx] = (features[idx] - ml_mean[idx]) / ml_std[idx]; | |||
| 429 | } | |||
| 430 | ||||
| 431 | float score = 0.0f; | |||
| 432 | ||||
| 433 | av1_nn_predict(features, nn_config, 1, &score); | |||
| 434 | ||||
| 435 | if (score > split_only_thresh) { | |||
| 436 | av1_set_square_split_only(part_state); | |||
| 437 | } | |||
| 438 | ||||
| 439 | if (cpi->sf.part_sf.simple_motion_search_split >= 2 && | |||
| 440 | score < no_split_thresh) { | |||
| 441 | av1_disable_square_split_partition(part_state); | |||
| 442 | } | |||
| 443 | ||||
| 444 | // If the score is very low, prune rectangular split since it is unlikely to | |||
| 445 | // occur. | |||
| 446 | if (cpi->sf.part_sf.simple_motion_search_rect_split) { | |||
| 447 | const float scale = res_idx >= 2 ? 3.0f : 2.0f; | |||
| 448 | const float rect_split_thresh = | |||
| 449 | scale * av1_simple_motion_search_no_split_thresh | |||
| 450 | [cpi->sf.part_sf.simple_motion_search_rect_split][res_idx] | |||
| 451 | [bsize_idx]; | |||
| 452 | if (score < rect_split_thresh) { | |||
| 453 | part_state->do_rectangular_split = 0; | |||
| 454 | } | |||
| 455 | } | |||
| 456 | } | |||
| 457 | ||||
| 458 | // Given a list of ref frames in refs, performs simple_motion_search on each of | |||
| 459 | // the refs and returns the ref with the smallest sse. Returns -1 if none of the | |||
| 460 | // ref in the list is available. Also stores the best sse and var in best_sse, | |||
| 461 | // best_var, respectively. If save_mv is 0, don't update mv_ref_fulls in | |||
| 462 | // sms_tree. If save_mv is 1, update mv_ref_fulls under sms_tree and the | |||
| 463 | // subtrees. | |||
| 464 | static int simple_motion_search_get_best_ref( | |||
| 465 | AV1_COMP *const cpi, MACROBLOCK *x, SIMPLE_MOTION_DATA_TREE *sms_tree, | |||
| 466 | int mi_row, int mi_col, BLOCK_SIZE bsize, const int *const refs, | |||
| 467 | int num_refs, int use_subpixel, int save_mv, unsigned int *best_sse, | |||
| 468 | unsigned int *best_var) { | |||
| 469 | const AV1_COMMON *const cm = &cpi->common; | |||
| 470 | int best_ref = -1; | |||
| 471 | ||||
| 472 | if (mi_col >= cm->mi_params.mi_cols || mi_row >= cm->mi_params.mi_rows) { | |||
| 473 | // If the whole block is outside of the image, set the var and sse to 0. | |||
| 474 | *best_var = 0; | |||
| 475 | *best_sse = 0; | |||
| 476 | ||||
| 477 | return best_ref; | |||
| 478 | } | |||
| 479 | ||||
| 480 | // Otherwise do loop through the reference frames and find the one with the | |||
| 481 | // minimum SSE | |||
| 482 | const int num_planes = 1; | |||
| 483 | ||||
| 484 | *best_sse = INT_MAX2147483647; | |||
| 485 | ||||
| 486 | for (int ref_idx = 0; ref_idx < num_refs; ref_idx++) { | |||
| 487 | const int ref = refs[ref_idx]; | |||
| 488 | ||||
| 489 | if (cpi->ref_frame_flags & av1_ref_frame_flag_list[ref]) { | |||
| 490 | const FULLPEL_MV *start_mvs = sms_tree->start_mvs; | |||
| 491 | unsigned int curr_sse = 0, curr_var = 0; | |||
| 492 | const int_mv best_mv = av1_simple_motion_search_sse_var( | |||
| 493 | cpi, x, mi_row, mi_col, bsize, ref, start_mvs[ref], num_planes, | |||
| 494 | use_subpixel, &curr_sse, &curr_var); | |||
| 495 | if (curr_sse < *best_sse) { | |||
| 496 | *best_sse = curr_sse; | |||
| 497 | *best_var = curr_var; | |||
| 498 | best_ref = ref; | |||
| 499 | } | |||
| 500 | ||||
| 501 | if (save_mv) { | |||
| 502 | sms_tree->start_mvs[ref].row = best_mv.as_mv.row / 8; | |||
| 503 | sms_tree->start_mvs[ref].col = best_mv.as_mv.col / 8; | |||
| 504 | ||||
| 505 | if (bsize >= BLOCK_8X8) { | |||
| 506 | for (int r_idx = 0; r_idx < SUB_PARTITIONS_SPLIT4; r_idx++) { | |||
| 507 | // Propagate the new motion vectors to a lower level | |||
| 508 | SIMPLE_MOTION_DATA_TREE *sub_tree = sms_tree->split[r_idx]; | |||
| 509 | sub_tree->start_mvs[ref] = sms_tree->start_mvs[ref]; | |||
| 510 | } | |||
| 511 | } | |||
| 512 | } | |||
| 513 | } | |||
| 514 | } | |||
| 515 | ||||
| 516 | return best_ref; | |||
| 517 | } | |||
| 518 | ||||
| 519 | // Collects features using simple_motion_search and store them in features. The | |||
| 520 | // features are also cached in SIMPLE_MOTION_DATA_TREE. By default, the features | |||
| 521 | // collected are the sse and var from the subblocks flagged by features_to_get. | |||
| 522 | // Furthermore, if features is not NULL, then 7 more features are appended to | |||
| 523 | // the end of features: | |||
| 524 | // - log(1.0 + dc_q ** 2) | |||
| 525 | // - whether an above macroblock exists | |||
| 526 | // - width of above macroblock | |||
| 527 | // - height of above macroblock | |||
| 528 | // - whether a left marcoblock exists | |||
| 529 | // - width of left macroblock | |||
| 530 | // - height of left macroblock | |||
| 531 | static inline void simple_motion_search_prune_part_features( | |||
| 532 | AV1_COMP *const cpi, MACROBLOCK *x, SIMPLE_MOTION_DATA_TREE *sms_tree, | |||
| 533 | int mi_row, int mi_col, BLOCK_SIZE bsize, float *features, | |||
| 534 | int features_to_get) { | |||
| 535 | const int w_mi = mi_size_wide[bsize]; | |||
| 536 | const int h_mi = mi_size_high[bsize]; | |||
| 537 | assert(mi_size_wide[bsize] == mi_size_high[bsize])((void) sizeof ((mi_size_wide[bsize] == mi_size_high[bsize]) ? 1 : 0), __extension__ ({ if (mi_size_wide[bsize] == mi_size_high [bsize]) ; else __assert_fail ("mi_size_wide[bsize] == mi_size_high[bsize]" , "/root/firefox-clang/third_party/aom/av1/encoder/partition_strategy.c" , 537, __extension__ __PRETTY_FUNCTION__); })); | |||
| 538 | assert(bsize >= BLOCK_8X8)((void) sizeof ((bsize >= BLOCK_8X8) ? 1 : 0), __extension__ ({ if (bsize >= BLOCK_8X8) ; else __assert_fail ("bsize >= BLOCK_8X8" , "/root/firefox-clang/third_party/aom/av1/encoder/partition_strategy.c" , 538, __extension__ __PRETTY_FUNCTION__); })); | |||
| 539 | assert(cpi->ref_frame_flags & av1_ref_frame_flag_list[LAST_FRAME] ||((void) sizeof ((cpi->ref_frame_flags & av1_ref_frame_flag_list [LAST_FRAME] || cpi->ref_frame_flags & av1_ref_frame_flag_list [ALTREF_FRAME]) ? 1 : 0), __extension__ ({ if (cpi->ref_frame_flags & av1_ref_frame_flag_list[LAST_FRAME] || cpi->ref_frame_flags & av1_ref_frame_flag_list[ALTREF_FRAME]) ; else __assert_fail ("cpi->ref_frame_flags & av1_ref_frame_flag_list[LAST_FRAME] || cpi->ref_frame_flags & av1_ref_frame_flag_list[ALTREF_FRAME]" , "/root/firefox-clang/third_party/aom/av1/encoder/partition_strategy.c" , 540, __extension__ __PRETTY_FUNCTION__); })) | |||
| 540 | cpi->ref_frame_flags & av1_ref_frame_flag_list[ALTREF_FRAME])((void) sizeof ((cpi->ref_frame_flags & av1_ref_frame_flag_list [LAST_FRAME] || cpi->ref_frame_flags & av1_ref_frame_flag_list [ALTREF_FRAME]) ? 1 : 0), __extension__ ({ if (cpi->ref_frame_flags & av1_ref_frame_flag_list[LAST_FRAME] || cpi->ref_frame_flags & av1_ref_frame_flag_list[ALTREF_FRAME]) ; else __assert_fail ("cpi->ref_frame_flags & av1_ref_frame_flag_list[LAST_FRAME] || cpi->ref_frame_flags & av1_ref_frame_flag_list[ALTREF_FRAME]" , "/root/firefox-clang/third_party/aom/av1/encoder/partition_strategy.c" , 540, __extension__ __PRETTY_FUNCTION__); })); | |||
| 541 | ||||
| 542 | // Setting up motion search | |||
| 543 | const int ref_list[] = { cpi->rc.is_src_frame_alt_ref ? ALTREF_FRAME | |||
| 544 | : LAST_FRAME }; | |||
| 545 | const int num_refs = 1; | |||
| 546 | const int use_subpixel = 1; | |||
| 547 | ||||
| 548 | // Doing whole block first to update the mv | |||
| 549 | if (!sms_tree->sms_none_valid && features_to_get & FEATURE_SMS_NONE_FLAG1) { | |||
| 550 | simple_motion_search_get_best_ref(cpi, x, sms_tree, mi_row, mi_col, bsize, | |||
| 551 | ref_list, num_refs, use_subpixel, 1, | |||
| 552 | &sms_tree->sms_none_feat[0], | |||
| 553 | &sms_tree->sms_none_feat[1]); | |||
| 554 | sms_tree->sms_none_valid = 1; | |||
| 555 | } | |||
| 556 | ||||
| 557 | // Split subblocks | |||
| 558 | if (features_to_get & FEATURE_SMS_SPLIT_FLAG(1 << 1)) { | |||
| 559 | const BLOCK_SIZE subsize = get_partition_subsize(bsize, PARTITION_SPLIT); | |||
| 560 | for (int r_idx = 0; r_idx < SUB_PARTITIONS_SPLIT4; r_idx++) { | |||
| 561 | const int sub_mi_col = mi_col + (r_idx & 1) * w_mi / 2; | |||
| 562 | const int sub_mi_row = mi_row + (r_idx >> 1) * h_mi / 2; | |||
| 563 | SIMPLE_MOTION_DATA_TREE *sub_tree = sms_tree->split[r_idx]; | |||
| 564 | ||||
| 565 | if (!sub_tree->sms_none_valid) { | |||
| 566 | simple_motion_search_get_best_ref( | |||
| 567 | cpi, x, sub_tree, sub_mi_row, sub_mi_col, subsize, ref_list, | |||
| 568 | num_refs, use_subpixel, 1, &sub_tree->sms_none_feat[0], | |||
| 569 | &sub_tree->sms_none_feat[1]); | |||
| 570 | sub_tree->sms_none_valid = 1; | |||
| 571 | } | |||
| 572 | } | |||
| 573 | } | |||
| 574 | ||||
| 575 | // Rectangular subblocks | |||
| 576 | if (!sms_tree->sms_rect_valid && features_to_get & FEATURE_SMS_RECT_FLAG(1 << 2)) { | |||
| 577 | // Horz subblock | |||
| 578 | BLOCK_SIZE subsize = get_partition_subsize(bsize, PARTITION_HORZ); | |||
| 579 | for (int r_idx = 0; r_idx < SUB_PARTITIONS_RECT2; r_idx++) { | |||
| 580 | const int sub_mi_col = mi_col + 0; | |||
| 581 | const int sub_mi_row = mi_row + r_idx * h_mi / 2; | |||
| 582 | ||||
| 583 | simple_motion_search_get_best_ref( | |||
| 584 | cpi, x, sms_tree, sub_mi_row, sub_mi_col, subsize, ref_list, num_refs, | |||
| 585 | use_subpixel, 0, &sms_tree->sms_rect_feat[2 * r_idx], | |||
| 586 | &sms_tree->sms_rect_feat[2 * r_idx + 1]); | |||
| 587 | } | |||
| 588 | ||||
| 589 | // Vert subblock | |||
| 590 | subsize = get_partition_subsize(bsize, PARTITION_VERT); | |||
| 591 | for (int r_idx = 0; r_idx < SUB_PARTITIONS_RECT2; r_idx++) { | |||
| 592 | const int sub_mi_col = mi_col + r_idx * w_mi / 2; | |||
| 593 | const int sub_mi_row = mi_row + 0; | |||
| 594 | ||||
| 595 | simple_motion_search_get_best_ref( | |||
| 596 | cpi, x, sms_tree, sub_mi_row, sub_mi_col, subsize, ref_list, num_refs, | |||
| 597 | use_subpixel, 0, &sms_tree->sms_rect_feat[4 + 2 * r_idx], | |||
| 598 | &sms_tree->sms_rect_feat[4 + 2 * r_idx + 1]); | |||
| 599 | } | |||
| 600 | sms_tree->sms_rect_valid = 1; | |||
| 601 | } | |||
| 602 | ||||
| 603 | if (!features) return; | |||
| 604 | ||||
| 605 | int f_idx = 0; | |||
| 606 | if (features_to_get & FEATURE_SMS_NONE_FLAG1) { | |||
| 607 | for (int sub_idx = 0; sub_idx < 2; sub_idx++) { | |||
| 608 | features[f_idx++] = log1pf((float)sms_tree->sms_none_feat[sub_idx]); | |||
| 609 | } | |||
| 610 | } | |||
| 611 | ||||
| 612 | if (features_to_get & FEATURE_SMS_SPLIT_FLAG(1 << 1)) { | |||
| 613 | for (int sub_idx = 0; sub_idx < SUB_PARTITIONS_SPLIT4; sub_idx++) { | |||
| 614 | SIMPLE_MOTION_DATA_TREE *sub_tree = sms_tree->split[sub_idx]; | |||
| 615 | features[f_idx++] = log1pf((float)sub_tree->sms_none_feat[0]); | |||
| 616 | features[f_idx++] = log1pf((float)sub_tree->sms_none_feat[1]); | |||
| 617 | } | |||
| 618 | } | |||
| 619 | ||||
| 620 | if (features_to_get & FEATURE_SMS_RECT_FLAG(1 << 2)) { | |||
| 621 | for (int sub_idx = 0; sub_idx < 8; sub_idx++) { | |||
| 622 | features[f_idx++] = log1pf((float)sms_tree->sms_rect_feat[sub_idx]); | |||
| 623 | } | |||
| 624 | } | |||
| 625 | ||||
| 626 | const MACROBLOCKD *xd = &x->e_mbd; | |||
| 627 | set_offsets_for_motion_search(cpi, x, mi_row, mi_col, bsize); | |||
| 628 | ||||
| 629 | // Q_INDEX | |||
| 630 | const int dc_q = av1_dc_quant_QTX(x->qindex, 0, xd->bd) >> (xd->bd - 8); | |||
| 631 | features[f_idx++] = log1pf((float)(dc_q * dc_q) / 256.0f); | |||
| 632 | ||||
| 633 | // Neighbor stuff | |||
| 634 | const int has_above = !!xd->above_mbmi; | |||
| 635 | const int has_left = !!xd->left_mbmi; | |||
| 636 | const BLOCK_SIZE above_bsize = has_above ? xd->above_mbmi->bsize : bsize; | |||
| 637 | const BLOCK_SIZE left_bsize = has_left ? xd->left_mbmi->bsize : bsize; | |||
| 638 | features[f_idx++] = (float)has_above; | |||
| 639 | features[f_idx++] = (float)mi_size_wide_log2[above_bsize]; | |||
| 640 | features[f_idx++] = (float)mi_size_high_log2[above_bsize]; | |||
| 641 | features[f_idx++] = (float)has_left; | |||
| 642 | features[f_idx++] = (float)mi_size_wide_log2[left_bsize]; | |||
| 643 | features[f_idx++] = (float)mi_size_high_log2[left_bsize]; | |||
| 644 | } | |||
| 645 | ||||
| 646 | // Performs a simple_motion_search with two reference frames and extract | |||
| 647 | // the variance of residues. Then use the features to determine whether we want | |||
| 648 | // to prune some partitions. | |||
| 649 | static void simple_motion_search_prune_rect(AV1_COMP *const cpi, MACROBLOCK *x, | |||
| 650 | SIMPLE_MOTION_DATA_TREE *sms_tree, | |||
| 651 | PartitionSearchState *part_state) { | |||
| 652 | const AV1_COMMON *const cm = &cpi->common; | |||
| 653 | const PartitionBlkParams *blk_params = &part_state->part_blk_params; | |||
| 654 | const int mi_row = blk_params->mi_row, mi_col = blk_params->mi_col; | |||
| 655 | const BLOCK_SIZE bsize = blk_params->bsize; | |||
| 656 | ||||
| 657 | const int bsize_idx = convert_bsize_to_idx(bsize); | |||
| 658 | const int is_720p_or_larger = AOMMIN(cm->width, cm->height)(((cm->width) < (cm->height)) ? (cm->width) : (cm ->height)) >= 720; | |||
| 659 | const int is_480p_or_larger = AOMMIN(cm->width, cm->height)(((cm->width) < (cm->height)) ? (cm->width) : (cm ->height)) >= 480; | |||
| 660 | // res_idx is 0 for lowres, 1 for 48p, 2 for 720p+ | |||
| 661 | const int res_idx = is_480p_or_larger + is_720p_or_larger; | |||
| 662 | ||||
| 663 | // Get model parameters | |||
| 664 | const NN_CONFIG *nn_config = | |||
| 665 | av1_simple_motion_search_prune_rect_nn_config[bsize_idx]; | |||
| 666 | const float *ml_mean = av1_simple_motion_search_prune_rect_mean[bsize_idx], | |||
| 667 | *ml_std = av1_simple_motion_search_prune_rect_std[bsize_idx]; | |||
| 668 | ||||
| 669 | const int agg = get_simple_motion_search_prune_agg( | |||
| 670 | x->qindex, cpi->sf.part_sf.simple_motion_search_prune_agg, 1); | |||
| 671 | if (agg < 0) { | |||
| 672 | return; | |||
| 673 | } | |||
| 674 | ||||
| 675 | const float prune_thresh = | |||
| 676 | av1_simple_motion_search_prune_rect_thresh[agg][res_idx][bsize_idx]; | |||
| 677 | ||||
| 678 | // If there is no valid threshold, return immediately. | |||
| 679 | if (!nn_config || prune_thresh == 0.0f) { | |||
| 680 | return; | |||
| 681 | } | |||
| 682 | ||||
| 683 | // Get features | |||
| 684 | float features[FEATURE_SIZE_SMS_PRUNE_PART25] = { 0.0f }; | |||
| 685 | simple_motion_search_prune_part_features(cpi, x, sms_tree, mi_row, mi_col, | |||
| 686 | bsize, features, | |||
| 687 | FEATURE_SMS_PRUNE_PART_FLAG(1 | (1 << 1) | (1 << 2))); | |||
| 688 | ||||
| 689 | // Note: it is intended to not normalize the features here, to keep it | |||
| 690 | // consistent for all features collected and passed to the external model. | |||
| 691 | if (cpi->sf.part_sf.simple_motion_search_prune_rect && | |||
| 692 | !frame_is_intra_only(cm) && | |||
| 693 | (part_state->partition_rect_allowed[HORZ] || | |||
| 694 | part_state->partition_rect_allowed[VERT]) && | |||
| 695 | bsize >= BLOCK_8X8 && !av1_superres_scaled(cm)) { | |||
| 696 | // Write features to file | |||
| 697 | write_features_to_file( | |||
| 698 | cpi->oxcf.partition_info_path, cpi->ext_part_controller.test_mode, | |||
| 699 | features, FEATURE_SIZE_SMS_PRUNE_PART25, 1, bsize, mi_row, mi_col); | |||
| 700 | ||||
| 701 | if (ext_ml_model_decision_before_none_part2( | |||
| 702 | cpi, features, &part_state->prune_rect_part[HORZ], | |||
| 703 | &part_state->prune_rect_part[VERT])) { | |||
| 704 | return; | |||
| 705 | } | |||
| 706 | } | |||
| 707 | ||||
| 708 | for (int f_idx = 0; f_idx < FEATURE_SIZE_SMS_PRUNE_PART25; f_idx++) { | |||
| 709 | features[f_idx] = (features[f_idx] - ml_mean[f_idx]) / ml_std[f_idx]; | |||
| 710 | } | |||
| 711 | ||||
| 712 | // Get probabilities | |||
| 713 | float scores[EXT_PARTITION_TYPES] = { 0.0f }, | |||
| 714 | probs[EXT_PARTITION_TYPES] = { 0.0f }; | |||
| 715 | const int num_classes = (bsize == BLOCK_128X128 || bsize == BLOCK_8X8) | |||
| 716 | ? PARTITION_TYPES | |||
| 717 | : EXT_PARTITION_TYPES; | |||
| 718 | ||||
| 719 | av1_nn_predict(features, nn_config, 1, scores); | |||
| 720 | ||||
| 721 | av1_nn_softmax(scores, probs, num_classes); | |||
| 722 | ||||
| 723 | // Determine if we should prune rectangular partitions. | |||
| 724 | if (probs[PARTITION_HORZ] <= prune_thresh) { | |||
| 725 | part_state->prune_rect_part[HORZ] = 1; | |||
| 726 | } | |||
| 727 | if (probs[PARTITION_VERT] <= prune_thresh) { | |||
| 728 | part_state->prune_rect_part[VERT] = 1; | |||
| 729 | } | |||
| 730 | } | |||
| 731 | ||||
| 732 | // Early terminates PARTITION_NONE using simple_motion_search features and the | |||
| 733 | // rate, distortion, and rdcost of PARTITION_NONE. This is only called when: | |||
| 734 | // - The frame is a show frame | |||
| 735 | // - The frame is not intra only | |||
| 736 | // - The current bsize is > BLOCK_8X8 | |||
| 737 | // - blk_row + blk_height/2 < total_rows and blk_col + blk_width/2 < total_cols | |||
| 738 | void av1_simple_motion_search_early_term_none( | |||
| 739 | AV1_COMP *const cpi, MACROBLOCK *x, SIMPLE_MOTION_DATA_TREE *sms_tree, | |||
| 740 | const RD_STATS *none_rdc, PartitionSearchState *part_state) { | |||
| 741 | const PartitionBlkParams *blk_params = &part_state->part_blk_params; | |||
| 742 | const int mi_row = blk_params->mi_row, mi_col = blk_params->mi_col; | |||
| 743 | const BLOCK_SIZE bsize = blk_params->bsize; | |||
| 744 | ||||
| 745 | float features[FEATURE_SIZE_SMS_TERM_NONE28] = { 0.0f }; | |||
| 746 | simple_motion_search_prune_part_features(cpi, x, sms_tree, mi_row, mi_col, | |||
| 747 | bsize, features, | |||
| 748 | FEATURE_SMS_PRUNE_PART_FLAG(1 | (1 << 1) | (1 << 2))); | |||
| 749 | int f_idx = FEATURE_SIZE_SMS_PRUNE_PART25; | |||
| 750 | ||||
| 751 | features[f_idx++] = log1pf((float)none_rdc->rate); | |||
| 752 | features[f_idx++] = log1pf((float)none_rdc->dist); | |||
| 753 | features[f_idx++] = log1pf((float)none_rdc->rdcost); | |||
| 754 | ||||
| 755 | assert(f_idx == FEATURE_SIZE_SMS_TERM_NONE)((void) sizeof ((f_idx == 28) ? 1 : 0), __extension__ ({ if ( f_idx == 28) ; else __assert_fail ("f_idx == FEATURE_SIZE_SMS_TERM_NONE" , "/root/firefox-clang/third_party/aom/av1/encoder/partition_strategy.c" , 755, __extension__ __PRETTY_FUNCTION__); })); | |||
| 756 | ||||
| 757 | const float *ml_mean = NULL((void*)0); | |||
| 758 | const float *ml_std = NULL((void*)0); | |||
| 759 | const float *ml_model = NULL((void*)0); | |||
| 760 | ||||
| 761 | if (bsize == BLOCK_128X128) { | |||
| 762 | ml_mean = av1_simple_motion_search_term_none_mean_128; | |||
| 763 | ml_std = av1_simple_motion_search_term_none_std_128; | |||
| 764 | ml_model = av1_simple_motion_search_term_none_model_128; | |||
| 765 | } else if (bsize == BLOCK_64X64) { | |||
| 766 | ml_mean = av1_simple_motion_search_term_none_mean_64; | |||
| 767 | ml_std = av1_simple_motion_search_term_none_std_64; | |||
| 768 | ml_model = av1_simple_motion_search_term_none_model_64; | |||
| 769 | } else if (bsize == BLOCK_32X32) { | |||
| 770 | ml_mean = av1_simple_motion_search_term_none_mean_32; | |||
| 771 | ml_std = av1_simple_motion_search_term_none_std_32; | |||
| 772 | ml_model = av1_simple_motion_search_term_none_model_32; | |||
| 773 | } else if (bsize == BLOCK_16X16) { | |||
| 774 | ml_mean = av1_simple_motion_search_term_none_mean_16; | |||
| 775 | ml_std = av1_simple_motion_search_term_none_std_16; | |||
| 776 | ml_model = av1_simple_motion_search_term_none_model_16; | |||
| 777 | } else { | |||
| 778 | assert(0 && "Unexpected block size in simple_motion_term_none")((void) sizeof ((0 && "Unexpected block size in simple_motion_term_none" ) ? 1 : 0), __extension__ ({ if (0 && "Unexpected block size in simple_motion_term_none" ) ; else __assert_fail ("0 && \"Unexpected block size in simple_motion_term_none\"" , "/root/firefox-clang/third_party/aom/av1/encoder/partition_strategy.c" , 778, __extension__ __PRETTY_FUNCTION__); })); | |||
| 779 | } | |||
| 780 | ||||
| 781 | // Write features to file | |||
| 782 | write_features_to_file(cpi->oxcf.partition_info_path, | |||
| 783 | cpi->ext_part_controller.test_mode, features, | |||
| 784 | FEATURE_SIZE_SMS_TERM_NONE28, 3, bsize, mi_row, mi_col); | |||
| 785 | ||||
| 786 | if (ext_ml_model_decision_after_none_part2( | |||
| 787 | cpi, features, &part_state->terminate_partition_search)) { | |||
| 788 | return; | |||
| 789 | } | |||
| 790 | ||||
| 791 | if (ml_model) { | |||
| 792 | float score = 0.0f; | |||
| 793 | for (f_idx = 0; f_idx < FEATURE_SIZE_SMS_TERM_NONE28; f_idx++) { | |||
| 794 | score += | |||
| 795 | ml_model[f_idx] * (features[f_idx] - ml_mean[f_idx]) / ml_std[f_idx]; | |||
| 796 | } | |||
| 797 | score += ml_model[FEATURE_SIZE_SMS_TERM_NONE28]; | |||
| 798 | ||||
| 799 | if (score >= 0.0f) { | |||
| 800 | part_state->terminate_partition_search = 1; | |||
| 801 | } | |||
| 802 | } | |||
| 803 | } | |||
| 804 | ||||
| 805 | void av1_get_max_min_partition_features(AV1_COMP *const cpi, MACROBLOCK *x, | |||
| 806 | int mi_row, int mi_col, | |||
| 807 | float *features) { | |||
| 808 | AV1_COMMON *const cm = &cpi->common; | |||
| 809 | MACROBLOCKD *xd = &x->e_mbd; | |||
| 810 | const BLOCK_SIZE sb_size = cm->seq_params->sb_size; | |||
| 811 | ||||
| 812 | // Currently this only allows 128X128 SB size. May extend it to 64X64 SB size. | |||
| 813 | assert(sb_size == BLOCK_128X128)((void) sizeof ((sb_size == BLOCK_128X128) ? 1 : 0), __extension__ ({ if (sb_size == BLOCK_128X128) ; else __assert_fail ("sb_size == BLOCK_128X128" , "/root/firefox-clang/third_party/aom/av1/encoder/partition_strategy.c" , 813, __extension__ __PRETTY_FUNCTION__); })); | |||
| 814 | ||||
| 815 | int f_idx = 0; | |||
| 816 | ||||
| 817 | const int dc_q = av1_dc_quant_QTX(x->qindex, 0, xd->bd) >> (xd->bd - 8); | |||
| 818 | const float log_q_sq = log1pf((float)(dc_q * dc_q) / 256.0f); | |||
| 819 | ||||
| 820 | // Perform full-pixel single motion search in Y plane of 16x16 mbs in the sb | |||
| 821 | float sum_mv_row_sq = 0; | |||
| 822 | float sum_mv_row = 0; | |||
| 823 | float min_abs_mv_row = FLT_MAX3.40282347e+38F; | |||
| 824 | float max_abs_mv_row = 0; | |||
| 825 | ||||
| 826 | float sum_mv_col_sq = 0; | |||
| 827 | float sum_mv_col = 0; | |||
| 828 | float min_abs_mv_col = FLT_MAX3.40282347e+38F; | |||
| 829 | float max_abs_mv_col = 0; | |||
| 830 | ||||
| 831 | float sum_log_sse_sq = 0; | |||
| 832 | float sum_log_sse = 0; | |||
| 833 | float min_log_sse = FLT_MAX3.40282347e+38F; | |||
| 834 | float max_log_sse = 0; | |||
| 835 | ||||
| 836 | const BLOCK_SIZE mb_size = BLOCK_16X16; | |||
| 837 | const int mb_rows = block_size_high[sb_size] / block_size_high[mb_size]; | |||
| 838 | const int mb_cols = block_size_wide[sb_size] / block_size_wide[mb_size]; | |||
| 839 | const int mb_in_mi_size_high_log2 = mi_size_high_log2[mb_size]; | |||
| 840 | const int mb_in_mi_size_wide_log2 = mi_size_wide_log2[mb_size]; | |||
| 841 | ||||
| 842 | for (int mb_row = 0; mb_row < mb_rows; mb_row++) | |||
| 843 | for (int mb_col = 0; mb_col < mb_cols; mb_col++) { | |||
| 844 | const int this_mi_row = mi_row + (mb_row << mb_in_mi_size_high_log2); | |||
| 845 | const int this_mi_col = mi_col + (mb_col << mb_in_mi_size_wide_log2); | |||
| 846 | unsigned int sse = 0; | |||
| 847 | unsigned int var = 0; | |||
| 848 | const FULLPEL_MV start_mv = kZeroFullMv; | |||
| 849 | const MV_REFERENCE_FRAME ref = | |||
| 850 | cpi->rc.is_src_frame_alt_ref ? ALTREF_FRAME : LAST_FRAME; | |||
| 851 | const int_mv best_mv = av1_simple_motion_search_sse_var( | |||
| 852 | cpi, x, this_mi_row, this_mi_col, mb_size, ref, start_mv, 1, 0, &sse, | |||
| 853 | &var); | |||
| 854 | ||||
| 855 | const float mv_row = (float)(best_mv.as_mv.row / 8); | |||
| 856 | const float mv_col = (float)(best_mv.as_mv.col / 8); | |||
| 857 | const float log_sse = log1pf((float)sse); | |||
| 858 | const float abs_mv_row = fabsf(mv_row); | |||
| 859 | const float abs_mv_col = fabsf(mv_col); | |||
| 860 | ||||
| 861 | sum_mv_row_sq += mv_row * mv_row; | |||
| 862 | sum_mv_row += mv_row; | |||
| 863 | sum_mv_col_sq += mv_col * mv_col; | |||
| 864 | sum_mv_col += mv_col; | |||
| 865 | ||||
| 866 | if (abs_mv_row < min_abs_mv_row) min_abs_mv_row = abs_mv_row; | |||
| 867 | if (abs_mv_row > max_abs_mv_row) max_abs_mv_row = abs_mv_row; | |||
| 868 | if (abs_mv_col < min_abs_mv_col) min_abs_mv_col = abs_mv_col; | |||
| 869 | if (abs_mv_col > max_abs_mv_col) max_abs_mv_col = abs_mv_col; | |||
| 870 | ||||
| 871 | sum_log_sse_sq += log_sse * log_sse; | |||
| 872 | sum_log_sse += log_sse; | |||
| 873 | if (log_sse < min_log_sse) min_log_sse = log_sse; | |||
| 874 | if (log_sse > max_log_sse) max_log_sse = log_sse; | |||
| 875 | } | |||
| 876 | const int blks = mb_rows * mb_cols; | |||
| 877 | const float avg_mv_row = sum_mv_row / (float)blks; | |||
| 878 | const float var_mv_row = | |||
| 879 | sum_mv_row_sq / (float)blks - avg_mv_row * avg_mv_row; | |||
| 880 | ||||
| 881 | const float avg_mv_col = sum_mv_col / (float)blks; | |||
| 882 | const float var_mv_col = | |||
| 883 | sum_mv_col_sq / (float)blks - avg_mv_col * avg_mv_col; | |||
| 884 | ||||
| 885 | const float avg_log_sse = sum_log_sse / (float)blks; | |||
| 886 | const float var_log_sse = | |||
| 887 | sum_log_sse_sq / (float)blks - avg_log_sse * avg_log_sse; | |||
| 888 | ||||
| 889 | features[f_idx++] = avg_log_sse; | |||
| 890 | features[f_idx++] = avg_mv_col; | |||
| 891 | features[f_idx++] = avg_mv_row; | |||
| 892 | features[f_idx++] = log_q_sq; | |||
| 893 | features[f_idx++] = max_abs_mv_col; | |||
| 894 | features[f_idx++] = max_abs_mv_row; | |||
| 895 | features[f_idx++] = max_log_sse; | |||
| 896 | features[f_idx++] = min_abs_mv_col; | |||
| 897 | features[f_idx++] = min_abs_mv_row; | |||
| 898 | features[f_idx++] = min_log_sse; | |||
| 899 | features[f_idx++] = var_log_sse; | |||
| 900 | features[f_idx++] = var_mv_col; | |||
| 901 | features[f_idx++] = var_mv_row; | |||
| 902 | ||||
| 903 | assert(f_idx == FEATURE_SIZE_MAX_MIN_PART_PRED)((void) sizeof ((f_idx == 13) ? 1 : 0), __extension__ ({ if ( f_idx == 13) ; else __assert_fail ("f_idx == FEATURE_SIZE_MAX_MIN_PART_PRED" , "/root/firefox-clang/third_party/aom/av1/encoder/partition_strategy.c" , 903, __extension__ __PRETTY_FUNCTION__); })); | |||
| 904 | } | |||
| 905 | ||||
| 906 | // Convert result index to block size. | |||
| 907 | // result idx block size | |||
| 908 | // 0 BLOCK_16X16 | |||
| 909 | // 1 BLOCK_32X32 | |||
| 910 | // 2 BLOCK_64X64 | |||
| 911 | // 3 BLOCK_128X128 | |||
| 912 | static BLOCK_SIZE get_block_size(int idx) { | |||
| 913 | return (BLOCK_SIZE)((idx + 2) * 3); | |||
| 914 | } | |||
| 915 | ||||
| 916 | BLOCK_SIZE av1_predict_max_partition(const AV1_COMP *const cpi, | |||
| 917 | const MACROBLOCK *const x, | |||
| 918 | const float *features) { | |||
| 919 | float scores[MAX_NUM_CLASSES_MAX_MIN_PART_PRED4] = { 0.0f }; | |||
| 920 | const NN_CONFIG *nn_config = &av1_max_part_pred_nn_config; | |||
| 921 | ||||
| 922 | assert(cpi->sf.part_sf.auto_max_partition_based_on_simple_motion !=((void) sizeof ((cpi->sf.part_sf.auto_max_partition_based_on_simple_motion != NOT_IN_USE) ? 1 : 0), __extension__ ({ if (cpi->sf.part_sf .auto_max_partition_based_on_simple_motion != NOT_IN_USE) ; else __assert_fail ("cpi->sf.part_sf.auto_max_partition_based_on_simple_motion != NOT_IN_USE" , "/root/firefox-clang/third_party/aom/av1/encoder/partition_strategy.c" , 923, __extension__ __PRETTY_FUNCTION__); })) | |||
| 923 | NOT_IN_USE)((void) sizeof ((cpi->sf.part_sf.auto_max_partition_based_on_simple_motion != NOT_IN_USE) ? 1 : 0), __extension__ ({ if (cpi->sf.part_sf .auto_max_partition_based_on_simple_motion != NOT_IN_USE) ; else __assert_fail ("cpi->sf.part_sf.auto_max_partition_based_on_simple_motion != NOT_IN_USE" , "/root/firefox-clang/third_party/aom/av1/encoder/partition_strategy.c" , 923, __extension__ __PRETTY_FUNCTION__); })); | |||
| 924 | ||||
| 925 | av1_nn_predict(features, nn_config, 1, scores); | |||
| 926 | ||||
| 927 | int result = MAX_NUM_CLASSES_MAX_MIN_PART_PRED4 - 1; | |||
| 928 | if (cpi->sf.part_sf.auto_max_partition_based_on_simple_motion == | |||
| 929 | DIRECT_PRED) { | |||
| 930 | result = 0; | |||
| 931 | float max_score = scores[0]; | |||
| 932 | for (int i = 1; i < MAX_NUM_CLASSES_MAX_MIN_PART_PRED4; ++i) { | |||
| 933 | if (scores[i] > max_score) { | |||
| 934 | max_score = scores[i]; | |||
| 935 | result = i; | |||
| 936 | } | |||
| 937 | } | |||
| 938 | return get_block_size(result); | |||
| 939 | } | |||
| 940 | ||||
| 941 | float probs[MAX_NUM_CLASSES_MAX_MIN_PART_PRED4] = { 0.0f }; | |||
| 942 | av1_nn_softmax(scores, probs, MAX_NUM_CLASSES_MAX_MIN_PART_PRED4); | |||
| 943 | ||||
| 944 | if (cpi->sf.part_sf.auto_max_partition_based_on_simple_motion == | |||
| 945 | RELAXED_PRED) { | |||
| 946 | for (result = MAX_NUM_CLASSES_MAX_MIN_PART_PRED4 - 1; result >= 0; | |||
| 947 | --result) { | |||
| 948 | if (result < MAX_NUM_CLASSES_MAX_MIN_PART_PRED4 - 1) { | |||
| 949 | probs[result] += probs[result + 1]; | |||
| 950 | } | |||
| 951 | if (probs[result] > 0.2) break; | |||
| 952 | } | |||
| 953 | } else if (cpi->sf.part_sf.auto_max_partition_based_on_simple_motion == | |||
| 954 | ADAPT_PRED) { | |||
| 955 | const BLOCK_SIZE sb_size = cpi->common.seq_params->sb_size; | |||
| 956 | // TODO(debargha): x->source_variance is unavailable at this point, | |||
| 957 | // so compute. The redundant recomputation later can be removed. | |||
| 958 | const unsigned int source_variance = av1_get_perpixel_variance_facade( | |||
| 959 | cpi, &x->e_mbd, &x->plane[0].src, sb_size, AOM_PLANE_Y0); | |||
| 960 | if (source_variance > 16) { | |||
| 961 | const double thresh = source_variance < 128 ? 0.05 : 0.1; | |||
| 962 | for (result = MAX_NUM_CLASSES_MAX_MIN_PART_PRED4 - 1; result >= 0; | |||
| 963 | --result) { | |||
| 964 | if (result < MAX_NUM_CLASSES_MAX_MIN_PART_PRED4 - 1) { | |||
| 965 | probs[result] += probs[result + 1]; | |||
| 966 | } | |||
| 967 | if (probs[result] > thresh) break; | |||
| 968 | } | |||
| 969 | } | |||
| 970 | } | |||
| 971 | ||||
| 972 | return get_block_size(result); | |||
| 973 | } | |||
| 974 | ||||
| 975 | // Get the minimum partition block width and height(in log scale) under a | |||
| 976 | // SIMPLE_MOTION_DATA_TREE. | |||
| 977 | static inline void get_min_bsize(const SIMPLE_MOTION_DATA_TREE *sms_tree, | |||
| 978 | int *min_bw, int *min_bh) { | |||
| 979 | if (!sms_tree) return; | |||
| 980 | ||||
| 981 | const BLOCK_SIZE bsize = sms_tree->block_size; | |||
| 982 | if (bsize == BLOCK_4X4) { | |||
| 983 | *min_bw = 0; | |||
| 984 | *min_bh = 0; | |||
| 985 | return; | |||
| 986 | } | |||
| 987 | ||||
| 988 | PARTITION_TYPE part_type = sms_tree->partitioning; | |||
| 989 | if (part_type == PARTITION_INVALID) return; | |||
| 990 | ||||
| 991 | if (part_type == PARTITION_SPLIT) { | |||
| 992 | for (int i = 0; i < SUB_PARTITIONS_SPLIT4; ++i) { | |||
| 993 | get_min_bsize(sms_tree->split[i], min_bw, min_bh); | |||
| 994 | } | |||
| 995 | } else { | |||
| 996 | if (part_type == PARTITION_HORZ_A || part_type == PARTITION_HORZ_B || | |||
| 997 | part_type == PARTITION_VERT_A || part_type == PARTITION_VERT_B) | |||
| 998 | part_type = PARTITION_SPLIT; | |||
| 999 | const BLOCK_SIZE subsize = get_partition_subsize(bsize, part_type); | |||
| 1000 | if (subsize != BLOCK_INVALID) { | |||
| 1001 | *min_bw = AOMMIN(*min_bw, mi_size_wide_log2[subsize])(((*min_bw) < (mi_size_wide_log2[subsize])) ? (*min_bw) : ( mi_size_wide_log2[subsize])); | |||
| 1002 | *min_bh = AOMMIN(*min_bh, mi_size_high_log2[subsize])(((*min_bh) < (mi_size_high_log2[subsize])) ? (*min_bh) : ( mi_size_high_log2[subsize])); | |||
| 1003 | } | |||
| 1004 | } | |||
| 1005 | } | |||
| 1006 | ||||
| 1007 | static inline void add_rd_feature(int64_t rd, int64_t best_rd, float *features, | |||
| 1008 | int *feature_idx) { | |||
| 1009 | const int rd_valid = rd > 0 && rd < INT64_MAX(9223372036854775807L); | |||
| 1010 | const float rd_ratio = rd_valid ? (float)rd / best_rd : 1.0f; | |||
| 1011 | features[(*feature_idx)++] = (float)rd_valid; | |||
| 1012 | features[(*feature_idx)++] = rd_ratio; | |||
| 1013 | } | |||
| 1014 | ||||
| 1015 | #define FEATURES 31 | |||
| 1016 | void av1_ml_early_term_after_split(AV1_COMP *const cpi, MACROBLOCK *const x, | |||
| 1017 | SIMPLE_MOTION_DATA_TREE *const sms_tree, | |||
| 1018 | int64_t best_rd, int64_t part_none_rd, | |||
| 1019 | int64_t part_split_rd, | |||
| 1020 | int64_t *split_block_rd, | |||
| 1021 | PartitionSearchState *part_state) { | |||
| 1022 | const PartitionBlkParams *blk_params = &part_state->part_blk_params; | |||
| 1023 | const int mi_row = blk_params->mi_row, mi_col = blk_params->mi_col; | |||
| 1024 | const BLOCK_SIZE bsize = blk_params->bsize; | |||
| 1025 | ||||
| 1026 | if (best_rd <= 0 || best_rd == INT64_MAX(9223372036854775807L) || | |||
| 1027 | part_state->terminate_partition_search) | |||
| 1028 | return; | |||
| 1029 | ||||
| 1030 | const AV1_COMMON *const cm = &cpi->common; | |||
| 1031 | const int is_480p_or_larger = AOMMIN(cm->width, cm->height)(((cm->width) < (cm->height)) ? (cm->width) : (cm ->height)) >= 480; | |||
| 1032 | const NN_CONFIG *nn_config = NULL((void*)0); | |||
| 1033 | float thresh = -1e6; | |||
| 1034 | switch (bsize) { | |||
| 1035 | case BLOCK_128X128: break; | |||
| 1036 | case BLOCK_64X64: | |||
| 1037 | nn_config = &av1_early_term_after_split_nnconfig_64; | |||
| 1038 | thresh = is_480p_or_larger ? -2.0f : -1.2f; | |||
| 1039 | break; | |||
| 1040 | case BLOCK_32X32: | |||
| 1041 | nn_config = &av1_early_term_after_split_nnconfig_32; | |||
| 1042 | thresh = is_480p_or_larger ? -2.6f : -2.3f; | |||
| 1043 | break; | |||
| 1044 | case BLOCK_16X16: | |||
| 1045 | nn_config = &av1_early_term_after_split_nnconfig_16; | |||
| 1046 | thresh = is_480p_or_larger ? -2.0f : -2.4f; | |||
| 1047 | break; | |||
| 1048 | case BLOCK_8X8: | |||
| 1049 | nn_config = &av1_early_term_after_split_nnconfig_8; | |||
| 1050 | thresh = is_480p_or_larger ? -1.0f : -1.4f; | |||
| 1051 | break; | |||
| 1052 | case BLOCK_4X4: break; | |||
| 1053 | default: | |||
| 1054 | assert(0 && "Invalid block size in av1_ml_early_term_after_split().")((void) sizeof ((0 && "Invalid block size in av1_ml_early_term_after_split()." ) ? 1 : 0), __extension__ ({ if (0 && "Invalid block size in av1_ml_early_term_after_split()." ) ; else __assert_fail ("0 && \"Invalid block size in av1_ml_early_term_after_split().\"" , "/root/firefox-clang/third_party/aom/av1/encoder/partition_strategy.c" , 1054, __extension__ __PRETTY_FUNCTION__); })); | |||
| 1055 | break; | |||
| 1056 | } | |||
| 1057 | if (!nn_config) return; | |||
| 1058 | ||||
| 1059 | // Use more conservative threshold for level 1. | |||
| 1060 | if (cpi->sf.part_sf.ml_early_term_after_part_split_level < 2) thresh -= 0.3f; | |||
| 1061 | ||||
| 1062 | const MACROBLOCKD *const xd = &x->e_mbd; | |||
| 1063 | const int dc_q = av1_dc_quant_QTX(x->qindex, 0, xd->bd) >> (xd->bd - 8); | |||
| 1064 | const int bs = block_size_wide[bsize]; | |||
| 1065 | int f_idx = 0; | |||
| 1066 | float features[FEATURES] = { 0.0f }; | |||
| 1067 | ||||
| 1068 | features[f_idx++] = log1pf((float)dc_q / 4.0f); | |||
| 1069 | features[f_idx++] = log1pf((float)best_rd / bs / bs / 1024.0f); | |||
| 1070 | ||||
| 1071 | add_rd_feature(part_none_rd, best_rd, features, &f_idx); | |||
| 1072 | add_rd_feature(part_split_rd, best_rd, features, &f_idx); | |||
| 1073 | ||||
| 1074 | for (int i = 0; i < SUB_PARTITIONS_SPLIT4; ++i) { | |||
| 1075 | add_rd_feature(split_block_rd[i], best_rd, features, &f_idx); | |||
| 1076 | int min_bw = MAX_SB_SIZE_LOG27; | |||
| 1077 | int min_bh = MAX_SB_SIZE_LOG27; | |||
| 1078 | get_min_bsize(sms_tree->split[i], &min_bw, &min_bh); | |||
| 1079 | features[f_idx++] = (float)min_bw; | |||
| 1080 | features[f_idx++] = (float)min_bh; | |||
| 1081 | } | |||
| 1082 | ||||
| 1083 | simple_motion_search_prune_part_features(cpi, x, sms_tree, mi_row, mi_col, | |||
| 1084 | bsize, NULL((void*)0), | |||
| 1085 | FEATURE_SMS_PRUNE_PART_FLAG(1 | (1 << 1) | (1 << 2))); | |||
| 1086 | ||||
| 1087 | features[f_idx++] = log1pf((float)sms_tree->sms_none_feat[1]); | |||
| 1088 | ||||
| 1089 | features[f_idx++] = log1pf((float)sms_tree->split[0]->sms_none_feat[1]); | |||
| 1090 | features[f_idx++] = log1pf((float)sms_tree->split[1]->sms_none_feat[1]); | |||
| 1091 | features[f_idx++] = log1pf((float)sms_tree->split[2]->sms_none_feat[1]); | |||
| 1092 | features[f_idx++] = log1pf((float)sms_tree->split[3]->sms_none_feat[1]); | |||
| 1093 | ||||
| 1094 | features[f_idx++] = log1pf((float)sms_tree->sms_rect_feat[1]); | |||
| 1095 | features[f_idx++] = log1pf((float)sms_tree->sms_rect_feat[3]); | |||
| 1096 | features[f_idx++] = log1pf((float)sms_tree->sms_rect_feat[5]); | |||
| 1097 | features[f_idx++] = log1pf((float)sms_tree->sms_rect_feat[7]); | |||
| 1098 | ||||
| 1099 | assert(f_idx == FEATURES)((void) sizeof ((f_idx == FEATURES) ? 1 : 0), __extension__ ( { if (f_idx == FEATURES) ; else __assert_fail ("f_idx == FEATURES" , "/root/firefox-clang/third_party/aom/av1/encoder/partition_strategy.c" , 1099, __extension__ __PRETTY_FUNCTION__); })); | |||
| 1100 | ||||
| 1101 | // Write features to file | |||
| 1102 | write_features_to_file(cpi->oxcf.partition_info_path, | |||
| 1103 | cpi->ext_part_controller.test_mode, features, FEATURES, | |||
| 1104 | 4, bsize, mi_row, mi_col); | |||
| 1105 | ||||
| 1106 | if (ext_ml_model_decision_after_split( | |||
| 1107 | cpi, features, &part_state->terminate_partition_search)) { | |||
| 1108 | return; | |||
| 1109 | } | |||
| 1110 | ||||
| 1111 | float score = 0.0f; | |||
| 1112 | av1_nn_predict(features, nn_config, 1, &score); | |||
| 1113 | // Score is indicator of confidence that we should NOT terminate. | |||
| 1114 | if (score < thresh) { | |||
| 1115 | part_state->terminate_partition_search = 1; | |||
| 1116 | } | |||
| 1117 | } | |||
| 1118 | #undef FEATURES | |||
| 1119 | ||||
| 1120 | void av1_ml_prune_rect_partition(AV1_COMP *const cpi, const MACROBLOCK *const x, | |||
| 1121 | int64_t best_rd, int64_t none_rd, | |||
| 1122 | const int64_t *split_rd, | |||
| 1123 | PartitionSearchState *part_state) { | |||
| 1124 | const PartitionBlkParams *blk_params = &part_state->part_blk_params; | |||
| 1125 | const int mi_row = blk_params->mi_row, mi_col = blk_params->mi_col; | |||
| 1126 | const BLOCK_SIZE bsize = blk_params->bsize; | |||
| 1127 | ||||
| 1128 | if (bsize < BLOCK_8X8 || best_rd >= 1000000000) return; | |||
| 1129 | best_rd = AOMMAX(best_rd, 1)(((best_rd) > (1)) ? (best_rd) : (1)); | |||
| 1130 | const NN_CONFIG *nn_config = NULL((void*)0); | |||
| 1131 | const float prob_thresholds[5] = { 0.01f, 0.01f, 0.004f, 0.002f, 0.002f }; | |||
| 1132 | float cur_thresh = 0.0f; | |||
| 1133 | switch (bsize) { | |||
| 1134 | case BLOCK_8X8: | |||
| 1135 | nn_config = &av1_rect_partition_nnconfig_8; | |||
| 1136 | cur_thresh = prob_thresholds[0]; | |||
| 1137 | break; | |||
| 1138 | case BLOCK_16X16: | |||
| 1139 | nn_config = &av1_rect_partition_nnconfig_16; | |||
| 1140 | cur_thresh = prob_thresholds[1]; | |||
| 1141 | break; | |||
| 1142 | case BLOCK_32X32: | |||
| 1143 | nn_config = &av1_rect_partition_nnconfig_32; | |||
| 1144 | cur_thresh = prob_thresholds[2]; | |||
| 1145 | break; | |||
| 1146 | case BLOCK_64X64: | |||
| 1147 | nn_config = &av1_rect_partition_nnconfig_64; | |||
| 1148 | cur_thresh = prob_thresholds[3]; | |||
| 1149 | break; | |||
| 1150 | case BLOCK_128X128: | |||
| 1151 | nn_config = &av1_rect_partition_nnconfig_128; | |||
| 1152 | cur_thresh = prob_thresholds[4]; | |||
| 1153 | break; | |||
| 1154 | default: assert(0 && "Unexpected bsize.")((void) sizeof ((0 && "Unexpected bsize.") ? 1 : 0), __extension__ ({ if (0 && "Unexpected bsize.") ; else __assert_fail ("0 && \"Unexpected bsize.\"", "/root/firefox-clang/third_party/aom/av1/encoder/partition_strategy.c" , 1154, __extension__ __PRETTY_FUNCTION__); })); | |||
| 1155 | } | |||
| 1156 | if (!nn_config) return; | |||
| 1157 | ||||
| 1158 | // 1. Compute input features | |||
| 1159 | float features[9]; | |||
| 1160 | ||||
| 1161 | // RD cost ratios | |||
| 1162 | for (int i = 0; i < 5; i++) features[i] = 1.0f; | |||
| 1163 | if (none_rd > 0 && none_rd < 1000000000) | |||
| 1164 | features[0] = (float)none_rd / (float)best_rd; | |||
| 1165 | for (int i = 0; i < SUB_PARTITIONS_SPLIT4; i++) { | |||
| 1166 | if (split_rd[i] > 0 && split_rd[i] < 1000000000) | |||
| 1167 | features[1 + i] = (float)split_rd[i] / (float)best_rd; | |||
| 1168 | } | |||
| 1169 | ||||
| 1170 | // Variance ratios | |||
| 1171 | const MACROBLOCKD *const xd = &x->e_mbd; | |||
| 1172 | int whole_block_variance; | |||
| 1173 | whole_block_variance = av1_get_perpixel_variance_facade( | |||
| 1174 | cpi, xd, &x->plane[0].src, bsize, AOM_PLANE_Y0); | |||
| 1175 | whole_block_variance = AOMMAX(whole_block_variance, 1)(((whole_block_variance) > (1)) ? (whole_block_variance) : (1)); | |||
| 1176 | ||||
| 1177 | int split_variance[SUB_PARTITIONS_SPLIT4]; | |||
| 1178 | const BLOCK_SIZE subsize = get_partition_subsize(bsize, PARTITION_SPLIT); | |||
| 1179 | struct buf_2d buf; | |||
| 1180 | buf.stride = x->plane[0].src.stride; | |||
| 1181 | const int bw = block_size_wide[bsize]; | |||
| 1182 | for (int i = 0; i < SUB_PARTITIONS_SPLIT4; ++i) { | |||
| 1183 | const int x_idx = (i & 1) * bw / 2; | |||
| 1184 | const int y_idx = (i >> 1) * bw / 2; | |||
| 1185 | buf.buf = x->plane[0].src.buf + x_idx + y_idx * buf.stride; | |||
| 1186 | split_variance[i] = | |||
| 1187 | av1_get_perpixel_variance_facade(cpi, xd, &buf, subsize, AOM_PLANE_Y0); | |||
| 1188 | } | |||
| 1189 | ||||
| 1190 | for (int i = 0; i < SUB_PARTITIONS_SPLIT4; i++) | |||
| 1191 | features[5 + i] = (float)split_variance[i] / (float)whole_block_variance; | |||
| 1192 | ||||
| 1193 | // Write features to file | |||
| 1194 | write_features_to_file(cpi->oxcf.partition_info_path, | |||
| 1195 | cpi->ext_part_controller.test_mode, features, | |||
| 1196 | /*feature_size=*/9, 5, bsize, mi_row, mi_col); | |||
| 1197 | ||||
| 1198 | if (ext_ml_model_decision_after_split_part2( | |||
| 1199 | &cpi->ext_part_controller, frame_is_intra_only(&cpi->common), | |||
| 1200 | features, &part_state->prune_rect_part[HORZ], | |||
| 1201 | &part_state->prune_rect_part[VERT])) { | |||
| 1202 | return; | |||
| 1203 | } | |||
| 1204 | ||||
| 1205 | // 2. Do the prediction and prune 0-2 partitions based on their probabilities | |||
| 1206 | float raw_scores[3] = { 0.0f }; | |||
| 1207 | av1_nn_predict(features, nn_config, 1, raw_scores); | |||
| 1208 | float probs[3] = { 0.0f }; | |||
| 1209 | av1_nn_softmax(raw_scores, probs, 3); | |||
| 1210 | ||||
| 1211 | // probs[0] is the probability of the fact that both rectangular partitions | |||
| 1212 | // are worse than current best_rd | |||
| 1213 | if (probs[1] <= cur_thresh) part_state->prune_rect_part[HORZ] = 1; | |||
| 1214 | if (probs[2] <= cur_thresh) part_state->prune_rect_part[VERT] = 1; | |||
| 1215 | } | |||
| 1216 | ||||
| 1217 | // Use a ML model to predict if horz_a, horz_b, vert_a, and vert_b should be | |||
| 1218 | // considered. | |||
| 1219 | static void ml_prune_ab_partition(AV1_COMP *const cpi, int part_ctx, | |||
| 1220 | int var_ctx, int64_t best_rd, | |||
| 1221 | PartitionSearchState *part_state, | |||
| 1222 | int *ab_partitions_allowed) { | |||
| 1223 | const PartitionBlkParams blk_params = part_state->part_blk_params; | |||
| 1224 | const int mi_row = blk_params.mi_row; | |||
| 1225 | const int mi_col = blk_params.mi_col; | |||
| 1226 | const BLOCK_SIZE bsize = blk_params.bsize; | |||
| 1227 | ||||
| 1228 | if (bsize < BLOCK_8X8 || best_rd >= 1000000000) return; | |||
| 1229 | const NN_CONFIG *nn_config = NULL((void*)0); | |||
| 1230 | switch (bsize) { | |||
| 1231 | case BLOCK_8X8: nn_config = NULL((void*)0); break; | |||
| 1232 | case BLOCK_16X16: nn_config = &av1_ab_partition_nnconfig_16; break; | |||
| 1233 | case BLOCK_32X32: nn_config = &av1_ab_partition_nnconfig_32; break; | |||
| 1234 | case BLOCK_64X64: nn_config = &av1_ab_partition_nnconfig_64; break; | |||
| 1235 | case BLOCK_128X128: nn_config = &av1_ab_partition_nnconfig_128; break; | |||
| 1236 | default: assert(0 && "Unexpected bsize.")((void) sizeof ((0 && "Unexpected bsize.") ? 1 : 0), __extension__ ({ if (0 && "Unexpected bsize.") ; else __assert_fail ("0 && \"Unexpected bsize.\"", "/root/firefox-clang/third_party/aom/av1/encoder/partition_strategy.c" , 1236, __extension__ __PRETTY_FUNCTION__); })); | |||
| 1237 | } | |||
| 1238 | if (!nn_config) return; | |||
| 1239 | ||||
| 1240 | // Generate features. | |||
| 1241 | float features[10]; | |||
| 1242 | int feature_index = 0; | |||
| 1243 | features[feature_index++] = (float)part_ctx; | |||
| 1244 | features[feature_index++] = (float)var_ctx; | |||
| 1245 | const int rdcost = (int)AOMMIN(INT_MAX, best_rd)(((2147483647) < (best_rd)) ? (2147483647) : (best_rd)); | |||
| 1246 | int sub_block_rdcost[8] = { 0 }; | |||
| 1247 | int rd_index = 0; | |||
| 1248 | for (int i = 0; i < SUB_PARTITIONS_RECT2; ++i) { | |||
| 1249 | const int64_t *horz_rd = part_state->rect_part_rd[HORZ]; | |||
| 1250 | if (horz_rd[i] > 0 && horz_rd[i] < 1000000000) | |||
| 1251 | sub_block_rdcost[rd_index] = (int)horz_rd[i]; | |||
| 1252 | ++rd_index; | |||
| 1253 | } | |||
| 1254 | for (int i = 0; i < SUB_PARTITIONS_RECT2; ++i) { | |||
| 1255 | const int64_t *vert_rd = part_state->rect_part_rd[VERT]; | |||
| 1256 | if (vert_rd[i] > 0 && vert_rd[i] < 1000000000) | |||
| 1257 | sub_block_rdcost[rd_index] = (int)vert_rd[i]; | |||
| 1258 | ++rd_index; | |||
| 1259 | } | |||
| 1260 | for (int i = 0; i < SUB_PARTITIONS_SPLIT4; ++i) { | |||
| 1261 | const int64_t *split_rd = part_state->split_rd; | |||
| 1262 | if (split_rd[i] > 0 && split_rd[i] < 1000000000) | |||
| 1263 | sub_block_rdcost[rd_index] = (int)split_rd[i]; | |||
| 1264 | ++rd_index; | |||
| 1265 | } | |||
| 1266 | for (int i = 0; i < 8; ++i) { | |||
| 1267 | // Ratio between the sub-block RD and the whole-block RD. | |||
| 1268 | float rd_ratio = 1.0f; | |||
| 1269 | if (sub_block_rdcost[i] > 0 && sub_block_rdcost[i] < rdcost) | |||
| 1270 | rd_ratio = (float)sub_block_rdcost[i] / (float)rdcost; | |||
| 1271 | features[feature_index++] = rd_ratio; | |||
| 1272 | } | |||
| 1273 | assert(feature_index == 10)((void) sizeof ((feature_index == 10) ? 1 : 0), __extension__ ({ if (feature_index == 10) ; else __assert_fail ("feature_index == 10" , "/root/firefox-clang/third_party/aom/av1/encoder/partition_strategy.c" , 1273, __extension__ __PRETTY_FUNCTION__); })); | |||
| 1274 | ||||
| 1275 | // Write features to file | |||
| 1276 | if (!frame_is_intra_only(&cpi->common)) { | |||
| 1277 | write_features_to_file(cpi->oxcf.partition_info_path, | |||
| 1278 | cpi->ext_part_controller.test_mode, features, | |||
| 1279 | /*feature_size=*/10, 6, bsize, mi_row, mi_col); | |||
| 1280 | } | |||
| 1281 | ||||
| 1282 | if (ext_ml_model_decision_after_rect( | |||
| 1283 | &cpi->ext_part_controller, frame_is_intra_only(&cpi->common), | |||
| 1284 | features, &ab_partitions_allowed[HORZ_A], | |||
| 1285 | &ab_partitions_allowed[HORZ_B], &ab_partitions_allowed[VERT_A], | |||
| 1286 | &ab_partitions_allowed[VERT_B])) { | |||
| 1287 | return; | |||
| 1288 | } | |||
| 1289 | ||||
| 1290 | // Calculate scores using the NN model. | |||
| 1291 | float score[16] = { 0.0f }; | |||
| 1292 | av1_nn_predict(features, nn_config, 1, score); | |||
| 1293 | int int_score[16]; | |||
| 1294 | int max_score = -1000; | |||
| 1295 | for (int i = 0; i < 16; ++i) { | |||
| 1296 | int_score[i] = (int)(100 * score[i]); | |||
| 1297 | max_score = AOMMAX(int_score[i], max_score)(((int_score[i]) > (max_score)) ? (int_score[i]) : (max_score )); | |||
| 1298 | } | |||
| 1299 | ||||
| 1300 | // Make decisions based on the model scores. | |||
| 1301 | int thresh = max_score; | |||
| 1302 | switch (bsize) { | |||
| 1303 | case BLOCK_16X16: thresh -= 150; break; | |||
| 1304 | case BLOCK_32X32: thresh -= 100; break; | |||
| 1305 | default: break; | |||
| 1306 | } | |||
| 1307 | av1_zero_array(ab_partitions_allowed, NUM_AB_PARTS)memset(ab_partitions_allowed, 0, NUM_AB_PARTS * sizeof(*(ab_partitions_allowed ))); | |||
| 1308 | for (int i = 0; i < 16; ++i) { | |||
| 1309 | if (int_score[i] >= thresh) { | |||
| 1310 | if ((i >> 0) & 1) ab_partitions_allowed[HORZ_A] = 1; | |||
| 1311 | if ((i >> 1) & 1) ab_partitions_allowed[HORZ_B] = 1; | |||
| 1312 | if ((i >> 2) & 1) ab_partitions_allowed[VERT_A] = 1; | |||
| 1313 | if ((i >> 3) & 1) ab_partitions_allowed[VERT_B] = 1; | |||
| 1314 | } | |||
| 1315 | } | |||
| 1316 | } | |||
| 1317 | ||||
| 1318 | #define FEATURES 18 | |||
| 1319 | #define LABELS 4 | |||
| 1320 | // Use a ML model to predict if horz4 and vert4 should be considered. | |||
| 1321 | void av1_ml_prune_4_partition(AV1_COMP *const cpi, MACROBLOCK *const x, | |||
| 1322 | int part_ctx, int64_t best_rd, | |||
| 1323 | PartitionSearchState *part_state, | |||
| 1324 | int *part4_allowed, | |||
| 1325 | unsigned int pb_source_variance) { | |||
| 1326 | const PartitionBlkParams blk_params = part_state->part_blk_params; | |||
| 1327 | const int mi_row = blk_params.mi_row; | |||
| 1328 | const int mi_col = blk_params.mi_col; | |||
| 1329 | const BLOCK_SIZE bsize = blk_params.bsize; | |||
| 1330 | ||||
| 1331 | int64_t(*rect_part_rd)[SUB_PARTITIONS_RECT2] = part_state->rect_part_rd; | |||
| 1332 | int64_t *split_rd = part_state->split_rd; | |||
| 1333 | if (ext_ml_model_decision_after_part_ab( | |||
| 1334 | cpi, x, bsize, part_ctx, best_rd, rect_part_rd, split_rd, | |||
| 1335 | &part4_allowed[HORZ4], &part4_allowed[VERT4], pb_source_variance, | |||
| 1336 | mi_row, mi_col)) | |||
| 1337 | return; | |||
| 1338 | ||||
| 1339 | if (best_rd >= 1000000000) return; | |||
| 1340 | int64_t *horz_rd = rect_part_rd[HORZ4]; | |||
| 1341 | int64_t *vert_rd = rect_part_rd[VERT4]; | |||
| 1342 | const NN_CONFIG *nn_config = NULL((void*)0); | |||
| 1343 | // 4-way partitions are only allowed for these three square block sizes. | |||
| 1344 | switch (bsize) { | |||
| 1345 | case BLOCK_16X16: nn_config = &av1_4_partition_nnconfig_16; break; | |||
| 1346 | case BLOCK_32X32: nn_config = &av1_4_partition_nnconfig_32; break; | |||
| 1347 | case BLOCK_64X64: nn_config = &av1_4_partition_nnconfig_64; break; | |||
| 1348 | default: assert(0 && "Unexpected bsize.")((void) sizeof ((0 && "Unexpected bsize.") ? 1 : 0), __extension__ ({ if (0 && "Unexpected bsize.") ; else __assert_fail ("0 && \"Unexpected bsize.\"", "/root/firefox-clang/third_party/aom/av1/encoder/partition_strategy.c" , 1348, __extension__ __PRETTY_FUNCTION__); })); | |||
| 1349 | } | |||
| 1350 | if (!nn_config) return; | |||
| 1351 | ||||
| 1352 | // Generate features. | |||
| 1353 | float features[FEATURES]; | |||
| 1354 | int feature_index = 0; | |||
| 1355 | features[feature_index++] = (float)part_ctx; | |||
| 1356 | features[feature_index++] = (float)get_unsigned_bits(pb_source_variance); | |||
| 1357 | ||||
| 1358 | const int rdcost = (int)AOMMIN(INT_MAX, best_rd)(((2147483647) < (best_rd)) ? (2147483647) : (best_rd)); | |||
| 1359 | int sub_block_rdcost[8] = { 0 }; | |||
| 1360 | int rd_index = 0; | |||
| 1361 | for (int i = 0; i < SUB_PARTITIONS_RECT2; ++i) { | |||
| 1362 | if (horz_rd[i] > 0 && horz_rd[i] < 1000000000) | |||
| 1363 | sub_block_rdcost[rd_index] = (int)horz_rd[i]; | |||
| 1364 | ++rd_index; | |||
| 1365 | } | |||
| 1366 | for (int i = 0; i < SUB_PARTITIONS_RECT2; ++i) { | |||
| 1367 | if (vert_rd[i] > 0 && vert_rd[i] < 1000000000) | |||
| 1368 | sub_block_rdcost[rd_index] = (int)vert_rd[i]; | |||
| 1369 | ++rd_index; | |||
| 1370 | } | |||
| 1371 | for (int i = 0; i < SUB_PARTITIONS_SPLIT4; ++i) { | |||
| 1372 | if (split_rd[i] > 0 && split_rd[i] < 1000000000) | |||
| 1373 | sub_block_rdcost[rd_index] = (int)split_rd[i]; | |||
| 1374 | ++rd_index; | |||
| 1375 | } | |||
| 1376 | for (int i = 0; i < 8; ++i) { | |||
| 1377 | // Ratio between the sub-block RD and the whole-block RD. | |||
| 1378 | float rd_ratio = 1.0f; | |||
| 1379 | if (sub_block_rdcost[i] > 0 && sub_block_rdcost[i] < rdcost) | |||
| 1380 | rd_ratio = (float)sub_block_rdcost[i] / (float)rdcost; | |||
| 1381 | features[feature_index++] = rd_ratio; | |||
| 1382 | } | |||
| 1383 | ||||
| 1384 | // Get variance of the 1:4 and 4:1 sub-blocks. | |||
| 1385 | unsigned int horz_4_source_var[SUB_PARTITIONS_PART44] = { 0 }; | |||
| 1386 | unsigned int vert_4_source_var[SUB_PARTITIONS_PART44] = { 0 }; | |||
| 1387 | { | |||
| 1388 | BLOCK_SIZE horz_4_bs = get_partition_subsize(bsize, PARTITION_HORZ_4); | |||
| 1389 | BLOCK_SIZE vert_4_bs = get_partition_subsize(bsize, PARTITION_VERT_4); | |||
| 1390 | ||||
| 1391 | assert(horz_4_bs != BLOCK_INVALID)((void) sizeof ((horz_4_bs != BLOCK_INVALID) ? 1 : 0), __extension__ ({ if (horz_4_bs != BLOCK_INVALID) ; else __assert_fail ("horz_4_bs != BLOCK_INVALID" , "/root/firefox-clang/third_party/aom/av1/encoder/partition_strategy.c" , 1391, __extension__ __PRETTY_FUNCTION__); })); | |||
| 1392 | assert(vert_4_bs != BLOCK_INVALID)((void) sizeof ((vert_4_bs != BLOCK_INVALID) ? 1 : 0), __extension__ ({ if (vert_4_bs != BLOCK_INVALID) ; else __assert_fail ("vert_4_bs != BLOCK_INVALID" , "/root/firefox-clang/third_party/aom/av1/encoder/partition_strategy.c" , 1392, __extension__ __PRETTY_FUNCTION__); })); | |||
| 1393 | ||||
| 1394 | av1_setup_src_planes(x, cpi->source, mi_row, mi_col, | |||
| 1395 | av1_num_planes(&cpi->common), bsize); | |||
| 1396 | const int src_stride = x->plane[0].src.stride; | |||
| 1397 | uint8_t *src = x->plane[0].src.buf; | |||
| 1398 | const MACROBLOCKD *const xd = &x->e_mbd; | |||
| 1399 | ||||
| 1400 | struct buf_2d horz_4_src, vert_4_src; | |||
| 1401 | horz_4_src.stride = src_stride; | |||
| 1402 | vert_4_src.stride = src_stride; | |||
| 1403 | ||||
| 1404 | for (int i = 0; i < SUB_PARTITIONS_PART44; ++i) { | |||
| 1405 | horz_4_src.buf = src + i * block_size_high[horz_4_bs] * src_stride; | |||
| 1406 | vert_4_src.buf = src + i * block_size_wide[vert_4_bs]; | |||
| 1407 | ||||
| 1408 | horz_4_source_var[i] = av1_get_perpixel_variance_facade( | |||
| 1409 | cpi, xd, &horz_4_src, horz_4_bs, AOM_PLANE_Y0); | |||
| 1410 | vert_4_source_var[i] = av1_get_perpixel_variance_facade( | |||
| 1411 | cpi, xd, &vert_4_src, vert_4_bs, AOM_PLANE_Y0); | |||
| 1412 | } | |||
| 1413 | } | |||
| 1414 | ||||
| 1415 | const float denom = (float)(pb_source_variance + 1); | |||
| 1416 | const float low_b = 0.1f; | |||
| 1417 | const float high_b = 10.0f; | |||
| 1418 | for (int i = 0; i < SUB_PARTITIONS_PART44; ++i) { | |||
| 1419 | // Ratio between the 4:1 sub-block variance and the whole-block variance. | |||
| 1420 | float var_ratio = (float)(horz_4_source_var[i] + 1) / denom; | |||
| 1421 | if (var_ratio < low_b) var_ratio = low_b; | |||
| 1422 | if (var_ratio > high_b) var_ratio = high_b; | |||
| 1423 | features[feature_index++] = var_ratio; | |||
| 1424 | } | |||
| 1425 | for (int i = 0; i < SUB_PARTITIONS_PART44; ++i) { | |||
| 1426 | // Ratio between the 1:4 sub-block RD and the whole-block RD. | |||
| 1427 | float var_ratio = (float)(vert_4_source_var[i] + 1) / denom; | |||
| 1428 | if (var_ratio < low_b) var_ratio = low_b; | |||
| 1429 | if (var_ratio > high_b) var_ratio = high_b; | |||
| 1430 | features[feature_index++] = var_ratio; | |||
| 1431 | } | |||
| 1432 | assert(feature_index == FEATURES)((void) sizeof ((feature_index == FEATURES) ? 1 : 0), __extension__ ({ if (feature_index == FEATURES) ; else __assert_fail ("feature_index == FEATURES" , "/root/firefox-clang/third_party/aom/av1/encoder/partition_strategy.c" , 1432, __extension__ __PRETTY_FUNCTION__); })); | |||
| 1433 | ||||
| 1434 | // Write features to file | |||
| 1435 | if (!frame_is_intra_only(&cpi->common)) { | |||
| 1436 | write_features_to_file(cpi->oxcf.partition_info_path, | |||
| 1437 | cpi->ext_part_controller.test_mode, features, | |||
| 1438 | FEATURES, 7, bsize, mi_row, mi_col); | |||
| 1439 | } | |||
| 1440 | ||||
| 1441 | // Calculate scores using the NN model. | |||
| 1442 | float score[LABELS] = { 0.0f }; | |||
| 1443 | av1_nn_predict(features, nn_config, 1, score); | |||
| 1444 | int int_score[LABELS]; | |||
| 1445 | int max_score = -1000; | |||
| 1446 | for (int i = 0; i < LABELS; ++i) { | |||
| 1447 | int_score[i] = (int)(100 * score[i]); | |||
| 1448 | max_score = AOMMAX(int_score[i], max_score)(((int_score[i]) > (max_score)) ? (int_score[i]) : (max_score )); | |||
| 1449 | } | |||
| 1450 | ||||
| 1451 | // Make decisions based on the model scores. | |||
| 1452 | int thresh = max_score; | |||
| 1453 | switch (bsize) { | |||
| 1454 | case BLOCK_16X16: thresh -= 500; break; | |||
| 1455 | case BLOCK_32X32: thresh -= 500; break; | |||
| 1456 | case BLOCK_64X64: thresh -= 200; break; | |||
| 1457 | default: break; | |||
| 1458 | } | |||
| 1459 | av1_zero_array(part4_allowed, NUM_PART4_TYPES)memset(part4_allowed, 0, NUM_PART4_TYPES * sizeof(*(part4_allowed ))); | |||
| 1460 | for (int i = 0; i < LABELS; ++i) { | |||
| 1461 | if (int_score[i] >= thresh) { | |||
| 1462 | if ((i >> 0) & 1) part4_allowed[HORZ4] = 1; | |||
| 1463 | if ((i >> 1) & 1) part4_allowed[VERT4] = 1; | |||
| 1464 | } | |||
| 1465 | } | |||
| 1466 | } | |||
| 1467 | #undef FEATURES | |||
| 1468 | #undef LABELS | |||
| 1469 | ||||
| 1470 | #define FEATURES 4 | |||
| 1471 | void av1_ml_predict_breakout(AV1_COMP *const cpi, const MACROBLOCK *const x, | |||
| 1472 | const RD_STATS *const rd_stats, | |||
| 1473 | unsigned int pb_source_variance, int bit_depth, | |||
| 1474 | PartitionSearchState *part_state) { | |||
| 1475 | const PartitionBlkParams *blk_params = &part_state->part_blk_params; | |||
| 1476 | const int mi_row = blk_params->mi_row, mi_col = blk_params->mi_col; | |||
| 1477 | const BLOCK_SIZE bsize = blk_params->bsize; | |||
| 1478 | ||||
| 1479 | const NN_CONFIG *nn_config = NULL((void*)0); | |||
| 1480 | int thresh = 0; | |||
| 1481 | switch (bsize) { | |||
| 1482 | case BLOCK_8X8: | |||
| 1483 | nn_config = &av1_partition_breakout_nnconfig_8; | |||
| 1484 | thresh = cpi->sf.part_sf.ml_partition_search_breakout_thresh[0]; | |||
| 1485 | break; | |||
| 1486 | case BLOCK_16X16: | |||
| 1487 | nn_config = &av1_partition_breakout_nnconfig_16; | |||
| 1488 | thresh = cpi->sf.part_sf.ml_partition_search_breakout_thresh[1]; | |||
| 1489 | break; | |||
| 1490 | case BLOCK_32X32: | |||
| 1491 | nn_config = &av1_partition_breakout_nnconfig_32; | |||
| 1492 | thresh = cpi->sf.part_sf.ml_partition_search_breakout_thresh[2]; | |||
| 1493 | break; | |||
| 1494 | case BLOCK_64X64: | |||
| 1495 | nn_config = &av1_partition_breakout_nnconfig_64; | |||
| 1496 | thresh = cpi->sf.part_sf.ml_partition_search_breakout_thresh[3]; | |||
| 1497 | break; | |||
| 1498 | case BLOCK_128X128: | |||
| 1499 | nn_config = &av1_partition_breakout_nnconfig_128; | |||
| 1500 | thresh = cpi->sf.part_sf.ml_partition_search_breakout_thresh[4]; | |||
| 1501 | break; | |||
| 1502 | default: assert(0 && "Unexpected bsize.")((void) sizeof ((0 && "Unexpected bsize.") ? 1 : 0), __extension__ ({ if (0 && "Unexpected bsize.") ; else __assert_fail ("0 && \"Unexpected bsize.\"", "/root/firefox-clang/third_party/aom/av1/encoder/partition_strategy.c" , 1502, __extension__ __PRETTY_FUNCTION__); })); | |||
| 1503 | } | |||
| 1504 | if (!nn_config || thresh < 0) return; | |||
| 1505 | ||||
| 1506 | const float ml_predict_breakout_thresh_scale[3] = { 1.15f, 1.05f, 1.0f }; | |||
| 1507 | thresh = (int)((float)thresh * | |||
| 1508 | ml_predict_breakout_thresh_scale | |||
| 1509 | [cpi->sf.part_sf.ml_predict_breakout_level - 1]); | |||
| 1510 | ||||
| 1511 | // Generate feature values. | |||
| 1512 | float features[FEATURES]; | |||
| 1513 | int feature_index = 0; | |||
| 1514 | ||||
| 1515 | const int num_pels_log2 = num_pels_log2_lookup[bsize]; | |||
| 1516 | float rate_f = (float)AOMMIN(rd_stats->rate, INT_MAX)(((rd_stats->rate) < (2147483647)) ? (rd_stats->rate ) : (2147483647)); | |||
| 1517 | rate_f = ((float)x->rdmult / 128.0f / 512.0f / (float)(1 << num_pels_log2)) * | |||
| 1518 | rate_f; | |||
| 1519 | features[feature_index++] = rate_f; | |||
| 1520 | ||||
| 1521 | const float dist_f = | |||
| 1522 | (float)(AOMMIN(rd_stats->dist, INT_MAX)(((rd_stats->dist) < (2147483647)) ? (rd_stats->dist ) : (2147483647)) >> num_pels_log2); | |||
| 1523 | features[feature_index++] = dist_f; | |||
| 1524 | ||||
| 1525 | features[feature_index++] = (float)pb_source_variance; | |||
| 1526 | ||||
| 1527 | const int dc_q = (int)x->plane[0].dequant_QTX[0] >> (bit_depth - 8); | |||
| 1528 | features[feature_index++] = (float)(dc_q * dc_q) / 256.0f; | |||
| 1529 | assert(feature_index == FEATURES)((void) sizeof ((feature_index == FEATURES) ? 1 : 0), __extension__ ({ if (feature_index == FEATURES) ; else __assert_fail ("feature_index == FEATURES" , "/root/firefox-clang/third_party/aom/av1/encoder/partition_strategy.c" , 1529, __extension__ __PRETTY_FUNCTION__); })); | |||
| 1530 | ||||
| 1531 | // Write features to file | |||
| 1532 | write_features_to_file(cpi->oxcf.partition_info_path, | |||
| 1533 | cpi->ext_part_controller.test_mode, features, FEATURES, | |||
| 1534 | 2, bsize, mi_row, mi_col); | |||
| 1535 | ||||
| 1536 | if (ext_ml_model_decision_after_none(&cpi->ext_part_controller, | |||
| 1537 | frame_is_intra_only(&cpi->common), | |||
| 1538 | features, &part_state->do_square_split, | |||
| 1539 | &part_state->do_rectangular_split)) { | |||
| 1540 | return; | |||
| 1541 | } | |||
| 1542 | ||||
| 1543 | // Calculate score using the NN model. | |||
| 1544 | float score = 0.0f; | |||
| 1545 | av1_nn_predict(features, nn_config, 1, &score); | |||
| 1546 | ||||
| 1547 | // Make decision. | |||
| 1548 | if ((int)(score * 100) >= thresh) { | |||
| 1549 | part_state->do_square_split = 0; | |||
| 1550 | part_state->do_rectangular_split = 0; | |||
| 1551 | } | |||
| 1552 | } | |||
| 1553 | #undef FEATURES | |||
| 1554 | ||||
| 1555 | void av1_prune_partitions_before_search(AV1_COMP *const cpi, | |||
| 1556 | MACROBLOCK *const x, | |||
| 1557 | SIMPLE_MOTION_DATA_TREE *const sms_tree, | |||
| 1558 | PartitionSearchState *part_state) { | |||
| 1559 | const AV1_COMMON *const cm = &cpi->common; | |||
| 1560 | const CommonModeInfoParams *const mi_params = &cm->mi_params; | |||
| 1561 | ||||
| 1562 | const PartitionBlkParams *blk_params = &part_state->part_blk_params; | |||
| 1563 | const BLOCK_SIZE bsize = blk_params->bsize; | |||
| 1564 | ||||
| 1565 | #if CONFIG_THREE_PASS0 | |||
| 1566 | if (cpi->third_pass_ctx) { | |||
| 1567 | int mi_row = blk_params->mi_row; | |||
| 1568 | int mi_col = blk_params->mi_col; | |||
| 1569 | double ratio_h, ratio_w; | |||
| 1570 | av1_get_third_pass_ratio(cpi->third_pass_ctx, 0, cm->height, cm->width, | |||
| 1571 | &ratio_h, &ratio_w); | |||
| 1572 | THIRD_PASS_MI_INFO *this_mi = av1_get_third_pass_mi( | |||
| 1573 | cpi->third_pass_ctx, 0, mi_row, mi_col, ratio_h, ratio_w); | |||
| 1574 | BLOCK_SIZE third_pass_bsize = | |||
| 1575 | av1_get_third_pass_adjusted_blk_size(this_mi, ratio_h, ratio_w); | |||
| 1576 | // check the actual partition of this block in the second pass | |||
| 1577 | PARTITION_TYPE third_pass_part = | |||
| 1578 | av1_third_pass_get_sb_part_type(cpi->third_pass_ctx, this_mi); | |||
| 1579 | ||||
| 1580 | int is_edge = (mi_row + mi_size_high[bsize] >= cm->mi_params.mi_rows) || | |||
| 1581 | (mi_col + mi_size_wide[bsize] >= cm->mi_params.mi_cols); | |||
| 1582 | ||||
| 1583 | if (!is_edge && block_size_wide[bsize] >= 16) { | |||
| 1584 | // If in second pass we used rectangular partition, then do not search for | |||
| 1585 | // rectangular partition in the different direction. | |||
| 1586 | if (third_pass_part != PARTITION_NONE) { | |||
| 1587 | if (third_pass_part == PARTITION_HORZ || | |||
| 1588 | third_pass_part == PARTITION_HORZ_4 || | |||
| 1589 | third_pass_part == PARTITION_HORZ_A || | |||
| 1590 | third_pass_part == PARTITION_HORZ_B) { | |||
| 1591 | part_state->partition_rect_allowed[VERT] = 0; | |||
| 1592 | } else if (third_pass_part == PARTITION_VERT || | |||
| 1593 | third_pass_part == PARTITION_VERT_4 || | |||
| 1594 | third_pass_part == PARTITION_VERT_A || | |||
| 1595 | third_pass_part == PARTITION_VERT_B) { | |||
| 1596 | part_state->partition_rect_allowed[HORZ] = 0; | |||
| 1597 | } | |||
| 1598 | } | |||
| 1599 | ||||
| 1600 | int minSize = AOMMIN(block_size_wide[third_pass_bsize],(((block_size_wide[third_pass_bsize]) < (block_size_high[third_pass_bsize ])) ? (block_size_wide[third_pass_bsize]) : (block_size_high[ third_pass_bsize])) | |||
| 1601 | block_size_high[third_pass_bsize])(((block_size_wide[third_pass_bsize]) < (block_size_high[third_pass_bsize ])) ? (block_size_wide[third_pass_bsize]) : (block_size_high[ third_pass_bsize])); | |||
| 1602 | int maxSize = AOMMAX(block_size_wide[third_pass_bsize],(((block_size_wide[third_pass_bsize]) > (block_size_high[third_pass_bsize ])) ? (block_size_wide[third_pass_bsize]) : (block_size_high[ third_pass_bsize])) | |||
| 1603 | block_size_high[third_pass_bsize])(((block_size_wide[third_pass_bsize]) > (block_size_high[third_pass_bsize ])) ? (block_size_wide[third_pass_bsize]) : (block_size_high[ third_pass_bsize])); | |||
| 1604 | if (block_size_wide[bsize] < minSize / 4) { | |||
| 1605 | // Current partition is too small, just terminate | |||
| 1606 | part_state->terminate_partition_search = 1; | |||
| 1607 | return; | |||
| 1608 | } else if (block_size_wide[bsize] < minSize / 2) { | |||
| 1609 | if (third_pass_part != PARTITION_NONE) { | |||
| 1610 | // Current partition is very small, and in second pass we used | |||
| 1611 | // rectangular partition. Terminate the search here then. | |||
| 1612 | part_state->terminate_partition_search = 1; | |||
| 1613 | return; | |||
| 1614 | } else { | |||
| 1615 | // Partition is small, but we still check this partition, only disable | |||
| 1616 | // further splits. | |||
| 1617 | // TODO(any): check why this is not covered by the termination for < | |||
| 1618 | // minSize/4. | |||
| 1619 | av1_disable_square_split_partition(part_state); | |||
| 1620 | av1_disable_rect_partitions(part_state); | |||
| 1621 | return; | |||
| 1622 | } | |||
| 1623 | } else if (block_size_wide[bsize] > maxSize) { | |||
| 1624 | // Partition is larger than in the second pass. Only allow split. | |||
| 1625 | av1_set_square_split_only(part_state); | |||
| 1626 | return; | |||
| 1627 | } else if (block_size_wide[bsize] >= minSize && | |||
| 1628 | block_size_wide[bsize] <= maxSize) { | |||
| 1629 | // Partition is within a range where it is very likely to find a good | |||
| 1630 | // choice, so do not prune anything. | |||
| 1631 | return; | |||
| 1632 | } | |||
| 1633 | } | |||
| 1634 | } | |||
| 1635 | #endif // CONFIG_THREE_PASS | |||
| 1636 | ||||
| 1637 | // Prune rectangular partitions for larger blocks. | |||
| 1638 | if (bsize > cpi->sf.part_sf.rect_partition_eval_thresh) { | |||
| 1639 | part_state->do_rectangular_split = 0; | |||
| 1640 | part_state->partition_rect_allowed[HORZ] = 0; | |||
| 1641 | part_state->partition_rect_allowed[VERT] = 0; | |||
| 1642 | } | |||
| 1643 | ||||
| 1644 | // Prune rectangular, AB and 4-way partition based on q index and block size | |||
| 1645 | if (cpi->sf.part_sf.prune_rectangular_split_based_on_qidx == 1) { | |||
| 1646 | if (bsize == BLOCK_8X8 && x->qindex < 35) | |||
| 1647 | av1_disable_rect_partitions(part_state); | |||
| 1648 | ||||
| 1649 | } else if (cpi->sf.part_sf.prune_rectangular_split_based_on_qidx == 2) { | |||
| 1650 | // Enumeration difference between two square partitions | |||
| 1651 | const int sqr_bsize_step = BLOCK_32X32 - BLOCK_16X16; | |||
| 1652 | int max_bsize = | |||
| 1653 | BLOCK_32X32 - (x->qindex * 3 / QINDEX_RANGE(255 - 0 + 1)) * sqr_bsize_step; | |||
| 1654 | max_bsize = AOMMAX(max_bsize, BLOCK_4X4)(((max_bsize) > (BLOCK_4X4)) ? (max_bsize) : (BLOCK_4X4)); | |||
| 1655 | const BLOCK_SIZE max_prune_bsize = | |||
| 1656 | (BLOCK_SIZE)AOMMIN(max_bsize, BLOCK_32X32)(((max_bsize) < (BLOCK_32X32)) ? (max_bsize) : (BLOCK_32X32 )); | |||
| 1657 | ||||
| 1658 | // Prune partition | |||
| 1659 | // qidx 0 to 85: prune bsize below BLOCK_32X32 | |||
| 1660 | // qidx 86 to 170: prune bsize below BLOCK_16X16 | |||
| 1661 | // qidx 171 to 255: prune bsize below BLOCK_8X8 | |||
| 1662 | if (bsize < max_prune_bsize) { | |||
| 1663 | av1_disable_rect_partitions(part_state); | |||
| 1664 | } | |||
| 1665 | } | |||
| 1666 | ||||
| 1667 | if (cpi->sf.part_sf.prune_sub_8x8_partition_level && (bsize == BLOCK_8X8)) { | |||
| 1668 | const MACROBLOCKD *const xd = &x->e_mbd; | |||
| 1669 | int prune_sub_8x8; | |||
| 1670 | if (cpi->sf.part_sf.prune_sub_8x8_partition_level == 2) { | |||
| 1671 | prune_sub_8x8 = 1; | |||
| 1672 | } else { | |||
| 1673 | assert(cpi->sf.part_sf.prune_sub_8x8_partition_level == 1)((void) sizeof ((cpi->sf.part_sf.prune_sub_8x8_partition_level == 1) ? 1 : 0), __extension__ ({ if (cpi->sf.part_sf.prune_sub_8x8_partition_level == 1) ; else __assert_fail ("cpi->sf.part_sf.prune_sub_8x8_partition_level == 1" , "/root/firefox-clang/third_party/aom/av1/encoder/partition_strategy.c" , 1673, __extension__ __PRETTY_FUNCTION__); })); | |||
| 1674 | // Prune if both neighbors are available and either is > BLOCK_8X8 | |||
| 1675 | prune_sub_8x8 = xd->left_available && xd->up_available && | |||
| 1676 | (xd->left_mbmi->bsize > BLOCK_8X8 || | |||
| 1677 | xd->above_mbmi->bsize > BLOCK_8X8); | |||
| 1678 | } | |||
| 1679 | if (prune_sub_8x8) { | |||
| 1680 | av1_disable_all_splits(part_state); | |||
| 1681 | } | |||
| 1682 | } | |||
| 1683 | ||||
| 1684 | // A CNN-based speed feature pruning out either split or all non-split | |||
| 1685 | // partition in INTRA frame coding. | |||
| 1686 | const int try_intra_cnn_based_part_prune = | |||
| 1687 | frame_is_intra_only(cm) && | |||
| 1688 | cpi->sf.part_sf.intra_cnn_based_part_prune_level && | |||
| 1689 | cm->seq_params->sb_size >= BLOCK_64X64 && bsize <= BLOCK_64X64 && | |||
| 1690 | blk_params->bsize_at_least_8x8 && | |||
| 1691 | av1_is_whole_blk_in_frame(blk_params, mi_params); | |||
| 1692 | ||||
| 1693 | if (try_intra_cnn_based_part_prune) { | |||
| 1694 | intra_mode_cnn_partition(&cpi->common, x, x->part_search_info.quad_tree_idx, | |||
| 1695 | cpi->sf.part_sf.intra_cnn_based_part_prune_level, | |||
| 1696 | part_state); | |||
| 1697 | } | |||
| 1698 | ||||
| 1699 | // Use simple motion search to prune out split or non-split partitions. This | |||
| 1700 | // must be done prior to PARTITION_SPLIT to propagate the initial mvs to a | |||
| 1701 | // smaller blocksize. | |||
| 1702 | const int try_split_only = | |||
| 1703 | cpi->sf.part_sf.simple_motion_search_split && | |||
| 1704 | part_state->do_square_split && blk_params->bsize_at_least_8x8 && | |||
| 1705 | av1_is_whole_blk_in_frame(blk_params, mi_params) && | |||
| 1706 | !frame_is_intra_only(cm) && !av1_superres_scaled(cm); | |||
| 1707 | ||||
| 1708 | if (try_split_only) { | |||
| 1709 | simple_motion_search_based_split(cpi, x, sms_tree, part_state); | |||
| 1710 | } | |||
| 1711 | ||||
| 1712 | // Use simple motion search to prune out rectangular partition in some | |||
| 1713 | // direction. The results are stored in prune_horz and prune_vert in order to | |||
| 1714 | // bypass future related pruning checks if a pruning decision has been made. | |||
| 1715 | ||||
| 1716 | // We want to search at least one partition mode, so don't prune if NONE and | |||
| 1717 | // SPLIT are disabled. | |||
| 1718 | const int non_rect_part_allowed = | |||
| 1719 | part_state->do_square_split || part_state->partition_none_allowed; | |||
| 1720 | // Only run the model if the partitions are not already pruned. | |||
| 1721 | const int rect_part_allowed = part_state->do_rectangular_split && | |||
| 1722 | ((part_state->partition_rect_allowed[HORZ] && | |||
| 1723 | !part_state->prune_rect_part[HORZ]) || | |||
| 1724 | (part_state->partition_rect_allowed[VERT] && | |||
| 1725 | !part_state->prune_rect_part[VERT])); | |||
| 1726 | ||||
| 1727 | const int try_prune_rect = cpi->sf.part_sf.simple_motion_search_prune_rect && | |||
| 1728 | !frame_is_intra_only(cm) && | |||
| 1729 | non_rect_part_allowed && rect_part_allowed && | |||
| 1730 | !av1_superres_scaled(cm); | |||
| 1731 | ||||
| 1732 | if (try_prune_rect) { | |||
| 1733 | simple_motion_search_prune_rect(cpi, x, sms_tree, part_state); | |||
| 1734 | } | |||
| 1735 | } | |||
| 1736 | ||||
| 1737 | #ifndef NDEBUG | |||
| 1738 | static inline int is_bsize_square(BLOCK_SIZE bsize) { | |||
| 1739 | return block_size_wide[bsize] == block_size_high[bsize]; | |||
| 1740 | } | |||
| 1741 | #endif // NDEBUG | |||
| 1742 | ||||
| 1743 | void av1_prune_partitions_by_max_min_bsize(SuperBlockEnc *sb_enc, | |||
| 1744 | PartitionSearchState *part_state) { | |||
| 1745 | assert(is_bsize_square(sb_enc->max_partition_size))((void) sizeof ((is_bsize_square(sb_enc->max_partition_size )) ? 1 : 0), __extension__ ({ if (is_bsize_square(sb_enc-> max_partition_size)) ; else __assert_fail ("is_bsize_square(sb_enc->max_partition_size)" , "/root/firefox-clang/third_party/aom/av1/encoder/partition_strategy.c" , 1745, __extension__ __PRETTY_FUNCTION__); })); | |||
| 1746 | assert(is_bsize_square(sb_enc->min_partition_size))((void) sizeof ((is_bsize_square(sb_enc->min_partition_size )) ? 1 : 0), __extension__ ({ if (is_bsize_square(sb_enc-> min_partition_size)) ; else __assert_fail ("is_bsize_square(sb_enc->min_partition_size)" , "/root/firefox-clang/third_party/aom/av1/encoder/partition_strategy.c" , 1746, __extension__ __PRETTY_FUNCTION__); })); | |||
| 1747 | assert(sb_enc->min_partition_size <= sb_enc->max_partition_size)((void) sizeof ((sb_enc->min_partition_size <= sb_enc-> max_partition_size) ? 1 : 0), __extension__ ({ if (sb_enc-> min_partition_size <= sb_enc->max_partition_size) ; else __assert_fail ("sb_enc->min_partition_size <= sb_enc->max_partition_size" , "/root/firefox-clang/third_party/aom/av1/encoder/partition_strategy.c" , 1747, __extension__ __PRETTY_FUNCTION__); })); | |||
| 1748 | const PartitionBlkParams *blk_params = &part_state->part_blk_params; | |||
| 1749 | const BLOCK_SIZE bsize = blk_params->bsize; | |||
| 1750 | assert(is_bsize_square(bsize))((void) sizeof ((is_bsize_square(bsize)) ? 1 : 0), __extension__ ({ if (is_bsize_square(bsize)) ; else __assert_fail ("is_bsize_square(bsize)" , "/root/firefox-clang/third_party/aom/av1/encoder/partition_strategy.c" , 1750, __extension__ __PRETTY_FUNCTION__); })); | |||
| 1751 | const int max_partition_size_1d = block_size_wide[sb_enc->max_partition_size]; | |||
| 1752 | const int min_partition_size_1d = block_size_wide[sb_enc->min_partition_size]; | |||
| 1753 | const int bsize_1d = block_size_wide[bsize]; | |||
| 1754 | assert(min_partition_size_1d <= max_partition_size_1d)((void) sizeof ((min_partition_size_1d <= max_partition_size_1d ) ? 1 : 0), __extension__ ({ if (min_partition_size_1d <= max_partition_size_1d ) ; else __assert_fail ("min_partition_size_1d <= max_partition_size_1d" , "/root/firefox-clang/third_party/aom/av1/encoder/partition_strategy.c" , 1754, __extension__ __PRETTY_FUNCTION__); })); | |||
| 1755 | const int is_le_min_sq_part = bsize_1d <= min_partition_size_1d; | |||
| 1756 | const int is_gt_max_sq_part = bsize_1d > max_partition_size_1d; | |||
| 1757 | if (is_gt_max_sq_part) { | |||
| 1758 | // If current block size is larger than max, only allow split. | |||
| 1759 | av1_set_square_split_only(part_state); | |||
| 1760 | } else if (is_le_min_sq_part) { | |||
| 1761 | // If current block size is less or equal to min, only allow none if valid | |||
| 1762 | // block large enough; only allow split otherwise. | |||
| 1763 | av1_disable_rect_partitions(part_state); | |||
| 1764 | ||||
| 1765 | // only disable square split when current block is not at the picture | |||
| 1766 | // boundary. otherwise, inherit the square split flag from previous logic | |||
| 1767 | if (av1_blk_has_rows_and_cols(blk_params)) { | |||
| 1768 | part_state->do_square_split = 0; | |||
| 1769 | } | |||
| 1770 | part_state->partition_none_allowed = !(part_state->do_square_split); | |||
| 1771 | } | |||
| 1772 | } | |||
| 1773 | ||||
| 1774 | // Decide whether to evaluate the AB partition specified by part_type based on | |||
| 1775 | // split and HORZ/VERT info | |||
| 1776 | static int evaluate_ab_partition_based_on_split( | |||
| 1777 | const PC_TREE *pc_tree, PARTITION_TYPE rect_part, | |||
| 1778 | const RD_RECT_PART_WIN_INFO *rect_part_win_info, int qindex, int split_idx1, | |||
| 1779 | int split_idx2) { | |||
| 1780 | int num_win = 0; | |||
| 1781 | // Threshold for number of winners | |||
| 1782 | // Conservative pruning for high quantizers | |||
| 1783 | const int num_win_thresh = AOMMIN(3 * (2 * (MAXQ - qindex) / MAXQ), 3)(((3 * (2 * (255 - qindex) / 255)) < (3)) ? (3 * (2 * (255 - qindex) / 255)) : (3)); | |||
| 1784 | int sub_part_win = | |||
| 1785 | (rect_part_win_info == NULL((void*)0)) ? (pc_tree->partitioning == rect_part) | |||
| 1786 | : (rect_part == PARTITION_HORZ) ? rect_part_win_info->rect_part_win[HORZ] | |||
| 1787 | : rect_part_win_info->rect_part_win[VERT]; | |||
| 1788 | num_win += (sub_part_win) ? 1 : 0; | |||
| 1789 | if (pc_tree->split[split_idx1]) { | |||
| 1790 | num_win += | |||
| 1791 | (pc_tree->split[split_idx1]->partitioning == PARTITION_NONE) ? 1 : 0; | |||
| 1792 | } else { | |||
| 1793 | num_win += 1; | |||
| 1794 | } | |||
| 1795 | if (pc_tree->split[split_idx2]) { | |||
| 1796 | num_win += | |||
| 1797 | (pc_tree->split[split_idx2]->partitioning == PARTITION_NONE) ? 1 : 0; | |||
| 1798 | } else { | |||
| 1799 | num_win += 1; | |||
| 1800 | } | |||
| 1801 | if (num_win < num_win_thresh) { | |||
| 1802 | return 0; | |||
| 1803 | } | |||
| 1804 | return 1; | |||
| 1805 | } | |||
| 1806 | ||||
| 1807 | void av1_prune_ab_partitions(AV1_COMP *cpi, const MACROBLOCK *x, | |||
| 1808 | const PC_TREE *pc_tree, int pb_source_variance, | |||
| 1809 | int64_t best_rdcost, | |||
| 1810 | const RD_RECT_PART_WIN_INFO *rect_part_win_info, | |||
| 1811 | bool_Bool ext_partition_allowed, | |||
| 1812 | PartitionSearchState *part_state, | |||
| 1813 | int *ab_partitions_allowed) { | |||
| 1814 | int64_t *horz_rd = part_state->rect_part_rd[HORZ]; | |||
| 1815 | int64_t *vert_rd = part_state->rect_part_rd[VERT]; | |||
| 1816 | int64_t *split_rd = part_state->split_rd; | |||
| 1817 | const PartitionCfg *const part_cfg = &cpi->oxcf.part_cfg; | |||
| 1818 | // The standard AB partitions are allowed initially if ext-partition-types are | |||
| 1819 | // allowed. | |||
| 1820 | int horzab_partition_allowed = ext_partition_allowed && | |||
| 1821 | part_cfg->enable_ab_partitions && | |||
| 1822 | part_state->partition_rect_allowed[HORZ]; | |||
| 1823 | int vertab_partition_allowed = ext_partition_allowed && | |||
| 1824 | part_cfg->enable_ab_partitions && | |||
| 1825 | part_state->partition_rect_allowed[VERT]; | |||
| 1826 | ||||
| 1827 | // Pruning: pruning out AB partitions on one main direction based on the | |||
| 1828 | // current best partition and source variance. | |||
| 1829 | if (cpi->sf.part_sf.prune_ext_partition_types_search_level) { | |||
| 1830 | if (cpi->sf.part_sf.prune_ext_partition_types_search_level == 1) { | |||
| 1831 | // TODO(debargha,huisu@google.com): may need to tune the threshold for | |||
| 1832 | // pb_source_variance. | |||
| 1833 | horzab_partition_allowed &= (pc_tree->partitioning == PARTITION_HORZ || | |||
| 1834 | (pc_tree->partitioning == PARTITION_NONE && | |||
| 1835 | pb_source_variance < 32) || | |||
| 1836 | pc_tree->partitioning == PARTITION_SPLIT); | |||
| 1837 | vertab_partition_allowed &= (pc_tree->partitioning == PARTITION_VERT || | |||
| 1838 | (pc_tree->partitioning == PARTITION_NONE && | |||
| 1839 | pb_source_variance < 32) || | |||
| 1840 | pc_tree->partitioning == PARTITION_SPLIT); | |||
| 1841 | } else { | |||
| 1842 | horzab_partition_allowed &= (pc_tree->partitioning == PARTITION_HORZ || | |||
| 1843 | pc_tree->partitioning == PARTITION_SPLIT); | |||
| 1844 | vertab_partition_allowed &= (pc_tree->partitioning == PARTITION_VERT || | |||
| 1845 | pc_tree->partitioning == PARTITION_SPLIT); | |||
| 1846 | } | |||
| 1847 | horz_rd[0] = (horz_rd[0] < INT64_MAX(9223372036854775807L) ? horz_rd[0] : 0); | |||
| 1848 | horz_rd[1] = (horz_rd[1] < INT64_MAX(9223372036854775807L) ? horz_rd[1] : 0); | |||
| 1849 | vert_rd[0] = (vert_rd[0] < INT64_MAX(9223372036854775807L) ? vert_rd[0] : 0); | |||
| 1850 | vert_rd[1] = (vert_rd[1] < INT64_MAX(9223372036854775807L) ? vert_rd[1] : 0); | |||
| 1851 | split_rd[0] = (split_rd[0] < INT64_MAX(9223372036854775807L) ? split_rd[0] : 0); | |||
| 1852 | split_rd[1] = (split_rd[1] < INT64_MAX(9223372036854775807L) ? split_rd[1] : 0); | |||
| 1853 | split_rd[2] = (split_rd[2] < INT64_MAX(9223372036854775807L) ? split_rd[2] : 0); | |||
| 1854 | split_rd[3] = (split_rd[3] < INT64_MAX(9223372036854775807L) ? split_rd[3] : 0); | |||
| 1855 | } | |||
| 1856 | ||||
| 1857 | // Pruning: pruning out horz_a or horz_b if the combined rdcost of its | |||
| 1858 | // subblocks estimated from previous partitions is much higher than the best | |||
| 1859 | // rd so far. | |||
| 1860 | ab_partitions_allowed[HORZ_A] = horzab_partition_allowed; | |||
| 1861 | ab_partitions_allowed[HORZ_B] = horzab_partition_allowed; | |||
| 1862 | if (cpi->sf.part_sf.prune_ext_partition_types_search_level) { | |||
| 1863 | const int64_t horz_a_rd = horz_rd[1] + split_rd[0] + split_rd[1]; | |||
| 1864 | const int64_t horz_b_rd = horz_rd[0] + split_rd[2] + split_rd[3]; | |||
| 1865 | switch (cpi->sf.part_sf.prune_ext_partition_types_search_level) { | |||
| 1866 | case 1: | |||
| 1867 | ab_partitions_allowed[HORZ_A] &= (horz_a_rd / 16 * 14 < best_rdcost); | |||
| 1868 | ab_partitions_allowed[HORZ_B] &= (horz_b_rd / 16 * 14 < best_rdcost); | |||
| 1869 | break; | |||
| 1870 | case 2: | |||
| 1871 | default: | |||
| 1872 | ab_partitions_allowed[HORZ_A] &= (horz_a_rd / 16 * 15 < best_rdcost); | |||
| 1873 | ab_partitions_allowed[HORZ_B] &= (horz_b_rd / 16 * 15 < best_rdcost); | |||
| 1874 | break; | |||
| 1875 | } | |||
| 1876 | } | |||
| 1877 | ||||
| 1878 | // Pruning: pruning out vert_a or vert_b if the combined rdcost of its | |||
| 1879 | // subblocks estimated from previous partitions is much higher than the best | |||
| 1880 | // rd so far. | |||
| 1881 | ab_partitions_allowed[VERT_A] = vertab_partition_allowed; | |||
| 1882 | ab_partitions_allowed[VERT_B] = vertab_partition_allowed; | |||
| 1883 | if (cpi->sf.part_sf.prune_ext_partition_types_search_level) { | |||
| 1884 | const int64_t vert_a_rd = vert_rd[1] + split_rd[0] + split_rd[2]; | |||
| 1885 | const int64_t vert_b_rd = vert_rd[0] + split_rd[1] + split_rd[3]; | |||
| 1886 | switch (cpi->sf.part_sf.prune_ext_partition_types_search_level) { | |||
| 1887 | case 1: | |||
| 1888 | ab_partitions_allowed[VERT_A] &= (vert_a_rd / 16 * 14 < best_rdcost); | |||
| 1889 | ab_partitions_allowed[VERT_B] &= (vert_b_rd / 16 * 14 < best_rdcost); | |||
| 1890 | break; | |||
| 1891 | case 2: | |||
| 1892 | default: | |||
| 1893 | ab_partitions_allowed[VERT_A] &= (vert_a_rd / 16 * 15 < best_rdcost); | |||
| 1894 | ab_partitions_allowed[VERT_B] &= (vert_b_rd / 16 * 15 < best_rdcost); | |||
| 1895 | break; | |||
| 1896 | } | |||
| 1897 | } | |||
| 1898 | ||||
| 1899 | // Pruning: pruning out some ab partitions using a DNN taking rd costs of | |||
| 1900 | // sub-blocks from previous basic partition types. | |||
| 1901 | if (cpi->sf.part_sf.ml_prune_partition && ext_partition_allowed && | |||
| 1902 | part_state->partition_rect_allowed[HORZ] && | |||
| 1903 | part_state->partition_rect_allowed[VERT]) { | |||
| 1904 | // TODO(huisu@google.com): x->source_variance may not be the current | |||
| 1905 | // block's variance. The correct one to use is pb_source_variance. Need to | |||
| 1906 | // re-train the model to fix it. | |||
| 1907 | ml_prune_ab_partition(cpi, pc_tree->partitioning, | |||
| 1908 | get_unsigned_bits(x->source_variance), best_rdcost, | |||
| 1909 | part_state, ab_partitions_allowed); | |||
| 1910 | } | |||
| 1911 | ||||
| 1912 | // Pruning: pruning AB partitions based on the number of horz/vert wins | |||
| 1913 | // in the current block and sub-blocks in PARTITION_SPLIT. | |||
| 1914 | if (cpi->sf.part_sf.prune_ext_part_using_split_info >= 2 && | |||
| 1915 | ab_partitions_allowed[HORZ_A]) { | |||
| 1916 | ab_partitions_allowed[HORZ_A] &= evaluate_ab_partition_based_on_split( | |||
| 1917 | pc_tree, PARTITION_HORZ, rect_part_win_info, x->qindex, 0, 1); | |||
| 1918 | } | |||
| 1919 | if (cpi->sf.part_sf.prune_ext_part_using_split_info >= 2 && | |||
| 1920 | ab_partitions_allowed[HORZ_B]) { | |||
| 1921 | ab_partitions_allowed[HORZ_B] &= evaluate_ab_partition_based_on_split( | |||
| 1922 | pc_tree, PARTITION_HORZ, rect_part_win_info, x->qindex, 2, 3); | |||
| 1923 | } | |||
| 1924 | if (cpi->sf.part_sf.prune_ext_part_using_split_info >= 2 && | |||
| 1925 | ab_partitions_allowed[VERT_A]) { | |||
| 1926 | ab_partitions_allowed[VERT_A] &= evaluate_ab_partition_based_on_split( | |||
| 1927 | pc_tree, PARTITION_VERT, rect_part_win_info, x->qindex, 0, 2); | |||
| 1928 | } | |||
| 1929 | if (cpi->sf.part_sf.prune_ext_part_using_split_info >= 2 && | |||
| 1930 | ab_partitions_allowed[VERT_B]) { | |||
| 1931 | ab_partitions_allowed[VERT_B] &= evaluate_ab_partition_based_on_split( | |||
| 1932 | pc_tree, PARTITION_VERT, rect_part_win_info, x->qindex, 1, 3); | |||
| 1933 | } | |||
| 1934 | } | |||
| 1935 | ||||
| 1936 | // Prepare features for the external model. Specifically, features after | |||
| 1937 | // ab partition is searched. | |||
| 1938 | static void prepare_features_after_part_ab( | |||
| 1939 | const AV1_COMP *const cpi, MACROBLOCK *const x, BLOCK_SIZE bsize, | |||
| 1940 | int part_ctx, int64_t best_rd, | |||
| 1941 | int64_t rect_part_rd[NUM_RECT_PARTS][SUB_PARTITIONS_RECT2], | |||
| 1942 | int64_t split_rd[SUB_PARTITIONS_SPLIT4], unsigned int pb_source_variance, | |||
| 1943 | int mi_row, int mi_col, aom_partition_features_t *const features) { | |||
| 1944 | int64_t *horz_rd = rect_part_rd[HORZ]; | |||
| 1945 | int64_t *vert_rd = rect_part_rd[VERT]; | |||
| 1946 | ||||
| 1947 | // Generate features. | |||
| 1948 | int feature_index = 0; | |||
| 1949 | features->after_part_ab.f[feature_index++] = (float)part_ctx; | |||
| 1950 | features->after_part_ab.f[feature_index++] = | |||
| 1951 | (float)get_unsigned_bits(pb_source_variance); | |||
| 1952 | ||||
| 1953 | const int rdcost = (int)AOMMIN(INT_MAX, best_rd)(((2147483647) < (best_rd)) ? (2147483647) : (best_rd)); | |||
| 1954 | int sub_block_rdcost[8] = { 0 }; | |||
| 1955 | int rd_index = 0; | |||
| 1956 | for (int i = 0; i < SUB_PARTITIONS_RECT2; ++i) { | |||
| 1957 | if (horz_rd[i] > 0 && horz_rd[i] < 1000000000) | |||
| 1958 | sub_block_rdcost[rd_index] = (int)horz_rd[i]; | |||
| 1959 | ++rd_index; | |||
| 1960 | } | |||
| 1961 | for (int i = 0; i < SUB_PARTITIONS_RECT2; ++i) { | |||
| 1962 | if (vert_rd[i] > 0 && vert_rd[i] < 1000000000) | |||
| 1963 | sub_block_rdcost[rd_index] = (int)vert_rd[i]; | |||
| 1964 | ++rd_index; | |||
| 1965 | } | |||
| 1966 | for (int i = 0; i < SUB_PARTITIONS_SPLIT4; ++i) { | |||
| 1967 | if (split_rd[i] > 0 && split_rd[i] < 1000000000) | |||
| 1968 | sub_block_rdcost[rd_index] = (int)split_rd[i]; | |||
| 1969 | ++rd_index; | |||
| 1970 | } | |||
| 1971 | for (int i = 0; i < 8; ++i) { | |||
| 1972 | // Ratio between the sub-block RD and the whole-block RD. | |||
| 1973 | float rd_ratio = 1.0f; | |||
| 1974 | if (sub_block_rdcost[i] > 0 && sub_block_rdcost[i] < rdcost) | |||
| 1975 | rd_ratio = (float)sub_block_rdcost[i] / (float)rdcost; | |||
| 1976 | features->after_part_ab.f[feature_index++] = rd_ratio; | |||
| 1977 | } | |||
| 1978 | ||||
| 1979 | // 4-way partitions are only allowed for these three square block sizes. | |||
| 1980 | assert(bsize == BLOCK_16X16 || bsize == BLOCK_32X32 || bsize == BLOCK_64X64)((void) sizeof ((bsize == BLOCK_16X16 || bsize == BLOCK_32X32 || bsize == BLOCK_64X64) ? 1 : 0), __extension__ ({ if (bsize == BLOCK_16X16 || bsize == BLOCK_32X32 || bsize == BLOCK_64X64 ) ; else __assert_fail ("bsize == BLOCK_16X16 || bsize == BLOCK_32X32 || bsize == BLOCK_64X64" , "/root/firefox-clang/third_party/aom/av1/encoder/partition_strategy.c" , 1980, __extension__ __PRETTY_FUNCTION__); })); | |||
| 1981 | ||||
| 1982 | // Get variance of the 1:4 and 4:1 sub-blocks. | |||
| 1983 | unsigned int horz_4_source_var[SUB_PARTITIONS_PART44] = { 0 }; | |||
| 1984 | unsigned int vert_4_source_var[SUB_PARTITIONS_PART44] = { 0 }; | |||
| 1985 | { | |||
| 1986 | BLOCK_SIZE horz_4_bs = get_partition_subsize(bsize, PARTITION_HORZ_4); | |||
| 1987 | BLOCK_SIZE vert_4_bs = get_partition_subsize(bsize, PARTITION_VERT_4); | |||
| 1988 | ||||
| 1989 | assert(horz_4_bs != BLOCK_INVALID)((void) sizeof ((horz_4_bs != BLOCK_INVALID) ? 1 : 0), __extension__ ({ if (horz_4_bs != BLOCK_INVALID) ; else __assert_fail ("horz_4_bs != BLOCK_INVALID" , "/root/firefox-clang/third_party/aom/av1/encoder/partition_strategy.c" , 1989, __extension__ __PRETTY_FUNCTION__); })); | |||
| 1990 | assert(vert_4_bs != BLOCK_INVALID)((void) sizeof ((vert_4_bs != BLOCK_INVALID) ? 1 : 0), __extension__ ({ if (vert_4_bs != BLOCK_INVALID) ; else __assert_fail ("vert_4_bs != BLOCK_INVALID" , "/root/firefox-clang/third_party/aom/av1/encoder/partition_strategy.c" , 1990, __extension__ __PRETTY_FUNCTION__); })); | |||
| 1991 | ||||
| 1992 | av1_setup_src_planes(x, cpi->source, mi_row, mi_col, | |||
| 1993 | av1_num_planes(&cpi->common), bsize); | |||
| 1994 | const int src_stride = x->plane[0].src.stride; | |||
| 1995 | uint8_t *src = x->plane[0].src.buf; | |||
| 1996 | const MACROBLOCKD *const xd = &x->e_mbd; | |||
| 1997 | ||||
| 1998 | struct buf_2d horz_4_src, vert_4_src; | |||
| 1999 | horz_4_src.stride = src_stride; | |||
| 2000 | vert_4_src.stride = src_stride; | |||
| 2001 | ||||
| 2002 | for (int i = 0; i < SUB_PARTITIONS_PART44; ++i) { | |||
| 2003 | horz_4_src.buf = src + i * block_size_high[horz_4_bs] * src_stride; | |||
| 2004 | vert_4_src.buf = src + i * block_size_wide[vert_4_bs]; | |||
| 2005 | ||||
| 2006 | horz_4_source_var[i] = av1_get_perpixel_variance_facade( | |||
| 2007 | cpi, xd, &horz_4_src, horz_4_bs, AOM_PLANE_Y0); | |||
| 2008 | vert_4_source_var[i] = av1_get_perpixel_variance_facade( | |||
| 2009 | cpi, xd, &vert_4_src, vert_4_bs, AOM_PLANE_Y0); | |||
| 2010 | } | |||
| 2011 | } | |||
| 2012 | ||||
| 2013 | const float denom = (float)(pb_source_variance + 1); | |||
| 2014 | const float low_b = 0.1f; | |||
| 2015 | const float high_b = 10.0f; | |||
| 2016 | for (int i = 0; i < SUB_PARTITIONS_PART44; ++i) { | |||
| 2017 | // Ratio between the 4:1 sub-block variance and the whole-block variance. | |||
| 2018 | float var_ratio = (float)(horz_4_source_var[i] + 1) / denom; | |||
| 2019 | if (var_ratio < low_b) var_ratio = low_b; | |||
| 2020 | if (var_ratio > high_b) var_ratio = high_b; | |||
| 2021 | features->after_part_ab.f[feature_index++] = var_ratio; | |||
| 2022 | } | |||
| 2023 | for (int i = 0; i < SUB_PARTITIONS_PART44; ++i) { | |||
| 2024 | // Ratio between the 1:4 sub-block RD and the whole-block RD. | |||
| 2025 | float var_ratio = (float)(vert_4_source_var[i] + 1) / denom; | |||
| 2026 | if (var_ratio < low_b) var_ratio = low_b; | |||
| 2027 | if (var_ratio > high_b) var_ratio = high_b; | |||
| 2028 | features->after_part_ab.f[feature_index++] = var_ratio; | |||
| 2029 | } | |||
| 2030 | assert(feature_index == 18)((void) sizeof ((feature_index == 18) ? 1 : 0), __extension__ ({ if (feature_index == 18) ; else __assert_fail ("feature_index == 18" , "/root/firefox-clang/third_party/aom/av1/encoder/partition_strategy.c" , 2030, __extension__ __PRETTY_FUNCTION__); })); | |||
| 2031 | } | |||
| 2032 | ||||
| 2033 | // If the external partition model is used, we let it determine partition | |||
| 2034 | // decisions before partition none. Specifically, these parameters: | |||
| 2035 | // partition_none_allowed | |||
| 2036 | // partition_horz_allowed | |||
| 2037 | // partition_vert_allowed | |||
| 2038 | // do_rectangular_split | |||
| 2039 | // do_square_split | |||
| 2040 | static bool_Bool ext_ml_model_decision_before_none( | |||
| 2041 | AV1_COMP *cpi, const float features_from_motion[FEATURE_SIZE_SMS_SPLIT17], | |||
| 2042 | int *partition_none_allowed, int *partition_horz_allowed, | |||
| 2043 | int *partition_vert_allowed, int *do_rectangular_split, | |||
| 2044 | int *do_square_split) { | |||
| 2045 | ExtPartController *const ext_part_controller = &cpi->ext_part_controller; | |||
| 2046 | if (!ext_part_controller->ready) return false0; | |||
| 2047 | ||||
| 2048 | // Setup features. | |||
| 2049 | aom_partition_features_t features; | |||
| 2050 | features.id = AOM_EXT_PART_FEATURE_BEFORE_NONE; | |||
| 2051 | for (int i = 0; i < FEATURE_SIZE_SMS_SPLIT17; ++i) { | |||
| 2052 | features.before_part_none.f[i] = features_from_motion[i]; | |||
| 2053 | } | |||
| 2054 | ||||
| 2055 | // Send necessary features to the external model. | |||
| 2056 | av1_ext_part_send_features(ext_part_controller, &features); | |||
| 2057 | ||||
| 2058 | // Get partition decisions from the external model. | |||
| 2059 | aom_partition_decision_t decision; | |||
| 2060 | const bool_Bool valid_decision = | |||
| 2061 | av1_ext_part_get_partition_decision(ext_part_controller, &decision); | |||
| 2062 | if (!valid_decision) return false0; | |||
| 2063 | ||||
| 2064 | // Populate decisions | |||
| 2065 | *partition_none_allowed = decision.partition_none_allowed; | |||
| 2066 | *partition_horz_allowed = decision.partition_rect_allowed[HORZ]; | |||
| 2067 | *partition_vert_allowed = decision.partition_rect_allowed[VERT]; | |||
| 2068 | *do_rectangular_split = decision.do_rectangular_split; | |||
| 2069 | *do_square_split = decision.do_square_split; | |||
| 2070 | ||||
| 2071 | return true1; | |||
| 2072 | } | |||
| 2073 | ||||
| 2074 | // If the external partition model is used, we let it determine partition | |||
| 2075 | // decisions before partition none. Specifically, these parameters: | |||
| 2076 | // prune_horz | |||
| 2077 | // prune_vert | |||
| 2078 | static bool_Bool ext_ml_model_decision_before_none_part2( | |||
| 2079 | AV1_COMP *cpi, | |||
| 2080 | const float features_from_motion[FEATURE_SIZE_SMS_PRUNE_PART25], | |||
| 2081 | int *prune_horz, int *prune_vert) { | |||
| 2082 | ExtPartController *const ext_part_controller = &cpi->ext_part_controller; | |||
| 2083 | if (!ext_part_controller->ready) return false0; | |||
| 2084 | ||||
| 2085 | // Setup features. | |||
| 2086 | aom_partition_features_t features; | |||
| 2087 | features.id = AOM_EXT_PART_FEATURE_BEFORE_NONE_PART2; | |||
| 2088 | for (int i = 0; i < FEATURE_SIZE_SMS_PRUNE_PART25; ++i) { | |||
| 2089 | features.before_part_none.f_part2[i] = features_from_motion[i]; | |||
| 2090 | } | |||
| 2091 | ||||
| 2092 | // Send necessary features to the external model. | |||
| 2093 | av1_ext_part_send_features(ext_part_controller, &features); | |||
| 2094 | ||||
| 2095 | // Get partition decisions from the external model. | |||
| 2096 | aom_partition_decision_t decision; | |||
| 2097 | const bool_Bool valid_decision = | |||
| 2098 | av1_ext_part_get_partition_decision(ext_part_controller, &decision); | |||
| 2099 | if (!valid_decision) return false0; | |||
| 2100 | ||||
| 2101 | // Populate decisions | |||
| 2102 | *prune_horz = decision.prune_rect_part[HORZ]; | |||
| 2103 | *prune_vert = decision.prune_rect_part[VERT]; | |||
| 2104 | ||||
| 2105 | return true1; | |||
| 2106 | } | |||
| 2107 | ||||
| 2108 | // If the external partition model is used, we let it determine partition | |||
| 2109 | // decisions after none partition. Specifically, these parameters: | |||
| 2110 | // do_square_split | |||
| 2111 | // do_rectangular_split | |||
| 2112 | bool_Bool ext_ml_model_decision_after_none( | |||
| 2113 | ExtPartController *const ext_part_controller, const int is_intra_frame, | |||
| 2114 | const float *const features_after_none, int *do_square_split, | |||
| 2115 | int *do_rectangular_split) { | |||
| 2116 | if (!ext_part_controller->ready || is_intra_frame) return false0; | |||
| 2117 | ||||
| 2118 | // Setup features. | |||
| 2119 | aom_partition_features_t features; | |||
| 2120 | features.id = AOM_EXT_PART_FEATURE_AFTER_NONE; | |||
| 2121 | for (int i = 0; i < 4; ++i) { | |||
| 2122 | features.after_part_none.f[i] = features_after_none[i]; | |||
| 2123 | } | |||
| 2124 | ||||
| 2125 | // Send necessary features to the external model. | |||
| 2126 | av1_ext_part_send_features(ext_part_controller, &features); | |||
| 2127 | ||||
| 2128 | // Get partition decisions from the external model. | |||
| 2129 | aom_partition_decision_t decision; | |||
| 2130 | const bool_Bool valid_decision = | |||
| 2131 | av1_ext_part_get_partition_decision(ext_part_controller, &decision); | |||
| 2132 | if (!valid_decision) return false0; | |||
| 2133 | ||||
| 2134 | // Populate decisions | |||
| 2135 | *do_square_split = decision.do_square_split; | |||
| 2136 | *do_rectangular_split = decision.do_rectangular_split; | |||
| 2137 | ||||
| 2138 | return true1; | |||
| 2139 | } | |||
| 2140 | ||||
| 2141 | // If the external partition model is used, we let it determine partition | |||
| 2142 | // decisions after none partition. Specifically, these parameters: | |||
| 2143 | // terminate_partition_search | |||
| 2144 | bool_Bool ext_ml_model_decision_after_none_part2( | |||
| 2145 | AV1_COMP *const cpi, const float *const features_terminate, | |||
| 2146 | int *terminate_partition_search) { | |||
| 2147 | AV1_COMMON *const cm = &cpi->common; | |||
| 2148 | ExtPartController *const ext_part_controller = &cpi->ext_part_controller; | |||
| 2149 | if (!ext_part_controller->ready || frame_is_intra_only(cm)) return false0; | |||
| 2150 | ||||
| 2151 | // Setup features. | |||
| 2152 | aom_partition_features_t features; | |||
| 2153 | features.id = AOM_EXT_PART_FEATURE_AFTER_NONE_PART2; | |||
| 2154 | for (int i = 0; i < FEATURE_SIZE_SMS_TERM_NONE28; ++i) { | |||
| 2155 | features.after_part_none.f_terminate[i] = features_terminate[i]; | |||
| 2156 | } | |||
| 2157 | ||||
| 2158 | // Send necessary features to the external model. | |||
| 2159 | av1_ext_part_send_features(ext_part_controller, &features); | |||
| 2160 | ||||
| 2161 | // Get partition decisions from the external model. | |||
| 2162 | aom_partition_decision_t decision; | |||
| 2163 | const bool_Bool valid_decision = | |||
| 2164 | av1_ext_part_get_partition_decision(ext_part_controller, &decision); | |||
| 2165 | if (!valid_decision) return false0; | |||
| 2166 | ||||
| 2167 | // Populate decisions | |||
| 2168 | *terminate_partition_search = decision.terminate_partition_search; | |||
| 2169 | ||||
| 2170 | return true1; | |||
| 2171 | } | |||
| 2172 | ||||
| 2173 | // If the external partition model is used, we let it determine partition | |||
| 2174 | // decisions after none partition. Specifically, these parameters: | |||
| 2175 | // terminate_partition_search | |||
| 2176 | bool_Bool ext_ml_model_decision_after_split(AV1_COMP *const cpi, | |||
| 2177 | const float *const features_terminate, | |||
| 2178 | int *terminate_partition_search) { | |||
| 2179 | const AV1_COMMON *const cm = &cpi->common; | |||
| 2180 | ExtPartController *const ext_part_controller = &cpi->ext_part_controller; | |||
| 2181 | if (frame_is_intra_only(cm) || !cpi->ext_part_controller.ready) { | |||
| 2182 | return false0; | |||
| 2183 | } | |||
| 2184 | ||||
| 2185 | // Setup features. | |||
| 2186 | aom_partition_features_t features; | |||
| 2187 | features.id = AOM_EXT_PART_FEATURE_AFTER_SPLIT; | |||
| 2188 | for (int i = 0; i < 31; ++i) { | |||
| 2189 | features.after_part_split.f_terminate[i] = features_terminate[i]; | |||
| 2190 | } | |||
| 2191 | ||||
| 2192 | // Send necessary features to the external model. | |||
| 2193 | av1_ext_part_send_features(ext_part_controller, &features); | |||
| 2194 | ||||
| 2195 | // Get partition decisions from the external model. | |||
| 2196 | aom_partition_decision_t decision; | |||
| 2197 | const bool_Bool valid_decision = | |||
| 2198 | av1_ext_part_get_partition_decision(ext_part_controller, &decision); | |||
| 2199 | if (!valid_decision) return false0; | |||
| 2200 | ||||
| 2201 | // Populate decisions | |||
| 2202 | *terminate_partition_search = decision.terminate_partition_search; | |||
| 2203 | ||||
| 2204 | return true1; | |||
| 2205 | } | |||
| 2206 | ||||
| 2207 | // If the external partition model is used, we let it determine partition | |||
| 2208 | // decisions after none partition. Specifically, these parameters: | |||
| 2209 | // prune_rect_part[HORZ] | |||
| 2210 | // prune_rect_part[VERT] | |||
| 2211 | bool_Bool ext_ml_model_decision_after_split_part2( | |||
| 2212 | ExtPartController *const ext_part_controller, const int is_intra_frame, | |||
| 2213 | const float *const features_prune, int *prune_rect_part_horz, | |||
| 2214 | int *prune_rect_part_vert) { | |||
| 2215 | if (is_intra_frame || !ext_part_controller->ready) { | |||
| 2216 | return false0; | |||
| 2217 | } | |||
| 2218 | ||||
| 2219 | // Setup features. | |||
| 2220 | aom_partition_features_t features; | |||
| 2221 | features.id = AOM_EXT_PART_FEATURE_AFTER_SPLIT_PART2; | |||
| 2222 | for (int i = 0; i < 9; ++i) { | |||
| 2223 | features.after_part_split.f_prune_rect[i] = features_prune[i]; | |||
| 2224 | } | |||
| 2225 | ||||
| 2226 | // Send necessary features to the external model. | |||
| 2227 | av1_ext_part_send_features(ext_part_controller, &features); | |||
| 2228 | ||||
| 2229 | // Get partition decisions from the external model. | |||
| 2230 | aom_partition_decision_t decision; | |||
| 2231 | const bool_Bool valid_decision = | |||
| 2232 | av1_ext_part_get_partition_decision(ext_part_controller, &decision); | |||
| 2233 | if (!valid_decision) return false0; | |||
| 2234 | ||||
| 2235 | // Populate decisions | |||
| 2236 | *prune_rect_part_horz = decision.prune_rect_part[0]; | |||
| 2237 | *prune_rect_part_vert = decision.prune_rect_part[1]; | |||
| 2238 | ||||
| 2239 | return true1; | |||
| 2240 | } | |||
| 2241 | ||||
| 2242 | // If the external partition model is used, we let it determine partition | |||
| 2243 | // decisions after rectangular partition. Specifically, these parameters: | |||
| 2244 | // horza_partition_allowed | |||
| 2245 | // horzb_partition_allowed | |||
| 2246 | // verta_partition_allowed | |||
| 2247 | // vertb_partition_allowed | |||
| 2248 | static bool_Bool ext_ml_model_decision_after_rect( | |||
| 2249 | ExtPartController *const ext_part_controller, const int is_intra_frame, | |||
| 2250 | const float *const features_after_rect, int *horza_partition_allowed, | |||
| 2251 | int *horzb_partition_allowed, int *verta_partition_allowed, | |||
| 2252 | int *vertb_partition_allowed) { | |||
| 2253 | if (is_intra_frame || !ext_part_controller->ready) return false0; | |||
| 2254 | ||||
| 2255 | // Setup features. | |||
| 2256 | aom_partition_features_t features; | |||
| 2257 | features.id = AOM_EXT_PART_FEATURE_AFTER_RECT; | |||
| 2258 | for (int i = 0; i < 10; ++i) { | |||
| 2259 | features.after_part_rect.f[i] = features_after_rect[i]; | |||
| 2260 | } | |||
| 2261 | ||||
| 2262 | // Send necessary features to the external model. | |||
| 2263 | av1_ext_part_send_features(ext_part_controller, &features); | |||
| 2264 | ||||
| 2265 | // Get partition decisions from the external model. | |||
| 2266 | aom_partition_decision_t decision; | |||
| 2267 | const bool_Bool valid_decision = | |||
| 2268 | av1_ext_part_get_partition_decision(ext_part_controller, &decision); | |||
| 2269 | if (!valid_decision) return false0; | |||
| 2270 | ||||
| 2271 | // Populate decisions | |||
| 2272 | *horza_partition_allowed = decision.horza_partition_allowed; | |||
| 2273 | *horzb_partition_allowed = decision.horzb_partition_allowed; | |||
| 2274 | *verta_partition_allowed = decision.verta_partition_allowed; | |||
| 2275 | *vertb_partition_allowed = decision.vertb_partition_allowed; | |||
| 2276 | ||||
| 2277 | return true1; | |||
| 2278 | } | |||
| 2279 | ||||
| 2280 | // If the external partition model is used, we let it determine partition | |||
| 2281 | // decisions after AB partition. Specifically, these parameters: | |||
| 2282 | // partition_vert4_allowed | |||
| 2283 | // partition_horz4_allowed | |||
| 2284 | static bool_Bool ext_ml_model_decision_after_part_ab( | |||
| 2285 | AV1_COMP *const cpi, MACROBLOCK *const x, BLOCK_SIZE bsize, int part_ctx, | |||
| 2286 | int64_t best_rd, int64_t rect_part_rd[NUM_RECT_PARTS][SUB_PARTITIONS_RECT2], | |||
| 2287 | int64_t split_rd[SUB_PARTITIONS_SPLIT4], int *const partition_horz4_allowed, | |||
| 2288 | int *const partition_vert4_allowed, unsigned int pb_source_variance, | |||
| 2289 | int mi_row, int mi_col) { | |||
| 2290 | const AV1_COMMON *const cm = &cpi->common; | |||
| 2291 | ExtPartController *const ext_part_controller = &cpi->ext_part_controller; | |||
| 2292 | ||||
| 2293 | if (!frame_is_intra_only(cm) && ext_part_controller->ready) { | |||
| 2294 | // Setup features. | |||
| 2295 | aom_partition_features_t features; | |||
| 2296 | features.id = AOM_EXT_PART_FEATURE_AFTER_AB; | |||
| 2297 | prepare_features_after_part_ab(cpi, x, bsize, part_ctx, best_rd, | |||
| 2298 | rect_part_rd, split_rd, pb_source_variance, | |||
| 2299 | mi_row, mi_col, &features); | |||
| 2300 | ||||
| 2301 | // Send necessary features to the external model. | |||
| 2302 | av1_ext_part_send_features(ext_part_controller, &features); | |||
| 2303 | ||||
| 2304 | // Get partition decisions from the external model. | |||
| 2305 | aom_partition_decision_t decision; | |||
| 2306 | const bool_Bool valid_decision = | |||
| 2307 | av1_ext_part_get_partition_decision(ext_part_controller, &decision); | |||
| 2308 | if (!valid_decision) return false0; | |||
| 2309 | ||||
| 2310 | // Populate decisions | |||
| 2311 | *partition_horz4_allowed = decision.partition_horz4_allowed; | |||
| 2312 | *partition_vert4_allowed = decision.partition_vert4_allowed; | |||
| 2313 | ||||
| 2314 | return true1; | |||
| 2315 | } | |||
| 2316 | ||||
| 2317 | return false0; | |||
| 2318 | } | |||
| 2319 | ||||
| 2320 | // This function resembles "av1_setup_sms_tree()" in context_tree.c | |||
| 2321 | // with function signature change. | |||
| 2322 | static SIMPLE_MOTION_DATA_TREE *setup_sms_tree( | |||
| 2323 | AV1_COMP *const cpi, SIMPLE_MOTION_DATA_TREE *sms_tree) { | |||
| 2324 | AV1_COMMON *const cm = &cpi->common; | |||
| 2325 | const int stat_generation_stage = is_stat_generation_stage(cpi); | |||
| 2326 | const int is_sb_size_128 = cm->seq_params->sb_size == BLOCK_128X128; | |||
| 2327 | const int tree_nodes = | |||
| 2328 | av1_get_pc_tree_nodes(is_sb_size_128, stat_generation_stage); | |||
| 2329 | int sms_tree_index = 0; | |||
| 2330 | SIMPLE_MOTION_DATA_TREE *this_sms; | |||
| 2331 | int square_index = 1; | |||
| 2332 | int nodes; | |||
| 2333 | this_sms = &sms_tree[0]; | |||
| 2334 | ||||
| 2335 | if (!stat_generation_stage) { | |||
| 2336 | const int leaf_factor = is_sb_size_128 ? 4 : 1; | |||
| 2337 | const int leaf_nodes = 256 * leaf_factor; | |||
| 2338 | ||||
| 2339 | // Sets up all the leaf nodes in the tree. | |||
| 2340 | for (sms_tree_index = 0; sms_tree_index < leaf_nodes; ++sms_tree_index) { | |||
| 2341 | SIMPLE_MOTION_DATA_TREE *const tree = &sms_tree[sms_tree_index]; | |||
| 2342 | tree->block_size = square[0]; | |||
| 2343 | } | |||
| 2344 | ||||
| 2345 | // Each node has 4 leaf nodes, fill each block_size level of the tree | |||
| 2346 | // from leafs to the root. | |||
| 2347 | for (nodes = leaf_nodes >> 2; nodes > 0; nodes >>= 2) { | |||
| 2348 | for (int i = 0; i < nodes; ++i) { | |||
| 2349 | SIMPLE_MOTION_DATA_TREE *const tree = &sms_tree[sms_tree_index]; | |||
| 2350 | tree->block_size = square[square_index]; | |||
| 2351 | for (int j = 0; j < 4; j++) tree->split[j] = this_sms++; | |||
| 2352 | ++sms_tree_index; | |||
| 2353 | } | |||
| 2354 | ++square_index; | |||
| 2355 | } | |||
| 2356 | } else { | |||
| 2357 | // Allocation for firstpass/LAP stage | |||
| 2358 | // TODO(Mufaddal): refactor square_index to use a common block_size macro | |||
| 2359 | // from firstpass.c | |||
| 2360 | SIMPLE_MOTION_DATA_TREE *const tree = &sms_tree[sms_tree_index]; | |||
| 2361 | square_index = 2; | |||
| 2362 | tree->block_size = square[square_index]; | |||
| 2363 | } | |||
| 2364 | ||||
| 2365 | // Set up the root node for the largest superblock size | |||
| 2366 | return &sms_tree[tree_nodes - 1]; | |||
| 2367 | } | |||
| 2368 | ||||
| 2369 | static void write_motion_feature_to_file( | |||
| 2370 | const char *const path, const int sb_counter, const unsigned int *block_sse, | |||
| 2371 | const unsigned int *block_var, const int num_blocks, const BLOCK_SIZE bsize, | |||
| 2372 | const BLOCK_SIZE fixed_block_size, const int mi_row, const int mi_col) { | |||
| 2373 | char filename[256]; | |||
| 2374 | snprintf(filename, sizeof(filename), "%s/motion_search_feature_sb%d", path, | |||
| 2375 | sb_counter); | |||
| 2376 | FILE *pfile = fopen(filename, "w"); | |||
| 2377 | fprintf(pfile, "%d,%d,%d,%d,%d\n", mi_row, mi_col, bsize, | |||
| ||||
| 2378 | block_size_wide[fixed_block_size], num_blocks); | |||
| 2379 | for (int i = 0; i < num_blocks; ++i) { | |||
| 2380 | fprintf(pfile, "%d", block_sse[i]); | |||
| 2381 | if (i < num_blocks - 1) fprintf(pfile, ","); | |||
| 2382 | } | |||
| 2383 | fprintf(pfile, "\n"); | |||
| 2384 | for (int i = 0; i < num_blocks; ++i) { | |||
| 2385 | fprintf(pfile, "%d", block_var[i]); | |||
| 2386 | if (i < num_blocks - 1) fprintf(pfile, ","); | |||
| 2387 | } | |||
| 2388 | fprintf(pfile, "\n"); | |||
| 2389 | fclose(pfile); | |||
| 2390 | } | |||
| 2391 | ||||
| 2392 | void av1_collect_motion_search_features_sb(AV1_COMP *const cpi, ThreadData *td, | |||
| 2393 | TileDataEnc *tile_data, | |||
| 2394 | const int mi_row, const int mi_col, | |||
| 2395 | const BLOCK_SIZE bsize, | |||
| 2396 | aom_partition_features_t *features) { | |||
| 2397 | const AV1_COMMON *const cm = &cpi->common; | |||
| 2398 | if (frame_is_intra_only(cm)) return; | |||
| ||||
| 2399 | ||||
| 2400 | MACROBLOCK *const x = &td->mb; | |||
| 2401 | const BLOCK_SIZE fixed_block_size = BLOCK_16X16; | |||
| 2402 | const int col_step = mi_size_wide[fixed_block_size]; | |||
| 2403 | const int row_step = mi_size_high[fixed_block_size]; | |||
| 2404 | SIMPLE_MOTION_DATA_TREE *sms_tree = NULL((void*)0); | |||
| 2405 | const int stat_generation_stage = is_stat_generation_stage(cpi); | |||
| 2406 | const int is_sb_size_128 = cm->seq_params->sb_size == BLOCK_128X128; | |||
| 2407 | const int tree_nodes = | |||
| 2408 | av1_get_pc_tree_nodes(is_sb_size_128, stat_generation_stage); | |||
| 2409 | CHECK_MEM_ERROR(cm, sms_tree, aom_calloc(tree_nodes, sizeof(*sms_tree)))do { sms_tree = (aom_calloc(tree_nodes, sizeof(*sms_tree))); if (!sms_tree) aom_internal_error((cm)->error, AOM_CODEC_MEM_ERROR , "Failed to allocate " "sms_tree"); } while (0); | |||
| 2410 | SIMPLE_MOTION_DATA_TREE *sms_root = setup_sms_tree(cpi, sms_tree); | |||
| 2411 | TileInfo *const tile_info = &tile_data->tile_info; | |||
| 2412 | av1_set_offsets_without_segment_id(cpi, tile_info, x, mi_row, mi_col, bsize); | |||
| 2413 | av1_init_simple_motion_search_mvs_for_sb(cpi, NULL((void*)0), x, sms_root, mi_row, | |||
| 2414 | mi_col); | |||
| 2415 | av1_reset_simple_motion_tree_partition(sms_root, bsize); | |||
| 2416 | const int ref_list[] = { cpi->rc.is_src_frame_alt_ref
| |||
| 2417 | : LAST_FRAME }; | |||
| 2418 | const int mi_width = | |||
| 2419 | AOMMIN(mi_size_wide[bsize], cm->mi_params.mi_cols - mi_col)(((mi_size_wide[bsize]) < (cm->mi_params.mi_cols - mi_col )) ? (mi_size_wide[bsize]) : (cm->mi_params.mi_cols - mi_col )); | |||
| 2420 | const int mi_height = | |||
| 2421 | AOMMIN(mi_size_high[bsize], cm->mi_params.mi_rows - mi_row)(((mi_size_high[bsize]) < (cm->mi_params.mi_rows - mi_row )) ? (mi_size_high[bsize]) : (cm->mi_params.mi_rows - mi_row )); | |||
| 2422 | const int col_steps = (mi_width / col_step) + ((mi_width % col_step) > 0); | |||
| 2423 | const int row_steps = (mi_height / row_step) + ((mi_height % row_step) > 0); | |||
| 2424 | const int num_blocks = col_steps * row_steps; | |||
| 2425 | unsigned int *block_sse = aom_calloc(num_blocks, sizeof(*block_sse)); | |||
| 2426 | unsigned int *block_var = aom_calloc(num_blocks, sizeof(*block_var)); | |||
| 2427 | if (!(block_sse && block_var)) { | |||
| 2428 | aom_free(sms_tree); | |||
| 2429 | aom_free(block_sse); | |||
| 2430 | aom_free(block_var); | |||
| 2431 | aom_internal_error(cm->error, AOM_CODEC_MEM_ERROR, | |||
| 2432 | "Error allocating block_sse & block_var"); | |||
| 2433 | } | |||
| 2434 | int idx = 0; | |||
| 2435 | ||||
| 2436 | for (int row = mi_row; | |||
| 2437 | row < AOMMIN(mi_row + mi_size_high[bsize], cm->mi_params.mi_rows)(((mi_row + mi_size_high[bsize]) < (cm->mi_params.mi_rows )) ? (mi_row + mi_size_high[bsize]) : (cm->mi_params.mi_rows )); | |||
| 2438 | row += row_step) { | |||
| 2439 | for (int col = mi_col; | |||
| 2440 | col < AOMMIN(mi_col + mi_size_wide[bsize], cm->mi_params.mi_cols)(((mi_col + mi_size_wide[bsize]) < (cm->mi_params.mi_cols )) ? (mi_col + mi_size_wide[bsize]) : (cm->mi_params.mi_cols )); | |||
| 2441 | col += col_step) { | |||
| 2442 | simple_motion_search_get_best_ref( | |||
| 2443 | cpi, x, sms_root, row, col, fixed_block_size, ref_list, | |||
| 2444 | /*num_refs=*/1, /*use_subpixel=*/1, | |||
| 2445 | /*save_mv=*/1, &block_sse[idx], &block_var[idx]); | |||
| 2446 | ++idx; | |||
| 2447 | } | |||
| 2448 | } | |||
| 2449 | if (features == NULL((void*)0)) { | |||
| 2450 | write_motion_feature_to_file(cpi->oxcf.partition_info_path, cpi->sb_counter, | |||
| 2451 | block_sse, block_var, idx, bsize, | |||
| 2452 | fixed_block_size, mi_row, mi_col); | |||
| 2453 | } else { | |||
| 2454 | features->sb_features.motion_features.unit_length = | |||
| 2455 | block_size_wide[fixed_block_size]; | |||
| 2456 | features->sb_features.motion_features.num_units = idx; | |||
| 2457 | for (int i = 0; i < idx; ++i) { | |||
| 2458 | features->sb_features.motion_features.block_sse[i] = block_sse[i]; | |||
| 2459 | features->sb_features.motion_features.block_var[i] = block_var[i]; | |||
| 2460 | } | |||
| 2461 | } | |||
| 2462 | ||||
| 2463 | aom_free(block_sse); | |||
| 2464 | aom_free(block_var); | |||
| 2465 | aom_free(sms_tree); | |||
| 2466 | } | |||
| 2467 | ||||
| 2468 | #if CONFIG_PARTITION_SEARCH_ORDER0 | |||
| 2469 | void av1_prepare_motion_search_features_block( | |||
| 2470 | AV1_COMP *const cpi, ThreadData *td, TileDataEnc *tile_data, | |||
| 2471 | const int mi_row, const int mi_col, const BLOCK_SIZE bsize, | |||
| 2472 | const int valid_partition_types, unsigned int *block_sse, | |||
| 2473 | unsigned int *block_var, unsigned int sub_block_sse[4], | |||
| 2474 | unsigned int sub_block_var[4], unsigned int horz_block_sse[2], | |||
| 2475 | unsigned int horz_block_var[2], unsigned int vert_block_sse[2], | |||
| 2476 | unsigned int vert_block_var[2]) { | |||
| 2477 | const AV1_COMMON *const cm = &cpi->common; | |||
| 2478 | if (frame_is_intra_only(cm)) return; | |||
| 2479 | MACROBLOCK *const x = &td->mb; | |||
| 2480 | SIMPLE_MOTION_DATA_TREE *sms_tree = NULL((void*)0); | |||
| 2481 | const int stat_generation_stage = is_stat_generation_stage(cpi); | |||
| 2482 | const int is_sb_size_128 = cm->seq_params->sb_size == BLOCK_128X128; | |||
| 2483 | const int tree_nodes = | |||
| 2484 | av1_get_pc_tree_nodes(is_sb_size_128, stat_generation_stage); | |||
| 2485 | CHECK_MEM_ERROR(cm, sms_tree, aom_calloc(tree_nodes, sizeof(*sms_tree)))do { sms_tree = (aom_calloc(tree_nodes, sizeof(*sms_tree))); if (!sms_tree) aom_internal_error((cm)->error, AOM_CODEC_MEM_ERROR , "Failed to allocate " "sms_tree"); } while (0); | |||
| 2486 | SIMPLE_MOTION_DATA_TREE *sms_root = setup_sms_tree(cpi, sms_tree); | |||
| 2487 | TileInfo *const tile_info = &tile_data->tile_info; | |||
| 2488 | av1_set_offsets_without_segment_id(cpi, tile_info, x, mi_row, mi_col, bsize); | |||
| 2489 | av1_reset_simple_motion_tree_partition(sms_root, bsize); | |||
| 2490 | const int ref_list[] = { cpi->rc.is_src_frame_alt_ref ? ALTREF_FRAME | |||
| 2491 | : LAST_FRAME }; | |||
| 2492 | const int sub_mi_width = mi_size_wide[bsize] / 2; | |||
| 2493 | const int sub_mi_height = sub_mi_width; | |||
| 2494 | simple_motion_search_get_best_ref( | |||
| 2495 | cpi, x, sms_root, mi_row, mi_col, bsize, ref_list, /*num_refs=*/1, | |||
| 2496 | /*use_subpixel=*/1, /*save_mv=*/1, block_sse, block_var); | |||
| 2497 | // Split to 4 sub blocks. | |||
| 2498 | if (valid_partition_types & (1 << PARTITION_SPLIT)) { | |||
| 2499 | const BLOCK_SIZE subsize = get_partition_subsize(bsize, PARTITION_SPLIT); | |||
| 2500 | for (int i = 0; i < 4; ++i) { | |||
| 2501 | const int row = mi_row + (i >> 1) * sub_mi_height; | |||
| 2502 | const int col = mi_col + (i & 1) * sub_mi_width; | |||
| 2503 | simple_motion_search_get_best_ref(cpi, x, sms_root, row, col, subsize, | |||
| 2504 | ref_list, /*num_refs=*/1, | |||
| 2505 | /*use_subpixel=*/1, /*save_mv=*/1, | |||
| 2506 | &sub_block_sse[i], &sub_block_var[i]); | |||
| 2507 | } | |||
| 2508 | } | |||
| 2509 | // Horizontal split | |||
| 2510 | if (valid_partition_types & (1 << PARTITION_HORZ)) { | |||
| 2511 | const BLOCK_SIZE subsize = get_partition_subsize(bsize, PARTITION_HORZ); | |||
| 2512 | for (int i = 0; i < 2; ++i) { | |||
| 2513 | const int row = mi_row + (i & 1) * sub_mi_height; | |||
| 2514 | const int col = mi_col; | |||
| 2515 | simple_motion_search_get_best_ref(cpi, x, sms_root, row, col, subsize, | |||
| 2516 | ref_list, /*num_refs=*/1, | |||
| 2517 | /*use_subpixel=*/1, /*save_mv=*/1, | |||
| 2518 | &horz_block_sse[i], &horz_block_var[i]); | |||
| 2519 | } | |||
| 2520 | } | |||
| 2521 | // Vertical split | |||
| 2522 | if (valid_partition_types & (1 << PARTITION_VERT)) { | |||
| 2523 | const BLOCK_SIZE subsize = get_partition_subsize(bsize, PARTITION_VERT); | |||
| 2524 | for (int i = 0; i < 2; ++i) { | |||
| 2525 | const int row = mi_row; | |||
| 2526 | const int col = mi_col + (i & 1) * sub_mi_width; | |||
| 2527 | simple_motion_search_get_best_ref(cpi, x, sms_root, row, col, subsize, | |||
| 2528 | ref_list, /*num_refs=*/1, | |||
| 2529 | /*use_subpixel=*/1, /*save_mv=*/1, | |||
| 2530 | &vert_block_sse[i], &vert_block_var[i]); | |||
| 2531 | } | |||
| 2532 | } | |||
| 2533 | ||||
| 2534 | aom_free(sms_tree); | |||
| 2535 | } | |||
| 2536 | #endif // CONFIG_PARTITION_SEARCH_ORDER | |||
| 2537 | #endif // !CONFIG_REALTIME_ONLY | |||
| 2538 | ||||
| 2539 | static inline void init_simple_motion_search_mvs( | |||
| 2540 | SIMPLE_MOTION_DATA_TREE *sms_tree, const FULLPEL_MV *start_mvs) { | |||
| 2541 | memcpy(sms_tree->start_mvs, start_mvs, sizeof(sms_tree->start_mvs)); | |||
| 2542 | av1_zero(sms_tree->sms_none_feat)memset(&(sms_tree->sms_none_feat), 0, sizeof(sms_tree-> sms_none_feat)); | |||
| 2543 | av1_zero(sms_tree->sms_rect_feat)memset(&(sms_tree->sms_rect_feat), 0, sizeof(sms_tree-> sms_rect_feat)); | |||
| 2544 | av1_zero(sms_tree->sms_none_valid)memset(&(sms_tree->sms_none_valid), 0, sizeof(sms_tree ->sms_none_valid)); | |||
| 2545 | av1_zero(sms_tree->sms_rect_valid)memset(&(sms_tree->sms_rect_valid), 0, sizeof(sms_tree ->sms_rect_valid)); | |||
| 2546 | ||||
| 2547 | if (sms_tree->block_size >= BLOCK_8X8) { | |||
| 2548 | init_simple_motion_search_mvs(sms_tree->split[0], start_mvs); | |||
| 2549 | init_simple_motion_search_mvs(sms_tree->split[1], start_mvs); | |||
| 2550 | init_simple_motion_search_mvs(sms_tree->split[2], start_mvs); | |||
| 2551 | init_simple_motion_search_mvs(sms_tree->split[3], start_mvs); | |||
| 2552 | } | |||
| 2553 | } | |||
| 2554 | ||||
| 2555 | void av1_init_simple_motion_search_mvs_for_sb(const AV1_COMP *cpi, | |||
| 2556 | const TileInfo *tile_info, | |||
| 2557 | MACROBLOCK *x, | |||
| 2558 | SIMPLE_MOTION_DATA_TREE *sms_root, | |||
| 2559 | int mi_row, int mi_col) { | |||
| 2560 | // Use the NEARESTMV of the sb as the start mv | |||
| 2561 | const AV1_COMMON *cm = &cpi->common; | |||
| 2562 | MACROBLOCKD *const xd = &x->e_mbd; | |||
| 2563 | FULLPEL_MV ref_mvs[REF_FRAMES]; | |||
| 2564 | const BLOCK_SIZE sb_size = cm->seq_params->sb_size; | |||
| 2565 | av1_zero(ref_mvs)memset(&(ref_mvs), 0, sizeof(ref_mvs)); | |||
| 2566 | // If tile_info is NULL, assume that the offsets have already been set. | |||
| 2567 | if (tile_info) { | |||
| 2568 | av1_set_offsets_without_segment_id(cpi, tile_info, x, mi_row, mi_col, | |||
| 2569 | sb_size); | |||
| 2570 | } | |||
| 2571 | ||||
| 2572 | MB_MODE_INFO_EXT mbmi_ext; | |||
| 2573 | const int ref_frame = | |||
| 2574 | cpi->rc.is_src_frame_alt_ref ? ALTREF_FRAME : LAST_FRAME; | |||
| 2575 | av1_find_mv_refs(cm, xd, xd->mi[0], ref_frame, mbmi_ext.ref_mv_count, | |||
| 2576 | xd->ref_mv_stack, xd->weight, NULL((void*)0), mbmi_ext.global_mvs, | |||
| 2577 | mbmi_ext.mode_context); | |||
| 2578 | if (mbmi_ext.ref_mv_count[ref_frame] > 0) { | |||
| 2579 | ref_mvs[ref_frame] = | |||
| 2580 | get_fullmv_from_mv(&xd->ref_mv_stack[ref_frame][0].this_mv.as_mv); | |||
| 2581 | } else { | |||
| 2582 | ref_mvs[ref_frame] = | |||
| 2583 | get_fullmv_from_mv(&mbmi_ext.global_mvs[ref_frame].as_mv); | |||
| 2584 | } | |||
| 2585 | ||||
| 2586 | init_simple_motion_search_mvs(sms_root, ref_mvs); | |||
| 2587 | } |