alloccommon.c (19421B)
1 /* 2 * 3 * Copyright (c) 2016, Alliance for Open Media. All rights reserved. 4 * 5 * This source code is subject to the terms of the BSD 2 Clause License and 6 * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License 7 * was not distributed with this source code in the LICENSE file, you can 8 * obtain it at www.aomedia.org/license/software. If the Alliance for Open 9 * Media Patent License 1.0 was not distributed with this source code in the 10 * PATENTS file, you can obtain it at www.aomedia.org/license/patent. 11 */ 12 13 #include "config/aom_config.h" 14 15 #include "aom_mem/aom_mem.h" 16 #include "aom_scale/yv12config.h" 17 #include "aom_util/aom_pthread.h" 18 19 #include "av1/common/alloccommon.h" 20 #include "av1/common/av1_common_int.h" 21 #include "av1/common/blockd.h" 22 #include "av1/common/cdef_block.h" 23 #include "av1/common/entropymode.h" 24 #include "av1/common/entropymv.h" 25 #include "av1/common/enums.h" 26 #include "av1/common/restoration.h" 27 #include "av1/common/thread_common.h" 28 29 int av1_get_MBs(int width, int height) { 30 const int aligned_width = ALIGN_POWER_OF_TWO(width, 3); 31 const int aligned_height = ALIGN_POWER_OF_TWO(height, 3); 32 const int mi_cols = aligned_width >> MI_SIZE_LOG2; 33 const int mi_rows = aligned_height >> MI_SIZE_LOG2; 34 35 const int mb_cols = ROUND_POWER_OF_TWO(mi_cols, 2); 36 const int mb_rows = ROUND_POWER_OF_TWO(mi_rows, 2); 37 return mb_rows * mb_cols; 38 } 39 40 void av1_free_ref_frame_buffers(BufferPool *pool) { 41 int i; 42 43 for (i = 0; i < pool->num_frame_bufs; ++i) { 44 if (pool->frame_bufs[i].ref_count > 0 && 45 pool->frame_bufs[i].raw_frame_buffer.data != NULL) { 46 pool->release_fb_cb(pool->cb_priv, &pool->frame_bufs[i].raw_frame_buffer); 47 pool->frame_bufs[i].raw_frame_buffer.data = NULL; 48 pool->frame_bufs[i].raw_frame_buffer.size = 0; 49 pool->frame_bufs[i].raw_frame_buffer.priv = NULL; 50 pool->frame_bufs[i].ref_count = 0; 51 } 52 aom_free(pool->frame_bufs[i].mvs); 53 pool->frame_bufs[i].mvs = NULL; 54 aom_free(pool->frame_bufs[i].seg_map); 55 pool->frame_bufs[i].seg_map = NULL; 56 aom_free_frame_buffer(&pool->frame_bufs[i].buf); 57 } 58 aom_free(pool->frame_bufs); 59 pool->frame_bufs = NULL; 60 pool->num_frame_bufs = 0; 61 } 62 63 static inline void free_cdef_linebuf_conditional( 64 AV1_COMMON *const cm, const size_t *new_linebuf_size) { 65 CdefInfo *cdef_info = &cm->cdef_info; 66 for (int plane = 0; plane < MAX_MB_PLANE; plane++) { 67 if (new_linebuf_size[plane] != cdef_info->allocated_linebuf_size[plane]) { 68 aom_free(cdef_info->linebuf[plane]); 69 cdef_info->linebuf[plane] = NULL; 70 } 71 } 72 } 73 74 static inline void free_cdef_bufs_conditional(AV1_COMMON *const cm, 75 uint16_t **colbuf, 76 uint16_t **srcbuf, 77 const size_t *new_colbuf_size, 78 const size_t new_srcbuf_size) { 79 CdefInfo *cdef_info = &cm->cdef_info; 80 if (new_srcbuf_size != cdef_info->allocated_srcbuf_size) { 81 aom_free(*srcbuf); 82 *srcbuf = NULL; 83 } 84 for (int plane = 0; plane < MAX_MB_PLANE; plane++) { 85 if (new_colbuf_size[plane] != cdef_info->allocated_colbuf_size[plane]) { 86 aom_free(colbuf[plane]); 87 colbuf[plane] = NULL; 88 } 89 } 90 } 91 92 static inline void free_cdef_bufs(uint16_t **colbuf, uint16_t **srcbuf) { 93 aom_free(*srcbuf); 94 *srcbuf = NULL; 95 for (int plane = 0; plane < MAX_MB_PLANE; plane++) { 96 aom_free(colbuf[plane]); 97 colbuf[plane] = NULL; 98 } 99 } 100 101 static inline void free_cdef_row_sync(AV1CdefRowSync **cdef_row_mt, 102 const int num_mi_rows) { 103 if (*cdef_row_mt == NULL) return; 104 #if CONFIG_MULTITHREAD 105 for (int row_idx = 0; row_idx < num_mi_rows; row_idx++) { 106 if ((*cdef_row_mt)[row_idx].row_mutex_ != NULL) { 107 pthread_mutex_destroy((*cdef_row_mt)[row_idx].row_mutex_); 108 aom_free((*cdef_row_mt)[row_idx].row_mutex_); 109 } 110 if ((*cdef_row_mt)[row_idx].row_cond_ != NULL) { 111 pthread_cond_destroy((*cdef_row_mt)[row_idx].row_cond_); 112 aom_free((*cdef_row_mt)[row_idx].row_cond_); 113 } 114 } 115 #else 116 (void)num_mi_rows; 117 #endif // CONFIG_MULTITHREAD 118 aom_free(*cdef_row_mt); 119 *cdef_row_mt = NULL; 120 } 121 122 void av1_free_cdef_buffers(AV1_COMMON *const cm, 123 AV1CdefWorkerData **cdef_worker, 124 AV1CdefSync *cdef_sync) { 125 CdefInfo *cdef_info = &cm->cdef_info; 126 const int num_mi_rows = cdef_info->allocated_mi_rows; 127 128 for (int plane = 0; plane < MAX_MB_PLANE; plane++) { 129 aom_free(cdef_info->linebuf[plane]); 130 cdef_info->linebuf[plane] = NULL; 131 } 132 // De-allocation of column buffer & source buffer (worker_0). 133 free_cdef_bufs(cdef_info->colbuf, &cdef_info->srcbuf); 134 135 free_cdef_row_sync(&cdef_sync->cdef_row_mt, num_mi_rows); 136 137 if (cdef_info->allocated_num_workers < 2) return; 138 if (*cdef_worker != NULL) { 139 for (int idx = cdef_info->allocated_num_workers - 1; idx >= 1; idx--) { 140 // De-allocation of column buffer & source buffer for remaining workers. 141 free_cdef_bufs((*cdef_worker)[idx].colbuf, &(*cdef_worker)[idx].srcbuf); 142 } 143 aom_free(*cdef_worker); 144 *cdef_worker = NULL; 145 } 146 } 147 148 static inline void alloc_cdef_linebuf(AV1_COMMON *const cm, uint16_t **linebuf, 149 const int num_planes) { 150 CdefInfo *cdef_info = &cm->cdef_info; 151 for (int plane = 0; plane < num_planes; plane++) { 152 if (linebuf[plane] == NULL) 153 CHECK_MEM_ERROR(cm, linebuf[plane], 154 aom_malloc(cdef_info->allocated_linebuf_size[plane])); 155 } 156 } 157 158 static inline void alloc_cdef_bufs(AV1_COMMON *const cm, uint16_t **colbuf, 159 uint16_t **srcbuf, const int num_planes) { 160 CdefInfo *cdef_info = &cm->cdef_info; 161 if (*srcbuf == NULL) 162 CHECK_MEM_ERROR(cm, *srcbuf, 163 aom_memalign(16, cdef_info->allocated_srcbuf_size)); 164 165 for (int plane = 0; plane < num_planes; plane++) { 166 if (colbuf[plane] == NULL) 167 CHECK_MEM_ERROR(cm, colbuf[plane], 168 aom_malloc(cdef_info->allocated_colbuf_size[plane])); 169 } 170 } 171 172 static inline void alloc_cdef_row_sync(AV1_COMMON *const cm, 173 AV1CdefRowSync **cdef_row_mt, 174 const int num_mi_rows) { 175 if (*cdef_row_mt != NULL) return; 176 177 CHECK_MEM_ERROR(cm, *cdef_row_mt, 178 aom_calloc(num_mi_rows, sizeof(**cdef_row_mt))); 179 #if CONFIG_MULTITHREAD 180 for (int row_idx = 0; row_idx < num_mi_rows; row_idx++) { 181 CHECK_MEM_ERROR(cm, (*cdef_row_mt)[row_idx].row_mutex_, 182 aom_malloc(sizeof(*(*cdef_row_mt)[row_idx].row_mutex_))); 183 pthread_mutex_init((*cdef_row_mt)[row_idx].row_mutex_, NULL); 184 185 CHECK_MEM_ERROR(cm, (*cdef_row_mt)[row_idx].row_cond_, 186 aom_malloc(sizeof(*(*cdef_row_mt)[row_idx].row_cond_))); 187 pthread_cond_init((*cdef_row_mt)[row_idx].row_cond_, NULL); 188 } 189 #endif // CONFIG_MULTITHREAD 190 } 191 192 void av1_alloc_cdef_buffers(AV1_COMMON *const cm, 193 AV1CdefWorkerData **cdef_worker, 194 AV1CdefSync *cdef_sync, int num_workers, 195 int init_worker) { 196 const int num_planes = av1_num_planes(cm); 197 size_t new_linebuf_size[MAX_MB_PLANE] = { 0 }; 198 size_t new_colbuf_size[MAX_MB_PLANE] = { 0 }; 199 size_t new_srcbuf_size = 0; 200 CdefInfo *const cdef_info = &cm->cdef_info; 201 // Check for configuration change 202 const int num_mi_rows = 203 (cm->mi_params.mi_rows + MI_SIZE_64X64 - 1) / MI_SIZE_64X64; 204 const int is_num_workers_changed = 205 cdef_info->allocated_num_workers != num_workers; 206 const int is_cdef_enabled = 207 cm->seq_params->enable_cdef && !cm->tiles.single_tile_decoding; 208 209 // num-bufs=3 represents ping-pong buffers for top linebuf, 210 // followed by bottom linebuf. 211 // ping-pong is to avoid top linebuf over-write by consecutive row. 212 int num_bufs = 3; 213 if (num_workers > 1) 214 num_bufs = (cm->mi_params.mi_rows + MI_SIZE_64X64 - 1) / MI_SIZE_64X64; 215 216 if (is_cdef_enabled) { 217 // Calculate src buffer size 218 new_srcbuf_size = sizeof(*cdef_info->srcbuf) * CDEF_INBUF_SIZE; 219 for (int plane = 0; plane < num_planes; plane++) { 220 const int shift = 221 plane == AOM_PLANE_Y ? 0 : cm->seq_params->subsampling_x; 222 // Calculate top and bottom line buffer size 223 const int luma_stride = 224 ALIGN_POWER_OF_TWO(cm->mi_params.mi_cols << MI_SIZE_LOG2, 4); 225 new_linebuf_size[plane] = sizeof(*cdef_info->linebuf) * num_bufs * 226 (CDEF_VBORDER << 1) * (luma_stride >> shift); 227 // Calculate column buffer size 228 const int block_height = 229 (CDEF_BLOCKSIZE << (MI_SIZE_LOG2 - shift)) * 2 * CDEF_VBORDER; 230 new_colbuf_size[plane] = 231 sizeof(*cdef_info->colbuf[plane]) * block_height * CDEF_HBORDER; 232 } 233 } 234 235 // Free src, line and column buffers for worker 0 in case of reallocation 236 free_cdef_linebuf_conditional(cm, new_linebuf_size); 237 free_cdef_bufs_conditional(cm, cdef_info->colbuf, &cdef_info->srcbuf, 238 new_colbuf_size, new_srcbuf_size); 239 240 // The flag init_worker indicates if cdef_worker has to be allocated for the 241 // frame. This is passed as 1 always from decoder. At encoder side, it is 0 242 // when called for parallel frames during FPMT (where cdef_worker is shared 243 // across parallel frames) and 1 otherwise. 244 if (*cdef_worker != NULL && init_worker) { 245 if (is_num_workers_changed) { 246 // Free src and column buffers for remaining workers in case of change in 247 // num_workers 248 for (int idx = cdef_info->allocated_num_workers - 1; idx >= 1; idx--) 249 free_cdef_bufs((*cdef_worker)[idx].colbuf, &(*cdef_worker)[idx].srcbuf); 250 251 aom_free(*cdef_worker); 252 *cdef_worker = NULL; 253 } else if (num_workers > 1) { 254 // Free src and column buffers for remaining workers in case of 255 // reallocation 256 for (int idx = num_workers - 1; idx >= 1; idx--) 257 free_cdef_bufs_conditional(cm, (*cdef_worker)[idx].colbuf, 258 &(*cdef_worker)[idx].srcbuf, new_colbuf_size, 259 new_srcbuf_size); 260 } 261 } 262 263 if (cdef_info->allocated_mi_rows != num_mi_rows) 264 free_cdef_row_sync(&cdef_sync->cdef_row_mt, cdef_info->allocated_mi_rows); 265 266 // Store allocated sizes for reallocation 267 cdef_info->allocated_srcbuf_size = new_srcbuf_size; 268 av1_copy(cdef_info->allocated_colbuf_size, new_colbuf_size); 269 av1_copy(cdef_info->allocated_linebuf_size, new_linebuf_size); 270 // Store configuration to check change in configuration 271 cdef_info->allocated_mi_rows = num_mi_rows; 272 cdef_info->allocated_num_workers = num_workers; 273 274 if (!is_cdef_enabled) return; 275 276 // Memory allocation of column buffer & source buffer (worker_0). 277 alloc_cdef_bufs(cm, cdef_info->colbuf, &cdef_info->srcbuf, num_planes); 278 alloc_cdef_linebuf(cm, cdef_info->linebuf, num_planes); 279 280 if (num_workers < 2) return; 281 282 if (init_worker) { 283 if (*cdef_worker == NULL) 284 CHECK_MEM_ERROR(cm, *cdef_worker, 285 aom_calloc(num_workers, sizeof(**cdef_worker))); 286 287 // Memory allocation of column buffer & source buffer for remaining workers. 288 for (int idx = num_workers - 1; idx >= 1; idx--) 289 alloc_cdef_bufs(cm, (*cdef_worker)[idx].colbuf, 290 &(*cdef_worker)[idx].srcbuf, num_planes); 291 } 292 293 alloc_cdef_row_sync(cm, &cdef_sync->cdef_row_mt, 294 cdef_info->allocated_mi_rows); 295 } 296 297 #if !CONFIG_REALTIME_ONLY || CONFIG_AV1_DECODER 298 // Allocate buffers which are independent of restoration_unit_size 299 void av1_alloc_restoration_buffers(AV1_COMMON *cm, bool is_sgr_enabled) { 300 const int num_planes = av1_num_planes(cm); 301 302 if (cm->rst_tmpbuf == NULL && is_sgr_enabled) { 303 CHECK_MEM_ERROR(cm, cm->rst_tmpbuf, 304 (int32_t *)aom_memalign(16, RESTORATION_TMPBUF_SIZE)); 305 } 306 307 if (cm->rlbs == NULL) { 308 CHECK_MEM_ERROR(cm, cm->rlbs, aom_malloc(sizeof(RestorationLineBuffers))); 309 } 310 311 // For striped loop restoration, we divide each plane into "stripes", 312 // of height 64 luma pixels but with an offset by RESTORATION_UNIT_OFFSET 313 // luma pixels to match the output from CDEF. We will need to store 2 * 314 // RESTORATION_CTX_VERT lines of data for each stripe. 315 int mi_h = cm->mi_params.mi_rows; 316 const int ext_h = RESTORATION_UNIT_OFFSET + (mi_h << MI_SIZE_LOG2); 317 const int num_stripes = (ext_h + 63) / 64; 318 319 // Now we need to allocate enough space to store the line buffers for the 320 // stripes 321 const int frame_w = cm->superres_upscaled_width; 322 const int use_highbd = cm->seq_params->use_highbitdepth; 323 324 for (int p = 0; p < num_planes; ++p) { 325 const int is_uv = p > 0; 326 const int ss_x = is_uv && cm->seq_params->subsampling_x; 327 const int plane_w = ((frame_w + ss_x) >> ss_x) + 2 * RESTORATION_EXTRA_HORZ; 328 const int stride = ALIGN_POWER_OF_TWO(plane_w, 5); 329 const int buf_size = num_stripes * stride * RESTORATION_CTX_VERT 330 << use_highbd; 331 RestorationStripeBoundaries *boundaries = &cm->rst_info[p].boundaries; 332 333 if (buf_size != boundaries->stripe_boundary_size || 334 boundaries->stripe_boundary_above == NULL || 335 boundaries->stripe_boundary_below == NULL) { 336 aom_free(boundaries->stripe_boundary_above); 337 aom_free(boundaries->stripe_boundary_below); 338 339 CHECK_MEM_ERROR(cm, boundaries->stripe_boundary_above, 340 (uint8_t *)aom_memalign(32, buf_size)); 341 CHECK_MEM_ERROR(cm, boundaries->stripe_boundary_below, 342 (uint8_t *)aom_memalign(32, buf_size)); 343 344 boundaries->stripe_boundary_size = buf_size; 345 } 346 boundaries->stripe_boundary_stride = stride; 347 } 348 } 349 350 void av1_free_restoration_buffers(AV1_COMMON *cm) { 351 int p; 352 for (p = 0; p < MAX_MB_PLANE; ++p) 353 av1_free_restoration_struct(&cm->rst_info[p]); 354 aom_free(cm->rst_tmpbuf); 355 cm->rst_tmpbuf = NULL; 356 aom_free(cm->rlbs); 357 cm->rlbs = NULL; 358 for (p = 0; p < MAX_MB_PLANE; ++p) { 359 RestorationStripeBoundaries *boundaries = &cm->rst_info[p].boundaries; 360 aom_free(boundaries->stripe_boundary_above); 361 aom_free(boundaries->stripe_boundary_below); 362 boundaries->stripe_boundary_above = NULL; 363 boundaries->stripe_boundary_below = NULL; 364 } 365 366 aom_free_frame_buffer(&cm->rst_frame); 367 } 368 #endif // !CONFIG_REALTIME_ONLY || CONFIG_AV1_DECODER 369 370 void av1_free_above_context_buffers(CommonContexts *above_contexts) { 371 int i; 372 const int num_planes = above_contexts->num_planes; 373 374 for (int tile_row = 0; tile_row < above_contexts->num_tile_rows; tile_row++) { 375 for (i = 0; i < num_planes; i++) { 376 if (above_contexts->entropy[i] == NULL) break; 377 aom_free(above_contexts->entropy[i][tile_row]); 378 above_contexts->entropy[i][tile_row] = NULL; 379 } 380 if (above_contexts->partition != NULL) { 381 aom_free(above_contexts->partition[tile_row]); 382 above_contexts->partition[tile_row] = NULL; 383 } 384 385 if (above_contexts->txfm != NULL) { 386 aom_free(above_contexts->txfm[tile_row]); 387 above_contexts->txfm[tile_row] = NULL; 388 } 389 } 390 for (i = 0; i < num_planes; i++) { 391 aom_free(above_contexts->entropy[i]); 392 above_contexts->entropy[i] = NULL; 393 } 394 aom_free(above_contexts->partition); 395 above_contexts->partition = NULL; 396 397 aom_free(above_contexts->txfm); 398 above_contexts->txfm = NULL; 399 400 above_contexts->num_tile_rows = 0; 401 above_contexts->num_mi_cols = 0; 402 above_contexts->num_planes = 0; 403 } 404 405 void av1_free_context_buffers(AV1_COMMON *cm) { 406 if (cm->mi_params.free_mi != NULL) cm->mi_params.free_mi(&cm->mi_params); 407 408 av1_free_above_context_buffers(&cm->above_contexts); 409 } 410 411 int av1_alloc_above_context_buffers(CommonContexts *above_contexts, 412 int num_tile_rows, int num_mi_cols, 413 int num_planes) { 414 const int aligned_mi_cols = 415 ALIGN_POWER_OF_TWO(num_mi_cols, MAX_MIB_SIZE_LOG2); 416 417 // Allocate above context buffers 418 above_contexts->num_tile_rows = num_tile_rows; 419 above_contexts->num_mi_cols = aligned_mi_cols; 420 above_contexts->num_planes = num_planes; 421 for (int plane_idx = 0; plane_idx < num_planes; plane_idx++) { 422 above_contexts->entropy[plane_idx] = (ENTROPY_CONTEXT **)aom_calloc( 423 num_tile_rows, sizeof(above_contexts->entropy[0])); 424 if (!above_contexts->entropy[plane_idx]) return 1; 425 } 426 427 above_contexts->partition = (PARTITION_CONTEXT **)aom_calloc( 428 num_tile_rows, sizeof(above_contexts->partition)); 429 if (!above_contexts->partition) return 1; 430 431 above_contexts->txfm = 432 (TXFM_CONTEXT **)aom_calloc(num_tile_rows, sizeof(above_contexts->txfm)); 433 if (!above_contexts->txfm) return 1; 434 435 for (int tile_row = 0; tile_row < num_tile_rows; tile_row++) { 436 for (int plane_idx = 0; plane_idx < num_planes; plane_idx++) { 437 above_contexts->entropy[plane_idx][tile_row] = 438 (ENTROPY_CONTEXT *)aom_calloc( 439 aligned_mi_cols, sizeof(*above_contexts->entropy[0][tile_row])); 440 if (!above_contexts->entropy[plane_idx][tile_row]) return 1; 441 } 442 443 above_contexts->partition[tile_row] = (PARTITION_CONTEXT *)aom_calloc( 444 aligned_mi_cols, sizeof(*above_contexts->partition[tile_row])); 445 if (!above_contexts->partition[tile_row]) return 1; 446 447 above_contexts->txfm[tile_row] = (TXFM_CONTEXT *)aom_calloc( 448 aligned_mi_cols, sizeof(*above_contexts->txfm[tile_row])); 449 if (!above_contexts->txfm[tile_row]) return 1; 450 } 451 452 return 0; 453 } 454 455 // Allocate the dynamically allocated arrays in 'mi_params' assuming 456 // 'mi_params->set_mb_mi()' was already called earlier to initialize the rest of 457 // the struct members. 458 static int alloc_mi(CommonModeInfoParams *mi_params) { 459 const int aligned_mi_rows = calc_mi_size(mi_params->mi_rows); 460 const int mi_grid_size = mi_params->mi_stride * aligned_mi_rows; 461 const int alloc_size_1d = mi_size_wide[mi_params->mi_alloc_bsize]; 462 const int alloc_mi_size = 463 mi_params->mi_alloc_stride * (aligned_mi_rows / alloc_size_1d); 464 465 if (mi_params->mi_alloc_size < alloc_mi_size || 466 mi_params->mi_grid_size < mi_grid_size) { 467 mi_params->free_mi(mi_params); 468 469 mi_params->mi_alloc = 470 aom_calloc(alloc_mi_size, sizeof(*mi_params->mi_alloc)); 471 if (!mi_params->mi_alloc) return 1; 472 mi_params->mi_alloc_size = alloc_mi_size; 473 474 mi_params->mi_grid_base = (MB_MODE_INFO **)aom_calloc( 475 mi_grid_size, sizeof(*mi_params->mi_grid_base)); 476 if (!mi_params->mi_grid_base) return 1; 477 478 mi_params->tx_type_map = 479 aom_calloc(mi_grid_size, sizeof(*mi_params->tx_type_map)); 480 if (!mi_params->tx_type_map) return 1; 481 mi_params->mi_grid_size = mi_grid_size; 482 } 483 484 return 0; 485 } 486 487 int av1_alloc_context_buffers(AV1_COMMON *cm, int width, int height, 488 BLOCK_SIZE min_partition_size) { 489 CommonModeInfoParams *const mi_params = &cm->mi_params; 490 mi_params->set_mb_mi(mi_params, width, height, min_partition_size); 491 if (alloc_mi(mi_params)) goto fail; 492 return 0; 493 494 fail: 495 // clear the mi_* values to force a realloc on resync 496 mi_params->set_mb_mi(mi_params, 0, 0, BLOCK_4X4); 497 av1_free_context_buffers(cm); 498 return 1; 499 } 500 501 void av1_remove_common(AV1_COMMON *cm) { 502 av1_free_context_buffers(cm); 503 504 aom_free(cm->fc); 505 cm->fc = NULL; 506 aom_free(cm->default_frame_context); 507 cm->default_frame_context = NULL; 508 } 509 510 void av1_init_mi_buffers(CommonModeInfoParams *mi_params) { 511 mi_params->setup_mi(mi_params); 512 }