1/* 2 * Copyright (c) 2010 The WebM project authors. All Rights Reserved. 3 * 4 * Use of this source code is governed by a BSD-style license 5 * that can be found in the LICENSE file in the root of the source 6 * tree. An additional intellectual property rights grant can be found 7 * in the file PATENTS. All contributing project authors may 8 * be found in the AUTHORS file in the root of the source tree. 9 */ 10 11#include <math.h> 12#include <stdio.h> 13#include <limits.h> 14 15#include "./vp9_rtcd.h" 16#include "./vpx_config.h" 17#include "./vpx_dsp_rtcd.h" 18#include "./vpx_scale_rtcd.h" 19#include "vpx/internal/vpx_psnr.h" 20#include "vpx_dsp/vpx_dsp_common.h" 21#include "vpx_dsp/vpx_filter.h" 22#if CONFIG_INTERNAL_STATS 23#include "vpx_dsp/ssim.h" 24#endif 25#include "vpx_ports/mem.h" 26#include "vpx_ports/system_state.h" 27#include "vpx_ports/vpx_timer.h" 28 29#include "vp9/common/vp9_alloccommon.h" 30#include "vp9/common/vp9_filter.h" 31#include "vp9/common/vp9_idct.h" 32#if CONFIG_VP9_POSTPROC 33#include "vp9/common/vp9_postproc.h" 34#endif 35#include "vp9/common/vp9_reconinter.h" 36#include "vp9/common/vp9_reconintra.h" 37#include "vp9/common/vp9_tile_common.h" 38 39#include "vp9/encoder/vp9_aq_complexity.h" 40#include "vp9/encoder/vp9_aq_cyclicrefresh.h" 41#include "vp9/encoder/vp9_aq_variance.h" 42#include "vp9/encoder/vp9_bitstream.h" 43#include "vp9/encoder/vp9_context_tree.h" 44#include "vp9/encoder/vp9_encodeframe.h" 45#include "vp9/encoder/vp9_encodemv.h" 46#include "vp9/encoder/vp9_encoder.h" 47#include "vp9/encoder/vp9_ethread.h" 48#include "vp9/encoder/vp9_firstpass.h" 49#include "vp9/encoder/vp9_mbgraph.h" 50#include "vp9/encoder/vp9_picklpf.h" 51#include "vp9/encoder/vp9_ratectrl.h" 52#include "vp9/encoder/vp9_rd.h" 53#include "vp9/encoder/vp9_resize.h" 54#include "vp9/encoder/vp9_segmentation.h" 55#include "vp9/encoder/vp9_skin_detection.h" 56#include "vp9/encoder/vp9_speed_features.h" 57#include "vp9/encoder/vp9_svc_layercontext.h" 58#include "vp9/encoder/vp9_temporal_filter.h" 59 60#define AM_SEGMENT_ID_INACTIVE 7 61#define AM_SEGMENT_ID_ACTIVE 0 62 63#define SHARP_FILTER_QTHRESH 0 /* Q threshold for 8-tap sharp filter */ 64 65#define ALTREF_HIGH_PRECISION_MV 1 // Whether to use high precision mv 66 // for altref computation. 67#define HIGH_PRECISION_MV_QTHRESH 200 // Q threshold for high precision 68 // mv. Choose a very high value for 69 // now so that HIGH_PRECISION is always 70 // chosen. 71// #define OUTPUT_YUV_REC 72 73#ifdef OUTPUT_YUV_DENOISED 74FILE *yuv_denoised_file = NULL; 75#endif 76#ifdef OUTPUT_YUV_SKINMAP 77FILE *yuv_skinmap_file = NULL; 78#endif 79#ifdef OUTPUT_YUV_REC 80FILE *yuv_rec_file; 81#endif 82 83#if 0 84FILE *framepsnr; 85FILE *kf_list; 86FILE *keyfile; 87#endif 88 89static INLINE void Scale2Ratio(VPX_SCALING mode, int *hr, int *hs) { 90 switch (mode) { 91 case NORMAL: 92 *hr = 1; 93 *hs = 1; 94 break; 95 case FOURFIVE: 96 *hr = 4; 97 *hs = 5; 98 break; 99 case THREEFIVE: 100 *hr = 3; 101 *hs = 5; 102 break; 103 case ONETWO: 104 *hr = 1; 105 *hs = 2; 106 break; 107 default: 108 *hr = 1; 109 *hs = 1; 110 assert(0); 111 break; 112 } 113} 114 115// Mark all inactive blocks as active. Other segmentation features may be set 116// so memset cannot be used, instead only inactive blocks should be reset. 117static void suppress_active_map(VP9_COMP *cpi) { 118 unsigned char *const seg_map = cpi->segmentation_map; 119 int i; 120 if (cpi->active_map.enabled || cpi->active_map.update) 121 for (i = 0; i < cpi->common.mi_rows * cpi->common.mi_cols; ++i) 122 if (seg_map[i] == AM_SEGMENT_ID_INACTIVE) 123 seg_map[i] = AM_SEGMENT_ID_ACTIVE; 124} 125 126static void apply_active_map(VP9_COMP *cpi) { 127 struct segmentation *const seg = &cpi->common.seg; 128 unsigned char *const seg_map = cpi->segmentation_map; 129 const unsigned char *const active_map = cpi->active_map.map; 130 int i; 131 132 assert(AM_SEGMENT_ID_ACTIVE == CR_SEGMENT_ID_BASE); 133 134 if (frame_is_intra_only(&cpi->common)) { 135 cpi->active_map.enabled = 0; 136 cpi->active_map.update = 1; 137 } 138 139 if (cpi->active_map.update) { 140 if (cpi->active_map.enabled) { 141 for (i = 0; i < cpi->common.mi_rows * cpi->common.mi_cols; ++i) 142 if (seg_map[i] == AM_SEGMENT_ID_ACTIVE) seg_map[i] = active_map[i]; 143 vp9_enable_segmentation(seg); 144 vp9_enable_segfeature(seg, AM_SEGMENT_ID_INACTIVE, SEG_LVL_SKIP); 145 vp9_enable_segfeature(seg, AM_SEGMENT_ID_INACTIVE, SEG_LVL_ALT_LF); 146 // Setting the data to -MAX_LOOP_FILTER will result in the computed loop 147 // filter level being zero regardless of the value of seg->abs_delta. 148 vp9_set_segdata(seg, AM_SEGMENT_ID_INACTIVE, 149 SEG_LVL_ALT_LF, -MAX_LOOP_FILTER); 150 } else { 151 vp9_disable_segfeature(seg, AM_SEGMENT_ID_INACTIVE, SEG_LVL_SKIP); 152 vp9_disable_segfeature(seg, AM_SEGMENT_ID_INACTIVE, SEG_LVL_ALT_LF); 153 if (seg->enabled) { 154 seg->update_data = 1; 155 seg->update_map = 1; 156 } 157 } 158 cpi->active_map.update = 0; 159 } 160} 161 162int vp9_set_active_map(VP9_COMP* cpi, 163 unsigned char* new_map_16x16, 164 int rows, 165 int cols) { 166 if (rows == cpi->common.mb_rows && cols == cpi->common.mb_cols) { 167 unsigned char *const active_map_8x8 = cpi->active_map.map; 168 const int mi_rows = cpi->common.mi_rows; 169 const int mi_cols = cpi->common.mi_cols; 170 cpi->active_map.update = 1; 171 if (new_map_16x16) { 172 int r, c; 173 for (r = 0; r < mi_rows; ++r) { 174 for (c = 0; c < mi_cols; ++c) { 175 active_map_8x8[r * mi_cols + c] = 176 new_map_16x16[(r >> 1) * cols + (c >> 1)] 177 ? AM_SEGMENT_ID_ACTIVE 178 : AM_SEGMENT_ID_INACTIVE; 179 } 180 } 181 cpi->active_map.enabled = 1; 182 } else { 183 cpi->active_map.enabled = 0; 184 } 185 return 0; 186 } else { 187 return -1; 188 } 189} 190 191int vp9_get_active_map(VP9_COMP* cpi, 192 unsigned char* new_map_16x16, 193 int rows, 194 int cols) { 195 if (rows == cpi->common.mb_rows && cols == cpi->common.mb_cols && 196 new_map_16x16) { 197 unsigned char* const seg_map_8x8 = cpi->segmentation_map; 198 const int mi_rows = cpi->common.mi_rows; 199 const int mi_cols = cpi->common.mi_cols; 200 memset(new_map_16x16, !cpi->active_map.enabled, rows * cols); 201 if (cpi->active_map.enabled) { 202 int r, c; 203 for (r = 0; r < mi_rows; ++r) { 204 for (c = 0; c < mi_cols; ++c) { 205 // Cyclic refresh segments are considered active despite not having 206 // AM_SEGMENT_ID_ACTIVE 207 new_map_16x16[(r >> 1) * cols + (c >> 1)] |= 208 seg_map_8x8[r * mi_cols + c] != AM_SEGMENT_ID_INACTIVE; 209 } 210 } 211 } 212 return 0; 213 } else { 214 return -1; 215 } 216} 217 218void vp9_set_high_precision_mv(VP9_COMP *cpi, int allow_high_precision_mv) { 219 MACROBLOCK *const mb = &cpi->td.mb; 220 cpi->common.allow_high_precision_mv = allow_high_precision_mv; 221 if (cpi->common.allow_high_precision_mv) { 222 mb->mvcost = mb->nmvcost_hp; 223 mb->mvsadcost = mb->nmvsadcost_hp; 224 } else { 225 mb->mvcost = mb->nmvcost; 226 mb->mvsadcost = mb->nmvsadcost; 227 } 228} 229 230static void setup_frame(VP9_COMP *cpi) { 231 VP9_COMMON *const cm = &cpi->common; 232 // Set up entropy context depending on frame type. The decoder mandates 233 // the use of the default context, index 0, for keyframes and inter 234 // frames where the error_resilient_mode or intra_only flag is set. For 235 // other inter-frames the encoder currently uses only two contexts; 236 // context 1 for ALTREF frames and context 0 for the others. 237 if (frame_is_intra_only(cm) || cm->error_resilient_mode) { 238 vp9_setup_past_independence(cm); 239 } else { 240 if (!cpi->use_svc) 241 cm->frame_context_idx = cpi->refresh_alt_ref_frame; 242 } 243 244 if (cm->frame_type == KEY_FRAME) { 245 if (!is_two_pass_svc(cpi)) 246 cpi->refresh_golden_frame = 1; 247 cpi->refresh_alt_ref_frame = 1; 248 vp9_zero(cpi->interp_filter_selected); 249 } else { 250 *cm->fc = cm->frame_contexts[cm->frame_context_idx]; 251 vp9_zero(cpi->interp_filter_selected[0]); 252 } 253} 254 255static void vp9_enc_setup_mi(VP9_COMMON *cm) { 256 int i; 257 cm->mi = cm->mip + cm->mi_stride + 1; 258 memset(cm->mip, 0, cm->mi_stride * (cm->mi_rows + 1) * sizeof(*cm->mip)); 259 cm->prev_mi = cm->prev_mip + cm->mi_stride + 1; 260 // Clear top border row 261 memset(cm->prev_mip, 0, sizeof(*cm->prev_mip) * cm->mi_stride); 262 // Clear left border column 263 for (i = 1; i < cm->mi_rows + 1; ++i) 264 memset(&cm->prev_mip[i * cm->mi_stride], 0, sizeof(*cm->prev_mip)); 265 266 cm->mi_grid_visible = cm->mi_grid_base + cm->mi_stride + 1; 267 cm->prev_mi_grid_visible = cm->prev_mi_grid_base + cm->mi_stride + 1; 268 269 memset(cm->mi_grid_base, 0, 270 cm->mi_stride * (cm->mi_rows + 1) * sizeof(*cm->mi_grid_base)); 271} 272 273static int vp9_enc_alloc_mi(VP9_COMMON *cm, int mi_size) { 274 cm->mip = vpx_calloc(mi_size, sizeof(*cm->mip)); 275 if (!cm->mip) 276 return 1; 277 cm->prev_mip = vpx_calloc(mi_size, sizeof(*cm->prev_mip)); 278 if (!cm->prev_mip) 279 return 1; 280 cm->mi_alloc_size = mi_size; 281 282 cm->mi_grid_base = (MODE_INFO **)vpx_calloc(mi_size, sizeof(MODE_INFO*)); 283 if (!cm->mi_grid_base) 284 return 1; 285 cm->prev_mi_grid_base = (MODE_INFO **)vpx_calloc(mi_size, sizeof(MODE_INFO*)); 286 if (!cm->prev_mi_grid_base) 287 return 1; 288 289 return 0; 290} 291 292static void vp9_enc_free_mi(VP9_COMMON *cm) { 293 vpx_free(cm->mip); 294 cm->mip = NULL; 295 vpx_free(cm->prev_mip); 296 cm->prev_mip = NULL; 297 vpx_free(cm->mi_grid_base); 298 cm->mi_grid_base = NULL; 299 vpx_free(cm->prev_mi_grid_base); 300 cm->prev_mi_grid_base = NULL; 301} 302 303static void vp9_swap_mi_and_prev_mi(VP9_COMMON *cm) { 304 // Current mip will be the prev_mip for the next frame. 305 MODE_INFO **temp_base = cm->prev_mi_grid_base; 306 MODE_INFO *temp = cm->prev_mip; 307 cm->prev_mip = cm->mip; 308 cm->mip = temp; 309 310 // Update the upper left visible macroblock ptrs. 311 cm->mi = cm->mip + cm->mi_stride + 1; 312 cm->prev_mi = cm->prev_mip + cm->mi_stride + 1; 313 314 cm->prev_mi_grid_base = cm->mi_grid_base; 315 cm->mi_grid_base = temp_base; 316 cm->mi_grid_visible = cm->mi_grid_base + cm->mi_stride + 1; 317 cm->prev_mi_grid_visible = cm->prev_mi_grid_base + cm->mi_stride + 1; 318} 319 320void vp9_initialize_enc(void) { 321 static volatile int init_done = 0; 322 323 if (!init_done) { 324 vp9_rtcd(); 325 vpx_dsp_rtcd(); 326 vpx_scale_rtcd(); 327 vp9_init_intra_predictors(); 328 vp9_init_me_luts(); 329 vp9_rc_init_minq_luts(); 330 vp9_entropy_mv_init(); 331 vp9_temporal_filter_init(); 332 init_done = 1; 333 } 334} 335 336static void dealloc_compressor_data(VP9_COMP *cpi) { 337 VP9_COMMON *const cm = &cpi->common; 338 int i; 339 340 vpx_free(cpi->mbmi_ext_base); 341 cpi->mbmi_ext_base = NULL; 342 343 vpx_free(cpi->tile_data); 344 cpi->tile_data = NULL; 345 346 // Delete sementation map 347 vpx_free(cpi->segmentation_map); 348 cpi->segmentation_map = NULL; 349 vpx_free(cpi->coding_context.last_frame_seg_map_copy); 350 cpi->coding_context.last_frame_seg_map_copy = NULL; 351 352 vpx_free(cpi->nmvcosts[0]); 353 vpx_free(cpi->nmvcosts[1]); 354 cpi->nmvcosts[0] = NULL; 355 cpi->nmvcosts[1] = NULL; 356 357 vpx_free(cpi->nmvcosts_hp[0]); 358 vpx_free(cpi->nmvcosts_hp[1]); 359 cpi->nmvcosts_hp[0] = NULL; 360 cpi->nmvcosts_hp[1] = NULL; 361 362 vpx_free(cpi->nmvsadcosts[0]); 363 vpx_free(cpi->nmvsadcosts[1]); 364 cpi->nmvsadcosts[0] = NULL; 365 cpi->nmvsadcosts[1] = NULL; 366 367 vpx_free(cpi->nmvsadcosts_hp[0]); 368 vpx_free(cpi->nmvsadcosts_hp[1]); 369 cpi->nmvsadcosts_hp[0] = NULL; 370 cpi->nmvsadcosts_hp[1] = NULL; 371 372 vp9_cyclic_refresh_free(cpi->cyclic_refresh); 373 cpi->cyclic_refresh = NULL; 374 375 vpx_free(cpi->active_map.map); 376 cpi->active_map.map = NULL; 377 378 vp9_free_ref_frame_buffers(cm->buffer_pool); 379#if CONFIG_VP9_POSTPROC 380 vp9_free_postproc_buffers(cm); 381#endif 382 vp9_free_context_buffers(cm); 383 384 vpx_free_frame_buffer(&cpi->last_frame_uf); 385 vpx_free_frame_buffer(&cpi->scaled_source); 386 vpx_free_frame_buffer(&cpi->scaled_last_source); 387 vpx_free_frame_buffer(&cpi->alt_ref_buffer); 388 vp9_lookahead_destroy(cpi->lookahead); 389 390 vpx_free(cpi->tile_tok[0][0]); 391 cpi->tile_tok[0][0] = 0; 392 393 vp9_free_pc_tree(&cpi->td); 394 395 for (i = 0; i < cpi->svc.number_spatial_layers; ++i) { 396 LAYER_CONTEXT *const lc = &cpi->svc.layer_context[i]; 397 vpx_free(lc->rc_twopass_stats_in.buf); 398 lc->rc_twopass_stats_in.buf = NULL; 399 lc->rc_twopass_stats_in.sz = 0; 400 } 401 402 if (cpi->source_diff_var != NULL) { 403 vpx_free(cpi->source_diff_var); 404 cpi->source_diff_var = NULL; 405 } 406 407 for (i = 0; i < MAX_LAG_BUFFERS; ++i) { 408 vpx_free_frame_buffer(&cpi->svc.scaled_frames[i]); 409 } 410 memset(&cpi->svc.scaled_frames[0], 0, 411 MAX_LAG_BUFFERS * sizeof(cpi->svc.scaled_frames[0])); 412 413 vpx_free_frame_buffer(&cpi->svc.empty_frame.img); 414 memset(&cpi->svc.empty_frame, 0, sizeof(cpi->svc.empty_frame)); 415 416 vp9_free_svc_cyclic_refresh(cpi); 417} 418 419static void save_coding_context(VP9_COMP *cpi) { 420 CODING_CONTEXT *const cc = &cpi->coding_context; 421 VP9_COMMON *cm = &cpi->common; 422 423 // Stores a snapshot of key state variables which can subsequently be 424 // restored with a call to vp9_restore_coding_context. These functions are 425 // intended for use in a re-code loop in vp9_compress_frame where the 426 // quantizer value is adjusted between loop iterations. 427 vp9_copy(cc->nmvjointcost, cpi->td.mb.nmvjointcost); 428 429 memcpy(cc->nmvcosts[0], cpi->nmvcosts[0], 430 MV_VALS * sizeof(*cpi->nmvcosts[0])); 431 memcpy(cc->nmvcosts[1], cpi->nmvcosts[1], 432 MV_VALS * sizeof(*cpi->nmvcosts[1])); 433 memcpy(cc->nmvcosts_hp[0], cpi->nmvcosts_hp[0], 434 MV_VALS * sizeof(*cpi->nmvcosts_hp[0])); 435 memcpy(cc->nmvcosts_hp[1], cpi->nmvcosts_hp[1], 436 MV_VALS * sizeof(*cpi->nmvcosts_hp[1])); 437 438 vp9_copy(cc->segment_pred_probs, cm->seg.pred_probs); 439 440 memcpy(cpi->coding_context.last_frame_seg_map_copy, 441 cm->last_frame_seg_map, (cm->mi_rows * cm->mi_cols)); 442 443 vp9_copy(cc->last_ref_lf_deltas, cm->lf.last_ref_deltas); 444 vp9_copy(cc->last_mode_lf_deltas, cm->lf.last_mode_deltas); 445 446 cc->fc = *cm->fc; 447} 448 449static void restore_coding_context(VP9_COMP *cpi) { 450 CODING_CONTEXT *const cc = &cpi->coding_context; 451 VP9_COMMON *cm = &cpi->common; 452 453 // Restore key state variables to the snapshot state stored in the 454 // previous call to vp9_save_coding_context. 455 vp9_copy(cpi->td.mb.nmvjointcost, cc->nmvjointcost); 456 457 memcpy(cpi->nmvcosts[0], cc->nmvcosts[0], MV_VALS * sizeof(*cc->nmvcosts[0])); 458 memcpy(cpi->nmvcosts[1], cc->nmvcosts[1], MV_VALS * sizeof(*cc->nmvcosts[1])); 459 memcpy(cpi->nmvcosts_hp[0], cc->nmvcosts_hp[0], 460 MV_VALS * sizeof(*cc->nmvcosts_hp[0])); 461 memcpy(cpi->nmvcosts_hp[1], cc->nmvcosts_hp[1], 462 MV_VALS * sizeof(*cc->nmvcosts_hp[1])); 463 464 vp9_copy(cm->seg.pred_probs, cc->segment_pred_probs); 465 466 memcpy(cm->last_frame_seg_map, 467 cpi->coding_context.last_frame_seg_map_copy, 468 (cm->mi_rows * cm->mi_cols)); 469 470 vp9_copy(cm->lf.last_ref_deltas, cc->last_ref_lf_deltas); 471 vp9_copy(cm->lf.last_mode_deltas, cc->last_mode_lf_deltas); 472 473 *cm->fc = cc->fc; 474} 475 476static void configure_static_seg_features(VP9_COMP *cpi) { 477 VP9_COMMON *const cm = &cpi->common; 478 const RATE_CONTROL *const rc = &cpi->rc; 479 struct segmentation *const seg = &cm->seg; 480 481 int high_q = (int)(rc->avg_q > 48.0); 482 int qi_delta; 483 484 // Disable and clear down for KF 485 if (cm->frame_type == KEY_FRAME) { 486 // Clear down the global segmentation map 487 memset(cpi->segmentation_map, 0, cm->mi_rows * cm->mi_cols); 488 seg->update_map = 0; 489 seg->update_data = 0; 490 cpi->static_mb_pct = 0; 491 492 // Disable segmentation 493 vp9_disable_segmentation(seg); 494 495 // Clear down the segment features. 496 vp9_clearall_segfeatures(seg); 497 } else if (cpi->refresh_alt_ref_frame) { 498 // If this is an alt ref frame 499 // Clear down the global segmentation map 500 memset(cpi->segmentation_map, 0, cm->mi_rows * cm->mi_cols); 501 seg->update_map = 0; 502 seg->update_data = 0; 503 cpi->static_mb_pct = 0; 504 505 // Disable segmentation and individual segment features by default 506 vp9_disable_segmentation(seg); 507 vp9_clearall_segfeatures(seg); 508 509 // Scan frames from current to arf frame. 510 // This function re-enables segmentation if appropriate. 511 vp9_update_mbgraph_stats(cpi); 512 513 // If segmentation was enabled set those features needed for the 514 // arf itself. 515 if (seg->enabled) { 516 seg->update_map = 1; 517 seg->update_data = 1; 518 519 qi_delta = vp9_compute_qdelta(rc, rc->avg_q, rc->avg_q * 0.875, 520 cm->bit_depth); 521 vp9_set_segdata(seg, 1, SEG_LVL_ALT_Q, qi_delta - 2); 522 vp9_set_segdata(seg, 1, SEG_LVL_ALT_LF, -2); 523 524 vp9_enable_segfeature(seg, 1, SEG_LVL_ALT_Q); 525 vp9_enable_segfeature(seg, 1, SEG_LVL_ALT_LF); 526 527 // Where relevant assume segment data is delta data 528 seg->abs_delta = SEGMENT_DELTADATA; 529 } 530 } else if (seg->enabled) { 531 // All other frames if segmentation has been enabled 532 533 // First normal frame in a valid gf or alt ref group 534 if (rc->frames_since_golden == 0) { 535 // Set up segment features for normal frames in an arf group 536 if (rc->source_alt_ref_active) { 537 seg->update_map = 0; 538 seg->update_data = 1; 539 seg->abs_delta = SEGMENT_DELTADATA; 540 541 qi_delta = vp9_compute_qdelta(rc, rc->avg_q, rc->avg_q * 1.125, 542 cm->bit_depth); 543 vp9_set_segdata(seg, 1, SEG_LVL_ALT_Q, qi_delta + 2); 544 vp9_enable_segfeature(seg, 1, SEG_LVL_ALT_Q); 545 546 vp9_set_segdata(seg, 1, SEG_LVL_ALT_LF, -2); 547 vp9_enable_segfeature(seg, 1, SEG_LVL_ALT_LF); 548 549 // Segment coding disabled for compred testing 550 if (high_q || (cpi->static_mb_pct == 100)) { 551 vp9_set_segdata(seg, 1, SEG_LVL_REF_FRAME, ALTREF_FRAME); 552 vp9_enable_segfeature(seg, 1, SEG_LVL_REF_FRAME); 553 vp9_enable_segfeature(seg, 1, SEG_LVL_SKIP); 554 } 555 } else { 556 // Disable segmentation and clear down features if alt ref 557 // is not active for this group 558 559 vp9_disable_segmentation(seg); 560 561 memset(cpi->segmentation_map, 0, cm->mi_rows * cm->mi_cols); 562 563 seg->update_map = 0; 564 seg->update_data = 0; 565 566 vp9_clearall_segfeatures(seg); 567 } 568 } else if (rc->is_src_frame_alt_ref) { 569 // Special case where we are coding over the top of a previous 570 // alt ref frame. 571 // Segment coding disabled for compred testing 572 573 // Enable ref frame features for segment 0 as well 574 vp9_enable_segfeature(seg, 0, SEG_LVL_REF_FRAME); 575 vp9_enable_segfeature(seg, 1, SEG_LVL_REF_FRAME); 576 577 // All mbs should use ALTREF_FRAME 578 vp9_clear_segdata(seg, 0, SEG_LVL_REF_FRAME); 579 vp9_set_segdata(seg, 0, SEG_LVL_REF_FRAME, ALTREF_FRAME); 580 vp9_clear_segdata(seg, 1, SEG_LVL_REF_FRAME); 581 vp9_set_segdata(seg, 1, SEG_LVL_REF_FRAME, ALTREF_FRAME); 582 583 // Skip all MBs if high Q (0,0 mv and skip coeffs) 584 if (high_q) { 585 vp9_enable_segfeature(seg, 0, SEG_LVL_SKIP); 586 vp9_enable_segfeature(seg, 1, SEG_LVL_SKIP); 587 } 588 // Enable data update 589 seg->update_data = 1; 590 } else { 591 // All other frames. 592 593 // No updates.. leave things as they are. 594 seg->update_map = 0; 595 seg->update_data = 0; 596 } 597 } 598} 599 600static void update_reference_segmentation_map(VP9_COMP *cpi) { 601 VP9_COMMON *const cm = &cpi->common; 602 MODE_INFO **mi_8x8_ptr = cm->mi_grid_visible; 603 uint8_t *cache_ptr = cm->last_frame_seg_map; 604 int row, col; 605 606 for (row = 0; row < cm->mi_rows; row++) { 607 MODE_INFO **mi_8x8 = mi_8x8_ptr; 608 uint8_t *cache = cache_ptr; 609 for (col = 0; col < cm->mi_cols; col++, mi_8x8++, cache++) 610 cache[0] = mi_8x8[0]->mbmi.segment_id; 611 mi_8x8_ptr += cm->mi_stride; 612 cache_ptr += cm->mi_cols; 613 } 614} 615 616static void alloc_raw_frame_buffers(VP9_COMP *cpi) { 617 VP9_COMMON *cm = &cpi->common; 618 const VP9EncoderConfig *oxcf = &cpi->oxcf; 619 620 if (!cpi->lookahead) 621 cpi->lookahead = vp9_lookahead_init(oxcf->width, oxcf->height, 622 cm->subsampling_x, cm->subsampling_y, 623#if CONFIG_VP9_HIGHBITDEPTH 624 cm->use_highbitdepth, 625#endif 626 oxcf->lag_in_frames); 627 if (!cpi->lookahead) 628 vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR, 629 "Failed to allocate lag buffers"); 630 631 // TODO(agrange) Check if ARF is enabled and skip allocation if not. 632 if (vpx_realloc_frame_buffer(&cpi->alt_ref_buffer, 633 oxcf->width, oxcf->height, 634 cm->subsampling_x, cm->subsampling_y, 635#if CONFIG_VP9_HIGHBITDEPTH 636 cm->use_highbitdepth, 637#endif 638 VP9_ENC_BORDER_IN_PIXELS, cm->byte_alignment, 639 NULL, NULL, NULL)) 640 vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR, 641 "Failed to allocate altref buffer"); 642} 643 644static void alloc_util_frame_buffers(VP9_COMP *cpi) { 645 VP9_COMMON *const cm = &cpi->common; 646 if (vpx_realloc_frame_buffer(&cpi->last_frame_uf, 647 cm->width, cm->height, 648 cm->subsampling_x, cm->subsampling_y, 649#if CONFIG_VP9_HIGHBITDEPTH 650 cm->use_highbitdepth, 651#endif 652 VP9_ENC_BORDER_IN_PIXELS, cm->byte_alignment, 653 NULL, NULL, NULL)) 654 vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR, 655 "Failed to allocate last frame buffer"); 656 657 if (vpx_realloc_frame_buffer(&cpi->scaled_source, 658 cm->width, cm->height, 659 cm->subsampling_x, cm->subsampling_y, 660#if CONFIG_VP9_HIGHBITDEPTH 661 cm->use_highbitdepth, 662#endif 663 VP9_ENC_BORDER_IN_PIXELS, cm->byte_alignment, 664 NULL, NULL, NULL)) 665 vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR, 666 "Failed to allocate scaled source buffer"); 667 668 if (vpx_realloc_frame_buffer(&cpi->scaled_last_source, 669 cm->width, cm->height, 670 cm->subsampling_x, cm->subsampling_y, 671#if CONFIG_VP9_HIGHBITDEPTH 672 cm->use_highbitdepth, 673#endif 674 VP9_ENC_BORDER_IN_PIXELS, cm->byte_alignment, 675 NULL, NULL, NULL)) 676 vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR, 677 "Failed to allocate scaled last source buffer"); 678} 679 680 681static int alloc_context_buffers_ext(VP9_COMP *cpi) { 682 VP9_COMMON *cm = &cpi->common; 683 int mi_size = cm->mi_cols * cm->mi_rows; 684 685 cpi->mbmi_ext_base = vpx_calloc(mi_size, sizeof(*cpi->mbmi_ext_base)); 686 if (!cpi->mbmi_ext_base) 687 return 1; 688 689 return 0; 690} 691 692static void alloc_compressor_data(VP9_COMP *cpi) { 693 VP9_COMMON *cm = &cpi->common; 694 695 vp9_alloc_context_buffers(cm, cm->width, cm->height); 696 697 alloc_context_buffers_ext(cpi); 698 699 vpx_free(cpi->tile_tok[0][0]); 700 701 { 702 unsigned int tokens = get_token_alloc(cm->mb_rows, cm->mb_cols); 703 CHECK_MEM_ERROR(cm, cpi->tile_tok[0][0], 704 vpx_calloc(tokens, sizeof(*cpi->tile_tok[0][0]))); 705 } 706 707 vp9_setup_pc_tree(&cpi->common, &cpi->td); 708} 709 710void vp9_new_framerate(VP9_COMP *cpi, double framerate) { 711 cpi->framerate = framerate < 0.1 ? 30 : framerate; 712 vp9_rc_update_framerate(cpi); 713} 714 715static void set_tile_limits(VP9_COMP *cpi) { 716 VP9_COMMON *const cm = &cpi->common; 717 718 int min_log2_tile_cols, max_log2_tile_cols; 719 vp9_get_tile_n_bits(cm->mi_cols, &min_log2_tile_cols, &max_log2_tile_cols); 720 721 if (is_two_pass_svc(cpi) && 722 (cpi->svc.encode_empty_frame_state == ENCODING || 723 cpi->svc.number_spatial_layers > 1)) { 724 cm->log2_tile_cols = 0; 725 cm->log2_tile_rows = 0; 726 } else { 727 cm->log2_tile_cols = clamp(cpi->oxcf.tile_columns, 728 min_log2_tile_cols, max_log2_tile_cols); 729 cm->log2_tile_rows = cpi->oxcf.tile_rows; 730 } 731} 732 733static void update_frame_size(VP9_COMP *cpi) { 734 VP9_COMMON *const cm = &cpi->common; 735 MACROBLOCKD *const xd = &cpi->td.mb.e_mbd; 736 737 vp9_set_mb_mi(cm, cm->width, cm->height); 738 vp9_init_context_buffers(cm); 739 vp9_init_macroblockd(cm, xd, NULL); 740 cpi->td.mb.mbmi_ext_base = cpi->mbmi_ext_base; 741 memset(cpi->mbmi_ext_base, 0, 742 cm->mi_rows * cm->mi_cols * sizeof(*cpi->mbmi_ext_base)); 743 744 set_tile_limits(cpi); 745 746 if (is_two_pass_svc(cpi)) { 747 if (vpx_realloc_frame_buffer(&cpi->alt_ref_buffer, 748 cm->width, cm->height, 749 cm->subsampling_x, cm->subsampling_y, 750#if CONFIG_VP9_HIGHBITDEPTH 751 cm->use_highbitdepth, 752#endif 753 VP9_ENC_BORDER_IN_PIXELS, cm->byte_alignment, 754 NULL, NULL, NULL)) 755 vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR, 756 "Failed to reallocate alt_ref_buffer"); 757 } 758} 759 760static void init_buffer_indices(VP9_COMP *cpi) { 761 cpi->lst_fb_idx = 0; 762 cpi->gld_fb_idx = 1; 763 cpi->alt_fb_idx = 2; 764} 765 766static void init_config(struct VP9_COMP *cpi, VP9EncoderConfig *oxcf) { 767 VP9_COMMON *const cm = &cpi->common; 768 769 cpi->oxcf = *oxcf; 770 cpi->framerate = oxcf->init_framerate; 771 772 cm->profile = oxcf->profile; 773 cm->bit_depth = oxcf->bit_depth; 774#if CONFIG_VP9_HIGHBITDEPTH 775 cm->use_highbitdepth = oxcf->use_highbitdepth; 776#endif 777 cm->color_space = oxcf->color_space; 778 cm->color_range = oxcf->color_range; 779 780 cm->width = oxcf->width; 781 cm->height = oxcf->height; 782 alloc_compressor_data(cpi); 783 784 cpi->svc.temporal_layering_mode = oxcf->temporal_layering_mode; 785 786 // Single thread case: use counts in common. 787 cpi->td.counts = &cm->counts; 788 789 // Spatial scalability. 790 cpi->svc.number_spatial_layers = oxcf->ss_number_layers; 791 // Temporal scalability. 792 cpi->svc.number_temporal_layers = oxcf->ts_number_layers; 793 794 if ((cpi->svc.number_temporal_layers > 1 && cpi->oxcf.rc_mode == VPX_CBR) || 795 ((cpi->svc.number_temporal_layers > 1 || 796 cpi->svc.number_spatial_layers > 1) && 797 cpi->oxcf.pass != 1)) { 798 vp9_init_layer_context(cpi); 799 } 800 801 // change includes all joint functionality 802 vp9_change_config(cpi, oxcf); 803 804 cpi->static_mb_pct = 0; 805 cpi->ref_frame_flags = 0; 806 807 init_buffer_indices(cpi); 808} 809 810static void set_rc_buffer_sizes(RATE_CONTROL *rc, 811 const VP9EncoderConfig *oxcf) { 812 const int64_t bandwidth = oxcf->target_bandwidth; 813 const int64_t starting = oxcf->starting_buffer_level_ms; 814 const int64_t optimal = oxcf->optimal_buffer_level_ms; 815 const int64_t maximum = oxcf->maximum_buffer_size_ms; 816 817 rc->starting_buffer_level = starting * bandwidth / 1000; 818 rc->optimal_buffer_level = (optimal == 0) ? bandwidth / 8 819 : optimal * bandwidth / 1000; 820 rc->maximum_buffer_size = (maximum == 0) ? bandwidth / 8 821 : maximum * bandwidth / 1000; 822} 823 824#if CONFIG_VP9_HIGHBITDEPTH 825#define HIGHBD_BFP(BT, SDF, SDAF, VF, SVF, SVAF, SDX3F, SDX8F, SDX4DF) \ 826 cpi->fn_ptr[BT].sdf = SDF; \ 827 cpi->fn_ptr[BT].sdaf = SDAF; \ 828 cpi->fn_ptr[BT].vf = VF; \ 829 cpi->fn_ptr[BT].svf = SVF; \ 830 cpi->fn_ptr[BT].svaf = SVAF; \ 831 cpi->fn_ptr[BT].sdx3f = SDX3F; \ 832 cpi->fn_ptr[BT].sdx8f = SDX8F; \ 833 cpi->fn_ptr[BT].sdx4df = SDX4DF; 834 835#define MAKE_BFP_SAD_WRAPPER(fnname) \ 836static unsigned int fnname##_bits8(const uint8_t *src_ptr, \ 837 int source_stride, \ 838 const uint8_t *ref_ptr, \ 839 int ref_stride) { \ 840 return fnname(src_ptr, source_stride, ref_ptr, ref_stride); \ 841} \ 842static unsigned int fnname##_bits10(const uint8_t *src_ptr, \ 843 int source_stride, \ 844 const uint8_t *ref_ptr, \ 845 int ref_stride) { \ 846 return fnname(src_ptr, source_stride, ref_ptr, ref_stride) >> 2; \ 847} \ 848static unsigned int fnname##_bits12(const uint8_t *src_ptr, \ 849 int source_stride, \ 850 const uint8_t *ref_ptr, \ 851 int ref_stride) { \ 852 return fnname(src_ptr, source_stride, ref_ptr, ref_stride) >> 4; \ 853} 854 855#define MAKE_BFP_SADAVG_WRAPPER(fnname) static unsigned int \ 856fnname##_bits8(const uint8_t *src_ptr, \ 857 int source_stride, \ 858 const uint8_t *ref_ptr, \ 859 int ref_stride, \ 860 const uint8_t *second_pred) { \ 861 return fnname(src_ptr, source_stride, ref_ptr, ref_stride, second_pred); \ 862} \ 863static unsigned int fnname##_bits10(const uint8_t *src_ptr, \ 864 int source_stride, \ 865 const uint8_t *ref_ptr, \ 866 int ref_stride, \ 867 const uint8_t *second_pred) { \ 868 return fnname(src_ptr, source_stride, ref_ptr, ref_stride, \ 869 second_pred) >> 2; \ 870} \ 871static unsigned int fnname##_bits12(const uint8_t *src_ptr, \ 872 int source_stride, \ 873 const uint8_t *ref_ptr, \ 874 int ref_stride, \ 875 const uint8_t *second_pred) { \ 876 return fnname(src_ptr, source_stride, ref_ptr, ref_stride, \ 877 second_pred) >> 4; \ 878} 879 880#define MAKE_BFP_SAD3_WRAPPER(fnname) \ 881static void fnname##_bits8(const uint8_t *src_ptr, \ 882 int source_stride, \ 883 const uint8_t *ref_ptr, \ 884 int ref_stride, \ 885 unsigned int *sad_array) { \ 886 fnname(src_ptr, source_stride, ref_ptr, ref_stride, sad_array); \ 887} \ 888static void fnname##_bits10(const uint8_t *src_ptr, \ 889 int source_stride, \ 890 const uint8_t *ref_ptr, \ 891 int ref_stride, \ 892 unsigned int *sad_array) { \ 893 int i; \ 894 fnname(src_ptr, source_stride, ref_ptr, ref_stride, sad_array); \ 895 for (i = 0; i < 3; i++) \ 896 sad_array[i] >>= 2; \ 897} \ 898static void fnname##_bits12(const uint8_t *src_ptr, \ 899 int source_stride, \ 900 const uint8_t *ref_ptr, \ 901 int ref_stride, \ 902 unsigned int *sad_array) { \ 903 int i; \ 904 fnname(src_ptr, source_stride, ref_ptr, ref_stride, sad_array); \ 905 for (i = 0; i < 3; i++) \ 906 sad_array[i] >>= 4; \ 907} 908 909#define MAKE_BFP_SAD8_WRAPPER(fnname) \ 910static void fnname##_bits8(const uint8_t *src_ptr, \ 911 int source_stride, \ 912 const uint8_t *ref_ptr, \ 913 int ref_stride, \ 914 unsigned int *sad_array) { \ 915 fnname(src_ptr, source_stride, ref_ptr, ref_stride, sad_array); \ 916} \ 917static void fnname##_bits10(const uint8_t *src_ptr, \ 918 int source_stride, \ 919 const uint8_t *ref_ptr, \ 920 int ref_stride, \ 921 unsigned int *sad_array) { \ 922 int i; \ 923 fnname(src_ptr, source_stride, ref_ptr, ref_stride, sad_array); \ 924 for (i = 0; i < 8; i++) \ 925 sad_array[i] >>= 2; \ 926} \ 927static void fnname##_bits12(const uint8_t *src_ptr, \ 928 int source_stride, \ 929 const uint8_t *ref_ptr, \ 930 int ref_stride, \ 931 unsigned int *sad_array) { \ 932 int i; \ 933 fnname(src_ptr, source_stride, ref_ptr, ref_stride, sad_array); \ 934 for (i = 0; i < 8; i++) \ 935 sad_array[i] >>= 4; \ 936} 937#define MAKE_BFP_SAD4D_WRAPPER(fnname) \ 938static void fnname##_bits8(const uint8_t *src_ptr, \ 939 int source_stride, \ 940 const uint8_t* const ref_ptr[], \ 941 int ref_stride, \ 942 unsigned int *sad_array) { \ 943 fnname(src_ptr, source_stride, ref_ptr, ref_stride, sad_array); \ 944} \ 945static void fnname##_bits10(const uint8_t *src_ptr, \ 946 int source_stride, \ 947 const uint8_t* const ref_ptr[], \ 948 int ref_stride, \ 949 unsigned int *sad_array) { \ 950 int i; \ 951 fnname(src_ptr, source_stride, ref_ptr, ref_stride, sad_array); \ 952 for (i = 0; i < 4; i++) \ 953 sad_array[i] >>= 2; \ 954} \ 955static void fnname##_bits12(const uint8_t *src_ptr, \ 956 int source_stride, \ 957 const uint8_t* const ref_ptr[], \ 958 int ref_stride, \ 959 unsigned int *sad_array) { \ 960 int i; \ 961 fnname(src_ptr, source_stride, ref_ptr, ref_stride, sad_array); \ 962 for (i = 0; i < 4; i++) \ 963 sad_array[i] >>= 4; \ 964} 965 966MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad32x16) 967MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad32x16_avg) 968MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad32x16x4d) 969MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad16x32) 970MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad16x32_avg) 971MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad16x32x4d) 972MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad64x32) 973MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad64x32_avg) 974MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad64x32x4d) 975MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad32x64) 976MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad32x64_avg) 977MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad32x64x4d) 978MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad32x32) 979MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad32x32_avg) 980MAKE_BFP_SAD3_WRAPPER(vpx_highbd_sad32x32x3) 981MAKE_BFP_SAD8_WRAPPER(vpx_highbd_sad32x32x8) 982MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad32x32x4d) 983MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad64x64) 984MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad64x64_avg) 985MAKE_BFP_SAD3_WRAPPER(vpx_highbd_sad64x64x3) 986MAKE_BFP_SAD8_WRAPPER(vpx_highbd_sad64x64x8) 987MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad64x64x4d) 988MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad16x16) 989MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad16x16_avg) 990MAKE_BFP_SAD3_WRAPPER(vpx_highbd_sad16x16x3) 991MAKE_BFP_SAD8_WRAPPER(vpx_highbd_sad16x16x8) 992MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad16x16x4d) 993MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad16x8) 994MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad16x8_avg) 995MAKE_BFP_SAD3_WRAPPER(vpx_highbd_sad16x8x3) 996MAKE_BFP_SAD8_WRAPPER(vpx_highbd_sad16x8x8) 997MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad16x8x4d) 998MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad8x16) 999MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad8x16_avg) 1000MAKE_BFP_SAD3_WRAPPER(vpx_highbd_sad8x16x3) 1001MAKE_BFP_SAD8_WRAPPER(vpx_highbd_sad8x16x8) 1002MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad8x16x4d) 1003MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad8x8) 1004MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad8x8_avg) 1005MAKE_BFP_SAD3_WRAPPER(vpx_highbd_sad8x8x3) 1006MAKE_BFP_SAD8_WRAPPER(vpx_highbd_sad8x8x8) 1007MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad8x8x4d) 1008MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad8x4) 1009MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad8x4_avg) 1010MAKE_BFP_SAD8_WRAPPER(vpx_highbd_sad8x4x8) 1011MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad8x4x4d) 1012MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad4x8) 1013MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad4x8_avg) 1014MAKE_BFP_SAD8_WRAPPER(vpx_highbd_sad4x8x8) 1015MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad4x8x4d) 1016MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad4x4) 1017MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad4x4_avg) 1018MAKE_BFP_SAD3_WRAPPER(vpx_highbd_sad4x4x3) 1019MAKE_BFP_SAD8_WRAPPER(vpx_highbd_sad4x4x8) 1020MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad4x4x4d) 1021 1022static void highbd_set_var_fns(VP9_COMP *const cpi) { 1023 VP9_COMMON *const cm = &cpi->common; 1024 if (cm->use_highbitdepth) { 1025 switch (cm->bit_depth) { 1026 case VPX_BITS_8: 1027 HIGHBD_BFP(BLOCK_32X16, 1028 vpx_highbd_sad32x16_bits8, 1029 vpx_highbd_sad32x16_avg_bits8, 1030 vpx_highbd_8_variance32x16, 1031 vpx_highbd_8_sub_pixel_variance32x16, 1032 vpx_highbd_8_sub_pixel_avg_variance32x16, 1033 NULL, 1034 NULL, 1035 vpx_highbd_sad32x16x4d_bits8) 1036 1037 HIGHBD_BFP(BLOCK_16X32, 1038 vpx_highbd_sad16x32_bits8, 1039 vpx_highbd_sad16x32_avg_bits8, 1040 vpx_highbd_8_variance16x32, 1041 vpx_highbd_8_sub_pixel_variance16x32, 1042 vpx_highbd_8_sub_pixel_avg_variance16x32, 1043 NULL, 1044 NULL, 1045 vpx_highbd_sad16x32x4d_bits8) 1046 1047 HIGHBD_BFP(BLOCK_64X32, 1048 vpx_highbd_sad64x32_bits8, 1049 vpx_highbd_sad64x32_avg_bits8, 1050 vpx_highbd_8_variance64x32, 1051 vpx_highbd_8_sub_pixel_variance64x32, 1052 vpx_highbd_8_sub_pixel_avg_variance64x32, 1053 NULL, 1054 NULL, 1055 vpx_highbd_sad64x32x4d_bits8) 1056 1057 HIGHBD_BFP(BLOCK_32X64, 1058 vpx_highbd_sad32x64_bits8, 1059 vpx_highbd_sad32x64_avg_bits8, 1060 vpx_highbd_8_variance32x64, 1061 vpx_highbd_8_sub_pixel_variance32x64, 1062 vpx_highbd_8_sub_pixel_avg_variance32x64, 1063 NULL, 1064 NULL, 1065 vpx_highbd_sad32x64x4d_bits8) 1066 1067 HIGHBD_BFP(BLOCK_32X32, 1068 vpx_highbd_sad32x32_bits8, 1069 vpx_highbd_sad32x32_avg_bits8, 1070 vpx_highbd_8_variance32x32, 1071 vpx_highbd_8_sub_pixel_variance32x32, 1072 vpx_highbd_8_sub_pixel_avg_variance32x32, 1073 vpx_highbd_sad32x32x3_bits8, 1074 vpx_highbd_sad32x32x8_bits8, 1075 vpx_highbd_sad32x32x4d_bits8) 1076 1077 HIGHBD_BFP(BLOCK_64X64, 1078 vpx_highbd_sad64x64_bits8, 1079 vpx_highbd_sad64x64_avg_bits8, 1080 vpx_highbd_8_variance64x64, 1081 vpx_highbd_8_sub_pixel_variance64x64, 1082 vpx_highbd_8_sub_pixel_avg_variance64x64, 1083 vpx_highbd_sad64x64x3_bits8, 1084 vpx_highbd_sad64x64x8_bits8, 1085 vpx_highbd_sad64x64x4d_bits8) 1086 1087 HIGHBD_BFP(BLOCK_16X16, 1088 vpx_highbd_sad16x16_bits8, 1089 vpx_highbd_sad16x16_avg_bits8, 1090 vpx_highbd_8_variance16x16, 1091 vpx_highbd_8_sub_pixel_variance16x16, 1092 vpx_highbd_8_sub_pixel_avg_variance16x16, 1093 vpx_highbd_sad16x16x3_bits8, 1094 vpx_highbd_sad16x16x8_bits8, 1095 vpx_highbd_sad16x16x4d_bits8) 1096 1097 HIGHBD_BFP(BLOCK_16X8, 1098 vpx_highbd_sad16x8_bits8, 1099 vpx_highbd_sad16x8_avg_bits8, 1100 vpx_highbd_8_variance16x8, 1101 vpx_highbd_8_sub_pixel_variance16x8, 1102 vpx_highbd_8_sub_pixel_avg_variance16x8, 1103 vpx_highbd_sad16x8x3_bits8, 1104 vpx_highbd_sad16x8x8_bits8, 1105 vpx_highbd_sad16x8x4d_bits8) 1106 1107 HIGHBD_BFP(BLOCK_8X16, 1108 vpx_highbd_sad8x16_bits8, 1109 vpx_highbd_sad8x16_avg_bits8, 1110 vpx_highbd_8_variance8x16, 1111 vpx_highbd_8_sub_pixel_variance8x16, 1112 vpx_highbd_8_sub_pixel_avg_variance8x16, 1113 vpx_highbd_sad8x16x3_bits8, 1114 vpx_highbd_sad8x16x8_bits8, 1115 vpx_highbd_sad8x16x4d_bits8) 1116 1117 HIGHBD_BFP(BLOCK_8X8, 1118 vpx_highbd_sad8x8_bits8, 1119 vpx_highbd_sad8x8_avg_bits8, 1120 vpx_highbd_8_variance8x8, 1121 vpx_highbd_8_sub_pixel_variance8x8, 1122 vpx_highbd_8_sub_pixel_avg_variance8x8, 1123 vpx_highbd_sad8x8x3_bits8, 1124 vpx_highbd_sad8x8x8_bits8, 1125 vpx_highbd_sad8x8x4d_bits8) 1126 1127 HIGHBD_BFP(BLOCK_8X4, 1128 vpx_highbd_sad8x4_bits8, 1129 vpx_highbd_sad8x4_avg_bits8, 1130 vpx_highbd_8_variance8x4, 1131 vpx_highbd_8_sub_pixel_variance8x4, 1132 vpx_highbd_8_sub_pixel_avg_variance8x4, 1133 NULL, 1134 vpx_highbd_sad8x4x8_bits8, 1135 vpx_highbd_sad8x4x4d_bits8) 1136 1137 HIGHBD_BFP(BLOCK_4X8, 1138 vpx_highbd_sad4x8_bits8, 1139 vpx_highbd_sad4x8_avg_bits8, 1140 vpx_highbd_8_variance4x8, 1141 vpx_highbd_8_sub_pixel_variance4x8, 1142 vpx_highbd_8_sub_pixel_avg_variance4x8, 1143 NULL, 1144 vpx_highbd_sad4x8x8_bits8, 1145 vpx_highbd_sad4x8x4d_bits8) 1146 1147 HIGHBD_BFP(BLOCK_4X4, 1148 vpx_highbd_sad4x4_bits8, 1149 vpx_highbd_sad4x4_avg_bits8, 1150 vpx_highbd_8_variance4x4, 1151 vpx_highbd_8_sub_pixel_variance4x4, 1152 vpx_highbd_8_sub_pixel_avg_variance4x4, 1153 vpx_highbd_sad4x4x3_bits8, 1154 vpx_highbd_sad4x4x8_bits8, 1155 vpx_highbd_sad4x4x4d_bits8) 1156 break; 1157 1158 case VPX_BITS_10: 1159 HIGHBD_BFP(BLOCK_32X16, 1160 vpx_highbd_sad32x16_bits10, 1161 vpx_highbd_sad32x16_avg_bits10, 1162 vpx_highbd_10_variance32x16, 1163 vpx_highbd_10_sub_pixel_variance32x16, 1164 vpx_highbd_10_sub_pixel_avg_variance32x16, 1165 NULL, 1166 NULL, 1167 vpx_highbd_sad32x16x4d_bits10) 1168 1169 HIGHBD_BFP(BLOCK_16X32, 1170 vpx_highbd_sad16x32_bits10, 1171 vpx_highbd_sad16x32_avg_bits10, 1172 vpx_highbd_10_variance16x32, 1173 vpx_highbd_10_sub_pixel_variance16x32, 1174 vpx_highbd_10_sub_pixel_avg_variance16x32, 1175 NULL, 1176 NULL, 1177 vpx_highbd_sad16x32x4d_bits10) 1178 1179 HIGHBD_BFP(BLOCK_64X32, 1180 vpx_highbd_sad64x32_bits10, 1181 vpx_highbd_sad64x32_avg_bits10, 1182 vpx_highbd_10_variance64x32, 1183 vpx_highbd_10_sub_pixel_variance64x32, 1184 vpx_highbd_10_sub_pixel_avg_variance64x32, 1185 NULL, 1186 NULL, 1187 vpx_highbd_sad64x32x4d_bits10) 1188 1189 HIGHBD_BFP(BLOCK_32X64, 1190 vpx_highbd_sad32x64_bits10, 1191 vpx_highbd_sad32x64_avg_bits10, 1192 vpx_highbd_10_variance32x64, 1193 vpx_highbd_10_sub_pixel_variance32x64, 1194 vpx_highbd_10_sub_pixel_avg_variance32x64, 1195 NULL, 1196 NULL, 1197 vpx_highbd_sad32x64x4d_bits10) 1198 1199 HIGHBD_BFP(BLOCK_32X32, 1200 vpx_highbd_sad32x32_bits10, 1201 vpx_highbd_sad32x32_avg_bits10, 1202 vpx_highbd_10_variance32x32, 1203 vpx_highbd_10_sub_pixel_variance32x32, 1204 vpx_highbd_10_sub_pixel_avg_variance32x32, 1205 vpx_highbd_sad32x32x3_bits10, 1206 vpx_highbd_sad32x32x8_bits10, 1207 vpx_highbd_sad32x32x4d_bits10) 1208 1209 HIGHBD_BFP(BLOCK_64X64, 1210 vpx_highbd_sad64x64_bits10, 1211 vpx_highbd_sad64x64_avg_bits10, 1212 vpx_highbd_10_variance64x64, 1213 vpx_highbd_10_sub_pixel_variance64x64, 1214 vpx_highbd_10_sub_pixel_avg_variance64x64, 1215 vpx_highbd_sad64x64x3_bits10, 1216 vpx_highbd_sad64x64x8_bits10, 1217 vpx_highbd_sad64x64x4d_bits10) 1218 1219 HIGHBD_BFP(BLOCK_16X16, 1220 vpx_highbd_sad16x16_bits10, 1221 vpx_highbd_sad16x16_avg_bits10, 1222 vpx_highbd_10_variance16x16, 1223 vpx_highbd_10_sub_pixel_variance16x16, 1224 vpx_highbd_10_sub_pixel_avg_variance16x16, 1225 vpx_highbd_sad16x16x3_bits10, 1226 vpx_highbd_sad16x16x8_bits10, 1227 vpx_highbd_sad16x16x4d_bits10) 1228 1229 HIGHBD_BFP(BLOCK_16X8, 1230 vpx_highbd_sad16x8_bits10, 1231 vpx_highbd_sad16x8_avg_bits10, 1232 vpx_highbd_10_variance16x8, 1233 vpx_highbd_10_sub_pixel_variance16x8, 1234 vpx_highbd_10_sub_pixel_avg_variance16x8, 1235 vpx_highbd_sad16x8x3_bits10, 1236 vpx_highbd_sad16x8x8_bits10, 1237 vpx_highbd_sad16x8x4d_bits10) 1238 1239 HIGHBD_BFP(BLOCK_8X16, 1240 vpx_highbd_sad8x16_bits10, 1241 vpx_highbd_sad8x16_avg_bits10, 1242 vpx_highbd_10_variance8x16, 1243 vpx_highbd_10_sub_pixel_variance8x16, 1244 vpx_highbd_10_sub_pixel_avg_variance8x16, 1245 vpx_highbd_sad8x16x3_bits10, 1246 vpx_highbd_sad8x16x8_bits10, 1247 vpx_highbd_sad8x16x4d_bits10) 1248 1249 HIGHBD_BFP(BLOCK_8X8, 1250 vpx_highbd_sad8x8_bits10, 1251 vpx_highbd_sad8x8_avg_bits10, 1252 vpx_highbd_10_variance8x8, 1253 vpx_highbd_10_sub_pixel_variance8x8, 1254 vpx_highbd_10_sub_pixel_avg_variance8x8, 1255 vpx_highbd_sad8x8x3_bits10, 1256 vpx_highbd_sad8x8x8_bits10, 1257 vpx_highbd_sad8x8x4d_bits10) 1258 1259 HIGHBD_BFP(BLOCK_8X4, 1260 vpx_highbd_sad8x4_bits10, 1261 vpx_highbd_sad8x4_avg_bits10, 1262 vpx_highbd_10_variance8x4, 1263 vpx_highbd_10_sub_pixel_variance8x4, 1264 vpx_highbd_10_sub_pixel_avg_variance8x4, 1265 NULL, 1266 vpx_highbd_sad8x4x8_bits10, 1267 vpx_highbd_sad8x4x4d_bits10) 1268 1269 HIGHBD_BFP(BLOCK_4X8, 1270 vpx_highbd_sad4x8_bits10, 1271 vpx_highbd_sad4x8_avg_bits10, 1272 vpx_highbd_10_variance4x8, 1273 vpx_highbd_10_sub_pixel_variance4x8, 1274 vpx_highbd_10_sub_pixel_avg_variance4x8, 1275 NULL, 1276 vpx_highbd_sad4x8x8_bits10, 1277 vpx_highbd_sad4x8x4d_bits10) 1278 1279 HIGHBD_BFP(BLOCK_4X4, 1280 vpx_highbd_sad4x4_bits10, 1281 vpx_highbd_sad4x4_avg_bits10, 1282 vpx_highbd_10_variance4x4, 1283 vpx_highbd_10_sub_pixel_variance4x4, 1284 vpx_highbd_10_sub_pixel_avg_variance4x4, 1285 vpx_highbd_sad4x4x3_bits10, 1286 vpx_highbd_sad4x4x8_bits10, 1287 vpx_highbd_sad4x4x4d_bits10) 1288 break; 1289 1290 case VPX_BITS_12: 1291 HIGHBD_BFP(BLOCK_32X16, 1292 vpx_highbd_sad32x16_bits12, 1293 vpx_highbd_sad32x16_avg_bits12, 1294 vpx_highbd_12_variance32x16, 1295 vpx_highbd_12_sub_pixel_variance32x16, 1296 vpx_highbd_12_sub_pixel_avg_variance32x16, 1297 NULL, 1298 NULL, 1299 vpx_highbd_sad32x16x4d_bits12) 1300 1301 HIGHBD_BFP(BLOCK_16X32, 1302 vpx_highbd_sad16x32_bits12, 1303 vpx_highbd_sad16x32_avg_bits12, 1304 vpx_highbd_12_variance16x32, 1305 vpx_highbd_12_sub_pixel_variance16x32, 1306 vpx_highbd_12_sub_pixel_avg_variance16x32, 1307 NULL, 1308 NULL, 1309 vpx_highbd_sad16x32x4d_bits12) 1310 1311 HIGHBD_BFP(BLOCK_64X32, 1312 vpx_highbd_sad64x32_bits12, 1313 vpx_highbd_sad64x32_avg_bits12, 1314 vpx_highbd_12_variance64x32, 1315 vpx_highbd_12_sub_pixel_variance64x32, 1316 vpx_highbd_12_sub_pixel_avg_variance64x32, 1317 NULL, 1318 NULL, 1319 vpx_highbd_sad64x32x4d_bits12) 1320 1321 HIGHBD_BFP(BLOCK_32X64, 1322 vpx_highbd_sad32x64_bits12, 1323 vpx_highbd_sad32x64_avg_bits12, 1324 vpx_highbd_12_variance32x64, 1325 vpx_highbd_12_sub_pixel_variance32x64, 1326 vpx_highbd_12_sub_pixel_avg_variance32x64, 1327 NULL, 1328 NULL, 1329 vpx_highbd_sad32x64x4d_bits12) 1330 1331 HIGHBD_BFP(BLOCK_32X32, 1332 vpx_highbd_sad32x32_bits12, 1333 vpx_highbd_sad32x32_avg_bits12, 1334 vpx_highbd_12_variance32x32, 1335 vpx_highbd_12_sub_pixel_variance32x32, 1336 vpx_highbd_12_sub_pixel_avg_variance32x32, 1337 vpx_highbd_sad32x32x3_bits12, 1338 vpx_highbd_sad32x32x8_bits12, 1339 vpx_highbd_sad32x32x4d_bits12) 1340 1341 HIGHBD_BFP(BLOCK_64X64, 1342 vpx_highbd_sad64x64_bits12, 1343 vpx_highbd_sad64x64_avg_bits12, 1344 vpx_highbd_12_variance64x64, 1345 vpx_highbd_12_sub_pixel_variance64x64, 1346 vpx_highbd_12_sub_pixel_avg_variance64x64, 1347 vpx_highbd_sad64x64x3_bits12, 1348 vpx_highbd_sad64x64x8_bits12, 1349 vpx_highbd_sad64x64x4d_bits12) 1350 1351 HIGHBD_BFP(BLOCK_16X16, 1352 vpx_highbd_sad16x16_bits12, 1353 vpx_highbd_sad16x16_avg_bits12, 1354 vpx_highbd_12_variance16x16, 1355 vpx_highbd_12_sub_pixel_variance16x16, 1356 vpx_highbd_12_sub_pixel_avg_variance16x16, 1357 vpx_highbd_sad16x16x3_bits12, 1358 vpx_highbd_sad16x16x8_bits12, 1359 vpx_highbd_sad16x16x4d_bits12) 1360 1361 HIGHBD_BFP(BLOCK_16X8, 1362 vpx_highbd_sad16x8_bits12, 1363 vpx_highbd_sad16x8_avg_bits12, 1364 vpx_highbd_12_variance16x8, 1365 vpx_highbd_12_sub_pixel_variance16x8, 1366 vpx_highbd_12_sub_pixel_avg_variance16x8, 1367 vpx_highbd_sad16x8x3_bits12, 1368 vpx_highbd_sad16x8x8_bits12, 1369 vpx_highbd_sad16x8x4d_bits12) 1370 1371 HIGHBD_BFP(BLOCK_8X16, 1372 vpx_highbd_sad8x16_bits12, 1373 vpx_highbd_sad8x16_avg_bits12, 1374 vpx_highbd_12_variance8x16, 1375 vpx_highbd_12_sub_pixel_variance8x16, 1376 vpx_highbd_12_sub_pixel_avg_variance8x16, 1377 vpx_highbd_sad8x16x3_bits12, 1378 vpx_highbd_sad8x16x8_bits12, 1379 vpx_highbd_sad8x16x4d_bits12) 1380 1381 HIGHBD_BFP(BLOCK_8X8, 1382 vpx_highbd_sad8x8_bits12, 1383 vpx_highbd_sad8x8_avg_bits12, 1384 vpx_highbd_12_variance8x8, 1385 vpx_highbd_12_sub_pixel_variance8x8, 1386 vpx_highbd_12_sub_pixel_avg_variance8x8, 1387 vpx_highbd_sad8x8x3_bits12, 1388 vpx_highbd_sad8x8x8_bits12, 1389 vpx_highbd_sad8x8x4d_bits12) 1390 1391 HIGHBD_BFP(BLOCK_8X4, 1392 vpx_highbd_sad8x4_bits12, 1393 vpx_highbd_sad8x4_avg_bits12, 1394 vpx_highbd_12_variance8x4, 1395 vpx_highbd_12_sub_pixel_variance8x4, 1396 vpx_highbd_12_sub_pixel_avg_variance8x4, 1397 NULL, 1398 vpx_highbd_sad8x4x8_bits12, 1399 vpx_highbd_sad8x4x4d_bits12) 1400 1401 HIGHBD_BFP(BLOCK_4X8, 1402 vpx_highbd_sad4x8_bits12, 1403 vpx_highbd_sad4x8_avg_bits12, 1404 vpx_highbd_12_variance4x8, 1405 vpx_highbd_12_sub_pixel_variance4x8, 1406 vpx_highbd_12_sub_pixel_avg_variance4x8, 1407 NULL, 1408 vpx_highbd_sad4x8x8_bits12, 1409 vpx_highbd_sad4x8x4d_bits12) 1410 1411 HIGHBD_BFP(BLOCK_4X4, 1412 vpx_highbd_sad4x4_bits12, 1413 vpx_highbd_sad4x4_avg_bits12, 1414 vpx_highbd_12_variance4x4, 1415 vpx_highbd_12_sub_pixel_variance4x4, 1416 vpx_highbd_12_sub_pixel_avg_variance4x4, 1417 vpx_highbd_sad4x4x3_bits12, 1418 vpx_highbd_sad4x4x8_bits12, 1419 vpx_highbd_sad4x4x4d_bits12) 1420 break; 1421 1422 default: 1423 assert(0 && "cm->bit_depth should be VPX_BITS_8, " 1424 "VPX_BITS_10 or VPX_BITS_12"); 1425 } 1426 } 1427} 1428#endif // CONFIG_VP9_HIGHBITDEPTH 1429 1430static void realloc_segmentation_maps(VP9_COMP *cpi) { 1431 VP9_COMMON *const cm = &cpi->common; 1432 1433 // Create the encoder segmentation map and set all entries to 0 1434 vpx_free(cpi->segmentation_map); 1435 CHECK_MEM_ERROR(cm, cpi->segmentation_map, 1436 vpx_calloc(cm->mi_rows * cm->mi_cols, 1)); 1437 1438 // Create a map used for cyclic background refresh. 1439 if (cpi->cyclic_refresh) 1440 vp9_cyclic_refresh_free(cpi->cyclic_refresh); 1441 CHECK_MEM_ERROR(cm, cpi->cyclic_refresh, 1442 vp9_cyclic_refresh_alloc(cm->mi_rows, cm->mi_cols)); 1443 1444 // Create a map used to mark inactive areas. 1445 vpx_free(cpi->active_map.map); 1446 CHECK_MEM_ERROR(cm, cpi->active_map.map, 1447 vpx_calloc(cm->mi_rows * cm->mi_cols, 1)); 1448 1449 // And a place holder structure is the coding context 1450 // for use if we want to save and restore it 1451 vpx_free(cpi->coding_context.last_frame_seg_map_copy); 1452 CHECK_MEM_ERROR(cm, cpi->coding_context.last_frame_seg_map_copy, 1453 vpx_calloc(cm->mi_rows * cm->mi_cols, 1)); 1454} 1455 1456void vp9_change_config(struct VP9_COMP *cpi, const VP9EncoderConfig *oxcf) { 1457 VP9_COMMON *const cm = &cpi->common; 1458 RATE_CONTROL *const rc = &cpi->rc; 1459 int last_w = cpi->oxcf.width; 1460 int last_h = cpi->oxcf.height; 1461 1462 if (cm->profile != oxcf->profile) 1463 cm->profile = oxcf->profile; 1464 cm->bit_depth = oxcf->bit_depth; 1465 cm->color_space = oxcf->color_space; 1466 cm->color_range = oxcf->color_range; 1467 1468 if (cm->profile <= PROFILE_1) 1469 assert(cm->bit_depth == VPX_BITS_8); 1470 else 1471 assert(cm->bit_depth > VPX_BITS_8); 1472 1473 cpi->oxcf = *oxcf; 1474#if CONFIG_VP9_HIGHBITDEPTH 1475 cpi->td.mb.e_mbd.bd = (int)cm->bit_depth; 1476#endif // CONFIG_VP9_HIGHBITDEPTH 1477 1478 rc->baseline_gf_interval = (MIN_GF_INTERVAL + MAX_GF_INTERVAL) / 2; 1479 1480 cpi->refresh_golden_frame = 0; 1481 cpi->refresh_last_frame = 1; 1482 cm->refresh_frame_context = 1; 1483 cm->reset_frame_context = 0; 1484 1485 vp9_reset_segment_features(&cm->seg); 1486 vp9_set_high_precision_mv(cpi, 0); 1487 1488 { 1489 int i; 1490 1491 for (i = 0; i < MAX_SEGMENTS; i++) 1492 cpi->segment_encode_breakout[i] = cpi->oxcf.encode_breakout; 1493 } 1494 cpi->encode_breakout = cpi->oxcf.encode_breakout; 1495 1496 set_rc_buffer_sizes(rc, &cpi->oxcf); 1497 1498 // Under a configuration change, where maximum_buffer_size may change, 1499 // keep buffer level clipped to the maximum allowed buffer size. 1500 rc->bits_off_target = VPXMIN(rc->bits_off_target, rc->maximum_buffer_size); 1501 rc->buffer_level = VPXMIN(rc->buffer_level, rc->maximum_buffer_size); 1502 1503 // Set up frame rate and related parameters rate control values. 1504 vp9_new_framerate(cpi, cpi->framerate); 1505 1506 // Set absolute upper and lower quality limits 1507 rc->worst_quality = cpi->oxcf.worst_allowed_q; 1508 rc->best_quality = cpi->oxcf.best_allowed_q; 1509 1510 cm->interp_filter = cpi->sf.default_interp_filter; 1511 1512 if (cpi->oxcf.render_width > 0 && cpi->oxcf.render_height > 0) { 1513 cm->render_width = cpi->oxcf.render_width; 1514 cm->render_height = cpi->oxcf.render_height; 1515 } else { 1516 cm->render_width = cpi->oxcf.width; 1517 cm->render_height = cpi->oxcf.height; 1518 } 1519 if (last_w != cpi->oxcf.width || last_h != cpi->oxcf.height) { 1520 cm->width = cpi->oxcf.width; 1521 cm->height = cpi->oxcf.height; 1522 } 1523 1524 if (cpi->initial_width) { 1525 int new_mi_size = 0; 1526 vp9_set_mb_mi(cm, cm->width, cm->height); 1527 new_mi_size = cm->mi_stride * calc_mi_size(cm->mi_rows); 1528 if (cm->mi_alloc_size < new_mi_size) { 1529 vp9_free_context_buffers(cm); 1530 alloc_compressor_data(cpi); 1531 realloc_segmentation_maps(cpi); 1532 cpi->initial_width = cpi->initial_height = 0; 1533 } 1534 } 1535 update_frame_size(cpi); 1536 1537 if ((cpi->svc.number_temporal_layers > 1 && 1538 cpi->oxcf.rc_mode == VPX_CBR) || 1539 ((cpi->svc.number_temporal_layers > 1 || 1540 cpi->svc.number_spatial_layers > 1) && 1541 cpi->oxcf.pass != 1)) { 1542 vp9_update_layer_context_change_config(cpi, 1543 (int)cpi->oxcf.target_bandwidth); 1544 } 1545 1546 cpi->alt_ref_source = NULL; 1547 rc->is_src_frame_alt_ref = 0; 1548 1549#if 0 1550 // Experimental RD Code 1551 cpi->frame_distortion = 0; 1552 cpi->last_frame_distortion = 0; 1553#endif 1554 1555 set_tile_limits(cpi); 1556 1557 cpi->ext_refresh_frame_flags_pending = 0; 1558 cpi->ext_refresh_frame_context_pending = 0; 1559 1560#if CONFIG_VP9_HIGHBITDEPTH 1561 highbd_set_var_fns(cpi); 1562#endif 1563} 1564 1565#ifndef M_LOG2_E 1566#define M_LOG2_E 0.693147180559945309417 1567#endif 1568#define log2f(x) (log (x) / (float) M_LOG2_E) 1569 1570static void cal_nmvjointsadcost(int *mvjointsadcost) { 1571 mvjointsadcost[0] = 600; 1572 mvjointsadcost[1] = 300; 1573 mvjointsadcost[2] = 300; 1574 mvjointsadcost[3] = 300; 1575} 1576 1577static void cal_nmvsadcosts(int *mvsadcost[2]) { 1578 int i = 1; 1579 1580 mvsadcost[0][0] = 0; 1581 mvsadcost[1][0] = 0; 1582 1583 do { 1584 double z = 256 * (2 * (log2f(8 * i) + .6)); 1585 mvsadcost[0][i] = (int)z; 1586 mvsadcost[1][i] = (int)z; 1587 mvsadcost[0][-i] = (int)z; 1588 mvsadcost[1][-i] = (int)z; 1589 } while (++i <= MV_MAX); 1590} 1591 1592static void cal_nmvsadcosts_hp(int *mvsadcost[2]) { 1593 int i = 1; 1594 1595 mvsadcost[0][0] = 0; 1596 mvsadcost[1][0] = 0; 1597 1598 do { 1599 double z = 256 * (2 * (log2f(8 * i) + .6)); 1600 mvsadcost[0][i] = (int)z; 1601 mvsadcost[1][i] = (int)z; 1602 mvsadcost[0][-i] = (int)z; 1603 mvsadcost[1][-i] = (int)z; 1604 } while (++i <= MV_MAX); 1605} 1606 1607 1608VP9_COMP *vp9_create_compressor(VP9EncoderConfig *oxcf, 1609 BufferPool *const pool) { 1610 unsigned int i; 1611 VP9_COMP *volatile const cpi = vpx_memalign(32, sizeof(VP9_COMP)); 1612 VP9_COMMON *volatile const cm = cpi != NULL ? &cpi->common : NULL; 1613 1614 if (!cm) 1615 return NULL; 1616 1617 vp9_zero(*cpi); 1618 1619 if (setjmp(cm->error.jmp)) { 1620 cm->error.setjmp = 0; 1621 vp9_remove_compressor(cpi); 1622 return 0; 1623 } 1624 1625 cm->error.setjmp = 1; 1626 cm->alloc_mi = vp9_enc_alloc_mi; 1627 cm->free_mi = vp9_enc_free_mi; 1628 cm->setup_mi = vp9_enc_setup_mi; 1629 1630 CHECK_MEM_ERROR(cm, cm->fc, 1631 (FRAME_CONTEXT *)vpx_calloc(1, sizeof(*cm->fc))); 1632 CHECK_MEM_ERROR(cm, cm->frame_contexts, 1633 (FRAME_CONTEXT *)vpx_calloc(FRAME_CONTEXTS, 1634 sizeof(*cm->frame_contexts))); 1635 1636 cpi->use_svc = 0; 1637 cpi->resize_state = 0; 1638 cpi->resize_avg_qp = 0; 1639 cpi->resize_buffer_underflow = 0; 1640 cpi->common.buffer_pool = pool; 1641 1642 cpi->rc.high_source_sad = 0; 1643 1644 init_config(cpi, oxcf); 1645 vp9_rc_init(&cpi->oxcf, oxcf->pass, &cpi->rc); 1646 1647 cm->current_video_frame = 0; 1648 cpi->partition_search_skippable_frame = 0; 1649 cpi->tile_data = NULL; 1650 1651 realloc_segmentation_maps(cpi); 1652 1653 CHECK_MEM_ERROR(cm, cpi->nmvcosts[0], 1654 vpx_calloc(MV_VALS, sizeof(*cpi->nmvcosts[0]))); 1655 CHECK_MEM_ERROR(cm, cpi->nmvcosts[1], 1656 vpx_calloc(MV_VALS, sizeof(*cpi->nmvcosts[1]))); 1657 CHECK_MEM_ERROR(cm, cpi->nmvcosts_hp[0], 1658 vpx_calloc(MV_VALS, sizeof(*cpi->nmvcosts_hp[0]))); 1659 CHECK_MEM_ERROR(cm, cpi->nmvcosts_hp[1], 1660 vpx_calloc(MV_VALS, sizeof(*cpi->nmvcosts_hp[1]))); 1661 CHECK_MEM_ERROR(cm, cpi->nmvsadcosts[0], 1662 vpx_calloc(MV_VALS, sizeof(*cpi->nmvsadcosts[0]))); 1663 CHECK_MEM_ERROR(cm, cpi->nmvsadcosts[1], 1664 vpx_calloc(MV_VALS, sizeof(*cpi->nmvsadcosts[1]))); 1665 CHECK_MEM_ERROR(cm, cpi->nmvsadcosts_hp[0], 1666 vpx_calloc(MV_VALS, sizeof(*cpi->nmvsadcosts_hp[0]))); 1667 CHECK_MEM_ERROR(cm, cpi->nmvsadcosts_hp[1], 1668 vpx_calloc(MV_VALS, sizeof(*cpi->nmvsadcosts_hp[1]))); 1669 1670 for (i = 0; i < (sizeof(cpi->mbgraph_stats) / 1671 sizeof(cpi->mbgraph_stats[0])); i++) { 1672 CHECK_MEM_ERROR(cm, cpi->mbgraph_stats[i].mb_stats, 1673 vpx_calloc(cm->MBs * 1674 sizeof(*cpi->mbgraph_stats[i].mb_stats), 1)); 1675 } 1676 1677#if CONFIG_FP_MB_STATS 1678 cpi->use_fp_mb_stats = 0; 1679 if (cpi->use_fp_mb_stats) { 1680 // a place holder used to store the first pass mb stats in the first pass 1681 CHECK_MEM_ERROR(cm, cpi->twopass.frame_mb_stats_buf, 1682 vpx_calloc(cm->MBs * sizeof(uint8_t), 1)); 1683 } else { 1684 cpi->twopass.frame_mb_stats_buf = NULL; 1685 } 1686#endif 1687 1688 cpi->refresh_alt_ref_frame = 0; 1689 cpi->multi_arf_last_grp_enabled = 0; 1690 1691 cpi->b_calculate_psnr = CONFIG_INTERNAL_STATS; 1692#if CONFIG_INTERNAL_STATS 1693 cpi->b_calculate_ssimg = 0; 1694 cpi->b_calculate_blockiness = 1; 1695 cpi->b_calculate_consistency = 1; 1696 cpi->total_inconsistency = 0; 1697 cpi->psnr.worst = 100.0; 1698 cpi->worst_ssim = 100.0; 1699 1700 cpi->count = 0; 1701 cpi->bytes = 0; 1702 1703 if (cpi->b_calculate_psnr) { 1704 cpi->total_sq_error = 0; 1705 cpi->total_samples = 0; 1706 1707 cpi->totalp_sq_error = 0; 1708 cpi->totalp_samples = 0; 1709 1710 cpi->tot_recode_hits = 0; 1711 cpi->summed_quality = 0; 1712 cpi->summed_weights = 0; 1713 cpi->summedp_quality = 0; 1714 cpi->summedp_weights = 0; 1715 } 1716 1717 if (cpi->b_calculate_ssimg) { 1718 cpi->ssimg.worst= 100.0; 1719 } 1720 cpi->fastssim.worst = 100.0; 1721 1722 cpi->psnrhvs.worst = 100.0; 1723 1724 if (cpi->b_calculate_blockiness) { 1725 cpi->total_blockiness = 0; 1726 cpi->worst_blockiness = 0.0; 1727 } 1728 1729 if (cpi->b_calculate_consistency) { 1730 cpi->ssim_vars = vpx_malloc(sizeof(*cpi->ssim_vars) * 1731 4 * cpi->common.mi_rows * cpi->common.mi_cols); 1732 cpi->worst_consistency = 100.0; 1733 } 1734 1735#endif 1736 1737 cpi->first_time_stamp_ever = INT64_MAX; 1738 1739 cal_nmvjointsadcost(cpi->td.mb.nmvjointsadcost); 1740 cpi->td.mb.nmvcost[0] = &cpi->nmvcosts[0][MV_MAX]; 1741 cpi->td.mb.nmvcost[1] = &cpi->nmvcosts[1][MV_MAX]; 1742 cpi->td.mb.nmvsadcost[0] = &cpi->nmvsadcosts[0][MV_MAX]; 1743 cpi->td.mb.nmvsadcost[1] = &cpi->nmvsadcosts[1][MV_MAX]; 1744 cal_nmvsadcosts(cpi->td.mb.nmvsadcost); 1745 1746 cpi->td.mb.nmvcost_hp[0] = &cpi->nmvcosts_hp[0][MV_MAX]; 1747 cpi->td.mb.nmvcost_hp[1] = &cpi->nmvcosts_hp[1][MV_MAX]; 1748 cpi->td.mb.nmvsadcost_hp[0] = &cpi->nmvsadcosts_hp[0][MV_MAX]; 1749 cpi->td.mb.nmvsadcost_hp[1] = &cpi->nmvsadcosts_hp[1][MV_MAX]; 1750 cal_nmvsadcosts_hp(cpi->td.mb.nmvsadcost_hp); 1751 1752#if CONFIG_VP9_TEMPORAL_DENOISING 1753#ifdef OUTPUT_YUV_DENOISED 1754 yuv_denoised_file = fopen("denoised.yuv", "ab"); 1755#endif 1756#endif 1757#ifdef OUTPUT_YUV_SKINMAP 1758 yuv_skinmap_file = fopen("skinmap.yuv", "ab"); 1759#endif 1760#ifdef OUTPUT_YUV_REC 1761 yuv_rec_file = fopen("rec.yuv", "wb"); 1762#endif 1763 1764#if 0 1765 framepsnr = fopen("framepsnr.stt", "a"); 1766 kf_list = fopen("kf_list.stt", "w"); 1767#endif 1768 1769 cpi->allow_encode_breakout = ENCODE_BREAKOUT_ENABLED; 1770 1771 if (oxcf->pass == 1) { 1772 vp9_init_first_pass(cpi); 1773 } else if (oxcf->pass == 2) { 1774 const size_t packet_sz = sizeof(FIRSTPASS_STATS); 1775 const int packets = (int)(oxcf->two_pass_stats_in.sz / packet_sz); 1776 1777 if (cpi->svc.number_spatial_layers > 1 1778 || cpi->svc.number_temporal_layers > 1) { 1779 FIRSTPASS_STATS *const stats = oxcf->two_pass_stats_in.buf; 1780 FIRSTPASS_STATS *stats_copy[VPX_SS_MAX_LAYERS] = {0}; 1781 int i; 1782 1783 for (i = 0; i < oxcf->ss_number_layers; ++i) { 1784 FIRSTPASS_STATS *const last_packet_for_layer = 1785 &stats[packets - oxcf->ss_number_layers + i]; 1786 const int layer_id = (int)last_packet_for_layer->spatial_layer_id; 1787 const int packets_in_layer = (int)last_packet_for_layer->count + 1; 1788 if (layer_id >= 0 && layer_id < oxcf->ss_number_layers) { 1789 LAYER_CONTEXT *const lc = &cpi->svc.layer_context[layer_id]; 1790 1791 vpx_free(lc->rc_twopass_stats_in.buf); 1792 1793 lc->rc_twopass_stats_in.sz = packets_in_layer * packet_sz; 1794 CHECK_MEM_ERROR(cm, lc->rc_twopass_stats_in.buf, 1795 vpx_malloc(lc->rc_twopass_stats_in.sz)); 1796 lc->twopass.stats_in_start = lc->rc_twopass_stats_in.buf; 1797 lc->twopass.stats_in = lc->twopass.stats_in_start; 1798 lc->twopass.stats_in_end = lc->twopass.stats_in_start 1799 + packets_in_layer - 1; 1800 stats_copy[layer_id] = lc->rc_twopass_stats_in.buf; 1801 } 1802 } 1803 1804 for (i = 0; i < packets; ++i) { 1805 const int layer_id = (int)stats[i].spatial_layer_id; 1806 if (layer_id >= 0 && layer_id < oxcf->ss_number_layers 1807 && stats_copy[layer_id] != NULL) { 1808 *stats_copy[layer_id] = stats[i]; 1809 ++stats_copy[layer_id]; 1810 } 1811 } 1812 1813 vp9_init_second_pass_spatial_svc(cpi); 1814 } else { 1815#if CONFIG_FP_MB_STATS 1816 if (cpi->use_fp_mb_stats) { 1817 const size_t psz = cpi->common.MBs * sizeof(uint8_t); 1818 const int ps = (int)(oxcf->firstpass_mb_stats_in.sz / psz); 1819 1820 cpi->twopass.firstpass_mb_stats.mb_stats_start = 1821 oxcf->firstpass_mb_stats_in.buf; 1822 cpi->twopass.firstpass_mb_stats.mb_stats_end = 1823 cpi->twopass.firstpass_mb_stats.mb_stats_start + 1824 (ps - 1) * cpi->common.MBs * sizeof(uint8_t); 1825 } 1826#endif 1827 1828 cpi->twopass.stats_in_start = oxcf->two_pass_stats_in.buf; 1829 cpi->twopass.stats_in = cpi->twopass.stats_in_start; 1830 cpi->twopass.stats_in_end = &cpi->twopass.stats_in[packets - 1]; 1831 1832 vp9_init_second_pass(cpi); 1833 } 1834 } 1835 1836 vp9_set_speed_features_framesize_independent(cpi); 1837 vp9_set_speed_features_framesize_dependent(cpi); 1838 1839 // Allocate memory to store variances for a frame. 1840 CHECK_MEM_ERROR(cm, cpi->source_diff_var, 1841 vpx_calloc(cm->MBs, sizeof(diff))); 1842 cpi->source_var_thresh = 0; 1843 cpi->frames_till_next_var_check = 0; 1844 1845#define BFP(BT, SDF, SDAF, VF, SVF, SVAF, SDX3F, SDX8F, SDX4DF)\ 1846 cpi->fn_ptr[BT].sdf = SDF; \ 1847 cpi->fn_ptr[BT].sdaf = SDAF; \ 1848 cpi->fn_ptr[BT].vf = VF; \ 1849 cpi->fn_ptr[BT].svf = SVF; \ 1850 cpi->fn_ptr[BT].svaf = SVAF; \ 1851 cpi->fn_ptr[BT].sdx3f = SDX3F; \ 1852 cpi->fn_ptr[BT].sdx8f = SDX8F; \ 1853 cpi->fn_ptr[BT].sdx4df = SDX4DF; 1854 1855 BFP(BLOCK_32X16, vpx_sad32x16, vpx_sad32x16_avg, 1856 vpx_variance32x16, vpx_sub_pixel_variance32x16, 1857 vpx_sub_pixel_avg_variance32x16, NULL, NULL, vpx_sad32x16x4d) 1858 1859 BFP(BLOCK_16X32, vpx_sad16x32, vpx_sad16x32_avg, 1860 vpx_variance16x32, vpx_sub_pixel_variance16x32, 1861 vpx_sub_pixel_avg_variance16x32, NULL, NULL, vpx_sad16x32x4d) 1862 1863 BFP(BLOCK_64X32, vpx_sad64x32, vpx_sad64x32_avg, 1864 vpx_variance64x32, vpx_sub_pixel_variance64x32, 1865 vpx_sub_pixel_avg_variance64x32, NULL, NULL, vpx_sad64x32x4d) 1866 1867 BFP(BLOCK_32X64, vpx_sad32x64, vpx_sad32x64_avg, 1868 vpx_variance32x64, vpx_sub_pixel_variance32x64, 1869 vpx_sub_pixel_avg_variance32x64, NULL, NULL, vpx_sad32x64x4d) 1870 1871 BFP(BLOCK_32X32, vpx_sad32x32, vpx_sad32x32_avg, 1872 vpx_variance32x32, vpx_sub_pixel_variance32x32, 1873 vpx_sub_pixel_avg_variance32x32, vpx_sad32x32x3, vpx_sad32x32x8, 1874 vpx_sad32x32x4d) 1875 1876 BFP(BLOCK_64X64, vpx_sad64x64, vpx_sad64x64_avg, 1877 vpx_variance64x64, vpx_sub_pixel_variance64x64, 1878 vpx_sub_pixel_avg_variance64x64, vpx_sad64x64x3, vpx_sad64x64x8, 1879 vpx_sad64x64x4d) 1880 1881 BFP(BLOCK_16X16, vpx_sad16x16, vpx_sad16x16_avg, 1882 vpx_variance16x16, vpx_sub_pixel_variance16x16, 1883 vpx_sub_pixel_avg_variance16x16, vpx_sad16x16x3, vpx_sad16x16x8, 1884 vpx_sad16x16x4d) 1885 1886 BFP(BLOCK_16X8, vpx_sad16x8, vpx_sad16x8_avg, 1887 vpx_variance16x8, vpx_sub_pixel_variance16x8, 1888 vpx_sub_pixel_avg_variance16x8, 1889 vpx_sad16x8x3, vpx_sad16x8x8, vpx_sad16x8x4d) 1890 1891 BFP(BLOCK_8X16, vpx_sad8x16, vpx_sad8x16_avg, 1892 vpx_variance8x16, vpx_sub_pixel_variance8x16, 1893 vpx_sub_pixel_avg_variance8x16, 1894 vpx_sad8x16x3, vpx_sad8x16x8, vpx_sad8x16x4d) 1895 1896 BFP(BLOCK_8X8, vpx_sad8x8, vpx_sad8x8_avg, 1897 vpx_variance8x8, vpx_sub_pixel_variance8x8, 1898 vpx_sub_pixel_avg_variance8x8, 1899 vpx_sad8x8x3, vpx_sad8x8x8, vpx_sad8x8x4d) 1900 1901 BFP(BLOCK_8X4, vpx_sad8x4, vpx_sad8x4_avg, 1902 vpx_variance8x4, vpx_sub_pixel_variance8x4, 1903 vpx_sub_pixel_avg_variance8x4, NULL, vpx_sad8x4x8, vpx_sad8x4x4d) 1904 1905 BFP(BLOCK_4X8, vpx_sad4x8, vpx_sad4x8_avg, 1906 vpx_variance4x8, vpx_sub_pixel_variance4x8, 1907 vpx_sub_pixel_avg_variance4x8, NULL, vpx_sad4x8x8, vpx_sad4x8x4d) 1908 1909 BFP(BLOCK_4X4, vpx_sad4x4, vpx_sad4x4_avg, 1910 vpx_variance4x4, vpx_sub_pixel_variance4x4, 1911 vpx_sub_pixel_avg_variance4x4, 1912 vpx_sad4x4x3, vpx_sad4x4x8, vpx_sad4x4x4d) 1913 1914#if CONFIG_VP9_HIGHBITDEPTH 1915 highbd_set_var_fns(cpi); 1916#endif 1917 1918 /* vp9_init_quantizer() is first called here. Add check in 1919 * vp9_frame_init_quantizer() so that vp9_init_quantizer is only 1920 * called later when needed. This will avoid unnecessary calls of 1921 * vp9_init_quantizer() for every frame. 1922 */ 1923 vp9_init_quantizer(cpi); 1924 1925 vp9_loop_filter_init(cm); 1926 1927 cm->error.setjmp = 0; 1928 1929 return cpi; 1930} 1931#define SNPRINT(H, T) \ 1932 snprintf((H) + strlen(H), sizeof(H) - strlen(H), (T)) 1933 1934#define SNPRINT2(H, T, V) \ 1935 snprintf((H) + strlen(H), sizeof(H) - strlen(H), (T), (V)) 1936 1937void vp9_remove_compressor(VP9_COMP *cpi) { 1938 VP9_COMMON *cm; 1939 unsigned int i; 1940 int t; 1941 1942 if (!cpi) 1943 return; 1944 1945 cm = &cpi->common; 1946 if (cm->current_video_frame > 0) { 1947#if CONFIG_INTERNAL_STATS 1948 vpx_clear_system_state(); 1949 1950 if (cpi->oxcf.pass != 1) { 1951 char headings[512] = {0}; 1952 char results[512] = {0}; 1953 FILE *f = fopen("opsnr.stt", "a"); 1954 double time_encoded = (cpi->last_end_time_stamp_seen 1955 - cpi->first_time_stamp_ever) / 10000000.000; 1956 double total_encode_time = (cpi->time_receive_data + 1957 cpi->time_compress_data) / 1000.000; 1958 const double dr = 1959 (double)cpi->bytes * (double) 8 / (double)1000 / time_encoded; 1960 const double peak = (double)((1 << cpi->oxcf.input_bit_depth) - 1); 1961 1962 if (cpi->b_calculate_psnr) { 1963 const double total_psnr = 1964 vpx_sse_to_psnr((double)cpi->total_samples, peak, 1965 (double)cpi->total_sq_error); 1966 const double totalp_psnr = 1967 vpx_sse_to_psnr((double)cpi->totalp_samples, peak, 1968 (double)cpi->totalp_sq_error); 1969 const double total_ssim = 100 * pow(cpi->summed_quality / 1970 cpi->summed_weights, 8.0); 1971 const double totalp_ssim = 100 * pow(cpi->summedp_quality / 1972 cpi->summedp_weights, 8.0); 1973 1974 snprintf(headings, sizeof(headings), 1975 "Bitrate\tAVGPsnr\tGLBPsnr\tAVPsnrP\tGLPsnrP\t" 1976 "VPXSSIM\tVPSSIMP\tFASTSIM\tPSNRHVS\t" 1977 "WstPsnr\tWstSsim\tWstFast\tWstHVS"); 1978 snprintf(results, sizeof(results), 1979 "%7.2f\t%7.3f\t%7.3f\t%7.3f\t%7.3f\t" 1980 "%7.3f\t%7.3f\t%7.3f\t%7.3f\t" 1981 "%7.3f\t%7.3f\t%7.3f\t%7.3f", 1982 dr, cpi->psnr.stat[ALL] / cpi->count, total_psnr, 1983 cpi->psnrp.stat[ALL] / cpi->count, totalp_psnr, 1984 total_ssim, totalp_ssim, 1985 cpi->fastssim.stat[ALL] / cpi->count, 1986 cpi->psnrhvs.stat[ALL] / cpi->count, 1987 cpi->psnr.worst, cpi->worst_ssim, cpi->fastssim.worst, 1988 cpi->psnrhvs.worst); 1989 1990 if (cpi->b_calculate_blockiness) { 1991 SNPRINT(headings, "\t Block\tWstBlck"); 1992 SNPRINT2(results, "\t%7.3f", cpi->total_blockiness / cpi->count); 1993 SNPRINT2(results, "\t%7.3f", cpi->worst_blockiness); 1994 } 1995 1996 if (cpi->b_calculate_consistency) { 1997 double consistency = 1998 vpx_sse_to_psnr((double)cpi->totalp_samples, peak, 1999 (double)cpi->total_inconsistency); 2000 2001 SNPRINT(headings, "\tConsist\tWstCons"); 2002 SNPRINT2(results, "\t%7.3f", consistency); 2003 SNPRINT2(results, "\t%7.3f", cpi->worst_consistency); 2004 } 2005 2006 if (cpi->b_calculate_ssimg) { 2007 SNPRINT(headings, "\t SSIMG\tWtSSIMG"); 2008 SNPRINT2(results, "\t%7.3f", cpi->ssimg.stat[ALL] / cpi->count); 2009 SNPRINT2(results, "\t%7.3f", cpi->ssimg.worst); 2010 } 2011 2012 fprintf(f, "%s\t Time\n", headings); 2013 fprintf(f, "%s\t%8.0f\n", results, total_encode_time); 2014 } 2015 2016 fclose(f); 2017 } 2018 2019#endif 2020 2021#if 0 2022 { 2023 printf("\n_pick_loop_filter_level:%d\n", cpi->time_pick_lpf / 1000); 2024 printf("\n_frames recive_data encod_mb_row compress_frame Total\n"); 2025 printf("%6d %10ld %10ld %10ld %10ld\n", cpi->common.current_video_frame, 2026 cpi->time_receive_data / 1000, cpi->time_encode_sb_row / 1000, 2027 cpi->time_compress_data / 1000, 2028 (cpi->time_receive_data + cpi->time_compress_data) / 1000); 2029 } 2030#endif 2031 } 2032 2033#if CONFIG_VP9_TEMPORAL_DENOISING 2034 vp9_denoiser_free(&(cpi->denoiser)); 2035#endif 2036 2037 for (t = 0; t < cpi->num_workers; ++t) { 2038 VPxWorker *const worker = &cpi->workers[t]; 2039 EncWorkerData *const thread_data = &cpi->tile_thr_data[t]; 2040 2041 // Deallocate allocated threads. 2042 vpx_get_worker_interface()->end(worker); 2043 2044 // Deallocate allocated thread data. 2045 if (t < cpi->num_workers - 1) { 2046 vpx_free(thread_data->td->counts); 2047 vp9_free_pc_tree(thread_data->td); 2048 vpx_free(thread_data->td); 2049 } 2050 } 2051 vpx_free(cpi->tile_thr_data); 2052 vpx_free(cpi->workers); 2053 2054 if (cpi->num_workers > 1) 2055 vp9_loop_filter_dealloc(&cpi->lf_row_sync); 2056 2057 dealloc_compressor_data(cpi); 2058 2059 for (i = 0; i < sizeof(cpi->mbgraph_stats) / 2060 sizeof(cpi->mbgraph_stats[0]); ++i) { 2061 vpx_free(cpi->mbgraph_stats[i].mb_stats); 2062 } 2063 2064#if CONFIG_FP_MB_STATS 2065 if (cpi->use_fp_mb_stats) { 2066 vpx_free(cpi->twopass.frame_mb_stats_buf); 2067 cpi->twopass.frame_mb_stats_buf = NULL; 2068 } 2069#endif 2070 2071 vp9_remove_common(cm); 2072 vp9_free_ref_frame_buffers(cm->buffer_pool); 2073#if CONFIG_VP9_POSTPROC 2074 vp9_free_postproc_buffers(cm); 2075#endif 2076 vpx_free(cpi); 2077 2078#if CONFIG_VP9_TEMPORAL_DENOISING 2079#ifdef OUTPUT_YUV_DENOISED 2080 fclose(yuv_denoised_file); 2081#endif 2082#endif 2083#ifdef OUTPUT_YUV_SKINMAP 2084 fclose(yuv_skinmap_file); 2085#endif 2086#ifdef OUTPUT_YUV_REC 2087 fclose(yuv_rec_file); 2088#endif 2089 2090#if 0 2091 2092 if (keyfile) 2093 fclose(keyfile); 2094 2095 if (framepsnr) 2096 fclose(framepsnr); 2097 2098 if (kf_list) 2099 fclose(kf_list); 2100 2101#endif 2102} 2103 2104/* TODO(yaowu): The block_variance calls the unoptimized versions of variance() 2105 * and highbd_8_variance(). It should not. 2106 */ 2107static void encoder_variance(const uint8_t *a, int a_stride, 2108 const uint8_t *b, int b_stride, 2109 int w, int h, unsigned int *sse, int *sum) { 2110 int i, j; 2111 2112 *sum = 0; 2113 *sse = 0; 2114 2115 for (i = 0; i < h; i++) { 2116 for (j = 0; j < w; j++) { 2117 const int diff = a[j] - b[j]; 2118 *sum += diff; 2119 *sse += diff * diff; 2120 } 2121 2122 a += a_stride; 2123 b += b_stride; 2124 } 2125} 2126 2127#if CONFIG_VP9_HIGHBITDEPTH 2128static void encoder_highbd_variance64(const uint8_t *a8, int a_stride, 2129 const uint8_t *b8, int b_stride, 2130 int w, int h, uint64_t *sse, 2131 uint64_t *sum) { 2132 int i, j; 2133 2134 uint16_t *a = CONVERT_TO_SHORTPTR(a8); 2135 uint16_t *b = CONVERT_TO_SHORTPTR(b8); 2136 *sum = 0; 2137 *sse = 0; 2138 2139 for (i = 0; i < h; i++) { 2140 for (j = 0; j < w; j++) { 2141 const int diff = a[j] - b[j]; 2142 *sum += diff; 2143 *sse += diff * diff; 2144 } 2145 a += a_stride; 2146 b += b_stride; 2147 } 2148} 2149 2150static void encoder_highbd_8_variance(const uint8_t *a8, int a_stride, 2151 const uint8_t *b8, int b_stride, 2152 int w, int h, 2153 unsigned int *sse, int *sum) { 2154 uint64_t sse_long = 0; 2155 uint64_t sum_long = 0; 2156 encoder_highbd_variance64(a8, a_stride, b8, b_stride, w, h, 2157 &sse_long, &sum_long); 2158 *sse = (unsigned int)sse_long; 2159 *sum = (int)sum_long; 2160} 2161#endif // CONFIG_VP9_HIGHBITDEPTH 2162 2163static int64_t get_sse(const uint8_t *a, int a_stride, 2164 const uint8_t *b, int b_stride, 2165 int width, int height) { 2166 const int dw = width % 16; 2167 const int dh = height % 16; 2168 int64_t total_sse = 0; 2169 unsigned int sse = 0; 2170 int sum = 0; 2171 int x, y; 2172 2173 if (dw > 0) { 2174 encoder_variance(&a[width - dw], a_stride, &b[width - dw], b_stride, 2175 dw, height, &sse, &sum); 2176 total_sse += sse; 2177 } 2178 2179 if (dh > 0) { 2180 encoder_variance(&a[(height - dh) * a_stride], a_stride, 2181 &b[(height - dh) * b_stride], b_stride, 2182 width - dw, dh, &sse, &sum); 2183 total_sse += sse; 2184 } 2185 2186 for (y = 0; y < height / 16; ++y) { 2187 const uint8_t *pa = a; 2188 const uint8_t *pb = b; 2189 for (x = 0; x < width / 16; ++x) { 2190 vpx_mse16x16(pa, a_stride, pb, b_stride, &sse); 2191 total_sse += sse; 2192 2193 pa += 16; 2194 pb += 16; 2195 } 2196 2197 a += 16 * a_stride; 2198 b += 16 * b_stride; 2199 } 2200 2201 return total_sse; 2202} 2203 2204#if CONFIG_VP9_HIGHBITDEPTH 2205static int64_t highbd_get_sse_shift(const uint8_t *a8, int a_stride, 2206 const uint8_t *b8, int b_stride, 2207 int width, int height, 2208 unsigned int input_shift) { 2209 const uint16_t *a = CONVERT_TO_SHORTPTR(a8); 2210 const uint16_t *b = CONVERT_TO_SHORTPTR(b8); 2211 int64_t total_sse = 0; 2212 int x, y; 2213 for (y = 0; y < height; ++y) { 2214 for (x = 0; x < width; ++x) { 2215 int64_t diff; 2216 diff = (a[x] >> input_shift) - (b[x] >> input_shift); 2217 total_sse += diff * diff; 2218 } 2219 a += a_stride; 2220 b += b_stride; 2221 } 2222 return total_sse; 2223} 2224 2225static int64_t highbd_get_sse(const uint8_t *a, int a_stride, 2226 const uint8_t *b, int b_stride, 2227 int width, int height) { 2228 int64_t total_sse = 0; 2229 int x, y; 2230 const int dw = width % 16; 2231 const int dh = height % 16; 2232 unsigned int sse = 0; 2233 int sum = 0; 2234 if (dw > 0) { 2235 encoder_highbd_8_variance(&a[width - dw], a_stride, 2236 &b[width - dw], b_stride, 2237 dw, height, &sse, &sum); 2238 total_sse += sse; 2239 } 2240 if (dh > 0) { 2241 encoder_highbd_8_variance(&a[(height - dh) * a_stride], a_stride, 2242 &b[(height - dh) * b_stride], b_stride, 2243 width - dw, dh, &sse, &sum); 2244 total_sse += sse; 2245 } 2246 for (y = 0; y < height / 16; ++y) { 2247 const uint8_t *pa = a; 2248 const uint8_t *pb = b; 2249 for (x = 0; x < width / 16; ++x) { 2250 vpx_highbd_8_mse16x16(pa, a_stride, pb, b_stride, &sse); 2251 total_sse += sse; 2252 pa += 16; 2253 pb += 16; 2254 } 2255 a += 16 * a_stride; 2256 b += 16 * b_stride; 2257 } 2258 return total_sse; 2259} 2260#endif // CONFIG_VP9_HIGHBITDEPTH 2261 2262typedef struct { 2263 double psnr[4]; // total/y/u/v 2264 uint64_t sse[4]; // total/y/u/v 2265 uint32_t samples[4]; // total/y/u/v 2266} PSNR_STATS; 2267 2268#if CONFIG_VP9_HIGHBITDEPTH 2269static void calc_highbd_psnr(const YV12_BUFFER_CONFIG *a, 2270 const YV12_BUFFER_CONFIG *b, 2271 PSNR_STATS *psnr, 2272 unsigned int bit_depth, 2273 unsigned int in_bit_depth) { 2274 const int widths[3] = 2275 {a->y_crop_width, a->uv_crop_width, a->uv_crop_width }; 2276 const int heights[3] = 2277 {a->y_crop_height, a->uv_crop_height, a->uv_crop_height}; 2278 const uint8_t *a_planes[3] = {a->y_buffer, a->u_buffer, a->v_buffer }; 2279 const int a_strides[3] = {a->y_stride, a->uv_stride, a->uv_stride}; 2280 const uint8_t *b_planes[3] = {b->y_buffer, b->u_buffer, b->v_buffer }; 2281 const int b_strides[3] = {b->y_stride, b->uv_stride, b->uv_stride}; 2282 int i; 2283 uint64_t total_sse = 0; 2284 uint32_t total_samples = 0; 2285 const double peak = (double)((1 << in_bit_depth) - 1); 2286 const unsigned int input_shift = bit_depth - in_bit_depth; 2287 2288 for (i = 0; i < 3; ++i) { 2289 const int w = widths[i]; 2290 const int h = heights[i]; 2291 const uint32_t samples = w * h; 2292 uint64_t sse; 2293 if (a->flags & YV12_FLAG_HIGHBITDEPTH) { 2294 if (input_shift) { 2295 sse = highbd_get_sse_shift(a_planes[i], a_strides[i], 2296 b_planes[i], b_strides[i], w, h, 2297 input_shift); 2298 } else { 2299 sse = highbd_get_sse(a_planes[i], a_strides[i], 2300 b_planes[i], b_strides[i], w, h); 2301 } 2302 } else { 2303 sse = get_sse(a_planes[i], a_strides[i], 2304 b_planes[i], b_strides[i], 2305 w, h); 2306 } 2307 psnr->sse[1 + i] = sse; 2308 psnr->samples[1 + i] = samples; 2309 psnr->psnr[1 + i] = vpx_sse_to_psnr(samples, peak, (double)sse); 2310 2311 total_sse += sse; 2312 total_samples += samples; 2313 } 2314 2315 psnr->sse[0] = total_sse; 2316 psnr->samples[0] = total_samples; 2317 psnr->psnr[0] = vpx_sse_to_psnr((double)total_samples, peak, 2318 (double)total_sse); 2319} 2320 2321#else // !CONFIG_VP9_HIGHBITDEPTH 2322 2323static void calc_psnr(const YV12_BUFFER_CONFIG *a, const YV12_BUFFER_CONFIG *b, 2324 PSNR_STATS *psnr) { 2325 static const double peak = 255.0; 2326 const int widths[3] = { 2327 a->y_crop_width, a->uv_crop_width, a->uv_crop_width}; 2328 const int heights[3] = { 2329 a->y_crop_height, a->uv_crop_height, a->uv_crop_height}; 2330 const uint8_t *a_planes[3] = {a->y_buffer, a->u_buffer, a->v_buffer}; 2331 const int a_strides[3] = {a->y_stride, a->uv_stride, a->uv_stride}; 2332 const uint8_t *b_planes[3] = {b->y_buffer, b->u_buffer, b->v_buffer}; 2333 const int b_strides[3] = {b->y_stride, b->uv_stride, b->uv_stride}; 2334 int i; 2335 uint64_t total_sse = 0; 2336 uint32_t total_samples = 0; 2337 2338 for (i = 0; i < 3; ++i) { 2339 const int w = widths[i]; 2340 const int h = heights[i]; 2341 const uint32_t samples = w * h; 2342 const uint64_t sse = get_sse(a_planes[i], a_strides[i], 2343 b_planes[i], b_strides[i], 2344 w, h); 2345 psnr->sse[1 + i] = sse; 2346 psnr->samples[1 + i] = samples; 2347 psnr->psnr[1 + i] = vpx_sse_to_psnr(samples, peak, (double)sse); 2348 2349 total_sse += sse; 2350 total_samples += samples; 2351 } 2352 2353 psnr->sse[0] = total_sse; 2354 psnr->samples[0] = total_samples; 2355 psnr->psnr[0] = vpx_sse_to_psnr((double)total_samples, peak, 2356 (double)total_sse); 2357} 2358#endif // CONFIG_VP9_HIGHBITDEPTH 2359 2360static void generate_psnr_packet(VP9_COMP *cpi) { 2361 struct vpx_codec_cx_pkt pkt; 2362 int i; 2363 PSNR_STATS psnr; 2364#if CONFIG_VP9_HIGHBITDEPTH 2365 calc_highbd_psnr(cpi->Source, cpi->common.frame_to_show, &psnr, 2366 cpi->td.mb.e_mbd.bd, cpi->oxcf.input_bit_depth); 2367#else 2368 calc_psnr(cpi->Source, cpi->common.frame_to_show, &psnr); 2369#endif 2370 2371 for (i = 0; i < 4; ++i) { 2372 pkt.data.psnr.samples[i] = psnr.samples[i]; 2373 pkt.data.psnr.sse[i] = psnr.sse[i]; 2374 pkt.data.psnr.psnr[i] = psnr.psnr[i]; 2375 } 2376 pkt.kind = VPX_CODEC_PSNR_PKT; 2377 if (cpi->use_svc) 2378 cpi->svc.layer_context[cpi->svc.spatial_layer_id * 2379 cpi->svc.number_temporal_layers].psnr_pkt = pkt.data.psnr; 2380 else 2381 vpx_codec_pkt_list_add(cpi->output_pkt_list, &pkt); 2382} 2383 2384int vp9_use_as_reference(VP9_COMP *cpi, int ref_frame_flags) { 2385 if (ref_frame_flags > 7) 2386 return -1; 2387 2388 cpi->ref_frame_flags = ref_frame_flags; 2389 return 0; 2390} 2391 2392void vp9_update_reference(VP9_COMP *cpi, int ref_frame_flags) { 2393 cpi->ext_refresh_golden_frame = (ref_frame_flags & VP9_GOLD_FLAG) != 0; 2394 cpi->ext_refresh_alt_ref_frame = (ref_frame_flags & VP9_ALT_FLAG) != 0; 2395 cpi->ext_refresh_last_frame = (ref_frame_flags & VP9_LAST_FLAG) != 0; 2396 cpi->ext_refresh_frame_flags_pending = 1; 2397} 2398 2399static YV12_BUFFER_CONFIG *get_vp9_ref_frame_buffer(VP9_COMP *cpi, 2400 VP9_REFFRAME ref_frame_flag) { 2401 MV_REFERENCE_FRAME ref_frame = NONE; 2402 if (ref_frame_flag == VP9_LAST_FLAG) 2403 ref_frame = LAST_FRAME; 2404 else if (ref_frame_flag == VP9_GOLD_FLAG) 2405 ref_frame = GOLDEN_FRAME; 2406 else if (ref_frame_flag == VP9_ALT_FLAG) 2407 ref_frame = ALTREF_FRAME; 2408 2409 return ref_frame == NONE ? NULL : get_ref_frame_buffer(cpi, ref_frame); 2410} 2411 2412int vp9_copy_reference_enc(VP9_COMP *cpi, VP9_REFFRAME ref_frame_flag, 2413 YV12_BUFFER_CONFIG *sd) { 2414 YV12_BUFFER_CONFIG *cfg = get_vp9_ref_frame_buffer(cpi, ref_frame_flag); 2415 if (cfg) { 2416 vp8_yv12_copy_frame(cfg, sd); 2417 return 0; 2418 } else { 2419 return -1; 2420 } 2421} 2422 2423int vp9_set_reference_enc(VP9_COMP *cpi, VP9_REFFRAME ref_frame_flag, 2424 YV12_BUFFER_CONFIG *sd) { 2425 YV12_BUFFER_CONFIG *cfg = get_vp9_ref_frame_buffer(cpi, ref_frame_flag); 2426 if (cfg) { 2427 vp8_yv12_copy_frame(sd, cfg); 2428 return 0; 2429 } else { 2430 return -1; 2431 } 2432} 2433 2434int vp9_update_entropy(VP9_COMP * cpi, int update) { 2435 cpi->ext_refresh_frame_context = update; 2436 cpi->ext_refresh_frame_context_pending = 1; 2437 return 0; 2438} 2439 2440#if defined(OUTPUT_YUV_DENOISED) || defined(OUTPUT_YUV_SKINMAP) 2441// The denoiser buffer is allocated as a YUV 440 buffer. This function writes it 2442// as YUV 420. We simply use the top-left pixels of the UV buffers, since we do 2443// not denoise the UV channels at this time. If ever we implement UV channel 2444// denoising we will have to modify this. 2445void vp9_write_yuv_frame_420(YV12_BUFFER_CONFIG *s, FILE *f) { 2446 uint8_t *src = s->y_buffer; 2447 int h = s->y_height; 2448 2449 do { 2450 fwrite(src, s->y_width, 1, f); 2451 src += s->y_stride; 2452 } while (--h); 2453 2454 src = s->u_buffer; 2455 h = s->uv_height; 2456 2457 do { 2458 fwrite(src, s->uv_width, 1, f); 2459 src += s->uv_stride; 2460 } while (--h); 2461 2462 src = s->v_buffer; 2463 h = s->uv_height; 2464 2465 do { 2466 fwrite(src, s->uv_width, 1, f); 2467 src += s->uv_stride; 2468 } while (--h); 2469} 2470#endif 2471 2472#ifdef OUTPUT_YUV_REC 2473void vp9_write_yuv_rec_frame(VP9_COMMON *cm) { 2474 YV12_BUFFER_CONFIG *s = cm->frame_to_show; 2475 uint8_t *src = s->y_buffer; 2476 int h = cm->height; 2477 2478#if CONFIG_VP9_HIGHBITDEPTH 2479 if (s->flags & YV12_FLAG_HIGHBITDEPTH) { 2480 uint16_t *src16 = CONVERT_TO_SHORTPTR(s->y_buffer); 2481 2482 do { 2483 fwrite(src16, s->y_width, 2, yuv_rec_file); 2484 src16 += s->y_stride; 2485 } while (--h); 2486 2487 src16 = CONVERT_TO_SHORTPTR(s->u_buffer); 2488 h = s->uv_height; 2489 2490 do { 2491 fwrite(src16, s->uv_width, 2, yuv_rec_file); 2492 src16 += s->uv_stride; 2493 } while (--h); 2494 2495 src16 = CONVERT_TO_SHORTPTR(s->v_buffer); 2496 h = s->uv_height; 2497 2498 do { 2499 fwrite(src16, s->uv_width, 2, yuv_rec_file); 2500 src16 += s->uv_stride; 2501 } while (--h); 2502 2503 fflush(yuv_rec_file); 2504 return; 2505 } 2506#endif // CONFIG_VP9_HIGHBITDEPTH 2507 2508 do { 2509 fwrite(src, s->y_width, 1, yuv_rec_file); 2510 src += s->y_stride; 2511 } while (--h); 2512 2513 src = s->u_buffer; 2514 h = s->uv_height; 2515 2516 do { 2517 fwrite(src, s->uv_width, 1, yuv_rec_file); 2518 src += s->uv_stride; 2519 } while (--h); 2520 2521 src = s->v_buffer; 2522 h = s->uv_height; 2523 2524 do { 2525 fwrite(src, s->uv_width, 1, yuv_rec_file); 2526 src += s->uv_stride; 2527 } while (--h); 2528 2529 fflush(yuv_rec_file); 2530} 2531#endif 2532 2533#if CONFIG_VP9_HIGHBITDEPTH 2534static void scale_and_extend_frame_nonnormative(const YV12_BUFFER_CONFIG *src, 2535 YV12_BUFFER_CONFIG *dst, 2536 int bd) { 2537#else 2538static void scale_and_extend_frame_nonnormative(const YV12_BUFFER_CONFIG *src, 2539 YV12_BUFFER_CONFIG *dst) { 2540#endif // CONFIG_VP9_HIGHBITDEPTH 2541 // TODO(dkovalev): replace YV12_BUFFER_CONFIG with vpx_image_t 2542 int i; 2543 const uint8_t *const srcs[3] = {src->y_buffer, src->u_buffer, src->v_buffer}; 2544 const int src_strides[3] = {src->y_stride, src->uv_stride, src->uv_stride}; 2545 const int src_widths[3] = {src->y_crop_width, src->uv_crop_width, 2546 src->uv_crop_width }; 2547 const int src_heights[3] = {src->y_crop_height, src->uv_crop_height, 2548 src->uv_crop_height}; 2549 uint8_t *const dsts[3] = {dst->y_buffer, dst->u_buffer, dst->v_buffer}; 2550 const int dst_strides[3] = {dst->y_stride, dst->uv_stride, dst->uv_stride}; 2551 const int dst_widths[3] = {dst->y_crop_width, dst->uv_crop_width, 2552 dst->uv_crop_width}; 2553 const int dst_heights[3] = {dst->y_crop_height, dst->uv_crop_height, 2554 dst->uv_crop_height}; 2555 2556 for (i = 0; i < MAX_MB_PLANE; ++i) { 2557#if CONFIG_VP9_HIGHBITDEPTH 2558 if (src->flags & YV12_FLAG_HIGHBITDEPTH) { 2559 vp9_highbd_resize_plane(srcs[i], src_heights[i], src_widths[i], 2560 src_strides[i], dsts[i], dst_heights[i], 2561 dst_widths[i], dst_strides[i], bd); 2562 } else { 2563 vp9_resize_plane(srcs[i], src_heights[i], src_widths[i], src_strides[i], 2564 dsts[i], dst_heights[i], dst_widths[i], dst_strides[i]); 2565 } 2566#else 2567 vp9_resize_plane(srcs[i], src_heights[i], src_widths[i], src_strides[i], 2568 dsts[i], dst_heights[i], dst_widths[i], dst_strides[i]); 2569#endif // CONFIG_VP9_HIGHBITDEPTH 2570 } 2571 vpx_extend_frame_borders(dst); 2572} 2573 2574#if CONFIG_VP9_HIGHBITDEPTH 2575static void scale_and_extend_frame(const YV12_BUFFER_CONFIG *src, 2576 YV12_BUFFER_CONFIG *dst, int bd) { 2577#else 2578static void scale_and_extend_frame(const YV12_BUFFER_CONFIG *src, 2579 YV12_BUFFER_CONFIG *dst) { 2580#endif // CONFIG_VP9_HIGHBITDEPTH 2581 const int src_w = src->y_crop_width; 2582 const int src_h = src->y_crop_height; 2583 const int dst_w = dst->y_crop_width; 2584 const int dst_h = dst->y_crop_height; 2585 const uint8_t *const srcs[3] = {src->y_buffer, src->u_buffer, src->v_buffer}; 2586 const int src_strides[3] = {src->y_stride, src->uv_stride, src->uv_stride}; 2587 uint8_t *const dsts[3] = {dst->y_buffer, dst->u_buffer, dst->v_buffer}; 2588 const int dst_strides[3] = {dst->y_stride, dst->uv_stride, dst->uv_stride}; 2589 const InterpKernel *const kernel = vp9_filter_kernels[EIGHTTAP]; 2590 int x, y, i; 2591 2592 for (y = 0; y < dst_h; y += 16) { 2593 for (x = 0; x < dst_w; x += 16) { 2594 for (i = 0; i < MAX_MB_PLANE; ++i) { 2595 const int factor = (i == 0 || i == 3 ? 1 : 2); 2596 const int x_q4 = x * (16 / factor) * src_w / dst_w; 2597 const int y_q4 = y * (16 / factor) * src_h / dst_h; 2598 const int src_stride = src_strides[i]; 2599 const int dst_stride = dst_strides[i]; 2600 const uint8_t *src_ptr = srcs[i] + (y / factor) * src_h / dst_h * 2601 src_stride + (x / factor) * src_w / dst_w; 2602 uint8_t *dst_ptr = dsts[i] + (y / factor) * dst_stride + (x / factor); 2603 2604#if CONFIG_VP9_HIGHBITDEPTH 2605 if (src->flags & YV12_FLAG_HIGHBITDEPTH) { 2606 vpx_highbd_convolve8(src_ptr, src_stride, dst_ptr, dst_stride, 2607 kernel[x_q4 & 0xf], 16 * src_w / dst_w, 2608 kernel[y_q4 & 0xf], 16 * src_h / dst_h, 2609 16 / factor, 16 / factor, bd); 2610 } else { 2611 vpx_scaled_2d(src_ptr, src_stride, dst_ptr, dst_stride, 2612 kernel[x_q4 & 0xf], 16 * src_w / dst_w, 2613 kernel[y_q4 & 0xf], 16 * src_h / dst_h, 2614 16 / factor, 16 / factor); 2615 } 2616#else 2617 vpx_scaled_2d(src_ptr, src_stride, dst_ptr, dst_stride, 2618 kernel[x_q4 & 0xf], 16 * src_w / dst_w, 2619 kernel[y_q4 & 0xf], 16 * src_h / dst_h, 2620 16 / factor, 16 / factor); 2621#endif // CONFIG_VP9_HIGHBITDEPTH 2622 } 2623 } 2624 } 2625 2626 vpx_extend_frame_borders(dst); 2627} 2628 2629static int scale_down(VP9_COMP *cpi, int q) { 2630 RATE_CONTROL *const rc = &cpi->rc; 2631 GF_GROUP *const gf_group = &cpi->twopass.gf_group; 2632 int scale = 0; 2633 assert(frame_is_kf_gf_arf(cpi)); 2634 2635 if (rc->frame_size_selector == UNSCALED && 2636 q >= rc->rf_level_maxq[gf_group->rf_level[gf_group->index]]) { 2637 const int max_size_thresh = (int)(rate_thresh_mult[SCALE_STEP1] 2638 * VPXMAX(rc->this_frame_target, rc->avg_frame_bandwidth)); 2639 scale = rc->projected_frame_size > max_size_thresh ? 1 : 0; 2640 } 2641 return scale; 2642} 2643 2644// Function to test for conditions that indicate we should loop 2645// back and recode a frame. 2646static int recode_loop_test(VP9_COMP *cpi, 2647 int high_limit, int low_limit, 2648 int q, int maxq, int minq) { 2649 const RATE_CONTROL *const rc = &cpi->rc; 2650 const VP9EncoderConfig *const oxcf = &cpi->oxcf; 2651 const int frame_is_kfgfarf = frame_is_kf_gf_arf(cpi); 2652 int force_recode = 0; 2653 2654 if ((rc->projected_frame_size >= rc->max_frame_bandwidth) || 2655 (cpi->sf.recode_loop == ALLOW_RECODE) || 2656 (frame_is_kfgfarf && 2657 (cpi->sf.recode_loop == ALLOW_RECODE_KFARFGF))) { 2658 if (frame_is_kfgfarf && 2659 (oxcf->resize_mode == RESIZE_DYNAMIC) && 2660 scale_down(cpi, q)) { 2661 // Code this group at a lower resolution. 2662 cpi->resize_pending = 1; 2663 return 1; 2664 } 2665 2666 // TODO(agrange) high_limit could be greater than the scale-down threshold. 2667 if ((rc->projected_frame_size > high_limit && q < maxq) || 2668 (rc->projected_frame_size < low_limit && q > minq)) { 2669 force_recode = 1; 2670 } else if (cpi->oxcf.rc_mode == VPX_CQ) { 2671 // Deal with frame undershoot and whether or not we are 2672 // below the automatically set cq level. 2673 if (q > oxcf->cq_level && 2674 rc->projected_frame_size < ((rc->this_frame_target * 7) >> 3)) { 2675 force_recode = 1; 2676 } 2677 } 2678 } 2679 return force_recode; 2680} 2681 2682void vp9_update_reference_frames(VP9_COMP *cpi) { 2683 VP9_COMMON * const cm = &cpi->common; 2684 BufferPool *const pool = cm->buffer_pool; 2685 2686 // At this point the new frame has been encoded. 2687 // If any buffer copy / swapping is signaled it should be done here. 2688 if (cm->frame_type == KEY_FRAME) { 2689 ref_cnt_fb(pool->frame_bufs, 2690 &cm->ref_frame_map[cpi->gld_fb_idx], cm->new_fb_idx); 2691 ref_cnt_fb(pool->frame_bufs, 2692 &cm->ref_frame_map[cpi->alt_fb_idx], cm->new_fb_idx); 2693 } else if (vp9_preserve_existing_gf(cpi)) { 2694 // We have decided to preserve the previously existing golden frame as our 2695 // new ARF frame. However, in the short term in function 2696 // vp9_bitstream.c::get_refresh_mask() we left it in the GF slot and, if 2697 // we're updating the GF with the current decoded frame, we save it to the 2698 // ARF slot instead. 2699 // We now have to update the ARF with the current frame and swap gld_fb_idx 2700 // and alt_fb_idx so that, overall, we've stored the old GF in the new ARF 2701 // slot and, if we're updating the GF, the current frame becomes the new GF. 2702 int tmp; 2703 2704 ref_cnt_fb(pool->frame_bufs, 2705 &cm->ref_frame_map[cpi->alt_fb_idx], cm->new_fb_idx); 2706 2707 tmp = cpi->alt_fb_idx; 2708 cpi->alt_fb_idx = cpi->gld_fb_idx; 2709 cpi->gld_fb_idx = tmp; 2710 2711 if (is_two_pass_svc(cpi)) { 2712 cpi->svc.layer_context[0].gold_ref_idx = cpi->gld_fb_idx; 2713 cpi->svc.layer_context[0].alt_ref_idx = cpi->alt_fb_idx; 2714 } 2715 } else { /* For non key/golden frames */ 2716 if (cpi->refresh_alt_ref_frame) { 2717 int arf_idx = cpi->alt_fb_idx; 2718 if ((cpi->oxcf.pass == 2) && cpi->multi_arf_allowed) { 2719 const GF_GROUP *const gf_group = &cpi->twopass.gf_group; 2720 arf_idx = gf_group->arf_update_idx[gf_group->index]; 2721 } 2722 2723 ref_cnt_fb(pool->frame_bufs, 2724 &cm->ref_frame_map[arf_idx], cm->new_fb_idx); 2725 memcpy(cpi->interp_filter_selected[ALTREF_FRAME], 2726 cpi->interp_filter_selected[0], 2727 sizeof(cpi->interp_filter_selected[0])); 2728 } 2729 2730 if (cpi->refresh_golden_frame) { 2731 ref_cnt_fb(pool->frame_bufs, 2732 &cm->ref_frame_map[cpi->gld_fb_idx], cm->new_fb_idx); 2733 if (!cpi->rc.is_src_frame_alt_ref) 2734 memcpy(cpi->interp_filter_selected[GOLDEN_FRAME], 2735 cpi->interp_filter_selected[0], 2736 sizeof(cpi->interp_filter_selected[0])); 2737 else 2738 memcpy(cpi->interp_filter_selected[GOLDEN_FRAME], 2739 cpi->interp_filter_selected[ALTREF_FRAME], 2740 sizeof(cpi->interp_filter_selected[ALTREF_FRAME])); 2741 } 2742 } 2743 2744 if (cpi->refresh_last_frame) { 2745 ref_cnt_fb(pool->frame_bufs, 2746 &cm->ref_frame_map[cpi->lst_fb_idx], cm->new_fb_idx); 2747 if (!cpi->rc.is_src_frame_alt_ref) 2748 memcpy(cpi->interp_filter_selected[LAST_FRAME], 2749 cpi->interp_filter_selected[0], 2750 sizeof(cpi->interp_filter_selected[0])); 2751 } 2752#if CONFIG_VP9_TEMPORAL_DENOISING 2753 if (cpi->oxcf.noise_sensitivity > 0) { 2754 vp9_denoiser_update_frame_info(&cpi->denoiser, 2755 *cpi->Source, 2756 cpi->common.frame_type, 2757 cpi->refresh_alt_ref_frame, 2758 cpi->refresh_golden_frame, 2759 cpi->refresh_last_frame, 2760 cpi->resize_pending); 2761 } 2762#endif 2763} 2764 2765static void loopfilter_frame(VP9_COMP *cpi, VP9_COMMON *cm) { 2766 MACROBLOCKD *xd = &cpi->td.mb.e_mbd; 2767 struct loopfilter *lf = &cm->lf; 2768 2769 if (xd->lossless) { 2770 lf->filter_level = 0; 2771 } else { 2772 struct vpx_usec_timer timer; 2773 2774 vpx_clear_system_state(); 2775 2776 vpx_usec_timer_start(&timer); 2777 2778 vp9_pick_filter_level(cpi->Source, cpi, cpi->sf.lpf_pick); 2779 2780 vpx_usec_timer_mark(&timer); 2781 cpi->time_pick_lpf += vpx_usec_timer_elapsed(&timer); 2782 } 2783 2784 if (lf->filter_level > 0) { 2785 vp9_build_mask_frame(cm, lf->filter_level, 0); 2786 2787 if (cpi->num_workers > 1) 2788 vp9_loop_filter_frame_mt(cm->frame_to_show, cm, xd->plane, 2789 lf->filter_level, 0, 0, 2790 cpi->workers, cpi->num_workers, 2791 &cpi->lf_row_sync); 2792 else 2793 vp9_loop_filter_frame(cm->frame_to_show, cm, xd, lf->filter_level, 0, 0); 2794 } 2795 2796 vpx_extend_frame_inner_borders(cm->frame_to_show); 2797} 2798 2799static INLINE void alloc_frame_mvs(const VP9_COMMON *cm, 2800 int buffer_idx) { 2801 RefCntBuffer *const new_fb_ptr = &cm->buffer_pool->frame_bufs[buffer_idx]; 2802 if (new_fb_ptr->mvs == NULL || 2803 new_fb_ptr->mi_rows < cm->mi_rows || 2804 new_fb_ptr->mi_cols < cm->mi_cols) { 2805 vpx_free(new_fb_ptr->mvs); 2806 new_fb_ptr->mvs = 2807 (MV_REF *)vpx_calloc(cm->mi_rows * cm->mi_cols, 2808 sizeof(*new_fb_ptr->mvs)); 2809 new_fb_ptr->mi_rows = cm->mi_rows; 2810 new_fb_ptr->mi_cols = cm->mi_cols; 2811 } 2812} 2813 2814void vp9_scale_references(VP9_COMP *cpi) { 2815 VP9_COMMON *cm = &cpi->common; 2816 MV_REFERENCE_FRAME ref_frame; 2817 const VP9_REFFRAME ref_mask[3] = {VP9_LAST_FLAG, VP9_GOLD_FLAG, VP9_ALT_FLAG}; 2818 2819 for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ++ref_frame) { 2820 // Need to convert from VP9_REFFRAME to index into ref_mask (subtract 1). 2821 if (cpi->ref_frame_flags & ref_mask[ref_frame - 1]) { 2822 BufferPool *const pool = cm->buffer_pool; 2823 const YV12_BUFFER_CONFIG *const ref = get_ref_frame_buffer(cpi, 2824 ref_frame); 2825 2826 if (ref == NULL) { 2827 cpi->scaled_ref_idx[ref_frame - 1] = INVALID_IDX; 2828 continue; 2829 } 2830 2831#if CONFIG_VP9_HIGHBITDEPTH 2832 if (ref->y_crop_width != cm->width || ref->y_crop_height != cm->height) { 2833 RefCntBuffer *new_fb_ptr = NULL; 2834 int force_scaling = 0; 2835 int new_fb = cpi->scaled_ref_idx[ref_frame - 1]; 2836 if (new_fb == INVALID_IDX) { 2837 new_fb = get_free_fb(cm); 2838 force_scaling = 1; 2839 } 2840 if (new_fb == INVALID_IDX) 2841 return; 2842 new_fb_ptr = &pool->frame_bufs[new_fb]; 2843 if (force_scaling || 2844 new_fb_ptr->buf.y_crop_width != cm->width || 2845 new_fb_ptr->buf.y_crop_height != cm->height) { 2846 vpx_realloc_frame_buffer(&new_fb_ptr->buf, 2847 cm->width, cm->height, 2848 cm->subsampling_x, cm->subsampling_y, 2849 cm->use_highbitdepth, 2850 VP9_ENC_BORDER_IN_PIXELS, cm->byte_alignment, 2851 NULL, NULL, NULL); 2852 scale_and_extend_frame(ref, &new_fb_ptr->buf, (int)cm->bit_depth); 2853 cpi->scaled_ref_idx[ref_frame - 1] = new_fb; 2854 alloc_frame_mvs(cm, new_fb); 2855 } 2856#else 2857 if (ref->y_crop_width != cm->width || ref->y_crop_height != cm->height) { 2858 RefCntBuffer *new_fb_ptr = NULL; 2859 int force_scaling = 0; 2860 int new_fb = cpi->scaled_ref_idx[ref_frame - 1]; 2861 if (new_fb == INVALID_IDX) { 2862 new_fb = get_free_fb(cm); 2863 force_scaling = 1; 2864 } 2865 if (new_fb == INVALID_IDX) 2866 return; 2867 new_fb_ptr = &pool->frame_bufs[new_fb]; 2868 if (force_scaling || 2869 new_fb_ptr->buf.y_crop_width != cm->width || 2870 new_fb_ptr->buf.y_crop_height != cm->height) { 2871 vpx_realloc_frame_buffer(&new_fb_ptr->buf, 2872 cm->width, cm->height, 2873 cm->subsampling_x, cm->subsampling_y, 2874 VP9_ENC_BORDER_IN_PIXELS, cm->byte_alignment, 2875 NULL, NULL, NULL); 2876 scale_and_extend_frame(ref, &new_fb_ptr->buf); 2877 cpi->scaled_ref_idx[ref_frame - 1] = new_fb; 2878 alloc_frame_mvs(cm, new_fb); 2879 } 2880#endif // CONFIG_VP9_HIGHBITDEPTH 2881 } else { 2882 const int buf_idx = get_ref_frame_buf_idx(cpi, ref_frame); 2883 RefCntBuffer *const buf = &pool->frame_bufs[buf_idx]; 2884 buf->buf.y_crop_width = ref->y_crop_width; 2885 buf->buf.y_crop_height = ref->y_crop_height; 2886 cpi->scaled_ref_idx[ref_frame - 1] = buf_idx; 2887 ++buf->ref_count; 2888 } 2889 } else { 2890 if (cpi->oxcf.pass != 0 || cpi->use_svc) 2891 cpi->scaled_ref_idx[ref_frame - 1] = INVALID_IDX; 2892 } 2893 } 2894} 2895 2896static void release_scaled_references(VP9_COMP *cpi) { 2897 VP9_COMMON *cm = &cpi->common; 2898 int i; 2899 if (cpi->oxcf.pass == 0 && !cpi->use_svc) { 2900 // Only release scaled references under certain conditions: 2901 // if reference will be updated, or if scaled reference has same resolution. 2902 int refresh[3]; 2903 refresh[0] = (cpi->refresh_last_frame) ? 1 : 0; 2904 refresh[1] = (cpi->refresh_golden_frame) ? 1 : 0; 2905 refresh[2] = (cpi->refresh_alt_ref_frame) ? 1 : 0; 2906 for (i = LAST_FRAME; i <= ALTREF_FRAME; ++i) { 2907 const int idx = cpi->scaled_ref_idx[i - 1]; 2908 RefCntBuffer *const buf = idx != INVALID_IDX ? 2909 &cm->buffer_pool->frame_bufs[idx] : NULL; 2910 const YV12_BUFFER_CONFIG *const ref = get_ref_frame_buffer(cpi, i); 2911 if (buf != NULL && 2912 (refresh[i - 1] || 2913 (buf->buf.y_crop_width == ref->y_crop_width && 2914 buf->buf.y_crop_height == ref->y_crop_height))) { 2915 --buf->ref_count; 2916 cpi->scaled_ref_idx[i -1] = INVALID_IDX; 2917 } 2918 } 2919 } else { 2920 for (i = 0; i < MAX_REF_FRAMES; ++i) { 2921 const int idx = cpi->scaled_ref_idx[i]; 2922 RefCntBuffer *const buf = idx != INVALID_IDX ? 2923 &cm->buffer_pool->frame_bufs[idx] : NULL; 2924 if (buf != NULL) { 2925 --buf->ref_count; 2926 cpi->scaled_ref_idx[i] = INVALID_IDX; 2927 } 2928 } 2929 } 2930} 2931 2932static void full_to_model_count(unsigned int *model_count, 2933 unsigned int *full_count) { 2934 int n; 2935 model_count[ZERO_TOKEN] = full_count[ZERO_TOKEN]; 2936 model_count[ONE_TOKEN] = full_count[ONE_TOKEN]; 2937 model_count[TWO_TOKEN] = full_count[TWO_TOKEN]; 2938 for (n = THREE_TOKEN; n < EOB_TOKEN; ++n) 2939 model_count[TWO_TOKEN] += full_count[n]; 2940 model_count[EOB_MODEL_TOKEN] = full_count[EOB_TOKEN]; 2941} 2942 2943static void full_to_model_counts(vp9_coeff_count_model *model_count, 2944 vp9_coeff_count *full_count) { 2945 int i, j, k, l; 2946 2947 for (i = 0; i < PLANE_TYPES; ++i) 2948 for (j = 0; j < REF_TYPES; ++j) 2949 for (k = 0; k < COEF_BANDS; ++k) 2950 for (l = 0; l < BAND_COEFF_CONTEXTS(k); ++l) 2951 full_to_model_count(model_count[i][j][k][l], full_count[i][j][k][l]); 2952} 2953 2954#if 0 && CONFIG_INTERNAL_STATS 2955static void output_frame_level_debug_stats(VP9_COMP *cpi) { 2956 VP9_COMMON *const cm = &cpi->common; 2957 FILE *const f = fopen("tmp.stt", cm->current_video_frame ? "a" : "w"); 2958 int64_t recon_err; 2959 2960 vpx_clear_system_state(); 2961 2962 recon_err = vp9_get_y_sse(cpi->Source, get_frame_new_buffer(cm)); 2963 2964 if (cpi->twopass.total_left_stats.coded_error != 0.0) 2965 fprintf(f, "%10u %dx%d %d %d %10d %10d %10d %10d" 2966 "%10"PRId64" %10"PRId64" %5d %5d %10"PRId64" " 2967 "%10"PRId64" %10"PRId64" %10d " 2968 "%7.2lf %7.2lf %7.2lf %7.2lf %7.2lf" 2969 "%6d %6d %5d %5d %5d " 2970 "%10"PRId64" %10.3lf" 2971 "%10lf %8u %10"PRId64" %10d %10d %10d\n", 2972 cpi->common.current_video_frame, 2973 cm->width, cm->height, 2974 cpi->rc.source_alt_ref_pending, 2975 cpi->rc.source_alt_ref_active, 2976 cpi->rc.this_frame_target, 2977 cpi->rc.projected_frame_size, 2978 cpi->rc.projected_frame_size / cpi->common.MBs, 2979 (cpi->rc.projected_frame_size - cpi->rc.this_frame_target), 2980 cpi->rc.vbr_bits_off_target, 2981 cpi->rc.vbr_bits_off_target_fast, 2982 cpi->twopass.extend_minq, 2983 cpi->twopass.extend_minq_fast, 2984 cpi->rc.total_target_vs_actual, 2985 (cpi->rc.starting_buffer_level - cpi->rc.bits_off_target), 2986 cpi->rc.total_actual_bits, cm->base_qindex, 2987 vp9_convert_qindex_to_q(cm->base_qindex, cm->bit_depth), 2988 (double)vp9_dc_quant(cm->base_qindex, 0, cm->bit_depth) / 4.0, 2989 vp9_convert_qindex_to_q(cpi->twopass.active_worst_quality, 2990 cm->bit_depth), 2991 cpi->rc.avg_q, 2992 vp9_convert_qindex_to_q(cpi->oxcf.cq_level, cm->bit_depth), 2993 cpi->refresh_last_frame, cpi->refresh_golden_frame, 2994 cpi->refresh_alt_ref_frame, cm->frame_type, cpi->rc.gfu_boost, 2995 cpi->twopass.bits_left, 2996 cpi->twopass.total_left_stats.coded_error, 2997 cpi->twopass.bits_left / 2998 (1 + cpi->twopass.total_left_stats.coded_error), 2999 cpi->tot_recode_hits, recon_err, cpi->rc.kf_boost, 3000 cpi->twopass.kf_zeromotion_pct, 3001 cpi->twopass.fr_content_type); 3002 3003 fclose(f); 3004 3005 if (0) { 3006 FILE *const fmodes = fopen("Modes.stt", "a"); 3007 int i; 3008 3009 fprintf(fmodes, "%6d:%1d:%1d:%1d ", cpi->common.current_video_frame, 3010 cm->frame_type, cpi->refresh_golden_frame, 3011 cpi->refresh_alt_ref_frame); 3012 3013 for (i = 0; i < MAX_MODES; ++i) 3014 fprintf(fmodes, "%5d ", cpi->mode_chosen_counts[i]); 3015 3016 fprintf(fmodes, "\n"); 3017 3018 fclose(fmodes); 3019 } 3020} 3021#endif 3022 3023static void set_mv_search_params(VP9_COMP *cpi) { 3024 const VP9_COMMON *const cm = &cpi->common; 3025 const unsigned int max_mv_def = VPXMIN(cm->width, cm->height); 3026 3027 // Default based on max resolution. 3028 cpi->mv_step_param = vp9_init_search_range(max_mv_def); 3029 3030 if (cpi->sf.mv.auto_mv_step_size) { 3031 if (frame_is_intra_only(cm)) { 3032 // Initialize max_mv_magnitude for use in the first INTER frame 3033 // after a key/intra-only frame. 3034 cpi->max_mv_magnitude = max_mv_def; 3035 } else { 3036 if (cm->show_frame) { 3037 // Allow mv_steps to correspond to twice the max mv magnitude found 3038 // in the previous frame, capped by the default max_mv_magnitude based 3039 // on resolution. 3040 cpi->mv_step_param = vp9_init_search_range( 3041 VPXMIN(max_mv_def, 2 * cpi->max_mv_magnitude)); 3042 } 3043 cpi->max_mv_magnitude = 0; 3044 } 3045 } 3046} 3047 3048static void set_size_independent_vars(VP9_COMP *cpi) { 3049 vp9_set_speed_features_framesize_independent(cpi); 3050 vp9_set_rd_speed_thresholds(cpi); 3051 vp9_set_rd_speed_thresholds_sub8x8(cpi); 3052 cpi->common.interp_filter = cpi->sf.default_interp_filter; 3053} 3054 3055static void set_size_dependent_vars(VP9_COMP *cpi, int *q, 3056 int *bottom_index, int *top_index) { 3057 VP9_COMMON *const cm = &cpi->common; 3058 const VP9EncoderConfig *const oxcf = &cpi->oxcf; 3059 3060 // Setup variables that depend on the dimensions of the frame. 3061 vp9_set_speed_features_framesize_dependent(cpi); 3062 3063 // Decide q and q bounds. 3064 *q = vp9_rc_pick_q_and_bounds(cpi, bottom_index, top_index); 3065 3066 if (!frame_is_intra_only(cm)) { 3067 vp9_set_high_precision_mv(cpi, (*q) < HIGH_PRECISION_MV_QTHRESH); 3068 } 3069 3070 // Configure experimental use of segmentation for enhanced coding of 3071 // static regions if indicated. 3072 // Only allowed in the second pass of a two pass encode, as it requires 3073 // lagged coding, and if the relevant speed feature flag is set. 3074 if (oxcf->pass == 2 && cpi->sf.static_segmentation) 3075 configure_static_seg_features(cpi); 3076 3077#if CONFIG_VP9_POSTPROC 3078 if (oxcf->noise_sensitivity > 0) { 3079 int l = 0; 3080 switch (oxcf->noise_sensitivity) { 3081 case 1: 3082 l = 20; 3083 break; 3084 case 2: 3085 l = 40; 3086 break; 3087 case 3: 3088 l = 60; 3089 break; 3090 case 4: 3091 case 5: 3092 l = 100; 3093 break; 3094 case 6: 3095 l = 150; 3096 break; 3097 } 3098 vp9_denoise(cpi->Source, cpi->Source, l); 3099 } 3100#endif // CONFIG_VP9_POSTPROC 3101} 3102 3103#if CONFIG_VP9_TEMPORAL_DENOISING 3104static void setup_denoiser_buffer(VP9_COMP *cpi) { 3105 VP9_COMMON *const cm = &cpi->common; 3106 if (cpi->oxcf.noise_sensitivity > 0 && 3107 !cpi->denoiser.frame_buffer_initialized) { 3108 vp9_denoiser_alloc(&(cpi->denoiser), cm->width, cm->height, 3109 cm->subsampling_x, cm->subsampling_y, 3110#if CONFIG_VP9_HIGHBITDEPTH 3111 cm->use_highbitdepth, 3112#endif 3113 VP9_ENC_BORDER_IN_PIXELS); 3114 } 3115} 3116#endif 3117 3118static void init_motion_estimation(VP9_COMP *cpi) { 3119 int y_stride = cpi->scaled_source.y_stride; 3120 3121 if (cpi->sf.mv.search_method == NSTEP) { 3122 vp9_init3smotion_compensation(&cpi->ss_cfg, y_stride); 3123 } else if (cpi->sf.mv.search_method == DIAMOND) { 3124 vp9_init_dsmotion_compensation(&cpi->ss_cfg, y_stride); 3125 } 3126} 3127 3128static void set_frame_size(VP9_COMP *cpi) { 3129 int ref_frame; 3130 VP9_COMMON *const cm = &cpi->common; 3131 VP9EncoderConfig *const oxcf = &cpi->oxcf; 3132 MACROBLOCKD *const xd = &cpi->td.mb.e_mbd; 3133 3134 if (oxcf->pass == 2 && 3135 oxcf->rc_mode == VPX_VBR && 3136 ((oxcf->resize_mode == RESIZE_FIXED && cm->current_video_frame == 0) || 3137 (oxcf->resize_mode == RESIZE_DYNAMIC && cpi->resize_pending))) { 3138 calculate_coded_size( 3139 cpi, &oxcf->scaled_frame_width, &oxcf->scaled_frame_height); 3140 3141 // There has been a change in frame size. 3142 vp9_set_size_literal(cpi, oxcf->scaled_frame_width, 3143 oxcf->scaled_frame_height); 3144 } 3145 3146 if (oxcf->pass == 0 && 3147 oxcf->rc_mode == VPX_CBR && 3148 !cpi->use_svc && 3149 oxcf->resize_mode == RESIZE_DYNAMIC && 3150 cpi->resize_pending != 0) { 3151 oxcf->scaled_frame_width = 3152 (oxcf->width * cpi->resize_scale_num) / cpi->resize_scale_den; 3153 oxcf->scaled_frame_height = 3154 (oxcf->height * cpi->resize_scale_num) /cpi->resize_scale_den; 3155 // There has been a change in frame size. 3156 vp9_set_size_literal(cpi, 3157 oxcf->scaled_frame_width, 3158 oxcf->scaled_frame_height); 3159 3160 // TODO(agrange) Scale cpi->max_mv_magnitude if frame-size has changed. 3161 set_mv_search_params(cpi); 3162 3163#if CONFIG_VP9_TEMPORAL_DENOISING 3164 // Reset the denoiser on the resized frame. 3165 if (cpi->oxcf.noise_sensitivity > 0) { 3166 vp9_denoiser_free(&(cpi->denoiser)); 3167 setup_denoiser_buffer(cpi); 3168 // Dynamic resize is only triggered for non-SVC, so we can force 3169 // golden frame update here as temporary fix to denoiser. 3170 cpi->refresh_golden_frame = 1; 3171 } 3172#endif 3173 } 3174 3175 if ((oxcf->pass == 2) && 3176 (!cpi->use_svc || 3177 (is_two_pass_svc(cpi) && 3178 cpi->svc.encode_empty_frame_state != ENCODING))) { 3179 vp9_set_target_rate(cpi); 3180 } 3181 3182 alloc_frame_mvs(cm, cm->new_fb_idx); 3183 3184 // Reset the frame pointers to the current frame size. 3185 vpx_realloc_frame_buffer(get_frame_new_buffer(cm), 3186 cm->width, cm->height, 3187 cm->subsampling_x, cm->subsampling_y, 3188#if CONFIG_VP9_HIGHBITDEPTH 3189 cm->use_highbitdepth, 3190#endif 3191 VP9_ENC_BORDER_IN_PIXELS, cm->byte_alignment, 3192 NULL, NULL, NULL); 3193 3194 alloc_util_frame_buffers(cpi); 3195 init_motion_estimation(cpi); 3196 3197 for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ++ref_frame) { 3198 RefBuffer *const ref_buf = &cm->frame_refs[ref_frame - 1]; 3199 const int buf_idx = get_ref_frame_buf_idx(cpi, ref_frame); 3200 3201 ref_buf->idx = buf_idx; 3202 3203 if (buf_idx != INVALID_IDX) { 3204 YV12_BUFFER_CONFIG *const buf = &cm->buffer_pool->frame_bufs[buf_idx].buf; 3205 ref_buf->buf = buf; 3206#if CONFIG_VP9_HIGHBITDEPTH 3207 vp9_setup_scale_factors_for_frame(&ref_buf->sf, 3208 buf->y_crop_width, buf->y_crop_height, 3209 cm->width, cm->height, 3210 (buf->flags & YV12_FLAG_HIGHBITDEPTH) ? 3211 1 : 0); 3212#else 3213 vp9_setup_scale_factors_for_frame(&ref_buf->sf, 3214 buf->y_crop_width, buf->y_crop_height, 3215 cm->width, cm->height); 3216#endif // CONFIG_VP9_HIGHBITDEPTH 3217 if (vp9_is_scaled(&ref_buf->sf)) 3218 vpx_extend_frame_borders(buf); 3219 } else { 3220 ref_buf->buf = NULL; 3221 } 3222 } 3223 3224 set_ref_ptrs(cm, xd, LAST_FRAME, LAST_FRAME); 3225} 3226 3227static void encode_without_recode_loop(VP9_COMP *cpi, 3228 size_t *size, 3229 uint8_t *dest) { 3230 VP9_COMMON *const cm = &cpi->common; 3231 int q = 0, bottom_index = 0, top_index = 0; // Dummy variables. 3232 3233 vpx_clear_system_state(); 3234 3235 set_frame_size(cpi); 3236 3237 cpi->Source = vp9_scale_if_required(cm, 3238 cpi->un_scaled_source, 3239 &cpi->scaled_source, 3240 (cpi->oxcf.pass == 0)); 3241 3242 // Avoid scaling last_source unless its needed. 3243 // Last source is currently only used for screen-content mode, 3244 // or if partition_search_type == SOURCE_VAR_BASED_PARTITION. 3245 if (cpi->unscaled_last_source != NULL && 3246 (cpi->oxcf.content == VP9E_CONTENT_SCREEN || 3247 cpi->sf.partition_search_type == SOURCE_VAR_BASED_PARTITION)) 3248 cpi->Last_Source = vp9_scale_if_required(cm, 3249 cpi->unscaled_last_source, 3250 &cpi->scaled_last_source, 3251 (cpi->oxcf.pass == 0)); 3252 3253#if CONFIG_VP9_TEMPORAL_DENOISING 3254 if (cpi->oxcf.noise_sensitivity > 0 && 3255 cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ) { 3256 vp9_denoiser_update_noise_estimate(cpi); 3257 } 3258#endif 3259 3260 if (cpi->oxcf.pass == 0 && 3261 cpi->oxcf.rc_mode == VPX_CBR && 3262 cpi->resize_state == 0 && 3263 cm->frame_type != KEY_FRAME && 3264 cpi->oxcf.content == VP9E_CONTENT_SCREEN) 3265 vp9_avg_source_sad(cpi); 3266 3267 if (frame_is_intra_only(cm) == 0) { 3268 vp9_scale_references(cpi); 3269 } 3270 3271 set_size_independent_vars(cpi); 3272 set_size_dependent_vars(cpi, &q, &bottom_index, &top_index); 3273 3274 vp9_set_quantizer(cm, q); 3275 vp9_set_variance_partition_thresholds(cpi, q); 3276 3277 setup_frame(cpi); 3278 3279 suppress_active_map(cpi); 3280 // Variance adaptive and in frame q adjustment experiments are mutually 3281 // exclusive. 3282 if (cpi->oxcf.aq_mode == VARIANCE_AQ) { 3283 vp9_vaq_frame_setup(cpi); 3284 } else if (cpi->oxcf.aq_mode == COMPLEXITY_AQ) { 3285 vp9_setup_in_frame_q_adj(cpi); 3286 } else if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ) { 3287 vp9_cyclic_refresh_setup(cpi); 3288 } 3289 apply_active_map(cpi); 3290 3291 // transform / motion compensation build reconstruction frame 3292 vp9_encode_frame(cpi); 3293 3294 // Check if we should drop this frame because of high overshoot. 3295 // Only for frames where high temporal-source sad is detected. 3296 if (cpi->oxcf.pass == 0 && 3297 cpi->oxcf.rc_mode == VPX_CBR && 3298 cpi->resize_state == 0 && 3299 cm->frame_type != KEY_FRAME && 3300 cpi->oxcf.content == VP9E_CONTENT_SCREEN && 3301 cpi->rc.high_source_sad == 1) { 3302 int frame_size = 0; 3303 // Get an estimate of the encoded frame size. 3304 save_coding_context(cpi); 3305 vp9_pack_bitstream(cpi, dest, size); 3306 restore_coding_context(cpi); 3307 frame_size = (int)(*size) << 3; 3308 // Check if encoded frame will overshoot too much, and if so, set the q and 3309 // adjust some rate control parameters, and return to re-encode the frame. 3310 if (vp9_encodedframe_overshoot(cpi, frame_size, &q)) { 3311 vpx_clear_system_state(); 3312 vp9_set_quantizer(cm, q); 3313 vp9_set_variance_partition_thresholds(cpi, q); 3314 suppress_active_map(cpi); 3315 // Turn-off cyclic refresh for re-encoded frame. 3316 if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ) { 3317 unsigned char *const seg_map = cpi->segmentation_map; 3318 memset(seg_map, 0, cm->mi_rows * cm->mi_cols); 3319 vp9_disable_segmentation(&cm->seg); 3320 } 3321 apply_active_map(cpi); 3322 vp9_encode_frame(cpi); 3323 } 3324 } 3325 3326 // Update some stats from cyclic refresh, and check if we should not update 3327 // golden reference, for non-SVC 1 pass CBR. 3328 if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ && 3329 cm->frame_type != KEY_FRAME && 3330 !cpi->use_svc && 3331 cpi->ext_refresh_frame_flags_pending == 0 && 3332 (cpi->oxcf.pass == 0 && cpi->oxcf.rc_mode == VPX_CBR)) 3333 vp9_cyclic_refresh_check_golden_update(cpi); 3334 3335 // Update the skip mb flag probabilities based on the distribution 3336 // seen in the last encoder iteration. 3337 // update_base_skip_probs(cpi); 3338 vpx_clear_system_state(); 3339} 3340 3341static void encode_with_recode_loop(VP9_COMP *cpi, 3342 size_t *size, 3343 uint8_t *dest) { 3344 VP9_COMMON *const cm = &cpi->common; 3345 RATE_CONTROL *const rc = &cpi->rc; 3346 int bottom_index, top_index; 3347 int loop_count = 0; 3348 int loop_at_this_size = 0; 3349 int loop = 0; 3350 int overshoot_seen = 0; 3351 int undershoot_seen = 0; 3352 int frame_over_shoot_limit; 3353 int frame_under_shoot_limit; 3354 int q = 0, q_low = 0, q_high = 0; 3355 3356 set_size_independent_vars(cpi); 3357 3358 do { 3359 vpx_clear_system_state(); 3360 3361 set_frame_size(cpi); 3362 3363 if (loop_count == 0 || cpi->resize_pending != 0) { 3364 set_size_dependent_vars(cpi, &q, &bottom_index, &top_index); 3365 3366 // TODO(agrange) Scale cpi->max_mv_magnitude if frame-size has changed. 3367 set_mv_search_params(cpi); 3368 3369 // Reset the loop state for new frame size. 3370 overshoot_seen = 0; 3371 undershoot_seen = 0; 3372 3373 // Reconfiguration for change in frame size has concluded. 3374 cpi->resize_pending = 0; 3375 3376 q_low = bottom_index; 3377 q_high = top_index; 3378 3379 loop_at_this_size = 0; 3380 } 3381 3382 // Decide frame size bounds first time through. 3383 if (loop_count == 0) { 3384 vp9_rc_compute_frame_size_bounds(cpi, rc->this_frame_target, 3385 &frame_under_shoot_limit, 3386 &frame_over_shoot_limit); 3387 } 3388 3389 cpi->Source = vp9_scale_if_required(cm, cpi->un_scaled_source, 3390 &cpi->scaled_source, 3391 (cpi->oxcf.pass == 0)); 3392 3393 if (cpi->unscaled_last_source != NULL) 3394 cpi->Last_Source = vp9_scale_if_required(cm, cpi->unscaled_last_source, 3395 &cpi->scaled_last_source, 3396 (cpi->oxcf.pass == 0)); 3397 3398 if (frame_is_intra_only(cm) == 0) { 3399 if (loop_count > 0) { 3400 release_scaled_references(cpi); 3401 } 3402 vp9_scale_references(cpi); 3403 } 3404 3405 vp9_set_quantizer(cm, q); 3406 3407 if (loop_count == 0) 3408 setup_frame(cpi); 3409 3410 // Variance adaptive and in frame q adjustment experiments are mutually 3411 // exclusive. 3412 if (cpi->oxcf.aq_mode == VARIANCE_AQ) { 3413 vp9_vaq_frame_setup(cpi); 3414 } else if (cpi->oxcf.aq_mode == COMPLEXITY_AQ) { 3415 vp9_setup_in_frame_q_adj(cpi); 3416 } 3417 3418 // transform / motion compensation build reconstruction frame 3419 vp9_encode_frame(cpi); 3420 3421 // Update the skip mb flag probabilities based on the distribution 3422 // seen in the last encoder iteration. 3423 // update_base_skip_probs(cpi); 3424 3425 vpx_clear_system_state(); 3426 3427 // Dummy pack of the bitstream using up to date stats to get an 3428 // accurate estimate of output frame size to determine if we need 3429 // to recode. 3430 if (cpi->sf.recode_loop >= ALLOW_RECODE_KFARFGF) { 3431 save_coding_context(cpi); 3432 if (!cpi->sf.use_nonrd_pick_mode) 3433 vp9_pack_bitstream(cpi, dest, size); 3434 3435 rc->projected_frame_size = (int)(*size) << 3; 3436 restore_coding_context(cpi); 3437 3438 if (frame_over_shoot_limit == 0) 3439 frame_over_shoot_limit = 1; 3440 } 3441 3442 if (cpi->oxcf.rc_mode == VPX_Q) { 3443 loop = 0; 3444 } else { 3445 if ((cm->frame_type == KEY_FRAME) && 3446 rc->this_key_frame_forced && 3447 (rc->projected_frame_size < rc->max_frame_bandwidth)) { 3448 int last_q = q; 3449 int64_t kf_err; 3450 3451 int64_t high_err_target = cpi->ambient_err; 3452 int64_t low_err_target = cpi->ambient_err >> 1; 3453 3454#if CONFIG_VP9_HIGHBITDEPTH 3455 if (cm->use_highbitdepth) { 3456 kf_err = vp9_highbd_get_y_sse(cpi->Source, get_frame_new_buffer(cm)); 3457 } else { 3458 kf_err = vp9_get_y_sse(cpi->Source, get_frame_new_buffer(cm)); 3459 } 3460#else 3461 kf_err = vp9_get_y_sse(cpi->Source, get_frame_new_buffer(cm)); 3462#endif // CONFIG_VP9_HIGHBITDEPTH 3463 3464 // Prevent possible divide by zero error below for perfect KF 3465 kf_err += !kf_err; 3466 3467 // The key frame is not good enough or we can afford 3468 // to make it better without undue risk of popping. 3469 if ((kf_err > high_err_target && 3470 rc->projected_frame_size <= frame_over_shoot_limit) || 3471 (kf_err > low_err_target && 3472 rc->projected_frame_size <= frame_under_shoot_limit)) { 3473 // Lower q_high 3474 q_high = q > q_low ? q - 1 : q_low; 3475 3476 // Adjust Q 3477 q = (int)((q * high_err_target) / kf_err); 3478 q = VPXMIN(q, (q_high + q_low) >> 1); 3479 } else if (kf_err < low_err_target && 3480 rc->projected_frame_size >= frame_under_shoot_limit) { 3481 // The key frame is much better than the previous frame 3482 // Raise q_low 3483 q_low = q < q_high ? q + 1 : q_high; 3484 3485 // Adjust Q 3486 q = (int)((q * low_err_target) / kf_err); 3487 q = VPXMIN(q, (q_high + q_low + 1) >> 1); 3488 } 3489 3490 // Clamp Q to upper and lower limits: 3491 q = clamp(q, q_low, q_high); 3492 3493 loop = q != last_q; 3494 } else if (recode_loop_test( 3495 cpi, frame_over_shoot_limit, frame_under_shoot_limit, 3496 q, VPXMAX(q_high, top_index), bottom_index)) { 3497 // Is the projected frame size out of range and are we allowed 3498 // to attempt to recode. 3499 int last_q = q; 3500 int retries = 0; 3501 3502 if (cpi->resize_pending == 1) { 3503 // Change in frame size so go back around the recode loop. 3504 cpi->rc.frame_size_selector = 3505 SCALE_STEP1 - cpi->rc.frame_size_selector; 3506 cpi->rc.next_frame_size_selector = cpi->rc.frame_size_selector; 3507 3508#if CONFIG_INTERNAL_STATS 3509 ++cpi->tot_recode_hits; 3510#endif 3511 ++loop_count; 3512 loop = 1; 3513 continue; 3514 } 3515 3516 // Frame size out of permitted range: 3517 // Update correction factor & compute new Q to try... 3518 3519 // Frame is too large 3520 if (rc->projected_frame_size > rc->this_frame_target) { 3521 // Special case if the projected size is > the max allowed. 3522 if (rc->projected_frame_size >= rc->max_frame_bandwidth) 3523 q_high = rc->worst_quality; 3524 3525 // Raise Qlow as to at least the current value 3526 q_low = q < q_high ? q + 1 : q_high; 3527 3528 if (undershoot_seen || loop_at_this_size > 1) { 3529 // Update rate_correction_factor unless 3530 vp9_rc_update_rate_correction_factors(cpi); 3531 3532 q = (q_high + q_low + 1) / 2; 3533 } else { 3534 // Update rate_correction_factor unless 3535 vp9_rc_update_rate_correction_factors(cpi); 3536 3537 q = vp9_rc_regulate_q(cpi, rc->this_frame_target, 3538 bottom_index, VPXMAX(q_high, top_index)); 3539 3540 while (q < q_low && retries < 10) { 3541 vp9_rc_update_rate_correction_factors(cpi); 3542 q = vp9_rc_regulate_q(cpi, rc->this_frame_target, 3543 bottom_index, VPXMAX(q_high, top_index)); 3544 retries++; 3545 } 3546 } 3547 3548 overshoot_seen = 1; 3549 } else { 3550 // Frame is too small 3551 q_high = q > q_low ? q - 1 : q_low; 3552 3553 if (overshoot_seen || loop_at_this_size > 1) { 3554 vp9_rc_update_rate_correction_factors(cpi); 3555 q = (q_high + q_low) / 2; 3556 } else { 3557 vp9_rc_update_rate_correction_factors(cpi); 3558 q = vp9_rc_regulate_q(cpi, rc->this_frame_target, 3559 bottom_index, top_index); 3560 // Special case reset for qlow for constrained quality. 3561 // This should only trigger where there is very substantial 3562 // undershoot on a frame and the auto cq level is above 3563 // the user passsed in value. 3564 if (cpi->oxcf.rc_mode == VPX_CQ && 3565 q < q_low) { 3566 q_low = q; 3567 } 3568 3569 while (q > q_high && retries < 10) { 3570 vp9_rc_update_rate_correction_factors(cpi); 3571 q = vp9_rc_regulate_q(cpi, rc->this_frame_target, 3572 bottom_index, top_index); 3573 retries++; 3574 } 3575 } 3576 3577 undershoot_seen = 1; 3578 } 3579 3580 // Clamp Q to upper and lower limits: 3581 q = clamp(q, q_low, q_high); 3582 3583 loop = (q != last_q); 3584 } else { 3585 loop = 0; 3586 } 3587 } 3588 3589 // Special case for overlay frame. 3590 if (rc->is_src_frame_alt_ref && 3591 rc->projected_frame_size < rc->max_frame_bandwidth) 3592 loop = 0; 3593 3594 if (loop) { 3595 ++loop_count; 3596 ++loop_at_this_size; 3597 3598#if CONFIG_INTERNAL_STATS 3599 ++cpi->tot_recode_hits; 3600#endif 3601 } 3602 } while (loop); 3603} 3604 3605static int get_ref_frame_flags(const VP9_COMP *cpi) { 3606 const int *const map = cpi->common.ref_frame_map; 3607 const int gold_is_last = map[cpi->gld_fb_idx] == map[cpi->lst_fb_idx]; 3608 const int alt_is_last = map[cpi->alt_fb_idx] == map[cpi->lst_fb_idx]; 3609 const int gold_is_alt = map[cpi->gld_fb_idx] == map[cpi->alt_fb_idx]; 3610 int flags = VP9_ALT_FLAG | VP9_GOLD_FLAG | VP9_LAST_FLAG; 3611 3612 if (gold_is_last) 3613 flags &= ~VP9_GOLD_FLAG; 3614 3615 if (cpi->rc.frames_till_gf_update_due == INT_MAX && 3616 (cpi->svc.number_temporal_layers == 1 && 3617 cpi->svc.number_spatial_layers == 1)) 3618 flags &= ~VP9_GOLD_FLAG; 3619 3620 if (alt_is_last) 3621 flags &= ~VP9_ALT_FLAG; 3622 3623 if (gold_is_alt) 3624 flags &= ~VP9_ALT_FLAG; 3625 3626 return flags; 3627} 3628 3629static void set_ext_overrides(VP9_COMP *cpi) { 3630 // Overrides the defaults with the externally supplied values with 3631 // vp9_update_reference() and vp9_update_entropy() calls 3632 // Note: The overrides are valid only for the next frame passed 3633 // to encode_frame_to_data_rate() function 3634 if (cpi->ext_refresh_frame_context_pending) { 3635 cpi->common.refresh_frame_context = cpi->ext_refresh_frame_context; 3636 cpi->ext_refresh_frame_context_pending = 0; 3637 } 3638 if (cpi->ext_refresh_frame_flags_pending) { 3639 cpi->refresh_last_frame = cpi->ext_refresh_last_frame; 3640 cpi->refresh_golden_frame = cpi->ext_refresh_golden_frame; 3641 cpi->refresh_alt_ref_frame = cpi->ext_refresh_alt_ref_frame; 3642 } 3643} 3644 3645YV12_BUFFER_CONFIG *vp9_scale_if_required(VP9_COMMON *cm, 3646 YV12_BUFFER_CONFIG *unscaled, 3647 YV12_BUFFER_CONFIG *scaled, 3648 int use_normative_scaler) { 3649 if (cm->mi_cols * MI_SIZE != unscaled->y_width || 3650 cm->mi_rows * MI_SIZE != unscaled->y_height) { 3651#if CONFIG_VP9_HIGHBITDEPTH 3652 if (use_normative_scaler) 3653 scale_and_extend_frame(unscaled, scaled, (int)cm->bit_depth); 3654 else 3655 scale_and_extend_frame_nonnormative(unscaled, scaled, (int)cm->bit_depth); 3656#else 3657 if (use_normative_scaler) 3658 scale_and_extend_frame(unscaled, scaled); 3659 else 3660 scale_and_extend_frame_nonnormative(unscaled, scaled); 3661#endif // CONFIG_VP9_HIGHBITDEPTH 3662 return scaled; 3663 } else { 3664 return unscaled; 3665 } 3666} 3667 3668static void set_arf_sign_bias(VP9_COMP *cpi) { 3669 VP9_COMMON *const cm = &cpi->common; 3670 int arf_sign_bias; 3671 3672 if ((cpi->oxcf.pass == 2) && cpi->multi_arf_allowed) { 3673 const GF_GROUP *const gf_group = &cpi->twopass.gf_group; 3674 arf_sign_bias = cpi->rc.source_alt_ref_active && 3675 (!cpi->refresh_alt_ref_frame || 3676 (gf_group->rf_level[gf_group->index] == GF_ARF_LOW)); 3677 } else { 3678 arf_sign_bias = 3679 (cpi->rc.source_alt_ref_active && !cpi->refresh_alt_ref_frame); 3680 } 3681 cm->ref_frame_sign_bias[ALTREF_FRAME] = arf_sign_bias; 3682} 3683 3684static int setup_interp_filter_search_mask(VP9_COMP *cpi) { 3685 INTERP_FILTER ifilter; 3686 int ref_total[MAX_REF_FRAMES] = {0}; 3687 MV_REFERENCE_FRAME ref; 3688 int mask = 0; 3689 if (cpi->common.last_frame_type == KEY_FRAME || 3690 cpi->refresh_alt_ref_frame) 3691 return mask; 3692 for (ref = LAST_FRAME; ref <= ALTREF_FRAME; ++ref) 3693 for (ifilter = EIGHTTAP; ifilter <= EIGHTTAP_SHARP; ++ifilter) 3694 ref_total[ref] += cpi->interp_filter_selected[ref][ifilter]; 3695 3696 for (ifilter = EIGHTTAP; ifilter <= EIGHTTAP_SHARP; ++ifilter) { 3697 if ((ref_total[LAST_FRAME] && 3698 cpi->interp_filter_selected[LAST_FRAME][ifilter] == 0) && 3699 (ref_total[GOLDEN_FRAME] == 0 || 3700 cpi->interp_filter_selected[GOLDEN_FRAME][ifilter] * 50 3701 < ref_total[GOLDEN_FRAME]) && 3702 (ref_total[ALTREF_FRAME] == 0 || 3703 cpi->interp_filter_selected[ALTREF_FRAME][ifilter] * 50 3704 < ref_total[ALTREF_FRAME])) 3705 mask |= 1 << ifilter; 3706 } 3707 return mask; 3708} 3709 3710static void encode_frame_to_data_rate(VP9_COMP *cpi, 3711 size_t *size, 3712 uint8_t *dest, 3713 unsigned int *frame_flags) { 3714 VP9_COMMON *const cm = &cpi->common; 3715 const VP9EncoderConfig *const oxcf = &cpi->oxcf; 3716 struct segmentation *const seg = &cm->seg; 3717 TX_SIZE t; 3718 3719 set_ext_overrides(cpi); 3720 vpx_clear_system_state(); 3721 3722 // Set the arf sign bias for this frame. 3723 set_arf_sign_bias(cpi); 3724 3725 // Set default state for segment based loop filter update flags. 3726 cm->lf.mode_ref_delta_update = 0; 3727 3728 if (cpi->oxcf.pass == 2 && 3729 cpi->sf.adaptive_interp_filter_search) 3730 cpi->sf.interp_filter_search_mask = 3731 setup_interp_filter_search_mask(cpi); 3732 3733 // Set various flags etc to special state if it is a key frame. 3734 if (frame_is_intra_only(cm)) { 3735 // Reset the loop filter deltas and segmentation map. 3736 vp9_reset_segment_features(&cm->seg); 3737 3738 // If segmentation is enabled force a map update for key frames. 3739 if (seg->enabled) { 3740 seg->update_map = 1; 3741 seg->update_data = 1; 3742 } 3743 3744 // The alternate reference frame cannot be active for a key frame. 3745 cpi->rc.source_alt_ref_active = 0; 3746 3747 cm->error_resilient_mode = oxcf->error_resilient_mode; 3748 cm->frame_parallel_decoding_mode = oxcf->frame_parallel_decoding_mode; 3749 3750 // By default, encoder assumes decoder can use prev_mi. 3751 if (cm->error_resilient_mode) { 3752 cm->frame_parallel_decoding_mode = 1; 3753 cm->reset_frame_context = 0; 3754 cm->refresh_frame_context = 0; 3755 } else if (cm->intra_only) { 3756 // Only reset the current context. 3757 cm->reset_frame_context = 2; 3758 } 3759 } 3760 if (is_two_pass_svc(cpi) && cm->error_resilient_mode == 0) { 3761 // Use context 0 for intra only empty frame, but the last frame context 3762 // for other empty frames. 3763 if (cpi->svc.encode_empty_frame_state == ENCODING) { 3764 if (cpi->svc.encode_intra_empty_frame != 0) 3765 cm->frame_context_idx = 0; 3766 else 3767 cm->frame_context_idx = FRAME_CONTEXTS - 1; 3768 } else { 3769 cm->frame_context_idx = 3770 cpi->svc.spatial_layer_id * cpi->svc.number_temporal_layers + 3771 cpi->svc.temporal_layer_id; 3772 } 3773 3774 cm->frame_parallel_decoding_mode = oxcf->frame_parallel_decoding_mode; 3775 3776 // The probs will be updated based on the frame type of its previous 3777 // frame if frame_parallel_decoding_mode is 0. The type may vary for 3778 // the frame after a key frame in base layer since we may drop enhancement 3779 // layers. So set frame_parallel_decoding_mode to 1 in this case. 3780 if (cm->frame_parallel_decoding_mode == 0) { 3781 if (cpi->svc.number_temporal_layers == 1) { 3782 if (cpi->svc.spatial_layer_id == 0 && 3783 cpi->svc.layer_context[0].last_frame_type == KEY_FRAME) 3784 cm->frame_parallel_decoding_mode = 1; 3785 } else if (cpi->svc.spatial_layer_id == 0) { 3786 // Find the 2nd frame in temporal base layer and 1st frame in temporal 3787 // enhancement layers from the key frame. 3788 int i; 3789 for (i = 0; i < cpi->svc.number_temporal_layers; ++i) { 3790 if (cpi->svc.layer_context[0].frames_from_key_frame == 1 << i) { 3791 cm->frame_parallel_decoding_mode = 1; 3792 break; 3793 } 3794 } 3795 } 3796 } 3797 } 3798 3799 // For 1 pass CBR, check if we are dropping this frame. 3800 // Never drop on key frame. 3801 if (oxcf->pass == 0 && 3802 oxcf->rc_mode == VPX_CBR && 3803 cm->frame_type != KEY_FRAME) { 3804 if (vp9_rc_drop_frame(cpi)) { 3805 vp9_rc_postencode_update_drop_frame(cpi); 3806 ++cm->current_video_frame; 3807 cpi->ext_refresh_frame_flags_pending = 0; 3808 return; 3809 } 3810 } 3811 3812 vpx_clear_system_state(); 3813 3814#if CONFIG_INTERNAL_STATS 3815 memset(cpi->mode_chosen_counts, 0, 3816 MAX_MODES * sizeof(*cpi->mode_chosen_counts)); 3817#endif 3818 3819 if (cpi->sf.recode_loop == DISALLOW_RECODE) { 3820 encode_without_recode_loop(cpi, size, dest); 3821 } else { 3822 encode_with_recode_loop(cpi, size, dest); 3823 } 3824 3825#if CONFIG_VP9_TEMPORAL_DENOISING 3826#ifdef OUTPUT_YUV_DENOISED 3827 if (oxcf->noise_sensitivity > 0) { 3828 vp9_write_yuv_frame_420(&cpi->denoiser.running_avg_y[INTRA_FRAME], 3829 yuv_denoised_file); 3830 } 3831#endif 3832#endif 3833#ifdef OUTPUT_YUV_SKINMAP 3834 if (cpi->common.current_video_frame > 1) { 3835 vp9_compute_skin_map(cpi, yuv_skinmap_file); 3836 } 3837#endif 3838 3839 // Special case code to reduce pulsing when key frames are forced at a 3840 // fixed interval. Note the reconstruction error if it is the frame before 3841 // the force key frame 3842 if (cpi->rc.next_key_frame_forced && cpi->rc.frames_to_key == 1) { 3843#if CONFIG_VP9_HIGHBITDEPTH 3844 if (cm->use_highbitdepth) { 3845 cpi->ambient_err = vp9_highbd_get_y_sse(cpi->Source, 3846 get_frame_new_buffer(cm)); 3847 } else { 3848 cpi->ambient_err = vp9_get_y_sse(cpi->Source, get_frame_new_buffer(cm)); 3849 } 3850#else 3851 cpi->ambient_err = vp9_get_y_sse(cpi->Source, get_frame_new_buffer(cm)); 3852#endif // CONFIG_VP9_HIGHBITDEPTH 3853 } 3854 3855 // If the encoder forced a KEY_FRAME decision 3856 if (cm->frame_type == KEY_FRAME) 3857 cpi->refresh_last_frame = 1; 3858 3859 cm->frame_to_show = get_frame_new_buffer(cm); 3860 cm->frame_to_show->color_space = cm->color_space; 3861 cm->frame_to_show->color_range = cm->color_range; 3862 cm->frame_to_show->render_width = cm->render_width; 3863 cm->frame_to_show->render_height = cm->render_height; 3864 3865 // Pick the loop filter level for the frame. 3866 loopfilter_frame(cpi, cm); 3867 3868 // build the bitstream 3869 vp9_pack_bitstream(cpi, dest, size); 3870 3871 if (cm->seg.update_map) 3872 update_reference_segmentation_map(cpi); 3873 3874 if (frame_is_intra_only(cm) == 0) { 3875 release_scaled_references(cpi); 3876 } 3877 vp9_update_reference_frames(cpi); 3878 3879 for (t = TX_4X4; t <= TX_32X32; t++) 3880 full_to_model_counts(cpi->td.counts->coef[t], 3881 cpi->td.rd_counts.coef_counts[t]); 3882 3883 if (!cm->error_resilient_mode && !cm->frame_parallel_decoding_mode) 3884 vp9_adapt_coef_probs(cm); 3885 3886 if (!frame_is_intra_only(cm)) { 3887 if (!cm->error_resilient_mode && !cm->frame_parallel_decoding_mode) { 3888 vp9_adapt_mode_probs(cm); 3889 vp9_adapt_mv_probs(cm, cm->allow_high_precision_mv); 3890 } 3891 } 3892 3893 cpi->ext_refresh_frame_flags_pending = 0; 3894 3895 if (cpi->refresh_golden_frame == 1) 3896 cpi->frame_flags |= FRAMEFLAGS_GOLDEN; 3897 else 3898 cpi->frame_flags &= ~FRAMEFLAGS_GOLDEN; 3899 3900 if (cpi->refresh_alt_ref_frame == 1) 3901 cpi->frame_flags |= FRAMEFLAGS_ALTREF; 3902 else 3903 cpi->frame_flags &= ~FRAMEFLAGS_ALTREF; 3904 3905 cpi->ref_frame_flags = get_ref_frame_flags(cpi); 3906 3907 cm->last_frame_type = cm->frame_type; 3908 3909 if (!(is_two_pass_svc(cpi) && cpi->svc.encode_empty_frame_state == ENCODING)) 3910 vp9_rc_postencode_update(cpi, *size); 3911 3912#if 0 3913 output_frame_level_debug_stats(cpi); 3914#endif 3915 3916 if (cm->frame_type == KEY_FRAME) { 3917 // Tell the caller that the frame was coded as a key frame 3918 *frame_flags = cpi->frame_flags | FRAMEFLAGS_KEY; 3919 } else { 3920 *frame_flags = cpi->frame_flags & ~FRAMEFLAGS_KEY; 3921 } 3922 3923 // Clear the one shot update flags for segmentation map and mode/ref loop 3924 // filter deltas. 3925 cm->seg.update_map = 0; 3926 cm->seg.update_data = 0; 3927 cm->lf.mode_ref_delta_update = 0; 3928 3929 // keep track of the last coded dimensions 3930 cm->last_width = cm->width; 3931 cm->last_height = cm->height; 3932 3933 // reset to normal state now that we are done. 3934 if (!cm->show_existing_frame) 3935 cm->last_show_frame = cm->show_frame; 3936 3937 if (cm->show_frame) { 3938 vp9_swap_mi_and_prev_mi(cm); 3939 // Don't increment frame counters if this was an altref buffer 3940 // update not a real frame 3941 ++cm->current_video_frame; 3942 if (cpi->use_svc) 3943 vp9_inc_frame_in_layer(cpi); 3944 } 3945 cm->prev_frame = cm->cur_frame; 3946 3947 if (cpi->use_svc) 3948 cpi->svc.layer_context[cpi->svc.spatial_layer_id * 3949 cpi->svc.number_temporal_layers + 3950 cpi->svc.temporal_layer_id].last_frame_type = 3951 cm->frame_type; 3952} 3953 3954static void SvcEncode(VP9_COMP *cpi, size_t *size, uint8_t *dest, 3955 unsigned int *frame_flags) { 3956 vp9_rc_get_svc_params(cpi); 3957 encode_frame_to_data_rate(cpi, size, dest, frame_flags); 3958} 3959 3960static void Pass0Encode(VP9_COMP *cpi, size_t *size, uint8_t *dest, 3961 unsigned int *frame_flags) { 3962 if (cpi->oxcf.rc_mode == VPX_CBR) { 3963 vp9_rc_get_one_pass_cbr_params(cpi); 3964 } else { 3965 vp9_rc_get_one_pass_vbr_params(cpi); 3966 } 3967 encode_frame_to_data_rate(cpi, size, dest, frame_flags); 3968} 3969 3970static void Pass2Encode(VP9_COMP *cpi, size_t *size, 3971 uint8_t *dest, unsigned int *frame_flags) { 3972 cpi->allow_encode_breakout = ENCODE_BREAKOUT_ENABLED; 3973 encode_frame_to_data_rate(cpi, size, dest, frame_flags); 3974 3975 if (!(is_two_pass_svc(cpi) && cpi->svc.encode_empty_frame_state == ENCODING)) 3976 vp9_twopass_postencode_update(cpi); 3977} 3978 3979static void init_ref_frame_bufs(VP9_COMMON *cm) { 3980 int i; 3981 BufferPool *const pool = cm->buffer_pool; 3982 cm->new_fb_idx = INVALID_IDX; 3983 for (i = 0; i < REF_FRAMES; ++i) { 3984 cm->ref_frame_map[i] = INVALID_IDX; 3985 pool->frame_bufs[i].ref_count = 0; 3986 } 3987} 3988 3989static void check_initial_width(VP9_COMP *cpi, 3990#if CONFIG_VP9_HIGHBITDEPTH 3991 int use_highbitdepth, 3992#endif 3993 int subsampling_x, int subsampling_y) { 3994 VP9_COMMON *const cm = &cpi->common; 3995 3996 if (!cpi->initial_width || 3997#if CONFIG_VP9_HIGHBITDEPTH 3998 cm->use_highbitdepth != use_highbitdepth || 3999#endif 4000 cm->subsampling_x != subsampling_x || 4001 cm->subsampling_y != subsampling_y) { 4002 cm->subsampling_x = subsampling_x; 4003 cm->subsampling_y = subsampling_y; 4004#if CONFIG_VP9_HIGHBITDEPTH 4005 cm->use_highbitdepth = use_highbitdepth; 4006#endif 4007 4008 alloc_raw_frame_buffers(cpi); 4009 init_ref_frame_bufs(cm); 4010 alloc_util_frame_buffers(cpi); 4011 4012 init_motion_estimation(cpi); // TODO(agrange) This can be removed. 4013 4014 cpi->initial_width = cm->width; 4015 cpi->initial_height = cm->height; 4016 cpi->initial_mbs = cm->MBs; 4017 } 4018} 4019 4020int vp9_receive_raw_frame(VP9_COMP *cpi, unsigned int frame_flags, 4021 YV12_BUFFER_CONFIG *sd, int64_t time_stamp, 4022 int64_t end_time) { 4023 VP9_COMMON *cm = &cpi->common; 4024 struct vpx_usec_timer timer; 4025 int res = 0; 4026 const int subsampling_x = sd->subsampling_x; 4027 const int subsampling_y = sd->subsampling_y; 4028#if CONFIG_VP9_HIGHBITDEPTH 4029 const int use_highbitdepth = sd->flags & YV12_FLAG_HIGHBITDEPTH; 4030 check_initial_width(cpi, use_highbitdepth, subsampling_x, subsampling_y); 4031#else 4032 check_initial_width(cpi, subsampling_x, subsampling_y); 4033#endif // CONFIG_VP9_HIGHBITDEPTH 4034 4035#if CONFIG_VP9_TEMPORAL_DENOISING 4036 setup_denoiser_buffer(cpi); 4037#endif 4038 vpx_usec_timer_start(&timer); 4039 4040 if (vp9_lookahead_push(cpi->lookahead, sd, time_stamp, end_time, 4041#if CONFIG_VP9_HIGHBITDEPTH 4042 use_highbitdepth, 4043#endif // CONFIG_VP9_HIGHBITDEPTH 4044 frame_flags)) 4045 res = -1; 4046 vpx_usec_timer_mark(&timer); 4047 cpi->time_receive_data += vpx_usec_timer_elapsed(&timer); 4048 4049 if ((cm->profile == PROFILE_0 || cm->profile == PROFILE_2) && 4050 (subsampling_x != 1 || subsampling_y != 1)) { 4051 vpx_internal_error(&cm->error, VPX_CODEC_INVALID_PARAM, 4052 "Non-4:2:0 color format requires profile 1 or 3"); 4053 res = -1; 4054 } 4055 if ((cm->profile == PROFILE_1 || cm->profile == PROFILE_3) && 4056 (subsampling_x == 1 && subsampling_y == 1)) { 4057 vpx_internal_error(&cm->error, VPX_CODEC_INVALID_PARAM, 4058 "4:2:0 color format requires profile 0 or 2"); 4059 res = -1; 4060 } 4061 4062 return res; 4063} 4064 4065 4066static int frame_is_reference(const VP9_COMP *cpi) { 4067 const VP9_COMMON *cm = &cpi->common; 4068 4069 return cm->frame_type == KEY_FRAME || 4070 cpi->refresh_last_frame || 4071 cpi->refresh_golden_frame || 4072 cpi->refresh_alt_ref_frame || 4073 cm->refresh_frame_context || 4074 cm->lf.mode_ref_delta_update || 4075 cm->seg.update_map || 4076 cm->seg.update_data; 4077} 4078 4079static void adjust_frame_rate(VP9_COMP *cpi, 4080 const struct lookahead_entry *source) { 4081 int64_t this_duration; 4082 int step = 0; 4083 4084 if (source->ts_start == cpi->first_time_stamp_ever) { 4085 this_duration = source->ts_end - source->ts_start; 4086 step = 1; 4087 } else { 4088 int64_t last_duration = cpi->last_end_time_stamp_seen 4089 - cpi->last_time_stamp_seen; 4090 4091 this_duration = source->ts_end - cpi->last_end_time_stamp_seen; 4092 4093 // do a step update if the duration changes by 10% 4094 if (last_duration) 4095 step = (int)((this_duration - last_duration) * 10 / last_duration); 4096 } 4097 4098 if (this_duration) { 4099 if (step) { 4100 vp9_new_framerate(cpi, 10000000.0 / this_duration); 4101 } else { 4102 // Average this frame's rate into the last second's average 4103 // frame rate. If we haven't seen 1 second yet, then average 4104 // over the whole interval seen. 4105 const double interval = VPXMIN( 4106 (double)(source->ts_end - cpi->first_time_stamp_ever), 10000000.0); 4107 double avg_duration = 10000000.0 / cpi->framerate; 4108 avg_duration *= (interval - avg_duration + this_duration); 4109 avg_duration /= interval; 4110 4111 vp9_new_framerate(cpi, 10000000.0 / avg_duration); 4112 } 4113 } 4114 cpi->last_time_stamp_seen = source->ts_start; 4115 cpi->last_end_time_stamp_seen = source->ts_end; 4116} 4117 4118// Returns 0 if this is not an alt ref else the offset of the source frame 4119// used as the arf midpoint. 4120static int get_arf_src_index(VP9_COMP *cpi) { 4121 RATE_CONTROL *const rc = &cpi->rc; 4122 int arf_src_index = 0; 4123 if (is_altref_enabled(cpi)) { 4124 if (cpi->oxcf.pass == 2) { 4125 const GF_GROUP *const gf_group = &cpi->twopass.gf_group; 4126 if (gf_group->update_type[gf_group->index] == ARF_UPDATE) { 4127 arf_src_index = gf_group->arf_src_offset[gf_group->index]; 4128 } 4129 } else if (rc->source_alt_ref_pending) { 4130 arf_src_index = rc->frames_till_gf_update_due; 4131 } 4132 } 4133 return arf_src_index; 4134} 4135 4136static void check_src_altref(VP9_COMP *cpi, 4137 const struct lookahead_entry *source) { 4138 RATE_CONTROL *const rc = &cpi->rc; 4139 4140 if (cpi->oxcf.pass == 2) { 4141 const GF_GROUP *const gf_group = &cpi->twopass.gf_group; 4142 rc->is_src_frame_alt_ref = 4143 (gf_group->update_type[gf_group->index] == OVERLAY_UPDATE); 4144 } else { 4145 rc->is_src_frame_alt_ref = cpi->alt_ref_source && 4146 (source == cpi->alt_ref_source); 4147 } 4148 4149 if (rc->is_src_frame_alt_ref) { 4150 // Current frame is an ARF overlay frame. 4151 cpi->alt_ref_source = NULL; 4152 4153 // Don't refresh the last buffer for an ARF overlay frame. It will 4154 // become the GF so preserve last as an alternative prediction option. 4155 cpi->refresh_last_frame = 0; 4156 } 4157} 4158 4159#if CONFIG_INTERNAL_STATS 4160extern double vp9_get_blockiness(const uint8_t *img1, int img1_pitch, 4161 const uint8_t *img2, int img2_pitch, 4162 int width, int height); 4163 4164static void adjust_image_stat(double y, double u, double v, double all, 4165 ImageStat *s) { 4166 s->stat[Y] += y; 4167 s->stat[U] += u; 4168 s->stat[V] += v; 4169 s->stat[ALL] += all; 4170 s->worst = VPXMIN(s->worst, all); 4171} 4172#endif // CONFIG_INTERNAL_STATS 4173 4174int vp9_get_compressed_data(VP9_COMP *cpi, unsigned int *frame_flags, 4175 size_t *size, uint8_t *dest, 4176 int64_t *time_stamp, int64_t *time_end, int flush) { 4177 const VP9EncoderConfig *const oxcf = &cpi->oxcf; 4178 VP9_COMMON *const cm = &cpi->common; 4179 BufferPool *const pool = cm->buffer_pool; 4180 RATE_CONTROL *const rc = &cpi->rc; 4181 struct vpx_usec_timer cmptimer; 4182 YV12_BUFFER_CONFIG *force_src_buffer = NULL; 4183 struct lookahead_entry *last_source = NULL; 4184 struct lookahead_entry *source = NULL; 4185 int arf_src_index; 4186 int i; 4187 4188 if (is_two_pass_svc(cpi)) { 4189#if CONFIG_SPATIAL_SVC 4190 vp9_svc_start_frame(cpi); 4191 // Use a small empty frame instead of a real frame 4192 if (cpi->svc.encode_empty_frame_state == ENCODING) 4193 source = &cpi->svc.empty_frame; 4194#endif 4195 if (oxcf->pass == 2) 4196 vp9_restore_layer_context(cpi); 4197 } else if (is_one_pass_cbr_svc(cpi)) { 4198 vp9_one_pass_cbr_svc_start_layer(cpi); 4199 } 4200 4201 vpx_usec_timer_start(&cmptimer); 4202 4203 vp9_set_high_precision_mv(cpi, ALTREF_HIGH_PRECISION_MV); 4204 4205 // Is multi-arf enabled. 4206 // Note that at the moment multi_arf is only configured for 2 pass VBR and 4207 // will not work properly with svc. 4208 if ((oxcf->pass == 2) && !cpi->use_svc && 4209 (cpi->oxcf.enable_auto_arf > 1)) 4210 cpi->multi_arf_allowed = 1; 4211 else 4212 cpi->multi_arf_allowed = 0; 4213 4214 // Normal defaults 4215 cm->reset_frame_context = 0; 4216 cm->refresh_frame_context = 1; 4217 if (!is_one_pass_cbr_svc(cpi)) { 4218 cpi->refresh_last_frame = 1; 4219 cpi->refresh_golden_frame = 0; 4220 cpi->refresh_alt_ref_frame = 0; 4221 } 4222 4223 // Should we encode an arf frame. 4224 arf_src_index = get_arf_src_index(cpi); 4225 4226 // Skip alt frame if we encode the empty frame 4227 if (is_two_pass_svc(cpi) && source != NULL) 4228 arf_src_index = 0; 4229 4230 if (arf_src_index) { 4231 assert(arf_src_index <= rc->frames_to_key); 4232 4233 if ((source = vp9_lookahead_peek(cpi->lookahead, arf_src_index)) != NULL) { 4234 cpi->alt_ref_source = source; 4235 4236#if CONFIG_SPATIAL_SVC 4237 if (is_two_pass_svc(cpi) && cpi->svc.spatial_layer_id > 0) { 4238 int i; 4239 // Reference a hidden frame from a lower layer 4240 for (i = cpi->svc.spatial_layer_id - 1; i >= 0; --i) { 4241 if (oxcf->ss_enable_auto_arf[i]) { 4242 cpi->gld_fb_idx = cpi->svc.layer_context[i].alt_ref_idx; 4243 break; 4244 } 4245 } 4246 } 4247 cpi->svc.layer_context[cpi->svc.spatial_layer_id].has_alt_frame = 1; 4248#endif 4249 4250 if (oxcf->arnr_max_frames > 0) { 4251 // Produce the filtered ARF frame. 4252 vp9_temporal_filter(cpi, arf_src_index); 4253 vpx_extend_frame_borders(&cpi->alt_ref_buffer); 4254 force_src_buffer = &cpi->alt_ref_buffer; 4255 } 4256 4257 cm->show_frame = 0; 4258 cm->intra_only = 0; 4259 cpi->refresh_alt_ref_frame = 1; 4260 cpi->refresh_golden_frame = 0; 4261 cpi->refresh_last_frame = 0; 4262 rc->is_src_frame_alt_ref = 0; 4263 rc->source_alt_ref_pending = 0; 4264 } else { 4265 rc->source_alt_ref_pending = 0; 4266 } 4267 } 4268 4269 if (!source) { 4270 // Get last frame source. 4271 if (cm->current_video_frame > 0) { 4272 if ((last_source = vp9_lookahead_peek(cpi->lookahead, -1)) == NULL) 4273 return -1; 4274 } 4275 4276 // Read in the source frame. 4277 if (cpi->use_svc) 4278 source = vp9_svc_lookahead_pop(cpi, cpi->lookahead, flush); 4279 else 4280 source = vp9_lookahead_pop(cpi->lookahead, flush); 4281 4282 if (source != NULL) { 4283 cm->show_frame = 1; 4284 cm->intra_only = 0; 4285 // if the flags indicate intra frame, but if the current picture is for 4286 // non-zero spatial layer, it should not be an intra picture. 4287 // TODO(Won Kap): this needs to change if per-layer intra frame is 4288 // allowed. 4289 if ((source->flags & VPX_EFLAG_FORCE_KF) && 4290 cpi->svc.spatial_layer_id > cpi->svc.first_spatial_layer_to_encode) { 4291 source->flags &= ~(unsigned int)(VPX_EFLAG_FORCE_KF); 4292 } 4293 4294 // Check to see if the frame should be encoded as an arf overlay. 4295 check_src_altref(cpi, source); 4296 } 4297 } 4298 4299 if (source) { 4300 cpi->un_scaled_source = cpi->Source = force_src_buffer ? force_src_buffer 4301 : &source->img; 4302 4303 cpi->unscaled_last_source = last_source != NULL ? &last_source->img : NULL; 4304 4305 *time_stamp = source->ts_start; 4306 *time_end = source->ts_end; 4307 *frame_flags = (source->flags & VPX_EFLAG_FORCE_KF) ? FRAMEFLAGS_KEY : 0; 4308 4309 } else { 4310 *size = 0; 4311 if (flush && oxcf->pass == 1 && !cpi->twopass.first_pass_done) { 4312 vp9_end_first_pass(cpi); /* get last stats packet */ 4313 cpi->twopass.first_pass_done = 1; 4314 } 4315 return -1; 4316 } 4317 4318 if (source->ts_start < cpi->first_time_stamp_ever) { 4319 cpi->first_time_stamp_ever = source->ts_start; 4320 cpi->last_end_time_stamp_seen = source->ts_start; 4321 } 4322 4323 // Clear down mmx registers 4324 vpx_clear_system_state(); 4325 4326 // adjust frame rates based on timestamps given 4327 if (cm->show_frame) { 4328 adjust_frame_rate(cpi, source); 4329 } 4330 4331 if (is_one_pass_cbr_svc(cpi)) { 4332 vp9_update_temporal_layer_framerate(cpi); 4333 vp9_restore_layer_context(cpi); 4334 } 4335 4336 // Find a free buffer for the new frame, releasing the reference previously 4337 // held. 4338 if (cm->new_fb_idx != INVALID_IDX) { 4339 --pool->frame_bufs[cm->new_fb_idx].ref_count; 4340 } 4341 cm->new_fb_idx = get_free_fb(cm); 4342 4343 if (cm->new_fb_idx == INVALID_IDX) 4344 return -1; 4345 4346 cm->cur_frame = &pool->frame_bufs[cm->new_fb_idx]; 4347 4348 if (!cpi->use_svc && cpi->multi_arf_allowed) { 4349 if (cm->frame_type == KEY_FRAME) { 4350 init_buffer_indices(cpi); 4351 } else if (oxcf->pass == 2) { 4352 const GF_GROUP *const gf_group = &cpi->twopass.gf_group; 4353 cpi->alt_fb_idx = gf_group->arf_ref_idx[gf_group->index]; 4354 } 4355 } 4356 4357 // Start with a 0 size frame. 4358 *size = 0; 4359 4360 cpi->frame_flags = *frame_flags; 4361 4362 if ((oxcf->pass == 2) && 4363 (!cpi->use_svc || 4364 (is_two_pass_svc(cpi) && 4365 cpi->svc.encode_empty_frame_state != ENCODING))) { 4366 vp9_rc_get_second_pass_params(cpi); 4367 } else if (oxcf->pass == 1) { 4368 set_frame_size(cpi); 4369 } 4370 4371 if (cpi->oxcf.pass != 0 || 4372 cpi->use_svc || 4373 frame_is_intra_only(cm) == 1) { 4374 for (i = 0; i < MAX_REF_FRAMES; ++i) 4375 cpi->scaled_ref_idx[i] = INVALID_IDX; 4376 } 4377 4378 if (oxcf->pass == 1 && 4379 (!cpi->use_svc || is_two_pass_svc(cpi))) { 4380 const int lossless = is_lossless_requested(oxcf); 4381#if CONFIG_VP9_HIGHBITDEPTH 4382 if (cpi->oxcf.use_highbitdepth) 4383 cpi->td.mb.fwd_txm4x4 = lossless ? 4384 vp9_highbd_fwht4x4 : vpx_highbd_fdct4x4; 4385 else 4386 cpi->td.mb.fwd_txm4x4 = lossless ? vp9_fwht4x4 : vpx_fdct4x4; 4387 cpi->td.mb.highbd_itxm_add = lossless ? vp9_highbd_iwht4x4_add : 4388 vp9_highbd_idct4x4_add; 4389#else 4390 cpi->td.mb.fwd_txm4x4 = lossless ? vp9_fwht4x4 : vpx_fdct4x4; 4391#endif // CONFIG_VP9_HIGHBITDEPTH 4392 cpi->td.mb.itxm_add = lossless ? vp9_iwht4x4_add : vp9_idct4x4_add; 4393 vp9_first_pass(cpi, source); 4394 } else if (oxcf->pass == 2 && 4395 (!cpi->use_svc || is_two_pass_svc(cpi))) { 4396 Pass2Encode(cpi, size, dest, frame_flags); 4397 } else if (cpi->use_svc) { 4398 SvcEncode(cpi, size, dest, frame_flags); 4399 } else { 4400 // One pass encode 4401 Pass0Encode(cpi, size, dest, frame_flags); 4402 } 4403 4404 if (cm->refresh_frame_context) 4405 cm->frame_contexts[cm->frame_context_idx] = *cm->fc; 4406 4407 // No frame encoded, or frame was dropped, release scaled references. 4408 if ((*size == 0) && (frame_is_intra_only(cm) == 0)) { 4409 release_scaled_references(cpi); 4410 } 4411 4412 if (*size > 0) { 4413 cpi->droppable = !frame_is_reference(cpi); 4414 } 4415 4416 // Save layer specific state. 4417 if (is_one_pass_cbr_svc(cpi) || 4418 ((cpi->svc.number_temporal_layers > 1 || 4419 cpi->svc.number_spatial_layers > 1) && 4420 oxcf->pass == 2)) { 4421 vp9_save_layer_context(cpi); 4422 } 4423 4424 vpx_usec_timer_mark(&cmptimer); 4425 cpi->time_compress_data += vpx_usec_timer_elapsed(&cmptimer); 4426 4427 if (cpi->b_calculate_psnr && oxcf->pass != 1 && cm->show_frame) 4428 generate_psnr_packet(cpi); 4429 4430#if CONFIG_INTERNAL_STATS 4431 4432 if (oxcf->pass != 1) { 4433 double samples = 0.0; 4434 cpi->bytes += (int)(*size); 4435 4436 if (cm->show_frame) { 4437 cpi->count++; 4438 4439 if (cpi->b_calculate_psnr) { 4440 YV12_BUFFER_CONFIG *orig = cpi->Source; 4441 YV12_BUFFER_CONFIG *recon = cpi->common.frame_to_show; 4442 YV12_BUFFER_CONFIG *pp = &cm->post_proc_buffer; 4443 PSNR_STATS psnr; 4444#if CONFIG_VP9_HIGHBITDEPTH 4445 calc_highbd_psnr(orig, recon, &psnr, cpi->td.mb.e_mbd.bd, 4446 cpi->oxcf.input_bit_depth); 4447#else 4448 calc_psnr(orig, recon, &psnr); 4449#endif // CONFIG_VP9_HIGHBITDEPTH 4450 4451 adjust_image_stat(psnr.psnr[1], psnr.psnr[2], psnr.psnr[3], 4452 psnr.psnr[0], &cpi->psnr); 4453 cpi->total_sq_error += psnr.sse[0]; 4454 cpi->total_samples += psnr.samples[0]; 4455 samples = psnr.samples[0]; 4456 4457 { 4458 PSNR_STATS psnr2; 4459 double frame_ssim2 = 0, weight = 0; 4460#if CONFIG_VP9_POSTPROC 4461 if (vpx_alloc_frame_buffer(&cm->post_proc_buffer, 4462 recon->y_crop_width, recon->y_crop_height, 4463 cm->subsampling_x, cm->subsampling_y, 4464#if CONFIG_VP9_HIGHBITDEPTH 4465 cm->use_highbitdepth, 4466#endif 4467 VP9_ENC_BORDER_IN_PIXELS, 4468 cm->byte_alignment) < 0) { 4469 vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR, 4470 "Failed to allocate post processing buffer"); 4471 } 4472 4473 vp9_deblock(cm->frame_to_show, &cm->post_proc_buffer, 4474 cm->lf.filter_level * 10 / 6); 4475#endif 4476 vpx_clear_system_state(); 4477 4478#if CONFIG_VP9_HIGHBITDEPTH 4479 calc_highbd_psnr(orig, pp, &psnr2, cpi->td.mb.e_mbd.bd, 4480 cpi->oxcf.input_bit_depth); 4481#else 4482 calc_psnr(orig, pp, &psnr2); 4483#endif // CONFIG_VP9_HIGHBITDEPTH 4484 4485 cpi->totalp_sq_error += psnr2.sse[0]; 4486 cpi->totalp_samples += psnr2.samples[0]; 4487 adjust_image_stat(psnr2.psnr[1], psnr2.psnr[2], psnr2.psnr[3], 4488 psnr2.psnr[0], &cpi->psnrp); 4489 4490#if CONFIG_VP9_HIGHBITDEPTH 4491 if (cm->use_highbitdepth) { 4492 frame_ssim2 = vpx_highbd_calc_ssim(orig, recon, &weight, 4493 (int)cm->bit_depth); 4494 } else { 4495 frame_ssim2 = vpx_calc_ssim(orig, recon, &weight); 4496 } 4497#else 4498 frame_ssim2 = vpx_calc_ssim(orig, recon, &weight); 4499#endif // CONFIG_VP9_HIGHBITDEPTH 4500 4501 cpi->worst_ssim = VPXMIN(cpi->worst_ssim, frame_ssim2); 4502 cpi->summed_quality += frame_ssim2 * weight; 4503 cpi->summed_weights += weight; 4504 4505#if CONFIG_VP9_HIGHBITDEPTH 4506 if (cm->use_highbitdepth) { 4507 frame_ssim2 = vpx_highbd_calc_ssim( 4508 orig, &cm->post_proc_buffer, &weight, (int)cm->bit_depth); 4509 } else { 4510 frame_ssim2 = vpx_calc_ssim(orig, &cm->post_proc_buffer, &weight); 4511 } 4512#else 4513 frame_ssim2 = vpx_calc_ssim(orig, &cm->post_proc_buffer, &weight); 4514#endif // CONFIG_VP9_HIGHBITDEPTH 4515 4516 cpi->summedp_quality += frame_ssim2 * weight; 4517 cpi->summedp_weights += weight; 4518#if 0 4519 { 4520 FILE *f = fopen("q_used.stt", "a"); 4521 fprintf(f, "%5d : Y%f7.3:U%f7.3:V%f7.3:F%f7.3:S%7.3f\n", 4522 cpi->common.current_video_frame, y2, u2, v2, 4523 frame_psnr2, frame_ssim2); 4524 fclose(f); 4525 } 4526#endif 4527 } 4528 } 4529 if (cpi->b_calculate_blockiness) { 4530#if CONFIG_VP9_HIGHBITDEPTH 4531 if (!cm->use_highbitdepth) 4532#endif 4533 { 4534 double frame_blockiness = vp9_get_blockiness( 4535 cpi->Source->y_buffer, cpi->Source->y_stride, 4536 cm->frame_to_show->y_buffer, cm->frame_to_show->y_stride, 4537 cpi->Source->y_width, cpi->Source->y_height); 4538 cpi->worst_blockiness = 4539 VPXMAX(cpi->worst_blockiness, frame_blockiness); 4540 cpi->total_blockiness += frame_blockiness; 4541 } 4542 } 4543 4544 if (cpi->b_calculate_consistency) { 4545#if CONFIG_VP9_HIGHBITDEPTH 4546 if (!cm->use_highbitdepth) 4547#endif 4548 { 4549 double this_inconsistency = vpx_get_ssim_metrics( 4550 cpi->Source->y_buffer, cpi->Source->y_stride, 4551 cm->frame_to_show->y_buffer, cm->frame_to_show->y_stride, 4552 cpi->Source->y_width, cpi->Source->y_height, cpi->ssim_vars, 4553 &cpi->metrics, 1); 4554 4555 const double peak = (double)((1 << cpi->oxcf.input_bit_depth) - 1); 4556 double consistency = vpx_sse_to_psnr(samples, peak, 4557 (double)cpi->total_inconsistency); 4558 if (consistency > 0.0) 4559 cpi->worst_consistency = 4560 VPXMIN(cpi->worst_consistency, consistency); 4561 cpi->total_inconsistency += this_inconsistency; 4562 } 4563 } 4564 4565 if (cpi->b_calculate_ssimg) { 4566 double y, u, v, frame_all; 4567#if CONFIG_VP9_HIGHBITDEPTH 4568 if (cm->use_highbitdepth) { 4569 frame_all = vpx_highbd_calc_ssimg(cpi->Source, cm->frame_to_show, &y, 4570 &u, &v, (int)cm->bit_depth); 4571 } else { 4572 frame_all = vpx_calc_ssimg(cpi->Source, cm->frame_to_show, &y, &u, 4573 &v); 4574 } 4575#else 4576 frame_all = vpx_calc_ssimg(cpi->Source, cm->frame_to_show, &y, &u, &v); 4577#endif // CONFIG_VP9_HIGHBITDEPTH 4578 adjust_image_stat(y, u, v, frame_all, &cpi->ssimg); 4579 } 4580#if CONFIG_VP9_HIGHBITDEPTH 4581 if (!cm->use_highbitdepth) 4582#endif 4583 { 4584 double y, u, v, frame_all; 4585 frame_all = vpx_calc_fastssim(cpi->Source, cm->frame_to_show, &y, &u, 4586 &v); 4587 adjust_image_stat(y, u, v, frame_all, &cpi->fastssim); 4588 /* TODO(JBB): add 10/12 bit support */ 4589 } 4590#if CONFIG_VP9_HIGHBITDEPTH 4591 if (!cm->use_highbitdepth) 4592#endif 4593 { 4594 double y, u, v, frame_all; 4595 frame_all = vpx_psnrhvs(cpi->Source, cm->frame_to_show, &y, &u, &v); 4596 adjust_image_stat(y, u, v, frame_all, &cpi->psnrhvs); 4597 } 4598 } 4599 } 4600 4601#endif 4602 4603 if (is_two_pass_svc(cpi)) { 4604 if (cpi->svc.encode_empty_frame_state == ENCODING) { 4605 cpi->svc.encode_empty_frame_state = ENCODED; 4606 cpi->svc.encode_intra_empty_frame = 0; 4607 } 4608 4609 if (cm->show_frame) { 4610 ++cpi->svc.spatial_layer_to_encode; 4611 if (cpi->svc.spatial_layer_to_encode >= cpi->svc.number_spatial_layers) 4612 cpi->svc.spatial_layer_to_encode = 0; 4613 4614 // May need the empty frame after an visible frame. 4615 cpi->svc.encode_empty_frame_state = NEED_TO_ENCODE; 4616 } 4617 } else if (is_one_pass_cbr_svc(cpi)) { 4618 if (cm->show_frame) { 4619 ++cpi->svc.spatial_layer_to_encode; 4620 if (cpi->svc.spatial_layer_to_encode >= cpi->svc.number_spatial_layers) 4621 cpi->svc.spatial_layer_to_encode = 0; 4622 } 4623 } 4624 vpx_clear_system_state(); 4625 return 0; 4626} 4627 4628int vp9_get_preview_raw_frame(VP9_COMP *cpi, YV12_BUFFER_CONFIG *dest, 4629 vp9_ppflags_t *flags) { 4630 VP9_COMMON *cm = &cpi->common; 4631#if !CONFIG_VP9_POSTPROC 4632 (void)flags; 4633#endif 4634 4635 if (!cm->show_frame) { 4636 return -1; 4637 } else { 4638 int ret; 4639#if CONFIG_VP9_POSTPROC 4640 ret = vp9_post_proc_frame(cm, dest, flags); 4641#else 4642 if (cm->frame_to_show) { 4643 *dest = *cm->frame_to_show; 4644 dest->y_width = cm->width; 4645 dest->y_height = cm->height; 4646 dest->uv_width = cm->width >> cm->subsampling_x; 4647 dest->uv_height = cm->height >> cm->subsampling_y; 4648 ret = 0; 4649 } else { 4650 ret = -1; 4651 } 4652#endif // !CONFIG_VP9_POSTPROC 4653 vpx_clear_system_state(); 4654 return ret; 4655 } 4656} 4657 4658int vp9_set_internal_size(VP9_COMP *cpi, 4659 VPX_SCALING horiz_mode, VPX_SCALING vert_mode) { 4660 VP9_COMMON *cm = &cpi->common; 4661 int hr = 0, hs = 0, vr = 0, vs = 0; 4662 4663 if (horiz_mode > ONETWO || vert_mode > ONETWO) 4664 return -1; 4665 4666 Scale2Ratio(horiz_mode, &hr, &hs); 4667 Scale2Ratio(vert_mode, &vr, &vs); 4668 4669 // always go to the next whole number 4670 cm->width = (hs - 1 + cpi->oxcf.width * hr) / hs; 4671 cm->height = (vs - 1 + cpi->oxcf.height * vr) / vs; 4672 if (cm->current_video_frame) { 4673 assert(cm->width <= cpi->initial_width); 4674 assert(cm->height <= cpi->initial_height); 4675 } 4676 4677 update_frame_size(cpi); 4678 4679 return 0; 4680} 4681 4682int vp9_set_size_literal(VP9_COMP *cpi, unsigned int width, 4683 unsigned int height) { 4684 VP9_COMMON *cm = &cpi->common; 4685#if CONFIG_VP9_HIGHBITDEPTH 4686 check_initial_width(cpi, cm->use_highbitdepth, 1, 1); 4687#else 4688 check_initial_width(cpi, 1, 1); 4689#endif // CONFIG_VP9_HIGHBITDEPTH 4690 4691#if CONFIG_VP9_TEMPORAL_DENOISING 4692 setup_denoiser_buffer(cpi); 4693#endif 4694 4695 if (width) { 4696 cm->width = width; 4697 if (cm->width > cpi->initial_width) { 4698 cm->width = cpi->initial_width; 4699 printf("Warning: Desired width too large, changed to %d\n", cm->width); 4700 } 4701 } 4702 4703 if (height) { 4704 cm->height = height; 4705 if (cm->height > cpi->initial_height) { 4706 cm->height = cpi->initial_height; 4707 printf("Warning: Desired height too large, changed to %d\n", cm->height); 4708 } 4709 } 4710 assert(cm->width <= cpi->initial_width); 4711 assert(cm->height <= cpi->initial_height); 4712 4713 update_frame_size(cpi); 4714 4715 return 0; 4716} 4717 4718void vp9_set_svc(VP9_COMP *cpi, int use_svc) { 4719 cpi->use_svc = use_svc; 4720 return; 4721} 4722 4723int64_t vp9_get_y_sse(const YV12_BUFFER_CONFIG *a, 4724 const YV12_BUFFER_CONFIG *b) { 4725 assert(a->y_crop_width == b->y_crop_width); 4726 assert(a->y_crop_height == b->y_crop_height); 4727 4728 return get_sse(a->y_buffer, a->y_stride, b->y_buffer, b->y_stride, 4729 a->y_crop_width, a->y_crop_height); 4730} 4731 4732#if CONFIG_VP9_HIGHBITDEPTH 4733int64_t vp9_highbd_get_y_sse(const YV12_BUFFER_CONFIG *a, 4734 const YV12_BUFFER_CONFIG *b) { 4735 assert(a->y_crop_width == b->y_crop_width); 4736 assert(a->y_crop_height == b->y_crop_height); 4737 assert((a->flags & YV12_FLAG_HIGHBITDEPTH) != 0); 4738 assert((b->flags & YV12_FLAG_HIGHBITDEPTH) != 0); 4739 4740 return highbd_get_sse(a->y_buffer, a->y_stride, b->y_buffer, b->y_stride, 4741 a->y_crop_width, a->y_crop_height); 4742} 4743#endif // CONFIG_VP9_HIGHBITDEPTH 4744 4745int vp9_get_quantizer(VP9_COMP *cpi) { 4746 return cpi->common.base_qindex; 4747} 4748 4749void vp9_apply_encoding_flags(VP9_COMP *cpi, vpx_enc_frame_flags_t flags) { 4750 if (flags & (VP8_EFLAG_NO_REF_LAST | VP8_EFLAG_NO_REF_GF | 4751 VP8_EFLAG_NO_REF_ARF)) { 4752 int ref = 7; 4753 4754 if (flags & VP8_EFLAG_NO_REF_LAST) 4755 ref ^= VP9_LAST_FLAG; 4756 4757 if (flags & VP8_EFLAG_NO_REF_GF) 4758 ref ^= VP9_GOLD_FLAG; 4759 4760 if (flags & VP8_EFLAG_NO_REF_ARF) 4761 ref ^= VP9_ALT_FLAG; 4762 4763 vp9_use_as_reference(cpi, ref); 4764 } 4765 4766 if (flags & (VP8_EFLAG_NO_UPD_LAST | VP8_EFLAG_NO_UPD_GF | 4767 VP8_EFLAG_NO_UPD_ARF | VP8_EFLAG_FORCE_GF | 4768 VP8_EFLAG_FORCE_ARF)) { 4769 int upd = 7; 4770 4771 if (flags & VP8_EFLAG_NO_UPD_LAST) 4772 upd ^= VP9_LAST_FLAG; 4773 4774 if (flags & VP8_EFLAG_NO_UPD_GF) 4775 upd ^= VP9_GOLD_FLAG; 4776 4777 if (flags & VP8_EFLAG_NO_UPD_ARF) 4778 upd ^= VP9_ALT_FLAG; 4779 4780 vp9_update_reference(cpi, upd); 4781 } 4782 4783 if (flags & VP8_EFLAG_NO_UPD_ENTROPY) { 4784 vp9_update_entropy(cpi, 0); 4785 } 4786} 4787