granicus.if.org Git - libvpx/blob - vp9/encoder/vp9_encoder.c

   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_dsp/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_360.h"
  40 #include "vp9/encoder/vp9_aq_complexity.h"
  41 #include "vp9/encoder/vp9_aq_cyclicrefresh.h"
  42 #include "vp9/encoder/vp9_aq_variance.h"
  43 #include "vp9/encoder/vp9_bitstream.h"
  44 #include "vp9/encoder/vp9_context_tree.h"
  45 #include "vp9/encoder/vp9_encodeframe.h"
  46 #include "vp9/encoder/vp9_encodemv.h"
  47 #include "vp9/encoder/vp9_encoder.h"
  48 #include "vp9/encoder/vp9_extend.h"
  49 #include "vp9/encoder/vp9_ethread.h"
  50 #include "vp9/encoder/vp9_firstpass.h"
  51 #include "vp9/encoder/vp9_mbgraph.h"
  52 #include "vp9/encoder/vp9_noise_estimate.h"
  53 #include "vp9/encoder/vp9_picklpf.h"
  54 #include "vp9/encoder/vp9_ratectrl.h"
  55 #include "vp9/encoder/vp9_rd.h"
  56 #include "vp9/encoder/vp9_resize.h"
  57 #include "vp9/encoder/vp9_segmentation.h"
  58 #include "vp9/encoder/vp9_skin_detection.h"
  59 #include "vp9/encoder/vp9_speed_features.h"
  60 #include "vp9/encoder/vp9_svc_layercontext.h"
  61 #include "vp9/encoder/vp9_temporal_filter.h"
  62
  63 #define AM_SEGMENT_ID_INACTIVE 7
  64 #define AM_SEGMENT_ID_ACTIVE 0
  65
  66 #define ALTREF_HIGH_PRECISION_MV 1     // Whether to use high precision mv
  67                                        //  for altref computation.
  68 #define HIGH_PRECISION_MV_QTHRESH 200  // Q threshold for high precision
  69                                        // mv. Choose a very high value for
  70                                        // now so that HIGH_PRECISION is always
  71                                        // chosen.
  72 // #define OUTPUT_YUV_REC
  73
  74 #ifdef OUTPUT_YUV_DENOISED
  75 FILE *yuv_denoised_file = NULL;
  76 #endif
  77 #ifdef OUTPUT_YUV_SKINMAP
  78 FILE *yuv_skinmap_file = NULL;
  79 #endif
  80 #ifdef OUTPUT_YUV_REC
  81 FILE *yuv_rec_file;
  82 #endif
  83
  84 #if 0
  85 FILE *framepsnr;
  86 FILE *kf_list;
  87 FILE *keyfile;
  88 #endif
  89
  90 #ifdef ENABLE_KF_DENOISE
  91 // Test condition for spatial denoise of source.
  92 static int is_spatial_denoise_enabled(VP9_COMP *cpi) {
  93   VP9_COMMON *const cm = &cpi->common;
  94   const VP9EncoderConfig *const oxcf = &cpi->oxcf;
  95
  96   return (oxcf->pass != 1) && !is_lossless_requested(&cpi->oxcf) &&
  97          frame_is_intra_only(cm);
  98 }
  99 #endif
 100
 101 // Test for whether to calculate metrics for the frame.
 102 static int is_psnr_calc_enabled(VP9_COMP *cpi) {
 103   VP9_COMMON *const cm = &cpi->common;
 104   const VP9EncoderConfig *const oxcf = &cpi->oxcf;
 105
 106   return cpi->b_calculate_psnr && (oxcf->pass != 1) && cm->show_frame;
 107 }
 108
 109 /* clang-format off */
 110 static const Vp9LevelSpec vp9_level_defs[VP9_LEVELS] = {
 111   { LEVEL_1,   829440,      36864,    200,    400,   2, 1,  4,  8 },
 112   { LEVEL_1_1, 2764800,     73728,    800,    1000,  2, 1,  4,  8 },
 113   { LEVEL_2,   4608000,     122880,   1800,   1500,  2, 1,  4,  8 },
 114   { LEVEL_2_1, 9216000,     245760,   3600,   2800,  2, 2,  4,  8 },
 115   { LEVEL_3,   20736000,    552960,   7200,   6000,  2, 4,  4,  8 },
 116   { LEVEL_3_1, 36864000,    983040,   12000,  10000, 2, 4,  4,  8 },
 117   { LEVEL_4,   83558400,    2228224,  18000,  16000, 4, 4,  4,  8 },
 118   { LEVEL_4_1, 160432128,   2228224,  30000,  18000, 4, 4,  5,  6 },
 119   { LEVEL_5,   311951360,   8912896,  60000,  36000, 6, 8,  6,  4 },
 120   { LEVEL_5_1, 588251136,   8912896,  120000, 46000, 8, 8,  10, 4 },
 121   // TODO(huisu): update max_cpb_size for level 5_2 ~ 6_2 when
 122   // they are finalized (currently TBD).
 123   { LEVEL_5_2, 1176502272,  8912896,  180000, 0,     8, 8,  10, 4 },
 124   { LEVEL_6,   1176502272,  35651584, 180000, 0,     8, 16, 10, 4 },
 125   { LEVEL_6_1, 2353004544u, 35651584, 240000, 0,     8, 16, 10, 4 },
 126   { LEVEL_6_2, 4706009088u, 35651584, 480000, 0,     8, 16, 10, 4 },
 127 };
 128 /* clang-format on */
 129
 130 static INLINE void Scale2Ratio(VPX_SCALING mode, int *hr, int *hs) {
 131   switch (mode) {
 132     case NORMAL:
 133       *hr = 1;
 134       *hs = 1;
 135       break;
 136     case FOURFIVE:
 137       *hr = 4;
 138       *hs = 5;
 139       break;
 140     case THREEFIVE:
 141       *hr = 3;
 142       *hs = 5;
 143       break;
 144     case ONETWO:
 145       *hr = 1;
 146       *hs = 2;
 147       break;
 148     default:
 149       *hr = 1;
 150       *hs = 1;
 151       assert(0);
 152       break;
 153   }
 154 }
 155
 156 // Mark all inactive blocks as active. Other segmentation features may be set
 157 // so memset cannot be used, instead only inactive blocks should be reset.
 158 static void suppress_active_map(VP9_COMP *cpi) {
 159   unsigned char *const seg_map = cpi->segmentation_map;
 160
 161   if (cpi->active_map.enabled || cpi->active_map.update) {
 162     const int rows = cpi->common.mi_rows;
 163     const int cols = cpi->common.mi_cols;
 164     int i;
 165
 166     for (i = 0; i < rows * cols; ++i)
 167       if (seg_map[i] == AM_SEGMENT_ID_INACTIVE)
 168         seg_map[i] = AM_SEGMENT_ID_ACTIVE;
 169   }
 170 }
 171
 172 static void apply_active_map(VP9_COMP *cpi) {
 173   struct segmentation *const seg = &cpi->common.seg;
 174   unsigned char *const seg_map = cpi->segmentation_map;
 175   const unsigned char *const active_map = cpi->active_map.map;
 176   int i;
 177
 178   assert(AM_SEGMENT_ID_ACTIVE == CR_SEGMENT_ID_BASE);
 179
 180   if (frame_is_intra_only(&cpi->common)) {
 181     cpi->active_map.enabled = 0;
 182     cpi->active_map.update = 1;
 183   }
 184
 185   if (cpi->active_map.update) {
 186     if (cpi->active_map.enabled) {
 187       for (i = 0; i < cpi->common.mi_rows * cpi->common.mi_cols; ++i)
 188         if (seg_map[i] == AM_SEGMENT_ID_ACTIVE) seg_map[i] = active_map[i];
 189       vp9_enable_segmentation(seg);
 190       vp9_enable_segfeature(seg, AM_SEGMENT_ID_INACTIVE, SEG_LVL_SKIP);
 191       vp9_enable_segfeature(seg, AM_SEGMENT_ID_INACTIVE, SEG_LVL_ALT_LF);
 192       // Setting the data to -MAX_LOOP_FILTER will result in the computed loop
 193       // filter level being zero regardless of the value of seg->abs_delta.
 194       vp9_set_segdata(seg, AM_SEGMENT_ID_INACTIVE, SEG_LVL_ALT_LF,
 195                       -MAX_LOOP_FILTER);
 196     } else {
 197       vp9_disable_segfeature(seg, AM_SEGMENT_ID_INACTIVE, SEG_LVL_SKIP);
 198       vp9_disable_segfeature(seg, AM_SEGMENT_ID_INACTIVE, SEG_LVL_ALT_LF);
 199       if (seg->enabled) {
 200         seg->update_data = 1;
 201         seg->update_map = 1;
 202       }
 203     }
 204     cpi->active_map.update = 0;
 205   }
 206 }
 207
 208 static void init_level_info(Vp9LevelInfo *level_info) {
 209   Vp9LevelStats *const level_stats = &level_info->level_stats;
 210   Vp9LevelSpec *const level_spec = &level_info->level_spec;
 211
 212   memset(level_stats, 0, sizeof(*level_stats));
 213   memset(level_spec, 0, sizeof(*level_spec));
 214   level_spec->level = LEVEL_UNKNOWN;
 215   level_spec->min_altref_distance = INT_MAX;
 216 }
 217
 218 VP9_LEVEL vp9_get_level(const Vp9LevelSpec *const level_spec) {
 219   int i;
 220   const Vp9LevelSpec *this_level;
 221
 222   vpx_clear_system_state();
 223
 224   for (i = 0; i < VP9_LEVELS; ++i) {
 225     this_level = &vp9_level_defs[i];
 226     if ((double)level_spec->max_luma_sample_rate * (1 + SAMPLE_RATE_GRACE_P) >
 227             (double)this_level->max_luma_sample_rate ||
 228         level_spec->max_luma_picture_size > this_level->max_luma_picture_size ||
 229         level_spec->average_bitrate > this_level->average_bitrate ||
 230         level_spec->max_cpb_size > this_level->max_cpb_size ||
 231         level_spec->compression_ratio < this_level->compression_ratio ||
 232         level_spec->max_col_tiles > this_level->max_col_tiles ||
 233         level_spec->min_altref_distance < this_level->min_altref_distance ||
 234         level_spec->max_ref_frame_buffers > this_level->max_ref_frame_buffers)
 235       continue;
 236     break;
 237   }
 238   return (i == VP9_LEVELS) ? LEVEL_UNKNOWN : vp9_level_defs[i].level;
 239 }
 240
 241 int vp9_set_active_map(VP9_COMP *cpi, unsigned char *new_map_16x16, int rows,
 242                        int cols) {
 243   if (rows == cpi->common.mb_rows && cols == cpi->common.mb_cols) {
 244     unsigned char *const active_map_8x8 = cpi->active_map.map;
 245     const int mi_rows = cpi->common.mi_rows;
 246     const int mi_cols = cpi->common.mi_cols;
 247     cpi->active_map.update = 1;
 248     if (new_map_16x16) {
 249       int r, c;
 250       for (r = 0; r < mi_rows; ++r) {
 251         for (c = 0; c < mi_cols; ++c) {
 252           active_map_8x8[r * mi_cols + c] =
 253               new_map_16x16[(r >> 1) * cols + (c >> 1)]
 254                   ? AM_SEGMENT_ID_ACTIVE
 255                   : AM_SEGMENT_ID_INACTIVE;
 256         }
 257       }
 258       cpi->active_map.enabled = 1;
 259     } else {
 260       cpi->active_map.enabled = 0;
 261     }
 262     return 0;
 263   } else {
 264     return -1;
 265   }
 266 }
 267
 268 int vp9_get_active_map(VP9_COMP *cpi, unsigned char *new_map_16x16, int rows,
 269                        int cols) {
 270   if (rows == cpi->common.mb_rows && cols == cpi->common.mb_cols &&
 271       new_map_16x16) {
 272     unsigned char *const seg_map_8x8 = cpi->segmentation_map;
 273     const int mi_rows = cpi->common.mi_rows;
 274     const int mi_cols = cpi->common.mi_cols;
 275     memset(new_map_16x16, !cpi->active_map.enabled, rows * cols);
 276     if (cpi->active_map.enabled) {
 277       int r, c;
 278       for (r = 0; r < mi_rows; ++r) {
 279         for (c = 0; c < mi_cols; ++c) {
 280           // Cyclic refresh segments are considered active despite not having
 281           // AM_SEGMENT_ID_ACTIVE
 282           new_map_16x16[(r >> 1) * cols + (c >> 1)] |=
 283               seg_map_8x8[r * mi_cols + c] != AM_SEGMENT_ID_INACTIVE;
 284         }
 285       }
 286     }
 287     return 0;
 288   } else {
 289     return -1;
 290   }
 291 }
 292
 293 void vp9_set_high_precision_mv(VP9_COMP *cpi, int allow_high_precision_mv) {
 294   MACROBLOCK *const mb = &cpi->td.mb;
 295   cpi->common.allow_high_precision_mv = allow_high_precision_mv;
 296   if (cpi->common.allow_high_precision_mv) {
 297     mb->mvcost = mb->nmvcost_hp;
 298     mb->mvsadcost = mb->nmvsadcost_hp;
 299   } else {
 300     mb->mvcost = mb->nmvcost;
 301     mb->mvsadcost = mb->nmvsadcost;
 302   }
 303 }
 304
 305 static void setup_frame(VP9_COMP *cpi) {
 306   VP9_COMMON *const cm = &cpi->common;
 307   // Set up entropy context depending on frame type. The decoder mandates
 308   // the use of the default context, index 0, for keyframes and inter
 309   // frames where the error_resilient_mode or intra_only flag is set. For
 310   // other inter-frames the encoder currently uses only two contexts;
 311   // context 1 for ALTREF frames and context 0 for the others.
 312   if (frame_is_intra_only(cm) || cm->error_resilient_mode) {
 313     vp9_setup_past_independence(cm);
 314   } else {
 315     if (!cpi->use_svc) cm->frame_context_idx = cpi->refresh_alt_ref_frame;
 316   }
 317
 318   if (cm->frame_type == KEY_FRAME) {
 319     if (!is_two_pass_svc(cpi)) cpi->refresh_golden_frame = 1;
 320     cpi->refresh_alt_ref_frame = 1;
 321     vp9_zero(cpi->interp_filter_selected);
 322   } else {
 323     *cm->fc = cm->frame_contexts[cm->frame_context_idx];
 324     vp9_zero(cpi->interp_filter_selected[0]);
 325   }
 326 }
 327
 328 static void vp9_enc_setup_mi(VP9_COMMON *cm) {
 329   int i;
 330   cm->mi = cm->mip + cm->mi_stride + 1;
 331   memset(cm->mip, 0, cm->mi_stride * (cm->mi_rows + 1) * sizeof(*cm->mip));
 332   cm->prev_mi = cm->prev_mip + cm->mi_stride + 1;
 333   // Clear top border row
 334   memset(cm->prev_mip, 0, sizeof(*cm->prev_mip) * cm->mi_stride);
 335   // Clear left border column
 336   for (i = 1; i < cm->mi_rows + 1; ++i)
 337     memset(&cm->prev_mip[i * cm->mi_stride], 0, sizeof(*cm->prev_mip));
 338
 339   cm->mi_grid_visible = cm->mi_grid_base + cm->mi_stride + 1;
 340   cm->prev_mi_grid_visible = cm->prev_mi_grid_base + cm->mi_stride + 1;
 341
 342   memset(cm->mi_grid_base, 0,
 343          cm->mi_stride * (cm->mi_rows + 1) * sizeof(*cm->mi_grid_base));
 344 }
 345
 346 static int vp9_enc_alloc_mi(VP9_COMMON *cm, int mi_size) {
 347   cm->mip = vpx_calloc(mi_size, sizeof(*cm->mip));
 348   if (!cm->mip) return 1;
 349   cm->prev_mip = vpx_calloc(mi_size, sizeof(*cm->prev_mip));
 350   if (!cm->prev_mip) return 1;
 351   cm->mi_alloc_size = mi_size;
 352
 353   cm->mi_grid_base = (MODE_INFO **)vpx_calloc(mi_size, sizeof(MODE_INFO *));
 354   if (!cm->mi_grid_base) return 1;
 355   cm->prev_mi_grid_base =
 356       (MODE_INFO **)vpx_calloc(mi_size, sizeof(MODE_INFO *));
 357   if (!cm->prev_mi_grid_base) return 1;
 358
 359   return 0;
 360 }
 361
 362 static void vp9_enc_free_mi(VP9_COMMON *cm) {
 363   vpx_free(cm->mip);
 364   cm->mip = NULL;
 365   vpx_free(cm->prev_mip);
 366   cm->prev_mip = NULL;
 367   vpx_free(cm->mi_grid_base);
 368   cm->mi_grid_base = NULL;
 369   vpx_free(cm->prev_mi_grid_base);
 370   cm->prev_mi_grid_base = NULL;
 371 }
 372
 373 static void vp9_swap_mi_and_prev_mi(VP9_COMMON *cm) {
 374   // Current mip will be the prev_mip for the next frame.
 375   MODE_INFO **temp_base = cm->prev_mi_grid_base;
 376   MODE_INFO *temp = cm->prev_mip;
 377   cm->prev_mip = cm->mip;
 378   cm->mip = temp;
 379
 380   // Update the upper left visible macroblock ptrs.
 381   cm->mi = cm->mip + cm->mi_stride + 1;
 382   cm->prev_mi = cm->prev_mip + cm->mi_stride + 1;
 383
 384   cm->prev_mi_grid_base = cm->mi_grid_base;
 385   cm->mi_grid_base = temp_base;
 386   cm->mi_grid_visible = cm->mi_grid_base + cm->mi_stride + 1;
 387   cm->prev_mi_grid_visible = cm->prev_mi_grid_base + cm->mi_stride + 1;
 388 }
 389
 390 void vp9_initialize_enc(void) {
 391   static volatile int init_done = 0;
 392
 393   if (!init_done) {
 394     vp9_rtcd();
 395     vpx_dsp_rtcd();
 396     vpx_scale_rtcd();
 397     vp9_init_intra_predictors();
 398     vp9_init_me_luts();
 399     vp9_rc_init_minq_luts();
 400     vp9_entropy_mv_init();
 401     vp9_temporal_filter_init();
 402     init_done = 1;
 403   }
 404 }
 405
 406 static void dealloc_compressor_data(VP9_COMP *cpi) {
 407   VP9_COMMON *const cm = &cpi->common;
 408   int i;
 409
 410   vpx_free(cpi->mbmi_ext_base);
 411   cpi->mbmi_ext_base = NULL;
 412
 413   vpx_free(cpi->tile_data);
 414   cpi->tile_data = NULL;
 415
 416   vpx_free(cpi->segmentation_map);
 417   cpi->segmentation_map = NULL;
 418   vpx_free(cpi->coding_context.last_frame_seg_map_copy);
 419   cpi->coding_context.last_frame_seg_map_copy = NULL;
 420
 421   vpx_free(cpi->nmvcosts[0]);
 422   vpx_free(cpi->nmvcosts[1]);
 423   cpi->nmvcosts[0] = NULL;
 424   cpi->nmvcosts[1] = NULL;
 425
 426   vpx_free(cpi->nmvcosts_hp[0]);
 427   vpx_free(cpi->nmvcosts_hp[1]);
 428   cpi->nmvcosts_hp[0] = NULL;
 429   cpi->nmvcosts_hp[1] = NULL;
 430
 431   vpx_free(cpi->nmvsadcosts[0]);
 432   vpx_free(cpi->nmvsadcosts[1]);
 433   cpi->nmvsadcosts[0] = NULL;
 434   cpi->nmvsadcosts[1] = NULL;
 435
 436   vpx_free(cpi->nmvsadcosts_hp[0]);
 437   vpx_free(cpi->nmvsadcosts_hp[1]);
 438   cpi->nmvsadcosts_hp[0] = NULL;
 439   cpi->nmvsadcosts_hp[1] = NULL;
 440
 441   vp9_cyclic_refresh_free(cpi->cyclic_refresh);
 442   cpi->cyclic_refresh = NULL;
 443
 444   vpx_free(cpi->active_map.map);
 445   cpi->active_map.map = NULL;
 446
 447   vpx_free(cpi->consec_zero_mv);
 448   cpi->consec_zero_mv = NULL;
 449
 450   vp9_free_ref_frame_buffers(cm->buffer_pool);
 451 #if CONFIG_VP9_POSTPROC
 452   vp9_free_postproc_buffers(cm);
 453 #endif
 454   vp9_free_context_buffers(cm);
 455
 456   vpx_free_frame_buffer(&cpi->last_frame_uf);
 457   vpx_free_frame_buffer(&cpi->scaled_source);
 458   vpx_free_frame_buffer(&cpi->scaled_last_source);
 459   vpx_free_frame_buffer(&cpi->alt_ref_buffer);
 460 #ifdef ENABLE_KF_DENOISE
 461   vpx_free_frame_buffer(&cpi->raw_unscaled_source);
 462   vpx_free_frame_buffer(&cpi->raw_scaled_source);
 463 #endif
 464
 465   vp9_lookahead_destroy(cpi->lookahead);
 466
 467   vpx_free(cpi->tile_tok[0][0]);
 468   cpi->tile_tok[0][0] = 0;
 469
 470   vp9_free_pc_tree(&cpi->td);
 471
 472   for (i = 0; i < cpi->svc.number_spatial_layers; ++i) {
 473     LAYER_CONTEXT *const lc = &cpi->svc.layer_context[i];
 474     vpx_free(lc->rc_twopass_stats_in.buf);
 475     lc->rc_twopass_stats_in.buf = NULL;
 476     lc->rc_twopass_stats_in.sz = 0;
 477   }
 478
 479   if (cpi->source_diff_var != NULL) {
 480     vpx_free(cpi->source_diff_var);
 481     cpi->source_diff_var = NULL;
 482   }
 483
 484   for (i = 0; i < MAX_LAG_BUFFERS; ++i) {
 485     vpx_free_frame_buffer(&cpi->svc.scaled_frames[i]);
 486   }
 487   memset(&cpi->svc.scaled_frames[0], 0,
 488          MAX_LAG_BUFFERS * sizeof(cpi->svc.scaled_frames[0]));
 489
 490   vpx_free_frame_buffer(&cpi->svc.scaled_temp);
 491   memset(&cpi->svc.scaled_temp, 0, sizeof(cpi->svc.scaled_temp));
 492
 493   vpx_free_frame_buffer(&cpi->svc.empty_frame.img);
 494   memset(&cpi->svc.empty_frame, 0, sizeof(cpi->svc.empty_frame));
 495
 496   vp9_free_svc_cyclic_refresh(cpi);
 497 }
 498
 499 static void save_coding_context(VP9_COMP *cpi) {
 500   CODING_CONTEXT *const cc = &cpi->coding_context;
 501   VP9_COMMON *cm = &cpi->common;
 502
 503   // Stores a snapshot of key state variables which can subsequently be
 504   // restored with a call to vp9_restore_coding_context. These functions are
 505   // intended for use in a re-code loop in vp9_compress_frame where the
 506   // quantizer value is adjusted between loop iterations.
 507   vp9_copy(cc->nmvjointcost, cpi->td.mb.nmvjointcost);
 508
 509   memcpy(cc->nmvcosts[0], cpi->nmvcosts[0],
 510          MV_VALS * sizeof(*cpi->nmvcosts[0]));
 511   memcpy(cc->nmvcosts[1], cpi->nmvcosts[1],
 512          MV_VALS * sizeof(*cpi->nmvcosts[1]));
 513   memcpy(cc->nmvcosts_hp[0], cpi->nmvcosts_hp[0],
 514          MV_VALS * sizeof(*cpi->nmvcosts_hp[0]));
 515   memcpy(cc->nmvcosts_hp[1], cpi->nmvcosts_hp[1],
 516          MV_VALS * sizeof(*cpi->nmvcosts_hp[1]));
 517
 518   vp9_copy(cc->segment_pred_probs, cm->seg.pred_probs);
 519
 520   memcpy(cpi->coding_context.last_frame_seg_map_copy, cm->last_frame_seg_map,
 521          (cm->mi_rows * cm->mi_cols));
 522
 523   vp9_copy(cc->last_ref_lf_deltas, cm->lf.last_ref_deltas);
 524   vp9_copy(cc->last_mode_lf_deltas, cm->lf.last_mode_deltas);
 525
 526   cc->fc = *cm->fc;
 527 }
 528
 529 static void restore_coding_context(VP9_COMP *cpi) {
 530   CODING_CONTEXT *const cc = &cpi->coding_context;
 531   VP9_COMMON *cm = &cpi->common;
 532
 533   // Restore key state variables to the snapshot state stored in the
 534   // previous call to vp9_save_coding_context.
 535   vp9_copy(cpi->td.mb.nmvjointcost, cc->nmvjointcost);
 536
 537   memcpy(cpi->nmvcosts[0], cc->nmvcosts[0], MV_VALS * sizeof(*cc->nmvcosts[0]));
 538   memcpy(cpi->nmvcosts[1], cc->nmvcosts[1], MV_VALS * sizeof(*cc->nmvcosts[1]));
 539   memcpy(cpi->nmvcosts_hp[0], cc->nmvcosts_hp[0],
 540          MV_VALS * sizeof(*cc->nmvcosts_hp[0]));
 541   memcpy(cpi->nmvcosts_hp[1], cc->nmvcosts_hp[1],
 542          MV_VALS * sizeof(*cc->nmvcosts_hp[1]));
 543
 544   vp9_copy(cm->seg.pred_probs, cc->segment_pred_probs);
 545
 546   memcpy(cm->last_frame_seg_map, cpi->coding_context.last_frame_seg_map_copy,
 547          (cm->mi_rows * cm->mi_cols));
 548
 549   vp9_copy(cm->lf.last_ref_deltas, cc->last_ref_lf_deltas);
 550   vp9_copy(cm->lf.last_mode_deltas, cc->last_mode_lf_deltas);
 551
 552   *cm->fc = cc->fc;
 553 }
 554
 555 static void configure_static_seg_features(VP9_COMP *cpi) {
 556   VP9_COMMON *const cm = &cpi->common;
 557   const RATE_CONTROL *const rc = &cpi->rc;
 558   struct segmentation *const seg = &cm->seg;
 559
 560   int high_q = (int)(rc->avg_q > 48.0);
 561   int qi_delta;
 562
 563   // Disable and clear down for KF
 564   if (cm->frame_type == KEY_FRAME) {
 565     // Clear down the global segmentation map
 566     memset(cpi->segmentation_map, 0, cm->mi_rows * cm->mi_cols);
 567     seg->update_map = 0;
 568     seg->update_data = 0;
 569     cpi->static_mb_pct = 0;
 570
 571     // Disable segmentation
 572     vp9_disable_segmentation(seg);
 573
 574     // Clear down the segment features.
 575     vp9_clearall_segfeatures(seg);
 576   } else if (cpi->refresh_alt_ref_frame) {
 577     // If this is an alt ref frame
 578     // Clear down the global segmentation map
 579     memset(cpi->segmentation_map, 0, cm->mi_rows * cm->mi_cols);
 580     seg->update_map = 0;
 581     seg->update_data = 0;
 582     cpi->static_mb_pct = 0;
 583
 584     // Disable segmentation and individual segment features by default
 585     vp9_disable_segmentation(seg);
 586     vp9_clearall_segfeatures(seg);
 587
 588     // Scan frames from current to arf frame.
 589     // This function re-enables segmentation if appropriate.
 590     vp9_update_mbgraph_stats(cpi);
 591
 592     // If segmentation was enabled set those features needed for the
 593     // arf itself.
 594     if (seg->enabled) {
 595       seg->update_map = 1;
 596       seg->update_data = 1;
 597
 598       qi_delta =
 599           vp9_compute_qdelta(rc, rc->avg_q, rc->avg_q * 0.875, cm->bit_depth);
 600       vp9_set_segdata(seg, 1, SEG_LVL_ALT_Q, qi_delta - 2);
 601       vp9_set_segdata(seg, 1, SEG_LVL_ALT_LF, -2);
 602
 603       vp9_enable_segfeature(seg, 1, SEG_LVL_ALT_Q);
 604       vp9_enable_segfeature(seg, 1, SEG_LVL_ALT_LF);
 605
 606       // Where relevant assume segment data is delta data
 607       seg->abs_delta = SEGMENT_DELTADATA;
 608     }
 609   } else if (seg->enabled) {
 610     // All other frames if segmentation has been enabled
 611
 612     // First normal frame in a valid gf or alt ref group
 613     if (rc->frames_since_golden == 0) {
 614       // Set up segment features for normal frames in an arf group
 615       if (rc->source_alt_ref_active) {
 616         seg->update_map = 0;
 617         seg->update_data = 1;
 618         seg->abs_delta = SEGMENT_DELTADATA;
 619
 620         qi_delta =
 621             vp9_compute_qdelta(rc, rc->avg_q, rc->avg_q * 1.125, cm->bit_depth);
 622         vp9_set_segdata(seg, 1, SEG_LVL_ALT_Q, qi_delta + 2);
 623         vp9_enable_segfeature(seg, 1, SEG_LVL_ALT_Q);
 624
 625         vp9_set_segdata(seg, 1, SEG_LVL_ALT_LF, -2);
 626         vp9_enable_segfeature(seg, 1, SEG_LVL_ALT_LF);
 627
 628         // Segment coding disabled for compred testing
 629         if (high_q || (cpi->static_mb_pct == 100)) {
 630           vp9_set_segdata(seg, 1, SEG_LVL_REF_FRAME, ALTREF_FRAME);
 631           vp9_enable_segfeature(seg, 1, SEG_LVL_REF_FRAME);
 632           vp9_enable_segfeature(seg, 1, SEG_LVL_SKIP);
 633         }
 634       } else {
 635         // Disable segmentation and clear down features if alt ref
 636         // is not active for this group
 637
 638         vp9_disable_segmentation(seg);
 639
 640         memset(cpi->segmentation_map, 0, cm->mi_rows * cm->mi_cols);
 641
 642         seg->update_map = 0;
 643         seg->update_data = 0;
 644
 645         vp9_clearall_segfeatures(seg);
 646       }
 647     } else if (rc->is_src_frame_alt_ref) {
 648       // Special case where we are coding over the top of a previous
 649       // alt ref frame.
 650       // Segment coding disabled for compred testing
 651
 652       // Enable ref frame features for segment 0 as well
 653       vp9_enable_segfeature(seg, 0, SEG_LVL_REF_FRAME);
 654       vp9_enable_segfeature(seg, 1, SEG_LVL_REF_FRAME);
 655
 656       // All mbs should use ALTREF_FRAME
 657       vp9_clear_segdata(seg, 0, SEG_LVL_REF_FRAME);
 658       vp9_set_segdata(seg, 0, SEG_LVL_REF_FRAME, ALTREF_FRAME);
 659       vp9_clear_segdata(seg, 1, SEG_LVL_REF_FRAME);
 660       vp9_set_segdata(seg, 1, SEG_LVL_REF_FRAME, ALTREF_FRAME);
 661
 662       // Skip all MBs if high Q (0,0 mv and skip coeffs)
 663       if (high_q) {
 664         vp9_enable_segfeature(seg, 0, SEG_LVL_SKIP);
 665         vp9_enable_segfeature(seg, 1, SEG_LVL_SKIP);
 666       }
 667       // Enable data update
 668       seg->update_data = 1;
 669     } else {
 670       // All other frames.
 671
 672       // No updates.. leave things as they are.
 673       seg->update_map = 0;
 674       seg->update_data = 0;
 675     }
 676   }
 677 }
 678
 679 static void update_reference_segmentation_map(VP9_COMP *cpi) {
 680   VP9_COMMON *const cm = &cpi->common;
 681   MODE_INFO **mi_8x8_ptr = cm->mi_grid_visible;
 682   uint8_t *cache_ptr = cm->last_frame_seg_map;
 683   int row, col;
 684
 685   for (row = 0; row < cm->mi_rows; row++) {
 686     MODE_INFO **mi_8x8 = mi_8x8_ptr;
 687     uint8_t *cache = cache_ptr;
 688     for (col = 0; col < cm->mi_cols; col++, mi_8x8++, cache++)
 689       cache[0] = mi_8x8[0]->segment_id;
 690     mi_8x8_ptr += cm->mi_stride;
 691     cache_ptr += cm->mi_cols;
 692   }
 693 }
 694
 695 static void alloc_raw_frame_buffers(VP9_COMP *cpi) {
 696   VP9_COMMON *cm = &cpi->common;
 697   const VP9EncoderConfig *oxcf = &cpi->oxcf;
 698
 699   if (!cpi->lookahead)
 700     cpi->lookahead = vp9_lookahead_init(oxcf->width, oxcf->height,
 701                                         cm->subsampling_x, cm->subsampling_y,
 702 #if CONFIG_VP9_HIGHBITDEPTH
 703                                         cm->use_highbitdepth,
 704 #endif
 705                                         oxcf->lag_in_frames);
 706   if (!cpi->lookahead)
 707     vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
 708                        "Failed to allocate lag buffers");
 709
 710   // TODO(agrange) Check if ARF is enabled and skip allocation if not.
 711   if (vpx_realloc_frame_buffer(&cpi->alt_ref_buffer, oxcf->width, oxcf->height,
 712                                cm->subsampling_x, cm->subsampling_y,
 713 #if CONFIG_VP9_HIGHBITDEPTH
 714                                cm->use_highbitdepth,
 715 #endif
 716                                VP9_ENC_BORDER_IN_PIXELS, cm->byte_alignment,
 717                                NULL, NULL, NULL))
 718     vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
 719                        "Failed to allocate altref buffer");
 720 }
 721
 722 static void alloc_util_frame_buffers(VP9_COMP *cpi) {
 723   VP9_COMMON *const cm = &cpi->common;
 724   if (vpx_realloc_frame_buffer(&cpi->last_frame_uf, cm->width, cm->height,
 725                                cm->subsampling_x, cm->subsampling_y,
 726 #if CONFIG_VP9_HIGHBITDEPTH
 727                                cm->use_highbitdepth,
 728 #endif
 729                                VP9_ENC_BORDER_IN_PIXELS, cm->byte_alignment,
 730                                NULL, NULL, NULL))
 731     vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
 732                        "Failed to allocate last frame buffer");
 733
 734   if (vpx_realloc_frame_buffer(&cpi->scaled_source, cm->width, cm->height,
 735                                cm->subsampling_x, cm->subsampling_y,
 736 #if CONFIG_VP9_HIGHBITDEPTH
 737                                cm->use_highbitdepth,
 738 #endif
 739                                VP9_ENC_BORDER_IN_PIXELS, cm->byte_alignment,
 740                                NULL, NULL, NULL))
 741     vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
 742                        "Failed to allocate scaled source buffer");
 743
 744   // For 1 pass cbr: allocate scaled_frame that may be used as an intermediate
 745   // buffer for a 2 stage down-sampling: two stages of 1:2 down-sampling for a
 746   // target of 1/4x1/4.
 747   if (is_one_pass_cbr_svc(cpi) && !cpi->svc.scaled_temp_is_alloc) {
 748     cpi->svc.scaled_temp_is_alloc = 1;
 749     if (vpx_realloc_frame_buffer(
 750             &cpi->svc.scaled_temp, cm->width >> 1, cm->height >> 1,
 751             cm->subsampling_x, cm->subsampling_y,
 752 #if CONFIG_VP9_HIGHBITDEPTH
 753             cm->use_highbitdepth,
 754 #endif
 755             VP9_ENC_BORDER_IN_PIXELS, cm->byte_alignment, NULL, NULL, NULL))
 756       vpx_internal_error(&cpi->common.error, VPX_CODEC_MEM_ERROR,
 757                          "Failed to allocate scaled_frame for svc ");
 758   }
 759
 760   if (vpx_realloc_frame_buffer(&cpi->scaled_last_source, cm->width, cm->height,
 761                                cm->subsampling_x, cm->subsampling_y,
 762 #if CONFIG_VP9_HIGHBITDEPTH
 763                                cm->use_highbitdepth,
 764 #endif
 765                                VP9_ENC_BORDER_IN_PIXELS, cm->byte_alignment,
 766                                NULL, NULL, NULL))
 767     vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
 768                        "Failed to allocate scaled last source buffer");
 769 #ifdef ENABLE_KF_DENOISE
 770   if (vpx_realloc_frame_buffer(&cpi->raw_unscaled_source, cm->width, cm->height,
 771                                cm->subsampling_x, cm->subsampling_y,
 772 #if CONFIG_VP9_HIGHBITDEPTH
 773                                cm->use_highbitdepth,
 774 #endif
 775                                VP9_ENC_BORDER_IN_PIXELS, cm->byte_alignment,
 776                                NULL, NULL, NULL))
 777     vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
 778                        "Failed to allocate unscaled raw source frame buffer");
 779
 780   if (vpx_realloc_frame_buffer(&cpi->raw_scaled_source, cm->width, cm->height,
 781                                cm->subsampling_x, cm->subsampling_y,
 782 #if CONFIG_VP9_HIGHBITDEPTH
 783                                cm->use_highbitdepth,
 784 #endif
 785                                VP9_ENC_BORDER_IN_PIXELS, cm->byte_alignment,
 786                                NULL, NULL, NULL))
 787     vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
 788                        "Failed to allocate scaled raw source frame buffer");
 789 #endif
 790 }
 791
 792 static int alloc_context_buffers_ext(VP9_COMP *cpi) {
 793   VP9_COMMON *cm = &cpi->common;
 794   int mi_size = cm->mi_cols * cm->mi_rows;
 795
 796   cpi->mbmi_ext_base = vpx_calloc(mi_size, sizeof(*cpi->mbmi_ext_base));
 797   if (!cpi->mbmi_ext_base) return 1;
 798
 799   return 0;
 800 }
 801
 802 static void alloc_compressor_data(VP9_COMP *cpi) {
 803   VP9_COMMON *cm = &cpi->common;
 804
 805   vp9_alloc_context_buffers(cm, cm->width, cm->height);
 806
 807   alloc_context_buffers_ext(cpi);
 808
 809   vpx_free(cpi->tile_tok[0][0]);
 810
 811   {
 812     unsigned int tokens = get_token_alloc(cm->mb_rows, cm->mb_cols);
 813     CHECK_MEM_ERROR(cm, cpi->tile_tok[0][0],
 814                     vpx_calloc(tokens, sizeof(*cpi->tile_tok[0][0])));
 815   }
 816
 817   vp9_setup_pc_tree(&cpi->common, &cpi->td);
 818 }
 819
 820 void vp9_new_framerate(VP9_COMP *cpi, double framerate) {
 821   cpi->framerate = framerate < 0.1 ? 30 : framerate;
 822   vp9_rc_update_framerate(cpi);
 823 }
 824
 825 static void set_tile_limits(VP9_COMP *cpi) {
 826   VP9_COMMON *const cm = &cpi->common;
 827
 828   int min_log2_tile_cols, max_log2_tile_cols;
 829   vp9_get_tile_n_bits(cm->mi_cols, &min_log2_tile_cols, &max_log2_tile_cols);
 830
 831   if (is_two_pass_svc(cpi) && (cpi->svc.encode_empty_frame_state == ENCODING ||
 832                                cpi->svc.number_spatial_layers > 1)) {
 833     cm->log2_tile_cols = 0;
 834     cm->log2_tile_rows = 0;
 835   } else {
 836     cm->log2_tile_cols =
 837         clamp(cpi->oxcf.tile_columns, min_log2_tile_cols, max_log2_tile_cols);
 838     cm->log2_tile_rows = cpi->oxcf.tile_rows;
 839   }
 840 }
 841
 842 static void update_frame_size(VP9_COMP *cpi) {
 843   VP9_COMMON *const cm = &cpi->common;
 844   MACROBLOCKD *const xd = &cpi->td.mb.e_mbd;
 845
 846   vp9_set_mb_mi(cm, cm->width, cm->height);
 847   vp9_init_context_buffers(cm);
 848   vp9_init_macroblockd(cm, xd, NULL);
 849   cpi->td.mb.mbmi_ext_base = cpi->mbmi_ext_base;
 850   memset(cpi->mbmi_ext_base, 0,
 851          cm->mi_rows * cm->mi_cols * sizeof(*cpi->mbmi_ext_base));
 852
 853   set_tile_limits(cpi);
 854
 855   if (is_two_pass_svc(cpi)) {
 856     if (vpx_realloc_frame_buffer(&cpi->alt_ref_buffer, cm->width, cm->height,
 857                                  cm->subsampling_x, cm->subsampling_y,
 858 #if CONFIG_VP9_HIGHBITDEPTH
 859                                  cm->use_highbitdepth,
 860 #endif
 861                                  VP9_ENC_BORDER_IN_PIXELS, cm->byte_alignment,
 862                                  NULL, NULL, NULL))
 863       vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
 864                          "Failed to reallocate alt_ref_buffer");
 865   }
 866 }
 867
 868 static void init_buffer_indices(VP9_COMP *cpi) {
 869   cpi->lst_fb_idx = 0;
 870   cpi->gld_fb_idx = 1;
 871   cpi->alt_fb_idx = 2;
 872 }
 873
 874 static void init_config(struct VP9_COMP *cpi, VP9EncoderConfig *oxcf) {
 875   VP9_COMMON *const cm = &cpi->common;
 876
 877   cpi->oxcf = *oxcf;
 878   cpi->framerate = oxcf->init_framerate;
 879   cm->profile = oxcf->profile;
 880   cm->bit_depth = oxcf->bit_depth;
 881 #if CONFIG_VP9_HIGHBITDEPTH
 882   cm->use_highbitdepth = oxcf->use_highbitdepth;
 883 #endif
 884   cm->color_space = oxcf->color_space;
 885   cm->color_range = oxcf->color_range;
 886
 887   cpi->target_level = oxcf->target_level;
 888   cpi->keep_level_stats = oxcf->target_level != LEVEL_MAX;
 889
 890   cm->width = oxcf->width;
 891   cm->height = oxcf->height;
 892   alloc_compressor_data(cpi);
 893
 894   cpi->svc.temporal_layering_mode = oxcf->temporal_layering_mode;
 895
 896   // Single thread case: use counts in common.
 897   cpi->td.counts = &cm->counts;
 898
 899   // Spatial scalability.
 900   cpi->svc.number_spatial_layers = oxcf->ss_number_layers;
 901   // Temporal scalability.
 902   cpi->svc.number_temporal_layers = oxcf->ts_number_layers;
 903
 904   if ((cpi->svc.number_temporal_layers > 1 && cpi->oxcf.rc_mode == VPX_CBR) ||
 905       ((cpi->svc.number_temporal_layers > 1 ||
 906         cpi->svc.number_spatial_layers > 1) &&
 907        cpi->oxcf.pass != 1)) {
 908     vp9_init_layer_context(cpi);
 909   }
 910
 911   // change includes all joint functionality
 912   vp9_change_config(cpi, oxcf);
 913
 914   cpi->static_mb_pct = 0;
 915   cpi->ref_frame_flags = 0;
 916
 917   init_buffer_indices(cpi);
 918
 919   vp9_noise_estimate_init(&cpi->noise_estimate, cm->width, cm->height);
 920 }
 921
 922 static void set_rc_buffer_sizes(RATE_CONTROL *rc,
 923                                 const VP9EncoderConfig *oxcf) {
 924   const int64_t bandwidth = oxcf->target_bandwidth;
 925   const int64_t starting = oxcf->starting_buffer_level_ms;
 926   const int64_t optimal = oxcf->optimal_buffer_level_ms;
 927   const int64_t maximum = oxcf->maximum_buffer_size_ms;
 928
 929   rc->starting_buffer_level = starting * bandwidth / 1000;
 930   rc->optimal_buffer_level =
 931       (optimal == 0) ? bandwidth / 8 : optimal * bandwidth / 1000;
 932   rc->maximum_buffer_size =
 933       (maximum == 0) ? bandwidth / 8 : maximum * bandwidth / 1000;
 934 }
 935
 936 #if CONFIG_VP9_HIGHBITDEPTH
 937 #define HIGHBD_BFP(BT, SDF, SDAF, VF, SVF, SVAF, SDX3F, SDX8F, SDX4DF) \
 938   cpi->fn_ptr[BT].sdf = SDF;                                           \
 939   cpi->fn_ptr[BT].sdaf = SDAF;                                         \
 940   cpi->fn_ptr[BT].vf = VF;                                             \
 941   cpi->fn_ptr[BT].svf = SVF;                                           \
 942   cpi->fn_ptr[BT].svaf = SVAF;                                         \
 943   cpi->fn_ptr[BT].sdx3f = SDX3F;                                       \
 944   cpi->fn_ptr[BT].sdx8f = SDX8F;                                       \
 945   cpi->fn_ptr[BT].sdx4df = SDX4DF;
 946
 947 #define MAKE_BFP_SAD_WRAPPER(fnname)                                           \
 948   static unsigned int fnname##_bits8(const uint8_t *src_ptr,                   \
 949                                      int source_stride,                        \
 950                                      const uint8_t *ref_ptr, int ref_stride) { \
 951     return fnname(src_ptr, source_stride, ref_ptr, ref_stride);                \
 952   }                                                                            \
 953   static unsigned int fnname##_bits10(                                         \
 954       const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr,       \
 955       int ref_stride) {                                                        \
 956     return fnname(src_ptr, source_stride, ref_ptr, ref_stride) >> 2;           \
 957   }                                                                            \
 958   static unsigned int fnname##_bits12(                                         \
 959       const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr,       \
 960       int ref_stride) {                                                        \
 961     return fnname(src_ptr, source_stride, ref_ptr, ref_stride) >> 4;           \
 962   }
 963
 964 #define MAKE_BFP_SADAVG_WRAPPER(fnname)                                        \
 965   static unsigned int fnname##_bits8(                                          \
 966       const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr,       \
 967       int ref_stride, const uint8_t *second_pred) {                            \
 968     return fnname(src_ptr, source_stride, ref_ptr, ref_stride, second_pred);   \
 969   }                                                                            \
 970   static unsigned int fnname##_bits10(                                         \
 971       const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr,       \
 972       int ref_stride, const uint8_t *second_pred) {                            \
 973     return fnname(src_ptr, source_stride, ref_ptr, ref_stride, second_pred) >> \
 974            2;                                                                  \
 975   }                                                                            \
 976   static unsigned int fnname##_bits12(                                         \
 977       const uint8_t *src_ptr, int source_stride, const uint8_t *ref_ptr,       \
 978       int ref_stride, const uint8_t *second_pred) {                            \
 979     return fnname(src_ptr, source_stride, ref_ptr, ref_stride, second_pred) >> \
 980            4;                                                                  \
 981   }
 982
 983 #define MAKE_BFP_SAD3_WRAPPER(fnname)                                    \
 984   static void fnname##_bits8(const uint8_t *src_ptr, int source_stride,  \
 985                              const uint8_t *ref_ptr, int ref_stride,     \
 986                              unsigned int *sad_array) {                  \
 987     fnname(src_ptr, source_stride, ref_ptr, ref_stride, sad_array);      \
 988   }                                                                      \
 989   static void fnname##_bits10(const uint8_t *src_ptr, int source_stride, \
 990                               const uint8_t *ref_ptr, int ref_stride,    \
 991                               unsigned int *sad_array) {                 \
 992     int i;                                                               \
 993     fnname(src_ptr, source_stride, ref_ptr, ref_stride, sad_array);      \
 994     for (i = 0; i < 3; i++) sad_array[i] >>= 2;                          \
 995   }                                                                      \
 996   static void fnname##_bits12(const uint8_t *src_ptr, int source_stride, \
 997                               const uint8_t *ref_ptr, int ref_stride,    \
 998                               unsigned int *sad_array) {                 \
 999     int i;                                                               \
1000     fnname(src_ptr, source_stride, ref_ptr, ref_stride, sad_array);      \
1001     for (i = 0; i < 3; i++) sad_array[i] >>= 4;                          \
1002   }
1003
1004 #define MAKE_BFP_SAD8_WRAPPER(fnname)                                    \
1005   static void fnname##_bits8(const uint8_t *src_ptr, int source_stride,  \
1006                              const uint8_t *ref_ptr, int ref_stride,     \
1007                              unsigned int *sad_array) {                  \
1008     fnname(src_ptr, source_stride, ref_ptr, ref_stride, sad_array);      \
1009   }                                                                      \
1010   static void fnname##_bits10(const uint8_t *src_ptr, int source_stride, \
1011                               const uint8_t *ref_ptr, int ref_stride,    \
1012                               unsigned int *sad_array) {                 \
1013     int i;                                                               \
1014     fnname(src_ptr, source_stride, ref_ptr, ref_stride, sad_array);      \
1015     for (i = 0; i < 8; i++) sad_array[i] >>= 2;                          \
1016   }                                                                      \
1017   static void fnname##_bits12(const uint8_t *src_ptr, int source_stride, \
1018                               const uint8_t *ref_ptr, int ref_stride,    \
1019                               unsigned int *sad_array) {                 \
1020     int i;                                                               \
1021     fnname(src_ptr, source_stride, ref_ptr, ref_stride, sad_array);      \
1022     for (i = 0; i < 8; i++) sad_array[i] >>= 4;                          \
1023   }
1024 #define MAKE_BFP_SAD4D_WRAPPER(fnname)                                        \
1025   static void fnname##_bits8(const uint8_t *src_ptr, int source_stride,       \
1026                              const uint8_t *const ref_ptr[], int ref_stride,  \
1027                              unsigned int *sad_array) {                       \
1028     fnname(src_ptr, source_stride, ref_ptr, ref_stride, sad_array);           \
1029   }                                                                           \
1030   static void fnname##_bits10(const uint8_t *src_ptr, int source_stride,      \
1031                               const uint8_t *const ref_ptr[], int ref_stride, \
1032                               unsigned int *sad_array) {                      \
1033     int i;                                                                    \
1034     fnname(src_ptr, source_stride, ref_ptr, ref_stride, sad_array);           \
1035     for (i = 0; i < 4; i++) sad_array[i] >>= 2;                               \
1036   }                                                                           \
1037   static void fnname##_bits12(const uint8_t *src_ptr, int source_stride,      \
1038                               const uint8_t *const ref_ptr[], int ref_stride, \
1039                               unsigned int *sad_array) {                      \
1040     int i;                                                                    \
1041     fnname(src_ptr, source_stride, ref_ptr, ref_stride, sad_array);           \
1042     for (i = 0; i < 4; i++) sad_array[i] >>= 4;                               \
1043   }
1044
1045 MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad32x16)
1046 MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad32x16_avg)
1047 MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad32x16x4d)
1048 MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad16x32)
1049 MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad16x32_avg)
1050 MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad16x32x4d)
1051 MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad64x32)
1052 MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad64x32_avg)
1053 MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad64x32x4d)
1054 MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad32x64)
1055 MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad32x64_avg)
1056 MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad32x64x4d)
1057 MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad32x32)
1058 MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad32x32_avg)
1059 MAKE_BFP_SAD3_WRAPPER(vpx_highbd_sad32x32x3)
1060 MAKE_BFP_SAD8_WRAPPER(vpx_highbd_sad32x32x8)
1061 MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad32x32x4d)
1062 MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad64x64)
1063 MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad64x64_avg)
1064 MAKE_BFP_SAD3_WRAPPER(vpx_highbd_sad64x64x3)
1065 MAKE_BFP_SAD8_WRAPPER(vpx_highbd_sad64x64x8)
1066 MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad64x64x4d)
1067 MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad16x16)
1068 MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad16x16_avg)
1069 MAKE_BFP_SAD3_WRAPPER(vpx_highbd_sad16x16x3)
1070 MAKE_BFP_SAD8_WRAPPER(vpx_highbd_sad16x16x8)
1071 MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad16x16x4d)
1072 MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad16x8)
1073 MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad16x8_avg)
1074 MAKE_BFP_SAD3_WRAPPER(vpx_highbd_sad16x8x3)
1075 MAKE_BFP_SAD8_WRAPPER(vpx_highbd_sad16x8x8)
1076 MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad16x8x4d)
1077 MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad8x16)
1078 MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad8x16_avg)
1079 MAKE_BFP_SAD3_WRAPPER(vpx_highbd_sad8x16x3)
1080 MAKE_BFP_SAD8_WRAPPER(vpx_highbd_sad8x16x8)
1081 MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad8x16x4d)
1082 MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad8x8)
1083 MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad8x8_avg)
1084 MAKE_BFP_SAD3_WRAPPER(vpx_highbd_sad8x8x3)
1085 MAKE_BFP_SAD8_WRAPPER(vpx_highbd_sad8x8x8)
1086 MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad8x8x4d)
1087 MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad8x4)
1088 MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad8x4_avg)
1089 MAKE_BFP_SAD8_WRAPPER(vpx_highbd_sad8x4x8)
1090 MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad8x4x4d)
1091 MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad4x8)
1092 MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad4x8_avg)
1093 MAKE_BFP_SAD8_WRAPPER(vpx_highbd_sad4x8x8)
1094 MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad4x8x4d)
1095 MAKE_BFP_SAD_WRAPPER(vpx_highbd_sad4x4)
1096 MAKE_BFP_SADAVG_WRAPPER(vpx_highbd_sad4x4_avg)
1097 MAKE_BFP_SAD3_WRAPPER(vpx_highbd_sad4x4x3)
1098 MAKE_BFP_SAD8_WRAPPER(vpx_highbd_sad4x4x8)
1099 MAKE_BFP_SAD4D_WRAPPER(vpx_highbd_sad4x4x4d)
1100
1101 static void highbd_set_var_fns(VP9_COMP *const cpi) {
1102   VP9_COMMON *const cm = &cpi->common;
1103   if (cm->use_highbitdepth) {
1104     switch (cm->bit_depth) {
1105       case VPX_BITS_8:
1106         HIGHBD_BFP(BLOCK_32X16, vpx_highbd_sad32x16_bits8,
1107                    vpx_highbd_sad32x16_avg_bits8, vpx_highbd_8_variance32x16,
1108                    vpx_highbd_8_sub_pixel_variance32x16,
1109                    vpx_highbd_8_sub_pixel_avg_variance32x16, NULL, NULL,
1110                    vpx_highbd_sad32x16x4d_bits8)
1111
1112         HIGHBD_BFP(BLOCK_16X32, vpx_highbd_sad16x32_bits8,
1113                    vpx_highbd_sad16x32_avg_bits8, vpx_highbd_8_variance16x32,
1114                    vpx_highbd_8_sub_pixel_variance16x32,
1115                    vpx_highbd_8_sub_pixel_avg_variance16x32, NULL, NULL,
1116                    vpx_highbd_sad16x32x4d_bits8)
1117
1118         HIGHBD_BFP(BLOCK_64X32, vpx_highbd_sad64x32_bits8,
1119                    vpx_highbd_sad64x32_avg_bits8, vpx_highbd_8_variance64x32,
1120                    vpx_highbd_8_sub_pixel_variance64x32,
1121                    vpx_highbd_8_sub_pixel_avg_variance64x32, NULL, NULL,
1122                    vpx_highbd_sad64x32x4d_bits8)
1123
1124         HIGHBD_BFP(BLOCK_32X64, vpx_highbd_sad32x64_bits8,
1125                    vpx_highbd_sad32x64_avg_bits8, vpx_highbd_8_variance32x64,
1126                    vpx_highbd_8_sub_pixel_variance32x64,
1127                    vpx_highbd_8_sub_pixel_avg_variance32x64, NULL, NULL,
1128                    vpx_highbd_sad32x64x4d_bits8)
1129
1130         HIGHBD_BFP(BLOCK_32X32, vpx_highbd_sad32x32_bits8,
1131                    vpx_highbd_sad32x32_avg_bits8, vpx_highbd_8_variance32x32,
1132                    vpx_highbd_8_sub_pixel_variance32x32,
1133                    vpx_highbd_8_sub_pixel_avg_variance32x32,
1134                    vpx_highbd_sad32x32x3_bits8, vpx_highbd_sad32x32x8_bits8,
1135                    vpx_highbd_sad32x32x4d_bits8)
1136
1137         HIGHBD_BFP(BLOCK_64X64, vpx_highbd_sad64x64_bits8,
1138                    vpx_highbd_sad64x64_avg_bits8, vpx_highbd_8_variance64x64,
1139                    vpx_highbd_8_sub_pixel_variance64x64,
1140                    vpx_highbd_8_sub_pixel_avg_variance64x64,
1141                    vpx_highbd_sad64x64x3_bits8, vpx_highbd_sad64x64x8_bits8,
1142                    vpx_highbd_sad64x64x4d_bits8)
1143
1144         HIGHBD_BFP(BLOCK_16X16, vpx_highbd_sad16x16_bits8,
1145                    vpx_highbd_sad16x16_avg_bits8, vpx_highbd_8_variance16x16,
1146                    vpx_highbd_8_sub_pixel_variance16x16,
1147                    vpx_highbd_8_sub_pixel_avg_variance16x16,
1148                    vpx_highbd_sad16x16x3_bits8, vpx_highbd_sad16x16x8_bits8,
1149                    vpx_highbd_sad16x16x4d_bits8)
1150
1151         HIGHBD_BFP(
1152             BLOCK_16X8, vpx_highbd_sad16x8_bits8, vpx_highbd_sad16x8_avg_bits8,
1153             vpx_highbd_8_variance16x8, vpx_highbd_8_sub_pixel_variance16x8,
1154             vpx_highbd_8_sub_pixel_avg_variance16x8, vpx_highbd_sad16x8x3_bits8,
1155             vpx_highbd_sad16x8x8_bits8, vpx_highbd_sad16x8x4d_bits8)
1156
1157         HIGHBD_BFP(
1158             BLOCK_8X16, vpx_highbd_sad8x16_bits8, vpx_highbd_sad8x16_avg_bits8,
1159             vpx_highbd_8_variance8x16, vpx_highbd_8_sub_pixel_variance8x16,
1160             vpx_highbd_8_sub_pixel_avg_variance8x16, vpx_highbd_sad8x16x3_bits8,
1161             vpx_highbd_sad8x16x8_bits8, vpx_highbd_sad8x16x4d_bits8)
1162
1163         HIGHBD_BFP(
1164             BLOCK_8X8, vpx_highbd_sad8x8_bits8, vpx_highbd_sad8x8_avg_bits8,
1165             vpx_highbd_8_variance8x8, vpx_highbd_8_sub_pixel_variance8x8,
1166             vpx_highbd_8_sub_pixel_avg_variance8x8, vpx_highbd_sad8x8x3_bits8,
1167             vpx_highbd_sad8x8x8_bits8, vpx_highbd_sad8x8x4d_bits8)
1168
1169         HIGHBD_BFP(BLOCK_8X4, vpx_highbd_sad8x4_bits8,
1170                    vpx_highbd_sad8x4_avg_bits8, vpx_highbd_8_variance8x4,
1171                    vpx_highbd_8_sub_pixel_variance8x4,
1172                    vpx_highbd_8_sub_pixel_avg_variance8x4, NULL,
1173                    vpx_highbd_sad8x4x8_bits8, vpx_highbd_sad8x4x4d_bits8)
1174
1175         HIGHBD_BFP(BLOCK_4X8, vpx_highbd_sad4x8_bits8,
1176                    vpx_highbd_sad4x8_avg_bits8, vpx_highbd_8_variance4x8,
1177                    vpx_highbd_8_sub_pixel_variance4x8,
1178                    vpx_highbd_8_sub_pixel_avg_variance4x8, NULL,
1179                    vpx_highbd_sad4x8x8_bits8, vpx_highbd_sad4x8x4d_bits8)
1180
1181         HIGHBD_BFP(
1182             BLOCK_4X4, vpx_highbd_sad4x4_bits8, vpx_highbd_sad4x4_avg_bits8,
1183             vpx_highbd_8_variance4x4, vpx_highbd_8_sub_pixel_variance4x4,
1184             vpx_highbd_8_sub_pixel_avg_variance4x4, vpx_highbd_sad4x4x3_bits8,
1185             vpx_highbd_sad4x4x8_bits8, vpx_highbd_sad4x4x4d_bits8)
1186         break;
1187
1188       case VPX_BITS_10:
1189         HIGHBD_BFP(BLOCK_32X16, vpx_highbd_sad32x16_bits10,
1190                    vpx_highbd_sad32x16_avg_bits10, vpx_highbd_10_variance32x16,
1191                    vpx_highbd_10_sub_pixel_variance32x16,
1192                    vpx_highbd_10_sub_pixel_avg_variance32x16, NULL, NULL,
1193                    vpx_highbd_sad32x16x4d_bits10)
1194
1195         HIGHBD_BFP(BLOCK_16X32, vpx_highbd_sad16x32_bits10,
1196                    vpx_highbd_sad16x32_avg_bits10, vpx_highbd_10_variance16x32,
1197                    vpx_highbd_10_sub_pixel_variance16x32,
1198                    vpx_highbd_10_sub_pixel_avg_variance16x32, NULL, NULL,
1199                    vpx_highbd_sad16x32x4d_bits10)
1200
1201         HIGHBD_BFP(BLOCK_64X32, vpx_highbd_sad64x32_bits10,
1202                    vpx_highbd_sad64x32_avg_bits10, vpx_highbd_10_variance64x32,
1203                    vpx_highbd_10_sub_pixel_variance64x32,
1204                    vpx_highbd_10_sub_pixel_avg_variance64x32, NULL, NULL,
1205                    vpx_highbd_sad64x32x4d_bits10)
1206
1207         HIGHBD_BFP(BLOCK_32X64, vpx_highbd_sad32x64_bits10,
1208                    vpx_highbd_sad32x64_avg_bits10, vpx_highbd_10_variance32x64,
1209                    vpx_highbd_10_sub_pixel_variance32x64,
1210                    vpx_highbd_10_sub_pixel_avg_variance32x64, NULL, NULL,
1211                    vpx_highbd_sad32x64x4d_bits10)
1212
1213         HIGHBD_BFP(BLOCK_32X32, vpx_highbd_sad32x32_bits10,
1214                    vpx_highbd_sad32x32_avg_bits10, vpx_highbd_10_variance32x32,
1215                    vpx_highbd_10_sub_pixel_variance32x32,
1216                    vpx_highbd_10_sub_pixel_avg_variance32x32,
1217                    vpx_highbd_sad32x32x3_bits10, vpx_highbd_sad32x32x8_bits10,
1218                    vpx_highbd_sad32x32x4d_bits10)
1219
1220         HIGHBD_BFP(BLOCK_64X64, vpx_highbd_sad64x64_bits10,
1221                    vpx_highbd_sad64x64_avg_bits10, vpx_highbd_10_variance64x64,
1222                    vpx_highbd_10_sub_pixel_variance64x64,
1223                    vpx_highbd_10_sub_pixel_avg_variance64x64,
1224                    vpx_highbd_sad64x64x3_bits10, vpx_highbd_sad64x64x8_bits10,
1225                    vpx_highbd_sad64x64x4d_bits10)
1226
1227         HIGHBD_BFP(BLOCK_16X16, vpx_highbd_sad16x16_bits10,
1228                    vpx_highbd_sad16x16_avg_bits10, vpx_highbd_10_variance16x16,
1229                    vpx_highbd_10_sub_pixel_variance16x16,
1230                    vpx_highbd_10_sub_pixel_avg_variance16x16,
1231                    vpx_highbd_sad16x16x3_bits10, vpx_highbd_sad16x16x8_bits10,
1232                    vpx_highbd_sad16x16x4d_bits10)
1233
1234         HIGHBD_BFP(BLOCK_16X8, vpx_highbd_sad16x8_bits10,
1235                    vpx_highbd_sad16x8_avg_bits10, vpx_highbd_10_variance16x8,
1236                    vpx_highbd_10_sub_pixel_variance16x8,
1237                    vpx_highbd_10_sub_pixel_avg_variance16x8,
1238                    vpx_highbd_sad16x8x3_bits10, vpx_highbd_sad16x8x8_bits10,
1239                    vpx_highbd_sad16x8x4d_bits10)
1240
1241         HIGHBD_BFP(BLOCK_8X16, vpx_highbd_sad8x16_bits10,
1242                    vpx_highbd_sad8x16_avg_bits10, vpx_highbd_10_variance8x16,
1243                    vpx_highbd_10_sub_pixel_variance8x16,
1244                    vpx_highbd_10_sub_pixel_avg_variance8x16,
1245                    vpx_highbd_sad8x16x3_bits10, vpx_highbd_sad8x16x8_bits10,
1246                    vpx_highbd_sad8x16x4d_bits10)
1247
1248         HIGHBD_BFP(
1249             BLOCK_8X8, vpx_highbd_sad8x8_bits10, vpx_highbd_sad8x8_avg_bits10,
1250             vpx_highbd_10_variance8x8, vpx_highbd_10_sub_pixel_variance8x8,
1251             vpx_highbd_10_sub_pixel_avg_variance8x8, vpx_highbd_sad8x8x3_bits10,
1252             vpx_highbd_sad8x8x8_bits10, vpx_highbd_sad8x8x4d_bits10)
1253
1254         HIGHBD_BFP(BLOCK_8X4, vpx_highbd_sad8x4_bits10,
1255                    vpx_highbd_sad8x4_avg_bits10, vpx_highbd_10_variance8x4,
1256                    vpx_highbd_10_sub_pixel_variance8x4,
1257                    vpx_highbd_10_sub_pixel_avg_variance8x4, NULL,
1258                    vpx_highbd_sad8x4x8_bits10, vpx_highbd_sad8x4x4d_bits10)
1259
1260         HIGHBD_BFP(BLOCK_4X8, vpx_highbd_sad4x8_bits10,
1261                    vpx_highbd_sad4x8_avg_bits10, vpx_highbd_10_variance4x8,
1262                    vpx_highbd_10_sub_pixel_variance4x8,
1263                    vpx_highbd_10_sub_pixel_avg_variance4x8, NULL,
1264                    vpx_highbd_sad4x8x8_bits10, vpx_highbd_sad4x8x4d_bits10)
1265
1266         HIGHBD_BFP(
1267             BLOCK_4X4, vpx_highbd_sad4x4_bits10, vpx_highbd_sad4x4_avg_bits10,
1268             vpx_highbd_10_variance4x4, vpx_highbd_10_sub_pixel_variance4x4,
1269             vpx_highbd_10_sub_pixel_avg_variance4x4, vpx_highbd_sad4x4x3_bits10,
1270             vpx_highbd_sad4x4x8_bits10, vpx_highbd_sad4x4x4d_bits10)
1271         break;
1272
1273       case VPX_BITS_12:
1274         HIGHBD_BFP(BLOCK_32X16, vpx_highbd_sad32x16_bits12,
1275                    vpx_highbd_sad32x16_avg_bits12, vpx_highbd_12_variance32x16,
1276                    vpx_highbd_12_sub_pixel_variance32x16,
1277                    vpx_highbd_12_sub_pixel_avg_variance32x16, NULL, NULL,
1278                    vpx_highbd_sad32x16x4d_bits12)
1279
1280         HIGHBD_BFP(BLOCK_16X32, vpx_highbd_sad16x32_bits12,
1281                    vpx_highbd_sad16x32_avg_bits12, vpx_highbd_12_variance16x32,
1282                    vpx_highbd_12_sub_pixel_variance16x32,
1283                    vpx_highbd_12_sub_pixel_avg_variance16x32, NULL, NULL,
1284                    vpx_highbd_sad16x32x4d_bits12)
1285
1286         HIGHBD_BFP(BLOCK_64X32, vpx_highbd_sad64x32_bits12,
1287                    vpx_highbd_sad64x32_avg_bits12, vpx_highbd_12_variance64x32,
1288                    vpx_highbd_12_sub_pixel_variance64x32,
1289                    vpx_highbd_12_sub_pixel_avg_variance64x32, NULL, NULL,
1290                    vpx_highbd_sad64x32x4d_bits12)
1291
1292         HIGHBD_BFP(BLOCK_32X64, vpx_highbd_sad32x64_bits12,
1293                    vpx_highbd_sad32x64_avg_bits12, vpx_highbd_12_variance32x64,
1294                    vpx_highbd_12_sub_pixel_variance32x64,
1295                    vpx_highbd_12_sub_pixel_avg_variance32x64, NULL, NULL,
1296                    vpx_highbd_sad32x64x4d_bits12)
1297
1298         HIGHBD_BFP(BLOCK_32X32, vpx_highbd_sad32x32_bits12,
1299                    vpx_highbd_sad32x32_avg_bits12, vpx_highbd_12_variance32x32,
1300                    vpx_highbd_12_sub_pixel_variance32x32,
1301                    vpx_highbd_12_sub_pixel_avg_variance32x32,
1302                    vpx_highbd_sad32x32x3_bits12, vpx_highbd_sad32x32x8_bits12,
1303                    vpx_highbd_sad32x32x4d_bits12)
1304
1305         HIGHBD_BFP(BLOCK_64X64, vpx_highbd_sad64x64_bits12,
1306                    vpx_highbd_sad64x64_avg_bits12, vpx_highbd_12_variance64x64,
1307                    vpx_highbd_12_sub_pixel_variance64x64,
1308                    vpx_highbd_12_sub_pixel_avg_variance64x64,
1309                    vpx_highbd_sad64x64x3_bits12, vpx_highbd_sad64x64x8_bits12,
1310                    vpx_highbd_sad64x64x4d_bits12)
1311
1312         HIGHBD_BFP(BLOCK_16X16, vpx_highbd_sad16x16_bits12,
1313                    vpx_highbd_sad16x16_avg_bits12, vpx_highbd_12_variance16x16,
1314                    vpx_highbd_12_sub_pixel_variance16x16,
1315                    vpx_highbd_12_sub_pixel_avg_variance16x16,
1316                    vpx_highbd_sad16x16x3_bits12, vpx_highbd_sad16x16x8_bits12,
1317                    vpx_highbd_sad16x16x4d_bits12)
1318
1319         HIGHBD_BFP(BLOCK_16X8, vpx_highbd_sad16x8_bits12,
1320                    vpx_highbd_sad16x8_avg_bits12, vpx_highbd_12_variance16x8,
1321                    vpx_highbd_12_sub_pixel_variance16x8,
1322                    vpx_highbd_12_sub_pixel_avg_variance16x8,
1323                    vpx_highbd_sad16x8x3_bits12, vpx_highbd_sad16x8x8_bits12,
1324                    vpx_highbd_sad16x8x4d_bits12)
1325
1326         HIGHBD_BFP(BLOCK_8X16, vpx_highbd_sad8x16_bits12,
1327                    vpx_highbd_sad8x16_avg_bits12, vpx_highbd_12_variance8x16,
1328                    vpx_highbd_12_sub_pixel_variance8x16,
1329                    vpx_highbd_12_sub_pixel_avg_variance8x16,
1330                    vpx_highbd_sad8x16x3_bits12, vpx_highbd_sad8x16x8_bits12,
1331                    vpx_highbd_sad8x16x4d_bits12)
1332
1333         HIGHBD_BFP(
1334             BLOCK_8X8, vpx_highbd_sad8x8_bits12, vpx_highbd_sad8x8_avg_bits12,
1335             vpx_highbd_12_variance8x8, vpx_highbd_12_sub_pixel_variance8x8,
1336             vpx_highbd_12_sub_pixel_avg_variance8x8, vpx_highbd_sad8x8x3_bits12,
1337             vpx_highbd_sad8x8x8_bits12, vpx_highbd_sad8x8x4d_bits12)
1338
1339         HIGHBD_BFP(BLOCK_8X4, vpx_highbd_sad8x4_bits12,
1340                    vpx_highbd_sad8x4_avg_bits12, vpx_highbd_12_variance8x4,
1341                    vpx_highbd_12_sub_pixel_variance8x4,
1342                    vpx_highbd_12_sub_pixel_avg_variance8x4, NULL,
1343                    vpx_highbd_sad8x4x8_bits12, vpx_highbd_sad8x4x4d_bits12)
1344
1345         HIGHBD_BFP(BLOCK_4X8, vpx_highbd_sad4x8_bits12,
1346                    vpx_highbd_sad4x8_avg_bits12, vpx_highbd_12_variance4x8,
1347                    vpx_highbd_12_sub_pixel_variance4x8,
1348                    vpx_highbd_12_sub_pixel_avg_variance4x8, NULL,
1349                    vpx_highbd_sad4x8x8_bits12, vpx_highbd_sad4x8x4d_bits12)
1350
1351         HIGHBD_BFP(
1352             BLOCK_4X4, vpx_highbd_sad4x4_bits12, vpx_highbd_sad4x4_avg_bits12,
1353             vpx_highbd_12_variance4x4, vpx_highbd_12_sub_pixel_variance4x4,
1354             vpx_highbd_12_sub_pixel_avg_variance4x4, vpx_highbd_sad4x4x3_bits12,
1355             vpx_highbd_sad4x4x8_bits12, vpx_highbd_sad4x4x4d_bits12)
1356         break;
1357
1358       default:
1359         assert(0 &&
1360                "cm->bit_depth should be VPX_BITS_8, "
1361                "VPX_BITS_10 or VPX_BITS_12");
1362     }
1363   }
1364 }
1365 #endif  // CONFIG_VP9_HIGHBITDEPTH
1366
1367 static void realloc_segmentation_maps(VP9_COMP *cpi) {
1368   VP9_COMMON *const cm = &cpi->common;
1369
1370   // Create the encoder segmentation map and set all entries to 0
1371   vpx_free(cpi->segmentation_map);
1372   CHECK_MEM_ERROR(cm, cpi->segmentation_map,
1373                   vpx_calloc(cm->mi_rows * cm->mi_cols, 1));
1374
1375   // Create a map used for cyclic background refresh.
1376   if (cpi->cyclic_refresh) vp9_cyclic_refresh_free(cpi->cyclic_refresh);
1377   CHECK_MEM_ERROR(cm, cpi->cyclic_refresh,
1378                   vp9_cyclic_refresh_alloc(cm->mi_rows, cm->mi_cols));
1379
1380   // Create a map used to mark inactive areas.
1381   vpx_free(cpi->active_map.map);
1382   CHECK_MEM_ERROR(cm, cpi->active_map.map,
1383                   vpx_calloc(cm->mi_rows * cm->mi_cols, 1));
1384
1385   // And a place holder structure is the coding context
1386   // for use if we want to save and restore it
1387   vpx_free(cpi->coding_context.last_frame_seg_map_copy);
1388   CHECK_MEM_ERROR(cm, cpi->coding_context.last_frame_seg_map_copy,
1389                   vpx_calloc(cm->mi_rows * cm->mi_cols, 1));
1390 }
1391
1392 void vp9_change_config(struct VP9_COMP *cpi, const VP9EncoderConfig *oxcf) {
1393   VP9_COMMON *const cm = &cpi->common;
1394   RATE_CONTROL *const rc = &cpi->rc;
1395   int last_w = cpi->oxcf.width;
1396   int last_h = cpi->oxcf.height;
1397
1398   if (cm->profile != oxcf->profile) cm->profile = oxcf->profile;
1399   cm->bit_depth = oxcf->bit_depth;
1400   cm->color_space = oxcf->color_space;
1401   cm->color_range = oxcf->color_range;
1402
1403   cpi->target_level = oxcf->target_level;
1404   cpi->keep_level_stats = oxcf->target_level != LEVEL_MAX;
1405
1406   if (cm->profile <= PROFILE_1)
1407     assert(cm->bit_depth == VPX_BITS_8);
1408   else
1409     assert(cm->bit_depth > VPX_BITS_8);
1410
1411   cpi->oxcf = *oxcf;
1412 #if CONFIG_VP9_HIGHBITDEPTH
1413   cpi->td.mb.e_mbd.bd = (int)cm->bit_depth;
1414 #endif  // CONFIG_VP9_HIGHBITDEPTH
1415
1416   if ((oxcf->pass == 0) && (oxcf->rc_mode == VPX_Q)) {
1417     rc->baseline_gf_interval = FIXED_GF_INTERVAL;
1418   } else {
1419     rc->baseline_gf_interval = (MIN_GF_INTERVAL + MAX_GF_INTERVAL) / 2;
1420   }
1421
1422   cpi->refresh_golden_frame = 0;
1423   cpi->refresh_last_frame = 1;
1424   cm->refresh_frame_context = 1;
1425   cm->reset_frame_context = 0;
1426
1427   vp9_reset_segment_features(&cm->seg);
1428   vp9_set_high_precision_mv(cpi, 0);
1429
1430   {
1431     int i;
1432
1433     for (i = 0; i < MAX_SEGMENTS; i++)
1434       cpi->segment_encode_breakout[i] = cpi->oxcf.encode_breakout;
1435   }
1436   cpi->encode_breakout = cpi->oxcf.encode_breakout;
1437
1438   set_rc_buffer_sizes(rc, &cpi->oxcf);
1439
1440   // Under a configuration change, where maximum_buffer_size may change,
1441   // keep buffer level clipped to the maximum allowed buffer size.
1442   rc->bits_off_target = VPXMIN(rc->bits_off_target, rc->maximum_buffer_size);
1443   rc->buffer_level = VPXMIN(rc->buffer_level, rc->maximum_buffer_size);
1444
1445   // Set up frame rate and related parameters rate control values.
1446   vp9_new_framerate(cpi, cpi->framerate);
1447
1448   // Set absolute upper and lower quality limits
1449   rc->worst_quality = cpi->oxcf.worst_allowed_q;
1450   rc->best_quality = cpi->oxcf.best_allowed_q;
1451
1452   cm->interp_filter = cpi->sf.default_interp_filter;
1453
1454   if (cpi->oxcf.render_width > 0 && cpi->oxcf.render_height > 0) {
1455     cm->render_width = cpi->oxcf.render_width;
1456     cm->render_height = cpi->oxcf.render_height;
1457   } else {
1458     cm->render_width = cpi->oxcf.width;
1459     cm->render_height = cpi->oxcf.height;
1460   }
1461   if (last_w != cpi->oxcf.width || last_h != cpi->oxcf.height) {
1462     cm->width = cpi->oxcf.width;
1463     cm->height = cpi->oxcf.height;
1464     cpi->external_resize = 1;
1465   }
1466
1467   if (cpi->initial_width) {
1468     int new_mi_size = 0;
1469     vp9_set_mb_mi(cm, cm->width, cm->height);
1470     new_mi_size = cm->mi_stride * calc_mi_size(cm->mi_rows);
1471     if (cm->mi_alloc_size < new_mi_size) {
1472       vp9_free_context_buffers(cm);
1473       alloc_compressor_data(cpi);
1474       realloc_segmentation_maps(cpi);
1475       cpi->initial_width = cpi->initial_height = 0;
1476       cpi->external_resize = 0;
1477     } else if (cm->mi_alloc_size == new_mi_size &&
1478                (cpi->oxcf.width > last_w || cpi->oxcf.height > last_h)) {
1479       vp9_alloc_loop_filter(cm);
1480     }
1481   }
1482
1483   update_frame_size(cpi);
1484
1485   if (last_w != cpi->oxcf.width || last_h != cpi->oxcf.height) {
1486     memset(cpi->consec_zero_mv, 0,
1487            cm->mi_rows * cm->mi_cols * sizeof(*cpi->consec_zero_mv));
1488     if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ)
1489       vp9_cyclic_refresh_reset_resize(cpi);
1490   }
1491
1492   if ((cpi->svc.number_temporal_layers > 1 && cpi->oxcf.rc_mode == VPX_CBR) ||
1493       ((cpi->svc.number_temporal_layers > 1 ||
1494         cpi->svc.number_spatial_layers > 1) &&
1495        cpi->oxcf.pass != 1)) {
1496     vp9_update_layer_context_change_config(cpi,
1497                                            (int)cpi->oxcf.target_bandwidth);
1498   }
1499
1500   cpi->alt_ref_source = NULL;
1501   rc->is_src_frame_alt_ref = 0;
1502
1503 #if 0
1504   // Experimental RD Code
1505   cpi->frame_distortion = 0;
1506   cpi->last_frame_distortion = 0;
1507 #endif
1508
1509   set_tile_limits(cpi);
1510
1511   cpi->ext_refresh_frame_flags_pending = 0;
1512   cpi->ext_refresh_frame_context_pending = 0;
1513
1514 #if CONFIG_VP9_HIGHBITDEPTH
1515   highbd_set_var_fns(cpi);
1516 #endif
1517 }
1518
1519 #ifndef M_LOG2_E
1520 #define M_LOG2_E 0.693147180559945309417
1521 #endif
1522 #define log2f(x) (log(x) / (float)M_LOG2_E)
1523
1524 /***********************************************************************
1525  * Read before modifying 'cal_nmvjointsadcost' or 'cal_nmvsadcosts'    *
1526  ***********************************************************************
1527  * The following 2 functions ('cal_nmvjointsadcost' and                *
1528  * 'cal_nmvsadcosts') are used to calculate cost lookup tables         *
1529  * used by 'vp9_diamond_search_sad'. The C implementation of the       *
1530  * function is generic, but the AVX intrinsics optimised version       *
1531  * relies on the following properties of the computed tables:          *
1532  * For cal_nmvjointsadcost:                                            *
1533  *   - mvjointsadcost[1] == mvjointsadcost[2] == mvjointsadcost[3]     *
1534  * For cal_nmvsadcosts:                                                *
1535  *   - For all i: mvsadcost[0][i] == mvsadcost[1][i]                   *
1536  *         (Equal costs for both components)                           *
1537  *   - For all i: mvsadcost[0][i] == mvsadcost[0][-i]                  *
1538  *         (Cost function is even)                                     *
1539  * If these do not hold, then the AVX optimised version of the         *
1540  * 'vp9_diamond_search_sad' function cannot be used as it is, in which *
1541  * case you can revert to using the C function instead.                *
1542  ***********************************************************************/
1543
1544 static void cal_nmvjointsadcost(int *mvjointsadcost) {
1545   /*********************************************************************
1546    * Warning: Read the comments above before modifying this function   *
1547    *********************************************************************/
1548   mvjointsadcost[0] = 600;
1549   mvjointsadcost[1] = 300;
1550   mvjointsadcost[2] = 300;
1551   mvjointsadcost[3] = 300;
1552 }
1553
1554 static void cal_nmvsadcosts(int *mvsadcost[2]) {
1555   /*********************************************************************
1556    * Warning: Read the comments above before modifying this function   *
1557    *********************************************************************/
1558   int i = 1;
1559
1560   mvsadcost[0][0] = 0;
1561   mvsadcost[1][0] = 0;
1562
1563   do {
1564     double z = 256 * (2 * (log2f(8 * i) + .6));
1565     mvsadcost[0][i] = (int)z;
1566     mvsadcost[1][i] = (int)z;
1567     mvsadcost[0][-i] = (int)z;
1568     mvsadcost[1][-i] = (int)z;
1569   } while (++i <= MV_MAX);
1570 }
1571
1572 static void cal_nmvsadcosts_hp(int *mvsadcost[2]) {
1573   int i = 1;
1574
1575   mvsadcost[0][0] = 0;
1576   mvsadcost[1][0] = 0;
1577
1578   do {
1579     double z = 256 * (2 * (log2f(8 * i) + .6));
1580     mvsadcost[0][i] = (int)z;
1581     mvsadcost[1][i] = (int)z;
1582     mvsadcost[0][-i] = (int)z;
1583     mvsadcost[1][-i] = (int)z;
1584   } while (++i <= MV_MAX);
1585 }
1586
1587 VP9_COMP *vp9_create_compressor(VP9EncoderConfig *oxcf,
1588                                 BufferPool *const pool) {
1589   unsigned int i;
1590   VP9_COMP *volatile const cpi = vpx_memalign(32, sizeof(VP9_COMP));
1591   VP9_COMMON *volatile const cm = cpi != NULL ? &cpi->common : NULL;
1592
1593   if (!cm) return NULL;
1594
1595   vp9_zero(*cpi);
1596
1597   if (setjmp(cm->error.jmp)) {
1598     cm->error.setjmp = 0;
1599     vp9_remove_compressor(cpi);
1600     return 0;
1601   }
1602
1603   cm->error.setjmp = 1;
1604   cm->alloc_mi = vp9_enc_alloc_mi;
1605   cm->free_mi = vp9_enc_free_mi;
1606   cm->setup_mi = vp9_enc_setup_mi;
1607
1608   CHECK_MEM_ERROR(cm, cm->fc, (FRAME_CONTEXT *)vpx_calloc(1, sizeof(*cm->fc)));
1609   CHECK_MEM_ERROR(
1610       cm, cm->frame_contexts,
1611       (FRAME_CONTEXT *)vpx_calloc(FRAME_CONTEXTS, sizeof(*cm->frame_contexts)));
1612
1613   cpi->use_svc = 0;
1614   cpi->resize_state = 0;
1615   cpi->external_resize = 0;
1616   cpi->resize_avg_qp = 0;
1617   cpi->resize_buffer_underflow = 0;
1618   cpi->use_skin_detection = 0;
1619   cpi->common.buffer_pool = pool;
1620
1621   init_config(cpi, oxcf);
1622   vp9_rc_init(&cpi->oxcf, oxcf->pass, &cpi->rc);
1623
1624   cm->current_video_frame = 0;
1625   cpi->partition_search_skippable_frame = 0;
1626   cpi->tile_data = NULL;
1627
1628   realloc_segmentation_maps(cpi);
1629
1630   CHECK_MEM_ERROR(
1631       cm, cpi->consec_zero_mv,
1632       vpx_calloc(cm->mi_rows * cm->mi_cols, sizeof(*cpi->consec_zero_mv)));
1633
1634   CHECK_MEM_ERROR(cm, cpi->nmvcosts[0],
1635                   vpx_calloc(MV_VALS, sizeof(*cpi->nmvcosts[0])));
1636   CHECK_MEM_ERROR(cm, cpi->nmvcosts[1],
1637                   vpx_calloc(MV_VALS, sizeof(*cpi->nmvcosts[1])));
1638   CHECK_MEM_ERROR(cm, cpi->nmvcosts_hp[0],
1639                   vpx_calloc(MV_VALS, sizeof(*cpi->nmvcosts_hp[0])));
1640   CHECK_MEM_ERROR(cm, cpi->nmvcosts_hp[1],
1641                   vpx_calloc(MV_VALS, sizeof(*cpi->nmvcosts_hp[1])));
1642   CHECK_MEM_ERROR(cm, cpi->nmvsadcosts[0],
1643                   vpx_calloc(MV_VALS, sizeof(*cpi->nmvsadcosts[0])));
1644   CHECK_MEM_ERROR(cm, cpi->nmvsadcosts[1],
1645                   vpx_calloc(MV_VALS, sizeof(*cpi->nmvsadcosts[1])));
1646   CHECK_MEM_ERROR(cm, cpi->nmvsadcosts_hp[0],
1647                   vpx_calloc(MV_VALS, sizeof(*cpi->nmvsadcosts_hp[0])));
1648   CHECK_MEM_ERROR(cm, cpi->nmvsadcosts_hp[1],
1649                   vpx_calloc(MV_VALS, sizeof(*cpi->nmvsadcosts_hp[1])));
1650
1651   for (i = 0; i < (sizeof(cpi->mbgraph_stats) / sizeof(cpi->mbgraph_stats[0]));
1652        i++) {
1653     CHECK_MEM_ERROR(
1654         cm, cpi->mbgraph_stats[i].mb_stats,
1655         vpx_calloc(cm->MBs * sizeof(*cpi->mbgraph_stats[i].mb_stats), 1));
1656   }
1657
1658 #if CONFIG_FP_MB_STATS
1659   cpi->use_fp_mb_stats = 0;
1660   if (cpi->use_fp_mb_stats) {
1661     // a place holder used to store the first pass mb stats in the first pass
1662     CHECK_MEM_ERROR(cm, cpi->twopass.frame_mb_stats_buf,
1663                     vpx_calloc(cm->MBs * sizeof(uint8_t), 1));
1664   } else {
1665     cpi->twopass.frame_mb_stats_buf = NULL;
1666   }
1667 #endif
1668
1669   cpi->refresh_alt_ref_frame = 0;
1670   cpi->multi_arf_last_grp_enabled = 0;
1671
1672   cpi->b_calculate_psnr = CONFIG_INTERNAL_STATS;
1673
1674   init_level_info(&cpi->level_info);
1675
1676 #if CONFIG_INTERNAL_STATS
1677   cpi->b_calculate_blockiness = 1;
1678   cpi->b_calculate_consistency = 1;
1679   cpi->total_inconsistency = 0;
1680   cpi->psnr.worst = 100.0;
1681   cpi->worst_ssim = 100.0;
1682
1683   cpi->count = 0;
1684   cpi->bytes = 0;
1685
1686   if (cpi->b_calculate_psnr) {
1687     cpi->total_sq_error = 0;
1688     cpi->total_samples = 0;
1689
1690     cpi->totalp_sq_error = 0;
1691     cpi->totalp_samples = 0;
1692
1693     cpi->tot_recode_hits = 0;
1694     cpi->summed_quality = 0;
1695     cpi->summed_weights = 0;
1696     cpi->summedp_quality = 0;
1697     cpi->summedp_weights = 0;
1698   }
1699
1700   cpi->fastssim.worst = 100.0;
1701
1702   cpi->psnrhvs.worst = 100.0;
1703
1704   if (cpi->b_calculate_blockiness) {
1705     cpi->total_blockiness = 0;
1706     cpi->worst_blockiness = 0.0;
1707   }
1708
1709   if (cpi->b_calculate_consistency) {
1710     CHECK_MEM_ERROR(cm, cpi->ssim_vars,
1711                     vpx_malloc(sizeof(*cpi->ssim_vars) * 4 *
1712                                cpi->common.mi_rows * cpi->common.mi_cols));
1713     cpi->worst_consistency = 100.0;
1714   }
1715
1716 #endif
1717
1718   cpi->first_time_stamp_ever = INT64_MAX;
1719
1720   /*********************************************************************
1721    * Warning: Read the comments around 'cal_nmvjointsadcost' and       *
1722    * 'cal_nmvsadcosts' before modifying how these tables are computed. *
1723    *********************************************************************/
1724   cal_nmvjointsadcost(cpi->td.mb.nmvjointsadcost);
1725   cpi->td.mb.nmvcost[0] = &cpi->nmvcosts[0][MV_MAX];
1726   cpi->td.mb.nmvcost[1] = &cpi->nmvcosts[1][MV_MAX];
1727   cpi->td.mb.nmvsadcost[0] = &cpi->nmvsadcosts[0][MV_MAX];
1728   cpi->td.mb.nmvsadcost[1] = &cpi->nmvsadcosts[1][MV_MAX];
1729   cal_nmvsadcosts(cpi->td.mb.nmvsadcost);
1730
1731   cpi->td.mb.nmvcost_hp[0] = &cpi->nmvcosts_hp[0][MV_MAX];
1732   cpi->td.mb.nmvcost_hp[1] = &cpi->nmvcosts_hp[1][MV_MAX];
1733   cpi->td.mb.nmvsadcost_hp[0] = &cpi->nmvsadcosts_hp[0][MV_MAX];
1734   cpi->td.mb.nmvsadcost_hp[1] = &cpi->nmvsadcosts_hp[1][MV_MAX];
1735   cal_nmvsadcosts_hp(cpi->td.mb.nmvsadcost_hp);
1736
1737 #if CONFIG_VP9_TEMPORAL_DENOISING
1738 #ifdef OUTPUT_YUV_DENOISED
1739   yuv_denoised_file = fopen("denoised.yuv", "ab");
1740 #endif
1741 #endif
1742 #ifdef OUTPUT_YUV_SKINMAP
1743   yuv_skinmap_file = fopen("skinmap.yuv", "ab");
1744 #endif
1745 #ifdef OUTPUT_YUV_REC
1746   yuv_rec_file = fopen("rec.yuv", "wb");
1747 #endif
1748
1749 #if 0
1750   framepsnr = fopen("framepsnr.stt", "a");
1751   kf_list = fopen("kf_list.stt", "w");
1752 #endif
1753
1754   cpi->allow_encode_breakout = ENCODE_BREAKOUT_ENABLED;
1755
1756   if (oxcf->pass == 1) {
1757     vp9_init_first_pass(cpi);
1758   } else if (oxcf->pass == 2) {
1759     const size_t packet_sz = sizeof(FIRSTPASS_STATS);
1760     const int packets = (int)(oxcf->two_pass_stats_in.sz / packet_sz);
1761
1762     if (cpi->svc.number_spatial_layers > 1 ||
1763         cpi->svc.number_temporal_layers > 1) {
1764       FIRSTPASS_STATS *const stats = oxcf->two_pass_stats_in.buf;
1765       FIRSTPASS_STATS *stats_copy[VPX_SS_MAX_LAYERS] = { 0 };
1766       int i;
1767
1768       for (i = 0; i < oxcf->ss_number_layers; ++i) {
1769         FIRSTPASS_STATS *const last_packet_for_layer =
1770             &stats[packets - oxcf->ss_number_layers + i];
1771         const int layer_id = (int)last_packet_for_layer->spatial_layer_id;
1772         const int packets_in_layer = (int)last_packet_for_layer->count + 1;
1773         if (layer_id >= 0 && layer_id < oxcf->ss_number_layers) {
1774           LAYER_CONTEXT *const lc = &cpi->svc.layer_context[layer_id];
1775
1776           vpx_free(lc->rc_twopass_stats_in.buf);
1777
1778           lc->rc_twopass_stats_in.sz = packets_in_layer * packet_sz;
1779           CHECK_MEM_ERROR(cm, lc->rc_twopass_stats_in.buf,
1780                           vpx_malloc(lc->rc_twopass_stats_in.sz));
1781           lc->twopass.stats_in_start = lc->rc_twopass_stats_in.buf;
1782           lc->twopass.stats_in = lc->twopass.stats_in_start;
1783           lc->twopass.stats_in_end =
1784               lc->twopass.stats_in_start + packets_in_layer - 1;
1785           stats_copy[layer_id] = lc->rc_twopass_stats_in.buf;
1786         }
1787       }
1788
1789       for (i = 0; i < packets; ++i) {
1790         const int layer_id = (int)stats[i].spatial_layer_id;
1791         if (layer_id >= 0 && layer_id < oxcf->ss_number_layers &&
1792             stats_copy[layer_id] != NULL) {
1793           *stats_copy[layer_id] = stats[i];
1794           ++stats_copy[layer_id];
1795         }
1796       }
1797
1798       vp9_init_second_pass_spatial_svc(cpi);
1799     } else {
1800 #if CONFIG_FP_MB_STATS
1801       if (cpi->use_fp_mb_stats) {
1802         const size_t psz = cpi->common.MBs * sizeof(uint8_t);
1803         const int ps = (int)(oxcf->firstpass_mb_stats_in.sz / psz);
1804
1805         cpi->twopass.firstpass_mb_stats.mb_stats_start =
1806             oxcf->firstpass_mb_stats_in.buf;
1807         cpi->twopass.firstpass_mb_stats.mb_stats_end =
1808             cpi->twopass.firstpass_mb_stats.mb_stats_start +
1809             (ps - 1) * cpi->common.MBs * sizeof(uint8_t);
1810       }
1811 #endif
1812
1813       cpi->twopass.stats_in_start = oxcf->two_pass_stats_in.buf;
1814       cpi->twopass.stats_in = cpi->twopass.stats_in_start;
1815       cpi->twopass.stats_in_end = &cpi->twopass.stats_in[packets - 1];
1816
1817       vp9_init_second_pass(cpi);
1818     }
1819   }
1820
1821   vp9_set_speed_features_framesize_independent(cpi);
1822   vp9_set_speed_features_framesize_dependent(cpi);
1823
1824   // Allocate memory to store variances for a frame.
1825   CHECK_MEM_ERROR(cm, cpi->source_diff_var, vpx_calloc(cm->MBs, sizeof(diff)));
1826   cpi->source_var_thresh = 0;
1827   cpi->frames_till_next_var_check = 0;
1828
1829 #define BFP(BT, SDF, SDAF, VF, SVF, SVAF, SDX3F, SDX8F, SDX4DF) \
1830   cpi->fn_ptr[BT].sdf = SDF;                                    \
1831   cpi->fn_ptr[BT].sdaf = SDAF;                                  \
1832   cpi->fn_ptr[BT].vf = VF;                                      \
1833   cpi->fn_ptr[BT].svf = SVF;                                    \
1834   cpi->fn_ptr[BT].svaf = SVAF;                                  \
1835   cpi->fn_ptr[BT].sdx3f = SDX3F;                                \
1836   cpi->fn_ptr[BT].sdx8f = SDX8F;                                \
1837   cpi->fn_ptr[BT].sdx4df = SDX4DF;
1838
1839   BFP(BLOCK_32X16, vpx_sad32x16, vpx_sad32x16_avg, vpx_variance32x16,
1840       vpx_sub_pixel_variance32x16, vpx_sub_pixel_avg_variance32x16, NULL, NULL,
1841       vpx_sad32x16x4d)
1842
1843   BFP(BLOCK_16X32, vpx_sad16x32, vpx_sad16x32_avg, vpx_variance16x32,
1844       vpx_sub_pixel_variance16x32, vpx_sub_pixel_avg_variance16x32, NULL, NULL,
1845       vpx_sad16x32x4d)
1846
1847   BFP(BLOCK_64X32, vpx_sad64x32, vpx_sad64x32_avg, vpx_variance64x32,
1848       vpx_sub_pixel_variance64x32, vpx_sub_pixel_avg_variance64x32, NULL, NULL,
1849       vpx_sad64x32x4d)
1850
1851   BFP(BLOCK_32X64, vpx_sad32x64, vpx_sad32x64_avg, vpx_variance32x64,
1852       vpx_sub_pixel_variance32x64, vpx_sub_pixel_avg_variance32x64, NULL, NULL,
1853       vpx_sad32x64x4d)
1854
1855   BFP(BLOCK_32X32, vpx_sad32x32, vpx_sad32x32_avg, vpx_variance32x32,
1856       vpx_sub_pixel_variance32x32, vpx_sub_pixel_avg_variance32x32,
1857       vpx_sad32x32x3, vpx_sad32x32x8, vpx_sad32x32x4d)
1858
1859   BFP(BLOCK_64X64, vpx_sad64x64, vpx_sad64x64_avg, vpx_variance64x64,
1860       vpx_sub_pixel_variance64x64, vpx_sub_pixel_avg_variance64x64,
1861       vpx_sad64x64x3, vpx_sad64x64x8, vpx_sad64x64x4d)
1862
1863   BFP(BLOCK_16X16, vpx_sad16x16, vpx_sad16x16_avg, vpx_variance16x16,
1864       vpx_sub_pixel_variance16x16, vpx_sub_pixel_avg_variance16x16,
1865       vpx_sad16x16x3, vpx_sad16x16x8, vpx_sad16x16x4d)
1866
1867   BFP(BLOCK_16X8, vpx_sad16x8, vpx_sad16x8_avg, vpx_variance16x8,
1868       vpx_sub_pixel_variance16x8, vpx_sub_pixel_avg_variance16x8, vpx_sad16x8x3,
1869       vpx_sad16x8x8, vpx_sad16x8x4d)
1870
1871   BFP(BLOCK_8X16, vpx_sad8x16, vpx_sad8x16_avg, vpx_variance8x16,
1872       vpx_sub_pixel_variance8x16, vpx_sub_pixel_avg_variance8x16, vpx_sad8x16x3,
1873       vpx_sad8x16x8, vpx_sad8x16x4d)
1874
1875   BFP(BLOCK_8X8, vpx_sad8x8, vpx_sad8x8_avg, vpx_variance8x8,
1876       vpx_sub_pixel_variance8x8, vpx_sub_pixel_avg_variance8x8, vpx_sad8x8x3,
1877       vpx_sad8x8x8, vpx_sad8x8x4d)
1878
1879   BFP(BLOCK_8X4, vpx_sad8x4, vpx_sad8x4_avg, vpx_variance8x4,
1880       vpx_sub_pixel_variance8x4, vpx_sub_pixel_avg_variance8x4, NULL,
1881       vpx_sad8x4x8, vpx_sad8x4x4d)
1882
1883   BFP(BLOCK_4X8, vpx_sad4x8, vpx_sad4x8_avg, vpx_variance4x8,
1884       vpx_sub_pixel_variance4x8, vpx_sub_pixel_avg_variance4x8, NULL,
1885       vpx_sad4x8x8, vpx_sad4x8x4d)
1886
1887   BFP(BLOCK_4X4, vpx_sad4x4, vpx_sad4x4_avg, vpx_variance4x4,
1888       vpx_sub_pixel_variance4x4, vpx_sub_pixel_avg_variance4x4, vpx_sad4x4x3,
1889       vpx_sad4x4x8, vpx_sad4x4x4d)
1890
1891 #if CONFIG_VP9_HIGHBITDEPTH
1892   highbd_set_var_fns(cpi);
1893 #endif
1894
1895   /* vp9_init_quantizer() is first called here. Add check in
1896    * vp9_frame_init_quantizer() so that vp9_init_quantizer is only
1897    * called later when needed. This will avoid unnecessary calls of
1898    * vp9_init_quantizer() for every frame.
1899    */
1900   vp9_init_quantizer(cpi);
1901
1902   vp9_loop_filter_init(cm);
1903
1904   cm->error.setjmp = 0;
1905
1906   return cpi;
1907 }
1908
1909 #if CONFIG_INTERNAL_STATS
1910 #define SNPRINT(H, T) snprintf((H) + strlen(H), sizeof(H) - strlen(H), (T))
1911
1912 #define SNPRINT2(H, T, V) \
1913   snprintf((H) + strlen(H), sizeof(H) - strlen(H), (T), (V))
1914 #endif  // CONFIG_INTERNAL_STATS
1915
1916 void vp9_remove_compressor(VP9_COMP *cpi) {
1917   VP9_COMMON *cm;
1918   unsigned int i;
1919   int t;
1920
1921   if (!cpi) return;
1922
1923   cm = &cpi->common;
1924   if (cm->current_video_frame > 0) {
1925 #if CONFIG_INTERNAL_STATS
1926     vpx_clear_system_state();
1927
1928     if (cpi->oxcf.pass != 1) {
1929       char headings[512] = { 0 };
1930       char results[512] = { 0 };
1931       FILE *f = fopen("opsnr.stt", "a");
1932       double time_encoded =
1933           (cpi->last_end_time_stamp_seen - cpi->first_time_stamp_ever) /
1934           10000000.000;
1935       double total_encode_time =
1936           (cpi->time_receive_data + cpi->time_compress_data) / 1000.000;
1937       const double dr =
1938           (double)cpi->bytes * (double)8 / (double)1000 / time_encoded;
1939       const double peak = (double)((1 << cpi->oxcf.input_bit_depth) - 1);
1940       const double target_rate = (double)cpi->oxcf.target_bandwidth / 1000;
1941       const double rate_err = ((100.0 * (dr - target_rate)) / target_rate);
1942
1943       if (cpi->b_calculate_psnr) {
1944         const double total_psnr = vpx_sse_to_psnr(
1945             (double)cpi->total_samples, peak, (double)cpi->total_sq_error);
1946         const double totalp_psnr = vpx_sse_to_psnr(
1947             (double)cpi->totalp_samples, peak, (double)cpi->totalp_sq_error);
1948         const double total_ssim =
1949             100 * pow(cpi->summed_quality / cpi->summed_weights, 8.0);
1950         const double totalp_ssim =
1951             100 * pow(cpi->summedp_quality / cpi->summedp_weights, 8.0);
1952
1953         snprintf(headings, sizeof(headings),
1954                  "Bitrate\tAVGPsnr\tGLBPsnr\tAVPsnrP\tGLPsnrP\t"
1955                  "VPXSSIM\tVPSSIMP\tFASTSIM\tPSNRHVS\t"
1956                  "WstPsnr\tWstSsim\tWstFast\tWstHVS");
1957         snprintf(results, sizeof(results),
1958                  "%7.2f\t%7.3f\t%7.3f\t%7.3f\t%7.3f\t"
1959                  "%7.3f\t%7.3f\t%7.3f\t%7.3f\t"
1960                  "%7.3f\t%7.3f\t%7.3f\t%7.3f",
1961                  dr, cpi->psnr.stat[ALL] / cpi->count, total_psnr,
1962                  cpi->psnrp.stat[ALL] / cpi->count, totalp_psnr, total_ssim,
1963                  totalp_ssim, cpi->fastssim.stat[ALL] / cpi->count,
1964                  cpi->psnrhvs.stat[ALL] / cpi->count, cpi->psnr.worst,
1965                  cpi->worst_ssim, cpi->fastssim.worst, cpi->psnrhvs.worst);
1966
1967         if (cpi->b_calculate_blockiness) {
1968           SNPRINT(headings, "\t  Block\tWstBlck");
1969           SNPRINT2(results, "\t%7.3f", cpi->total_blockiness / cpi->count);
1970           SNPRINT2(results, "\t%7.3f", cpi->worst_blockiness);
1971         }
1972
1973         if (cpi->b_calculate_consistency) {
1974           double consistency =
1975               vpx_sse_to_psnr((double)cpi->totalp_samples, peak,
1976                               (double)cpi->total_inconsistency);
1977
1978           SNPRINT(headings, "\tConsist\tWstCons");
1979           SNPRINT2(results, "\t%7.3f", consistency);
1980           SNPRINT2(results, "\t%7.3f", cpi->worst_consistency);
1981         }
1982         fprintf(f, "%s\t    Time  Rc-Err Abs Err\n", headings);
1983         fprintf(f, "%s\t%8.0f %7.2f %7.2f\n", results, total_encode_time,
1984                 rate_err, fabs(rate_err));
1985       }
1986
1987       fclose(f);
1988     }
1989
1990 #endif
1991
1992 #if 0
1993     {
1994       printf("\n_pick_loop_filter_level:%d\n", cpi->time_pick_lpf / 1000);
1995       printf("\n_frames recive_data encod_mb_row compress_frame  Total\n");
1996       printf("%6d %10ld %10ld %10ld %10ld\n", cpi->common.current_video_frame,
1997              cpi->time_receive_data / 1000, cpi->time_encode_sb_row / 1000,
1998              cpi->time_compress_data / 1000,
1999              (cpi->time_receive_data + cpi->time_compress_data) / 1000);
2000     }
2001 #endif
2002   }
2003
2004 #if CONFIG_VP9_TEMPORAL_DENOISING
2005   vp9_denoiser_free(&(cpi->denoiser));
2006 #endif
2007
2008   for (t = 0; t < cpi->num_workers; ++t) {
2009     VPxWorker *const worker = &cpi->workers[t];
2010     EncWorkerData *const thread_data = &cpi->tile_thr_data[t];
2011
2012     // Deallocate allocated threads.
2013     vpx_get_worker_interface()->end(worker);
2014
2015     // Deallocate allocated thread data.
2016     if (t < cpi->num_workers - 1) {
2017       vpx_free(thread_data->td->counts);
2018       vp9_free_pc_tree(thread_data->td);
2019       vpx_free(thread_data->td);
2020     }
2021   }
2022   vpx_free(cpi->tile_thr_data);
2023   vpx_free(cpi->workers);
2024
2025   if (cpi->num_workers > 1) vp9_loop_filter_dealloc(&cpi->lf_row_sync);
2026
2027   dealloc_compressor_data(cpi);
2028
2029   for (i = 0; i < sizeof(cpi->mbgraph_stats) / sizeof(cpi->mbgraph_stats[0]);
2030        ++i) {
2031     vpx_free(cpi->mbgraph_stats[i].mb_stats);
2032   }
2033
2034 #if CONFIG_FP_MB_STATS
2035   if (cpi->use_fp_mb_stats) {
2036     vpx_free(cpi->twopass.frame_mb_stats_buf);
2037     cpi->twopass.frame_mb_stats_buf = NULL;
2038   }
2039 #endif
2040
2041   vp9_remove_common(cm);
2042   vp9_free_ref_frame_buffers(cm->buffer_pool);
2043 #if CONFIG_VP9_POSTPROC
2044   vp9_free_postproc_buffers(cm);
2045 #endif
2046   vpx_free(cpi);
2047
2048 #if CONFIG_VP9_TEMPORAL_DENOISING
2049 #ifdef OUTPUT_YUV_DENOISED
2050   fclose(yuv_denoised_file);
2051 #endif
2052 #endif
2053 #ifdef OUTPUT_YUV_SKINMAP
2054   fclose(yuv_skinmap_file);
2055 #endif
2056 #ifdef OUTPUT_YUV_REC
2057   fclose(yuv_rec_file);
2058 #endif
2059
2060 #if 0
2061
2062   if (keyfile)
2063     fclose(keyfile);
2064
2065   if (framepsnr)
2066     fclose(framepsnr);
2067
2068   if (kf_list)
2069     fclose(kf_list);
2070
2071 #endif
2072 }
2073
2074 static void generate_psnr_packet(VP9_COMP *cpi) {
2075   struct vpx_codec_cx_pkt pkt;
2076   int i;
2077   PSNR_STATS psnr;
2078 #if CONFIG_VP9_HIGHBITDEPTH
2079   vpx_calc_highbd_psnr(cpi->raw_source_frame, cpi->common.frame_to_show, &psnr,
2080                        cpi->td.mb.e_mbd.bd, cpi->oxcf.input_bit_depth);
2081 #else
2082   vpx_calc_psnr(cpi->raw_source_frame, cpi->common.frame_to_show, &psnr);
2083 #endif
2084
2085   for (i = 0; i < 4; ++i) {
2086     pkt.data.psnr.samples[i] = psnr.samples[i];
2087     pkt.data.psnr.sse[i] = psnr.sse[i];
2088     pkt.data.psnr.psnr[i] = psnr.psnr[i];
2089   }
2090   pkt.kind = VPX_CODEC_PSNR_PKT;
2091   if (cpi->use_svc)
2092     cpi->svc.layer_context[cpi->svc.spatial_layer_id *
2093                            cpi->svc.number_temporal_layers]
2094         .psnr_pkt = pkt.data.psnr;
2095   else
2096     vpx_codec_pkt_list_add(cpi->output_pkt_list, &pkt);
2097 }
2098
2099 int vp9_use_as_reference(VP9_COMP *cpi, int ref_frame_flags) {
2100   if (ref_frame_flags > 7) return -1;
2101
2102   cpi->ref_frame_flags = ref_frame_flags;
2103   return 0;
2104 }
2105
2106 void vp9_update_reference(VP9_COMP *cpi, int ref_frame_flags) {
2107   cpi->ext_refresh_golden_frame = (ref_frame_flags & VP9_GOLD_FLAG) != 0;
2108   cpi->ext_refresh_alt_ref_frame = (ref_frame_flags & VP9_ALT_FLAG) != 0;
2109   cpi->ext_refresh_last_frame = (ref_frame_flags & VP9_LAST_FLAG) != 0;
2110   cpi->ext_refresh_frame_flags_pending = 1;
2111 }
2112
2113 static YV12_BUFFER_CONFIG *get_vp9_ref_frame_buffer(
2114     VP9_COMP *cpi, VP9_REFFRAME ref_frame_flag) {
2115   MV_REFERENCE_FRAME ref_frame = NONE;
2116   if (ref_frame_flag == VP9_LAST_FLAG)
2117     ref_frame = LAST_FRAME;
2118   else if (ref_frame_flag == VP9_GOLD_FLAG)
2119     ref_frame = GOLDEN_FRAME;
2120   else if (ref_frame_flag == VP9_ALT_FLAG)
2121     ref_frame = ALTREF_FRAME;
2122
2123   return ref_frame == NONE ? NULL : get_ref_frame_buffer(cpi, ref_frame);
2124 }
2125
2126 int vp9_copy_reference_enc(VP9_COMP *cpi, VP9_REFFRAME ref_frame_flag,
2127                            YV12_BUFFER_CONFIG *sd) {
2128   YV12_BUFFER_CONFIG *cfg = get_vp9_ref_frame_buffer(cpi, ref_frame_flag);
2129   if (cfg) {
2130     vp8_yv12_copy_frame(cfg, sd);
2131     return 0;
2132   } else {
2133     return -1;
2134   }
2135 }
2136
2137 int vp9_set_reference_enc(VP9_COMP *cpi, VP9_REFFRAME ref_frame_flag,
2138                           YV12_BUFFER_CONFIG *sd) {
2139   YV12_BUFFER_CONFIG *cfg = get_vp9_ref_frame_buffer(cpi, ref_frame_flag);
2140   if (cfg) {
2141     vp8_yv12_copy_frame(sd, cfg);
2142     return 0;
2143   } else {
2144     return -1;
2145   }
2146 }
2147
2148 int vp9_update_entropy(VP9_COMP *cpi, int update) {
2149   cpi->ext_refresh_frame_context = update;
2150   cpi->ext_refresh_frame_context_pending = 1;
2151   return 0;
2152 }
2153
2154 #if defined(OUTPUT_YUV_DENOISED) || defined(OUTPUT_YUV_SKINMAP)
2155 // The denoiser buffer is allocated as a YUV 440 buffer. This function writes it
2156 // as YUV 420. We simply use the top-left pixels of the UV buffers, since we do
2157 // not denoise the UV channels at this time. If ever we implement UV channel
2158 // denoising we will have to modify this.
2159 void vp9_write_yuv_frame_420(YV12_BUFFER_CONFIG *s, FILE *f) {
2160   uint8_t *src = s->y_buffer;
2161   int h = s->y_height;
2162
2163   do {
2164     fwrite(src, s->y_width, 1, f);
2165     src += s->y_stride;
2166   } while (--h);
2167
2168   src = s->u_buffer;
2169   h = s->uv_height;
2170
2171   do {
2172     fwrite(src, s->uv_width, 1, f);
2173     src += s->uv_stride;
2174   } while (--h);
2175
2176   src = s->v_buffer;
2177   h = s->uv_height;
2178
2179   do {
2180     fwrite(src, s->uv_width, 1, f);
2181     src += s->uv_stride;
2182   } while (--h);
2183 }
2184 #endif
2185
2186 #ifdef OUTPUT_YUV_REC
2187 void vp9_write_yuv_rec_frame(VP9_COMMON *cm) {
2188   YV12_BUFFER_CONFIG *s = cm->frame_to_show;
2189   uint8_t *src = s->y_buffer;
2190   int h = cm->height;
2191
2192 #if CONFIG_VP9_HIGHBITDEPTH
2193   if (s->flags & YV12_FLAG_HIGHBITDEPTH) {
2194     uint16_t *src16 = CONVERT_TO_SHORTPTR(s->y_buffer);
2195
2196     do {
2197       fwrite(src16, s->y_width, 2, yuv_rec_file);
2198       src16 += s->y_stride;
2199     } while (--h);
2200
2201     src16 = CONVERT_TO_SHORTPTR(s->u_buffer);
2202     h = s->uv_height;
2203
2204     do {
2205       fwrite(src16, s->uv_width, 2, yuv_rec_file);
2206       src16 += s->uv_stride;
2207     } while (--h);
2208
2209     src16 = CONVERT_TO_SHORTPTR(s->v_buffer);
2210     h = s->uv_height;
2211
2212     do {
2213       fwrite(src16, s->uv_width, 2, yuv_rec_file);
2214       src16 += s->uv_stride;
2215     } while (--h);
2216
2217     fflush(yuv_rec_file);
2218     return;
2219   }
2220 #endif  // CONFIG_VP9_HIGHBITDEPTH
2221
2222   do {
2223     fwrite(src, s->y_width, 1, yuv_rec_file);
2224     src += s->y_stride;
2225   } while (--h);
2226
2227   src = s->u_buffer;
2228   h = s->uv_height;
2229
2230   do {
2231     fwrite(src, s->uv_width, 1, yuv_rec_file);
2232     src += s->uv_stride;
2233   } while (--h);
2234
2235   src = s->v_buffer;
2236   h = s->uv_height;
2237
2238   do {
2239     fwrite(src, s->uv_width, 1, yuv_rec_file);
2240     src += s->uv_stride;
2241   } while (--h);
2242
2243   fflush(yuv_rec_file);
2244 }
2245 #endif
2246
2247 #if CONFIG_VP9_HIGHBITDEPTH
2248 static void scale_and_extend_frame_nonnormative(const YV12_BUFFER_CONFIG *src,
2249                                                 YV12_BUFFER_CONFIG *dst,
2250                                                 int bd) {
2251 #else
2252 static void scale_and_extend_frame_nonnormative(const YV12_BUFFER_CONFIG *src,
2253                                                 YV12_BUFFER_CONFIG *dst) {
2254 #endif  // CONFIG_VP9_HIGHBITDEPTH
2255   // TODO(dkovalev): replace YV12_BUFFER_CONFIG with vpx_image_t
2256   int i;
2257   const uint8_t *const srcs[3] = { src->y_buffer, src->u_buffer,
2258                                    src->v_buffer };
2259   const int src_strides[3] = { src->y_stride, src->uv_stride, src->uv_stride };
2260   const int src_widths[3] = { src->y_crop_width, src->uv_crop_width,
2261                               src->uv_crop_width };
2262   const int src_heights[3] = { src->y_crop_height, src->uv_crop_height,
2263                                src->uv_crop_height };
2264   uint8_t *const dsts[3] = { dst->y_buffer, dst->u_buffer, dst->v_buffer };
2265   const int dst_strides[3] = { dst->y_stride, dst->uv_stride, dst->uv_stride };
2266   const int dst_widths[3] = { dst->y_crop_width, dst->uv_crop_width,
2267                               dst->uv_crop_width };
2268   const int dst_heights[3] = { dst->y_crop_height, dst->uv_crop_height,
2269                                dst->uv_crop_height };
2270
2271   for (i = 0; i < MAX_MB_PLANE; ++i) {
2272 #if CONFIG_VP9_HIGHBITDEPTH
2273     if (src->flags & YV12_FLAG_HIGHBITDEPTH) {
2274       vp9_highbd_resize_plane(srcs[i], src_heights[i], src_widths[i],
2275                               src_strides[i], dsts[i], dst_heights[i],
2276                               dst_widths[i], dst_strides[i], bd);
2277     } else {
2278       vp9_resize_plane(srcs[i], src_heights[i], src_widths[i], src_strides[i],
2279                        dsts[i], dst_heights[i], dst_widths[i], dst_strides[i]);
2280     }
2281 #else
2282     vp9_resize_plane(srcs[i], src_heights[i], src_widths[i], src_strides[i],
2283                      dsts[i], dst_heights[i], dst_widths[i], dst_strides[i]);
2284 #endif  // CONFIG_VP9_HIGHBITDEPTH
2285   }
2286   vpx_extend_frame_borders(dst);
2287 }
2288
2289 #if CONFIG_VP9_HIGHBITDEPTH
2290 static void scale_and_extend_frame(const YV12_BUFFER_CONFIG *src,
2291                                    YV12_BUFFER_CONFIG *dst, int bd) {
2292   const int src_w = src->y_crop_width;
2293   const int src_h = src->y_crop_height;
2294   const int dst_w = dst->y_crop_width;
2295   const int dst_h = dst->y_crop_height;
2296   const uint8_t *const srcs[3] = { src->y_buffer, src->u_buffer,
2297                                    src->v_buffer };
2298   const int src_strides[3] = { src->y_stride, src->uv_stride, src->uv_stride };
2299   uint8_t *const dsts[3] = { dst->y_buffer, dst->u_buffer, dst->v_buffer };
2300   const int dst_strides[3] = { dst->y_stride, dst->uv_stride, dst->uv_stride };
2301   const InterpKernel *const kernel = vp9_filter_kernels[EIGHTTAP];
2302   int x, y, i;
2303
2304   for (i = 0; i < MAX_MB_PLANE; ++i) {
2305     const int factor = (i == 0 || i == 3 ? 1 : 2);
2306     const int src_stride = src_strides[i];
2307     const int dst_stride = dst_strides[i];
2308     for (y = 0; y < dst_h; y += 16) {
2309       const int y_q4 = y * (16 / factor) * src_h / dst_h;
2310       for (x = 0; x < dst_w; x += 16) {
2311         const int x_q4 = x * (16 / factor) * src_w / dst_w;
2312         const uint8_t *src_ptr = srcs[i] +
2313                                  (y / factor) * src_h / dst_h * src_stride +
2314                                  (x / factor) * src_w / dst_w;
2315         uint8_t *dst_ptr = dsts[i] + (y / factor) * dst_stride + (x / factor);
2316
2317         if (src->flags & YV12_FLAG_HIGHBITDEPTH) {
2318           vpx_highbd_convolve8(src_ptr, src_stride, dst_ptr, dst_stride,
2319                                kernel[x_q4 & 0xf], 16 * src_w / dst_w,
2320                                kernel[y_q4 & 0xf], 16 * src_h / dst_h,
2321                                16 / factor, 16 / factor, bd);
2322         } else {
2323           vpx_scaled_2d(src_ptr, src_stride, dst_ptr, dst_stride,
2324                         kernel[x_q4 & 0xf], 16 * src_w / dst_w,
2325                         kernel[y_q4 & 0xf], 16 * src_h / dst_h, 16 / factor,
2326                         16 / factor);
2327         }
2328       }
2329     }
2330   }
2331
2332   vpx_extend_frame_borders(dst);
2333 }
2334 #else
2335 void vp9_scale_and_extend_frame_c(const YV12_BUFFER_CONFIG *src,
2336                                   YV12_BUFFER_CONFIG *dst) {
2337   const int src_w = src->y_crop_width;
2338   const int src_h = src->y_crop_height;
2339   const int dst_w = dst->y_crop_width;
2340   const int dst_h = dst->y_crop_height;
2341   const uint8_t *const srcs[3] = { src->y_buffer, src->u_buffer,
2342                                    src->v_buffer };
2343   const int src_strides[3] = { src->y_stride, src->uv_stride, src->uv_stride };
2344   uint8_t *const dsts[3] = { dst->y_buffer, dst->u_buffer, dst->v_buffer };
2345   const int dst_strides[3] = { dst->y_stride, dst->uv_stride, dst->uv_stride };
2346   const InterpKernel *const kernel = vp9_filter_kernels[EIGHTTAP];
2347   int x, y, i;
2348
2349   for (i = 0; i < MAX_MB_PLANE; ++i) {
2350     const int factor = (i == 0 || i == 3 ? 1 : 2);
2351     const int src_stride = src_strides[i];
2352     const int dst_stride = dst_strides[i];
2353     for (y = 0; y < dst_h; y += 16) {
2354       const int y_q4 = y * (16 / factor) * src_h / dst_h;
2355       for (x = 0; x < dst_w; x += 16) {
2356         const int x_q4 = x * (16 / factor) * src_w / dst_w;
2357         const uint8_t *src_ptr = srcs[i] +
2358                                  (y / factor) * src_h / dst_h * src_stride +
2359                                  (x / factor) * src_w / dst_w;
2360         uint8_t *dst_ptr = dsts[i] + (y / factor) * dst_stride + (x / factor);
2361
2362         vpx_scaled_2d(src_ptr, src_stride, dst_ptr, dst_stride,
2363                       kernel[x_q4 & 0xf], 16 * src_w / dst_w,
2364                       kernel[y_q4 & 0xf], 16 * src_h / dst_h, 16 / factor,
2365                       16 / factor);
2366       }
2367     }
2368   }
2369
2370   vpx_extend_frame_borders(dst);
2371 }
2372 #endif  // CONFIG_VP9_HIGHBITDEPTH
2373
2374 static int scale_down(VP9_COMP *cpi, int q) {
2375   RATE_CONTROL *const rc = &cpi->rc;
2376   GF_GROUP *const gf_group = &cpi->twopass.gf_group;
2377   int scale = 0;
2378   assert(frame_is_kf_gf_arf(cpi));
2379
2380   if (rc->frame_size_selector == UNSCALED &&
2381       q >= rc->rf_level_maxq[gf_group->rf_level[gf_group->index]]) {
2382     const int max_size_thresh =
2383         (int)(rate_thresh_mult[SCALE_STEP1] *
2384               VPXMAX(rc->this_frame_target, rc->avg_frame_bandwidth));
2385     scale = rc->projected_frame_size > max_size_thresh ? 1 : 0;
2386   }
2387   return scale;
2388 }
2389
2390 static int big_rate_miss(VP9_COMP *cpi, int high_limit, int low_limit) {
2391   const RATE_CONTROL *const rc = &cpi->rc;
2392
2393   return (rc->projected_frame_size > ((high_limit * 3) / 2)) ||
2394          (rc->projected_frame_size < (low_limit / 2));
2395 }
2396
2397 // Function to test for conditions that indicate we should loop
2398 // back and recode a frame.
2399 static int recode_loop_test(VP9_COMP *cpi, int high_limit, int low_limit, int q,
2400                             int maxq, int minq) {
2401   const RATE_CONTROL *const rc = &cpi->rc;
2402   const VP9EncoderConfig *const oxcf = &cpi->oxcf;
2403   const int frame_is_kfgfarf = frame_is_kf_gf_arf(cpi);
2404   int force_recode = 0;
2405
2406   if ((rc->projected_frame_size >= rc->max_frame_bandwidth) ||
2407       big_rate_miss(cpi, high_limit, low_limit) ||
2408       (cpi->sf.recode_loop == ALLOW_RECODE) ||
2409       (frame_is_kfgfarf && (cpi->sf.recode_loop == ALLOW_RECODE_KFARFGF))) {
2410     if (frame_is_kfgfarf && (oxcf->resize_mode == RESIZE_DYNAMIC) &&
2411         scale_down(cpi, q)) {
2412       // Code this group at a lower resolution.
2413       cpi->resize_pending = 1;
2414       return 1;
2415     }
2416
2417     // TODO(agrange) high_limit could be greater than the scale-down threshold.
2418     if ((rc->projected_frame_size > high_limit && q < maxq) ||
2419         (rc->projected_frame_size < low_limit && q > minq)) {
2420       force_recode = 1;
2421     } else if (cpi->oxcf.rc_mode == VPX_CQ) {
2422       // Deal with frame undershoot and whether or not we are
2423       // below the automatically set cq level.
2424       if (q > oxcf->cq_level &&
2425           rc->projected_frame_size < ((rc->this_frame_target * 7) >> 3)) {
2426         force_recode = 1;
2427       }
2428     }
2429   }
2430   return force_recode;
2431 }
2432
2433 void vp9_update_reference_frames(VP9_COMP *cpi) {
2434   VP9_COMMON *const cm = &cpi->common;
2435   BufferPool *const pool = cm->buffer_pool;
2436
2437   // At this point the new frame has been encoded.
2438   // If any buffer copy / swapping is signaled it should be done here.
2439   if (cm->frame_type == KEY_FRAME) {
2440     ref_cnt_fb(pool->frame_bufs, &cm->ref_frame_map[cpi->gld_fb_idx],
2441                cm->new_fb_idx);
2442     ref_cnt_fb(pool->frame_bufs, &cm->ref_frame_map[cpi->alt_fb_idx],
2443                cm->new_fb_idx);
2444   } else if (vp9_preserve_existing_gf(cpi)) {
2445     // We have decided to preserve the previously existing golden frame as our
2446     // new ARF frame. However, in the short term in function
2447     // vp9_get_refresh_mask() we left it in the GF slot and, if
2448     // we're updating the GF with the current decoded frame, we save it to the
2449     // ARF slot instead.
2450     // We now have to update the ARF with the current frame and swap gld_fb_idx
2451     // and alt_fb_idx so that, overall, we've stored the old GF in the new ARF
2452     // slot and, if we're updating the GF, the current frame becomes the new GF.
2453     int tmp;
2454
2455     ref_cnt_fb(pool->frame_bufs, &cm->ref_frame_map[cpi->alt_fb_idx],
2456                cm->new_fb_idx);
2457
2458     tmp = cpi->alt_fb_idx;
2459     cpi->alt_fb_idx = cpi->gld_fb_idx;
2460     cpi->gld_fb_idx = tmp;
2461
2462     if (is_two_pass_svc(cpi)) {
2463       cpi->svc.layer_context[0].gold_ref_idx = cpi->gld_fb_idx;
2464       cpi->svc.layer_context[0].alt_ref_idx = cpi->alt_fb_idx;
2465     }
2466   } else { /* For non key/golden frames */
2467     if (cpi->refresh_alt_ref_frame) {
2468       int arf_idx = cpi->alt_fb_idx;
2469       if ((cpi->oxcf.pass == 2) && cpi->multi_arf_allowed) {
2470         const GF_GROUP *const gf_group = &cpi->twopass.gf_group;
2471         arf_idx = gf_group->arf_update_idx[gf_group->index];
2472       }
2473
2474       ref_cnt_fb(pool->frame_bufs, &cm->ref_frame_map[arf_idx], cm->new_fb_idx);
2475       memcpy(cpi->interp_filter_selected[ALTREF_FRAME],
2476              cpi->interp_filter_selected[0],
2477              sizeof(cpi->interp_filter_selected[0]));
2478     }
2479
2480     if (cpi->refresh_golden_frame) {
2481       ref_cnt_fb(pool->frame_bufs, &cm->ref_frame_map[cpi->gld_fb_idx],
2482                  cm->new_fb_idx);
2483       if (!cpi->rc.is_src_frame_alt_ref)
2484         memcpy(cpi->interp_filter_selected[GOLDEN_FRAME],
2485                cpi->interp_filter_selected[0],
2486                sizeof(cpi->interp_filter_selected[0]));
2487       else
2488         memcpy(cpi->interp_filter_selected[GOLDEN_FRAME],
2489                cpi->interp_filter_selected[ALTREF_FRAME],
2490                sizeof(cpi->interp_filter_selected[ALTREF_FRAME]));
2491     }
2492   }
2493
2494   if (cpi->refresh_last_frame) {
2495     ref_cnt_fb(pool->frame_bufs, &cm->ref_frame_map[cpi->lst_fb_idx],
2496                cm->new_fb_idx);
2497     if (!cpi->rc.is_src_frame_alt_ref)
2498       memcpy(cpi->interp_filter_selected[LAST_FRAME],
2499              cpi->interp_filter_selected[0],
2500              sizeof(cpi->interp_filter_selected[0]));
2501   }
2502 #if CONFIG_VP9_TEMPORAL_DENOISING
2503   if (cpi->oxcf.noise_sensitivity > 0 &&
2504       cpi->denoiser.denoising_level > kDenLowLow) {
2505     vp9_denoiser_update_frame_info(
2506         &cpi->denoiser, *cpi->Source, cpi->common.frame_type,
2507         cpi->refresh_alt_ref_frame, cpi->refresh_golden_frame,
2508         cpi->refresh_last_frame, cpi->resize_pending);
2509   }
2510 #endif
2511   if (is_one_pass_cbr_svc(cpi)) {
2512     // Keep track of frame index for each reference frame.
2513     SVC *const svc = &cpi->svc;
2514     if (cm->frame_type == KEY_FRAME) {
2515       svc->ref_frame_index[cpi->lst_fb_idx] = svc->current_superframe;
2516       svc->ref_frame_index[cpi->gld_fb_idx] = svc->current_superframe;
2517       svc->ref_frame_index[cpi->alt_fb_idx] = svc->current_superframe;
2518     } else {
2519       if (cpi->refresh_last_frame)
2520         svc->ref_frame_index[cpi->lst_fb_idx] = svc->current_superframe;
2521       if (cpi->refresh_golden_frame)
2522         svc->ref_frame_index[cpi->gld_fb_idx] = svc->current_superframe;
2523       if (cpi->refresh_alt_ref_frame)
2524         svc->ref_frame_index[cpi->alt_fb_idx] = svc->current_superframe;
2525     }
2526   }
2527 }
2528
2529 static void loopfilter_frame(VP9_COMP *cpi, VP9_COMMON *cm) {
2530   MACROBLOCKD *xd = &cpi->td.mb.e_mbd;
2531   struct loopfilter *lf = &cm->lf;
2532
2533   if (xd->lossless) {
2534     lf->filter_level = 0;
2535     lf->last_filt_level = 0;
2536   } else {
2537     struct vpx_usec_timer timer;
2538
2539     vpx_clear_system_state();
2540
2541     vpx_usec_timer_start(&timer);
2542
2543     if (!cpi->rc.is_src_frame_alt_ref) {
2544       if ((cpi->common.frame_type == KEY_FRAME) &&
2545           (!cpi->rc.this_key_frame_forced)) {
2546         lf->last_filt_level = 0;
2547       }
2548       vp9_pick_filter_level(cpi->Source, cpi, cpi->sf.lpf_pick);
2549       lf->last_filt_level = lf->filter_level;
2550     } else {
2551       lf->filter_level = 0;
2552     }
2553
2554     vpx_usec_timer_mark(&timer);
2555     cpi->time_pick_lpf += vpx_usec_timer_elapsed(&timer);
2556   }
2557
2558   if (lf->filter_level > 0) {
2559     vp9_build_mask_frame(cm, lf->filter_level, 0);
2560
2561     if (cpi->num_workers > 1)
2562       vp9_loop_filter_frame_mt(cm->frame_to_show, cm, xd->plane,
2563                                lf->filter_level, 0, 0, cpi->workers,
2564                                cpi->num_workers, &cpi->lf_row_sync);
2565     else
2566       vp9_loop_filter_frame(cm->frame_to_show, cm, xd, lf->filter_level, 0, 0);
2567   }
2568
2569   vpx_extend_frame_inner_borders(cm->frame_to_show);
2570 }
2571
2572 static INLINE void alloc_frame_mvs(VP9_COMMON *const cm, int buffer_idx) {
2573   RefCntBuffer *const new_fb_ptr = &cm->buffer_pool->frame_bufs[buffer_idx];
2574   if (new_fb_ptr->mvs == NULL || new_fb_ptr->mi_rows < cm->mi_rows ||
2575       new_fb_ptr->mi_cols < cm->mi_cols) {
2576     vpx_free(new_fb_ptr->mvs);
2577     CHECK_MEM_ERROR(cm, new_fb_ptr->mvs,
2578                     (MV_REF *)vpx_calloc(cm->mi_rows * cm->mi_cols,
2579                                          sizeof(*new_fb_ptr->mvs)));
2580     new_fb_ptr->mi_rows = cm->mi_rows;
2581     new_fb_ptr->mi_cols = cm->mi_cols;
2582   }
2583 }
2584
2585 void vp9_scale_references(VP9_COMP *cpi) {
2586   VP9_COMMON *cm = &cpi->common;
2587   MV_REFERENCE_FRAME ref_frame;
2588   const VP9_REFFRAME ref_mask[3] = { VP9_LAST_FLAG, VP9_GOLD_FLAG,
2589                                      VP9_ALT_FLAG };
2590
2591   for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ++ref_frame) {
2592     // Need to convert from VP9_REFFRAME to index into ref_mask (subtract 1).
2593     if (cpi->ref_frame_flags & ref_mask[ref_frame - 1]) {
2594       BufferPool *const pool = cm->buffer_pool;
2595       const YV12_BUFFER_CONFIG *const ref =
2596           get_ref_frame_buffer(cpi, ref_frame);
2597
2598       if (ref == NULL) {
2599         cpi->scaled_ref_idx[ref_frame - 1] = INVALID_IDX;
2600         continue;
2601       }
2602
2603 #if CONFIG_VP9_HIGHBITDEPTH
2604       if (ref->y_crop_width != cm->width || ref->y_crop_height != cm->height) {
2605         RefCntBuffer *new_fb_ptr = NULL;
2606         int force_scaling = 0;
2607         int new_fb = cpi->scaled_ref_idx[ref_frame - 1];
2608         if (new_fb == INVALID_IDX) {
2609           new_fb = get_free_fb(cm);
2610           force_scaling = 1;
2611         }
2612         if (new_fb == INVALID_IDX) return;
2613         new_fb_ptr = &pool->frame_bufs[new_fb];
2614         if (force_scaling || new_fb_ptr->buf.y_crop_width != cm->width ||
2615             new_fb_ptr->buf.y_crop_height != cm->height) {
2616           if (vpx_realloc_frame_buffer(&new_fb_ptr->buf, cm->width, cm->height,
2617                                        cm->subsampling_x, cm->subsampling_y,
2618                                        cm->use_highbitdepth,
2619                                        VP9_ENC_BORDER_IN_PIXELS,
2620                                        cm->byte_alignment, NULL, NULL, NULL))
2621             vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
2622                                "Failed to allocate frame buffer");
2623           scale_and_extend_frame(ref, &new_fb_ptr->buf, (int)cm->bit_depth);
2624           cpi->scaled_ref_idx[ref_frame - 1] = new_fb;
2625           alloc_frame_mvs(cm, new_fb);
2626         }
2627 #else
2628       if (ref->y_crop_width != cm->width || ref->y_crop_height != cm->height) {
2629         RefCntBuffer *new_fb_ptr = NULL;
2630         int force_scaling = 0;
2631         int new_fb = cpi->scaled_ref_idx[ref_frame - 1];
2632         if (new_fb == INVALID_IDX) {
2633           new_fb = get_free_fb(cm);
2634           force_scaling = 1;
2635         }
2636         if (new_fb == INVALID_IDX) return;
2637         new_fb_ptr = &pool->frame_bufs[new_fb];
2638         if (force_scaling || new_fb_ptr->buf.y_crop_width != cm->width ||
2639             new_fb_ptr->buf.y_crop_height != cm->height) {
2640           if (vpx_realloc_frame_buffer(&new_fb_ptr->buf, cm->width, cm->height,
2641                                        cm->subsampling_x, cm->subsampling_y,
2642                                        VP9_ENC_BORDER_IN_PIXELS,
2643                                        cm->byte_alignment, NULL, NULL, NULL))
2644             vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
2645                                "Failed to allocate frame buffer");
2646           vp9_scale_and_extend_frame(ref, &new_fb_ptr->buf);
2647           cpi->scaled_ref_idx[ref_frame - 1] = new_fb;
2648           alloc_frame_mvs(cm, new_fb);
2649         }
2650 #endif  // CONFIG_VP9_HIGHBITDEPTH
2651       } else {
2652         int buf_idx;
2653         RefCntBuffer *buf = NULL;
2654         if (cpi->oxcf.pass == 0 && !cpi->use_svc) {
2655           // Check for release of scaled reference.
2656           buf_idx = cpi->scaled_ref_idx[ref_frame - 1];
2657           buf = (buf_idx != INVALID_IDX) ? &pool->frame_bufs[buf_idx] : NULL;
2658           if (buf != NULL) {
2659             --buf->ref_count;
2660             cpi->scaled_ref_idx[ref_frame - 1] = INVALID_IDX;
2661           }
2662         }
2663         buf_idx = get_ref_frame_buf_idx(cpi, ref_frame);
2664         buf = &pool->frame_bufs[buf_idx];
2665         buf->buf.y_crop_width = ref->y_crop_width;
2666         buf->buf.y_crop_height = ref->y_crop_height;
2667         cpi->scaled_ref_idx[ref_frame - 1] = buf_idx;
2668         ++buf->ref_count;
2669       }
2670     } else {
2671       if (cpi->oxcf.pass != 0 || cpi->use_svc)
2672         cpi->scaled_ref_idx[ref_frame - 1] = INVALID_IDX;
2673     }
2674   }
2675 }
2676
2677 static void release_scaled_references(VP9_COMP *cpi) {
2678   VP9_COMMON *cm = &cpi->common;
2679   int i;
2680   if (cpi->oxcf.pass == 0 && !cpi->use_svc) {
2681     // Only release scaled references under certain conditions:
2682     // if reference will be updated, or if scaled reference has same resolution.
2683     int refresh[3];
2684     refresh[0] = (cpi->refresh_last_frame) ? 1 : 0;
2685     refresh[1] = (cpi->refresh_golden_frame) ? 1 : 0;
2686     refresh[2] = (cpi->refresh_alt_ref_frame) ? 1 : 0;
2687     for (i = LAST_FRAME; i <= ALTREF_FRAME; ++i) {
2688       const int idx = cpi->scaled_ref_idx[i - 1];
2689       RefCntBuffer *const buf =
2690           idx != INVALID_IDX ? &cm->buffer_pool->frame_bufs[idx] : NULL;
2691       const YV12_BUFFER_CONFIG *const ref = get_ref_frame_buffer(cpi, i);
2692       if (buf != NULL &&
2693           (refresh[i - 1] || (buf->buf.y_crop_width == ref->y_crop_width &&
2694                               buf->buf.y_crop_height == ref->y_crop_height))) {
2695         --buf->ref_count;
2696         cpi->scaled_ref_idx[i - 1] = INVALID_IDX;
2697       }
2698     }
2699   } else {
2700     for (i = 0; i < MAX_REF_FRAMES; ++i) {
2701       const int idx = cpi->scaled_ref_idx[i];
2702       RefCntBuffer *const buf =
2703           idx != INVALID_IDX ? &cm->buffer_pool->frame_bufs[idx] : NULL;
2704       if (buf != NULL) {
2705         --buf->ref_count;
2706         cpi->scaled_ref_idx[i] = INVALID_IDX;
2707       }
2708     }
2709   }
2710 }
2711
2712 static void full_to_model_count(unsigned int *model_count,
2713                                 unsigned int *full_count) {
2714   int n;
2715   model_count[ZERO_TOKEN] = full_count[ZERO_TOKEN];
2716   model_count[ONE_TOKEN] = full_count[ONE_TOKEN];
2717   model_count[TWO_TOKEN] = full_count[TWO_TOKEN];
2718   for (n = THREE_TOKEN; n < EOB_TOKEN; ++n)
2719     model_count[TWO_TOKEN] += full_count[n];
2720   model_count[EOB_MODEL_TOKEN] = full_count[EOB_TOKEN];
2721 }
2722
2723 static void full_to_model_counts(vp9_coeff_count_model *model_count,
2724                                  vp9_coeff_count *full_count) {
2725   int i, j, k, l;
2726
2727   for (i = 0; i < PLANE_TYPES; ++i)
2728     for (j = 0; j < REF_TYPES; ++j)
2729       for (k = 0; k < COEF_BANDS; ++k)
2730         for (l = 0; l < BAND_COEFF_CONTEXTS(k); ++l)
2731           full_to_model_count(model_count[i][j][k][l], full_count[i][j][k][l]);
2732 }
2733
2734 #if 0 && CONFIG_INTERNAL_STATS
2735 static void output_frame_level_debug_stats(VP9_COMP *cpi) {
2736   VP9_COMMON *const cm = &cpi->common;
2737   FILE *const f = fopen("tmp.stt", cm->current_video_frame ? "a" : "w");
2738   int64_t recon_err;
2739
2740   vpx_clear_system_state();
2741
2742 #if CONFIG_VP9_HIGHBITDEPTH
2743   if (cm->use_highbitdepth) {
2744     recon_err = vpx_highbd_get_y_sse(cpi->Source, get_frame_new_buffer(cm));
2745   } else {
2746     recon_err = vpx_get_y_sse(cpi->Source, get_frame_new_buffer(cm));
2747   }
2748 #else
2749   recon_err = vpx_get_y_sse(cpi->Source, get_frame_new_buffer(cm));
2750 #endif  // CONFIG_VP9_HIGHBITDEPTH
2751
2752
2753   if (cpi->twopass.total_left_stats.coded_error != 0.0) {
2754     double dc_quant_devisor;
2755 #if CONFIG_VP9_HIGHBITDEPTH
2756     switch (cm->bit_depth) {
2757       case VPX_BITS_8:
2758         dc_quant_devisor = 4.0;
2759         break;
2760       case VPX_BITS_10:
2761         dc_quant_devisor = 16.0;
2762         break;
2763       case VPX_BITS_12:
2764         dc_quant_devisor = 64.0;
2765         break;
2766       default:
2767         assert(0 && "bit_depth must be VPX_BITS_8, VPX_BITS_10 or VPX_BITS_12");
2768         break;
2769     }
2770 #else
2771     dc_quant_devisor = 4.0;
2772 #endif
2773
2774     fprintf(f, "%10u %dx%d %10d %10d %d %d %10d %10d %10d %10d"
2775        "%10"PRId64" %10"PRId64" %5d %5d %10"PRId64" "
2776        "%10"PRId64" %10"PRId64" %10d "
2777        "%7.2lf %7.2lf %7.2lf %7.2lf %7.2lf"
2778         "%6d %6d %5d %5d %5d "
2779         "%10"PRId64" %10.3lf"
2780         "%10lf %8u %10"PRId64" %10d %10d %10d %10d %10d\n",
2781         cpi->common.current_video_frame,
2782         cm->width, cm->height,
2783         cpi->td.rd_counts.m_search_count,
2784         cpi->td.rd_counts.ex_search_count,
2785         cpi->rc.source_alt_ref_pending,
2786         cpi->rc.source_alt_ref_active,
2787         cpi->rc.this_frame_target,
2788         cpi->rc.projected_frame_size,
2789         cpi->rc.projected_frame_size / cpi->common.MBs,
2790         (cpi->rc.projected_frame_size - cpi->rc.this_frame_target),
2791         cpi->rc.vbr_bits_off_target,
2792         cpi->rc.vbr_bits_off_target_fast,
2793         cpi->twopass.extend_minq,
2794         cpi->twopass.extend_minq_fast,
2795         cpi->rc.total_target_vs_actual,
2796         (cpi->rc.starting_buffer_level - cpi->rc.bits_off_target),
2797         cpi->rc.total_actual_bits, cm->base_qindex,
2798         vp9_convert_qindex_to_q(cm->base_qindex, cm->bit_depth),
2799         (double)vp9_dc_quant(cm->base_qindex, 0, cm->bit_depth) /
2800             dc_quant_devisor,
2801         vp9_convert_qindex_to_q(cpi->twopass.active_worst_quality,
2802                                 cm->bit_depth),
2803         cpi->rc.avg_q,
2804         vp9_convert_qindex_to_q(cpi->oxcf.cq_level, cm->bit_depth),
2805         cpi->refresh_last_frame, cpi->refresh_golden_frame,
2806         cpi->refresh_alt_ref_frame, cm->frame_type, cpi->rc.gfu_boost,
2807         cpi->twopass.bits_left,
2808         cpi->twopass.total_left_stats.coded_error,
2809         cpi->twopass.bits_left /
2810             (1 + cpi->twopass.total_left_stats.coded_error),
2811         cpi->tot_recode_hits, recon_err, cpi->rc.kf_boost,
2812         cpi->twopass.kf_zeromotion_pct,
2813         cpi->twopass.fr_content_type,
2814         cm->lf.filter_level,
2815         cm->seg.aq_av_offset);
2816   }
2817   fclose(f);
2818
2819   if (0) {
2820     FILE *const fmodes = fopen("Modes.stt", "a");
2821     int i;
2822
2823     fprintf(fmodes, "%6d:%1d:%1d:%1d ", cpi->common.current_video_frame,
2824             cm->frame_type, cpi->refresh_golden_frame,
2825             cpi->refresh_alt_ref_frame);
2826
2827     for (i = 0; i < MAX_MODES; ++i)
2828       fprintf(fmodes, "%5d ", cpi->mode_chosen_counts[i]);
2829
2830     fprintf(fmodes, "\n");
2831
2832     fclose(fmodes);
2833   }
2834 }
2835 #endif
2836
2837 static void set_mv_search_params(VP9_COMP *cpi) {
2838   const VP9_COMMON *const cm = &cpi->common;
2839   const unsigned int max_mv_def = VPXMIN(cm->width, cm->height);
2840
2841   // Default based on max resolution.
2842   cpi->mv_step_param = vp9_init_search_range(max_mv_def);
2843
2844   if (cpi->sf.mv.auto_mv_step_size) {
2845     if (frame_is_intra_only(cm)) {
2846       // Initialize max_mv_magnitude for use in the first INTER frame
2847       // after a key/intra-only frame.
2848       cpi->max_mv_magnitude = max_mv_def;
2849     } else {
2850       if (cm->show_frame) {
2851         // Allow mv_steps to correspond to twice the max mv magnitude found
2852         // in the previous frame, capped by the default max_mv_magnitude based
2853         // on resolution.
2854         cpi->mv_step_param = vp9_init_search_range(
2855             VPXMIN(max_mv_def, 2 * cpi->max_mv_magnitude));
2856       }
2857       cpi->max_mv_magnitude = 0;
2858     }
2859   }
2860 }
2861
2862 static void set_size_independent_vars(VP9_COMP *cpi) {
2863   vp9_set_speed_features_framesize_independent(cpi);
2864   vp9_set_rd_speed_thresholds(cpi);
2865   vp9_set_rd_speed_thresholds_sub8x8(cpi);
2866   cpi->common.interp_filter = cpi->sf.default_interp_filter;
2867 }
2868
2869 static void set_size_dependent_vars(VP9_COMP *cpi, int *q, int *bottom_index,
2870                                     int *top_index) {
2871   VP9_COMMON *const cm = &cpi->common;
2872   const VP9EncoderConfig *const oxcf = &cpi->oxcf;
2873
2874   // Setup variables that depend on the dimensions of the frame.
2875   vp9_set_speed_features_framesize_dependent(cpi);
2876
2877   // Decide q and q bounds.
2878   *q = vp9_rc_pick_q_and_bounds(cpi, bottom_index, top_index);
2879
2880   if (!frame_is_intra_only(cm)) {
2881     vp9_set_high_precision_mv(cpi, (*q) < HIGH_PRECISION_MV_QTHRESH);
2882   }
2883
2884   // Configure experimental use of segmentation for enhanced coding of
2885   // static regions if indicated.
2886   // Only allowed in the second pass of a two pass encode, as it requires
2887   // lagged coding, and if the relevant speed feature flag is set.
2888   if (oxcf->pass == 2 && cpi->sf.static_segmentation)
2889     configure_static_seg_features(cpi);
2890
2891 #if CONFIG_VP9_POSTPROC && !(CONFIG_VP9_TEMPORAL_DENOISING)
2892   if (oxcf->noise_sensitivity > 0) {
2893     int l = 0;
2894     switch (oxcf->noise_sensitivity) {
2895       case 1: l = 20; break;
2896       case 2: l = 40; break;
2897       case 3: l = 60; break;
2898       case 4:
2899       case 5: l = 100; break;
2900       case 6: l = 150; break;
2901     }
2902     if (!cpi->common.postproc_state.limits) {
2903       cpi->common.postproc_state.limits = vpx_calloc(
2904           cpi->common.width, sizeof(*cpi->common.postproc_state.limits));
2905     }
2906     vp9_denoise(cpi->Source, cpi->Source, l, cpi->common.postproc_state.limits);
2907   }
2908 #endif  // CONFIG_VP9_POSTPROC
2909 }
2910
2911 #if CONFIG_VP9_TEMPORAL_DENOISING
2912 static void setup_denoiser_buffer(VP9_COMP *cpi) {
2913   VP9_COMMON *const cm = &cpi->common;
2914   if (cpi->oxcf.noise_sensitivity > 0 &&
2915       !cpi->denoiser.frame_buffer_initialized) {
2916     if (vp9_denoiser_alloc(&cpi->denoiser, cm->width, cm->height,
2917                            cm->subsampling_x, cm->subsampling_y,
2918 #if CONFIG_VP9_HIGHBITDEPTH
2919                            cm->use_highbitdepth,
2920 #endif
2921                            VP9_ENC_BORDER_IN_PIXELS))
2922       vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
2923                          "Failed to allocate denoiser");
2924   }
2925 }
2926 #endif
2927
2928 static void init_motion_estimation(VP9_COMP *cpi) {
2929   int y_stride = cpi->scaled_source.y_stride;
2930
2931   if (cpi->sf.mv.search_method == NSTEP) {
2932     vp9_init3smotion_compensation(&cpi->ss_cfg, y_stride);
2933   } else if (cpi->sf.mv.search_method == DIAMOND) {
2934     vp9_init_dsmotion_compensation(&cpi->ss_cfg, y_stride);
2935   }
2936 }
2937
2938 static void set_frame_size(VP9_COMP *cpi) {
2939   int ref_frame;
2940   VP9_COMMON *const cm = &cpi->common;
2941   VP9EncoderConfig *const oxcf = &cpi->oxcf;
2942   MACROBLOCKD *const xd = &cpi->td.mb.e_mbd;
2943
2944   if (oxcf->pass == 2 && oxcf->rc_mode == VPX_VBR &&
2945       ((oxcf->resize_mode == RESIZE_FIXED && cm->current_video_frame == 0) ||
2946        (oxcf->resize_mode == RESIZE_DYNAMIC && cpi->resize_pending))) {
2947     calculate_coded_size(cpi, &oxcf->scaled_frame_width,
2948                          &oxcf->scaled_frame_height);
2949
2950     // There has been a change in frame size.
2951     vp9_set_size_literal(cpi, oxcf->scaled_frame_width,
2952                          oxcf->scaled_frame_height);
2953   }
2954
2955   if (oxcf->pass == 0 && oxcf->rc_mode == VPX_CBR && !cpi->use_svc &&
2956       oxcf->resize_mode == RESIZE_DYNAMIC && cpi->resize_pending != 0) {
2957     oxcf->scaled_frame_width =
2958         (oxcf->width * cpi->resize_scale_num) / cpi->resize_scale_den;
2959     oxcf->scaled_frame_height =
2960         (oxcf->height * cpi->resize_scale_num) / cpi->resize_scale_den;
2961     // There has been a change in frame size.
2962     vp9_set_size_literal(cpi, oxcf->scaled_frame_width,
2963                          oxcf->scaled_frame_height);
2964
2965     // TODO(agrange) Scale cpi->max_mv_magnitude if frame-size has changed.
2966     set_mv_search_params(cpi);
2967
2968     vp9_noise_estimate_init(&cpi->noise_estimate, cm->width, cm->height);
2969 #if CONFIG_VP9_TEMPORAL_DENOISING
2970     // Reset the denoiser on the resized frame.
2971     if (cpi->oxcf.noise_sensitivity > 0) {
2972       vp9_denoiser_free(&(cpi->denoiser));
2973       setup_denoiser_buffer(cpi);
2974       // Dynamic resize is only triggered for non-SVC, so we can force
2975       // golden frame update here as temporary fix to denoiser.
2976       cpi->refresh_golden_frame = 1;
2977     }
2978 #endif
2979   }
2980
2981   if ((oxcf->pass == 2) &&
2982       (!cpi->use_svc || (is_two_pass_svc(cpi) &&
2983                          cpi->svc.encode_empty_frame_state != ENCODING))) {
2984     vp9_set_target_rate(cpi);
2985   }
2986
2987   alloc_frame_mvs(cm, cm->new_fb_idx);
2988
2989   // Reset the frame pointers to the current frame size.
2990   if (vpx_realloc_frame_buffer(get_frame_new_buffer(cm), cm->width, cm->height,
2991                                cm->subsampling_x, cm->subsampling_y,
2992 #if CONFIG_VP9_HIGHBITDEPTH
2993                                cm->use_highbitdepth,
2994 #endif
2995                                VP9_ENC_BORDER_IN_PIXELS, cm->byte_alignment,
2996                                NULL, NULL, NULL))
2997     vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
2998                        "Failed to allocate frame buffer");
2999
3000   alloc_util_frame_buffers(cpi);
3001   init_motion_estimation(cpi);
3002
3003   for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ++ref_frame) {
3004     RefBuffer *const ref_buf = &cm->frame_refs[ref_frame - 1];
3005     const int buf_idx = get_ref_frame_buf_idx(cpi, ref_frame);
3006
3007     ref_buf->idx = buf_idx;
3008
3009     if (buf_idx != INVALID_IDX) {
3010       YV12_BUFFER_CONFIG *const buf = &cm->buffer_pool->frame_bufs[buf_idx].buf;
3011       ref_buf->buf = buf;
3012 #if CONFIG_VP9_HIGHBITDEPTH
3013       vp9_setup_scale_factors_for_frame(
3014           &ref_buf->sf, buf->y_crop_width, buf->y_crop_height, cm->width,
3015           cm->height, (buf->flags & YV12_FLAG_HIGHBITDEPTH) ? 1 : 0);
3016 #else
3017       vp9_setup_scale_factors_for_frame(&ref_buf->sf, buf->y_crop_width,
3018                                         buf->y_crop_height, cm->width,
3019                                         cm->height);
3020 #endif  // CONFIG_VP9_HIGHBITDEPTH
3021       if (vp9_is_scaled(&ref_buf->sf)) vpx_extend_frame_borders(buf);
3022     } else {
3023       ref_buf->buf = NULL;
3024     }
3025   }
3026
3027   set_ref_ptrs(cm, xd, LAST_FRAME, LAST_FRAME);
3028 }
3029
3030 static void encode_without_recode_loop(VP9_COMP *cpi, size_t *size,
3031                                        uint8_t *dest) {
3032   VP9_COMMON *const cm = &cpi->common;
3033   int q = 0, bottom_index = 0, top_index = 0;  // Dummy variables.
3034
3035   vpx_clear_system_state();
3036
3037   set_frame_size(cpi);
3038
3039   if (is_one_pass_cbr_svc(cpi) &&
3040       cpi->un_scaled_source->y_width == cm->width << 2 &&
3041       cpi->un_scaled_source->y_height == cm->height << 2 &&
3042       cpi->svc.scaled_temp.y_width == cm->width << 1 &&
3043       cpi->svc.scaled_temp.y_height == cm->height << 1) {
3044     // For svc, if it is a 1/4x1/4 downscaling, do a two-stage scaling to take
3045     // advantage of the 1:2 optimized scaler. In the process, the 1/2x1/2
3046     // result will be saved in scaled_temp and might be used later.
3047     cpi->Source = vp9_svc_twostage_scale(
3048         cm, cpi->un_scaled_source, &cpi->scaled_source, &cpi->svc.scaled_temp);
3049     cpi->svc.scaled_one_half = 1;
3050   } else if (is_one_pass_cbr_svc(cpi) &&
3051              cpi->un_scaled_source->y_width == cm->width << 1 &&
3052              cpi->un_scaled_source->y_height == cm->height << 1 &&
3053              cpi->svc.scaled_one_half) {
3054     // If the spatial layer is 1/2x1/2 and the scaling is already done in the
3055     // two-stage scaling, use the result directly.
3056     cpi->Source = &cpi->svc.scaled_temp;
3057     cpi->svc.scaled_one_half = 0;
3058   } else {
3059     cpi->Source = vp9_scale_if_required(
3060         cm, cpi->un_scaled_source, &cpi->scaled_source, (cpi->oxcf.pass == 0));
3061   }
3062   // Unfiltered raw source used in metrics calculation if the source
3063   // has been filtered.
3064   if (is_psnr_calc_enabled(cpi)) {
3065 #ifdef ENABLE_KF_DENOISE
3066     if (is_spatial_denoise_enabled(cpi)) {
3067       cpi->raw_source_frame =
3068           vp9_scale_if_required(cm, &cpi->raw_unscaled_source,
3069                                 &cpi->raw_scaled_source, (cpi->oxcf.pass == 0));
3070     } else {
3071       cpi->raw_source_frame = cpi->Source;
3072     }
3073 #else
3074     cpi->raw_source_frame = cpi->Source;
3075 #endif
3076   }
3077
3078   // Avoid scaling last_source unless its needed.
3079   // Last source is needed if vp9_avg_source_sad() is used, or if
3080   // partition_search_type == SOURCE_VAR_BASED_PARTITION, or if noise
3081   // estimation is enabled.
3082   if (cpi->unscaled_last_source != NULL &&
3083       (cpi->oxcf.content == VP9E_CONTENT_SCREEN ||
3084        (cpi->oxcf.pass == 0 && cpi->oxcf.rc_mode == VPX_VBR &&
3085         cpi->oxcf.mode == REALTIME && cpi->oxcf.speed >= 5) ||
3086        cpi->sf.partition_search_type == SOURCE_VAR_BASED_PARTITION ||
3087        cpi->noise_estimate.enabled))
3088     cpi->Last_Source =
3089         vp9_scale_if_required(cm, cpi->unscaled_last_source,
3090                               &cpi->scaled_last_source, (cpi->oxcf.pass == 0));
3091
3092   if (cm->frame_type == KEY_FRAME || cpi->resize_pending != 0) {
3093     memset(cpi->consec_zero_mv, 0,
3094            cm->mi_rows * cm->mi_cols * sizeof(*cpi->consec_zero_mv));
3095   }
3096
3097   vp9_update_noise_estimate(cpi);
3098
3099   if (cpi->oxcf.pass == 0 && cpi->oxcf.mode == REALTIME &&
3100       cpi->oxcf.speed >= 5 && cpi->resize_state == 0 &&
3101       (cpi->oxcf.content == VP9E_CONTENT_SCREEN ||
3102        cpi->oxcf.rc_mode == VPX_VBR))
3103     vp9_avg_source_sad(cpi);
3104
3105   // For 1 pass SVC, since only ZEROMV is allowed for upsampled reference
3106   // frame (i.e, svc->force_zero_mode_spatial_ref = 0), we can avoid this
3107   // frame-level upsampling.
3108   if (frame_is_intra_only(cm) == 0 && !is_one_pass_cbr_svc(cpi)) {
3109     vp9_scale_references(cpi);
3110   }
3111
3112   set_size_independent_vars(cpi);
3113   set_size_dependent_vars(cpi, &q, &bottom_index, &top_index);
3114
3115   if (cpi->oxcf.speed >= 5 && cpi->oxcf.pass == 0 &&
3116       cpi->oxcf.rc_mode == VPX_CBR &&
3117       cpi->oxcf.content != VP9E_CONTENT_SCREEN &&
3118       cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ) {
3119     cpi->use_skin_detection = 1;
3120   }
3121
3122   vp9_set_quantizer(cm, q);
3123   vp9_set_variance_partition_thresholds(cpi, q);
3124
3125   setup_frame(cpi);
3126
3127   suppress_active_map(cpi);
3128   // Variance adaptive and in frame q adjustment experiments are mutually
3129   // exclusive.
3130   if (cpi->oxcf.aq_mode == VARIANCE_AQ) {
3131     vp9_vaq_frame_setup(cpi);
3132   } else if (cpi->oxcf.aq_mode == EQUATOR360_AQ) {
3133     vp9_360aq_frame_setup(cpi);
3134   } else if (cpi->oxcf.aq_mode == COMPLEXITY_AQ) {
3135     vp9_setup_in_frame_q_adj(cpi);
3136   } else if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ) {
3137     vp9_cyclic_refresh_setup(cpi);
3138   }
3139   apply_active_map(cpi);
3140
3141   vp9_encode_frame(cpi);
3142
3143   // Check if we should drop this frame because of high overshoot.
3144   // Only for frames where high temporal-source SAD is detected.
3145   if (cpi->oxcf.pass == 0 && cpi->oxcf.rc_mode == VPX_CBR &&
3146       cpi->resize_state == 0 && cm->frame_type != KEY_FRAME &&
3147       cpi->oxcf.content == VP9E_CONTENT_SCREEN &&
3148       cpi->rc.high_source_sad == 1) {
3149     int frame_size = 0;
3150     // Get an estimate of the encoded frame size.
3151     save_coding_context(cpi);
3152     vp9_pack_bitstream(cpi, dest, size);
3153     restore_coding_context(cpi);
3154     frame_size = (int)(*size) << 3;
3155     // Check if encoded frame will overshoot too much, and if so, set the q and
3156     // adjust some rate control parameters, and return to re-encode the frame.
3157     if (vp9_encodedframe_overshoot(cpi, frame_size, &q)) {
3158       vpx_clear_system_state();
3159       vp9_set_quantizer(cm, q);
3160       vp9_set_variance_partition_thresholds(cpi, q);
3161       suppress_active_map(cpi);
3162       // Turn-off cyclic refresh for re-encoded frame.
3163       if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ) {
3164         unsigned char *const seg_map = cpi->segmentation_map;
3165         memset(seg_map, 0, cm->mi_rows * cm->mi_cols);
3166         vp9_disable_segmentation(&cm->seg);
3167       }
3168       apply_active_map(cpi);
3169       vp9_encode_frame(cpi);
3170     }
3171   }
3172
3173   // Update some stats from cyclic refresh, and check if we should not update
3174   // golden reference, for non-SVC 1 pass CBR.
3175   if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ && cm->frame_type != KEY_FRAME &&
3176       !cpi->use_svc && cpi->ext_refresh_frame_flags_pending == 0 &&
3177       (cpi->oxcf.pass == 0 && cpi->oxcf.rc_mode == VPX_CBR))
3178     vp9_cyclic_refresh_check_golden_update(cpi);
3179
3180   // Update the skip mb flag probabilities based on the distribution
3181   // seen in the last encoder iteration.
3182   // update_base_skip_probs(cpi);
3183   vpx_clear_system_state();
3184 }
3185
3186 static void encode_with_recode_loop(VP9_COMP *cpi, size_t *size,
3187                                     uint8_t *dest) {
3188   VP9_COMMON *const cm = &cpi->common;
3189   RATE_CONTROL *const rc = &cpi->rc;
3190   int bottom_index, top_index;
3191   int loop_count = 0;
3192   int loop_at_this_size = 0;
3193   int loop = 0;
3194   int overshoot_seen = 0;
3195   int undershoot_seen = 0;
3196   int frame_over_shoot_limit;
3197   int frame_under_shoot_limit;
3198   int q = 0, q_low = 0, q_high = 0;
3199
3200   set_size_independent_vars(cpi);
3201
3202   do {
3203     vpx_clear_system_state();
3204
3205     set_frame_size(cpi);
3206
3207     if (loop_count == 0 || cpi->resize_pending != 0) {
3208       set_size_dependent_vars(cpi, &q, &bottom_index, &top_index);
3209
3210       // TODO(agrange) Scale cpi->max_mv_magnitude if frame-size has changed.
3211       set_mv_search_params(cpi);
3212
3213       // Reset the loop state for new frame size.
3214       overshoot_seen = 0;
3215       undershoot_seen = 0;
3216
3217       // Reconfiguration for change in frame size has concluded.
3218       cpi->resize_pending = 0;
3219
3220       q_low = bottom_index;
3221       q_high = top_index;
3222
3223       loop_at_this_size = 0;
3224     }
3225
3226     // Decide frame size bounds first time through.
3227     if (loop_count == 0) {
3228       vp9_rc_compute_frame_size_bounds(cpi, rc->this_frame_target,
3229                                        &frame_under_shoot_limit,
3230                                        &frame_over_shoot_limit);
3231     }
3232
3233     cpi->Source = vp9_scale_if_required(
3234         cm, cpi->un_scaled_source, &cpi->scaled_source, (cpi->oxcf.pass == 0));
3235
3236     // Unfiltered raw source used in metrics calculation if the source
3237     // has been filtered.
3238     if (is_psnr_calc_enabled(cpi)) {
3239 #ifdef ENABLE_KF_DENOISE
3240       if (is_spatial_denoise_enabled(cpi)) {
3241         cpi->raw_source_frame = vp9_scale_if_required(
3242             cm, &cpi->raw_unscaled_source, &cpi->raw_scaled_source,
3243             (cpi->oxcf.pass == 0));
3244       } else {
3245         cpi->raw_source_frame = cpi->Source;
3246       }
3247 #else
3248       cpi->raw_source_frame = cpi->Source;
3249 #endif
3250     }
3251
3252     if (cpi->unscaled_last_source != NULL)
3253       cpi->Last_Source = vp9_scale_if_required(cm, cpi->unscaled_last_source,
3254                                                &cpi->scaled_last_source,
3255                                                (cpi->oxcf.pass == 0));
3256
3257     if (frame_is_intra_only(cm) == 0) {
3258       if (loop_count > 0) {
3259         release_scaled_references(cpi);
3260       }
3261       vp9_scale_references(cpi);
3262     }
3263
3264     vp9_set_quantizer(cm, q);
3265
3266     if (loop_count == 0) setup_frame(cpi);
3267
3268     // Variance adaptive and in frame q adjustment experiments are mutually
3269     // exclusive.
3270     if (cpi->oxcf.aq_mode == VARIANCE_AQ) {
3271       vp9_vaq_frame_setup(cpi);
3272     } else if (cpi->oxcf.aq_mode == EQUATOR360_AQ) {
3273       vp9_360aq_frame_setup(cpi);
3274     } else if (cpi->oxcf.aq_mode == COMPLEXITY_AQ) {
3275       vp9_setup_in_frame_q_adj(cpi);
3276     }
3277
3278     vp9_encode_frame(cpi);
3279
3280     // Update the skip mb flag probabilities based on the distribution
3281     // seen in the last encoder iteration.
3282     // update_base_skip_probs(cpi);
3283
3284     vpx_clear_system_state();
3285
3286     // Dummy pack of the bitstream using up to date stats to get an
3287     // accurate estimate of output frame size to determine if we need
3288     // to recode.
3289     if (cpi->sf.recode_loop >= ALLOW_RECODE_KFARFGF) {
3290       save_coding_context(cpi);
3291       if (!cpi->sf.use_nonrd_pick_mode) vp9_pack_bitstream(cpi, dest, size);
3292
3293       rc->projected_frame_size = (int)(*size) << 3;
3294       restore_coding_context(cpi);
3295
3296       if (frame_over_shoot_limit == 0) frame_over_shoot_limit = 1;
3297     }
3298
3299     if (cpi->oxcf.rc_mode == VPX_Q) {
3300       loop = 0;
3301     } else {
3302       if ((cm->frame_type == KEY_FRAME) && rc->this_key_frame_forced &&
3303           (rc->projected_frame_size < rc->max_frame_bandwidth)) {
3304         int last_q = q;
3305         int64_t kf_err;
3306
3307         int64_t high_err_target = cpi->ambient_err;
3308         int64_t low_err_target = cpi->ambient_err >> 1;
3309
3310 #if CONFIG_VP9_HIGHBITDEPTH
3311         if (cm->use_highbitdepth) {
3312           kf_err = vpx_highbd_get_y_sse(cpi->Source, get_frame_new_buffer(cm));
3313         } else {
3314           kf_err = vpx_get_y_sse(cpi->Source, get_frame_new_buffer(cm));
3315         }
3316 #else
3317         kf_err = vpx_get_y_sse(cpi->Source, get_frame_new_buffer(cm));
3318 #endif  // CONFIG_VP9_HIGHBITDEPTH
3319
3320         // Prevent possible divide by zero error below for perfect KF
3321         kf_err += !kf_err;
3322
3323         // The key frame is not good enough or we can afford
3324         // to make it better without undue risk of popping.
3325         if ((kf_err > high_err_target &&
3326              rc->projected_frame_size <= frame_over_shoot_limit) ||
3327             (kf_err > low_err_target &&
3328              rc->projected_frame_size <= frame_under_shoot_limit)) {
3329           // Lower q_high
3330           q_high = q > q_low ? q - 1 : q_low;
3331
3332           // Adjust Q
3333           q = (int)((q * high_err_target) / kf_err);
3334           q = VPXMIN(q, (q_high + q_low) >> 1);
3335         } else if (kf_err < low_err_target &&
3336                    rc->projected_frame_size >= frame_under_shoot_limit) {
3337           // The key frame is much better than the previous frame
3338           // Raise q_low
3339           q_low = q < q_high ? q + 1 : q_high;
3340
3341           // Adjust Q
3342           q = (int)((q * low_err_target) / kf_err);
3343           q = VPXMIN(q, (q_high + q_low + 1) >> 1);
3344         }
3345
3346         // Clamp Q to upper and lower limits:
3347         q = clamp(q, q_low, q_high);
3348
3349         loop = q != last_q;
3350       } else if (recode_loop_test(cpi, frame_over_shoot_limit,
3351                                   frame_under_shoot_limit, q,
3352                                   VPXMAX(q_high, top_index), bottom_index)) {
3353         // Is the projected frame size out of range and are we allowed
3354         // to attempt to recode.
3355         int last_q = q;
3356         int retries = 0;
3357
3358         if (cpi->resize_pending == 1) {
3359           // Change in frame size so go back around the recode loop.
3360           cpi->rc.frame_size_selector =
3361               SCALE_STEP1 - cpi->rc.frame_size_selector;
3362           cpi->rc.next_frame_size_selector = cpi->rc.frame_size_selector;
3363
3364 #if CONFIG_INTERNAL_STATS
3365           ++cpi->tot_recode_hits;
3366 #endif
3367           ++loop_count;
3368           loop = 1;
3369           continue;
3370         }
3371
3372         // Frame size out of permitted range:
3373         // Update correction factor & compute new Q to try...
3374
3375         // Frame is too large
3376         if (rc->projected_frame_size > rc->this_frame_target) {
3377           // Special case if the projected size is > the max allowed.
3378           if (rc->projected_frame_size >= rc->max_frame_bandwidth)
3379             q_high = rc->worst_quality;
3380
3381           // Raise Qlow as to at least the current value
3382           q_low = q < q_high ? q + 1 : q_high;
3383
3384           if (undershoot_seen || loop_at_this_size > 1) {
3385             // Update rate_correction_factor unless
3386             vp9_rc_update_rate_correction_factors(cpi);
3387
3388             q = (q_high + q_low + 1) / 2;
3389           } else {
3390             // Update rate_correction_factor unless
3391             vp9_rc_update_rate_correction_factors(cpi);
3392
3393             q = vp9_rc_regulate_q(cpi, rc->this_frame_target, bottom_index,
3394                                   VPXMAX(q_high, top_index));
3395
3396             while (q < q_low && retries < 10) {
3397               vp9_rc_update_rate_correction_factors(cpi);
3398               q = vp9_rc_regulate_q(cpi, rc->this_frame_target, bottom_index,
3399                                     VPXMAX(q_high, top_index));
3400               retries++;
3401             }
3402           }
3403
3404           overshoot_seen = 1;
3405         } else {
3406           // Frame is too small
3407           q_high = q > q_low ? q - 1 : q_low;
3408
3409           if (overshoot_seen || loop_at_this_size > 1) {
3410             vp9_rc_update_rate_correction_factors(cpi);
3411             q = (q_high + q_low) / 2;
3412           } else {
3413             vp9_rc_update_rate_correction_factors(cpi);
3414             q = vp9_rc_regulate_q(cpi, rc->this_frame_target, bottom_index,
3415                                   top_index);
3416             // Special case reset for qlow for constrained quality.
3417             // This should only trigger where there is very substantial
3418             // undershoot on a frame and the auto cq level is above
3419             // the user passsed in value.
3420             if (cpi->oxcf.rc_mode == VPX_CQ && q < q_low) {
3421               q_low = q;
3422             }
3423
3424             while (q > q_high && retries < 10) {
3425               vp9_rc_update_rate_correction_factors(cpi);
3426               q = vp9_rc_regulate_q(cpi, rc->this_frame_target, bottom_index,
3427                                     top_index);
3428               retries++;
3429             }
3430           }
3431
3432           undershoot_seen = 1;
3433         }
3434
3435         // Clamp Q to upper and lower limits:
3436         q = clamp(q, q_low, q_high);
3437
3438         loop = (q != last_q);
3439       } else {
3440         loop = 0;
3441       }
3442     }
3443
3444     // Special case for overlay frame.
3445     if (rc->is_src_frame_alt_ref &&
3446         rc->projected_frame_size < rc->max_frame_bandwidth)
3447       loop = 0;
3448
3449     if (loop) {
3450       ++loop_count;
3451       ++loop_at_this_size;
3452
3453 #if CONFIG_INTERNAL_STATS
3454       ++cpi->tot_recode_hits;
3455 #endif
3456     }
3457   } while (loop);
3458 }
3459
3460 static int get_ref_frame_flags(const VP9_COMP *cpi) {
3461   const int *const map = cpi->common.ref_frame_map;
3462   const int gold_is_last = map[cpi->gld_fb_idx] == map[cpi->lst_fb_idx];
3463   const int alt_is_last = map[cpi->alt_fb_idx] == map[cpi->lst_fb_idx];
3464   const int gold_is_alt = map[cpi->gld_fb_idx] == map[cpi->alt_fb_idx];
3465   int flags = VP9_ALT_FLAG | VP9_GOLD_FLAG | VP9_LAST_FLAG;
3466
3467   if (gold_is_last) flags &= ~VP9_GOLD_FLAG;
3468
3469   if (cpi->rc.frames_till_gf_update_due == INT_MAX &&
3470       (cpi->svc.number_temporal_layers == 1 &&
3471        cpi->svc.number_spatial_layers == 1))
3472     flags &= ~VP9_GOLD_FLAG;
3473
3474   if (alt_is_last) flags &= ~VP9_ALT_FLAG;
3475
3476   if (gold_is_alt) flags &= ~VP9_ALT_FLAG;
3477
3478   return flags;
3479 }
3480
3481 static void set_ext_overrides(VP9_COMP *cpi) {
3482   // Overrides the defaults with the externally supplied values with
3483   // vp9_update_reference() and vp9_update_entropy() calls
3484   // Note: The overrides are valid only for the next frame passed
3485   // to encode_frame_to_data_rate() function
3486   if (cpi->ext_refresh_frame_context_pending) {
3487     cpi->common.refresh_frame_context = cpi->ext_refresh_frame_context;
3488     cpi->ext_refresh_frame_context_pending = 0;
3489   }
3490   if (cpi->ext_refresh_frame_flags_pending) {
3491     cpi->refresh_last_frame = cpi->ext_refresh_last_frame;
3492     cpi->refresh_golden_frame = cpi->ext_refresh_golden_frame;
3493     cpi->refresh_alt_ref_frame = cpi->ext_refresh_alt_ref_frame;
3494   }
3495 }
3496
3497 YV12_BUFFER_CONFIG *vp9_svc_twostage_scale(VP9_COMMON *cm,
3498                                            YV12_BUFFER_CONFIG *unscaled,
3499                                            YV12_BUFFER_CONFIG *scaled,
3500                                            YV12_BUFFER_CONFIG *scaled_temp) {
3501   if (cm->mi_cols * MI_SIZE != unscaled->y_width ||
3502       cm->mi_rows * MI_SIZE != unscaled->y_height) {
3503 #if CONFIG_VP9_HIGHBITDEPTH
3504     scale_and_extend_frame(unscaled, scaled_temp, (int)cm->bit_depth);
3505     scale_and_extend_frame(scaled_temp, scaled, (int)cm->bit_depth);
3506 #else
3507     vp9_scale_and_extend_frame(unscaled, scaled_temp);
3508     vp9_scale_and_extend_frame(scaled_temp, scaled);
3509 #endif  // CONFIG_VP9_HIGHBITDEPTH
3510     return scaled;
3511   } else {
3512     return unscaled;
3513   }
3514 }
3515
3516 YV12_BUFFER_CONFIG *vp9_scale_if_required(VP9_COMMON *cm,
3517                                           YV12_BUFFER_CONFIG *unscaled,
3518                                           YV12_BUFFER_CONFIG *scaled,
3519                                           int use_normative_scaler) {
3520   if (cm->mi_cols * MI_SIZE != unscaled->y_width ||
3521       cm->mi_rows * MI_SIZE != unscaled->y_height) {
3522 #if CONFIG_VP9_HIGHBITDEPTH
3523     if (use_normative_scaler && unscaled->y_width <= (scaled->y_width << 1) &&
3524         unscaled->y_height <= (scaled->y_height << 1))
3525       scale_and_extend_frame(unscaled, scaled, (int)cm->bit_depth);
3526     else
3527       scale_and_extend_frame_nonnormative(unscaled, scaled, (int)cm->bit_depth);
3528 #else
3529     if (use_normative_scaler && unscaled->y_width <= (scaled->y_width << 1) &&
3530         unscaled->y_height <= (scaled->y_height << 1))
3531       vp9_scale_and_extend_frame(unscaled, scaled);
3532     else
3533       scale_and_extend_frame_nonnormative(unscaled, scaled);
3534 #endif  // CONFIG_VP9_HIGHBITDEPTH
3535     return scaled;
3536   } else {
3537     return unscaled;
3538   }
3539 }
3540
3541 static void set_arf_sign_bias(VP9_COMP *cpi) {
3542   VP9_COMMON *const cm = &cpi->common;
3543   int arf_sign_bias;
3544
3545   if ((cpi->oxcf.pass == 2) && cpi->multi_arf_allowed) {
3546     const GF_GROUP *const gf_group = &cpi->twopass.gf_group;
3547     arf_sign_bias = cpi->rc.source_alt_ref_active &&
3548                     (!cpi->refresh_alt_ref_frame ||
3549                      (gf_group->rf_level[gf_group->index] == GF_ARF_LOW));
3550   } else {
3551     arf_sign_bias =
3552         (cpi->rc.source_alt_ref_active && !cpi->refresh_alt_ref_frame);
3553   }
3554   cm->ref_frame_sign_bias[ALTREF_FRAME] = arf_sign_bias;
3555 }
3556
3557 static int setup_interp_filter_search_mask(VP9_COMP *cpi) {
3558   INTERP_FILTER ifilter;
3559   int ref_total[MAX_REF_FRAMES] = { 0 };
3560   MV_REFERENCE_FRAME ref;
3561   int mask = 0;
3562   if (cpi->common.last_frame_type == KEY_FRAME || cpi->refresh_alt_ref_frame)
3563     return mask;
3564   for (ref = LAST_FRAME; ref <= ALTREF_FRAME; ++ref)
3565     for (ifilter = EIGHTTAP; ifilter <= EIGHTTAP_SHARP; ++ifilter)
3566       ref_total[ref] += cpi->interp_filter_selected[ref][ifilter];
3567
3568   for (ifilter = EIGHTTAP; ifilter <= EIGHTTAP_SHARP; ++ifilter) {
3569     if ((ref_total[LAST_FRAME] &&
3570          cpi->interp_filter_selected[LAST_FRAME][ifilter] == 0) &&
3571         (ref_total[GOLDEN_FRAME] == 0 ||
3572          cpi->interp_filter_selected[GOLDEN_FRAME][ifilter] * 50 <
3573              ref_total[GOLDEN_FRAME]) &&
3574         (ref_total[ALTREF_FRAME] == 0 ||
3575          cpi->interp_filter_selected[ALTREF_FRAME][ifilter] * 50 <
3576              ref_total[ALTREF_FRAME]))
3577       mask |= 1 << ifilter;
3578   }
3579   return mask;
3580 }
3581
3582 #ifdef ENABLE_KF_DENOISE
3583 // Baseline Kernal weights for denoise
3584 static uint8_t dn_kernal_3[9] = { 1, 2, 1, 2, 4, 2, 1, 2, 1 };
3585 static uint8_t dn_kernal_5[25] = { 1, 1, 1, 1, 1, 1, 1, 2, 1, 1, 1, 2, 4,
3586                                    2, 1, 1, 1, 2, 1, 1, 1, 1, 1, 1, 1 };
3587
3588 static INLINE void add_denoise_point(int centre_val, int data_val, int thresh,
3589                                      uint8_t point_weight, int *sum_val,
3590                                      int *sum_weight) {
3591   if (abs(centre_val - data_val) <= thresh) {
3592     *sum_weight += point_weight;
3593     *sum_val += (int)data_val * (int)point_weight;
3594   }
3595 }
3596
3597 static void spatial_denoise_point(uint8_t *src_ptr, const int stride,
3598                                   const int strength) {
3599   int sum_weight = 0;
3600   int sum_val = 0;
3601   int thresh = strength;
3602   int kernal_size = 5;
3603   int half_k_size = 2;
3604   int i, j;
3605   int max_diff = 0;
3606   uint8_t *tmp_ptr;
3607   uint8_t *kernal_ptr;
3608
3609   // Find the maximum deviation from the source point in the locale.
3610   tmp_ptr = src_ptr - (stride * (half_k_size + 1)) - (half_k_size + 1);
3611   for (i = 0; i < kernal_size + 2; ++i) {
3612     for (j = 0; j < kernal_size + 2; ++j) {
3613       max_diff = VPXMAX(max_diff, abs((int)*src_ptr - (int)tmp_ptr[j]));
3614     }
3615     tmp_ptr += stride;
3616   }
3617
3618   // Select the kernal size.
3619   if (max_diff > (strength + (strength >> 1))) {
3620     kernal_size = 3;
3621     half_k_size = 1;
3622     thresh = thresh >> 1;
3623   }
3624   kernal_ptr = (kernal_size == 3) ? dn_kernal_3 : dn_kernal_5;
3625
3626   // Apply the kernal
3627   tmp_ptr = src_ptr - (stride * half_k_size) - half_k_size;
3628   for (i = 0; i < kernal_size; ++i) {
3629     for (j = 0; j < kernal_size; ++j) {
3630       add_denoise_point((int)*src_ptr, (int)tmp_ptr[j], thresh, *kernal_ptr,
3631                         &sum_val, &sum_weight);
3632       ++kernal_ptr;
3633     }
3634     tmp_ptr += stride;
3635   }
3636
3637   // Update the source value with the new filtered value
3638   *src_ptr = (uint8_t)((sum_val + (sum_weight >> 1)) / sum_weight);
3639 }
3640
3641 #if CONFIG_VP9_HIGHBITDEPTH
3642 static void highbd_spatial_denoise_point(uint16_t *src_ptr, const int stride,
3643                                          const int strength) {
3644   int sum_weight = 0;
3645   int sum_val = 0;
3646   int thresh = strength;
3647   int kernal_size = 5;
3648   int half_k_size = 2;
3649   int i, j;
3650   int max_diff = 0;
3651   uint16_t *tmp_ptr;
3652   uint8_t *kernal_ptr;
3653
3654   // Find the maximum deviation from the source point in the locale.
3655   tmp_ptr = src_ptr - (stride * (half_k_size + 1)) - (half_k_size + 1);
3656   for (i = 0; i < kernal_size + 2; ++i) {
3657     for (j = 0; j < kernal_size + 2; ++j) {
3658       max_diff = VPXMAX(max_diff, abs((int)src_ptr - (int)tmp_ptr[j]));
3659     }
3660     tmp_ptr += stride;
3661   }
3662
3663   // Select the kernal size.
3664   if (max_diff > (strength + (strength >> 1))) {
3665     kernal_size = 3;
3666     half_k_size = 1;
3667     thresh = thresh >> 1;
3668   }
3669   kernal_ptr = (kernal_size == 3) ? dn_kernal_3 : dn_kernal_5;
3670
3671   // Apply the kernal
3672   tmp_ptr = src_ptr - (stride * half_k_size) - half_k_size;
3673   for (i = 0; i < kernal_size; ++i) {
3674     for (j = 0; j < kernal_size; ++j) {
3675       add_denoise_point((int)*src_ptr, (int)tmp_ptr[j], thresh, *kernal_ptr,
3676                         &sum_val, &sum_weight);
3677       ++kernal_ptr;
3678     }
3679     tmp_ptr += stride;
3680   }
3681
3682   // Update the source value with the new filtered value
3683   *src_ptr = (uint16_t)((sum_val + (sum_weight >> 1)) / sum_weight);
3684 }
3685 #endif  // CONFIG_VP9_HIGHBITDEPTH
3686
3687 // Apply thresholded spatial noise supression to a given buffer.
3688 static void spatial_denoise_buffer(VP9_COMP *cpi, uint8_t *buffer,
3689                                    const int stride, const int width,
3690                                    const int height, const int strength) {
3691   VP9_COMMON *const cm = &cpi->common;
3692   uint8_t *src_ptr = buffer;
3693   int row;
3694   int col;
3695
3696   for (row = 0; row < height; ++row) {
3697     for (col = 0; col < width; ++col) {
3698 #if CONFIG_VP9_HIGHBITDEPTH
3699       if (cm->use_highbitdepth)
3700         highbd_spatial_denoise_point(CONVERT_TO_SHORTPTR(&src_ptr[col]), stride,
3701                                      strength);
3702       else
3703         spatial_denoise_point(&src_ptr[col], stride, strength);
3704 #else
3705       spatial_denoise_point(&src_ptr[col], stride, strength);
3706 #endif  // CONFIG_VP9_HIGHBITDEPTH
3707     }
3708     src_ptr += stride;
3709   }
3710 }
3711
3712 // Apply thresholded spatial noise supression to source.
3713 static void spatial_denoise_frame(VP9_COMP *cpi) {
3714   YV12_BUFFER_CONFIG *src = cpi->Source;
3715   const VP9EncoderConfig *const oxcf = &cpi->oxcf;
3716   TWO_PASS *const twopass = &cpi->twopass;
3717   VP9_COMMON *const cm = &cpi->common;
3718
3719   // Base the filter strength on the current active max Q.
3720   const int q = (int)(vp9_convert_qindex_to_q(twopass->active_worst_quality,
3721                                               cm->bit_depth));
3722   int strength =
3723       VPXMAX(oxcf->arnr_strength >> 2, VPXMIN(oxcf->arnr_strength, (q >> 4)));
3724
3725   // Denoise each of Y,U and V buffers.
3726   spatial_denoise_buffer(cpi, src->y_buffer, src->y_stride, src->y_width,
3727                          src->y_height, strength);
3728
3729   strength += (strength >> 1);
3730   spatial_denoise_buffer(cpi, src->u_buffer, src->uv_stride, src->uv_width,
3731                          src->uv_height, strength << 1);
3732
3733   spatial_denoise_buffer(cpi, src->v_buffer, src->uv_stride, src->uv_width,
3734                          src->uv_height, strength << 1);
3735 }
3736 #endif  // ENABLE_KF_DENOISE
3737
3738 static void encode_frame_to_data_rate(VP9_COMP *cpi, size_t *size,
3739                                       uint8_t *dest,
3740                                       unsigned int *frame_flags) {
3741   VP9_COMMON *const cm = &cpi->common;
3742   const VP9EncoderConfig *const oxcf = &cpi->oxcf;
3743   struct segmentation *const seg = &cm->seg;
3744   TX_SIZE t;
3745
3746   set_ext_overrides(cpi);
3747   vpx_clear_system_state();
3748
3749 #ifdef ENABLE_KF_DENOISE
3750   // Spatial denoise of key frame.
3751   if (is_spatial_denoise_enabled(cpi)) spatial_denoise_frame(cpi);
3752 #endif
3753
3754   // Set the arf sign bias for this frame.
3755   set_arf_sign_bias(cpi);
3756
3757   // Set default state for segment based loop filter update flags.
3758   cm->lf.mode_ref_delta_update = 0;
3759
3760   if (cpi->oxcf.pass == 2 && cpi->sf.adaptive_interp_filter_search)
3761     cpi->sf.interp_filter_search_mask = setup_interp_filter_search_mask(cpi);
3762
3763   // Set various flags etc to special state if it is a key frame.
3764   if (frame_is_intra_only(cm)) {
3765     // Reset the loop filter deltas and segmentation map.
3766     vp9_reset_segment_features(&cm->seg);
3767
3768     // If segmentation is enabled force a map update for key frames.
3769     if (seg->enabled) {
3770       seg->update_map = 1;
3771       seg->update_data = 1;
3772     }
3773
3774     // The alternate reference frame cannot be active for a key frame.
3775     cpi->rc.source_alt_ref_active = 0;
3776
3777     cm->error_resilient_mode = oxcf->error_resilient_mode;
3778     cm->frame_parallel_decoding_mode = oxcf->frame_parallel_decoding_mode;
3779
3780     // By default, encoder assumes decoder can use prev_mi.
3781     if (cm->error_resilient_mode) {
3782       cm->frame_parallel_decoding_mode = 1;
3783       cm->reset_frame_context = 0;
3784       cm->refresh_frame_context = 0;
3785     } else if (cm->intra_only) {
3786       // Only reset the current context.
3787       cm->reset_frame_context = 2;
3788     }
3789   }
3790   if (is_two_pass_svc(cpi) && cm->error_resilient_mode == 0) {
3791     // Use context 0 for intra only empty frame, but the last frame context
3792     // for other empty frames.
3793     if (cpi->svc.encode_empty_frame_state == ENCODING) {
3794       if (cpi->svc.encode_intra_empty_frame != 0)
3795         cm->frame_context_idx = 0;
3796       else
3797         cm->frame_context_idx = FRAME_CONTEXTS - 1;
3798     } else {
3799       cm->frame_context_idx =
3800           cpi->svc.spatial_layer_id * cpi->svc.number_temporal_layers +
3801           cpi->svc.temporal_layer_id;
3802     }
3803
3804     cm->frame_parallel_decoding_mode = oxcf->frame_parallel_decoding_mode;
3805
3806     // The probs will be updated based on the frame type of its previous
3807     // frame if frame_parallel_decoding_mode is 0. The type may vary for
3808     // the frame after a key frame in base layer since we may drop enhancement
3809     // layers. So set frame_parallel_decoding_mode to 1 in this case.
3810     if (cm->frame_parallel_decoding_mode == 0) {
3811       if (cpi->svc.number_temporal_layers == 1) {
3812         if (cpi->svc.spatial_layer_id == 0 &&
3813             cpi->svc.layer_context[0].last_frame_type == KEY_FRAME)
3814           cm->frame_parallel_decoding_mode = 1;
3815       } else if (cpi->svc.spatial_layer_id == 0) {
3816         // Find the 2nd frame in temporal base layer and 1st frame in temporal
3817         // enhancement layers from the key frame.
3818         int i;
3819         for (i = 0; i < cpi->svc.number_temporal_layers; ++i) {
3820           if (cpi->svc.layer_context[0].frames_from_key_frame == 1 << i) {
3821             cm->frame_parallel_decoding_mode = 1;
3822             break;
3823           }
3824         }
3825       }
3826     }
3827   }
3828
3829   // For 1 pass CBR, check if we are dropping this frame.
3830   // For spatial layers, for now only check for frame-dropping on first spatial
3831   // layer, and if decision is to drop, we drop whole super-frame.
3832   if (oxcf->pass == 0 && oxcf->rc_mode == VPX_CBR &&
3833       cm->frame_type != KEY_FRAME) {
3834     if (vp9_rc_drop_frame(cpi) ||
3835         (is_one_pass_cbr_svc(cpi) && cpi->svc.rc_drop_superframe == 1)) {
3836       vp9_rc_postencode_update_drop_frame(cpi);
3837       ++cm->current_video_frame;
3838       cpi->ext_refresh_frame_flags_pending = 0;
3839       cpi->svc.rc_drop_superframe = 1;
3840       // TODO(marpan): Advancing the svc counters on dropped frames can break
3841       // the referencing scheme for the fixed svc patterns defined in
3842       // vp9_one_pass_cbr_svc_start_layer(). Look into fixing this issue, but
3843       // for now, don't advance the svc frame counters on dropped frame.
3844       // if (cpi->use_svc)
3845       //   vp9_inc_frame_in_layer(cpi);
3846       return;
3847     }
3848   }
3849
3850   vpx_clear_system_state();
3851
3852 #if CONFIG_INTERNAL_STATS
3853   memset(cpi->mode_chosen_counts, 0,
3854          MAX_MODES * sizeof(*cpi->mode_chosen_counts));
3855 #endif
3856
3857   if (cpi->sf.recode_loop == DISALLOW_RECODE) {
3858     encode_without_recode_loop(cpi, size, dest);
3859   } else {
3860     encode_with_recode_loop(cpi, size, dest);
3861   }
3862
3863 #if CONFIG_VP9_TEMPORAL_DENOISING
3864 #ifdef OUTPUT_YUV_DENOISED
3865   if (oxcf->noise_sensitivity > 0) {
3866     vp9_write_yuv_frame_420(&cpi->denoiser.running_avg_y[INTRA_FRAME],
3867                             yuv_denoised_file);
3868   }
3869 #endif
3870 #endif
3871 #ifdef OUTPUT_YUV_SKINMAP
3872   if (cpi->common.current_video_frame > 1) {
3873     vp9_compute_skin_map(cpi, yuv_skinmap_file);
3874   }
3875 #endif
3876
3877   // Special case code to reduce pulsing when key frames are forced at a
3878   // fixed interval. Note the reconstruction error if it is the frame before
3879   // the force key frame
3880   if (cpi->rc.next_key_frame_forced && cpi->rc.frames_to_key == 1) {
3881 #if CONFIG_VP9_HIGHBITDEPTH
3882     if (cm->use_highbitdepth) {
3883       cpi->ambient_err =
3884           vpx_highbd_get_y_sse(cpi->Source, get_frame_new_buffer(cm));
3885     } else {
3886       cpi->ambient_err = vpx_get_y_sse(cpi->Source, get_frame_new_buffer(cm));
3887     }
3888 #else
3889     cpi->ambient_err = vpx_get_y_sse(cpi->Source, get_frame_new_buffer(cm));
3890 #endif  // CONFIG_VP9_HIGHBITDEPTH
3891   }
3892
3893   // If the encoder forced a KEY_FRAME decision
3894   if (cm->frame_type == KEY_FRAME) cpi->refresh_last_frame = 1;
3895
3896   cm->frame_to_show = get_frame_new_buffer(cm);
3897   cm->frame_to_show->color_space = cm->color_space;
3898   cm->frame_to_show->color_range = cm->color_range;
3899   cm->frame_to_show->render_width = cm->render_width;
3900   cm->frame_to_show->render_height = cm->render_height;
3901
3902   // Pick the loop filter level for the frame.
3903   loopfilter_frame(cpi, cm);
3904
3905   // build the bitstream
3906   vp9_pack_bitstream(cpi, dest, size);
3907
3908   if (cm->seg.update_map) update_reference_segmentation_map(cpi);
3909
3910   if (frame_is_intra_only(cm) == 0) {
3911     release_scaled_references(cpi);
3912   }
3913   vp9_update_reference_frames(cpi);
3914
3915   for (t = TX_4X4; t <= TX_32X32; t++)
3916     full_to_model_counts(cpi->td.counts->coef[t],
3917                          cpi->td.rd_counts.coef_counts[t]);
3918
3919   if (!cm->error_resilient_mode && !cm->frame_parallel_decoding_mode)
3920     vp9_adapt_coef_probs(cm);
3921
3922   if (!frame_is_intra_only(cm)) {
3923     if (!cm->error_resilient_mode && !cm->frame_parallel_decoding_mode) {
3924       vp9_adapt_mode_probs(cm);
3925       vp9_adapt_mv_probs(cm, cm->allow_high_precision_mv);
3926     }
3927   }
3928
3929   cpi->ext_refresh_frame_flags_pending = 0;
3930
3931   if (cpi->refresh_golden_frame == 1)
3932     cpi->frame_flags |= FRAMEFLAGS_GOLDEN;
3933   else
3934     cpi->frame_flags &= ~FRAMEFLAGS_GOLDEN;
3935
3936   if (cpi->refresh_alt_ref_frame == 1)
3937     cpi->frame_flags |= FRAMEFLAGS_ALTREF;
3938   else
3939     cpi->frame_flags &= ~FRAMEFLAGS_ALTREF;
3940
3941   cpi->ref_frame_flags = get_ref_frame_flags(cpi);
3942
3943   cm->last_frame_type = cm->frame_type;
3944
3945   if (!(is_two_pass_svc(cpi) && cpi->svc.encode_empty_frame_state == ENCODING))
3946     vp9_rc_postencode_update(cpi, *size);
3947
3948 #if 0
3949   output_frame_level_debug_stats(cpi);
3950 #endif
3951
3952   if (cm->frame_type == KEY_FRAME) {
3953     // Tell the caller that the frame was coded as a key frame
3954     *frame_flags = cpi->frame_flags | FRAMEFLAGS_KEY;
3955   } else {
3956     *frame_flags = cpi->frame_flags & ~FRAMEFLAGS_KEY;
3957   }
3958
3959   // Clear the one shot update flags for segmentation map and mode/ref loop
3960   // filter deltas.
3961   cm->seg.update_map = 0;
3962   cm->seg.update_data = 0;
3963   cm->lf.mode_ref_delta_update = 0;
3964
3965   // keep track of the last coded dimensions
3966   cm->last_width = cm->width;
3967   cm->last_height = cm->height;
3968
3969   // reset to normal state now that we are done.
3970   if (!cm->show_existing_frame) cm->last_show_frame = cm->show_frame;
3971
3972   if (cm->show_frame) {
3973     vp9_swap_mi_and_prev_mi(cm);
3974     // Don't increment frame counters if this was an altref buffer
3975     // update not a real frame
3976     ++cm->current_video_frame;
3977     if (cpi->use_svc) vp9_inc_frame_in_layer(cpi);
3978   }
3979   cm->prev_frame = cm->cur_frame;
3980
3981   if (cpi->use_svc)
3982     cpi->svc.layer_context[cpi->svc.spatial_layer_id *
3983                                cpi->svc.number_temporal_layers +
3984                            cpi->svc.temporal_layer_id]
3985         .last_frame_type = cm->frame_type;
3986 }
3987
3988 static void SvcEncode(VP9_COMP *cpi, size_t *size, uint8_t *dest,
3989                       unsigned int *frame_flags) {
3990   vp9_rc_get_svc_params(cpi);
3991   encode_frame_to_data_rate(cpi, size, dest, frame_flags);
3992 }
3993
3994 static void Pass0Encode(VP9_COMP *cpi, size_t *size, uint8_t *dest,
3995                         unsigned int *frame_flags) {
3996   if (cpi->oxcf.rc_mode == VPX_CBR) {
3997     vp9_rc_get_one_pass_cbr_params(cpi);
3998   } else {
3999     vp9_rc_get_one_pass_vbr_params(cpi);
4000   }
4001   encode_frame_to_data_rate(cpi, size, dest, frame_flags);
4002 }
4003
4004 static void Pass2Encode(VP9_COMP *cpi, size_t *size, uint8_t *dest,
4005                         unsigned int *frame_flags) {
4006   cpi->allow_encode_breakout = ENCODE_BREAKOUT_ENABLED;
4007   encode_frame_to_data_rate(cpi, size, dest, frame_flags);
4008
4009   if (!(is_two_pass_svc(cpi) && cpi->svc.encode_empty_frame_state == ENCODING))
4010     vp9_twopass_postencode_update(cpi);
4011 }
4012
4013 static void init_ref_frame_bufs(VP9_COMMON *cm) {
4014   int i;
4015   BufferPool *const pool = cm->buffer_pool;
4016   cm->new_fb_idx = INVALID_IDX;
4017   for (i = 0; i < REF_FRAMES; ++i) {
4018     cm->ref_frame_map[i] = INVALID_IDX;
4019     pool->frame_bufs[i].ref_count = 0;
4020   }
4021 }
4022
4023 static void check_initial_width(VP9_COMP *cpi,
4024 #if CONFIG_VP9_HIGHBITDEPTH
4025                                 int use_highbitdepth,
4026 #endif
4027                                 int subsampling_x, int subsampling_y) {
4028   VP9_COMMON *const cm = &cpi->common;
4029
4030   if (!cpi->initial_width ||
4031 #if CONFIG_VP9_HIGHBITDEPTH
4032       cm->use_highbitdepth != use_highbitdepth ||
4033 #endif
4034       cm->subsampling_x != subsampling_x ||
4035       cm->subsampling_y != subsampling_y) {
4036     cm->subsampling_x = subsampling_x;
4037     cm->subsampling_y = subsampling_y;
4038 #if CONFIG_VP9_HIGHBITDEPTH
4039     cm->use_highbitdepth = use_highbitdepth;
4040 #endif
4041
4042     alloc_raw_frame_buffers(cpi);
4043     init_ref_frame_bufs(cm);
4044     alloc_util_frame_buffers(cpi);
4045
4046     init_motion_estimation(cpi);  // TODO(agrange) This can be removed.
4047
4048     cpi->initial_width = cm->width;
4049     cpi->initial_height = cm->height;
4050     cpi->initial_mbs = cm->MBs;
4051   }
4052 }
4053
4054 int vp9_receive_raw_frame(VP9_COMP *cpi, unsigned int frame_flags,
4055                           YV12_BUFFER_CONFIG *sd, int64_t time_stamp,
4056                           int64_t end_time) {
4057   VP9_COMMON *const cm = &cpi->common;
4058   struct vpx_usec_timer timer;
4059   int res = 0;
4060   const int subsampling_x = sd->subsampling_x;
4061   const int subsampling_y = sd->subsampling_y;
4062 #if CONFIG_VP9_HIGHBITDEPTH
4063   const int use_highbitdepth = (sd->flags & YV12_FLAG_HIGHBITDEPTH) != 0;
4064 #endif
4065
4066 #if CONFIG_VP9_HIGHBITDEPTH
4067   check_initial_width(cpi, use_highbitdepth, subsampling_x, subsampling_y);
4068 #else
4069   check_initial_width(cpi, subsampling_x, subsampling_y);
4070 #endif  // CONFIG_VP9_HIGHBITDEPTH
4071
4072 #if CONFIG_VP9_TEMPORAL_DENOISING
4073   setup_denoiser_buffer(cpi);
4074 #endif
4075   vpx_usec_timer_start(&timer);
4076
4077   if (vp9_lookahead_push(cpi->lookahead, sd, time_stamp, end_time,
4078 #if CONFIG_VP9_HIGHBITDEPTH
4079                          use_highbitdepth,
4080 #endif  // CONFIG_VP9_HIGHBITDEPTH
4081                          frame_flags))
4082     res = -1;
4083   vpx_usec_timer_mark(&timer);
4084   cpi->time_receive_data += vpx_usec_timer_elapsed(&timer);
4085
4086   if ((cm->profile == PROFILE_0 || cm->profile == PROFILE_2) &&
4087       (subsampling_x != 1 || subsampling_y != 1)) {
4088     vpx_internal_error(&cm->error, VPX_CODEC_INVALID_PARAM,
4089                        "Non-4:2:0 color format requires profile 1 or 3");
4090     res = -1;
4091   }
4092   if ((cm->profile == PROFILE_1 || cm->profile == PROFILE_3) &&
4093       (subsampling_x == 1 && subsampling_y == 1)) {
4094     vpx_internal_error(&cm->error, VPX_CODEC_INVALID_PARAM,
4095                        "4:2:0 color format requires profile 0 or 2");
4096     res = -1;
4097   }
4098
4099   return res;
4100 }
4101
4102 static int frame_is_reference(const VP9_COMP *cpi) {
4103   const VP9_COMMON *cm = &cpi->common;
4104
4105   return cm->frame_type == KEY_FRAME || cpi->refresh_last_frame ||
4106          cpi->refresh_golden_frame || cpi->refresh_alt_ref_frame ||
4107          cm->refresh_frame_context || cm->lf.mode_ref_delta_update ||
4108          cm->seg.update_map || cm->seg.update_data;
4109 }
4110
4111 static void adjust_frame_rate(VP9_COMP *cpi,
4112                               const struct lookahead_entry *source) {
4113   int64_t this_duration;
4114   int step = 0;
4115
4116   if (source->ts_start == cpi->first_time_stamp_ever) {
4117     this_duration = source->ts_end - source->ts_start;
4118     step = 1;
4119   } else {
4120     int64_t last_duration =
4121         cpi->last_end_time_stamp_seen - cpi->last_time_stamp_seen;
4122
4123     this_duration = source->ts_end - cpi->last_end_time_stamp_seen;
4124
4125     // do a step update if the duration changes by 10%
4126     if (last_duration)
4127       step = (int)((this_duration - last_duration) * 10 / last_duration);
4128   }
4129
4130   if (this_duration) {
4131     if (step) {
4132       vp9_new_framerate(cpi, 10000000.0 / this_duration);
4133     } else {
4134       // Average this frame's rate into the last second's average
4135       // frame rate. If we haven't seen 1 second yet, then average
4136       // over the whole interval seen.
4137       const double interval = VPXMIN(
4138           (double)(source->ts_end - cpi->first_time_stamp_ever), 10000000.0);
4139       double avg_duration = 10000000.0 / cpi->framerate;
4140       avg_duration *= (interval - avg_duration + this_duration);
4141       avg_duration /= interval;
4142
4143       vp9_new_framerate(cpi, 10000000.0 / avg_duration);
4144     }
4145   }
4146   cpi->last_time_stamp_seen = source->ts_start;
4147   cpi->last_end_time_stamp_seen = source->ts_end;
4148 }
4149
4150 // Returns 0 if this is not an alt ref else the offset of the source frame
4151 // used as the arf midpoint.
4152 static int get_arf_src_index(VP9_COMP *cpi) {
4153   RATE_CONTROL *const rc = &cpi->rc;
4154   int arf_src_index = 0;
4155   if (is_altref_enabled(cpi)) {
4156     if (cpi->oxcf.pass == 2) {
4157       const GF_GROUP *const gf_group = &cpi->twopass.gf_group;
4158       if (gf_group->update_type[gf_group->index] == ARF_UPDATE) {
4159         arf_src_index = gf_group->arf_src_offset[gf_group->index];
4160       }
4161     } else if (rc->source_alt_ref_pending) {
4162       arf_src_index = rc->frames_till_gf_update_due;
4163     }
4164   }
4165   return arf_src_index;
4166 }
4167
4168 static void check_src_altref(VP9_COMP *cpi,
4169                              const struct lookahead_entry *source) {
4170   RATE_CONTROL *const rc = &cpi->rc;
4171
4172   if (cpi->oxcf.pass == 2) {
4173     const GF_GROUP *const gf_group = &cpi->twopass.gf_group;
4174     rc->is_src_frame_alt_ref =
4175         (gf_group->update_type[gf_group->index] == OVERLAY_UPDATE);
4176   } else {
4177     rc->is_src_frame_alt_ref =
4178         cpi->alt_ref_source && (source == cpi->alt_ref_source);
4179   }
4180
4181   if (rc->is_src_frame_alt_ref) {
4182     // Current frame is an ARF overlay frame.
4183     cpi->alt_ref_source = NULL;
4184
4185     // Don't refresh the last buffer for an ARF overlay frame. It will
4186     // become the GF so preserve last as an alternative prediction option.
4187     cpi->refresh_last_frame = 0;
4188   }
4189 }
4190
4191 #if CONFIG_INTERNAL_STATS
4192 extern double vp9_get_blockiness(const uint8_t *img1, int img1_pitch,
4193                                  const uint8_t *img2, int img2_pitch, int width,
4194                                  int height);
4195
4196 static void adjust_image_stat(double y, double u, double v, double all,
4197                               ImageStat *s) {
4198   s->stat[Y] += y;
4199   s->stat[U] += u;
4200   s->stat[V] += v;
4201   s->stat[ALL] += all;
4202   s->worst = VPXMIN(s->worst, all);
4203 }
4204 #endif  // CONFIG_INTERNAL_STATS
4205
4206 static void update_level_info(VP9_COMP *cpi, size_t *size, int arf_src_index) {
4207   VP9_COMMON *const cm = &cpi->common;
4208   Vp9LevelInfo *const level_info = &cpi->level_info;
4209   Vp9LevelSpec *const level_spec = &level_info->level_spec;
4210   Vp9LevelStats *const level_stats = &level_info->level_stats;
4211   int i, idx;
4212   uint64_t luma_samples, dur_end;
4213   const uint32_t luma_pic_size = cm->width * cm->height;
4214   double cpb_data_size;
4215
4216   vpx_clear_system_state();
4217
4218   // update level_stats
4219   level_stats->total_compressed_size += *size;
4220   if (cm->show_frame) {
4221     level_stats->total_uncompressed_size +=
4222         luma_pic_size +
4223         2 * (luma_pic_size >> (cm->subsampling_x + cm->subsampling_y));
4224     level_stats->time_encoded =
4225         (cpi->last_end_time_stamp_seen - cpi->first_time_stamp_ever) /
4226         (double)TICKS_PER_SEC;
4227   }
4228
4229   if (arf_src_index > 0) {
4230     if (!level_stats->seen_first_altref) {
4231       level_stats->seen_first_altref = 1;
4232     } else if (level_stats->frames_since_last_altref <
4233                level_spec->min_altref_distance) {
4234       level_spec->min_altref_distance = level_stats->frames_since_last_altref;
4235     }
4236     level_stats->frames_since_last_altref = 0;
4237   } else {
4238     ++level_stats->frames_since_last_altref;
4239   }
4240
4241   if (level_stats->frame_window_buffer.len < FRAME_WINDOW_SIZE - 1) {
4242     idx = (level_stats->frame_window_buffer.start +
4243            level_stats->frame_window_buffer.len++) %
4244           FRAME_WINDOW_SIZE;
4245   } else {
4246     idx = level_stats->frame_window_buffer.start;
4247     level_stats->frame_window_buffer.start = (idx + 1) % FRAME_WINDOW_SIZE;
4248   }
4249   level_stats->frame_window_buffer.buf[idx].ts = cpi->last_time_stamp_seen;
4250   level_stats->frame_window_buffer.buf[idx].size = (uint32_t)(*size);
4251   level_stats->frame_window_buffer.buf[idx].luma_samples = luma_pic_size;
4252
4253   if (cm->frame_type == KEY_FRAME) {
4254     level_stats->ref_refresh_map = 0;
4255   } else {
4256     int count = 0;
4257     level_stats->ref_refresh_map |= vp9_get_refresh_mask(cpi);
4258     // Also need to consider the case where the encoder refers to a buffer
4259     // that has been implicitly refreshed after encoding a keyframe.
4260     if (!cm->intra_only) {
4261       level_stats->ref_refresh_map |= (1 << cpi->lst_fb_idx);
4262       level_stats->ref_refresh_map |= (1 << cpi->gld_fb_idx);
4263       level_stats->ref_refresh_map |= (1 << cpi->alt_fb_idx);
4264     }
4265     for (i = 0; i < REF_FRAMES; ++i) {
4266       count += (level_stats->ref_refresh_map >> i) & 1;
4267     }
4268     if (count > level_spec->max_ref_frame_buffers) {
4269       level_spec->max_ref_frame_buffers = count;
4270     }
4271   }
4272
4273   // update average_bitrate
4274   level_spec->average_bitrate = (double)level_stats->total_compressed_size /
4275                                 125.0 / level_stats->time_encoded;
4276
4277   // update max_luma_sample_rate
4278   luma_samples = 0;
4279   for (i = 0; i < level_stats->frame_window_buffer.len; ++i) {
4280     idx = (level_stats->frame_window_buffer.start +
4281            level_stats->frame_window_buffer.len - 1 - i) %
4282           FRAME_WINDOW_SIZE;
4283     if (i == 0) {
4284       dur_end = level_stats->frame_window_buffer.buf[idx].ts;
4285     }
4286     if (dur_end - level_stats->frame_window_buffer.buf[idx].ts >=
4287         TICKS_PER_SEC) {
4288       break;
4289     }
4290     luma_samples += level_stats->frame_window_buffer.buf[idx].luma_samples;
4291   }
4292   if (luma_samples > level_spec->max_luma_sample_rate) {
4293     level_spec->max_luma_sample_rate = luma_samples;
4294   }
4295
4296   // update max_cpb_size
4297   cpb_data_size = 0;
4298   for (i = 0; i < CPB_WINDOW_SIZE; ++i) {
4299     if (i >= level_stats->frame_window_buffer.len) break;
4300     idx = (level_stats->frame_window_buffer.start +
4301            level_stats->frame_window_buffer.len - 1 - i) %
4302           FRAME_WINDOW_SIZE;
4303     cpb_data_size += level_stats->frame_window_buffer.buf[idx].size;
4304   }
4305   cpb_data_size = cpb_data_size / 125.0;
4306   if (cpb_data_size > level_spec->max_cpb_size) {
4307     level_spec->max_cpb_size = cpb_data_size;
4308   }
4309
4310   // update max_luma_picture_size
4311   if (luma_pic_size > level_spec->max_luma_picture_size) {
4312     level_spec->max_luma_picture_size = luma_pic_size;
4313   }
4314
4315   // update compression_ratio
4316   level_spec->compression_ratio = (double)level_stats->total_uncompressed_size *
4317                                   cm->bit_depth /
4318                                   level_stats->total_compressed_size / 8.0;
4319
4320   // update max_col_tiles
4321   if (level_spec->max_col_tiles < (1 << cm->log2_tile_cols)) {
4322     level_spec->max_col_tiles = (1 << cm->log2_tile_cols);
4323   }
4324 }
4325
4326 int vp9_get_compressed_data(VP9_COMP *cpi, unsigned int *frame_flags,
4327                             size_t *size, uint8_t *dest, int64_t *time_stamp,
4328                             int64_t *time_end, int flush) {
4329   const VP9EncoderConfig *const oxcf = &cpi->oxcf;
4330   VP9_COMMON *const cm = &cpi->common;
4331   BufferPool *const pool = cm->buffer_pool;
4332   RATE_CONTROL *const rc = &cpi->rc;
4333   struct vpx_usec_timer cmptimer;
4334   YV12_BUFFER_CONFIG *force_src_buffer = NULL;
4335   struct lookahead_entry *last_source = NULL;
4336   struct lookahead_entry *source = NULL;
4337   int arf_src_index;
4338   int i;
4339
4340   if (is_two_pass_svc(cpi)) {
4341 #if CONFIG_SPATIAL_SVC
4342     vp9_svc_start_frame(cpi);
4343     // Use a small empty frame instead of a real frame
4344     if (cpi->svc.encode_empty_frame_state == ENCODING)
4345       source = &cpi->svc.empty_frame;
4346 #endif
4347     if (oxcf->pass == 2) vp9_restore_layer_context(cpi);
4348   } else if (is_one_pass_cbr_svc(cpi)) {
4349     vp9_one_pass_cbr_svc_start_layer(cpi);
4350   }
4351
4352   vpx_usec_timer_start(&cmptimer);
4353
4354   vp9_set_high_precision_mv(cpi, ALTREF_HIGH_PRECISION_MV);
4355
4356   // Is multi-arf enabled.
4357   // Note that at the moment multi_arf is only configured for 2 pass VBR and
4358   // will not work properly with svc.
4359   if ((oxcf->pass == 2) && !cpi->use_svc && (cpi->oxcf.enable_auto_arf > 1))
4360     cpi->multi_arf_allowed = 1;
4361   else
4362     cpi->multi_arf_allowed = 0;
4363
4364   // Normal defaults
4365   cm->reset_frame_context = 0;
4366   cm->refresh_frame_context = 1;
4367   if (!is_one_pass_cbr_svc(cpi)) {
4368     cpi->refresh_last_frame = 1;
4369     cpi->refresh_golden_frame = 0;
4370     cpi->refresh_alt_ref_frame = 0;
4371   }
4372
4373   // Should we encode an arf frame.
4374   arf_src_index = get_arf_src_index(cpi);
4375
4376   // Skip alt frame if we encode the empty frame
4377   if (is_two_pass_svc(cpi) && source != NULL) arf_src_index = 0;
4378
4379   if (arf_src_index) {
4380     for (i = 0; i <= arf_src_index; ++i) {
4381       struct lookahead_entry *e = vp9_lookahead_peek(cpi->lookahead, i);
4382       // Avoid creating an alt-ref if there's a forced keyframe pending.
4383       if (e == NULL) {
4384         break;
4385       } else if (e->flags == VPX_EFLAG_FORCE_KF) {
4386         arf_src_index = 0;
4387         flush = 1;
4388         break;
4389       }
4390     }
4391   }
4392
4393   if (arf_src_index) {
4394     assert(arf_src_index <= rc->frames_to_key);
4395
4396     if ((source = vp9_lookahead_peek(cpi->lookahead, arf_src_index)) != NULL) {
4397       cpi->alt_ref_source = source;
4398
4399 #if CONFIG_SPATIAL_SVC
4400       if (is_two_pass_svc(cpi) && cpi->svc.spatial_layer_id > 0) {
4401         int i;
4402         // Reference a hidden frame from a lower layer
4403         for (i = cpi->svc.spatial_layer_id - 1; i >= 0; --i) {
4404           if (oxcf->ss_enable_auto_arf[i]) {
4405             cpi->gld_fb_idx = cpi->svc.layer_context[i].alt_ref_idx;
4406             break;
4407           }
4408         }
4409       }
4410       cpi->svc.layer_context[cpi->svc.spatial_layer_id].has_alt_frame = 1;
4411 #endif
4412
4413       if ((oxcf->arnr_max_frames > 0) && (oxcf->arnr_strength > 0)) {
4414         // Produce the filtered ARF frame.
4415         vp9_temporal_filter(cpi, arf_src_index);
4416         vpx_extend_frame_borders(&cpi->alt_ref_buffer);
4417         force_src_buffer = &cpi->alt_ref_buffer;
4418       }
4419
4420       cm->show_frame = 0;
4421       cm->intra_only = 0;
4422       cpi->refresh_alt_ref_frame = 1;
4423       cpi->refresh_golden_frame = 0;
4424       cpi->refresh_last_frame = 0;
4425       rc->is_src_frame_alt_ref = 0;
4426       rc->source_alt_ref_pending = 0;
4427     } else {
4428       rc->source_alt_ref_pending = 0;
4429     }
4430   }
4431
4432   if (!source) {
4433     // Get last frame source.
4434     if (cm->current_video_frame > 0) {
4435       if ((last_source = vp9_lookahead_peek(cpi->lookahead, -1)) == NULL)
4436         return -1;
4437     }
4438
4439     // Read in the source frame.
4440     if (cpi->use_svc)
4441       source = vp9_svc_lookahead_pop(cpi, cpi->lookahead, flush);
4442     else
4443       source = vp9_lookahead_pop(cpi->lookahead, flush);
4444
4445     if (source != NULL) {
4446       cm->show_frame = 1;
4447       cm->intra_only = 0;
4448       // if the flags indicate intra frame, but if the current picture is for
4449       // non-zero spatial layer, it should not be an intra picture.
4450       // TODO(Won Kap): this needs to change if per-layer intra frame is
4451       // allowed.
4452       if ((source->flags & VPX_EFLAG_FORCE_KF) &&
4453           cpi->svc.spatial_layer_id > cpi->svc.first_spatial_layer_to_encode) {
4454         source->flags &= ~(unsigned int)(VPX_EFLAG_FORCE_KF);
4455       }
4456
4457       // Check to see if the frame should be encoded as an arf overlay.
4458       check_src_altref(cpi, source);
4459     }
4460   }
4461
4462   if (source) {
4463     cpi->un_scaled_source = cpi->Source =
4464         force_src_buffer ? force_src_buffer : &source->img;
4465
4466 #ifdef ENABLE_KF_DENOISE
4467     // Copy of raw source for metrics calculation.
4468     if (is_psnr_calc_enabled(cpi))
4469       vp9_copy_and_extend_frame(cpi->Source, &cpi->raw_unscaled_source);
4470 #endif
4471
4472     cpi->unscaled_last_source = last_source != NULL ? &last_source->img : NULL;
4473
4474     *time_stamp = source->ts_start;
4475     *time_end = source->ts_end;
4476     *frame_flags = (source->flags & VPX_EFLAG_FORCE_KF) ? FRAMEFLAGS_KEY : 0;
4477
4478   } else {
4479     *size = 0;
4480     if (flush && oxcf->pass == 1 && !cpi->twopass.first_pass_done) {
4481       vp9_end_first_pass(cpi); /* get last stats packet */
4482       cpi->twopass.first_pass_done = 1;
4483     }
4484     return -1;
4485   }
4486
4487   if (source->ts_start < cpi->first_time_stamp_ever) {
4488     cpi->first_time_stamp_ever = source->ts_start;
4489     cpi->last_end_time_stamp_seen = source->ts_start;
4490   }
4491
4492   // Clear down mmx registers
4493   vpx_clear_system_state();
4494
4495   // adjust frame rates based on timestamps given
4496   if (cm->show_frame) {
4497     adjust_frame_rate(cpi, source);
4498   }
4499
4500   if (is_one_pass_cbr_svc(cpi)) {
4501     vp9_update_temporal_layer_framerate(cpi);
4502     vp9_restore_layer_context(cpi);
4503   }
4504
4505   // Find a free buffer for the new frame, releasing the reference previously
4506   // held.
4507   if (cm->new_fb_idx != INVALID_IDX) {
4508     --pool->frame_bufs[cm->new_fb_idx].ref_count;
4509   }
4510   cm->new_fb_idx = get_free_fb(cm);
4511
4512   if (cm->new_fb_idx == INVALID_IDX) return -1;
4513
4514   cm->cur_frame = &pool->frame_bufs[cm->new_fb_idx];
4515
4516   if (!cpi->use_svc && cpi->multi_arf_allowed) {
4517     if (cm->frame_type == KEY_FRAME) {
4518       init_buffer_indices(cpi);
4519     } else if (oxcf->pass == 2) {
4520       const GF_GROUP *const gf_group = &cpi->twopass.gf_group;
4521       cpi->alt_fb_idx = gf_group->arf_ref_idx[gf_group->index];
4522     }
4523   }
4524
4525   // Start with a 0 size frame.
4526   *size = 0;
4527
4528   cpi->frame_flags = *frame_flags;
4529
4530   if ((oxcf->pass == 2) &&
4531       (!cpi->use_svc || (is_two_pass_svc(cpi) &&
4532                          cpi->svc.encode_empty_frame_state != ENCODING))) {
4533     vp9_rc_get_second_pass_params(cpi);
4534   } else if (oxcf->pass == 1) {
4535     set_frame_size(cpi);
4536   }
4537
4538   if (cpi->oxcf.pass != 0 || cpi->use_svc || frame_is_intra_only(cm) == 1) {
4539     for (i = 0; i < MAX_REF_FRAMES; ++i) cpi->scaled_ref_idx[i] = INVALID_IDX;
4540   }
4541
4542   if (oxcf->pass == 1 && (!cpi->use_svc || is_two_pass_svc(cpi))) {
4543     const int lossless = is_lossless_requested(oxcf);
4544 #if CONFIG_VP9_HIGHBITDEPTH
4545     if (cpi->oxcf.use_highbitdepth)
4546       cpi->td.mb.fwd_txm4x4 =
4547           lossless ? vp9_highbd_fwht4x4 : vpx_highbd_fdct4x4;
4548     else
4549       cpi->td.mb.fwd_txm4x4 = lossless ? vp9_fwht4x4 : vpx_fdct4x4;
4550     cpi->td.mb.highbd_itxm_add =
4551         lossless ? vp9_highbd_iwht4x4_add : vp9_highbd_idct4x4_add;
4552 #else
4553     cpi->td.mb.fwd_txm4x4 = lossless ? vp9_fwht4x4 : vpx_fdct4x4;
4554 #endif  // CONFIG_VP9_HIGHBITDEPTH
4555     cpi->td.mb.itxm_add = lossless ? vp9_iwht4x4_add : vp9_idct4x4_add;
4556     vp9_first_pass(cpi, source);
4557   } else if (oxcf->pass == 2 && (!cpi->use_svc || is_two_pass_svc(cpi))) {
4558     Pass2Encode(cpi, size, dest, frame_flags);
4559   } else if (cpi->use_svc) {
4560     SvcEncode(cpi, size, dest, frame_flags);
4561   } else {
4562     // One pass encode
4563     Pass0Encode(cpi, size, dest, frame_flags);
4564   }
4565
4566   if (cm->refresh_frame_context)
4567     cm->frame_contexts[cm->frame_context_idx] = *cm->fc;
4568
4569   // No frame encoded, or frame was dropped, release scaled references.
4570   if ((*size == 0) && (frame_is_intra_only(cm) == 0)) {
4571     release_scaled_references(cpi);
4572   }
4573
4574   if (*size > 0) {
4575     cpi->droppable = !frame_is_reference(cpi);
4576   }
4577
4578   // Save layer specific state.
4579   if (is_one_pass_cbr_svc(cpi) || ((cpi->svc.number_temporal_layers > 1 ||
4580                                     cpi->svc.number_spatial_layers > 1) &&
4581                                    oxcf->pass == 2)) {
4582     vp9_save_layer_context(cpi);
4583   }
4584
4585   vpx_usec_timer_mark(&cmptimer);
4586   cpi->time_compress_data += vpx_usec_timer_elapsed(&cmptimer);
4587
4588   // Should we calculate metrics for the frame.
4589   if (is_psnr_calc_enabled(cpi)) generate_psnr_packet(cpi);
4590
4591   if (cpi->keep_level_stats && oxcf->pass != 1)
4592     update_level_info(cpi, size, arf_src_index);
4593
4594 #if CONFIG_INTERNAL_STATS
4595
4596   if (oxcf->pass != 1) {
4597     double samples = 0.0;
4598     cpi->bytes += (int)(*size);
4599
4600     if (cm->show_frame) {
4601       uint32_t bit_depth = 8;
4602       uint32_t in_bit_depth = 8;
4603       cpi->count++;
4604 #if CONFIG_VP9_HIGHBITDEPTH
4605       if (cm->use_highbitdepth) {
4606         in_bit_depth = cpi->oxcf.input_bit_depth;
4607         bit_depth = cm->bit_depth;
4608       }
4609 #endif
4610
4611       if (cpi->b_calculate_psnr) {
4612         YV12_BUFFER_CONFIG *orig = cpi->raw_source_frame;
4613         YV12_BUFFER_CONFIG *recon = cpi->common.frame_to_show;
4614         YV12_BUFFER_CONFIG *pp = &cm->post_proc_buffer;
4615         PSNR_STATS psnr;
4616 #if CONFIG_VP9_HIGHBITDEPTH
4617         vpx_calc_highbd_psnr(orig, recon, &psnr, cpi->td.mb.e_mbd.bd,
4618                              in_bit_depth);
4619 #else
4620         vpx_calc_psnr(orig, recon, &psnr);
4621 #endif  // CONFIG_VP9_HIGHBITDEPTH
4622
4623         adjust_image_stat(psnr.psnr[1], psnr.psnr[2], psnr.psnr[3],
4624                           psnr.psnr[0], &cpi->psnr);
4625         cpi->total_sq_error += psnr.sse[0];
4626         cpi->total_samples += psnr.samples[0];
4627         samples = psnr.samples[0];
4628
4629         {
4630           PSNR_STATS psnr2;
4631           double frame_ssim2 = 0, weight = 0;
4632 #if CONFIG_VP9_POSTPROC
4633           if (vpx_alloc_frame_buffer(
4634                   pp, recon->y_crop_width, recon->y_crop_height,
4635                   cm->subsampling_x, cm->subsampling_y,
4636 #if CONFIG_VP9_HIGHBITDEPTH
4637                   cm->use_highbitdepth,
4638 #endif
4639                   VP9_ENC_BORDER_IN_PIXELS, cm->byte_alignment) < 0) {
4640             vpx_internal_error(&cm->error, VPX_CODEC_MEM_ERROR,
4641                                "Failed to allocate post processing buffer");
4642           }
4643           {
4644             vp9_ppflags_t ppflags;
4645             ppflags.post_proc_flag = VP9D_DEBLOCK;
4646             ppflags.deblocking_level = 0;  // not used in vp9_post_proc_frame()
4647             ppflags.noise_level = 0;       // not used in vp9_post_proc_frame()
4648             vp9_post_proc_frame(cm, pp, &ppflags);
4649           }
4650 #endif
4651           vpx_clear_system_state();
4652
4653 #if CONFIG_VP9_HIGHBITDEPTH
4654           vpx_calc_highbd_psnr(orig, pp, &psnr2, cpi->td.mb.e_mbd.bd,
4655                                cpi->oxcf.input_bit_depth);
4656 #else
4657           vpx_calc_psnr(orig, pp, &psnr2);
4658 #endif  // CONFIG_VP9_HIGHBITDEPTH
4659
4660           cpi->totalp_sq_error += psnr2.sse[0];
4661           cpi->totalp_samples += psnr2.samples[0];
4662           adjust_image_stat(psnr2.psnr[1], psnr2.psnr[2], psnr2.psnr[3],
4663                             psnr2.psnr[0], &cpi->psnrp);
4664
4665 #if CONFIG_VP9_HIGHBITDEPTH
4666           if (cm->use_highbitdepth) {
4667             frame_ssim2 = vpx_highbd_calc_ssim(orig, recon, &weight, bit_depth,
4668                                                in_bit_depth);
4669           } else {
4670             frame_ssim2 = vpx_calc_ssim(orig, recon, &weight);
4671           }
4672 #else
4673           frame_ssim2 = vpx_calc_ssim(orig, recon, &weight);
4674 #endif  // CONFIG_VP9_HIGHBITDEPTH
4675
4676           cpi->worst_ssim = VPXMIN(cpi->worst_ssim, frame_ssim2);
4677           cpi->summed_quality += frame_ssim2 * weight;
4678           cpi->summed_weights += weight;
4679
4680 #if CONFIG_VP9_HIGHBITDEPTH
4681           if (cm->use_highbitdepth) {
4682             frame_ssim2 = vpx_highbd_calc_ssim(orig, pp, &weight, bit_depth,
4683                                                in_bit_depth);
4684           } else {
4685             frame_ssim2 = vpx_calc_ssim(orig, pp, &weight);
4686           }
4687 #else
4688           frame_ssim2 = vpx_calc_ssim(orig, pp, &weight);
4689 #endif  // CONFIG_VP9_HIGHBITDEPTH
4690
4691           cpi->summedp_quality += frame_ssim2 * weight;
4692           cpi->summedp_weights += weight;
4693 #if 0
4694           {
4695             FILE *f = fopen("q_used.stt", "a");
4696             fprintf(f, "%5d : Y%f7.3:U%f7.3:V%f7.3:F%f7.3:S%7.3f\n",
4697                     cpi->common.current_video_frame, y2, u2, v2,
4698                     frame_psnr2, frame_ssim2);
4699             fclose(f);
4700           }
4701 #endif
4702         }
4703       }
4704       if (cpi->b_calculate_blockiness) {
4705 #if CONFIG_VP9_HIGHBITDEPTH
4706         if (!cm->use_highbitdepth)
4707 #endif
4708         {
4709           double frame_blockiness = vp9_get_blockiness(
4710               cpi->Source->y_buffer, cpi->Source->y_stride,
4711               cm->frame_to_show->y_buffer, cm->frame_to_show->y_stride,
4712               cpi->Source->y_width, cpi->Source->y_height);
4713           cpi->worst_blockiness =
4714               VPXMAX(cpi->worst_blockiness, frame_blockiness);
4715           cpi->total_blockiness += frame_blockiness;
4716         }
4717       }
4718
4719       if (cpi->b_calculate_consistency) {
4720 #if CONFIG_VP9_HIGHBITDEPTH
4721         if (!cm->use_highbitdepth)
4722 #endif
4723         {
4724           double this_inconsistency = vpx_get_ssim_metrics(
4725               cpi->Source->y_buffer, cpi->Source->y_stride,
4726               cm->frame_to_show->y_buffer, cm->frame_to_show->y_stride,
4727               cpi->Source->y_width, cpi->Source->y_height, cpi->ssim_vars,
4728               &cpi->metrics, 1);
4729
4730           const double peak = (double)((1 << cpi->oxcf.input_bit_depth) - 1);
4731           double consistency =
4732               vpx_sse_to_psnr(samples, peak, (double)cpi->total_inconsistency);
4733           if (consistency > 0.0)
4734             cpi->worst_consistency =
4735                 VPXMIN(cpi->worst_consistency, consistency);
4736           cpi->total_inconsistency += this_inconsistency;
4737         }
4738       }
4739
4740       {
4741         double y, u, v, frame_all;
4742         frame_all = vpx_calc_fastssim(cpi->Source, cm->frame_to_show, &y, &u,
4743                                       &v, bit_depth, in_bit_depth);
4744         adjust_image_stat(y, u, v, frame_all, &cpi->fastssim);
4745       }
4746       {
4747         double y, u, v, frame_all;
4748         frame_all = vpx_psnrhvs(cpi->Source, cm->frame_to_show, &y, &u, &v,
4749                                 bit_depth, in_bit_depth);
4750         adjust_image_stat(y, u, v, frame_all, &cpi->psnrhvs);
4751       }
4752     }
4753   }
4754
4755 #endif
4756
4757   if (is_two_pass_svc(cpi)) {
4758     if (cpi->svc.encode_empty_frame_state == ENCODING) {
4759       cpi->svc.encode_empty_frame_state = ENCODED;
4760       cpi->svc.encode_intra_empty_frame = 0;
4761     }
4762
4763     if (cm->show_frame) {
4764       ++cpi->svc.spatial_layer_to_encode;
4765       if (cpi->svc.spatial_layer_to_encode >= cpi->svc.number_spatial_layers)
4766         cpi->svc.spatial_layer_to_encode = 0;
4767
4768       // May need the empty frame after an visible frame.
4769       cpi->svc.encode_empty_frame_state = NEED_TO_ENCODE;
4770     }
4771   } else if (is_one_pass_cbr_svc(cpi)) {
4772     if (cm->show_frame) {
4773       ++cpi->svc.spatial_layer_to_encode;
4774       if (cpi->svc.spatial_layer_to_encode >= cpi->svc.number_spatial_layers)
4775         cpi->svc.spatial_layer_to_encode = 0;
4776     }
4777   }
4778   vpx_clear_system_state();
4779   return 0;
4780 }
4781
4782 int vp9_get_preview_raw_frame(VP9_COMP *cpi, YV12_BUFFER_CONFIG *dest,
4783                               vp9_ppflags_t *flags) {
4784   VP9_COMMON *cm = &cpi->common;
4785 #if !CONFIG_VP9_POSTPROC
4786   (void)flags;
4787 #endif
4788
4789   if (!cm->show_frame) {
4790     return -1;
4791   } else {
4792     int ret;
4793 #if CONFIG_VP9_POSTPROC
4794     ret = vp9_post_proc_frame(cm, dest, flags);
4795 #else
4796     if (cm->frame_to_show) {
4797       *dest = *cm->frame_to_show;
4798       dest->y_width = cm->width;
4799       dest->y_height = cm->height;
4800       dest->uv_width = cm->width >> cm->subsampling_x;
4801       dest->uv_height = cm->height >> cm->subsampling_y;
4802       ret = 0;
4803     } else {
4804       ret = -1;
4805     }
4806 #endif  // !CONFIG_VP9_POSTPROC
4807     vpx_clear_system_state();
4808     return ret;
4809   }
4810 }
4811
4812 int vp9_set_internal_size(VP9_COMP *cpi, VPX_SCALING horiz_mode,
4813                           VPX_SCALING vert_mode) {
4814   VP9_COMMON *cm = &cpi->common;
4815   int hr = 0, hs = 0, vr = 0, vs = 0;
4816
4817   if (horiz_mode > ONETWO || vert_mode > ONETWO) return -1;
4818
4819   Scale2Ratio(horiz_mode, &hr, &hs);
4820   Scale2Ratio(vert_mode, &vr, &vs);
4821
4822   // always go to the next whole number
4823   cm->width = (hs - 1 + cpi->oxcf.width * hr) / hs;
4824   cm->height = (vs - 1 + cpi->oxcf.height * vr) / vs;
4825   if (cm->current_video_frame) {
4826     assert(cm->width <= cpi->initial_width);
4827     assert(cm->height <= cpi->initial_height);
4828   }
4829
4830   update_frame_size(cpi);
4831
4832   return 0;
4833 }
4834
4835 int vp9_set_size_literal(VP9_COMP *cpi, unsigned int width,
4836                          unsigned int height) {
4837   VP9_COMMON *cm = &cpi->common;
4838 #if CONFIG_VP9_HIGHBITDEPTH
4839   check_initial_width(cpi, cm->use_highbitdepth, 1, 1);
4840 #else
4841   check_initial_width(cpi, 1, 1);
4842 #endif  // CONFIG_VP9_HIGHBITDEPTH
4843
4844 #if CONFIG_VP9_TEMPORAL_DENOISING
4845   setup_denoiser_buffer(cpi);
4846 #endif
4847
4848   if (width) {
4849     cm->width = width;
4850     if (cm->width > cpi->initial_width) {
4851       cm->width = cpi->initial_width;
4852       printf("Warning: Desired width too large, changed to %d\n", cm->width);
4853     }
4854   }
4855
4856   if (height) {
4857     cm->height = height;
4858     if (cm->height > cpi->initial_height) {
4859       cm->height = cpi->initial_height;
4860       printf("Warning: Desired height too large, changed to %d\n", cm->height);
4861     }
4862   }
4863   assert(cm->width <= cpi->initial_width);
4864   assert(cm->height <= cpi->initial_height);
4865
4866   update_frame_size(cpi);
4867
4868   return 0;
4869 }
4870
4871 void vp9_set_svc(VP9_COMP *cpi, int use_svc) {
4872   cpi->use_svc = use_svc;
4873   return;
4874 }
4875
4876 int vp9_get_quantizer(VP9_COMP *cpi) { return cpi->common.base_qindex; }
4877
4878 void vp9_apply_encoding_flags(VP9_COMP *cpi, vpx_enc_frame_flags_t flags) {
4879   if (flags &
4880       (VP8_EFLAG_NO_REF_LAST | VP8_EFLAG_NO_REF_GF | VP8_EFLAG_NO_REF_ARF)) {
4881     int ref = 7;
4882
4883     if (flags & VP8_EFLAG_NO_REF_LAST) ref ^= VP9_LAST_FLAG;
4884
4885     if (flags & VP8_EFLAG_NO_REF_GF) ref ^= VP9_GOLD_FLAG;
4886
4887     if (flags & VP8_EFLAG_NO_REF_ARF) ref ^= VP9_ALT_FLAG;
4888
4889     vp9_use_as_reference(cpi, ref);
4890   }
4891
4892   if (flags &
4893       (VP8_EFLAG_NO_UPD_LAST | VP8_EFLAG_NO_UPD_GF | VP8_EFLAG_NO_UPD_ARF |
4894        VP8_EFLAG_FORCE_GF | VP8_EFLAG_FORCE_ARF)) {
4895     int upd = 7;
4896
4897     if (flags & VP8_EFLAG_NO_UPD_LAST) upd ^= VP9_LAST_FLAG;
4898
4899     if (flags & VP8_EFLAG_NO_UPD_GF) upd ^= VP9_GOLD_FLAG;
4900
4901     if (flags & VP8_EFLAG_NO_UPD_ARF) upd ^= VP9_ALT_FLAG;
4902
4903     vp9_update_reference(cpi, upd);
4904   }
4905
4906   if (flags & VP8_EFLAG_NO_UPD_ENTROPY) {
4907     vp9_update_entropy(cpi, 0);
4908   }
4909 }