vp9_encodeframe.c 122 KB
Newer Older
John Koleszar's avatar
John Koleszar committed
1
/*
2
 *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
John Koleszar's avatar
John Koleszar committed
3
 *
4
 *  Use of this source code is governed by a BSD-style license
5 6
 *  that can be found in the LICENSE file in the root of the source
 *  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.
John Koleszar's avatar
John Koleszar committed
9 10
 */

Dmitry Kovalev's avatar
Dmitry Kovalev committed
11 12 13 14
#include <limits.h>
#include <math.h>
#include <stdio.h>

Jim Bankoski's avatar
Jim Bankoski committed
15
#include "./vp9_rtcd.h"
Dmitry Kovalev's avatar
Dmitry Kovalev committed
16 17 18 19
#include "./vpx_config.h"

#include "vpx_ports/vpx_timer.h"

20
#include "vp9/common/vp9_common.h"
Yaowu Xu's avatar
Yaowu Xu committed
21
#include "vp9/common/vp9_entropy.h"
22
#include "vp9/common/vp9_entropymode.h"
23
#include "vp9/common/vp9_idct.h"
Dmitry Kovalev's avatar
Dmitry Kovalev committed
24 25 26
#include "vp9/common/vp9_mvref_common.h"
#include "vp9/common/vp9_pred_common.h"
#include "vp9/common/vp9_quant_common.h"
27
#include "vp9/common/vp9_reconintra.h"
Dmitry Kovalev's avatar
Dmitry Kovalev committed
28
#include "vp9/common/vp9_reconinter.h"
29
#include "vp9/common/vp9_seg_common.h"
30
#include "vp9/common/vp9_systemdependent.h"
31
#include "vp9/common/vp9_tile_common.h"
32

33
#include "vp9/encoder/vp9_aq_complexity.h"
Marco Paniconi's avatar
Marco Paniconi committed
34 35
#include "vp9/encoder/vp9_aq_cyclicrefresh.h"
#include "vp9/encoder/vp9_aq_variance.h"
Dmitry Kovalev's avatar
Dmitry Kovalev committed
36 37 38
#include "vp9/encoder/vp9_encodeframe.h"
#include "vp9/encoder/vp9_encodemb.h"
#include "vp9/encoder/vp9_encodemv.h"
39
#include "vp9/encoder/vp9_extend.h"
40
#include "vp9/encoder/vp9_pickmode.h"
Dmitry Kovalev's avatar
Dmitry Kovalev committed
41 42
#include "vp9/encoder/vp9_rdopt.h"
#include "vp9/encoder/vp9_segmentation.h"
43
#include "vp9/encoder/vp9_tokenize.h"
44

45 46 47 48 49 50
#define GF_ZEROMV_ZBIN_BOOST 0
#define LF_ZEROMV_ZBIN_BOOST 0
#define MV_ZBIN_BOOST        0
#define SPLIT_MV_ZBIN_BOOST  0
#define INTRA_ZBIN_BOOST     0

51
static INLINE uint8_t *get_sb_index(MACROBLOCK *x, BLOCK_SIZE subsize) {
52 53 54 55 56
  switch (subsize) {
    case BLOCK_64X64:
    case BLOCK_64X32:
    case BLOCK_32X64:
    case BLOCK_32X32:
57
      return &x->sb_index;
58 59 60
    case BLOCK_32X16:
    case BLOCK_16X32:
    case BLOCK_16X16:
61
      return &x->mb_index;
62 63 64
    case BLOCK_16X8:
    case BLOCK_8X16:
    case BLOCK_8X8:
65
      return &x->b_index;
66 67 68
    case BLOCK_8X4:
    case BLOCK_4X8:
    case BLOCK_4X4:
69
      return &x->ab_index;
70 71 72 73 74 75
    default:
      assert(0);
      return NULL;
  }
}

Jim Bankoski's avatar
Jim Bankoski committed
76
static void encode_superblock(VP9_COMP *cpi, TOKENEXTRA **t, int output_enabled,
77
                              int mi_row, int mi_col, BLOCK_SIZE bsize);
78

79
static void adjust_act_zbin(VP9_COMP *cpi, MACROBLOCK *x);
80

81 82 83 84
// activity_avg must be positive, or flat regions could get a zero weight
//  (infinite lambda), which confounds analysis.
// This also avoids the need for divide by zero checks in
//  vp9_activity_masking().
85
#define ACTIVITY_AVG_MIN 64
86

87
// Motion vector component magnitude threshold for defining fast motion.
88
#define FAST_MOTION_MV_THRESH 24
89

90 91 92 93
// This is used as a reference when computing the source variance for the
//  purposes of activity masking.
// Eventually this should be replaced by custom no-reference routines,
//  which will be faster.
94 95 96 97 98 99 100 101 102 103 104
static const uint8_t VP9_VAR_OFFS[64] = {
  128, 128, 128, 128, 128, 128, 128, 128,
  128, 128, 128, 128, 128, 128, 128, 128,
  128, 128, 128, 128, 128, 128, 128, 128,
  128, 128, 128, 128, 128, 128, 128, 128,
  128, 128, 128, 128, 128, 128, 128, 128,
  128, 128, 128, 128, 128, 128, 128, 128,
  128, 128, 128, 128, 128, 128, 128, 128,
  128, 128, 128, 128, 128, 128, 128, 128
};

105 106
static unsigned int get_sby_perpixel_variance(VP9_COMP *cpi,
                                              MACROBLOCK *x,
107
                                              BLOCK_SIZE bs) {
108
  unsigned int var, sse;
109
  var = cpi->fn_ptr[bs].vf(x->plane[0].src.buf, x->plane[0].src.stride,
110
                           VP9_VAR_OFFS, 0, &sse);
111
  return ROUND_POWER_OF_TWO(var, num_pels_log2_lookup[bs]);
112 113
}

114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136
static unsigned int get_sby_perpixel_diff_variance(VP9_COMP *cpi,
                                                   MACROBLOCK *x,
                                                   int mi_row,
                                                   int mi_col,
                                                   BLOCK_SIZE bs) {
  const YV12_BUFFER_CONFIG *yv12 = get_ref_frame_buffer(cpi, LAST_FRAME);
  int offset = (mi_row * MI_SIZE) * yv12->y_stride + (mi_col * MI_SIZE);
  unsigned int var, sse;
  var = cpi->fn_ptr[bs].vf(x->plane[0].src.buf,
                           x->plane[0].src.stride,
                           yv12->y_buffer + offset,
                           yv12->y_stride,
                           &sse);
  return ROUND_POWER_OF_TWO(var, num_pels_log2_lookup[bs]);
}

static BLOCK_SIZE get_rd_var_based_fixed_partition(VP9_COMP *cpi,
                                                   int mi_row,
                                                   int mi_col) {
  unsigned int var = get_sby_perpixel_diff_variance(cpi, &cpi->mb,
                                                    mi_row, mi_col,
                                                    BLOCK_64X64);
  if (var < 8)
137
    return BLOCK_64X64;
138
  else if (var < 128)
139
    return BLOCK_32X32;
140 141 142 143
  else if (var < 2048)
    return BLOCK_16X16;
  else
    return BLOCK_8X8;
144 145
}

146 147 148 149 150 151
static BLOCK_SIZE get_nonrd_var_based_fixed_partition(VP9_COMP *cpi,
                                                      int mi_row,
                                                      int mi_col) {
  unsigned int var = get_sby_perpixel_diff_variance(cpi, &cpi->mb,
                                                    mi_row, mi_col,
                                                    BLOCK_64X64);
152
  if (var < 4)
153
    return BLOCK_64X64;
154
  else if (var < 10)
155 156
    return BLOCK_32X32;
  else
157
    return BLOCK_16X16;
158 159
}

160 161
// Lighter version of set_offsets that only sets the mode info
// pointers.
162
static INLINE void set_modeinfo_offsets(VP9_COMMON *const cm,
163 164 165
                                        MACROBLOCKD *const xd,
                                        int mi_row,
                                        int mi_col) {
166 167 168
  const int idx_str = xd->mi_stride * mi_row + mi_col;
  xd->mi = cm->mi_grid_visible + idx_str;
  xd->mi[0] = cm->mi + idx_str;
169 170
}

Alex Converse's avatar
Alex Converse committed
171
static int is_block_in_mb_map(const VP9_COMP *cpi, int mi_row, int mi_col,
172
                              BLOCK_SIZE bsize) {
Alex Converse's avatar
Alex Converse committed
173
  const VP9_COMMON *const cm = &cpi->common;
174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196
  const int mb_rows = cm->mb_rows;
  const int mb_cols = cm->mb_cols;
  const int mb_row = mi_row >> 1;
  const int mb_col = mi_col >> 1;
  const int mb_width = num_8x8_blocks_wide_lookup[bsize] >> 1;
  const int mb_height = num_8x8_blocks_high_lookup[bsize] >> 1;
  int r, c;
  if (bsize <= BLOCK_16X16) {
    return cpi->active_map[mb_row * mb_cols + mb_col];
  }
  for (r = 0; r < mb_height; ++r) {
    for (c = 0; c < mb_width; ++c) {
      int row = mb_row + r;
      int col = mb_col + c;
      if (row >= mb_rows || col >= mb_cols)
        continue;
      if (cpi->active_map[row * mb_cols + col])
        return 1;
    }
  }
  return 0;
}

Alex Converse's avatar
Alex Converse committed
197 198 199 200 201 202 203 204 205 206
static int check_active_map(const VP9_COMP *cpi, const MACROBLOCK *x,
                            int mi_row, int mi_col,
                            BLOCK_SIZE bsize) {
  if (cpi->active_map_enabled && !x->e_mbd.lossless) {
    return is_block_in_mb_map(cpi, mi_row, mi_col, bsize);
  } else {
    return 1;
  }
}

207 208 209 210 211 212 213 214 215 216 217 218 219
static void set_offsets(VP9_COMP *cpi, const TileInfo *const tile,
                        int mi_row, int mi_col, BLOCK_SIZE bsize) {
  MACROBLOCK *const x = &cpi->mb;
  VP9_COMMON *const cm = &cpi->common;
  MACROBLOCKD *const xd = &x->e_mbd;
  MB_MODE_INFO *mbmi;
  const int mi_width = num_8x8_blocks_wide_lookup[bsize];
  const int mi_height = num_8x8_blocks_high_lookup[bsize];
  const int mb_row = mi_row >> 1;
  const int mb_col = mi_col >> 1;
  const int idx_map = mb_row * cm->mb_cols + mb_col;
  const struct segmentation *const seg = &cm->seg;

220
  set_skip_context(xd, mi_row, mi_col);
221 222 223

  // Activity map pointer
  x->mb_activity_ptr = &cpi->mb_activity_map[idx_map];
Alex Converse's avatar
Alex Converse committed
224
  x->in_active_map = check_active_map(cpi, x, mi_row, mi_col, bsize);
225 226 227

  set_modeinfo_offsets(cm, xd, mi_row, mi_col);

228
  mbmi = &xd->mi[0]->mbmi;
229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283

  // Set up destination pointers.
  vp9_setup_dst_planes(xd, get_frame_new_buffer(cm), mi_row, mi_col);

  // Set up limit values for MV components.
  // Mv beyond the range do not produce new/different prediction block.
  x->mv_row_min = -(((mi_row + mi_height) * MI_SIZE) + VP9_INTERP_EXTEND);
  x->mv_col_min = -(((mi_col + mi_width) * MI_SIZE) + VP9_INTERP_EXTEND);
  x->mv_row_max = (cm->mi_rows - mi_row) * MI_SIZE + VP9_INTERP_EXTEND;
  x->mv_col_max = (cm->mi_cols - mi_col) * MI_SIZE + VP9_INTERP_EXTEND;

  // Set up distance of MB to edge of frame in 1/8th pel units.
  assert(!(mi_col & (mi_width - 1)) && !(mi_row & (mi_height - 1)));
  set_mi_row_col(xd, tile, mi_row, mi_height, mi_col, mi_width,
                 cm->mi_rows, cm->mi_cols);

  // Set up source buffers.
  vp9_setup_src_planes(x, cpi->Source, mi_row, mi_col);

  // R/D setup.
  x->rddiv = cpi->RDDIV;
  x->rdmult = cpi->RDMULT;

  // Setup segment ID.
  if (seg->enabled) {
    if (cpi->oxcf.aq_mode != VARIANCE_AQ) {
      const uint8_t *const map = seg->update_map ? cpi->segmentation_map
                                                 : cm->last_frame_seg_map;
      mbmi->segment_id = vp9_get_segment_id(cm, map, bsize, mi_row, mi_col);
    }
    vp9_init_plane_quantizers(cpi, x);

    if (seg->enabled && cpi->seg0_cnt > 0 &&
        !vp9_segfeature_active(seg, 0, SEG_LVL_REF_FRAME) &&
        vp9_segfeature_active(seg, 1, SEG_LVL_REF_FRAME)) {
      cpi->seg0_progress = (cpi->seg0_idx << 16) / cpi->seg0_cnt;
    } else {
      const int y = mb_row & ~3;
      const int x = mb_col & ~3;
      const int p16 = ((mb_row & 1) << 1) + (mb_col & 1);
      const int p32 = ((mb_row & 2) << 2) + ((mb_col & 2) << 1);
      const int tile_progress = tile->mi_col_start * cm->mb_rows >> 1;
      const int mb_cols = (tile->mi_col_end - tile->mi_col_start) >> 1;

      cpi->seg0_progress = ((y * mb_cols + x * 4 + p32 + p16 + tile_progress)
          << 16) / cm->MBs;
    }

    x->encode_breakout = cpi->segment_encode_breakout[mbmi->segment_id];
  } else {
    mbmi->segment_id = 0;
    x->encode_breakout = cpi->encode_breakout;
  }
}

284
static void duplicate_mode_info_in_sb(VP9_COMMON * const cm,
285 286 287 288 289 290 291 292 293 294
                                     MACROBLOCKD *const xd,
                                     int mi_row,
                                     int mi_col,
                                     BLOCK_SIZE bsize) {
  const int block_width = num_8x8_blocks_wide_lookup[bsize];
  const int block_height = num_8x8_blocks_high_lookup[bsize];
  int i, j;
  for (j = 0; j < block_height; ++j)
    for (i = 0; i < block_width; ++i) {
      if (mi_row + j < cm->mi_rows && mi_col + i < cm->mi_cols)
295
        xd->mi[j * xd->mi_stride + i] = xd->mi[0];
296 297 298 299 300 301 302 303 304 305
    }
}

static void set_block_size(VP9_COMP * const cpi,
                           const TileInfo *const tile,
                           int mi_row, int mi_col,
                           BLOCK_SIZE bsize) {
  if (cpi->common.mi_cols > mi_col && cpi->common.mi_rows > mi_row) {
    MACROBLOCKD *const xd = &cpi->mb.e_mbd;
    set_modeinfo_offsets(&cpi->common, xd, mi_row, mi_col);
306
    xd->mi[0]->mbmi.sb_type = bsize;
307
    duplicate_mode_info_in_sb(&cpi->common, xd, mi_row, mi_col, bsize);
308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502
  }
}

typedef struct {
  int64_t sum_square_error;
  int64_t sum_error;
  int count;
  int variance;
} var;

typedef struct {
  var none;
  var horz[2];
  var vert[2];
} partition_variance;

typedef struct {
  partition_variance part_variances;
  var split[4];
} v8x8;

typedef struct {
  partition_variance part_variances;
  v8x8 split[4];
} v16x16;

typedef struct {
  partition_variance part_variances;
  v16x16 split[4];
} v32x32;

typedef struct {
  partition_variance part_variances;
  v32x32 split[4];
} v64x64;

typedef struct {
  partition_variance *part_variances;
  var *split[4];
} variance_node;

typedef enum {
  V16X16,
  V32X32,
  V64X64,
} TREE_LEVEL;

static void tree_to_node(void *data, BLOCK_SIZE bsize, variance_node *node) {
  int i;
  switch (bsize) {
    case BLOCK_64X64: {
      v64x64 *vt = (v64x64 *) data;
      node->part_variances = &vt->part_variances;
      for (i = 0; i < 4; i++)
        node->split[i] = &vt->split[i].part_variances.none;
      break;
    }
    case BLOCK_32X32: {
      v32x32 *vt = (v32x32 *) data;
      node->part_variances = &vt->part_variances;
      for (i = 0; i < 4; i++)
        node->split[i] = &vt->split[i].part_variances.none;
      break;
    }
    case BLOCK_16X16: {
      v16x16 *vt = (v16x16 *) data;
      node->part_variances = &vt->part_variances;
      for (i = 0; i < 4; i++)
        node->split[i] = &vt->split[i].part_variances.none;
      break;
    }
    case BLOCK_8X8: {
      v8x8 *vt = (v8x8 *) data;
      node->part_variances = &vt->part_variances;
      for (i = 0; i < 4; i++)
        node->split[i] = &vt->split[i];
      break;
    }
    default: {
      assert(0);
    }
  }
}

// Set variance values given sum square error, sum error, count.
static void fill_variance(int64_t s2, int64_t s, int c, var *v) {
  v->sum_square_error = s2;
  v->sum_error = s;
  v->count = c;
  if (c > 0)
    v->variance = (int)(256 *
                        (v->sum_square_error - v->sum_error * v->sum_error /
                         v->count) / v->count);
  else
    v->variance = 0;
}

void sum_2_variances(const var *a, const var *b, var *r) {
  fill_variance(a->sum_square_error + b->sum_square_error,
                a->sum_error + b->sum_error, a->count + b->count, r);
}

static void fill_variance_tree(void *data, BLOCK_SIZE bsize) {
  variance_node node;
  tree_to_node(data, bsize, &node);
  sum_2_variances(node.split[0], node.split[1], &node.part_variances->horz[0]);
  sum_2_variances(node.split[2], node.split[3], &node.part_variances->horz[1]);
  sum_2_variances(node.split[0], node.split[2], &node.part_variances->vert[0]);
  sum_2_variances(node.split[1], node.split[3], &node.part_variances->vert[1]);
  sum_2_variances(&node.part_variances->vert[0], &node.part_variances->vert[1],
                  &node.part_variances->none);
}

static int set_vt_partitioning(VP9_COMP *cpi,
                               void *data,
                               const TileInfo *const tile,
                               BLOCK_SIZE bsize,
                               int mi_row,
                               int mi_col,
                               int mi_size) {
  VP9_COMMON * const cm = &cpi->common;
  variance_node vt;
  const int block_width = num_8x8_blocks_wide_lookup[bsize];
  const int block_height = num_8x8_blocks_high_lookup[bsize];
  // TODO(debargha): Choose this more intelligently.
  const int64_t threshold_multiplier = 25;
  int64_t threshold = threshold_multiplier * cpi->common.base_qindex;
  assert(block_height == block_width);

  tree_to_node(data, bsize, &vt);

  // Split none is available only if we have more than half a block size
  // in width and height inside the visible image.
  if (mi_col + block_width / 2 < cm->mi_cols &&
      mi_row + block_height / 2 < cm->mi_rows &&
      vt.part_variances->none.variance < threshold) {
    set_block_size(cpi, tile, mi_row, mi_col, bsize);
    return 1;
  }

  // Vertical split is available on all but the bottom border.
  if (mi_row + block_height / 2 < cm->mi_rows &&
      vt.part_variances->vert[0].variance < threshold &&
      vt.part_variances->vert[1].variance < threshold) {
    BLOCK_SIZE subsize = get_subsize(bsize, PARTITION_VERT);
    set_block_size(cpi, tile, mi_row, mi_col, subsize);
    set_block_size(cpi, tile, mi_row, mi_col + block_width / 2, subsize);
    return 1;
  }

  // Horizontal split is available on all but the right border.
  if (mi_col + block_width / 2 < cm->mi_cols &&
      vt.part_variances->horz[0].variance < threshold &&
      vt.part_variances->horz[1].variance < threshold) {
    BLOCK_SIZE subsize = get_subsize(bsize, PARTITION_HORZ);
    set_block_size(cpi, tile, mi_row, mi_col, subsize);
    set_block_size(cpi, tile, mi_row + block_height / 2, mi_col, subsize);
    return 1;
  }
  return 0;
}

// TODO(debargha): Fix this function and make it work as expected.
static void choose_partitioning(VP9_COMP *cpi,
                                const TileInfo *const tile,
                                int mi_row, int mi_col) {
  VP9_COMMON * const cm = &cpi->common;
  MACROBLOCK *x = &cpi->mb;
  MACROBLOCKD *xd = &cpi->mb.e_mbd;

  int i, j, k;
  v64x64 vt;
  uint8_t *s;
  const uint8_t *d;
  int sp;
  int dp;
  int pixels_wide = 64, pixels_high = 64;
  int_mv nearest_mv, near_mv;
  const YV12_BUFFER_CONFIG *yv12 = get_ref_frame_buffer(cpi, LAST_FRAME);
  const struct scale_factors *const sf = &cm->frame_refs[LAST_FRAME - 1].sf;

  vp9_zero(vt);
  set_offsets(cpi, tile, mi_row, mi_col, BLOCK_64X64);

  if (xd->mb_to_right_edge < 0)
    pixels_wide += (xd->mb_to_right_edge >> 3);
  if (xd->mb_to_bottom_edge < 0)
    pixels_high += (xd->mb_to_bottom_edge >> 3);

  s = x->plane[0].src.buf;
  sp = x->plane[0].src.stride;

  if (cm->frame_type != KEY_FRAME) {
    vp9_setup_pre_planes(xd, 0, yv12, mi_row, mi_col, sf);

503 504
    xd->mi[0]->mbmi.ref_frame[0] = LAST_FRAME;
    xd->mi[0]->mbmi.sb_type = BLOCK_64X64;
505
    vp9_find_best_ref_mvs(xd, cm->allow_high_precision_mv,
506
                          xd->mi[0]->mbmi.ref_mvs[LAST_FRAME],
507 508
                          &nearest_mv, &near_mv);

509
    xd->mi[0]->mbmi.mv[0] = nearest_mv;
510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601
    vp9_build_inter_predictors_sby(xd, mi_row, mi_col, BLOCK_64X64);

    d = xd->plane[0].dst.buf;
    dp = xd->plane[0].dst.stride;
  } else {
    d = VP9_VAR_OFFS;
    dp = 0;
  }

  // Fill in the entire tree of 8x8 variances for splits.
  for (i = 0; i < 4; i++) {
    const int x32_idx = ((i & 1) << 5);
    const int y32_idx = ((i >> 1) << 5);
    for (j = 0; j < 4; j++) {
      const int x16_idx = x32_idx + ((j & 1) << 4);
      const int y16_idx = y32_idx + ((j >> 1) << 4);
      v16x16 *vst = &vt.split[i].split[j];
      for (k = 0; k < 4; k++) {
        int x_idx = x16_idx + ((k & 1) << 3);
        int y_idx = y16_idx + ((k >> 1) << 3);
        unsigned int sse = 0;
        int sum = 0;
        if (x_idx < pixels_wide && y_idx < pixels_high)
          vp9_get_sse_sum_8x8(s + y_idx * sp + x_idx, sp,
                              d + y_idx * dp + x_idx, dp, &sse, &sum);
        fill_variance(sse, sum, 64, &vst->split[k].part_variances.none);
      }
    }
  }
  // Fill the rest of the variance tree by summing split partition values.
  for (i = 0; i < 4; i++) {
    for (j = 0; j < 4; j++) {
      fill_variance_tree(&vt.split[i].split[j], BLOCK_16X16);
    }
    fill_variance_tree(&vt.split[i], BLOCK_32X32);
  }
  fill_variance_tree(&vt, BLOCK_64X64);

  // Now go through the entire structure,  splitting every block size until
  // we get to one that's got a variance lower than our threshold,  or we
  // hit 8x8.
  if (!set_vt_partitioning(cpi, &vt, tile, BLOCK_64X64,
                           mi_row, mi_col, 8)) {
    for (i = 0; i < 4; ++i) {
      const int x32_idx = ((i & 1) << 2);
      const int y32_idx = ((i >> 1) << 2);
      if (!set_vt_partitioning(cpi, &vt.split[i], tile, BLOCK_32X32,
                               (mi_row + y32_idx), (mi_col + x32_idx), 4)) {
        for (j = 0; j < 4; ++j) {
          const int x16_idx = ((j & 1) << 1);
          const int y16_idx = ((j >> 1) << 1);
          // NOTE: This is a temporary hack to disable 8x8 partitions,
          // since it works really bad - possibly due to a bug
#define DISABLE_8X8_VAR_BASED_PARTITION
#ifdef DISABLE_8X8_VAR_BASED_PARTITION
          if (mi_row + y32_idx + y16_idx + 1 < cm->mi_rows &&
              mi_row + x32_idx + x16_idx + 1 < cm->mi_cols) {
            set_block_size(cpi, tile,
                           (mi_row + y32_idx + y16_idx),
                           (mi_col + x32_idx + x16_idx),
                           BLOCK_16X16);
          } else {
            for (k = 0; k < 4; ++k) {
              const int x8_idx = (k & 1);
              const int y8_idx = (k >> 1);
              set_block_size(cpi, tile,
                             (mi_row + y32_idx + y16_idx + y8_idx),
                             (mi_col + x32_idx + x16_idx + x8_idx),
                             BLOCK_8X8);
            }
          }
#else
          if (!set_vt_partitioning(cpi, &vt.split[i].split[j], tile,
                                   BLOCK_16X16,
                                   (mi_row + y32_idx + y16_idx),
                                   (mi_col + x32_idx + x16_idx), 2)) {
            for (k = 0; k < 4; ++k) {
              const int x8_idx = (k & 1);
              const int y8_idx = (k >> 1);
              set_block_size(cpi, tile,
                             (mi_row + y32_idx + y16_idx + y8_idx),
                             (mi_col + x32_idx + x16_idx + x8_idx),
                             BLOCK_8X8);
            }
          }
#endif
        }
      }
    }
  }
}

602
// Original activity measure from Tim T's code.
603
static unsigned int tt_activity_measure(MACROBLOCK *x) {
John Koleszar's avatar
John Koleszar committed
604
  unsigned int sse;
605 606 607 608 609 610 611 612 613
  // TODO: This could also be done over smaller areas (8x8), but that would
  // require extensive changes elsewhere, as lambda is assumed to be fixed
  // over an entire MB in most of the code.
  // Another option is to compute four 8x8 variances, and pick a single
  // lambda using a non-linear combination (e.g., the smallest, or second
  // smallest, etc.).
  const unsigned int act = vp9_variance16x16(x->plane[0].src.buf,
                                             x->plane[0].src.stride,
                                             VP9_VAR_OFFS, 0, &sse) << 4;
614
  // If the region is flat, lower the activity some more.
615
  return act < (8 << 12) ? MIN(act, 5 << 12) : act;
616 617
}

618
// Stub for alternative experimental activity measures.
619 620
static unsigned int alt_activity_measure(MACROBLOCK *x, int use_dc_pred) {
  return vp9_encode_intra(x, use_dc_pred);
621 622 623 624
}

// Measure the activity of the current macroblock
// What we measure here is TBD so abstracted to this function
625
#define ALT_ACT_MEASURE 1
626
static unsigned int mb_activity_measure(MACROBLOCK *x, int mb_row, int mb_col) {
John Koleszar's avatar
John Koleszar committed
627
  unsigned int mb_activity;
628

John Koleszar's avatar
John Koleszar committed
629
  if (ALT_ACT_MEASURE) {
630
    const int use_dc_pred = (mb_col || mb_row) && (!mb_col || !mb_row);
631

John Koleszar's avatar
John Koleszar committed
632
    // Or use and alternative.
633
    mb_activity = alt_activity_measure(x, use_dc_pred);
John Koleszar's avatar
John Koleszar committed
634 635
  } else {
    // Original activity measure from Tim T's code.
636
    mb_activity = tt_activity_measure(x);
John Koleszar's avatar
John Koleszar committed
637
  }
638

639
  return MAX(mb_activity, ACTIVITY_AVG_MIN);
640 641 642
}

// Calculate an "average" mb activity value for the frame
643
#define ACT_MEDIAN 0
644
static void calc_av_activity(VP9_COMP *cpi, int64_t activity_sum) {
645
#if ACT_MEDIAN
John Koleszar's avatar
John Koleszar committed
646 647 648 649 650 651 652 653
  // Find median: Simple n^2 algorithm for experimentation
  {
    unsigned int median;
    unsigned int i, j;
    unsigned int *sortlist;
    unsigned int tmp;

    // Create a list to sort to
654 655
    CHECK_MEM_ERROR(&cpi->common, sortlist, vpx_calloc(sizeof(unsigned int),
                    cpi->common.MBs));
John Koleszar's avatar
John Koleszar committed
656 657 658

    // Copy map to sort list
    vpx_memcpy(sortlist, cpi->mb_activity_map,
Jim Bankoski's avatar
Jim Bankoski committed
659
        sizeof(unsigned int) * cpi->common.MBs);
John Koleszar's avatar
John Koleszar committed
660 661 662 663 664 665 666 667 668

    // Ripple each value down to its correct position
    for (i = 1; i < cpi->common.MBs; i ++) {
      for (j = i; j > 0; j --) {
        if (sortlist[j] < sortlist[j - 1]) {
          // Swap values
          tmp = sortlist[j - 1];
          sortlist[j - 1] = sortlist[j];
          sortlist[j] = tmp;
669 670 671
        } else {
          break;
        }
John Koleszar's avatar
John Koleszar committed
672 673
      }
    }
674

John Koleszar's avatar
John Koleszar committed
675 676
    // Even number MBs so estimate median as mean of two either side.
    median = (1 + sortlist[cpi->common.MBs >> 1] +
Jim Bankoski's avatar
Jim Bankoski committed
677
        sortlist[(cpi->common.MBs >> 1) + 1]) >> 1;
678

John Koleszar's avatar
John Koleszar committed
679
    cpi->activity_avg = median;
680

John Koleszar's avatar
John Koleszar committed
681 682
    vpx_free(sortlist);
  }
683
#else
John Koleszar's avatar
John Koleszar committed
684
  // Simple mean for now
Jim Bankoski's avatar
Jim Bankoski committed
685
  cpi->activity_avg = (unsigned int) (activity_sum / cpi->common.MBs);
Dmitry Kovalev's avatar
Dmitry Kovalev committed
686
#endif  // ACT_MEDIAN
687

688 689
  if (cpi->activity_avg < ACTIVITY_AVG_MIN)
    cpi->activity_avg = ACTIVITY_AVG_MIN;
690

John Koleszar's avatar
John Koleszar committed
691 692 693
  // Experimental code: return fixed value normalized for several clips
  if (ALT_ACT_MEASURE)
    cpi->activity_avg = 100000;
694 695
}

696
#define USE_ACT_INDEX   0
697
#define OUTPUT_NORM_ACT_STATS   0
698 699

#if USE_ACT_INDEX
700
// Calculate an activity index for each mb
701 702
static void calc_activity_index(VP9_COMP *cpi, MACROBLOCK *x) {
  VP9_COMMON *const cm = &cpi->common;
John Koleszar's avatar
John Koleszar committed
703
  int mb_row, mb_col;
704

John Koleszar's avatar
John Koleszar committed
705 706 707
  int64_t act;
  int64_t a;
  int64_t b;
708 709

#if OUTPUT_NORM_ACT_STATS
John Koleszar's avatar
John Koleszar committed
710 711
  FILE *f = fopen("norm_act.stt", "a");
  fprintf(f, "\n%12d\n", cpi->activity_avg);
712 713
#endif

John Koleszar's avatar
John Koleszar committed
714 715
  // Reset pointers to start of activity map
  x->mb_activity_ptr = cpi->mb_activity_map;
716

John Koleszar's avatar
John Koleszar committed
717 718 719 720 721 722
  // Calculate normalized mb activity number.
  for (mb_row = 0; mb_row < cm->mb_rows; mb_row++) {
    // for each macroblock col in image
    for (mb_col = 0; mb_col < cm->mb_cols; mb_col++) {
      // Read activity from the map
      act = *(x->mb_activity_ptr);
723

John Koleszar's avatar
John Koleszar committed
724 725 726
      // Calculate a normalized activity number
      a = act + 4 * cpi->activity_avg;
      b = 4 * act + cpi->activity_avg;
727

John Koleszar's avatar
John Koleszar committed
728
      if (b >= a)
Jim Bankoski's avatar
Jim Bankoski committed
729
      *(x->activity_ptr) = (int)((b + (a >> 1)) / a) - 1;
John Koleszar's avatar
John Koleszar committed
730
      else
Jim Bankoski's avatar
Jim Bankoski committed
731
      *(x->activity_ptr) = 1 - (int)((a + (b >> 1)) / b);
732 733

#if OUTPUT_NORM_ACT_STATS
John Koleszar's avatar
John Koleszar committed
734
      fprintf(f, " %6d", *(x->mb_activity_ptr));
735
#endif
John Koleszar's avatar
John Koleszar committed
736 737 738
      // Increment activity map pointers
      x->mb_activity_ptr++;
    }
739 740

#if OUTPUT_NORM_ACT_STATS
John Koleszar's avatar
John Koleszar committed
741
    fprintf(f, "\n");
742
#endif
John Koleszar's avatar
John Koleszar committed
743
  }
744 745

#if OUTPUT_NORM_ACT_STATS
John Koleszar's avatar
John Koleszar committed
746
  fclose(f);
747 748
#endif
}
Dmitry Kovalev's avatar
Dmitry Kovalev committed
749
#endif  // USE_ACT_INDEX
750 751 752

// Loop through all MBs. Note activity of each, average activity and
// calculate a normalized activity for each
753
static void build_activity_map(VP9_COMP *cpi) {
Dmitry Kovalev's avatar
Dmitry Kovalev committed
754
  MACROBLOCK *const x = &cpi->mb;
John Koleszar's avatar
John Koleszar committed
755
  MACROBLOCKD *xd = &x->e_mbd;
Dmitry Kovalev's avatar
Dmitry Kovalev committed
756
  VP9_COMMON *const cm = &cpi->common;
757

758
#if ALT_ACT_MEASURE
759
  YV12_BUFFER_CONFIG *new_yv12 = get_frame_new_buffer(cm);
John Koleszar's avatar
John Koleszar committed
760 761
  int recon_yoffset;
  int recon_y_stride = new_yv12->y_stride;
762 763
#endif

John Koleszar's avatar
John Koleszar committed
764 765 766
  int mb_row, mb_col;
  unsigned int mb_activity;
  int64_t activity_sum = 0;
767

768 769
  x->mb_activity_ptr = cpi->mb_activity_map;

John Koleszar's avatar
John Koleszar committed
770 771
  // for each macroblock row in image
  for (mb_row = 0; mb_row < cm->mb_rows; mb_row++) {
772
#if ALT_ACT_MEASURE
John Koleszar's avatar
John Koleszar committed
773 774 775
    // reset above block coeffs
    xd->up_available = (mb_row != 0);
    recon_yoffset = (mb_row * recon_y_stride * 16);
776
#endif
John Koleszar's avatar
John Koleszar committed
777 778
    // for each macroblock col in image
    for (mb_col = 0; mb_col < cm->mb_cols; mb_col++) {
779
#if ALT_ACT_MEASURE
780
      xd->plane[0].dst.buf = new_yv12->y_buffer + recon_yoffset;
John Koleszar's avatar
John Koleszar committed
781 782
      xd->left_available = (mb_col != 0);
      recon_yoffset += 16;
783
#endif
784

John Koleszar's avatar
John Koleszar committed
785
      // measure activity
786
      mb_activity = mb_activity_measure(x, mb_row, mb_col);
787

John Koleszar's avatar
John Koleszar committed
788 789
      // Keep frame sum
      activity_sum += mb_activity;
790

John Koleszar's avatar
John Koleszar committed
791 792
      // Store MB level activity details.
      *x->mb_activity_ptr = mb_activity;
793

John Koleszar's avatar
John Koleszar committed
794 795
      // Increment activity map pointer
      x->mb_activity_ptr++;
796

John Koleszar's avatar
John Koleszar committed
797
      // adjust to the next column of source macroblocks
John Koleszar's avatar
John Koleszar committed
798
      x->plane[0].src.buf += 16;
John Koleszar's avatar
John Koleszar committed
799
    }
800

John Koleszar's avatar
John Koleszar committed
801
    // adjust to the next row of mbs
John Koleszar's avatar
John Koleszar committed
802
    x->plane[0].src.buf += 16 * x->plane[0].src.stride - 16 * cm->mb_cols;
John Koleszar's avatar
John Koleszar committed
803
  }
804

John Koleszar's avatar
John Koleszar committed
805 806
  // Calculate an "average" MB activity
  calc_av_activity(cpi, activity_sum);
807

808
#if USE_ACT_INDEX
John Koleszar's avatar
John Koleszar committed
809 810
  // Calculate an activity index number of each mb
  calc_activity_index(cpi, x);
811
#endif
812 813
}

814
// Macroblock activity masking
815
static void activity_masking(VP9_COMP *cpi, MACROBLOCK *x) {
816
#if USE_ACT_INDEX
John Koleszar's avatar
John Koleszar committed
817 818 819
  x->rdmult += *(x->mb_activity_ptr) * (x->rdmult >> 2);
  x->errorperbit = x->rdmult * 100 / (110 * x->rddiv);
  x->errorperbit += (x->errorperbit == 0);
820
#else
821
  const int64_t act = *(x->mb_activity_ptr);
822

John Koleszar's avatar
John Koleszar committed
823
  // Apply the masking to the RD multiplier.
824 825
  const int64_t a = act + (2 * cpi->activity_avg);
  const int64_t b = (2 * act) + cpi->activity_avg;
826

Jim Bankoski's avatar
Jim Bankoski committed
827
  x->rdmult = (unsigned int) (((int64_t) x->rdmult * b + (a >> 1)) / a);
John Koleszar's avatar
John Koleszar committed
828 829
  x->errorperbit = x->rdmult * 100 / (110 * x->rddiv);
  x->errorperbit += (x->errorperbit == 0);
830
#endif
831

John Koleszar's avatar
John Koleszar committed
832 833
  // Activity based Zbin adjustment
  adjust_act_zbin(cpi, x);
834
}
John Koleszar's avatar
John Koleszar committed
835

Jim Bankoski's avatar
Jim Bankoski committed
836
static void update_state(VP9_COMP *cpi, PICK_MODE_CONTEXT *ctx,
837 838
                         int mi_row, int mi_col, BLOCK_SIZE bsize,
                         int output_enabled) {
Ronald S. Bultje's avatar
Ronald S. Bultje committed
839
  int i, x_idx, y;
840 841 842
  VP9_COMMON *const cm = &cpi->common;
  MACROBLOCK *const x = &cpi->mb;
  MACROBLOCKD *const xd = &x->e_mbd;
843 844
  struct macroblock_plane *const p = x->plane;
  struct macroblockd_plane *const pd = xd->plane;
John Koleszar's avatar
John Koleszar committed
845
  MODE_INFO *mi = &ctx->mic;
846 847
  MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
  MODE_INFO *mi_addr = xd->mi[0];
848
  const struct segmentation *const seg = &cm->seg;
849

850
  const int mis = cm->mi_stride;
Jim Bankoski's avatar
Jim Bankoski committed
851
  const int mi_width = num_8x8_blocks_wide_lookup[bsize];
Jim Bankoski's avatar
Jim Bankoski committed
852
  const int mi_height = num_8x8_blocks_high_lookup[bsize];
853
  int max_plane;
Adrian Grange's avatar
Adrian Grange committed
854

855
  assert(mi->mbmi.sb_type == bsize);
856

857 858 859 860 861
  *mi_addr = *mi;

  // For in frame adaptive Q, check for reseting the segment_id and updating
  // the cyclic refresh map.
  if ((cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ) && seg->enabled &&
862
      output_enabled) {
863
    vp9_cyclic_refresh_update_segment(cpi, &xd->mi[0]->mbmi,
864
                                      mi_row, mi_col, bsize, 1);
865
    vp9_init_plane_quantizers(cpi, x);
866
  }
867

868 869
  max_plane = is_inter_block(mbmi) ? MAX_MB_PLANE : 1;
  for (i = 0; i < max_plane; ++i) {
870
    p[i].coeff = ctx->coeff_pbuf[i][1];
871
    p[i].qcoeff = ctx->qcoeff_pbuf[i][1];
872
    pd[i].dqcoeff = ctx->dqcoeff_pbuf[i][1];
873
    p[i].eobs = ctx->eobs_pbuf[i][1];
874 875
  }

876 877
  for (i = max_plane; i < MAX_MB_PLANE; ++i) {
    p[i].coeff = ctx->coeff_pbuf[i][2];
878
    p[i].qcoeff = ctx->qcoeff_pbuf[i][2];
879
    pd[i].dqcoeff = ctx->dqcoeff_pbuf[i][2];
880
    p[i].eobs = ctx->eobs_pbuf[i][2];
881 882
  }

John Koleszar's avatar
John Koleszar committed
883 884
  // Restore the coding context of the MB to that that was in place
  // when the mode was picked for it
885 886
  for (y = 0; y < mi_height; y++)
    for (x_idx = 0; x_idx < mi_width; x_idx++)
James Zern's avatar
James Zern committed
887
      if ((xd->mb_to_right_edge >> (3 + MI_SIZE_LOG2)) + mi_width > x_idx
888
        && (xd->mb_to_bottom_edge >> (3 + MI_SIZE_LOG2)) + mi_height > y) {
889
        xd->mi[x_idx + y * mis] = mi_addr;
890
      }
891

892
  if (cpi->oxcf.aq_mode)
Dmitry Kovalev's avatar
Dmitry Kovalev committed
893
    vp9_init_plane_quantizers(cpi, x);
894

895 896
  // FIXME(rbultje) I'm pretty sure this should go to the end of this block
  // (i.e. after the output_enabled)
897 898
  if (bsize < BLOCK_32X32) {
    if (bsize < BLOCK_16X16)
899 900
      ctx->tx_rd_diff[ALLOW_16X16] = ctx->tx_rd_diff[ALLOW_8X8];
    ctx->tx_rd_diff[ALLOW_32X32] = ctx->tx_rd_diff[ALLOW_16X16];
Ronald S. Bultje's avatar
Ronald S. Bultje committed
901
  }
Adrian Grange's avatar
Adrian Grange committed
902

903
  if (is_inter_block(mbmi) && mbmi->sb_type < BLOCK_8X8) {
904 905
    mbmi->mv[0].as_int = mi->bmi[3].as_mv[0].as_int;
    mbmi->mv[1].as_int = mi->bmi[3].as_mv[1].as_int;
John Koleszar's avatar
John Koleszar committed
906 907
  }

908
  x->skip = ctx->skip;
909
  vpx_memcpy(x->zcoeff_blk[mbmi->tx_size], ctx->zcoeff_blk,
910
             sizeof(uint8_t) * ctx->num_4x4_blk);
911

Ronald S. Bultje's avatar
Ronald S. Bultje committed
912 913 914
  if (!output_enabled)
    return;

915
  if (!vp9_segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_SKIP)) {
916
    for (i = 0; i < TX_MODES; i++)
917
      cpi->rd_tx_select_diff[i] += ctx->tx_rd_diff[i];
918 919
  }

920
#if CONFIG_INTERNAL_STATS
921
  if (frame_is_intra_only(cm)) {
John Koleszar's avatar
John Koleszar committed
922
    static const int kf_mode_index[] = {
923 924 925 926
      THR_DC        /*DC_PRED*/,
      THR_V_PRED    /*V_PRED*/,
      THR_H_PRED    /*H_PRED*/,
      THR_D45_PRED  /*D45_PRED*/,
John Koleszar's avatar
John Koleszar committed
927 928 929
      THR_D135_PRED /*D135_PRED*/,
      THR_D117_PRED /*D117_PRED*/,
      THR_D153_PRED /*D153_PRED*/,
Dmitry Kovalev's avatar
Dmitry Kovalev committed
930
      THR_D207_PRED /*D207_PRED*/,
931 932
      THR_D63_PRED  /*D63_PRED*/,
      THR_TM        /*TM_PRED*/,
John Koleszar's avatar
John Koleszar committed
933
    };
934
    ++cpi->mode_chosen_counts[kf_mode_index[mbmi->mode]];
John Koleszar's avatar
John Koleszar committed
935 936
  } else {
    // Note how often each mode chosen as best
937 938 939 940
    ++cpi->mode_chosen_counts[ctx->best_mode_index];
  }
#endif
  if (!frame_is_intra_only(cm)) {
941
    if (is_inter_block(mbmi)) {
942
      vp9_update_mv_count(cm, xd);
943 944 945 946 947

      if (cm->interp_filter == SWITCHABLE) {
        const int ctx = vp9_get_pred_context_switchable_interp(xd);
        ++cm->counts.switchable_interp[ctx][mbmi->interp_filter];
      }
948
    }
Adrian Grange's avatar
Adrian Grange committed
949

950 951 952
    cpi->rd_comp_pred_diff[SINGLE_REFERENCE] += ctx->single_pred_diff;
    cpi->rd_comp_pred_diff[COMPOUND_REFERENCE] += ctx->comp_pred_diff;
    cpi->rd_comp_pred_diff[REFERENCE_MODE_SELECT] += ctx->hybrid_pred_diff;
953

954
    for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i)
955
      cpi->rd_filter_diff[i] += ctx->best_filter_diff[i];
John Koleszar's avatar
John Koleszar committed
956
  }
Adrian Grange's avatar
Adrian Grange committed
957 958
}

Jim Bankoski's avatar
Jim Bankoski committed
959
void vp9_setup_src_planes(MACROBLOCK *x, const YV12_BUFFER_CONFIG *src,
960 961 962 963 964
                          int mi_row, int mi_col) {
  uint8_t *const buffers[4] = {src->y_buffer, src->u_buffer, src->v_buffer,
                               src->alpha_buffer};
  const int strides[4] = {src->y_stride, src->uv_stride, src->uv_stride,
                          src->alpha_stride};
John Koleszar's avatar