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 <limits.h>
12#include <math.h>
13#include <stdio.h>
14
15#include "./vp9_rtcd.h"
16#include "./vpx_dsp_rtcd.h"
17#include "./vpx_config.h"
18
19#include "vpx_dsp/vpx_dsp_common.h"
20#include "vpx_ports/mem.h"
21#include "vpx_ports/vpx_timer.h"
22#include "vpx_ports/system_state.h"
23
24#include "vp9/common/vp9_common.h"
25#include "vp9/common/vp9_entropy.h"
26#include "vp9/common/vp9_entropymode.h"
27#include "vp9/common/vp9_idct.h"
28#include "vp9/common/vp9_mvref_common.h"
29#include "vp9/common/vp9_pred_common.h"
30#include "vp9/common/vp9_quant_common.h"
31#include "vp9/common/vp9_reconintra.h"
32#include "vp9/common/vp9_reconinter.h"
33#include "vp9/common/vp9_seg_common.h"
34#include "vp9/common/vp9_tile_common.h"
35
36#include "vp9/encoder/vp9_aq_360.h"
37#include "vp9/encoder/vp9_aq_complexity.h"
38#include "vp9/encoder/vp9_aq_cyclicrefresh.h"
39#include "vp9/encoder/vp9_aq_variance.h"
40#include "vp9/encoder/vp9_encodeframe.h"
41#include "vp9/encoder/vp9_encodemb.h"
42#include "vp9/encoder/vp9_encodemv.h"
43#include "vp9/encoder/vp9_ethread.h"
44#include "vp9/encoder/vp9_extend.h"
45#include "vp9/encoder/vp9_pickmode.h"
46#include "vp9/encoder/vp9_rd.h"
47#include "vp9/encoder/vp9_rdopt.h"
48#include "vp9/encoder/vp9_segmentation.h"
49#include "vp9/encoder/vp9_tokenize.h"
50
51static void encode_superblock(VP9_COMP *cpi, ThreadData *td, TOKENEXTRA **t,
52                              int output_enabled, int mi_row, int mi_col,
53                              BLOCK_SIZE bsize, PICK_MODE_CONTEXT *ctx);
54
55// Machine learning-based early termination parameters.
56static const double train_mean[24] = {
57  303501.697372, 3042630.372158, 24.694696, 1.392182,
58  689.413511,    162.027012,     1.478213,  0.0,
59  135382.260230, 912738.513263,  28.845217, 1.515230,
60  544.158492,    131.807995,     1.436863,  0.0,
61  43682.377587,  208131.711766,  28.084737, 1.356677,
62  138.254122,    119.522553,     1.252322,  0.0
63};
64
65static const double train_stdm[24] = {
66  673689.212982, 5996652.516628, 0.024449, 1.989792,
67  985.880847,    0.014638,       2.001898, 0.0,
68  208798.775332, 1812548.443284, 0.018693, 1.838009,
69  396.986910,    0.015657,       1.332541, 0.0,
70  55888.847031,  448587.962714,  0.017900, 1.904776,
71  98.652832,     0.016598,       1.320992, 0.0
72};
73
74// Error tolerance: 0.01%-0.0.05%-0.1%
75static const double classifiers[24] = {
76  0.111736, 0.289977, 0.042219, 0.204765, 0.120410, -0.143863,
77  0.282376, 0.847811, 0.637161, 0.131570, 0.018636, 0.202134,
78  0.112797, 0.028162, 0.182450, 1.124367, 0.386133, 0.083700,
79  0.050028, 0.150873, 0.061119, 0.109318, 0.127255, 0.625211
80};
81
82// This is used as a reference when computing the source variance for the
83//  purpose of activity masking.
84// Eventually this should be replaced by custom no-reference routines,
85//  which will be faster.
86static const uint8_t VP9_VAR_OFFS[64] = {
87  128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128,
88  128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128,
89  128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128,
90  128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128,
91  128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128
92};
93
94#if CONFIG_VP9_HIGHBITDEPTH
95static const uint16_t VP9_HIGH_VAR_OFFS_8[64] = {
96  128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128,
97  128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128,
98  128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128,
99  128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128,
100  128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128
101};
102
103static const uint16_t VP9_HIGH_VAR_OFFS_10[64] = {
104  128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4,
105  128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4,
106  128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4,
107  128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4,
108  128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4,
109  128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4,
110  128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4,
111  128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4
112};
113
114static const uint16_t VP9_HIGH_VAR_OFFS_12[64] = {
115  128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16,
116  128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16,
117  128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16,
118  128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16,
119  128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16,
120  128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16,
121  128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16,
122  128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16,
123  128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16,
124  128 * 16
125};
126#endif  // CONFIG_VP9_HIGHBITDEPTH
127
128unsigned int vp9_get_sby_variance(VP9_COMP *cpi, const struct buf_2d *ref,
129                                  BLOCK_SIZE bs) {
130  unsigned int sse;
131  const unsigned int var =
132      cpi->fn_ptr[bs].vf(ref->buf, ref->stride, VP9_VAR_OFFS, 0, &sse);
133  return var;
134}
135
136#if CONFIG_VP9_HIGHBITDEPTH
137unsigned int vp9_high_get_sby_variance(VP9_COMP *cpi, const struct buf_2d *ref,
138                                       BLOCK_SIZE bs, int bd) {
139  unsigned int var, sse;
140  switch (bd) {
141    case 10:
142      var =
143          cpi->fn_ptr[bs].vf(ref->buf, ref->stride,
144                             CONVERT_TO_BYTEPTR(VP9_HIGH_VAR_OFFS_10), 0, &sse);
145      break;
146    case 12:
147      var =
148          cpi->fn_ptr[bs].vf(ref->buf, ref->stride,
149                             CONVERT_TO_BYTEPTR(VP9_HIGH_VAR_OFFS_12), 0, &sse);
150      break;
151    case 8:
152    default:
153      var =
154          cpi->fn_ptr[bs].vf(ref->buf, ref->stride,
155                             CONVERT_TO_BYTEPTR(VP9_HIGH_VAR_OFFS_8), 0, &sse);
156      break;
157  }
158  return var;
159}
160#endif  // CONFIG_VP9_HIGHBITDEPTH
161
162unsigned int vp9_get_sby_perpixel_variance(VP9_COMP *cpi,
163                                           const struct buf_2d *ref,
164                                           BLOCK_SIZE bs) {
165  return ROUND_POWER_OF_TWO(vp9_get_sby_variance(cpi, ref, bs),
166                            num_pels_log2_lookup[bs]);
167}
168
169#if CONFIG_VP9_HIGHBITDEPTH
170unsigned int vp9_high_get_sby_perpixel_variance(VP9_COMP *cpi,
171                                                const struct buf_2d *ref,
172                                                BLOCK_SIZE bs, int bd) {
173  return (unsigned int)ROUND64_POWER_OF_TWO(
174      (int64_t)vp9_high_get_sby_variance(cpi, ref, bs, bd),
175      num_pels_log2_lookup[bs]);
176}
177#endif  // CONFIG_VP9_HIGHBITDEPTH
178
179static unsigned int get_sby_perpixel_diff_variance(VP9_COMP *cpi,
180                                                   const struct buf_2d *ref,
181                                                   int mi_row, int mi_col,
182                                                   BLOCK_SIZE bs) {
183  unsigned int sse, var;
184  uint8_t *last_y;
185  const YV12_BUFFER_CONFIG *last = get_ref_frame_buffer(cpi, LAST_FRAME);
186
187  assert(last != NULL);
188  last_y =
189      &last->y_buffer[mi_row * MI_SIZE * last->y_stride + mi_col * MI_SIZE];
190  var = cpi->fn_ptr[bs].vf(ref->buf, ref->stride, last_y, last->y_stride, &sse);
191  return ROUND_POWER_OF_TWO(var, num_pels_log2_lookup[bs]);
192}
193
194static BLOCK_SIZE get_rd_var_based_fixed_partition(VP9_COMP *cpi, MACROBLOCK *x,
195                                                   int mi_row, int mi_col) {
196  unsigned int var = get_sby_perpixel_diff_variance(
197      cpi, &x->plane[0].src, mi_row, mi_col, BLOCK_64X64);
198  if (var < 8)
199    return BLOCK_64X64;
200  else if (var < 128)
201    return BLOCK_32X32;
202  else if (var < 2048)
203    return BLOCK_16X16;
204  else
205    return BLOCK_8X8;
206}
207
208// Lighter version of set_offsets that only sets the mode info
209// pointers.
210static INLINE void set_mode_info_offsets(VP9_COMMON *const cm,
211                                         MACROBLOCK *const x,
212                                         MACROBLOCKD *const xd, int mi_row,
213                                         int mi_col) {
214  const int idx_str = xd->mi_stride * mi_row + mi_col;
215  xd->mi = cm->mi_grid_visible + idx_str;
216  xd->mi[0] = cm->mi + idx_str;
217  x->mbmi_ext = x->mbmi_ext_base + (mi_row * cm->mi_cols + mi_col);
218}
219
220static void set_offsets(VP9_COMP *cpi, const TileInfo *const tile,
221                        MACROBLOCK *const x, int mi_row, int mi_col,
222                        BLOCK_SIZE bsize) {
223  VP9_COMMON *const cm = &cpi->common;
224  MACROBLOCKD *const xd = &x->e_mbd;
225  MODE_INFO *mi;
226  const int mi_width = num_8x8_blocks_wide_lookup[bsize];
227  const int mi_height = num_8x8_blocks_high_lookup[bsize];
228  const struct segmentation *const seg = &cm->seg;
229  MvLimits *const mv_limits = &x->mv_limits;
230
231  set_skip_context(xd, mi_row, mi_col);
232
233  set_mode_info_offsets(cm, x, xd, mi_row, mi_col);
234
235  mi = xd->mi[0];
236
237  // Set up destination pointers.
238  vp9_setup_dst_planes(xd->plane, get_frame_new_buffer(cm), mi_row, mi_col);
239
240  // Set up limit values for MV components.
241  // Mv beyond the range do not produce new/different prediction block.
242  mv_limits->row_min = -(((mi_row + mi_height) * MI_SIZE) + VP9_INTERP_EXTEND);
243  mv_limits->col_min = -(((mi_col + mi_width) * MI_SIZE) + VP9_INTERP_EXTEND);
244  mv_limits->row_max = (cm->mi_rows - mi_row) * MI_SIZE + VP9_INTERP_EXTEND;
245  mv_limits->col_max = (cm->mi_cols - mi_col) * MI_SIZE + VP9_INTERP_EXTEND;
246
247  // Set up distance of MB to edge of frame in 1/8th pel units.
248  assert(!(mi_col & (mi_width - 1)) && !(mi_row & (mi_height - 1)));
249  set_mi_row_col(xd, tile, mi_row, mi_height, mi_col, mi_width, cm->mi_rows,
250                 cm->mi_cols);
251
252  // Set up source buffers.
253  vp9_setup_src_planes(x, cpi->Source, mi_row, mi_col);
254
255  // R/D setup.
256  x->rddiv = cpi->rd.RDDIV;
257  x->rdmult = cpi->rd.RDMULT;
258
259  // Setup segment ID.
260  if (seg->enabled) {
261    if (cpi->oxcf.aq_mode != VARIANCE_AQ && cpi->oxcf.aq_mode != LOOKAHEAD_AQ &&
262        cpi->oxcf.aq_mode != EQUATOR360_AQ) {
263      const uint8_t *const map =
264          seg->update_map ? cpi->segmentation_map : cm->last_frame_seg_map;
265      mi->segment_id = get_segment_id(cm, map, bsize, mi_row, mi_col);
266    }
267    vp9_init_plane_quantizers(cpi, x);
268
269    x->encode_breakout = cpi->segment_encode_breakout[mi->segment_id];
270  } else {
271    mi->segment_id = 0;
272    x->encode_breakout = cpi->encode_breakout;
273  }
274
275  // required by vp9_append_sub8x8_mvs_for_idx() and vp9_find_best_ref_mvs()
276  xd->tile = *tile;
277}
278
279static void duplicate_mode_info_in_sb(VP9_COMMON *cm, MACROBLOCKD *xd,
280                                      int mi_row, int mi_col,
281                                      BLOCK_SIZE bsize) {
282  const int block_width =
283      VPXMIN(num_8x8_blocks_wide_lookup[bsize], cm->mi_cols - mi_col);
284  const int block_height =
285      VPXMIN(num_8x8_blocks_high_lookup[bsize], cm->mi_rows - mi_row);
286  const int mi_stride = xd->mi_stride;
287  MODE_INFO *const src_mi = xd->mi[0];
288  int i, j;
289
290  for (j = 0; j < block_height; ++j)
291    for (i = 0; i < block_width; ++i) xd->mi[j * mi_stride + i] = src_mi;
292}
293
294static void set_block_size(VP9_COMP *const cpi, MACROBLOCK *const x,
295                           MACROBLOCKD *const xd, int mi_row, int mi_col,
296                           BLOCK_SIZE bsize) {
297  if (cpi->common.mi_cols > mi_col && cpi->common.mi_rows > mi_row) {
298    set_mode_info_offsets(&cpi->common, x, xd, mi_row, mi_col);
299    xd->mi[0]->sb_type = bsize;
300  }
301}
302
303typedef struct {
304  // This struct is used for computing variance in choose_partitioning(), where
305  // the max number of samples within a superblock is 16x16 (with 4x4 avg). Even
306  // in high bitdepth, uint32_t is enough for sum_square_error (2^12 * 2^12 * 16
307  // * 16 = 2^32).
308  uint32_t sum_square_error;
309  int32_t sum_error;
310  int log2_count;
311  int variance;
312} var;
313
314typedef struct {
315  var none;
316  var horz[2];
317  var vert[2];
318} partition_variance;
319
320typedef struct {
321  partition_variance part_variances;
322  var split[4];
323} v4x4;
324
325typedef struct {
326  partition_variance part_variances;
327  v4x4 split[4];
328} v8x8;
329
330typedef struct {
331  partition_variance part_variances;
332  v8x8 split[4];
333} v16x16;
334
335typedef struct {
336  partition_variance part_variances;
337  v16x16 split[4];
338} v32x32;
339
340typedef struct {
341  partition_variance part_variances;
342  v32x32 split[4];
343} v64x64;
344
345typedef struct {
346  partition_variance *part_variances;
347  var *split[4];
348} variance_node;
349
350typedef enum {
351  V16X16,
352  V32X32,
353  V64X64,
354} TREE_LEVEL;
355
356static void tree_to_node(void *data, BLOCK_SIZE bsize, variance_node *node) {
357  int i;
358  node->part_variances = NULL;
359  switch (bsize) {
360    case BLOCK_64X64: {
361      v64x64 *vt = (v64x64 *)data;
362      node->part_variances = &vt->part_variances;
363      for (i = 0; i < 4; i++)
364        node->split[i] = &vt->split[i].part_variances.none;
365      break;
366    }
367    case BLOCK_32X32: {
368      v32x32 *vt = (v32x32 *)data;
369      node->part_variances = &vt->part_variances;
370      for (i = 0; i < 4; i++)
371        node->split[i] = &vt->split[i].part_variances.none;
372      break;
373    }
374    case BLOCK_16X16: {
375      v16x16 *vt = (v16x16 *)data;
376      node->part_variances = &vt->part_variances;
377      for (i = 0; i < 4; i++)
378        node->split[i] = &vt->split[i].part_variances.none;
379      break;
380    }
381    case BLOCK_8X8: {
382      v8x8 *vt = (v8x8 *)data;
383      node->part_variances = &vt->part_variances;
384      for (i = 0; i < 4; i++)
385        node->split[i] = &vt->split[i].part_variances.none;
386      break;
387    }
388    case BLOCK_4X4: {
389      v4x4 *vt = (v4x4 *)data;
390      node->part_variances = &vt->part_variances;
391      for (i = 0; i < 4; i++) node->split[i] = &vt->split[i];
392      break;
393    }
394    default: {
395      assert(0);
396      break;
397    }
398  }
399}
400
401// Set variance values given sum square error, sum error, count.
402static void fill_variance(uint32_t s2, int32_t s, int c, var *v) {
403  v->sum_square_error = s2;
404  v->sum_error = s;
405  v->log2_count = c;
406}
407
408static void get_variance(var *v) {
409  v->variance =
410      (int)(256 * (v->sum_square_error -
411                   ((v->sum_error * v->sum_error) >> v->log2_count)) >>
412            v->log2_count);
413}
414
415static void sum_2_variances(const var *a, const var *b, var *r) {
416  assert(a->log2_count == b->log2_count);
417  fill_variance(a->sum_square_error + b->sum_square_error,
418                a->sum_error + b->sum_error, a->log2_count + 1, r);
419}
420
421static void fill_variance_tree(void *data, BLOCK_SIZE bsize) {
422  variance_node node;
423  memset(&node, 0, sizeof(node));
424  tree_to_node(data, bsize, &node);
425  sum_2_variances(node.split[0], node.split[1], &node.part_variances->horz[0]);
426  sum_2_variances(node.split[2], node.split[3], &node.part_variances->horz[1]);
427  sum_2_variances(node.split[0], node.split[2], &node.part_variances->vert[0]);
428  sum_2_variances(node.split[1], node.split[3], &node.part_variances->vert[1]);
429  sum_2_variances(&node.part_variances->vert[0], &node.part_variances->vert[1],
430                  &node.part_variances->none);
431}
432
433static int set_vt_partitioning(VP9_COMP *cpi, MACROBLOCK *const x,
434                               MACROBLOCKD *const xd, void *data,
435                               BLOCK_SIZE bsize, int mi_row, int mi_col,
436                               int64_t threshold, BLOCK_SIZE bsize_min,
437                               int force_split) {
438  VP9_COMMON *const cm = &cpi->common;
439  variance_node vt;
440  const int block_width = num_8x8_blocks_wide_lookup[bsize];
441  const int block_height = num_8x8_blocks_high_lookup[bsize];
442
443  assert(block_height == block_width);
444  tree_to_node(data, bsize, &vt);
445
446  if (force_split == 1) return 0;
447
448  // For bsize=bsize_min (16x16/8x8 for 8x8/4x4 downsampling), select if
449  // variance is below threshold, otherwise split will be selected.
450  // No check for vert/horiz split as too few samples for variance.
451  if (bsize == bsize_min) {
452    // Variance already computed to set the force_split.
453    if (cm->frame_type == KEY_FRAME) get_variance(&vt.part_variances->none);
454    if (mi_col + block_width / 2 < cm->mi_cols &&
455        mi_row + block_height / 2 < cm->mi_rows &&
456        vt.part_variances->none.variance < threshold) {
457      set_block_size(cpi, x, xd, mi_row, mi_col, bsize);
458      return 1;
459    }
460    return 0;
461  } else if (bsize > bsize_min) {
462    // Variance already computed to set the force_split.
463    if (cm->frame_type == KEY_FRAME) get_variance(&vt.part_variances->none);
464    // For key frame: take split for bsize above 32X32 or very high variance.
465    if (cm->frame_type == KEY_FRAME &&
466        (bsize > BLOCK_32X32 ||
467         vt.part_variances->none.variance > (threshold << 4))) {
468      return 0;
469    }
470    // If variance is low, take the bsize (no split).
471    if (mi_col + block_width / 2 < cm->mi_cols &&
472        mi_row + block_height / 2 < cm->mi_rows &&
473        vt.part_variances->none.variance < threshold) {
474      set_block_size(cpi, x, xd, mi_row, mi_col, bsize);
475      return 1;
476    }
477
478    // Check vertical split.
479    if (mi_row + block_height / 2 < cm->mi_rows) {
480      BLOCK_SIZE subsize = get_subsize(bsize, PARTITION_VERT);
481      get_variance(&vt.part_variances->vert[0]);
482      get_variance(&vt.part_variances->vert[1]);
483      if (vt.part_variances->vert[0].variance < threshold &&
484          vt.part_variances->vert[1].variance < threshold &&
485          get_plane_block_size(subsize, &xd->plane[1]) < BLOCK_INVALID) {
486        set_block_size(cpi, x, xd, mi_row, mi_col, subsize);
487        set_block_size(cpi, x, xd, mi_row, mi_col + block_width / 2, subsize);
488        return 1;
489      }
490    }
491    // Check horizontal split.
492    if (mi_col + block_width / 2 < cm->mi_cols) {
493      BLOCK_SIZE subsize = get_subsize(bsize, PARTITION_HORZ);
494      get_variance(&vt.part_variances->horz[0]);
495      get_variance(&vt.part_variances->horz[1]);
496      if (vt.part_variances->horz[0].variance < threshold &&
497          vt.part_variances->horz[1].variance < threshold &&
498          get_plane_block_size(subsize, &xd->plane[1]) < BLOCK_INVALID) {
499        set_block_size(cpi, x, xd, mi_row, mi_col, subsize);
500        set_block_size(cpi, x, xd, mi_row + block_height / 2, mi_col, subsize);
501        return 1;
502      }
503    }
504
505    return 0;
506  }
507  return 0;
508}
509
510static int64_t scale_part_thresh_sumdiff(int64_t threshold_base, int speed,
511                                         int width, int height,
512                                         int content_state) {
513  if (speed >= 8) {
514    if (width <= 640 && height <= 480)
515      return (5 * threshold_base) >> 2;
516    else if ((content_state == kLowSadLowSumdiff) ||
517             (content_state == kHighSadLowSumdiff) ||
518             (content_state == kLowVarHighSumdiff))
519      return (5 * threshold_base) >> 2;
520  } else if (speed == 7) {
521    if ((content_state == kLowSadLowSumdiff) ||
522        (content_state == kHighSadLowSumdiff) ||
523        (content_state == kLowVarHighSumdiff)) {
524      return (5 * threshold_base) >> 2;
525    }
526  }
527  return threshold_base;
528}
529
530// Set the variance split thresholds for following the block sizes:
531// 0 - threshold_64x64, 1 - threshold_32x32, 2 - threshold_16x16,
532// 3 - vbp_threshold_8x8. vbp_threshold_8x8 (to split to 4x4 partition) is
533// currently only used on key frame.
534static void set_vbp_thresholds(VP9_COMP *cpi, int64_t thresholds[], int q,
535                               int content_state) {
536  VP9_COMMON *const cm = &cpi->common;
537  const int is_key_frame = (cm->frame_type == KEY_FRAME);
538  const int threshold_multiplier = is_key_frame ? 20 : 1;
539  int64_t threshold_base =
540      (int64_t)(threshold_multiplier * cpi->y_dequant[q][1]);
541
542  if (is_key_frame) {
543    thresholds[0] = threshold_base;
544    thresholds[1] = threshold_base >> 2;
545    thresholds[2] = threshold_base >> 2;
546    thresholds[3] = threshold_base << 2;
547  } else {
548    // Increase base variance threshold based on estimated noise level.
549    if (cpi->noise_estimate.enabled && cm->width >= 640 && cm->height >= 480) {
550      NOISE_LEVEL noise_level =
551          vp9_noise_estimate_extract_level(&cpi->noise_estimate);
552      if (noise_level == kHigh)
553        threshold_base = 3 * threshold_base;
554      else if (noise_level == kMedium)
555        threshold_base = threshold_base << 1;
556      else if (noise_level < kLow)
557        threshold_base = (7 * threshold_base) >> 3;
558    }
559#if CONFIG_VP9_TEMPORAL_DENOISING
560    if (cpi->oxcf.noise_sensitivity > 0 && denoise_svc(cpi) &&
561        cpi->oxcf.speed > 5 && cpi->denoiser.denoising_level >= kDenLow)
562      threshold_base =
563          vp9_scale_part_thresh(threshold_base, cpi->denoiser.denoising_level,
564                                content_state, cpi->svc.temporal_layer_id);
565    else
566      threshold_base =
567          scale_part_thresh_sumdiff(threshold_base, cpi->oxcf.speed, cm->width,
568                                    cm->height, content_state);
569#else
570    // Increase base variance threshold based on content_state/sum_diff level.
571    threshold_base = scale_part_thresh_sumdiff(
572        threshold_base, cpi->oxcf.speed, cm->width, cm->height, content_state);
573#endif
574    thresholds[0] = threshold_base;
575    thresholds[2] = threshold_base << cpi->oxcf.speed;
576    if (cm->width >= 1280 && cm->height >= 720 && cpi->oxcf.speed < 7)
577      thresholds[2] = thresholds[2] << 1;
578    if (cm->width <= 352 && cm->height <= 288) {
579      thresholds[0] = threshold_base >> 3;
580      thresholds[1] = threshold_base >> 1;
581      thresholds[2] = threshold_base << 3;
582    } else if (cm->width < 1280 && cm->height < 720) {
583      thresholds[1] = (5 * threshold_base) >> 2;
584    } else if (cm->width < 1920 && cm->height < 1080) {
585      thresholds[1] = threshold_base << 1;
586    } else {
587      thresholds[1] = (5 * threshold_base) >> 1;
588    }
589  }
590}
591
592void vp9_set_variance_partition_thresholds(VP9_COMP *cpi, int q,
593                                           int content_state) {
594  VP9_COMMON *const cm = &cpi->common;
595  SPEED_FEATURES *const sf = &cpi->sf;
596  const int is_key_frame = (cm->frame_type == KEY_FRAME);
597  if (sf->partition_search_type != VAR_BASED_PARTITION &&
598      sf->partition_search_type != REFERENCE_PARTITION) {
599    return;
600  } else {
601    set_vbp_thresholds(cpi, cpi->vbp_thresholds, q, content_state);
602    // The thresholds below are not changed locally.
603    if (is_key_frame) {
604      cpi->vbp_threshold_sad = 0;
605      cpi->vbp_threshold_copy = 0;
606      cpi->vbp_bsize_min = BLOCK_8X8;
607    } else {
608      if (cm->width <= 352 && cm->height <= 288)
609        cpi->vbp_threshold_sad = 10;
610      else
611        cpi->vbp_threshold_sad = (cpi->y_dequant[q][1] << 1) > 1000
612                                     ? (cpi->y_dequant[q][1] << 1)
613                                     : 1000;
614      cpi->vbp_bsize_min = BLOCK_16X16;
615      if (cm->width <= 352 && cm->height <= 288)
616        cpi->vbp_threshold_copy = 4000;
617      else if (cm->width <= 640 && cm->height <= 360)
618        cpi->vbp_threshold_copy = 8000;
619      else
620        cpi->vbp_threshold_copy = (cpi->y_dequant[q][1] << 3) > 8000
621                                      ? (cpi->y_dequant[q][1] << 3)
622                                      : 8000;
623    }
624    cpi->vbp_threshold_minmax = 15 + (q >> 3);
625  }
626}
627
628// Compute the minmax over the 8x8 subblocks.
629static int compute_minmax_8x8(const uint8_t *s, int sp, const uint8_t *d,
630                              int dp, int x16_idx, int y16_idx,
631#if CONFIG_VP9_HIGHBITDEPTH
632                              int highbd_flag,
633#endif
634                              int pixels_wide, int pixels_high) {
635  int k;
636  int minmax_max = 0;
637  int minmax_min = 255;
638  // Loop over the 4 8x8 subblocks.
639  for (k = 0; k < 4; k++) {
640    int x8_idx = x16_idx + ((k & 1) << 3);
641    int y8_idx = y16_idx + ((k >> 1) << 3);
642    int min = 0;
643    int max = 0;
644    if (x8_idx < pixels_wide && y8_idx < pixels_high) {
645#if CONFIG_VP9_HIGHBITDEPTH
646      if (highbd_flag & YV12_FLAG_HIGHBITDEPTH) {
647        vpx_highbd_minmax_8x8(s + y8_idx * sp + x8_idx, sp,
648                              d + y8_idx * dp + x8_idx, dp, &min, &max);
649      } else {
650        vpx_minmax_8x8(s + y8_idx * sp + x8_idx, sp, d + y8_idx * dp + x8_idx,
651                       dp, &min, &max);
652      }
653#else
654      vpx_minmax_8x8(s + y8_idx * sp + x8_idx, sp, d + y8_idx * dp + x8_idx, dp,
655                     &min, &max);
656#endif
657      if ((max - min) > minmax_max) minmax_max = (max - min);
658      if ((max - min) < minmax_min) minmax_min = (max - min);
659    }
660  }
661  return (minmax_max - minmax_min);
662}
663
664static void fill_variance_4x4avg(const uint8_t *s, int sp, const uint8_t *d,
665                                 int dp, int x8_idx, int y8_idx, v8x8 *vst,
666#if CONFIG_VP9_HIGHBITDEPTH
667                                 int highbd_flag,
668#endif
669                                 int pixels_wide, int pixels_high,
670                                 int is_key_frame) {
671  int k;
672  for (k = 0; k < 4; k++) {
673    int x4_idx = x8_idx + ((k & 1) << 2);
674    int y4_idx = y8_idx + ((k >> 1) << 2);
675    unsigned int sse = 0;
676    int sum = 0;
677    if (x4_idx < pixels_wide && y4_idx < pixels_high) {
678      int s_avg;
679      int d_avg = 128;
680#if CONFIG_VP9_HIGHBITDEPTH
681      if (highbd_flag & YV12_FLAG_HIGHBITDEPTH) {
682        s_avg = vpx_highbd_avg_4x4(s + y4_idx * sp + x4_idx, sp);
683        if (!is_key_frame)
684          d_avg = vpx_highbd_avg_4x4(d + y4_idx * dp + x4_idx, dp);
685      } else {
686        s_avg = vpx_avg_4x4(s + y4_idx * sp + x4_idx, sp);
687        if (!is_key_frame) d_avg = vpx_avg_4x4(d + y4_idx * dp + x4_idx, dp);
688      }
689#else
690      s_avg = vpx_avg_4x4(s + y4_idx * sp + x4_idx, sp);
691      if (!is_key_frame) d_avg = vpx_avg_4x4(d + y4_idx * dp + x4_idx, dp);
692#endif
693      sum = s_avg - d_avg;
694      sse = sum * sum;
695    }
696    fill_variance(sse, sum, 0, &vst->split[k].part_variances.none);
697  }
698}
699
700static void fill_variance_8x8avg(const uint8_t *s, int sp, const uint8_t *d,
701                                 int dp, int x16_idx, int y16_idx, v16x16 *vst,
702#if CONFIG_VP9_HIGHBITDEPTH
703                                 int highbd_flag,
704#endif
705                                 int pixels_wide, int pixels_high,
706                                 int is_key_frame) {
707  int k;
708  for (k = 0; k < 4; k++) {
709    int x8_idx = x16_idx + ((k & 1) << 3);
710    int y8_idx = y16_idx + ((k >> 1) << 3);
711    unsigned int sse = 0;
712    int sum = 0;
713    if (x8_idx < pixels_wide && y8_idx < pixels_high) {
714      int s_avg;
715      int d_avg = 128;
716#if CONFIG_VP9_HIGHBITDEPTH
717      if (highbd_flag & YV12_FLAG_HIGHBITDEPTH) {
718        s_avg = vpx_highbd_avg_8x8(s + y8_idx * sp + x8_idx, sp);
719        if (!is_key_frame)
720          d_avg = vpx_highbd_avg_8x8(d + y8_idx * dp + x8_idx, dp);
721      } else {
722        s_avg = vpx_avg_8x8(s + y8_idx * sp + x8_idx, sp);
723        if (!is_key_frame) d_avg = vpx_avg_8x8(d + y8_idx * dp + x8_idx, dp);
724      }
725#else
726      s_avg = vpx_avg_8x8(s + y8_idx * sp + x8_idx, sp);
727      if (!is_key_frame) d_avg = vpx_avg_8x8(d + y8_idx * dp + x8_idx, dp);
728#endif
729      sum = s_avg - d_avg;
730      sse = sum * sum;
731    }
732    fill_variance(sse, sum, 0, &vst->split[k].part_variances.none);
733  }
734}
735
736// Check if most of the superblock is skin content, and if so, force split to
737// 32x32, and set x->sb_is_skin for use in mode selection.
738static int skin_sb_split(VP9_COMP *cpi, MACROBLOCK *x, const int low_res,
739                         int mi_row, int mi_col, int *force_split) {
740  VP9_COMMON *const cm = &cpi->common;
741#if CONFIG_VP9_HIGHBITDEPTH
742  if (cm->use_highbitdepth) return 0;
743#endif
744  // Avoid checking superblocks on/near boundary and avoid low resolutions.
745  // Note superblock may still pick 64X64 if y_sad is very small
746  // (i.e., y_sad < cpi->vbp_threshold_sad) below. For now leave this as is.
747  if (!low_res && (mi_col >= 8 && mi_col + 8 < cm->mi_cols && mi_row >= 8 &&
748                   mi_row + 8 < cm->mi_rows)) {
749    int num_16x16_skin = 0;
750    int num_16x16_nonskin = 0;
751    uint8_t *ysignal = x->plane[0].src.buf;
752    uint8_t *usignal = x->plane[1].src.buf;
753    uint8_t *vsignal = x->plane[2].src.buf;
754    int sp = x->plane[0].src.stride;
755    int spuv = x->plane[1].src.stride;
756    const int block_index = mi_row * cm->mi_cols + mi_col;
757    const int bw = num_8x8_blocks_wide_lookup[BLOCK_64X64];
758    const int bh = num_8x8_blocks_high_lookup[BLOCK_64X64];
759    const int xmis = VPXMIN(cm->mi_cols - mi_col, bw);
760    const int ymis = VPXMIN(cm->mi_rows - mi_row, bh);
761    // Loop through the 16x16 sub-blocks.
762    int i, j;
763    for (i = 0; i < ymis; i += 2) {
764      for (j = 0; j < xmis; j += 2) {
765        int bl_index = block_index + i * cm->mi_cols + j;
766        int is_skin = cpi->skin_map[bl_index];
767        num_16x16_skin += is_skin;
768        num_16x16_nonskin += (1 - is_skin);
769        if (num_16x16_nonskin > 3) {
770          // Exit loop if at least 4 of the 16x16 blocks are not skin.
771          i = ymis;
772          break;
773        }
774        ysignal += 16;
775        usignal += 8;
776        vsignal += 8;
777      }
778      ysignal += (sp << 4) - 64;
779      usignal += (spuv << 3) - 32;
780      vsignal += (spuv << 3) - 32;
781    }
782    if (num_16x16_skin > 12) {
783      *force_split = 1;
784      return 1;
785    }
786  }
787  return 0;
788}
789
790static void set_low_temp_var_flag(VP9_COMP *cpi, MACROBLOCK *x, MACROBLOCKD *xd,
791                                  v64x64 *vt, int64_t thresholds[],
792                                  MV_REFERENCE_FRAME ref_frame_partition,
793                                  int mi_col, int mi_row) {
794  int i, j;
795  VP9_COMMON *const cm = &cpi->common;
796  const int mv_thr = cm->width > 640 ? 8 : 4;
797  // Check temporal variance for bsize >= 16x16, if LAST_FRAME was selected and
798  // int_pro mv is small. If the temporal variance is small set the flag
799  // variance_low for the block. The variance threshold can be adjusted, the
800  // higher the more aggressive.
801  if (ref_frame_partition == LAST_FRAME &&
802      (cpi->sf.short_circuit_low_temp_var == 1 ||
803       (xd->mi[0]->mv[0].as_mv.col < mv_thr &&
804        xd->mi[0]->mv[0].as_mv.col > -mv_thr &&
805        xd->mi[0]->mv[0].as_mv.row < mv_thr &&
806        xd->mi[0]->mv[0].as_mv.row > -mv_thr))) {
807    if (xd->mi[0]->sb_type == BLOCK_64X64) {
808      if ((vt->part_variances).none.variance < (thresholds[0] >> 1))
809        x->variance_low[0] = 1;
810    } else if (xd->mi[0]->sb_type == BLOCK_64X32) {
811      for (i = 0; i < 2; i++) {
812        if (vt->part_variances.horz[i].variance < (thresholds[0] >> 2))
813          x->variance_low[i + 1] = 1;
814      }
815    } else if (xd->mi[0]->sb_type == BLOCK_32X64) {
816      for (i = 0; i < 2; i++) {
817        if (vt->part_variances.vert[i].variance < (thresholds[0] >> 2))
818          x->variance_low[i + 3] = 1;
819      }
820    } else {
821      for (i = 0; i < 4; i++) {
822        const int idx[4][2] = { { 0, 0 }, { 0, 4 }, { 4, 0 }, { 4, 4 } };
823        const int idx_str =
824            cm->mi_stride * (mi_row + idx[i][0]) + mi_col + idx[i][1];
825        MODE_INFO **this_mi = cm->mi_grid_visible + idx_str;
826
827        if (cm->mi_cols <= mi_col + idx[i][1] ||
828            cm->mi_rows <= mi_row + idx[i][0])
829          continue;
830
831        if ((*this_mi)->sb_type == BLOCK_32X32) {
832          int64_t threshold_32x32 = (cpi->sf.short_circuit_low_temp_var == 1 ||
833                                     cpi->sf.short_circuit_low_temp_var == 3)
834                                        ? ((5 * thresholds[1]) >> 3)
835                                        : (thresholds[1] >> 1);
836          if (vt->split[i].part_variances.none.variance < threshold_32x32)
837            x->variance_low[i + 5] = 1;
838        } else if (cpi->sf.short_circuit_low_temp_var >= 2) {
839          // For 32x16 and 16x32 blocks, the flag is set on each 16x16 block
840          // inside.
841          if ((*this_mi)->sb_type == BLOCK_16X16 ||
842              (*this_mi)->sb_type == BLOCK_32X16 ||
843              (*this_mi)->sb_type == BLOCK_16X32) {
844            for (j = 0; j < 4; j++) {
845              if (vt->split[i].split[j].part_variances.none.variance <
846                  (thresholds[2] >> 8))
847                x->variance_low[(i << 2) + j + 9] = 1;
848            }
849          }
850        }
851      }
852    }
853  }
854}
855
856static void copy_partitioning_helper(VP9_COMP *cpi, MACROBLOCK *x,
857                                     MACROBLOCKD *xd, BLOCK_SIZE bsize,
858                                     int mi_row, int mi_col) {
859  VP9_COMMON *const cm = &cpi->common;
860  BLOCK_SIZE *prev_part = cpi->prev_partition;
861  int start_pos = mi_row * cm->mi_stride + mi_col;
862
863  const int bsl = b_width_log2_lookup[bsize];
864  const int bs = (1 << bsl) >> 2;
865  BLOCK_SIZE subsize;
866  PARTITION_TYPE partition;
867
868  if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols) return;
869
870  partition = partition_lookup[bsl][prev_part[start_pos]];
871  subsize = get_subsize(bsize, partition);
872
873  if (subsize < BLOCK_8X8) {
874    set_block_size(cpi, x, xd, mi_row, mi_col, bsize);
875  } else {
876    switch (partition) {
877      case PARTITION_NONE:
878        set_block_size(cpi, x, xd, mi_row, mi_col, bsize);
879        break;
880      case PARTITION_HORZ:
881        set_block_size(cpi, x, xd, mi_row, mi_col, subsize);
882        set_block_size(cpi, x, xd, mi_row + bs, mi_col, subsize);
883        break;
884      case PARTITION_VERT:
885        set_block_size(cpi, x, xd, mi_row, mi_col, subsize);
886        set_block_size(cpi, x, xd, mi_row, mi_col + bs, subsize);
887        break;
888      case PARTITION_SPLIT:
889        copy_partitioning_helper(cpi, x, xd, subsize, mi_row, mi_col);
890        copy_partitioning_helper(cpi, x, xd, subsize, mi_row + bs, mi_col);
891        copy_partitioning_helper(cpi, x, xd, subsize, mi_row, mi_col + bs);
892        copy_partitioning_helper(cpi, x, xd, subsize, mi_row + bs, mi_col + bs);
893        break;
894      default: assert(0);
895    }
896  }
897}
898
899static int copy_partitioning(VP9_COMP *cpi, MACROBLOCK *x, MACROBLOCKD *xd,
900                             int mi_row, int mi_col, int segment_id,
901                             int sb_offset) {
902  int svc_copy_allowed = 1;
903  int frames_since_key_thresh = 1;
904  if (cpi->use_svc) {
905    // For SVC, don't allow copy if base spatial layer is key frame, or if
906    // frame is not a temporal enhancement layer frame.
907    int layer = LAYER_IDS_TO_IDX(0, cpi->svc.temporal_layer_id,
908                                 cpi->svc.number_temporal_layers);
909    const LAYER_CONTEXT *lc = &cpi->svc.layer_context[layer];
910    if (lc->is_key_frame || !cpi->svc.non_reference_frame) svc_copy_allowed = 0;
911    frames_since_key_thresh = cpi->svc.number_spatial_layers << 1;
912  }
913  if (cpi->rc.frames_since_key > frames_since_key_thresh && svc_copy_allowed &&
914      !cpi->resize_pending && segment_id == CR_SEGMENT_ID_BASE &&
915      cpi->prev_segment_id[sb_offset] == CR_SEGMENT_ID_BASE &&
916      cpi->copied_frame_cnt[sb_offset] < cpi->max_copied_frame) {
917    if (cpi->prev_partition != NULL) {
918      copy_partitioning_helper(cpi, x, xd, BLOCK_64X64, mi_row, mi_col);
919      cpi->copied_frame_cnt[sb_offset] += 1;
920      memcpy(x->variance_low, &(cpi->prev_variance_low[sb_offset * 25]),
921             sizeof(x->variance_low));
922      return 1;
923    }
924  }
925
926  return 0;
927}
928
929static int scale_partitioning_svc(VP9_COMP *cpi, MACROBLOCK *x, MACROBLOCKD *xd,
930                                  BLOCK_SIZE bsize, int mi_row, int mi_col,
931                                  int mi_row_high, int mi_col_high) {
932  VP9_COMMON *const cm = &cpi->common;
933  SVC *const svc = &cpi->svc;
934  BLOCK_SIZE *prev_part = svc->prev_partition_svc;
935  // Variables with _high are for higher resolution.
936  int bsize_high = 0;
937  int subsize_high = 0;
938  const int bsl_high = b_width_log2_lookup[bsize];
939  const int bs_high = (1 << bsl_high) >> 2;
940  const int has_rows = (mi_row_high + bs_high) < cm->mi_rows;
941  const int has_cols = (mi_col_high + bs_high) < cm->mi_cols;
942
943  const int row_boundary_block_scale_factor[BLOCK_SIZES] = {
944    13, 13, 13, 1, 0, 1, 1, 0, 1, 1, 0, 1, 0
945  };
946  const int col_boundary_block_scale_factor[BLOCK_SIZES] = {
947    13, 13, 13, 2, 2, 0, 2, 2, 0, 2, 2, 0, 0
948  };
949  int start_pos;
950  BLOCK_SIZE bsize_low;
951  PARTITION_TYPE partition_high;
952
953  if (mi_row_high >= cm->mi_rows || mi_col_high >= cm->mi_cols) return 0;
954  if (mi_row >= (cm->mi_rows >> 1) || mi_col >= (cm->mi_cols >> 1)) return 0;
955
956  // Find corresponding (mi_col/mi_row) block down-scaled by 2x2.
957  start_pos = mi_row * (svc->mi_stride[svc->spatial_layer_id - 1]) + mi_col;
958  bsize_low = prev_part[start_pos];
959  // The block size is too big for boundaries. Do variance based partitioning.
960  if ((!has_rows || !has_cols) && bsize_low > BLOCK_16X16) return 1;
961
962  // For reference frames: return 1 (do variance-based partitioning) if the
963  // superblock is not low source sad and lower-resoln bsize is below 32x32.
964  if (!cpi->svc.non_reference_frame && !x->skip_low_source_sad &&
965      bsize_low < BLOCK_32X32)
966    return 1;
967
968  // Scale up block size by 2x2. Force 64x64 for size larger than 32x32.
969  if (bsize_low < BLOCK_32X32) {
970    bsize_high = bsize_low + 3;
971  } else if (bsize_low >= BLOCK_32X32) {
972    bsize_high = BLOCK_64X64;
973  }
974  // Scale up blocks on boundary.
975  if (!has_cols && has_rows) {
976    bsize_high = bsize_low + row_boundary_block_scale_factor[bsize_low];
977  } else if (has_cols && !has_rows) {
978    bsize_high = bsize_low + col_boundary_block_scale_factor[bsize_low];
979  } else if (!has_cols && !has_rows) {
980    bsize_high = bsize_low;
981  }
982
983  partition_high = partition_lookup[bsl_high][bsize_high];
984  subsize_high = get_subsize(bsize, partition_high);
985
986  if (subsize_high < BLOCK_8X8) {
987    set_block_size(cpi, x, xd, mi_row_high, mi_col_high, bsize_high);
988  } else {
989    const int bsl = b_width_log2_lookup[bsize];
990    const int bs = (1 << bsl) >> 2;
991    switch (partition_high) {
992      case PARTITION_NONE:
993        set_block_size(cpi, x, xd, mi_row_high, mi_col_high, bsize_high);
994        break;
995      case PARTITION_HORZ:
996        set_block_size(cpi, x, xd, mi_row_high, mi_col_high, subsize_high);
997        if (subsize_high < BLOCK_64X64)
998          set_block_size(cpi, x, xd, mi_row_high + bs_high, mi_col_high,
999                         subsize_high);
1000        break;
1001      case PARTITION_VERT:
1002        set_block_size(cpi, x, xd, mi_row_high, mi_col_high, subsize_high);
1003        if (subsize_high < BLOCK_64X64)
1004          set_block_size(cpi, x, xd, mi_row_high, mi_col_high + bs_high,
1005                         subsize_high);
1006        break;
1007      case PARTITION_SPLIT:
1008        if (scale_partitioning_svc(cpi, x, xd, subsize_high, mi_row, mi_col,
1009                                   mi_row_high, mi_col_high))
1010          return 1;
1011        if (scale_partitioning_svc(cpi, x, xd, subsize_high, mi_row + (bs >> 1),
1012                                   mi_col, mi_row_high + bs_high, mi_col_high))
1013          return 1;
1014        if (scale_partitioning_svc(cpi, x, xd, subsize_high, mi_row,
1015                                   mi_col + (bs >> 1), mi_row_high,
1016                                   mi_col_high + bs_high))
1017          return 1;
1018        if (scale_partitioning_svc(cpi, x, xd, subsize_high, mi_row + (bs >> 1),
1019                                   mi_col + (bs >> 1), mi_row_high + bs_high,
1020                                   mi_col_high + bs_high))
1021          return 1;
1022        break;
1023      default: assert(0);
1024    }
1025  }
1026
1027  return 0;
1028}
1029
1030static void update_partition_svc(VP9_COMP *cpi, BLOCK_SIZE bsize, int mi_row,
1031                                 int mi_col) {
1032  VP9_COMMON *const cm = &cpi->common;
1033  BLOCK_SIZE *prev_part = cpi->svc.prev_partition_svc;
1034  int start_pos = mi_row * cm->mi_stride + mi_col;
1035  const int bsl = b_width_log2_lookup[bsize];
1036  const int bs = (1 << bsl) >> 2;
1037  BLOCK_SIZE subsize;
1038  PARTITION_TYPE partition;
1039  const MODE_INFO *mi = NULL;
1040  int xx, yy;
1041
1042  if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols) return;
1043
1044  mi = cm->mi_grid_visible[start_pos];
1045  partition = partition_lookup[bsl][mi->sb_type];
1046  subsize = get_subsize(bsize, partition);
1047  if (subsize < BLOCK_8X8) {
1048    prev_part[start_pos] = bsize;
1049  } else {
1050    switch (partition) {
1051      case PARTITION_NONE:
1052        prev_part[start_pos] = bsize;
1053        if (bsize == BLOCK_64X64) {
1054          for (xx = 0; xx < 8; xx += 4)
1055            for (yy = 0; yy < 8; yy += 4) {
1056              if ((mi_row + xx < cm->mi_rows) && (mi_col + yy < cm->mi_cols))
1057                prev_part[start_pos + xx * cm->mi_stride + yy] = bsize;
1058            }
1059        }
1060        break;
1061      case PARTITION_HORZ:
1062        prev_part[start_pos] = subsize;
1063        if (mi_row + bs < cm->mi_rows)
1064          prev_part[start_pos + bs * cm->mi_stride] = subsize;
1065        break;
1066      case PARTITION_VERT:
1067        prev_part[start_pos] = subsize;
1068        if (mi_col + bs < cm->mi_cols) prev_part[start_pos + bs] = subsize;
1069        break;
1070      case PARTITION_SPLIT:
1071        update_partition_svc(cpi, subsize, mi_row, mi_col);
1072        update_partition_svc(cpi, subsize, mi_row + bs, mi_col);
1073        update_partition_svc(cpi, subsize, mi_row, mi_col + bs);
1074        update_partition_svc(cpi, subsize, mi_row + bs, mi_col + bs);
1075        break;
1076      default: assert(0);
1077    }
1078  }
1079}
1080
1081static void update_prev_partition_helper(VP9_COMP *cpi, BLOCK_SIZE bsize,
1082                                         int mi_row, int mi_col) {
1083  VP9_COMMON *const cm = &cpi->common;
1084  BLOCK_SIZE *prev_part = cpi->prev_partition;
1085  int start_pos = mi_row * cm->mi_stride + mi_col;
1086  const int bsl = b_width_log2_lookup[bsize];
1087  const int bs = (1 << bsl) >> 2;
1088  BLOCK_SIZE subsize;
1089  PARTITION_TYPE partition;
1090  const MODE_INFO *mi = NULL;
1091
1092  if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols) return;
1093
1094  mi = cm->mi_grid_visible[start_pos];
1095  partition = partition_lookup[bsl][mi->sb_type];
1096  subsize = get_subsize(bsize, partition);
1097  if (subsize < BLOCK_8X8) {
1098    prev_part[start_pos] = bsize;
1099  } else {
1100    switch (partition) {
1101      case PARTITION_NONE: prev_part[start_pos] = bsize; break;
1102      case PARTITION_HORZ:
1103        prev_part[start_pos] = subsize;
1104        if (mi_row + bs < cm->mi_rows)
1105          prev_part[start_pos + bs * cm->mi_stride] = subsize;
1106        break;
1107      case PARTITION_VERT:
1108        prev_part[start_pos] = subsize;
1109        if (mi_col + bs < cm->mi_cols) prev_part[start_pos + bs] = subsize;
1110        break;
1111      case PARTITION_SPLIT:
1112        update_prev_partition_helper(cpi, subsize, mi_row, mi_col);
1113        update_prev_partition_helper(cpi, subsize, mi_row + bs, mi_col);
1114        update_prev_partition_helper(cpi, subsize, mi_row, mi_col + bs);
1115        update_prev_partition_helper(cpi, subsize, mi_row + bs, mi_col + bs);
1116        break;
1117      default: assert(0);
1118    }
1119  }
1120}
1121
1122static void update_prev_partition(VP9_COMP *cpi, MACROBLOCK *x, int segment_id,
1123                                  int mi_row, int mi_col, int sb_offset) {
1124  update_prev_partition_helper(cpi, BLOCK_64X64, mi_row, mi_col);
1125  cpi->prev_segment_id[sb_offset] = segment_id;
1126  memcpy(&(cpi->prev_variance_low[sb_offset * 25]), x->variance_low,
1127         sizeof(x->variance_low));
1128  // Reset the counter for copy partitioning
1129  cpi->copied_frame_cnt[sb_offset] = 0;
1130}
1131
1132static void chroma_check(VP9_COMP *cpi, MACROBLOCK *x, int bsize,
1133                         unsigned int y_sad, int is_key_frame) {
1134  int i;
1135  MACROBLOCKD *xd = &x->e_mbd;
1136
1137  if (is_key_frame) return;
1138
1139  // For speed >= 8, avoid the chroma check if y_sad is above threshold.
1140  if (cpi->oxcf.speed >= 8) {
1141    if (y_sad > cpi->vbp_thresholds[1] &&
1142        (!cpi->noise_estimate.enabled ||
1143         vp9_noise_estimate_extract_level(&cpi->noise_estimate) < kMedium))
1144      return;
1145  }
1146
1147  for (i = 1; i <= 2; ++i) {
1148    unsigned int uv_sad = UINT_MAX;
1149    struct macroblock_plane *p = &x->plane[i];
1150    struct macroblockd_plane *pd = &xd->plane[i];
1151    const BLOCK_SIZE bs = get_plane_block_size(bsize, pd);
1152
1153    if (bs != BLOCK_INVALID)
1154      uv_sad = cpi->fn_ptr[bs].sdf(p->src.buf, p->src.stride, pd->dst.buf,
1155                                   pd->dst.stride);
1156
1157    // TODO(marpan): Investigate if we should lower this threshold if
1158    // superblock is detected as skin.
1159    x->color_sensitivity[i - 1] = uv_sad > (y_sad >> 2);
1160  }
1161}
1162
1163static uint64_t avg_source_sad(VP9_COMP *cpi, MACROBLOCK *x, int shift,
1164                               int sb_offset) {
1165  unsigned int tmp_sse;
1166  uint64_t tmp_sad;
1167  unsigned int tmp_variance;
1168  const BLOCK_SIZE bsize = BLOCK_64X64;
1169  uint8_t *src_y = cpi->Source->y_buffer;
1170  int src_ystride = cpi->Source->y_stride;
1171  uint8_t *last_src_y = cpi->Last_Source->y_buffer;
1172  int last_src_ystride = cpi->Last_Source->y_stride;
1173  uint64_t avg_source_sad_threshold = 10000;
1174  uint64_t avg_source_sad_threshold2 = 12000;
1175#if CONFIG_VP9_HIGHBITDEPTH
1176  if (cpi->common.use_highbitdepth) return 0;
1177#endif
1178  src_y += shift;
1179  last_src_y += shift;
1180  tmp_sad =
1181      cpi->fn_ptr[bsize].sdf(src_y, src_ystride, last_src_y, last_src_ystride);
1182  tmp_variance = vpx_variance64x64(src_y, src_ystride, last_src_y,
1183                                   last_src_ystride, &tmp_sse);
1184  // Note: tmp_sse - tmp_variance = ((sum * sum) >> 12)
1185  if (tmp_sad < avg_source_sad_threshold)
1186    x->content_state_sb = ((tmp_sse - tmp_variance) < 25) ? kLowSadLowSumdiff
1187                                                          : kLowSadHighSumdiff;
1188  else
1189    x->content_state_sb = ((tmp_sse - tmp_variance) < 25) ? kHighSadLowSumdiff
1190                                                          : kHighSadHighSumdiff;
1191
1192  // Detect large lighting change.
1193  if (cpi->oxcf.content != VP9E_CONTENT_SCREEN &&
1194      cpi->oxcf.rc_mode == VPX_CBR && tmp_variance < (tmp_sse >> 3) &&
1195      (tmp_sse - tmp_variance) > 10000)
1196    x->content_state_sb = kLowVarHighSumdiff;
1197  else if (tmp_sad > (avg_source_sad_threshold << 1))
1198    x->content_state_sb = kVeryHighSad;
1199
1200  if (cpi->content_state_sb_fd != NULL) {
1201    if (tmp_sad < avg_source_sad_threshold2) {
1202      // Cap the increment to 255.
1203      if (cpi->content_state_sb_fd[sb_offset] < 255)
1204        cpi->content_state_sb_fd[sb_offset]++;
1205    } else {
1206      cpi->content_state_sb_fd[sb_offset] = 0;
1207    }
1208  }
1209  return tmp_sad;
1210}
1211
1212// This function chooses partitioning based on the variance between source and
1213// reconstructed last, where variance is computed for down-sampled inputs.
1214static int choose_partitioning(VP9_COMP *cpi, const TileInfo *const tile,
1215                               MACROBLOCK *x, int mi_row, int mi_col) {
1216  VP9_COMMON *const cm = &cpi->common;
1217  MACROBLOCKD *xd = &x->e_mbd;
1218  int i, j, k, m;
1219  v64x64 vt;
1220  v16x16 *vt2 = NULL;
1221  int force_split[21];
1222  int avg_32x32;
1223  int max_var_32x32 = 0;
1224  int min_var_32x32 = INT_MAX;
1225  int var_32x32;
1226  int avg_16x16[4];
1227  int maxvar_16x16[4];
1228  int minvar_16x16[4];
1229  int64_t threshold_4x4avg;
1230  NOISE_LEVEL noise_level = kLow;
1231  int content_state = 0;
1232  uint8_t *s;
1233  const uint8_t *d;
1234  int sp;
1235  int dp;
1236  int compute_minmax_variance = 1;
1237  unsigned int y_sad = UINT_MAX;
1238  BLOCK_SIZE bsize = BLOCK_64X64;
1239  // Ref frame used in partitioning.
1240  MV_REFERENCE_FRAME ref_frame_partition = LAST_FRAME;
1241  int pixels_wide = 64, pixels_high = 64;
1242  int64_t thresholds[4] = { cpi->vbp_thresholds[0], cpi->vbp_thresholds[1],
1243                            cpi->vbp_thresholds[2], cpi->vbp_thresholds[3] };
1244
1245  // For the variance computation under SVC mode, we treat the frame as key if
1246  // the reference (base layer frame) is key frame (i.e., is_key_frame == 1).
1247  const int is_key_frame =
1248      (cm->frame_type == KEY_FRAME ||
1249       (is_one_pass_cbr_svc(cpi) &&
1250        cpi->svc.layer_context[cpi->svc.temporal_layer_id].is_key_frame));
1251  // Always use 4x4 partition for key frame.
1252  const int use_4x4_partition = cm->frame_type == KEY_FRAME;
1253  const int low_res = (cm->width <= 352 && cm->height <= 288);
1254  int variance4x4downsample[16];
1255  int segment_id;
1256  int sb_offset = (cm->mi_stride >> 3) * (mi_row >> 3) + (mi_col >> 3);
1257
1258  set_offsets(cpi, tile, x, mi_row, mi_col, BLOCK_64X64);
1259  segment_id = xd->mi[0]->segment_id;
1260
1261  if (cpi->oxcf.speed >= 8 || (cpi->use_svc && cpi->svc.non_reference_frame))
1262    compute_minmax_variance = 0;
1263
1264  memset(x->variance_low, 0, sizeof(x->variance_low));
1265
1266  if (cpi->sf.use_source_sad && !is_key_frame) {
1267    int sb_offset2 = ((cm->mi_cols + 7) >> 3) * (mi_row >> 3) + (mi_col >> 3);
1268    content_state = x->content_state_sb;
1269    x->skip_low_source_sad = (content_state == kLowSadLowSumdiff ||
1270                              content_state == kLowSadHighSumdiff)
1271                                 ? 1
1272                                 : 0;
1273    x->lowvar_highsumdiff = (content_state == kLowVarHighSumdiff) ? 1 : 0;
1274    if (cpi->content_state_sb_fd != NULL)
1275      x->last_sb_high_content = cpi->content_state_sb_fd[sb_offset2];
1276
1277    // For SVC on top spatial layer: use/scale the partition from
1278    // the lower spatial resolution if svc_use_lowres_part is enabled.
1279    if (cpi->sf.svc_use_lowres_part &&
1280        cpi->svc.spatial_layer_id == cpi->svc.number_spatial_layers - 1 &&
1281        cpi->svc.prev_partition_svc != NULL && content_state != kVeryHighSad) {
1282      if (!scale_partitioning_svc(cpi, x, xd, BLOCK_64X64, mi_row >> 1,
1283                                  mi_col >> 1, mi_row, mi_col)) {
1284        if (cpi->sf.copy_partition_flag) {
1285          update_prev_partition(cpi, x, segment_id, mi_row, mi_col, sb_offset);
1286        }
1287        return 0;
1288      }
1289    }
1290    // If source_sad is low copy the partition without computing the y_sad.
1291    if (x->skip_low_source_sad && cpi->sf.copy_partition_flag &&
1292        copy_partitioning(cpi, x, xd, mi_row, mi_col, segment_id, sb_offset)) {
1293      x->sb_use_mv_part = 1;
1294      if (cpi->sf.svc_use_lowres_part &&
1295          cpi->svc.spatial_layer_id == cpi->svc.number_spatial_layers - 2)
1296        update_partition_svc(cpi, BLOCK_64X64, mi_row, mi_col);
1297      return 0;
1298    }
1299  }
1300
1301  if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ && cm->seg.enabled &&
1302      cyclic_refresh_segment_id_boosted(segment_id)) {
1303    int q = vp9_get_qindex(&cm->seg, segment_id, cm->base_qindex);
1304    set_vbp_thresholds(cpi, thresholds, q, content_state);
1305  } else {
1306    set_vbp_thresholds(cpi, thresholds, cm->base_qindex, content_state);
1307  }
1308
1309  // For non keyframes, disable 4x4 average for low resolution when speed = 8
1310  threshold_4x4avg = (cpi->oxcf.speed < 8) ? thresholds[1] << 1 : INT64_MAX;
1311
1312  if (xd->mb_to_right_edge < 0) pixels_wide += (xd->mb_to_right_edge >> 3);
1313  if (xd->mb_to_bottom_edge < 0) pixels_high += (xd->mb_to_bottom_edge >> 3);
1314
1315  s = x->plane[0].src.buf;
1316  sp = x->plane[0].src.stride;
1317
1318  // Index for force_split: 0 for 64x64, 1-4 for 32x32 blocks,
1319  // 5-20 for the 16x16 blocks.
1320  force_split[0] = 0;
1321
1322  if (!is_key_frame) {
1323    // In the case of spatial/temporal scalable coding, the assumption here is
1324    // that the temporal reference frame will always be of type LAST_FRAME.
1325    // TODO(marpan): If that assumption is broken, we need to revisit this code.
1326    MODE_INFO *mi = xd->mi[0];
1327    YV12_BUFFER_CONFIG *yv12 = get_ref_frame_buffer(cpi, LAST_FRAME);
1328
1329    const YV12_BUFFER_CONFIG *yv12_g = NULL;
1330    unsigned int y_sad_g, y_sad_thr, y_sad_last;
1331    bsize = BLOCK_32X32 + (mi_col + 4 < cm->mi_cols) * 2 +
1332            (mi_row + 4 < cm->mi_rows);
1333
1334    assert(yv12 != NULL);
1335
1336    if (!(is_one_pass_cbr_svc(cpi) && cpi->svc.spatial_layer_id)) {
1337      // For now, GOLDEN will not be used for non-zero spatial layers, since
1338      // it may not be a temporal reference.
1339      yv12_g = get_ref_frame_buffer(cpi, GOLDEN_FRAME);
1340    }
1341
1342    // Only compute y_sad_g (sad for golden reference) for speed < 8.
1343    if (cpi->oxcf.speed < 8 && yv12_g && yv12_g != yv12 &&
1344        (cpi->ref_frame_flags & VP9_GOLD_FLAG)) {
1345      vp9_setup_pre_planes(xd, 0, yv12_g, mi_row, mi_col,
1346                           &cm->frame_refs[GOLDEN_FRAME - 1].sf);
1347      y_sad_g = cpi->fn_ptr[bsize].sdf(
1348          x->plane[0].src.buf, x->plane[0].src.stride, xd->plane[0].pre[0].buf,
1349          xd->plane[0].pre[0].stride);
1350    } else {
1351      y_sad_g = UINT_MAX;
1352    }
1353
1354    if (cpi->oxcf.lag_in_frames > 0 && cpi->oxcf.rc_mode == VPX_VBR &&
1355        cpi->rc.is_src_frame_alt_ref) {
1356      yv12 = get_ref_frame_buffer(cpi, ALTREF_FRAME);
1357      vp9_setup_pre_planes(xd, 0, yv12, mi_row, mi_col,
1358                           &cm->frame_refs[ALTREF_FRAME - 1].sf);
1359      mi->ref_frame[0] = ALTREF_FRAME;
1360      y_sad_g = UINT_MAX;
1361    } else {
1362      vp9_setup_pre_planes(xd, 0, yv12, mi_row, mi_col,
1363                           &cm->frame_refs[LAST_FRAME - 1].sf);
1364      mi->ref_frame[0] = LAST_FRAME;
1365    }
1366    mi->ref_frame[1] = NONE;
1367    mi->sb_type = BLOCK_64X64;
1368    mi->mv[0].as_int = 0;
1369    mi->interp_filter = BILINEAR;
1370
1371    if (cpi->oxcf.speed >= 8 && !low_res &&
1372        x->content_state_sb != kVeryHighSad) {
1373      y_sad = cpi->fn_ptr[bsize].sdf(
1374          x->plane[0].src.buf, x->plane[0].src.stride, xd->plane[0].pre[0].buf,
1375          xd->plane[0].pre[0].stride);
1376    } else {
1377      y_sad = vp9_int_pro_motion_estimation(cpi, x, bsize, mi_row, mi_col);
1378      x->sb_use_mv_part = 1;
1379      x->sb_mvcol_part = mi->mv[0].as_mv.col;
1380      x->sb_mvrow_part = mi->mv[0].as_mv.row;
1381    }
1382
1383    y_sad_last = y_sad;
1384    // Pick ref frame for partitioning, bias last frame when y_sad_g and y_sad
1385    // are close if short_circuit_low_temp_var is on.
1386    y_sad_thr = cpi->sf.short_circuit_low_temp_var ? (y_sad * 7) >> 3 : y_sad;
1387    if (y_sad_g < y_sad_thr) {
1388      vp9_setup_pre_planes(xd, 0, yv12_g, mi_row, mi_col,
1389                           &cm->frame_refs[GOLDEN_FRAME - 1].sf);
1390      mi->ref_frame[0] = GOLDEN_FRAME;
1391      mi->mv[0].as_int = 0;
1392      y_sad = y_sad_g;
1393      ref_frame_partition = GOLDEN_FRAME;
1394    } else {
1395      x->pred_mv[LAST_FRAME] = mi->mv[0].as_mv;
1396      ref_frame_partition = LAST_FRAME;
1397    }
1398
1399    set_ref_ptrs(cm, xd, mi->ref_frame[0], mi->ref_frame[1]);
1400    vp9_build_inter_predictors_sb(xd, mi_row, mi_col, BLOCK_64X64);
1401
1402    if (cpi->use_skin_detection)
1403      x->sb_is_skin =
1404          skin_sb_split(cpi, x, low_res, mi_row, mi_col, force_split);
1405
1406    d = xd->plane[0].dst.buf;
1407    dp = xd->plane[0].dst.stride;
1408
1409    // If the y_sad is very small, take 64x64 as partition and exit.
1410    // Don't check on boosted segment for now, as 64x64 is suppressed there.
1411    if (segment_id == CR_SEGMENT_ID_BASE && y_sad < cpi->vbp_threshold_sad) {
1412      const int block_width = num_8x8_blocks_wide_lookup[BLOCK_64X64];
1413      const int block_height = num_8x8_blocks_high_lookup[BLOCK_64X64];
1414      if (mi_col + block_width / 2 < cm->mi_cols &&
1415          mi_row + block_height / 2 < cm->mi_rows) {
1416        set_block_size(cpi, x, xd, mi_row, mi_col, BLOCK_64X64);
1417        x->variance_low[0] = 1;
1418        chroma_check(cpi, x, bsize, y_sad, is_key_frame);
1419        if (cpi->sf.svc_use_lowres_part &&
1420            cpi->svc.spatial_layer_id == cpi->svc.number_spatial_layers - 2)
1421          update_partition_svc(cpi, BLOCK_64X64, mi_row, mi_col);
1422        if (cpi->sf.copy_partition_flag) {
1423          update_prev_partition(cpi, x, segment_id, mi_row, mi_col, sb_offset);
1424        }
1425        return 0;
1426      }
1427    }
1428
1429    // If the y_sad is small enough, copy the partition of the superblock in the
1430    // last frame to current frame only if the last frame is not a keyframe.
1431    // Stop the copy every cpi->max_copied_frame to refresh the partition.
1432    // TODO(jianj) : tune the threshold.
1433    if (cpi->sf.copy_partition_flag && y_sad_last < cpi->vbp_threshold_copy &&
1434        copy_partitioning(cpi, x, xd, mi_row, mi_col, segment_id, sb_offset)) {
1435      chroma_check(cpi, x, bsize, y_sad, is_key_frame);
1436      if (cpi->sf.svc_use_lowres_part &&
1437          cpi->svc.spatial_layer_id == cpi->svc.number_spatial_layers - 2)
1438        update_partition_svc(cpi, BLOCK_64X64, mi_row, mi_col);
1439      return 0;
1440    }
1441  } else {
1442    d = VP9_VAR_OFFS;
1443    dp = 0;
1444#if CONFIG_VP9_HIGHBITDEPTH
1445    if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
1446      switch (xd->bd) {
1447        case 10: d = CONVERT_TO_BYTEPTR(VP9_HIGH_VAR_OFFS_10); break;
1448        case 12: d = CONVERT_TO_BYTEPTR(VP9_HIGH_VAR_OFFS_12); break;
1449        case 8:
1450        default: d = CONVERT_TO_BYTEPTR(VP9_HIGH_VAR_OFFS_8); break;
1451      }
1452    }
1453#endif  // CONFIG_VP9_HIGHBITDEPTH
1454  }
1455
1456  if (low_res && threshold_4x4avg < INT64_MAX)
1457    CHECK_MEM_ERROR(cm, vt2, vpx_calloc(16, sizeof(*vt2)));
1458  // Fill in the entire tree of 8x8 (or 4x4 under some conditions) variances
1459  // for splits.
1460  for (i = 0; i < 4; i++) {
1461    const int x32_idx = ((i & 1) << 5);
1462    const int y32_idx = ((i >> 1) << 5);
1463    const int i2 = i << 2;
1464    force_split[i + 1] = 0;
1465    avg_16x16[i] = 0;
1466    maxvar_16x16[i] = 0;
1467    minvar_16x16[i] = INT_MAX;
1468    for (j = 0; j < 4; j++) {
1469      const int x16_idx = x32_idx + ((j & 1) << 4);
1470      const int y16_idx = y32_idx + ((j >> 1) << 4);
1471      const int split_index = 5 + i2 + j;
1472      v16x16 *vst = &vt.split[i].split[j];
1473      force_split[split_index] = 0;
1474      variance4x4downsample[i2 + j] = 0;
1475      if (!is_key_frame) {
1476        fill_variance_8x8avg(s, sp, d, dp, x16_idx, y16_idx, vst,
1477#if CONFIG_VP9_HIGHBITDEPTH
1478                             xd->cur_buf->flags,
1479#endif
1480                             pixels_wide, pixels_high, is_key_frame);
1481        fill_variance_tree(&vt.split[i].split[j], BLOCK_16X16);
1482        get_variance(&vt.split[i].split[j].part_variances.none);
1483        avg_16x16[i] += vt.split[i].split[j].part_variances.none.variance;
1484        if (vt.split[i].split[j].part_variances.none.variance < minvar_16x16[i])
1485          minvar_16x16[i] = vt.split[i].split[j].part_variances.none.variance;
1486        if (vt.split[i].split[j].part_variances.none.variance > maxvar_16x16[i])
1487          maxvar_16x16[i] = vt.split[i].split[j].part_variances.none.variance;
1488        if (vt.split[i].split[j].part_variances.none.variance > thresholds[2]) {
1489          // 16X16 variance is above threshold for split, so force split to 8x8
1490          // for this 16x16 block (this also forces splits for upper levels).
1491          force_split[split_index] = 1;
1492          force_split[i + 1] = 1;
1493          force_split[0] = 1;
1494        } else if (compute_minmax_variance &&
1495                   vt.split[i].split[j].part_variances.none.variance >
1496                       thresholds[1] &&
1497                   !cyclic_refresh_segment_id_boosted(segment_id)) {
1498          // We have some nominal amount of 16x16 variance (based on average),
1499          // compute the minmax over the 8x8 sub-blocks, and if above threshold,
1500          // force split to 8x8 block for this 16x16 block.
1501          int minmax = compute_minmax_8x8(s, sp, d, dp, x16_idx, y16_idx,
1502#if CONFIG_VP9_HIGHBITDEPTH
1503                                          xd->cur_buf->flags,
1504#endif
1505                                          pixels_wide, pixels_high);
1506          int thresh_minmax = (int)cpi->vbp_threshold_minmax;
1507          if (x->content_state_sb == kVeryHighSad)
1508            thresh_minmax = thresh_minmax << 1;
1509          if (minmax > thresh_minmax) {
1510            force_split[split_index] = 1;
1511            force_split[i + 1] = 1;
1512            force_split[0] = 1;
1513          }
1514        }
1515      }
1516      if (is_key_frame || (low_res &&
1517                           vt.split[i].split[j].part_variances.none.variance >
1518                               threshold_4x4avg)) {
1519        force_split[split_index] = 0;
1520        // Go down to 4x4 down-sampling for variance.
1521        variance4x4downsample[i2 + j] = 1;
1522        for (k = 0; k < 4; k++) {
1523          int x8_idx = x16_idx + ((k & 1) << 3);
1524          int y8_idx = y16_idx + ((k >> 1) << 3);
1525          v8x8 *vst2 = is_key_frame ? &vst->split[k] : &vt2[i2 + j].split[k];
1526          fill_variance_4x4avg(s, sp, d, dp, x8_idx, y8_idx, vst2,
1527#if CONFIG_VP9_HIGHBITDEPTH
1528                               xd->cur_buf->flags,
1529#endif
1530                               pixels_wide, pixels_high, is_key_frame);
1531        }
1532      }
1533    }
1534  }
1535  if (cpi->noise_estimate.enabled)
1536    noise_level = vp9_noise_estimate_extract_level(&cpi->noise_estimate);
1537  // Fill the rest of the variance tree by summing split partition values.
1538  avg_32x32 = 0;
1539  for (i = 0; i < 4; i++) {
1540    const int i2 = i << 2;
1541    for (j = 0; j < 4; j++) {
1542      if (variance4x4downsample[i2 + j] == 1) {
1543        v16x16 *vtemp = (!is_key_frame) ? &vt2[i2 + j] : &vt.split[i].split[j];
1544        for (m = 0; m < 4; m++) fill_variance_tree(&vtemp->split[m], BLOCK_8X8);
1545        fill_variance_tree(vtemp, BLOCK_16X16);
1546        // If variance of this 16x16 block is above the threshold, force block
1547        // to split. This also forces a split on the upper levels.
1548        get_variance(&vtemp->part_variances.none);
1549        if (vtemp->part_variances.none.variance > thresholds[2]) {
1550          force_split[5 + i2 + j] = 1;
1551          force_split[i + 1] = 1;
1552          force_split[0] = 1;
1553        }
1554      }
1555    }
1556    fill_variance_tree(&vt.split[i], BLOCK_32X32);
1557    // If variance of this 32x32 block is above the threshold, or if its above
1558    // (some threshold of) the average variance over the sub-16x16 blocks, then
1559    // force this block to split. This also forces a split on the upper
1560    // (64x64) level.
1561    if (!force_split[i + 1]) {
1562      get_variance(&vt.split[i].part_variances.none);
1563      var_32x32 = vt.split[i].part_variances.none.variance;
1564      max_var_32x32 = VPXMAX(var_32x32, max_var_32x32);
1565      min_var_32x32 = VPXMIN(var_32x32, min_var_32x32);
1566      if (vt.split[i].part_variances.none.variance > thresholds[1] ||
1567          (!is_key_frame &&
1568           vt.split[i].part_variances.none.variance > (thresholds[1] >> 1) &&
1569           vt.split[i].part_variances.none.variance > (avg_16x16[i] >> 1))) {
1570        force_split[i + 1] = 1;
1571        force_split[0] = 1;
1572      } else if (!is_key_frame && noise_level < kLow && cm->height <= 360 &&
1573                 (maxvar_16x16[i] - minvar_16x16[i]) > (thresholds[1] >> 1) &&
1574                 maxvar_16x16[i] > thresholds[1]) {
1575        force_split[i + 1] = 1;
1576        force_split[0] = 1;
1577      }
1578      avg_32x32 += var_32x32;
1579    }
1580  }
1581  if (!force_split[0]) {
1582    fill_variance_tree(&vt, BLOCK_64X64);
1583    get_variance(&vt.part_variances.none);
1584    // If variance of this 64x64 block is above (some threshold of) the average
1585    // variance over the sub-32x32 blocks, then force this block to split.
1586    // Only checking this for noise level >= medium for now.
1587    if (!is_key_frame && noise_level >= kMedium &&
1588        vt.part_variances.none.variance > (9 * avg_32x32) >> 5)
1589      force_split[0] = 1;
1590    // Else if the maximum 32x32 variance minus the miniumum 32x32 variance in
1591    // a 64x64 block is greater than threshold and the maximum 32x32 variance is
1592    // above a miniumum threshold, then force the split of a 64x64 block
1593    // Only check this for low noise.
1594    else if (!is_key_frame && noise_level < kMedium &&
1595             (max_var_32x32 - min_var_32x32) > 3 * (thresholds[0] >> 3) &&
1596             max_var_32x32 > thresholds[0] >> 1)
1597      force_split[0] = 1;
1598  }
1599
1600  // Now go through the entire structure, splitting every block size until
1601  // we get to one that's got a variance lower than our threshold.
1602  if (mi_col + 8 > cm->mi_cols || mi_row + 8 > cm->mi_rows ||
1603      !set_vt_partitioning(cpi, x, xd, &vt, BLOCK_64X64, mi_row, mi_col,
1604                           thresholds[0], BLOCK_16X16, force_split[0])) {
1605    for (i = 0; i < 4; ++i) {
1606      const int x32_idx = ((i & 1) << 2);
1607      const int y32_idx = ((i >> 1) << 2);
1608      const int i2 = i << 2;
1609      if (!set_vt_partitioning(cpi, x, xd, &vt.split[i], BLOCK_32X32,
1610                               (mi_row + y32_idx), (mi_col + x32_idx),
1611                               thresholds[1], BLOCK_16X16,
1612                               force_split[i + 1])) {
1613        for (j = 0; j < 4; ++j) {
1614          const int x16_idx = ((j & 1) << 1);
1615          const int y16_idx = ((j >> 1) << 1);
1616          // For inter frames: if variance4x4downsample[] == 1 for this 16x16
1617          // block, then the variance is based on 4x4 down-sampling, so use vt2
1618          // in set_vt_partioning(), otherwise use vt.
1619          v16x16 *vtemp = (!is_key_frame && variance4x4downsample[i2 + j] == 1)
1620                              ? &vt2[i2 + j]
1621                              : &vt.split[i].split[j];
1622          if (!set_vt_partitioning(
1623                  cpi, x, xd, vtemp, BLOCK_16X16, mi_row + y32_idx + y16_idx,
1624                  mi_col + x32_idx + x16_idx, thresholds[2], cpi->vbp_bsize_min,
1625                  force_split[5 + i2 + j])) {
1626            for (k = 0; k < 4; ++k) {
1627              const int x8_idx = (k & 1);
1628              const int y8_idx = (k >> 1);
1629              if (use_4x4_partition) {
1630                if (!set_vt_partitioning(cpi, x, xd, &vtemp->split[k],
1631                                         BLOCK_8X8,
1632                                         mi_row + y32_idx + y16_idx + y8_idx,
1633                                         mi_col + x32_idx + x16_idx + x8_idx,
1634                                         thresholds[3], BLOCK_8X8, 0)) {
1635                  set_block_size(
1636                      cpi, x, xd, (mi_row + y32_idx + y16_idx + y8_idx),
1637                      (mi_col + x32_idx + x16_idx + x8_idx), BLOCK_4X4);
1638                }
1639              } else {
1640                set_block_size(
1641                    cpi, x, xd, (mi_row + y32_idx + y16_idx + y8_idx),
1642                    (mi_col + x32_idx + x16_idx + x8_idx), BLOCK_8X8);
1643              }
1644            }
1645          }
1646        }
1647      }
1648    }
1649  }
1650
1651  if (cm->frame_type != KEY_FRAME && cpi->sf.copy_partition_flag) {
1652    update_prev_partition(cpi, x, segment_id, mi_row, mi_col, sb_offset);
1653  }
1654
1655  if (cm->frame_type != KEY_FRAME && cpi->sf.svc_use_lowres_part &&
1656      cpi->svc.spatial_layer_id == cpi->svc.number_spatial_layers - 2)
1657    update_partition_svc(cpi, BLOCK_64X64, mi_row, mi_col);
1658
1659  if (cpi->sf.short_circuit_low_temp_var) {
1660    set_low_temp_var_flag(cpi, x, xd, &vt, thresholds, ref_frame_partition,
1661                          mi_col, mi_row);
1662  }
1663
1664  chroma_check(cpi, x, bsize, y_sad, is_key_frame);
1665  if (vt2) vpx_free(vt2);
1666  return 0;
1667}
1668
1669static void update_state(VP9_COMP *cpi, ThreadData *td, PICK_MODE_CONTEXT *ctx,
1670                         int mi_row, int mi_col, BLOCK_SIZE bsize,
1671                         int output_enabled) {
1672  int i, x_idx, y;
1673  VP9_COMMON *const cm = &cpi->common;
1674  RD_COUNTS *const rdc = &td->rd_counts;
1675  MACROBLOCK *const x = &td->mb;
1676  MACROBLOCKD *const xd = &x->e_mbd;
1677  struct macroblock_plane *const p = x->plane;
1678  struct macroblockd_plane *const pd = xd->plane;
1679  MODE_INFO *mi = &ctx->mic;
1680  MODE_INFO *const xdmi = xd->mi[0];
1681  MODE_INFO *mi_addr = xd->mi[0];
1682  const struct segmentation *const seg = &cm->seg;
1683  const int bw = num_8x8_blocks_wide_lookup[mi->sb_type];
1684  const int bh = num_8x8_blocks_high_lookup[mi->sb_type];
1685  const int x_mis = VPXMIN(bw, cm->mi_cols - mi_col);
1686  const int y_mis = VPXMIN(bh, cm->mi_rows - mi_row);
1687  MV_REF *const frame_mvs = cm->cur_frame->mvs + mi_row * cm->mi_cols + mi_col;
1688  int w, h;
1689
1690  const int mis = cm->mi_stride;
1691  const int mi_width = num_8x8_blocks_wide_lookup[bsize];
1692  const int mi_height = num_8x8_blocks_high_lookup[bsize];
1693  int max_plane;
1694
1695  assert(mi->sb_type == bsize);
1696
1697  *mi_addr = *mi;
1698  *x->mbmi_ext = ctx->mbmi_ext;
1699
1700  // If segmentation in use
1701  if (seg->enabled) {
1702    // For in frame complexity AQ copy the segment id from the segment map.
1703    if (cpi->oxcf.aq_mode == COMPLEXITY_AQ) {
1704      const uint8_t *const map =
1705          seg->update_map ? cpi->segmentation_map : cm->last_frame_seg_map;
1706      mi_addr->segment_id = get_segment_id(cm, map, bsize, mi_row, mi_col);
1707    }
1708    // Else for cyclic refresh mode update the segment map, set the segment id
1709    // and then update the quantizer.
1710    if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ) {
1711      vp9_cyclic_refresh_update_segment(cpi, xd->mi[0], mi_row, mi_col, bsize,
1712                                        ctx->rate, ctx->dist, x->skip, p);
1713    }
1714  }
1715
1716  max_plane = is_inter_block(xdmi) ? MAX_MB_PLANE : 1;
1717  for (i = 0; i < max_plane; ++i) {
1718    p[i].coeff = ctx->coeff_pbuf[i][1];
1719    p[i].qcoeff = ctx->qcoeff_pbuf[i][1];
1720    pd[i].dqcoeff = ctx->dqcoeff_pbuf[i][1];
1721    p[i].eobs = ctx->eobs_pbuf[i][1];
1722  }
1723
1724  for (i = max_plane; i < MAX_MB_PLANE; ++i) {
1725    p[i].coeff = ctx->coeff_pbuf[i][2];
1726    p[i].qcoeff = ctx->qcoeff_pbuf[i][2];
1727    pd[i].dqcoeff = ctx->dqcoeff_pbuf[i][2];
1728    p[i].eobs = ctx->eobs_pbuf[i][2];
1729  }
1730
1731  // Restore the coding context of the MB to that that was in place
1732  // when the mode was picked for it
1733  for (y = 0; y < mi_height; y++)
1734    for (x_idx = 0; x_idx < mi_width; x_idx++)
1735      if ((xd->mb_to_right_edge >> (3 + MI_SIZE_LOG2)) + mi_width > x_idx &&
1736          (xd->mb_to_bottom_edge >> (3 + MI_SIZE_LOG2)) + mi_height > y) {
1737        xd->mi[x_idx + y * mis] = mi_addr;
1738      }
1739
1740  if (cpi->oxcf.aq_mode != NO_AQ) vp9_init_plane_quantizers(cpi, x);
1741
1742  if (is_inter_block(xdmi) && xdmi->sb_type < BLOCK_8X8) {
1743    xdmi->mv[0].as_int = mi->bmi[3].as_mv[0].as_int;
1744    xdmi->mv[1].as_int = mi->bmi[3].as_mv[1].as_int;
1745  }
1746
1747  x->skip = ctx->skip;
1748  memcpy(x->zcoeff_blk[xdmi->tx_size], ctx->zcoeff_blk,
1749         sizeof(ctx->zcoeff_blk[0]) * ctx->num_4x4_blk);
1750
1751  if (!output_enabled) return;
1752
1753#if CONFIG_INTERNAL_STATS
1754  if (frame_is_intra_only(cm)) {
1755    static const int kf_mode_index[] = {
1756      THR_DC /*DC_PRED*/,          THR_V_PRED /*V_PRED*/,
1757      THR_H_PRED /*H_PRED*/,       THR_D45_PRED /*D45_PRED*/,
1758      THR_D135_PRED /*D135_PRED*/, THR_D117_PRED /*D117_PRED*/,
1759      THR_D153_PRED /*D153_PRED*/, THR_D207_PRED /*D207_PRED*/,
1760      THR_D63_PRED /*D63_PRED*/,   THR_TM /*TM_PRED*/,
1761    };
1762    ++cpi->mode_chosen_counts[kf_mode_index[xdmi->mode]];
1763  } else {
1764    // Note how often each mode chosen as best
1765    ++cpi->mode_chosen_counts[ctx->best_mode_index];
1766  }
1767#endif
1768  if (!frame_is_intra_only(cm)) {
1769    if (is_inter_block(xdmi)) {
1770      vp9_update_mv_count(td);
1771
1772      if (cm->interp_filter == SWITCHABLE) {
1773        const int ctx = get_pred_context_switchable_interp(xd);
1774        ++td->counts->switchable_interp[ctx][xdmi->interp_filter];
1775      }
1776    }
1777
1778    rdc->comp_pred_diff[SINGLE_REFERENCE] += ctx->single_pred_diff;
1779    rdc->comp_pred_diff[COMPOUND_REFERENCE] += ctx->comp_pred_diff;
1780    rdc->comp_pred_diff[REFERENCE_MODE_SELECT] += ctx->hybrid_pred_diff;
1781
1782    for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i)
1783      rdc->filter_diff[i] += ctx->best_filter_diff[i];
1784  }
1785
1786  for (h = 0; h < y_mis; ++h) {
1787    MV_REF *const frame_mv = frame_mvs + h * cm->mi_cols;
1788    for (w = 0; w < x_mis; ++w) {
1789      MV_REF *const mv = frame_mv + w;
1790      mv->ref_frame[0] = mi->ref_frame[0];
1791      mv->ref_frame[1] = mi->ref_frame[1];
1792      mv->mv[0].as_int = mi->mv[0].as_int;
1793      mv->mv[1].as_int = mi->mv[1].as_int;
1794    }
1795  }
1796}
1797
1798void vp9_setup_src_planes(MACROBLOCK *x, const YV12_BUFFER_CONFIG *src,
1799                          int mi_row, int mi_col) {
1800  uint8_t *const buffers[3] = { src->y_buffer, src->u_buffer, src->v_buffer };
1801  const int strides[3] = { src->y_stride, src->uv_stride, src->uv_stride };
1802  int i;
1803
1804  // Set current frame pointer.
1805  x->e_mbd.cur_buf = src;
1806
1807  for (i = 0; i < MAX_MB_PLANE; i++)
1808    setup_pred_plane(&x->plane[i].src, buffers[i], strides[i], mi_row, mi_col,
1809                     NULL, x->e_mbd.plane[i].subsampling_x,
1810                     x->e_mbd.plane[i].subsampling_y);
1811}
1812
1813static void set_mode_info_seg_skip(MACROBLOCK *x, TX_MODE tx_mode,
1814                                   RD_COST *rd_cost, BLOCK_SIZE bsize) {
1815  MACROBLOCKD *const xd = &x->e_mbd;
1816  MODE_INFO *const mi = xd->mi[0];
1817  INTERP_FILTER filter_ref;
1818
1819  filter_ref = get_pred_context_switchable_interp(xd);
1820  if (filter_ref == SWITCHABLE_FILTERS) filter_ref = EIGHTTAP;
1821
1822  mi->sb_type = bsize;
1823  mi->mode = ZEROMV;
1824  mi->tx_size =
1825      VPXMIN(max_txsize_lookup[bsize], tx_mode_to_biggest_tx_size[tx_mode]);
1826  mi->skip = 1;
1827  mi->uv_mode = DC_PRED;
1828  mi->ref_frame[0] = LAST_FRAME;
1829  mi->ref_frame[1] = NONE;
1830  mi->mv[0].as_int = 0;
1831  mi->interp_filter = filter_ref;
1832
1833  xd->mi[0]->bmi[0].as_mv[0].as_int = 0;
1834  x->skip = 1;
1835
1836  vp9_rd_cost_init(rd_cost);
1837}
1838
1839static int set_segment_rdmult(VP9_COMP *const cpi, MACROBLOCK *const x,
1840                              int8_t segment_id) {
1841  int segment_qindex;
1842  VP9_COMMON *const cm = &cpi->common;
1843  vp9_init_plane_quantizers(cpi, x);
1844  vpx_clear_system_state();
1845  segment_qindex = vp9_get_qindex(&cm->seg, segment_id, cm->base_qindex);
1846  return vp9_compute_rd_mult(cpi, segment_qindex + cm->y_dc_delta_q);
1847}
1848
1849static void rd_pick_sb_modes(VP9_COMP *cpi, TileDataEnc *tile_data,
1850                             MACROBLOCK *const x, int mi_row, int mi_col,
1851                             RD_COST *rd_cost, BLOCK_SIZE bsize,
1852                             PICK_MODE_CONTEXT *ctx, int64_t best_rd) {
1853  VP9_COMMON *const cm = &cpi->common;
1854  TileInfo *const tile_info = &tile_data->tile_info;
1855  MACROBLOCKD *const xd = &x->e_mbd;
1856  MODE_INFO *mi;
1857  struct macroblock_plane *const p = x->plane;
1858  struct macroblockd_plane *const pd = xd->plane;
1859  const AQ_MODE aq_mode = cpi->oxcf.aq_mode;
1860  int i, orig_rdmult;
1861
1862  vpx_clear_system_state();
1863
1864  // Use the lower precision, but faster, 32x32 fdct for mode selection.
1865  x->use_lp32x32fdct = 1;
1866
1867  set_offsets(cpi, tile_info, x, mi_row, mi_col, bsize);
1868  mi = xd->mi[0];
1869  mi->sb_type = bsize;
1870
1871  for (i = 0; i < MAX_MB_PLANE; ++i) {
1872    p[i].coeff = ctx->coeff_pbuf[i][0];
1873    p[i].qcoeff = ctx->qcoeff_pbuf[i][0];
1874    pd[i].dqcoeff = ctx->dqcoeff_pbuf[i][0];
1875    p[i].eobs = ctx->eobs_pbuf[i][0];
1876  }
1877  ctx->is_coded = 0;
1878  ctx->skippable = 0;
1879  ctx->pred_pixel_ready = 0;
1880  x->skip_recode = 0;
1881
1882  // Set to zero to make sure we do not use the previous encoded frame stats
1883  mi->skip = 0;
1884
1885#if CONFIG_VP9_HIGHBITDEPTH
1886  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
1887    x->source_variance = vp9_high_get_sby_perpixel_variance(
1888        cpi, &x->plane[0].src, bsize, xd->bd);
1889  } else {
1890    x->source_variance =
1891        vp9_get_sby_perpixel_variance(cpi, &x->plane[0].src, bsize);
1892  }
1893#else
1894  x->source_variance =
1895      vp9_get_sby_perpixel_variance(cpi, &x->plane[0].src, bsize);
1896#endif  // CONFIG_VP9_HIGHBITDEPTH
1897
1898  // Save rdmult before it might be changed, so it can be restored later.
1899  orig_rdmult = x->rdmult;
1900
1901  if ((cpi->sf.tx_domain_thresh > 0.0) || (cpi->sf.quant_opt_thresh > 0.0)) {
1902    double logvar = vp9_log_block_var(cpi, x, bsize);
1903    // Check block complexity as part of descision on using pixel or transform
1904    // domain distortion in rd tests.
1905    x->block_tx_domain = cpi->sf.allow_txfm_domain_distortion &&
1906                         (logvar >= cpi->sf.tx_domain_thresh);
1907
1908    // Check block complexity as part of descision on using quantized
1909    // coefficient optimisation inside the rd loop.
1910    x->block_qcoeff_opt =
1911        cpi->sf.allow_quant_coeff_opt && (logvar <= cpi->sf.quant_opt_thresh);
1912  } else {
1913    x->block_tx_domain = cpi->sf.allow_txfm_domain_distortion;
1914    x->block_qcoeff_opt = cpi->sf.allow_quant_coeff_opt;
1915  }
1916
1917  if (aq_mode == VARIANCE_AQ) {
1918    const int energy =
1919        bsize <= BLOCK_16X16 ? x->mb_energy : vp9_block_energy(cpi, x, bsize);
1920
1921    if (cm->frame_type == KEY_FRAME || cpi->refresh_alt_ref_frame ||
1922        cpi->force_update_segmentation ||
1923        (cpi->refresh_golden_frame && !cpi->rc.is_src_frame_alt_ref)) {
1924      mi->segment_id = vp9_vaq_segment_id(energy);
1925    } else {
1926      const uint8_t *const map =
1927          cm->seg.update_map ? cpi->segmentation_map : cm->last_frame_seg_map;
1928      mi->segment_id = get_segment_id(cm, map, bsize, mi_row, mi_col);
1929    }
1930    x->rdmult = set_segment_rdmult(cpi, x, mi->segment_id);
1931  } else if (aq_mode == LOOKAHEAD_AQ) {
1932    const uint8_t *const map = cpi->segmentation_map;
1933
1934    // I do not change rdmult here consciously.
1935    mi->segment_id = get_segment_id(cm, map, bsize, mi_row, mi_col);
1936  } else if (aq_mode == EQUATOR360_AQ) {
1937    if (cm->frame_type == KEY_FRAME || cpi->force_update_segmentation) {
1938      mi->segment_id = vp9_360aq_segment_id(mi_row, cm->mi_rows);
1939    } else {
1940      const uint8_t *const map =
1941          cm->seg.update_map ? cpi->segmentation_map : cm->last_frame_seg_map;
1942      mi->segment_id = get_segment_id(cm, map, bsize, mi_row, mi_col);
1943    }
1944    x->rdmult = set_segment_rdmult(cpi, x, mi->segment_id);
1945  } else if (aq_mode == COMPLEXITY_AQ) {
1946    x->rdmult = set_segment_rdmult(cpi, x, mi->segment_id);
1947  } else if (aq_mode == CYCLIC_REFRESH_AQ) {
1948    const uint8_t *const map =
1949        cm->seg.update_map ? cpi->segmentation_map : cm->last_frame_seg_map;
1950    // If segment is boosted, use rdmult for that segment.
1951    if (cyclic_refresh_segment_id_boosted(
1952            get_segment_id(cm, map, bsize, mi_row, mi_col)))
1953      x->rdmult = vp9_cyclic_refresh_get_rdmult(cpi->cyclic_refresh);
1954  }
1955
1956  // Find best coding mode & reconstruct the MB so it is available
1957  // as a predictor for MBs that follow in the SB
1958  if (frame_is_intra_only(cm)) {
1959    vp9_rd_pick_intra_mode_sb(cpi, x, rd_cost, bsize, ctx, best_rd);
1960  } else {
1961    if (bsize >= BLOCK_8X8) {
1962      if (segfeature_active(&cm->seg, mi->segment_id, SEG_LVL_SKIP))
1963        vp9_rd_pick_inter_mode_sb_seg_skip(cpi, tile_data, x, rd_cost, bsize,
1964                                           ctx, best_rd);
1965      else
1966        vp9_rd_pick_inter_mode_sb(cpi, tile_data, x, mi_row, mi_col, rd_cost,
1967                                  bsize, ctx, best_rd);
1968    } else {
1969      vp9_rd_pick_inter_mode_sub8x8(cpi, tile_data, x, mi_row, mi_col, rd_cost,
1970                                    bsize, ctx, best_rd);
1971    }
1972  }
1973
1974  // Examine the resulting rate and for AQ mode 2 make a segment choice.
1975  if ((rd_cost->rate != INT_MAX) && (aq_mode == COMPLEXITY_AQ) &&
1976      (bsize >= BLOCK_16X16) &&
1977      (cm->frame_type == KEY_FRAME || cpi->refresh_alt_ref_frame ||
1978       (cpi->refresh_golden_frame && !cpi->rc.is_src_frame_alt_ref))) {
1979    vp9_caq_select_segment(cpi, x, bsize, mi_row, mi_col, rd_cost->rate);
1980  }
1981
1982  x->rdmult = orig_rdmult;
1983
1984  // TODO(jingning) The rate-distortion optimization flow needs to be
1985  // refactored to provide proper exit/return handle.
1986  if (rd_cost->rate == INT_MAX) rd_cost->rdcost = INT64_MAX;
1987
1988  ctx->rate = rd_cost->rate;
1989  ctx->dist = rd_cost->dist;
1990}
1991
1992static void update_stats(VP9_COMMON *cm, ThreadData *td) {
1993  const MACROBLOCK *x = &td->mb;
1994  const MACROBLOCKD *const xd = &x->e_mbd;
1995  const MODE_INFO *const mi = xd->mi[0];
1996  const MB_MODE_INFO_EXT *const mbmi_ext = x->mbmi_ext;
1997  const BLOCK_SIZE bsize = mi->sb_type;
1998
1999  if (!frame_is_intra_only(cm)) {
2000    FRAME_COUNTS *const counts = td->counts;
2001    const int inter_block = is_inter_block(mi);
2002    const int seg_ref_active =
2003        segfeature_active(&cm->seg, mi->segment_id, SEG_LVL_REF_FRAME);
2004    if (!seg_ref_active) {
2005      counts->intra_inter[get_intra_inter_context(xd)][inter_block]++;
2006      // If the segment reference feature is enabled we have only a single
2007      // reference frame allowed for the segment so exclude it from
2008      // the reference frame counts used to work out probabilities.
2009      if (inter_block) {
2010        const MV_REFERENCE_FRAME ref0 = mi->ref_frame[0];
2011        if (cm->reference_mode == REFERENCE_MODE_SELECT)
2012          counts->comp_inter[vp9_get_reference_mode_context(cm, xd)]
2013                            [has_second_ref(mi)]++;
2014
2015        if (has_second_ref(mi)) {
2016          counts->comp_ref[vp9_get_pred_context_comp_ref_p(cm, xd)]
2017                          [ref0 == GOLDEN_FRAME]++;
2018        } else {
2019          counts->single_ref[vp9_get_pred_context_single_ref_p1(xd)][0]
2020                            [ref0 != LAST_FRAME]++;
2021          if (ref0 != LAST_FRAME)
2022            counts->single_ref[vp9_get_pred_context_single_ref_p2(xd)][1]
2023                              [ref0 != GOLDEN_FRAME]++;
2024        }
2025      }
2026    }
2027    if (inter_block &&
2028        !segfeature_active(&cm->seg, mi->segment_id, SEG_LVL_SKIP)) {
2029      const int mode_ctx = mbmi_ext->mode_context[mi->ref_frame[0]];
2030      if (bsize >= BLOCK_8X8) {
2031        const PREDICTION_MODE mode = mi->mode;
2032        ++counts->inter_mode[mode_ctx][INTER_OFFSET(mode)];
2033      } else {
2034        const int num_4x4_w = num_4x4_blocks_wide_lookup[bsize];
2035        const int num_4x4_h = num_4x4_blocks_high_lookup[bsize];
2036        int idx, idy;
2037        for (idy = 0; idy < 2; idy += num_4x4_h) {
2038          for (idx = 0; idx < 2; idx += num_4x4_w) {
2039            const int j = idy * 2 + idx;
2040            const PREDICTION_MODE b_mode = mi->bmi[j].as_mode;
2041            ++counts->inter_mode[mode_ctx][INTER_OFFSET(b_mode)];
2042          }
2043        }
2044      }
2045    }
2046  }
2047}
2048
2049static void restore_context(MACROBLOCK *const x, int mi_row, int mi_col,
2050                            ENTROPY_CONTEXT a[16 * MAX_MB_PLANE],
2051                            ENTROPY_CONTEXT l[16 * MAX_MB_PLANE],
2052                            PARTITION_CONTEXT sa[8], PARTITION_CONTEXT sl[8],
2053                            BLOCK_SIZE bsize) {
2054  MACROBLOCKD *const xd = &x->e_mbd;
2055  int p;
2056  const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
2057  const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
2058  int mi_width = num_8x8_blocks_wide_lookup[bsize];
2059  int mi_height = num_8x8_blocks_high_lookup[bsize];
2060  for (p = 0; p < MAX_MB_PLANE; p++) {
2061    memcpy(xd->above_context[p] + ((mi_col * 2) >> xd->plane[p].subsampling_x),
2062           a + num_4x4_blocks_wide * p,
2063           (sizeof(ENTROPY_CONTEXT) * num_4x4_blocks_wide) >>
2064               xd->plane[p].subsampling_x);
2065    memcpy(xd->left_context[p] +
2066               ((mi_row & MI_MASK) * 2 >> xd->plane[p].subsampling_y),
2067           l + num_4x4_blocks_high * p,
2068           (sizeof(ENTROPY_CONTEXT) * num_4x4_blocks_high) >>
2069               xd->plane[p].subsampling_y);
2070  }
2071  memcpy(xd->above_seg_context + mi_col, sa,
2072         sizeof(*xd->above_seg_context) * mi_width);
2073  memcpy(xd->left_seg_context + (mi_row & MI_MASK), sl,
2074         sizeof(xd->left_seg_context[0]) * mi_height);
2075}
2076
2077static void save_context(MACROBLOCK *const x, int mi_row, int mi_col,
2078                         ENTROPY_CONTEXT a[16 * MAX_MB_PLANE],
2079                         ENTROPY_CONTEXT l[16 * MAX_MB_PLANE],
2080                         PARTITION_CONTEXT sa[8], PARTITION_CONTEXT sl[8],
2081                         BLOCK_SIZE bsize) {
2082  const MACROBLOCKD *const xd = &x->e_mbd;
2083  int p;
2084  const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
2085  const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
2086  int mi_width = num_8x8_blocks_wide_lookup[bsize];
2087  int mi_height = num_8x8_blocks_high_lookup[bsize];
2088
2089  // buffer the above/left context information of the block in search.
2090  for (p = 0; p < MAX_MB_PLANE; ++p) {
2091    memcpy(a + num_4x4_blocks_wide * p,
2092           xd->above_context[p] + (mi_col * 2 >> xd->plane[p].subsampling_x),
2093           (sizeof(ENTROPY_CONTEXT) * num_4x4_blocks_wide) >>
2094               xd->plane[p].subsampling_x);
2095    memcpy(l + num_4x4_blocks_high * p,
2096           xd->left_context[p] +
2097               ((mi_row & MI_MASK) * 2 >> xd->plane[p].subsampling_y),
2098           (sizeof(ENTROPY_CONTEXT) * num_4x4_blocks_high) >>
2099               xd->plane[p].subsampling_y);
2100  }
2101  memcpy(sa, xd->above_seg_context + mi_col,
2102         sizeof(*xd->above_seg_context) * mi_width);
2103  memcpy(sl, xd->left_seg_context + (mi_row & MI_MASK),
2104         sizeof(xd->left_seg_context[0]) * mi_height);
2105}
2106
2107static void encode_b(VP9_COMP *cpi, const TileInfo *const tile, ThreadData *td,
2108                     TOKENEXTRA **tp, int mi_row, int mi_col,
2109                     int output_enabled, BLOCK_SIZE bsize,
2110                     PICK_MODE_CONTEXT *ctx) {
2111  MACROBLOCK *const x = &td->mb;
2112  set_offsets(cpi, tile, x, mi_row, mi_col, bsize);
2113  update_state(cpi, td, ctx, mi_row, mi_col, bsize, output_enabled);
2114  encode_superblock(cpi, td, tp, output_enabled, mi_row, mi_col, bsize, ctx);
2115
2116  if (output_enabled) {
2117    update_stats(&cpi->common, td);
2118
2119    (*tp)->token = EOSB_TOKEN;
2120    (*tp)++;
2121  }
2122}
2123
2124static void encode_sb(VP9_COMP *cpi, ThreadData *td, const TileInfo *const tile,
2125                      TOKENEXTRA **tp, int mi_row, int mi_col,
2126                      int output_enabled, BLOCK_SIZE bsize, PC_TREE *pc_tree) {
2127  VP9_COMMON *const cm = &cpi->common;
2128  MACROBLOCK *const x = &td->mb;
2129  MACROBLOCKD *const xd = &x->e_mbd;
2130
2131  const int bsl = b_width_log2_lookup[bsize], hbs = (1 << bsl) / 4;
2132  int ctx;
2133  PARTITION_TYPE partition;
2134  BLOCK_SIZE subsize = bsize;
2135
2136  if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols) return;
2137
2138  if (bsize >= BLOCK_8X8) {
2139    ctx = partition_plane_context(xd, mi_row, mi_col, bsize);
2140    subsize = get_subsize(bsize, pc_tree->partitioning);
2141  } else {
2142    ctx = 0;
2143    subsize = BLOCK_4X4;
2144  }
2145
2146  partition = partition_lookup[bsl][subsize];
2147  if (output_enabled && bsize != BLOCK_4X4)
2148    td->counts->partition[ctx][partition]++;
2149
2150  switch (partition) {
2151    case PARTITION_NONE:
2152      encode_b(cpi, tile, td, tp, mi_row, mi_col, output_enabled, subsize,
2153               &pc_tree->none);
2154      break;
2155    case PARTITION_VERT:
2156      encode_b(cpi, tile, td, tp, mi_row, mi_col, output_enabled, subsize,
2157               &pc_tree->vertical[0]);
2158      if (mi_col + hbs < cm->mi_cols && bsize > BLOCK_8X8) {
2159        encode_b(cpi, tile, td, tp, mi_row, mi_col + hbs, output_enabled,
2160                 subsize, &pc_tree->vertical[1]);
2161      }
2162      break;
2163    case PARTITION_HORZ:
2164      encode_b(cpi, tile, td, tp, mi_row, mi_col, output_enabled, subsize,
2165               &pc_tree->horizontal[0]);
2166      if (mi_row + hbs < cm->mi_rows && bsize > BLOCK_8X8) {
2167        encode_b(cpi, tile, td, tp, mi_row + hbs, mi_col, output_enabled,
2168                 subsize, &pc_tree->horizontal[1]);
2169      }
2170      break;
2171    case PARTITION_SPLIT:
2172      if (bsize == BLOCK_8X8) {
2173        encode_b(cpi, tile, td, tp, mi_row, mi_col, output_enabled, subsize,
2174                 pc_tree->leaf_split[0]);
2175      } else {
2176        encode_sb(cpi, td, tile, tp, mi_row, mi_col, output_enabled, subsize,
2177                  pc_tree->split[0]);
2178        encode_sb(cpi, td, tile, tp, mi_row, mi_col + hbs, output_enabled,
2179                  subsize, pc_tree->split[1]);
2180        encode_sb(cpi, td, tile, tp, mi_row + hbs, mi_col, output_enabled,
2181                  subsize, pc_tree->split[2]);
2182        encode_sb(cpi, td, tile, tp, mi_row + hbs, mi_col + hbs, output_enabled,
2183                  subsize, pc_tree->split[3]);
2184      }
2185      break;
2186    default: assert(0 && "Invalid partition type."); break;
2187  }
2188
2189  if (partition != PARTITION_SPLIT || bsize == BLOCK_8X8)
2190    update_partition_context(xd, mi_row, mi_col, subsize, bsize);
2191}
2192
2193// Check to see if the given partition size is allowed for a specified number
2194// of 8x8 block rows and columns remaining in the image.
2195// If not then return the largest allowed partition size
2196static BLOCK_SIZE find_partition_size(BLOCK_SIZE bsize, int rows_left,
2197                                      int cols_left, int *bh, int *bw) {
2198  if (rows_left <= 0 || cols_left <= 0) {
2199    return VPXMIN(bsize, BLOCK_8X8);
2200  } else {
2201    for (; bsize > 0; bsize -= 3) {
2202      *bh = num_8x8_blocks_high_lookup[bsize];
2203      *bw = num_8x8_blocks_wide_lookup[bsize];
2204      if ((*bh <= rows_left) && (*bw <= cols_left)) {
2205        break;
2206      }
2207    }
2208  }
2209  return bsize;
2210}
2211
2212static void set_partial_b64x64_partition(MODE_INFO *mi, int mis, int bh_in,
2213                                         int bw_in, int row8x8_remaining,
2214                                         int col8x8_remaining, BLOCK_SIZE bsize,
2215                                         MODE_INFO **mi_8x8) {
2216  int bh = bh_in;
2217  int r, c;
2218  for (r = 0; r < MI_BLOCK_SIZE; r += bh) {
2219    int bw = bw_in;
2220    for (c = 0; c < MI_BLOCK_SIZE; c += bw) {
2221      const int index = r * mis + c;
2222      mi_8x8[index] = mi + index;
2223      mi_8x8[index]->sb_type = find_partition_size(
2224          bsize, row8x8_remaining - r, col8x8_remaining - c, &bh, &bw);
2225    }
2226  }
2227}
2228
2229// This function attempts to set all mode info entries in a given SB64
2230// to the same block partition size.
2231// However, at the bottom and right borders of the image the requested size
2232// may not be allowed in which case this code attempts to choose the largest
2233// allowable partition.
2234static void set_fixed_partitioning(VP9_COMP *cpi, const TileInfo *const tile,
2235                                   MODE_INFO **mi_8x8, int mi_row, int mi_col,
2236                                   BLOCK_SIZE bsize) {
2237  VP9_COMMON *const cm = &cpi->common;
2238  const int mis = cm->mi_stride;
2239  const int row8x8_remaining = tile->mi_row_end - mi_row;
2240  const int col8x8_remaining = tile->mi_col_end - mi_col;
2241  int block_row, block_col;
2242  MODE_INFO *mi_upper_left = cm->mi + mi_row * mis + mi_col;
2243  int bh = num_8x8_blocks_high_lookup[bsize];
2244  int bw = num_8x8_blocks_wide_lookup[bsize];
2245
2246  assert((row8x8_remaining > 0) && (col8x8_remaining > 0));
2247
2248  // Apply the requested partition size to the SB64 if it is all "in image"
2249  if ((col8x8_remaining >= MI_BLOCK_SIZE) &&
2250      (row8x8_remaining >= MI_BLOCK_SIZE)) {
2251    for (block_row = 0; block_row < MI_BLOCK_SIZE; block_row += bh) {
2252      for (block_col = 0; block_col < MI_BLOCK_SIZE; block_col += bw) {
2253        int index = block_row * mis + block_col;
2254        mi_8x8[index] = mi_upper_left + index;
2255        mi_8x8[index]->sb_type = bsize;
2256      }
2257    }
2258  } else {
2259    // Else this is a partial SB64.
2260    set_partial_b64x64_partition(mi_upper_left, mis, bh, bw, row8x8_remaining,
2261                                 col8x8_remaining, bsize, mi_8x8);
2262  }
2263}
2264
2265static const struct {
2266  int row;
2267  int col;
2268} coord_lookup[16] = {
2269  // 32x32 index = 0
2270  { 0, 0 },
2271  { 0, 2 },
2272  { 2, 0 },
2273  { 2, 2 },
2274  // 32x32 index = 1
2275  { 0, 4 },
2276  { 0, 6 },
2277  { 2, 4 },
2278  { 2, 6 },
2279  // 32x32 index = 2
2280  { 4, 0 },
2281  { 4, 2 },
2282  { 6, 0 },
2283  { 6, 2 },
2284  // 32x32 index = 3
2285  { 4, 4 },
2286  { 4, 6 },
2287  { 6, 4 },
2288  { 6, 6 },
2289};
2290
2291static void set_source_var_based_partition(VP9_COMP *cpi,
2292                                           const TileInfo *const tile,
2293                                           MACROBLOCK *const x,
2294                                           MODE_INFO **mi_8x8, int mi_row,
2295                                           int mi_col) {
2296  VP9_COMMON *const cm = &cpi->common;
2297  const int mis = cm->mi_stride;
2298  const int row8x8_remaining = tile->mi_row_end - mi_row;
2299  const int col8x8_remaining = tile->mi_col_end - mi_col;
2300  MODE_INFO *mi_upper_left = cm->mi + mi_row * mis + mi_col;
2301
2302  vp9_setup_src_planes(x, cpi->Source, mi_row, mi_col);
2303
2304  assert((row8x8_remaining > 0) && (col8x8_remaining > 0));
2305
2306  // In-image SB64
2307  if ((col8x8_remaining >= MI_BLOCK_SIZE) &&
2308      (row8x8_remaining >= MI_BLOCK_SIZE)) {
2309    int i, j;
2310    int index;
2311    diff d32[4];
2312    const int offset = (mi_row >> 1) * cm->mb_cols + (mi_col >> 1);
2313    int is_larger_better = 0;
2314    int use32x32 = 0;
2315    unsigned int thr = cpi->source_var_thresh;
2316
2317    memset(d32, 0, 4 * sizeof(diff));
2318
2319    for (i = 0; i < 4; i++) {
2320      diff *d16[4];
2321
2322      for (j = 0; j < 4; j++) {
2323        int b_mi_row = coord_lookup[i * 4 + j].row;
2324        int b_mi_col = coord_lookup[i * 4 + j].col;
2325        int boffset = b_mi_row / 2 * cm->mb_cols + b_mi_col / 2;
2326
2327        d16[j] = cpi->source_diff_var + offset + boffset;
2328
2329        index = b_mi_row * mis + b_mi_col;
2330        mi_8x8[index] = mi_upper_left + index;
2331        mi_8x8[index]->sb_type = BLOCK_16X16;
2332
2333        // TODO(yunqingwang): If d16[j].var is very large, use 8x8 partition
2334        // size to further improve quality.
2335      }
2336
2337      is_larger_better = (d16[0]->var < thr) && (d16[1]->var < thr) &&
2338                         (d16[2]->var < thr) && (d16[3]->var < thr);
2339
2340      // Use 32x32 partition
2341      if (is_larger_better) {
2342        use32x32 += 1;
2343
2344        for (j = 0; j < 4; j++) {
2345          d32[i].sse += d16[j]->sse;
2346          d32[i].sum += d16[j]->sum;
2347        }
2348
2349        d32[i].var =
2350            (unsigned int)(d32[i].sse -
2351                           (unsigned int)(((int64_t)d32[i].sum * d32[i].sum) >>
2352                                          10));
2353
2354        index = coord_lookup[i * 4].row * mis + coord_lookup[i * 4].col;
2355        mi_8x8[index] = mi_upper_left + index;
2356        mi_8x8[index]->sb_type = BLOCK_32X32;
2357      }
2358    }
2359
2360    if (use32x32 == 4) {
2361      thr <<= 1;
2362      is_larger_better = (d32[0].var < thr) && (d32[1].var < thr) &&
2363                         (d32[2].var < thr) && (d32[3].var < thr);
2364
2365      // Use 64x64 partition
2366      if (is_larger_better) {
2367        mi_8x8[0] = mi_upper_left;
2368        mi_8x8[0]->sb_type = BLOCK_64X64;
2369      }
2370    }
2371  } else {  // partial in-image SB64
2372    int bh = num_8x8_blocks_high_lookup[BLOCK_16X16];
2373    int bw = num_8x8_blocks_wide_lookup[BLOCK_16X16];
2374    set_partial_b64x64_partition(mi_upper_left, mis, bh, bw, row8x8_remaining,
2375                                 col8x8_remaining, BLOCK_16X16, mi_8x8);
2376  }
2377}
2378
2379static void update_state_rt(VP9_COMP *cpi, ThreadData *td,
2380                            PICK_MODE_CONTEXT *ctx, int mi_row, int mi_col,
2381                            int bsize) {
2382  VP9_COMMON *const cm = &cpi->common;
2383  MACROBLOCK *const x = &td->mb;
2384  MACROBLOCKD *const xd = &x->e_mbd;
2385  MODE_INFO *const mi = xd->mi[0];
2386  struct macroblock_plane *const p = x->plane;
2387  const struct segmentation *const seg = &cm->seg;
2388  const int bw = num_8x8_blocks_wide_lookup[mi->sb_type];
2389  const int bh = num_8x8_blocks_high_lookup[mi->sb_type];
2390  const int x_mis = VPXMIN(bw, cm->mi_cols - mi_col);
2391  const int y_mis = VPXMIN(bh, cm->mi_rows - mi_row);
2392
2393  *(xd->mi[0]) = ctx->mic;
2394  *(x->mbmi_ext) = ctx->mbmi_ext;
2395
2396  if (seg->enabled && cpi->oxcf.aq_mode != NO_AQ) {
2397    // For in frame complexity AQ or variance AQ, copy segment_id from
2398    // segmentation_map.
2399    if (cpi->oxcf.aq_mode != CYCLIC_REFRESH_AQ) {
2400      const uint8_t *const map =
2401          seg->update_map ? cpi->segmentation_map : cm->last_frame_seg_map;
2402      mi->segment_id = get_segment_id(cm, map, bsize, mi_row, mi_col);
2403    } else {
2404      // Setting segmentation map for cyclic_refresh.
2405      vp9_cyclic_refresh_update_segment(cpi, mi, mi_row, mi_col, bsize,
2406                                        ctx->rate, ctx->dist, x->skip, p);
2407    }
2408    vp9_init_plane_quantizers(cpi, x);
2409  }
2410
2411  if (is_inter_block(mi)) {
2412    vp9_update_mv_count(td);
2413    if (cm->interp_filter == SWITCHABLE) {
2414      const int pred_ctx = get_pred_context_switchable_interp(xd);
2415      ++td->counts->switchable_interp[pred_ctx][mi->interp_filter];
2416    }
2417
2418    if (mi->sb_type < BLOCK_8X8) {
2419      mi->mv[0].as_int = mi->bmi[3].as_mv[0].as_int;
2420      mi->mv[1].as_int = mi->bmi[3].as_mv[1].as_int;
2421    }
2422  }
2423
2424  if (cm->use_prev_frame_mvs || !cm->error_resilient_mode ||
2425      (cpi->svc.use_base_mv && cpi->svc.number_spatial_layers > 1 &&
2426       cpi->svc.spatial_layer_id != cpi->svc.number_spatial_layers - 1)) {
2427    MV_REF *const frame_mvs =
2428        cm->cur_frame->mvs + mi_row * cm->mi_cols + mi_col;
2429    int w, h;
2430
2431    for (h = 0; h < y_mis; ++h) {
2432      MV_REF *const frame_mv = frame_mvs + h * cm->mi_cols;
2433      for (w = 0; w < x_mis; ++w) {
2434        MV_REF *const mv = frame_mv + w;
2435        mv->ref_frame[0] = mi->ref_frame[0];
2436        mv->ref_frame[1] = mi->ref_frame[1];
2437        mv->mv[0].as_int = mi->mv[0].as_int;
2438        mv->mv[1].as_int = mi->mv[1].as_int;
2439      }
2440    }
2441  }
2442
2443  x->skip = ctx->skip;
2444  x->skip_txfm[0] = mi->segment_id ? 0 : ctx->skip_txfm[0];
2445}
2446
2447static void encode_b_rt(VP9_COMP *cpi, ThreadData *td,
2448                        const TileInfo *const tile, TOKENEXTRA **tp, int mi_row,
2449                        int mi_col, int output_enabled, BLOCK_SIZE bsize,
2450                        PICK_MODE_CONTEXT *ctx) {
2451  MACROBLOCK *const x = &td->mb;
2452  set_offsets(cpi, tile, x, mi_row, mi_col, bsize);
2453  update_state_rt(cpi, td, ctx, mi_row, mi_col, bsize);
2454
2455  encode_superblock(cpi, td, tp, output_enabled, mi_row, mi_col, bsize, ctx);
2456  update_stats(&cpi->common, td);
2457
2458  (*tp)->token = EOSB_TOKEN;
2459  (*tp)++;
2460}
2461
2462static void encode_sb_rt(VP9_COMP *cpi, ThreadData *td,
2463                         const TileInfo *const tile, TOKENEXTRA **tp,
2464                         int mi_row, int mi_col, int output_enabled,
2465                         BLOCK_SIZE bsize, PC_TREE *pc_tree) {
2466  VP9_COMMON *const cm = &cpi->common;
2467  MACROBLOCK *const x = &td->mb;
2468  MACROBLOCKD *const xd = &x->e_mbd;
2469
2470  const int bsl = b_width_log2_lookup[bsize], hbs = (1 << bsl) / 4;
2471  int ctx;
2472  PARTITION_TYPE partition;
2473  BLOCK_SIZE subsize;
2474
2475  if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols) return;
2476
2477  if (bsize >= BLOCK_8X8) {
2478    const int idx_str = xd->mi_stride * mi_row + mi_col;
2479    MODE_INFO **mi_8x8 = cm->mi_grid_visible + idx_str;
2480    ctx = partition_plane_context(xd, mi_row, mi_col, bsize);
2481    subsize = mi_8x8[0]->sb_type;
2482  } else {
2483    ctx = 0;
2484    subsize = BLOCK_4X4;
2485  }
2486
2487  partition = partition_lookup[bsl][subsize];
2488  if (output_enabled && bsize != BLOCK_4X4)
2489    td->counts->partition[ctx][partition]++;
2490
2491  switch (partition) {
2492    case PARTITION_NONE:
2493      encode_b_rt(cpi, td, tile, tp, mi_row, mi_col, output_enabled, subsize,
2494                  &pc_tree->none);
2495      break;
2496    case PARTITION_VERT:
2497      encode_b_rt(cpi, td, tile, tp, mi_row, mi_col, output_enabled, subsize,
2498                  &pc_tree->vertical[0]);
2499      if (mi_col + hbs < cm->mi_cols && bsize > BLOCK_8X8) {
2500        encode_b_rt(cpi, td, tile, tp, mi_row, mi_col + hbs, output_enabled,
2501                    subsize, &pc_tree->vertical[1]);
2502      }
2503      break;
2504    case PARTITION_HORZ:
2505      encode_b_rt(cpi, td, tile, tp, mi_row, mi_col, output_enabled, subsize,
2506                  &pc_tree->horizontal[0]);
2507      if (mi_row + hbs < cm->mi_rows && bsize > BLOCK_8X8) {
2508        encode_b_rt(cpi, td, tile, tp, mi_row + hbs, mi_col, output_enabled,
2509                    subsize, &pc_tree->horizontal[1]);
2510      }
2511      break;
2512    case PARTITION_SPLIT:
2513      subsize = get_subsize(bsize, PARTITION_SPLIT);
2514      encode_sb_rt(cpi, td, tile, tp, mi_row, mi_col, output_enabled, subsize,
2515                   pc_tree->split[0]);
2516      encode_sb_rt(cpi, td, tile, tp, mi_row, mi_col + hbs, output_enabled,
2517                   subsize, pc_tree->split[1]);
2518      encode_sb_rt(cpi, td, tile, tp, mi_row + hbs, mi_col, output_enabled,
2519                   subsize, pc_tree->split[2]);
2520      encode_sb_rt(cpi, td, tile, tp, mi_row + hbs, mi_col + hbs,
2521                   output_enabled, subsize, pc_tree->split[3]);
2522      break;
2523    default: assert(0 && "Invalid partition type."); break;
2524  }
2525
2526  if (partition != PARTITION_SPLIT || bsize == BLOCK_8X8)
2527    update_partition_context(xd, mi_row, mi_col, subsize, bsize);
2528}
2529
2530static void rd_use_partition(VP9_COMP *cpi, ThreadData *td,
2531                             TileDataEnc *tile_data, MODE_INFO **mi_8x8,
2532                             TOKENEXTRA **tp, int mi_row, int mi_col,
2533                             BLOCK_SIZE bsize, int *rate, int64_t *dist,
2534                             int do_recon, PC_TREE *pc_tree) {
2535  VP9_COMMON *const cm = &cpi->common;
2536  TileInfo *const tile_info = &tile_data->tile_info;
2537  MACROBLOCK *const x = &td->mb;
2538  MACROBLOCKD *const xd = &x->e_mbd;
2539  const int mis = cm->mi_stride;
2540  const int bsl = b_width_log2_lookup[bsize];
2541  const int mi_step = num_4x4_blocks_wide_lookup[bsize] / 2;
2542  const int bss = (1 << bsl) / 4;
2543  int i, pl;
2544  PARTITION_TYPE partition = PARTITION_NONE;
2545  BLOCK_SIZE subsize;
2546  ENTROPY_CONTEXT l[16 * MAX_MB_PLANE], a[16 * MAX_MB_PLANE];
2547  PARTITION_CONTEXT sl[8], sa[8];
2548  RD_COST last_part_rdc, none_rdc, chosen_rdc;
2549  BLOCK_SIZE sub_subsize = BLOCK_4X4;
2550  int splits_below = 0;
2551  BLOCK_SIZE bs_type = mi_8x8[0]->sb_type;
2552  int do_partition_search = 1;
2553  PICK_MODE_CONTEXT *ctx = &pc_tree->none;
2554
2555  if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols) return;
2556
2557  assert(num_4x4_blocks_wide_lookup[bsize] ==
2558         num_4x4_blocks_high_lookup[bsize]);
2559
2560  vp9_rd_cost_reset(&last_part_rdc);
2561  vp9_rd_cost_reset(&none_rdc);
2562  vp9_rd_cost_reset(&chosen_rdc);
2563
2564  partition = partition_lookup[bsl][bs_type];
2565  subsize = get_subsize(bsize, partition);
2566
2567  pc_tree->partitioning = partition;
2568  save_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
2569
2570  if (bsize == BLOCK_16X16 && cpi->oxcf.aq_mode != NO_AQ) {
2571    set_offsets(cpi, tile_info, x, mi_row, mi_col, bsize);
2572    x->mb_energy = vp9_block_energy(cpi, x, bsize);
2573  }
2574
2575  if (do_partition_search &&
2576      cpi->sf.partition_search_type == SEARCH_PARTITION &&
2577      cpi->sf.adjust_partitioning_from_last_frame) {
2578    // Check if any of the sub blocks are further split.
2579    if (partition == PARTITION_SPLIT && subsize > BLOCK_8X8) {
2580      sub_subsize = get_subsize(subsize, PARTITION_SPLIT);
2581      splits_below = 1;
2582      for (i = 0; i < 4; i++) {
2583        int jj = i >> 1, ii = i & 0x01;
2584        MODE_INFO *this_mi = mi_8x8[jj * bss * mis + ii * bss];
2585        if (this_mi && this_mi->sb_type >= sub_subsize) {
2586          splits_below = 0;
2587        }
2588      }
2589    }
2590
2591    // If partition is not none try none unless each of the 4 splits are split
2592    // even further..
2593    if (partition != PARTITION_NONE && !splits_below &&
2594        mi_row + (mi_step >> 1) < cm->mi_rows &&
2595        mi_col + (mi_step >> 1) < cm->mi_cols) {
2596      pc_tree->partitioning = PARTITION_NONE;
2597      rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &none_rdc, bsize, ctx,
2598                       INT64_MAX);
2599
2600      pl = partition_plane_context(xd, mi_row, mi_col, bsize);
2601
2602      if (none_rdc.rate < INT_MAX) {
2603        none_rdc.rate += cpi->partition_cost[pl][PARTITION_NONE];
2604        none_rdc.rdcost =
2605            RDCOST(x->rdmult, x->rddiv, none_rdc.rate, none_rdc.dist);
2606      }
2607
2608      restore_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
2609      mi_8x8[0]->sb_type = bs_type;
2610      pc_tree->partitioning = partition;
2611    }
2612  }
2613
2614  switch (partition) {
2615    case PARTITION_NONE:
2616      rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &last_part_rdc, bsize,
2617                       ctx, INT64_MAX);
2618      break;
2619    case PARTITION_HORZ:
2620      rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &last_part_rdc,
2621                       subsize, &pc_tree->horizontal[0], INT64_MAX);
2622      if (last_part_rdc.rate != INT_MAX && bsize >= BLOCK_8X8 &&
2623          mi_row + (mi_step >> 1) < cm->mi_rows) {
2624        RD_COST tmp_rdc;
2625        PICK_MODE_CONTEXT *ctx = &pc_tree->horizontal[0];
2626        vp9_rd_cost_init(&tmp_rdc);
2627        update_state(cpi, td, ctx, mi_row, mi_col, subsize, 0);
2628        encode_superblock(cpi, td, tp, 0, mi_row, mi_col, subsize, ctx);
2629        rd_pick_sb_modes(cpi, tile_data, x, mi_row + (mi_step >> 1), mi_col,
2630                         &tmp_rdc, subsize, &pc_tree->horizontal[1], INT64_MAX);
2631        if (tmp_rdc.rate == INT_MAX || tmp_rdc.dist == INT64_MAX) {
2632          vp9_rd_cost_reset(&last_part_rdc);
2633          break;
2634        }
2635        last_part_rdc.rate += tmp_rdc.rate;
2636        last_part_rdc.dist += tmp_rdc.dist;
2637        last_part_rdc.rdcost += tmp_rdc.rdcost;
2638      }
2639      break;
2640    case PARTITION_VERT:
2641      rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &last_part_rdc,
2642                       subsize, &pc_tree->vertical[0], INT64_MAX);
2643      if (last_part_rdc.rate != INT_MAX && bsize >= BLOCK_8X8 &&
2644          mi_col + (mi_step >> 1) < cm->mi_cols) {
2645        RD_COST tmp_rdc;
2646        PICK_MODE_CONTEXT *ctx = &pc_tree->vertical[0];
2647        vp9_rd_cost_init(&tmp_rdc);
2648        update_state(cpi, td, ctx, mi_row, mi_col, subsize, 0);
2649        encode_superblock(cpi, td, tp, 0, mi_row, mi_col, subsize, ctx);
2650        rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col + (mi_step >> 1),
2651                         &tmp_rdc, subsize,
2652                         &pc_tree->vertical[bsize > BLOCK_8X8], INT64_MAX);
2653        if (tmp_rdc.rate == INT_MAX || tmp_rdc.dist == INT64_MAX) {
2654          vp9_rd_cost_reset(&last_part_rdc);
2655          break;
2656        }
2657        last_part_rdc.rate += tmp_rdc.rate;
2658        last_part_rdc.dist += tmp_rdc.dist;
2659        last_part_rdc.rdcost += tmp_rdc.rdcost;
2660      }
2661      break;
2662    case PARTITION_SPLIT:
2663      if (bsize == BLOCK_8X8) {
2664        rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &last_part_rdc,
2665                         subsize, pc_tree->leaf_split[0], INT64_MAX);
2666        break;
2667      }
2668      last_part_rdc.rate = 0;
2669      last_part_rdc.dist = 0;
2670      last_part_rdc.rdcost = 0;
2671      for (i = 0; i < 4; i++) {
2672        int x_idx = (i & 1) * (mi_step >> 1);
2673        int y_idx = (i >> 1) * (mi_step >> 1);
2674        int jj = i >> 1, ii = i & 0x01;
2675        RD_COST tmp_rdc;
2676        if ((mi_row + y_idx >= cm->mi_rows) || (mi_col + x_idx >= cm->mi_cols))
2677          continue;
2678
2679        vp9_rd_cost_init(&tmp_rdc);
2680        rd_use_partition(cpi, td, tile_data, mi_8x8 + jj * bss * mis + ii * bss,
2681                         tp, mi_row + y_idx, mi_col + x_idx, subsize,
2682                         &tmp_rdc.rate, &tmp_rdc.dist, i != 3,
2683                         pc_tree->split[i]);
2684        if (tmp_rdc.rate == INT_MAX || tmp_rdc.dist == INT64_MAX) {
2685          vp9_rd_cost_reset(&last_part_rdc);
2686          break;
2687        }
2688        last_part_rdc.rate += tmp_rdc.rate;
2689        last_part_rdc.dist += tmp_rdc.dist;
2690      }
2691      break;
2692    default: assert(0); break;
2693  }
2694
2695  pl = partition_plane_context(xd, mi_row, mi_col, bsize);
2696  if (last_part_rdc.rate < INT_MAX) {
2697    last_part_rdc.rate += cpi->partition_cost[pl][partition];
2698    last_part_rdc.rdcost =
2699        RDCOST(x->rdmult, x->rddiv, last_part_rdc.rate, last_part_rdc.dist);
2700  }
2701
2702  if (do_partition_search && cpi->sf.adjust_partitioning_from_last_frame &&
2703      cpi->sf.partition_search_type == SEARCH_PARTITION &&
2704      partition != PARTITION_SPLIT && bsize > BLOCK_8X8 &&
2705      (mi_row + mi_step < cm->mi_rows ||
2706       mi_row + (mi_step >> 1) == cm->mi_rows) &&
2707      (mi_col + mi_step < cm->mi_cols ||
2708       mi_col + (mi_step >> 1) == cm->mi_cols)) {
2709    BLOCK_SIZE split_subsize = get_subsize(bsize, PARTITION_SPLIT);
2710    chosen_rdc.rate = 0;
2711    chosen_rdc.dist = 0;
2712    restore_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
2713    pc_tree->partitioning = PARTITION_SPLIT;
2714
2715    // Split partition.
2716    for (i = 0; i < 4; i++) {
2717      int x_idx = (i & 1) * (mi_step >> 1);
2718      int y_idx = (i >> 1) * (mi_step >> 1);
2719      RD_COST tmp_rdc;
2720      ENTROPY_CONTEXT l[16 * MAX_MB_PLANE], a[16 * MAX_MB_PLANE];
2721      PARTITION_CONTEXT sl[8], sa[8];
2722
2723      if ((mi_row + y_idx >= cm->mi_rows) || (mi_col + x_idx >= cm->mi_cols))
2724        continue;
2725
2726      save_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
2727      pc_tree->split[i]->partitioning = PARTITION_NONE;
2728      rd_pick_sb_modes(cpi, tile_data, x, mi_row + y_idx, mi_col + x_idx,
2729                       &tmp_rdc, split_subsize, &pc_tree->split[i]->none,
2730                       INT64_MAX);
2731
2732      restore_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
2733
2734      if (tmp_rdc.rate == INT_MAX || tmp_rdc.dist == INT64_MAX) {
2735        vp9_rd_cost_reset(&chosen_rdc);
2736        break;
2737      }
2738
2739      chosen_rdc.rate += tmp_rdc.rate;
2740      chosen_rdc.dist += tmp_rdc.dist;
2741
2742      if (i != 3)
2743        encode_sb(cpi, td, tile_info, tp, mi_row + y_idx, mi_col + x_idx, 0,
2744                  split_subsize, pc_tree->split[i]);
2745
2746      pl = partition_plane_context(xd, mi_row + y_idx, mi_col + x_idx,
2747                                   split_subsize);
2748      chosen_rdc.rate += cpi->partition_cost[pl][PARTITION_NONE];
2749    }
2750    pl = partition_plane_context(xd, mi_row, mi_col, bsize);
2751    if (chosen_rdc.rate < INT_MAX) {
2752      chosen_rdc.rate += cpi->partition_cost[pl][PARTITION_SPLIT];
2753      chosen_rdc.rdcost =
2754          RDCOST(x->rdmult, x->rddiv, chosen_rdc.rate, chosen_rdc.dist);
2755    }
2756  }
2757
2758  // If last_part is better set the partitioning to that.
2759  if (last_part_rdc.rdcost < chosen_rdc.rdcost) {
2760    mi_8x8[0]->sb_type = bsize;
2761    if (bsize >= BLOCK_8X8) pc_tree->partitioning = partition;
2762    chosen_rdc = last_part_rdc;
2763  }
2764  // If none was better set the partitioning to that.
2765  if (none_rdc.rdcost < chosen_rdc.rdcost) {
2766    if (bsize >= BLOCK_8X8) pc_tree->partitioning = PARTITION_NONE;
2767    chosen_rdc = none_rdc;
2768  }
2769
2770  restore_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
2771
2772  // We must have chosen a partitioning and encoding or we'll fail later on.
2773  // No other opportunities for success.
2774  if (bsize == BLOCK_64X64)
2775    assert(chosen_rdc.rate < INT_MAX && chosen_rdc.dist < INT64_MAX);
2776
2777  if (do_recon) {
2778    int output_enabled = (bsize == BLOCK_64X64);
2779    encode_sb(cpi, td, tile_info, tp, mi_row, mi_col, output_enabled, bsize,
2780              pc_tree);
2781  }
2782
2783  *rate = chosen_rdc.rate;
2784  *dist = chosen_rdc.dist;
2785}
2786
2787static const BLOCK_SIZE min_partition_size[BLOCK_SIZES] = {
2788  BLOCK_4X4,   BLOCK_4X4,   BLOCK_4X4,  BLOCK_4X4, BLOCK_4X4,
2789  BLOCK_4X4,   BLOCK_8X8,   BLOCK_8X8,  BLOCK_8X8, BLOCK_16X16,
2790  BLOCK_16X16, BLOCK_16X16, BLOCK_16X16
2791};
2792
2793static const BLOCK_SIZE max_partition_size[BLOCK_SIZES] = {
2794  BLOCK_8X8,   BLOCK_16X16, BLOCK_16X16, BLOCK_16X16, BLOCK_32X32,
2795  BLOCK_32X32, BLOCK_32X32, BLOCK_64X64, BLOCK_64X64, BLOCK_64X64,
2796  BLOCK_64X64, BLOCK_64X64, BLOCK_64X64
2797};
2798
2799// Look at all the mode_info entries for blocks that are part of this
2800// partition and find the min and max values for sb_type.
2801// At the moment this is designed to work on a 64x64 SB but could be
2802// adjusted to use a size parameter.
2803//
2804// The min and max are assumed to have been initialized prior to calling this
2805// function so repeat calls can accumulate a min and max of more than one sb64.
2806static void get_sb_partition_size_range(MACROBLOCKD *xd, MODE_INFO **mi_8x8,
2807                                        BLOCK_SIZE *min_block_size,
2808                                        BLOCK_SIZE *max_block_size,
2809                                        int bs_hist[BLOCK_SIZES]) {
2810  int sb_width_in_blocks = MI_BLOCK_SIZE;
2811  int sb_height_in_blocks = MI_BLOCK_SIZE;
2812  int i, j;
2813  int index = 0;
2814
2815  // Check the sb_type for each block that belongs to this region.
2816  for (i = 0; i < sb_height_in_blocks; ++i) {
2817    for (j = 0; j < sb_width_in_blocks; ++j) {
2818      MODE_INFO *mi = mi_8x8[index + j];
2819      BLOCK_SIZE sb_type = mi ? mi->sb_type : 0;
2820      bs_hist[sb_type]++;
2821      *min_block_size = VPXMIN(*min_block_size, sb_type);
2822      *max_block_size = VPXMAX(*max_block_size, sb_type);
2823    }
2824    index += xd->mi_stride;
2825  }
2826}
2827
2828// Next square block size less or equal than current block size.
2829static const BLOCK_SIZE next_square_size[BLOCK_SIZES] = {
2830  BLOCK_4X4,   BLOCK_4X4,   BLOCK_4X4,   BLOCK_8X8,   BLOCK_8X8,
2831  BLOCK_8X8,   BLOCK_16X16, BLOCK_16X16, BLOCK_16X16, BLOCK_32X32,
2832  BLOCK_32X32, BLOCK_32X32, BLOCK_64X64
2833};
2834
2835// Look at neighboring blocks and set a min and max partition size based on
2836// what they chose.
2837static void rd_auto_partition_range(VP9_COMP *cpi, const TileInfo *const tile,
2838                                    MACROBLOCKD *const xd, int mi_row,
2839                                    int mi_col, BLOCK_SIZE *min_block_size,
2840                                    BLOCK_SIZE *max_block_size) {
2841  VP9_COMMON *const cm = &cpi->common;
2842  MODE_INFO **mi = xd->mi;
2843  const int left_in_image = !!xd->left_mi;
2844  const int above_in_image = !!xd->above_mi;
2845  const int row8x8_remaining = tile->mi_row_end - mi_row;
2846  const int col8x8_remaining = tile->mi_col_end - mi_col;
2847  int bh, bw;
2848  BLOCK_SIZE min_size = BLOCK_4X4;
2849  BLOCK_SIZE max_size = BLOCK_64X64;
2850  int bs_hist[BLOCK_SIZES] = { 0 };
2851
2852  // Trap case where we do not have a prediction.
2853  if (left_in_image || above_in_image || cm->frame_type != KEY_FRAME) {
2854    // Default "min to max" and "max to min"
2855    min_size = BLOCK_64X64;
2856    max_size = BLOCK_4X4;
2857
2858    // NOTE: each call to get_sb_partition_size_range() uses the previous
2859    // passed in values for min and max as a starting point.
2860    // Find the min and max partition used in previous frame at this location
2861    if (cm->frame_type != KEY_FRAME) {
2862      MODE_INFO **prev_mi =
2863          &cm->prev_mi_grid_visible[mi_row * xd->mi_stride + mi_col];
2864      get_sb_partition_size_range(xd, prev_mi, &min_size, &max_size, bs_hist);
2865    }
2866    // Find the min and max partition sizes used in the left SB64
2867    if (left_in_image) {
2868      MODE_INFO **left_sb64_mi = &mi[-MI_BLOCK_SIZE];
2869      get_sb_partition_size_range(xd, left_sb64_mi, &min_size, &max_size,
2870                                  bs_hist);
2871    }
2872    // Find the min and max partition sizes used in the above SB64.
2873    if (above_in_image) {
2874      MODE_INFO **above_sb64_mi = &mi[-xd->mi_stride * MI_BLOCK_SIZE];
2875      get_sb_partition_size_range(xd, above_sb64_mi, &min_size, &max_size,
2876                                  bs_hist);
2877    }
2878
2879    // Adjust observed min and max for "relaxed" auto partition case.
2880    if (cpi->sf.auto_min_max_partition_size == RELAXED_NEIGHBORING_MIN_MAX) {
2881      min_size = min_partition_size[min_size];
2882      max_size = max_partition_size[max_size];
2883    }
2884  }
2885
2886  // Check border cases where max and min from neighbors may not be legal.
2887  max_size = find_partition_size(max_size, row8x8_remaining, col8x8_remaining,
2888                                 &bh, &bw);
2889  // Test for blocks at the edge of the active image.
2890  // This may be the actual edge of the image or where there are formatting
2891  // bars.
2892  if (vp9_active_edge_sb(cpi, mi_row, mi_col)) {
2893    min_size = BLOCK_4X4;
2894  } else {
2895    min_size =
2896        VPXMIN(cpi->sf.rd_auto_partition_min_limit, VPXMIN(min_size, max_size));
2897  }
2898
2899  // When use_square_partition_only is true, make sure at least one square
2900  // partition is allowed by selecting the next smaller square size as
2901  // *min_block_size.
2902  if (cpi->sf.use_square_partition_only &&
2903      next_square_size[max_size] < min_size) {
2904    min_size = next_square_size[max_size];
2905  }
2906
2907  *min_block_size = min_size;
2908  *max_block_size = max_size;
2909}
2910
2911// TODO(jingning) refactor functions setting partition search range
2912static void set_partition_range(VP9_COMMON *cm, MACROBLOCKD *xd, int mi_row,
2913                                int mi_col, BLOCK_SIZE bsize,
2914                                BLOCK_SIZE *min_bs, BLOCK_SIZE *max_bs) {
2915  int mi_width = num_8x8_blocks_wide_lookup[bsize];
2916  int mi_height = num_8x8_blocks_high_lookup[bsize];
2917  int idx, idy;
2918
2919  MODE_INFO *mi;
2920  const int idx_str = cm->mi_stride * mi_row + mi_col;
2921  MODE_INFO **prev_mi = &cm->prev_mi_grid_visible[idx_str];
2922  BLOCK_SIZE bs, min_size, max_size;
2923
2924  min_size = BLOCK_64X64;
2925  max_size = BLOCK_4X4;
2926
2927  if (prev_mi) {
2928    for (idy = 0; idy < mi_height; ++idy) {
2929      for (idx = 0; idx < mi_width; ++idx) {
2930        mi = prev_mi[idy * cm->mi_stride + idx];
2931        bs = mi ? mi->sb_type : bsize;
2932        min_size = VPXMIN(min_size, bs);
2933        max_size = VPXMAX(max_size, bs);
2934      }
2935    }
2936  }
2937
2938  if (xd->left_mi) {
2939    for (idy = 0; idy < mi_height; ++idy) {
2940      mi = xd->mi[idy * cm->mi_stride - 1];
2941      bs = mi ? mi->sb_type : bsize;
2942      min_size = VPXMIN(min_size, bs);
2943      max_size = VPXMAX(max_size, bs);
2944    }
2945  }
2946
2947  if (xd->above_mi) {
2948    for (idx = 0; idx < mi_width; ++idx) {
2949      mi = xd->mi[idx - cm->mi_stride];
2950      bs = mi ? mi->sb_type : bsize;
2951      min_size = VPXMIN(min_size, bs);
2952      max_size = VPXMAX(max_size, bs);
2953    }
2954  }
2955
2956  if (min_size == max_size) {
2957    min_size = min_partition_size[min_size];
2958    max_size = max_partition_size[max_size];
2959  }
2960
2961  *min_bs = min_size;
2962  *max_bs = max_size;
2963}
2964
2965static INLINE void store_pred_mv(MACROBLOCK *x, PICK_MODE_CONTEXT *ctx) {
2966  memcpy(ctx->pred_mv, x->pred_mv, sizeof(x->pred_mv));
2967}
2968
2969static INLINE void load_pred_mv(MACROBLOCK *x, PICK_MODE_CONTEXT *ctx) {
2970  memcpy(x->pred_mv, ctx->pred_mv, sizeof(x->pred_mv));
2971}
2972
2973#if CONFIG_FP_MB_STATS
2974const int num_16x16_blocks_wide_lookup[BLOCK_SIZES] = { 1, 1, 1, 1, 1, 1, 1,
2975                                                        1, 2, 2, 2, 4, 4 };
2976const int num_16x16_blocks_high_lookup[BLOCK_SIZES] = { 1, 1, 1, 1, 1, 1, 1,
2977                                                        2, 1, 2, 4, 2, 4 };
2978const int qindex_skip_threshold_lookup[BLOCK_SIZES] = {
2979  0, 10, 10, 30, 40, 40, 60, 80, 80, 90, 100, 100, 120
2980};
2981const int qindex_split_threshold_lookup[BLOCK_SIZES] = {
2982  0, 3, 3, 7, 15, 15, 30, 40, 40, 60, 80, 80, 120
2983};
2984const int complexity_16x16_blocks_threshold[BLOCK_SIZES] = {
2985  1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 4, 4, 6
2986};
2987
2988typedef enum {
2989  MV_ZERO = 0,
2990  MV_LEFT = 1,
2991  MV_UP = 2,
2992  MV_RIGHT = 3,
2993  MV_DOWN = 4,
2994  MV_INVALID
2995} MOTION_DIRECTION;
2996
2997static INLINE MOTION_DIRECTION get_motion_direction_fp(uint8_t fp_byte) {
2998  if (fp_byte & FPMB_MOTION_ZERO_MASK) {
2999    return MV_ZERO;
3000  } else if (fp_byte & FPMB_MOTION_LEFT_MASK) {
3001    return MV_LEFT;
3002  } else if (fp_byte & FPMB_MOTION_RIGHT_MASK) {
3003    return MV_RIGHT;
3004  } else if (fp_byte & FPMB_MOTION_UP_MASK) {
3005    return MV_UP;
3006  } else {
3007    return MV_DOWN;
3008  }
3009}
3010
3011static INLINE int get_motion_inconsistency(MOTION_DIRECTION this_mv,
3012                                           MOTION_DIRECTION that_mv) {
3013  if (this_mv == that_mv) {
3014    return 0;
3015  } else {
3016    return abs(this_mv - that_mv) == 2 ? 2 : 1;
3017  }
3018}
3019#endif
3020
3021// Calculate the score used in machine-learning based partition search early
3022// termination.
3023static double compute_score(VP9_COMMON *const cm, MACROBLOCKD *const xd,
3024                            PICK_MODE_CONTEXT *ctx, int mi_row, int mi_col,
3025                            BLOCK_SIZE bsize) {
3026  const double *clf;
3027  const double *mean;
3028  const double *sd;
3029  const int mag_mv =
3030      abs(ctx->mic.mv[0].as_mv.col) + abs(ctx->mic.mv[0].as_mv.row);
3031  const int left_in_image = !!xd->left_mi;
3032  const int above_in_image = !!xd->above_mi;
3033  MODE_INFO **prev_mi =
3034      &cm->prev_mi_grid_visible[mi_col + cm->mi_stride * mi_row];
3035  int above_par = 0;  // above_partitioning
3036  int left_par = 0;   // left_partitioning
3037  int last_par = 0;   // last_partitioning
3038  BLOCK_SIZE context_size;
3039  double score;
3040  int offset = 0;
3041
3042  assert(b_width_log2_lookup[bsize] == b_height_log2_lookup[bsize]);
3043
3044  if (above_in_image) {
3045    context_size = xd->above_mi->sb_type;
3046    if (context_size < bsize)
3047      above_par = 2;
3048    else if (context_size == bsize)
3049      above_par = 1;
3050  }
3051
3052  if (left_in_image) {
3053    context_size = xd->left_mi->sb_type;
3054    if (context_size < bsize)
3055      left_par = 2;
3056    else if (context_size == bsize)
3057      left_par = 1;
3058  }
3059
3060  if (prev_mi) {
3061    context_size = prev_mi[0]->sb_type;
3062    if (context_size < bsize)
3063      last_par = 2;
3064    else if (context_size == bsize)
3065      last_par = 1;
3066  }
3067
3068  if (bsize == BLOCK_64X64)
3069    offset = 0;
3070  else if (bsize == BLOCK_32X32)
3071    offset = 8;
3072  else if (bsize == BLOCK_16X16)
3073    offset = 16;
3074
3075  // early termination score calculation
3076  clf = &classifiers[offset];
3077  mean = &train_mean[offset];
3078  sd = &train_stdm[offset];
3079  score = clf[0] * (((double)ctx->rate - mean[0]) / sd[0]) +
3080          clf[1] * (((double)ctx->dist - mean[1]) / sd[1]) +
3081          clf[2] * (((double)mag_mv / 2 - mean[2]) * sd[2]) +
3082          clf[3] * (((double)(left_par + above_par) / 2 - mean[3]) * sd[3]) +
3083          clf[4] * (((double)ctx->sum_y_eobs - mean[4]) / sd[4]) +
3084          clf[5] * (((double)cm->base_qindex - mean[5]) * sd[5]) +
3085          clf[6] * (((double)last_par - mean[6]) * sd[6]) + clf[7];
3086  return score;
3087}
3088
3089// TODO(jingning,jimbankoski,rbultje): properly skip partition types that are
3090// unlikely to be selected depending on previous rate-distortion optimization
3091// results, for encoding speed-up.
3092static void rd_pick_partition(VP9_COMP *cpi, ThreadData *td,
3093                              TileDataEnc *tile_data, TOKENEXTRA **tp,
3094                              int mi_row, int mi_col, BLOCK_SIZE bsize,
3095                              RD_COST *rd_cost, int64_t best_rd,
3096                              PC_TREE *pc_tree) {
3097  VP9_COMMON *const cm = &cpi->common;
3098  TileInfo *const tile_info = &tile_data->tile_info;
3099  MACROBLOCK *const x = &td->mb;
3100  MACROBLOCKD *const xd = &x->e_mbd;
3101  const int mi_step = num_8x8_blocks_wide_lookup[bsize] / 2;
3102  ENTROPY_CONTEXT l[16 * MAX_MB_PLANE], a[16 * MAX_MB_PLANE];
3103  PARTITION_CONTEXT sl[8], sa[8];
3104  TOKENEXTRA *tp_orig = *tp;
3105  PICK_MODE_CONTEXT *ctx = &pc_tree->none;
3106  int i;
3107  const int pl = partition_plane_context(xd, mi_row, mi_col, bsize);
3108  BLOCK_SIZE subsize;
3109  RD_COST this_rdc, sum_rdc, best_rdc;
3110  int do_split = bsize >= BLOCK_8X8;
3111  int do_rect = 1;
3112  INTERP_FILTER pred_interp_filter;
3113
3114  // Override skipping rectangular partition operations for edge blocks
3115  const int force_horz_split = (mi_row + mi_step >= cm->mi_rows);
3116  const int force_vert_split = (mi_col + mi_step >= cm->mi_cols);
3117  const int xss = x->e_mbd.plane[1].subsampling_x;
3118  const int yss = x->e_mbd.plane[1].subsampling_y;
3119
3120  BLOCK_SIZE min_size = x->min_partition_size;
3121  BLOCK_SIZE max_size = x->max_partition_size;
3122
3123#if CONFIG_FP_MB_STATS
3124  unsigned int src_diff_var = UINT_MAX;
3125  int none_complexity = 0;
3126#endif
3127
3128  int partition_none_allowed = !force_horz_split && !force_vert_split;
3129  int partition_horz_allowed =
3130      !force_vert_split && yss <= xss && bsize >= BLOCK_8X8;
3131  int partition_vert_allowed =
3132      !force_horz_split && xss <= yss && bsize >= BLOCK_8X8;
3133
3134  int64_t dist_breakout_thr = cpi->sf.partition_search_breakout_thr.dist;
3135  int rate_breakout_thr = cpi->sf.partition_search_breakout_thr.rate;
3136
3137  (void)*tp_orig;
3138
3139  assert(num_8x8_blocks_wide_lookup[bsize] ==
3140         num_8x8_blocks_high_lookup[bsize]);
3141
3142  // Adjust dist breakout threshold according to the partition size.
3143  dist_breakout_thr >>=
3144      8 - (b_width_log2_lookup[bsize] + b_height_log2_lookup[bsize]);
3145  rate_breakout_thr *= num_pels_log2_lookup[bsize];
3146
3147  vp9_rd_cost_init(&this_rdc);
3148  vp9_rd_cost_init(&sum_rdc);
3149  vp9_rd_cost_reset(&best_rdc);
3150  best_rdc.rdcost = best_rd;
3151
3152  set_offsets(cpi, tile_info, x, mi_row, mi_col, bsize);
3153
3154  if (bsize == BLOCK_16X16 && cpi->oxcf.aq_mode != NO_AQ &&
3155      cpi->oxcf.aq_mode != LOOKAHEAD_AQ)
3156    x->mb_energy = vp9_block_energy(cpi, x, bsize);
3157
3158  if (cpi->sf.cb_partition_search && bsize == BLOCK_16X16) {
3159    int cb_partition_search_ctrl =
3160        ((pc_tree->index == 0 || pc_tree->index == 3) +
3161         get_chessboard_index(cm->current_video_frame)) &
3162        0x1;
3163
3164    if (cb_partition_search_ctrl && bsize > min_size && bsize < max_size)
3165      set_partition_range(cm, xd, mi_row, mi_col, bsize, &min_size, &max_size);
3166  }
3167
3168  // Determine partition types in search according to the speed features.
3169  // The threshold set here has to be of square block size.
3170  if (cpi->sf.auto_min_max_partition_size) {
3171    partition_none_allowed &= (bsize <= max_size && bsize >= min_size);
3172    partition_horz_allowed &=
3173        ((bsize <= max_size && bsize > min_size) || force_horz_split);
3174    partition_vert_allowed &=
3175        ((bsize <= max_size && bsize > min_size) || force_vert_split);
3176    do_split &= bsize > min_size;
3177  }
3178
3179  if (cpi->sf.use_square_partition_only &&
3180      bsize > cpi->sf.use_square_only_threshold) {
3181    if (cpi->use_svc) {
3182      if (!vp9_active_h_edge(cpi, mi_row, mi_step) || x->e_mbd.lossless)
3183        partition_horz_allowed &= force_horz_split;
3184      if (!vp9_active_v_edge(cpi, mi_row, mi_step) || x->e_mbd.lossless)
3185        partition_vert_allowed &= force_vert_split;
3186    } else {
3187      partition_horz_allowed &= force_horz_split;
3188      partition_vert_allowed &= force_vert_split;
3189    }
3190  }
3191
3192  save_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
3193
3194#if CONFIG_FP_MB_STATS
3195  if (cpi->use_fp_mb_stats) {
3196    set_offsets(cpi, tile_info, x, mi_row, mi_col, bsize);
3197    src_diff_var = get_sby_perpixel_diff_variance(cpi, &x->plane[0].src, mi_row,
3198                                                  mi_col, bsize);
3199  }
3200#endif
3201
3202#if CONFIG_FP_MB_STATS
3203  // Decide whether we shall split directly and skip searching NONE by using
3204  // the first pass block statistics
3205  if (cpi->use_fp_mb_stats && bsize >= BLOCK_32X32 && do_split &&
3206      partition_none_allowed && src_diff_var > 4 &&
3207      cm->base_qindex < qindex_split_threshold_lookup[bsize]) {
3208    int mb_row = mi_row >> 1;
3209    int mb_col = mi_col >> 1;
3210    int mb_row_end =
3211        VPXMIN(mb_row + num_16x16_blocks_high_lookup[bsize], cm->mb_rows);
3212    int mb_col_end =
3213        VPXMIN(mb_col + num_16x16_blocks_wide_lookup[bsize], cm->mb_cols);
3214    int r, c;
3215
3216    // compute a complexity measure, basically measure inconsistency of motion
3217    // vectors obtained from the first pass in the current block
3218    for (r = mb_row; r < mb_row_end; r++) {
3219      for (c = mb_col; c < mb_col_end; c++) {
3220        const int mb_index = r * cm->mb_cols + c;
3221
3222        MOTION_DIRECTION this_mv;
3223        MOTION_DIRECTION right_mv;
3224        MOTION_DIRECTION bottom_mv;
3225
3226        this_mv =
3227            get_motion_direction_fp(cpi->twopass.this_frame_mb_stats[mb_index]);
3228
3229        // to its right
3230        if (c != mb_col_end - 1) {
3231          right_mv = get_motion_direction_fp(
3232              cpi->twopass.this_frame_mb_stats[mb_index + 1]);
3233          none_complexity += get_motion_inconsistency(this_mv, right_mv);
3234        }
3235
3236        // to its bottom
3237        if (r != mb_row_end - 1) {
3238          bottom_mv = get_motion_direction_fp(
3239              cpi->twopass.this_frame_mb_stats[mb_index + cm->mb_cols]);
3240          none_complexity += get_motion_inconsistency(this_mv, bottom_mv);
3241        }
3242
3243        // do not count its left and top neighbors to avoid double counting
3244      }
3245    }
3246
3247    if (none_complexity > complexity_16x16_blocks_threshold[bsize]) {
3248      partition_none_allowed = 0;
3249    }
3250  }
3251#endif
3252
3253  // PARTITION_NONE
3254  if (partition_none_allowed) {
3255    rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &this_rdc, bsize, ctx,
3256                     best_rdc.rdcost);
3257    if (this_rdc.rate != INT_MAX) {
3258      if (bsize >= BLOCK_8X8) {
3259        this_rdc.rate += cpi->partition_cost[pl][PARTITION_NONE];
3260        this_rdc.rdcost =
3261            RDCOST(x->rdmult, x->rddiv, this_rdc.rate, this_rdc.dist);
3262      }
3263
3264      if (this_rdc.rdcost < best_rdc.rdcost) {
3265        MODE_INFO *mi = xd->mi[0];
3266
3267        best_rdc = this_rdc;
3268        if (bsize >= BLOCK_8X8) pc_tree->partitioning = PARTITION_NONE;
3269
3270        if (!cpi->sf.ml_partition_search_early_termination) {
3271          // If all y, u, v transform blocks in this partition are skippable,
3272          // and the dist & rate are within the thresholds, the partition search
3273          // is terminated for current branch of the partition search tree.
3274          if (!x->e_mbd.lossless && ctx->skippable &&
3275              ((best_rdc.dist < (dist_breakout_thr >> 2)) ||
3276               (best_rdc.dist < dist_breakout_thr &&
3277                best_rdc.rate < rate_breakout_thr))) {
3278            do_split = 0;
3279            do_rect = 0;
3280          }
3281        } else {
3282          // Currently, the machine-learning based partition search early
3283          // termination is only used while bsize is 16x16, 32x32 or 64x64,
3284          // VPXMIN(cm->width, cm->height) >= 480, and speed = 0.
3285          if (!x->e_mbd.lossless &&
3286              !segfeature_active(&cm->seg, mi->segment_id, SEG_LVL_SKIP) &&
3287              ctx->mic.mode >= INTRA_MODES && bsize >= BLOCK_16X16) {
3288            if (compute_score(cm, xd, ctx, mi_row, mi_col, bsize) < 0.0) {
3289              do_split = 0;
3290              do_rect = 0;
3291            }
3292          }
3293        }
3294
3295#if CONFIG_FP_MB_STATS
3296        // Check if every 16x16 first pass block statistics has zero
3297        // motion and the corresponding first pass residue is small enough.
3298        // If that is the case, check the difference variance between the
3299        // current frame and the last frame. If the variance is small enough,
3300        // stop further splitting in RD optimization
3301        if (cpi->use_fp_mb_stats && do_split != 0 &&
3302            cm->base_qindex > qindex_skip_threshold_lookup[bsize]) {
3303          int mb_row = mi_row >> 1;
3304          int mb_col = mi_col >> 1;
3305          int mb_row_end =
3306              VPXMIN(mb_row + num_16x16_blocks_high_lookup[bsize], cm->mb_rows);
3307          int mb_col_end =
3308              VPXMIN(mb_col + num_16x16_blocks_wide_lookup[bsize], cm->mb_cols);
3309          int r, c;
3310
3311          int skip = 1;
3312          for (r = mb_row; r < mb_row_end; r++) {
3313            for (c = mb_col; c < mb_col_end; c++) {
3314              const int mb_index = r * cm->mb_cols + c;
3315              if (!(cpi->twopass.this_frame_mb_stats[mb_index] &
3316                    FPMB_MOTION_ZERO_MASK) ||
3317                  !(cpi->twopass.this_frame_mb_stats[mb_index] &
3318                    FPMB_ERROR_SMALL_MASK)) {
3319                skip = 0;
3320                break;
3321              }
3322            }
3323            if (skip == 0) {
3324              break;
3325            }
3326          }
3327
3328          if (skip) {
3329            if (src_diff_var == UINT_MAX) {
3330              set_offsets(cpi, tile_info, x, mi_row, mi_col, bsize);
3331              src_diff_var = get_sby_perpixel_diff_variance(
3332                  cpi, &x->plane[0].src, mi_row, mi_col, bsize);
3333            }
3334            if (src_diff_var < 8) {
3335              do_split = 0;
3336              do_rect = 0;
3337            }
3338          }
3339        }
3340#endif
3341      }
3342    }
3343    restore_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
3344  }
3345
3346  // store estimated motion vector
3347  if (cpi->sf.adaptive_motion_search) store_pred_mv(x, ctx);
3348
3349  // If the interp_filter is marked as SWITCHABLE_FILTERS, it was for an
3350  // intra block and used for context purposes.
3351  if (ctx->mic.interp_filter == SWITCHABLE_FILTERS) {
3352    pred_interp_filter = EIGHTTAP;
3353  } else {
3354    pred_interp_filter = ctx->mic.interp_filter;
3355  }
3356
3357  // PARTITION_SPLIT
3358  // TODO(jingning): use the motion vectors given by the above search as
3359  // the starting point of motion search in the following partition type check.
3360  if (do_split) {
3361    subsize = get_subsize(bsize, PARTITION_SPLIT);
3362    if (bsize == BLOCK_8X8) {
3363      i = 4;
3364      if (cpi->sf.adaptive_pred_interp_filter && partition_none_allowed)
3365        pc_tree->leaf_split[0]->pred_interp_filter = pred_interp_filter;
3366      rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &sum_rdc, subsize,
3367                       pc_tree->leaf_split[0], best_rdc.rdcost);
3368
3369      if (sum_rdc.rate == INT_MAX) sum_rdc.rdcost = INT64_MAX;
3370    } else {
3371      for (i = 0; i < 4 && sum_rdc.rdcost < best_rdc.rdcost; ++i) {
3372        const int x_idx = (i & 1) * mi_step;
3373        const int y_idx = (i >> 1) * mi_step;
3374
3375        if (mi_row + y_idx >= cm->mi_rows || mi_col + x_idx >= cm->mi_cols)
3376          continue;
3377
3378        if (cpi->sf.adaptive_motion_search) load_pred_mv(x, ctx);
3379
3380        pc_tree->split[i]->index = i;
3381        rd_pick_partition(cpi, td, tile_data, tp, mi_row + y_idx,
3382                          mi_col + x_idx, subsize, &this_rdc,
3383                          best_rdc.rdcost - sum_rdc.rdcost, pc_tree->split[i]);
3384
3385        if (this_rdc.rate == INT_MAX) {
3386          sum_rdc.rdcost = INT64_MAX;
3387          break;
3388        } else {
3389          sum_rdc.rate += this_rdc.rate;
3390          sum_rdc.dist += this_rdc.dist;
3391          sum_rdc.rdcost += this_rdc.rdcost;
3392        }
3393      }
3394    }
3395
3396    if (sum_rdc.rdcost < best_rdc.rdcost && i == 4) {
3397      sum_rdc.rate += cpi->partition_cost[pl][PARTITION_SPLIT];
3398      sum_rdc.rdcost = RDCOST(x->rdmult, x->rddiv, sum_rdc.rate, sum_rdc.dist);
3399
3400      if (sum_rdc.rdcost < best_rdc.rdcost) {
3401        best_rdc = sum_rdc;
3402        pc_tree->partitioning = PARTITION_SPLIT;
3403
3404        // Rate and distortion based partition search termination clause.
3405        if (!cpi->sf.ml_partition_search_early_termination &&
3406            !x->e_mbd.lossless && ((best_rdc.dist < (dist_breakout_thr >> 2)) ||
3407                                   (best_rdc.dist < dist_breakout_thr &&
3408                                    best_rdc.rate < rate_breakout_thr))) {
3409          do_rect = 0;
3410        }
3411      }
3412    } else {
3413      // skip rectangular partition test when larger block size
3414      // gives better rd cost
3415      if ((cpi->sf.less_rectangular_check) &&
3416          ((bsize > cpi->sf.use_square_only_threshold) ||
3417           (best_rdc.dist < dist_breakout_thr)))
3418        do_rect &= !partition_none_allowed;
3419    }
3420    restore_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
3421  }
3422
3423  // PARTITION_HORZ
3424  if (partition_horz_allowed &&
3425      (do_rect || vp9_active_h_edge(cpi, mi_row, mi_step))) {
3426    subsize = get_subsize(bsize, PARTITION_HORZ);
3427    if (cpi->sf.adaptive_motion_search) load_pred_mv(x, ctx);
3428    if (cpi->sf.adaptive_pred_interp_filter && bsize == BLOCK_8X8 &&
3429        partition_none_allowed)
3430      pc_tree->horizontal[0].pred_interp_filter = pred_interp_filter;
3431    rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &sum_rdc, subsize,
3432                     &pc_tree->horizontal[0], best_rdc.rdcost);
3433
3434    if (sum_rdc.rdcost < best_rdc.rdcost && mi_row + mi_step < cm->mi_rows &&
3435        bsize > BLOCK_8X8) {
3436      PICK_MODE_CONTEXT *ctx = &pc_tree->horizontal[0];
3437      update_state(cpi, td, ctx, mi_row, mi_col, subsize, 0);
3438      encode_superblock(cpi, td, tp, 0, mi_row, mi_col, subsize, ctx);
3439
3440      if (cpi->sf.adaptive_motion_search) load_pred_mv(x, ctx);
3441      if (cpi->sf.adaptive_pred_interp_filter && bsize == BLOCK_8X8 &&
3442          partition_none_allowed)
3443        pc_tree->horizontal[1].pred_interp_filter = pred_interp_filter;
3444      rd_pick_sb_modes(cpi, tile_data, x, mi_row + mi_step, mi_col, &this_rdc,
3445                       subsize, &pc_tree->horizontal[1],
3446                       best_rdc.rdcost - sum_rdc.rdcost);
3447      if (this_rdc.rate == INT_MAX) {
3448        sum_rdc.rdcost = INT64_MAX;
3449      } else {
3450        sum_rdc.rate += this_rdc.rate;
3451        sum_rdc.dist += this_rdc.dist;
3452        sum_rdc.rdcost += this_rdc.rdcost;
3453      }
3454    }
3455
3456    if (sum_rdc.rdcost < best_rdc.rdcost) {
3457      sum_rdc.rate += cpi->partition_cost[pl][PARTITION_HORZ];
3458      sum_rdc.rdcost = RDCOST(x->rdmult, x->rddiv, sum_rdc.rate, sum_rdc.dist);
3459      if (sum_rdc.rdcost < best_rdc.rdcost) {
3460        best_rdc = sum_rdc;
3461        pc_tree->partitioning = PARTITION_HORZ;
3462
3463        if ((cpi->sf.less_rectangular_check) &&
3464            (bsize > cpi->sf.use_square_only_threshold))
3465          do_rect = 0;
3466      }
3467    }
3468    restore_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
3469  }
3470
3471  // PARTITION_VERT
3472  if (partition_vert_allowed &&
3473      (do_rect || vp9_active_v_edge(cpi, mi_col, mi_step))) {
3474    subsize = get_subsize(bsize, PARTITION_VERT);
3475
3476    if (cpi->sf.adaptive_motion_search) load_pred_mv(x, ctx);
3477    if (cpi->sf.adaptive_pred_interp_filter && bsize == BLOCK_8X8 &&
3478        partition_none_allowed)
3479      pc_tree->vertical[0].pred_interp_filter = pred_interp_filter;
3480    rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &sum_rdc, subsize,
3481                     &pc_tree->vertical[0], best_rdc.rdcost);
3482    if (sum_rdc.rdcost < best_rdc.rdcost && mi_col + mi_step < cm->mi_cols &&
3483        bsize > BLOCK_8X8) {
3484      update_state(cpi, td, &pc_tree->vertical[0], mi_row, mi_col, subsize, 0);
3485      encode_superblock(cpi, td, tp, 0, mi_row, mi_col, subsize,
3486                        &pc_tree->vertical[0]);
3487
3488      if (cpi->sf.adaptive_motion_search) load_pred_mv(x, ctx);
3489      if (cpi->sf.adaptive_pred_interp_filter && bsize == BLOCK_8X8 &&
3490          partition_none_allowed)
3491        pc_tree->vertical[1].pred_interp_filter = pred_interp_filter;
3492      rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col + mi_step, &this_rdc,
3493                       subsize, &pc_tree->vertical[1],
3494                       best_rdc.rdcost - sum_rdc.rdcost);
3495      if (this_rdc.rate == INT_MAX) {
3496        sum_rdc.rdcost = INT64_MAX;
3497      } else {
3498        sum_rdc.rate += this_rdc.rate;
3499        sum_rdc.dist += this_rdc.dist;
3500        sum_rdc.rdcost += this_rdc.rdcost;
3501      }
3502    }
3503
3504    if (sum_rdc.rdcost < best_rdc.rdcost) {
3505      sum_rdc.rate += cpi->partition_cost[pl][PARTITION_VERT];
3506      sum_rdc.rdcost = RDCOST(x->rdmult, x->rddiv, sum_rdc.rate, sum_rdc.dist);
3507      if (sum_rdc.rdcost < best_rdc.rdcost) {
3508        best_rdc = sum_rdc;
3509        pc_tree->partitioning = PARTITION_VERT;
3510      }
3511    }
3512    restore_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
3513  }
3514
3515  // TODO(jbb): This code added so that we avoid static analysis
3516  // warning related to the fact that best_rd isn't used after this
3517  // point.  This code should be refactored so that the duplicate
3518  // checks occur in some sub function and thus are used...
3519  (void)best_rd;
3520  *rd_cost = best_rdc;
3521
3522  if (best_rdc.rate < INT_MAX && best_rdc.dist < INT64_MAX &&
3523      pc_tree->index != 3) {
3524    int output_enabled = (bsize == BLOCK_64X64);
3525    encode_sb(cpi, td, tile_info, tp, mi_row, mi_col, output_enabled, bsize,
3526              pc_tree);
3527  }
3528
3529  if (bsize == BLOCK_64X64) {
3530    assert(tp_orig < *tp);
3531    assert(best_rdc.rate < INT_MAX);
3532    assert(best_rdc.dist < INT64_MAX);
3533  } else {
3534    assert(tp_orig == *tp);
3535  }
3536}
3537
3538static void encode_rd_sb_row(VP9_COMP *cpi, ThreadData *td,
3539                             TileDataEnc *tile_data, int mi_row,
3540                             TOKENEXTRA **tp) {
3541  VP9_COMMON *const cm = &cpi->common;
3542  TileInfo *const tile_info = &tile_data->tile_info;
3543  MACROBLOCK *const x = &td->mb;
3544  MACROBLOCKD *const xd = &x->e_mbd;
3545  SPEED_FEATURES *const sf = &cpi->sf;
3546  const int mi_col_start = tile_info->mi_col_start;
3547  const int mi_col_end = tile_info->mi_col_end;
3548  int mi_col;
3549  const int sb_row = mi_row >> MI_BLOCK_SIZE_LOG2;
3550  const int num_sb_cols =
3551      get_num_cols(tile_data->tile_info, MI_BLOCK_SIZE_LOG2);
3552  int sb_col_in_tile;
3553
3554  // Initialize the left context for the new SB row
3555  memset(&xd->left_context, 0, sizeof(xd->left_context));
3556  memset(xd->left_seg_context, 0, sizeof(xd->left_seg_context));
3557
3558  // Code each SB in the row
3559  for (mi_col = mi_col_start, sb_col_in_tile = 0; mi_col < mi_col_end;
3560       mi_col += MI_BLOCK_SIZE, sb_col_in_tile++) {
3561    const struct segmentation *const seg = &cm->seg;
3562    int dummy_rate;
3563    int64_t dummy_dist;
3564    RD_COST dummy_rdc;
3565    int i;
3566    int seg_skip = 0;
3567
3568    const int idx_str = cm->mi_stride * mi_row + mi_col;
3569    MODE_INFO **mi = cm->mi_grid_visible + idx_str;
3570
3571    (*(cpi->row_mt_sync_read_ptr))(&tile_data->row_mt_sync, sb_row,
3572                                   sb_col_in_tile);
3573
3574    if (sf->adaptive_pred_interp_filter) {
3575      for (i = 0; i < 64; ++i) td->leaf_tree[i].pred_interp_filter = SWITCHABLE;
3576
3577      for (i = 0; i < 64; ++i) {
3578        td->pc_tree[i].vertical[0].pred_interp_filter = SWITCHABLE;
3579        td->pc_tree[i].vertical[1].pred_interp_filter = SWITCHABLE;
3580        td->pc_tree[i].horizontal[0].pred_interp_filter = SWITCHABLE;
3581        td->pc_tree[i].horizontal[1].pred_interp_filter = SWITCHABLE;
3582      }
3583    }
3584
3585    vp9_zero(x->pred_mv);
3586    td->pc_root->index = 0;
3587
3588    if (seg->enabled) {
3589      const uint8_t *const map =
3590          seg->update_map ? cpi->segmentation_map : cm->last_frame_seg_map;
3591      int segment_id = get_segment_id(cm, map, BLOCK_64X64, mi_row, mi_col);
3592      seg_skip = segfeature_active(seg, segment_id, SEG_LVL_SKIP);
3593    }
3594
3595    x->source_variance = UINT_MAX;
3596    if (sf->partition_search_type == FIXED_PARTITION || seg_skip) {
3597      const BLOCK_SIZE bsize =
3598          seg_skip ? BLOCK_64X64 : sf->always_this_block_size;
3599      set_offsets(cpi, tile_info, x, mi_row, mi_col, BLOCK_64X64);
3600      set_fixed_partitioning(cpi, tile_info, mi, mi_row, mi_col, bsize);
3601      rd_use_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col, BLOCK_64X64,
3602                       &dummy_rate, &dummy_dist, 1, td->pc_root);
3603    } else if (cpi->partition_search_skippable_frame) {
3604      BLOCK_SIZE bsize;
3605      set_offsets(cpi, tile_info, x, mi_row, mi_col, BLOCK_64X64);
3606      bsize = get_rd_var_based_fixed_partition(cpi, x, mi_row, mi_col);
3607      set_fixed_partitioning(cpi, tile_info, mi, mi_row, mi_col, bsize);
3608      rd_use_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col, BLOCK_64X64,
3609                       &dummy_rate, &dummy_dist, 1, td->pc_root);
3610    } else if (sf->partition_search_type == VAR_BASED_PARTITION &&
3611               cm->frame_type != KEY_FRAME) {
3612      choose_partitioning(cpi, tile_info, x, mi_row, mi_col);
3613      rd_use_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col, BLOCK_64X64,
3614                       &dummy_rate, &dummy_dist, 1, td->pc_root);
3615    } else {
3616      // If required set upper and lower partition size limits
3617      if (sf->auto_min_max_partition_size) {
3618        set_offsets(cpi, tile_info, x, mi_row, mi_col, BLOCK_64X64);
3619        rd_auto_partition_range(cpi, tile_info, xd, mi_row, mi_col,
3620                                &x->min_partition_size, &x->max_partition_size);
3621      }
3622      rd_pick_partition(cpi, td, tile_data, tp, mi_row, mi_col, BLOCK_64X64,
3623                        &dummy_rdc, INT64_MAX, td->pc_root);
3624    }
3625    (*(cpi->row_mt_sync_write_ptr))(&tile_data->row_mt_sync, sb_row,
3626                                    sb_col_in_tile, num_sb_cols);
3627  }
3628}
3629
3630static void init_encode_frame_mb_context(VP9_COMP *cpi) {
3631  MACROBLOCK *const x = &cpi->td.mb;
3632  VP9_COMMON *const cm = &cpi->common;
3633  MACROBLOCKD *const xd = &x->e_mbd;
3634  const int aligned_mi_cols = mi_cols_aligned_to_sb(cm->mi_cols);
3635
3636  // Copy data over into macro block data structures.
3637  vp9_setup_src_planes(x, cpi->Source, 0, 0);
3638
3639  vp9_setup_block_planes(&x->e_mbd, cm->subsampling_x, cm->subsampling_y);
3640
3641  // Note: this memset assumes above_context[0], [1] and [2]
3642  // are allocated as part of the same buffer.
3643  memset(xd->above_context[0], 0,
3644         sizeof(*xd->above_context[0]) * 2 * aligned_mi_cols * MAX_MB_PLANE);
3645  memset(xd->above_seg_context, 0,
3646         sizeof(*xd->above_seg_context) * aligned_mi_cols);
3647}
3648
3649static int check_dual_ref_flags(VP9_COMP *cpi) {
3650  const int ref_flags = cpi->ref_frame_flags;
3651
3652  if (segfeature_active(&cpi->common.seg, 1, SEG_LVL_REF_FRAME)) {
3653    return 0;
3654  } else {
3655    return (!!(ref_flags & VP9_GOLD_FLAG) + !!(ref_flags & VP9_LAST_FLAG) +
3656            !!(ref_flags & VP9_ALT_FLAG)) >= 2;
3657  }
3658}
3659
3660static void reset_skip_tx_size(VP9_COMMON *cm, TX_SIZE max_tx_size) {
3661  int mi_row, mi_col;
3662  const int mis = cm->mi_stride;
3663  MODE_INFO **mi_ptr = cm->mi_grid_visible;
3664
3665  for (mi_row = 0; mi_row < cm->mi_rows; ++mi_row, mi_ptr += mis) {
3666    for (mi_col = 0; mi_col < cm->mi_cols; ++mi_col) {
3667      if (mi_ptr[mi_col]->tx_size > max_tx_size)
3668        mi_ptr[mi_col]->tx_size = max_tx_size;
3669    }
3670  }
3671}
3672
3673static MV_REFERENCE_FRAME get_frame_type(const VP9_COMP *cpi) {
3674  if (frame_is_intra_only(&cpi->common))
3675    return INTRA_FRAME;
3676  else if (cpi->rc.is_src_frame_alt_ref && cpi->refresh_golden_frame)
3677    return ALTREF_FRAME;
3678  else if (cpi->refresh_golden_frame || cpi->refresh_alt_ref_frame)
3679    return GOLDEN_FRAME;
3680  else
3681    return LAST_FRAME;
3682}
3683
3684static TX_MODE select_tx_mode(const VP9_COMP *cpi, MACROBLOCKD *const xd) {
3685  if (xd->lossless) return ONLY_4X4;
3686  if (cpi->common.frame_type == KEY_FRAME && cpi->sf.use_nonrd_pick_mode)
3687    return ALLOW_16X16;
3688  if (cpi->sf.tx_size_search_method == USE_LARGESTALL)
3689    return ALLOW_32X32;
3690  else if (cpi->sf.tx_size_search_method == USE_FULL_RD ||
3691           cpi->sf.tx_size_search_method == USE_TX_8X8)
3692    return TX_MODE_SELECT;
3693  else
3694    return cpi->common.tx_mode;
3695}
3696
3697static void hybrid_intra_mode_search(VP9_COMP *cpi, MACROBLOCK *const x,
3698                                     RD_COST *rd_cost, BLOCK_SIZE bsize,
3699                                     PICK_MODE_CONTEXT *ctx) {
3700  if (!cpi->sf.nonrd_keyframe && bsize < BLOCK_16X16)
3701    vp9_rd_pick_intra_mode_sb(cpi, x, rd_cost, bsize, ctx, INT64_MAX);
3702  else
3703    vp9_pick_intra_mode(cpi, x, rd_cost, bsize, ctx);
3704}
3705
3706static void nonrd_pick_sb_modes(VP9_COMP *cpi, TileDataEnc *tile_data,
3707                                MACROBLOCK *const x, int mi_row, int mi_col,
3708                                RD_COST *rd_cost, BLOCK_SIZE bsize,
3709                                PICK_MODE_CONTEXT *ctx) {
3710  VP9_COMMON *const cm = &cpi->common;
3711  TileInfo *const tile_info = &tile_data->tile_info;
3712  MACROBLOCKD *const xd = &x->e_mbd;
3713  MODE_INFO *mi;
3714  ENTROPY_CONTEXT l[16 * MAX_MB_PLANE], a[16 * MAX_MB_PLANE];
3715  BLOCK_SIZE bs = VPXMAX(bsize, BLOCK_8X8);  // processing unit block size
3716  const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bs];
3717  const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bs];
3718  int plane;
3719
3720  set_offsets(cpi, tile_info, x, mi_row, mi_col, bsize);
3721  mi = xd->mi[0];
3722  mi->sb_type = bsize;
3723
3724  for (plane = 0; plane < MAX_MB_PLANE; ++plane) {
3725    struct macroblockd_plane *pd = &xd->plane[plane];
3726    memcpy(a + num_4x4_blocks_wide * plane, pd->above_context,
3727           (sizeof(a[0]) * num_4x4_blocks_wide) >> pd->subsampling_x);
3728    memcpy(l + num_4x4_blocks_high * plane, pd->left_context,
3729           (sizeof(l[0]) * num_4x4_blocks_high) >> pd->subsampling_y);
3730  }
3731
3732  if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ && cm->seg.enabled)
3733    if (cyclic_refresh_segment_id_boosted(mi->segment_id))
3734      x->rdmult = vp9_cyclic_refresh_get_rdmult(cpi->cyclic_refresh);
3735
3736  if (cm->frame_type == KEY_FRAME)
3737    hybrid_intra_mode_search(cpi, x, rd_cost, bsize, ctx);
3738  else if (segfeature_active(&cm->seg, mi->segment_id, SEG_LVL_SKIP))
3739    set_mode_info_seg_skip(x, cm->tx_mode, rd_cost, bsize);
3740  else if (bsize >= BLOCK_8X8)
3741    vp9_pick_inter_mode(cpi, x, tile_data, mi_row, mi_col, rd_cost, bsize, ctx);
3742  else
3743    vp9_pick_inter_mode_sub8x8(cpi, x, mi_row, mi_col, rd_cost, bsize, ctx);
3744
3745  duplicate_mode_info_in_sb(cm, xd, mi_row, mi_col, bsize);
3746
3747  for (plane = 0; plane < MAX_MB_PLANE; ++plane) {
3748    struct macroblockd_plane *pd = &xd->plane[plane];
3749    memcpy(pd->above_context, a + num_4x4_blocks_wide * plane,
3750           (sizeof(a[0]) * num_4x4_blocks_wide) >> pd->subsampling_x);
3751    memcpy(pd->left_context, l + num_4x4_blocks_high * plane,
3752           (sizeof(l[0]) * num_4x4_blocks_high) >> pd->subsampling_y);
3753  }
3754
3755  if (rd_cost->rate == INT_MAX) vp9_rd_cost_reset(rd_cost);
3756
3757  ctx->rate = rd_cost->rate;
3758  ctx->dist = rd_cost->dist;
3759}
3760
3761static void fill_mode_info_sb(VP9_COMMON *cm, MACROBLOCK *x, int mi_row,
3762                              int mi_col, BLOCK_SIZE bsize, PC_TREE *pc_tree) {
3763  MACROBLOCKD *xd = &x->e_mbd;
3764  int bsl = b_width_log2_lookup[bsize], hbs = (1 << bsl) / 4;
3765  PARTITION_TYPE partition = pc_tree->partitioning;
3766  BLOCK_SIZE subsize = get_subsize(bsize, partition);
3767
3768  assert(bsize >= BLOCK_8X8);
3769
3770  if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols) return;
3771
3772  switch (partition) {
3773    case PARTITION_NONE:
3774      set_mode_info_offsets(cm, x, xd, mi_row, mi_col);
3775      *(xd->mi[0]) = pc_tree->none.mic;
3776      *(x->mbmi_ext) = pc_tree->none.mbmi_ext;
3777      duplicate_mode_info_in_sb(cm, xd, mi_row, mi_col, bsize);
3778      break;
3779    case PARTITION_VERT:
3780      set_mode_info_offsets(cm, x, xd, mi_row, mi_col);
3781      *(xd->mi[0]) = pc_tree->vertical[0].mic;
3782      *(x->mbmi_ext) = pc_tree->vertical[0].mbmi_ext;
3783      duplicate_mode_info_in_sb(cm, xd, mi_row, mi_col, subsize);
3784
3785      if (mi_col + hbs < cm->mi_cols) {
3786        set_mode_info_offsets(cm, x, xd, mi_row, mi_col + hbs);
3787        *(xd->mi[0]) = pc_tree->vertical[1].mic;
3788        *(x->mbmi_ext) = pc_tree->vertical[1].mbmi_ext;
3789        duplicate_mode_info_in_sb(cm, xd, mi_row, mi_col + hbs, subsize);
3790      }
3791      break;
3792    case PARTITION_HORZ:
3793      set_mode_info_offsets(cm, x, xd, mi_row, mi_col);
3794      *(xd->mi[0]) = pc_tree->horizontal[0].mic;
3795      *(x->mbmi_ext) = pc_tree->horizontal[0].mbmi_ext;
3796      duplicate_mode_info_in_sb(cm, xd, mi_row, mi_col, subsize);
3797      if (mi_row + hbs < cm->mi_rows) {
3798        set_mode_info_offsets(cm, x, xd, mi_row + hbs, mi_col);
3799        *(xd->mi[0]) = pc_tree->horizontal[1].mic;
3800        *(x->mbmi_ext) = pc_tree->horizontal[1].mbmi_ext;
3801        duplicate_mode_info_in_sb(cm, xd, mi_row + hbs, mi_col, subsize);
3802      }
3803      break;
3804    case PARTITION_SPLIT: {
3805      fill_mode_info_sb(cm, x, mi_row, mi_col, subsize, pc_tree->split[0]);
3806      fill_mode_info_sb(cm, x, mi_row, mi_col + hbs, subsize,
3807                        pc_tree->split[1]);
3808      fill_mode_info_sb(cm, x, mi_row + hbs, mi_col, subsize,
3809                        pc_tree->split[2]);
3810      fill_mode_info_sb(cm, x, mi_row + hbs, mi_col + hbs, subsize,
3811                        pc_tree->split[3]);
3812      break;
3813    }
3814    default: break;
3815  }
3816}
3817
3818// Reset the prediction pixel ready flag recursively.
3819static void pred_pixel_ready_reset(PC_TREE *pc_tree, BLOCK_SIZE bsize) {
3820  pc_tree->none.pred_pixel_ready = 0;
3821  pc_tree->horizontal[0].pred_pixel_ready = 0;
3822  pc_tree->horizontal[1].pred_pixel_ready = 0;
3823  pc_tree->vertical[0].pred_pixel_ready = 0;
3824  pc_tree->vertical[1].pred_pixel_ready = 0;
3825
3826  if (bsize > BLOCK_8X8) {
3827    BLOCK_SIZE subsize = get_subsize(bsize, PARTITION_SPLIT);
3828    int i;
3829    for (i = 0; i < 4; ++i) pred_pixel_ready_reset(pc_tree->split[i], subsize);
3830  }
3831}
3832
3833static void nonrd_pick_partition(VP9_COMP *cpi, ThreadData *td,
3834                                 TileDataEnc *tile_data, TOKENEXTRA **tp,
3835                                 int mi_row, int mi_col, BLOCK_SIZE bsize,
3836                                 RD_COST *rd_cost, int do_recon,
3837                                 int64_t best_rd, PC_TREE *pc_tree) {
3838  const SPEED_FEATURES *const sf = &cpi->sf;
3839  VP9_COMMON *const cm = &cpi->common;
3840  TileInfo *const tile_info = &tile_data->tile_info;
3841  MACROBLOCK *const x = &td->mb;
3842  MACROBLOCKD *const xd = &x->e_mbd;
3843  const int ms = num_8x8_blocks_wide_lookup[bsize] / 2;
3844  TOKENEXTRA *tp_orig = *tp;
3845  PICK_MODE_CONTEXT *ctx = &pc_tree->none;
3846  int i;
3847  BLOCK_SIZE subsize = bsize;
3848  RD_COST this_rdc, sum_rdc, best_rdc;
3849  int do_split = bsize >= BLOCK_8X8;
3850  int do_rect = 1;
3851  // Override skipping rectangular partition operations for edge blocks
3852  const int force_horz_split = (mi_row + ms >= cm->mi_rows);
3853  const int force_vert_split = (mi_col + ms >= cm->mi_cols);
3854  const int xss = x->e_mbd.plane[1].subsampling_x;
3855  const int yss = x->e_mbd.plane[1].subsampling_y;
3856
3857  int partition_none_allowed = !force_horz_split && !force_vert_split;
3858  int partition_horz_allowed =
3859      !force_vert_split && yss <= xss && bsize >= BLOCK_8X8;
3860  int partition_vert_allowed =
3861      !force_horz_split && xss <= yss && bsize >= BLOCK_8X8;
3862  (void)*tp_orig;
3863
3864  // Avoid checking for rectangular partitions for speed >= 6.
3865  if (cpi->oxcf.speed >= 6) do_rect = 0;
3866
3867  assert(num_8x8_blocks_wide_lookup[bsize] ==
3868         num_8x8_blocks_high_lookup[bsize]);
3869
3870  vp9_rd_cost_init(&sum_rdc);
3871  vp9_rd_cost_reset(&best_rdc);
3872  best_rdc.rdcost = best_rd;
3873
3874  // Determine partition types in search according to the speed features.
3875  // The threshold set here has to be of square block size.
3876  if (sf->auto_min_max_partition_size) {
3877    partition_none_allowed &=
3878        (bsize <= x->max_partition_size && bsize >= x->min_partition_size);
3879    partition_horz_allowed &=
3880        ((bsize <= x->max_partition_size && bsize > x->min_partition_size) ||
3881         force_horz_split);
3882    partition_vert_allowed &=
3883        ((bsize <= x->max_partition_size && bsize > x->min_partition_size) ||
3884         force_vert_split);
3885    do_split &= bsize > x->min_partition_size;
3886  }
3887  if (sf->use_square_partition_only) {
3888    partition_horz_allowed &= force_horz_split;
3889    partition_vert_allowed &= force_vert_split;
3890  }
3891
3892  ctx->pred_pixel_ready =
3893      !(partition_vert_allowed || partition_horz_allowed || do_split);
3894
3895  // PARTITION_NONE
3896  if (partition_none_allowed) {
3897    nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &this_rdc, bsize,
3898                        ctx);
3899    ctx->mic = *xd->mi[0];
3900    ctx->mbmi_ext = *x->mbmi_ext;
3901    ctx->skip_txfm[0] = x->skip_txfm[0];
3902    ctx->skip = x->skip;
3903
3904    if (this_rdc.rate != INT_MAX) {
3905      int pl = partition_plane_context(xd, mi_row, mi_col, bsize);
3906      this_rdc.rate += cpi->partition_cost[pl][PARTITION_NONE];
3907      this_rdc.rdcost =
3908          RDCOST(x->rdmult, x->rddiv, this_rdc.rate, this_rdc.dist);
3909      if (this_rdc.rdcost < best_rdc.rdcost) {
3910        int64_t dist_breakout_thr = sf->partition_search_breakout_thr.dist;
3911        int64_t rate_breakout_thr = sf->partition_search_breakout_thr.rate;
3912
3913        dist_breakout_thr >>=
3914            8 - (b_width_log2_lookup[bsize] + b_height_log2_lookup[bsize]);
3915
3916        rate_breakout_thr *= num_pels_log2_lookup[bsize];
3917
3918        best_rdc = this_rdc;
3919        if (bsize >= BLOCK_8X8) pc_tree->partitioning = PARTITION_NONE;
3920
3921        if (!x->e_mbd.lossless && this_rdc.rate < rate_breakout_thr &&
3922            this_rdc.dist < dist_breakout_thr) {
3923          do_split = 0;
3924          do_rect = 0;
3925        }
3926      }
3927    }
3928  }
3929
3930  // store estimated motion vector
3931  store_pred_mv(x, ctx);
3932
3933  // PARTITION_SPLIT
3934  if (do_split) {
3935    int pl = partition_plane_context(xd, mi_row, mi_col, bsize);
3936    sum_rdc.rate += cpi->partition_cost[pl][PARTITION_SPLIT];
3937    sum_rdc.rdcost = RDCOST(x->rdmult, x->rddiv, sum_rdc.rate, sum_rdc.dist);
3938    subsize = get_subsize(bsize, PARTITION_SPLIT);
3939    for (i = 0; i < 4 && sum_rdc.rdcost < best_rdc.rdcost; ++i) {
3940      const int x_idx = (i & 1) * ms;
3941      const int y_idx = (i >> 1) * ms;
3942
3943      if (mi_row + y_idx >= cm->mi_rows || mi_col + x_idx >= cm->mi_cols)
3944        continue;
3945      load_pred_mv(x, ctx);
3946      nonrd_pick_partition(cpi, td, tile_data, tp, mi_row + y_idx,
3947                           mi_col + x_idx, subsize, &this_rdc, 0,
3948                           best_rdc.rdcost - sum_rdc.rdcost, pc_tree->split[i]);
3949
3950      if (this_rdc.rate == INT_MAX) {
3951        vp9_rd_cost_reset(&sum_rdc);
3952      } else {
3953        sum_rdc.rate += this_rdc.rate;
3954        sum_rdc.dist += this_rdc.dist;
3955        sum_rdc.rdcost += this_rdc.rdcost;
3956      }
3957    }
3958
3959    if (sum_rdc.rdcost < best_rdc.rdcost) {
3960      best_rdc = sum_rdc;
3961      pc_tree->partitioning = PARTITION_SPLIT;
3962    } else {
3963      // skip rectangular partition test when larger block size
3964      // gives better rd cost
3965      if (sf->less_rectangular_check) do_rect &= !partition_none_allowed;
3966    }
3967  }
3968
3969  // PARTITION_HORZ
3970  if (partition_horz_allowed && do_rect) {
3971    subsize = get_subsize(bsize, PARTITION_HORZ);
3972    if (sf->adaptive_motion_search) load_pred_mv(x, ctx);
3973    pc_tree->horizontal[0].pred_pixel_ready = 1;
3974    nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &sum_rdc, subsize,
3975                        &pc_tree->horizontal[0]);
3976
3977    pc_tree->horizontal[0].mic = *xd->mi[0];
3978    pc_tree->horizontal[0].mbmi_ext = *x->mbmi_ext;
3979    pc_tree->horizontal[0].skip_txfm[0] = x->skip_txfm[0];
3980    pc_tree->horizontal[0].skip = x->skip;
3981
3982    if (sum_rdc.rdcost < best_rdc.rdcost && mi_row + ms < cm->mi_rows) {
3983      load_pred_mv(x, ctx);
3984      pc_tree->horizontal[1].pred_pixel_ready = 1;
3985      nonrd_pick_sb_modes(cpi, tile_data, x, mi_row + ms, mi_col, &this_rdc,
3986                          subsize, &pc_tree->horizontal[1]);
3987
3988      pc_tree->horizontal[1].mic = *xd->mi[0];
3989      pc_tree->horizontal[1].mbmi_ext = *x->mbmi_ext;
3990      pc_tree->horizontal[1].skip_txfm[0] = x->skip_txfm[0];
3991      pc_tree->horizontal[1].skip = x->skip;
3992
3993      if (this_rdc.rate == INT_MAX) {
3994        vp9_rd_cost_reset(&sum_rdc);
3995      } else {
3996        int pl = partition_plane_context(xd, mi_row, mi_col, bsize);
3997        this_rdc.rate += cpi->partition_cost[pl][PARTITION_HORZ];
3998        sum_rdc.rate += this_rdc.rate;
3999        sum_rdc.dist += this_rdc.dist;
4000        sum_rdc.rdcost =
4001            RDCOST(x->rdmult, x->rddiv, sum_rdc.rate, sum_rdc.dist);
4002      }
4003    }
4004
4005    if (sum_rdc.rdcost < best_rdc.rdcost) {
4006      best_rdc = sum_rdc;
4007      pc_tree->partitioning = PARTITION_HORZ;
4008    } else {
4009      pred_pixel_ready_reset(pc_tree, bsize);
4010    }
4011  }
4012
4013  // PARTITION_VERT
4014  if (partition_vert_allowed && do_rect) {
4015    subsize = get_subsize(bsize, PARTITION_VERT);
4016    if (sf->adaptive_motion_search) load_pred_mv(x, ctx);
4017    pc_tree->vertical[0].pred_pixel_ready = 1;
4018    nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &sum_rdc, subsize,
4019                        &pc_tree->vertical[0]);
4020    pc_tree->vertical[0].mic = *xd->mi[0];
4021    pc_tree->vertical[0].mbmi_ext = *x->mbmi_ext;
4022    pc_tree->vertical[0].skip_txfm[0] = x->skip_txfm[0];
4023    pc_tree->vertical[0].skip = x->skip;
4024
4025    if (sum_rdc.rdcost < best_rdc.rdcost && mi_col + ms < cm->mi_cols) {
4026      load_pred_mv(x, ctx);
4027      pc_tree->vertical[1].pred_pixel_ready = 1;
4028      nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col + ms, &this_rdc,
4029                          subsize, &pc_tree->vertical[1]);
4030      pc_tree->vertical[1].mic = *xd->mi[0];
4031      pc_tree->vertical[1].mbmi_ext = *x->mbmi_ext;
4032      pc_tree->vertical[1].skip_txfm[0] = x->skip_txfm[0];
4033      pc_tree->vertical[1].skip = x->skip;
4034
4035      if (this_rdc.rate == INT_MAX) {
4036        vp9_rd_cost_reset(&sum_rdc);
4037      } else {
4038        int pl = partition_plane_context(xd, mi_row, mi_col, bsize);
4039        sum_rdc.rate += cpi->partition_cost[pl][PARTITION_VERT];
4040        sum_rdc.rate += this_rdc.rate;
4041        sum_rdc.dist += this_rdc.dist;
4042        sum_rdc.rdcost =
4043            RDCOST(x->rdmult, x->rddiv, sum_rdc.rate, sum_rdc.dist);
4044      }
4045    }
4046
4047    if (sum_rdc.rdcost < best_rdc.rdcost) {
4048      best_rdc = sum_rdc;
4049      pc_tree->partitioning = PARTITION_VERT;
4050    } else {
4051      pred_pixel_ready_reset(pc_tree, bsize);
4052    }
4053  }
4054
4055  *rd_cost = best_rdc;
4056
4057  if (best_rdc.rate == INT_MAX) {
4058    vp9_rd_cost_reset(rd_cost);
4059    return;
4060  }
4061
4062  // update mode info array
4063  fill_mode_info_sb(cm, x, mi_row, mi_col, bsize, pc_tree);
4064
4065  if (best_rdc.rate < INT_MAX && best_rdc.dist < INT64_MAX && do_recon) {
4066    int output_enabled = (bsize == BLOCK_64X64);
4067    encode_sb_rt(cpi, td, tile_info, tp, mi_row, mi_col, output_enabled, bsize,
4068                 pc_tree);
4069  }
4070
4071  if (bsize == BLOCK_64X64 && do_recon) {
4072    assert(tp_orig < *tp);
4073    assert(best_rdc.rate < INT_MAX);
4074    assert(best_rdc.dist < INT64_MAX);
4075  } else {
4076    assert(tp_orig == *tp);
4077  }
4078}
4079
4080static void nonrd_select_partition(VP9_COMP *cpi, ThreadData *td,
4081                                   TileDataEnc *tile_data, MODE_INFO **mi,
4082                                   TOKENEXTRA **tp, int mi_row, int mi_col,
4083                                   BLOCK_SIZE bsize, int output_enabled,
4084                                   RD_COST *rd_cost, PC_TREE *pc_tree) {
4085  VP9_COMMON *const cm = &cpi->common;
4086  TileInfo *const tile_info = &tile_data->tile_info;
4087  MACROBLOCK *const x = &td->mb;
4088  MACROBLOCKD *const xd = &x->e_mbd;
4089  const int bsl = b_width_log2_lookup[bsize], hbs = (1 << bsl) / 4;
4090  const int mis = cm->mi_stride;
4091  PARTITION_TYPE partition;
4092  BLOCK_SIZE subsize;
4093  RD_COST this_rdc;
4094  BLOCK_SIZE subsize_ref =
4095      (cpi->sf.adapt_partition_source_sad) ? BLOCK_8X8 : BLOCK_16X16;
4096
4097  vp9_rd_cost_reset(&this_rdc);
4098  if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols) return;
4099
4100  subsize = (bsize >= BLOCK_8X8) ? mi[0]->sb_type : BLOCK_4X4;
4101  partition = partition_lookup[bsl][subsize];
4102
4103  if (bsize == BLOCK_32X32 && subsize == BLOCK_32X32) {
4104    x->max_partition_size = BLOCK_32X32;
4105    x->min_partition_size = BLOCK_16X16;
4106    nonrd_pick_partition(cpi, td, tile_data, tp, mi_row, mi_col, bsize, rd_cost,
4107                         0, INT64_MAX, pc_tree);
4108  } else if (bsize == BLOCK_32X32 && partition != PARTITION_NONE &&
4109             subsize >= subsize_ref) {
4110    x->max_partition_size = BLOCK_32X32;
4111    x->min_partition_size = BLOCK_8X8;
4112    nonrd_pick_partition(cpi, td, tile_data, tp, mi_row, mi_col, bsize, rd_cost,
4113                         0, INT64_MAX, pc_tree);
4114  } else if (bsize == BLOCK_16X16 && partition != PARTITION_NONE) {
4115    x->max_partition_size = BLOCK_16X16;
4116    x->min_partition_size = BLOCK_8X8;
4117    nonrd_pick_partition(cpi, td, tile_data, tp, mi_row, mi_col, bsize, rd_cost,
4118                         0, INT64_MAX, pc_tree);
4119  } else {
4120    switch (partition) {
4121      case PARTITION_NONE:
4122        pc_tree->none.pred_pixel_ready = 1;
4123        nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, rd_cost, subsize,
4124                            &pc_tree->none);
4125        pc_tree->none.mic = *xd->mi[0];
4126        pc_tree->none.mbmi_ext = *x->mbmi_ext;
4127        pc_tree->none.skip_txfm[0] = x->skip_txfm[0];
4128        pc_tree->none.skip = x->skip;
4129        break;
4130      case PARTITION_VERT:
4131        pc_tree->vertical[0].pred_pixel_ready = 1;
4132        nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, rd_cost, subsize,
4133                            &pc_tree->vertical[0]);
4134        pc_tree->vertical[0].mic = *xd->mi[0];
4135        pc_tree->vertical[0].mbmi_ext = *x->mbmi_ext;
4136        pc_tree->vertical[0].skip_txfm[0] = x->skip_txfm[0];
4137        pc_tree->vertical[0].skip = x->skip;
4138        if (mi_col + hbs < cm->mi_cols) {
4139          pc_tree->vertical[1].pred_pixel_ready = 1;
4140          nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col + hbs,
4141                              &this_rdc, subsize, &pc_tree->vertical[1]);
4142          pc_tree->vertical[1].mic = *xd->mi[0];
4143          pc_tree->vertical[1].mbmi_ext = *x->mbmi_ext;
4144          pc_tree->vertical[1].skip_txfm[0] = x->skip_txfm[0];
4145          pc_tree->vertical[1].skip = x->skip;
4146          if (this_rdc.rate != INT_MAX && this_rdc.dist != INT64_MAX &&
4147              rd_cost->rate != INT_MAX && rd_cost->dist != INT64_MAX) {
4148            rd_cost->rate += this_rdc.rate;
4149            rd_cost->dist += this_rdc.dist;
4150          }
4151        }
4152        break;
4153      case PARTITION_HORZ:
4154        pc_tree->horizontal[0].pred_pixel_ready = 1;
4155        nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, rd_cost, subsize,
4156                            &pc_tree->horizontal[0]);
4157        pc_tree->horizontal[0].mic = *xd->mi[0];
4158        pc_tree->horizontal[0].mbmi_ext = *x->mbmi_ext;
4159        pc_tree->horizontal[0].skip_txfm[0] = x->skip_txfm[0];
4160        pc_tree->horizontal[0].skip = x->skip;
4161        if (mi_row + hbs < cm->mi_rows) {
4162          pc_tree->horizontal[1].pred_pixel_ready = 1;
4163          nonrd_pick_sb_modes(cpi, tile_data, x, mi_row + hbs, mi_col,
4164                              &this_rdc, subsize, &pc_tree->horizontal[1]);
4165          pc_tree->horizontal[1].mic = *xd->mi[0];
4166          pc_tree->horizontal[1].mbmi_ext = *x->mbmi_ext;
4167          pc_tree->horizontal[1].skip_txfm[0] = x->skip_txfm[0];
4168          pc_tree->horizontal[1].skip = x->skip;
4169          if (this_rdc.rate != INT_MAX && this_rdc.dist != INT64_MAX &&
4170              rd_cost->rate != INT_MAX && rd_cost->dist != INT64_MAX) {
4171            rd_cost->rate += this_rdc.rate;
4172            rd_cost->dist += this_rdc.dist;
4173          }
4174        }
4175        break;
4176      case PARTITION_SPLIT:
4177        subsize = get_subsize(bsize, PARTITION_SPLIT);
4178        nonrd_select_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col,
4179                               subsize, output_enabled, rd_cost,
4180                               pc_tree->split[0]);
4181        nonrd_select_partition(cpi, td, tile_data, mi + hbs, tp, mi_row,
4182                               mi_col + hbs, subsize, output_enabled, &this_rdc,
4183                               pc_tree->split[1]);
4184        if (this_rdc.rate != INT_MAX && this_rdc.dist != INT64_MAX &&
4185            rd_cost->rate != INT_MAX && rd_cost->dist != INT64_MAX) {
4186          rd_cost->rate += this_rdc.rate;
4187          rd_cost->dist += this_rdc.dist;
4188        }
4189        nonrd_select_partition(cpi, td, tile_data, mi + hbs * mis, tp,
4190                               mi_row + hbs, mi_col, subsize, output_enabled,
4191                               &this_rdc, pc_tree->split[2]);
4192        if (this_rdc.rate != INT_MAX && this_rdc.dist != INT64_MAX &&
4193            rd_cost->rate != INT_MAX && rd_cost->dist != INT64_MAX) {
4194          rd_cost->rate += this_rdc.rate;
4195          rd_cost->dist += this_rdc.dist;
4196        }
4197        nonrd_select_partition(cpi, td, tile_data, mi + hbs * mis + hbs, tp,
4198                               mi_row + hbs, mi_col + hbs, subsize,
4199                               output_enabled, &this_rdc, pc_tree->split[3]);
4200        if (this_rdc.rate != INT_MAX && this_rdc.dist != INT64_MAX &&
4201            rd_cost->rate != INT_MAX && rd_cost->dist != INT64_MAX) {
4202          rd_cost->rate += this_rdc.rate;
4203          rd_cost->dist += this_rdc.dist;
4204        }
4205        break;
4206      default: assert(0 && "Invalid partition type."); break;
4207    }
4208  }
4209
4210  if (bsize == BLOCK_64X64 && output_enabled)
4211    encode_sb_rt(cpi, td, tile_info, tp, mi_row, mi_col, 1, bsize, pc_tree);
4212}
4213
4214static void nonrd_use_partition(VP9_COMP *cpi, ThreadData *td,
4215                                TileDataEnc *tile_data, MODE_INFO **mi,
4216                                TOKENEXTRA **tp, int mi_row, int mi_col,
4217                                BLOCK_SIZE bsize, int output_enabled,
4218                                RD_COST *dummy_cost, PC_TREE *pc_tree) {
4219  VP9_COMMON *const cm = &cpi->common;
4220  TileInfo *tile_info = &tile_data->tile_info;
4221  MACROBLOCK *const x = &td->mb;
4222  MACROBLOCKD *const xd = &x->e_mbd;
4223  const int bsl = b_width_log2_lookup[bsize], hbs = (1 << bsl) / 4;
4224  const int mis = cm->mi_stride;
4225  PARTITION_TYPE partition;
4226  BLOCK_SIZE subsize;
4227
4228  if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols) return;
4229
4230  subsize = (bsize >= BLOCK_8X8) ? mi[0]->sb_type : BLOCK_4X4;
4231  partition = partition_lookup[bsl][subsize];
4232
4233  if (output_enabled && bsize != BLOCK_4X4) {
4234    int ctx = partition_plane_context(xd, mi_row, mi_col, bsize);
4235    td->counts->partition[ctx][partition]++;
4236  }
4237
4238  switch (partition) {
4239    case PARTITION_NONE:
4240      pc_tree->none.pred_pixel_ready = 1;
4241      nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, dummy_cost,
4242                          subsize, &pc_tree->none);
4243      pc_tree->none.mic = *xd->mi[0];
4244      pc_tree->none.mbmi_ext = *x->mbmi_ext;
4245      pc_tree->none.skip_txfm[0] = x->skip_txfm[0];
4246      pc_tree->none.skip = x->skip;
4247      encode_b_rt(cpi, td, tile_info, tp, mi_row, mi_col, output_enabled,
4248                  subsize, &pc_tree->none);
4249      break;
4250    case PARTITION_VERT:
4251      pc_tree->vertical[0].pred_pixel_ready = 1;
4252      nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, dummy_cost,
4253                          subsize, &pc_tree->vertical[0]);
4254      pc_tree->vertical[0].mic = *xd->mi[0];
4255      pc_tree->vertical[0].mbmi_ext = *x->mbmi_ext;
4256      pc_tree->vertical[0].skip_txfm[0] = x->skip_txfm[0];
4257      pc_tree->vertical[0].skip = x->skip;
4258      encode_b_rt(cpi, td, tile_info, tp, mi_row, mi_col, output_enabled,
4259                  subsize, &pc_tree->vertical[0]);
4260      if (mi_col + hbs < cm->mi_cols && bsize > BLOCK_8X8) {
4261        pc_tree->vertical[1].pred_pixel_ready = 1;
4262        nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col + hbs, dummy_cost,
4263                            subsize, &pc_tree->vertical[1]);
4264        pc_tree->vertical[1].mic = *xd->mi[0];
4265        pc_tree->vertical[1].mbmi_ext = *x->mbmi_ext;
4266        pc_tree->vertical[1].skip_txfm[0] = x->skip_txfm[0];
4267        pc_tree->vertical[1].skip = x->skip;
4268        encode_b_rt(cpi, td, tile_info, tp, mi_row, mi_col + hbs,
4269                    output_enabled, subsize, &pc_tree->vertical[1]);
4270      }
4271      break;
4272    case PARTITION_HORZ:
4273      pc_tree->horizontal[0].pred_pixel_ready = 1;
4274      nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, dummy_cost,
4275                          subsize, &pc_tree->horizontal[0]);
4276      pc_tree->horizontal[0].mic = *xd->mi[0];
4277      pc_tree->horizontal[0].mbmi_ext = *x->mbmi_ext;
4278      pc_tree->horizontal[0].skip_txfm[0] = x->skip_txfm[0];
4279      pc_tree->horizontal[0].skip = x->skip;
4280      encode_b_rt(cpi, td, tile_info, tp, mi_row, mi_col, output_enabled,
4281                  subsize, &pc_tree->horizontal[0]);
4282
4283      if (mi_row + hbs < cm->mi_rows && bsize > BLOCK_8X8) {
4284        pc_tree->horizontal[1].pred_pixel_ready = 1;
4285        nonrd_pick_sb_modes(cpi, tile_data, x, mi_row + hbs, mi_col, dummy_cost,
4286                            subsize, &pc_tree->horizontal[1]);
4287        pc_tree->horizontal[1].mic = *xd->mi[0];
4288        pc_tree->horizontal[1].mbmi_ext = *x->mbmi_ext;
4289        pc_tree->horizontal[1].skip_txfm[0] = x->skip_txfm[0];
4290        pc_tree->horizontal[1].skip = x->skip;
4291        encode_b_rt(cpi, td, tile_info, tp, mi_row + hbs, mi_col,
4292                    output_enabled, subsize, &pc_tree->horizontal[1]);
4293      }
4294      break;
4295    case PARTITION_SPLIT:
4296      subsize = get_subsize(bsize, PARTITION_SPLIT);
4297      if (bsize == BLOCK_8X8) {
4298        nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, dummy_cost,
4299                            subsize, pc_tree->leaf_split[0]);
4300        encode_b_rt(cpi, td, tile_info, tp, mi_row, mi_col, output_enabled,
4301                    subsize, pc_tree->leaf_split[0]);
4302      } else {
4303        nonrd_use_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col, subsize,
4304                            output_enabled, dummy_cost, pc_tree->split[0]);
4305        nonrd_use_partition(cpi, td, tile_data, mi + hbs, tp, mi_row,
4306                            mi_col + hbs, subsize, output_enabled, dummy_cost,
4307                            pc_tree->split[1]);
4308        nonrd_use_partition(cpi, td, tile_data, mi + hbs * mis, tp,
4309                            mi_row + hbs, mi_col, subsize, output_enabled,
4310                            dummy_cost, pc_tree->split[2]);
4311        nonrd_use_partition(cpi, td, tile_data, mi + hbs * mis + hbs, tp,
4312                            mi_row + hbs, mi_col + hbs, subsize, output_enabled,
4313                            dummy_cost, pc_tree->split[3]);
4314      }
4315      break;
4316    default: assert(0 && "Invalid partition type."); break;
4317  }
4318
4319  if (partition != PARTITION_SPLIT || bsize == BLOCK_8X8)
4320    update_partition_context(xd, mi_row, mi_col, subsize, bsize);
4321}
4322
4323static void encode_nonrd_sb_row(VP9_COMP *cpi, ThreadData *td,
4324                                TileDataEnc *tile_data, int mi_row,
4325                                TOKENEXTRA **tp) {
4326  SPEED_FEATURES *const sf = &cpi->sf;
4327  VP9_COMMON *const cm = &cpi->common;
4328  TileInfo *const tile_info = &tile_data->tile_info;
4329  MACROBLOCK *const x = &td->mb;
4330  MACROBLOCKD *const xd = &x->e_mbd;
4331  const int mi_col_start = tile_info->mi_col_start;
4332  const int mi_col_end = tile_info->mi_col_end;
4333  int mi_col;
4334  const int sb_row = mi_row >> MI_BLOCK_SIZE_LOG2;
4335  const int num_sb_cols =
4336      get_num_cols(tile_data->tile_info, MI_BLOCK_SIZE_LOG2);
4337  int sb_col_in_tile;
4338
4339  // Initialize the left context for the new SB row
4340  memset(&xd->left_context, 0, sizeof(xd->left_context));
4341  memset(xd->left_seg_context, 0, sizeof(xd->left_seg_context));
4342
4343  // Code each SB in the row
4344  for (mi_col = mi_col_start, sb_col_in_tile = 0; mi_col < mi_col_end;
4345       mi_col += MI_BLOCK_SIZE, ++sb_col_in_tile) {
4346    const struct segmentation *const seg = &cm->seg;
4347    RD_COST dummy_rdc;
4348    const int idx_str = cm->mi_stride * mi_row + mi_col;
4349    MODE_INFO **mi = cm->mi_grid_visible + idx_str;
4350    PARTITION_SEARCH_TYPE partition_search_type = sf->partition_search_type;
4351    BLOCK_SIZE bsize = BLOCK_64X64;
4352    int seg_skip = 0;
4353
4354    (*(cpi->row_mt_sync_read_ptr))(&tile_data->row_mt_sync, sb_row,
4355                                   sb_col_in_tile);
4356
4357    if (cpi->use_skin_detection) {
4358      vp9_compute_skin_sb(cpi, BLOCK_16X16, mi_row, mi_col);
4359    }
4360
4361    x->source_variance = UINT_MAX;
4362    vp9_zero(x->pred_mv);
4363    vp9_rd_cost_init(&dummy_rdc);
4364    x->color_sensitivity[0] = 0;
4365    x->color_sensitivity[1] = 0;
4366    x->sb_is_skin = 0;
4367    x->skip_low_source_sad = 0;
4368    x->lowvar_highsumdiff = 0;
4369    x->content_state_sb = 0;
4370    x->sb_use_mv_part = 0;
4371    x->sb_mvcol_part = 0;
4372    x->sb_mvrow_part = 0;
4373    x->sb_pickmode_part = 0;
4374    x->arf_frame_usage = 0;
4375    x->lastgolden_frame_usage = 0;
4376
4377    if (seg->enabled) {
4378      const uint8_t *const map =
4379          seg->update_map ? cpi->segmentation_map : cm->last_frame_seg_map;
4380      int segment_id = get_segment_id(cm, map, BLOCK_64X64, mi_row, mi_col);
4381      seg_skip = segfeature_active(seg, segment_id, SEG_LVL_SKIP);
4382      if (seg_skip) {
4383        partition_search_type = FIXED_PARTITION;
4384      }
4385    }
4386
4387    if (cpi->compute_source_sad_onepass && cpi->sf.use_source_sad) {
4388      int shift = cpi->Source->y_stride * (mi_row << 3) + (mi_col << 3);
4389      int sb_offset2 = ((cm->mi_cols + 7) >> 3) * (mi_row >> 3) + (mi_col >> 3);
4390      int64_t source_sad = avg_source_sad(cpi, x, shift, sb_offset2);
4391      if (sf->adapt_partition_source_sad &&
4392          (cpi->oxcf.rc_mode == VPX_VBR && !cpi->rc.is_src_frame_alt_ref &&
4393           source_sad > sf->adapt_partition_thresh &&
4394           (cpi->refresh_golden_frame || cpi->refresh_alt_ref_frame)))
4395        partition_search_type = REFERENCE_PARTITION;
4396    }
4397
4398    // Set the partition type of the 64X64 block
4399    switch (partition_search_type) {
4400      case VAR_BASED_PARTITION:
4401        // TODO(jingning, marpan): The mode decision and encoding process
4402        // support both intra and inter sub8x8 block coding for RTC mode.
4403        // Tune the thresholds accordingly to use sub8x8 block coding for
4404        // coding performance improvement.
4405        choose_partitioning(cpi, tile_info, x, mi_row, mi_col);
4406        nonrd_use_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col,
4407                            BLOCK_64X64, 1, &dummy_rdc, td->pc_root);
4408        break;
4409      case SOURCE_VAR_BASED_PARTITION:
4410        set_source_var_based_partition(cpi, tile_info, x, mi, mi_row, mi_col);
4411        nonrd_use_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col,
4412                            BLOCK_64X64, 1, &dummy_rdc, td->pc_root);
4413        break;
4414      case FIXED_PARTITION:
4415        if (!seg_skip) bsize = sf->always_this_block_size;
4416        set_fixed_partitioning(cpi, tile_info, mi, mi_row, mi_col, bsize);
4417        nonrd_use_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col,
4418                            BLOCK_64X64, 1, &dummy_rdc, td->pc_root);
4419        break;
4420      case REFERENCE_PARTITION:
4421        x->sb_pickmode_part = 1;
4422        set_offsets(cpi, tile_info, x, mi_row, mi_col, BLOCK_64X64);
4423        // Use nonrd_pick_partition on scene-cut for VBR mode.
4424        // nonrd_pick_partition does not support 4x4 partition, so avoid it
4425        // on key frame for now.
4426        if ((cpi->oxcf.rc_mode == VPX_VBR && cpi->rc.high_source_sad &&
4427             cpi->oxcf.speed < 6 && cm->frame_type != KEY_FRAME &&
4428             (cpi->refresh_golden_frame || cpi->refresh_alt_ref_frame))) {
4429          // Use lower max_partition_size for low resoultions.
4430          if (cm->width <= 352 && cm->height <= 288)
4431            x->max_partition_size = BLOCK_32X32;
4432          else
4433            x->max_partition_size = BLOCK_64X64;
4434          x->min_partition_size = BLOCK_8X8;
4435          nonrd_pick_partition(cpi, td, tile_data, tp, mi_row, mi_col,
4436                               BLOCK_64X64, &dummy_rdc, 1, INT64_MAX,
4437                               td->pc_root);
4438        } else {
4439          choose_partitioning(cpi, tile_info, x, mi_row, mi_col);
4440          // TODO(marpan): Seems like nonrd_select_partition does not support
4441          // 4x4 partition. Since 4x4 is used on key frame, use this switch
4442          // for now.
4443          if (cm->frame_type == KEY_FRAME)
4444            nonrd_use_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col,
4445                                BLOCK_64X64, 1, &dummy_rdc, td->pc_root);
4446          else
4447            nonrd_select_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col,
4448                                   BLOCK_64X64, 1, &dummy_rdc, td->pc_root);
4449        }
4450
4451        break;
4452      default: assert(0); break;
4453    }
4454
4455    // Update ref_frame usage for inter frame if this group is ARF group.
4456    if (!cpi->rc.is_src_frame_alt_ref && !cpi->refresh_golden_frame &&
4457        !cpi->refresh_alt_ref_frame && cpi->rc.alt_ref_gf_group &&
4458        cpi->sf.use_altref_onepass) {
4459      int sboffset = ((cm->mi_cols + 7) >> 3) * (mi_row >> 3) + (mi_col >> 3);
4460      if (cpi->count_arf_frame_usage != NULL)
4461        cpi->count_arf_frame_usage[sboffset] = x->arf_frame_usage;
4462      if (cpi->count_lastgolden_frame_usage != NULL)
4463        cpi->count_lastgolden_frame_usage[sboffset] = x->lastgolden_frame_usage;
4464    }
4465
4466    (*(cpi->row_mt_sync_write_ptr))(&tile_data->row_mt_sync, sb_row,
4467                                    sb_col_in_tile, num_sb_cols);
4468  }
4469}
4470// end RTC play code
4471
4472static INLINE uint32_t variance(const diff *const d) {
4473  return d->sse - (uint32_t)(((int64_t)d->sum * d->sum) >> 8);
4474}
4475
4476#if CONFIG_VP9_HIGHBITDEPTH
4477static INLINE uint32_t variance_highbd(diff *const d) {
4478  const int64_t var = (int64_t)d->sse - (((int64_t)d->sum * d->sum) >> 8);
4479  return (var >= 0) ? (uint32_t)var : 0;
4480}
4481#endif  // CONFIG_VP9_HIGHBITDEPTH
4482
4483static int set_var_thresh_from_histogram(VP9_COMP *cpi) {
4484  const SPEED_FEATURES *const sf = &cpi->sf;
4485  const VP9_COMMON *const cm = &cpi->common;
4486
4487  const uint8_t *src = cpi->Source->y_buffer;
4488  const uint8_t *last_src = cpi->Last_Source->y_buffer;
4489  const int src_stride = cpi->Source->y_stride;
4490  const int last_stride = cpi->Last_Source->y_stride;
4491
4492  // Pick cutoff threshold
4493  const int cutoff = (VPXMIN(cm->width, cm->height) >= 720)
4494                         ? (cm->MBs * VAR_HIST_LARGE_CUT_OFF / 100)
4495                         : (cm->MBs * VAR_HIST_SMALL_CUT_OFF / 100);
4496  DECLARE_ALIGNED(16, int, hist[VAR_HIST_BINS]);
4497  diff *var16 = cpi->source_diff_var;
4498
4499  int sum = 0;
4500  int i, j;
4501
4502  memset(hist, 0, VAR_HIST_BINS * sizeof(hist[0]));
4503
4504  for (i = 0; i < cm->mb_rows; i++) {
4505    for (j = 0; j < cm->mb_cols; j++) {
4506#if CONFIG_VP9_HIGHBITDEPTH
4507      if (cm->use_highbitdepth) {
4508        switch (cm->bit_depth) {
4509          case VPX_BITS_8:
4510            vpx_highbd_8_get16x16var(src, src_stride, last_src, last_stride,
4511                                     &var16->sse, &var16->sum);
4512            var16->var = variance(var16);
4513            break;
4514          case VPX_BITS_10:
4515            vpx_highbd_10_get16x16var(src, src_stride, last_src, last_stride,
4516                                      &var16->sse, &var16->sum);
4517            var16->var = variance_highbd(var16);
4518            break;
4519          case VPX_BITS_12:
4520            vpx_highbd_12_get16x16var(src, src_stride, last_src, last_stride,
4521                                      &var16->sse, &var16->sum);
4522            var16->var = variance_highbd(var16);
4523            break;
4524          default:
4525            assert(0 &&
4526                   "cm->bit_depth should be VPX_BITS_8, VPX_BITS_10"
4527                   " or VPX_BITS_12");
4528            return -1;
4529        }
4530      } else {
4531        vpx_get16x16var(src, src_stride, last_src, last_stride, &var16->sse,
4532                        &var16->sum);
4533        var16->var = variance(var16);
4534      }
4535#else
4536      vpx_get16x16var(src, src_stride, last_src, last_stride, &var16->sse,
4537                      &var16->sum);
4538      var16->var = variance(var16);
4539#endif  // CONFIG_VP9_HIGHBITDEPTH
4540
4541      if (var16->var >= VAR_HIST_MAX_BG_VAR)
4542        hist[VAR_HIST_BINS - 1]++;
4543      else
4544        hist[var16->var / VAR_HIST_FACTOR]++;
4545
4546      src += 16;
4547      last_src += 16;
4548      var16++;
4549    }
4550
4551    src = src - cm->mb_cols * 16 + 16 * src_stride;
4552    last_src = last_src - cm->mb_cols * 16 + 16 * last_stride;
4553  }
4554
4555  cpi->source_var_thresh = 0;
4556
4557  if (hist[VAR_HIST_BINS - 1] < cutoff) {
4558    for (i = 0; i < VAR_HIST_BINS - 1; i++) {
4559      sum += hist[i];
4560
4561      if (sum > cutoff) {
4562        cpi->source_var_thresh = (i + 1) * VAR_HIST_FACTOR;
4563        return 0;
4564      }
4565    }
4566  }
4567
4568  return sf->search_type_check_frequency;
4569}
4570
4571static void source_var_based_partition_search_method(VP9_COMP *cpi) {
4572  VP9_COMMON *const cm = &cpi->common;
4573  SPEED_FEATURES *const sf = &cpi->sf;
4574
4575  if (cm->frame_type == KEY_FRAME) {
4576    // For key frame, use SEARCH_PARTITION.
4577    sf->partition_search_type = SEARCH_PARTITION;
4578  } else if (cm->intra_only) {
4579    sf->partition_search_type = FIXED_PARTITION;
4580  } else {
4581    if (cm->last_width != cm->width || cm->last_height != cm->height) {
4582      if (cpi->source_diff_var) vpx_free(cpi->source_diff_var);
4583
4584      CHECK_MEM_ERROR(cm, cpi->source_diff_var,
4585                      vpx_calloc(cm->MBs, sizeof(diff)));
4586    }
4587
4588    if (!cpi->frames_till_next_var_check)
4589      cpi->frames_till_next_var_check = set_var_thresh_from_histogram(cpi);
4590
4591    if (cpi->frames_till_next_var_check > 0) {
4592      sf->partition_search_type = FIXED_PARTITION;
4593      cpi->frames_till_next_var_check--;
4594    }
4595  }
4596}
4597
4598static int get_skip_encode_frame(const VP9_COMMON *cm, ThreadData *const td) {
4599  unsigned int intra_count = 0, inter_count = 0;
4600  int j;
4601
4602  for (j = 0; j < INTRA_INTER_CONTEXTS; ++j) {
4603    intra_count += td->counts->intra_inter[j][0];
4604    inter_count += td->counts->intra_inter[j][1];
4605  }
4606
4607  return (intra_count << 2) < inter_count && cm->frame_type != KEY_FRAME &&
4608         cm->show_frame;
4609}
4610
4611void vp9_init_tile_data(VP9_COMP *cpi) {
4612  VP9_COMMON *const cm = &cpi->common;
4613  const int tile_cols = 1 << cm->log2_tile_cols;
4614  const int tile_rows = 1 << cm->log2_tile_rows;
4615  int tile_col, tile_row;
4616  TOKENEXTRA *pre_tok = cpi->tile_tok[0][0];
4617  TOKENLIST *tplist = cpi->tplist[0][0];
4618  int tile_tok = 0;
4619  int tplist_count = 0;
4620
4621  if (cpi->tile_data == NULL || cpi->allocated_tiles < tile_cols * tile_rows) {
4622    if (cpi->tile_data != NULL) vpx_free(cpi->tile_data);
4623    CHECK_MEM_ERROR(cm, cpi->tile_data, vpx_malloc(tile_cols * tile_rows *
4624                                                   sizeof(*cpi->tile_data)));
4625    cpi->allocated_tiles = tile_cols * tile_rows;
4626
4627    for (tile_row = 0; tile_row < tile_rows; ++tile_row)
4628      for (tile_col = 0; tile_col < tile_cols; ++tile_col) {
4629        TileDataEnc *tile_data =
4630            &cpi->tile_data[tile_row * tile_cols + tile_col];
4631        int i, j;
4632        for (i = 0; i < BLOCK_SIZES; ++i) {
4633          for (j = 0; j < MAX_MODES; ++j) {
4634            tile_data->thresh_freq_fact[i][j] = RD_THRESH_INIT_FACT;
4635            tile_data->mode_map[i][j] = j;
4636          }
4637        }
4638#if CONFIG_MULTITHREAD
4639        tile_data->row_base_thresh_freq_fact = NULL;
4640#endif
4641      }
4642  }
4643
4644  for (tile_row = 0; tile_row < tile_rows; ++tile_row) {
4645    for (tile_col = 0; tile_col < tile_cols; ++tile_col) {
4646      TileDataEnc *this_tile = &cpi->tile_data[tile_row * tile_cols + tile_col];
4647      TileInfo *tile_info = &this_tile->tile_info;
4648      vp9_tile_init(tile_info, cm, tile_row, tile_col);
4649
4650      cpi->tile_tok[tile_row][tile_col] = pre_tok + tile_tok;
4651      pre_tok = cpi->tile_tok[tile_row][tile_col];
4652      tile_tok = allocated_tokens(*tile_info);
4653
4654      cpi->tplist[tile_row][tile_col] = tplist + tplist_count;
4655      tplist = cpi->tplist[tile_row][tile_col];
4656      tplist_count = get_num_vert_units(*tile_info, MI_BLOCK_SIZE_LOG2);
4657    }
4658  }
4659}
4660
4661void vp9_encode_sb_row(VP9_COMP *cpi, ThreadData *td, int tile_row,
4662                       int tile_col, int mi_row) {
4663  VP9_COMMON *const cm = &cpi->common;
4664  const int tile_cols = 1 << cm->log2_tile_cols;
4665  TileDataEnc *this_tile = &cpi->tile_data[tile_row * tile_cols + tile_col];
4666  const TileInfo *const tile_info = &this_tile->tile_info;
4667  TOKENEXTRA *tok = NULL;
4668  int tile_sb_row;
4669  int tile_mb_cols = (tile_info->mi_col_end - tile_info->mi_col_start + 1) >> 1;
4670
4671  tile_sb_row = mi_cols_aligned_to_sb(mi_row - tile_info->mi_row_start) >>
4672                MI_BLOCK_SIZE_LOG2;
4673  get_start_tok(cpi, tile_row, tile_col, mi_row, &tok);
4674  cpi->tplist[tile_row][tile_col][tile_sb_row].start = tok;
4675
4676  if (cpi->sf.use_nonrd_pick_mode)
4677    encode_nonrd_sb_row(cpi, td, this_tile, mi_row, &tok);
4678  else
4679    encode_rd_sb_row(cpi, td, this_tile, mi_row, &tok);
4680
4681  cpi->tplist[tile_row][tile_col][tile_sb_row].stop = tok;
4682  cpi->tplist[tile_row][tile_col][tile_sb_row].count =
4683      (unsigned int)(cpi->tplist[tile_row][tile_col][tile_sb_row].stop -
4684                     cpi->tplist[tile_row][tile_col][tile_sb_row].start);
4685  assert(tok - cpi->tplist[tile_row][tile_col][tile_sb_row].start <=
4686         get_token_alloc(MI_BLOCK_SIZE >> 1, tile_mb_cols));
4687
4688  (void)tile_mb_cols;
4689}
4690
4691void vp9_encode_tile(VP9_COMP *cpi, ThreadData *td, int tile_row,
4692                     int tile_col) {
4693  VP9_COMMON *const cm = &cpi->common;
4694  const int tile_cols = 1 << cm->log2_tile_cols;
4695  TileDataEnc *this_tile = &cpi->tile_data[tile_row * tile_cols + tile_col];
4696  const TileInfo *const tile_info = &this_tile->tile_info;
4697  const int mi_row_start = tile_info->mi_row_start;
4698  const int mi_row_end = tile_info->mi_row_end;
4699  int mi_row;
4700
4701  for (mi_row = mi_row_start; mi_row < mi_row_end; mi_row += MI_BLOCK_SIZE)
4702    vp9_encode_sb_row(cpi, td, tile_row, tile_col, mi_row);
4703}
4704
4705static void encode_tiles(VP9_COMP *cpi) {
4706  VP9_COMMON *const cm = &cpi->common;
4707  const int tile_cols = 1 << cm->log2_tile_cols;
4708  const int tile_rows = 1 << cm->log2_tile_rows;
4709  int tile_col, tile_row;
4710
4711  vp9_init_tile_data(cpi);
4712
4713  for (tile_row = 0; tile_row < tile_rows; ++tile_row)
4714    for (tile_col = 0; tile_col < tile_cols; ++tile_col)
4715      vp9_encode_tile(cpi, &cpi->td, tile_row, tile_col);
4716}
4717
4718#if CONFIG_FP_MB_STATS
4719static int input_fpmb_stats(FIRSTPASS_MB_STATS *firstpass_mb_stats,
4720                            VP9_COMMON *cm, uint8_t **this_frame_mb_stats) {
4721  uint8_t *mb_stats_in = firstpass_mb_stats->mb_stats_start +
4722                         cm->current_video_frame * cm->MBs * sizeof(uint8_t);
4723
4724  if (mb_stats_in > firstpass_mb_stats->mb_stats_end) return EOF;
4725
4726  *this_frame_mb_stats = mb_stats_in;
4727
4728  return 1;
4729}
4730#endif
4731
4732static void encode_frame_internal(VP9_COMP *cpi) {
4733  SPEED_FEATURES *const sf = &cpi->sf;
4734  ThreadData *const td = &cpi->td;
4735  MACROBLOCK *const x = &td->mb;
4736  VP9_COMMON *const cm = &cpi->common;
4737  MACROBLOCKD *const xd = &x->e_mbd;
4738
4739  xd->mi = cm->mi_grid_visible;
4740  xd->mi[0] = cm->mi;
4741
4742  vp9_zero(*td->counts);
4743  vp9_zero(cpi->td.rd_counts);
4744
4745  xd->lossless = cm->base_qindex == 0 && cm->y_dc_delta_q == 0 &&
4746                 cm->uv_dc_delta_q == 0 && cm->uv_ac_delta_q == 0;
4747
4748#if CONFIG_VP9_HIGHBITDEPTH
4749  if (cm->use_highbitdepth)
4750    x->fwd_txfm4x4 = xd->lossless ? vp9_highbd_fwht4x4 : vpx_highbd_fdct4x4;
4751  else
4752    x->fwd_txfm4x4 = xd->lossless ? vp9_fwht4x4 : vpx_fdct4x4;
4753  x->highbd_inv_txfm_add =
4754      xd->lossless ? vp9_highbd_iwht4x4_add : vp9_highbd_idct4x4_add;
4755#else
4756  x->fwd_txfm4x4 = xd->lossless ? vp9_fwht4x4 : vpx_fdct4x4;
4757#endif  // CONFIG_VP9_HIGHBITDEPTH
4758  x->inv_txfm_add = xd->lossless ? vp9_iwht4x4_add : vp9_idct4x4_add;
4759
4760  if (xd->lossless) x->optimize = 0;
4761
4762  cm->tx_mode = select_tx_mode(cpi, xd);
4763
4764  vp9_frame_init_quantizer(cpi);
4765
4766  vp9_initialize_rd_consts(cpi);
4767  vp9_initialize_me_consts(cpi, x, cm->base_qindex);
4768  init_encode_frame_mb_context(cpi);
4769  cm->use_prev_frame_mvs =
4770      !cm->error_resilient_mode && cm->width == cm->last_width &&
4771      cm->height == cm->last_height && !cm->intra_only && cm->last_show_frame;
4772  // Special case: set prev_mi to NULL when the previous mode info
4773  // context cannot be used.
4774  cm->prev_mi =
4775      cm->use_prev_frame_mvs ? cm->prev_mip + cm->mi_stride + 1 : NULL;
4776
4777  x->quant_fp = cpi->sf.use_quant_fp;
4778  vp9_zero(x->skip_txfm);
4779  if (sf->use_nonrd_pick_mode) {
4780    // Initialize internal buffer pointers for rtc coding, where non-RD
4781    // mode decision is used and hence no buffer pointer swap needed.
4782    int i;
4783    struct macroblock_plane *const p = x->plane;
4784    struct macroblockd_plane *const pd = xd->plane;
4785    PICK_MODE_CONTEXT *ctx = &cpi->td.pc_root->none;
4786
4787    for (i = 0; i < MAX_MB_PLANE; ++i) {
4788      p[i].coeff = ctx->coeff_pbuf[i][0];
4789      p[i].qcoeff = ctx->qcoeff_pbuf[i][0];
4790      pd[i].dqcoeff = ctx->dqcoeff_pbuf[i][0];
4791      p[i].eobs = ctx->eobs_pbuf[i][0];
4792    }
4793    vp9_zero(x->zcoeff_blk);
4794
4795    if (cm->frame_type != KEY_FRAME && cpi->rc.frames_since_golden == 0 &&
4796        !(cpi->oxcf.lag_in_frames > 0 && cpi->oxcf.rc_mode == VPX_VBR) &&
4797        !cpi->use_svc)
4798      cpi->ref_frame_flags &= (~VP9_GOLD_FLAG);
4799
4800    if (sf->partition_search_type == SOURCE_VAR_BASED_PARTITION)
4801      source_var_based_partition_search_method(cpi);
4802  }
4803
4804  {
4805    struct vpx_usec_timer emr_timer;
4806    vpx_usec_timer_start(&emr_timer);
4807
4808#if CONFIG_FP_MB_STATS
4809    if (cpi->use_fp_mb_stats) {
4810      input_fpmb_stats(&cpi->twopass.firstpass_mb_stats, cm,
4811                       &cpi->twopass.this_frame_mb_stats);
4812    }
4813#endif
4814
4815    if (!cpi->row_mt) {
4816      cpi->row_mt_sync_read_ptr = vp9_row_mt_sync_read_dummy;
4817      cpi->row_mt_sync_write_ptr = vp9_row_mt_sync_write_dummy;
4818      // If allowed, encoding tiles in parallel with one thread handling one
4819      // tile when row based multi-threading is disabled.
4820      if (VPXMIN(cpi->oxcf.max_threads, 1 << cm->log2_tile_cols) > 1)
4821        vp9_encode_tiles_mt(cpi);
4822      else
4823        encode_tiles(cpi);
4824    } else {
4825      cpi->row_mt_sync_read_ptr = vp9_row_mt_sync_read;
4826      cpi->row_mt_sync_write_ptr = vp9_row_mt_sync_write;
4827      vp9_encode_tiles_row_mt(cpi);
4828    }
4829
4830    vpx_usec_timer_mark(&emr_timer);
4831    cpi->time_encode_sb_row += vpx_usec_timer_elapsed(&emr_timer);
4832  }
4833
4834  sf->skip_encode_frame =
4835      sf->skip_encode_sb ? get_skip_encode_frame(cm, td) : 0;
4836
4837#if 0
4838  // Keep record of the total distortion this time around for future use
4839  cpi->last_frame_distortion = cpi->frame_distortion;
4840#endif
4841}
4842
4843static INTERP_FILTER get_interp_filter(
4844    const int64_t threshes[SWITCHABLE_FILTER_CONTEXTS], int is_alt_ref) {
4845  if (!is_alt_ref && threshes[EIGHTTAP_SMOOTH] > threshes[EIGHTTAP] &&
4846      threshes[EIGHTTAP_SMOOTH] > threshes[EIGHTTAP_SHARP] &&
4847      threshes[EIGHTTAP_SMOOTH] > threshes[SWITCHABLE - 1]) {
4848    return EIGHTTAP_SMOOTH;
4849  } else if (threshes[EIGHTTAP_SHARP] > threshes[EIGHTTAP] &&
4850             threshes[EIGHTTAP_SHARP] > threshes[SWITCHABLE - 1]) {
4851    return EIGHTTAP_SHARP;
4852  } else if (threshes[EIGHTTAP] > threshes[SWITCHABLE - 1]) {
4853    return EIGHTTAP;
4854  } else {
4855    return SWITCHABLE;
4856  }
4857}
4858
4859static int compute_frame_aq_offset(struct VP9_COMP *cpi) {
4860  VP9_COMMON *const cm = &cpi->common;
4861  MODE_INFO **mi_8x8_ptr = cm->mi_grid_visible;
4862  struct segmentation *const seg = &cm->seg;
4863
4864  int mi_row, mi_col;
4865  int sum_delta = 0;
4866  int map_index = 0;
4867  int qdelta_index;
4868  int segment_id;
4869
4870  for (mi_row = 0; mi_row < cm->mi_rows; mi_row++) {
4871    MODE_INFO **mi_8x8 = mi_8x8_ptr;
4872    for (mi_col = 0; mi_col < cm->mi_cols; mi_col++, mi_8x8++) {
4873      segment_id = mi_8x8[0]->segment_id;
4874      qdelta_index = get_segdata(seg, segment_id, SEG_LVL_ALT_Q);
4875      sum_delta += qdelta_index;
4876      map_index++;
4877    }
4878    mi_8x8_ptr += cm->mi_stride;
4879  }
4880
4881  return sum_delta / (cm->mi_rows * cm->mi_cols);
4882}
4883
4884void vp9_encode_frame(VP9_COMP *cpi) {
4885  VP9_COMMON *const cm = &cpi->common;
4886
4887  // In the longer term the encoder should be generalized to match the
4888  // decoder such that we allow compound where one of the 3 buffers has a
4889  // different sign bias and that buffer is then the fixed ref. However, this
4890  // requires further work in the rd loop. For now the only supported encoder
4891  // side behavior is where the ALT ref buffer has opposite sign bias to
4892  // the other two.
4893  if (!frame_is_intra_only(cm)) {
4894    if ((cm->ref_frame_sign_bias[ALTREF_FRAME] ==
4895         cm->ref_frame_sign_bias[GOLDEN_FRAME]) ||
4896        (cm->ref_frame_sign_bias[ALTREF_FRAME] ==
4897         cm->ref_frame_sign_bias[LAST_FRAME])) {
4898      cpi->allow_comp_inter_inter = 0;
4899    } else {
4900      cpi->allow_comp_inter_inter = 1;
4901      cm->comp_fixed_ref = ALTREF_FRAME;
4902      cm->comp_var_ref[0] = LAST_FRAME;
4903      cm->comp_var_ref[1] = GOLDEN_FRAME;
4904    }
4905  }
4906
4907  if (cpi->sf.frame_parameter_update) {
4908    int i;
4909    RD_OPT *const rd_opt = &cpi->rd;
4910    FRAME_COUNTS *counts = cpi->td.counts;
4911    RD_COUNTS *const rdc = &cpi->td.rd_counts;
4912
4913    // This code does a single RD pass over the whole frame assuming
4914    // either compound, single or hybrid prediction as per whatever has
4915    // worked best for that type of frame in the past.
4916    // It also predicts whether another coding mode would have worked
4917    // better than this coding mode. If that is the case, it remembers
4918    // that for subsequent frames.
4919    // It also does the same analysis for transform size selection.
4920    const MV_REFERENCE_FRAME frame_type = get_frame_type(cpi);
4921    int64_t *const mode_thrs = rd_opt->prediction_type_threshes[frame_type];
4922    int64_t *const filter_thrs = rd_opt->filter_threshes[frame_type];
4923    const int is_alt_ref = frame_type == ALTREF_FRAME;
4924
4925    /* prediction (compound, single or hybrid) mode selection */
4926    if (is_alt_ref || !cpi->allow_comp_inter_inter)
4927      cm->reference_mode = SINGLE_REFERENCE;
4928    else if (mode_thrs[COMPOUND_REFERENCE] > mode_thrs[SINGLE_REFERENCE] &&
4929             mode_thrs[COMPOUND_REFERENCE] > mode_thrs[REFERENCE_MODE_SELECT] &&
4930             check_dual_ref_flags(cpi) && cpi->static_mb_pct == 100)
4931      cm->reference_mode = COMPOUND_REFERENCE;
4932    else if (mode_thrs[SINGLE_REFERENCE] > mode_thrs[REFERENCE_MODE_SELECT])
4933      cm->reference_mode = SINGLE_REFERENCE;
4934    else
4935      cm->reference_mode = REFERENCE_MODE_SELECT;
4936
4937    if (cm->interp_filter == SWITCHABLE)
4938      cm->interp_filter = get_interp_filter(filter_thrs, is_alt_ref);
4939
4940    encode_frame_internal(cpi);
4941
4942    for (i = 0; i < REFERENCE_MODES; ++i)
4943      mode_thrs[i] = (mode_thrs[i] + rdc->comp_pred_diff[i] / cm->MBs) / 2;
4944
4945    for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i)
4946      filter_thrs[i] = (filter_thrs[i] + rdc->filter_diff[i] / cm->MBs) / 2;
4947
4948    if (cm->reference_mode == REFERENCE_MODE_SELECT) {
4949      int single_count_zero = 0;
4950      int comp_count_zero = 0;
4951
4952      for (i = 0; i < COMP_INTER_CONTEXTS; i++) {
4953        single_count_zero += counts->comp_inter[i][0];
4954        comp_count_zero += counts->comp_inter[i][1];
4955      }
4956
4957      if (comp_count_zero == 0) {
4958        cm->reference_mode = SINGLE_REFERENCE;
4959        vp9_zero(counts->comp_inter);
4960      } else if (single_count_zero == 0) {
4961        cm->reference_mode = COMPOUND_REFERENCE;
4962        vp9_zero(counts->comp_inter);
4963      }
4964    }
4965
4966    if (cm->tx_mode == TX_MODE_SELECT) {
4967      int count4x4 = 0;
4968      int count8x8_lp = 0, count8x8_8x8p = 0;
4969      int count16x16_16x16p = 0, count16x16_lp = 0;
4970      int count32x32 = 0;
4971
4972      for (i = 0; i < TX_SIZE_CONTEXTS; ++i) {
4973        count4x4 += counts->tx.p32x32[i][TX_4X4];
4974        count4x4 += counts->tx.p16x16[i][TX_4X4];
4975        count4x4 += counts->tx.p8x8[i][TX_4X4];
4976
4977        count8x8_lp += counts->tx.p32x32[i][TX_8X8];
4978        count8x8_lp += counts->tx.p16x16[i][TX_8X8];
4979        count8x8_8x8p += counts->tx.p8x8[i][TX_8X8];
4980
4981        count16x16_16x16p += counts->tx.p16x16[i][TX_16X16];
4982        count16x16_lp += counts->tx.p32x32[i][TX_16X16];
4983        count32x32 += counts->tx.p32x32[i][TX_32X32];
4984      }
4985      if (count4x4 == 0 && count16x16_lp == 0 && count16x16_16x16p == 0 &&
4986          count32x32 == 0) {
4987        cm->tx_mode = ALLOW_8X8;
4988        reset_skip_tx_size(cm, TX_8X8);
4989      } else if (count8x8_8x8p == 0 && count16x16_16x16p == 0 &&
4990                 count8x8_lp == 0 && count16x16_lp == 0 && count32x32 == 0) {
4991        cm->tx_mode = ONLY_4X4;
4992        reset_skip_tx_size(cm, TX_4X4);
4993      } else if (count8x8_lp == 0 && count16x16_lp == 0 && count4x4 == 0) {
4994        cm->tx_mode = ALLOW_32X32;
4995      } else if (count32x32 == 0 && count8x8_lp == 0 && count4x4 == 0) {
4996        cm->tx_mode = ALLOW_16X16;
4997        reset_skip_tx_size(cm, TX_16X16);
4998      }
4999    }
5000  } else {
5001    FRAME_COUNTS *counts = cpi->td.counts;
5002    cm->reference_mode = SINGLE_REFERENCE;
5003    if (cpi->allow_comp_inter_inter && cpi->sf.use_compound_nonrd_pickmode &&
5004        cpi->rc.alt_ref_gf_group && !cpi->rc.is_src_frame_alt_ref &&
5005        cm->frame_type != KEY_FRAME)
5006      cm->reference_mode = REFERENCE_MODE_SELECT;
5007
5008    encode_frame_internal(cpi);
5009
5010    if (cm->reference_mode == REFERENCE_MODE_SELECT) {
5011      int single_count_zero = 0;
5012      int comp_count_zero = 0;
5013      int i;
5014      for (i = 0; i < COMP_INTER_CONTEXTS; i++) {
5015        single_count_zero += counts->comp_inter[i][0];
5016        comp_count_zero += counts->comp_inter[i][1];
5017      }
5018      if (comp_count_zero == 0) {
5019        cm->reference_mode = SINGLE_REFERENCE;
5020        vp9_zero(counts->comp_inter);
5021      } else if (single_count_zero == 0) {
5022        cm->reference_mode = COMPOUND_REFERENCE;
5023        vp9_zero(counts->comp_inter);
5024      }
5025    }
5026  }
5027
5028  // If segmented AQ is enabled compute the average AQ weighting.
5029  if (cm->seg.enabled && (cpi->oxcf.aq_mode != NO_AQ) &&
5030      (cm->seg.update_map || cm->seg.update_data)) {
5031    cm->seg.aq_av_offset = compute_frame_aq_offset(cpi);
5032  }
5033}
5034
5035static void sum_intra_stats(FRAME_COUNTS *counts, const MODE_INFO *mi) {
5036  const PREDICTION_MODE y_mode = mi->mode;
5037  const PREDICTION_MODE uv_mode = mi->uv_mode;
5038  const BLOCK_SIZE bsize = mi->sb_type;
5039
5040  if (bsize < BLOCK_8X8) {
5041    int idx, idy;
5042    const int num_4x4_w = num_4x4_blocks_wide_lookup[bsize];
5043    const int num_4x4_h = num_4x4_blocks_high_lookup[bsize];
5044    for (idy = 0; idy < 2; idy += num_4x4_h)
5045      for (idx = 0; idx < 2; idx += num_4x4_w)
5046        ++counts->y_mode[0][mi->bmi[idy * 2 + idx].as_mode];
5047  } else {
5048    ++counts->y_mode[size_group_lookup[bsize]][y_mode];
5049  }
5050
5051  ++counts->uv_mode[y_mode][uv_mode];
5052}
5053
5054static void update_zeromv_cnt(VP9_COMP *const cpi, const MODE_INFO *const mi,
5055                              int mi_row, int mi_col, BLOCK_SIZE bsize) {
5056  const VP9_COMMON *const cm = &cpi->common;
5057  MV mv = mi->mv[0].as_mv;
5058  const int bw = num_8x8_blocks_wide_lookup[bsize];
5059  const int bh = num_8x8_blocks_high_lookup[bsize];
5060  const int xmis = VPXMIN(cm->mi_cols - mi_col, bw);
5061  const int ymis = VPXMIN(cm->mi_rows - mi_row, bh);
5062  const int block_index = mi_row * cm->mi_cols + mi_col;
5063  int x, y;
5064  for (y = 0; y < ymis; y++)
5065    for (x = 0; x < xmis; x++) {
5066      int map_offset = block_index + y * cm->mi_cols + x;
5067      if (is_inter_block(mi) && mi->segment_id <= CR_SEGMENT_ID_BOOST2) {
5068        if (abs(mv.row) < 8 && abs(mv.col) < 8) {
5069          if (cpi->consec_zero_mv[map_offset] < 255)
5070            cpi->consec_zero_mv[map_offset]++;
5071        } else {
5072          cpi->consec_zero_mv[map_offset] = 0;
5073        }
5074      }
5075    }
5076}
5077
5078static void encode_superblock(VP9_COMP *cpi, ThreadData *td, TOKENEXTRA **t,
5079                              int output_enabled, int mi_row, int mi_col,
5080                              BLOCK_SIZE bsize, PICK_MODE_CONTEXT *ctx) {
5081  VP9_COMMON *const cm = &cpi->common;
5082  MACROBLOCK *const x = &td->mb;
5083  MACROBLOCKD *const xd = &x->e_mbd;
5084  MODE_INFO *mi = xd->mi[0];
5085  const int seg_skip =
5086      segfeature_active(&cm->seg, mi->segment_id, SEG_LVL_SKIP);
5087  x->skip_recode = !x->select_tx_size && mi->sb_type >= BLOCK_8X8 &&
5088                   cpi->oxcf.aq_mode != COMPLEXITY_AQ &&
5089                   cpi->oxcf.aq_mode != CYCLIC_REFRESH_AQ &&
5090                   cpi->sf.allow_skip_recode;
5091
5092  if (!x->skip_recode && !cpi->sf.use_nonrd_pick_mode)
5093    memset(x->skip_txfm, 0, sizeof(x->skip_txfm));
5094
5095  x->skip_optimize = ctx->is_coded;
5096  ctx->is_coded = 1;
5097  x->use_lp32x32fdct = cpi->sf.use_lp32x32fdct;
5098  x->skip_encode = (!output_enabled && cpi->sf.skip_encode_frame &&
5099                    x->q_index < QIDX_SKIP_THRESH);
5100
5101  if (x->skip_encode) return;
5102
5103  if (!is_inter_block(mi)) {
5104    int plane;
5105#if CONFIG_BETTER_HW_COMPATIBILITY && CONFIG_VP9_HIGHBITDEPTH
5106    if ((xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) &&
5107        (xd->above_mi == NULL || xd->left_mi == NULL) &&
5108        need_top_left[mi->uv_mode])
5109      assert(0);
5110#endif  // CONFIG_BETTER_HW_COMPATIBILITY && CONFIG_VP9_HIGHBITDEPTH
5111    mi->skip = 1;
5112    for (plane = 0; plane < MAX_MB_PLANE; ++plane)
5113      vp9_encode_intra_block_plane(x, VPXMAX(bsize, BLOCK_8X8), plane, 1);
5114    if (output_enabled) sum_intra_stats(td->counts, mi);
5115    vp9_tokenize_sb(cpi, td, t, !output_enabled, seg_skip,
5116                    VPXMAX(bsize, BLOCK_8X8));
5117  } else {
5118    int ref;
5119    const int is_compound = has_second_ref(mi);
5120    set_ref_ptrs(cm, xd, mi->ref_frame[0], mi->ref_frame[1]);
5121    for (ref = 0; ref < 1 + is_compound; ++ref) {
5122      YV12_BUFFER_CONFIG *cfg = get_ref_frame_buffer(cpi, mi->ref_frame[ref]);
5123      assert(cfg != NULL);
5124      vp9_setup_pre_planes(xd, ref, cfg, mi_row, mi_col,
5125                           &xd->block_refs[ref]->sf);
5126    }
5127    if (!(cpi->sf.reuse_inter_pred_sby && ctx->pred_pixel_ready) || seg_skip)
5128      vp9_build_inter_predictors_sby(xd, mi_row, mi_col,
5129                                     VPXMAX(bsize, BLOCK_8X8));
5130
5131    vp9_build_inter_predictors_sbuv(xd, mi_row, mi_col,
5132                                    VPXMAX(bsize, BLOCK_8X8));
5133
5134    vp9_encode_sb(x, VPXMAX(bsize, BLOCK_8X8));
5135    vp9_tokenize_sb(cpi, td, t, !output_enabled, seg_skip,
5136                    VPXMAX(bsize, BLOCK_8X8));
5137  }
5138
5139  if (seg_skip) {
5140    assert(mi->skip);
5141  }
5142
5143  if (output_enabled) {
5144    if (cm->tx_mode == TX_MODE_SELECT && mi->sb_type >= BLOCK_8X8 &&
5145        !(is_inter_block(mi) && mi->skip)) {
5146      ++get_tx_counts(max_txsize_lookup[bsize], get_tx_size_context(xd),
5147                      &td->counts->tx)[mi->tx_size];
5148    } else {
5149      // The new intra coding scheme requires no change of transform size
5150      if (is_inter_block(mi)) {
5151        mi->tx_size = VPXMIN(tx_mode_to_biggest_tx_size[cm->tx_mode],
5152                             max_txsize_lookup[bsize]);
5153      } else {
5154        mi->tx_size = (bsize >= BLOCK_8X8) ? mi->tx_size : TX_4X4;
5155      }
5156    }
5157
5158    ++td->counts->tx.tx_totals[mi->tx_size];
5159    ++td->counts->tx.tx_totals[get_uv_tx_size(mi, &xd->plane[1])];
5160    if (cm->seg.enabled && cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ)
5161      vp9_cyclic_refresh_update_sb_postencode(cpi, mi, mi_row, mi_col, bsize);
5162    if (cpi->oxcf.pass == 0 && cpi->svc.temporal_layer_id == 0)
5163      update_zeromv_cnt(cpi, mi, mi_row, mi_col, bsize);
5164  }
5165}
5166