xref: /aosp_15_r20/external/libaom/av1/encoder/var_based_part.c (revision 77c1e3ccc04c968bd2bc212e87364f250e820521)
1 /*
2  * Copyright (c) 2019, Alliance for Open Media. All rights reserved.
3  *
4  * This source code is subject to the terms of the BSD 2 Clause License and
5  * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
6  * was not distributed with this source code in the LICENSE file, you can
7  * obtain it at www.aomedia.org/license/software. If the Alliance for Open
8  * Media Patent License 1.0 was not distributed with this source code in the
9  * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
10  */
11 
12 #include <limits.h>
13 #include <math.h>
14 #include <stdbool.h>
15 #include <stdio.h>
16 
17 #include "config/aom_config.h"
18 #include "config/aom_dsp_rtcd.h"
19 #include "config/av1_rtcd.h"
20 
21 #include "aom_dsp/aom_dsp_common.h"
22 #include "aom_dsp/binary_codes_writer.h"
23 #include "aom_ports/mem.h"
24 #include "aom_ports/aom_timer.h"
25 
26 #include "av1/common/reconinter.h"
27 #include "av1/common/blockd.h"
28 
29 #include "av1/encoder/encodeframe.h"
30 #include "av1/encoder/encodeframe_utils.h"
31 #include "av1/encoder/var_based_part.h"
32 #include "av1/encoder/reconinter_enc.h"
33 #include "av1/encoder/rdopt_utils.h"
34 
35 // Possible values for the force_split variable while evaluating variance based
36 // partitioning.
37 enum {
38   // Evaluate all partition types
39   PART_EVAL_ALL = 0,
40   // Force PARTITION_SPLIT
41   PART_EVAL_ONLY_SPLIT = 1,
42   // Force PARTITION_NONE
43   PART_EVAL_ONLY_NONE = 2
44 } UENUM1BYTE(PART_EVAL_STATUS);
45 
46 typedef struct {
47   VPVariance *part_variances;
48   VPartVar *split[4];
49 } variance_node;
50 
tree_to_node(void * data,BLOCK_SIZE bsize,variance_node * node)51 static inline void tree_to_node(void *data, BLOCK_SIZE bsize,
52                                 variance_node *node) {
53   node->part_variances = NULL;
54   switch (bsize) {
55     case BLOCK_128X128: {
56       VP128x128 *vt = (VP128x128 *)data;
57       node->part_variances = &vt->part_variances;
58       for (int split_idx = 0; split_idx < 4; split_idx++)
59         node->split[split_idx] = &vt->split[split_idx].part_variances.none;
60       break;
61     }
62     case BLOCK_64X64: {
63       VP64x64 *vt = (VP64x64 *)data;
64       node->part_variances = &vt->part_variances;
65       for (int split_idx = 0; split_idx < 4; split_idx++)
66         node->split[split_idx] = &vt->split[split_idx].part_variances.none;
67       break;
68     }
69     case BLOCK_32X32: {
70       VP32x32 *vt = (VP32x32 *)data;
71       node->part_variances = &vt->part_variances;
72       for (int split_idx = 0; split_idx < 4; split_idx++)
73         node->split[split_idx] = &vt->split[split_idx].part_variances.none;
74       break;
75     }
76     case BLOCK_16X16: {
77       VP16x16 *vt = (VP16x16 *)data;
78       node->part_variances = &vt->part_variances;
79       for (int split_idx = 0; split_idx < 4; split_idx++)
80         node->split[split_idx] = &vt->split[split_idx].part_variances.none;
81       break;
82     }
83     case BLOCK_8X8: {
84       VP8x8 *vt = (VP8x8 *)data;
85       node->part_variances = &vt->part_variances;
86       for (int split_idx = 0; split_idx < 4; split_idx++)
87         node->split[split_idx] = &vt->split[split_idx].part_variances.none;
88       break;
89     }
90     default: {
91       VP4x4 *vt = (VP4x4 *)data;
92       assert(bsize == BLOCK_4X4);
93       node->part_variances = &vt->part_variances;
94       for (int split_idx = 0; split_idx < 4; split_idx++)
95         node->split[split_idx] = &vt->split[split_idx];
96       break;
97     }
98   }
99 }
100 
101 // Set variance values given sum square error, sum error, count.
fill_variance(uint32_t s2,int32_t s,int c,VPartVar * v)102 static inline void fill_variance(uint32_t s2, int32_t s, int c, VPartVar *v) {
103   v->sum_square_error = s2;
104   v->sum_error = s;
105   v->log2_count = c;
106 }
107 
get_variance(VPartVar * v)108 static inline void get_variance(VPartVar *v) {
109   v->variance =
110       (int)(256 * (v->sum_square_error -
111                    (uint32_t)(((int64_t)v->sum_error * v->sum_error) >>
112                               v->log2_count)) >>
113             v->log2_count);
114 }
115 
sum_2_variances(const VPartVar * a,const VPartVar * b,VPartVar * r)116 static inline void sum_2_variances(const VPartVar *a, const VPartVar *b,
117                                    VPartVar *r) {
118   assert(a->log2_count == b->log2_count);
119   fill_variance(a->sum_square_error + b->sum_square_error,
120                 a->sum_error + b->sum_error, a->log2_count + 1, r);
121 }
122 
fill_variance_tree(void * data,BLOCK_SIZE bsize)123 static inline void fill_variance_tree(void *data, BLOCK_SIZE bsize) {
124   variance_node node;
125   memset(&node, 0, sizeof(node));
126   tree_to_node(data, bsize, &node);
127   sum_2_variances(node.split[0], node.split[1], &node.part_variances->horz[0]);
128   sum_2_variances(node.split[2], node.split[3], &node.part_variances->horz[1]);
129   sum_2_variances(node.split[0], node.split[2], &node.part_variances->vert[0]);
130   sum_2_variances(node.split[1], node.split[3], &node.part_variances->vert[1]);
131   sum_2_variances(&node.part_variances->vert[0], &node.part_variances->vert[1],
132                   &node.part_variances->none);
133 }
134 
set_block_size(AV1_COMP * const cpi,int mi_row,int mi_col,BLOCK_SIZE bsize)135 static inline void set_block_size(AV1_COMP *const cpi, int mi_row, int mi_col,
136                                   BLOCK_SIZE bsize) {
137   if (cpi->common.mi_params.mi_cols > mi_col &&
138       cpi->common.mi_params.mi_rows > mi_row) {
139     CommonModeInfoParams *mi_params = &cpi->common.mi_params;
140     const int mi_grid_idx = get_mi_grid_idx(mi_params, mi_row, mi_col);
141     const int mi_alloc_idx = get_alloc_mi_idx(mi_params, mi_row, mi_col);
142     MB_MODE_INFO *mi = mi_params->mi_grid_base[mi_grid_idx] =
143         &mi_params->mi_alloc[mi_alloc_idx];
144     mi->bsize = bsize;
145   }
146 }
147 
set_vt_partitioning(AV1_COMP * cpi,MACROBLOCKD * const xd,const TileInfo * const tile,void * data,BLOCK_SIZE bsize,int mi_row,int mi_col,int64_t threshold,BLOCK_SIZE bsize_min,PART_EVAL_STATUS force_split)148 static int set_vt_partitioning(AV1_COMP *cpi, MACROBLOCKD *const xd,
149                                const TileInfo *const tile, void *data,
150                                BLOCK_SIZE bsize, int mi_row, int mi_col,
151                                int64_t threshold, BLOCK_SIZE bsize_min,
152                                PART_EVAL_STATUS force_split) {
153   AV1_COMMON *const cm = &cpi->common;
154   variance_node vt;
155   const int block_width = mi_size_wide[bsize];
156   const int block_height = mi_size_high[bsize];
157   int bs_width_check = block_width;
158   int bs_height_check = block_height;
159   int bs_width_vert_check = block_width >> 1;
160   int bs_height_horiz_check = block_height >> 1;
161   // On the right and bottom boundary we only need to check
162   // if half the bsize fits, because boundary is extended
163   // up to 64. So do this check only for sb_size = 64X64.
164   if (cm->seq_params->sb_size == BLOCK_64X64) {
165     if (tile->mi_col_end == cm->mi_params.mi_cols) {
166       bs_width_check = (block_width >> 1) + 1;
167       bs_width_vert_check = (block_width >> 2) + 1;
168     }
169     if (tile->mi_row_end == cm->mi_params.mi_rows) {
170       bs_height_check = (block_height >> 1) + 1;
171       bs_height_horiz_check = (block_height >> 2) + 1;
172     }
173   }
174 
175   assert(block_height == block_width);
176   tree_to_node(data, bsize, &vt);
177 
178   if (mi_col + bs_width_check <= tile->mi_col_end &&
179       mi_row + bs_height_check <= tile->mi_row_end &&
180       force_split == PART_EVAL_ONLY_NONE) {
181     set_block_size(cpi, mi_row, mi_col, bsize);
182     return 1;
183   }
184   if (force_split == PART_EVAL_ONLY_SPLIT) return 0;
185 
186   // For bsize=bsize_min (16x16/8x8 for 8x8/4x4 downsampling), select if
187   // variance is below threshold, otherwise split will be selected.
188   // No check for vert/horiz split as too few samples for variance.
189   if (bsize == bsize_min) {
190     // Variance already computed to set the force_split.
191     if (frame_is_intra_only(cm)) get_variance(&vt.part_variances->none);
192     if (mi_col + bs_width_check <= tile->mi_col_end &&
193         mi_row + bs_height_check <= tile->mi_row_end &&
194         vt.part_variances->none.variance < threshold) {
195       set_block_size(cpi, mi_row, mi_col, bsize);
196       return 1;
197     }
198     return 0;
199   } else if (bsize > bsize_min) {
200     // Variance already computed to set the force_split.
201     if (frame_is_intra_only(cm)) get_variance(&vt.part_variances->none);
202     // For key frame: take split for bsize above 32X32 or very high variance.
203     if (frame_is_intra_only(cm) &&
204         (bsize > BLOCK_32X32 ||
205          vt.part_variances->none.variance > (threshold << 4))) {
206       return 0;
207     }
208     // If variance is low, take the bsize (no split).
209     if (mi_col + bs_width_check <= tile->mi_col_end &&
210         mi_row + bs_height_check <= tile->mi_row_end &&
211         vt.part_variances->none.variance < threshold) {
212       set_block_size(cpi, mi_row, mi_col, bsize);
213       return 1;
214     }
215     // Check vertical split.
216     if (mi_row + bs_height_check <= tile->mi_row_end &&
217         mi_col + bs_width_vert_check <= tile->mi_col_end) {
218       BLOCK_SIZE subsize = get_partition_subsize(bsize, PARTITION_VERT);
219       BLOCK_SIZE plane_bsize =
220           get_plane_block_size(subsize, xd->plane[AOM_PLANE_U].subsampling_x,
221                                xd->plane[AOM_PLANE_U].subsampling_y);
222       get_variance(&vt.part_variances->vert[0]);
223       get_variance(&vt.part_variances->vert[1]);
224       if (vt.part_variances->vert[0].variance < threshold &&
225           vt.part_variances->vert[1].variance < threshold &&
226           plane_bsize < BLOCK_INVALID) {
227         set_block_size(cpi, mi_row, mi_col, subsize);
228         set_block_size(cpi, mi_row, mi_col + block_width / 2, subsize);
229         return 1;
230       }
231     }
232     // Check horizontal split.
233     if (mi_col + bs_width_check <= tile->mi_col_end &&
234         mi_row + bs_height_horiz_check <= tile->mi_row_end) {
235       BLOCK_SIZE subsize = get_partition_subsize(bsize, PARTITION_HORZ);
236       BLOCK_SIZE plane_bsize =
237           get_plane_block_size(subsize, xd->plane[AOM_PLANE_U].subsampling_x,
238                                xd->plane[AOM_PLANE_U].subsampling_y);
239       get_variance(&vt.part_variances->horz[0]);
240       get_variance(&vt.part_variances->horz[1]);
241       if (vt.part_variances->horz[0].variance < threshold &&
242           vt.part_variances->horz[1].variance < threshold &&
243           plane_bsize < BLOCK_INVALID) {
244         set_block_size(cpi, mi_row, mi_col, subsize);
245         set_block_size(cpi, mi_row + block_height / 2, mi_col, subsize);
246         return 1;
247       }
248     }
249     return 0;
250   }
251   return 0;
252 }
253 
all_blks_inside(int x16_idx,int y16_idx,int pixels_wide,int pixels_high)254 static inline int all_blks_inside(int x16_idx, int y16_idx, int pixels_wide,
255                                   int pixels_high) {
256   int all_inside = 1;
257   for (int idx = 0; idx < 4; idx++) {
258     all_inside &= ((x16_idx + GET_BLK_IDX_X(idx, 3)) < pixels_wide);
259     all_inside &= ((y16_idx + GET_BLK_IDX_Y(idx, 3)) < pixels_high);
260   }
261   return all_inside;
262 }
263 
264 #if CONFIG_AV1_HIGHBITDEPTH
265 // TODO(yunqingwang): Perform average of four 8x8 blocks similar to lowbd
fill_variance_8x8avg_highbd(const uint8_t * src_buf,int src_stride,const uint8_t * dst_buf,int dst_stride,int x16_idx,int y16_idx,VP16x16 * vst,int pixels_wide,int pixels_high)266 static inline void fill_variance_8x8avg_highbd(
267     const uint8_t *src_buf, int src_stride, const uint8_t *dst_buf,
268     int dst_stride, int x16_idx, int y16_idx, VP16x16 *vst, int pixels_wide,
269     int pixels_high) {
270   for (int idx = 0; idx < 4; idx++) {
271     const int x8_idx = x16_idx + GET_BLK_IDX_X(idx, 3);
272     const int y8_idx = y16_idx + GET_BLK_IDX_Y(idx, 3);
273     unsigned int sse = 0;
274     int sum = 0;
275     if (x8_idx < pixels_wide && y8_idx < pixels_high) {
276       int src_avg = aom_highbd_avg_8x8(src_buf + y8_idx * src_stride + x8_idx,
277                                        src_stride);
278       int dst_avg = aom_highbd_avg_8x8(dst_buf + y8_idx * dst_stride + x8_idx,
279                                        dst_stride);
280 
281       sum = src_avg - dst_avg;
282       sse = sum * sum;
283     }
284     fill_variance(sse, sum, 0, &vst->split[idx].part_variances.none);
285   }
286 }
287 #endif
288 
fill_variance_8x8avg_lowbd(const uint8_t * src_buf,int src_stride,const uint8_t * dst_buf,int dst_stride,int x16_idx,int y16_idx,VP16x16 * vst,int pixels_wide,int pixels_high)289 static inline void fill_variance_8x8avg_lowbd(
290     const uint8_t *src_buf, int src_stride, const uint8_t *dst_buf,
291     int dst_stride, int x16_idx, int y16_idx, VP16x16 *vst, int pixels_wide,
292     int pixels_high) {
293   unsigned int sse[4] = { 0 };
294   int sum[4] = { 0 };
295 
296   if (all_blks_inside(x16_idx, y16_idx, pixels_wide, pixels_high)) {
297     int src_avg[4];
298     int dst_avg[4];
299     aom_avg_8x8_quad(src_buf, src_stride, x16_idx, y16_idx, src_avg);
300     aom_avg_8x8_quad(dst_buf, dst_stride, x16_idx, y16_idx, dst_avg);
301     for (int idx = 0; idx < 4; idx++) {
302       sum[idx] = src_avg[idx] - dst_avg[idx];
303       sse[idx] = sum[idx] * sum[idx];
304     }
305   } else {
306     for (int idx = 0; idx < 4; idx++) {
307       const int x8_idx = x16_idx + GET_BLK_IDX_X(idx, 3);
308       const int y8_idx = y16_idx + GET_BLK_IDX_Y(idx, 3);
309       if (x8_idx < pixels_wide && y8_idx < pixels_high) {
310         int src_avg =
311             aom_avg_8x8(src_buf + y8_idx * src_stride + x8_idx, src_stride);
312         int dst_avg =
313             aom_avg_8x8(dst_buf + y8_idx * dst_stride + x8_idx, dst_stride);
314         sum[idx] = src_avg - dst_avg;
315         sse[idx] = sum[idx] * sum[idx];
316       }
317     }
318   }
319 
320   for (int idx = 0; idx < 4; idx++) {
321     fill_variance(sse[idx], sum[idx], 0, &vst->split[idx].part_variances.none);
322   }
323 }
324 
325 // Obtain parameters required to calculate variance (such as sum, sse, etc,.)
326 // at 8x8 sub-block level for a given 16x16 block.
327 // The function can be called only when is_key_frame is false since sum is
328 // computed between source and reference frames.
fill_variance_8x8avg(const uint8_t * src_buf,int src_stride,const uint8_t * dst_buf,int dst_stride,int x16_idx,int y16_idx,VP16x16 * vst,int highbd_flag,int pixels_wide,int pixels_high)329 static inline void fill_variance_8x8avg(const uint8_t *src_buf, int src_stride,
330                                         const uint8_t *dst_buf, int dst_stride,
331                                         int x16_idx, int y16_idx, VP16x16 *vst,
332                                         int highbd_flag, int pixels_wide,
333                                         int pixels_high) {
334 #if CONFIG_AV1_HIGHBITDEPTH
335   if (highbd_flag) {
336     fill_variance_8x8avg_highbd(src_buf, src_stride, dst_buf, dst_stride,
337                                 x16_idx, y16_idx, vst, pixels_wide,
338                                 pixels_high);
339     return;
340   }
341 #else
342   (void)highbd_flag;
343 #endif  // CONFIG_AV1_HIGHBITDEPTH
344   fill_variance_8x8avg_lowbd(src_buf, src_stride, dst_buf, dst_stride, x16_idx,
345                              y16_idx, vst, pixels_wide, pixels_high);
346 }
347 
compute_minmax_8x8(const uint8_t * src_buf,int src_stride,const uint8_t * dst_buf,int dst_stride,int x16_idx,int y16_idx,int highbd_flag,int pixels_wide,int pixels_high)348 static int compute_minmax_8x8(const uint8_t *src_buf, int src_stride,
349                               const uint8_t *dst_buf, int dst_stride,
350                               int x16_idx, int y16_idx,
351 #if CONFIG_AV1_HIGHBITDEPTH
352                               int highbd_flag,
353 #endif
354                               int pixels_wide, int pixels_high) {
355   int minmax_max = 0;
356   int minmax_min = 255;
357   // Loop over the 4 8x8 subblocks.
358   for (int idx = 0; idx < 4; idx++) {
359     const int x8_idx = x16_idx + GET_BLK_IDX_X(idx, 3);
360     const int y8_idx = y16_idx + GET_BLK_IDX_Y(idx, 3);
361     int min = 0;
362     int max = 0;
363     if (x8_idx < pixels_wide && y8_idx < pixels_high) {
364 #if CONFIG_AV1_HIGHBITDEPTH
365       if (highbd_flag & YV12_FLAG_HIGHBITDEPTH) {
366         aom_highbd_minmax_8x8(
367             src_buf + y8_idx * src_stride + x8_idx, src_stride,
368             dst_buf + y8_idx * dst_stride + x8_idx, dst_stride, &min, &max);
369       } else {
370         aom_minmax_8x8(src_buf + y8_idx * src_stride + x8_idx, src_stride,
371                        dst_buf + y8_idx * dst_stride + x8_idx, dst_stride, &min,
372                        &max);
373       }
374 #else
375       aom_minmax_8x8(src_buf + y8_idx * src_stride + x8_idx, src_stride,
376                      dst_buf + y8_idx * dst_stride + x8_idx, dst_stride, &min,
377                      &max);
378 #endif
379       if ((max - min) > minmax_max) minmax_max = (max - min);
380       if ((max - min) < minmax_min) minmax_min = (max - min);
381     }
382   }
383   return (minmax_max - minmax_min);
384 }
385 
386 // Function to compute average and variance of 4x4 sub-block.
387 // The function can be called only when is_key_frame is true since sum is
388 // computed using source frame only.
fill_variance_4x4avg(const uint8_t * src_buf,int src_stride,int x8_idx,int y8_idx,VP8x8 * vst,int highbd_flag,int pixels_wide,int pixels_high,int border_offset_4x4)389 static inline void fill_variance_4x4avg(const uint8_t *src_buf, int src_stride,
390                                         int x8_idx, int y8_idx, VP8x8 *vst,
391 #if CONFIG_AV1_HIGHBITDEPTH
392                                         int highbd_flag,
393 #endif
394                                         int pixels_wide, int pixels_high,
395                                         int border_offset_4x4) {
396   for (int idx = 0; idx < 4; idx++) {
397     const int x4_idx = x8_idx + GET_BLK_IDX_X(idx, 2);
398     const int y4_idx = y8_idx + GET_BLK_IDX_Y(idx, 2);
399     unsigned int sse = 0;
400     int sum = 0;
401     if (x4_idx < pixels_wide - border_offset_4x4 &&
402         y4_idx < pixels_high - border_offset_4x4) {
403       int src_avg;
404       int dst_avg = 128;
405 #if CONFIG_AV1_HIGHBITDEPTH
406       if (highbd_flag & YV12_FLAG_HIGHBITDEPTH) {
407         src_avg = aom_highbd_avg_4x4(src_buf + y4_idx * src_stride + x4_idx,
408                                      src_stride);
409       } else {
410         src_avg =
411             aom_avg_4x4(src_buf + y4_idx * src_stride + x4_idx, src_stride);
412       }
413 #else
414       src_avg = aom_avg_4x4(src_buf + y4_idx * src_stride + x4_idx, src_stride);
415 #endif
416 
417       sum = src_avg - dst_avg;
418       sse = sum * sum;
419     }
420     fill_variance(sse, sum, 0, &vst->split[idx].part_variances.none);
421   }
422 }
423 
scale_part_thresh_content(int64_t threshold_base,int speed,int non_reference_frame,int is_static)424 static int64_t scale_part_thresh_content(int64_t threshold_base, int speed,
425                                          int non_reference_frame,
426                                          int is_static) {
427   int64_t threshold = threshold_base;
428   if (non_reference_frame && !is_static) threshold = (3 * threshold) >> 1;
429   if (speed >= 8) {
430     return (5 * threshold) >> 2;
431   }
432   return threshold;
433 }
434 
435 // Tune thresholds less or more aggressively to prefer larger partitions
tune_thresh_based_on_qindex(AV1_COMP * cpi,int64_t thresholds[],uint64_t block_sad,int current_qindex,int num_pixels,bool is_segment_id_boosted,int source_sad_nonrd,int lighting_change)436 static inline void tune_thresh_based_on_qindex(
437     AV1_COMP *cpi, int64_t thresholds[], uint64_t block_sad, int current_qindex,
438     int num_pixels, bool is_segment_id_boosted, int source_sad_nonrd,
439     int lighting_change) {
440   double weight;
441   if (cpi->sf.rt_sf.prefer_large_partition_blocks >= 3) {
442     const int win = 20;
443     if (current_qindex < QINDEX_LARGE_BLOCK_THR - win)
444       weight = 1.0;
445     else if (current_qindex > QINDEX_LARGE_BLOCK_THR + win)
446       weight = 0.0;
447     else
448       weight =
449           1.0 - (current_qindex - QINDEX_LARGE_BLOCK_THR + win) / (2 * win);
450     if (num_pixels > RESOLUTION_480P) {
451       for (int i = 0; i < 4; i++) {
452         thresholds[i] <<= 1;
453       }
454     }
455     if (num_pixels <= RESOLUTION_288P) {
456       thresholds[3] = INT64_MAX;
457       if (is_segment_id_boosted == false) {
458         thresholds[1] <<= 2;
459         thresholds[2] <<= (source_sad_nonrd <= kLowSad) ? 5 : 4;
460       } else {
461         thresholds[1] <<= 1;
462         thresholds[2] <<= 3;
463       }
464       // Allow for split to 8x8 for superblocks where part of it has
465       // moving boundary. So allow for sb with source_sad above threshold,
466       // and avoid very large source_sad or high source content, to avoid
467       // too many 8x8 within superblock.
468       uint64_t avg_source_sad_thresh = 25000;
469       uint64_t block_sad_low = 25000;
470       uint64_t block_sad_high = 50000;
471       if (cpi->svc.temporal_layer_id == 0 &&
472           cpi->svc.number_temporal_layers > 1) {
473         // Increase the sad thresholds for base TL0, as reference/LAST is
474         // 2/4 frames behind (for 2/3 #TL).
475         avg_source_sad_thresh = 40000;
476         block_sad_high = 70000;
477       }
478       if (is_segment_id_boosted == false &&
479           cpi->rc.avg_source_sad < avg_source_sad_thresh &&
480           block_sad > block_sad_low && block_sad < block_sad_high &&
481           !lighting_change) {
482         thresholds[2] = (3 * thresholds[2]) >> 2;
483         thresholds[3] = thresholds[2] << 3;
484       }
485       // Condition the increase of partition thresholds on the segment
486       // and the content. Avoid the increase for superblocks which have
487       // high source sad, unless the whole frame has very high motion
488       // (i.e, cpi->rc.avg_source_sad is very large, in which case all blocks
489       // have high source sad).
490     } else if (num_pixels > RESOLUTION_480P && is_segment_id_boosted == false &&
491                (source_sad_nonrd != kHighSad ||
492                 cpi->rc.avg_source_sad > 50000)) {
493       thresholds[0] = (3 * thresholds[0]) >> 1;
494       thresholds[3] = INT64_MAX;
495       if (current_qindex > QINDEX_LARGE_BLOCK_THR) {
496         thresholds[1] =
497             (int)((1 - weight) * (thresholds[1] << 1) + weight * thresholds[1]);
498         thresholds[2] =
499             (int)((1 - weight) * (thresholds[2] << 1) + weight * thresholds[2]);
500       }
501     } else if (current_qindex > QINDEX_LARGE_BLOCK_THR &&
502                is_segment_id_boosted == false &&
503                (source_sad_nonrd != kHighSad ||
504                 cpi->rc.avg_source_sad > 50000)) {
505       thresholds[1] =
506           (int)((1 - weight) * (thresholds[1] << 2) + weight * thresholds[1]);
507       thresholds[2] =
508           (int)((1 - weight) * (thresholds[2] << 4) + weight * thresholds[2]);
509       thresholds[3] = INT64_MAX;
510     }
511   } else if (cpi->sf.rt_sf.prefer_large_partition_blocks >= 2) {
512     thresholds[1] <<= (source_sad_nonrd <= kLowSad) ? 2 : 0;
513     thresholds[2] =
514         (source_sad_nonrd <= kLowSad) ? (3 * thresholds[2]) : thresholds[2];
515   } else if (cpi->sf.rt_sf.prefer_large_partition_blocks >= 1) {
516     const int fac = (source_sad_nonrd <= kLowSad) ? 2 : 1;
517     if (current_qindex < QINDEX_LARGE_BLOCK_THR - 45)
518       weight = 1.0;
519     else if (current_qindex > QINDEX_LARGE_BLOCK_THR + 45)
520       weight = 0.0;
521     else
522       weight = 1.0 - (current_qindex - QINDEX_LARGE_BLOCK_THR + 45) / (2 * 45);
523     thresholds[1] =
524         (int)((1 - weight) * (thresholds[1] << 1) + weight * thresholds[1]);
525     thresholds[2] =
526         (int)((1 - weight) * (thresholds[2] << 1) + weight * thresholds[2]);
527     thresholds[3] =
528         (int)((1 - weight) * (thresholds[3] << fac) + weight * thresholds[3]);
529   }
530   if (cpi->sf.part_sf.disable_8x8_part_based_on_qidx && (current_qindex < 128))
531     thresholds[3] = INT64_MAX;
532 }
533 
set_vbp_thresholds_key_frame(AV1_COMP * cpi,int64_t thresholds[],int64_t threshold_base,int threshold_left_shift,int num_pixels)534 static void set_vbp_thresholds_key_frame(AV1_COMP *cpi, int64_t thresholds[],
535                                          int64_t threshold_base,
536                                          int threshold_left_shift,
537                                          int num_pixels) {
538   if (cpi->sf.rt_sf.force_large_partition_blocks_intra) {
539     const int shift_steps =
540         threshold_left_shift - (cpi->oxcf.mode == ALLINTRA ? 7 : 8);
541     assert(shift_steps >= 0);
542     threshold_base <<= shift_steps;
543   }
544   thresholds[0] = threshold_base;
545   thresholds[1] = threshold_base;
546   if (num_pixels < RESOLUTION_720P) {
547     thresholds[2] = threshold_base / 3;
548     thresholds[3] = threshold_base >> 1;
549   } else {
550     int shift_val = 2;
551     if (cpi->sf.rt_sf.force_large_partition_blocks_intra) {
552       shift_val = 0;
553     }
554 
555     thresholds[2] = threshold_base >> shift_val;
556     thresholds[3] = threshold_base >> shift_val;
557   }
558   thresholds[4] = threshold_base << 2;
559 }
560 
tune_thresh_based_on_resolution(AV1_COMP * cpi,int64_t thresholds[],int64_t threshold_base,int current_qindex,int source_sad_rd,int num_pixels)561 static inline void tune_thresh_based_on_resolution(
562     AV1_COMP *cpi, int64_t thresholds[], int64_t threshold_base,
563     int current_qindex, int source_sad_rd, int num_pixels) {
564   if (num_pixels >= RESOLUTION_720P) thresholds[3] = thresholds[3] << 1;
565   if (num_pixels <= RESOLUTION_288P) {
566     const int qindex_thr[5][2] = {
567       { 200, 220 }, { 140, 170 }, { 120, 150 }, { 200, 210 }, { 170, 220 },
568     };
569     int th_idx = 0;
570     if (cpi->sf.rt_sf.var_part_based_on_qidx >= 1)
571       th_idx =
572           (source_sad_rd <= kLowSad) ? cpi->sf.rt_sf.var_part_based_on_qidx : 0;
573     if (cpi->sf.rt_sf.var_part_based_on_qidx >= 3)
574       th_idx = cpi->sf.rt_sf.var_part_based_on_qidx;
575     const int qindex_low_thr = qindex_thr[th_idx][0];
576     const int qindex_high_thr = qindex_thr[th_idx][1];
577     if (current_qindex >= qindex_high_thr) {
578       threshold_base = (5 * threshold_base) >> 1;
579       thresholds[1] = threshold_base >> 3;
580       thresholds[2] = threshold_base << 2;
581       thresholds[3] = threshold_base << 5;
582     } else if (current_qindex < qindex_low_thr) {
583       thresholds[1] = threshold_base >> 3;
584       thresholds[2] = threshold_base >> 1;
585       thresholds[3] = threshold_base << 3;
586     } else {
587       int64_t qi_diff_low = current_qindex - qindex_low_thr;
588       int64_t qi_diff_high = qindex_high_thr - current_qindex;
589       int64_t threshold_diff = qindex_high_thr - qindex_low_thr;
590       int64_t threshold_base_high = (5 * threshold_base) >> 1;
591 
592       threshold_diff = threshold_diff > 0 ? threshold_diff : 1;
593       threshold_base =
594           (qi_diff_low * threshold_base_high + qi_diff_high * threshold_base) /
595           threshold_diff;
596       thresholds[1] = threshold_base >> 3;
597       thresholds[2] = ((qi_diff_low * threshold_base) +
598                        qi_diff_high * (threshold_base >> 1)) /
599                       threshold_diff;
600       thresholds[3] = ((qi_diff_low * (threshold_base << 5)) +
601                        qi_diff_high * (threshold_base << 3)) /
602                       threshold_diff;
603     }
604   } else if (num_pixels < RESOLUTION_720P) {
605     thresholds[2] = (5 * threshold_base) >> 2;
606   } else if (num_pixels < RESOLUTION_1080P) {
607     thresholds[2] = threshold_base << 1;
608   } else {
609     // num_pixels >= RESOLUTION_1080P
610     if (cpi->oxcf.tune_cfg.content == AOM_CONTENT_SCREEN) {
611       if (num_pixels < RESOLUTION_1440P) {
612         thresholds[2] = (5 * threshold_base) >> 1;
613       } else {
614         thresholds[2] = (7 * threshold_base) >> 1;
615       }
616     } else {
617       if (cpi->oxcf.speed > 7) {
618         thresholds[2] = 6 * threshold_base;
619       } else {
620         thresholds[2] = 3 * threshold_base;
621       }
622     }
623   }
624 }
625 
626 // Increase the base partition threshold, based on content and noise level.
tune_base_thresh_content(AV1_COMP * cpi,int64_t threshold_base,int content_lowsumdiff,int source_sad_nonrd,int num_pixels)627 static inline int64_t tune_base_thresh_content(AV1_COMP *cpi,
628                                                int64_t threshold_base,
629                                                int content_lowsumdiff,
630                                                int source_sad_nonrd,
631                                                int num_pixels) {
632   AV1_COMMON *const cm = &cpi->common;
633   int64_t updated_thresh_base = threshold_base;
634   if (cpi->noise_estimate.enabled && content_lowsumdiff &&
635       num_pixels > RESOLUTION_480P && cm->current_frame.frame_number > 60) {
636     NOISE_LEVEL noise_level =
637         av1_noise_estimate_extract_level(&cpi->noise_estimate);
638     if (noise_level == kHigh)
639       updated_thresh_base = (5 * updated_thresh_base) >> 1;
640     else if (noise_level == kMedium &&
641              !cpi->sf.rt_sf.prefer_large_partition_blocks)
642       updated_thresh_base = (5 * updated_thresh_base) >> 2;
643   }
644   updated_thresh_base = scale_part_thresh_content(
645       updated_thresh_base, cpi->oxcf.speed,
646       cpi->ppi->rtc_ref.non_reference_frame, cpi->rc.frame_source_sad == 0);
647   if (cpi->oxcf.speed >= 11 && source_sad_nonrd > kLowSad &&
648       cpi->rc.high_motion_content_screen_rtc)
649     updated_thresh_base = updated_thresh_base << 5;
650   return updated_thresh_base;
651 }
652 
set_vbp_thresholds(AV1_COMP * cpi,int64_t thresholds[],uint64_t blk_sad,int qindex,int content_lowsumdiff,int source_sad_nonrd,int source_sad_rd,bool is_segment_id_boosted,int lighting_change)653 static inline void set_vbp_thresholds(AV1_COMP *cpi, int64_t thresholds[],
654                                       uint64_t blk_sad, int qindex,
655                                       int content_lowsumdiff,
656                                       int source_sad_nonrd, int source_sad_rd,
657                                       bool is_segment_id_boosted,
658                                       int lighting_change) {
659   AV1_COMMON *const cm = &cpi->common;
660   const int is_key_frame = frame_is_intra_only(cm);
661   const int threshold_multiplier = is_key_frame ? 120 : 1;
662   const int ac_q = av1_ac_quant_QTX(qindex, 0, cm->seq_params->bit_depth);
663   int64_t threshold_base = (int64_t)(threshold_multiplier * ac_q);
664   const int current_qindex = cm->quant_params.base_qindex;
665   const int threshold_left_shift = cpi->sf.rt_sf.var_part_split_threshold_shift;
666   const int num_pixels = cm->width * cm->height;
667 
668   if (is_key_frame) {
669     set_vbp_thresholds_key_frame(cpi, thresholds, threshold_base,
670                                  threshold_left_shift, num_pixels);
671     return;
672   }
673 
674   threshold_base = tune_base_thresh_content(
675       cpi, threshold_base, content_lowsumdiff, source_sad_nonrd, num_pixels);
676   thresholds[0] = threshold_base >> 1;
677   thresholds[1] = threshold_base;
678   thresholds[3] = threshold_base << threshold_left_shift;
679 
680   tune_thresh_based_on_resolution(cpi, thresholds, threshold_base,
681                                   current_qindex, source_sad_rd, num_pixels);
682 
683   tune_thresh_based_on_qindex(cpi, thresholds, blk_sad, current_qindex,
684                               num_pixels, is_segment_id_boosted,
685                               source_sad_nonrd, lighting_change);
686 }
687 
688 // Set temporal variance low flag for superblock 64x64.
689 // Only first 25 in the array are used in this case.
set_low_temp_var_flag_64x64(CommonModeInfoParams * mi_params,PartitionSearchInfo * part_info,MACROBLOCKD * xd,VP64x64 * vt,const int64_t thresholds[],int mi_col,int mi_row)690 static inline void set_low_temp_var_flag_64x64(CommonModeInfoParams *mi_params,
691                                                PartitionSearchInfo *part_info,
692                                                MACROBLOCKD *xd, VP64x64 *vt,
693                                                const int64_t thresholds[],
694                                                int mi_col, int mi_row) {
695   if (xd->mi[0]->bsize == BLOCK_64X64) {
696     if ((vt->part_variances).none.variance < (thresholds[0] >> 1))
697       part_info->variance_low[0] = 1;
698   } else if (xd->mi[0]->bsize == BLOCK_64X32) {
699     for (int part_idx = 0; part_idx < 2; part_idx++) {
700       if (vt->part_variances.horz[part_idx].variance < (thresholds[0] >> 2))
701         part_info->variance_low[part_idx + 1] = 1;
702     }
703   } else if (xd->mi[0]->bsize == BLOCK_32X64) {
704     for (int part_idx = 0; part_idx < 2; part_idx++) {
705       if (vt->part_variances.vert[part_idx].variance < (thresholds[0] >> 2))
706         part_info->variance_low[part_idx + 3] = 1;
707     }
708   } else {
709     static const int idx[4][2] = { { 0, 0 }, { 0, 8 }, { 8, 0 }, { 8, 8 } };
710     for (int lvl1_idx = 0; lvl1_idx < 4; lvl1_idx++) {
711       const int idx_str = mi_params->mi_stride * (mi_row + idx[lvl1_idx][0]) +
712                           mi_col + idx[lvl1_idx][1];
713       MB_MODE_INFO **this_mi = mi_params->mi_grid_base + idx_str;
714 
715       if (mi_params->mi_cols <= mi_col + idx[lvl1_idx][1] ||
716           mi_params->mi_rows <= mi_row + idx[lvl1_idx][0])
717         continue;
718 
719       if (*this_mi == NULL) continue;
720 
721       if ((*this_mi)->bsize == BLOCK_32X32) {
722         int64_t threshold_32x32 = (5 * thresholds[1]) >> 3;
723         if (vt->split[lvl1_idx].part_variances.none.variance < threshold_32x32)
724           part_info->variance_low[lvl1_idx + 5] = 1;
725       } else {
726         // For 32x16 and 16x32 blocks, the flag is set on each 16x16 block
727         // inside.
728         if ((*this_mi)->bsize == BLOCK_16X16 ||
729             (*this_mi)->bsize == BLOCK_32X16 ||
730             (*this_mi)->bsize == BLOCK_16X32) {
731           for (int lvl2_idx = 0; lvl2_idx < 4; lvl2_idx++) {
732             if (vt->split[lvl1_idx]
733                     .split[lvl2_idx]
734                     .part_variances.none.variance < (thresholds[2] >> 8))
735               part_info->variance_low[(lvl1_idx << 2) + lvl2_idx + 9] = 1;
736           }
737         }
738       }
739     }
740   }
741 }
742 
set_low_temp_var_flag_128x128(CommonModeInfoParams * mi_params,PartitionSearchInfo * part_info,MACROBLOCKD * xd,VP128x128 * vt,const int64_t thresholds[],int mi_col,int mi_row)743 static inline void set_low_temp_var_flag_128x128(
744     CommonModeInfoParams *mi_params, PartitionSearchInfo *part_info,
745     MACROBLOCKD *xd, VP128x128 *vt, const int64_t thresholds[], int mi_col,
746     int mi_row) {
747   if (xd->mi[0]->bsize == BLOCK_128X128) {
748     if (vt->part_variances.none.variance < (thresholds[0] >> 1))
749       part_info->variance_low[0] = 1;
750   } else if (xd->mi[0]->bsize == BLOCK_128X64) {
751     for (int part_idx = 0; part_idx < 2; part_idx++) {
752       if (vt->part_variances.horz[part_idx].variance < (thresholds[0] >> 2))
753         part_info->variance_low[part_idx + 1] = 1;
754     }
755   } else if (xd->mi[0]->bsize == BLOCK_64X128) {
756     for (int part_idx = 0; part_idx < 2; part_idx++) {
757       if (vt->part_variances.vert[part_idx].variance < (thresholds[0] >> 2))
758         part_info->variance_low[part_idx + 3] = 1;
759     }
760   } else {
761     static const int idx64[4][2] = {
762       { 0, 0 }, { 0, 16 }, { 16, 0 }, { 16, 16 }
763     };
764     static const int idx32[4][2] = { { 0, 0 }, { 0, 8 }, { 8, 0 }, { 8, 8 } };
765     for (int lvl1_idx = 0; lvl1_idx < 4; lvl1_idx++) {
766       const int idx_str = mi_params->mi_stride * (mi_row + idx64[lvl1_idx][0]) +
767                           mi_col + idx64[lvl1_idx][1];
768       MB_MODE_INFO **mi_64 = mi_params->mi_grid_base + idx_str;
769       if (*mi_64 == NULL) continue;
770       if (mi_params->mi_cols <= mi_col + idx64[lvl1_idx][1] ||
771           mi_params->mi_rows <= mi_row + idx64[lvl1_idx][0])
772         continue;
773       const int64_t threshold_64x64 = (5 * thresholds[1]) >> 3;
774       if ((*mi_64)->bsize == BLOCK_64X64) {
775         if (vt->split[lvl1_idx].part_variances.none.variance < threshold_64x64)
776           part_info->variance_low[5 + lvl1_idx] = 1;
777       } else if ((*mi_64)->bsize == BLOCK_64X32) {
778         for (int part_idx = 0; part_idx < 2; part_idx++)
779           if (vt->split[lvl1_idx].part_variances.horz[part_idx].variance <
780               (threshold_64x64 >> 1))
781             part_info->variance_low[9 + (lvl1_idx << 1) + part_idx] = 1;
782       } else if ((*mi_64)->bsize == BLOCK_32X64) {
783         for (int part_idx = 0; part_idx < 2; part_idx++)
784           if (vt->split[lvl1_idx].part_variances.vert[part_idx].variance <
785               (threshold_64x64 >> 1))
786             part_info->variance_low[17 + (lvl1_idx << 1) + part_idx] = 1;
787       } else {
788         for (int lvl2_idx = 0; lvl2_idx < 4; lvl2_idx++) {
789           const int idx_str1 =
790               mi_params->mi_stride * idx32[lvl2_idx][0] + idx32[lvl2_idx][1];
791           MB_MODE_INFO **mi_32 = mi_params->mi_grid_base + idx_str + idx_str1;
792           if (*mi_32 == NULL) continue;
793 
794           if (mi_params->mi_cols <=
795                   mi_col + idx64[lvl1_idx][1] + idx32[lvl2_idx][1] ||
796               mi_params->mi_rows <=
797                   mi_row + idx64[lvl1_idx][0] + idx32[lvl2_idx][0])
798             continue;
799           const int64_t threshold_32x32 = (5 * thresholds[2]) >> 3;
800           if ((*mi_32)->bsize == BLOCK_32X32) {
801             if (vt->split[lvl1_idx]
802                     .split[lvl2_idx]
803                     .part_variances.none.variance < threshold_32x32)
804               part_info->variance_low[25 + (lvl1_idx << 2) + lvl2_idx] = 1;
805           } else {
806             // For 32x16 and 16x32 blocks, the flag is set on each 16x16 block
807             // inside.
808             if ((*mi_32)->bsize == BLOCK_16X16 ||
809                 (*mi_32)->bsize == BLOCK_32X16 ||
810                 (*mi_32)->bsize == BLOCK_16X32) {
811               for (int lvl3_idx = 0; lvl3_idx < 4; lvl3_idx++) {
812                 VPartVar *none_var = &vt->split[lvl1_idx]
813                                           .split[lvl2_idx]
814                                           .split[lvl3_idx]
815                                           .part_variances.none;
816                 if (none_var->variance < (thresholds[3] >> 8))
817                   part_info->variance_low[41 + (lvl1_idx << 4) +
818                                           (lvl2_idx << 2) + lvl3_idx] = 1;
819               }
820             }
821           }
822         }
823       }
824     }
825   }
826 }
827 
set_low_temp_var_flag(AV1_COMP * cpi,PartitionSearchInfo * part_info,MACROBLOCKD * xd,VP128x128 * vt,int64_t thresholds[],MV_REFERENCE_FRAME ref_frame_partition,int mi_col,int mi_row,const bool is_small_sb)828 static inline void set_low_temp_var_flag(
829     AV1_COMP *cpi, PartitionSearchInfo *part_info, MACROBLOCKD *xd,
830     VP128x128 *vt, int64_t thresholds[], MV_REFERENCE_FRAME ref_frame_partition,
831     int mi_col, int mi_row, const bool is_small_sb) {
832   AV1_COMMON *const cm = &cpi->common;
833   // Check temporal variance for bsize >= 16x16, if LAST_FRAME was selected.
834   // If the temporal variance is small set the flag
835   // variance_low for the block. The variance threshold can be adjusted, the
836   // higher the more aggressive.
837   if (ref_frame_partition == LAST_FRAME) {
838     if (is_small_sb)
839       set_low_temp_var_flag_64x64(&cm->mi_params, part_info, xd,
840                                   &(vt->split[0]), thresholds, mi_col, mi_row);
841     else
842       set_low_temp_var_flag_128x128(&cm->mi_params, part_info, xd, vt,
843                                     thresholds, mi_col, mi_row);
844   }
845 }
846 
847 static const int pos_shift_16x16[4][4] = {
848   { 9, 10, 13, 14 }, { 11, 12, 15, 16 }, { 17, 18, 21, 22 }, { 19, 20, 23, 24 }
849 };
850 
av1_get_force_skip_low_temp_var_small_sb(const uint8_t * variance_low,int mi_row,int mi_col,BLOCK_SIZE bsize)851 int av1_get_force_skip_low_temp_var_small_sb(const uint8_t *variance_low,
852                                              int mi_row, int mi_col,
853                                              BLOCK_SIZE bsize) {
854   // Relative indices of MB inside the superblock.
855   const int mi_x = mi_row & 0xF;
856   const int mi_y = mi_col & 0xF;
857   // Relative indices of 16x16 block inside the superblock.
858   const int i = mi_x >> 2;
859   const int j = mi_y >> 2;
860   int force_skip_low_temp_var = 0;
861   // Set force_skip_low_temp_var based on the block size and block offset.
862   switch (bsize) {
863     case BLOCK_64X64: force_skip_low_temp_var = variance_low[0]; break;
864     case BLOCK_64X32:
865       if (!mi_y && !mi_x) {
866         force_skip_low_temp_var = variance_low[1];
867       } else if (!mi_y && mi_x) {
868         force_skip_low_temp_var = variance_low[2];
869       }
870       break;
871     case BLOCK_32X64:
872       if (!mi_y && !mi_x) {
873         force_skip_low_temp_var = variance_low[3];
874       } else if (mi_y && !mi_x) {
875         force_skip_low_temp_var = variance_low[4];
876       }
877       break;
878     case BLOCK_32X32:
879       if (!mi_y && !mi_x) {
880         force_skip_low_temp_var = variance_low[5];
881       } else if (mi_y && !mi_x) {
882         force_skip_low_temp_var = variance_low[6];
883       } else if (!mi_y && mi_x) {
884         force_skip_low_temp_var = variance_low[7];
885       } else if (mi_y && mi_x) {
886         force_skip_low_temp_var = variance_low[8];
887       }
888       break;
889     case BLOCK_32X16:
890     case BLOCK_16X32:
891     case BLOCK_16X16:
892       force_skip_low_temp_var = variance_low[pos_shift_16x16[i][j]];
893       break;
894     default: break;
895   }
896 
897   return force_skip_low_temp_var;
898 }
899 
av1_get_force_skip_low_temp_var(const uint8_t * variance_low,int mi_row,int mi_col,BLOCK_SIZE bsize)900 int av1_get_force_skip_low_temp_var(const uint8_t *variance_low, int mi_row,
901                                     int mi_col, BLOCK_SIZE bsize) {
902   int force_skip_low_temp_var = 0;
903   int x, y;
904   x = (mi_col & 0x1F) >> 4;
905   // y = (mi_row & 0x1F) >> 4;
906   // const int idx64 = (y << 1) + x;
907   y = (mi_row & 0x17) >> 3;
908   const int idx64 = y + x;
909 
910   x = (mi_col & 0xF) >> 3;
911   // y = (mi_row & 0xF) >> 3;
912   // const int idx32 = (y << 1) + x;
913   y = (mi_row & 0xB) >> 2;
914   const int idx32 = y + x;
915 
916   x = (mi_col & 0x7) >> 2;
917   // y = (mi_row & 0x7) >> 2;
918   // const int idx16 = (y << 1) + x;
919   y = (mi_row & 0x5) >> 1;
920   const int idx16 = y + x;
921   // Set force_skip_low_temp_var based on the block size and block offset.
922   switch (bsize) {
923     case BLOCK_128X128: force_skip_low_temp_var = variance_low[0]; break;
924     case BLOCK_128X64:
925       assert((mi_col & 0x1F) == 0);
926       force_skip_low_temp_var = variance_low[1 + ((mi_row & 0x1F) != 0)];
927       break;
928     case BLOCK_64X128:
929       assert((mi_row & 0x1F) == 0);
930       force_skip_low_temp_var = variance_low[3 + ((mi_col & 0x1F) != 0)];
931       break;
932     case BLOCK_64X64:
933       // Location of this 64x64 block inside the 128x128 superblock
934       force_skip_low_temp_var = variance_low[5 + idx64];
935       break;
936     case BLOCK_64X32:
937       x = (mi_col & 0x1F) >> 4;
938       y = (mi_row & 0x1F) >> 3;
939       /*
940       .---------------.---------------.
941       | x=0,y=0,idx=0 | x=0,y=0,idx=2 |
942       :---------------+---------------:
943       | x=0,y=1,idx=1 | x=1,y=1,idx=3 |
944       :---------------+---------------:
945       | x=0,y=2,idx=4 | x=1,y=2,idx=6 |
946       :---------------+---------------:
947       | x=0,y=3,idx=5 | x=1,y=3,idx=7 |
948       '---------------'---------------'
949       */
950       const int idx64x32 = (x << 1) + (y % 2) + ((y >> 1) << 2);
951       force_skip_low_temp_var = variance_low[9 + idx64x32];
952       break;
953     case BLOCK_32X64:
954       x = (mi_col & 0x1F) >> 3;
955       y = (mi_row & 0x1F) >> 4;
956       const int idx32x64 = (y << 2) + x;
957       force_skip_low_temp_var = variance_low[17 + idx32x64];
958       break;
959     case BLOCK_32X32:
960       force_skip_low_temp_var = variance_low[25 + (idx64 << 2) + idx32];
961       break;
962     case BLOCK_32X16:
963     case BLOCK_16X32:
964     case BLOCK_16X16:
965       force_skip_low_temp_var =
966           variance_low[41 + (idx64 << 4) + (idx32 << 2) + idx16];
967       break;
968     default: break;
969   }
970   return force_skip_low_temp_var;
971 }
972 
av1_set_variance_partition_thresholds(AV1_COMP * cpi,int qindex,int content_lowsumdiff)973 void av1_set_variance_partition_thresholds(AV1_COMP *cpi, int qindex,
974                                            int content_lowsumdiff) {
975   SPEED_FEATURES *const sf = &cpi->sf;
976   if (sf->part_sf.partition_search_type != VAR_BASED_PARTITION) {
977     return;
978   } else {
979     set_vbp_thresholds(cpi, cpi->vbp_info.thresholds, 0, qindex,
980                        content_lowsumdiff, 0, 0, 0, 0);
981     // The threshold below is not changed locally.
982     cpi->vbp_info.threshold_minmax = 15 + (qindex >> 3);
983   }
984 }
985 
chroma_check(AV1_COMP * cpi,MACROBLOCK * x,BLOCK_SIZE bsize,unsigned int y_sad,unsigned int y_sad_g,unsigned int y_sad_alt,bool is_key_frame,bool zero_motion,unsigned int * uv_sad)986 static inline void chroma_check(AV1_COMP *cpi, MACROBLOCK *x, BLOCK_SIZE bsize,
987                                 unsigned int y_sad, unsigned int y_sad_g,
988                                 unsigned int y_sad_alt, bool is_key_frame,
989                                 bool zero_motion, unsigned int *uv_sad) {
990   MACROBLOCKD *xd = &x->e_mbd;
991   const int source_sad_nonrd = x->content_state_sb.source_sad_nonrd;
992   int shift_upper_limit = 1;
993   int shift_lower_limit = 3;
994   int fac_uv = 6;
995   if (is_key_frame || cpi->oxcf.tool_cfg.enable_monochrome) return;
996 
997   // Use lower threshold (more conservative in setting color flag) for
998   // higher resolutions non-screen, which tend to have more camera noise.
999   // Since this may be used to skip compound mode in nonrd pickmode, which
1000   // is generally more effective for higher resolutions, better to be more
1001   // conservative.
1002   if (cpi->oxcf.tune_cfg.content != AOM_CONTENT_SCREEN) {
1003     if (cpi->common.width * cpi->common.height >= RESOLUTION_1080P)
1004       fac_uv = 3;
1005     else
1006       fac_uv = 5;
1007   }
1008   if (cpi->oxcf.tune_cfg.content == AOM_CONTENT_SCREEN &&
1009       cpi->rc.high_source_sad) {
1010     shift_lower_limit = 7;
1011   } else if (cpi->oxcf.tune_cfg.content == AOM_CONTENT_SCREEN &&
1012              cpi->rc.percent_blocks_with_motion > 90 &&
1013              cpi->rc.frame_source_sad > 10000 && source_sad_nonrd > kLowSad) {
1014     shift_lower_limit = 8;
1015     shift_upper_limit = 3;
1016   } else if (source_sad_nonrd >= kMedSad && x->source_variance > 500 &&
1017              cpi->common.width * cpi->common.height >= 640 * 360) {
1018     shift_upper_limit = 2;
1019     shift_lower_limit = source_sad_nonrd > kMedSad ? 5 : 4;
1020   }
1021 
1022   MB_MODE_INFO *mi = xd->mi[0];
1023   const AV1_COMMON *const cm = &cpi->common;
1024   const YV12_BUFFER_CONFIG *yv12 = get_ref_frame_yv12_buf(cm, LAST_FRAME);
1025   const YV12_BUFFER_CONFIG *yv12_g = get_ref_frame_yv12_buf(cm, GOLDEN_FRAME);
1026   const YV12_BUFFER_CONFIG *yv12_alt = get_ref_frame_yv12_buf(cm, ALTREF_FRAME);
1027   const struct scale_factors *const sf =
1028       get_ref_scale_factors_const(cm, LAST_FRAME);
1029   struct buf_2d dst;
1030   unsigned int uv_sad_g = 0;
1031   unsigned int uv_sad_alt = 0;
1032 
1033   for (int plane = AOM_PLANE_U; plane < MAX_MB_PLANE; ++plane) {
1034     struct macroblock_plane *p = &x->plane[plane];
1035     struct macroblockd_plane *pd = &xd->plane[plane];
1036     const BLOCK_SIZE bs =
1037         get_plane_block_size(bsize, pd->subsampling_x, pd->subsampling_y);
1038 
1039     if (bs != BLOCK_INVALID) {
1040       // For last:
1041       if (zero_motion) {
1042         if (mi->ref_frame[0] == LAST_FRAME) {
1043           uv_sad[plane - 1] = cpi->ppi->fn_ptr[bs].sdf(
1044               p->src.buf, p->src.stride, pd->pre[0].buf, pd->pre[0].stride);
1045         } else {
1046           uint8_t *src = (plane == 1) ? yv12->u_buffer : yv12->v_buffer;
1047           setup_pred_plane(&dst, xd->mi[0]->bsize, src, yv12->uv_crop_width,
1048                            yv12->uv_crop_height, yv12->uv_stride, xd->mi_row,
1049                            xd->mi_col, sf, xd->plane[plane].subsampling_x,
1050                            xd->plane[plane].subsampling_y);
1051 
1052           uv_sad[plane - 1] = cpi->ppi->fn_ptr[bs].sdf(
1053               p->src.buf, p->src.stride, dst.buf, dst.stride);
1054         }
1055       } else {
1056         uv_sad[plane - 1] = cpi->ppi->fn_ptr[bs].sdf(
1057             p->src.buf, p->src.stride, pd->dst.buf, pd->dst.stride);
1058       }
1059 
1060       // For golden:
1061       if (y_sad_g != UINT_MAX) {
1062         uint8_t *src = (plane == 1) ? yv12_g->u_buffer : yv12_g->v_buffer;
1063         setup_pred_plane(&dst, xd->mi[0]->bsize, src, yv12_g->uv_crop_width,
1064                          yv12_g->uv_crop_height, yv12_g->uv_stride, xd->mi_row,
1065                          xd->mi_col, sf, xd->plane[plane].subsampling_x,
1066                          xd->plane[plane].subsampling_y);
1067         uv_sad_g = cpi->ppi->fn_ptr[bs].sdf(p->src.buf, p->src.stride, dst.buf,
1068                                             dst.stride);
1069       }
1070 
1071       // For altref:
1072       if (y_sad_alt != UINT_MAX) {
1073         uint8_t *src = (plane == 1) ? yv12_alt->u_buffer : yv12_alt->v_buffer;
1074         setup_pred_plane(&dst, xd->mi[0]->bsize, src, yv12_alt->uv_crop_width,
1075                          yv12_alt->uv_crop_height, yv12_alt->uv_stride,
1076                          xd->mi_row, xd->mi_col, sf,
1077                          xd->plane[plane].subsampling_x,
1078                          xd->plane[plane].subsampling_y);
1079         uv_sad_alt = cpi->ppi->fn_ptr[bs].sdf(p->src.buf, p->src.stride,
1080                                               dst.buf, dst.stride);
1081       }
1082     }
1083 
1084     if (uv_sad[plane - 1] > (y_sad >> shift_upper_limit))
1085       x->color_sensitivity_sb[COLOR_SENS_IDX(plane)] = 1;
1086     else if (uv_sad[plane - 1] < (y_sad >> shift_lower_limit))
1087       x->color_sensitivity_sb[COLOR_SENS_IDX(plane)] = 0;
1088     // Borderline case: to be refined at coding block level in nonrd_pickmode,
1089     // for coding block size < sb_size.
1090     else
1091       x->color_sensitivity_sb[COLOR_SENS_IDX(plane)] = 2;
1092 
1093     x->color_sensitivity_sb_g[COLOR_SENS_IDX(plane)] =
1094         uv_sad_g > y_sad_g / fac_uv;
1095     x->color_sensitivity_sb_alt[COLOR_SENS_IDX(plane)] =
1096         uv_sad_alt > y_sad_alt / fac_uv;
1097   }
1098 }
1099 
fill_variance_tree_leaves(AV1_COMP * cpi,MACROBLOCK * x,VP128x128 * vt,PART_EVAL_STATUS * force_split,int avg_16x16[][4],int maxvar_16x16[][4],int minvar_16x16[][4],int64_t * thresholds,const uint8_t * src_buf,int src_stride,const uint8_t * dst_buf,int dst_stride,bool is_key_frame,const bool is_small_sb)1100 static void fill_variance_tree_leaves(
1101     AV1_COMP *cpi, MACROBLOCK *x, VP128x128 *vt, PART_EVAL_STATUS *force_split,
1102     int avg_16x16[][4], int maxvar_16x16[][4], int minvar_16x16[][4],
1103     int64_t *thresholds, const uint8_t *src_buf, int src_stride,
1104     const uint8_t *dst_buf, int dst_stride, bool is_key_frame,
1105     const bool is_small_sb) {
1106   MACROBLOCKD *xd = &x->e_mbd;
1107   const int num_64x64_blocks = is_small_sb ? 1 : 4;
1108   // TODO(kyslov) Bring back compute_minmax_variance with content type detection
1109   const int compute_minmax_variance = 0;
1110   const int segment_id = xd->mi[0]->segment_id;
1111   int pixels_wide = 128, pixels_high = 128;
1112   int border_offset_4x4 = 0;
1113   int temporal_denoising = cpi->sf.rt_sf.use_rtc_tf;
1114   // dst_buf pointer is not used for is_key_frame, so it should be NULL.
1115   assert(IMPLIES(is_key_frame, dst_buf == NULL));
1116   if (is_small_sb) {
1117     pixels_wide = 64;
1118     pixels_high = 64;
1119   }
1120   if (xd->mb_to_right_edge < 0) pixels_wide += (xd->mb_to_right_edge >> 3);
1121   if (xd->mb_to_bottom_edge < 0) pixels_high += (xd->mb_to_bottom_edge >> 3);
1122 #if CONFIG_AV1_TEMPORAL_DENOISING
1123   temporal_denoising |= cpi->oxcf.noise_sensitivity;
1124 #endif
1125   // For temporal filtering or temporal denoiser enabled: since the source
1126   // is modified we need to avoid 4x4 avg along superblock boundary, since
1127   // simd code will load 8 pixels for 4x4 avg and so can access source
1128   // data outside superblock (while its being modified by temporal filter).
1129   // Temporal filtering is never done on key frames.
1130   if (!is_key_frame && temporal_denoising) border_offset_4x4 = 4;
1131   for (int blk64_idx = 0; blk64_idx < num_64x64_blocks; blk64_idx++) {
1132     const int x64_idx = GET_BLK_IDX_X(blk64_idx, 6);
1133     const int y64_idx = GET_BLK_IDX_Y(blk64_idx, 6);
1134     const int blk64_scale_idx = blk64_idx << 2;
1135     force_split[blk64_idx + 1] = PART_EVAL_ALL;
1136 
1137     for (int lvl1_idx = 0; lvl1_idx < 4; lvl1_idx++) {
1138       const int x32_idx = x64_idx + GET_BLK_IDX_X(lvl1_idx, 5);
1139       const int y32_idx = y64_idx + GET_BLK_IDX_Y(lvl1_idx, 5);
1140       const int lvl1_scale_idx = (blk64_scale_idx + lvl1_idx) << 2;
1141       force_split[5 + blk64_scale_idx + lvl1_idx] = PART_EVAL_ALL;
1142       avg_16x16[blk64_idx][lvl1_idx] = 0;
1143       maxvar_16x16[blk64_idx][lvl1_idx] = 0;
1144       minvar_16x16[blk64_idx][lvl1_idx] = INT_MAX;
1145       for (int lvl2_idx = 0; lvl2_idx < 4; lvl2_idx++) {
1146         const int x16_idx = x32_idx + GET_BLK_IDX_X(lvl2_idx, 4);
1147         const int y16_idx = y32_idx + GET_BLK_IDX_Y(lvl2_idx, 4);
1148         const int split_index = 21 + lvl1_scale_idx + lvl2_idx;
1149         VP16x16 *vst = &vt->split[blk64_idx].split[lvl1_idx].split[lvl2_idx];
1150         force_split[split_index] = PART_EVAL_ALL;
1151         if (is_key_frame) {
1152           // Go down to 4x4 down-sampling for variance.
1153           for (int lvl3_idx = 0; lvl3_idx < 4; lvl3_idx++) {
1154             const int x8_idx = x16_idx + GET_BLK_IDX_X(lvl3_idx, 3);
1155             const int y8_idx = y16_idx + GET_BLK_IDX_Y(lvl3_idx, 3);
1156             VP8x8 *vst2 = &vst->split[lvl3_idx];
1157             fill_variance_4x4avg(src_buf, src_stride, x8_idx, y8_idx, vst2,
1158 #if CONFIG_AV1_HIGHBITDEPTH
1159                                  xd->cur_buf->flags,
1160 #endif
1161                                  pixels_wide, pixels_high, border_offset_4x4);
1162           }
1163         } else {
1164           fill_variance_8x8avg(src_buf, src_stride, dst_buf, dst_stride,
1165                                x16_idx, y16_idx, vst, is_cur_buf_hbd(xd),
1166                                pixels_wide, pixels_high);
1167 
1168           fill_variance_tree(vst, BLOCK_16X16);
1169           VPartVar *none_var = &vt->split[blk64_idx]
1170                                     .split[lvl1_idx]
1171                                     .split[lvl2_idx]
1172                                     .part_variances.none;
1173           get_variance(none_var);
1174           const int val_none_var = none_var->variance;
1175           avg_16x16[blk64_idx][lvl1_idx] += val_none_var;
1176           minvar_16x16[blk64_idx][lvl1_idx] =
1177               AOMMIN(minvar_16x16[blk64_idx][lvl1_idx], val_none_var);
1178           maxvar_16x16[blk64_idx][lvl1_idx] =
1179               AOMMAX(maxvar_16x16[blk64_idx][lvl1_idx], val_none_var);
1180           if (val_none_var > thresholds[3]) {
1181             // 16X16 variance is above threshold for split, so force split to
1182             // 8x8 for this 16x16 block (this also forces splits for upper
1183             // levels).
1184             force_split[split_index] = PART_EVAL_ONLY_SPLIT;
1185             force_split[5 + blk64_scale_idx + lvl1_idx] = PART_EVAL_ONLY_SPLIT;
1186             force_split[blk64_idx + 1] = PART_EVAL_ONLY_SPLIT;
1187             force_split[0] = PART_EVAL_ONLY_SPLIT;
1188           } else if (!cyclic_refresh_segment_id_boosted(segment_id) &&
1189                      compute_minmax_variance && val_none_var > thresholds[2]) {
1190             // We have some nominal amount of 16x16 variance (based on average),
1191             // compute the minmax over the 8x8 sub-blocks, and if above
1192             // threshold, force split to 8x8 block for this 16x16 block.
1193             int minmax = compute_minmax_8x8(src_buf, src_stride, dst_buf,
1194                                             dst_stride, x16_idx, y16_idx,
1195 #if CONFIG_AV1_HIGHBITDEPTH
1196                                             xd->cur_buf->flags,
1197 #endif
1198                                             pixels_wide, pixels_high);
1199             const int thresh_minmax = (int)cpi->vbp_info.threshold_minmax;
1200             if (minmax > thresh_minmax) {
1201               force_split[split_index] = PART_EVAL_ONLY_SPLIT;
1202               force_split[5 + blk64_scale_idx + lvl1_idx] =
1203                   PART_EVAL_ONLY_SPLIT;
1204               force_split[blk64_idx + 1] = PART_EVAL_ONLY_SPLIT;
1205               force_split[0] = PART_EVAL_ONLY_SPLIT;
1206             }
1207           }
1208         }
1209       }
1210     }
1211   }
1212 }
1213 
set_ref_frame_for_partition(AV1_COMP * cpi,MACROBLOCK * x,MACROBLOCKD * xd,MV_REFERENCE_FRAME * ref_frame_partition,MB_MODE_INFO * mi,unsigned int * y_sad,unsigned int * y_sad_g,unsigned int * y_sad_alt,const YV12_BUFFER_CONFIG * yv12_g,const YV12_BUFFER_CONFIG * yv12_alt,int mi_row,int mi_col,int num_planes)1214 static inline void set_ref_frame_for_partition(
1215     AV1_COMP *cpi, MACROBLOCK *x, MACROBLOCKD *xd,
1216     MV_REFERENCE_FRAME *ref_frame_partition, MB_MODE_INFO *mi,
1217     unsigned int *y_sad, unsigned int *y_sad_g, unsigned int *y_sad_alt,
1218     const YV12_BUFFER_CONFIG *yv12_g, const YV12_BUFFER_CONFIG *yv12_alt,
1219     int mi_row, int mi_col, int num_planes) {
1220   AV1_COMMON *const cm = &cpi->common;
1221   const bool is_set_golden_ref_frame =
1222       *y_sad_g < 0.9 * *y_sad && *y_sad_g < *y_sad_alt;
1223   const bool is_set_altref_ref_frame =
1224       *y_sad_alt < 0.9 * *y_sad && *y_sad_alt < *y_sad_g;
1225 
1226   if (is_set_golden_ref_frame) {
1227     av1_setup_pre_planes(xd, 0, yv12_g, mi_row, mi_col,
1228                          get_ref_scale_factors(cm, GOLDEN_FRAME), num_planes);
1229     mi->ref_frame[0] = GOLDEN_FRAME;
1230     mi->mv[0].as_int = 0;
1231     *y_sad = *y_sad_g;
1232     *ref_frame_partition = GOLDEN_FRAME;
1233     x->nonrd_prune_ref_frame_search = 0;
1234     x->sb_me_partition = 0;
1235   } else if (is_set_altref_ref_frame) {
1236     av1_setup_pre_planes(xd, 0, yv12_alt, mi_row, mi_col,
1237                          get_ref_scale_factors(cm, ALTREF_FRAME), num_planes);
1238     mi->ref_frame[0] = ALTREF_FRAME;
1239     mi->mv[0].as_int = 0;
1240     *y_sad = *y_sad_alt;
1241     *ref_frame_partition = ALTREF_FRAME;
1242     x->nonrd_prune_ref_frame_search = 0;
1243     x->sb_me_partition = 0;
1244   } else {
1245     *ref_frame_partition = LAST_FRAME;
1246     x->nonrd_prune_ref_frame_search =
1247         cpi->sf.rt_sf.nonrd_prune_ref_frame_search;
1248   }
1249 }
1250 
mv_distance(const FULLPEL_MV * mv0,const FULLPEL_MV * mv1)1251 static AOM_FORCE_INLINE int mv_distance(const FULLPEL_MV *mv0,
1252                                         const FULLPEL_MV *mv1) {
1253   return abs(mv0->row - mv1->row) + abs(mv0->col - mv1->col);
1254 }
1255 
evaluate_neighbour_mvs(AV1_COMP * cpi,MACROBLOCK * x,unsigned int * y_sad,bool is_small_sb,int est_motion)1256 static inline void evaluate_neighbour_mvs(AV1_COMP *cpi, MACROBLOCK *x,
1257                                           unsigned int *y_sad, bool is_small_sb,
1258                                           int est_motion) {
1259   const int source_sad_nonrd = x->content_state_sb.source_sad_nonrd;
1260   // TODO([email protected]): test if this condition works with other
1261   // speeds.
1262   if (est_motion > 2 && source_sad_nonrd > kMedSad) return;
1263 
1264   MACROBLOCKD *xd = &x->e_mbd;
1265   BLOCK_SIZE bsize = is_small_sb ? BLOCK_64X64 : BLOCK_128X128;
1266   MB_MODE_INFO *mi = xd->mi[0];
1267 
1268   unsigned int above_y_sad = UINT_MAX;
1269   unsigned int left_y_sad = UINT_MAX;
1270   FULLPEL_MV above_mv = kZeroFullMv;
1271   FULLPEL_MV left_mv = kZeroFullMv;
1272   SubpelMvLimits subpel_mv_limits;
1273   const MV dummy_mv = { 0, 0 };
1274   av1_set_subpel_mv_search_range(&subpel_mv_limits, &x->mv_limits, &dummy_mv);
1275 
1276   // Current best MV
1277   FULLPEL_MV best_mv = get_fullmv_from_mv(&mi->mv[0].as_mv);
1278   const int multi = (est_motion > 2 && source_sad_nonrd > kLowSad) ? 7 : 8;
1279 
1280   if (xd->up_available) {
1281     const MB_MODE_INFO *above_mbmi = xd->above_mbmi;
1282     if (above_mbmi->mode >= INTRA_MODE_END &&
1283         above_mbmi->ref_frame[0] == LAST_FRAME) {
1284       MV temp = above_mbmi->mv[0].as_mv;
1285       clamp_mv(&temp, &subpel_mv_limits);
1286       above_mv = get_fullmv_from_mv(&temp);
1287 
1288       if (mv_distance(&best_mv, &above_mv) > 0) {
1289         uint8_t const *ref_buf =
1290             get_buf_from_fullmv(&xd->plane[0].pre[0], &above_mv);
1291         above_y_sad = cpi->ppi->fn_ptr[bsize].sdf(
1292             x->plane[0].src.buf, x->plane[0].src.stride, ref_buf,
1293             xd->plane[0].pre[0].stride);
1294       }
1295     }
1296   }
1297   if (xd->left_available) {
1298     const MB_MODE_INFO *left_mbmi = xd->left_mbmi;
1299     if (left_mbmi->mode >= INTRA_MODE_END &&
1300         left_mbmi->ref_frame[0] == LAST_FRAME) {
1301       MV temp = left_mbmi->mv[0].as_mv;
1302       clamp_mv(&temp, &subpel_mv_limits);
1303       left_mv = get_fullmv_from_mv(&temp);
1304 
1305       if (mv_distance(&best_mv, &left_mv) > 0 &&
1306           mv_distance(&above_mv, &left_mv) > 0) {
1307         uint8_t const *ref_buf =
1308             get_buf_from_fullmv(&xd->plane[0].pre[0], &left_mv);
1309         left_y_sad = cpi->ppi->fn_ptr[bsize].sdf(
1310             x->plane[0].src.buf, x->plane[0].src.stride, ref_buf,
1311             xd->plane[0].pre[0].stride);
1312       }
1313     }
1314   }
1315 
1316   if (above_y_sad < ((multi * *y_sad) >> 3) && above_y_sad < left_y_sad) {
1317     *y_sad = above_y_sad;
1318     mi->mv[0].as_mv = get_mv_from_fullmv(&above_mv);
1319     clamp_mv(&mi->mv[0].as_mv, &subpel_mv_limits);
1320   }
1321   if (left_y_sad < ((multi * *y_sad) >> 3) && left_y_sad < above_y_sad) {
1322     *y_sad = left_y_sad;
1323     mi->mv[0].as_mv = get_mv_from_fullmv(&left_mv);
1324     clamp_mv(&mi->mv[0].as_mv, &subpel_mv_limits);
1325   }
1326 }
1327 
setup_planes(AV1_COMP * cpi,MACROBLOCK * x,unsigned int * y_sad,unsigned int * y_sad_g,unsigned int * y_sad_alt,unsigned int * y_sad_last,MV_REFERENCE_FRAME * ref_frame_partition,struct scale_factors * sf_no_scale,int mi_row,int mi_col,bool is_small_sb,bool scaled_ref_last)1328 static void setup_planes(AV1_COMP *cpi, MACROBLOCK *x, unsigned int *y_sad,
1329                          unsigned int *y_sad_g, unsigned int *y_sad_alt,
1330                          unsigned int *y_sad_last,
1331                          MV_REFERENCE_FRAME *ref_frame_partition,
1332                          struct scale_factors *sf_no_scale, int mi_row,
1333                          int mi_col, bool is_small_sb, bool scaled_ref_last) {
1334   AV1_COMMON *const cm = &cpi->common;
1335   MACROBLOCKD *xd = &x->e_mbd;
1336   const int num_planes = av1_num_planes(cm);
1337   bool scaled_ref_golden = false;
1338   bool scaled_ref_alt = false;
1339   BLOCK_SIZE bsize = is_small_sb ? BLOCK_64X64 : BLOCK_128X128;
1340   MB_MODE_INFO *mi = xd->mi[0];
1341   const YV12_BUFFER_CONFIG *yv12 =
1342       scaled_ref_last ? av1_get_scaled_ref_frame(cpi, LAST_FRAME)
1343                       : get_ref_frame_yv12_buf(cm, LAST_FRAME);
1344   assert(yv12 != NULL);
1345   const YV12_BUFFER_CONFIG *yv12_g = NULL;
1346   const YV12_BUFFER_CONFIG *yv12_alt = NULL;
1347   // Check if LAST is a reference. For spatial layers always use it as
1348   // reference scaling.
1349   int use_last_ref = (cpi->ref_frame_flags & AOM_LAST_FLAG) ||
1350                      cpi->svc.number_spatial_layers > 1;
1351   int use_golden_ref = cpi->ref_frame_flags & AOM_GOLD_FLAG;
1352   int use_alt_ref = cpi->ppi->rtc_ref.set_ref_frame_config ||
1353                     cpi->sf.rt_sf.use_nonrd_altref_frame ||
1354                     (cpi->sf.rt_sf.use_comp_ref_nonrd &&
1355                      cpi->sf.rt_sf.ref_frame_comp_nonrd[2] == 1);
1356 
1357   // For 1 spatial layer: GOLDEN is another temporal reference.
1358   // Check if it should be used as reference for partitioning.
1359   if (cpi->svc.number_spatial_layers == 1 && use_golden_ref &&
1360       (x->content_state_sb.source_sad_nonrd != kZeroSad || !use_last_ref)) {
1361     yv12_g = get_ref_frame_yv12_buf(cm, GOLDEN_FRAME);
1362     if (yv12_g && (yv12_g->y_crop_height != cm->height ||
1363                    yv12_g->y_crop_width != cm->width)) {
1364       yv12_g = av1_get_scaled_ref_frame(cpi, GOLDEN_FRAME);
1365       scaled_ref_golden = true;
1366     }
1367     if (yv12_g && yv12_g != yv12) {
1368       av1_setup_pre_planes(
1369           xd, 0, yv12_g, mi_row, mi_col,
1370           scaled_ref_golden ? NULL : get_ref_scale_factors(cm, GOLDEN_FRAME),
1371           num_planes);
1372       *y_sad_g = cpi->ppi->fn_ptr[bsize].sdf(
1373           x->plane[AOM_PLANE_Y].src.buf, x->plane[AOM_PLANE_Y].src.stride,
1374           xd->plane[AOM_PLANE_Y].pre[0].buf,
1375           xd->plane[AOM_PLANE_Y].pre[0].stride);
1376     }
1377   }
1378 
1379   // For 1 spatial layer: ALTREF is another temporal reference.
1380   // Check if it should be used as reference for partitioning.
1381   if (cpi->svc.number_spatial_layers == 1 && use_alt_ref &&
1382       (cpi->ref_frame_flags & AOM_ALT_FLAG) &&
1383       (x->content_state_sb.source_sad_nonrd != kZeroSad || !use_last_ref)) {
1384     yv12_alt = get_ref_frame_yv12_buf(cm, ALTREF_FRAME);
1385     if (yv12_alt && (yv12_alt->y_crop_height != cm->height ||
1386                      yv12_alt->y_crop_width != cm->width)) {
1387       yv12_alt = av1_get_scaled_ref_frame(cpi, ALTREF_FRAME);
1388       scaled_ref_alt = true;
1389     }
1390     if (yv12_alt && yv12_alt != yv12) {
1391       av1_setup_pre_planes(
1392           xd, 0, yv12_alt, mi_row, mi_col,
1393           scaled_ref_alt ? NULL : get_ref_scale_factors(cm, ALTREF_FRAME),
1394           num_planes);
1395       *y_sad_alt = cpi->ppi->fn_ptr[bsize].sdf(
1396           x->plane[AOM_PLANE_Y].src.buf, x->plane[AOM_PLANE_Y].src.stride,
1397           xd->plane[AOM_PLANE_Y].pre[0].buf,
1398           xd->plane[AOM_PLANE_Y].pre[0].stride);
1399     }
1400   }
1401 
1402   if (use_last_ref) {
1403     const int source_sad_nonrd = x->content_state_sb.source_sad_nonrd;
1404     av1_setup_pre_planes(
1405         xd, 0, yv12, mi_row, mi_col,
1406         scaled_ref_last ? NULL : get_ref_scale_factors(cm, LAST_FRAME),
1407         num_planes);
1408     mi->ref_frame[0] = LAST_FRAME;
1409     mi->ref_frame[1] = NONE_FRAME;
1410     mi->bsize = cm->seq_params->sb_size;
1411     mi->mv[0].as_int = 0;
1412     mi->interp_filters = av1_broadcast_interp_filter(BILINEAR);
1413 
1414     int est_motion = cpi->sf.rt_sf.estimate_motion_for_var_based_partition;
1415     // TODO(b/290596301): Look into adjusting this condition.
1416     // There is regression on color content when
1417     // estimate_motion_for_var_based_partition = 3 and high motion,
1418     // so for now force it to 2 based on superblock sad.
1419     if (est_motion > 2 && source_sad_nonrd > kMedSad) est_motion = 2;
1420 
1421     if (est_motion == 1 || est_motion == 2) {
1422       if (xd->mb_to_right_edge >= 0 && xd->mb_to_bottom_edge >= 0) {
1423         // For screen only do int_pro_motion for spatial variance above
1424         // threshold and motion level above LowSad.
1425         if (x->source_variance > 100 && source_sad_nonrd > kLowSad) {
1426           int is_screen = cpi->oxcf.tune_cfg.content == AOM_CONTENT_SCREEN;
1427           int me_search_size_col =
1428               is_screen ? 96 : block_size_wide[cm->seq_params->sb_size] >> 1;
1429           // For screen use larger search size row motion to capture
1430           // vertical scroll, which can be larger motion.
1431           int me_search_size_row =
1432               is_screen ? 192 : block_size_high[cm->seq_params->sb_size] >> 1;
1433           unsigned int y_sad_zero;
1434           *y_sad = av1_int_pro_motion_estimation(
1435               cpi, x, cm->seq_params->sb_size, mi_row, mi_col, &kZeroMv,
1436               &y_sad_zero, me_search_size_col, me_search_size_row);
1437           // The logic below selects whether the motion estimated in the
1438           // int_pro_motion() will be used in nonrd_pickmode. Only do this
1439           // for screen for now.
1440           if (is_screen) {
1441             unsigned int thresh_sad =
1442                 (cm->seq_params->sb_size == BLOCK_128X128) ? 50000 : 20000;
1443             if (*y_sad < (y_sad_zero >> 1) && *y_sad < thresh_sad) {
1444               x->sb_me_partition = 1;
1445               x->sb_me_mv.as_int = mi->mv[0].as_int;
1446             } else {
1447               x->sb_me_partition = 0;
1448               // Fall back to using zero motion.
1449               *y_sad = y_sad_zero;
1450               mi->mv[0].as_int = 0;
1451             }
1452           }
1453         }
1454       }
1455     }
1456 
1457     if (*y_sad == UINT_MAX) {
1458       *y_sad = cpi->ppi->fn_ptr[bsize].sdf(
1459           x->plane[AOM_PLANE_Y].src.buf, x->plane[AOM_PLANE_Y].src.stride,
1460           xd->plane[AOM_PLANE_Y].pre[0].buf,
1461           xd->plane[AOM_PLANE_Y].pre[0].stride);
1462     }
1463 
1464     // Evaluate if neighbours' MVs give better predictions. Zero MV is tested
1465     // already, so only non-zero MVs are tested here. Here the neighbour blocks
1466     // are the first block above or left to this superblock.
1467     if (est_motion >= 2 && (xd->up_available || xd->left_available))
1468       evaluate_neighbour_mvs(cpi, x, y_sad, is_small_sb, est_motion);
1469 
1470     *y_sad_last = *y_sad;
1471   }
1472 
1473   // Pick the ref frame for partitioning, use golden or altref frame only if
1474   // its lower sad, bias to LAST with factor 0.9.
1475   set_ref_frame_for_partition(cpi, x, xd, ref_frame_partition, mi, y_sad,
1476                               y_sad_g, y_sad_alt, yv12_g, yv12_alt, mi_row,
1477                               mi_col, num_planes);
1478 
1479   // Only calculate the predictor for non-zero MV.
1480   if (mi->mv[0].as_int != 0) {
1481     if (!scaled_ref_last) {
1482       set_ref_ptrs(cm, xd, mi->ref_frame[0], mi->ref_frame[1]);
1483     } else {
1484       xd->block_ref_scale_factors[0] = sf_no_scale;
1485       xd->block_ref_scale_factors[1] = sf_no_scale;
1486     }
1487     av1_enc_build_inter_predictor(cm, xd, mi_row, mi_col, NULL,
1488                                   cm->seq_params->sb_size, AOM_PLANE_Y,
1489                                   num_planes - 1);
1490   }
1491 }
1492 
1493 // Decides whether to split or merge a 16x16 partition block in variance based
1494 // partitioning based on the 8x8 sub-block variances.
get_part_eval_based_on_sub_blk_var(VP16x16 * var_16x16_info,int64_t threshold16)1495 static inline PART_EVAL_STATUS get_part_eval_based_on_sub_blk_var(
1496     VP16x16 *var_16x16_info, int64_t threshold16) {
1497   int max_8x8_var = 0, min_8x8_var = INT_MAX;
1498   for (int split_idx = 0; split_idx < 4; split_idx++) {
1499     get_variance(&var_16x16_info->split[split_idx].part_variances.none);
1500     int this_8x8_var =
1501         var_16x16_info->split[split_idx].part_variances.none.variance;
1502     max_8x8_var = AOMMAX(this_8x8_var, max_8x8_var);
1503     min_8x8_var = AOMMIN(this_8x8_var, min_8x8_var);
1504   }
1505   // If the difference between maximum and minimum sub-block variances is high,
1506   // then only evaluate PARTITION_SPLIT for the 16x16 block. Otherwise, evaluate
1507   // only PARTITION_NONE. The shift factor for threshold16 has been derived
1508   // empirically.
1509   return ((max_8x8_var - min_8x8_var) > (threshold16 << 2))
1510              ? PART_EVAL_ONLY_SPLIT
1511              : PART_EVAL_ONLY_NONE;
1512 }
1513 
is_set_force_zeromv_skip_based_on_src_sad(int set_zeromv_skip_based_on_source_sad,SOURCE_SAD source_sad_nonrd)1514 static inline bool is_set_force_zeromv_skip_based_on_src_sad(
1515     int set_zeromv_skip_based_on_source_sad, SOURCE_SAD source_sad_nonrd) {
1516   if (set_zeromv_skip_based_on_source_sad == 0) return false;
1517 
1518   if (set_zeromv_skip_based_on_source_sad >= 3)
1519     return source_sad_nonrd <= kLowSad;
1520   else if (set_zeromv_skip_based_on_source_sad >= 2)
1521     return source_sad_nonrd <= kVeryLowSad;
1522   else if (set_zeromv_skip_based_on_source_sad >= 1)
1523     return source_sad_nonrd == kZeroSad;
1524 
1525   return false;
1526 }
1527 
set_force_zeromv_skip_for_sb(AV1_COMP * cpi,MACROBLOCK * x,const TileInfo * const tile,VP128x128 * vt,unsigned int * uv_sad,int mi_row,int mi_col,unsigned int y_sad,BLOCK_SIZE bsize)1528 static inline bool set_force_zeromv_skip_for_sb(
1529     AV1_COMP *cpi, MACROBLOCK *x, const TileInfo *const tile, VP128x128 *vt,
1530     unsigned int *uv_sad, int mi_row, int mi_col, unsigned int y_sad,
1531     BLOCK_SIZE bsize) {
1532   AV1_COMMON *const cm = &cpi->common;
1533   if (!is_set_force_zeromv_skip_based_on_src_sad(
1534           cpi->sf.rt_sf.set_zeromv_skip_based_on_source_sad,
1535           x->content_state_sb.source_sad_nonrd))
1536     return false;
1537   int shift = cpi->sf.rt_sf.increase_source_sad_thresh ? 1 : 0;
1538   const int block_width = mi_size_wide[cm->seq_params->sb_size];
1539   const int block_height = mi_size_high[cm->seq_params->sb_size];
1540   const unsigned int thresh_exit_part_y =
1541       cpi->zeromv_skip_thresh_exit_part[bsize] << shift;
1542   unsigned int thresh_exit_part_uv =
1543       CALC_CHROMA_THRESH_FOR_ZEROMV_SKIP(thresh_exit_part_y) << shift;
1544   // Be more aggressive in UV threshold if source_sad >= VeryLowSad
1545   // to suppreess visual artifact caused by the speed feature:
1546   // set_zeromv_skip_based_on_source_sad = 2. For now only for
1547   // part_early_exit_zeromv = 1.
1548   if (x->content_state_sb.source_sad_nonrd >= kVeryLowSad &&
1549       cpi->sf.rt_sf.part_early_exit_zeromv == 1)
1550     thresh_exit_part_uv = thresh_exit_part_uv >> 3;
1551   if (mi_col + block_width <= tile->mi_col_end &&
1552       mi_row + block_height <= tile->mi_row_end && y_sad < thresh_exit_part_y &&
1553       uv_sad[0] < thresh_exit_part_uv && uv_sad[1] < thresh_exit_part_uv) {
1554     set_block_size(cpi, mi_row, mi_col, bsize);
1555     x->force_zeromv_skip_for_sb = 1;
1556     aom_free(vt);
1557     // Partition shape is set here at SB level.
1558     // Exit needs to happen from av1_choose_var_based_partitioning().
1559     return true;
1560   } else if (x->content_state_sb.source_sad_nonrd == kZeroSad &&
1561              cpi->sf.rt_sf.part_early_exit_zeromv >= 2)
1562     x->force_zeromv_skip_for_sb = 2;
1563   return false;
1564 }
1565 
av1_choose_var_based_partitioning(AV1_COMP * cpi,const TileInfo * const tile,ThreadData * td,MACROBLOCK * x,int mi_row,int mi_col)1566 int av1_choose_var_based_partitioning(AV1_COMP *cpi, const TileInfo *const tile,
1567                                       ThreadData *td, MACROBLOCK *x, int mi_row,
1568                                       int mi_col) {
1569 #if CONFIG_COLLECT_COMPONENT_TIMING
1570   start_timing(cpi, choose_var_based_partitioning_time);
1571 #endif
1572   AV1_COMMON *const cm = &cpi->common;
1573   MACROBLOCKD *xd = &x->e_mbd;
1574   const int64_t *const vbp_thresholds = cpi->vbp_info.thresholds;
1575   PART_EVAL_STATUS force_split[85];
1576   int avg_64x64;
1577   int max_var_32x32[4];
1578   int min_var_32x32[4];
1579   int var_32x32;
1580   int var_64x64;
1581   int min_var_64x64 = INT_MAX;
1582   int max_var_64x64 = 0;
1583   int avg_16x16[4][4];
1584   int maxvar_16x16[4][4];
1585   int minvar_16x16[4][4];
1586   const uint8_t *src_buf;
1587   const uint8_t *dst_buf;
1588   int dst_stride;
1589   unsigned int uv_sad[MAX_MB_PLANE - 1];
1590   NOISE_LEVEL noise_level = kLow;
1591   bool is_zero_motion = true;
1592   bool scaled_ref_last = false;
1593   struct scale_factors sf_no_scale;
1594   av1_setup_scale_factors_for_frame(&sf_no_scale, cm->width, cm->height,
1595                                     cm->width, cm->height);
1596 
1597   bool is_key_frame =
1598       (frame_is_intra_only(cm) ||
1599        (cpi->ppi->use_svc &&
1600         cpi->svc.layer_context[cpi->svc.temporal_layer_id].is_key_frame));
1601 
1602   assert(cm->seq_params->sb_size == BLOCK_64X64 ||
1603          cm->seq_params->sb_size == BLOCK_128X128);
1604   const bool is_small_sb = (cm->seq_params->sb_size == BLOCK_64X64);
1605   const int num_64x64_blocks = is_small_sb ? 1 : 4;
1606 
1607   unsigned int y_sad = UINT_MAX;
1608   unsigned int y_sad_g = UINT_MAX;
1609   unsigned int y_sad_alt = UINT_MAX;
1610   unsigned int y_sad_last = UINT_MAX;
1611   BLOCK_SIZE bsize = is_small_sb ? BLOCK_64X64 : BLOCK_128X128;
1612 
1613   // Force skip encoding for all superblocks on slide change for
1614   // non_reference_frames.
1615   if (cpi->sf.rt_sf.skip_encoding_non_reference_slide_change &&
1616       cpi->rc.high_source_sad && cpi->ppi->rtc_ref.non_reference_frame) {
1617     MB_MODE_INFO **mi = cm->mi_params.mi_grid_base +
1618                         get_mi_grid_idx(&cm->mi_params, mi_row, mi_col);
1619     av1_set_fixed_partitioning(cpi, tile, mi, mi_row, mi_col, bsize);
1620     x->force_zeromv_skip_for_sb = 1;
1621     return 0;
1622   }
1623 
1624   // Ref frame used in partitioning.
1625   MV_REFERENCE_FRAME ref_frame_partition = LAST_FRAME;
1626 
1627   int64_t thresholds[5] = { vbp_thresholds[0], vbp_thresholds[1],
1628                             vbp_thresholds[2], vbp_thresholds[3],
1629                             vbp_thresholds[4] };
1630 
1631   const int segment_id = xd->mi[0]->segment_id;
1632   uint64_t blk_sad = 0;
1633   if (cpi->src_sad_blk_64x64 != NULL &&
1634       cpi->svc.spatial_layer_id == cpi->svc.number_spatial_layers - 1) {
1635     const int sb_size_by_mb = (cm->seq_params->sb_size == BLOCK_128X128)
1636                                   ? (cm->seq_params->mib_size >> 1)
1637                                   : cm->seq_params->mib_size;
1638     const int sb_cols =
1639         (cm->mi_params.mi_cols + sb_size_by_mb - 1) / sb_size_by_mb;
1640     const int sbi_col = mi_col / sb_size_by_mb;
1641     const int sbi_row = mi_row / sb_size_by_mb;
1642     blk_sad = cpi->src_sad_blk_64x64[sbi_col + sbi_row * sb_cols];
1643   }
1644 
1645   const bool is_segment_id_boosted =
1646       cpi->oxcf.q_cfg.aq_mode == CYCLIC_REFRESH_AQ && cm->seg.enabled &&
1647       cyclic_refresh_segment_id_boosted(segment_id);
1648   const int qindex =
1649       is_segment_id_boosted
1650           ? av1_get_qindex(&cm->seg, segment_id, cm->quant_params.base_qindex)
1651           : cm->quant_params.base_qindex;
1652   set_vbp_thresholds(
1653       cpi, thresholds, blk_sad, qindex, x->content_state_sb.low_sumdiff,
1654       x->content_state_sb.source_sad_nonrd, x->content_state_sb.source_sad_rd,
1655       is_segment_id_boosted, x->content_state_sb.lighting_change);
1656 
1657   src_buf = x->plane[AOM_PLANE_Y].src.buf;
1658   int src_stride = x->plane[AOM_PLANE_Y].src.stride;
1659 
1660   // Index for force_split: 0 for 64x64, 1-4 for 32x32 blocks,
1661   // 5-20 for the 16x16 blocks.
1662   force_split[0] = PART_EVAL_ALL;
1663   memset(x->part_search_info.variance_low, 0,
1664          sizeof(x->part_search_info.variance_low));
1665 
1666   // Check if LAST frame is NULL, and if so, treat this frame
1667   // as a key frame, for the purpose of the superblock partitioning.
1668   // LAST == NULL can happen in cases where enhancement spatial layers are
1669   // enabled dyanmically and the only reference is the spatial(GOLDEN).
1670   // If LAST frame has a different resolution: set the scaled_ref_last flag
1671   // and check if ref_scaled is NULL.
1672   if (!frame_is_intra_only(cm)) {
1673     const YV12_BUFFER_CONFIG *ref = get_ref_frame_yv12_buf(cm, LAST_FRAME);
1674     if (ref == NULL) {
1675       is_key_frame = true;
1676     } else if (ref->y_crop_height != cm->height ||
1677                ref->y_crop_width != cm->width) {
1678       scaled_ref_last = true;
1679       const YV12_BUFFER_CONFIG *ref_scaled =
1680           av1_get_scaled_ref_frame(cpi, LAST_FRAME);
1681       if (ref_scaled == NULL) is_key_frame = true;
1682     }
1683   }
1684 
1685   x->source_variance = UINT_MAX;
1686   // For nord_pickmode: compute source_variance, only for superblocks with
1687   // some motion for now. This input can then be used to bias the partitioning
1688   // or the chroma_check.
1689   if (cpi->sf.rt_sf.use_nonrd_pick_mode &&
1690       x->content_state_sb.source_sad_nonrd > kLowSad)
1691     x->source_variance = av1_get_perpixel_variance_facade(
1692         cpi, xd, &x->plane[0].src, cm->seq_params->sb_size, AOM_PLANE_Y);
1693 
1694   if (!is_key_frame) {
1695     setup_planes(cpi, x, &y_sad, &y_sad_g, &y_sad_alt, &y_sad_last,
1696                  &ref_frame_partition, &sf_no_scale, mi_row, mi_col,
1697                  is_small_sb, scaled_ref_last);
1698 
1699     MB_MODE_INFO *mi = xd->mi[0];
1700     // Use reference SB directly for zero mv.
1701     if (mi->mv[0].as_int != 0) {
1702       dst_buf = xd->plane[AOM_PLANE_Y].dst.buf;
1703       dst_stride = xd->plane[AOM_PLANE_Y].dst.stride;
1704       is_zero_motion = false;
1705     } else {
1706       dst_buf = xd->plane[AOM_PLANE_Y].pre[0].buf;
1707       dst_stride = xd->plane[AOM_PLANE_Y].pre[0].stride;
1708     }
1709   } else {
1710     dst_buf = NULL;
1711     dst_stride = 0;
1712   }
1713 
1714   // check and set the color sensitivity of sb.
1715   av1_zero(uv_sad);
1716   chroma_check(cpi, x, bsize, y_sad_last, y_sad_g, y_sad_alt, is_key_frame,
1717                is_zero_motion, uv_sad);
1718 
1719   x->force_zeromv_skip_for_sb = 0;
1720 
1721   VP128x128 *vt;
1722   AOM_CHECK_MEM_ERROR(xd->error_info, vt, aom_malloc(sizeof(*vt)));
1723   vt->split = td->vt64x64;
1724 
1725   // If the superblock is completely static (zero source sad) and
1726   // the y_sad (relative to LAST ref) is very small, take the sb_size partition
1727   // and exit, and force zeromv_last skip mode for nonrd_pickmode.
1728   // Only do this on the base segment (so the QP-boosted segment, if applied,
1729   // can still continue cleaning/ramping up the quality).
1730   // Condition on color uv_sad is also added.
1731   if (!is_key_frame && cpi->sf.rt_sf.part_early_exit_zeromv &&
1732       cpi->rc.frames_since_key > 30 && segment_id == CR_SEGMENT_ID_BASE &&
1733       ref_frame_partition == LAST_FRAME && xd->mi[0]->mv[0].as_int == 0) {
1734     // Exit here, if zero mv skip flag is set at SB level.
1735     if (set_force_zeromv_skip_for_sb(cpi, x, tile, vt, uv_sad, mi_row, mi_col,
1736                                      y_sad, bsize))
1737       return 0;
1738   }
1739 
1740   if (cpi->noise_estimate.enabled)
1741     noise_level = av1_noise_estimate_extract_level(&cpi->noise_estimate);
1742 
1743   // Fill in the entire tree of 8x8 (for inter frames) or 4x4 (for key frames)
1744   // variances for splits.
1745   fill_variance_tree_leaves(cpi, x, vt, force_split, avg_16x16, maxvar_16x16,
1746                             minvar_16x16, thresholds, src_buf, src_stride,
1747                             dst_buf, dst_stride, is_key_frame, is_small_sb);
1748 
1749   avg_64x64 = 0;
1750   for (int blk64_idx = 0; blk64_idx < num_64x64_blocks; ++blk64_idx) {
1751     max_var_32x32[blk64_idx] = 0;
1752     min_var_32x32[blk64_idx] = INT_MAX;
1753     const int blk64_scale_idx = blk64_idx << 2;
1754     for (int lvl1_idx = 0; lvl1_idx < 4; lvl1_idx++) {
1755       const int lvl1_scale_idx = (blk64_scale_idx + lvl1_idx) << 2;
1756       for (int lvl2_idx = 0; lvl2_idx < 4; lvl2_idx++) {
1757         if (!is_key_frame) continue;
1758         VP16x16 *vtemp = &vt->split[blk64_idx].split[lvl1_idx].split[lvl2_idx];
1759         for (int lvl3_idx = 0; lvl3_idx < 4; lvl3_idx++)
1760           fill_variance_tree(&vtemp->split[lvl3_idx], BLOCK_8X8);
1761         fill_variance_tree(vtemp, BLOCK_16X16);
1762         // If variance of this 16x16 block is above the threshold, force block
1763         // to split. This also forces a split on the upper levels.
1764         get_variance(&vtemp->part_variances.none);
1765         if (vtemp->part_variances.none.variance > thresholds[3]) {
1766           const int split_index = 21 + lvl1_scale_idx + lvl2_idx;
1767           force_split[split_index] =
1768               cpi->sf.rt_sf.vbp_prune_16x16_split_using_min_max_sub_blk_var
1769                   ? get_part_eval_based_on_sub_blk_var(vtemp, thresholds[3])
1770                   : PART_EVAL_ONLY_SPLIT;
1771           force_split[5 + blk64_scale_idx + lvl1_idx] = PART_EVAL_ONLY_SPLIT;
1772           force_split[blk64_idx + 1] = PART_EVAL_ONLY_SPLIT;
1773           force_split[0] = PART_EVAL_ONLY_SPLIT;
1774         }
1775       }
1776       fill_variance_tree(&vt->split[blk64_idx].split[lvl1_idx], BLOCK_32X32);
1777       // If variance of this 32x32 block is above the threshold, or if its above
1778       // (some threshold of) the average variance over the sub-16x16 blocks,
1779       // then force this block to split. This also forces a split on the upper
1780       // (64x64) level.
1781       uint64_t frame_sad_thresh = 20000;
1782       const int is_360p_or_smaller = cm->width * cm->height <= RESOLUTION_360P;
1783       if (cpi->svc.number_temporal_layers > 2 &&
1784           cpi->svc.temporal_layer_id == 0)
1785         frame_sad_thresh = frame_sad_thresh << 1;
1786       if (force_split[5 + blk64_scale_idx + lvl1_idx] == PART_EVAL_ALL) {
1787         get_variance(&vt->split[blk64_idx].split[lvl1_idx].part_variances.none);
1788         var_32x32 =
1789             vt->split[blk64_idx].split[lvl1_idx].part_variances.none.variance;
1790         max_var_32x32[blk64_idx] = AOMMAX(var_32x32, max_var_32x32[blk64_idx]);
1791         min_var_32x32[blk64_idx] = AOMMIN(var_32x32, min_var_32x32[blk64_idx]);
1792         const int max_min_var_16X16_diff = (maxvar_16x16[blk64_idx][lvl1_idx] -
1793                                             minvar_16x16[blk64_idx][lvl1_idx]);
1794 
1795         if (var_32x32 > thresholds[2] ||
1796             (!is_key_frame && var_32x32 > (thresholds[2] >> 1) &&
1797              var_32x32 > (avg_16x16[blk64_idx][lvl1_idx] >> 1))) {
1798           force_split[5 + blk64_scale_idx + lvl1_idx] = PART_EVAL_ONLY_SPLIT;
1799           force_split[blk64_idx + 1] = PART_EVAL_ONLY_SPLIT;
1800           force_split[0] = PART_EVAL_ONLY_SPLIT;
1801         } else if (!is_key_frame && is_360p_or_smaller &&
1802                    ((max_min_var_16X16_diff > (thresholds[2] >> 1) &&
1803                      maxvar_16x16[blk64_idx][lvl1_idx] > thresholds[2]) ||
1804                     (cpi->sf.rt_sf.prefer_large_partition_blocks &&
1805                      x->content_state_sb.source_sad_nonrd > kLowSad &&
1806                      cpi->rc.frame_source_sad < frame_sad_thresh &&
1807                      maxvar_16x16[blk64_idx][lvl1_idx] > (thresholds[2] >> 4) &&
1808                      maxvar_16x16[blk64_idx][lvl1_idx] >
1809                          (minvar_16x16[blk64_idx][lvl1_idx] << 2)))) {
1810           force_split[5 + blk64_scale_idx + lvl1_idx] = PART_EVAL_ONLY_SPLIT;
1811           force_split[blk64_idx + 1] = PART_EVAL_ONLY_SPLIT;
1812           force_split[0] = PART_EVAL_ONLY_SPLIT;
1813         }
1814       }
1815     }
1816     if (force_split[1 + blk64_idx] == PART_EVAL_ALL) {
1817       fill_variance_tree(&vt->split[blk64_idx], BLOCK_64X64);
1818       get_variance(&vt->split[blk64_idx].part_variances.none);
1819       var_64x64 = vt->split[blk64_idx].part_variances.none.variance;
1820       max_var_64x64 = AOMMAX(var_64x64, max_var_64x64);
1821       min_var_64x64 = AOMMIN(var_64x64, min_var_64x64);
1822       // If the difference of the max-min variances of sub-blocks or max
1823       // variance of a sub-block is above some threshold of then force this
1824       // block to split. Only checking this for noise level >= medium, if
1825       // encoder is in SVC or if we already forced large blocks.
1826       const int max_min_var_32x32_diff =
1827           max_var_32x32[blk64_idx] - min_var_32x32[blk64_idx];
1828       const int check_max_var = max_var_32x32[blk64_idx] > thresholds[1] >> 1;
1829       const bool check_noise_lvl = noise_level >= kMedium ||
1830                                    cpi->ppi->use_svc ||
1831                                    cpi->sf.rt_sf.prefer_large_partition_blocks;
1832       const int64_t set_threshold = 3 * (thresholds[1] >> 3);
1833 
1834       if (!is_key_frame && max_min_var_32x32_diff > set_threshold &&
1835           check_max_var && check_noise_lvl) {
1836         force_split[1 + blk64_idx] = PART_EVAL_ONLY_SPLIT;
1837         force_split[0] = PART_EVAL_ONLY_SPLIT;
1838       }
1839       avg_64x64 += var_64x64;
1840     }
1841     if (is_small_sb) force_split[0] = PART_EVAL_ONLY_SPLIT;
1842   }
1843 
1844   if (force_split[0] == PART_EVAL_ALL) {
1845     fill_variance_tree(vt, BLOCK_128X128);
1846     get_variance(&vt->part_variances.none);
1847     const int set_avg_64x64 = (9 * avg_64x64) >> 5;
1848     if (!is_key_frame && vt->part_variances.none.variance > set_avg_64x64)
1849       force_split[0] = PART_EVAL_ONLY_SPLIT;
1850 
1851     if (!is_key_frame &&
1852         (max_var_64x64 - min_var_64x64) > 3 * (thresholds[0] >> 3) &&
1853         max_var_64x64 > thresholds[0] >> 1)
1854       force_split[0] = PART_EVAL_ONLY_SPLIT;
1855   }
1856 
1857   if (mi_col + 32 > tile->mi_col_end || mi_row + 32 > tile->mi_row_end ||
1858       !set_vt_partitioning(cpi, xd, tile, vt, BLOCK_128X128, mi_row, mi_col,
1859                            thresholds[0], BLOCK_16X16, force_split[0])) {
1860     for (int blk64_idx = 0; blk64_idx < num_64x64_blocks; ++blk64_idx) {
1861       const int x64_idx = GET_BLK_IDX_X(blk64_idx, 4);
1862       const int y64_idx = GET_BLK_IDX_Y(blk64_idx, 4);
1863       const int blk64_scale_idx = blk64_idx << 2;
1864 
1865       // Now go through the entire structure, splitting every block size until
1866       // we get to one that's got a variance lower than our threshold.
1867       if (set_vt_partitioning(cpi, xd, tile, &vt->split[blk64_idx], BLOCK_64X64,
1868                               mi_row + y64_idx, mi_col + x64_idx, thresholds[1],
1869                               BLOCK_16X16, force_split[1 + blk64_idx]))
1870         continue;
1871       for (int lvl1_idx = 0; lvl1_idx < 4; ++lvl1_idx) {
1872         const int x32_idx = GET_BLK_IDX_X(lvl1_idx, 3);
1873         const int y32_idx = GET_BLK_IDX_Y(lvl1_idx, 3);
1874         const int lvl1_scale_idx = (blk64_scale_idx + lvl1_idx) << 2;
1875         if (set_vt_partitioning(
1876                 cpi, xd, tile, &vt->split[blk64_idx].split[lvl1_idx],
1877                 BLOCK_32X32, (mi_row + y64_idx + y32_idx),
1878                 (mi_col + x64_idx + x32_idx), thresholds[2], BLOCK_16X16,
1879                 force_split[5 + blk64_scale_idx + lvl1_idx]))
1880           continue;
1881         for (int lvl2_idx = 0; lvl2_idx < 4; ++lvl2_idx) {
1882           const int x16_idx = GET_BLK_IDX_X(lvl2_idx, 2);
1883           const int y16_idx = GET_BLK_IDX_Y(lvl2_idx, 2);
1884           const int split_index = 21 + lvl1_scale_idx + lvl2_idx;
1885           VP16x16 *vtemp =
1886               &vt->split[blk64_idx].split[lvl1_idx].split[lvl2_idx];
1887           if (set_vt_partitioning(cpi, xd, tile, vtemp, BLOCK_16X16,
1888                                   mi_row + y64_idx + y32_idx + y16_idx,
1889                                   mi_col + x64_idx + x32_idx + x16_idx,
1890                                   thresholds[3], BLOCK_8X8,
1891                                   force_split[split_index]))
1892             continue;
1893           for (int lvl3_idx = 0; lvl3_idx < 4; ++lvl3_idx) {
1894             const int x8_idx = GET_BLK_IDX_X(lvl3_idx, 1);
1895             const int y8_idx = GET_BLK_IDX_Y(lvl3_idx, 1);
1896             set_block_size(cpi, (mi_row + y64_idx + y32_idx + y16_idx + y8_idx),
1897                            (mi_col + x64_idx + x32_idx + x16_idx + x8_idx),
1898                            BLOCK_8X8);
1899           }
1900         }
1901       }
1902     }
1903   }
1904 
1905   if (cpi->sf.rt_sf.short_circuit_low_temp_var) {
1906     set_low_temp_var_flag(cpi, &x->part_search_info, xd, vt, thresholds,
1907                           ref_frame_partition, mi_col, mi_row, is_small_sb);
1908   }
1909 
1910   aom_free(vt);
1911 #if CONFIG_COLLECT_COMPONENT_TIMING
1912   end_timing(cpi, choose_var_based_partitioning_time);
1913 #endif
1914   return 0;
1915 }
1916