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/var_based_part.h"
31 #include "av1/encoder/reconinter_enc.h"
32 #include "av1/encoder/rdopt_utils.h"
33
34 // Possible values for the force_split variable while evaluating variance based
35 // partitioning.
36 enum {
37 // Evaluate all partition types
38 PART_EVAL_ALL = 0,
39 // Force PARTITION_SPLIT
40 PART_EVAL_ONLY_SPLIT = 1,
41 // Force PARTITION_NONE
42 PART_EVAL_ONLY_NONE = 2
43 } UENUM1BYTE(PART_EVAL_STATUS);
44
45 typedef struct {
46 VPVariance *part_variances;
47 VPartVar *split[4];
48 } variance_node;
49
tree_to_node(void * data,BLOCK_SIZE bsize,variance_node * node)50 static AOM_INLINE void tree_to_node(void *data, BLOCK_SIZE bsize,
51 variance_node *node) {
52 node->part_variances = NULL;
53 switch (bsize) {
54 case BLOCK_128X128: {
55 VP128x128 *vt = (VP128x128 *)data;
56 node->part_variances = &vt->part_variances;
57 for (int split_idx = 0; split_idx < 4; split_idx++)
58 node->split[split_idx] = &vt->split[split_idx].part_variances.none;
59 break;
60 }
61 case BLOCK_64X64: {
62 VP64x64 *vt = (VP64x64 *)data;
63 node->part_variances = &vt->part_variances;
64 for (int split_idx = 0; split_idx < 4; split_idx++)
65 node->split[split_idx] = &vt->split[split_idx].part_variances.none;
66 break;
67 }
68 case BLOCK_32X32: {
69 VP32x32 *vt = (VP32x32 *)data;
70 node->part_variances = &vt->part_variances;
71 for (int split_idx = 0; split_idx < 4; split_idx++)
72 node->split[split_idx] = &vt->split[split_idx].part_variances.none;
73 break;
74 }
75 case BLOCK_16X16: {
76 VP16x16 *vt = (VP16x16 *)data;
77 node->part_variances = &vt->part_variances;
78 for (int split_idx = 0; split_idx < 4; split_idx++)
79 node->split[split_idx] = &vt->split[split_idx].part_variances.none;
80 break;
81 }
82 case BLOCK_8X8: {
83 VP8x8 *vt = (VP8x8 *)data;
84 node->part_variances = &vt->part_variances;
85 for (int split_idx = 0; split_idx < 4; split_idx++)
86 node->split[split_idx] = &vt->split[split_idx].part_variances.none;
87 break;
88 }
89 default: {
90 VP4x4 *vt = (VP4x4 *)data;
91 assert(bsize == BLOCK_4X4);
92 node->part_variances = &vt->part_variances;
93 for (int split_idx = 0; split_idx < 4; split_idx++)
94 node->split[split_idx] = &vt->split[split_idx];
95 break;
96 }
97 }
98 }
99
100 // Set variance values given sum square error, sum error, count.
fill_variance(uint32_t s2,int32_t s,int c,VPartVar * v)101 static AOM_INLINE void fill_variance(uint32_t s2, int32_t s, int c,
102 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 AOM_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 AOM_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 AOM_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 AOM_INLINE void set_block_size(AV1_COMP *const cpi, int mi_row,
136 int mi_col, 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 AOM_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 AOM_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 AOM_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 AOM_INLINE void fill_variance_8x8avg(
330 const uint8_t *src_buf, int src_stride, const uint8_t *dst_buf,
331 int dst_stride, int x16_idx, int y16_idx, VP16x16 *vst, int highbd_flag,
332 int pixels_wide, int pixels_high) {
333 #if CONFIG_AV1_HIGHBITDEPTH
334 if (highbd_flag) {
335 fill_variance_8x8avg_highbd(src_buf, src_stride, dst_buf, dst_stride,
336 x16_idx, y16_idx, vst, pixels_wide,
337 pixels_high);
338 return;
339 }
340 #else
341 (void)highbd_flag;
342 #endif // CONFIG_AV1_HIGHBITDEPTH
343 fill_variance_8x8avg_lowbd(src_buf, src_stride, dst_buf, dst_stride, x16_idx,
344 y16_idx, vst, pixels_wide, pixels_high);
345 }
346
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)347 static int compute_minmax_8x8(const uint8_t *src_buf, int src_stride,
348 const uint8_t *dst_buf, int dst_stride,
349 int x16_idx, int y16_idx,
350 #if CONFIG_AV1_HIGHBITDEPTH
351 int highbd_flag,
352 #endif
353 int pixels_wide, int pixels_high) {
354 int minmax_max = 0;
355 int minmax_min = 255;
356 // Loop over the 4 8x8 subblocks.
357 for (int idx = 0; idx < 4; idx++) {
358 const int x8_idx = x16_idx + GET_BLK_IDX_X(idx, 3);
359 const int y8_idx = y16_idx + GET_BLK_IDX_Y(idx, 3);
360 int min = 0;
361 int max = 0;
362 if (x8_idx < pixels_wide && y8_idx < pixels_high) {
363 #if CONFIG_AV1_HIGHBITDEPTH
364 if (highbd_flag & YV12_FLAG_HIGHBITDEPTH) {
365 aom_highbd_minmax_8x8(
366 src_buf + y8_idx * src_stride + x8_idx, src_stride,
367 dst_buf + y8_idx * dst_stride + x8_idx, dst_stride, &min, &max);
368 } else {
369 aom_minmax_8x8(src_buf + y8_idx * src_stride + x8_idx, src_stride,
370 dst_buf + y8_idx * dst_stride + x8_idx, dst_stride, &min,
371 &max);
372 }
373 #else
374 aom_minmax_8x8(src_buf + y8_idx * src_stride + x8_idx, src_stride,
375 dst_buf + y8_idx * dst_stride + x8_idx, dst_stride, &min,
376 &max);
377 #endif
378 if ((max - min) > minmax_max) minmax_max = (max - min);
379 if ((max - min) < minmax_min) minmax_min = (max - min);
380 }
381 }
382 return (minmax_max - minmax_min);
383 }
384
385 // Function to compute average and variance of 4x4 sub-block.
386 // The function can be called only when is_key_frame is true since sum is
387 // 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)388 static AOM_INLINE void fill_variance_4x4avg(const uint8_t *src_buf,
389 int src_stride, int x8_idx,
390 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
424 // TODO(kyslov) Bring back threshold adjustment based on content state
scale_part_thresh_content(int64_t threshold_base,int speed,int width,int height,int non_reference_frame)425 static int64_t scale_part_thresh_content(int64_t threshold_base, int speed,
426 int width, int height,
427 int non_reference_frame) {
428 (void)width;
429 (void)height;
430 int64_t threshold = threshold_base;
431 if (non_reference_frame) threshold = (3 * threshold) >> 1;
432 if (speed >= 8) {
433 return (5 * threshold) >> 2;
434 }
435 return threshold;
436 }
437
438 // 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)439 static AOM_INLINE void tune_thresh_based_on_qindex(
440 AV1_COMP *cpi, int64_t thresholds[], uint64_t block_sad, int current_qindex,
441 int num_pixels, bool is_segment_id_boosted, int source_sad_nonrd,
442 int lighting_change) {
443 double weight;
444 if (cpi->sf.rt_sf.prefer_large_partition_blocks >= 3) {
445 const int win = 20;
446 if (current_qindex < QINDEX_LARGE_BLOCK_THR - win)
447 weight = 1.0;
448 else if (current_qindex > QINDEX_LARGE_BLOCK_THR + win)
449 weight = 0.0;
450 else
451 weight =
452 1.0 - (current_qindex - QINDEX_LARGE_BLOCK_THR + win) / (2 * win);
453 if (num_pixels > RESOLUTION_480P) {
454 for (int i = 0; i < 4; i++) {
455 thresholds[i] <<= 1;
456 }
457 }
458 if (num_pixels <= RESOLUTION_288P) {
459 thresholds[3] = INT64_MAX;
460 if (is_segment_id_boosted == false) {
461 thresholds[1] <<= 2;
462 thresholds[2] <<= (source_sad_nonrd <= kLowSad) ? 5 : 4;
463 } else {
464 thresholds[1] <<= 1;
465 thresholds[2] <<= 3;
466 }
467 // Allow for split to 8x8 for superblocks where part of it has
468 // moving boundary. So allow for sb with source_sad above threshold,
469 // and avoid very large source_sad or high source content, to avoid
470 // too many 8x8 within superblock.
471 uint64_t avg_source_sad_thresh = 25000;
472 uint64_t block_sad_low = 25000;
473 uint64_t block_sad_high = 50000;
474 if (cpi->svc.temporal_layer_id == 0 &&
475 cpi->svc.number_temporal_layers > 1) {
476 // Increase the sad thresholds for base TL0, as reference/LAST is
477 // 2/4 frames behind (for 2/3 #TL).
478 avg_source_sad_thresh = 40000;
479 block_sad_high = 70000;
480 }
481 if (is_segment_id_boosted == false &&
482 cpi->rc.avg_source_sad < avg_source_sad_thresh &&
483 block_sad > block_sad_low && block_sad < block_sad_high &&
484 !lighting_change) {
485 thresholds[2] = (3 * thresholds[2]) >> 2;
486 thresholds[3] = thresholds[2] << 3;
487 }
488 // Condition the increase of partition thresholds on the segment
489 // and the content. Avoid the increase for superblocks which have
490 // high source sad, unless the whole frame has very high motion
491 // (i.e, cpi->rc.avg_source_sad is very large, in which case all blocks
492 // have high source sad).
493 } else if (num_pixels > RESOLUTION_480P && is_segment_id_boosted == false &&
494 (source_sad_nonrd != kHighSad ||
495 cpi->rc.avg_source_sad > 50000)) {
496 thresholds[0] = (3 * thresholds[0]) >> 1;
497 thresholds[3] = INT64_MAX;
498 if (current_qindex > QINDEX_LARGE_BLOCK_THR) {
499 thresholds[1] =
500 (int)((1 - weight) * (thresholds[1] << 1) + weight * thresholds[1]);
501 thresholds[2] =
502 (int)((1 - weight) * (thresholds[2] << 1) + weight * thresholds[2]);
503 }
504 } else if (current_qindex > QINDEX_LARGE_BLOCK_THR &&
505 is_segment_id_boosted == false &&
506 (source_sad_nonrd != kHighSad ||
507 cpi->rc.avg_source_sad > 50000)) {
508 thresholds[1] =
509 (int)((1 - weight) * (thresholds[1] << 2) + weight * thresholds[1]);
510 thresholds[2] =
511 (int)((1 - weight) * (thresholds[2] << 4) + weight * thresholds[2]);
512 thresholds[3] = INT64_MAX;
513 }
514 } else if (cpi->sf.rt_sf.prefer_large_partition_blocks >= 2) {
515 thresholds[1] <<= (source_sad_nonrd <= kLowSad) ? 2 : 0;
516 thresholds[2] =
517 (source_sad_nonrd <= kLowSad) ? (3 * thresholds[2]) : thresholds[2];
518 } else if (cpi->sf.rt_sf.prefer_large_partition_blocks >= 1) {
519 const int fac = (source_sad_nonrd <= kLowSad) ? 2 : 1;
520 if (current_qindex < QINDEX_LARGE_BLOCK_THR - 45)
521 weight = 1.0;
522 else if (current_qindex > QINDEX_LARGE_BLOCK_THR + 45)
523 weight = 0.0;
524 else
525 weight = 1.0 - (current_qindex - QINDEX_LARGE_BLOCK_THR + 45) / (2 * 45);
526 thresholds[1] =
527 (int)((1 - weight) * (thresholds[1] << 1) + weight * thresholds[1]);
528 thresholds[2] =
529 (int)((1 - weight) * (thresholds[2] << 1) + weight * thresholds[2]);
530 thresholds[3] =
531 (int)((1 - weight) * (thresholds[3] << fac) + weight * thresholds[3]);
532 }
533 if (cpi->sf.part_sf.disable_8x8_part_based_on_qidx && (current_qindex < 128))
534 thresholds[3] = INT64_MAX;
535 }
536
set_vbp_thresholds_key_frame(AV1_COMP * cpi,int64_t thresholds[],int64_t threshold_base,int threshold_left_shift,int num_pixels)537 static void set_vbp_thresholds_key_frame(AV1_COMP *cpi, int64_t thresholds[],
538 int64_t threshold_base,
539 int threshold_left_shift,
540 int num_pixels) {
541 if (cpi->sf.rt_sf.force_large_partition_blocks_intra) {
542 const int shift_steps =
543 threshold_left_shift - (cpi->oxcf.mode == ALLINTRA ? 7 : 8);
544 assert(shift_steps >= 0);
545 threshold_base <<= shift_steps;
546 }
547 thresholds[0] = threshold_base;
548 thresholds[1] = threshold_base;
549 if (num_pixels < RESOLUTION_720P) {
550 thresholds[2] = threshold_base / 3;
551 thresholds[3] = threshold_base >> 1;
552 } else {
553 int shift_val = 2;
554 if (cpi->sf.rt_sf.force_large_partition_blocks_intra) {
555 shift_val = 0;
556 }
557
558 thresholds[2] = threshold_base >> shift_val;
559 thresholds[3] = threshold_base >> shift_val;
560 }
561 thresholds[4] = threshold_base << 2;
562 }
563
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)564 static AOM_INLINE void tune_thresh_based_on_resolution(
565 AV1_COMP *cpi, int64_t thresholds[], int64_t threshold_base,
566 int current_qindex, int source_sad_rd, int num_pixels) {
567 if (num_pixels >= RESOLUTION_720P) thresholds[3] = thresholds[3] << 1;
568 if (num_pixels <= RESOLUTION_288P) {
569 const int qindex_thr[5][2] = {
570 { 200, 220 }, { 140, 170 }, { 120, 150 }, { 200, 210 }, { 170, 220 },
571 };
572 int th_idx = 0;
573 if (cpi->sf.rt_sf.var_part_based_on_qidx >= 1)
574 th_idx =
575 (source_sad_rd <= kLowSad) ? cpi->sf.rt_sf.var_part_based_on_qidx : 0;
576 if (cpi->sf.rt_sf.var_part_based_on_qidx >= 3)
577 th_idx = cpi->sf.rt_sf.var_part_based_on_qidx;
578 const int qindex_low_thr = qindex_thr[th_idx][0];
579 const int qindex_high_thr = qindex_thr[th_idx][1];
580 if (current_qindex >= qindex_high_thr) {
581 threshold_base = (5 * threshold_base) >> 1;
582 thresholds[1] = threshold_base >> 3;
583 thresholds[2] = threshold_base << 2;
584 thresholds[3] = threshold_base << 5;
585 } else if (current_qindex < qindex_low_thr) {
586 thresholds[1] = threshold_base >> 3;
587 thresholds[2] = threshold_base >> 1;
588 thresholds[3] = threshold_base << 3;
589 } else {
590 int64_t qi_diff_low = current_qindex - qindex_low_thr;
591 int64_t qi_diff_high = qindex_high_thr - current_qindex;
592 int64_t threshold_diff = qindex_high_thr - qindex_low_thr;
593 int64_t threshold_base_high = (5 * threshold_base) >> 1;
594
595 threshold_diff = threshold_diff > 0 ? threshold_diff : 1;
596 threshold_base =
597 (qi_diff_low * threshold_base_high + qi_diff_high * threshold_base) /
598 threshold_diff;
599 thresholds[1] = threshold_base >> 3;
600 thresholds[2] = ((qi_diff_low * threshold_base) +
601 qi_diff_high * (threshold_base >> 1)) /
602 threshold_diff;
603 thresholds[3] = ((qi_diff_low * (threshold_base << 5)) +
604 qi_diff_high * (threshold_base << 3)) /
605 threshold_diff;
606 }
607 } else if (num_pixels < RESOLUTION_720P) {
608 thresholds[2] = (5 * threshold_base) >> 2;
609 } else if (num_pixels < RESOLUTION_1080P) {
610 thresholds[2] = threshold_base << 1;
611 } else {
612 // num_pixels >= RESOLUTION_1080P
613 if (cpi->oxcf.tune_cfg.content == AOM_CONTENT_SCREEN) {
614 if (num_pixels < RESOLUTION_1440P) {
615 thresholds[2] = (5 * threshold_base) >> 1;
616 } else {
617 thresholds[2] = (7 * threshold_base) >> 1;
618 }
619 } else {
620 if (cpi->oxcf.speed > 7) {
621 thresholds[2] = 6 * threshold_base;
622 } else {
623 thresholds[2] = 3 * threshold_base;
624 }
625 }
626 }
627 }
628
629 // Increase partition thresholds for noisy content. Apply it only for
630 // superblocks where sumdiff is low, as we assume the sumdiff of superblock
631 // whose only change is due to noise will be low (i.e, noise will average
632 // out over large block).
tune_thresh_noisy_content(AV1_COMP * cpi,int64_t threshold_base,int content_lowsumdiff,int num_pixels)633 static AOM_INLINE int64_t tune_thresh_noisy_content(AV1_COMP *cpi,
634 int64_t threshold_base,
635 int content_lowsumdiff,
636 int num_pixels) {
637 AV1_COMMON *const cm = &cpi->common;
638 int64_t updated_thresh_base = threshold_base;
639 if (cpi->noise_estimate.enabled && content_lowsumdiff &&
640 num_pixels > RESOLUTION_480P && cm->current_frame.frame_number > 60) {
641 NOISE_LEVEL noise_level =
642 av1_noise_estimate_extract_level(&cpi->noise_estimate);
643 if (noise_level == kHigh)
644 updated_thresh_base = (5 * updated_thresh_base) >> 1;
645 else if (noise_level == kMedium &&
646 !cpi->sf.rt_sf.prefer_large_partition_blocks)
647 updated_thresh_base = (5 * updated_thresh_base) >> 2;
648 }
649 // TODO(kyslov) Enable var based partition adjusment on temporal denoising
650 #if 0 // CONFIG_AV1_TEMPORAL_DENOISING
651 if (cpi->oxcf.noise_sensitivity > 0 && denoise_svc(cpi) &&
652 cpi->oxcf.speed > 5 && cpi->denoiser.denoising_level >= kDenLow)
653 updated_thresh_base =
654 av1_scale_part_thresh(updated_thresh_base, cpi->denoiser.denoising_level,
655 content_state, cpi->svc.temporal_layer_id);
656 else
657 threshold_base =
658 scale_part_thresh_content(updated_thresh_base, cpi->oxcf.speed, cm->width,
659 cm->height, cpi->ppi->rtc_ref.non_reference_frame);
660 #else
661 // Increase base variance threshold based on content_state/sum_diff level.
662 updated_thresh_base = scale_part_thresh_content(
663 updated_thresh_base, cpi->oxcf.speed, cm->width, cm->height,
664 cpi->ppi->rtc_ref.non_reference_frame);
665 #endif
666 return updated_thresh_base;
667 }
668
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)669 static AOM_INLINE void set_vbp_thresholds(
670 AV1_COMP *cpi, int64_t thresholds[], uint64_t blk_sad, int qindex,
671 int content_lowsumdiff, int source_sad_nonrd, int source_sad_rd,
672 bool is_segment_id_boosted, int lighting_change) {
673 AV1_COMMON *const cm = &cpi->common;
674 const int is_key_frame = frame_is_intra_only(cm);
675 const int threshold_multiplier = is_key_frame ? 120 : 1;
676 const int ac_q = av1_ac_quant_QTX(qindex, 0, cm->seq_params->bit_depth);
677 int64_t threshold_base = (int64_t)(threshold_multiplier * ac_q);
678 const int current_qindex = cm->quant_params.base_qindex;
679 const int threshold_left_shift = cpi->sf.rt_sf.var_part_split_threshold_shift;
680 const int num_pixels = cm->width * cm->height;
681
682 if (is_key_frame) {
683 set_vbp_thresholds_key_frame(cpi, thresholds, threshold_base,
684 threshold_left_shift, num_pixels);
685 return;
686 }
687
688 threshold_base = tune_thresh_noisy_content(cpi, threshold_base,
689 content_lowsumdiff, num_pixels);
690 thresholds[0] = threshold_base >> 1;
691 thresholds[1] = threshold_base;
692 thresholds[3] = threshold_base << threshold_left_shift;
693
694 tune_thresh_based_on_resolution(cpi, thresholds, threshold_base,
695 current_qindex, source_sad_rd, num_pixels);
696
697 tune_thresh_based_on_qindex(cpi, thresholds, blk_sad, current_qindex,
698 num_pixels, is_segment_id_boosted,
699 source_sad_nonrd, lighting_change);
700 }
701
702 // Set temporal variance low flag for superblock 64x64.
703 // 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)704 static AOM_INLINE void set_low_temp_var_flag_64x64(
705 CommonModeInfoParams *mi_params, PartitionSearchInfo *part_info,
706 MACROBLOCKD *xd, VP64x64 *vt, const int64_t thresholds[], int mi_col,
707 int mi_row) {
708 if (xd->mi[0]->bsize == BLOCK_64X64) {
709 if ((vt->part_variances).none.variance < (thresholds[0] >> 1))
710 part_info->variance_low[0] = 1;
711 } else if (xd->mi[0]->bsize == BLOCK_64X32) {
712 for (int part_idx = 0; part_idx < 2; part_idx++) {
713 if (vt->part_variances.horz[part_idx].variance < (thresholds[0] >> 2))
714 part_info->variance_low[part_idx + 1] = 1;
715 }
716 } else if (xd->mi[0]->bsize == BLOCK_32X64) {
717 for (int part_idx = 0; part_idx < 2; part_idx++) {
718 if (vt->part_variances.vert[part_idx].variance < (thresholds[0] >> 2))
719 part_info->variance_low[part_idx + 3] = 1;
720 }
721 } else {
722 static const int idx[4][2] = { { 0, 0 }, { 0, 8 }, { 8, 0 }, { 8, 8 } };
723 for (int lvl1_idx = 0; lvl1_idx < 4; lvl1_idx++) {
724 const int idx_str = mi_params->mi_stride * (mi_row + idx[lvl1_idx][0]) +
725 mi_col + idx[lvl1_idx][1];
726 MB_MODE_INFO **this_mi = mi_params->mi_grid_base + idx_str;
727
728 if (mi_params->mi_cols <= mi_col + idx[lvl1_idx][1] ||
729 mi_params->mi_rows <= mi_row + idx[lvl1_idx][0])
730 continue;
731
732 if (*this_mi == NULL) continue;
733
734 if ((*this_mi)->bsize == BLOCK_32X32) {
735 int64_t threshold_32x32 = (5 * thresholds[1]) >> 3;
736 if (vt->split[lvl1_idx].part_variances.none.variance < threshold_32x32)
737 part_info->variance_low[lvl1_idx + 5] = 1;
738 } else {
739 // For 32x16 and 16x32 blocks, the flag is set on each 16x16 block
740 // inside.
741 if ((*this_mi)->bsize == BLOCK_16X16 ||
742 (*this_mi)->bsize == BLOCK_32X16 ||
743 (*this_mi)->bsize == BLOCK_16X32) {
744 for (int lvl2_idx = 0; lvl2_idx < 4; lvl2_idx++) {
745 if (vt->split[lvl1_idx]
746 .split[lvl2_idx]
747 .part_variances.none.variance < (thresholds[2] >> 8))
748 part_info->variance_low[(lvl1_idx << 2) + lvl2_idx + 9] = 1;
749 }
750 }
751 }
752 }
753 }
754 }
755
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)756 static AOM_INLINE void set_low_temp_var_flag_128x128(
757 CommonModeInfoParams *mi_params, PartitionSearchInfo *part_info,
758 MACROBLOCKD *xd, VP128x128 *vt, const int64_t thresholds[], int mi_col,
759 int mi_row) {
760 if (xd->mi[0]->bsize == BLOCK_128X128) {
761 if (vt->part_variances.none.variance < (thresholds[0] >> 1))
762 part_info->variance_low[0] = 1;
763 } else if (xd->mi[0]->bsize == BLOCK_128X64) {
764 for (int part_idx = 0; part_idx < 2; part_idx++) {
765 if (vt->part_variances.horz[part_idx].variance < (thresholds[0] >> 2))
766 part_info->variance_low[part_idx + 1] = 1;
767 }
768 } else if (xd->mi[0]->bsize == BLOCK_64X128) {
769 for (int part_idx = 0; part_idx < 2; part_idx++) {
770 if (vt->part_variances.vert[part_idx].variance < (thresholds[0] >> 2))
771 part_info->variance_low[part_idx + 3] = 1;
772 }
773 } else {
774 static const int idx64[4][2] = {
775 { 0, 0 }, { 0, 16 }, { 16, 0 }, { 16, 16 }
776 };
777 static const int idx32[4][2] = { { 0, 0 }, { 0, 8 }, { 8, 0 }, { 8, 8 } };
778 for (int lvl1_idx = 0; lvl1_idx < 4; lvl1_idx++) {
779 const int idx_str = mi_params->mi_stride * (mi_row + idx64[lvl1_idx][0]) +
780 mi_col + idx64[lvl1_idx][1];
781 MB_MODE_INFO **mi_64 = mi_params->mi_grid_base + idx_str;
782 if (*mi_64 == NULL) continue;
783 if (mi_params->mi_cols <= mi_col + idx64[lvl1_idx][1] ||
784 mi_params->mi_rows <= mi_row + idx64[lvl1_idx][0])
785 continue;
786 const int64_t threshold_64x64 = (5 * thresholds[1]) >> 3;
787 if ((*mi_64)->bsize == BLOCK_64X64) {
788 if (vt->split[lvl1_idx].part_variances.none.variance < threshold_64x64)
789 part_info->variance_low[5 + lvl1_idx] = 1;
790 } else if ((*mi_64)->bsize == BLOCK_64X32) {
791 for (int part_idx = 0; part_idx < 2; part_idx++)
792 if (vt->split[lvl1_idx].part_variances.horz[part_idx].variance <
793 (threshold_64x64 >> 1))
794 part_info->variance_low[9 + (lvl1_idx << 1) + part_idx] = 1;
795 } else if ((*mi_64)->bsize == BLOCK_32X64) {
796 for (int part_idx = 0; part_idx < 2; part_idx++)
797 if (vt->split[lvl1_idx].part_variances.vert[part_idx].variance <
798 (threshold_64x64 >> 1))
799 part_info->variance_low[17 + (lvl1_idx << 1) + part_idx] = 1;
800 } else {
801 for (int lvl2_idx = 0; lvl2_idx < 4; lvl2_idx++) {
802 const int idx_str1 =
803 mi_params->mi_stride * idx32[lvl2_idx][0] + idx32[lvl2_idx][1];
804 MB_MODE_INFO **mi_32 = mi_params->mi_grid_base + idx_str + idx_str1;
805 if (*mi_32 == NULL) continue;
806
807 if (mi_params->mi_cols <=
808 mi_col + idx64[lvl1_idx][1] + idx32[lvl2_idx][1] ||
809 mi_params->mi_rows <=
810 mi_row + idx64[lvl1_idx][0] + idx32[lvl2_idx][0])
811 continue;
812 const int64_t threshold_32x32 = (5 * thresholds[2]) >> 3;
813 if ((*mi_32)->bsize == BLOCK_32X32) {
814 if (vt->split[lvl1_idx]
815 .split[lvl2_idx]
816 .part_variances.none.variance < threshold_32x32)
817 part_info->variance_low[25 + (lvl1_idx << 2) + lvl2_idx] = 1;
818 } else {
819 // For 32x16 and 16x32 blocks, the flag is set on each 16x16 block
820 // inside.
821 if ((*mi_32)->bsize == BLOCK_16X16 ||
822 (*mi_32)->bsize == BLOCK_32X16 ||
823 (*mi_32)->bsize == BLOCK_16X32) {
824 for (int lvl3_idx = 0; lvl3_idx < 4; lvl3_idx++) {
825 VPartVar *none_var = &vt->split[lvl1_idx]
826 .split[lvl2_idx]
827 .split[lvl3_idx]
828 .part_variances.none;
829 if (none_var->variance < (thresholds[3] >> 8))
830 part_info->variance_low[41 + (lvl1_idx << 4) +
831 (lvl2_idx << 2) + lvl3_idx] = 1;
832 }
833 }
834 }
835 }
836 }
837 }
838 }
839 }
840
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)841 static AOM_INLINE void set_low_temp_var_flag(
842 AV1_COMP *cpi, PartitionSearchInfo *part_info, MACROBLOCKD *xd,
843 VP128x128 *vt, int64_t thresholds[], MV_REFERENCE_FRAME ref_frame_partition,
844 int mi_col, int mi_row, const bool is_small_sb) {
845 AV1_COMMON *const cm = &cpi->common;
846 // Check temporal variance for bsize >= 16x16, if LAST_FRAME was selected.
847 // If the temporal variance is small set the flag
848 // variance_low for the block. The variance threshold can be adjusted, the
849 // higher the more aggressive.
850 if (ref_frame_partition == LAST_FRAME) {
851 if (is_small_sb)
852 set_low_temp_var_flag_64x64(&cm->mi_params, part_info, xd,
853 &(vt->split[0]), thresholds, mi_col, mi_row);
854 else
855 set_low_temp_var_flag_128x128(&cm->mi_params, part_info, xd, vt,
856 thresholds, mi_col, mi_row);
857 }
858 }
859
860 static const int pos_shift_16x16[4][4] = {
861 { 9, 10, 13, 14 }, { 11, 12, 15, 16 }, { 17, 18, 21, 22 }, { 19, 20, 23, 24 }
862 };
863
av1_get_force_skip_low_temp_var_small_sb(const uint8_t * variance_low,int mi_row,int mi_col,BLOCK_SIZE bsize)864 int av1_get_force_skip_low_temp_var_small_sb(const uint8_t *variance_low,
865 int mi_row, int mi_col,
866 BLOCK_SIZE bsize) {
867 // Relative indices of MB inside the superblock.
868 const int mi_x = mi_row & 0xF;
869 const int mi_y = mi_col & 0xF;
870 // Relative indices of 16x16 block inside the superblock.
871 const int i = mi_x >> 2;
872 const int j = mi_y >> 2;
873 int force_skip_low_temp_var = 0;
874 // Set force_skip_low_temp_var based on the block size and block offset.
875 switch (bsize) {
876 case BLOCK_64X64: force_skip_low_temp_var = variance_low[0]; break;
877 case BLOCK_64X32:
878 if (!mi_y && !mi_x) {
879 force_skip_low_temp_var = variance_low[1];
880 } else if (!mi_y && mi_x) {
881 force_skip_low_temp_var = variance_low[2];
882 }
883 break;
884 case BLOCK_32X64:
885 if (!mi_y && !mi_x) {
886 force_skip_low_temp_var = variance_low[3];
887 } else if (mi_y && !mi_x) {
888 force_skip_low_temp_var = variance_low[4];
889 }
890 break;
891 case BLOCK_32X32:
892 if (!mi_y && !mi_x) {
893 force_skip_low_temp_var = variance_low[5];
894 } else if (mi_y && !mi_x) {
895 force_skip_low_temp_var = variance_low[6];
896 } else if (!mi_y && mi_x) {
897 force_skip_low_temp_var = variance_low[7];
898 } else if (mi_y && mi_x) {
899 force_skip_low_temp_var = variance_low[8];
900 }
901 break;
902 case BLOCK_32X16:
903 case BLOCK_16X32:
904 case BLOCK_16X16:
905 force_skip_low_temp_var = variance_low[pos_shift_16x16[i][j]];
906 break;
907 default: break;
908 }
909
910 return force_skip_low_temp_var;
911 }
912
av1_get_force_skip_low_temp_var(const uint8_t * variance_low,int mi_row,int mi_col,BLOCK_SIZE bsize)913 int av1_get_force_skip_low_temp_var(const uint8_t *variance_low, int mi_row,
914 int mi_col, BLOCK_SIZE bsize) {
915 int force_skip_low_temp_var = 0;
916 int x, y;
917 x = (mi_col & 0x1F) >> 4;
918 // y = (mi_row & 0x1F) >> 4;
919 // const int idx64 = (y << 1) + x;
920 y = (mi_row & 0x17) >> 3;
921 const int idx64 = y + x;
922
923 x = (mi_col & 0xF) >> 3;
924 // y = (mi_row & 0xF) >> 3;
925 // const int idx32 = (y << 1) + x;
926 y = (mi_row & 0xB) >> 2;
927 const int idx32 = y + x;
928
929 x = (mi_col & 0x7) >> 2;
930 // y = (mi_row & 0x7) >> 2;
931 // const int idx16 = (y << 1) + x;
932 y = (mi_row & 0x5) >> 1;
933 const int idx16 = y + x;
934 // Set force_skip_low_temp_var based on the block size and block offset.
935 switch (bsize) {
936 case BLOCK_128X128: force_skip_low_temp_var = variance_low[0]; break;
937 case BLOCK_128X64:
938 assert((mi_col & 0x1F) == 0);
939 force_skip_low_temp_var = variance_low[1 + ((mi_row & 0x1F) != 0)];
940 break;
941 case BLOCK_64X128:
942 assert((mi_row & 0x1F) == 0);
943 force_skip_low_temp_var = variance_low[3 + ((mi_col & 0x1F) != 0)];
944 break;
945 case BLOCK_64X64:
946 // Location of this 64x64 block inside the 128x128 superblock
947 force_skip_low_temp_var = variance_low[5 + idx64];
948 break;
949 case BLOCK_64X32:
950 x = (mi_col & 0x1F) >> 4;
951 y = (mi_row & 0x1F) >> 3;
952 /*
953 .---------------.---------------.
954 | x=0,y=0,idx=0 | x=0,y=0,idx=2 |
955 :---------------+---------------:
956 | x=0,y=1,idx=1 | x=1,y=1,idx=3 |
957 :---------------+---------------:
958 | x=0,y=2,idx=4 | x=1,y=2,idx=6 |
959 :---------------+---------------:
960 | x=0,y=3,idx=5 | x=1,y=3,idx=7 |
961 '---------------'---------------'
962 */
963 const int idx64x32 = (x << 1) + (y % 2) + ((y >> 1) << 2);
964 force_skip_low_temp_var = variance_low[9 + idx64x32];
965 break;
966 case BLOCK_32X64:
967 x = (mi_col & 0x1F) >> 3;
968 y = (mi_row & 0x1F) >> 4;
969 const int idx32x64 = (y << 2) + x;
970 force_skip_low_temp_var = variance_low[17 + idx32x64];
971 break;
972 case BLOCK_32X32:
973 force_skip_low_temp_var = variance_low[25 + (idx64 << 2) + idx32];
974 break;
975 case BLOCK_32X16:
976 case BLOCK_16X32:
977 case BLOCK_16X16:
978 force_skip_low_temp_var =
979 variance_low[41 + (idx64 << 4) + (idx32 << 2) + idx16];
980 break;
981 default: break;
982 }
983 return force_skip_low_temp_var;
984 }
985
av1_set_variance_partition_thresholds(AV1_COMP * cpi,int qindex,int content_lowsumdiff)986 void av1_set_variance_partition_thresholds(AV1_COMP *cpi, int qindex,
987 int content_lowsumdiff) {
988 SPEED_FEATURES *const sf = &cpi->sf;
989 if (sf->part_sf.partition_search_type != VAR_BASED_PARTITION) {
990 return;
991 } else {
992 set_vbp_thresholds(cpi, cpi->vbp_info.thresholds, 0, qindex,
993 content_lowsumdiff, 0, 0, 0, 0);
994 // The threshold below is not changed locally.
995 cpi->vbp_info.threshold_minmax = 15 + (qindex >> 3);
996 }
997 }
998
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)999 static AOM_INLINE void chroma_check(AV1_COMP *cpi, MACROBLOCK *x,
1000 BLOCK_SIZE bsize, unsigned int y_sad,
1001 unsigned int y_sad_g,
1002 unsigned int y_sad_alt, bool is_key_frame,
1003 bool zero_motion, unsigned int *uv_sad) {
1004 MACROBLOCKD *xd = &x->e_mbd;
1005 const int source_sad_nonrd = x->content_state_sb.source_sad_nonrd;
1006 int shift_upper_limit = 1;
1007 int shift_lower_limit = 3;
1008 int fac_uv = 6;
1009 if (is_key_frame || cpi->oxcf.tool_cfg.enable_monochrome) return;
1010
1011 // Use lower threshold (more conservative in setting color flag) for
1012 // higher resolutions non-screen, which tend to have more camera noise.
1013 // Since this may be used to skip compound mode in nonrd pickmode, which
1014 // is generally more effective for higher resolutions, better to be more
1015 // conservative.
1016 if (cpi->oxcf.tune_cfg.content != AOM_CONTENT_SCREEN) {
1017 if (cpi->common.width * cpi->common.height >= RESOLUTION_1080P)
1018 fac_uv = 3;
1019 else
1020 fac_uv = 5;
1021 }
1022 if (cpi->oxcf.tune_cfg.content == AOM_CONTENT_SCREEN &&
1023 cpi->rc.high_source_sad) {
1024 shift_lower_limit = 7;
1025 } else if (source_sad_nonrd >= kMedSad && x->source_variance > 500 &&
1026 cpi->common.width * cpi->common.height >= 640 * 360) {
1027 shift_upper_limit = 2;
1028 shift_lower_limit = source_sad_nonrd > kMedSad ? 5 : 4;
1029 }
1030
1031 MB_MODE_INFO *mi = xd->mi[0];
1032 const AV1_COMMON *const cm = &cpi->common;
1033 const YV12_BUFFER_CONFIG *yv12 = get_ref_frame_yv12_buf(cm, LAST_FRAME);
1034 const YV12_BUFFER_CONFIG *yv12_g = get_ref_frame_yv12_buf(cm, GOLDEN_FRAME);
1035 const YV12_BUFFER_CONFIG *yv12_alt = get_ref_frame_yv12_buf(cm, ALTREF_FRAME);
1036 const struct scale_factors *const sf =
1037 get_ref_scale_factors_const(cm, LAST_FRAME);
1038 struct buf_2d dst;
1039 unsigned int uv_sad_g = 0;
1040 unsigned int uv_sad_alt = 0;
1041
1042 for (int plane = AOM_PLANE_U; plane < MAX_MB_PLANE; ++plane) {
1043 struct macroblock_plane *p = &x->plane[plane];
1044 struct macroblockd_plane *pd = &xd->plane[plane];
1045 const BLOCK_SIZE bs =
1046 get_plane_block_size(bsize, pd->subsampling_x, pd->subsampling_y);
1047
1048 if (bs != BLOCK_INVALID) {
1049 // For last:
1050 if (zero_motion) {
1051 if (mi->ref_frame[0] == LAST_FRAME) {
1052 uv_sad[plane - 1] = cpi->ppi->fn_ptr[bs].sdf(
1053 p->src.buf, p->src.stride, pd->pre[0].buf, pd->pre[0].stride);
1054 } else {
1055 uint8_t *src = (plane == 1) ? yv12->u_buffer : yv12->v_buffer;
1056 setup_pred_plane(&dst, xd->mi[0]->bsize, src, yv12->uv_crop_width,
1057 yv12->uv_crop_height, yv12->uv_stride, xd->mi_row,
1058 xd->mi_col, sf, xd->plane[plane].subsampling_x,
1059 xd->plane[plane].subsampling_y);
1060
1061 uv_sad[plane - 1] = cpi->ppi->fn_ptr[bs].sdf(
1062 p->src.buf, p->src.stride, dst.buf, dst.stride);
1063 }
1064 } else {
1065 uv_sad[plane - 1] = cpi->ppi->fn_ptr[bs].sdf(
1066 p->src.buf, p->src.stride, pd->dst.buf, pd->dst.stride);
1067 }
1068
1069 // For golden:
1070 if (y_sad_g != UINT_MAX) {
1071 uint8_t *src = (plane == 1) ? yv12_g->u_buffer : yv12_g->v_buffer;
1072 setup_pred_plane(&dst, xd->mi[0]->bsize, src, yv12_g->uv_crop_width,
1073 yv12_g->uv_crop_height, yv12_g->uv_stride, xd->mi_row,
1074 xd->mi_col, sf, xd->plane[plane].subsampling_x,
1075 xd->plane[plane].subsampling_y);
1076 uv_sad_g = cpi->ppi->fn_ptr[bs].sdf(p->src.buf, p->src.stride, dst.buf,
1077 dst.stride);
1078 }
1079
1080 // For altref:
1081 if (y_sad_alt != UINT_MAX) {
1082 uint8_t *src = (plane == 1) ? yv12_alt->u_buffer : yv12_alt->v_buffer;
1083 setup_pred_plane(&dst, xd->mi[0]->bsize, src, yv12_alt->uv_crop_width,
1084 yv12_alt->uv_crop_height, yv12_alt->uv_stride,
1085 xd->mi_row, xd->mi_col, sf,
1086 xd->plane[plane].subsampling_x,
1087 xd->plane[plane].subsampling_y);
1088 uv_sad_alt = cpi->ppi->fn_ptr[bs].sdf(p->src.buf, p->src.stride,
1089 dst.buf, dst.stride);
1090 }
1091 }
1092
1093 if (uv_sad[plane - 1] > (y_sad >> shift_upper_limit))
1094 x->color_sensitivity_sb[COLOR_SENS_IDX(plane)] = 1;
1095 else if (uv_sad[plane - 1] < (y_sad >> shift_lower_limit))
1096 x->color_sensitivity_sb[COLOR_SENS_IDX(plane)] = 0;
1097 // Borderline case: to be refined at coding block level in nonrd_pickmode,
1098 // for coding block size < sb_size.
1099 else
1100 x->color_sensitivity_sb[COLOR_SENS_IDX(plane)] = 2;
1101
1102 x->color_sensitivity_sb_g[COLOR_SENS_IDX(plane)] =
1103 uv_sad_g > y_sad_g / fac_uv;
1104 x->color_sensitivity_sb_alt[COLOR_SENS_IDX(plane)] =
1105 uv_sad_alt > y_sad_alt / fac_uv;
1106 }
1107 }
1108
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)1109 static void fill_variance_tree_leaves(
1110 AV1_COMP *cpi, MACROBLOCK *x, VP128x128 *vt, PART_EVAL_STATUS *force_split,
1111 int avg_16x16[][4], int maxvar_16x16[][4], int minvar_16x16[][4],
1112 int64_t *thresholds, const uint8_t *src_buf, int src_stride,
1113 const uint8_t *dst_buf, int dst_stride, bool is_key_frame,
1114 const bool is_small_sb) {
1115 MACROBLOCKD *xd = &x->e_mbd;
1116 const int num_64x64_blocks = is_small_sb ? 1 : 4;
1117 // TODO(kyslov) Bring back compute_minmax_variance with content type detection
1118 const int compute_minmax_variance = 0;
1119 const int segment_id = xd->mi[0]->segment_id;
1120 int pixels_wide = 128, pixels_high = 128;
1121 int border_offset_4x4 = 0;
1122 int temporal_denoising = cpi->sf.rt_sf.use_rtc_tf;
1123 // dst_buf pointer is not used for is_key_frame, so it should be NULL.
1124 assert(IMPLIES(is_key_frame, dst_buf == NULL));
1125 if (is_small_sb) {
1126 pixels_wide = 64;
1127 pixels_high = 64;
1128 }
1129 if (xd->mb_to_right_edge < 0) pixels_wide += (xd->mb_to_right_edge >> 3);
1130 if (xd->mb_to_bottom_edge < 0) pixels_high += (xd->mb_to_bottom_edge >> 3);
1131 #if CONFIG_AV1_TEMPORAL_DENOISING
1132 temporal_denoising |= cpi->oxcf.noise_sensitivity;
1133 #endif
1134 // For temporal filtering or temporal denoiser enabled: since the source
1135 // is modified we need to avoid 4x4 avg along superblock boundary, since
1136 // simd code will load 8 pixels for 4x4 avg and so can access source
1137 // data outside superblock (while its being modified by temporal filter).
1138 // Temporal filtering is never done on key frames.
1139 if (!is_key_frame && temporal_denoising) border_offset_4x4 = 4;
1140 for (int blk64_idx = 0; blk64_idx < num_64x64_blocks; blk64_idx++) {
1141 const int x64_idx = GET_BLK_IDX_X(blk64_idx, 6);
1142 const int y64_idx = GET_BLK_IDX_Y(blk64_idx, 6);
1143 const int blk64_scale_idx = blk64_idx << 2;
1144 force_split[blk64_idx + 1] = PART_EVAL_ALL;
1145
1146 for (int lvl1_idx = 0; lvl1_idx < 4; lvl1_idx++) {
1147 const int x32_idx = x64_idx + GET_BLK_IDX_X(lvl1_idx, 5);
1148 const int y32_idx = y64_idx + GET_BLK_IDX_Y(lvl1_idx, 5);
1149 const int lvl1_scale_idx = (blk64_scale_idx + lvl1_idx) << 2;
1150 force_split[5 + blk64_scale_idx + lvl1_idx] = PART_EVAL_ALL;
1151 avg_16x16[blk64_idx][lvl1_idx] = 0;
1152 maxvar_16x16[blk64_idx][lvl1_idx] = 0;
1153 minvar_16x16[blk64_idx][lvl1_idx] = INT_MAX;
1154 for (int lvl2_idx = 0; lvl2_idx < 4; lvl2_idx++) {
1155 const int x16_idx = x32_idx + GET_BLK_IDX_X(lvl2_idx, 4);
1156 const int y16_idx = y32_idx + GET_BLK_IDX_Y(lvl2_idx, 4);
1157 const int split_index = 21 + lvl1_scale_idx + lvl2_idx;
1158 VP16x16 *vst = &vt->split[blk64_idx].split[lvl1_idx].split[lvl2_idx];
1159 force_split[split_index] = PART_EVAL_ALL;
1160 if (is_key_frame) {
1161 // Go down to 4x4 down-sampling for variance.
1162 for (int lvl3_idx = 0; lvl3_idx < 4; lvl3_idx++) {
1163 const int x8_idx = x16_idx + GET_BLK_IDX_X(lvl3_idx, 3);
1164 const int y8_idx = y16_idx + GET_BLK_IDX_Y(lvl3_idx, 3);
1165 VP8x8 *vst2 = &vst->split[lvl3_idx];
1166 fill_variance_4x4avg(src_buf, src_stride, x8_idx, y8_idx, vst2,
1167 #if CONFIG_AV1_HIGHBITDEPTH
1168 xd->cur_buf->flags,
1169 #endif
1170 pixels_wide, pixels_high, border_offset_4x4);
1171 }
1172 } else {
1173 fill_variance_8x8avg(src_buf, src_stride, dst_buf, dst_stride,
1174 x16_idx, y16_idx, vst, is_cur_buf_hbd(xd),
1175 pixels_wide, pixels_high);
1176
1177 fill_variance_tree(vst, BLOCK_16X16);
1178 VPartVar *none_var = &vt->split[blk64_idx]
1179 .split[lvl1_idx]
1180 .split[lvl2_idx]
1181 .part_variances.none;
1182 get_variance(none_var);
1183 const int val_none_var = none_var->variance;
1184 avg_16x16[blk64_idx][lvl1_idx] += val_none_var;
1185 minvar_16x16[blk64_idx][lvl1_idx] =
1186 AOMMIN(minvar_16x16[blk64_idx][lvl1_idx], val_none_var);
1187 maxvar_16x16[blk64_idx][lvl1_idx] =
1188 AOMMAX(maxvar_16x16[blk64_idx][lvl1_idx], val_none_var);
1189 if (val_none_var > thresholds[3]) {
1190 // 16X16 variance is above threshold for split, so force split to
1191 // 8x8 for this 16x16 block (this also forces splits for upper
1192 // levels).
1193 force_split[split_index] = PART_EVAL_ONLY_SPLIT;
1194 force_split[5 + blk64_scale_idx + lvl1_idx] = PART_EVAL_ONLY_SPLIT;
1195 force_split[blk64_idx + 1] = PART_EVAL_ONLY_SPLIT;
1196 force_split[0] = PART_EVAL_ONLY_SPLIT;
1197 } else if (!cyclic_refresh_segment_id_boosted(segment_id) &&
1198 compute_minmax_variance && val_none_var > thresholds[2]) {
1199 // We have some nominal amount of 16x16 variance (based on average),
1200 // compute the minmax over the 8x8 sub-blocks, and if above
1201 // threshold, force split to 8x8 block for this 16x16 block.
1202 int minmax = compute_minmax_8x8(src_buf, src_stride, dst_buf,
1203 dst_stride, x16_idx, y16_idx,
1204 #if CONFIG_AV1_HIGHBITDEPTH
1205 xd->cur_buf->flags,
1206 #endif
1207 pixels_wide, pixels_high);
1208 const int thresh_minmax = (int)cpi->vbp_info.threshold_minmax;
1209 if (minmax > thresh_minmax) {
1210 force_split[split_index] = PART_EVAL_ONLY_SPLIT;
1211 force_split[5 + blk64_scale_idx + lvl1_idx] =
1212 PART_EVAL_ONLY_SPLIT;
1213 force_split[blk64_idx + 1] = PART_EVAL_ONLY_SPLIT;
1214 force_split[0] = PART_EVAL_ONLY_SPLIT;
1215 }
1216 }
1217 }
1218 }
1219 }
1220 }
1221 }
1222
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)1223 static AOM_INLINE void set_ref_frame_for_partition(
1224 AV1_COMP *cpi, MACROBLOCK *x, MACROBLOCKD *xd,
1225 MV_REFERENCE_FRAME *ref_frame_partition, MB_MODE_INFO *mi,
1226 unsigned int *y_sad, unsigned int *y_sad_g, unsigned int *y_sad_alt,
1227 const YV12_BUFFER_CONFIG *yv12_g, const YV12_BUFFER_CONFIG *yv12_alt,
1228 int mi_row, int mi_col, int num_planes) {
1229 AV1_COMMON *const cm = &cpi->common;
1230 const bool is_set_golden_ref_frame =
1231 *y_sad_g < 0.9 * *y_sad && *y_sad_g < *y_sad_alt;
1232 const bool is_set_altref_ref_frame =
1233 *y_sad_alt < 0.9 * *y_sad && *y_sad_alt < *y_sad_g;
1234
1235 if (is_set_golden_ref_frame) {
1236 av1_setup_pre_planes(xd, 0, yv12_g, mi_row, mi_col,
1237 get_ref_scale_factors(cm, GOLDEN_FRAME), num_planes);
1238 mi->ref_frame[0] = GOLDEN_FRAME;
1239 mi->mv[0].as_int = 0;
1240 *y_sad = *y_sad_g;
1241 *ref_frame_partition = GOLDEN_FRAME;
1242 x->nonrd_prune_ref_frame_search = 0;
1243 x->sb_me_partition = 0;
1244 } else if (is_set_altref_ref_frame) {
1245 av1_setup_pre_planes(xd, 0, yv12_alt, mi_row, mi_col,
1246 get_ref_scale_factors(cm, ALTREF_FRAME), num_planes);
1247 mi->ref_frame[0] = ALTREF_FRAME;
1248 mi->mv[0].as_int = 0;
1249 *y_sad = *y_sad_alt;
1250 *ref_frame_partition = ALTREF_FRAME;
1251 x->nonrd_prune_ref_frame_search = 0;
1252 x->sb_me_partition = 0;
1253 } else {
1254 *ref_frame_partition = LAST_FRAME;
1255 x->nonrd_prune_ref_frame_search =
1256 cpi->sf.rt_sf.nonrd_prune_ref_frame_search;
1257 }
1258 }
1259
mv_distance(const FULLPEL_MV * mv0,const FULLPEL_MV * mv1)1260 static AOM_FORCE_INLINE int mv_distance(const FULLPEL_MV *mv0,
1261 const FULLPEL_MV *mv1) {
1262 return abs(mv0->row - mv1->row) + abs(mv0->col - mv1->col);
1263 }
1264
evaluate_neighbour_mvs(AV1_COMP * cpi,MACROBLOCK * x,unsigned int * y_sad,bool is_small_sb,int est_motion)1265 static AOM_INLINE void evaluate_neighbour_mvs(AV1_COMP *cpi, MACROBLOCK *x,
1266 unsigned int *y_sad,
1267 bool is_small_sb,
1268 int est_motion) {
1269 const int source_sad_nonrd = x->content_state_sb.source_sad_nonrd;
1270 // TODO(yunqingwang@google.com): test if this condition works with other
1271 // speeds.
1272 if (est_motion > 2 && source_sad_nonrd > kMedSad) return;
1273
1274 MACROBLOCKD *xd = &x->e_mbd;
1275 BLOCK_SIZE bsize = is_small_sb ? BLOCK_64X64 : BLOCK_128X128;
1276 MB_MODE_INFO *mi = xd->mi[0];
1277
1278 unsigned int above_y_sad = UINT_MAX;
1279 unsigned int left_y_sad = UINT_MAX;
1280 FULLPEL_MV above_mv = kZeroFullMv;
1281 FULLPEL_MV left_mv = kZeroFullMv;
1282 SubpelMvLimits subpel_mv_limits;
1283 const MV dummy_mv = { 0, 0 };
1284 av1_set_subpel_mv_search_range(&subpel_mv_limits, &x->mv_limits, &dummy_mv);
1285
1286 // Current best MV
1287 FULLPEL_MV best_mv = get_fullmv_from_mv(&mi->mv[0].as_mv);
1288 const int multi = (est_motion > 2 && source_sad_nonrd > kLowSad) ? 7 : 8;
1289
1290 if (xd->up_available) {
1291 const MB_MODE_INFO *above_mbmi = xd->above_mbmi;
1292 if (above_mbmi->mode >= INTRA_MODE_END &&
1293 above_mbmi->ref_frame[0] == LAST_FRAME) {
1294 MV temp = above_mbmi->mv[0].as_mv;
1295 clamp_mv(&temp, &subpel_mv_limits);
1296 above_mv = get_fullmv_from_mv(&temp);
1297
1298 if (mv_distance(&best_mv, &above_mv) > 0) {
1299 uint8_t const *ref_buf =
1300 get_buf_from_fullmv(&xd->plane[0].pre[0], &above_mv);
1301 above_y_sad = cpi->ppi->fn_ptr[bsize].sdf(
1302 x->plane[0].src.buf, x->plane[0].src.stride, ref_buf,
1303 xd->plane[0].pre[0].stride);
1304 }
1305 }
1306 }
1307 if (xd->left_available) {
1308 const MB_MODE_INFO *left_mbmi = xd->left_mbmi;
1309 if (left_mbmi->mode >= INTRA_MODE_END &&
1310 left_mbmi->ref_frame[0] == LAST_FRAME) {
1311 MV temp = left_mbmi->mv[0].as_mv;
1312 clamp_mv(&temp, &subpel_mv_limits);
1313 left_mv = get_fullmv_from_mv(&temp);
1314
1315 if (mv_distance(&best_mv, &left_mv) > 0 &&
1316 mv_distance(&above_mv, &left_mv) > 0) {
1317 uint8_t const *ref_buf =
1318 get_buf_from_fullmv(&xd->plane[0].pre[0], &left_mv);
1319 left_y_sad = cpi->ppi->fn_ptr[bsize].sdf(
1320 x->plane[0].src.buf, x->plane[0].src.stride, ref_buf,
1321 xd->plane[0].pre[0].stride);
1322 }
1323 }
1324 }
1325
1326 if (above_y_sad < ((multi * *y_sad) >> 3) && above_y_sad < left_y_sad) {
1327 *y_sad = above_y_sad;
1328 mi->mv[0].as_mv = get_mv_from_fullmv(&above_mv);
1329 clamp_mv(&mi->mv[0].as_mv, &subpel_mv_limits);
1330 }
1331 if (left_y_sad < ((multi * *y_sad) >> 3) && left_y_sad < above_y_sad) {
1332 *y_sad = left_y_sad;
1333 mi->mv[0].as_mv = get_mv_from_fullmv(&left_mv);
1334 clamp_mv(&mi->mv[0].as_mv, &subpel_mv_limits);
1335 }
1336 }
1337
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)1338 static void setup_planes(AV1_COMP *cpi, MACROBLOCK *x, unsigned int *y_sad,
1339 unsigned int *y_sad_g, unsigned int *y_sad_alt,
1340 unsigned int *y_sad_last,
1341 MV_REFERENCE_FRAME *ref_frame_partition,
1342 struct scale_factors *sf_no_scale, int mi_row,
1343 int mi_col, bool is_small_sb, bool scaled_ref_last) {
1344 AV1_COMMON *const cm = &cpi->common;
1345 MACROBLOCKD *xd = &x->e_mbd;
1346 const int num_planes = av1_num_planes(cm);
1347 bool scaled_ref_golden = false;
1348 bool scaled_ref_alt = false;
1349 BLOCK_SIZE bsize = is_small_sb ? BLOCK_64X64 : BLOCK_128X128;
1350 MB_MODE_INFO *mi = xd->mi[0];
1351 const YV12_BUFFER_CONFIG *yv12 =
1352 scaled_ref_last ? av1_get_scaled_ref_frame(cpi, LAST_FRAME)
1353 : get_ref_frame_yv12_buf(cm, LAST_FRAME);
1354 assert(yv12 != NULL);
1355 const YV12_BUFFER_CONFIG *yv12_g = NULL;
1356 const YV12_BUFFER_CONFIG *yv12_alt = NULL;
1357 // Check if LAST is a reference. For spatial layers always use it as
1358 // reference scaling.
1359 int use_last_ref = (cpi->ref_frame_flags & AOM_LAST_FLAG) ||
1360 cpi->svc.number_spatial_layers > 1;
1361 int use_golden_ref = cpi->ref_frame_flags & AOM_GOLD_FLAG;
1362 int use_alt_ref = cpi->ppi->rtc_ref.set_ref_frame_config ||
1363 cpi->sf.rt_sf.use_nonrd_altref_frame ||
1364 (cpi->sf.rt_sf.use_comp_ref_nonrd &&
1365 cpi->sf.rt_sf.ref_frame_comp_nonrd[2] == 1);
1366
1367 // For 1 spatial layer: GOLDEN is another temporal reference.
1368 // Check if it should be used as reference for partitioning.
1369 if (cpi->svc.number_spatial_layers == 1 && use_golden_ref &&
1370 (x->content_state_sb.source_sad_nonrd != kZeroSad || !use_last_ref)) {
1371 yv12_g = get_ref_frame_yv12_buf(cm, GOLDEN_FRAME);
1372 if (yv12_g && (yv12_g->y_crop_height != cm->height ||
1373 yv12_g->y_crop_width != cm->width)) {
1374 yv12_g = av1_get_scaled_ref_frame(cpi, GOLDEN_FRAME);
1375 scaled_ref_golden = true;
1376 }
1377 if (yv12_g && yv12_g != yv12) {
1378 av1_setup_pre_planes(
1379 xd, 0, yv12_g, mi_row, mi_col,
1380 scaled_ref_golden ? NULL : get_ref_scale_factors(cm, GOLDEN_FRAME),
1381 num_planes);
1382 *y_sad_g = cpi->ppi->fn_ptr[bsize].sdf(
1383 x->plane[AOM_PLANE_Y].src.buf, x->plane[AOM_PLANE_Y].src.stride,
1384 xd->plane[AOM_PLANE_Y].pre[0].buf,
1385 xd->plane[AOM_PLANE_Y].pre[0].stride);
1386 }
1387 }
1388
1389 // For 1 spatial layer: ALTREF is another temporal reference.
1390 // Check if it should be used as reference for partitioning.
1391 if (cpi->svc.number_spatial_layers == 1 && use_alt_ref &&
1392 (cpi->ref_frame_flags & AOM_ALT_FLAG) &&
1393 (x->content_state_sb.source_sad_nonrd != kZeroSad || !use_last_ref)) {
1394 yv12_alt = get_ref_frame_yv12_buf(cm, ALTREF_FRAME);
1395 if (yv12_alt && (yv12_alt->y_crop_height != cm->height ||
1396 yv12_alt->y_crop_width != cm->width)) {
1397 yv12_alt = av1_get_scaled_ref_frame(cpi, ALTREF_FRAME);
1398 scaled_ref_alt = true;
1399 }
1400 if (yv12_alt && yv12_alt != yv12) {
1401 av1_setup_pre_planes(
1402 xd, 0, yv12_alt, mi_row, mi_col,
1403 scaled_ref_alt ? NULL : get_ref_scale_factors(cm, ALTREF_FRAME),
1404 num_planes);
1405 *y_sad_alt = cpi->ppi->fn_ptr[bsize].sdf(
1406 x->plane[AOM_PLANE_Y].src.buf, x->plane[AOM_PLANE_Y].src.stride,
1407 xd->plane[AOM_PLANE_Y].pre[0].buf,
1408 xd->plane[AOM_PLANE_Y].pre[0].stride);
1409 }
1410 }
1411
1412 if (use_last_ref) {
1413 const int source_sad_nonrd = x->content_state_sb.source_sad_nonrd;
1414 av1_setup_pre_planes(
1415 xd, 0, yv12, mi_row, mi_col,
1416 scaled_ref_last ? NULL : get_ref_scale_factors(cm, LAST_FRAME),
1417 num_planes);
1418 mi->ref_frame[0] = LAST_FRAME;
1419 mi->ref_frame[1] = NONE_FRAME;
1420 mi->bsize = cm->seq_params->sb_size;
1421 mi->mv[0].as_int = 0;
1422 mi->interp_filters = av1_broadcast_interp_filter(BILINEAR);
1423
1424 int est_motion = cpi->sf.rt_sf.estimate_motion_for_var_based_partition;
1425 // TODO(b/290596301): Look into adjusting this condition.
1426 // There is regression on color content when
1427 // estimate_motion_for_var_based_partition = 3 and high motion,
1428 // so for now force it to 2 based on superblock sad.
1429 if (est_motion > 2 && source_sad_nonrd > kMedSad) est_motion = 2;
1430
1431 if (est_motion == 1 || est_motion == 2) {
1432 if (xd->mb_to_right_edge >= 0 && xd->mb_to_bottom_edge >= 0) {
1433 // For screen only do int_pro_motion for spatial variance above
1434 // threshold and motion level above LowSad.
1435 if (x->source_variance > 100 && source_sad_nonrd > kLowSad) {
1436 int is_screen = cpi->oxcf.tune_cfg.content == AOM_CONTENT_SCREEN;
1437 int me_search_size_col =
1438 is_screen ? 96 : block_size_wide[cm->seq_params->sb_size] >> 1;
1439 // For screen use larger search size row motion to capture
1440 // vertical scroll, which can be larger motion.
1441 int me_search_size_row =
1442 is_screen ? 192 : block_size_high[cm->seq_params->sb_size] >> 1;
1443 unsigned int y_sad_zero;
1444 *y_sad = av1_int_pro_motion_estimation(
1445 cpi, x, cm->seq_params->sb_size, mi_row, mi_col, &kZeroMv,
1446 &y_sad_zero, me_search_size_col, me_search_size_row);
1447 // The logic below selects whether the motion estimated in the
1448 // int_pro_motion() will be used in nonrd_pickmode. Only do this
1449 // for screen for now.
1450 if (is_screen) {
1451 unsigned int thresh_sad =
1452 (cm->seq_params->sb_size == BLOCK_128X128) ? 50000 : 20000;
1453 if (*y_sad < (y_sad_zero >> 1) && *y_sad < thresh_sad) {
1454 x->sb_me_partition = 1;
1455 x->sb_me_mv.as_int = mi->mv[0].as_int;
1456 } else {
1457 x->sb_me_partition = 0;
1458 // Fall back to using zero motion.
1459 *y_sad = y_sad_zero;
1460 mi->mv[0].as_int = 0;
1461 }
1462 }
1463 }
1464 }
1465 }
1466
1467 if (*y_sad == UINT_MAX) {
1468 *y_sad = cpi->ppi->fn_ptr[bsize].sdf(
1469 x->plane[AOM_PLANE_Y].src.buf, x->plane[AOM_PLANE_Y].src.stride,
1470 xd->plane[AOM_PLANE_Y].pre[0].buf,
1471 xd->plane[AOM_PLANE_Y].pre[0].stride);
1472 }
1473
1474 // Evaluate if neighbours' MVs give better predictions. Zero MV is tested
1475 // already, so only non-zero MVs are tested here. Here the neighbour blocks
1476 // are the first block above or left to this superblock.
1477 if (est_motion >= 2 && (xd->up_available || xd->left_available))
1478 evaluate_neighbour_mvs(cpi, x, y_sad, is_small_sb, est_motion);
1479
1480 *y_sad_last = *y_sad;
1481 }
1482
1483 // Pick the ref frame for partitioning, use golden or altref frame only if
1484 // its lower sad, bias to LAST with factor 0.9.
1485 set_ref_frame_for_partition(cpi, x, xd, ref_frame_partition, mi, y_sad,
1486 y_sad_g, y_sad_alt, yv12_g, yv12_alt, mi_row,
1487 mi_col, num_planes);
1488
1489 // Only calculate the predictor for non-zero MV.
1490 if (mi->mv[0].as_int != 0) {
1491 if (!scaled_ref_last) {
1492 set_ref_ptrs(cm, xd, mi->ref_frame[0], mi->ref_frame[1]);
1493 } else {
1494 xd->block_ref_scale_factors[0] = sf_no_scale;
1495 xd->block_ref_scale_factors[1] = sf_no_scale;
1496 }
1497 av1_enc_build_inter_predictor(cm, xd, mi_row, mi_col, NULL,
1498 cm->seq_params->sb_size, AOM_PLANE_Y,
1499 num_planes - 1);
1500 }
1501 }
1502
1503 // Decides whether to split or merge a 16x16 partition block in variance based
1504 // partitioning based on the 8x8 sub-block variances.
get_part_eval_based_on_sub_blk_var(VP16x16 * var_16x16_info,int64_t threshold16)1505 static AOM_INLINE PART_EVAL_STATUS get_part_eval_based_on_sub_blk_var(
1506 VP16x16 *var_16x16_info, int64_t threshold16) {
1507 int max_8x8_var = 0, min_8x8_var = INT_MAX;
1508 for (int split_idx = 0; split_idx < 4; split_idx++) {
1509 get_variance(&var_16x16_info->split[split_idx].part_variances.none);
1510 int this_8x8_var =
1511 var_16x16_info->split[split_idx].part_variances.none.variance;
1512 max_8x8_var = AOMMAX(this_8x8_var, max_8x8_var);
1513 min_8x8_var = AOMMIN(this_8x8_var, min_8x8_var);
1514 }
1515 // If the difference between maximum and minimum sub-block variances is high,
1516 // then only evaluate PARTITION_SPLIT for the 16x16 block. Otherwise, evaluate
1517 // only PARTITION_NONE. The shift factor for threshold16 has been derived
1518 // empirically.
1519 return ((max_8x8_var - min_8x8_var) > (threshold16 << 2))
1520 ? PART_EVAL_ONLY_SPLIT
1521 : PART_EVAL_ONLY_NONE;
1522 }
1523
is_set_force_zeromv_skip_based_on_src_sad(int set_zeromv_skip_based_on_source_sad,SOURCE_SAD source_sad_nonrd)1524 static AOM_INLINE bool is_set_force_zeromv_skip_based_on_src_sad(
1525 int set_zeromv_skip_based_on_source_sad, SOURCE_SAD source_sad_nonrd) {
1526 if (set_zeromv_skip_based_on_source_sad == 0) return false;
1527
1528 if (set_zeromv_skip_based_on_source_sad >= 3)
1529 return source_sad_nonrd <= kLowSad;
1530 else if (set_zeromv_skip_based_on_source_sad >= 2)
1531 return source_sad_nonrd <= kVeryLowSad;
1532 else if (set_zeromv_skip_based_on_source_sad >= 1)
1533 return source_sad_nonrd == kZeroSad;
1534
1535 return false;
1536 }
1537
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)1538 static AOM_INLINE bool set_force_zeromv_skip_for_sb(
1539 AV1_COMP *cpi, MACROBLOCK *x, const TileInfo *const tile, VP128x128 *vt,
1540 unsigned int *uv_sad, int mi_row, int mi_col, unsigned int y_sad,
1541 BLOCK_SIZE bsize) {
1542 AV1_COMMON *const cm = &cpi->common;
1543 if (!is_set_force_zeromv_skip_based_on_src_sad(
1544 cpi->sf.rt_sf.set_zeromv_skip_based_on_source_sad,
1545 x->content_state_sb.source_sad_nonrd))
1546 return false;
1547 int shift = cpi->sf.rt_sf.increase_source_sad_thresh ? 1 : 0;
1548 const int block_width = mi_size_wide[cm->seq_params->sb_size];
1549 const int block_height = mi_size_high[cm->seq_params->sb_size];
1550 const unsigned int thresh_exit_part_y =
1551 cpi->zeromv_skip_thresh_exit_part[bsize] << shift;
1552 unsigned int thresh_exit_part_uv =
1553 CALC_CHROMA_THRESH_FOR_ZEROMV_SKIP(thresh_exit_part_y) << shift;
1554 // Be more aggressive in UV threshold if source_sad >= VeryLowSad
1555 // to suppreess visual artifact caused by the speed feature:
1556 // set_zeromv_skip_based_on_source_sad = 2. For now only for
1557 // part_early_exit_zeromv = 1.
1558 if (x->content_state_sb.source_sad_nonrd >= kVeryLowSad &&
1559 cpi->sf.rt_sf.part_early_exit_zeromv == 1)
1560 thresh_exit_part_uv = thresh_exit_part_uv >> 3;
1561 if (mi_col + block_width <= tile->mi_col_end &&
1562 mi_row + block_height <= tile->mi_row_end && y_sad < thresh_exit_part_y &&
1563 uv_sad[0] < thresh_exit_part_uv && uv_sad[1] < thresh_exit_part_uv) {
1564 set_block_size(cpi, mi_row, mi_col, bsize);
1565 x->force_zeromv_skip_for_sb = 1;
1566 aom_free(vt);
1567 // Partition shape is set here at SB level.
1568 // Exit needs to happen from av1_choose_var_based_partitioning().
1569 return true;
1570 } else if (x->content_state_sb.source_sad_nonrd == kZeroSad &&
1571 cpi->sf.rt_sf.part_early_exit_zeromv >= 2)
1572 x->force_zeromv_skip_for_sb = 2;
1573 return false;
1574 }
1575
av1_choose_var_based_partitioning(AV1_COMP * cpi,const TileInfo * const tile,ThreadData * td,MACROBLOCK * x,int mi_row,int mi_col)1576 int av1_choose_var_based_partitioning(AV1_COMP *cpi, const TileInfo *const tile,
1577 ThreadData *td, MACROBLOCK *x, int mi_row,
1578 int mi_col) {
1579 #if CONFIG_COLLECT_COMPONENT_TIMING
1580 start_timing(cpi, choose_var_based_partitioning_time);
1581 #endif
1582 AV1_COMMON *const cm = &cpi->common;
1583 MACROBLOCKD *xd = &x->e_mbd;
1584 const int64_t *const vbp_thresholds = cpi->vbp_info.thresholds;
1585 PART_EVAL_STATUS force_split[85];
1586 int avg_64x64;
1587 int max_var_32x32[4];
1588 int min_var_32x32[4];
1589 int var_32x32;
1590 int var_64x64;
1591 int min_var_64x64 = INT_MAX;
1592 int max_var_64x64 = 0;
1593 int avg_16x16[4][4];
1594 int maxvar_16x16[4][4];
1595 int minvar_16x16[4][4];
1596 const uint8_t *src_buf;
1597 const uint8_t *dst_buf;
1598 int dst_stride;
1599 unsigned int uv_sad[MAX_MB_PLANE - 1];
1600 NOISE_LEVEL noise_level = kLow;
1601 bool is_zero_motion = true;
1602 bool scaled_ref_last = false;
1603 struct scale_factors sf_no_scale;
1604 av1_setup_scale_factors_for_frame(&sf_no_scale, cm->width, cm->height,
1605 cm->width, cm->height);
1606
1607 bool is_key_frame =
1608 (frame_is_intra_only(cm) ||
1609 (cpi->ppi->use_svc &&
1610 cpi->svc.layer_context[cpi->svc.temporal_layer_id].is_key_frame));
1611
1612 assert(cm->seq_params->sb_size == BLOCK_64X64 ||
1613 cm->seq_params->sb_size == BLOCK_128X128);
1614 const bool is_small_sb = (cm->seq_params->sb_size == BLOCK_64X64);
1615 const int num_64x64_blocks = is_small_sb ? 1 : 4;
1616
1617 unsigned int y_sad = UINT_MAX;
1618 unsigned int y_sad_g = UINT_MAX;
1619 unsigned int y_sad_alt = UINT_MAX;
1620 unsigned int y_sad_last = UINT_MAX;
1621 BLOCK_SIZE bsize = is_small_sb ? BLOCK_64X64 : BLOCK_128X128;
1622
1623 // Ref frame used in partitioning.
1624 MV_REFERENCE_FRAME ref_frame_partition = LAST_FRAME;
1625
1626 int64_t thresholds[5] = { vbp_thresholds[0], vbp_thresholds[1],
1627 vbp_thresholds[2], vbp_thresholds[3],
1628 vbp_thresholds[4] };
1629
1630 const int segment_id = xd->mi[0]->segment_id;
1631 uint64_t blk_sad = 0;
1632 if (cpi->src_sad_blk_64x64 != NULL &&
1633 cpi->svc.spatial_layer_id == cpi->svc.number_spatial_layers - 1) {
1634 const int sb_size_by_mb = (cm->seq_params->sb_size == BLOCK_128X128)
1635 ? (cm->seq_params->mib_size >> 1)
1636 : cm->seq_params->mib_size;
1637 const int sb_cols =
1638 (cm->mi_params.mi_cols + sb_size_by_mb - 1) / sb_size_by_mb;
1639 const int sbi_col = mi_col / sb_size_by_mb;
1640 const int sbi_row = mi_row / sb_size_by_mb;
1641 blk_sad = cpi->src_sad_blk_64x64[sbi_col + sbi_row * sb_cols];
1642 }
1643
1644 const bool is_segment_id_boosted =
1645 cpi->oxcf.q_cfg.aq_mode == CYCLIC_REFRESH_AQ && cm->seg.enabled &&
1646 cyclic_refresh_segment_id_boosted(segment_id);
1647 const int qindex =
1648 is_segment_id_boosted
1649 ? av1_get_qindex(&cm->seg, segment_id, cm->quant_params.base_qindex)
1650 : cm->quant_params.base_qindex;
1651 set_vbp_thresholds(
1652 cpi, thresholds, blk_sad, qindex, x->content_state_sb.low_sumdiff,
1653 x->content_state_sb.source_sad_nonrd, x->content_state_sb.source_sad_rd,
1654 is_segment_id_boosted, x->content_state_sb.lighting_change);
1655
1656 src_buf = x->plane[AOM_PLANE_Y].src.buf;
1657 int src_stride = x->plane[AOM_PLANE_Y].src.stride;
1658
1659 // Index for force_split: 0 for 64x64, 1-4 for 32x32 blocks,
1660 // 5-20 for the 16x16 blocks.
1661 force_split[0] = PART_EVAL_ALL;
1662 memset(x->part_search_info.variance_low, 0,
1663 sizeof(x->part_search_info.variance_low));
1664
1665 // Check if LAST frame is NULL, and if so, treat this frame
1666 // as a key frame, for the purpose of the superblock partitioning.
1667 // LAST == NULL can happen in cases where enhancement spatial layers are
1668 // enabled dyanmically and the only reference is the spatial(GOLDEN).
1669 // If LAST frame has a different resolution: set the scaled_ref_last flag
1670 // and check if ref_scaled is NULL.
1671 if (!frame_is_intra_only(cm)) {
1672 const YV12_BUFFER_CONFIG *ref = get_ref_frame_yv12_buf(cm, LAST_FRAME);
1673 if (ref == NULL) {
1674 is_key_frame = true;
1675 } else if (ref->y_crop_height != cm->height ||
1676 ref->y_crop_width != cm->width) {
1677 scaled_ref_last = true;
1678 const YV12_BUFFER_CONFIG *ref_scaled =
1679 av1_get_scaled_ref_frame(cpi, LAST_FRAME);
1680 if (ref_scaled == NULL) is_key_frame = true;
1681 }
1682 }
1683
1684 x->source_variance = UINT_MAX;
1685 // For nord_pickmode: compute source_variance, only for superblocks with
1686 // some motion for now. This input can then be used to bias the partitioning
1687 // or the chroma_check.
1688 if (cpi->sf.rt_sf.use_nonrd_pick_mode &&
1689 x->content_state_sb.source_sad_nonrd > kLowSad)
1690 x->source_variance = av1_get_perpixel_variance_facade(
1691 cpi, xd, &x->plane[0].src, cm->seq_params->sb_size, AOM_PLANE_Y);
1692
1693 if (!is_key_frame) {
1694 setup_planes(cpi, x, &y_sad, &y_sad_g, &y_sad_alt, &y_sad_last,
1695 &ref_frame_partition, &sf_no_scale, mi_row, mi_col,
1696 is_small_sb, scaled_ref_last);
1697
1698 MB_MODE_INFO *mi = xd->mi[0];
1699 // Use reference SB directly for zero mv.
1700 if (mi->mv[0].as_int != 0) {
1701 dst_buf = xd->plane[AOM_PLANE_Y].dst.buf;
1702 dst_stride = xd->plane[AOM_PLANE_Y].dst.stride;
1703 is_zero_motion = false;
1704 } else {
1705 dst_buf = xd->plane[AOM_PLANE_Y].pre[0].buf;
1706 dst_stride = xd->plane[AOM_PLANE_Y].pre[0].stride;
1707 }
1708 } else {
1709 dst_buf = NULL;
1710 dst_stride = 0;
1711 }
1712
1713 // check and set the color sensitivity of sb.
1714 av1_zero(uv_sad);
1715 chroma_check(cpi, x, bsize, y_sad_last, y_sad_g, y_sad_alt, is_key_frame,
1716 is_zero_motion, uv_sad);
1717
1718 x->force_zeromv_skip_for_sb = 0;
1719
1720 VP128x128 *vt;
1721 AOM_CHECK_MEM_ERROR(xd->error_info, vt, aom_malloc(sizeof(*vt)));
1722 vt->split = td->vt64x64;
1723
1724 // If the superblock is completely static (zero source sad) and
1725 // the y_sad (relative to LAST ref) is very small, take the sb_size partition
1726 // and exit, and force zeromv_last skip mode for nonrd_pickmode.
1727 // Only do this on the base segment (so the QP-boosted segment, if applied,
1728 // can still continue cleaning/ramping up the quality).
1729 // Condition on color uv_sad is also added.
1730 if (!is_key_frame && cpi->sf.rt_sf.part_early_exit_zeromv &&
1731 cpi->rc.frames_since_key > 30 && segment_id == CR_SEGMENT_ID_BASE &&
1732 ref_frame_partition == LAST_FRAME && xd->mi[0]->mv[0].as_int == 0) {
1733 // Exit here, if zero mv skip flag is set at SB level.
1734 if (set_force_zeromv_skip_for_sb(cpi, x, tile, vt, uv_sad, mi_row, mi_col,
1735 y_sad, bsize))
1736 return 0;
1737 }
1738
1739 if (cpi->noise_estimate.enabled)
1740 noise_level = av1_noise_estimate_extract_level(&cpi->noise_estimate);
1741
1742 // Fill in the entire tree of 8x8 (for inter frames) or 4x4 (for key frames)
1743 // variances for splits.
1744 fill_variance_tree_leaves(cpi, x, vt, force_split, avg_16x16, maxvar_16x16,
1745 minvar_16x16, thresholds, src_buf, src_stride,
1746 dst_buf, dst_stride, is_key_frame, is_small_sb);
1747
1748 avg_64x64 = 0;
1749 for (int blk64_idx = 0; blk64_idx < num_64x64_blocks; ++blk64_idx) {
1750 max_var_32x32[blk64_idx] = 0;
1751 min_var_32x32[blk64_idx] = INT_MAX;
1752 const int blk64_scale_idx = blk64_idx << 2;
1753 for (int lvl1_idx = 0; lvl1_idx < 4; lvl1_idx++) {
1754 const int lvl1_scale_idx = (blk64_scale_idx + lvl1_idx) << 2;
1755 for (int lvl2_idx = 0; lvl2_idx < 4; lvl2_idx++) {
1756 if (!is_key_frame) continue;
1757 VP16x16 *vtemp = &vt->split[blk64_idx].split[lvl1_idx].split[lvl2_idx];
1758 for (int lvl3_idx = 0; lvl3_idx < 4; lvl3_idx++)
1759 fill_variance_tree(&vtemp->split[lvl3_idx], BLOCK_8X8);
1760 fill_variance_tree(vtemp, BLOCK_16X16);
1761 // If variance of this 16x16 block is above the threshold, force block
1762 // to split. This also forces a split on the upper levels.
1763 get_variance(&vtemp->part_variances.none);
1764 if (vtemp->part_variances.none.variance > thresholds[3]) {
1765 const int split_index = 21 + lvl1_scale_idx + lvl2_idx;
1766 force_split[split_index] =
1767 cpi->sf.rt_sf.vbp_prune_16x16_split_using_min_max_sub_blk_var
1768 ? get_part_eval_based_on_sub_blk_var(vtemp, thresholds[3])
1769 : PART_EVAL_ONLY_SPLIT;
1770 force_split[5 + blk64_scale_idx + lvl1_idx] = PART_EVAL_ONLY_SPLIT;
1771 force_split[blk64_idx + 1] = PART_EVAL_ONLY_SPLIT;
1772 force_split[0] = PART_EVAL_ONLY_SPLIT;
1773 }
1774 }
1775 fill_variance_tree(&vt->split[blk64_idx].split[lvl1_idx], BLOCK_32X32);
1776 // If variance of this 32x32 block is above the threshold, or if its above
1777 // (some threshold of) the average variance over the sub-16x16 blocks,
1778 // then force this block to split. This also forces a split on the upper
1779 // (64x64) level.
1780 uint64_t frame_sad_thresh = 20000;
1781 const int is_360p_or_smaller = cm->width * cm->height <= RESOLUTION_360P;
1782 if (cpi->svc.number_temporal_layers > 2 &&
1783 cpi->svc.temporal_layer_id == 0)
1784 frame_sad_thresh = frame_sad_thresh << 1;
1785 if (force_split[5 + blk64_scale_idx + lvl1_idx] == PART_EVAL_ALL) {
1786 get_variance(&vt->split[blk64_idx].split[lvl1_idx].part_variances.none);
1787 var_32x32 =
1788 vt->split[blk64_idx].split[lvl1_idx].part_variances.none.variance;
1789 max_var_32x32[blk64_idx] = AOMMAX(var_32x32, max_var_32x32[blk64_idx]);
1790 min_var_32x32[blk64_idx] = AOMMIN(var_32x32, min_var_32x32[blk64_idx]);
1791 const int max_min_var_16X16_diff = (maxvar_16x16[blk64_idx][lvl1_idx] -
1792 minvar_16x16[blk64_idx][lvl1_idx]);
1793
1794 if (var_32x32 > thresholds[2] ||
1795 (!is_key_frame && var_32x32 > (thresholds[2] >> 1) &&
1796 var_32x32 > (avg_16x16[blk64_idx][lvl1_idx] >> 1))) {
1797 force_split[5 + blk64_scale_idx + lvl1_idx] = PART_EVAL_ONLY_SPLIT;
1798 force_split[blk64_idx + 1] = PART_EVAL_ONLY_SPLIT;
1799 force_split[0] = PART_EVAL_ONLY_SPLIT;
1800 } else if (!is_key_frame && is_360p_or_smaller &&
1801 ((max_min_var_16X16_diff > (thresholds[2] >> 1) &&
1802 maxvar_16x16[blk64_idx][lvl1_idx] > thresholds[2]) ||
1803 (cpi->sf.rt_sf.prefer_large_partition_blocks &&
1804 x->content_state_sb.source_sad_nonrd > kLowSad &&
1805 cpi->rc.frame_source_sad < frame_sad_thresh &&
1806 maxvar_16x16[blk64_idx][lvl1_idx] > (thresholds[2] >> 4) &&
1807 maxvar_16x16[blk64_idx][lvl1_idx] >
1808 (minvar_16x16[blk64_idx][lvl1_idx] << 2)))) {
1809 force_split[5 + blk64_scale_idx + lvl1_idx] = PART_EVAL_ONLY_SPLIT;
1810 force_split[blk64_idx + 1] = PART_EVAL_ONLY_SPLIT;
1811 force_split[0] = PART_EVAL_ONLY_SPLIT;
1812 }
1813 }
1814 }
1815 if (force_split[1 + blk64_idx] == PART_EVAL_ALL) {
1816 fill_variance_tree(&vt->split[blk64_idx], BLOCK_64X64);
1817 get_variance(&vt->split[blk64_idx].part_variances.none);
1818 var_64x64 = vt->split[blk64_idx].part_variances.none.variance;
1819 max_var_64x64 = AOMMAX(var_64x64, max_var_64x64);
1820 min_var_64x64 = AOMMIN(var_64x64, min_var_64x64);
1821 // If the difference of the max-min variances of sub-blocks or max
1822 // variance of a sub-block is above some threshold of then force this
1823 // block to split. Only checking this for noise level >= medium, if
1824 // encoder is in SVC or if we already forced large blocks.
1825 const int max_min_var_32x32_diff =
1826 max_var_32x32[blk64_idx] - min_var_32x32[blk64_idx];
1827 const int check_max_var = max_var_32x32[blk64_idx] > thresholds[1] >> 1;
1828 const bool check_noise_lvl = noise_level >= kMedium ||
1829 cpi->ppi->use_svc ||
1830 cpi->sf.rt_sf.prefer_large_partition_blocks;
1831 const int64_t set_threshold = 3 * (thresholds[1] >> 3);
1832
1833 if (!is_key_frame && max_min_var_32x32_diff > set_threshold &&
1834 check_max_var && check_noise_lvl) {
1835 force_split[1 + blk64_idx] = PART_EVAL_ONLY_SPLIT;
1836 force_split[0] = PART_EVAL_ONLY_SPLIT;
1837 }
1838 avg_64x64 += var_64x64;
1839 }
1840 if (is_small_sb) force_split[0] = PART_EVAL_ONLY_SPLIT;
1841 }
1842
1843 if (force_split[0] == PART_EVAL_ALL) {
1844 fill_variance_tree(vt, BLOCK_128X128);
1845 get_variance(&vt->part_variances.none);
1846 const int set_avg_64x64 = (9 * avg_64x64) >> 5;
1847 if (!is_key_frame && vt->part_variances.none.variance > set_avg_64x64)
1848 force_split[0] = PART_EVAL_ONLY_SPLIT;
1849
1850 if (!is_key_frame &&
1851 (max_var_64x64 - min_var_64x64) > 3 * (thresholds[0] >> 3) &&
1852 max_var_64x64 > thresholds[0] >> 1)
1853 force_split[0] = PART_EVAL_ONLY_SPLIT;
1854 }
1855
1856 if (mi_col + 32 > tile->mi_col_end || mi_row + 32 > tile->mi_row_end ||
1857 !set_vt_partitioning(cpi, xd, tile, vt, BLOCK_128X128, mi_row, mi_col,
1858 thresholds[0], BLOCK_16X16, force_split[0])) {
1859 for (int blk64_idx = 0; blk64_idx < num_64x64_blocks; ++blk64_idx) {
1860 const int x64_idx = GET_BLK_IDX_X(blk64_idx, 4);
1861 const int y64_idx = GET_BLK_IDX_Y(blk64_idx, 4);
1862 const int blk64_scale_idx = blk64_idx << 2;
1863
1864 // Now go through the entire structure, splitting every block size until
1865 // we get to one that's got a variance lower than our threshold.
1866 if (set_vt_partitioning(cpi, xd, tile, &vt->split[blk64_idx], BLOCK_64X64,
1867 mi_row + y64_idx, mi_col + x64_idx, thresholds[1],
1868 BLOCK_16X16, force_split[1 + blk64_idx]))
1869 continue;
1870 for (int lvl1_idx = 0; lvl1_idx < 4; ++lvl1_idx) {
1871 const int x32_idx = GET_BLK_IDX_X(lvl1_idx, 3);
1872 const int y32_idx = GET_BLK_IDX_Y(lvl1_idx, 3);
1873 const int lvl1_scale_idx = (blk64_scale_idx + lvl1_idx) << 2;
1874 if (set_vt_partitioning(
1875 cpi, xd, tile, &vt->split[blk64_idx].split[lvl1_idx],
1876 BLOCK_32X32, (mi_row + y64_idx + y32_idx),
1877 (mi_col + x64_idx + x32_idx), thresholds[2], BLOCK_16X16,
1878 force_split[5 + blk64_scale_idx + lvl1_idx]))
1879 continue;
1880 for (int lvl2_idx = 0; lvl2_idx < 4; ++lvl2_idx) {
1881 const int x16_idx = GET_BLK_IDX_X(lvl2_idx, 2);
1882 const int y16_idx = GET_BLK_IDX_Y(lvl2_idx, 2);
1883 const int split_index = 21 + lvl1_scale_idx + lvl2_idx;
1884 VP16x16 *vtemp =
1885 &vt->split[blk64_idx].split[lvl1_idx].split[lvl2_idx];
1886 if (set_vt_partitioning(cpi, xd, tile, vtemp, BLOCK_16X16,
1887 mi_row + y64_idx + y32_idx + y16_idx,
1888 mi_col + x64_idx + x32_idx + x16_idx,
1889 thresholds[3], BLOCK_8X8,
1890 force_split[split_index]))
1891 continue;
1892 for (int lvl3_idx = 0; lvl3_idx < 4; ++lvl3_idx) {
1893 const int x8_idx = GET_BLK_IDX_X(lvl3_idx, 1);
1894 const int y8_idx = GET_BLK_IDX_Y(lvl3_idx, 1);
1895 set_block_size(cpi, (mi_row + y64_idx + y32_idx + y16_idx + y8_idx),
1896 (mi_col + x64_idx + x32_idx + x16_idx + x8_idx),
1897 BLOCK_8X8);
1898 }
1899 }
1900 }
1901 }
1902 }
1903
1904 if (cpi->sf.rt_sf.short_circuit_low_temp_var) {
1905 set_low_temp_var_flag(cpi, &x->part_search_info, xd, vt, thresholds,
1906 ref_frame_partition, mi_col, mi_row, is_small_sb);
1907 }
1908
1909 aom_free(vt);
1910 #if CONFIG_COLLECT_COMPONENT_TIMING
1911 end_timing(cpi, choose_var_based_partitioning_time);
1912 #endif
1913 return 0;
1914 }
1915