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
33 extern const uint8_t AV1_VAR_OFFS[];
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 AOM_INLINE void tree_to_node(void *data, BLOCK_SIZE bsize,
52 variance_node *node) {
53 int i;
54 node->part_variances = NULL;
55 switch (bsize) {
56 case BLOCK_128X128: {
57 VP128x128 *vt = (VP128x128 *)data;
58 node->part_variances = &vt->part_variances;
59 for (i = 0; i < 4; i++)
60 node->split[i] = &vt->split[i].part_variances.none;
61 break;
62 }
63 case BLOCK_64X64: {
64 VP64x64 *vt = (VP64x64 *)data;
65 node->part_variances = &vt->part_variances;
66 for (i = 0; i < 4; i++)
67 node->split[i] = &vt->split[i].part_variances.none;
68 break;
69 }
70 case BLOCK_32X32: {
71 VP32x32 *vt = (VP32x32 *)data;
72 node->part_variances = &vt->part_variances;
73 for (i = 0; i < 4; i++)
74 node->split[i] = &vt->split[i].part_variances.none;
75 break;
76 }
77 case BLOCK_16X16: {
78 VP16x16 *vt = (VP16x16 *)data;
79 node->part_variances = &vt->part_variances;
80 for (i = 0; i < 4; i++)
81 node->split[i] = &vt->split[i].part_variances.none;
82 break;
83 }
84 case BLOCK_8X8: {
85 VP8x8 *vt = (VP8x8 *)data;
86 node->part_variances = &vt->part_variances;
87 for (i = 0; i < 4; i++)
88 node->split[i] = &vt->split[i].part_variances.none;
89 break;
90 }
91 default: {
92 VP4x4 *vt = (VP4x4 *)data;
93 assert(bsize == BLOCK_4X4);
94 node->part_variances = &vt->part_variances;
95 for (i = 0; i < 4; i++) node->split[i] = &vt->split[i];
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 AOM_INLINE void fill_variance(uint32_t s2, int32_t s, int c,
103 VPartVar *v) {
104 v->sum_square_error = s2;
105 v->sum_error = s;
106 v->log2_count = c;
107 }
108
get_variance(VPartVar * v)109 static AOM_INLINE void get_variance(VPartVar *v) {
110 v->variance =
111 (int)(256 * (v->sum_square_error -
112 (uint32_t)(((int64_t)v->sum_error * v->sum_error) >>
113 v->log2_count)) >>
114 v->log2_count);
115 }
116
sum_2_variances(const VPartVar * a,const VPartVar * b,VPartVar * r)117 static AOM_INLINE void sum_2_variances(const VPartVar *a, const VPartVar *b,
118 VPartVar *r) {
119 assert(a->log2_count == b->log2_count);
120 fill_variance(a->sum_square_error + b->sum_square_error,
121 a->sum_error + b->sum_error, a->log2_count + 1, r);
122 }
123
fill_variance_tree(void * data,BLOCK_SIZE bsize)124 static AOM_INLINE void fill_variance_tree(void *data, BLOCK_SIZE bsize) {
125 variance_node node;
126 memset(&node, 0, sizeof(node));
127 tree_to_node(data, bsize, &node);
128 sum_2_variances(node.split[0], node.split[1], &node.part_variances->horz[0]);
129 sum_2_variances(node.split[2], node.split[3], &node.part_variances->horz[1]);
130 sum_2_variances(node.split[0], node.split[2], &node.part_variances->vert[0]);
131 sum_2_variances(node.split[1], node.split[3], &node.part_variances->vert[1]);
132 sum_2_variances(&node.part_variances->vert[0], &node.part_variances->vert[1],
133 &node.part_variances->none);
134 }
135
set_block_size(AV1_COMP * const cpi,int mi_row,int mi_col,BLOCK_SIZE bsize)136 static AOM_INLINE void set_block_size(AV1_COMP *const cpi, int mi_row,
137 int mi_col, BLOCK_SIZE bsize) {
138 if (cpi->common.mi_params.mi_cols > mi_col &&
139 cpi->common.mi_params.mi_rows > mi_row) {
140 CommonModeInfoParams *mi_params = &cpi->common.mi_params;
141 const int mi_grid_idx = get_mi_grid_idx(mi_params, mi_row, mi_col);
142 const int mi_alloc_idx = get_alloc_mi_idx(mi_params, mi_row, mi_col);
143 MB_MODE_INFO *mi = mi_params->mi_grid_base[mi_grid_idx] =
144 &mi_params->mi_alloc[mi_alloc_idx];
145 mi->bsize = bsize;
146 }
147 }
148
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)149 static int set_vt_partitioning(AV1_COMP *cpi, MACROBLOCKD *const xd,
150 const TileInfo *const tile, void *data,
151 BLOCK_SIZE bsize, int mi_row, int mi_col,
152 int64_t threshold, BLOCK_SIZE bsize_min,
153 PART_EVAL_STATUS force_split) {
154 AV1_COMMON *const cm = &cpi->common;
155 variance_node vt;
156 const int block_width = mi_size_wide[bsize];
157 const int block_height = mi_size_high[bsize];
158 int bs_width_check = block_width;
159 int bs_height_check = block_height;
160 int bs_width_vert_check = block_width >> 1;
161 int bs_height_horiz_check = block_height >> 1;
162 // On the right and bottom boundary we only need to check
163 // if half the bsize fits, because boundary is extended
164 // up to 64. So do this check only for sb_size = 64X64.
165 if (cm->seq_params->sb_size == BLOCK_64X64) {
166 if (tile->mi_col_end == cm->mi_params.mi_cols) {
167 bs_width_check = (block_width >> 1) + 1;
168 bs_width_vert_check = (block_width >> 2) + 1;
169 }
170 if (tile->mi_row_end == cm->mi_params.mi_rows) {
171 bs_height_check = (block_height >> 1) + 1;
172 bs_height_horiz_check = (block_height >> 2) + 1;
173 }
174 }
175
176 assert(block_height == block_width);
177 tree_to_node(data, bsize, &vt);
178
179 if (mi_col + bs_width_check <= tile->mi_col_end &&
180 mi_row + bs_height_check <= tile->mi_row_end &&
181 force_split == PART_EVAL_ONLY_NONE) {
182 set_block_size(cpi, mi_row, mi_col, bsize);
183 return 1;
184 }
185 if (force_split == PART_EVAL_ONLY_SPLIT) return 0;
186
187 // For bsize=bsize_min (16x16/8x8 for 8x8/4x4 downsampling), select if
188 // variance is below threshold, otherwise split will be selected.
189 // No check for vert/horiz split as too few samples for variance.
190 if (bsize == bsize_min) {
191 // Variance already computed to set the force_split.
192 if (frame_is_intra_only(cm)) get_variance(&vt.part_variances->none);
193 if (mi_col + bs_width_check <= tile->mi_col_end &&
194 mi_row + bs_height_check <= tile->mi_row_end &&
195 vt.part_variances->none.variance < threshold) {
196 set_block_size(cpi, mi_row, mi_col, bsize);
197 return 1;
198 }
199 return 0;
200 } else if (bsize > bsize_min) {
201 // Variance already computed to set the force_split.
202 if (frame_is_intra_only(cm)) get_variance(&vt.part_variances->none);
203 // For key frame: take split for bsize above 32X32 or very high variance.
204 if (frame_is_intra_only(cm) &&
205 (bsize > BLOCK_32X32 ||
206 vt.part_variances->none.variance > (threshold << 4))) {
207 return 0;
208 }
209 // If variance is low, take the bsize (no split).
210 if (mi_col + bs_width_check <= tile->mi_col_end &&
211 mi_row + bs_height_check <= tile->mi_row_end &&
212 vt.part_variances->none.variance < threshold) {
213 set_block_size(cpi, mi_row, mi_col, bsize);
214 return 1;
215 }
216 // Check vertical split.
217 if (mi_row + bs_height_check <= tile->mi_row_end &&
218 mi_col + bs_width_vert_check <= tile->mi_col_end) {
219 BLOCK_SIZE subsize = get_partition_subsize(bsize, PARTITION_VERT);
220 get_variance(&vt.part_variances->vert[0]);
221 get_variance(&vt.part_variances->vert[1]);
222 if (vt.part_variances->vert[0].variance < threshold &&
223 vt.part_variances->vert[1].variance < threshold &&
224 get_plane_block_size(subsize, xd->plane[1].subsampling_x,
225 xd->plane[1].subsampling_y) < BLOCK_INVALID) {
226 set_block_size(cpi, mi_row, mi_col, subsize);
227 set_block_size(cpi, mi_row, mi_col + block_width / 2, subsize);
228 return 1;
229 }
230 }
231 // Check horizontal split.
232 if (mi_col + bs_width_check <= tile->mi_col_end &&
233 mi_row + bs_height_horiz_check <= tile->mi_row_end) {
234 BLOCK_SIZE subsize = get_partition_subsize(bsize, PARTITION_HORZ);
235 get_variance(&vt.part_variances->horz[0]);
236 get_variance(&vt.part_variances->horz[1]);
237 if (vt.part_variances->horz[0].variance < threshold &&
238 vt.part_variances->horz[1].variance < threshold &&
239 get_plane_block_size(subsize, xd->plane[1].subsampling_x,
240 xd->plane[1].subsampling_y) < BLOCK_INVALID) {
241 set_block_size(cpi, mi_row, mi_col, subsize);
242 set_block_size(cpi, mi_row + block_height / 2, mi_col, subsize);
243 return 1;
244 }
245 }
246 return 0;
247 }
248 return 0;
249 }
250
all_blks_inside(int x16_idx,int y16_idx,int pixels_wide,int pixels_high)251 static AOM_INLINE int all_blks_inside(int x16_idx, int y16_idx, int pixels_wide,
252 int pixels_high) {
253 int all_inside = 1;
254 for (int k = 0; k < 4; k++) {
255 all_inside &= ((x16_idx + ((k & 1) << 3)) < pixels_wide);
256 all_inside &= ((y16_idx + ((k >> 1) << 3)) < pixels_high);
257 }
258 return all_inside;
259 }
260
261 #if CONFIG_AV1_HIGHBITDEPTH
262 // TODO(yunqingwang): Perform average of four 8x8 blocks similar to lowbd
fill_variance_8x8avg_highbd(const uint8_t * s,int sp,const uint8_t * d,int dp,int x16_idx,int y16_idx,VP16x16 * vst,int pixels_wide,int pixels_high,int is_key_frame)263 static AOM_INLINE void fill_variance_8x8avg_highbd(
264 const uint8_t *s, int sp, const uint8_t *d, int dp, int x16_idx,
265 int y16_idx, VP16x16 *vst, int pixels_wide, int pixels_high,
266 int is_key_frame) {
267 for (int k = 0; k < 4; k++) {
268 const int x8_idx = x16_idx + ((k & 1) << 3);
269 const int y8_idx = y16_idx + ((k >> 1) << 3);
270 unsigned int sse = 0;
271 int sum = 0;
272 if (x8_idx < pixels_wide && y8_idx < pixels_high) {
273 int s_avg;
274 int d_avg = 128;
275 s_avg = aom_highbd_avg_8x8(s + y8_idx * sp + x8_idx, sp);
276 if (!is_key_frame)
277 d_avg = aom_highbd_avg_8x8(d + y8_idx * dp + x8_idx, dp);
278
279 sum = s_avg - d_avg;
280 sse = sum * sum;
281 }
282 fill_variance(sse, sum, 0, &vst->split[k].part_variances.none);
283 }
284 }
285 #endif
286
fill_variance_8x8avg_lowbd(const uint8_t * s,int sp,const uint8_t * d,int dp,int x16_idx,int y16_idx,VP16x16 * vst,int pixels_wide,int pixels_high,int is_key_frame)287 static AOM_INLINE void fill_variance_8x8avg_lowbd(const uint8_t *s, int sp,
288 const uint8_t *d, int dp,
289 int x16_idx, int y16_idx,
290 VP16x16 *vst, int pixels_wide,
291 int pixels_high,
292 int is_key_frame) {
293 unsigned int sse[4] = { 0 };
294 int sum[4] = { 0 };
295 int d_avg[4] = { 128, 128, 128, 128 };
296 int s_avg[4];
297
298 if (all_blks_inside(x16_idx, y16_idx, pixels_wide, pixels_high)) {
299 aom_avg_8x8_quad(s, sp, x16_idx, y16_idx, s_avg);
300 if (!is_key_frame) aom_avg_8x8_quad(d, dp, x16_idx, y16_idx, d_avg);
301 for (int k = 0; k < 4; k++) {
302 sum[k] = s_avg[k] - d_avg[k];
303 sse[k] = sum[k] * sum[k];
304 }
305 } else {
306 for (int k = 0; k < 4; k++) {
307 const int x8_idx = x16_idx + ((k & 1) << 3);
308 const int y8_idx = y16_idx + ((k >> 1) << 3);
309 if (x8_idx < pixels_wide && y8_idx < pixels_high) {
310 s_avg[k] = aom_avg_8x8(s + y8_idx * sp + x8_idx, sp);
311 if (!is_key_frame) d_avg[k] = aom_avg_8x8(d + y8_idx * dp + x8_idx, dp);
312 sum[k] = s_avg[k] - d_avg[k];
313 sse[k] = sum[k] * sum[k];
314 }
315 }
316 }
317
318 for (int k = 0; k < 4; k++) {
319 fill_variance(sse[k], sum[k], 0, &vst->split[k].part_variances.none);
320 }
321 }
322
323 // Obtain parameters required to calculate variance (such as sum, sse, etc,.)
324 // at 8x8 sub-block level for a given 16x16 block.
fill_variance_8x8avg(const uint8_t * s,int sp,const uint8_t * d,int dp,int x16_idx,int y16_idx,VP16x16 * vst,int highbd_flag,int pixels_wide,int pixels_high,int is_key_frame)325 static AOM_INLINE void fill_variance_8x8avg(const uint8_t *s, int sp,
326 const uint8_t *d, int dp,
327 int x16_idx, int y16_idx,
328 VP16x16 *vst, int highbd_flag,
329 int pixels_wide, int pixels_high,
330 int is_key_frame) {
331 #if CONFIG_AV1_HIGHBITDEPTH
332 if (highbd_flag) {
333 fill_variance_8x8avg_highbd(s, sp, d, dp, x16_idx, y16_idx, vst,
334 pixels_wide, pixels_high, is_key_frame);
335 return;
336 }
337 #else
338 (void)highbd_flag;
339 #endif // CONFIG_AV1_HIGHBITDEPTH
340 fill_variance_8x8avg_lowbd(s, sp, d, dp, x16_idx, y16_idx, vst, pixels_wide,
341 pixels_high, is_key_frame);
342 }
343
compute_minmax_8x8(const uint8_t * s,int sp,const uint8_t * d,int dp,int x16_idx,int y16_idx,int highbd_flag,int pixels_wide,int pixels_high)344 static int compute_minmax_8x8(const uint8_t *s, int sp, const uint8_t *d,
345 int dp, int x16_idx, int y16_idx,
346 #if CONFIG_AV1_HIGHBITDEPTH
347 int highbd_flag,
348 #endif
349 int pixels_wide, int pixels_high) {
350 int k;
351 int minmax_max = 0;
352 int minmax_min = 255;
353 // Loop over the 4 8x8 subblocks.
354 for (k = 0; k < 4; k++) {
355 int x8_idx = x16_idx + ((k & 1) << 3);
356 int y8_idx = y16_idx + ((k >> 1) << 3);
357 int min = 0;
358 int max = 0;
359 if (x8_idx < pixels_wide && y8_idx < pixels_high) {
360 #if CONFIG_AV1_HIGHBITDEPTH
361 if (highbd_flag & YV12_FLAG_HIGHBITDEPTH) {
362 aom_highbd_minmax_8x8(s + y8_idx * sp + x8_idx, sp,
363 d + y8_idx * dp + x8_idx, dp, &min, &max);
364 } else {
365 aom_minmax_8x8(s + y8_idx * sp + x8_idx, sp, d + y8_idx * dp + x8_idx,
366 dp, &min, &max);
367 }
368 #else
369 aom_minmax_8x8(s + y8_idx * sp + x8_idx, sp, d + y8_idx * dp + x8_idx, dp,
370 &min, &max);
371 #endif
372 if ((max - min) > minmax_max) minmax_max = (max - min);
373 if ((max - min) < minmax_min) minmax_min = (max - min);
374 }
375 }
376 return (minmax_max - minmax_min);
377 }
378
fill_variance_4x4avg(const uint8_t * s,int sp,const uint8_t * d,int dp,int x8_idx,int y8_idx,VP8x8 * vst,int highbd_flag,int pixels_wide,int pixels_high,int is_key_frame,int border_offset_4x4)379 static AOM_INLINE void fill_variance_4x4avg(const uint8_t *s, int sp,
380 const uint8_t *d, int dp,
381 int x8_idx, int y8_idx, VP8x8 *vst,
382 #if CONFIG_AV1_HIGHBITDEPTH
383 int highbd_flag,
384 #endif
385 int pixels_wide, int pixels_high,
386 int is_key_frame,
387 int border_offset_4x4) {
388 int k;
389 for (k = 0; k < 4; k++) {
390 int x4_idx = x8_idx + ((k & 1) << 2);
391 int y4_idx = y8_idx + ((k >> 1) << 2);
392 unsigned int sse = 0;
393 int sum = 0;
394 if (x4_idx < pixels_wide - border_offset_4x4 &&
395 y4_idx < pixels_high - border_offset_4x4) {
396 int s_avg;
397 int d_avg = 128;
398 #if CONFIG_AV1_HIGHBITDEPTH
399 if (highbd_flag & YV12_FLAG_HIGHBITDEPTH) {
400 s_avg = aom_highbd_avg_4x4(s + y4_idx * sp + x4_idx, sp);
401 if (!is_key_frame)
402 d_avg = aom_highbd_avg_4x4(d + y4_idx * dp + x4_idx, dp);
403 } else {
404 s_avg = aom_avg_4x4(s + y4_idx * sp + x4_idx, sp);
405 if (!is_key_frame) d_avg = aom_avg_4x4(d + y4_idx * dp + x4_idx, dp);
406 }
407 #else
408 s_avg = aom_avg_4x4(s + y4_idx * sp + x4_idx, sp);
409 if (!is_key_frame) d_avg = aom_avg_4x4(d + y4_idx * dp + x4_idx, dp);
410 #endif
411
412 sum = s_avg - d_avg;
413 sse = sum * sum;
414 }
415 fill_variance(sse, sum, 0, &vst->split[k].part_variances.none);
416 }
417 }
418
419 // 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)420 static int64_t scale_part_thresh_content(int64_t threshold_base, int speed,
421 int width, int height,
422 int non_reference_frame) {
423 (void)width;
424 (void)height;
425 int64_t threshold = threshold_base;
426 if (non_reference_frame) threshold = (3 * threshold) >> 1;
427 if (speed >= 8) {
428 return (5 * threshold) >> 2;
429 }
430 return threshold;
431 }
432
tune_thresh_based_on_qindex_window(int qindex,int th,int win,int fac,int64_t thresholds[])433 static AOM_INLINE void tune_thresh_based_on_qindex_window(
434 int qindex, int th, int win, int fac, int64_t thresholds[]) {
435 double weight;
436
437 if (qindex < th - win)
438 weight = 1.0;
439 else if (qindex > th + win)
440 weight = 0.0;
441 else
442 weight = 1.0 - (qindex - th + win) / (2 * win);
443 thresholds[1] =
444 (int)((1 - weight) * (thresholds[1] << 1) + weight * thresholds[1]);
445 thresholds[2] =
446 (int)((1 - weight) * (thresholds[2] << 1) + weight * thresholds[2]);
447 thresholds[3] =
448 (int)((1 - weight) * (thresholds[3] << fac) + weight * thresholds[3]);
449 }
450
set_vbp_thresholds(AV1_COMP * cpi,int64_t thresholds[],int q,int content_lowsumdiff,int source_sad_nonrd,int source_sad_rd,int segment_id,uint64_t blk_sad,int lighting_change)451 static AOM_INLINE void set_vbp_thresholds(AV1_COMP *cpi, int64_t thresholds[],
452 int q, int content_lowsumdiff,
453 int source_sad_nonrd,
454 int source_sad_rd, int segment_id,
455 uint64_t blk_sad,
456 int lighting_change) {
457 AV1_COMMON *const cm = &cpi->common;
458 const int is_key_frame = frame_is_intra_only(cm);
459 const int threshold_multiplier = is_key_frame ? 120 : 1;
460 const int ac_q = av1_ac_quant_QTX(q, 0, cm->seq_params->bit_depth);
461 int64_t threshold_base = (int64_t)(threshold_multiplier * ac_q);
462 const int current_qindex = cm->quant_params.base_qindex;
463 const int threshold_left_shift = cpi->sf.rt_sf.var_part_split_threshold_shift;
464
465 if (is_key_frame) {
466 if (cpi->sf.rt_sf.force_large_partition_blocks_intra) {
467 const int shift_steps =
468 threshold_left_shift - (cpi->oxcf.mode == ALLINTRA ? 7 : 8);
469 assert(shift_steps >= 0);
470 threshold_base <<= shift_steps;
471 }
472 thresholds[0] = threshold_base;
473 thresholds[1] = threshold_base;
474 if (cm->width * cm->height < 1280 * 720) {
475 thresholds[2] = threshold_base / 3;
476 thresholds[3] = threshold_base >> 1;
477 } else {
478 int shift_val = 2;
479 if (cpi->sf.rt_sf.force_large_partition_blocks_intra) {
480 shift_val = 0;
481 }
482
483 thresholds[2] = threshold_base >> shift_val;
484 thresholds[3] = threshold_base >> shift_val;
485 }
486 thresholds[4] = threshold_base << 2;
487 return;
488 }
489
490 // Increase partition thresholds for noisy content. Apply it only for
491 // superblocks where sumdiff is low, as we assume the sumdiff of superblock
492 // whose only change is due to noise will be low (i.e, noise will average
493 // out over large block).
494 if (cpi->noise_estimate.enabled && content_lowsumdiff &&
495 (cm->width * cm->height > 640 * 480) &&
496 cm->current_frame.frame_number > 60) {
497 NOISE_LEVEL noise_level =
498 av1_noise_estimate_extract_level(&cpi->noise_estimate);
499 if (noise_level == kHigh)
500 threshold_base = (5 * threshold_base) >> 1;
501 else if (noise_level == kMedium &&
502 !cpi->sf.rt_sf.prefer_large_partition_blocks)
503 threshold_base = (5 * threshold_base) >> 2;
504 }
505 // TODO(kyslov) Enable var based partition adjusment on temporal denoising
506 #if 0 // CONFIG_AV1_TEMPORAL_DENOISING
507 if (cpi->oxcf.noise_sensitivity > 0 && denoise_svc(cpi) &&
508 cpi->oxcf.speed > 5 && cpi->denoiser.denoising_level >= kDenLow)
509 threshold_base =
510 av1_scale_part_thresh(threshold_base, cpi->denoiser.denoising_level,
511 content_state, cpi->svc.temporal_layer_id);
512 else
513 threshold_base =
514 scale_part_thresh_content(threshold_base, cpi->oxcf.speed, cm->width,
515 cm->height, cpi->ppi->rtc_ref.non_reference_frame);
516 #else
517 // Increase base variance threshold based on content_state/sum_diff level.
518 threshold_base = scale_part_thresh_content(
519 threshold_base, cpi->oxcf.speed, cm->width, cm->height,
520 cpi->ppi->rtc_ref.non_reference_frame);
521 #endif
522 thresholds[0] = threshold_base >> 1;
523 thresholds[1] = threshold_base;
524 thresholds[3] = threshold_base << threshold_left_shift;
525 if (cm->width >= 1280 && cm->height >= 720)
526 thresholds[3] = thresholds[3] << 1;
527 if (cm->width * cm->height <= 352 * 288) {
528 const int qindex_thr[5][2] = {
529 { 200, 220 }, { 140, 170 }, { 120, 150 }, { 200, 210 }, { 170, 220 },
530 };
531 int th_idx = 0;
532 if (cpi->sf.rt_sf.var_part_based_on_qidx >= 1)
533 th_idx =
534 (source_sad_rd <= kLowSad) ? cpi->sf.rt_sf.var_part_based_on_qidx : 0;
535 if (cpi->sf.rt_sf.var_part_based_on_qidx >= 3)
536 th_idx = cpi->sf.rt_sf.var_part_based_on_qidx;
537 const int qindex_low_thr = qindex_thr[th_idx][0];
538 const int qindex_high_thr = qindex_thr[th_idx][1];
539 if (current_qindex >= qindex_high_thr) {
540 threshold_base = (5 * threshold_base) >> 1;
541 thresholds[1] = threshold_base >> 3;
542 thresholds[2] = threshold_base << 2;
543 thresholds[3] = threshold_base << 5;
544 } else if (current_qindex < qindex_low_thr) {
545 thresholds[1] = threshold_base >> 3;
546 thresholds[2] = threshold_base >> 1;
547 thresholds[3] = threshold_base << 3;
548 } else {
549 int64_t qi_diff_low = current_qindex - qindex_low_thr;
550 int64_t qi_diff_high = qindex_high_thr - current_qindex;
551 int64_t threshold_diff = qindex_high_thr - qindex_low_thr;
552 int64_t threshold_base_high = (5 * threshold_base) >> 1;
553
554 threshold_diff = threshold_diff > 0 ? threshold_diff : 1;
555 threshold_base =
556 (qi_diff_low * threshold_base_high + qi_diff_high * threshold_base) /
557 threshold_diff;
558 thresholds[1] = threshold_base >> 3;
559 thresholds[2] = ((qi_diff_low * threshold_base) +
560 qi_diff_high * (threshold_base >> 1)) /
561 threshold_diff;
562 thresholds[3] = ((qi_diff_low * (threshold_base << 5)) +
563 qi_diff_high * (threshold_base << 3)) /
564 threshold_diff;
565 }
566 } else if (cm->width < 1280 && cm->height < 720) {
567 thresholds[2] = (5 * threshold_base) >> 2;
568 } else if (cm->width < 1920 && cm->height < 1080) {
569 thresholds[2] = threshold_base << 1;
570 } else if (cm->width < 2560 && cm->height < 1440) {
571 thresholds[2] = (5 * threshold_base) >> 1;
572 } else {
573 thresholds[2] = (7 * threshold_base) >> 1;
574 }
575 // Tune thresholds less or more aggressively to prefer larger partitions
576 if (cpi->sf.rt_sf.prefer_large_partition_blocks >= 3) {
577 double weight;
578 const int win = 20;
579 if (current_qindex < QINDEX_LARGE_BLOCK_THR - win)
580 weight = 1.0;
581 else if (current_qindex > QINDEX_LARGE_BLOCK_THR + win)
582 weight = 0.0;
583 else
584 weight =
585 1.0 - (current_qindex - QINDEX_LARGE_BLOCK_THR + win) / (2 * win);
586 if (cm->width * cm->height > 640 * 480) {
587 for (int i = 0; i < 4; i++) {
588 thresholds[i] <<= 1;
589 }
590 }
591 if (cm->width * cm->height <= 352 * 288) {
592 thresholds[3] = INT64_MAX;
593 if (segment_id == 0) {
594 thresholds[1] <<= 2;
595 thresholds[2] <<= (source_sad_nonrd <= kLowSad) ? 5 : 4;
596 } else {
597 thresholds[1] <<= 1;
598 thresholds[2] <<= 3;
599 }
600 // Allow for split to 8x8 for superblocks where part of it has
601 // moving boundary. So allow for sb with source_sad above threshold,
602 // and avoid very large source_sad or high source content, to avoid
603 // too many 8x8 within superblock.
604 if (segment_id == 0 && cpi->rc.avg_source_sad < 25000 &&
605 blk_sad > 25000 && blk_sad < 50000 && !lighting_change) {
606 thresholds[2] = (3 * thresholds[2]) >> 2;
607 thresholds[3] = thresholds[2] << 3;
608 }
609 // Condition the increase of partition thresholds on the segment
610 // and the content. Avoid the increase for superblocks which have
611 // high source sad, unless the whole frame has very high motion
612 // (i.e, cpi->rc.avg_source_sad is very large, in which case all blocks
613 // have high source sad).
614 } else if (cm->width * cm->height > 640 * 480 && segment_id == 0 &&
615 (source_sad_nonrd != kHighSad ||
616 cpi->rc.avg_source_sad > 50000)) {
617 thresholds[0] = (3 * thresholds[0]) >> 1;
618 thresholds[3] = INT64_MAX;
619 if (current_qindex > QINDEX_LARGE_BLOCK_THR) {
620 thresholds[1] =
621 (int)((1 - weight) * (thresholds[1] << 1) + weight * thresholds[1]);
622 thresholds[2] =
623 (int)((1 - weight) * (thresholds[2] << 1) + weight * thresholds[2]);
624 }
625 } else if (current_qindex > QINDEX_LARGE_BLOCK_THR && segment_id == 0 &&
626 (source_sad_nonrd != kHighSad ||
627 cpi->rc.avg_source_sad > 50000)) {
628 thresholds[1] =
629 (int)((1 - weight) * (thresholds[1] << 2) + weight * thresholds[1]);
630 thresholds[2] =
631 (int)((1 - weight) * (thresholds[2] << 4) + weight * thresholds[2]);
632 thresholds[3] = INT64_MAX;
633 }
634 } else if (cpi->sf.rt_sf.prefer_large_partition_blocks >= 2) {
635 thresholds[1] <<= (source_sad_nonrd <= kLowSad) ? 2 : 0;
636 thresholds[2] =
637 (source_sad_nonrd <= kLowSad) ? (3 * thresholds[2]) : thresholds[2];
638 } else if (cpi->sf.rt_sf.prefer_large_partition_blocks >= 1) {
639 const int fac = (source_sad_nonrd <= kLowSad) ? 2 : 1;
640 tune_thresh_based_on_qindex_window(current_qindex, QINDEX_LARGE_BLOCK_THR,
641 45, fac, thresholds);
642 }
643 if (cpi->sf.part_sf.disable_8x8_part_based_on_qidx && (current_qindex < 128))
644 thresholds[3] = INT64_MAX;
645 }
646
647 // Set temporal variance low flag for superblock 64x64.
648 // 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)649 static AOM_INLINE void set_low_temp_var_flag_64x64(
650 CommonModeInfoParams *mi_params, PartitionSearchInfo *part_info,
651 MACROBLOCKD *xd, VP64x64 *vt, const int64_t thresholds[], int mi_col,
652 int mi_row) {
653 if (xd->mi[0]->bsize == BLOCK_64X64) {
654 if ((vt->part_variances).none.variance < (thresholds[0] >> 1))
655 part_info->variance_low[0] = 1;
656 } else if (xd->mi[0]->bsize == BLOCK_64X32) {
657 for (int i = 0; i < 2; i++) {
658 if (vt->part_variances.horz[i].variance < (thresholds[0] >> 2))
659 part_info->variance_low[i + 1] = 1;
660 }
661 } else if (xd->mi[0]->bsize == BLOCK_32X64) {
662 for (int i = 0; i < 2; i++) {
663 if (vt->part_variances.vert[i].variance < (thresholds[0] >> 2))
664 part_info->variance_low[i + 3] = 1;
665 }
666 } else {
667 static const int idx[4][2] = { { 0, 0 }, { 0, 8 }, { 8, 0 }, { 8, 8 } };
668 for (int i = 0; i < 4; i++) {
669 const int idx_str =
670 mi_params->mi_stride * (mi_row + idx[i][0]) + mi_col + idx[i][1];
671 MB_MODE_INFO **this_mi = mi_params->mi_grid_base + idx_str;
672
673 if (mi_params->mi_cols <= mi_col + idx[i][1] ||
674 mi_params->mi_rows <= mi_row + idx[i][0])
675 continue;
676
677 if (*this_mi == NULL) continue;
678
679 if ((*this_mi)->bsize == BLOCK_32X32) {
680 int64_t threshold_32x32 = (5 * thresholds[1]) >> 3;
681 if (vt->split[i].part_variances.none.variance < threshold_32x32)
682 part_info->variance_low[i + 5] = 1;
683 } else {
684 // For 32x16 and 16x32 blocks, the flag is set on each 16x16 block
685 // inside.
686 if ((*this_mi)->bsize == BLOCK_16X16 ||
687 (*this_mi)->bsize == BLOCK_32X16 ||
688 (*this_mi)->bsize == BLOCK_16X32) {
689 for (int j = 0; j < 4; j++) {
690 if (vt->split[i].split[j].part_variances.none.variance <
691 (thresholds[2] >> 8))
692 part_info->variance_low[(i << 2) + j + 9] = 1;
693 }
694 }
695 }
696 }
697 }
698 }
699
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)700 static AOM_INLINE void set_low_temp_var_flag_128x128(
701 CommonModeInfoParams *mi_params, PartitionSearchInfo *part_info,
702 MACROBLOCKD *xd, VP128x128 *vt, const int64_t thresholds[], int mi_col,
703 int mi_row) {
704 if (xd->mi[0]->bsize == BLOCK_128X128) {
705 if (vt->part_variances.none.variance < (thresholds[0] >> 1))
706 part_info->variance_low[0] = 1;
707 } else if (xd->mi[0]->bsize == BLOCK_128X64) {
708 for (int i = 0; i < 2; i++) {
709 if (vt->part_variances.horz[i].variance < (thresholds[0] >> 2))
710 part_info->variance_low[i + 1] = 1;
711 }
712 } else if (xd->mi[0]->bsize == BLOCK_64X128) {
713 for (int i = 0; i < 2; i++) {
714 if (vt->part_variances.vert[i].variance < (thresholds[0] >> 2))
715 part_info->variance_low[i + 3] = 1;
716 }
717 } else {
718 static const int idx64[4][2] = {
719 { 0, 0 }, { 0, 16 }, { 16, 0 }, { 16, 16 }
720 };
721 static const int idx32[4][2] = { { 0, 0 }, { 0, 8 }, { 8, 0 }, { 8, 8 } };
722 for (int i = 0; i < 4; i++) {
723 const int idx_str =
724 mi_params->mi_stride * (mi_row + idx64[i][0]) + mi_col + idx64[i][1];
725 MB_MODE_INFO **mi_64 = mi_params->mi_grid_base + idx_str;
726 if (*mi_64 == NULL) continue;
727 if (mi_params->mi_cols <= mi_col + idx64[i][1] ||
728 mi_params->mi_rows <= mi_row + idx64[i][0])
729 continue;
730 const int64_t threshold_64x64 = (5 * thresholds[1]) >> 3;
731 if ((*mi_64)->bsize == BLOCK_64X64) {
732 if (vt->split[i].part_variances.none.variance < threshold_64x64)
733 part_info->variance_low[5 + i] = 1;
734 } else if ((*mi_64)->bsize == BLOCK_64X32) {
735 for (int j = 0; j < 2; j++)
736 if (vt->split[i].part_variances.horz[j].variance <
737 (threshold_64x64 >> 1))
738 part_info->variance_low[9 + (i << 1) + j] = 1;
739 } else if ((*mi_64)->bsize == BLOCK_32X64) {
740 for (int j = 0; j < 2; j++)
741 if (vt->split[i].part_variances.vert[j].variance <
742 (threshold_64x64 >> 1))
743 part_info->variance_low[17 + (i << 1) + j] = 1;
744 } else {
745 for (int k = 0; k < 4; k++) {
746 const int idx_str1 = mi_params->mi_stride * idx32[k][0] + idx32[k][1];
747 MB_MODE_INFO **mi_32 = mi_params->mi_grid_base + idx_str + idx_str1;
748 if (*mi_32 == NULL) continue;
749
750 if (mi_params->mi_cols <= mi_col + idx64[i][1] + idx32[k][1] ||
751 mi_params->mi_rows <= mi_row + idx64[i][0] + idx32[k][0])
752 continue;
753 const int64_t threshold_32x32 = (5 * thresholds[2]) >> 3;
754 if ((*mi_32)->bsize == BLOCK_32X32) {
755 if (vt->split[i].split[k].part_variances.none.variance <
756 threshold_32x32)
757 part_info->variance_low[25 + (i << 2) + k] = 1;
758 } else {
759 // For 32x16 and 16x32 blocks, the flag is set on each 16x16 block
760 // inside.
761 if ((*mi_32)->bsize == BLOCK_16X16 ||
762 (*mi_32)->bsize == BLOCK_32X16 ||
763 (*mi_32)->bsize == BLOCK_16X32) {
764 for (int j = 0; j < 4; j++) {
765 if (vt->split[i]
766 .split[k]
767 .split[j]
768 .part_variances.none.variance < (thresholds[3] >> 8))
769 part_info->variance_low[41 + (i << 4) + (k << 2) + j] = 1;
770 }
771 }
772 }
773 }
774 }
775 }
776 }
777 }
778
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)779 static AOM_INLINE void set_low_temp_var_flag(
780 AV1_COMP *cpi, PartitionSearchInfo *part_info, MACROBLOCKD *xd,
781 VP128x128 *vt, int64_t thresholds[], MV_REFERENCE_FRAME ref_frame_partition,
782 int mi_col, int mi_row) {
783 AV1_COMMON *const cm = &cpi->common;
784 // Check temporal variance for bsize >= 16x16, if LAST_FRAME was selected.
785 // If the temporal variance is small set the flag
786 // variance_low for the block. The variance threshold can be adjusted, the
787 // higher the more aggressive.
788 if (ref_frame_partition == LAST_FRAME) {
789 const int is_small_sb = (cm->seq_params->sb_size == BLOCK_64X64);
790 if (is_small_sb)
791 set_low_temp_var_flag_64x64(&cm->mi_params, part_info, xd,
792 &(vt->split[0]), thresholds, mi_col, mi_row);
793 else
794 set_low_temp_var_flag_128x128(&cm->mi_params, part_info, xd, vt,
795 thresholds, mi_col, mi_row);
796 }
797 }
798
799 static const int pos_shift_16x16[4][4] = {
800 { 9, 10, 13, 14 }, { 11, 12, 15, 16 }, { 17, 18, 21, 22 }, { 19, 20, 23, 24 }
801 };
802
av1_get_force_skip_low_temp_var_small_sb(const uint8_t * variance_low,int mi_row,int mi_col,BLOCK_SIZE bsize)803 int av1_get_force_skip_low_temp_var_small_sb(const uint8_t *variance_low,
804 int mi_row, int mi_col,
805 BLOCK_SIZE bsize) {
806 // Relative indices of MB inside the superblock.
807 const int mi_x = mi_row & 0xF;
808 const int mi_y = mi_col & 0xF;
809 // Relative indices of 16x16 block inside the superblock.
810 const int i = mi_x >> 2;
811 const int j = mi_y >> 2;
812 int force_skip_low_temp_var = 0;
813 // Set force_skip_low_temp_var based on the block size and block offset.
814 switch (bsize) {
815 case BLOCK_64X64: force_skip_low_temp_var = variance_low[0]; break;
816 case BLOCK_64X32:
817 if (!mi_y && !mi_x) {
818 force_skip_low_temp_var = variance_low[1];
819 } else if (!mi_y && mi_x) {
820 force_skip_low_temp_var = variance_low[2];
821 }
822 break;
823 case BLOCK_32X64:
824 if (!mi_y && !mi_x) {
825 force_skip_low_temp_var = variance_low[3];
826 } else if (mi_y && !mi_x) {
827 force_skip_low_temp_var = variance_low[4];
828 }
829 break;
830 case BLOCK_32X32:
831 if (!mi_y && !mi_x) {
832 force_skip_low_temp_var = variance_low[5];
833 } else if (mi_y && !mi_x) {
834 force_skip_low_temp_var = variance_low[6];
835 } else if (!mi_y && mi_x) {
836 force_skip_low_temp_var = variance_low[7];
837 } else if (mi_y && mi_x) {
838 force_skip_low_temp_var = variance_low[8];
839 }
840 break;
841 case BLOCK_32X16:
842 case BLOCK_16X32:
843 case BLOCK_16X16:
844 force_skip_low_temp_var = variance_low[pos_shift_16x16[i][j]];
845 break;
846 default: break;
847 }
848
849 return force_skip_low_temp_var;
850 }
851
av1_get_force_skip_low_temp_var(const uint8_t * variance_low,int mi_row,int mi_col,BLOCK_SIZE bsize)852 int av1_get_force_skip_low_temp_var(const uint8_t *variance_low, int mi_row,
853 int mi_col, BLOCK_SIZE bsize) {
854 int force_skip_low_temp_var = 0;
855 int x, y;
856 x = (mi_col & 0x1F) >> 4;
857 // y = (mi_row & 0x1F) >> 4;
858 // const int idx64 = (y << 1) + x;
859 y = (mi_row & 0x17) >> 3;
860 const int idx64 = y + x;
861
862 x = (mi_col & 0xF) >> 3;
863 // y = (mi_row & 0xF) >> 3;
864 // const int idx32 = (y << 1) + x;
865 y = (mi_row & 0xB) >> 2;
866 const int idx32 = y + x;
867
868 x = (mi_col & 0x7) >> 2;
869 // y = (mi_row & 0x7) >> 2;
870 // const int idx16 = (y << 1) + x;
871 y = (mi_row & 0x5) >> 1;
872 const int idx16 = y + x;
873 // Set force_skip_low_temp_var based on the block size and block offset.
874 switch (bsize) {
875 case BLOCK_128X128: force_skip_low_temp_var = variance_low[0]; break;
876 case BLOCK_128X64:
877 assert((mi_col & 0x1F) == 0);
878 force_skip_low_temp_var = variance_low[1 + ((mi_row & 0x1F) != 0)];
879 break;
880 case BLOCK_64X128:
881 assert((mi_row & 0x1F) == 0);
882 force_skip_low_temp_var = variance_low[3 + ((mi_col & 0x1F) != 0)];
883 break;
884 case BLOCK_64X64:
885 // Location of this 64x64 block inside the 128x128 superblock
886 force_skip_low_temp_var = variance_low[5 + idx64];
887 break;
888 case BLOCK_64X32:
889 x = (mi_col & 0x1F) >> 4;
890 y = (mi_row & 0x1F) >> 3;
891 /*
892 .---------------.---------------.
893 | x=0,y=0,idx=0 | x=0,y=0,idx=2 |
894 :---------------+---------------:
895 | x=0,y=1,idx=1 | x=1,y=1,idx=3 |
896 :---------------+---------------:
897 | x=0,y=2,idx=4 | x=1,y=2,idx=6 |
898 :---------------+---------------:
899 | x=0,y=3,idx=5 | x=1,y=3,idx=7 |
900 '---------------'---------------'
901 */
902 const int idx64x32 = (x << 1) + (y % 2) + ((y >> 1) << 2);
903 force_skip_low_temp_var = variance_low[9 + idx64x32];
904 break;
905 case BLOCK_32X64:
906 x = (mi_col & 0x1F) >> 3;
907 y = (mi_row & 0x1F) >> 4;
908 const int idx32x64 = (y << 2) + x;
909 force_skip_low_temp_var = variance_low[17 + idx32x64];
910 break;
911 case BLOCK_32X32:
912 force_skip_low_temp_var = variance_low[25 + (idx64 << 2) + idx32];
913 break;
914 case BLOCK_32X16:
915 case BLOCK_16X32:
916 case BLOCK_16X16:
917 force_skip_low_temp_var =
918 variance_low[41 + (idx64 << 4) + (idx32 << 2) + idx16];
919 break;
920 default: break;
921 }
922 return force_skip_low_temp_var;
923 }
924
av1_set_variance_partition_thresholds(AV1_COMP * cpi,int q,int content_lowsumdiff)925 void av1_set_variance_partition_thresholds(AV1_COMP *cpi, int q,
926 int content_lowsumdiff) {
927 SPEED_FEATURES *const sf = &cpi->sf;
928 if (sf->part_sf.partition_search_type != VAR_BASED_PARTITION) {
929 return;
930 } else {
931 set_vbp_thresholds(cpi, cpi->vbp_info.thresholds, q, content_lowsumdiff, 0,
932 0, 0, 0, 0);
933 // The threshold below is not changed locally.
934 cpi->vbp_info.threshold_minmax = 15 + (q >> 3);
935 }
936 }
937
chroma_check(AV1_COMP * cpi,MACROBLOCK * x,BLOCK_SIZE bsize,unsigned int y_sad,unsigned int y_sad_g,int is_key_frame,int zero_motion,unsigned int * uv_sad)938 static AOM_INLINE void chroma_check(AV1_COMP *cpi, MACROBLOCK *x,
939 BLOCK_SIZE bsize, unsigned int y_sad,
940 unsigned int y_sad_g, int is_key_frame,
941 int zero_motion, unsigned int *uv_sad) {
942 int i;
943 MACROBLOCKD *xd = &x->e_mbd;
944 int shift = 3;
945 if (is_key_frame || cpi->oxcf.tool_cfg.enable_monochrome) return;
946
947 if (cpi->oxcf.tune_cfg.content == AOM_CONTENT_SCREEN &&
948 cpi->rc.high_source_sad)
949 shift = 5;
950
951 MB_MODE_INFO *mi = xd->mi[0];
952 const AV1_COMMON *const cm = &cpi->common;
953 const YV12_BUFFER_CONFIG *yv12 = get_ref_frame_yv12_buf(cm, LAST_FRAME);
954 const YV12_BUFFER_CONFIG *yv12_g = get_ref_frame_yv12_buf(cm, GOLDEN_FRAME);
955 const struct scale_factors *const sf =
956 get_ref_scale_factors_const(cm, LAST_FRAME);
957 struct buf_2d dst;
958 unsigned int uv_sad_g = 0;
959
960 for (i = 1; i <= 2; ++i) {
961 struct macroblock_plane *p = &x->plane[i];
962 struct macroblockd_plane *pd = &xd->plane[i];
963 const BLOCK_SIZE bs =
964 get_plane_block_size(bsize, pd->subsampling_x, pd->subsampling_y);
965
966 if (bs != BLOCK_INVALID) {
967 // For last:
968 if (zero_motion) {
969 if (mi->ref_frame[0] == LAST_FRAME) {
970 uv_sad[i - 1] = cpi->ppi->fn_ptr[bs].sdf(
971 p->src.buf, p->src.stride, pd->pre[0].buf, pd->pre[0].stride);
972 } else {
973 uint8_t *src = (i == 1) ? yv12->u_buffer : yv12->v_buffer;
974 setup_pred_plane(&dst, xd->mi[0]->bsize, src, yv12->uv_crop_width,
975 yv12->uv_crop_height, yv12->uv_stride, xd->mi_row,
976 xd->mi_col, sf, xd->plane[i].subsampling_x,
977 xd->plane[i].subsampling_y);
978
979 uv_sad[i - 1] = cpi->ppi->fn_ptr[bs].sdf(p->src.buf, p->src.stride,
980 dst.buf, dst.stride);
981 }
982 } else {
983 uv_sad[i - 1] = cpi->ppi->fn_ptr[bs].sdf(p->src.buf, p->src.stride,
984 pd->dst.buf, pd->dst.stride);
985 }
986
987 // For golden:
988 if (y_sad_g != UINT_MAX) {
989 uint8_t *src = (i == 1) ? yv12_g->u_buffer : yv12_g->v_buffer;
990 setup_pred_plane(&dst, xd->mi[0]->bsize, src, yv12_g->uv_crop_width,
991 yv12_g->uv_crop_height, yv12_g->uv_stride, xd->mi_row,
992 xd->mi_col, sf, xd->plane[i].subsampling_x,
993 xd->plane[i].subsampling_y);
994 uv_sad_g = cpi->ppi->fn_ptr[bs].sdf(p->src.buf, p->src.stride, dst.buf,
995 dst.stride);
996 }
997 }
998
999 if (uv_sad[i - 1] > (y_sad >> 1))
1000 x->color_sensitivity_sb[i - 1] = 1;
1001 else if (uv_sad[i - 1] < (y_sad >> shift))
1002 x->color_sensitivity_sb[i - 1] = 0;
1003 // Borderline case: to be refined at coding block level in nonrd_pickmode,
1004 // for coding block size < sb_size.
1005 else
1006 x->color_sensitivity_sb[i - 1] = 2;
1007
1008 x->color_sensitivity_sb_g[i - 1] = uv_sad_g > y_sad_g / 6;
1009 }
1010 }
1011
fill_variance_tree_leaves(AV1_COMP * cpi,MACROBLOCK * x,VP128x128 * vt,VP16x16 * vt2,PART_EVAL_STATUS * force_split,int avg_16x16[][4],int maxvar_16x16[][4],int minvar_16x16[][4],int * variance4x4downsample,int64_t * thresholds,uint8_t * src,int src_stride,const uint8_t * dst,int dst_stride)1012 static void fill_variance_tree_leaves(
1013 AV1_COMP *cpi, MACROBLOCK *x, VP128x128 *vt, VP16x16 *vt2,
1014 PART_EVAL_STATUS *force_split, int avg_16x16[][4], int maxvar_16x16[][4],
1015 int minvar_16x16[][4], int *variance4x4downsample, int64_t *thresholds,
1016 uint8_t *src, int src_stride, const uint8_t *dst, int dst_stride) {
1017 AV1_COMMON *cm = &cpi->common;
1018 MACROBLOCKD *xd = &x->e_mbd;
1019 const int is_key_frame = frame_is_intra_only(cm);
1020 const int is_small_sb = (cm->seq_params->sb_size == BLOCK_64X64);
1021 const int num_64x64_blocks = is_small_sb ? 1 : 4;
1022 // TODO(kyslov) Bring back compute_minmax_variance with content type detection
1023 const int compute_minmax_variance = 0;
1024 const int segment_id = xd->mi[0]->segment_id;
1025 int pixels_wide = 128, pixels_high = 128;
1026 int border_offset_4x4 = 0;
1027 int temporal_denoising = cpi->sf.rt_sf.use_rtc_tf;
1028 if (is_small_sb) {
1029 pixels_wide = 64;
1030 pixels_high = 64;
1031 }
1032 if (xd->mb_to_right_edge < 0) pixels_wide += (xd->mb_to_right_edge >> 3);
1033 if (xd->mb_to_bottom_edge < 0) pixels_high += (xd->mb_to_bottom_edge >> 3);
1034 #if CONFIG_AV1_TEMPORAL_DENOISING
1035 temporal_denoising |= cpi->oxcf.noise_sensitivity;
1036 #endif
1037 // For temporal filtering or temporal denoiser enabled: since the source
1038 // is modified we need to avoid 4x4 avg along superblock boundary, since
1039 // simd code will load 8 pixels for 4x4 avg and so can access source
1040 // data outside superblock (while its being modified by temporal filter).
1041 // Temporal filtering is never done on key frames.
1042 if (!is_key_frame && temporal_denoising) border_offset_4x4 = 4;
1043 for (int m = 0; m < num_64x64_blocks; m++) {
1044 const int x64_idx = ((m & 1) << 6);
1045 const int y64_idx = ((m >> 1) << 6);
1046 const int m2 = m << 2;
1047 force_split[m + 1] = PART_EVAL_ALL;
1048
1049 for (int i = 0; i < 4; i++) {
1050 const int x32_idx = x64_idx + ((i & 1) << 5);
1051 const int y32_idx = y64_idx + ((i >> 1) << 5);
1052 const int i2 = (m2 + i) << 2;
1053 force_split[5 + m2 + i] = PART_EVAL_ALL;
1054 avg_16x16[m][i] = 0;
1055 maxvar_16x16[m][i] = 0;
1056 minvar_16x16[m][i] = INT_MAX;
1057 for (int j = 0; j < 4; j++) {
1058 const int x16_idx = x32_idx + ((j & 1) << 4);
1059 const int y16_idx = y32_idx + ((j >> 1) << 4);
1060 const int split_index = 21 + i2 + j;
1061 VP16x16 *vst = &vt->split[m].split[i].split[j];
1062 force_split[split_index] = PART_EVAL_ALL;
1063 variance4x4downsample[i2 + j] = 0;
1064 if (!is_key_frame) {
1065 fill_variance_8x8avg(src, src_stride, dst, dst_stride, x16_idx,
1066 y16_idx, vst, is_cur_buf_hbd(xd), pixels_wide,
1067 pixels_high, is_key_frame);
1068
1069 fill_variance_tree(&vt->split[m].split[i].split[j], BLOCK_16X16);
1070 get_variance(&vt->split[m].split[i].split[j].part_variances.none);
1071 avg_16x16[m][i] +=
1072 vt->split[m].split[i].split[j].part_variances.none.variance;
1073 if (vt->split[m].split[i].split[j].part_variances.none.variance <
1074 minvar_16x16[m][i])
1075 minvar_16x16[m][i] =
1076 vt->split[m].split[i].split[j].part_variances.none.variance;
1077 if (vt->split[m].split[i].split[j].part_variances.none.variance >
1078 maxvar_16x16[m][i])
1079 maxvar_16x16[m][i] =
1080 vt->split[m].split[i].split[j].part_variances.none.variance;
1081 if (vt->split[m].split[i].split[j].part_variances.none.variance >
1082 thresholds[3]) {
1083 // 16X16 variance is above threshold for split, so force split to
1084 // 8x8 for this 16x16 block (this also forces splits for upper
1085 // levels).
1086 force_split[split_index] = PART_EVAL_ONLY_SPLIT;
1087 force_split[5 + m2 + i] = PART_EVAL_ONLY_SPLIT;
1088 force_split[m + 1] = PART_EVAL_ONLY_SPLIT;
1089 force_split[0] = PART_EVAL_ONLY_SPLIT;
1090 } else if (!cyclic_refresh_segment_id_boosted(segment_id) &&
1091 compute_minmax_variance &&
1092 vt->split[m]
1093 .split[i]
1094 .split[j]
1095 .part_variances.none.variance > thresholds[2]) {
1096 // We have some nominal amount of 16x16 variance (based on average),
1097 // compute the minmax over the 8x8 sub-blocks, and if above
1098 // threshold, force split to 8x8 block for this 16x16 block.
1099 int minmax = compute_minmax_8x8(src, src_stride, dst, dst_stride,
1100 x16_idx, y16_idx,
1101 #if CONFIG_AV1_HIGHBITDEPTH
1102 xd->cur_buf->flags,
1103 #endif
1104 pixels_wide, pixels_high);
1105 int thresh_minmax = (int)cpi->vbp_info.threshold_minmax;
1106 if (minmax > thresh_minmax) {
1107 force_split[split_index] = PART_EVAL_ONLY_SPLIT;
1108 force_split[5 + m2 + i] = PART_EVAL_ONLY_SPLIT;
1109 force_split[m + 1] = PART_EVAL_ONLY_SPLIT;
1110 force_split[0] = PART_EVAL_ONLY_SPLIT;
1111 }
1112 }
1113 }
1114 if (is_key_frame) {
1115 force_split[split_index] = PART_EVAL_ALL;
1116 // Go down to 4x4 down-sampling for variance.
1117 variance4x4downsample[i2 + j] = 1;
1118 for (int k = 0; k < 4; k++) {
1119 int x8_idx = x16_idx + ((k & 1) << 3);
1120 int y8_idx = y16_idx + ((k >> 1) << 3);
1121 VP8x8 *vst2 = is_key_frame ? &vst->split[k] : &vt2[i2 + j].split[k];
1122 fill_variance_4x4avg(
1123 src, src_stride, dst, dst_stride, x8_idx, y8_idx, vst2,
1124 #if CONFIG_AV1_HIGHBITDEPTH
1125 xd->cur_buf->flags,
1126 #endif
1127 pixels_wide, pixels_high, is_key_frame, border_offset_4x4);
1128 }
1129 }
1130 }
1131 }
1132 }
1133 }
1134
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,int mi_row,int mi_col)1135 static void setup_planes(AV1_COMP *cpi, MACROBLOCK *x, unsigned int *y_sad,
1136 unsigned int *y_sad_g, unsigned int *y_sad_alt,
1137 unsigned int *y_sad_last,
1138 MV_REFERENCE_FRAME *ref_frame_partition, int mi_row,
1139 int mi_col) {
1140 AV1_COMMON *const cm = &cpi->common;
1141 MACROBLOCKD *xd = &x->e_mbd;
1142 const int num_planes = av1_num_planes(cm);
1143 const int is_small_sb = (cm->seq_params->sb_size == BLOCK_64X64);
1144 BLOCK_SIZE bsize = is_small_sb ? BLOCK_64X64 : BLOCK_128X128;
1145 MB_MODE_INFO *mi = xd->mi[0];
1146 const YV12_BUFFER_CONFIG *yv12 = get_ref_frame_yv12_buf(cm, LAST_FRAME);
1147 assert(yv12 != NULL);
1148 const YV12_BUFFER_CONFIG *yv12_g = NULL;
1149 const YV12_BUFFER_CONFIG *yv12_alt = NULL;
1150 // Check if LAST is a reference. For spatial layers always use it as
1151 // reference scaling (golden or altref being lower resolution) is not
1152 // handled/check here.
1153 int use_last_ref = (cpi->ref_frame_flags & AOM_LAST_FLAG) ||
1154 cpi->svc.number_spatial_layers > 1;
1155 int use_golden_ref = cpi->ref_frame_flags & AOM_GOLD_FLAG;
1156 int use_alt_ref = cpi->ppi->rtc_ref.set_ref_frame_config ||
1157 cpi->sf.rt_sf.use_nonrd_altref_frame;
1158
1159 // For 1 spatial layer: GOLDEN is another temporal reference.
1160 // Check if it should be used as reference for partitioning.
1161 if (cpi->svc.number_spatial_layers == 1 && use_golden_ref &&
1162 (x->content_state_sb.source_sad_nonrd != kZeroSad || !use_last_ref)) {
1163 yv12_g = get_ref_frame_yv12_buf(cm, GOLDEN_FRAME);
1164 if (yv12_g && yv12_g != yv12) {
1165 av1_setup_pre_planes(xd, 0, yv12_g, mi_row, mi_col,
1166 get_ref_scale_factors(cm, GOLDEN_FRAME), num_planes);
1167 *y_sad_g = cpi->ppi->fn_ptr[bsize].sdf(
1168 x->plane[0].src.buf, x->plane[0].src.stride, xd->plane[0].pre[0].buf,
1169 xd->plane[0].pre[0].stride);
1170 }
1171 }
1172
1173 // For 1 spatial layer: ALTREF is another temporal reference.
1174 // Check if it should be used as reference for partitioning.
1175 if (cpi->svc.number_spatial_layers == 1 && use_alt_ref &&
1176 (cpi->ref_frame_flags & AOM_ALT_FLAG) &&
1177 (x->content_state_sb.source_sad_nonrd != kZeroSad || !use_last_ref)) {
1178 yv12_alt = get_ref_frame_yv12_buf(cm, ALTREF_FRAME);
1179 if (yv12_alt && yv12_alt != yv12) {
1180 av1_setup_pre_planes(xd, 0, yv12_alt, mi_row, mi_col,
1181 get_ref_scale_factors(cm, ALTREF_FRAME), num_planes);
1182 *y_sad_alt = cpi->ppi->fn_ptr[bsize].sdf(
1183 x->plane[0].src.buf, x->plane[0].src.stride, xd->plane[0].pre[0].buf,
1184 xd->plane[0].pre[0].stride);
1185 }
1186 }
1187
1188 if (use_last_ref) {
1189 av1_setup_pre_planes(xd, 0, yv12, mi_row, mi_col,
1190 get_ref_scale_factors(cm, LAST_FRAME), num_planes);
1191 mi->ref_frame[0] = LAST_FRAME;
1192 mi->ref_frame[1] = NONE_FRAME;
1193 mi->bsize = cm->seq_params->sb_size;
1194 mi->mv[0].as_int = 0;
1195 mi->interp_filters = av1_broadcast_interp_filter(BILINEAR);
1196 if (cpi->sf.rt_sf.estimate_motion_for_var_based_partition) {
1197 if (xd->mb_to_right_edge >= 0 && xd->mb_to_bottom_edge >= 0) {
1198 const MV dummy_mv = { 0, 0 };
1199 *y_sad = av1_int_pro_motion_estimation(cpi, x, cm->seq_params->sb_size,
1200 mi_row, mi_col, &dummy_mv);
1201 }
1202 }
1203 if (*y_sad == UINT_MAX) {
1204 *y_sad = cpi->ppi->fn_ptr[bsize].sdf(
1205 x->plane[0].src.buf, x->plane[0].src.stride, xd->plane[0].pre[0].buf,
1206 xd->plane[0].pre[0].stride);
1207 }
1208 *y_sad_last = *y_sad;
1209 }
1210
1211 // Pick the ref frame for partitioning, use golden or altref frame only if
1212 // its lower sad, bias to LAST with factor 0.9.
1213 if (*y_sad_g < 0.9 * *y_sad && *y_sad_g < *y_sad_alt) {
1214 av1_setup_pre_planes(xd, 0, yv12_g, mi_row, mi_col,
1215 get_ref_scale_factors(cm, GOLDEN_FRAME), num_planes);
1216 mi->ref_frame[0] = GOLDEN_FRAME;
1217 mi->mv[0].as_int = 0;
1218 *y_sad = *y_sad_g;
1219 *ref_frame_partition = GOLDEN_FRAME;
1220 x->nonrd_prune_ref_frame_search = 0;
1221 } else if (*y_sad_alt < 0.9 * *y_sad && *y_sad_alt < *y_sad_g) {
1222 av1_setup_pre_planes(xd, 0, yv12_alt, mi_row, mi_col,
1223 get_ref_scale_factors(cm, ALTREF_FRAME), num_planes);
1224 mi->ref_frame[0] = ALTREF_FRAME;
1225 mi->mv[0].as_int = 0;
1226 *y_sad = *y_sad_alt;
1227 *ref_frame_partition = ALTREF_FRAME;
1228 x->nonrd_prune_ref_frame_search = 0;
1229 } else {
1230 *ref_frame_partition = LAST_FRAME;
1231 x->nonrd_prune_ref_frame_search =
1232 cpi->sf.rt_sf.nonrd_prune_ref_frame_search;
1233 }
1234
1235 // Only calculate the predictor for non-zero MV.
1236 if (mi->mv[0].as_int != 0) {
1237 set_ref_ptrs(cm, xd, mi->ref_frame[0], mi->ref_frame[1]);
1238 av1_enc_build_inter_predictor(cm, xd, mi_row, mi_col, NULL,
1239 cm->seq_params->sb_size, AOM_PLANE_Y,
1240 num_planes - 1);
1241 }
1242 }
1243
1244 // Decides whether to split or merge a 16x16 partition block in variance based
1245 // partitioning based on the 8x8 sub-block variances.
get_part_eval_based_on_sub_blk_var(VP16x16 * var_16x16_info,int64_t threshold16)1246 static AOM_INLINE PART_EVAL_STATUS get_part_eval_based_on_sub_blk_var(
1247 VP16x16 *var_16x16_info, int64_t threshold16) {
1248 int max_8x8_var = 0, min_8x8_var = INT_MAX;
1249 for (int k = 0; k < 4; k++) {
1250 get_variance(&var_16x16_info->split[k].part_variances.none);
1251 int this_8x8_var = var_16x16_info->split[k].part_variances.none.variance;
1252 max_8x8_var = AOMMAX(this_8x8_var, max_8x8_var);
1253 min_8x8_var = AOMMIN(this_8x8_var, min_8x8_var);
1254 }
1255 // If the difference between maximum and minimum sub-block variances is high,
1256 // then only evaluate PARTITION_SPLIT for the 16x16 block. Otherwise, evaluate
1257 // only PARTITION_NONE. The shift factor for threshold16 has been derived
1258 // empirically.
1259 return ((max_8x8_var - min_8x8_var) > (threshold16 << 2))
1260 ? PART_EVAL_ONLY_SPLIT
1261 : PART_EVAL_ONLY_NONE;
1262 }
1263
is_set_force_zeromv_skip_based_on_src_sad(int set_zeromv_skip_based_on_source_sad,SOURCE_SAD source_sad_nonrd)1264 static AOM_INLINE bool is_set_force_zeromv_skip_based_on_src_sad(
1265 int set_zeromv_skip_based_on_source_sad, SOURCE_SAD source_sad_nonrd) {
1266 if (set_zeromv_skip_based_on_source_sad == 0) return false;
1267
1268 if (set_zeromv_skip_based_on_source_sad >= 2)
1269 return source_sad_nonrd <= kVeryLowSad;
1270 else if (set_zeromv_skip_based_on_source_sad >= 1)
1271 return source_sad_nonrd == kZeroSad;
1272
1273 return false;
1274 }
1275
av1_choose_var_based_partitioning(AV1_COMP * cpi,const TileInfo * const tile,ThreadData * td,MACROBLOCK * x,int mi_row,int mi_col)1276 int av1_choose_var_based_partitioning(AV1_COMP *cpi, const TileInfo *const tile,
1277 ThreadData *td, MACROBLOCK *x, int mi_row,
1278 int mi_col) {
1279 #if CONFIG_COLLECT_COMPONENT_TIMING
1280 start_timing(cpi, choose_var_based_partitioning_time);
1281 #endif
1282 AV1_COMMON *const cm = &cpi->common;
1283 MACROBLOCKD *xd = &x->e_mbd;
1284 const int64_t *const vbp_thresholds = cpi->vbp_info.thresholds;
1285
1286 int i, j, k, m;
1287 VP128x128 *vt;
1288 VP16x16 *vt2 = NULL;
1289 PART_EVAL_STATUS force_split[85];
1290 int avg_64x64;
1291 int max_var_32x32[4];
1292 int min_var_32x32[4];
1293 int var_32x32;
1294 int var_64x64;
1295 int min_var_64x64 = INT_MAX;
1296 int max_var_64x64 = 0;
1297 int avg_16x16[4][4];
1298 int maxvar_16x16[4][4];
1299 int minvar_16x16[4][4];
1300 int64_t threshold_4x4avg;
1301 uint8_t *s;
1302 const uint8_t *d;
1303 int sp;
1304 int dp;
1305 unsigned int uv_sad[2];
1306 NOISE_LEVEL noise_level = kLow;
1307 int zero_motion = 1;
1308
1309 int is_key_frame =
1310 (frame_is_intra_only(cm) ||
1311 (cpi->ppi->use_svc &&
1312 cpi->svc.layer_context[cpi->svc.temporal_layer_id].is_key_frame));
1313
1314 assert(cm->seq_params->sb_size == BLOCK_64X64 ||
1315 cm->seq_params->sb_size == BLOCK_128X128);
1316 const int is_small_sb = (cm->seq_params->sb_size == BLOCK_64X64);
1317 const int num_64x64_blocks = is_small_sb ? 1 : 4;
1318
1319 unsigned int y_sad = UINT_MAX;
1320 unsigned int y_sad_g = UINT_MAX;
1321 unsigned int y_sad_alt = UINT_MAX;
1322 unsigned int y_sad_last = UINT_MAX;
1323 BLOCK_SIZE bsize = is_small_sb ? BLOCK_64X64 : BLOCK_128X128;
1324
1325 // Ref frame used in partitioning.
1326 MV_REFERENCE_FRAME ref_frame_partition = LAST_FRAME;
1327
1328 CHECK_MEM_ERROR(cm, vt, aom_malloc(sizeof(*vt)));
1329
1330 vt->split = td->vt64x64;
1331
1332 int64_t thresholds[5] = { vbp_thresholds[0], vbp_thresholds[1],
1333 vbp_thresholds[2], vbp_thresholds[3],
1334 vbp_thresholds[4] };
1335
1336 const int low_res = (cm->width <= 352 && cm->height <= 288);
1337 int variance4x4downsample[64];
1338 const int segment_id = xd->mi[0]->segment_id;
1339 uint64_t blk_sad = 0;
1340 if (cpi->src_sad_blk_64x64 != NULL && !cpi->ppi->use_svc) {
1341 const int sb_size_by_mb = (cm->seq_params->sb_size == BLOCK_128X128)
1342 ? (cm->seq_params->mib_size >> 1)
1343 : cm->seq_params->mib_size;
1344 const int sb_cols =
1345 (cm->mi_params.mi_cols + sb_size_by_mb - 1) / sb_size_by_mb;
1346 const int sbi_col = mi_col / sb_size_by_mb;
1347 const int sbi_row = mi_row / sb_size_by_mb;
1348 blk_sad = cpi->src_sad_blk_64x64[sbi_col + sbi_row * sb_cols];
1349 }
1350
1351 if (cpi->oxcf.q_cfg.aq_mode == CYCLIC_REFRESH_AQ && cm->seg.enabled &&
1352 cyclic_refresh_segment_id_boosted(segment_id)) {
1353 const int q =
1354 av1_get_qindex(&cm->seg, segment_id, cm->quant_params.base_qindex);
1355 set_vbp_thresholds(cpi, thresholds, q, x->content_state_sb.low_sumdiff,
1356 x->content_state_sb.source_sad_nonrd,
1357 x->content_state_sb.source_sad_rd, 1, blk_sad,
1358 x->content_state_sb.lighting_change);
1359 } else {
1360 set_vbp_thresholds(cpi, thresholds, cm->quant_params.base_qindex,
1361 x->content_state_sb.low_sumdiff,
1362 x->content_state_sb.source_sad_nonrd,
1363 x->content_state_sb.source_sad_rd, 0, blk_sad,
1364 x->content_state_sb.lighting_change);
1365 }
1366
1367 // For non keyframes, disable 4x4 average for low resolution when speed = 8
1368 threshold_4x4avg = INT64_MAX;
1369
1370 s = x->plane[0].src.buf;
1371 sp = x->plane[0].src.stride;
1372
1373 // Index for force_split: 0 for 64x64, 1-4 for 32x32 blocks,
1374 // 5-20 for the 16x16 blocks.
1375 force_split[0] = PART_EVAL_ALL;
1376 memset(x->part_search_info.variance_low, 0,
1377 sizeof(x->part_search_info.variance_low));
1378
1379 // Check if LAST frame is NULL or if the resolution of LAST is
1380 // different than the current frame resolution, and if so, treat this frame
1381 // as a key frame, for the purpose of the superblock partitioning.
1382 // LAST == NULL can happen in cases where enhancement spatial layers are
1383 // enabled dyanmically and the only reference is the spatial(GOLDEN).
1384 // TODO(marpan): Check se of scaled references for the different resoln.
1385 if (!frame_is_intra_only(cm)) {
1386 const YV12_BUFFER_CONFIG *const ref =
1387 get_ref_frame_yv12_buf(cm, LAST_FRAME);
1388 if (ref == NULL || ref->y_crop_height != cm->height ||
1389 ref->y_crop_width != cm->width) {
1390 is_key_frame = 1;
1391 }
1392 }
1393
1394 if (!is_key_frame) {
1395 setup_planes(cpi, x, &y_sad, &y_sad_g, &y_sad_alt, &y_sad_last,
1396 &ref_frame_partition, mi_row, mi_col);
1397
1398 MB_MODE_INFO *mi = xd->mi[0];
1399 // Use reference SB directly for zero mv.
1400 if (mi->mv[0].as_int != 0) {
1401 d = xd->plane[0].dst.buf;
1402 dp = xd->plane[0].dst.stride;
1403 zero_motion = 0;
1404 } else {
1405 d = xd->plane[0].pre[0].buf;
1406 dp = xd->plane[0].pre[0].stride;
1407 }
1408 } else {
1409 d = AV1_VAR_OFFS;
1410 dp = 0;
1411 }
1412
1413 uv_sad[0] = 0;
1414 uv_sad[1] = 0;
1415 chroma_check(cpi, x, bsize, y_sad_last, y_sad_g, is_key_frame, zero_motion,
1416 uv_sad);
1417
1418 x->force_zeromv_skip_for_sb = 0;
1419 const bool is_set_force_zeromv_skip =
1420 is_set_force_zeromv_skip_based_on_src_sad(
1421 cpi->sf.rt_sf.set_zeromv_skip_based_on_source_sad,
1422 x->content_state_sb.source_sad_nonrd);
1423
1424 // If the superblock is completely static (zero source sad) and
1425 // the y_sad (relative to LAST ref) is very small, take the sb_size partition
1426 // and exit, and force zeromv_last skip mode for nonrd_pickmode.
1427 // Only do this on the base segment (so the QP-boosted segment, if applied,
1428 // can still continue cleaning/ramping up the quality).
1429 // Condition on color uv_sad is also added.
1430 if (!is_key_frame && cpi->sf.rt_sf.part_early_exit_zeromv &&
1431 cpi->rc.frames_since_key > 30 && segment_id == CR_SEGMENT_ID_BASE &&
1432 is_set_force_zeromv_skip && ref_frame_partition == LAST_FRAME &&
1433 xd->mi[0]->mv[0].as_int == 0) {
1434 const int block_width = mi_size_wide[cm->seq_params->sb_size];
1435 const int block_height = mi_size_high[cm->seq_params->sb_size];
1436 const unsigned int thresh_exit_part_y =
1437 cpi->zeromv_skip_thresh_exit_part[bsize];
1438 const unsigned int thresh_exit_part_uv =
1439 CALC_CHROMA_THRESH_FOR_ZEROMV_SKIP(thresh_exit_part_y);
1440 if (mi_col + block_width <= tile->mi_col_end &&
1441 mi_row + block_height <= tile->mi_row_end &&
1442 y_sad < thresh_exit_part_y && uv_sad[0] < thresh_exit_part_uv &&
1443 uv_sad[1] < thresh_exit_part_uv) {
1444 set_block_size(cpi, mi_row, mi_col, bsize);
1445 x->force_zeromv_skip_for_sb = 1;
1446 if (vt2) aom_free(vt2);
1447 if (vt) aom_free(vt);
1448 return 0;
1449 } else if (x->content_state_sb.source_sad_nonrd == kZeroSad &&
1450 cpi->sf.rt_sf.part_early_exit_zeromv >= 2) {
1451 x->force_zeromv_skip_for_sb = 2;
1452 }
1453 }
1454
1455 if (cpi->noise_estimate.enabled)
1456 noise_level = av1_noise_estimate_extract_level(&cpi->noise_estimate);
1457
1458 if (low_res && threshold_4x4avg < INT64_MAX)
1459 CHECK_MEM_ERROR(cm, vt2, aom_malloc(sizeof(*vt2)));
1460 // Fill in the entire tree of 8x8 (or 4x4 under some conditions) variances
1461 // for splits.
1462 fill_variance_tree_leaves(cpi, x, vt, vt2, force_split, avg_16x16,
1463 maxvar_16x16, minvar_16x16, variance4x4downsample,
1464 thresholds, s, sp, d, dp);
1465
1466 avg_64x64 = 0;
1467 for (m = 0; m < num_64x64_blocks; ++m) {
1468 max_var_32x32[m] = 0;
1469 min_var_32x32[m] = INT_MAX;
1470 const int m2 = m << 2;
1471 for (i = 0; i < 4; i++) {
1472 const int i2 = (m2 + i) << 2;
1473 for (j = 0; j < 4; j++) {
1474 const int split_index = 21 + i2 + j;
1475 if (variance4x4downsample[i2 + j] == 1) {
1476 VP16x16 *vtemp =
1477 (!is_key_frame) ? &vt2[i2 + j] : &vt->split[m].split[i].split[j];
1478 for (k = 0; k < 4; k++)
1479 fill_variance_tree(&vtemp->split[k], BLOCK_8X8);
1480 fill_variance_tree(vtemp, BLOCK_16X16);
1481 // If variance of this 16x16 block is above the threshold, force block
1482 // to split. This also forces a split on the upper levels.
1483 get_variance(&vtemp->part_variances.none);
1484 if (vtemp->part_variances.none.variance > thresholds[3]) {
1485 force_split[split_index] =
1486 cpi->sf.rt_sf.vbp_prune_16x16_split_using_min_max_sub_blk_var
1487 ? get_part_eval_based_on_sub_blk_var(vtemp, thresholds[3])
1488 : PART_EVAL_ONLY_SPLIT;
1489 force_split[5 + m2 + i] = PART_EVAL_ONLY_SPLIT;
1490 force_split[m + 1] = PART_EVAL_ONLY_SPLIT;
1491 force_split[0] = PART_EVAL_ONLY_SPLIT;
1492 }
1493 }
1494 }
1495 fill_variance_tree(&vt->split[m].split[i], BLOCK_32X32);
1496 // If variance of this 32x32 block is above the threshold, or if its above
1497 // (some threshold of) the average variance over the sub-16x16 blocks,
1498 // then force this block to split. This also forces a split on the upper
1499 // (64x64) level.
1500 uint64_t frame_sad_thresh = 20000;
1501 if (cpi->svc.number_temporal_layers > 2 &&
1502 cpi->svc.temporal_layer_id == 0)
1503 frame_sad_thresh = frame_sad_thresh << 1;
1504 if (force_split[5 + m2 + i] == PART_EVAL_ALL) {
1505 get_variance(&vt->split[m].split[i].part_variances.none);
1506 var_32x32 = vt->split[m].split[i].part_variances.none.variance;
1507 max_var_32x32[m] = AOMMAX(var_32x32, max_var_32x32[m]);
1508 min_var_32x32[m] = AOMMIN(var_32x32, min_var_32x32[m]);
1509 if (vt->split[m].split[i].part_variances.none.variance >
1510 thresholds[2] ||
1511 (!is_key_frame &&
1512 vt->split[m].split[i].part_variances.none.variance >
1513 (thresholds[2] >> 1) &&
1514 vt->split[m].split[i].part_variances.none.variance >
1515 (avg_16x16[m][i] >> 1))) {
1516 force_split[5 + m2 + i] = PART_EVAL_ONLY_SPLIT;
1517 force_split[m + 1] = PART_EVAL_ONLY_SPLIT;
1518 force_split[0] = PART_EVAL_ONLY_SPLIT;
1519 } else if (!is_key_frame && (cm->width * cm->height <= 640 * 360) &&
1520 (((maxvar_16x16[m][i] - minvar_16x16[m][i]) >
1521 (thresholds[2] >> 1) &&
1522 maxvar_16x16[m][i] > thresholds[2]) ||
1523 (cpi->sf.rt_sf.prefer_large_partition_blocks &&
1524 x->content_state_sb.source_sad_nonrd > kLowSad &&
1525 cpi->rc.frame_source_sad < frame_sad_thresh &&
1526 maxvar_16x16[m][i] > (thresholds[2] >> 4) &&
1527 maxvar_16x16[m][i] > (minvar_16x16[m][i] << 2)))) {
1528 force_split[5 + m2 + i] = PART_EVAL_ONLY_SPLIT;
1529 force_split[m + 1] = PART_EVAL_ONLY_SPLIT;
1530 force_split[0] = PART_EVAL_ONLY_SPLIT;
1531 }
1532 }
1533 }
1534 if (force_split[1 + m] == PART_EVAL_ALL) {
1535 fill_variance_tree(&vt->split[m], BLOCK_64X64);
1536 get_variance(&vt->split[m].part_variances.none);
1537 var_64x64 = vt->split[m].part_variances.none.variance;
1538 max_var_64x64 = AOMMAX(var_64x64, max_var_64x64);
1539 min_var_64x64 = AOMMIN(var_64x64, min_var_64x64);
1540 // If the difference of the max-min variances of sub-blocks or max
1541 // variance of a sub-block is above some threshold of then force this
1542 // block to split. Only checking this for noise level >= medium, if
1543 // encoder is in SVC or if we already forced large blocks.
1544
1545 if (!is_key_frame &&
1546 (max_var_32x32[m] - min_var_32x32[m]) > 3 * (thresholds[1] >> 3) &&
1547 max_var_32x32[m] > thresholds[1] >> 1 &&
1548 (noise_level >= kMedium || cpi->ppi->use_svc ||
1549 cpi->sf.rt_sf.prefer_large_partition_blocks)) {
1550 force_split[1 + m] = PART_EVAL_ONLY_SPLIT;
1551 force_split[0] = PART_EVAL_ONLY_SPLIT;
1552 }
1553 avg_64x64 += var_64x64;
1554 }
1555 if (is_small_sb) force_split[0] = PART_EVAL_ONLY_SPLIT;
1556 }
1557
1558 if (force_split[0] == PART_EVAL_ALL) {
1559 fill_variance_tree(vt, BLOCK_128X128);
1560 get_variance(&vt->part_variances.none);
1561 if (!is_key_frame &&
1562 vt->part_variances.none.variance > (9 * avg_64x64) >> 5)
1563 force_split[0] = PART_EVAL_ONLY_SPLIT;
1564
1565 if (!is_key_frame &&
1566 (max_var_64x64 - min_var_64x64) > 3 * (thresholds[0] >> 3) &&
1567 max_var_64x64 > thresholds[0] >> 1)
1568 force_split[0] = PART_EVAL_ONLY_SPLIT;
1569 }
1570
1571 if (mi_col + 32 > tile->mi_col_end || mi_row + 32 > tile->mi_row_end ||
1572 !set_vt_partitioning(cpi, xd, tile, vt, BLOCK_128X128, mi_row, mi_col,
1573 thresholds[0], BLOCK_16X16, force_split[0])) {
1574 for (m = 0; m < num_64x64_blocks; ++m) {
1575 const int x64_idx = ((m & 1) << 4);
1576 const int y64_idx = ((m >> 1) << 4);
1577 const int m2 = m << 2;
1578
1579 // Now go through the entire structure, splitting every block size until
1580 // we get to one that's got a variance lower than our threshold.
1581 if (!set_vt_partitioning(cpi, xd, tile, &vt->split[m], BLOCK_64X64,
1582 mi_row + y64_idx, mi_col + x64_idx,
1583 thresholds[1], BLOCK_16X16,
1584 force_split[1 + m])) {
1585 for (i = 0; i < 4; ++i) {
1586 const int x32_idx = ((i & 1) << 3);
1587 const int y32_idx = ((i >> 1) << 3);
1588 const int i2 = (m2 + i) << 2;
1589 if (!set_vt_partitioning(cpi, xd, tile, &vt->split[m].split[i],
1590 BLOCK_32X32, (mi_row + y64_idx + y32_idx),
1591 (mi_col + x64_idx + x32_idx), thresholds[2],
1592 BLOCK_16X16, force_split[5 + m2 + i])) {
1593 for (j = 0; j < 4; ++j) {
1594 const int x16_idx = ((j & 1) << 2);
1595 const int y16_idx = ((j >> 1) << 2);
1596 const int split_index = 21 + i2 + j;
1597 // For inter frames: if variance4x4downsample[] == 1 for this
1598 // 16x16 block, then the variance is based on 4x4 down-sampling,
1599 // so use vt2 in set_vt_partioning(), otherwise use vt.
1600 VP16x16 *vtemp =
1601 (!is_key_frame && variance4x4downsample[i2 + j] == 1)
1602 ? &vt2[i2 + j]
1603 : &vt->split[m].split[i].split[j];
1604 if (!set_vt_partitioning(cpi, xd, tile, vtemp, BLOCK_16X16,
1605 mi_row + y64_idx + y32_idx + y16_idx,
1606 mi_col + x64_idx + x32_idx + x16_idx,
1607 thresholds[3], BLOCK_8X8,
1608 force_split[split_index])) {
1609 for (k = 0; k < 4; ++k) {
1610 const int x8_idx = (k & 1) << 1;
1611 const int y8_idx = (k >> 1) << 1;
1612 set_block_size(
1613 cpi, (mi_row + y64_idx + y32_idx + y16_idx + y8_idx),
1614 (mi_col + x64_idx + x32_idx + x16_idx + x8_idx),
1615 BLOCK_8X8);
1616 }
1617 }
1618 }
1619 }
1620 }
1621 }
1622 }
1623 }
1624
1625 if (cpi->sf.rt_sf.short_circuit_low_temp_var) {
1626 set_low_temp_var_flag(cpi, &x->part_search_info, xd, vt, thresholds,
1627 ref_frame_partition, mi_col, mi_row);
1628 }
1629
1630 if (vt2) aom_free(vt2);
1631 if (vt) aom_free(vt);
1632 #if CONFIG_COLLECT_COMPONENT_TIMING
1633 end_timing(cpi, choose_var_based_partitioning_time);
1634 #endif
1635 return 0;
1636 }
1637