2 * Copyright (c) 2010 The WebM project authors. All Rights Reserved.
4 * Use of this source code is governed by a BSD-style license
5 * that can be found in the LICENSE file in the root of the source
6 * tree. An additional intellectual property rights grant can be found
7 * in the file PATENTS. All contributing project authors may
8 * be found in the AUTHORS file in the root of the source tree.
16 #include "./vp9_rtcd.h"
17 #include "./vpx_dsp_rtcd.h"
18 #include "./vpx_config.h"
20 #include "vpx_dsp/vpx_dsp_common.h"
21 #include "vpx_ports/mem.h"
22 #include "vpx_ports/vpx_timer.h"
23 #include "vpx_ports/system_state.h"
25 #include "vp9/common/vp9_common.h"
26 #include "vp9/common/vp9_entropy.h"
27 #include "vp9/common/vp9_entropymode.h"
28 #include "vp9/common/vp9_idct.h"
29 #include "vp9/common/vp9_mvref_common.h"
30 #include "vp9/common/vp9_pred_common.h"
31 #include "vp9/common/vp9_quant_common.h"
32 #include "vp9/common/vp9_reconintra.h"
33 #include "vp9/common/vp9_reconinter.h"
34 #include "vp9/common/vp9_seg_common.h"
35 #include "vp9/common/vp9_tile_common.h"
37 #include "vp9/encoder/vp9_aq_360.h"
38 #include "vp9/encoder/vp9_aq_complexity.h"
39 #include "vp9/encoder/vp9_aq_cyclicrefresh.h"
40 #include "vp9/encoder/vp9_aq_variance.h"
41 #include "vp9/encoder/vp9_encodeframe.h"
42 #include "vp9/encoder/vp9_encodemb.h"
43 #include "vp9/encoder/vp9_encodemv.h"
44 #include "vp9/encoder/vp9_ethread.h"
45 #include "vp9/encoder/vp9_extend.h"
46 #include "vp9/encoder/vp9_multi_thread.h"
47 #include "vp9/encoder/vp9_partition_models.h"
48 #include "vp9/encoder/vp9_pickmode.h"
49 #include "vp9/encoder/vp9_rd.h"
50 #include "vp9/encoder/vp9_rdopt.h"
51 #include "vp9/encoder/vp9_segmentation.h"
52 #include "vp9/encoder/vp9_tokenize.h"
54 static void encode_superblock(VP9_COMP *cpi, ThreadData *td, TOKENEXTRA **t,
55 int output_enabled, int mi_row, int mi_col,
56 BLOCK_SIZE bsize, PICK_MODE_CONTEXT *ctx);
58 // This is used as a reference when computing the source variance for the
59 // purpose of activity masking.
60 // Eventually this should be replaced by custom no-reference routines,
61 // which will be faster.
62 static const uint8_t VP9_VAR_OFFS[64] = {
63 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128,
64 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128,
65 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128,
66 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128,
67 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128
70 #if CONFIG_VP9_HIGHBITDEPTH
71 static const uint16_t VP9_HIGH_VAR_OFFS_8[64] = {
72 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128,
73 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128,
74 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128,
75 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128,
76 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128
79 static const uint16_t VP9_HIGH_VAR_OFFS_10[64] = {
80 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4,
81 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4,
82 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4,
83 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4,
84 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4,
85 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4,
86 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4,
87 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4
90 static const uint16_t VP9_HIGH_VAR_OFFS_12[64] = {
91 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16,
92 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16,
93 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16,
94 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16,
95 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16,
96 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16,
97 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16,
98 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16,
99 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16,
102 #endif // CONFIG_VP9_HIGHBITDEPTH
104 unsigned int vp9_get_sby_variance(VP9_COMP *cpi, const struct buf_2d *ref,
107 const unsigned int var =
108 cpi->fn_ptr[bs].vf(ref->buf, ref->stride, VP9_VAR_OFFS, 0, &sse);
112 #if CONFIG_VP9_HIGHBITDEPTH
113 unsigned int vp9_high_get_sby_variance(VP9_COMP *cpi, const struct buf_2d *ref,
114 BLOCK_SIZE bs, int bd) {
115 unsigned int var, sse;
119 cpi->fn_ptr[bs].vf(ref->buf, ref->stride,
120 CONVERT_TO_BYTEPTR(VP9_HIGH_VAR_OFFS_10), 0, &sse);
124 cpi->fn_ptr[bs].vf(ref->buf, ref->stride,
125 CONVERT_TO_BYTEPTR(VP9_HIGH_VAR_OFFS_12), 0, &sse);
130 cpi->fn_ptr[bs].vf(ref->buf, ref->stride,
131 CONVERT_TO_BYTEPTR(VP9_HIGH_VAR_OFFS_8), 0, &sse);
136 #endif // CONFIG_VP9_HIGHBITDEPTH
138 unsigned int vp9_get_sby_perpixel_variance(VP9_COMP *cpi,
139 const struct buf_2d *ref,
141 return ROUND_POWER_OF_TWO(vp9_get_sby_variance(cpi, ref, bs),
142 num_pels_log2_lookup[bs]);
145 #if CONFIG_VP9_HIGHBITDEPTH
146 unsigned int vp9_high_get_sby_perpixel_variance(VP9_COMP *cpi,
147 const struct buf_2d *ref,
148 BLOCK_SIZE bs, int bd) {
149 return (unsigned int)ROUND64_POWER_OF_TWO(
150 (int64_t)vp9_high_get_sby_variance(cpi, ref, bs, bd),
151 num_pels_log2_lookup[bs]);
153 #endif // CONFIG_VP9_HIGHBITDEPTH
155 static unsigned int get_sby_perpixel_diff_variance(VP9_COMP *cpi,
156 const struct buf_2d *ref,
157 int mi_row, int mi_col,
159 unsigned int sse, var;
161 const YV12_BUFFER_CONFIG *last = get_ref_frame_buffer(cpi, LAST_FRAME);
163 assert(last != NULL);
165 &last->y_buffer[mi_row * MI_SIZE * last->y_stride + mi_col * MI_SIZE];
166 var = cpi->fn_ptr[bs].vf(ref->buf, ref->stride, last_y, last->y_stride, &sse);
167 return ROUND_POWER_OF_TWO(var, num_pels_log2_lookup[bs]);
170 static BLOCK_SIZE get_rd_var_based_fixed_partition(VP9_COMP *cpi, MACROBLOCK *x,
171 int mi_row, int mi_col) {
172 unsigned int var = get_sby_perpixel_diff_variance(
173 cpi, &x->plane[0].src, mi_row, mi_col, BLOCK_64X64);
184 static void set_segment_index(VP9_COMP *cpi, MACROBLOCK *const x, int mi_row,
185 int mi_col, BLOCK_SIZE bsize, int segment_index) {
186 VP9_COMMON *const cm = &cpi->common;
187 const struct segmentation *const seg = &cm->seg;
188 MACROBLOCKD *const xd = &x->e_mbd;
189 MODE_INFO *mi = xd->mi[0];
191 const AQ_MODE aq_mode = cpi->oxcf.aq_mode;
192 const uint8_t *const map =
193 seg->update_map ? cpi->segmentation_map : cm->last_frame_seg_map;
195 // Initialize the segmentation index as 0.
198 // Skip the rest if AQ mode is disabled.
199 if (!seg->enabled) return;
202 case CYCLIC_REFRESH_AQ:
203 mi->segment_id = get_segment_id(cm, map, bsize, mi_row, mi_col);
206 if (cm->frame_type == KEY_FRAME || cpi->refresh_alt_ref_frame ||
207 cpi->force_update_segmentation ||
208 (cpi->refresh_golden_frame && !cpi->rc.is_src_frame_alt_ref)) {
211 // Get sub block energy range
212 if (bsize >= BLOCK_32X32) {
213 vp9_get_sub_block_energy(cpi, x, mi_row, mi_col, bsize, &min_energy,
216 min_energy = bsize <= BLOCK_16X16 ? x->mb_energy
217 : vp9_block_energy(cpi, x, bsize);
219 mi->segment_id = vp9_vaq_segment_id(min_energy);
221 mi->segment_id = get_segment_id(cm, map, bsize, mi_row, mi_col);
225 mi->segment_id = get_segment_id(cm, map, bsize, mi_row, mi_col);
228 if (cm->frame_type == KEY_FRAME || cpi->force_update_segmentation)
229 mi->segment_id = vp9_360aq_segment_id(mi_row, cm->mi_rows);
231 mi->segment_id = get_segment_id(cm, map, bsize, mi_row, mi_col);
233 case PSNR_AQ: mi->segment_id = segment_index; break;
239 // Set segment index from ROI map if it's enabled.
240 if (cpi->roi.enabled)
241 mi->segment_id = get_segment_id(cm, map, bsize, mi_row, mi_col);
243 if (cpi->sf.enable_wiener_variance) mi->segment_id = x->segment_id;
245 vp9_init_plane_quantizers(cpi, x);
248 // Lighter version of set_offsets that only sets the mode info
250 static INLINE void set_mode_info_offsets(VP9_COMMON *const cm,
252 MACROBLOCKD *const xd, int mi_row,
254 const int idx_str = xd->mi_stride * mi_row + mi_col;
255 xd->mi = cm->mi_grid_visible + idx_str;
256 xd->mi[0] = cm->mi + idx_str;
259 x->mbmi_ext = x->mbmi_ext_base + (mi_row * cm->mi_cols + mi_col);
262 static void set_offsets(VP9_COMP *cpi, const TileInfo *const tile,
263 MACROBLOCK *const x, int mi_row, int mi_col,
265 VP9_COMMON *const cm = &cpi->common;
266 MACROBLOCKD *const xd = &x->e_mbd;
267 const int mi_width = num_8x8_blocks_wide_lookup[bsize];
268 const int mi_height = num_8x8_blocks_high_lookup[bsize];
269 MvLimits *const mv_limits = &x->mv_limits;
271 set_skip_context(xd, mi_row, mi_col);
273 set_mode_info_offsets(cm, x, xd, mi_row, mi_col);
275 // Set up destination pointers.
276 vp9_setup_dst_planes(xd->plane, get_frame_new_buffer(cm), mi_row, mi_col);
278 // Set up limit values for MV components.
279 // Mv beyond the range do not produce new/different prediction block.
280 mv_limits->row_min = -(((mi_row + mi_height) * MI_SIZE) + VP9_INTERP_EXTEND);
281 mv_limits->col_min = -(((mi_col + mi_width) * MI_SIZE) + VP9_INTERP_EXTEND);
282 mv_limits->row_max = (cm->mi_rows - mi_row) * MI_SIZE + VP9_INTERP_EXTEND;
283 mv_limits->col_max = (cm->mi_cols - mi_col) * MI_SIZE + VP9_INTERP_EXTEND;
285 // Set up distance of MB to edge of frame in 1/8th pel units.
286 assert(!(mi_col & (mi_width - 1)) && !(mi_row & (mi_height - 1)));
287 set_mi_row_col(xd, tile, mi_row, mi_height, mi_col, mi_width, cm->mi_rows,
290 // Set up source buffers.
291 vp9_setup_src_planes(x, cpi->Source, mi_row, mi_col);
294 x->rddiv = cpi->rd.RDDIV;
295 x->rdmult = cpi->rd.RDMULT;
297 // required by vp9_append_sub8x8_mvs_for_idx() and vp9_find_best_ref_mvs()
301 static void duplicate_mode_info_in_sb(VP9_COMMON *cm, MACROBLOCKD *xd,
302 int mi_row, int mi_col,
304 const int block_width =
305 VPXMIN(num_8x8_blocks_wide_lookup[bsize], cm->mi_cols - mi_col);
306 const int block_height =
307 VPXMIN(num_8x8_blocks_high_lookup[bsize], cm->mi_rows - mi_row);
308 const int mi_stride = xd->mi_stride;
309 MODE_INFO *const src_mi = xd->mi[0];
312 for (j = 0; j < block_height; ++j)
313 for (i = 0; i < block_width; ++i) xd->mi[j * mi_stride + i] = src_mi;
316 static void set_block_size(VP9_COMP *const cpi, MACROBLOCK *const x,
317 MACROBLOCKD *const xd, int mi_row, int mi_col,
319 if (cpi->common.mi_cols > mi_col && cpi->common.mi_rows > mi_row) {
320 set_mode_info_offsets(&cpi->common, x, xd, mi_row, mi_col);
321 xd->mi[0]->sb_type = bsize;
326 // This struct is used for computing variance in choose_partitioning(), where
327 // the max number of samples within a superblock is 16x16 (with 4x4 avg). Even
328 // in high bitdepth, uint32_t is enough for sum_square_error (2^12 * 2^12 * 16
330 uint32_t sum_square_error;
340 } partition_variance;
343 partition_variance part_variances;
348 partition_variance part_variances;
353 partition_variance part_variances;
358 partition_variance part_variances;
363 partition_variance part_variances;
368 partition_variance *part_variances;
378 static void tree_to_node(void *data, BLOCK_SIZE bsize, variance_node *node) {
380 node->part_variances = NULL;
383 v64x64 *vt = (v64x64 *)data;
384 node->part_variances = &vt->part_variances;
385 for (i = 0; i < 4; i++)
386 node->split[i] = &vt->split[i].part_variances.none;
390 v32x32 *vt = (v32x32 *)data;
391 node->part_variances = &vt->part_variances;
392 for (i = 0; i < 4; i++)
393 node->split[i] = &vt->split[i].part_variances.none;
397 v16x16 *vt = (v16x16 *)data;
398 node->part_variances = &vt->part_variances;
399 for (i = 0; i < 4; i++)
400 node->split[i] = &vt->split[i].part_variances.none;
404 v8x8 *vt = (v8x8 *)data;
405 node->part_variances = &vt->part_variances;
406 for (i = 0; i < 4; i++)
407 node->split[i] = &vt->split[i].part_variances.none;
411 v4x4 *vt = (v4x4 *)data;
412 assert(bsize == BLOCK_4X4);
413 node->part_variances = &vt->part_variances;
414 for (i = 0; i < 4; i++) node->split[i] = &vt->split[i];
420 // Set variance values given sum square error, sum error, count.
421 static void fill_variance(uint32_t s2, int32_t s, int c, var *v) {
422 v->sum_square_error = s2;
427 static void get_variance(var *v) {
429 (int)(256 * (v->sum_square_error -
430 (uint32_t)(((int64_t)v->sum_error * v->sum_error) >>
435 static void sum_2_variances(const var *a, const var *b, var *r) {
436 assert(a->log2_count == b->log2_count);
437 fill_variance(a->sum_square_error + b->sum_square_error,
438 a->sum_error + b->sum_error, a->log2_count + 1, r);
441 static void fill_variance_tree(void *data, BLOCK_SIZE bsize) {
443 memset(&node, 0, sizeof(node));
444 tree_to_node(data, bsize, &node);
445 sum_2_variances(node.split[0], node.split[1], &node.part_variances->horz[0]);
446 sum_2_variances(node.split[2], node.split[3], &node.part_variances->horz[1]);
447 sum_2_variances(node.split[0], node.split[2], &node.part_variances->vert[0]);
448 sum_2_variances(node.split[1], node.split[3], &node.part_variances->vert[1]);
449 sum_2_variances(&node.part_variances->vert[0], &node.part_variances->vert[1],
450 &node.part_variances->none);
453 static int set_vt_partitioning(VP9_COMP *cpi, MACROBLOCK *const x,
454 MACROBLOCKD *const xd, void *data,
455 BLOCK_SIZE bsize, int mi_row, int mi_col,
456 int64_t threshold, BLOCK_SIZE bsize_min,
458 VP9_COMMON *const cm = &cpi->common;
460 const int block_width = num_8x8_blocks_wide_lookup[bsize];
461 const int block_height = num_8x8_blocks_high_lookup[bsize];
463 assert(block_height == block_width);
464 tree_to_node(data, bsize, &vt);
466 if (force_split == 1) return 0;
468 // For bsize=bsize_min (16x16/8x8 for 8x8/4x4 downsampling), select if
469 // variance is below threshold, otherwise split will be selected.
470 // No check for vert/horiz split as too few samples for variance.
471 if (bsize == bsize_min) {
472 // Variance already computed to set the force_split.
473 if (frame_is_intra_only(cm)) get_variance(&vt.part_variances->none);
474 if (mi_col + block_width / 2 < cm->mi_cols &&
475 mi_row + block_height / 2 < cm->mi_rows &&
476 vt.part_variances->none.variance < threshold) {
477 set_block_size(cpi, x, xd, mi_row, mi_col, bsize);
481 } else if (bsize > bsize_min) {
482 // Variance already computed to set the force_split.
483 if (frame_is_intra_only(cm)) get_variance(&vt.part_variances->none);
484 // For key frame: take split for bsize above 32X32 or very high variance.
485 if (frame_is_intra_only(cm) &&
486 (bsize > BLOCK_32X32 ||
487 vt.part_variances->none.variance > (threshold << 4))) {
490 // If variance is low, take the bsize (no split).
491 if (mi_col + block_width / 2 < cm->mi_cols &&
492 mi_row + block_height / 2 < cm->mi_rows &&
493 vt.part_variances->none.variance < threshold) {
494 set_block_size(cpi, x, xd, mi_row, mi_col, bsize);
498 // Check vertical split.
499 if (mi_row + block_height / 2 < cm->mi_rows) {
500 BLOCK_SIZE subsize = get_subsize(bsize, PARTITION_VERT);
501 get_variance(&vt.part_variances->vert[0]);
502 get_variance(&vt.part_variances->vert[1]);
503 if (vt.part_variances->vert[0].variance < threshold &&
504 vt.part_variances->vert[1].variance < threshold &&
505 get_plane_block_size(subsize, &xd->plane[1]) < BLOCK_INVALID) {
506 set_block_size(cpi, x, xd, mi_row, mi_col, subsize);
507 set_block_size(cpi, x, xd, mi_row, mi_col + block_width / 2, subsize);
511 // Check horizontal split.
512 if (mi_col + block_width / 2 < cm->mi_cols) {
513 BLOCK_SIZE subsize = get_subsize(bsize, PARTITION_HORZ);
514 get_variance(&vt.part_variances->horz[0]);
515 get_variance(&vt.part_variances->horz[1]);
516 if (vt.part_variances->horz[0].variance < threshold &&
517 vt.part_variances->horz[1].variance < threshold &&
518 get_plane_block_size(subsize, &xd->plane[1]) < BLOCK_INVALID) {
519 set_block_size(cpi, x, xd, mi_row, mi_col, subsize);
520 set_block_size(cpi, x, xd, mi_row + block_height / 2, mi_col, subsize);
530 static int64_t scale_part_thresh_sumdiff(int64_t threshold_base, int speed,
531 int width, int height,
534 if (width <= 640 && height <= 480)
535 return (5 * threshold_base) >> 2;
536 else if ((content_state == kLowSadLowSumdiff) ||
537 (content_state == kHighSadLowSumdiff) ||
538 (content_state == kLowVarHighSumdiff))
539 return (5 * threshold_base) >> 2;
540 } else if (speed == 7) {
541 if ((content_state == kLowSadLowSumdiff) ||
542 (content_state == kHighSadLowSumdiff) ||
543 (content_state == kLowVarHighSumdiff)) {
544 return (5 * threshold_base) >> 2;
547 return threshold_base;
550 // Set the variance split thresholds for following the block sizes:
551 // 0 - threshold_64x64, 1 - threshold_32x32, 2 - threshold_16x16,
552 // 3 - vbp_threshold_8x8. vbp_threshold_8x8 (to split to 4x4 partition) is
553 // currently only used on key frame.
554 static void set_vbp_thresholds(VP9_COMP *cpi, int64_t thresholds[], int q,
556 VP9_COMMON *const cm = &cpi->common;
557 const int is_key_frame = frame_is_intra_only(cm);
558 const int threshold_multiplier =
559 is_key_frame ? 20 : cpi->sf.variance_part_thresh_mult;
560 int64_t threshold_base =
561 (int64_t)(threshold_multiplier * cpi->y_dequant[q][1]);
564 thresholds[0] = threshold_base;
565 thresholds[1] = threshold_base >> 2;
566 thresholds[2] = threshold_base >> 2;
567 thresholds[3] = threshold_base << 2;
569 // Increase base variance threshold based on estimated noise level.
570 if (cpi->noise_estimate.enabled && cm->width >= 640 && cm->height >= 480) {
571 NOISE_LEVEL noise_level =
572 vp9_noise_estimate_extract_level(&cpi->noise_estimate);
573 if (noise_level == kHigh)
574 threshold_base = 3 * threshold_base;
575 else if (noise_level == kMedium)
576 threshold_base = threshold_base << 1;
577 else if (noise_level < kLow)
578 threshold_base = (7 * threshold_base) >> 3;
580 #if CONFIG_VP9_TEMPORAL_DENOISING
581 if (cpi->oxcf.noise_sensitivity > 0 && denoise_svc(cpi) &&
582 cpi->oxcf.speed > 5 && cpi->denoiser.denoising_level >= kDenLow)
584 vp9_scale_part_thresh(threshold_base, cpi->denoiser.denoising_level,
585 content_state, cpi->svc.temporal_layer_id);
588 scale_part_thresh_sumdiff(threshold_base, cpi->oxcf.speed, cm->width,
589 cm->height, content_state);
591 // Increase base variance threshold based on content_state/sum_diff level.
592 threshold_base = scale_part_thresh_sumdiff(
593 threshold_base, cpi->oxcf.speed, cm->width, cm->height, content_state);
595 thresholds[0] = threshold_base;
596 thresholds[2] = threshold_base << cpi->oxcf.speed;
597 if (cm->width >= 1280 && cm->height >= 720 && cpi->oxcf.speed < 7)
598 thresholds[2] = thresholds[2] << 1;
599 if (cm->width <= 352 && cm->height <= 288) {
600 thresholds[0] = threshold_base >> 3;
601 thresholds[1] = threshold_base >> 1;
602 thresholds[2] = threshold_base << 3;
603 } else if (cm->width < 1280 && cm->height < 720) {
604 thresholds[1] = (5 * threshold_base) >> 2;
605 } else if (cm->width < 1920 && cm->height < 1080) {
606 thresholds[1] = threshold_base << 1;
608 thresholds[1] = (5 * threshold_base) >> 1;
610 if (cpi->sf.disable_16x16part_nonkey) thresholds[2] = INT64_MAX;
614 void vp9_set_variance_partition_thresholds(VP9_COMP *cpi, int q,
616 VP9_COMMON *const cm = &cpi->common;
617 SPEED_FEATURES *const sf = &cpi->sf;
618 const int is_key_frame = frame_is_intra_only(cm);
619 if (sf->partition_search_type != VAR_BASED_PARTITION &&
620 sf->partition_search_type != REFERENCE_PARTITION) {
623 set_vbp_thresholds(cpi, cpi->vbp_thresholds, q, content_state);
624 // The thresholds below are not changed locally.
626 cpi->vbp_threshold_sad = 0;
627 cpi->vbp_threshold_copy = 0;
628 cpi->vbp_bsize_min = BLOCK_8X8;
630 if (cm->width <= 352 && cm->height <= 288)
631 cpi->vbp_threshold_sad = 10;
633 cpi->vbp_threshold_sad = (cpi->y_dequant[q][1] << 1) > 1000
634 ? (cpi->y_dequant[q][1] << 1)
636 cpi->vbp_bsize_min = BLOCK_16X16;
637 if (cm->width <= 352 && cm->height <= 288)
638 cpi->vbp_threshold_copy = 4000;
639 else if (cm->width <= 640 && cm->height <= 360)
640 cpi->vbp_threshold_copy = 8000;
642 cpi->vbp_threshold_copy = (cpi->y_dequant[q][1] << 3) > 8000
643 ? (cpi->y_dequant[q][1] << 3)
645 if (cpi->rc.high_source_sad ||
646 (cpi->use_svc && cpi->svc.high_source_sad_superframe)) {
647 cpi->vbp_threshold_sad = 0;
648 cpi->vbp_threshold_copy = 0;
651 cpi->vbp_threshold_minmax = 15 + (q >> 3);
655 // Compute the minmax over the 8x8 subblocks.
656 static int compute_minmax_8x8(const uint8_t *s, int sp, const uint8_t *d,
657 int dp, int x16_idx, int y16_idx,
658 #if CONFIG_VP9_HIGHBITDEPTH
661 int pixels_wide, int pixels_high) {
664 int minmax_min = 255;
665 // Loop over the 4 8x8 subblocks.
666 for (k = 0; k < 4; k++) {
667 int x8_idx = x16_idx + ((k & 1) << 3);
668 int y8_idx = y16_idx + ((k >> 1) << 3);
671 if (x8_idx < pixels_wide && y8_idx < pixels_high) {
672 #if CONFIG_VP9_HIGHBITDEPTH
673 if (highbd_flag & YV12_FLAG_HIGHBITDEPTH) {
674 vpx_highbd_minmax_8x8(s + y8_idx * sp + x8_idx, sp,
675 d + y8_idx * dp + x8_idx, dp, &min, &max);
677 vpx_minmax_8x8(s + y8_idx * sp + x8_idx, sp, d + y8_idx * dp + x8_idx,
681 vpx_minmax_8x8(s + y8_idx * sp + x8_idx, sp, d + y8_idx * dp + x8_idx, dp,
684 if ((max - min) > minmax_max) minmax_max = (max - min);
685 if ((max - min) < minmax_min) minmax_min = (max - min);
688 return (minmax_max - minmax_min);
691 static void fill_variance_4x4avg(const uint8_t *s, int sp, const uint8_t *d,
692 int dp, int x8_idx, int y8_idx, v8x8 *vst,
693 #if CONFIG_VP9_HIGHBITDEPTH
696 int pixels_wide, int pixels_high,
699 for (k = 0; k < 4; k++) {
700 int x4_idx = x8_idx + ((k & 1) << 2);
701 int y4_idx = y8_idx + ((k >> 1) << 2);
702 unsigned int sse = 0;
704 if (x4_idx < pixels_wide && y4_idx < pixels_high) {
707 #if CONFIG_VP9_HIGHBITDEPTH
708 if (highbd_flag & YV12_FLAG_HIGHBITDEPTH) {
709 s_avg = vpx_highbd_avg_4x4(s + y4_idx * sp + x4_idx, sp);
711 d_avg = vpx_highbd_avg_4x4(d + y4_idx * dp + x4_idx, dp);
713 s_avg = vpx_avg_4x4(s + y4_idx * sp + x4_idx, sp);
714 if (!is_key_frame) d_avg = vpx_avg_4x4(d + y4_idx * dp + x4_idx, dp);
717 s_avg = vpx_avg_4x4(s + y4_idx * sp + x4_idx, sp);
718 if (!is_key_frame) d_avg = vpx_avg_4x4(d + y4_idx * dp + x4_idx, dp);
723 fill_variance(sse, sum, 0, &vst->split[k].part_variances.none);
727 static void fill_variance_8x8avg(const uint8_t *s, int sp, const uint8_t *d,
728 int dp, int x16_idx, int y16_idx, v16x16 *vst,
729 #if CONFIG_VP9_HIGHBITDEPTH
732 int pixels_wide, int pixels_high,
735 for (k = 0; k < 4; k++) {
736 int x8_idx = x16_idx + ((k & 1) << 3);
737 int y8_idx = y16_idx + ((k >> 1) << 3);
738 unsigned int sse = 0;
740 if (x8_idx < pixels_wide && y8_idx < pixels_high) {
743 #if CONFIG_VP9_HIGHBITDEPTH
744 if (highbd_flag & YV12_FLAG_HIGHBITDEPTH) {
745 s_avg = vpx_highbd_avg_8x8(s + y8_idx * sp + x8_idx, sp);
747 d_avg = vpx_highbd_avg_8x8(d + y8_idx * dp + x8_idx, dp);
749 s_avg = vpx_avg_8x8(s + y8_idx * sp + x8_idx, sp);
750 if (!is_key_frame) d_avg = vpx_avg_8x8(d + y8_idx * dp + x8_idx, dp);
753 s_avg = vpx_avg_8x8(s + y8_idx * sp + x8_idx, sp);
754 if (!is_key_frame) d_avg = vpx_avg_8x8(d + y8_idx * dp + x8_idx, dp);
759 fill_variance(sse, sum, 0, &vst->split[k].part_variances.none);
763 // Check if most of the superblock is skin content, and if so, force split to
764 // 32x32, and set x->sb_is_skin for use in mode selection.
765 static int skin_sb_split(VP9_COMP *cpi, MACROBLOCK *x, const int low_res,
766 int mi_row, int mi_col, int *force_split) {
767 VP9_COMMON *const cm = &cpi->common;
768 #if CONFIG_VP9_HIGHBITDEPTH
769 if (cm->use_highbitdepth) return 0;
771 // Avoid checking superblocks on/near boundary and avoid low resolutions.
772 // Note superblock may still pick 64X64 if y_sad is very small
773 // (i.e., y_sad < cpi->vbp_threshold_sad) below. For now leave this as is.
774 if (!low_res && (mi_col >= 8 && mi_col + 8 < cm->mi_cols && mi_row >= 8 &&
775 mi_row + 8 < cm->mi_rows)) {
776 int num_16x16_skin = 0;
777 int num_16x16_nonskin = 0;
778 uint8_t *ysignal = x->plane[0].src.buf;
779 uint8_t *usignal = x->plane[1].src.buf;
780 uint8_t *vsignal = x->plane[2].src.buf;
781 int sp = x->plane[0].src.stride;
782 int spuv = x->plane[1].src.stride;
783 const int block_index = mi_row * cm->mi_cols + mi_col;
784 const int bw = num_8x8_blocks_wide_lookup[BLOCK_64X64];
785 const int bh = num_8x8_blocks_high_lookup[BLOCK_64X64];
786 const int xmis = VPXMIN(cm->mi_cols - mi_col, bw);
787 const int ymis = VPXMIN(cm->mi_rows - mi_row, bh);
788 // Loop through the 16x16 sub-blocks.
790 for (i = 0; i < ymis; i += 2) {
791 for (j = 0; j < xmis; j += 2) {
792 int bl_index = block_index + i * cm->mi_cols + j;
793 int is_skin = cpi->skin_map[bl_index];
794 num_16x16_skin += is_skin;
795 num_16x16_nonskin += (1 - is_skin);
796 if (num_16x16_nonskin > 3) {
797 // Exit loop if at least 4 of the 16x16 blocks are not skin.
805 ysignal += (sp << 4) - 64;
806 usignal += (spuv << 3) - 32;
807 vsignal += (spuv << 3) - 32;
809 if (num_16x16_skin > 12) {
817 static void set_low_temp_var_flag(VP9_COMP *cpi, MACROBLOCK *x, MACROBLOCKD *xd,
818 v64x64 *vt, int64_t thresholds[],
819 MV_REFERENCE_FRAME ref_frame_partition,
820 int mi_col, int mi_row) {
822 VP9_COMMON *const cm = &cpi->common;
823 const int mv_thr = cm->width > 640 ? 8 : 4;
824 // Check temporal variance for bsize >= 16x16, if LAST_FRAME was selected and
825 // int_pro mv is small. If the temporal variance is small set the flag
826 // variance_low for the block. The variance threshold can be adjusted, the
827 // higher the more aggressive.
828 if (ref_frame_partition == LAST_FRAME &&
829 (cpi->sf.short_circuit_low_temp_var == 1 ||
830 (xd->mi[0]->mv[0].as_mv.col < mv_thr &&
831 xd->mi[0]->mv[0].as_mv.col > -mv_thr &&
832 xd->mi[0]->mv[0].as_mv.row < mv_thr &&
833 xd->mi[0]->mv[0].as_mv.row > -mv_thr))) {
834 if (xd->mi[0]->sb_type == BLOCK_64X64) {
835 if ((vt->part_variances).none.variance < (thresholds[0] >> 1))
836 x->variance_low[0] = 1;
837 } else if (xd->mi[0]->sb_type == BLOCK_64X32) {
838 for (i = 0; i < 2; i++) {
839 if (vt->part_variances.horz[i].variance < (thresholds[0] >> 2))
840 x->variance_low[i + 1] = 1;
842 } else if (xd->mi[0]->sb_type == BLOCK_32X64) {
843 for (i = 0; i < 2; i++) {
844 if (vt->part_variances.vert[i].variance < (thresholds[0] >> 2))
845 x->variance_low[i + 3] = 1;
848 for (i = 0; i < 4; i++) {
849 const int idx[4][2] = { { 0, 0 }, { 0, 4 }, { 4, 0 }, { 4, 4 } };
851 cm->mi_stride * (mi_row + idx[i][0]) + mi_col + idx[i][1];
852 MODE_INFO **this_mi = cm->mi_grid_visible + idx_str;
854 if (cm->mi_cols <= mi_col + idx[i][1] ||
855 cm->mi_rows <= mi_row + idx[i][0])
858 if ((*this_mi)->sb_type == BLOCK_32X32) {
859 int64_t threshold_32x32 = (cpi->sf.short_circuit_low_temp_var == 1 ||
860 cpi->sf.short_circuit_low_temp_var == 3)
861 ? ((5 * thresholds[1]) >> 3)
862 : (thresholds[1] >> 1);
863 if (vt->split[i].part_variances.none.variance < threshold_32x32)
864 x->variance_low[i + 5] = 1;
865 } else if (cpi->sf.short_circuit_low_temp_var >= 2) {
866 // For 32x16 and 16x32 blocks, the flag is set on each 16x16 block
868 if ((*this_mi)->sb_type == BLOCK_16X16 ||
869 (*this_mi)->sb_type == BLOCK_32X16 ||
870 (*this_mi)->sb_type == BLOCK_16X32) {
871 for (j = 0; j < 4; j++) {
872 if (vt->split[i].split[j].part_variances.none.variance <
873 (thresholds[2] >> 8))
874 x->variance_low[(i << 2) + j + 9] = 1;
883 static void copy_partitioning_helper(VP9_COMP *cpi, MACROBLOCK *x,
884 MACROBLOCKD *xd, BLOCK_SIZE bsize,
885 int mi_row, int mi_col) {
886 VP9_COMMON *const cm = &cpi->common;
887 BLOCK_SIZE *prev_part = cpi->prev_partition;
888 int start_pos = mi_row * cm->mi_stride + mi_col;
890 const int bsl = b_width_log2_lookup[bsize];
891 const int bs = (1 << bsl) >> 2;
893 PARTITION_TYPE partition;
895 if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols) return;
897 partition = partition_lookup[bsl][prev_part[start_pos]];
898 subsize = get_subsize(bsize, partition);
900 if (subsize < BLOCK_8X8) {
901 set_block_size(cpi, x, xd, mi_row, mi_col, bsize);
905 set_block_size(cpi, x, xd, mi_row, mi_col, bsize);
908 set_block_size(cpi, x, xd, mi_row, mi_col, subsize);
909 set_block_size(cpi, x, xd, mi_row + bs, mi_col, subsize);
912 set_block_size(cpi, x, xd, mi_row, mi_col, subsize);
913 set_block_size(cpi, x, xd, mi_row, mi_col + bs, subsize);
916 assert(partition == PARTITION_SPLIT);
917 copy_partitioning_helper(cpi, x, xd, subsize, mi_row, mi_col);
918 copy_partitioning_helper(cpi, x, xd, subsize, mi_row + bs, mi_col);
919 copy_partitioning_helper(cpi, x, xd, subsize, mi_row, mi_col + bs);
920 copy_partitioning_helper(cpi, x, xd, subsize, mi_row + bs, mi_col + bs);
926 static int copy_partitioning(VP9_COMP *cpi, MACROBLOCK *x, MACROBLOCKD *xd,
927 int mi_row, int mi_col, int segment_id,
929 int svc_copy_allowed = 1;
930 int frames_since_key_thresh = 1;
932 // For SVC, don't allow copy if base spatial layer is key frame, or if
933 // frame is not a temporal enhancement layer frame.
934 int layer = LAYER_IDS_TO_IDX(0, cpi->svc.temporal_layer_id,
935 cpi->svc.number_temporal_layers);
936 const LAYER_CONTEXT *lc = &cpi->svc.layer_context[layer];
937 if (lc->is_key_frame || !cpi->svc.non_reference_frame) svc_copy_allowed = 0;
938 frames_since_key_thresh = cpi->svc.number_spatial_layers << 1;
940 if (cpi->rc.frames_since_key > frames_since_key_thresh && svc_copy_allowed &&
941 !cpi->resize_pending && segment_id == CR_SEGMENT_ID_BASE &&
942 cpi->prev_segment_id[sb_offset] == CR_SEGMENT_ID_BASE &&
943 cpi->copied_frame_cnt[sb_offset] < cpi->max_copied_frame) {
944 if (cpi->prev_partition != NULL) {
945 copy_partitioning_helper(cpi, x, xd, BLOCK_64X64, mi_row, mi_col);
946 cpi->copied_frame_cnt[sb_offset] += 1;
947 memcpy(x->variance_low, &(cpi->prev_variance_low[sb_offset * 25]),
948 sizeof(x->variance_low));
956 static int scale_partitioning_svc(VP9_COMP *cpi, MACROBLOCK *x, MACROBLOCKD *xd,
957 BLOCK_SIZE bsize, int mi_row, int mi_col,
958 int mi_row_high, int mi_col_high) {
959 VP9_COMMON *const cm = &cpi->common;
960 SVC *const svc = &cpi->svc;
961 BLOCK_SIZE *prev_part = svc->prev_partition_svc;
962 // Variables with _high are for higher resolution.
964 int subsize_high = 0;
965 const int bsl_high = b_width_log2_lookup[bsize];
966 const int bs_high = (1 << bsl_high) >> 2;
967 const int has_rows = (mi_row_high + bs_high) < cm->mi_rows;
968 const int has_cols = (mi_col_high + bs_high) < cm->mi_cols;
970 const int row_boundary_block_scale_factor[BLOCK_SIZES] = { 13, 13, 13, 1, 0,
973 const int col_boundary_block_scale_factor[BLOCK_SIZES] = { 13, 13, 13, 2, 2,
977 BLOCK_SIZE bsize_low;
978 PARTITION_TYPE partition_high;
980 if (mi_row_high >= cm->mi_rows || mi_col_high >= cm->mi_cols) return 0;
981 if (mi_row >= svc->mi_rows[svc->spatial_layer_id - 1] ||
982 mi_col >= svc->mi_cols[svc->spatial_layer_id - 1])
985 // Find corresponding (mi_col/mi_row) block down-scaled by 2x2.
986 start_pos = mi_row * (svc->mi_stride[svc->spatial_layer_id - 1]) + mi_col;
987 bsize_low = prev_part[start_pos];
988 // The block size is too big for boundaries. Do variance based partitioning.
989 if ((!has_rows || !has_cols) && bsize_low > BLOCK_16X16) return 1;
991 // For reference frames: return 1 (do variance-based partitioning) if the
992 // superblock is not low source sad and lower-resoln bsize is below 32x32.
993 if (!cpi->svc.non_reference_frame && !x->skip_low_source_sad &&
994 bsize_low < BLOCK_32X32)
997 // Scale up block size by 2x2. Force 64x64 for size larger than 32x32.
998 if (bsize_low < BLOCK_32X32) {
999 bsize_high = bsize_low + 3;
1000 } else if (bsize_low >= BLOCK_32X32) {
1001 bsize_high = BLOCK_64X64;
1003 // Scale up blocks on boundary.
1004 if (!has_cols && has_rows) {
1005 bsize_high = bsize_low + row_boundary_block_scale_factor[bsize_low];
1006 } else if (has_cols && !has_rows) {
1007 bsize_high = bsize_low + col_boundary_block_scale_factor[bsize_low];
1008 } else if (!has_cols && !has_rows) {
1009 bsize_high = bsize_low;
1012 partition_high = partition_lookup[bsl_high][bsize_high];
1013 subsize_high = get_subsize(bsize, partition_high);
1015 if (subsize_high < BLOCK_8X8) {
1016 set_block_size(cpi, x, xd, mi_row_high, mi_col_high, bsize_high);
1018 const int bsl = b_width_log2_lookup[bsize];
1019 const int bs = (1 << bsl) >> 2;
1020 switch (partition_high) {
1021 case PARTITION_NONE:
1022 set_block_size(cpi, x, xd, mi_row_high, mi_col_high, bsize_high);
1024 case PARTITION_HORZ:
1025 set_block_size(cpi, x, xd, mi_row_high, mi_col_high, subsize_high);
1026 if (subsize_high < BLOCK_64X64)
1027 set_block_size(cpi, x, xd, mi_row_high + bs_high, mi_col_high,
1030 case PARTITION_VERT:
1031 set_block_size(cpi, x, xd, mi_row_high, mi_col_high, subsize_high);
1032 if (subsize_high < BLOCK_64X64)
1033 set_block_size(cpi, x, xd, mi_row_high, mi_col_high + bs_high,
1037 assert(partition_high == PARTITION_SPLIT);
1038 if (scale_partitioning_svc(cpi, x, xd, subsize_high, mi_row, mi_col,
1039 mi_row_high, mi_col_high))
1041 if (scale_partitioning_svc(cpi, x, xd, subsize_high, mi_row + (bs >> 1),
1042 mi_col, mi_row_high + bs_high, mi_col_high))
1044 if (scale_partitioning_svc(cpi, x, xd, subsize_high, mi_row,
1045 mi_col + (bs >> 1), mi_row_high,
1046 mi_col_high + bs_high))
1048 if (scale_partitioning_svc(cpi, x, xd, subsize_high, mi_row + (bs >> 1),
1049 mi_col + (bs >> 1), mi_row_high + bs_high,
1050 mi_col_high + bs_high))
1059 static void update_partition_svc(VP9_COMP *cpi, BLOCK_SIZE bsize, int mi_row,
1061 VP9_COMMON *const cm = &cpi->common;
1062 BLOCK_SIZE *prev_part = cpi->svc.prev_partition_svc;
1063 int start_pos = mi_row * cm->mi_stride + mi_col;
1064 const int bsl = b_width_log2_lookup[bsize];
1065 const int bs = (1 << bsl) >> 2;
1067 PARTITION_TYPE partition;
1068 const MODE_INFO *mi = NULL;
1071 if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols) return;
1073 mi = cm->mi_grid_visible[start_pos];
1074 partition = partition_lookup[bsl][mi->sb_type];
1075 subsize = get_subsize(bsize, partition);
1076 if (subsize < BLOCK_8X8) {
1077 prev_part[start_pos] = bsize;
1079 switch (partition) {
1080 case PARTITION_NONE:
1081 prev_part[start_pos] = bsize;
1082 if (bsize == BLOCK_64X64) {
1083 for (xx = 0; xx < 8; xx += 4)
1084 for (yy = 0; yy < 8; yy += 4) {
1085 if ((mi_row + xx < cm->mi_rows) && (mi_col + yy < cm->mi_cols))
1086 prev_part[start_pos + xx * cm->mi_stride + yy] = bsize;
1090 case PARTITION_HORZ:
1091 prev_part[start_pos] = subsize;
1092 if (mi_row + bs < cm->mi_rows)
1093 prev_part[start_pos + bs * cm->mi_stride] = subsize;
1095 case PARTITION_VERT:
1096 prev_part[start_pos] = subsize;
1097 if (mi_col + bs < cm->mi_cols) prev_part[start_pos + bs] = subsize;
1100 assert(partition == PARTITION_SPLIT);
1101 update_partition_svc(cpi, subsize, mi_row, mi_col);
1102 update_partition_svc(cpi, subsize, mi_row + bs, mi_col);
1103 update_partition_svc(cpi, subsize, mi_row, mi_col + bs);
1104 update_partition_svc(cpi, subsize, mi_row + bs, mi_col + bs);
1110 static void update_prev_partition_helper(VP9_COMP *cpi, BLOCK_SIZE bsize,
1111 int mi_row, int mi_col) {
1112 VP9_COMMON *const cm = &cpi->common;
1113 BLOCK_SIZE *prev_part = cpi->prev_partition;
1114 int start_pos = mi_row * cm->mi_stride + mi_col;
1115 const int bsl = b_width_log2_lookup[bsize];
1116 const int bs = (1 << bsl) >> 2;
1118 PARTITION_TYPE partition;
1119 const MODE_INFO *mi = NULL;
1121 if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols) return;
1123 mi = cm->mi_grid_visible[start_pos];
1124 partition = partition_lookup[bsl][mi->sb_type];
1125 subsize = get_subsize(bsize, partition);
1126 if (subsize < BLOCK_8X8) {
1127 prev_part[start_pos] = bsize;
1129 switch (partition) {
1130 case PARTITION_NONE: prev_part[start_pos] = bsize; break;
1131 case PARTITION_HORZ:
1132 prev_part[start_pos] = subsize;
1133 if (mi_row + bs < cm->mi_rows)
1134 prev_part[start_pos + bs * cm->mi_stride] = subsize;
1136 case PARTITION_VERT:
1137 prev_part[start_pos] = subsize;
1138 if (mi_col + bs < cm->mi_cols) prev_part[start_pos + bs] = subsize;
1141 assert(partition == PARTITION_SPLIT);
1142 update_prev_partition_helper(cpi, subsize, mi_row, mi_col);
1143 update_prev_partition_helper(cpi, subsize, mi_row + bs, mi_col);
1144 update_prev_partition_helper(cpi, subsize, mi_row, mi_col + bs);
1145 update_prev_partition_helper(cpi, subsize, mi_row + bs, mi_col + bs);
1151 static void update_prev_partition(VP9_COMP *cpi, MACROBLOCK *x, int segment_id,
1152 int mi_row, int mi_col, int sb_offset) {
1153 update_prev_partition_helper(cpi, BLOCK_64X64, mi_row, mi_col);
1154 cpi->prev_segment_id[sb_offset] = segment_id;
1155 memcpy(&(cpi->prev_variance_low[sb_offset * 25]), x->variance_low,
1156 sizeof(x->variance_low));
1157 // Reset the counter for copy partitioning
1158 cpi->copied_frame_cnt[sb_offset] = 0;
1161 static void chroma_check(VP9_COMP *cpi, MACROBLOCK *x, int bsize,
1162 unsigned int y_sad, int is_key_frame) {
1164 MACROBLOCKD *xd = &x->e_mbd;
1166 if (is_key_frame) return;
1168 // For speed >= 8, avoid the chroma check if y_sad is above threshold.
1169 if (cpi->oxcf.speed >= 8) {
1170 if (y_sad > cpi->vbp_thresholds[1] &&
1171 (!cpi->noise_estimate.enabled ||
1172 vp9_noise_estimate_extract_level(&cpi->noise_estimate) < kMedium))
1176 for (i = 1; i <= 2; ++i) {
1177 unsigned int uv_sad = UINT_MAX;
1178 struct macroblock_plane *p = &x->plane[i];
1179 struct macroblockd_plane *pd = &xd->plane[i];
1180 const BLOCK_SIZE bs = get_plane_block_size(bsize, pd);
1182 if (bs != BLOCK_INVALID)
1183 uv_sad = cpi->fn_ptr[bs].sdf(p->src.buf, p->src.stride, pd->dst.buf,
1186 // TODO(marpan): Investigate if we should lower this threshold if
1187 // superblock is detected as skin.
1188 x->color_sensitivity[i - 1] = uv_sad > (y_sad >> 2);
1192 static uint64_t avg_source_sad(VP9_COMP *cpi, MACROBLOCK *x, int shift,
1194 unsigned int tmp_sse;
1196 unsigned int tmp_variance;
1197 const BLOCK_SIZE bsize = BLOCK_64X64;
1198 uint8_t *src_y = cpi->Source->y_buffer;
1199 int src_ystride = cpi->Source->y_stride;
1200 uint8_t *last_src_y = cpi->Last_Source->y_buffer;
1201 int last_src_ystride = cpi->Last_Source->y_stride;
1202 uint64_t avg_source_sad_threshold = 10000;
1203 uint64_t avg_source_sad_threshold2 = 12000;
1204 #if CONFIG_VP9_HIGHBITDEPTH
1205 if (cpi->common.use_highbitdepth) return 0;
1208 last_src_y += shift;
1210 cpi->fn_ptr[bsize].sdf(src_y, src_ystride, last_src_y, last_src_ystride);
1211 tmp_variance = vpx_variance64x64(src_y, src_ystride, last_src_y,
1212 last_src_ystride, &tmp_sse);
1213 // Note: tmp_sse - tmp_variance = ((sum * sum) >> 12)
1214 if (tmp_sad < avg_source_sad_threshold)
1215 x->content_state_sb = ((tmp_sse - tmp_variance) < 25) ? kLowSadLowSumdiff
1216 : kLowSadHighSumdiff;
1218 x->content_state_sb = ((tmp_sse - tmp_variance) < 25) ? kHighSadLowSumdiff
1219 : kHighSadHighSumdiff;
1221 // Detect large lighting change.
1222 if (cpi->oxcf.content != VP9E_CONTENT_SCREEN &&
1223 cpi->oxcf.rc_mode == VPX_CBR && tmp_variance < (tmp_sse >> 3) &&
1224 (tmp_sse - tmp_variance) > 10000)
1225 x->content_state_sb = kLowVarHighSumdiff;
1226 else if (tmp_sad > (avg_source_sad_threshold << 1))
1227 x->content_state_sb = kVeryHighSad;
1229 if (cpi->content_state_sb_fd != NULL) {
1230 if (tmp_sad < avg_source_sad_threshold2) {
1231 // Cap the increment to 255.
1232 if (cpi->content_state_sb_fd[sb_offset] < 255)
1233 cpi->content_state_sb_fd[sb_offset]++;
1235 cpi->content_state_sb_fd[sb_offset] = 0;
1238 if (tmp_sad == 0) x->zero_temp_sad_source = 1;
1242 // This function chooses partitioning based on the variance between source and
1243 // reconstructed last, where variance is computed for down-sampled inputs.
1244 static int choose_partitioning(VP9_COMP *cpi, const TileInfo *const tile,
1245 MACROBLOCK *x, int mi_row, int mi_col) {
1246 VP9_COMMON *const cm = &cpi->common;
1247 MACROBLOCKD *xd = &x->e_mbd;
1251 int force_split[21];
1253 int max_var_32x32 = 0;
1254 int min_var_32x32 = INT_MAX;
1257 int maxvar_16x16[4];
1258 int minvar_16x16[4];
1259 int64_t threshold_4x4avg;
1260 NOISE_LEVEL noise_level = kLow;
1261 int content_state = 0;
1266 int compute_minmax_variance = 1;
1267 unsigned int y_sad = UINT_MAX;
1268 BLOCK_SIZE bsize = BLOCK_64X64;
1269 // Ref frame used in partitioning.
1270 MV_REFERENCE_FRAME ref_frame_partition = LAST_FRAME;
1271 int pixels_wide = 64, pixels_high = 64;
1272 int64_t thresholds[4] = { cpi->vbp_thresholds[0], cpi->vbp_thresholds[1],
1273 cpi->vbp_thresholds[2], cpi->vbp_thresholds[3] };
1274 int force_64_split = cpi->rc.high_source_sad ||
1275 (cpi->use_svc && cpi->svc.high_source_sad_superframe) ||
1276 (cpi->oxcf.content == VP9E_CONTENT_SCREEN &&
1277 cpi->compute_source_sad_onepass &&
1278 cpi->sf.use_source_sad && !x->zero_temp_sad_source);
1280 // For the variance computation under SVC mode, we treat the frame as key if
1281 // the reference (base layer frame) is key frame (i.e., is_key_frame == 1).
1283 (frame_is_intra_only(cm) ||
1284 (is_one_pass_cbr_svc(cpi) &&
1285 cpi->svc.layer_context[cpi->svc.temporal_layer_id].is_key_frame));
1286 // Always use 4x4 partition for key frame.
1287 const int use_4x4_partition = frame_is_intra_only(cm);
1288 const int low_res = (cm->width <= 352 && cm->height <= 288);
1289 int variance4x4downsample[16];
1291 int sb_offset = (cm->mi_stride >> 3) * (mi_row >> 3) + (mi_col >> 3);
1293 // For SVC: check if LAST frame is NULL or if the resolution of LAST is
1294 // different than the current frame resolution, and if so, treat this frame
1295 // as a key frame, for the purpose of the superblock partitioning.
1296 // LAST == NULL can happen in some cases where enhancement spatial layers are
1297 // enabled dyanmically in the stream and the only reference is the spatial
1298 // reference (GOLDEN).
1300 const YV12_BUFFER_CONFIG *const ref = get_ref_frame_buffer(cpi, LAST_FRAME);
1301 if (ref == NULL || ref->y_crop_height != cm->height ||
1302 ref->y_crop_width != cm->width)
1306 set_offsets(cpi, tile, x, mi_row, mi_col, BLOCK_64X64);
1307 set_segment_index(cpi, x, mi_row, mi_col, BLOCK_64X64, 0);
1308 segment_id = xd->mi[0]->segment_id;
1310 if (cpi->oxcf.speed >= 8 || (cpi->use_svc && cpi->svc.non_reference_frame))
1311 compute_minmax_variance = 0;
1313 memset(x->variance_low, 0, sizeof(x->variance_low));
1315 if (cpi->sf.use_source_sad && !is_key_frame) {
1316 int sb_offset2 = ((cm->mi_cols + 7) >> 3) * (mi_row >> 3) + (mi_col >> 3);
1317 content_state = x->content_state_sb;
1318 x->skip_low_source_sad = (content_state == kLowSadLowSumdiff ||
1319 content_state == kLowSadHighSumdiff)
1322 x->lowvar_highsumdiff = (content_state == kLowVarHighSumdiff) ? 1 : 0;
1323 if (cpi->content_state_sb_fd != NULL)
1324 x->last_sb_high_content = cpi->content_state_sb_fd[sb_offset2];
1326 // For SVC on top spatial layer: use/scale the partition from
1327 // the lower spatial resolution if svc_use_lowres_part is enabled.
1328 if (cpi->sf.svc_use_lowres_part &&
1329 cpi->svc.spatial_layer_id == cpi->svc.number_spatial_layers - 1 &&
1330 cpi->svc.prev_partition_svc != NULL && content_state != kVeryHighSad) {
1331 if (!scale_partitioning_svc(cpi, x, xd, BLOCK_64X64, mi_row >> 1,
1332 mi_col >> 1, mi_row, mi_col)) {
1333 if (cpi->sf.copy_partition_flag) {
1334 update_prev_partition(cpi, x, segment_id, mi_row, mi_col, sb_offset);
1339 // If source_sad is low copy the partition without computing the y_sad.
1340 if (x->skip_low_source_sad && cpi->sf.copy_partition_flag &&
1342 copy_partitioning(cpi, x, xd, mi_row, mi_col, segment_id, sb_offset)) {
1343 x->sb_use_mv_part = 1;
1344 if (cpi->sf.svc_use_lowres_part &&
1345 cpi->svc.spatial_layer_id == cpi->svc.number_spatial_layers - 2)
1346 update_partition_svc(cpi, BLOCK_64X64, mi_row, mi_col);
1351 if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ && cm->seg.enabled &&
1352 cyclic_refresh_segment_id_boosted(segment_id)) {
1353 int q = vp9_get_qindex(&cm->seg, segment_id, cm->base_qindex);
1354 set_vbp_thresholds(cpi, thresholds, q, content_state);
1356 set_vbp_thresholds(cpi, thresholds, cm->base_qindex, content_state);
1358 // Decrease 32x32 split threshold for screen on base layer, for scene
1359 // change/high motion frames.
1360 if (cpi->oxcf.content == VP9E_CONTENT_SCREEN &&
1361 cpi->svc.spatial_layer_id == 0 && force_64_split)
1362 thresholds[1] = 3 * thresholds[1] >> 2;
1364 // For non keyframes, disable 4x4 average for low resolution when speed = 8
1365 threshold_4x4avg = (cpi->oxcf.speed < 8) ? thresholds[1] << 1 : INT64_MAX;
1367 if (xd->mb_to_right_edge < 0) pixels_wide += (xd->mb_to_right_edge >> 3);
1368 if (xd->mb_to_bottom_edge < 0) pixels_high += (xd->mb_to_bottom_edge >> 3);
1370 s = x->plane[0].src.buf;
1371 sp = x->plane[0].src.stride;
1373 // Index for force_split: 0 for 64x64, 1-4 for 32x32 blocks,
1374 // 5-20 for the 16x16 blocks.
1375 force_split[0] = force_64_split;
1377 if (!is_key_frame) {
1378 // In the case of spatial/temporal scalable coding, the assumption here is
1379 // that the temporal reference frame will always be of type LAST_FRAME.
1380 // TODO(marpan): If that assumption is broken, we need to revisit this code.
1381 MODE_INFO *mi = xd->mi[0];
1382 YV12_BUFFER_CONFIG *yv12 = get_ref_frame_buffer(cpi, LAST_FRAME);
1384 const YV12_BUFFER_CONFIG *yv12_g = NULL;
1385 unsigned int y_sad_g, y_sad_thr, y_sad_last;
1386 bsize = BLOCK_32X32 + (mi_col + 4 < cm->mi_cols) * 2 +
1387 (mi_row + 4 < cm->mi_rows);
1389 assert(yv12 != NULL);
1391 if (!(is_one_pass_cbr_svc(cpi) && cpi->svc.spatial_layer_id) ||
1392 cpi->svc.use_gf_temporal_ref_current_layer) {
1393 // For now, GOLDEN will not be used for non-zero spatial layers, since
1394 // it may not be a temporal reference.
1395 yv12_g = get_ref_frame_buffer(cpi, GOLDEN_FRAME);
1398 // Only compute y_sad_g (sad for golden reference) for speed < 8.
1399 if (cpi->oxcf.speed < 8 && yv12_g && yv12_g != yv12 &&
1400 (cpi->ref_frame_flags & VP9_GOLD_FLAG)) {
1401 vp9_setup_pre_planes(xd, 0, yv12_g, mi_row, mi_col,
1402 &cm->frame_refs[GOLDEN_FRAME - 1].sf);
1403 y_sad_g = cpi->fn_ptr[bsize].sdf(
1404 x->plane[0].src.buf, x->plane[0].src.stride, xd->plane[0].pre[0].buf,
1405 xd->plane[0].pre[0].stride);
1410 if (cpi->oxcf.lag_in_frames > 0 && cpi->oxcf.rc_mode == VPX_VBR &&
1411 cpi->rc.is_src_frame_alt_ref) {
1412 yv12 = get_ref_frame_buffer(cpi, ALTREF_FRAME);
1413 vp9_setup_pre_planes(xd, 0, yv12, mi_row, mi_col,
1414 &cm->frame_refs[ALTREF_FRAME - 1].sf);
1415 mi->ref_frame[0] = ALTREF_FRAME;
1418 vp9_setup_pre_planes(xd, 0, yv12, mi_row, mi_col,
1419 &cm->frame_refs[LAST_FRAME - 1].sf);
1420 mi->ref_frame[0] = LAST_FRAME;
1422 mi->ref_frame[1] = NONE;
1423 mi->sb_type = BLOCK_64X64;
1424 mi->mv[0].as_int = 0;
1425 mi->interp_filter = BILINEAR;
1427 if (cpi->oxcf.speed >= 8 && !low_res &&
1428 x->content_state_sb != kVeryHighSad) {
1429 y_sad = cpi->fn_ptr[bsize].sdf(
1430 x->plane[0].src.buf, x->plane[0].src.stride, xd->plane[0].pre[0].buf,
1431 xd->plane[0].pre[0].stride);
1433 const MV dummy_mv = { 0, 0 };
1434 y_sad = vp9_int_pro_motion_estimation(cpi, x, bsize, mi_row, mi_col,
1436 x->sb_use_mv_part = 1;
1437 x->sb_mvcol_part = mi->mv[0].as_mv.col;
1438 x->sb_mvrow_part = mi->mv[0].as_mv.row;
1439 if (cpi->oxcf.content == VP9E_CONTENT_SCREEN &&
1440 cpi->svc.spatial_layer_id == cpi->svc.first_spatial_layer_to_encode &&
1441 cpi->svc.high_num_blocks_with_motion && !x->zero_temp_sad_source &&
1442 cm->width > 640 && cm->height > 480) {
1443 // Disable split below 16x16 block size when scroll motion (horz or
1444 // vert) is detected.
1445 // TODO(marpan/jianj): Improve this condition: issue is that search
1446 // range is hard-coded/limited in vp9_int_pro_motion_estimation() so
1447 // scroll motion may not be detected here.
1448 if (((abs(x->sb_mvrow_part) >= 48 && abs(x->sb_mvcol_part) <= 8) ||
1449 (abs(x->sb_mvcol_part) >= 48 && abs(x->sb_mvrow_part) <= 8)) &&
1451 compute_minmax_variance = 0;
1452 thresholds[2] = INT64_MAX;
1458 // Pick ref frame for partitioning, bias last frame when y_sad_g and y_sad
1459 // are close if short_circuit_low_temp_var is on.
1460 y_sad_thr = cpi->sf.short_circuit_low_temp_var ? (y_sad * 7) >> 3 : y_sad;
1461 if (y_sad_g < y_sad_thr) {
1462 vp9_setup_pre_planes(xd, 0, yv12_g, mi_row, mi_col,
1463 &cm->frame_refs[GOLDEN_FRAME - 1].sf);
1464 mi->ref_frame[0] = GOLDEN_FRAME;
1465 mi->mv[0].as_int = 0;
1467 ref_frame_partition = GOLDEN_FRAME;
1469 x->pred_mv[LAST_FRAME] = mi->mv[0].as_mv;
1470 ref_frame_partition = LAST_FRAME;
1473 set_ref_ptrs(cm, xd, mi->ref_frame[0], mi->ref_frame[1]);
1474 vp9_build_inter_predictors_sb(xd, mi_row, mi_col, BLOCK_64X64);
1476 if (cpi->use_skin_detection)
1478 skin_sb_split(cpi, x, low_res, mi_row, mi_col, force_split);
1480 d = xd->plane[0].dst.buf;
1481 dp = xd->plane[0].dst.stride;
1483 // If the y_sad is very small, take 64x64 as partition and exit.
1484 // Don't check on boosted segment for now, as 64x64 is suppressed there.
1485 if (segment_id == CR_SEGMENT_ID_BASE && y_sad < cpi->vbp_threshold_sad) {
1486 const int block_width = num_8x8_blocks_wide_lookup[BLOCK_64X64];
1487 const int block_height = num_8x8_blocks_high_lookup[BLOCK_64X64];
1488 if (mi_col + block_width / 2 < cm->mi_cols &&
1489 mi_row + block_height / 2 < cm->mi_rows) {
1490 set_block_size(cpi, x, xd, mi_row, mi_col, BLOCK_64X64);
1491 x->variance_low[0] = 1;
1492 chroma_check(cpi, x, bsize, y_sad, is_key_frame);
1493 if (cpi->sf.svc_use_lowres_part &&
1494 cpi->svc.spatial_layer_id == cpi->svc.number_spatial_layers - 2)
1495 update_partition_svc(cpi, BLOCK_64X64, mi_row, mi_col);
1496 if (cpi->sf.copy_partition_flag) {
1497 update_prev_partition(cpi, x, segment_id, mi_row, mi_col, sb_offset);
1503 // If the y_sad is small enough, copy the partition of the superblock in the
1504 // last frame to current frame only if the last frame is not a keyframe.
1505 // Stop the copy every cpi->max_copied_frame to refresh the partition.
1506 // TODO(jianj) : tune the threshold.
1507 if (cpi->sf.copy_partition_flag && y_sad_last < cpi->vbp_threshold_copy &&
1508 copy_partitioning(cpi, x, xd, mi_row, mi_col, segment_id, sb_offset)) {
1509 chroma_check(cpi, x, bsize, y_sad, is_key_frame);
1510 if (cpi->sf.svc_use_lowres_part &&
1511 cpi->svc.spatial_layer_id == cpi->svc.number_spatial_layers - 2)
1512 update_partition_svc(cpi, BLOCK_64X64, mi_row, mi_col);
1518 #if CONFIG_VP9_HIGHBITDEPTH
1519 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
1521 case 10: d = CONVERT_TO_BYTEPTR(VP9_HIGH_VAR_OFFS_10); break;
1522 case 12: d = CONVERT_TO_BYTEPTR(VP9_HIGH_VAR_OFFS_12); break;
1524 default: d = CONVERT_TO_BYTEPTR(VP9_HIGH_VAR_OFFS_8); break;
1527 #endif // CONFIG_VP9_HIGHBITDEPTH
1530 if (low_res && threshold_4x4avg < INT64_MAX)
1531 CHECK_MEM_ERROR(cm, vt2, vpx_calloc(16, sizeof(*vt2)));
1532 // Fill in the entire tree of 8x8 (or 4x4 under some conditions) variances
1534 for (i = 0; i < 4; i++) {
1535 const int x32_idx = ((i & 1) << 5);
1536 const int y32_idx = ((i >> 1) << 5);
1537 const int i2 = i << 2;
1538 force_split[i + 1] = 0;
1540 maxvar_16x16[i] = 0;
1541 minvar_16x16[i] = INT_MAX;
1542 for (j = 0; j < 4; j++) {
1543 const int x16_idx = x32_idx + ((j & 1) << 4);
1544 const int y16_idx = y32_idx + ((j >> 1) << 4);
1545 const int split_index = 5 + i2 + j;
1546 v16x16 *vst = &vt.split[i].split[j];
1547 force_split[split_index] = 0;
1548 variance4x4downsample[i2 + j] = 0;
1549 if (!is_key_frame) {
1550 fill_variance_8x8avg(s, sp, d, dp, x16_idx, y16_idx, vst,
1551 #if CONFIG_VP9_HIGHBITDEPTH
1554 pixels_wide, pixels_high, is_key_frame);
1555 fill_variance_tree(&vt.split[i].split[j], BLOCK_16X16);
1556 get_variance(&vt.split[i].split[j].part_variances.none);
1557 avg_16x16[i] += vt.split[i].split[j].part_variances.none.variance;
1558 if (vt.split[i].split[j].part_variances.none.variance < minvar_16x16[i])
1559 minvar_16x16[i] = vt.split[i].split[j].part_variances.none.variance;
1560 if (vt.split[i].split[j].part_variances.none.variance > maxvar_16x16[i])
1561 maxvar_16x16[i] = vt.split[i].split[j].part_variances.none.variance;
1562 if (vt.split[i].split[j].part_variances.none.variance > thresholds[2]) {
1563 // 16X16 variance is above threshold for split, so force split to 8x8
1564 // for this 16x16 block (this also forces splits for upper levels).
1565 force_split[split_index] = 1;
1566 force_split[i + 1] = 1;
1568 } else if (compute_minmax_variance &&
1569 vt.split[i].split[j].part_variances.none.variance >
1571 !cyclic_refresh_segment_id_boosted(segment_id)) {
1572 // We have some nominal amount of 16x16 variance (based on average),
1573 // compute the minmax over the 8x8 sub-blocks, and if above threshold,
1574 // force split to 8x8 block for this 16x16 block.
1575 int minmax = compute_minmax_8x8(s, sp, d, dp, x16_idx, y16_idx,
1576 #if CONFIG_VP9_HIGHBITDEPTH
1579 pixels_wide, pixels_high);
1580 int thresh_minmax = (int)cpi->vbp_threshold_minmax;
1581 if (x->content_state_sb == kVeryHighSad)
1582 thresh_minmax = thresh_minmax << 1;
1583 if (minmax > thresh_minmax) {
1584 force_split[split_index] = 1;
1585 force_split[i + 1] = 1;
1591 (low_res && vt.split[i].split[j].part_variances.none.variance >
1592 threshold_4x4avg)) {
1593 force_split[split_index] = 0;
1594 // Go down to 4x4 down-sampling for variance.
1595 variance4x4downsample[i2 + j] = 1;
1596 for (k = 0; k < 4; k++) {
1597 int x8_idx = x16_idx + ((k & 1) << 3);
1598 int y8_idx = y16_idx + ((k >> 1) << 3);
1599 v8x8 *vst2 = is_key_frame ? &vst->split[k] : &vt2[i2 + j].split[k];
1600 fill_variance_4x4avg(s, sp, d, dp, x8_idx, y8_idx, vst2,
1601 #if CONFIG_VP9_HIGHBITDEPTH
1604 pixels_wide, pixels_high, is_key_frame);
1609 if (cpi->noise_estimate.enabled)
1610 noise_level = vp9_noise_estimate_extract_level(&cpi->noise_estimate);
1611 // Fill the rest of the variance tree by summing split partition values.
1613 for (i = 0; i < 4; i++) {
1614 const int i2 = i << 2;
1615 for (j = 0; j < 4; j++) {
1616 if (variance4x4downsample[i2 + j] == 1) {
1617 v16x16 *vtemp = (!is_key_frame) ? &vt2[i2 + j] : &vt.split[i].split[j];
1618 for (m = 0; m < 4; m++) fill_variance_tree(&vtemp->split[m], BLOCK_8X8);
1619 fill_variance_tree(vtemp, BLOCK_16X16);
1620 // If variance of this 16x16 block is above the threshold, force block
1621 // to split. This also forces a split on the upper levels.
1622 get_variance(&vtemp->part_variances.none);
1623 if (vtemp->part_variances.none.variance > thresholds[2]) {
1624 force_split[5 + i2 + j] = 1;
1625 force_split[i + 1] = 1;
1630 fill_variance_tree(&vt.split[i], BLOCK_32X32);
1631 // If variance of this 32x32 block is above the threshold, or if its above
1632 // (some threshold of) the average variance over the sub-16x16 blocks, then
1633 // force this block to split. This also forces a split on the upper
1635 if (!force_split[i + 1]) {
1636 get_variance(&vt.split[i].part_variances.none);
1637 var_32x32 = vt.split[i].part_variances.none.variance;
1638 max_var_32x32 = VPXMAX(var_32x32, max_var_32x32);
1639 min_var_32x32 = VPXMIN(var_32x32, min_var_32x32);
1640 if (vt.split[i].part_variances.none.variance > thresholds[1] ||
1642 vt.split[i].part_variances.none.variance > (thresholds[1] >> 1) &&
1643 vt.split[i].part_variances.none.variance > (avg_16x16[i] >> 1))) {
1644 force_split[i + 1] = 1;
1646 } else if (!is_key_frame && noise_level < kLow && cm->height <= 360 &&
1647 (maxvar_16x16[i] - minvar_16x16[i]) > (thresholds[1] >> 1) &&
1648 maxvar_16x16[i] > thresholds[1]) {
1649 force_split[i + 1] = 1;
1652 avg_32x32 += var_32x32;
1655 if (!force_split[0]) {
1656 fill_variance_tree(&vt, BLOCK_64X64);
1657 get_variance(&vt.part_variances.none);
1658 // If variance of this 64x64 block is above (some threshold of) the average
1659 // variance over the sub-32x32 blocks, then force this block to split.
1660 // Only checking this for noise level >= medium for now.
1661 if (!is_key_frame && noise_level >= kMedium &&
1662 vt.part_variances.none.variance > (9 * avg_32x32) >> 5)
1664 // Else if the maximum 32x32 variance minus the miniumum 32x32 variance in
1665 // a 64x64 block is greater than threshold and the maximum 32x32 variance is
1666 // above a miniumum threshold, then force the split of a 64x64 block
1667 // Only check this for low noise.
1668 else if (!is_key_frame && noise_level < kMedium &&
1669 (max_var_32x32 - min_var_32x32) > 3 * (thresholds[0] >> 3) &&
1670 max_var_32x32 > thresholds[0] >> 1)
1674 // Now go through the entire structure, splitting every block size until
1675 // we get to one that's got a variance lower than our threshold.
1676 if (mi_col + 8 > cm->mi_cols || mi_row + 8 > cm->mi_rows ||
1677 !set_vt_partitioning(cpi, x, xd, &vt, BLOCK_64X64, mi_row, mi_col,
1678 thresholds[0], BLOCK_16X16, force_split[0])) {
1679 for (i = 0; i < 4; ++i) {
1680 const int x32_idx = ((i & 1) << 2);
1681 const int y32_idx = ((i >> 1) << 2);
1682 const int i2 = i << 2;
1683 if (!set_vt_partitioning(cpi, x, xd, &vt.split[i], BLOCK_32X32,
1684 (mi_row + y32_idx), (mi_col + x32_idx),
1685 thresholds[1], BLOCK_16X16,
1686 force_split[i + 1])) {
1687 for (j = 0; j < 4; ++j) {
1688 const int x16_idx = ((j & 1) << 1);
1689 const int y16_idx = ((j >> 1) << 1);
1690 // For inter frames: if variance4x4downsample[] == 1 for this 16x16
1691 // block, then the variance is based on 4x4 down-sampling, so use vt2
1692 // in set_vt_partioning(), otherwise use vt.
1693 v16x16 *vtemp = (!is_key_frame && variance4x4downsample[i2 + j] == 1)
1695 : &vt.split[i].split[j];
1696 if (!set_vt_partitioning(
1697 cpi, x, xd, vtemp, BLOCK_16X16, mi_row + y32_idx + y16_idx,
1698 mi_col + x32_idx + x16_idx, thresholds[2], cpi->vbp_bsize_min,
1699 force_split[5 + i2 + j])) {
1700 for (k = 0; k < 4; ++k) {
1701 const int x8_idx = (k & 1);
1702 const int y8_idx = (k >> 1);
1703 if (use_4x4_partition) {
1704 if (!set_vt_partitioning(cpi, x, xd, &vtemp->split[k],
1706 mi_row + y32_idx + y16_idx + y8_idx,
1707 mi_col + x32_idx + x16_idx + x8_idx,
1708 thresholds[3], BLOCK_8X8, 0)) {
1710 cpi, x, xd, (mi_row + y32_idx + y16_idx + y8_idx),
1711 (mi_col + x32_idx + x16_idx + x8_idx), BLOCK_4X4);
1715 cpi, x, xd, (mi_row + y32_idx + y16_idx + y8_idx),
1716 (mi_col + x32_idx + x16_idx + x8_idx), BLOCK_8X8);
1725 if (!frame_is_intra_only(cm) && cpi->sf.copy_partition_flag) {
1726 update_prev_partition(cpi, x, segment_id, mi_row, mi_col, sb_offset);
1729 if (!frame_is_intra_only(cm) && cpi->sf.svc_use_lowres_part &&
1730 cpi->svc.spatial_layer_id == cpi->svc.number_spatial_layers - 2)
1731 update_partition_svc(cpi, BLOCK_64X64, mi_row, mi_col);
1733 if (cpi->sf.short_circuit_low_temp_var) {
1734 set_low_temp_var_flag(cpi, x, xd, &vt, thresholds, ref_frame_partition,
1738 chroma_check(cpi, x, bsize, y_sad, is_key_frame);
1739 if (vt2) vpx_free(vt2);
1743 static void update_state(VP9_COMP *cpi, ThreadData *td, PICK_MODE_CONTEXT *ctx,
1744 int mi_row, int mi_col, BLOCK_SIZE bsize,
1745 int output_enabled) {
1747 VP9_COMMON *const cm = &cpi->common;
1748 RD_COUNTS *const rdc = &td->rd_counts;
1749 MACROBLOCK *const x = &td->mb;
1750 MACROBLOCKD *const xd = &x->e_mbd;
1751 struct macroblock_plane *const p = x->plane;
1752 struct macroblockd_plane *const pd = xd->plane;
1753 MODE_INFO *mi = &ctx->mic;
1754 MODE_INFO *const xdmi = xd->mi[0];
1755 MODE_INFO *mi_addr = xd->mi[0];
1756 const struct segmentation *const seg = &cm->seg;
1757 const int bw = num_8x8_blocks_wide_lookup[mi->sb_type];
1758 const int bh = num_8x8_blocks_high_lookup[mi->sb_type];
1759 const int x_mis = VPXMIN(bw, cm->mi_cols - mi_col);
1760 const int y_mis = VPXMIN(bh, cm->mi_rows - mi_row);
1761 MV_REF *const frame_mvs = cm->cur_frame->mvs + mi_row * cm->mi_cols + mi_col;
1764 const int mis = cm->mi_stride;
1765 const int mi_width = num_8x8_blocks_wide_lookup[bsize];
1766 const int mi_height = num_8x8_blocks_high_lookup[bsize];
1769 assert(mi->sb_type == bsize);
1772 *x->mbmi_ext = ctx->mbmi_ext;
1774 // If segmentation in use
1776 // For in frame complexity AQ copy the segment id from the segment map.
1777 if (cpi->oxcf.aq_mode == COMPLEXITY_AQ) {
1778 const uint8_t *const map =
1779 seg->update_map ? cpi->segmentation_map : cm->last_frame_seg_map;
1780 mi_addr->segment_id = get_segment_id(cm, map, bsize, mi_row, mi_col);
1782 // Else for cyclic refresh mode update the segment map, set the segment id
1783 // and then update the quantizer.
1784 if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ) {
1785 vp9_cyclic_refresh_update_segment(cpi, xd->mi[0], mi_row, mi_col, bsize,
1786 ctx->rate, ctx->dist, x->skip, p);
1790 max_plane = is_inter_block(xdmi) ? MAX_MB_PLANE : 1;
1791 for (i = 0; i < max_plane; ++i) {
1792 p[i].coeff = ctx->coeff_pbuf[i][1];
1793 p[i].qcoeff = ctx->qcoeff_pbuf[i][1];
1794 pd[i].dqcoeff = ctx->dqcoeff_pbuf[i][1];
1795 p[i].eobs = ctx->eobs_pbuf[i][1];
1798 for (i = max_plane; i < MAX_MB_PLANE; ++i) {
1799 p[i].coeff = ctx->coeff_pbuf[i][2];
1800 p[i].qcoeff = ctx->qcoeff_pbuf[i][2];
1801 pd[i].dqcoeff = ctx->dqcoeff_pbuf[i][2];
1802 p[i].eobs = ctx->eobs_pbuf[i][2];
1805 // Restore the coding context of the MB to that that was in place
1806 // when the mode was picked for it
1807 for (y = 0; y < mi_height; y++)
1808 for (x_idx = 0; x_idx < mi_width; x_idx++)
1809 if ((xd->mb_to_right_edge >> (3 + MI_SIZE_LOG2)) + mi_width > x_idx &&
1810 (xd->mb_to_bottom_edge >> (3 + MI_SIZE_LOG2)) + mi_height > y) {
1811 xd->mi[x_idx + y * mis] = mi_addr;
1814 if (cpi->oxcf.aq_mode != NO_AQ) vp9_init_plane_quantizers(cpi, x);
1816 if (is_inter_block(xdmi) && xdmi->sb_type < BLOCK_8X8) {
1817 xdmi->mv[0].as_int = mi->bmi[3].as_mv[0].as_int;
1818 xdmi->mv[1].as_int = mi->bmi[3].as_mv[1].as_int;
1821 x->skip = ctx->skip;
1822 memcpy(x->zcoeff_blk[xdmi->tx_size], ctx->zcoeff_blk,
1823 sizeof(ctx->zcoeff_blk[0]) * ctx->num_4x4_blk);
1825 if (!output_enabled) return;
1827 #if CONFIG_INTERNAL_STATS
1828 if (frame_is_intra_only(cm)) {
1829 static const int kf_mode_index[] = {
1830 THR_DC /*DC_PRED*/, THR_V_PRED /*V_PRED*/,
1831 THR_H_PRED /*H_PRED*/, THR_D45_PRED /*D45_PRED*/,
1832 THR_D135_PRED /*D135_PRED*/, THR_D117_PRED /*D117_PRED*/,
1833 THR_D153_PRED /*D153_PRED*/, THR_D207_PRED /*D207_PRED*/,
1834 THR_D63_PRED /*D63_PRED*/, THR_TM /*TM_PRED*/,
1836 ++cpi->mode_chosen_counts[kf_mode_index[xdmi->mode]];
1838 // Note how often each mode chosen as best
1839 ++cpi->mode_chosen_counts[ctx->best_mode_index];
1842 if (!frame_is_intra_only(cm)) {
1843 if (is_inter_block(xdmi)) {
1844 vp9_update_mv_count(td);
1846 if (cm->interp_filter == SWITCHABLE) {
1847 const int ctx = get_pred_context_switchable_interp(xd);
1848 ++td->counts->switchable_interp[ctx][xdmi->interp_filter];
1852 rdc->comp_pred_diff[SINGLE_REFERENCE] += ctx->single_pred_diff;
1853 rdc->comp_pred_diff[COMPOUND_REFERENCE] += ctx->comp_pred_diff;
1854 rdc->comp_pred_diff[REFERENCE_MODE_SELECT] += ctx->hybrid_pred_diff;
1856 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i)
1857 rdc->filter_diff[i] += ctx->best_filter_diff[i];
1860 for (h = 0; h < y_mis; ++h) {
1861 MV_REF *const frame_mv = frame_mvs + h * cm->mi_cols;
1862 for (w = 0; w < x_mis; ++w) {
1863 MV_REF *const mv = frame_mv + w;
1864 mv->ref_frame[0] = mi->ref_frame[0];
1865 mv->ref_frame[1] = mi->ref_frame[1];
1866 mv->mv[0].as_int = mi->mv[0].as_int;
1867 mv->mv[1].as_int = mi->mv[1].as_int;
1872 void vp9_setup_src_planes(MACROBLOCK *x, const YV12_BUFFER_CONFIG *src,
1873 int mi_row, int mi_col) {
1874 uint8_t *const buffers[3] = { src->y_buffer, src->u_buffer, src->v_buffer };
1875 const int strides[3] = { src->y_stride, src->uv_stride, src->uv_stride };
1878 // Set current frame pointer.
1879 x->e_mbd.cur_buf = src;
1881 for (i = 0; i < MAX_MB_PLANE; i++)
1882 setup_pred_plane(&x->plane[i].src, buffers[i], strides[i], mi_row, mi_col,
1883 NULL, x->e_mbd.plane[i].subsampling_x,
1884 x->e_mbd.plane[i].subsampling_y);
1887 static void set_mode_info_seg_skip(MACROBLOCK *x, TX_MODE tx_mode,
1888 RD_COST *rd_cost, BLOCK_SIZE bsize) {
1889 MACROBLOCKD *const xd = &x->e_mbd;
1890 MODE_INFO *const mi = xd->mi[0];
1891 INTERP_FILTER filter_ref;
1893 filter_ref = get_pred_context_switchable_interp(xd);
1894 if (filter_ref == SWITCHABLE_FILTERS) filter_ref = EIGHTTAP;
1896 mi->sb_type = bsize;
1899 VPXMIN(max_txsize_lookup[bsize], tx_mode_to_biggest_tx_size[tx_mode]);
1901 mi->uv_mode = DC_PRED;
1902 mi->ref_frame[0] = LAST_FRAME;
1903 mi->ref_frame[1] = NONE;
1904 mi->mv[0].as_int = 0;
1905 mi->interp_filter = filter_ref;
1907 xd->mi[0]->bmi[0].as_mv[0].as_int = 0;
1910 vp9_rd_cost_init(rd_cost);
1913 static void set_segment_rdmult(VP9_COMP *const cpi, MACROBLOCK *const x,
1914 int mi_row, int mi_col, BLOCK_SIZE bsize,
1916 VP9_COMMON *const cm = &cpi->common;
1917 const uint8_t *const map =
1918 cm->seg.update_map ? cpi->segmentation_map : cm->last_frame_seg_map;
1920 vp9_init_plane_quantizers(cpi, x);
1921 vpx_clear_system_state();
1923 if (aq_mode == NO_AQ || aq_mode == PSNR_AQ) {
1924 if (cpi->sf.enable_tpl_model || cpi->sf.enable_wiener_variance)
1925 x->rdmult = x->cb_rdmult;
1929 if (aq_mode == CYCLIC_REFRESH_AQ) {
1930 // If segment is boosted, use rdmult for that segment.
1931 if (cyclic_refresh_segment_id_boosted(
1932 get_segment_id(cm, map, bsize, mi_row, mi_col)))
1933 x->rdmult = vp9_cyclic_refresh_get_rdmult(cpi->cyclic_refresh);
1937 x->rdmult = vp9_compute_rd_mult(cpi, cm->base_qindex + cm->y_dc_delta_q);
1939 if (cpi->sf.enable_wiener_variance && cm->show_frame) {
1940 if (cm->seg.enabled)
1941 x->rdmult = vp9_compute_rd_mult(
1942 cpi, vp9_get_qindex(&cm->seg, x->e_mbd.mi[0]->segment_id,
1947 static void rd_pick_sb_modes(VP9_COMP *cpi, TileDataEnc *tile_data,
1948 MACROBLOCK *const x, int mi_row, int mi_col,
1949 RD_COST *rd_cost, BLOCK_SIZE bsize,
1950 PICK_MODE_CONTEXT *ctx, int64_t best_rd) {
1951 VP9_COMMON *const cm = &cpi->common;
1952 TileInfo *const tile_info = &tile_data->tile_info;
1953 MACROBLOCKD *const xd = &x->e_mbd;
1955 struct macroblock_plane *const p = x->plane;
1956 struct macroblockd_plane *const pd = xd->plane;
1957 const AQ_MODE aq_mode = cpi->oxcf.aq_mode;
1960 vpx_clear_system_state();
1962 // Use the lower precision, but faster, 32x32 fdct for mode selection.
1963 x->use_lp32x32fdct = 1;
1965 set_offsets(cpi, tile_info, x, mi_row, mi_col, bsize);
1967 mi->sb_type = bsize;
1969 for (i = 0; i < MAX_MB_PLANE; ++i) {
1970 p[i].coeff = ctx->coeff_pbuf[i][0];
1971 p[i].qcoeff = ctx->qcoeff_pbuf[i][0];
1972 pd[i].dqcoeff = ctx->dqcoeff_pbuf[i][0];
1973 p[i].eobs = ctx->eobs_pbuf[i][0];
1977 ctx->pred_pixel_ready = 0;
1980 // Set to zero to make sure we do not use the previous encoded frame stats
1983 #if CONFIG_VP9_HIGHBITDEPTH
1984 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
1985 x->source_variance = vp9_high_get_sby_perpixel_variance(
1986 cpi, &x->plane[0].src, bsize, xd->bd);
1988 x->source_variance =
1989 vp9_get_sby_perpixel_variance(cpi, &x->plane[0].src, bsize);
1992 x->source_variance =
1993 vp9_get_sby_perpixel_variance(cpi, &x->plane[0].src, bsize);
1994 #endif // CONFIG_VP9_HIGHBITDEPTH
1996 // Save rdmult before it might be changed, so it can be restored later.
1997 orig_rdmult = x->rdmult;
1999 if ((cpi->sf.tx_domain_thresh > 0.0) || (cpi->sf.quant_opt_thresh > 0.0)) {
2000 double logvar = vp9_log_block_var(cpi, x, bsize);
2001 // Check block complexity as part of descision on using pixel or transform
2002 // domain distortion in rd tests.
2003 x->block_tx_domain = cpi->sf.allow_txfm_domain_distortion &&
2004 (logvar >= cpi->sf.tx_domain_thresh);
2006 // Check block complexity as part of descision on using quantized
2007 // coefficient optimisation inside the rd loop.
2008 x->block_qcoeff_opt =
2009 cpi->sf.allow_quant_coeff_opt && (logvar <= cpi->sf.quant_opt_thresh);
2011 x->block_tx_domain = cpi->sf.allow_txfm_domain_distortion;
2012 x->block_qcoeff_opt = cpi->sf.allow_quant_coeff_opt;
2015 set_segment_index(cpi, x, mi_row, mi_col, bsize, 0);
2016 set_segment_rdmult(cpi, x, mi_row, mi_col, bsize, aq_mode);
2018 // Find best coding mode & reconstruct the MB so it is available
2019 // as a predictor for MBs that follow in the SB
2020 if (frame_is_intra_only(cm)) {
2021 vp9_rd_pick_intra_mode_sb(cpi, x, rd_cost, bsize, ctx, best_rd);
2023 if (bsize >= BLOCK_8X8) {
2024 if (segfeature_active(&cm->seg, mi->segment_id, SEG_LVL_SKIP))
2025 vp9_rd_pick_inter_mode_sb_seg_skip(cpi, tile_data, x, rd_cost, bsize,
2028 vp9_rd_pick_inter_mode_sb(cpi, tile_data, x, mi_row, mi_col, rd_cost,
2029 bsize, ctx, best_rd);
2031 vp9_rd_pick_inter_mode_sub8x8(cpi, tile_data, x, mi_row, mi_col, rd_cost,
2032 bsize, ctx, best_rd);
2036 // Examine the resulting rate and for AQ mode 2 make a segment choice.
2037 if ((rd_cost->rate != INT_MAX) && (aq_mode == COMPLEXITY_AQ) &&
2038 (bsize >= BLOCK_16X16) &&
2039 (cm->frame_type == KEY_FRAME || cpi->refresh_alt_ref_frame ||
2040 (cpi->refresh_golden_frame && !cpi->rc.is_src_frame_alt_ref))) {
2041 vp9_caq_select_segment(cpi, x, bsize, mi_row, mi_col, rd_cost->rate);
2044 // TODO(jingning) The rate-distortion optimization flow needs to be
2045 // refactored to provide proper exit/return handle.
2046 if (rd_cost->rate == INT_MAX)
2047 rd_cost->rdcost = INT64_MAX;
2049 rd_cost->rdcost = RDCOST(x->rdmult, x->rddiv, rd_cost->rate, rd_cost->dist);
2051 x->rdmult = orig_rdmult;
2053 ctx->rate = rd_cost->rate;
2054 ctx->dist = rd_cost->dist;
2057 static void update_stats(VP9_COMMON *cm, ThreadData *td) {
2058 const MACROBLOCK *x = &td->mb;
2059 const MACROBLOCKD *const xd = &x->e_mbd;
2060 const MODE_INFO *const mi = xd->mi[0];
2061 const MB_MODE_INFO_EXT *const mbmi_ext = x->mbmi_ext;
2062 const BLOCK_SIZE bsize = mi->sb_type;
2064 if (!frame_is_intra_only(cm)) {
2065 FRAME_COUNTS *const counts = td->counts;
2066 const int inter_block = is_inter_block(mi);
2067 const int seg_ref_active =
2068 segfeature_active(&cm->seg, mi->segment_id, SEG_LVL_REF_FRAME);
2069 if (!seg_ref_active) {
2070 counts->intra_inter[get_intra_inter_context(xd)][inter_block]++;
2071 // If the segment reference feature is enabled we have only a single
2072 // reference frame allowed for the segment so exclude it from
2073 // the reference frame counts used to work out probabilities.
2075 const MV_REFERENCE_FRAME ref0 = mi->ref_frame[0];
2076 if (cm->reference_mode == REFERENCE_MODE_SELECT)
2077 counts->comp_inter[vp9_get_reference_mode_context(cm, xd)]
2078 [has_second_ref(mi)]++;
2080 if (has_second_ref(mi)) {
2081 const int idx = cm->ref_frame_sign_bias[cm->comp_fixed_ref];
2082 const int ctx = vp9_get_pred_context_comp_ref_p(cm, xd);
2083 const int bit = mi->ref_frame[!idx] == cm->comp_var_ref[1];
2084 counts->comp_ref[ctx][bit]++;
2086 counts->single_ref[vp9_get_pred_context_single_ref_p1(xd)][0]
2087 [ref0 != LAST_FRAME]++;
2088 if (ref0 != LAST_FRAME)
2089 counts->single_ref[vp9_get_pred_context_single_ref_p2(xd)][1]
2090 [ref0 != GOLDEN_FRAME]++;
2095 !segfeature_active(&cm->seg, mi->segment_id, SEG_LVL_SKIP)) {
2096 const int mode_ctx = mbmi_ext->mode_context[mi->ref_frame[0]];
2097 if (bsize >= BLOCK_8X8) {
2098 const PREDICTION_MODE mode = mi->mode;
2099 ++counts->inter_mode[mode_ctx][INTER_OFFSET(mode)];
2101 const int num_4x4_w = num_4x4_blocks_wide_lookup[bsize];
2102 const int num_4x4_h = num_4x4_blocks_high_lookup[bsize];
2104 for (idy = 0; idy < 2; idy += num_4x4_h) {
2105 for (idx = 0; idx < 2; idx += num_4x4_w) {
2106 const int j = idy * 2 + idx;
2107 const PREDICTION_MODE b_mode = mi->bmi[j].as_mode;
2108 ++counts->inter_mode[mode_ctx][INTER_OFFSET(b_mode)];
2116 static void restore_context(MACROBLOCK *const x, int mi_row, int mi_col,
2117 ENTROPY_CONTEXT a[16 * MAX_MB_PLANE],
2118 ENTROPY_CONTEXT l[16 * MAX_MB_PLANE],
2119 PARTITION_CONTEXT sa[8], PARTITION_CONTEXT sl[8],
2121 MACROBLOCKD *const xd = &x->e_mbd;
2123 const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
2124 const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
2125 int mi_width = num_8x8_blocks_wide_lookup[bsize];
2126 int mi_height = num_8x8_blocks_high_lookup[bsize];
2127 for (p = 0; p < MAX_MB_PLANE; p++) {
2128 memcpy(xd->above_context[p] + ((mi_col * 2) >> xd->plane[p].subsampling_x),
2129 a + num_4x4_blocks_wide * p,
2130 (sizeof(ENTROPY_CONTEXT) * num_4x4_blocks_wide) >>
2131 xd->plane[p].subsampling_x);
2132 memcpy(xd->left_context[p] +
2133 ((mi_row & MI_MASK) * 2 >> xd->plane[p].subsampling_y),
2134 l + num_4x4_blocks_high * p,
2135 (sizeof(ENTROPY_CONTEXT) * num_4x4_blocks_high) >>
2136 xd->plane[p].subsampling_y);
2138 memcpy(xd->above_seg_context + mi_col, sa,
2139 sizeof(*xd->above_seg_context) * mi_width);
2140 memcpy(xd->left_seg_context + (mi_row & MI_MASK), sl,
2141 sizeof(xd->left_seg_context[0]) * mi_height);
2144 static void save_context(MACROBLOCK *const x, int mi_row, int mi_col,
2145 ENTROPY_CONTEXT a[16 * MAX_MB_PLANE],
2146 ENTROPY_CONTEXT l[16 * MAX_MB_PLANE],
2147 PARTITION_CONTEXT sa[8], PARTITION_CONTEXT sl[8],
2149 const MACROBLOCKD *const xd = &x->e_mbd;
2151 const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
2152 const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
2153 int mi_width = num_8x8_blocks_wide_lookup[bsize];
2154 int mi_height = num_8x8_blocks_high_lookup[bsize];
2156 // buffer the above/left context information of the block in search.
2157 for (p = 0; p < MAX_MB_PLANE; ++p) {
2158 memcpy(a + num_4x4_blocks_wide * p,
2159 xd->above_context[p] + (mi_col * 2 >> xd->plane[p].subsampling_x),
2160 (sizeof(ENTROPY_CONTEXT) * num_4x4_blocks_wide) >>
2161 xd->plane[p].subsampling_x);
2162 memcpy(l + num_4x4_blocks_high * p,
2163 xd->left_context[p] +
2164 ((mi_row & MI_MASK) * 2 >> xd->plane[p].subsampling_y),
2165 (sizeof(ENTROPY_CONTEXT) * num_4x4_blocks_high) >>
2166 xd->plane[p].subsampling_y);
2168 memcpy(sa, xd->above_seg_context + mi_col,
2169 sizeof(*xd->above_seg_context) * mi_width);
2170 memcpy(sl, xd->left_seg_context + (mi_row & MI_MASK),
2171 sizeof(xd->left_seg_context[0]) * mi_height);
2174 static void encode_b(VP9_COMP *cpi, const TileInfo *const tile, ThreadData *td,
2175 TOKENEXTRA **tp, int mi_row, int mi_col,
2176 int output_enabled, BLOCK_SIZE bsize,
2177 PICK_MODE_CONTEXT *ctx) {
2178 MACROBLOCK *const x = &td->mb;
2179 set_offsets(cpi, tile, x, mi_row, mi_col, bsize);
2181 if ((cpi->sf.enable_tpl_model || cpi->sf.enable_wiener_variance) &&
2182 cpi->oxcf.aq_mode == NO_AQ)
2183 x->rdmult = x->cb_rdmult;
2185 update_state(cpi, td, ctx, mi_row, mi_col, bsize, output_enabled);
2186 encode_superblock(cpi, td, tp, output_enabled, mi_row, mi_col, bsize, ctx);
2188 if (output_enabled) {
2189 update_stats(&cpi->common, td);
2191 (*tp)->token = EOSB_TOKEN;
2196 static void encode_sb(VP9_COMP *cpi, ThreadData *td, const TileInfo *const tile,
2197 TOKENEXTRA **tp, int mi_row, int mi_col,
2198 int output_enabled, BLOCK_SIZE bsize, PC_TREE *pc_tree) {
2199 VP9_COMMON *const cm = &cpi->common;
2200 MACROBLOCK *const x = &td->mb;
2201 MACROBLOCKD *const xd = &x->e_mbd;
2203 const int bsl = b_width_log2_lookup[bsize], hbs = (1 << bsl) / 4;
2205 PARTITION_TYPE partition;
2206 BLOCK_SIZE subsize = bsize;
2208 if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols) return;
2210 if (bsize >= BLOCK_8X8) {
2211 ctx = partition_plane_context(xd, mi_row, mi_col, bsize);
2212 subsize = get_subsize(bsize, pc_tree->partitioning);
2215 subsize = BLOCK_4X4;
2218 partition = partition_lookup[bsl][subsize];
2219 if (output_enabled && bsize != BLOCK_4X4)
2220 td->counts->partition[ctx][partition]++;
2222 switch (partition) {
2223 case PARTITION_NONE:
2224 encode_b(cpi, tile, td, tp, mi_row, mi_col, output_enabled, subsize,
2227 case PARTITION_VERT:
2228 encode_b(cpi, tile, td, tp, mi_row, mi_col, output_enabled, subsize,
2229 &pc_tree->vertical[0]);
2230 if (mi_col + hbs < cm->mi_cols && bsize > BLOCK_8X8) {
2231 encode_b(cpi, tile, td, tp, mi_row, mi_col + hbs, output_enabled,
2232 subsize, &pc_tree->vertical[1]);
2235 case PARTITION_HORZ:
2236 encode_b(cpi, tile, td, tp, mi_row, mi_col, output_enabled, subsize,
2237 &pc_tree->horizontal[0]);
2238 if (mi_row + hbs < cm->mi_rows && bsize > BLOCK_8X8) {
2239 encode_b(cpi, tile, td, tp, mi_row + hbs, mi_col, output_enabled,
2240 subsize, &pc_tree->horizontal[1]);
2244 assert(partition == PARTITION_SPLIT);
2245 if (bsize == BLOCK_8X8) {
2246 encode_b(cpi, tile, td, tp, mi_row, mi_col, output_enabled, subsize,
2247 pc_tree->leaf_split[0]);
2249 encode_sb(cpi, td, tile, tp, mi_row, mi_col, output_enabled, subsize,
2251 encode_sb(cpi, td, tile, tp, mi_row, mi_col + hbs, output_enabled,
2252 subsize, pc_tree->split[1]);
2253 encode_sb(cpi, td, tile, tp, mi_row + hbs, mi_col, output_enabled,
2254 subsize, pc_tree->split[2]);
2255 encode_sb(cpi, td, tile, tp, mi_row + hbs, mi_col + hbs, output_enabled,
2256 subsize, pc_tree->split[3]);
2261 if (partition != PARTITION_SPLIT || bsize == BLOCK_8X8)
2262 update_partition_context(xd, mi_row, mi_col, subsize, bsize);
2265 // Check to see if the given partition size is allowed for a specified number
2266 // of 8x8 block rows and columns remaining in the image.
2267 // If not then return the largest allowed partition size
2268 static BLOCK_SIZE find_partition_size(BLOCK_SIZE bsize, int rows_left,
2269 int cols_left, int *bh, int *bw) {
2270 if (rows_left <= 0 || cols_left <= 0) {
2271 return VPXMIN(bsize, BLOCK_8X8);
2273 for (; bsize > 0; bsize -= 3) {
2274 *bh = num_8x8_blocks_high_lookup[bsize];
2275 *bw = num_8x8_blocks_wide_lookup[bsize];
2276 if ((*bh <= rows_left) && (*bw <= cols_left)) {
2284 static void set_partial_b64x64_partition(MODE_INFO *mi, int mis, int bh_in,
2285 int bw_in, int row8x8_remaining,
2286 int col8x8_remaining, BLOCK_SIZE bsize,
2287 MODE_INFO **mi_8x8) {
2290 for (r = 0; r < MI_BLOCK_SIZE; r += bh) {
2292 for (c = 0; c < MI_BLOCK_SIZE; c += bw) {
2293 const int index = r * mis + c;
2294 mi_8x8[index] = mi + index;
2295 mi_8x8[index]->sb_type = find_partition_size(
2296 bsize, row8x8_remaining - r, col8x8_remaining - c, &bh, &bw);
2301 // This function attempts to set all mode info entries in a given SB64
2302 // to the same block partition size.
2303 // However, at the bottom and right borders of the image the requested size
2304 // may not be allowed in which case this code attempts to choose the largest
2305 // allowable partition.
2306 static void set_fixed_partitioning(VP9_COMP *cpi, const TileInfo *const tile,
2307 MODE_INFO **mi_8x8, int mi_row, int mi_col,
2309 VP9_COMMON *const cm = &cpi->common;
2310 const int mis = cm->mi_stride;
2311 const int row8x8_remaining = tile->mi_row_end - mi_row;
2312 const int col8x8_remaining = tile->mi_col_end - mi_col;
2313 int block_row, block_col;
2314 MODE_INFO *mi_upper_left = cm->mi + mi_row * mis + mi_col;
2315 int bh = num_8x8_blocks_high_lookup[bsize];
2316 int bw = num_8x8_blocks_wide_lookup[bsize];
2318 assert((row8x8_remaining > 0) && (col8x8_remaining > 0));
2320 // Apply the requested partition size to the SB64 if it is all "in image"
2321 if ((col8x8_remaining >= MI_BLOCK_SIZE) &&
2322 (row8x8_remaining >= MI_BLOCK_SIZE)) {
2323 for (block_row = 0; block_row < MI_BLOCK_SIZE; block_row += bh) {
2324 for (block_col = 0; block_col < MI_BLOCK_SIZE; block_col += bw) {
2325 int index = block_row * mis + block_col;
2326 mi_8x8[index] = mi_upper_left + index;
2327 mi_8x8[index]->sb_type = bsize;
2331 // Else this is a partial SB64.
2332 set_partial_b64x64_partition(mi_upper_left, mis, bh, bw, row8x8_remaining,
2333 col8x8_remaining, bsize, mi_8x8);
2337 static const struct {
2340 } coord_lookup[16] = {
2363 static void set_source_var_based_partition(VP9_COMP *cpi,
2364 const TileInfo *const tile,
2365 MACROBLOCK *const x,
2366 MODE_INFO **mi_8x8, int mi_row,
2368 VP9_COMMON *const cm = &cpi->common;
2369 const int mis = cm->mi_stride;
2370 const int row8x8_remaining = tile->mi_row_end - mi_row;
2371 const int col8x8_remaining = tile->mi_col_end - mi_col;
2372 MODE_INFO *mi_upper_left = cm->mi + mi_row * mis + mi_col;
2374 vp9_setup_src_planes(x, cpi->Source, mi_row, mi_col);
2376 assert((row8x8_remaining > 0) && (col8x8_remaining > 0));
2379 if ((col8x8_remaining >= MI_BLOCK_SIZE) &&
2380 (row8x8_remaining >= MI_BLOCK_SIZE)) {
2384 const int offset = (mi_row >> 1) * cm->mb_cols + (mi_col >> 1);
2385 int is_larger_better = 0;
2387 unsigned int thr = cpi->source_var_thresh;
2389 memset(d32, 0, 4 * sizeof(diff));
2391 for (i = 0; i < 4; i++) {
2394 for (j = 0; j < 4; j++) {
2395 int b_mi_row = coord_lookup[i * 4 + j].row;
2396 int b_mi_col = coord_lookup[i * 4 + j].col;
2397 int boffset = b_mi_row / 2 * cm->mb_cols + b_mi_col / 2;
2399 d16[j] = cpi->source_diff_var + offset + boffset;
2401 index = b_mi_row * mis + b_mi_col;
2402 mi_8x8[index] = mi_upper_left + index;
2403 mi_8x8[index]->sb_type = BLOCK_16X16;
2405 // TODO(yunqingwang): If d16[j].var is very large, use 8x8 partition
2406 // size to further improve quality.
2409 is_larger_better = (d16[0]->var < thr) && (d16[1]->var < thr) &&
2410 (d16[2]->var < thr) && (d16[3]->var < thr);
2412 // Use 32x32 partition
2413 if (is_larger_better) {
2416 for (j = 0; j < 4; j++) {
2417 d32[i].sse += d16[j]->sse;
2418 d32[i].sum += d16[j]->sum;
2422 (unsigned int)(d32[i].sse -
2423 (unsigned int)(((int64_t)d32[i].sum * d32[i].sum) >>
2426 index = coord_lookup[i * 4].row * mis + coord_lookup[i * 4].col;
2427 mi_8x8[index] = mi_upper_left + index;
2428 mi_8x8[index]->sb_type = BLOCK_32X32;
2432 if (use32x32 == 4) {
2434 is_larger_better = (d32[0].var < thr) && (d32[1].var < thr) &&
2435 (d32[2].var < thr) && (d32[3].var < thr);
2437 // Use 64x64 partition
2438 if (is_larger_better) {
2439 mi_8x8[0] = mi_upper_left;
2440 mi_8x8[0]->sb_type = BLOCK_64X64;
2443 } else { // partial in-image SB64
2444 int bh = num_8x8_blocks_high_lookup[BLOCK_16X16];
2445 int bw = num_8x8_blocks_wide_lookup[BLOCK_16X16];
2446 set_partial_b64x64_partition(mi_upper_left, mis, bh, bw, row8x8_remaining,
2447 col8x8_remaining, BLOCK_16X16, mi_8x8);
2451 static void update_state_rt(VP9_COMP *cpi, ThreadData *td,
2452 PICK_MODE_CONTEXT *ctx, int mi_row, int mi_col,
2454 VP9_COMMON *const cm = &cpi->common;
2455 MACROBLOCK *const x = &td->mb;
2456 MACROBLOCKD *const xd = &x->e_mbd;
2457 MODE_INFO *const mi = xd->mi[0];
2458 struct macroblock_plane *const p = x->plane;
2459 const struct segmentation *const seg = &cm->seg;
2460 const int bw = num_8x8_blocks_wide_lookup[mi->sb_type];
2461 const int bh = num_8x8_blocks_high_lookup[mi->sb_type];
2462 const int x_mis = VPXMIN(bw, cm->mi_cols - mi_col);
2463 const int y_mis = VPXMIN(bh, cm->mi_rows - mi_row);
2465 *(xd->mi[0]) = ctx->mic;
2466 *(x->mbmi_ext) = ctx->mbmi_ext;
2468 if (seg->enabled && (cpi->oxcf.aq_mode != NO_AQ || cpi->roi.enabled)) {
2469 // Setting segmentation map for cyclic_refresh.
2470 if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ) {
2471 vp9_cyclic_refresh_update_segment(cpi, mi, mi_row, mi_col, bsize,
2472 ctx->rate, ctx->dist, x->skip, p);
2474 const uint8_t *const map =
2475 seg->update_map ? cpi->segmentation_map : cm->last_frame_seg_map;
2476 mi->segment_id = get_segment_id(cm, map, bsize, mi_row, mi_col);
2478 vp9_init_plane_quantizers(cpi, x);
2481 if (is_inter_block(mi)) {
2482 vp9_update_mv_count(td);
2483 if (cm->interp_filter == SWITCHABLE) {
2484 const int pred_ctx = get_pred_context_switchable_interp(xd);
2485 ++td->counts->switchable_interp[pred_ctx][mi->interp_filter];
2488 if (mi->sb_type < BLOCK_8X8) {
2489 mi->mv[0].as_int = mi->bmi[3].as_mv[0].as_int;
2490 mi->mv[1].as_int = mi->bmi[3].as_mv[1].as_int;
2494 if (cm->use_prev_frame_mvs || !cm->error_resilient_mode ||
2495 (cpi->svc.use_base_mv && cpi->svc.number_spatial_layers > 1 &&
2496 cpi->svc.spatial_layer_id != cpi->svc.number_spatial_layers - 1)) {
2497 MV_REF *const frame_mvs =
2498 cm->cur_frame->mvs + mi_row * cm->mi_cols + mi_col;
2501 for (h = 0; h < y_mis; ++h) {
2502 MV_REF *const frame_mv = frame_mvs + h * cm->mi_cols;
2503 for (w = 0; w < x_mis; ++w) {
2504 MV_REF *const mv = frame_mv + w;
2505 mv->ref_frame[0] = mi->ref_frame[0];
2506 mv->ref_frame[1] = mi->ref_frame[1];
2507 mv->mv[0].as_int = mi->mv[0].as_int;
2508 mv->mv[1].as_int = mi->mv[1].as_int;
2513 x->skip = ctx->skip;
2514 x->skip_txfm[0] = (mi->segment_id || xd->lossless) ? 0 : ctx->skip_txfm[0];
2517 static void encode_b_rt(VP9_COMP *cpi, ThreadData *td,
2518 const TileInfo *const tile, TOKENEXTRA **tp, int mi_row,
2519 int mi_col, int output_enabled, BLOCK_SIZE bsize,
2520 PICK_MODE_CONTEXT *ctx) {
2521 MACROBLOCK *const x = &td->mb;
2522 set_offsets(cpi, tile, x, mi_row, mi_col, bsize);
2523 update_state_rt(cpi, td, ctx, mi_row, mi_col, bsize);
2525 encode_superblock(cpi, td, tp, output_enabled, mi_row, mi_col, bsize, ctx);
2526 update_stats(&cpi->common, td);
2528 (*tp)->token = EOSB_TOKEN;
2532 static void encode_sb_rt(VP9_COMP *cpi, ThreadData *td,
2533 const TileInfo *const tile, TOKENEXTRA **tp,
2534 int mi_row, int mi_col, int output_enabled,
2535 BLOCK_SIZE bsize, PC_TREE *pc_tree) {
2536 VP9_COMMON *const cm = &cpi->common;
2537 MACROBLOCK *const x = &td->mb;
2538 MACROBLOCKD *const xd = &x->e_mbd;
2540 const int bsl = b_width_log2_lookup[bsize], hbs = (1 << bsl) / 4;
2542 PARTITION_TYPE partition;
2545 if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols) return;
2547 if (bsize >= BLOCK_8X8) {
2548 const int idx_str = xd->mi_stride * mi_row + mi_col;
2549 MODE_INFO **mi_8x8 = cm->mi_grid_visible + idx_str;
2550 ctx = partition_plane_context(xd, mi_row, mi_col, bsize);
2551 subsize = mi_8x8[0]->sb_type;
2554 subsize = BLOCK_4X4;
2557 partition = partition_lookup[bsl][subsize];
2558 if (output_enabled && bsize != BLOCK_4X4)
2559 td->counts->partition[ctx][partition]++;
2561 switch (partition) {
2562 case PARTITION_NONE:
2563 encode_b_rt(cpi, td, tile, tp, mi_row, mi_col, output_enabled, subsize,
2566 case PARTITION_VERT:
2567 encode_b_rt(cpi, td, tile, tp, mi_row, mi_col, output_enabled, subsize,
2568 &pc_tree->vertical[0]);
2569 if (mi_col + hbs < cm->mi_cols && bsize > BLOCK_8X8) {
2570 encode_b_rt(cpi, td, tile, tp, mi_row, mi_col + hbs, output_enabled,
2571 subsize, &pc_tree->vertical[1]);
2574 case PARTITION_HORZ:
2575 encode_b_rt(cpi, td, tile, tp, mi_row, mi_col, output_enabled, subsize,
2576 &pc_tree->horizontal[0]);
2577 if (mi_row + hbs < cm->mi_rows && bsize > BLOCK_8X8) {
2578 encode_b_rt(cpi, td, tile, tp, mi_row + hbs, mi_col, output_enabled,
2579 subsize, &pc_tree->horizontal[1]);
2583 assert(partition == PARTITION_SPLIT);
2584 subsize = get_subsize(bsize, PARTITION_SPLIT);
2585 encode_sb_rt(cpi, td, tile, tp, mi_row, mi_col, output_enabled, subsize,
2587 encode_sb_rt(cpi, td, tile, tp, mi_row, mi_col + hbs, output_enabled,
2588 subsize, pc_tree->split[1]);
2589 encode_sb_rt(cpi, td, tile, tp, mi_row + hbs, mi_col, output_enabled,
2590 subsize, pc_tree->split[2]);
2591 encode_sb_rt(cpi, td, tile, tp, mi_row + hbs, mi_col + hbs,
2592 output_enabled, subsize, pc_tree->split[3]);
2596 if (partition != PARTITION_SPLIT || bsize == BLOCK_8X8)
2597 update_partition_context(xd, mi_row, mi_col, subsize, bsize);
2600 static void rd_use_partition(VP9_COMP *cpi, ThreadData *td,
2601 TileDataEnc *tile_data, MODE_INFO **mi_8x8,
2602 TOKENEXTRA **tp, int mi_row, int mi_col,
2603 BLOCK_SIZE bsize, int *rate, int64_t *dist,
2604 int do_recon, PC_TREE *pc_tree) {
2605 VP9_COMMON *const cm = &cpi->common;
2606 TileInfo *const tile_info = &tile_data->tile_info;
2607 MACROBLOCK *const x = &td->mb;
2608 MACROBLOCKD *const xd = &x->e_mbd;
2609 const int mis = cm->mi_stride;
2610 const int bsl = b_width_log2_lookup[bsize];
2611 const int mi_step = num_4x4_blocks_wide_lookup[bsize] / 2;
2612 const int bss = (1 << bsl) / 4;
2614 PARTITION_TYPE partition = PARTITION_NONE;
2616 ENTROPY_CONTEXT l[16 * MAX_MB_PLANE], a[16 * MAX_MB_PLANE];
2617 PARTITION_CONTEXT sl[8], sa[8];
2618 RD_COST last_part_rdc, none_rdc, chosen_rdc;
2619 BLOCK_SIZE sub_subsize = BLOCK_4X4;
2620 int splits_below = 0;
2621 BLOCK_SIZE bs_type = mi_8x8[0]->sb_type;
2622 int do_partition_search = 1;
2623 PICK_MODE_CONTEXT *ctx = &pc_tree->none;
2625 if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols) return;
2627 assert(num_4x4_blocks_wide_lookup[bsize] ==
2628 num_4x4_blocks_high_lookup[bsize]);
2630 vp9_rd_cost_reset(&last_part_rdc);
2631 vp9_rd_cost_reset(&none_rdc);
2632 vp9_rd_cost_reset(&chosen_rdc);
2634 partition = partition_lookup[bsl][bs_type];
2635 subsize = get_subsize(bsize, partition);
2637 pc_tree->partitioning = partition;
2638 save_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
2640 if (bsize == BLOCK_16X16 && cpi->oxcf.aq_mode != NO_AQ) {
2641 set_offsets(cpi, tile_info, x, mi_row, mi_col, bsize);
2642 x->mb_energy = vp9_block_energy(cpi, x, bsize);
2645 if (do_partition_search &&
2646 cpi->sf.partition_search_type == SEARCH_PARTITION &&
2647 cpi->sf.adjust_partitioning_from_last_frame) {
2648 // Check if any of the sub blocks are further split.
2649 if (partition == PARTITION_SPLIT && subsize > BLOCK_8X8) {
2650 sub_subsize = get_subsize(subsize, PARTITION_SPLIT);
2652 for (i = 0; i < 4; i++) {
2653 int jj = i >> 1, ii = i & 0x01;
2654 MODE_INFO *this_mi = mi_8x8[jj * bss * mis + ii * bss];
2655 if (this_mi && this_mi->sb_type >= sub_subsize) {
2661 // If partition is not none try none unless each of the 4 splits are split
2663 if (partition != PARTITION_NONE && !splits_below &&
2664 mi_row + (mi_step >> 1) < cm->mi_rows &&
2665 mi_col + (mi_step >> 1) < cm->mi_cols) {
2666 pc_tree->partitioning = PARTITION_NONE;
2667 rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &none_rdc, bsize, ctx,
2670 pl = partition_plane_context(xd, mi_row, mi_col, bsize);
2672 if (none_rdc.rate < INT_MAX) {
2673 none_rdc.rate += cpi->partition_cost[pl][PARTITION_NONE];
2675 RDCOST(x->rdmult, x->rddiv, none_rdc.rate, none_rdc.dist);
2678 restore_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
2679 mi_8x8[0]->sb_type = bs_type;
2680 pc_tree->partitioning = partition;
2684 switch (partition) {
2685 case PARTITION_NONE:
2686 rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &last_part_rdc, bsize,
2689 case PARTITION_HORZ:
2690 pc_tree->horizontal[0].skip_ref_frame_mask = 0;
2691 rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &last_part_rdc,
2692 subsize, &pc_tree->horizontal[0], INT64_MAX);
2693 if (last_part_rdc.rate != INT_MAX && bsize >= BLOCK_8X8 &&
2694 mi_row + (mi_step >> 1) < cm->mi_rows) {
2696 PICK_MODE_CONTEXT *ctx = &pc_tree->horizontal[0];
2697 vp9_rd_cost_init(&tmp_rdc);
2698 update_state(cpi, td, ctx, mi_row, mi_col, subsize, 0);
2699 encode_superblock(cpi, td, tp, 0, mi_row, mi_col, subsize, ctx);
2700 pc_tree->horizontal[1].skip_ref_frame_mask = 0;
2701 rd_pick_sb_modes(cpi, tile_data, x, mi_row + (mi_step >> 1), mi_col,
2702 &tmp_rdc, subsize, &pc_tree->horizontal[1], INT64_MAX);
2703 if (tmp_rdc.rate == INT_MAX || tmp_rdc.dist == INT64_MAX) {
2704 vp9_rd_cost_reset(&last_part_rdc);
2707 last_part_rdc.rate += tmp_rdc.rate;
2708 last_part_rdc.dist += tmp_rdc.dist;
2709 last_part_rdc.rdcost += tmp_rdc.rdcost;
2712 case PARTITION_VERT:
2713 pc_tree->vertical[0].skip_ref_frame_mask = 0;
2714 rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &last_part_rdc,
2715 subsize, &pc_tree->vertical[0], INT64_MAX);
2716 if (last_part_rdc.rate != INT_MAX && bsize >= BLOCK_8X8 &&
2717 mi_col + (mi_step >> 1) < cm->mi_cols) {
2719 PICK_MODE_CONTEXT *ctx = &pc_tree->vertical[0];
2720 vp9_rd_cost_init(&tmp_rdc);
2721 update_state(cpi, td, ctx, mi_row, mi_col, subsize, 0);
2722 encode_superblock(cpi, td, tp, 0, mi_row, mi_col, subsize, ctx);
2723 pc_tree->vertical[bsize > BLOCK_8X8].skip_ref_frame_mask = 0;
2724 rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col + (mi_step >> 1),
2726 &pc_tree->vertical[bsize > BLOCK_8X8], INT64_MAX);
2727 if (tmp_rdc.rate == INT_MAX || tmp_rdc.dist == INT64_MAX) {
2728 vp9_rd_cost_reset(&last_part_rdc);
2731 last_part_rdc.rate += tmp_rdc.rate;
2732 last_part_rdc.dist += tmp_rdc.dist;
2733 last_part_rdc.rdcost += tmp_rdc.rdcost;
2737 assert(partition == PARTITION_SPLIT);
2738 if (bsize == BLOCK_8X8) {
2739 rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &last_part_rdc,
2740 subsize, pc_tree->leaf_split[0], INT64_MAX);
2743 last_part_rdc.rate = 0;
2744 last_part_rdc.dist = 0;
2745 last_part_rdc.rdcost = 0;
2746 for (i = 0; i < 4; i++) {
2747 int x_idx = (i & 1) * (mi_step >> 1);
2748 int y_idx = (i >> 1) * (mi_step >> 1);
2749 int jj = i >> 1, ii = i & 0x01;
2751 if ((mi_row + y_idx >= cm->mi_rows) || (mi_col + x_idx >= cm->mi_cols))
2754 vp9_rd_cost_init(&tmp_rdc);
2755 rd_use_partition(cpi, td, tile_data, mi_8x8 + jj * bss * mis + ii * bss,
2756 tp, mi_row + y_idx, mi_col + x_idx, subsize,
2757 &tmp_rdc.rate, &tmp_rdc.dist, i != 3,
2759 if (tmp_rdc.rate == INT_MAX || tmp_rdc.dist == INT64_MAX) {
2760 vp9_rd_cost_reset(&last_part_rdc);
2763 last_part_rdc.rate += tmp_rdc.rate;
2764 last_part_rdc.dist += tmp_rdc.dist;
2769 pl = partition_plane_context(xd, mi_row, mi_col, bsize);
2770 if (last_part_rdc.rate < INT_MAX) {
2771 last_part_rdc.rate += cpi->partition_cost[pl][partition];
2772 last_part_rdc.rdcost =
2773 RDCOST(x->rdmult, x->rddiv, last_part_rdc.rate, last_part_rdc.dist);
2776 if (do_partition_search && cpi->sf.adjust_partitioning_from_last_frame &&
2777 cpi->sf.partition_search_type == SEARCH_PARTITION &&
2778 partition != PARTITION_SPLIT && bsize > BLOCK_8X8 &&
2779 (mi_row + mi_step < cm->mi_rows ||
2780 mi_row + (mi_step >> 1) == cm->mi_rows) &&
2781 (mi_col + mi_step < cm->mi_cols ||
2782 mi_col + (mi_step >> 1) == cm->mi_cols)) {
2783 BLOCK_SIZE split_subsize = get_subsize(bsize, PARTITION_SPLIT);
2784 chosen_rdc.rate = 0;
2785 chosen_rdc.dist = 0;
2786 restore_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
2787 pc_tree->partitioning = PARTITION_SPLIT;
2790 for (i = 0; i < 4; i++) {
2791 int x_idx = (i & 1) * (mi_step >> 1);
2792 int y_idx = (i >> 1) * (mi_step >> 1);
2794 ENTROPY_CONTEXT l[16 * MAX_MB_PLANE], a[16 * MAX_MB_PLANE];
2795 PARTITION_CONTEXT sl[8], sa[8];
2797 if ((mi_row + y_idx >= cm->mi_rows) || (mi_col + x_idx >= cm->mi_cols))
2800 save_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
2801 pc_tree->split[i]->partitioning = PARTITION_NONE;
2802 rd_pick_sb_modes(cpi, tile_data, x, mi_row + y_idx, mi_col + x_idx,
2803 &tmp_rdc, split_subsize, &pc_tree->split[i]->none,
2806 restore_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
2808 if (tmp_rdc.rate == INT_MAX || tmp_rdc.dist == INT64_MAX) {
2809 vp9_rd_cost_reset(&chosen_rdc);
2813 chosen_rdc.rate += tmp_rdc.rate;
2814 chosen_rdc.dist += tmp_rdc.dist;
2817 encode_sb(cpi, td, tile_info, tp, mi_row + y_idx, mi_col + x_idx, 0,
2818 split_subsize, pc_tree->split[i]);
2820 pl = partition_plane_context(xd, mi_row + y_idx, mi_col + x_idx,
2822 chosen_rdc.rate += cpi->partition_cost[pl][PARTITION_NONE];
2824 pl = partition_plane_context(xd, mi_row, mi_col, bsize);
2825 if (chosen_rdc.rate < INT_MAX) {
2826 chosen_rdc.rate += cpi->partition_cost[pl][PARTITION_SPLIT];
2828 RDCOST(x->rdmult, x->rddiv, chosen_rdc.rate, chosen_rdc.dist);
2832 // If last_part is better set the partitioning to that.
2833 if (last_part_rdc.rdcost < chosen_rdc.rdcost) {
2834 mi_8x8[0]->sb_type = bsize;
2835 if (bsize >= BLOCK_8X8) pc_tree->partitioning = partition;
2836 chosen_rdc = last_part_rdc;
2838 // If none was better set the partitioning to that.
2839 if (none_rdc.rdcost < chosen_rdc.rdcost) {
2840 if (bsize >= BLOCK_8X8) pc_tree->partitioning = PARTITION_NONE;
2841 chosen_rdc = none_rdc;
2844 restore_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
2846 // We must have chosen a partitioning and encoding or we'll fail later on.
2847 // No other opportunities for success.
2848 if (bsize == BLOCK_64X64)
2849 assert(chosen_rdc.rate < INT_MAX && chosen_rdc.dist < INT64_MAX);
2852 int output_enabled = (bsize == BLOCK_64X64);
2853 encode_sb(cpi, td, tile_info, tp, mi_row, mi_col, output_enabled, bsize,
2857 *rate = chosen_rdc.rate;
2858 *dist = chosen_rdc.dist;
2861 static const BLOCK_SIZE min_partition_size[BLOCK_SIZES] = {
2862 BLOCK_4X4, BLOCK_4X4, BLOCK_4X4, BLOCK_4X4, BLOCK_4X4,
2863 BLOCK_4X4, BLOCK_8X8, BLOCK_8X8, BLOCK_8X8, BLOCK_16X16,
2864 BLOCK_16X16, BLOCK_16X16, BLOCK_16X16
2867 static const BLOCK_SIZE max_partition_size[BLOCK_SIZES] = {
2868 BLOCK_8X8, BLOCK_16X16, BLOCK_16X16, BLOCK_16X16, BLOCK_32X32,
2869 BLOCK_32X32, BLOCK_32X32, BLOCK_64X64, BLOCK_64X64, BLOCK_64X64,
2870 BLOCK_64X64, BLOCK_64X64, BLOCK_64X64
2873 // Look at all the mode_info entries for blocks that are part of this
2874 // partition and find the min and max values for sb_type.
2875 // At the moment this is designed to work on a 64x64 SB but could be
2876 // adjusted to use a size parameter.
2878 // The min and max are assumed to have been initialized prior to calling this
2879 // function so repeat calls can accumulate a min and max of more than one sb64.
2880 static void get_sb_partition_size_range(MACROBLOCKD *xd, MODE_INFO **mi_8x8,
2881 BLOCK_SIZE *min_block_size,
2882 BLOCK_SIZE *max_block_size,
2883 int bs_hist[BLOCK_SIZES]) {
2884 int sb_width_in_blocks = MI_BLOCK_SIZE;
2885 int sb_height_in_blocks = MI_BLOCK_SIZE;
2889 // Check the sb_type for each block that belongs to this region.
2890 for (i = 0; i < sb_height_in_blocks; ++i) {
2891 for (j = 0; j < sb_width_in_blocks; ++j) {
2892 MODE_INFO *mi = mi_8x8[index + j];
2893 BLOCK_SIZE sb_type = mi ? mi->sb_type : 0;
2895 *min_block_size = VPXMIN(*min_block_size, sb_type);
2896 *max_block_size = VPXMAX(*max_block_size, sb_type);
2898 index += xd->mi_stride;
2902 // Next square block size less or equal than current block size.
2903 static const BLOCK_SIZE next_square_size[BLOCK_SIZES] = {
2904 BLOCK_4X4, BLOCK_4X4, BLOCK_4X4, BLOCK_8X8, BLOCK_8X8,
2905 BLOCK_8X8, BLOCK_16X16, BLOCK_16X16, BLOCK_16X16, BLOCK_32X32,
2906 BLOCK_32X32, BLOCK_32X32, BLOCK_64X64
2909 // Look at neighboring blocks and set a min and max partition size based on
2911 static void rd_auto_partition_range(VP9_COMP *cpi, const TileInfo *const tile,
2912 MACROBLOCKD *const xd, int mi_row,
2913 int mi_col, BLOCK_SIZE *min_block_size,
2914 BLOCK_SIZE *max_block_size) {
2915 VP9_COMMON *const cm = &cpi->common;
2916 MODE_INFO **mi = xd->mi;
2917 const int left_in_image = !!xd->left_mi;
2918 const int above_in_image = !!xd->above_mi;
2919 const int row8x8_remaining = tile->mi_row_end - mi_row;
2920 const int col8x8_remaining = tile->mi_col_end - mi_col;
2922 BLOCK_SIZE min_size = BLOCK_4X4;
2923 BLOCK_SIZE max_size = BLOCK_64X64;
2924 int bs_hist[BLOCK_SIZES] = { 0 };
2926 // Trap case where we do not have a prediction.
2927 if (left_in_image || above_in_image || cm->frame_type != KEY_FRAME) {
2928 // Default "min to max" and "max to min"
2929 min_size = BLOCK_64X64;
2930 max_size = BLOCK_4X4;
2932 // NOTE: each call to get_sb_partition_size_range() uses the previous
2933 // passed in values for min and max as a starting point.
2934 // Find the min and max partition used in previous frame at this location
2935 if (cm->frame_type != KEY_FRAME) {
2936 MODE_INFO **prev_mi =
2937 &cm->prev_mi_grid_visible[mi_row * xd->mi_stride + mi_col];
2938 get_sb_partition_size_range(xd, prev_mi, &min_size, &max_size, bs_hist);
2940 // Find the min and max partition sizes used in the left SB64
2941 if (left_in_image) {
2942 MODE_INFO **left_sb64_mi = &mi[-MI_BLOCK_SIZE];
2943 get_sb_partition_size_range(xd, left_sb64_mi, &min_size, &max_size,
2946 // Find the min and max partition sizes used in the above SB64.
2947 if (above_in_image) {
2948 MODE_INFO **above_sb64_mi = &mi[-xd->mi_stride * MI_BLOCK_SIZE];
2949 get_sb_partition_size_range(xd, above_sb64_mi, &min_size, &max_size,
2953 // Adjust observed min and max for "relaxed" auto partition case.
2954 if (cpi->sf.auto_min_max_partition_size == RELAXED_NEIGHBORING_MIN_MAX) {
2955 min_size = min_partition_size[min_size];
2956 max_size = max_partition_size[max_size];
2960 // Check border cases where max and min from neighbors may not be legal.
2961 max_size = find_partition_size(max_size, row8x8_remaining, col8x8_remaining,
2963 // Test for blocks at the edge of the active image.
2964 // This may be the actual edge of the image or where there are formatting
2966 if (vp9_active_edge_sb(cpi, mi_row, mi_col)) {
2967 min_size = BLOCK_4X4;
2970 VPXMIN(cpi->sf.rd_auto_partition_min_limit, VPXMIN(min_size, max_size));
2973 // When use_square_partition_only is true, make sure at least one square
2974 // partition is allowed by selecting the next smaller square size as
2976 if (cpi->sf.use_square_partition_only &&
2977 next_square_size[max_size] < min_size) {
2978 min_size = next_square_size[max_size];
2981 *min_block_size = min_size;
2982 *max_block_size = max_size;
2985 // TODO(jingning) refactor functions setting partition search range
2986 static void set_partition_range(VP9_COMMON *cm, MACROBLOCKD *xd, int mi_row,
2987 int mi_col, BLOCK_SIZE bsize,
2988 BLOCK_SIZE *min_bs, BLOCK_SIZE *max_bs) {
2989 int mi_width = num_8x8_blocks_wide_lookup[bsize];
2990 int mi_height = num_8x8_blocks_high_lookup[bsize];
2994 const int idx_str = cm->mi_stride * mi_row + mi_col;
2995 MODE_INFO **prev_mi = &cm->prev_mi_grid_visible[idx_str];
2996 BLOCK_SIZE bs, min_size, max_size;
2998 min_size = BLOCK_64X64;
2999 max_size = BLOCK_4X4;
3002 for (idy = 0; idy < mi_height; ++idy) {
3003 for (idx = 0; idx < mi_width; ++idx) {
3004 mi = prev_mi[idy * cm->mi_stride + idx];
3005 bs = mi ? mi->sb_type : bsize;
3006 min_size = VPXMIN(min_size, bs);
3007 max_size = VPXMAX(max_size, bs);
3013 for (idy = 0; idy < mi_height; ++idy) {
3014 mi = xd->mi[idy * cm->mi_stride - 1];
3015 bs = mi ? mi->sb_type : bsize;
3016 min_size = VPXMIN(min_size, bs);
3017 max_size = VPXMAX(max_size, bs);
3022 for (idx = 0; idx < mi_width; ++idx) {
3023 mi = xd->mi[idx - cm->mi_stride];
3024 bs = mi ? mi->sb_type : bsize;
3025 min_size = VPXMIN(min_size, bs);
3026 max_size = VPXMAX(max_size, bs);
3030 if (min_size == max_size) {
3031 min_size = min_partition_size[min_size];
3032 max_size = max_partition_size[max_size];
3039 static INLINE void store_pred_mv(MACROBLOCK *x, PICK_MODE_CONTEXT *ctx) {
3040 memcpy(ctx->pred_mv, x->pred_mv, sizeof(x->pred_mv));
3043 static INLINE void load_pred_mv(MACROBLOCK *x, PICK_MODE_CONTEXT *ctx) {
3044 memcpy(x->pred_mv, ctx->pred_mv, sizeof(x->pred_mv));
3047 #if CONFIG_FP_MB_STATS
3048 const int num_16x16_blocks_wide_lookup[BLOCK_SIZES] = { 1, 1, 1, 1, 1, 1, 1,
3050 const int num_16x16_blocks_high_lookup[BLOCK_SIZES] = { 1, 1, 1, 1, 1, 1, 1,
3052 const int qindex_skip_threshold_lookup[BLOCK_SIZES] = { 0, 10, 10, 30, 40,
3055 const int qindex_split_threshold_lookup[BLOCK_SIZES] = { 0, 3, 3, 7, 15,
3058 const int complexity_16x16_blocks_threshold[BLOCK_SIZES] = { 1, 1, 1, 1, 1,
3071 static INLINE MOTION_DIRECTION get_motion_direction_fp(uint8_t fp_byte) {
3072 if (fp_byte & FPMB_MOTION_ZERO_MASK) {
3074 } else if (fp_byte & FPMB_MOTION_LEFT_MASK) {
3076 } else if (fp_byte & FPMB_MOTION_RIGHT_MASK) {
3078 } else if (fp_byte & FPMB_MOTION_UP_MASK) {
3085 static INLINE int get_motion_inconsistency(MOTION_DIRECTION this_mv,
3086 MOTION_DIRECTION that_mv) {
3087 if (this_mv == that_mv) {
3090 return abs(this_mv - that_mv) == 2 ? 2 : 1;
3095 // Calculate prediction based on the given input features and neural net config.
3096 // Assume there are no more than NN_MAX_NODES_PER_LAYER nodes in each hidden
3098 static void nn_predict(const float *features, const NN_CONFIG *nn_config,
3100 int num_input_nodes = nn_config->num_inputs;
3102 float buf[2][NN_MAX_NODES_PER_LAYER];
3103 const float *input_nodes = features;
3105 // Propagate hidden layers.
3106 const int num_layers = nn_config->num_hidden_layers;
3108 assert(num_layers <= NN_MAX_HIDDEN_LAYERS);
3109 for (layer = 0; layer < num_layers; ++layer) {
3110 const float *weights = nn_config->weights[layer];
3111 const float *bias = nn_config->bias[layer];
3112 float *output_nodes = buf[buf_index];
3113 const int num_output_nodes = nn_config->num_hidden_nodes[layer];
3114 assert(num_output_nodes < NN_MAX_NODES_PER_LAYER);
3115 for (node = 0; node < num_output_nodes; ++node) {
3117 for (i = 0; i < num_input_nodes; ++i) val += weights[i] * input_nodes[i];
3119 // ReLU as activation function.
3120 val = VPXMAX(val, 0.0f);
3121 output_nodes[node] = val;
3122 weights += num_input_nodes;
3124 num_input_nodes = num_output_nodes;
3125 input_nodes = output_nodes;
3126 buf_index = 1 - buf_index;
3129 // Final output layer.
3131 const float *weights = nn_config->weights[num_layers];
3132 for (node = 0; node < nn_config->num_outputs; ++node) {
3133 const float *bias = nn_config->bias[num_layers];
3135 for (i = 0; i < num_input_nodes; ++i) val += weights[i] * input_nodes[i];
3136 output[node] = val + bias[node];
3137 weights += num_input_nodes;
3143 // Machine-learning based partition search early termination.
3144 // Return 1 to skip split and rect partitions.
3145 static int ml_pruning_partition(VP9_COMMON *const cm, MACROBLOCKD *const xd,
3146 PICK_MODE_CONTEXT *ctx, int mi_row, int mi_col,
3149 abs(ctx->mic.mv[0].as_mv.col) + abs(ctx->mic.mv[0].as_mv.row);
3150 const int left_in_image = !!xd->left_mi;
3151 const int above_in_image = !!xd->above_mi;
3152 MODE_INFO **prev_mi =
3153 &cm->prev_mi_grid_visible[mi_col + cm->mi_stride * mi_row];
3154 int above_par = 0; // above_partitioning
3155 int left_par = 0; // left_partitioning
3156 int last_par = 0; // last_partitioning
3159 BLOCK_SIZE context_size;
3160 const NN_CONFIG *nn_config = NULL;
3161 const float *mean, *sd, *linear_weights;
3162 float nn_score, linear_score;
3163 float features[FEATURES];
3165 assert(b_width_log2_lookup[bsize] == b_height_log2_lookup[bsize]);
3166 vpx_clear_system_state();
3171 nn_config = &vp9_partition_nnconfig_64x64;
3175 nn_config = &vp9_partition_nnconfig_32x32;
3179 nn_config = &vp9_partition_nnconfig_16x16;
3181 default: assert(0 && "Unexpected block size."); return 0;
3184 if (above_in_image) {
3185 context_size = xd->above_mi->sb_type;
3186 if (context_size < bsize)
3188 else if (context_size == bsize)
3192 if (left_in_image) {
3193 context_size = xd->left_mi->sb_type;
3194 if (context_size < bsize)
3196 else if (context_size == bsize)
3201 context_size = prev_mi[0]->sb_type;
3202 if (context_size < bsize)
3204 else if (context_size == bsize)
3208 mean = &vp9_partition_feature_mean[offset];
3209 sd = &vp9_partition_feature_std[offset];
3210 features[0] = ((float)ctx->rate - mean[0]) / sd[0];
3211 features[1] = ((float)ctx->dist - mean[1]) / sd[1];
3212 features[2] = ((float)mag_mv / 2 - mean[2]) * sd[2];
3213 features[3] = ((float)(left_par + above_par) / 2 - mean[3]) * sd[3];
3214 features[4] = ((float)ctx->sum_y_eobs - mean[4]) / sd[4];
3215 features[5] = ((float)cm->base_qindex - mean[5]) * sd[5];
3216 features[6] = ((float)last_par - mean[6]) * sd[6];
3218 // Predict using linear model.
3219 linear_weights = &vp9_partition_linear_weights[offset];
3220 linear_score = linear_weights[FEATURES];
3221 for (i = 0; i < FEATURES; ++i)
3222 linear_score += linear_weights[i] * features[i];
3223 if (linear_score > 0.1f) return 0;
3225 // Predict using neural net model.
3226 nn_predict(features, nn_config, &nn_score);
3228 if (linear_score < -0.0f && nn_score < 0.1f) return 1;
3229 if (nn_score < -0.0f && linear_score < 0.1f) return 1;
3235 // ML-based partition search breakout.
3236 static int ml_predict_breakout(VP9_COMP *const cpi, BLOCK_SIZE bsize,
3237 const MACROBLOCK *const x,
3238 const RD_COST *const rd_cost) {
3239 DECLARE_ALIGNED(16, static const uint8_t, vp9_64_zeros[64]) = { 0 };
3240 const VP9_COMMON *const cm = &cpi->common;
3241 float features[FEATURES];
3242 const float *linear_weights = NULL; // Linear model weights.
3243 float linear_score = 0.0f;
3244 const int qindex = cm->base_qindex;
3245 const int q_ctx = qindex >= 200 ? 0 : (qindex >= 150 ? 1 : 2);
3246 const int is_720p_or_larger = VPXMIN(cm->width, cm->height) >= 720;
3247 const int resolution_ctx = is_720p_or_larger ? 1 : 0;
3251 linear_weights = vp9_partition_breakout_weights_64[resolution_ctx][q_ctx];
3254 linear_weights = vp9_partition_breakout_weights_32[resolution_ctx][q_ctx];
3257 linear_weights = vp9_partition_breakout_weights_16[resolution_ctx][q_ctx];
3260 linear_weights = vp9_partition_breakout_weights_8[resolution_ctx][q_ctx];
3262 default: assert(0 && "Unexpected block size."); return 0;
3264 if (!linear_weights) return 0;
3266 { // Generate feature values.
3267 #if CONFIG_VP9_HIGHBITDEPTH
3269 vp9_ac_quant(cm->base_qindex, 0, cm->bit_depth) >> (x->e_mbd.bd - 8);
3271 const int ac_q = vp9_ac_quant(qindex, 0, cm->bit_depth);
3272 #endif // CONFIG_VP9_HIGHBITDEPTH
3273 const int num_pels_log2 = num_pels_log2_lookup[bsize];
3274 int feature_index = 0;
3275 unsigned int var, sse;
3276 float rate_f, dist_f;
3278 #if CONFIG_VP9_HIGHBITDEPTH
3279 if (x->e_mbd.cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
3281 vp9_high_get_sby_variance(cpi, &x->plane[0].src, bsize, x->e_mbd.bd);
3283 var = cpi->fn_ptr[bsize].vf(x->plane[0].src.buf, x->plane[0].src.stride,
3284 vp9_64_zeros, 0, &sse);
3287 var = cpi->fn_ptr[bsize].vf(x->plane[0].src.buf, x->plane[0].src.stride,
3288 vp9_64_zeros, 0, &sse);
3290 var = var >> num_pels_log2;
3292 vpx_clear_system_state();
3294 rate_f = (float)VPXMIN(rd_cost->rate, INT_MAX);
3295 dist_f = (float)(VPXMIN(rd_cost->dist, INT_MAX) >> num_pels_log2);
3297 ((float)x->rdmult / 128.0f / 512.0f / (float)(1 << num_pels_log2)) *
3300 features[feature_index++] = rate_f;
3301 features[feature_index++] = dist_f;
3302 features[feature_index++] = (float)var;
3303 features[feature_index++] = (float)ac_q;
3304 assert(feature_index == FEATURES);
3307 { // Calculate the output score.
3309 linear_score = linear_weights[FEATURES];
3310 for (i = 0; i < FEATURES; ++i)
3311 linear_score += linear_weights[i] * features[i];
3314 return linear_score >= cpi->sf.rd_ml_partition.search_breakout_thresh[q_ctx];
3320 static void ml_prune_rect_partition(VP9_COMP *const cpi, MACROBLOCK *const x,
3322 const PC_TREE *const pc_tree,
3323 int *allow_horz, int *allow_vert,
3325 const NN_CONFIG *nn_config = NULL;
3326 float score[LABELS] = {
3333 if (ref_rd <= 0 || ref_rd > 1000000000) return;
3336 case BLOCK_8X8: break;
3338 nn_config = &vp9_rect_part_nnconfig_16;
3339 thresh = cpi->sf.rd_ml_partition.prune_rect_thresh[1];
3342 nn_config = &vp9_rect_part_nnconfig_32;
3343 thresh = cpi->sf.rd_ml_partition.prune_rect_thresh[2];
3346 nn_config = &vp9_rect_part_nnconfig_64;
3347 thresh = cpi->sf.rd_ml_partition.prune_rect_thresh[3];
3349 default: assert(0 && "Unexpected block size."); return;
3351 if (!nn_config || thresh < 0) return;
3353 // Feature extraction and model score calculation.
3355 const VP9_COMMON *const cm = &cpi->common;
3356 #if CONFIG_VP9_HIGHBITDEPTH
3358 vp9_dc_quant(cm->base_qindex, 0, cm->bit_depth) >> (x->e_mbd.bd - 8);
3360 const int dc_q = vp9_dc_quant(cm->base_qindex, 0, cm->bit_depth);
3361 #endif // CONFIG_VP9_HIGHBITDEPTH
3362 const int bs = 4 * num_4x4_blocks_wide_lookup[bsize];
3363 int feature_index = 0;
3364 float features[FEATURES];
3366 features[feature_index++] = logf((float)dc_q + 1.0f);
3367 features[feature_index++] =
3368 (float)(pc_tree->partitioning == PARTITION_NONE);
3369 features[feature_index++] = logf((float)ref_rd / bs / bs + 1.0f);
3372 const float norm_factor = 1.0f / ((float)ref_rd + 1.0f);
3373 const int64_t none_rdcost = pc_tree->none.rdcost;
3374 float rd_ratio = 2.0f;
3375 if (none_rdcost > 0 && none_rdcost < 1000000000)
3376 rd_ratio = (float)none_rdcost * norm_factor;
3377 features[feature_index++] = VPXMIN(rd_ratio, 2.0f);
3379 for (i = 0; i < 4; ++i) {
3380 const int64_t this_rd = pc_tree->split[i]->none.rdcost;
3381 const int rd_valid = this_rd > 0 && this_rd < 1000000000;
3382 // Ratio between sub-block RD and whole block RD.
3383 features[feature_index++] =
3384 rd_valid ? (float)this_rd * norm_factor : 1.0f;
3388 assert(feature_index == FEATURES);
3389 nn_predict(features, nn_config, score);
3392 // Make decisions based on the model score.
3394 int max_score = -1000;
3395 int horz = 0, vert = 0;
3396 int int_score[LABELS];
3397 for (i = 0; i < LABELS; ++i) {
3398 int_score[i] = (int)(100 * score[i]);
3399 max_score = VPXMAX(int_score[i], max_score);
3401 thresh = max_score - thresh;
3402 for (i = 0; i < LABELS; ++i) {
3403 if (int_score[i] >= thresh) {
3404 if ((i >> 0) & 1) horz = 1;
3405 if ((i >> 1) & 1) vert = 1;
3408 *allow_horz = *allow_horz && horz;
3409 *allow_vert = *allow_vert && vert;
3415 // Perform fast and coarse motion search for the given block. This is a
3416 // pre-processing step for the ML based partition search speedup.
3417 static void simple_motion_search(const VP9_COMP *const cpi, MACROBLOCK *const x,
3418 BLOCK_SIZE bsize, int mi_row, int mi_col,
3419 MV ref_mv, MV_REFERENCE_FRAME ref,
3420 uint8_t *const pred_buf) {
3421 const VP9_COMMON *const cm = &cpi->common;
3422 MACROBLOCKD *const xd = &x->e_mbd;
3423 MODE_INFO *const mi = xd->mi[0];
3424 const YV12_BUFFER_CONFIG *const yv12 = get_ref_frame_buffer(cpi, ref);
3425 const int step_param = 1;
3426 const MvLimits tmp_mv_limits = x->mv_limits;
3427 const SEARCH_METHODS search_method = NSTEP;
3428 const int sadpb = x->sadperbit16;
3429 MV ref_mv_full = { ref_mv.row >> 3, ref_mv.col >> 3 };
3430 MV best_mv = { 0, 0 };
3433 assert(yv12 != NULL);
3435 vp9_setup_pre_planes(xd, 0, yv12, mi_row, mi_col,
3436 &cm->frame_refs[ref - 1].sf);
3437 mi->ref_frame[0] = ref;
3438 mi->ref_frame[1] = NONE;
3439 mi->sb_type = bsize;
3440 vp9_set_mv_search_range(&x->mv_limits, &ref_mv);
3441 vp9_full_pixel_search(cpi, x, bsize, &ref_mv_full, step_param, search_method,
3442 sadpb, cond_cost_list(cpi, cost_list), &ref_mv,
3446 x->mv_limits = tmp_mv_limits;
3447 mi->mv[0].as_mv = best_mv;
3449 set_ref_ptrs(cm, xd, mi->ref_frame[0], mi->ref_frame[1]);
3450 xd->plane[0].dst.buf = pred_buf;
3451 xd->plane[0].dst.stride = 64;
3452 vp9_build_inter_predictors_sby(xd, mi_row, mi_col, bsize);
3455 // Use a neural net model to prune partition-none and partition-split search.
3456 // Features used: QP; spatial block size contexts; variance of prediction
3457 // residue after simple_motion_search.
3459 static void ml_predict_var_rd_paritioning(const VP9_COMP *const cpi,
3460 MACROBLOCK *const x,
3461 PC_TREE *const pc_tree,
3462 BLOCK_SIZE bsize, int mi_row,
3463 int mi_col, int *none, int *split) {
3464 const VP9_COMMON *const cm = &cpi->common;
3465 const NN_CONFIG *nn_config = NULL;
3466 #if CONFIG_VP9_HIGHBITDEPTH
3467 MACROBLOCKD *xd = &x->e_mbd;
3468 DECLARE_ALIGNED(16, uint8_t, pred_buffer[64 * 64 * 2]);
3469 uint8_t *const pred_buf = (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH)
3470 ? (CONVERT_TO_BYTEPTR(pred_buffer))
3473 DECLARE_ALIGNED(16, uint8_t, pred_buffer[64 * 64]);
3474 uint8_t *const pred_buf = pred_buffer;
3475 #endif // CONFIG_VP9_HIGHBITDEPTH
3476 const int speed = cpi->oxcf.speed;
3477 float thresh = 0.0f;
3481 nn_config = &vp9_part_split_nnconfig_64;
3482 thresh = speed > 0 ? 2.8f : 3.0f;
3485 nn_config = &vp9_part_split_nnconfig_32;
3486 thresh = speed > 0 ? 3.5f : 3.0f;
3489 nn_config = &vp9_part_split_nnconfig_16;
3490 thresh = speed > 0 ? 3.8f : 4.0f;
3493 nn_config = &vp9_part_split_nnconfig_8;
3494 if (cm->width >= 720 && cm->height >= 720)
3495 thresh = speed > 0 ? 2.5f : 2.0f;
3497 thresh = speed > 0 ? 3.8f : 2.0f;
3499 default: assert(0 && "Unexpected block size."); return;
3502 if (!nn_config) return;
3504 // Do a simple single motion search to find a prediction for current block.
3505 // The variance of the residue will be used as input features.
3508 const MV_REFERENCE_FRAME ref =
3509 cpi->rc.is_src_frame_alt_ref ? ALTREF_FRAME : LAST_FRAME;
3510 // If bsize is 64x64, use zero MV as reference; otherwise, use MV result
3511 // of previous(larger) block as reference.
3512 if (bsize == BLOCK_64X64)
3513 ref_mv.row = ref_mv.col = 0;
3515 ref_mv = pc_tree->mv;
3516 vp9_setup_src_planes(x, cpi->Source, mi_row, mi_col);
3517 simple_motion_search(cpi, x, bsize, mi_row, mi_col, ref_mv, ref, pred_buf);
3518 pc_tree->mv = x->e_mbd.mi[0]->mv[0].as_mv;
3521 vpx_clear_system_state();
3524 float features[FEATURES] = { 0.0f };
3525 #if CONFIG_VP9_HIGHBITDEPTH
3527 vp9_dc_quant(cm->base_qindex, 0, cm->bit_depth) >> (xd->bd - 8);
3529 const int dc_q = vp9_dc_quant(cm->base_qindex, 0, cm->bit_depth);
3530 #endif // CONFIG_VP9_HIGHBITDEPTH
3531 int feature_idx = 0;
3534 // Generate model input features.
3535 features[feature_idx++] = logf((float)dc_q + 1.0f);
3537 // Get the variance of the residue as input features.
3539 const int bs = 4 * num_4x4_blocks_wide_lookup[bsize];
3540 const BLOCK_SIZE subsize = get_subsize(bsize, PARTITION_SPLIT);
3541 const uint8_t *pred = pred_buf;
3542 const uint8_t *src = x->plane[0].src.buf;
3543 const int src_stride = x->plane[0].src.stride;
3544 const int pred_stride = 64;
3546 // Variance of whole block.
3547 const unsigned int var =
3548 cpi->fn_ptr[bsize].vf(src, src_stride, pred, pred_stride, &sse);
3549 const float factor = (var == 0) ? 1.0f : (1.0f / (float)var);
3550 const MACROBLOCKD *const xd = &x->e_mbd;
3551 const int has_above = !!xd->above_mi;
3552 const int has_left = !!xd->left_mi;
3553 const BLOCK_SIZE above_bsize = has_above ? xd->above_mi->sb_type : bsize;
3554 const BLOCK_SIZE left_bsize = has_left ? xd->left_mi->sb_type : bsize;
3557 features[feature_idx++] = (float)has_above;
3558 features[feature_idx++] = (float)b_width_log2_lookup[above_bsize];
3559 features[feature_idx++] = (float)b_height_log2_lookup[above_bsize];
3560 features[feature_idx++] = (float)has_left;
3561 features[feature_idx++] = (float)b_width_log2_lookup[left_bsize];
3562 features[feature_idx++] = (float)b_height_log2_lookup[left_bsize];
3563 features[feature_idx++] = logf((float)var + 1.0f);
3564 for (i = 0; i < 4; ++i) {
3565 const int x_idx = (i & 1) * bs / 2;
3566 const int y_idx = (i >> 1) * bs / 2;
3567 const int src_offset = y_idx * src_stride + x_idx;
3568 const int pred_offset = y_idx * pred_stride + x_idx;
3569 // Variance of quarter block.
3570 const unsigned int sub_var =
3571 cpi->fn_ptr[subsize].vf(src + src_offset, src_stride,
3572 pred + pred_offset, pred_stride, &sse);
3573 const float var_ratio = (var == 0) ? 1.0f : factor * (float)sub_var;
3574 features[feature_idx++] = var_ratio;
3577 assert(feature_idx == FEATURES);
3579 // Feed the features into the model to get the confidence score.
3580 nn_predict(features, nn_config, &score);
3582 // Higher score means that the model has higher confidence that the split
3583 // partition is better than the non-split partition. So if the score is
3584 // high enough, we skip the none-split partition search; if the score is
3585 // low enough, we skip the split partition search.
3586 if (score > thresh) *none = 0;
3587 if (score < -thresh) *split = 0;
3592 static double log_wiener_var(int64_t wiener_variance) {
3593 return log(1.0 + wiener_variance) / log(2.0);
3596 static void build_kmeans_segmentation(VP9_COMP *cpi) {
3597 VP9_COMMON *cm = &cpi->common;
3598 BLOCK_SIZE bsize = BLOCK_64X64;
3599 KMEANS_DATA *kmeans_data;
3601 vp9_disable_segmentation(&cm->seg);
3602 if (cm->show_frame) {
3604 cpi->kmeans_data_size = 0;
3605 cpi->kmeans_ctr_num = 8;
3607 for (mi_row = 0; mi_row < cm->mi_rows; mi_row += MI_BLOCK_SIZE) {
3608 for (mi_col = 0; mi_col < cm->mi_cols; mi_col += MI_BLOCK_SIZE) {
3609 int mb_row_start = mi_row >> 1;
3610 int mb_col_start = mi_col >> 1;
3611 int mb_row_end = VPXMIN(
3612 (mi_row + num_8x8_blocks_high_lookup[bsize]) >> 1, cm->mb_rows);
3613 int mb_col_end = VPXMIN(
3614 (mi_col + num_8x8_blocks_wide_lookup[bsize]) >> 1, cm->mb_cols);
3616 int64_t wiener_variance = 0;
3618 for (row = mb_row_start; row < mb_row_end; ++row)
3619 for (col = mb_col_start; col < mb_col_end; ++col)
3620 wiener_variance += cpi->mb_wiener_variance[row * cm->mb_cols + col];
3623 (mb_row_end - mb_row_start) * (mb_col_end - mb_col_start);
3625 #if CONFIG_MULTITHREAD
3626 pthread_mutex_lock(&cpi->kmeans_mutex);
3627 #endif // CONFIG_MULTITHREAD
3629 kmeans_data = &cpi->kmeans_data_arr[cpi->kmeans_data_size++];
3630 kmeans_data->value = log_wiener_var(wiener_variance);
3631 kmeans_data->pos = mi_row * cpi->kmeans_data_stride + mi_col;
3632 #if CONFIG_MULTITHREAD
3633 pthread_mutex_unlock(&cpi->kmeans_mutex);
3634 #endif // CONFIG_MULTITHREAD
3638 vp9_kmeans(cpi->kmeans_ctr_ls, cpi->kmeans_boundary_ls,
3639 cpi->kmeans_count_ls, cpi->kmeans_ctr_num, cpi->kmeans_data_arr,
3640 cpi->kmeans_data_size);
3642 vp9_perceptual_aq_mode_setup(cpi, &cm->seg);
3646 static int wiener_var_segment(VP9_COMP *cpi, BLOCK_SIZE bsize, int mi_row,
3648 VP9_COMMON *cm = &cpi->common;
3649 int mb_row_start = mi_row >> 1;
3650 int mb_col_start = mi_col >> 1;
3652 VPXMIN((mi_row + num_8x8_blocks_high_lookup[bsize]) >> 1, cm->mb_rows);
3654 VPXMIN((mi_col + num_8x8_blocks_wide_lookup[bsize]) >> 1, cm->mb_cols);
3656 int64_t wiener_variance = 0;
3658 int8_t seg_hist[MAX_SEGMENTS] = { 0 };
3659 int8_t max_count = 0, max_index = -1;
3661 vpx_clear_system_state();
3663 assert(cpi->norm_wiener_variance > 0);
3665 for (row = mb_row_start; row < mb_row_end; ++row) {
3666 for (col = mb_col_start; col < mb_col_end; ++col) {
3667 wiener_variance = cpi->mb_wiener_variance[row * cm->mb_cols + col];
3669 vp9_get_group_idx(log_wiener_var(wiener_variance),
3670 cpi->kmeans_boundary_ls, cpi->kmeans_ctr_num);
3671 ++seg_hist[segment_id];
3675 for (idx = 0; idx < cpi->kmeans_ctr_num; ++idx) {
3676 if (seg_hist[idx] > max_count) {
3677 max_count = seg_hist[idx];
3682 assert(max_index >= 0);
3683 segment_id = max_index;
3688 static int get_rdmult_delta(VP9_COMP *cpi, BLOCK_SIZE bsize, int mi_row,
3689 int mi_col, int orig_rdmult) {
3690 const int gf_group_index = cpi->twopass.gf_group.index;
3691 TplDepFrame *tpl_frame = &cpi->tpl_stats[gf_group_index];
3692 TplDepStats *tpl_stats = tpl_frame->tpl_stats_ptr;
3693 int tpl_stride = tpl_frame->stride;
3694 int64_t intra_cost = 0;
3695 int64_t mc_dep_cost = 0;
3696 int mi_wide = num_8x8_blocks_wide_lookup[bsize];
3697 int mi_high = num_8x8_blocks_high_lookup[bsize];
3702 double r0, rk, beta;
3704 if (tpl_frame->is_valid == 0) return orig_rdmult;
3706 if (cpi->twopass.gf_group.layer_depth[gf_group_index] > 1) return orig_rdmult;
3708 if (gf_group_index >= MAX_ARF_GOP_SIZE) return orig_rdmult;
3710 for (row = mi_row; row < mi_row + mi_high; ++row) {
3711 for (col = mi_col; col < mi_col + mi_wide; ++col) {
3712 TplDepStats *this_stats = &tpl_stats[row * tpl_stride + col];
3714 if (row >= cpi->common.mi_rows || col >= cpi->common.mi_cols) continue;
3716 intra_cost += this_stats->intra_cost;
3717 mc_dep_cost += this_stats->mc_dep_cost;
3723 vpx_clear_system_state();
3726 rk = (double)intra_cost / mc_dep_cost;
3728 dr = vp9_get_adaptive_rdmult(cpi, beta);
3730 dr = VPXMIN(dr, orig_rdmult * 3 / 2);
3731 dr = VPXMAX(dr, orig_rdmult * 1 / 2);
3738 // TODO(jingning,jimbankoski,rbultje): properly skip partition types that are
3739 // unlikely to be selected depending on previous rate-distortion optimization
3740 // results, for encoding speed-up.
3741 static void rd_pick_partition(VP9_COMP *cpi, ThreadData *td,
3742 TileDataEnc *tile_data, TOKENEXTRA **tp,
3743 int mi_row, int mi_col, BLOCK_SIZE bsize,
3744 RD_COST *rd_cost, int64_t best_rd,
3746 VP9_COMMON *const cm = &cpi->common;
3747 TileInfo *const tile_info = &tile_data->tile_info;
3748 MACROBLOCK *const x = &td->mb;
3749 MACROBLOCKD *const xd = &x->e_mbd;
3750 const int mi_step = num_8x8_blocks_wide_lookup[bsize] / 2;
3751 ENTROPY_CONTEXT l[16 * MAX_MB_PLANE], a[16 * MAX_MB_PLANE];
3752 PARTITION_CONTEXT sl[8], sa[8];
3753 TOKENEXTRA *tp_orig = *tp;
3754 PICK_MODE_CONTEXT *const ctx = &pc_tree->none;
3756 const int pl = partition_plane_context(xd, mi_row, mi_col, bsize);
3758 RD_COST this_rdc, sum_rdc, best_rdc;
3759 int do_split = bsize >= BLOCK_8X8;
3761 INTERP_FILTER pred_interp_filter;
3763 // Override skipping rectangular partition operations for edge blocks
3764 const int force_horz_split = (mi_row + mi_step >= cm->mi_rows);
3765 const int force_vert_split = (mi_col + mi_step >= cm->mi_cols);
3766 const int xss = x->e_mbd.plane[1].subsampling_x;
3767 const int yss = x->e_mbd.plane[1].subsampling_y;
3769 BLOCK_SIZE min_size = x->min_partition_size;
3770 BLOCK_SIZE max_size = x->max_partition_size;
3772 #if CONFIG_FP_MB_STATS
3773 unsigned int src_diff_var = UINT_MAX;
3774 int none_complexity = 0;
3777 int partition_none_allowed = !force_horz_split && !force_vert_split;
3778 int partition_horz_allowed =
3779 !force_vert_split && yss <= xss && bsize >= BLOCK_8X8;
3780 int partition_vert_allowed =
3781 !force_horz_split && xss <= yss && bsize >= BLOCK_8X8;
3783 int64_t dist_breakout_thr = cpi->sf.partition_search_breakout_thr.dist;
3784 int rate_breakout_thr = cpi->sf.partition_search_breakout_thr.rate;
3786 int partition_mul = x->cb_rdmult;
3787 // Ref frames picked in the [i_th] quarter subblock during square partition
3788 // RD search. It may be used to prune ref frame selection of rect partitions.
3789 uint8_t ref_frames_used[4] = { 0, 0, 0, 0 };
3793 assert(num_8x8_blocks_wide_lookup[bsize] ==
3794 num_8x8_blocks_high_lookup[bsize]);
3796 dist_breakout_thr >>=
3797 8 - (b_width_log2_lookup[bsize] + b_height_log2_lookup[bsize]);
3799 rate_breakout_thr *= num_pels_log2_lookup[bsize];
3801 vp9_rd_cost_init(&this_rdc);
3802 vp9_rd_cost_init(&sum_rdc);
3803 vp9_rd_cost_reset(&best_rdc);
3804 best_rdc.rdcost = best_rd;
3806 set_offsets(cpi, tile_info, x, mi_row, mi_col, bsize);
3808 if (bsize == BLOCK_16X16 && cpi->oxcf.aq_mode != NO_AQ &&
3809 cpi->oxcf.aq_mode != LOOKAHEAD_AQ)
3810 x->mb_energy = vp9_block_energy(cpi, x, bsize);
3812 if (cpi->sf.cb_partition_search && bsize == BLOCK_16X16) {
3813 int cb_partition_search_ctrl =
3814 ((pc_tree->index == 0 || pc_tree->index == 3) +
3815 get_chessboard_index(cm->current_video_frame)) &
3818 if (cb_partition_search_ctrl && bsize > min_size && bsize < max_size)
3819 set_partition_range(cm, xd, mi_row, mi_col, bsize, &min_size, &max_size);
3822 // Get sub block energy range
3823 if (bsize >= BLOCK_16X16) {
3824 int min_energy, max_energy;
3825 vp9_get_sub_block_energy(cpi, x, mi_row, mi_col, bsize, &min_energy,
3827 must_split = (min_energy < -3) && (max_energy - min_energy > 2);
3830 // Determine partition types in search according to the speed features.
3831 // The threshold set here has to be of square block size.
3832 if (cpi->sf.auto_min_max_partition_size) {
3833 partition_none_allowed &= (bsize <= max_size);
3834 partition_horz_allowed &=
3835 ((bsize <= max_size && bsize > min_size) || force_horz_split);
3836 partition_vert_allowed &=
3837 ((bsize <= max_size && bsize > min_size) || force_vert_split);
3838 do_split &= bsize > min_size;
3841 if (cpi->sf.use_square_partition_only &&
3842 (bsize > cpi->sf.use_square_only_thresh_high ||
3843 bsize < cpi->sf.use_square_only_thresh_low)) {
3845 if (!vp9_active_h_edge(cpi, mi_row, mi_step) || x->e_mbd.lossless)
3846 partition_horz_allowed &= force_horz_split;
3847 if (!vp9_active_v_edge(cpi, mi_row, mi_step) || x->e_mbd.lossless)
3848 partition_vert_allowed &= force_vert_split;
3850 partition_horz_allowed &= force_horz_split;
3851 partition_vert_allowed &= force_vert_split;
3855 save_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
3857 #if CONFIG_FP_MB_STATS
3858 if (cpi->use_fp_mb_stats) {
3859 set_offsets(cpi, tile_info, x, mi_row, mi_col, bsize);
3860 src_diff_var = get_sby_perpixel_diff_variance(cpi, &x->plane[0].src, mi_row,
3865 #if CONFIG_FP_MB_STATS
3866 // Decide whether we shall split directly and skip searching NONE by using
3867 // the first pass block statistics
3868 if (cpi->use_fp_mb_stats && bsize >= BLOCK_32X32 && do_split &&
3869 partition_none_allowed && src_diff_var > 4 &&
3870 cm->base_qindex < qindex_split_threshold_lookup[bsize]) {
3871 int mb_row = mi_row >> 1;
3872 int mb_col = mi_col >> 1;
3874 VPXMIN(mb_row + num_16x16_blocks_high_lookup[bsize], cm->mb_rows);
3876 VPXMIN(mb_col + num_16x16_blocks_wide_lookup[bsize], cm->mb_cols);
3879 // compute a complexity measure, basically measure inconsistency of motion
3880 // vectors obtained from the first pass in the current block
3881 for (r = mb_row; r < mb_row_end; r++) {
3882 for (c = mb_col; c < mb_col_end; c++) {
3883 const int mb_index = r * cm->mb_cols + c;
3885 MOTION_DIRECTION this_mv;
3886 MOTION_DIRECTION right_mv;
3887 MOTION_DIRECTION bottom_mv;
3890 get_motion_direction_fp(cpi->twopass.this_frame_mb_stats[mb_index]);
3893 if (c != mb_col_end - 1) {
3894 right_mv = get_motion_direction_fp(
3895 cpi->twopass.this_frame_mb_stats[mb_index + 1]);
3896 none_complexity += get_motion_inconsistency(this_mv, right_mv);
3900 if (r != mb_row_end - 1) {
3901 bottom_mv = get_motion_direction_fp(
3902 cpi->twopass.this_frame_mb_stats[mb_index + cm->mb_cols]);
3903 none_complexity += get_motion_inconsistency(this_mv, bottom_mv);
3906 // do not count its left and top neighbors to avoid double counting
3910 if (none_complexity > complexity_16x16_blocks_threshold[bsize]) {
3911 partition_none_allowed = 0;
3916 pc_tree->partitioning = PARTITION_NONE;
3918 if (cpi->sf.rd_ml_partition.var_pruning && !frame_is_intra_only(cm)) {
3919 const int do_rd_ml_partition_var_pruning =
3920 partition_none_allowed && do_split &&
3921 mi_row + num_8x8_blocks_high_lookup[bsize] <= cm->mi_rows &&
3922 mi_col + num_8x8_blocks_wide_lookup[bsize] <= cm->mi_cols;
3923 if (do_rd_ml_partition_var_pruning) {
3924 ml_predict_var_rd_paritioning(cpi, x, pc_tree, bsize, mi_row, mi_col,
3925 &partition_none_allowed, &do_split);
3927 vp9_zero(pc_tree->mv);
3929 if (bsize > BLOCK_8X8) { // Store MV result as reference for subblocks.
3930 for (i = 0; i < 4; ++i) pc_tree->split[i]->mv = pc_tree->mv;
3935 if (partition_none_allowed) {
3936 rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &this_rdc, bsize, ctx,
3938 ctx->rdcost = this_rdc.rdcost;
3939 if (this_rdc.rate != INT_MAX) {
3940 if (cpi->sf.prune_ref_frame_for_rect_partitions) {
3941 const int ref1 = ctx->mic.ref_frame[0];
3942 const int ref2 = ctx->mic.ref_frame[1];
3943 for (i = 0; i < 4; ++i) {
3944 ref_frames_used[i] |= (1 << ref1);
3945 if (ref2 > 0) ref_frames_used[i] |= (1 << ref2);
3948 if (bsize >= BLOCK_8X8) {
3949 this_rdc.rdcost += RDCOST(partition_mul, x->rddiv,
3950 cpi->partition_cost[pl][PARTITION_NONE], 0);
3951 this_rdc.rate += cpi->partition_cost[pl][PARTITION_NONE];
3954 if (this_rdc.rdcost < best_rdc.rdcost) {
3955 MODE_INFO *mi = xd->mi[0];
3957 best_rdc = this_rdc;
3958 if (bsize >= BLOCK_8X8) pc_tree->partitioning = PARTITION_NONE;
3960 if (cpi->sf.rd_ml_partition.search_early_termination) {
3961 // Currently, the machine-learning based partition search early
3962 // termination is only used while bsize is 16x16, 32x32 or 64x64,
3963 // VPXMIN(cm->width, cm->height) >= 480, and speed = 0.
3964 if (!x->e_mbd.lossless &&
3965 !segfeature_active(&cm->seg, mi->segment_id, SEG_LVL_SKIP) &&
3966 ctx->mic.mode >= INTRA_MODES && bsize >= BLOCK_16X16) {
3967 if (ml_pruning_partition(cm, xd, ctx, mi_row, mi_col, bsize)) {
3974 if ((do_split || do_rect) && !x->e_mbd.lossless && ctx->skippable) {
3975 const int use_ml_based_breakout =
3976 cpi->sf.rd_ml_partition.search_breakout && cm->base_qindex >= 100;
3977 if (use_ml_based_breakout) {
3978 if (ml_predict_breakout(cpi, bsize, x, &this_rdc)) {
3983 if (!cpi->sf.rd_ml_partition.search_early_termination) {
3984 if ((best_rdc.dist < (dist_breakout_thr >> 2)) ||
3985 (best_rdc.dist < dist_breakout_thr &&
3986 best_rdc.rate < rate_breakout_thr)) {
3994 #if CONFIG_FP_MB_STATS
3995 // Check if every 16x16 first pass block statistics has zero
3996 // motion and the corresponding first pass residue is small enough.
3997 // If that is the case, check the difference variance between the
3998 // current frame and the last frame. If the variance is small enough,
3999 // stop further splitting in RD optimization
4000 if (cpi->use_fp_mb_stats && do_split != 0 &&
4001 cm->base_qindex > qindex_skip_threshold_lookup[bsize]) {
4002 int mb_row = mi_row >> 1;
4003 int mb_col = mi_col >> 1;
4005 VPXMIN(mb_row + num_16x16_blocks_high_lookup[bsize], cm->mb_rows);
4007 VPXMIN(mb_col + num_16x16_blocks_wide_lookup[bsize], cm->mb_cols);
4011 for (r = mb_row; r < mb_row_end; r++) {
4012 for (c = mb_col; c < mb_col_end; c++) {
4013 const int mb_index = r * cm->mb_cols + c;
4014 if (!(cpi->twopass.this_frame_mb_stats[mb_index] &
4015 FPMB_MOTION_ZERO_MASK) ||
4016 !(cpi->twopass.this_frame_mb_stats[mb_index] &
4017 FPMB_ERROR_SMALL_MASK)) {
4028 if (src_diff_var == UINT_MAX) {
4029 set_offsets(cpi, tile_info, x, mi_row, mi_col, bsize);
4030 src_diff_var = get_sby_perpixel_diff_variance(
4031 cpi, &x->plane[0].src, mi_row, mi_col, bsize);
4033 if (src_diff_var < 8) {
4042 restore_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
4044 vp9_zero(ctx->pred_mv);
4045 ctx->mic.interp_filter = EIGHTTAP;
4048 // store estimated motion vector
4049 store_pred_mv(x, ctx);
4051 // If the interp_filter is marked as SWITCHABLE_FILTERS, it was for an
4052 // intra block and used for context purposes.
4053 if (ctx->mic.interp_filter == SWITCHABLE_FILTERS) {
4054 pred_interp_filter = EIGHTTAP;
4056 pred_interp_filter = ctx->mic.interp_filter;
4060 // TODO(jingning): use the motion vectors given by the above search as
4061 // the starting point of motion search in the following partition type check.
4062 pc_tree->split[0]->none.rdcost = 0;
4063 pc_tree->split[1]->none.rdcost = 0;
4064 pc_tree->split[2]->none.rdcost = 0;
4065 pc_tree->split[3]->none.rdcost = 0;
4066 if (do_split || must_split) {
4067 subsize = get_subsize(bsize, PARTITION_SPLIT);
4068 load_pred_mv(x, ctx);
4069 if (bsize == BLOCK_8X8) {
4071 if (cpi->sf.adaptive_pred_interp_filter && partition_none_allowed)
4072 pc_tree->leaf_split[0]->pred_interp_filter = pred_interp_filter;
4073 rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &sum_rdc, subsize,
4074 pc_tree->leaf_split[0], best_rdc.rdcost);
4075 if (sum_rdc.rate == INT_MAX) {
4076 sum_rdc.rdcost = INT64_MAX;
4078 if (cpi->sf.prune_ref_frame_for_rect_partitions) {
4079 const int ref1 = pc_tree->leaf_split[0]->mic.ref_frame[0];
4080 const int ref2 = pc_tree->leaf_split[0]->mic.ref_frame[1];
4081 for (i = 0; i < 4; ++i) {
4082 ref_frames_used[i] |= (1 << ref1);
4083 if (ref2 > 0) ref_frames_used[i] |= (1 << ref2);
4088 for (i = 0; (i < 4) && ((sum_rdc.rdcost < best_rdc.rdcost) || must_split);
4090 const int x_idx = (i & 1) * mi_step;
4091 const int y_idx = (i >> 1) * mi_step;
4093 if (mi_row + y_idx >= cm->mi_rows || mi_col + x_idx >= cm->mi_cols)
4096 pc_tree->split[i]->index = i;
4097 if (cpi->sf.prune_ref_frame_for_rect_partitions)
4098 pc_tree->split[i]->none.rate = INT_MAX;
4099 rd_pick_partition(cpi, td, tile_data, tp, mi_row + y_idx,
4100 mi_col + x_idx, subsize, &this_rdc,
4101 // A must split test here increases the number of sub
4102 // partitions but hurts metrics results quite a bit,
4103 // so this extra test is commented out pending
4104 // further tests on whether it adds much in terms of
4106 // (must_split) ? best_rdc.rdcost
4107 // : best_rdc.rdcost - sum_rdc.rdcost,
4108 best_rdc.rdcost - sum_rdc.rdcost, pc_tree->split[i]);
4110 if (this_rdc.rate == INT_MAX) {
4111 sum_rdc.rdcost = INT64_MAX;
4114 if (cpi->sf.prune_ref_frame_for_rect_partitions &&
4115 pc_tree->split[i]->none.rate != INT_MAX) {
4116 const int ref1 = pc_tree->split[i]->none.mic.ref_frame[0];
4117 const int ref2 = pc_tree->split[i]->none.mic.ref_frame[1];
4118 ref_frames_used[i] |= (1 << ref1);
4119 if (ref2 > 0) ref_frames_used[i] |= (1 << ref2);
4121 sum_rdc.rate += this_rdc.rate;
4122 sum_rdc.dist += this_rdc.dist;
4123 sum_rdc.rdcost += this_rdc.rdcost;
4128 if (((sum_rdc.rdcost < best_rdc.rdcost) || must_split) && i == 4) {
4129 sum_rdc.rdcost += RDCOST(partition_mul, x->rddiv,
4130 cpi->partition_cost[pl][PARTITION_SPLIT], 0);
4131 sum_rdc.rate += cpi->partition_cost[pl][PARTITION_SPLIT];
4133 if ((sum_rdc.rdcost < best_rdc.rdcost) ||
4134 (must_split && (sum_rdc.dist < best_rdc.dist))) {
4136 pc_tree->partitioning = PARTITION_SPLIT;
4138 // Rate and distortion based partition search termination clause.
4139 if (!cpi->sf.rd_ml_partition.search_early_termination &&
4140 !x->e_mbd.lossless &&
4141 ((best_rdc.dist < (dist_breakout_thr >> 2)) ||
4142 (best_rdc.dist < dist_breakout_thr &&
4143 best_rdc.rate < rate_breakout_thr))) {
4148 // skip rectangular partition test when larger block size
4149 // gives better rd cost
4150 if (cpi->sf.less_rectangular_check &&
4151 (bsize > cpi->sf.use_square_only_thresh_high ||
4152 best_rdc.dist < dist_breakout_thr))
4153 do_rect &= !partition_none_allowed;
4155 restore_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
4158 pc_tree->horizontal[0].skip_ref_frame_mask = 0;
4159 pc_tree->horizontal[1].skip_ref_frame_mask = 0;
4160 pc_tree->vertical[0].skip_ref_frame_mask = 0;
4161 pc_tree->vertical[1].skip_ref_frame_mask = 0;
4162 if (cpi->sf.prune_ref_frame_for_rect_partitions) {
4163 uint8_t used_frames;
4164 used_frames = ref_frames_used[0] | ref_frames_used[1];
4165 if (used_frames) pc_tree->horizontal[0].skip_ref_frame_mask = ~used_frames;
4166 used_frames = ref_frames_used[2] | ref_frames_used[3];
4167 if (used_frames) pc_tree->horizontal[1].skip_ref_frame_mask = ~used_frames;
4168 used_frames = ref_frames_used[0] | ref_frames_used[2];
4169 if (used_frames) pc_tree->vertical[0].skip_ref_frame_mask = ~used_frames;
4170 used_frames = ref_frames_used[1] | ref_frames_used[3];
4171 if (used_frames) pc_tree->vertical[1].skip_ref_frame_mask = ~used_frames;
4175 const int do_ml_rect_partition_pruning =
4176 !frame_is_intra_only(cm) && !force_horz_split && !force_vert_split &&
4177 (partition_horz_allowed || partition_vert_allowed) && bsize > BLOCK_8X8;
4178 if (do_ml_rect_partition_pruning) {
4179 ml_prune_rect_partition(cpi, x, bsize, pc_tree, &partition_horz_allowed,
4180 &partition_vert_allowed, best_rdc.rdcost);
4185 if (partition_horz_allowed &&
4186 (do_rect || vp9_active_h_edge(cpi, mi_row, mi_step))) {
4187 const int part_mode_rate = cpi->partition_cost[pl][PARTITION_HORZ];
4188 const int64_t part_mode_rdcost =
4189 RDCOST(partition_mul, x->rddiv, part_mode_rate, 0);
4190 subsize = get_subsize(bsize, PARTITION_HORZ);
4191 load_pred_mv(x, ctx);
4192 if (cpi->sf.adaptive_pred_interp_filter && bsize == BLOCK_8X8 &&
4193 partition_none_allowed)
4194 pc_tree->horizontal[0].pred_interp_filter = pred_interp_filter;
4195 rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &sum_rdc, subsize,
4196 &pc_tree->horizontal[0],
4197 best_rdc.rdcost - part_mode_rdcost);
4198 if (sum_rdc.rdcost < INT64_MAX) {
4199 sum_rdc.rdcost += part_mode_rdcost;
4200 sum_rdc.rate += part_mode_rate;
4203 if (sum_rdc.rdcost < best_rdc.rdcost && mi_row + mi_step < cm->mi_rows &&
4204 bsize > BLOCK_8X8) {
4205 PICK_MODE_CONTEXT *ctx = &pc_tree->horizontal[0];
4206 update_state(cpi, td, ctx, mi_row, mi_col, subsize, 0);
4207 encode_superblock(cpi, td, tp, 0, mi_row, mi_col, subsize, ctx);
4208 if (cpi->sf.adaptive_pred_interp_filter && bsize == BLOCK_8X8 &&
4209 partition_none_allowed)
4210 pc_tree->horizontal[1].pred_interp_filter = pred_interp_filter;
4211 rd_pick_sb_modes(cpi, tile_data, x, mi_row + mi_step, mi_col, &this_rdc,
4212 subsize, &pc_tree->horizontal[1],
4213 best_rdc.rdcost - sum_rdc.rdcost);
4214 if (this_rdc.rate == INT_MAX) {
4215 sum_rdc.rdcost = INT64_MAX;
4217 sum_rdc.rate += this_rdc.rate;
4218 sum_rdc.dist += this_rdc.dist;
4219 sum_rdc.rdcost += this_rdc.rdcost;
4223 if (sum_rdc.rdcost < best_rdc.rdcost) {
4225 pc_tree->partitioning = PARTITION_HORZ;
4227 if (cpi->sf.less_rectangular_check &&
4228 bsize > cpi->sf.use_square_only_thresh_high)
4231 restore_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
4235 if (partition_vert_allowed &&
4236 (do_rect || vp9_active_v_edge(cpi, mi_col, mi_step))) {
4237 const int part_mode_rate = cpi->partition_cost[pl][PARTITION_VERT];
4238 const int64_t part_mode_rdcost =
4239 RDCOST(partition_mul, x->rddiv, part_mode_rate, 0);
4240 subsize = get_subsize(bsize, PARTITION_VERT);
4241 load_pred_mv(x, ctx);
4242 if (cpi->sf.adaptive_pred_interp_filter && bsize == BLOCK_8X8 &&
4243 partition_none_allowed)
4244 pc_tree->vertical[0].pred_interp_filter = pred_interp_filter;
4245 rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &sum_rdc, subsize,
4246 &pc_tree->vertical[0], best_rdc.rdcost - part_mode_rdcost);
4247 if (sum_rdc.rdcost < INT64_MAX) {
4248 sum_rdc.rdcost += part_mode_rdcost;
4249 sum_rdc.rate += part_mode_rate;
4252 if (sum_rdc.rdcost < best_rdc.rdcost && mi_col + mi_step < cm->mi_cols &&
4253 bsize > BLOCK_8X8) {
4254 update_state(cpi, td, &pc_tree->vertical[0], mi_row, mi_col, subsize, 0);
4255 encode_superblock(cpi, td, tp, 0, mi_row, mi_col, subsize,
4256 &pc_tree->vertical[0]);
4257 if (cpi->sf.adaptive_pred_interp_filter && bsize == BLOCK_8X8 &&
4258 partition_none_allowed)
4259 pc_tree->vertical[1].pred_interp_filter = pred_interp_filter;
4260 rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col + mi_step, &this_rdc,
4261 subsize, &pc_tree->vertical[1],
4262 best_rdc.rdcost - sum_rdc.rdcost);
4263 if (this_rdc.rate == INT_MAX) {
4264 sum_rdc.rdcost = INT64_MAX;
4266 sum_rdc.rate += this_rdc.rate;
4267 sum_rdc.dist += this_rdc.dist;
4268 sum_rdc.rdcost += this_rdc.rdcost;
4272 if (sum_rdc.rdcost < best_rdc.rdcost) {
4274 pc_tree->partitioning = PARTITION_VERT;
4276 restore_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
4279 // TODO(jbb): This code added so that we avoid static analysis
4280 // warning related to the fact that best_rd isn't used after this
4281 // point. This code should be refactored so that the duplicate
4282 // checks occur in some sub function and thus are used...
4284 *rd_cost = best_rdc;
4286 if (best_rdc.rate < INT_MAX && best_rdc.dist < INT64_MAX &&
4287 pc_tree->index != 3) {
4288 int output_enabled = (bsize == BLOCK_64X64);
4289 encode_sb(cpi, td, tile_info, tp, mi_row, mi_col, output_enabled, bsize,
4293 if (bsize == BLOCK_64X64) {
4294 assert(tp_orig < *tp);
4295 assert(best_rdc.rate < INT_MAX);
4296 assert(best_rdc.dist < INT64_MAX);
4298 assert(tp_orig == *tp);
4302 static void encode_rd_sb_row(VP9_COMP *cpi, ThreadData *td,
4303 TileDataEnc *tile_data, int mi_row,
4305 VP9_COMMON *const cm = &cpi->common;
4306 TileInfo *const tile_info = &tile_data->tile_info;
4307 MACROBLOCK *const x = &td->mb;
4308 MACROBLOCKD *const xd = &x->e_mbd;
4309 SPEED_FEATURES *const sf = &cpi->sf;
4310 const int mi_col_start = tile_info->mi_col_start;
4311 const int mi_col_end = tile_info->mi_col_end;
4313 const int sb_row = mi_row >> MI_BLOCK_SIZE_LOG2;
4314 const int num_sb_cols =
4315 get_num_cols(tile_data->tile_info, MI_BLOCK_SIZE_LOG2);
4318 // Initialize the left context for the new SB row
4319 memset(&xd->left_context, 0, sizeof(xd->left_context));
4320 memset(xd->left_seg_context, 0, sizeof(xd->left_seg_context));
4322 // Code each SB in the row
4323 for (mi_col = mi_col_start, sb_col_in_tile = 0; mi_col < mi_col_end;
4324 mi_col += MI_BLOCK_SIZE, sb_col_in_tile++) {
4325 const struct segmentation *const seg = &cm->seg;
4331 int orig_rdmult = cpi->rd.RDMULT;
4333 const int idx_str = cm->mi_stride * mi_row + mi_col;
4334 MODE_INFO **mi = cm->mi_grid_visible + idx_str;
4336 (*(cpi->row_mt_sync_read_ptr))(&tile_data->row_mt_sync, sb_row,
4339 if (sf->adaptive_pred_interp_filter) {
4340 for (i = 0; i < 64; ++i) td->leaf_tree[i].pred_interp_filter = SWITCHABLE;
4342 for (i = 0; i < 64; ++i) {
4343 td->pc_tree[i].vertical[0].pred_interp_filter = SWITCHABLE;
4344 td->pc_tree[i].vertical[1].pred_interp_filter = SWITCHABLE;
4345 td->pc_tree[i].horizontal[0].pred_interp_filter = SWITCHABLE;
4346 td->pc_tree[i].horizontal[1].pred_interp_filter = SWITCHABLE;
4350 for (i = 0; i < MAX_REF_FRAMES; ++i) {
4351 x->pred_mv[i].row = INT16_MAX;
4352 x->pred_mv[i].col = INT16_MAX;
4354 td->pc_root->index = 0;
4357 const uint8_t *const map =
4358 seg->update_map ? cpi->segmentation_map : cm->last_frame_seg_map;
4359 int segment_id = get_segment_id(cm, map, BLOCK_64X64, mi_row, mi_col);
4360 seg_skip = segfeature_active(seg, segment_id, SEG_LVL_SKIP);
4363 x->source_variance = UINT_MAX;
4365 x->cb_rdmult = orig_rdmult;
4367 if (sf->partition_search_type == FIXED_PARTITION || seg_skip) {
4368 const BLOCK_SIZE bsize =
4369 seg_skip ? BLOCK_64X64 : sf->always_this_block_size;
4370 set_offsets(cpi, tile_info, x, mi_row, mi_col, BLOCK_64X64);
4371 set_fixed_partitioning(cpi, tile_info, mi, mi_row, mi_col, bsize);
4372 rd_use_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col, BLOCK_64X64,
4373 &dummy_rate, &dummy_dist, 1, td->pc_root);
4374 } else if (cpi->partition_search_skippable_frame) {
4376 set_offsets(cpi, tile_info, x, mi_row, mi_col, BLOCK_64X64);
4377 bsize = get_rd_var_based_fixed_partition(cpi, x, mi_row, mi_col);
4378 set_fixed_partitioning(cpi, tile_info, mi, mi_row, mi_col, bsize);
4379 rd_use_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col, BLOCK_64X64,
4380 &dummy_rate, &dummy_dist, 1, td->pc_root);
4381 } else if (sf->partition_search_type == VAR_BASED_PARTITION &&
4382 cm->frame_type != KEY_FRAME) {
4383 choose_partitioning(cpi, tile_info, x, mi_row, mi_col);
4384 rd_use_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col, BLOCK_64X64,
4385 &dummy_rate, &dummy_dist, 1, td->pc_root);
4387 if (cpi->twopass.gf_group.index > 0 && cpi->sf.enable_tpl_model) {
4389 get_rdmult_delta(cpi, BLOCK_64X64, mi_row, mi_col, orig_rdmult);
4393 if (cpi->sf.enable_wiener_variance && cm->show_frame) {
4394 x->segment_id = wiener_var_segment(cpi, BLOCK_64X64, mi_row, mi_col);
4395 x->cb_rdmult = vp9_compute_rd_mult(
4396 cpi, vp9_get_qindex(&cm->seg, x->segment_id, cm->base_qindex));
4399 // If required set upper and lower partition size limits
4400 if (sf->auto_min_max_partition_size) {
4401 set_offsets(cpi, tile_info, x, mi_row, mi_col, BLOCK_64X64);
4402 rd_auto_partition_range(cpi, tile_info, xd, mi_row, mi_col,
4403 &x->min_partition_size, &x->max_partition_size);
4405 td->pc_root->none.rdcost = 0;
4406 rd_pick_partition(cpi, td, tile_data, tp, mi_row, mi_col, BLOCK_64X64,
4407 &dummy_rdc, INT64_MAX, td->pc_root);
4409 (*(cpi->row_mt_sync_write_ptr))(&tile_data->row_mt_sync, sb_row,
4410 sb_col_in_tile, num_sb_cols);
4414 static void init_encode_frame_mb_context(VP9_COMP *cpi) {
4415 MACROBLOCK *const x = &cpi->td.mb;
4416 VP9_COMMON *const cm = &cpi->common;
4417 MACROBLOCKD *const xd = &x->e_mbd;
4418 const int aligned_mi_cols = mi_cols_aligned_to_sb(cm->mi_cols);
4420 // Copy data over into macro block data structures.
4421 vp9_setup_src_planes(x, cpi->Source, 0, 0);
4423 vp9_setup_block_planes(&x->e_mbd, cm->subsampling_x, cm->subsampling_y);
4425 // Note: this memset assumes above_context[0], [1] and [2]
4426 // are allocated as part of the same buffer.
4427 memset(xd->above_context[0], 0,
4428 sizeof(*xd->above_context[0]) * 2 * aligned_mi_cols * MAX_MB_PLANE);
4429 memset(xd->above_seg_context, 0,
4430 sizeof(*xd->above_seg_context) * aligned_mi_cols);
4433 static int check_dual_ref_flags(VP9_COMP *cpi) {
4434 const int ref_flags = cpi->ref_frame_flags;
4436 if (segfeature_active(&cpi->common.seg, 1, SEG_LVL_REF_FRAME)) {
4439 return (!!(ref_flags & VP9_GOLD_FLAG) + !!(ref_flags & VP9_LAST_FLAG) +
4440 !!(ref_flags & VP9_ALT_FLAG)) >= 2;
4444 static void reset_skip_tx_size(VP9_COMMON *cm, TX_SIZE max_tx_size) {
4446 const int mis = cm->mi_stride;
4447 MODE_INFO **mi_ptr = cm->mi_grid_visible;
4449 for (mi_row = 0; mi_row < cm->mi_rows; ++mi_row, mi_ptr += mis) {
4450 for (mi_col = 0; mi_col < cm->mi_cols; ++mi_col) {
4451 if (mi_ptr[mi_col]->tx_size > max_tx_size)
4452 mi_ptr[mi_col]->tx_size = max_tx_size;
4457 static MV_REFERENCE_FRAME get_frame_type(const VP9_COMP *cpi) {
4458 if (frame_is_intra_only(&cpi->common))
4460 else if (cpi->rc.is_src_frame_alt_ref && cpi->refresh_golden_frame)
4461 return ALTREF_FRAME;
4462 else if (cpi->refresh_golden_frame || cpi->refresh_alt_ref_frame)
4463 return GOLDEN_FRAME;
4468 static TX_MODE select_tx_mode(const VP9_COMP *cpi, MACROBLOCKD *const xd) {
4469 if (xd->lossless) return ONLY_4X4;
4470 if (cpi->common.frame_type == KEY_FRAME && cpi->sf.use_nonrd_pick_mode)
4472 if (cpi->sf.tx_size_search_method == USE_LARGESTALL)
4474 else if (cpi->sf.tx_size_search_method == USE_FULL_RD ||
4475 cpi->sf.tx_size_search_method == USE_TX_8X8)
4476 return TX_MODE_SELECT;
4478 return cpi->common.tx_mode;
4481 static void hybrid_intra_mode_search(VP9_COMP *cpi, MACROBLOCK *const x,
4482 RD_COST *rd_cost, BLOCK_SIZE bsize,
4483 PICK_MODE_CONTEXT *ctx) {
4484 if (!cpi->sf.nonrd_keyframe && bsize < BLOCK_16X16)
4485 vp9_rd_pick_intra_mode_sb(cpi, x, rd_cost, bsize, ctx, INT64_MAX);
4487 vp9_pick_intra_mode(cpi, x, rd_cost, bsize, ctx);
4490 static void hybrid_search_svc_baseiskey(VP9_COMP *cpi, MACROBLOCK *const x,
4491 RD_COST *rd_cost, BLOCK_SIZE bsize,
4492 PICK_MODE_CONTEXT *ctx,
4493 TileDataEnc *tile_data, int mi_row,
4495 if (!cpi->sf.nonrd_keyframe && bsize <= BLOCK_8X8) {
4496 vp9_rd_pick_intra_mode_sb(cpi, x, rd_cost, bsize, ctx, INT64_MAX);
4498 if (cpi->svc.disable_inter_layer_pred == INTER_LAYER_PRED_OFF)
4499 vp9_pick_intra_mode(cpi, x, rd_cost, bsize, ctx);
4500 else if (bsize >= BLOCK_8X8)
4501 vp9_pick_inter_mode(cpi, x, tile_data, mi_row, mi_col, rd_cost, bsize,
4504 vp9_pick_inter_mode_sub8x8(cpi, x, mi_row, mi_col, rd_cost, bsize, ctx);
4508 static void hybrid_search_scene_change(VP9_COMP *cpi, MACROBLOCK *const x,
4509 RD_COST *rd_cost, BLOCK_SIZE bsize,
4510 PICK_MODE_CONTEXT *ctx,
4511 TileDataEnc *tile_data, int mi_row,
4513 if (!cpi->sf.nonrd_keyframe && bsize <= BLOCK_8X8) {
4514 vp9_rd_pick_intra_mode_sb(cpi, x, rd_cost, bsize, ctx, INT64_MAX);
4516 vp9_pick_inter_mode(cpi, x, tile_data, mi_row, mi_col, rd_cost, bsize, ctx);
4520 static void nonrd_pick_sb_modes(VP9_COMP *cpi, TileDataEnc *tile_data,
4521 MACROBLOCK *const x, int mi_row, int mi_col,
4522 RD_COST *rd_cost, BLOCK_SIZE bsize,
4523 PICK_MODE_CONTEXT *ctx) {
4524 VP9_COMMON *const cm = &cpi->common;
4525 TileInfo *const tile_info = &tile_data->tile_info;
4526 MACROBLOCKD *const xd = &x->e_mbd;
4528 ENTROPY_CONTEXT l[16 * MAX_MB_PLANE], a[16 * MAX_MB_PLANE];
4529 BLOCK_SIZE bs = VPXMAX(bsize, BLOCK_8X8); // processing unit block size
4530 const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bs];
4531 const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bs];
4534 set_offsets(cpi, tile_info, x, mi_row, mi_col, bsize);
4536 set_segment_index(cpi, x, mi_row, mi_col, bsize, 0);
4539 mi->sb_type = bsize;
4541 for (plane = 0; plane < MAX_MB_PLANE; ++plane) {
4542 struct macroblockd_plane *pd = &xd->plane[plane];
4543 memcpy(a + num_4x4_blocks_wide * plane, pd->above_context,
4544 (sizeof(a[0]) * num_4x4_blocks_wide) >> pd->subsampling_x);
4545 memcpy(l + num_4x4_blocks_high * plane, pd->left_context,
4546 (sizeof(l[0]) * num_4x4_blocks_high) >> pd->subsampling_y);
4549 if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ && cm->seg.enabled)
4550 if (cyclic_refresh_segment_id_boosted(mi->segment_id))
4551 x->rdmult = vp9_cyclic_refresh_get_rdmult(cpi->cyclic_refresh);
4553 if (frame_is_intra_only(cm))
4554 hybrid_intra_mode_search(cpi, x, rd_cost, bsize, ctx);
4555 else if (cpi->svc.layer_context[cpi->svc.temporal_layer_id].is_key_frame)
4556 hybrid_search_svc_baseiskey(cpi, x, rd_cost, bsize, ctx, tile_data, mi_row,
4558 else if (segfeature_active(&cm->seg, mi->segment_id, SEG_LVL_SKIP))
4559 set_mode_info_seg_skip(x, cm->tx_mode, rd_cost, bsize);
4560 else if (bsize >= BLOCK_8X8) {
4561 if (cpi->rc.hybrid_intra_scene_change)
4562 hybrid_search_scene_change(cpi, x, rd_cost, bsize, ctx, tile_data, mi_row,
4565 vp9_pick_inter_mode(cpi, x, tile_data, mi_row, mi_col, rd_cost, bsize,
4568 vp9_pick_inter_mode_sub8x8(cpi, x, mi_row, mi_col, rd_cost, bsize, ctx);
4571 duplicate_mode_info_in_sb(cm, xd, mi_row, mi_col, bsize);
4573 for (plane = 0; plane < MAX_MB_PLANE; ++plane) {
4574 struct macroblockd_plane *pd = &xd->plane[plane];
4575 memcpy(pd->above_context, a + num_4x4_blocks_wide * plane,
4576 (sizeof(a[0]) * num_4x4_blocks_wide) >> pd->subsampling_x);
4577 memcpy(pd->left_context, l + num_4x4_blocks_high * plane,
4578 (sizeof(l[0]) * num_4x4_blocks_high) >> pd->subsampling_y);
4581 if (rd_cost->rate == INT_MAX) vp9_rd_cost_reset(rd_cost);
4583 ctx->rate = rd_cost->rate;
4584 ctx->dist = rd_cost->dist;
4587 static void fill_mode_info_sb(VP9_COMMON *cm, MACROBLOCK *x, int mi_row,
4588 int mi_col, BLOCK_SIZE bsize, PC_TREE *pc_tree) {
4589 MACROBLOCKD *xd = &x->e_mbd;
4590 int bsl = b_width_log2_lookup[bsize], hbs = (1 << bsl) / 4;
4591 PARTITION_TYPE partition = pc_tree->partitioning;
4592 BLOCK_SIZE subsize = get_subsize(bsize, partition);
4594 assert(bsize >= BLOCK_8X8);
4596 if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols) return;
4598 switch (partition) {
4599 case PARTITION_NONE:
4600 set_mode_info_offsets(cm, x, xd, mi_row, mi_col);
4601 *(xd->mi[0]) = pc_tree->none.mic;
4602 *(x->mbmi_ext) = pc_tree->none.mbmi_ext;
4603 duplicate_mode_info_in_sb(cm, xd, mi_row, mi_col, bsize);
4605 case PARTITION_VERT:
4606 set_mode_info_offsets(cm, x, xd, mi_row, mi_col);
4607 *(xd->mi[0]) = pc_tree->vertical[0].mic;
4608 *(x->mbmi_ext) = pc_tree->vertical[0].mbmi_ext;
4609 duplicate_mode_info_in_sb(cm, xd, mi_row, mi_col, subsize);
4611 if (mi_col + hbs < cm->mi_cols) {
4612 set_mode_info_offsets(cm, x, xd, mi_row, mi_col + hbs);
4613 *(xd->mi[0]) = pc_tree->vertical[1].mic;
4614 *(x->mbmi_ext) = pc_tree->vertical[1].mbmi_ext;
4615 duplicate_mode_info_in_sb(cm, xd, mi_row, mi_col + hbs, subsize);
4618 case PARTITION_HORZ:
4619 set_mode_info_offsets(cm, x, xd, mi_row, mi_col);
4620 *(xd->mi[0]) = pc_tree->horizontal[0].mic;
4621 *(x->mbmi_ext) = pc_tree->horizontal[0].mbmi_ext;
4622 duplicate_mode_info_in_sb(cm, xd, mi_row, mi_col, subsize);
4623 if (mi_row + hbs < cm->mi_rows) {
4624 set_mode_info_offsets(cm, x, xd, mi_row + hbs, mi_col);
4625 *(xd->mi[0]) = pc_tree->horizontal[1].mic;
4626 *(x->mbmi_ext) = pc_tree->horizontal[1].mbmi_ext;
4627 duplicate_mode_info_in_sb(cm, xd, mi_row + hbs, mi_col, subsize);
4630 case PARTITION_SPLIT: {
4631 fill_mode_info_sb(cm, x, mi_row, mi_col, subsize, pc_tree->split[0]);
4632 fill_mode_info_sb(cm, x, mi_row, mi_col + hbs, subsize,
4634 fill_mode_info_sb(cm, x, mi_row + hbs, mi_col, subsize,
4636 fill_mode_info_sb(cm, x, mi_row + hbs, mi_col + hbs, subsize,
4644 // Reset the prediction pixel ready flag recursively.
4645 static void pred_pixel_ready_reset(PC_TREE *pc_tree, BLOCK_SIZE bsize) {
4646 pc_tree->none.pred_pixel_ready = 0;
4647 pc_tree->horizontal[0].pred_pixel_ready = 0;
4648 pc_tree->horizontal[1].pred_pixel_ready = 0;
4649 pc_tree->vertical[0].pred_pixel_ready = 0;
4650 pc_tree->vertical[1].pred_pixel_ready = 0;
4652 if (bsize > BLOCK_8X8) {
4653 BLOCK_SIZE subsize = get_subsize(bsize, PARTITION_SPLIT);
4655 for (i = 0; i < 4; ++i) pred_pixel_ready_reset(pc_tree->split[i], subsize);
4661 static int ml_predict_var_paritioning(VP9_COMP *cpi, MACROBLOCK *x,
4662 BLOCK_SIZE bsize, int mi_row,
4664 VP9_COMMON *const cm = &cpi->common;
4665 const NN_CONFIG *nn_config = NULL;
4668 case BLOCK_64X64: nn_config = &vp9_var_part_nnconfig_64; break;
4669 case BLOCK_32X32: nn_config = &vp9_var_part_nnconfig_32; break;
4670 case BLOCK_16X16: nn_config = &vp9_var_part_nnconfig_16; break;
4671 case BLOCK_8X8: break;
4672 default: assert(0 && "Unexpected block size."); return -1;
4675 if (!nn_config) return -1;
4677 vpx_clear_system_state();
4680 const float thresh = cpi->oxcf.speed <= 5 ? 1.25f : 0.0f;
4681 float features[FEATURES] = { 0.0f };
4682 const int dc_q = vp9_dc_quant(cm->base_qindex, 0, cm->bit_depth);
4683 int feature_idx = 0;
4684 float score[LABELS];
4686 features[feature_idx++] = logf((float)(dc_q * dc_q) / 256.0f + 1.0f);
4687 vp9_setup_src_planes(x, cpi->Source, mi_row, mi_col);
4689 const int bs = 4 * num_4x4_blocks_wide_lookup[bsize];
4690 const BLOCK_SIZE subsize = get_subsize(bsize, PARTITION_SPLIT);
4691 const int sb_offset_row = 8 * (mi_row & 7);
4692 const int sb_offset_col = 8 * (mi_col & 7);
4693 const uint8_t *pred = x->est_pred + sb_offset_row * 64 + sb_offset_col;
4694 const uint8_t *src = x->plane[0].src.buf;
4695 const int src_stride = x->plane[0].src.stride;
4696 const int pred_stride = 64;
4699 // Variance of whole block.
4700 const unsigned int var =
4701 cpi->fn_ptr[bsize].vf(src, src_stride, pred, pred_stride, &sse);
4702 const float factor = (var == 0) ? 1.0f : (1.0f / (float)var);
4704 features[feature_idx++] = logf((float)var + 1.0f);
4705 for (i = 0; i < 4; ++i) {
4706 const int x_idx = (i & 1) * bs / 2;
4707 const int y_idx = (i >> 1) * bs / 2;
4708 const int src_offset = y_idx * src_stride + x_idx;
4709 const int pred_offset = y_idx * pred_stride + x_idx;
4710 // Variance of quarter block.
4711 const unsigned int sub_var =
4712 cpi->fn_ptr[subsize].vf(src + src_offset, src_stride,
4713 pred + pred_offset, pred_stride, &sse);
4714 const float var_ratio = (var == 0) ? 1.0f : factor * (float)sub_var;
4715 features[feature_idx++] = var_ratio;
4719 assert(feature_idx == FEATURES);
4720 nn_predict(features, nn_config, score);
4721 if (score[0] > thresh) return PARTITION_SPLIT;
4722 if (score[0] < -thresh) return PARTITION_NONE;
4729 static void nonrd_pick_partition(VP9_COMP *cpi, ThreadData *td,
4730 TileDataEnc *tile_data, TOKENEXTRA **tp,
4731 int mi_row, int mi_col, BLOCK_SIZE bsize,
4732 RD_COST *rd_cost, int do_recon,
4733 int64_t best_rd, PC_TREE *pc_tree) {
4734 const SPEED_FEATURES *const sf = &cpi->sf;
4735 VP9_COMMON *const cm = &cpi->common;
4736 TileInfo *const tile_info = &tile_data->tile_info;
4737 MACROBLOCK *const x = &td->mb;
4738 MACROBLOCKD *const xd = &x->e_mbd;
4739 const int ms = num_8x8_blocks_wide_lookup[bsize] / 2;
4740 TOKENEXTRA *tp_orig = *tp;
4741 PICK_MODE_CONTEXT *ctx = &pc_tree->none;
4743 BLOCK_SIZE subsize = bsize;
4744 RD_COST this_rdc, sum_rdc, best_rdc;
4745 int do_split = bsize >= BLOCK_8X8;
4747 // Override skipping rectangular partition operations for edge blocks
4748 const int force_horz_split = (mi_row + ms >= cm->mi_rows);
4749 const int force_vert_split = (mi_col + ms >= cm->mi_cols);
4750 const int xss = x->e_mbd.plane[1].subsampling_x;
4751 const int yss = x->e_mbd.plane[1].subsampling_y;
4753 int partition_none_allowed = !force_horz_split && !force_vert_split;
4754 int partition_horz_allowed =
4755 !force_vert_split && yss <= xss && bsize >= BLOCK_8X8;
4756 int partition_vert_allowed =
4757 !force_horz_split && xss <= yss && bsize >= BLOCK_8X8;
4758 const int use_ml_based_partitioning =
4759 sf->partition_search_type == ML_BASED_PARTITION;
4763 // Avoid checking for rectangular partitions for speed >= 6.
4764 if (cpi->oxcf.speed >= 6) do_rect = 0;
4766 assert(num_8x8_blocks_wide_lookup[bsize] ==
4767 num_8x8_blocks_high_lookup[bsize]);
4769 vp9_rd_cost_init(&sum_rdc);
4770 vp9_rd_cost_reset(&best_rdc);
4771 best_rdc.rdcost = best_rd;
4773 // Determine partition types in search according to the speed features.
4774 // The threshold set here has to be of square block size.
4775 if (sf->auto_min_max_partition_size) {
4776 partition_none_allowed &=
4777 (bsize <= x->max_partition_size && bsize >= x->min_partition_size);
4778 partition_horz_allowed &=
4779 ((bsize <= x->max_partition_size && bsize > x->min_partition_size) ||
4781 partition_vert_allowed &=
4782 ((bsize <= x->max_partition_size && bsize > x->min_partition_size) ||
4784 do_split &= bsize > x->min_partition_size;
4786 if (sf->use_square_partition_only) {
4787 partition_horz_allowed &= force_horz_split;
4788 partition_vert_allowed &= force_vert_split;
4791 if (use_ml_based_partitioning) {
4792 if (partition_none_allowed || do_split) do_rect = 0;
4793 if (partition_none_allowed && do_split) {
4794 const int ml_predicted_partition =
4795 ml_predict_var_paritioning(cpi, x, bsize, mi_row, mi_col);
4796 if (ml_predicted_partition == PARTITION_NONE) do_split = 0;
4797 if (ml_predicted_partition == PARTITION_SPLIT) partition_none_allowed = 0;
4801 if (!partition_none_allowed && !do_split) do_rect = 1;
4803 ctx->pred_pixel_ready =
4804 !(partition_vert_allowed || partition_horz_allowed || do_split);
4807 if (partition_none_allowed) {
4808 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &this_rdc, bsize,
4810 ctx->mic = *xd->mi[0];
4811 ctx->mbmi_ext = *x->mbmi_ext;
4812 ctx->skip_txfm[0] = x->skip_txfm[0];
4813 ctx->skip = x->skip;
4815 if (this_rdc.rate != INT_MAX) {
4816 const int pl = partition_plane_context(xd, mi_row, mi_col, bsize);
4817 this_rdc.rate += cpi->partition_cost[pl][PARTITION_NONE];
4819 RDCOST(x->rdmult, x->rddiv, this_rdc.rate, this_rdc.dist);
4820 if (this_rdc.rdcost < best_rdc.rdcost) {
4821 best_rdc = this_rdc;
4822 if (bsize >= BLOCK_8X8) pc_tree->partitioning = PARTITION_NONE;
4824 if (!use_ml_based_partitioning) {
4825 int64_t dist_breakout_thr = sf->partition_search_breakout_thr.dist;
4826 int64_t rate_breakout_thr = sf->partition_search_breakout_thr.rate;
4827 dist_breakout_thr >>=
4828 8 - (b_width_log2_lookup[bsize] + b_height_log2_lookup[bsize]);
4829 rate_breakout_thr *= num_pels_log2_lookup[bsize];
4830 if (!x->e_mbd.lossless && this_rdc.rate < rate_breakout_thr &&
4831 this_rdc.dist < dist_breakout_thr) {
4840 // store estimated motion vector
4841 store_pred_mv(x, ctx);
4845 int pl = partition_plane_context(xd, mi_row, mi_col, bsize);
4846 sum_rdc.rate += cpi->partition_cost[pl][PARTITION_SPLIT];
4847 sum_rdc.rdcost = RDCOST(x->rdmult, x->rddiv, sum_rdc.rate, sum_rdc.dist);
4848 subsize = get_subsize(bsize, PARTITION_SPLIT);
4849 for (i = 0; i < 4 && sum_rdc.rdcost < best_rdc.rdcost; ++i) {
4850 const int x_idx = (i & 1) * ms;
4851 const int y_idx = (i >> 1) * ms;
4853 if (mi_row + y_idx >= cm->mi_rows || mi_col + x_idx >= cm->mi_cols)
4855 load_pred_mv(x, ctx);
4856 nonrd_pick_partition(cpi, td, tile_data, tp, mi_row + y_idx,
4857 mi_col + x_idx, subsize, &this_rdc, 0,
4858 best_rdc.rdcost - sum_rdc.rdcost, pc_tree->split[i]);
4860 if (this_rdc.rate == INT_MAX) {
4861 vp9_rd_cost_reset(&sum_rdc);
4863 sum_rdc.rate += this_rdc.rate;
4864 sum_rdc.dist += this_rdc.dist;
4865 sum_rdc.rdcost += this_rdc.rdcost;
4869 if (sum_rdc.rdcost < best_rdc.rdcost) {
4871 pc_tree->partitioning = PARTITION_SPLIT;
4873 // skip rectangular partition test when larger block size
4874 // gives better rd cost
4875 if (sf->less_rectangular_check) do_rect &= !partition_none_allowed;
4880 if (partition_horz_allowed && do_rect) {
4881 subsize = get_subsize(bsize, PARTITION_HORZ);
4882 load_pred_mv(x, ctx);
4883 pc_tree->horizontal[0].pred_pixel_ready = 1;
4884 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &sum_rdc, subsize,
4885 &pc_tree->horizontal[0]);
4887 pc_tree->horizontal[0].mic = *xd->mi[0];
4888 pc_tree->horizontal[0].mbmi_ext = *x->mbmi_ext;
4889 pc_tree->horizontal[0].skip_txfm[0] = x->skip_txfm[0];
4890 pc_tree->horizontal[0].skip = x->skip;
4892 if (sum_rdc.rdcost < best_rdc.rdcost && mi_row + ms < cm->mi_rows) {
4893 load_pred_mv(x, ctx);
4894 pc_tree->horizontal[1].pred_pixel_ready = 1;
4895 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row + ms, mi_col, &this_rdc,
4896 subsize, &pc_tree->horizontal[1]);
4898 pc_tree->horizontal[1].mic = *xd->mi[0];
4899 pc_tree->horizontal[1].mbmi_ext = *x->mbmi_ext;
4900 pc_tree->horizontal[1].skip_txfm[0] = x->skip_txfm[0];
4901 pc_tree->horizontal[1].skip = x->skip;
4903 if (this_rdc.rate == INT_MAX) {
4904 vp9_rd_cost_reset(&sum_rdc);
4906 int pl = partition_plane_context(xd, mi_row, mi_col, bsize);
4907 this_rdc.rate += cpi->partition_cost[pl][PARTITION_HORZ];
4908 sum_rdc.rate += this_rdc.rate;
4909 sum_rdc.dist += this_rdc.dist;
4911 RDCOST(x->rdmult, x->rddiv, sum_rdc.rate, sum_rdc.dist);
4915 if (sum_rdc.rdcost < best_rdc.rdcost) {
4917 pc_tree->partitioning = PARTITION_HORZ;
4919 pred_pixel_ready_reset(pc_tree, bsize);
4924 if (partition_vert_allowed && do_rect) {
4925 subsize = get_subsize(bsize, PARTITION_VERT);
4926 load_pred_mv(x, ctx);
4927 pc_tree->vertical[0].pred_pixel_ready = 1;
4928 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &sum_rdc, subsize,
4929 &pc_tree->vertical[0]);
4930 pc_tree->vertical[0].mic = *xd->mi[0];
4931 pc_tree->vertical[0].mbmi_ext = *x->mbmi_ext;
4932 pc_tree->vertical[0].skip_txfm[0] = x->skip_txfm[0];
4933 pc_tree->vertical[0].skip = x->skip;
4935 if (sum_rdc.rdcost < best_rdc.rdcost && mi_col + ms < cm->mi_cols) {
4936 load_pred_mv(x, ctx);
4937 pc_tree->vertical[1].pred_pixel_ready = 1;
4938 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col + ms, &this_rdc,
4939 subsize, &pc_tree->vertical[1]);
4940 pc_tree->vertical[1].mic = *xd->mi[0];
4941 pc_tree->vertical[1].mbmi_ext = *x->mbmi_ext;
4942 pc_tree->vertical[1].skip_txfm[0] = x->skip_txfm[0];
4943 pc_tree->vertical[1].skip = x->skip;
4945 if (this_rdc.rate == INT_MAX) {
4946 vp9_rd_cost_reset(&sum_rdc);
4948 int pl = partition_plane_context(xd, mi_row, mi_col, bsize);
4949 sum_rdc.rate += cpi->partition_cost[pl][PARTITION_VERT];
4950 sum_rdc.rate += this_rdc.rate;
4951 sum_rdc.dist += this_rdc.dist;
4953 RDCOST(x->rdmult, x->rddiv, sum_rdc.rate, sum_rdc.dist);
4957 if (sum_rdc.rdcost < best_rdc.rdcost) {
4959 pc_tree->partitioning = PARTITION_VERT;
4961 pred_pixel_ready_reset(pc_tree, bsize);
4965 *rd_cost = best_rdc;
4967 if (best_rdc.rate == INT_MAX) {
4968 vp9_rd_cost_reset(rd_cost);
4972 // update mode info array
4973 fill_mode_info_sb(cm, x, mi_row, mi_col, bsize, pc_tree);
4975 if (best_rdc.rate < INT_MAX && best_rdc.dist < INT64_MAX && do_recon) {
4976 int output_enabled = (bsize == BLOCK_64X64);
4977 encode_sb_rt(cpi, td, tile_info, tp, mi_row, mi_col, output_enabled, bsize,
4981 if (bsize == BLOCK_64X64 && do_recon) {
4982 assert(tp_orig < *tp);
4983 assert(best_rdc.rate < INT_MAX);
4984 assert(best_rdc.dist < INT64_MAX);
4986 assert(tp_orig == *tp);
4990 static void nonrd_select_partition(VP9_COMP *cpi, ThreadData *td,
4991 TileDataEnc *tile_data, MODE_INFO **mi,
4992 TOKENEXTRA **tp, int mi_row, int mi_col,
4993 BLOCK_SIZE bsize, int output_enabled,
4994 RD_COST *rd_cost, PC_TREE *pc_tree) {
4995 VP9_COMMON *const cm = &cpi->common;
4996 TileInfo *const tile_info = &tile_data->tile_info;
4997 MACROBLOCK *const x = &td->mb;
4998 MACROBLOCKD *const xd = &x->e_mbd;
4999 const int bsl = b_width_log2_lookup[bsize], hbs = (1 << bsl) / 4;
5000 const int mis = cm->mi_stride;
5001 PARTITION_TYPE partition;
5004 BLOCK_SIZE subsize_ref =
5005 (cpi->sf.adapt_partition_source_sad) ? BLOCK_8X8 : BLOCK_16X16;
5007 vp9_rd_cost_reset(&this_rdc);
5008 if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols) return;
5010 subsize = (bsize >= BLOCK_8X8) ? mi[0]->sb_type : BLOCK_4X4;
5011 partition = partition_lookup[bsl][subsize];
5013 if (bsize == BLOCK_32X32 && subsize == BLOCK_32X32) {
5014 x->max_partition_size = BLOCK_32X32;
5015 x->min_partition_size = BLOCK_16X16;
5016 nonrd_pick_partition(cpi, td, tile_data, tp, mi_row, mi_col, bsize, rd_cost,
5017 0, INT64_MAX, pc_tree);
5018 } else if (bsize == BLOCK_32X32 && partition != PARTITION_NONE &&
5019 subsize >= subsize_ref) {
5020 x->max_partition_size = BLOCK_32X32;
5021 x->min_partition_size = BLOCK_8X8;
5022 nonrd_pick_partition(cpi, td, tile_data, tp, mi_row, mi_col, bsize, rd_cost,
5023 0, INT64_MAX, pc_tree);
5024 } else if (bsize == BLOCK_16X16 && partition != PARTITION_NONE) {
5025 x->max_partition_size = BLOCK_16X16;
5026 x->min_partition_size = BLOCK_8X8;
5027 nonrd_pick_partition(cpi, td, tile_data, tp, mi_row, mi_col, bsize, rd_cost,
5028 0, INT64_MAX, pc_tree);
5030 switch (partition) {
5031 case PARTITION_NONE:
5032 pc_tree->none.pred_pixel_ready = 1;
5033 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, rd_cost, subsize,
5035 pc_tree->none.mic = *xd->mi[0];
5036 pc_tree->none.mbmi_ext = *x->mbmi_ext;
5037 pc_tree->none.skip_txfm[0] = x->skip_txfm[0];
5038 pc_tree->none.skip = x->skip;
5040 case PARTITION_VERT:
5041 pc_tree->vertical[0].pred_pixel_ready = 1;
5042 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, rd_cost, subsize,
5043 &pc_tree->vertical[0]);
5044 pc_tree->vertical[0].mic = *xd->mi[0];
5045 pc_tree->vertical[0].mbmi_ext = *x->mbmi_ext;
5046 pc_tree->vertical[0].skip_txfm[0] = x->skip_txfm[0];
5047 pc_tree->vertical[0].skip = x->skip;
5048 if (mi_col + hbs < cm->mi_cols) {
5049 pc_tree->vertical[1].pred_pixel_ready = 1;
5050 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col + hbs,
5051 &this_rdc, subsize, &pc_tree->vertical[1]);
5052 pc_tree->vertical[1].mic = *xd->mi[0];
5053 pc_tree->vertical[1].mbmi_ext = *x->mbmi_ext;
5054 pc_tree->vertical[1].skip_txfm[0] = x->skip_txfm[0];
5055 pc_tree->vertical[1].skip = x->skip;
5056 if (this_rdc.rate != INT_MAX && this_rdc.dist != INT64_MAX &&
5057 rd_cost->rate != INT_MAX && rd_cost->dist != INT64_MAX) {
5058 rd_cost->rate += this_rdc.rate;
5059 rd_cost->dist += this_rdc.dist;
5063 case PARTITION_HORZ:
5064 pc_tree->horizontal[0].pred_pixel_ready = 1;
5065 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, rd_cost, subsize,
5066 &pc_tree->horizontal[0]);
5067 pc_tree->horizontal[0].mic = *xd->mi[0];
5068 pc_tree->horizontal[0].mbmi_ext = *x->mbmi_ext;
5069 pc_tree->horizontal[0].skip_txfm[0] = x->skip_txfm[0];
5070 pc_tree->horizontal[0].skip = x->skip;
5071 if (mi_row + hbs < cm->mi_rows) {
5072 pc_tree->horizontal[1].pred_pixel_ready = 1;
5073 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row + hbs, mi_col,
5074 &this_rdc, subsize, &pc_tree->horizontal[1]);
5075 pc_tree->horizontal[1].mic = *xd->mi[0];
5076 pc_tree->horizontal[1].mbmi_ext = *x->mbmi_ext;
5077 pc_tree->horizontal[1].skip_txfm[0] = x->skip_txfm[0];
5078 pc_tree->horizontal[1].skip = x->skip;
5079 if (this_rdc.rate != INT_MAX && this_rdc.dist != INT64_MAX &&
5080 rd_cost->rate != INT_MAX && rd_cost->dist != INT64_MAX) {
5081 rd_cost->rate += this_rdc.rate;
5082 rd_cost->dist += this_rdc.dist;
5087 assert(partition == PARTITION_SPLIT);
5088 subsize = get_subsize(bsize, PARTITION_SPLIT);
5089 nonrd_select_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col,
5090 subsize, output_enabled, rd_cost,
5092 nonrd_select_partition(cpi, td, tile_data, mi + hbs, tp, mi_row,
5093 mi_col + hbs, subsize, output_enabled, &this_rdc,
5095 if (this_rdc.rate != INT_MAX && this_rdc.dist != INT64_MAX &&
5096 rd_cost->rate != INT_MAX && rd_cost->dist != INT64_MAX) {
5097 rd_cost->rate += this_rdc.rate;
5098 rd_cost->dist += this_rdc.dist;
5100 nonrd_select_partition(cpi, td, tile_data, mi + hbs * mis, tp,
5101 mi_row + hbs, mi_col, subsize, output_enabled,
5102 &this_rdc, pc_tree->split[2]);
5103 if (this_rdc.rate != INT_MAX && this_rdc.dist != INT64_MAX &&
5104 rd_cost->rate != INT_MAX && rd_cost->dist != INT64_MAX) {
5105 rd_cost->rate += this_rdc.rate;
5106 rd_cost->dist += this_rdc.dist;
5108 nonrd_select_partition(cpi, td, tile_data, mi + hbs * mis + hbs, tp,
5109 mi_row + hbs, mi_col + hbs, subsize,
5110 output_enabled, &this_rdc, pc_tree->split[3]);
5111 if (this_rdc.rate != INT_MAX && this_rdc.dist != INT64_MAX &&
5112 rd_cost->rate != INT_MAX && rd_cost->dist != INT64_MAX) {
5113 rd_cost->rate += this_rdc.rate;
5114 rd_cost->dist += this_rdc.dist;
5120 if (bsize == BLOCK_64X64 && output_enabled)
5121 encode_sb_rt(cpi, td, tile_info, tp, mi_row, mi_col, 1, bsize, pc_tree);
5124 static void nonrd_use_partition(VP9_COMP *cpi, ThreadData *td,
5125 TileDataEnc *tile_data, MODE_INFO **mi,
5126 TOKENEXTRA **tp, int mi_row, int mi_col,
5127 BLOCK_SIZE bsize, int output_enabled,
5128 RD_COST *dummy_cost, PC_TREE *pc_tree) {
5129 VP9_COMMON *const cm = &cpi->common;
5130 TileInfo *tile_info = &tile_data->tile_info;
5131 MACROBLOCK *const x = &td->mb;
5132 MACROBLOCKD *const xd = &x->e_mbd;
5133 const int bsl = b_width_log2_lookup[bsize], hbs = (1 << bsl) / 4;
5134 const int mis = cm->mi_stride;
5135 PARTITION_TYPE partition;
5138 if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols) return;
5140 subsize = (bsize >= BLOCK_8X8) ? mi[0]->sb_type : BLOCK_4X4;
5141 partition = partition_lookup[bsl][subsize];
5143 if (output_enabled && bsize != BLOCK_4X4) {
5144 int ctx = partition_plane_context(xd, mi_row, mi_col, bsize);
5145 td->counts->partition[ctx][partition]++;
5148 switch (partition) {
5149 case PARTITION_NONE:
5150 pc_tree->none.pred_pixel_ready = 1;
5151 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, dummy_cost,
5152 subsize, &pc_tree->none);
5153 pc_tree->none.mic = *xd->mi[0];
5154 pc_tree->none.mbmi_ext = *x->mbmi_ext;
5155 pc_tree->none.skip_txfm[0] = x->skip_txfm[0];
5156 pc_tree->none.skip = x->skip;
5157 encode_b_rt(cpi, td, tile_info, tp, mi_row, mi_col, output_enabled,
5158 subsize, &pc_tree->none);
5160 case PARTITION_VERT:
5161 pc_tree->vertical[0].pred_pixel_ready = 1;
5162 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, dummy_cost,
5163 subsize, &pc_tree->vertical[0]);
5164 pc_tree->vertical[0].mic = *xd->mi[0];
5165 pc_tree->vertical[0].mbmi_ext = *x->mbmi_ext;
5166 pc_tree->vertical[0].skip_txfm[0] = x->skip_txfm[0];
5167 pc_tree->vertical[0].skip = x->skip;
5168 encode_b_rt(cpi, td, tile_info, tp, mi_row, mi_col, output_enabled,
5169 subsize, &pc_tree->vertical[0]);
5170 if (mi_col + hbs < cm->mi_cols && bsize > BLOCK_8X8) {
5171 pc_tree->vertical[1].pred_pixel_ready = 1;
5172 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col + hbs, dummy_cost,
5173 subsize, &pc_tree->vertical[1]);
5174 pc_tree->vertical[1].mic = *xd->mi[0];
5175 pc_tree->vertical[1].mbmi_ext = *x->mbmi_ext;
5176 pc_tree->vertical[1].skip_txfm[0] = x->skip_txfm[0];
5177 pc_tree->vertical[1].skip = x->skip;
5178 encode_b_rt(cpi, td, tile_info, tp, mi_row, mi_col + hbs,
5179 output_enabled, subsize, &pc_tree->vertical[1]);
5182 case PARTITION_HORZ:
5183 pc_tree->horizontal[0].pred_pixel_ready = 1;
5184 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, dummy_cost,
5185 subsize, &pc_tree->horizontal[0]);
5186 pc_tree->horizontal[0].mic = *xd->mi[0];
5187 pc_tree->horizontal[0].mbmi_ext = *x->mbmi_ext;
5188 pc_tree->horizontal[0].skip_txfm[0] = x->skip_txfm[0];
5189 pc_tree->horizontal[0].skip = x->skip;
5190 encode_b_rt(cpi, td, tile_info, tp, mi_row, mi_col, output_enabled,
5191 subsize, &pc_tree->horizontal[0]);
5193 if (mi_row + hbs < cm->mi_rows && bsize > BLOCK_8X8) {
5194 pc_tree->horizontal[1].pred_pixel_ready = 1;
5195 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row + hbs, mi_col, dummy_cost,
5196 subsize, &pc_tree->horizontal[1]);
5197 pc_tree->horizontal[1].mic = *xd->mi[0];
5198 pc_tree->horizontal[1].mbmi_ext = *x->mbmi_ext;
5199 pc_tree->horizontal[1].skip_txfm[0] = x->skip_txfm[0];
5200 pc_tree->horizontal[1].skip = x->skip;
5201 encode_b_rt(cpi, td, tile_info, tp, mi_row + hbs, mi_col,
5202 output_enabled, subsize, &pc_tree->horizontal[1]);
5206 assert(partition == PARTITION_SPLIT);
5207 subsize = get_subsize(bsize, PARTITION_SPLIT);
5208 if (bsize == BLOCK_8X8) {
5209 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, dummy_cost,
5210 subsize, pc_tree->leaf_split[0]);
5211 encode_b_rt(cpi, td, tile_info, tp, mi_row, mi_col, output_enabled,
5212 subsize, pc_tree->leaf_split[0]);
5214 nonrd_use_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col, subsize,
5215 output_enabled, dummy_cost, pc_tree->split[0]);
5216 nonrd_use_partition(cpi, td, tile_data, mi + hbs, tp, mi_row,
5217 mi_col + hbs, subsize, output_enabled, dummy_cost,
5219 nonrd_use_partition(cpi, td, tile_data, mi + hbs * mis, tp,
5220 mi_row + hbs, mi_col, subsize, output_enabled,
5221 dummy_cost, pc_tree->split[2]);
5222 nonrd_use_partition(cpi, td, tile_data, mi + hbs * mis + hbs, tp,
5223 mi_row + hbs, mi_col + hbs, subsize, output_enabled,
5224 dummy_cost, pc_tree->split[3]);
5229 if (partition != PARTITION_SPLIT || bsize == BLOCK_8X8)
5230 update_partition_context(xd, mi_row, mi_col, subsize, bsize);
5233 // Get a prediction(stored in x->est_pred) for the whole 64x64 superblock.
5234 static void get_estimated_pred(VP9_COMP *cpi, const TileInfo *const tile,
5235 MACROBLOCK *x, int mi_row, int mi_col) {
5236 VP9_COMMON *const cm = &cpi->common;
5237 const int is_key_frame = frame_is_intra_only(cm);
5238 MACROBLOCKD *xd = &x->e_mbd;
5240 set_offsets(cpi, tile, x, mi_row, mi_col, BLOCK_64X64);
5242 if (!is_key_frame) {
5243 MODE_INFO *mi = xd->mi[0];
5244 YV12_BUFFER_CONFIG *yv12 = get_ref_frame_buffer(cpi, LAST_FRAME);
5245 const YV12_BUFFER_CONFIG *yv12_g = NULL;
5246 const BLOCK_SIZE bsize = BLOCK_32X32 + (mi_col + 4 < cm->mi_cols) * 2 +
5247 (mi_row + 4 < cm->mi_rows);
5248 unsigned int y_sad_g, y_sad_thr;
5249 unsigned int y_sad = UINT_MAX;
5251 assert(yv12 != NULL);
5253 if (!(is_one_pass_cbr_svc(cpi) && cpi->svc.spatial_layer_id) ||
5254 cpi->svc.use_gf_temporal_ref_current_layer) {
5255 // For now, GOLDEN will not be used for non-zero spatial layers, since
5256 // it may not be a temporal reference.
5257 yv12_g = get_ref_frame_buffer(cpi, GOLDEN_FRAME);
5260 // Only compute y_sad_g (sad for golden reference) for speed < 8.
5261 if (cpi->oxcf.speed < 8 && yv12_g && yv12_g != yv12 &&
5262 (cpi->ref_frame_flags & VP9_GOLD_FLAG)) {
5263 vp9_setup_pre_planes(xd, 0, yv12_g, mi_row, mi_col,
5264 &cm->frame_refs[GOLDEN_FRAME - 1].sf);
5265 y_sad_g = cpi->fn_ptr[bsize].sdf(
5266 x->plane[0].src.buf, x->plane[0].src.stride, xd->plane[0].pre[0].buf,
5267 xd->plane[0].pre[0].stride);
5272 if (cpi->oxcf.lag_in_frames > 0 && cpi->oxcf.rc_mode == VPX_VBR &&
5273 cpi->rc.is_src_frame_alt_ref) {
5274 yv12 = get_ref_frame_buffer(cpi, ALTREF_FRAME);
5275 vp9_setup_pre_planes(xd, 0, yv12, mi_row, mi_col,
5276 &cm->frame_refs[ALTREF_FRAME - 1].sf);
5277 mi->ref_frame[0] = ALTREF_FRAME;
5280 vp9_setup_pre_planes(xd, 0, yv12, mi_row, mi_col,
5281 &cm->frame_refs[LAST_FRAME - 1].sf);
5282 mi->ref_frame[0] = LAST_FRAME;
5284 mi->ref_frame[1] = NONE;
5285 mi->sb_type = BLOCK_64X64;
5286 mi->mv[0].as_int = 0;
5287 mi->interp_filter = BILINEAR;
5290 const MV dummy_mv = { 0, 0 };
5291 y_sad = vp9_int_pro_motion_estimation(cpi, x, bsize, mi_row, mi_col,
5293 x->sb_use_mv_part = 1;
5294 x->sb_mvcol_part = mi->mv[0].as_mv.col;
5295 x->sb_mvrow_part = mi->mv[0].as_mv.row;
5298 // Pick ref frame for partitioning, bias last frame when y_sad_g and y_sad
5299 // are close if short_circuit_low_temp_var is on.
5300 y_sad_thr = cpi->sf.short_circuit_low_temp_var ? (y_sad * 7) >> 3 : y_sad;
5301 if (y_sad_g < y_sad_thr) {
5302 vp9_setup_pre_planes(xd, 0, yv12_g, mi_row, mi_col,
5303 &cm->frame_refs[GOLDEN_FRAME - 1].sf);
5304 mi->ref_frame[0] = GOLDEN_FRAME;
5305 mi->mv[0].as_int = 0;
5307 x->pred_mv[LAST_FRAME] = mi->mv[0].as_mv;
5310 set_ref_ptrs(cm, xd, mi->ref_frame[0], mi->ref_frame[1]);
5311 xd->plane[0].dst.buf = x->est_pred;
5312 xd->plane[0].dst.stride = 64;
5313 vp9_build_inter_predictors_sb(xd, mi_row, mi_col, BLOCK_64X64);
5315 #if CONFIG_VP9_HIGHBITDEPTH
5317 case 8: memset(x->est_pred, 128, 64 * 64 * sizeof(x->est_pred[0])); break;
5319 memset(x->est_pred, 128 * 4, 64 * 64 * sizeof(x->est_pred[0]));
5322 memset(x->est_pred, 128 * 16, 64 * 64 * sizeof(x->est_pred[0]));
5326 memset(x->est_pred, 128, 64 * 64 * sizeof(x->est_pred[0]));
5327 #endif // CONFIG_VP9_HIGHBITDEPTH
5331 static void encode_nonrd_sb_row(VP9_COMP *cpi, ThreadData *td,
5332 TileDataEnc *tile_data, int mi_row,
5334 SPEED_FEATURES *const sf = &cpi->sf;
5335 VP9_COMMON *const cm = &cpi->common;
5336 TileInfo *const tile_info = &tile_data->tile_info;
5337 MACROBLOCK *const x = &td->mb;
5338 MACROBLOCKD *const xd = &x->e_mbd;
5339 const int mi_col_start = tile_info->mi_col_start;
5340 const int mi_col_end = tile_info->mi_col_end;
5342 const int sb_row = mi_row >> MI_BLOCK_SIZE_LOG2;
5343 const int num_sb_cols =
5344 get_num_cols(tile_data->tile_info, MI_BLOCK_SIZE_LOG2);
5347 // Initialize the left context for the new SB row
5348 memset(&xd->left_context, 0, sizeof(xd->left_context));
5349 memset(xd->left_seg_context, 0, sizeof(xd->left_seg_context));
5351 // Code each SB in the row
5352 for (mi_col = mi_col_start, sb_col_in_tile = 0; mi_col < mi_col_end;
5353 mi_col += MI_BLOCK_SIZE, ++sb_col_in_tile) {
5354 const struct segmentation *const seg = &cm->seg;
5356 const int idx_str = cm->mi_stride * mi_row + mi_col;
5357 MODE_INFO **mi = cm->mi_grid_visible + idx_str;
5358 PARTITION_SEARCH_TYPE partition_search_type = sf->partition_search_type;
5359 BLOCK_SIZE bsize = BLOCK_64X64;
5363 (*(cpi->row_mt_sync_read_ptr))(&tile_data->row_mt_sync, sb_row,
5366 if (cpi->use_skin_detection) {
5367 vp9_compute_skin_sb(cpi, BLOCK_16X16, mi_row, mi_col);
5370 x->source_variance = UINT_MAX;
5371 for (i = 0; i < MAX_REF_FRAMES; ++i) {
5372 x->pred_mv[i].row = INT16_MAX;
5373 x->pred_mv[i].col = INT16_MAX;
5375 vp9_rd_cost_init(&dummy_rdc);
5376 x->color_sensitivity[0] = 0;
5377 x->color_sensitivity[1] = 0;
5379 x->skip_low_source_sad = 0;
5380 x->lowvar_highsumdiff = 0;
5381 x->content_state_sb = 0;
5382 x->zero_temp_sad_source = 0;
5383 x->sb_use_mv_part = 0;
5384 x->sb_mvcol_part = 0;
5385 x->sb_mvrow_part = 0;
5386 x->sb_pickmode_part = 0;
5387 x->arf_frame_usage = 0;
5388 x->lastgolden_frame_usage = 0;
5391 const uint8_t *const map =
5392 seg->update_map ? cpi->segmentation_map : cm->last_frame_seg_map;
5393 int segment_id = get_segment_id(cm, map, BLOCK_64X64, mi_row, mi_col);
5394 seg_skip = segfeature_active(seg, segment_id, SEG_LVL_SKIP);
5396 partition_search_type = FIXED_PARTITION;
5400 if (cpi->compute_source_sad_onepass && cpi->sf.use_source_sad) {
5401 int shift = cpi->Source->y_stride * (mi_row << 3) + (mi_col << 3);
5402 int sb_offset2 = ((cm->mi_cols + 7) >> 3) * (mi_row >> 3) + (mi_col >> 3);
5403 int64_t source_sad = avg_source_sad(cpi, x, shift, sb_offset2);
5404 if (sf->adapt_partition_source_sad &&
5405 (cpi->oxcf.rc_mode == VPX_VBR && !cpi->rc.is_src_frame_alt_ref &&
5406 source_sad > sf->adapt_partition_thresh &&
5407 (cpi->refresh_golden_frame || cpi->refresh_alt_ref_frame)))
5408 partition_search_type = REFERENCE_PARTITION;
5411 // Set the partition type of the 64X64 block
5412 switch (partition_search_type) {
5413 case VAR_BASED_PARTITION:
5414 // TODO(jingning, marpan): The mode decision and encoding process
5415 // support both intra and inter sub8x8 block coding for RTC mode.
5416 // Tune the thresholds accordingly to use sub8x8 block coding for
5417 // coding performance improvement.
5418 choose_partitioning(cpi, tile_info, x, mi_row, mi_col);
5419 nonrd_use_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col,
5420 BLOCK_64X64, 1, &dummy_rdc, td->pc_root);
5422 case ML_BASED_PARTITION:
5423 get_estimated_pred(cpi, tile_info, x, mi_row, mi_col);
5424 x->max_partition_size = BLOCK_64X64;
5425 x->min_partition_size = BLOCK_8X8;
5426 x->sb_pickmode_part = 1;
5427 nonrd_pick_partition(cpi, td, tile_data, tp, mi_row, mi_col,
5428 BLOCK_64X64, &dummy_rdc, 1, INT64_MAX,
5431 case SOURCE_VAR_BASED_PARTITION:
5432 set_source_var_based_partition(cpi, tile_info, x, mi, mi_row, mi_col);
5433 nonrd_use_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col,
5434 BLOCK_64X64, 1, &dummy_rdc, td->pc_root);
5436 case FIXED_PARTITION:
5437 if (!seg_skip) bsize = sf->always_this_block_size;
5438 set_fixed_partitioning(cpi, tile_info, mi, mi_row, mi_col, bsize);
5439 nonrd_use_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col,
5440 BLOCK_64X64, 1, &dummy_rdc, td->pc_root);
5443 assert(partition_search_type == REFERENCE_PARTITION);
5444 x->sb_pickmode_part = 1;
5445 set_offsets(cpi, tile_info, x, mi_row, mi_col, BLOCK_64X64);
5446 // Use nonrd_pick_partition on scene-cut for VBR mode.
5447 // nonrd_pick_partition does not support 4x4 partition, so avoid it
5448 // on key frame for now.
5449 if ((cpi->oxcf.rc_mode == VPX_VBR && cpi->rc.high_source_sad &&
5450 cpi->oxcf.speed < 6 && !frame_is_intra_only(cm) &&
5451 (cpi->refresh_golden_frame || cpi->refresh_alt_ref_frame))) {
5452 // Use lower max_partition_size for low resoultions.
5453 if (cm->width <= 352 && cm->height <= 288)
5454 x->max_partition_size = BLOCK_32X32;
5456 x->max_partition_size = BLOCK_64X64;
5457 x->min_partition_size = BLOCK_8X8;
5458 nonrd_pick_partition(cpi, td, tile_data, tp, mi_row, mi_col,
5459 BLOCK_64X64, &dummy_rdc, 1, INT64_MAX,
5462 choose_partitioning(cpi, tile_info, x, mi_row, mi_col);
5463 // TODO(marpan): Seems like nonrd_select_partition does not support
5464 // 4x4 partition. Since 4x4 is used on key frame, use this switch
5466 if (frame_is_intra_only(cm))
5467 nonrd_use_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col,
5468 BLOCK_64X64, 1, &dummy_rdc, td->pc_root);
5470 nonrd_select_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col,
5471 BLOCK_64X64, 1, &dummy_rdc, td->pc_root);
5477 // Update ref_frame usage for inter frame if this group is ARF group.
5478 if (!cpi->rc.is_src_frame_alt_ref && !cpi->refresh_golden_frame &&
5479 !cpi->refresh_alt_ref_frame && cpi->rc.alt_ref_gf_group &&
5480 cpi->sf.use_altref_onepass) {
5481 int sboffset = ((cm->mi_cols + 7) >> 3) * (mi_row >> 3) + (mi_col >> 3);
5482 if (cpi->count_arf_frame_usage != NULL)
5483 cpi->count_arf_frame_usage[sboffset] = x->arf_frame_usage;
5484 if (cpi->count_lastgolden_frame_usage != NULL)
5485 cpi->count_lastgolden_frame_usage[sboffset] = x->lastgolden_frame_usage;
5488 (*(cpi->row_mt_sync_write_ptr))(&tile_data->row_mt_sync, sb_row,
5489 sb_col_in_tile, num_sb_cols);
5492 // end RTC play code
5494 static INLINE uint32_t variance(const diff *const d) {
5495 return d->sse - (uint32_t)(((int64_t)d->sum * d->sum) >> 8);
5498 #if CONFIG_VP9_HIGHBITDEPTH
5499 static INLINE uint32_t variance_highbd(diff *const d) {
5500 const int64_t var = (int64_t)d->sse - (((int64_t)d->sum * d->sum) >> 8);
5501 return (var >= 0) ? (uint32_t)var : 0;
5503 #endif // CONFIG_VP9_HIGHBITDEPTH
5505 static int set_var_thresh_from_histogram(VP9_COMP *cpi) {
5506 const SPEED_FEATURES *const sf = &cpi->sf;
5507 const VP9_COMMON *const cm = &cpi->common;
5509 const uint8_t *src = cpi->Source->y_buffer;
5510 const uint8_t *last_src = cpi->Last_Source->y_buffer;
5511 const int src_stride = cpi->Source->y_stride;
5512 const int last_stride = cpi->Last_Source->y_stride;
5514 // Pick cutoff threshold
5515 const int cutoff = (VPXMIN(cm->width, cm->height) >= 720)
5516 ? (cm->MBs * VAR_HIST_LARGE_CUT_OFF / 100)
5517 : (cm->MBs * VAR_HIST_SMALL_CUT_OFF / 100);
5518 DECLARE_ALIGNED(16, int, hist[VAR_HIST_BINS]);
5519 diff *var16 = cpi->source_diff_var;
5524 memset(hist, 0, VAR_HIST_BINS * sizeof(hist[0]));
5526 for (i = 0; i < cm->mb_rows; i++) {
5527 for (j = 0; j < cm->mb_cols; j++) {
5528 #if CONFIG_VP9_HIGHBITDEPTH
5529 if (cm->use_highbitdepth) {
5530 switch (cm->bit_depth) {
5532 vpx_highbd_8_get16x16var(src, src_stride, last_src, last_stride,
5533 &var16->sse, &var16->sum);
5534 var16->var = variance(var16);
5537 vpx_highbd_10_get16x16var(src, src_stride, last_src, last_stride,
5538 &var16->sse, &var16->sum);
5539 var16->var = variance_highbd(var16);
5542 assert(cm->bit_depth == VPX_BITS_12);
5543 vpx_highbd_12_get16x16var(src, src_stride, last_src, last_stride,
5544 &var16->sse, &var16->sum);
5545 var16->var = variance_highbd(var16);
5549 vpx_get16x16var(src, src_stride, last_src, last_stride, &var16->sse,
5551 var16->var = variance(var16);
5554 vpx_get16x16var(src, src_stride, last_src, last_stride, &var16->sse,
5556 var16->var = variance(var16);
5557 #endif // CONFIG_VP9_HIGHBITDEPTH
5559 if (var16->var >= VAR_HIST_MAX_BG_VAR)
5560 hist[VAR_HIST_BINS - 1]++;
5562 hist[var16->var / VAR_HIST_FACTOR]++;
5569 src = src - cm->mb_cols * 16 + 16 * src_stride;
5570 last_src = last_src - cm->mb_cols * 16 + 16 * last_stride;
5573 cpi->source_var_thresh = 0;
5575 if (hist[VAR_HIST_BINS - 1] < cutoff) {
5576 for (i = 0; i < VAR_HIST_BINS - 1; i++) {
5580 cpi->source_var_thresh = (i + 1) * VAR_HIST_FACTOR;
5586 return sf->search_type_check_frequency;
5589 static void source_var_based_partition_search_method(VP9_COMP *cpi) {
5590 VP9_COMMON *const cm = &cpi->common;
5591 SPEED_FEATURES *const sf = &cpi->sf;
5593 if (cm->frame_type == KEY_FRAME) {
5594 // For key frame, use SEARCH_PARTITION.
5595 sf->partition_search_type = SEARCH_PARTITION;
5596 } else if (cm->intra_only) {
5597 sf->partition_search_type = FIXED_PARTITION;
5599 if (cm->last_width != cm->width || cm->last_height != cm->height) {
5600 if (cpi->source_diff_var) vpx_free(cpi->source_diff_var);
5602 CHECK_MEM_ERROR(cm, cpi->source_diff_var,
5603 vpx_calloc(cm->MBs, sizeof(diff)));
5606 if (!cpi->frames_till_next_var_check)
5607 cpi->frames_till_next_var_check = set_var_thresh_from_histogram(cpi);
5609 if (cpi->frames_till_next_var_check > 0) {
5610 sf->partition_search_type = FIXED_PARTITION;
5611 cpi->frames_till_next_var_check--;
5616 static int get_skip_encode_frame(const VP9_COMMON *cm, ThreadData *const td) {
5617 unsigned int intra_count = 0, inter_count = 0;
5620 for (j = 0; j < INTRA_INTER_CONTEXTS; ++j) {
5621 intra_count += td->counts->intra_inter[j][0];
5622 inter_count += td->counts->intra_inter[j][1];
5625 return (intra_count << 2) < inter_count && cm->frame_type != KEY_FRAME &&
5629 void vp9_init_tile_data(VP9_COMP *cpi) {
5630 VP9_COMMON *const cm = &cpi->common;
5631 const int tile_cols = 1 << cm->log2_tile_cols;
5632 const int tile_rows = 1 << cm->log2_tile_rows;
5633 int tile_col, tile_row;
5634 TOKENEXTRA *pre_tok = cpi->tile_tok[0][0];
5635 TOKENLIST *tplist = cpi->tplist[0][0];
5637 int tplist_count = 0;
5639 if (cpi->tile_data == NULL || cpi->allocated_tiles < tile_cols * tile_rows) {
5640 if (cpi->tile_data != NULL) vpx_free(cpi->tile_data);
5643 vpx_malloc(tile_cols * tile_rows * sizeof(*cpi->tile_data)));
5644 cpi->allocated_tiles = tile_cols * tile_rows;
5646 for (tile_row = 0; tile_row < tile_rows; ++tile_row)
5647 for (tile_col = 0; tile_col < tile_cols; ++tile_col) {
5648 TileDataEnc *tile_data =
5649 &cpi->tile_data[tile_row * tile_cols + tile_col];
5651 for (i = 0; i < BLOCK_SIZES; ++i) {
5652 for (j = 0; j < MAX_MODES; ++j) {
5653 tile_data->thresh_freq_fact[i][j] = RD_THRESH_INIT_FACT;
5654 #if CONFIG_CONSISTENT_RECODE
5655 tile_data->thresh_freq_fact_prev[i][j] = RD_THRESH_INIT_FACT;
5657 tile_data->mode_map[i][j] = j;
5660 #if CONFIG_MULTITHREAD
5661 tile_data->row_base_thresh_freq_fact = NULL;
5666 for (tile_row = 0; tile_row < tile_rows; ++tile_row) {
5667 for (tile_col = 0; tile_col < tile_cols; ++tile_col) {
5668 TileDataEnc *this_tile = &cpi->tile_data[tile_row * tile_cols + tile_col];
5669 TileInfo *tile_info = &this_tile->tile_info;
5670 if (cpi->sf.adaptive_rd_thresh_row_mt &&
5671 this_tile->row_base_thresh_freq_fact == NULL)
5672 vp9_row_mt_alloc_rd_thresh(cpi, this_tile);
5673 vp9_tile_init(tile_info, cm, tile_row, tile_col);
5675 cpi->tile_tok[tile_row][tile_col] = pre_tok + tile_tok;
5676 pre_tok = cpi->tile_tok[tile_row][tile_col];
5677 tile_tok = allocated_tokens(*tile_info);
5679 cpi->tplist[tile_row][tile_col] = tplist + tplist_count;
5680 tplist = cpi->tplist[tile_row][tile_col];
5681 tplist_count = get_num_vert_units(*tile_info, MI_BLOCK_SIZE_LOG2);
5686 void vp9_encode_sb_row(VP9_COMP *cpi, ThreadData *td, int tile_row,
5687 int tile_col, int mi_row) {
5688 VP9_COMMON *const cm = &cpi->common;
5689 const int tile_cols = 1 << cm->log2_tile_cols;
5690 TileDataEnc *this_tile = &cpi->tile_data[tile_row * tile_cols + tile_col];
5691 const TileInfo *const tile_info = &this_tile->tile_info;
5692 TOKENEXTRA *tok = NULL;
5694 int tile_mb_cols = (tile_info->mi_col_end - tile_info->mi_col_start + 1) >> 1;
5696 tile_sb_row = mi_cols_aligned_to_sb(mi_row - tile_info->mi_row_start) >>
5698 get_start_tok(cpi, tile_row, tile_col, mi_row, &tok);
5699 cpi->tplist[tile_row][tile_col][tile_sb_row].start = tok;
5701 if (cpi->sf.use_nonrd_pick_mode)
5702 encode_nonrd_sb_row(cpi, td, this_tile, mi_row, &tok);
5704 encode_rd_sb_row(cpi, td, this_tile, mi_row, &tok);
5706 cpi->tplist[tile_row][tile_col][tile_sb_row].stop = tok;
5707 cpi->tplist[tile_row][tile_col][tile_sb_row].count =
5708 (unsigned int)(cpi->tplist[tile_row][tile_col][tile_sb_row].stop -
5709 cpi->tplist[tile_row][tile_col][tile_sb_row].start);
5710 assert(tok - cpi->tplist[tile_row][tile_col][tile_sb_row].start <=
5711 get_token_alloc(MI_BLOCK_SIZE >> 1, tile_mb_cols));
5716 void vp9_encode_tile(VP9_COMP *cpi, ThreadData *td, int tile_row,
5718 VP9_COMMON *const cm = &cpi->common;
5719 const int tile_cols = 1 << cm->log2_tile_cols;
5720 TileDataEnc *this_tile = &cpi->tile_data[tile_row * tile_cols + tile_col];
5721 const TileInfo *const tile_info = &this_tile->tile_info;
5722 const int mi_row_start = tile_info->mi_row_start;
5723 const int mi_row_end = tile_info->mi_row_end;
5726 for (mi_row = mi_row_start; mi_row < mi_row_end; mi_row += MI_BLOCK_SIZE)
5727 vp9_encode_sb_row(cpi, td, tile_row, tile_col, mi_row);
5730 static void encode_tiles(VP9_COMP *cpi) {
5731 VP9_COMMON *const cm = &cpi->common;
5732 const int tile_cols = 1 << cm->log2_tile_cols;
5733 const int tile_rows = 1 << cm->log2_tile_rows;
5734 int tile_col, tile_row;
5736 vp9_init_tile_data(cpi);
5738 for (tile_row = 0; tile_row < tile_rows; ++tile_row)
5739 for (tile_col = 0; tile_col < tile_cols; ++tile_col)
5740 vp9_encode_tile(cpi, &cpi->td, tile_row, tile_col);
5743 #if CONFIG_FP_MB_STATS
5744 static int input_fpmb_stats(FIRSTPASS_MB_STATS *firstpass_mb_stats,
5745 VP9_COMMON *cm, uint8_t **this_frame_mb_stats) {
5746 uint8_t *mb_stats_in = firstpass_mb_stats->mb_stats_start +
5747 cm->current_video_frame * cm->MBs * sizeof(uint8_t);
5749 if (mb_stats_in > firstpass_mb_stats->mb_stats_end) return EOF;
5751 *this_frame_mb_stats = mb_stats_in;
5757 static int compare_kmeans_data(const void *a, const void *b) {
5758 if (((const KMEANS_DATA *)a)->value > ((const KMEANS_DATA *)b)->value) {
5760 } else if (((const KMEANS_DATA *)a)->value <
5761 ((const KMEANS_DATA *)b)->value) {
5768 static void compute_boundary_ls(const double *ctr_ls, int k,
5769 double *boundary_ls) {
5770 // boundary_ls[j] is the upper bound of data centered at ctr_ls[j]
5772 for (j = 0; j < k - 1; ++j) {
5773 boundary_ls[j] = (ctr_ls[j] + ctr_ls[j + 1]) / 2.;
5775 boundary_ls[k - 1] = DBL_MAX;
5778 int vp9_get_group_idx(double value, double *boundary_ls, int k) {
5780 while (value >= boundary_ls[group_idx]) {
5782 if (group_idx == k - 1) {
5789 void vp9_kmeans(double *ctr_ls, double *boundary_ls, int *count_ls, int k,
5790 KMEANS_DATA *arr, int size) {
5799 vpx_clear_system_state();
5801 assert(k >= 2 && k <= MAX_KMEANS_GROUPS);
5803 qsort(arr, size, sizeof(*arr), compare_kmeans_data);
5806 max = arr[size - 1].value;
5808 // initialize the center points
5809 step = (max - min) * 1. / k;
5810 for (j = 0; j < k; ++j) {
5811 ctr_ls[j] = min + j * step + step / 2;
5814 for (itr = 0; itr < 10; ++itr) {
5815 compute_boundary_ls(ctr_ls, k, boundary_ls);
5819 for (i = 0; i < size; ++i) {
5820 while (arr[i].value >= boundary_ls[group_idx]) {
5822 if (group_idx == k - 1) {
5827 sum += arr[i].value;
5830 if (i + 1 == size || arr[i + 1].value >= boundary_ls[group_idx]) {
5832 ctr_ls[group_idx] = sum / count;
5840 // compute group_idx, boundary_ls and count_ls
5841 for (j = 0; j < k; ++j) {
5844 compute_boundary_ls(ctr_ls, k, boundary_ls);
5846 for (i = 0; i < size; ++i) {
5847 while (arr[i].value >= boundary_ls[group_idx]) {
5849 if (group_idx == k - 1) {
5853 arr[i].group_idx = group_idx;
5854 ++count_ls[group_idx];
5858 static void encode_frame_internal(VP9_COMP *cpi) {
5859 SPEED_FEATURES *const sf = &cpi->sf;
5860 ThreadData *const td = &cpi->td;
5861 MACROBLOCK *const x = &td->mb;
5862 VP9_COMMON *const cm = &cpi->common;
5863 MACROBLOCKD *const xd = &x->e_mbd;
5864 const int gf_group_index = cpi->twopass.gf_group.index;
5866 xd->mi = cm->mi_grid_visible;
5868 vp9_zero(*td->counts);
5869 vp9_zero(cpi->td.rd_counts);
5871 xd->lossless = cm->base_qindex == 0 && cm->y_dc_delta_q == 0 &&
5872 cm->uv_dc_delta_q == 0 && cm->uv_ac_delta_q == 0;
5874 #if CONFIG_VP9_HIGHBITDEPTH
5875 if (cm->use_highbitdepth)
5876 x->fwd_txfm4x4 = xd->lossless ? vp9_highbd_fwht4x4 : vpx_highbd_fdct4x4;
5878 x->fwd_txfm4x4 = xd->lossless ? vp9_fwht4x4 : vpx_fdct4x4;
5879 x->highbd_inv_txfm_add =
5880 xd->lossless ? vp9_highbd_iwht4x4_add : vp9_highbd_idct4x4_add;
5882 x->fwd_txfm4x4 = xd->lossless ? vp9_fwht4x4 : vpx_fdct4x4;
5883 #endif // CONFIG_VP9_HIGHBITDEPTH
5884 x->inv_txfm_add = xd->lossless ? vp9_iwht4x4_add : vp9_idct4x4_add;
5885 #if CONFIG_CONSISTENT_RECODE
5886 x->optimize = sf->optimize_coefficients == 1 && cpi->oxcf.pass != 1;
5888 if (xd->lossless) x->optimize = 0;
5889 x->sharpness = cpi->oxcf.sharpness;
5890 x->adjust_rdmult_by_segment = (cpi->oxcf.aq_mode == VARIANCE_AQ);
5892 cm->tx_mode = select_tx_mode(cpi, xd);
5894 vp9_frame_init_quantizer(cpi);
5896 vp9_initialize_rd_consts(cpi);
5897 vp9_initialize_me_consts(cpi, x, cm->base_qindex);
5898 init_encode_frame_mb_context(cpi);
5899 cm->use_prev_frame_mvs =
5900 !cm->error_resilient_mode && cm->width == cm->last_width &&
5901 cm->height == cm->last_height && !cm->intra_only && cm->last_show_frame;
5902 // Special case: set prev_mi to NULL when the previous mode info
5903 // context cannot be used.
5905 cm->use_prev_frame_mvs ? cm->prev_mip + cm->mi_stride + 1 : NULL;
5907 x->quant_fp = cpi->sf.use_quant_fp;
5908 vp9_zero(x->skip_txfm);
5909 if (sf->use_nonrd_pick_mode) {
5910 // Initialize internal buffer pointers for rtc coding, where non-RD
5911 // mode decision is used and hence no buffer pointer swap needed.
5913 struct macroblock_plane *const p = x->plane;
5914 struct macroblockd_plane *const pd = xd->plane;
5915 PICK_MODE_CONTEXT *ctx = &cpi->td.pc_root->none;
5917 for (i = 0; i < MAX_MB_PLANE; ++i) {
5918 p[i].coeff = ctx->coeff_pbuf[i][0];
5919 p[i].qcoeff = ctx->qcoeff_pbuf[i][0];
5920 pd[i].dqcoeff = ctx->dqcoeff_pbuf[i][0];
5921 p[i].eobs = ctx->eobs_pbuf[i][0];
5923 vp9_zero(x->zcoeff_blk);
5925 if (cm->frame_type != KEY_FRAME && cpi->rc.frames_since_golden == 0 &&
5926 !(cpi->oxcf.lag_in_frames > 0 && cpi->oxcf.rc_mode == VPX_VBR) &&
5928 cpi->ref_frame_flags &= (~VP9_GOLD_FLAG);
5930 if (sf->partition_search_type == SOURCE_VAR_BASED_PARTITION)
5931 source_var_based_partition_search_method(cpi);
5932 } else if (gf_group_index && gf_group_index < MAX_ARF_GOP_SIZE &&
5933 cpi->sf.enable_tpl_model) {
5934 TplDepFrame *tpl_frame = &cpi->tpl_stats[cpi->twopass.gf_group.index];
5935 TplDepStats *tpl_stats = tpl_frame->tpl_stats_ptr;
5937 int tpl_stride = tpl_frame->stride;
5938 int64_t intra_cost_base = 0;
5939 int64_t mc_dep_cost_base = 0;
5942 for (row = 0; row < cm->mi_rows && tpl_frame->is_valid; ++row) {
5943 for (col = 0; col < cm->mi_cols; ++col) {
5944 TplDepStats *this_stats = &tpl_stats[row * tpl_stride + col];
5945 intra_cost_base += this_stats->intra_cost;
5946 mc_dep_cost_base += this_stats->mc_dep_cost;
5950 vpx_clear_system_state();
5952 if (tpl_frame->is_valid)
5953 cpi->rd.r0 = (double)intra_cost_base / mc_dep_cost_base;
5956 // Frame segmentation
5957 if (cpi->sf.enable_wiener_variance) build_kmeans_segmentation(cpi);
5960 struct vpx_usec_timer emr_timer;
5961 vpx_usec_timer_start(&emr_timer);
5963 #if CONFIG_FP_MB_STATS
5964 if (cpi->use_fp_mb_stats) {
5965 input_fpmb_stats(&cpi->twopass.firstpass_mb_stats, cm,
5966 &cpi->twopass.this_frame_mb_stats);
5971 cpi->row_mt_sync_read_ptr = vp9_row_mt_sync_read_dummy;
5972 cpi->row_mt_sync_write_ptr = vp9_row_mt_sync_write_dummy;
5973 // If allowed, encoding tiles in parallel with one thread handling one
5974 // tile when row based multi-threading is disabled.
5975 if (VPXMIN(cpi->oxcf.max_threads, 1 << cm->log2_tile_cols) > 1)
5976 vp9_encode_tiles_mt(cpi);
5980 cpi->row_mt_sync_read_ptr = vp9_row_mt_sync_read;
5981 cpi->row_mt_sync_write_ptr = vp9_row_mt_sync_write;
5982 vp9_encode_tiles_row_mt(cpi);
5985 vpx_usec_timer_mark(&emr_timer);
5986 cpi->time_encode_sb_row += vpx_usec_timer_elapsed(&emr_timer);
5989 sf->skip_encode_frame =
5990 sf->skip_encode_sb ? get_skip_encode_frame(cm, td) : 0;
5993 // Keep record of the total distortion this time around for future use
5994 cpi->last_frame_distortion = cpi->frame_distortion;
5998 static INTERP_FILTER get_interp_filter(
5999 const int64_t threshes[SWITCHABLE_FILTER_CONTEXTS], int is_alt_ref) {
6000 if (!is_alt_ref && threshes[EIGHTTAP_SMOOTH] > threshes[EIGHTTAP] &&
6001 threshes[EIGHTTAP_SMOOTH] > threshes[EIGHTTAP_SHARP] &&
6002 threshes[EIGHTTAP_SMOOTH] > threshes[SWITCHABLE - 1]) {
6003 return EIGHTTAP_SMOOTH;
6004 } else if (threshes[EIGHTTAP_SHARP] > threshes[EIGHTTAP] &&
6005 threshes[EIGHTTAP_SHARP] > threshes[SWITCHABLE - 1]) {
6006 return EIGHTTAP_SHARP;
6007 } else if (threshes[EIGHTTAP] > threshes[SWITCHABLE - 1]) {
6014 static int compute_frame_aq_offset(struct VP9_COMP *cpi) {
6015 VP9_COMMON *const cm = &cpi->common;
6016 MODE_INFO **mi_8x8_ptr = cm->mi_grid_visible;
6017 struct segmentation *const seg = &cm->seg;
6025 for (mi_row = 0; mi_row < cm->mi_rows; mi_row++) {
6026 MODE_INFO **mi_8x8 = mi_8x8_ptr;
6027 for (mi_col = 0; mi_col < cm->mi_cols; mi_col++, mi_8x8++) {
6028 segment_id = mi_8x8[0]->segment_id;
6029 qdelta_index = get_segdata(seg, segment_id, SEG_LVL_ALT_Q);
6030 sum_delta += qdelta_index;
6033 mi_8x8_ptr += cm->mi_stride;
6036 return sum_delta / (cm->mi_rows * cm->mi_cols);
6039 #if CONFIG_CONSISTENT_RECODE
6040 static void restore_encode_params(VP9_COMP *cpi) {
6041 VP9_COMMON *const cm = &cpi->common;
6042 const int tile_cols = 1 << cm->log2_tile_cols;
6043 const int tile_rows = 1 << cm->log2_tile_rows;
6044 int tile_col, tile_row;
6046 RD_OPT *rd_opt = &cpi->rd;
6047 for (i = 0; i < MAX_REF_FRAMES; i++) {
6048 for (j = 0; j < REFERENCE_MODES; j++)
6049 rd_opt->prediction_type_threshes[i][j] =
6050 rd_opt->prediction_type_threshes_prev[i][j];
6052 for (j = 0; j < SWITCHABLE_FILTER_CONTEXTS; j++)
6053 rd_opt->filter_threshes[i][j] = rd_opt->filter_threshes_prev[i][j];
6056 if (cpi->tile_data != NULL) {
6057 for (tile_row = 0; tile_row < tile_rows; ++tile_row)
6058 for (tile_col = 0; tile_col < tile_cols; ++tile_col) {
6059 TileDataEnc *tile_data =
6060 &cpi->tile_data[tile_row * tile_cols + tile_col];
6061 for (i = 0; i < BLOCK_SIZES; ++i) {
6062 for (j = 0; j < MAX_MODES; ++j) {
6063 tile_data->thresh_freq_fact[i][j] =
6064 tile_data->thresh_freq_fact_prev[i][j];
6070 cm->interp_filter = cpi->sf.default_interp_filter;
6074 void vp9_encode_frame(VP9_COMP *cpi) {
6075 VP9_COMMON *const cm = &cpi->common;
6077 #if CONFIG_CONSISTENT_RECODE
6078 restore_encode_params(cpi);
6081 // In the longer term the encoder should be generalized to match the
6082 // decoder such that we allow compound where one of the 3 buffers has a
6083 // different sign bias and that buffer is then the fixed ref. However, this
6084 // requires further work in the rd loop. For now the only supported encoder
6085 // side behavior is where the ALT ref buffer has opposite sign bias to
6087 if (!frame_is_intra_only(cm)) {
6088 if (vp9_compound_reference_allowed(cm)) {
6089 cpi->allow_comp_inter_inter = 1;
6090 vp9_setup_compound_reference_mode(cm);
6092 cpi->allow_comp_inter_inter = 0;
6096 if (cpi->sf.frame_parameter_update) {
6098 RD_OPT *const rd_opt = &cpi->rd;
6099 FRAME_COUNTS *counts = cpi->td.counts;
6100 RD_COUNTS *const rdc = &cpi->td.rd_counts;
6102 // This code does a single RD pass over the whole frame assuming
6103 // either compound, single or hybrid prediction as per whatever has
6104 // worked best for that type of frame in the past.
6105 // It also predicts whether another coding mode would have worked
6106 // better than this coding mode. If that is the case, it remembers
6107 // that for subsequent frames.
6108 // It also does the same analysis for transform size selection.
6109 const MV_REFERENCE_FRAME frame_type = get_frame_type(cpi);
6110 int64_t *const mode_thrs = rd_opt->prediction_type_threshes[frame_type];
6111 int64_t *const filter_thrs = rd_opt->filter_threshes[frame_type];
6112 const int is_alt_ref = frame_type == ALTREF_FRAME;
6114 /* prediction (compound, single or hybrid) mode selection */
6115 if (is_alt_ref || !cpi->allow_comp_inter_inter)
6116 cm->reference_mode = SINGLE_REFERENCE;
6117 else if (mode_thrs[COMPOUND_REFERENCE] > mode_thrs[SINGLE_REFERENCE] &&
6118 mode_thrs[COMPOUND_REFERENCE] > mode_thrs[REFERENCE_MODE_SELECT] &&
6119 check_dual_ref_flags(cpi) && cpi->static_mb_pct == 100)
6120 cm->reference_mode = COMPOUND_REFERENCE;
6121 else if (mode_thrs[SINGLE_REFERENCE] > mode_thrs[REFERENCE_MODE_SELECT])
6122 cm->reference_mode = SINGLE_REFERENCE;
6124 cm->reference_mode = REFERENCE_MODE_SELECT;
6126 if (cm->interp_filter == SWITCHABLE)
6127 cm->interp_filter = get_interp_filter(filter_thrs, is_alt_ref);
6129 encode_frame_internal(cpi);
6131 for (i = 0; i < REFERENCE_MODES; ++i)
6132 mode_thrs[i] = (mode_thrs[i] + rdc->comp_pred_diff[i] / cm->MBs) / 2;
6134 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i)
6135 filter_thrs[i] = (filter_thrs[i] + rdc->filter_diff[i] / cm->MBs) / 2;
6137 if (cm->reference_mode == REFERENCE_MODE_SELECT) {
6138 int single_count_zero = 0;
6139 int comp_count_zero = 0;
6141 for (i = 0; i < COMP_INTER_CONTEXTS; i++) {
6142 single_count_zero += counts->comp_inter[i][0];
6143 comp_count_zero += counts->comp_inter[i][1];
6146 if (comp_count_zero == 0) {
6147 cm->reference_mode = SINGLE_REFERENCE;
6148 vp9_zero(counts->comp_inter);
6149 } else if (single_count_zero == 0) {
6150 cm->reference_mode = COMPOUND_REFERENCE;
6151 vp9_zero(counts->comp_inter);
6155 if (cm->tx_mode == TX_MODE_SELECT) {
6157 int count8x8_lp = 0, count8x8_8x8p = 0;
6158 int count16x16_16x16p = 0, count16x16_lp = 0;
6161 for (i = 0; i < TX_SIZE_CONTEXTS; ++i) {
6162 count4x4 += counts->tx.p32x32[i][TX_4X4];
6163 count4x4 += counts->tx.p16x16[i][TX_4X4];
6164 count4x4 += counts->tx.p8x8[i][TX_4X4];
6166 count8x8_lp += counts->tx.p32x32[i][TX_8X8];
6167 count8x8_lp += counts->tx.p16x16[i][TX_8X8];
6168 count8x8_8x8p += counts->tx.p8x8[i][TX_8X8];
6170 count16x16_16x16p += counts->tx.p16x16[i][TX_16X16];
6171 count16x16_lp += counts->tx.p32x32[i][TX_16X16];
6172 count32x32 += counts->tx.p32x32[i][TX_32X32];
6174 if (count4x4 == 0 && count16x16_lp == 0 && count16x16_16x16p == 0 &&
6176 cm->tx_mode = ALLOW_8X8;
6177 reset_skip_tx_size(cm, TX_8X8);
6178 } else if (count8x8_8x8p == 0 && count16x16_16x16p == 0 &&
6179 count8x8_lp == 0 && count16x16_lp == 0 && count32x32 == 0) {
6180 cm->tx_mode = ONLY_4X4;
6181 reset_skip_tx_size(cm, TX_4X4);
6182 } else if (count8x8_lp == 0 && count16x16_lp == 0 && count4x4 == 0) {
6183 cm->tx_mode = ALLOW_32X32;
6184 } else if (count32x32 == 0 && count8x8_lp == 0 && count4x4 == 0) {
6185 cm->tx_mode = ALLOW_16X16;
6186 reset_skip_tx_size(cm, TX_16X16);
6190 FRAME_COUNTS *counts = cpi->td.counts;
6191 cm->reference_mode = SINGLE_REFERENCE;
6192 if (cpi->allow_comp_inter_inter && cpi->sf.use_compound_nonrd_pickmode &&
6193 cpi->rc.alt_ref_gf_group && !cpi->rc.is_src_frame_alt_ref &&
6194 cm->frame_type != KEY_FRAME)
6195 cm->reference_mode = REFERENCE_MODE_SELECT;
6197 encode_frame_internal(cpi);
6199 if (cm->reference_mode == REFERENCE_MODE_SELECT) {
6200 int single_count_zero = 0;
6201 int comp_count_zero = 0;
6203 for (i = 0; i < COMP_INTER_CONTEXTS; i++) {
6204 single_count_zero += counts->comp_inter[i][0];
6205 comp_count_zero += counts->comp_inter[i][1];
6207 if (comp_count_zero == 0) {
6208 cm->reference_mode = SINGLE_REFERENCE;
6209 vp9_zero(counts->comp_inter);
6210 } else if (single_count_zero == 0) {
6211 cm->reference_mode = COMPOUND_REFERENCE;
6212 vp9_zero(counts->comp_inter);
6217 // If segmented AQ is enabled compute the average AQ weighting.
6218 if (cm->seg.enabled && (cpi->oxcf.aq_mode != NO_AQ) &&
6219 (cm->seg.update_map || cm->seg.update_data)) {
6220 cm->seg.aq_av_offset = compute_frame_aq_offset(cpi);
6224 static void sum_intra_stats(FRAME_COUNTS *counts, const MODE_INFO *mi) {
6225 const PREDICTION_MODE y_mode = mi->mode;
6226 const PREDICTION_MODE uv_mode = mi->uv_mode;
6227 const BLOCK_SIZE bsize = mi->sb_type;
6229 if (bsize < BLOCK_8X8) {
6231 const int num_4x4_w = num_4x4_blocks_wide_lookup[bsize];
6232 const int num_4x4_h = num_4x4_blocks_high_lookup[bsize];
6233 for (idy = 0; idy < 2; idy += num_4x4_h)
6234 for (idx = 0; idx < 2; idx += num_4x4_w)
6235 ++counts->y_mode[0][mi->bmi[idy * 2 + idx].as_mode];
6237 ++counts->y_mode[size_group_lookup[bsize]][y_mode];
6240 ++counts->uv_mode[y_mode][uv_mode];
6243 static void update_zeromv_cnt(VP9_COMP *const cpi, const MODE_INFO *const mi,
6244 int mi_row, int mi_col, BLOCK_SIZE bsize) {
6245 const VP9_COMMON *const cm = &cpi->common;
6246 MV mv = mi->mv[0].as_mv;
6247 const int bw = num_8x8_blocks_wide_lookup[bsize];
6248 const int bh = num_8x8_blocks_high_lookup[bsize];
6249 const int xmis = VPXMIN(cm->mi_cols - mi_col, bw);
6250 const int ymis = VPXMIN(cm->mi_rows - mi_row, bh);
6251 const int block_index = mi_row * cm->mi_cols + mi_col;
6253 for (y = 0; y < ymis; y++)
6254 for (x = 0; x < xmis; x++) {
6255 int map_offset = block_index + y * cm->mi_cols + x;
6256 if (mi->ref_frame[0] == LAST_FRAME && is_inter_block(mi) &&
6257 mi->segment_id <= CR_SEGMENT_ID_BOOST2) {
6258 if (abs(mv.row) < 8 && abs(mv.col) < 8) {
6259 if (cpi->consec_zero_mv[map_offset] < 255)
6260 cpi->consec_zero_mv[map_offset]++;
6262 cpi->consec_zero_mv[map_offset] = 0;
6268 static void encode_superblock(VP9_COMP *cpi, ThreadData *td, TOKENEXTRA **t,
6269 int output_enabled, int mi_row, int mi_col,
6270 BLOCK_SIZE bsize, PICK_MODE_CONTEXT *ctx) {
6271 VP9_COMMON *const cm = &cpi->common;
6272 MACROBLOCK *const x = &td->mb;
6273 MACROBLOCKD *const xd = &x->e_mbd;
6274 MODE_INFO *mi = xd->mi[0];
6275 const int seg_skip =
6276 segfeature_active(&cm->seg, mi->segment_id, SEG_LVL_SKIP);
6277 x->skip_recode = !x->select_tx_size && mi->sb_type >= BLOCK_8X8 &&
6278 cpi->oxcf.aq_mode != COMPLEXITY_AQ &&
6279 cpi->oxcf.aq_mode != CYCLIC_REFRESH_AQ &&
6280 cpi->sf.allow_skip_recode;
6282 if (!x->skip_recode && !cpi->sf.use_nonrd_pick_mode)
6283 memset(x->skip_txfm, 0, sizeof(x->skip_txfm));
6285 x->skip_optimize = ctx->is_coded;
6287 x->use_lp32x32fdct = cpi->sf.use_lp32x32fdct;
6288 x->skip_encode = (!output_enabled && cpi->sf.skip_encode_frame &&
6289 x->q_index < QIDX_SKIP_THRESH);
6291 if (x->skip_encode) return;
6293 if (!is_inter_block(mi)) {
6295 #if CONFIG_BETTER_HW_COMPATIBILITY && CONFIG_VP9_HIGHBITDEPTH
6296 if ((xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) &&
6297 (xd->above_mi == NULL || xd->left_mi == NULL) &&
6298 need_top_left[mi->uv_mode])
6300 #endif // CONFIG_BETTER_HW_COMPATIBILITY && CONFIG_VP9_HIGHBITDEPTH
6302 for (plane = 0; plane < MAX_MB_PLANE; ++plane)
6303 vp9_encode_intra_block_plane(x, VPXMAX(bsize, BLOCK_8X8), plane, 1);
6304 if (output_enabled) sum_intra_stats(td->counts, mi);
6305 vp9_tokenize_sb(cpi, td, t, !output_enabled, seg_skip,
6306 VPXMAX(bsize, BLOCK_8X8));
6309 const int is_compound = has_second_ref(mi);
6310 set_ref_ptrs(cm, xd, mi->ref_frame[0], mi->ref_frame[1]);
6311 for (ref = 0; ref < 1 + is_compound; ++ref) {
6312 YV12_BUFFER_CONFIG *cfg = get_ref_frame_buffer(cpi, mi->ref_frame[ref]);
6313 assert(cfg != NULL);
6314 vp9_setup_pre_planes(xd, ref, cfg, mi_row, mi_col,
6315 &xd->block_refs[ref]->sf);
6317 if (!(cpi->sf.reuse_inter_pred_sby && ctx->pred_pixel_ready) || seg_skip)
6318 vp9_build_inter_predictors_sby(xd, mi_row, mi_col,
6319 VPXMAX(bsize, BLOCK_8X8));
6321 vp9_build_inter_predictors_sbuv(xd, mi_row, mi_col,
6322 VPXMAX(bsize, BLOCK_8X8));
6324 vp9_encode_sb(x, VPXMAX(bsize, BLOCK_8X8));
6325 vp9_tokenize_sb(cpi, td, t, !output_enabled, seg_skip,
6326 VPXMAX(bsize, BLOCK_8X8));
6333 if (output_enabled) {
6334 if (cm->tx_mode == TX_MODE_SELECT && mi->sb_type >= BLOCK_8X8 &&
6335 !(is_inter_block(mi) && mi->skip)) {
6336 ++get_tx_counts(max_txsize_lookup[bsize], get_tx_size_context(xd),
6337 &td->counts->tx)[mi->tx_size];
6339 // The new intra coding scheme requires no change of transform size
6340 if (is_inter_block(mi)) {
6341 mi->tx_size = VPXMIN(tx_mode_to_biggest_tx_size[cm->tx_mode],
6342 max_txsize_lookup[bsize]);
6344 mi->tx_size = (bsize >= BLOCK_8X8) ? mi->tx_size : TX_4X4;
6348 ++td->counts->tx.tx_totals[mi->tx_size];
6349 ++td->counts->tx.tx_totals[get_uv_tx_size(mi, &xd->plane[1])];
6350 if (cm->seg.enabled && cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ)
6351 vp9_cyclic_refresh_update_sb_postencode(cpi, mi, mi_row, mi_col, bsize);
6352 if (cpi->oxcf.pass == 0 && cpi->svc.temporal_layer_id == 0 &&
6355 !cpi->svc.layer_context[cpi->svc.temporal_layer_id].is_key_frame &&
6356 cpi->svc.spatial_layer_id == cpi->svc.number_spatial_layers - 1)))
6357 update_zeromv_cnt(cpi, mi, mi_row, mi_col, bsize);