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.
15 #include "./vp9_rtcd.h"
16 #include "./vpx_config.h"
18 #include "vpx_ports/mem.h"
19 #include "vpx_ports/vpx_timer.h"
21 #include "vp9/common/vp9_common.h"
22 #include "vp9/common/vp9_entropy.h"
23 #include "vp9/common/vp9_entropymode.h"
24 #include "vp9/common/vp9_idct.h"
25 #include "vp9/common/vp9_mvref_common.h"
26 #include "vp9/common/vp9_pred_common.h"
27 #include "vp9/common/vp9_quant_common.h"
28 #include "vp9/common/vp9_reconintra.h"
29 #include "vp9/common/vp9_reconinter.h"
30 #include "vp9/common/vp9_seg_common.h"
31 #include "vp9/common/vp9_systemdependent.h"
32 #include "vp9/common/vp9_tile_common.h"
34 #include "vp9/encoder/vp9_aq_complexity.h"
35 #include "vp9/encoder/vp9_aq_cyclicrefresh.h"
36 #include "vp9/encoder/vp9_aq_variance.h"
37 #include "vp9/encoder/vp9_encodeframe.h"
38 #include "vp9/encoder/vp9_encodemb.h"
39 #include "vp9/encoder/vp9_encodemv.h"
40 #include "vp9/encoder/vp9_ethread.h"
41 #include "vp9/encoder/vp9_extend.h"
42 #include "vp9/encoder/vp9_pickmode.h"
43 #include "vp9/encoder/vp9_rd.h"
44 #include "vp9/encoder/vp9_rdopt.h"
45 #include "vp9/encoder/vp9_segmentation.h"
46 #include "vp9/encoder/vp9_tokenize.h"
48 static void encode_superblock(VP9_COMP *cpi, ThreadData * td,
49 TOKENEXTRA **t, int output_enabled,
50 int mi_row, int mi_col, BLOCK_SIZE bsize,
51 PICK_MODE_CONTEXT *ctx);
53 // This is used as a reference when computing the source variance for the
54 // purposes of activity masking.
55 // Eventually this should be replaced by custom no-reference routines,
56 // which will be faster.
57 static const uint8_t VP9_VAR_OFFS[64] = {
58 128, 128, 128, 128, 128, 128, 128, 128,
59 128, 128, 128, 128, 128, 128, 128, 128,
60 128, 128, 128, 128, 128, 128, 128, 128,
61 128, 128, 128, 128, 128, 128, 128, 128,
62 128, 128, 128, 128, 128, 128, 128, 128,
63 128, 128, 128, 128, 128, 128, 128, 128,
64 128, 128, 128, 128, 128, 128, 128, 128,
65 128, 128, 128, 128, 128, 128, 128, 128
68 #if CONFIG_VP9_HIGHBITDEPTH
69 static const uint16_t VP9_HIGH_VAR_OFFS_8[64] = {
70 128, 128, 128, 128, 128, 128, 128, 128,
71 128, 128, 128, 128, 128, 128, 128, 128,
72 128, 128, 128, 128, 128, 128, 128, 128,
73 128, 128, 128, 128, 128, 128, 128, 128,
74 128, 128, 128, 128, 128, 128, 128, 128,
75 128, 128, 128, 128, 128, 128, 128, 128,
76 128, 128, 128, 128, 128, 128, 128, 128,
77 128, 128, 128, 128, 128, 128, 128, 128
80 static const uint16_t VP9_HIGH_VAR_OFFS_10[64] = {
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,
88 128*4, 128*4, 128*4, 128*4, 128*4, 128*4, 128*4, 128*4
91 static const uint16_t VP9_HIGH_VAR_OFFS_12[64] = {
92 128*16, 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, 128*16,
94 128*16, 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, 128*16,
96 128*16, 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, 128*16,
98 128*16, 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, 128*16
101 #endif // CONFIG_VP9_HIGHBITDEPTH
103 unsigned int vp9_get_sby_perpixel_variance(VP9_COMP *cpi,
104 const struct buf_2d *ref,
107 const unsigned int var = cpi->fn_ptr[bs].vf(ref->buf, ref->stride,
108 VP9_VAR_OFFS, 0, &sse);
109 return ROUND_POWER_OF_TWO(var, num_pels_log2_lookup[bs]);
112 #if CONFIG_VP9_HIGHBITDEPTH
113 unsigned int vp9_high_get_sby_perpixel_variance(
114 VP9_COMP *cpi, const struct buf_2d *ref, BLOCK_SIZE bs, int bd) {
115 unsigned int var, sse;
118 var = cpi->fn_ptr[bs].vf(ref->buf, ref->stride,
119 CONVERT_TO_BYTEPTR(VP9_HIGH_VAR_OFFS_10),
123 var = cpi->fn_ptr[bs].vf(ref->buf, ref->stride,
124 CONVERT_TO_BYTEPTR(VP9_HIGH_VAR_OFFS_12),
129 var = cpi->fn_ptr[bs].vf(ref->buf, ref->stride,
130 CONVERT_TO_BYTEPTR(VP9_HIGH_VAR_OFFS_8),
134 return ROUND_POWER_OF_TWO(var, num_pels_log2_lookup[bs]);
136 #endif // CONFIG_VP9_HIGHBITDEPTH
138 static unsigned int get_sby_perpixel_diff_variance(VP9_COMP *cpi,
139 const struct buf_2d *ref,
140 int mi_row, int mi_col,
142 unsigned int sse, var;
144 const YV12_BUFFER_CONFIG *last = get_ref_frame_buffer(cpi, LAST_FRAME);
146 assert(last != NULL);
148 &last->y_buffer[mi_row * MI_SIZE * last->y_stride + mi_col * MI_SIZE];
149 var = cpi->fn_ptr[bs].vf(ref->buf, ref->stride, last_y, last->y_stride, &sse);
150 return ROUND_POWER_OF_TWO(var, num_pels_log2_lookup[bs]);
153 static BLOCK_SIZE get_rd_var_based_fixed_partition(VP9_COMP *cpi, MACROBLOCK *x,
156 unsigned int var = get_sby_perpixel_diff_variance(cpi, &x->plane[0].src,
169 // Lighter version of set_offsets that only sets the mode info
171 static INLINE void set_mode_info_offsets(VP9_COMMON *const cm,
172 MACROBLOCKD *const xd,
175 const int idx_str = xd->mi_stride * mi_row + mi_col;
176 xd->mi = cm->mi_grid_visible + idx_str;
177 xd->mi[0] = cm->mi + idx_str;
180 static void set_offsets(VP9_COMP *cpi, const TileInfo *const tile,
181 MACROBLOCK *const x, int mi_row, int mi_col,
183 VP9_COMMON *const cm = &cpi->common;
184 MACROBLOCKD *const xd = &x->e_mbd;
186 const int mi_width = num_8x8_blocks_wide_lookup[bsize];
187 const int mi_height = num_8x8_blocks_high_lookup[bsize];
188 const struct segmentation *const seg = &cm->seg;
190 set_skip_context(xd, mi_row, mi_col);
192 set_mode_info_offsets(cm, xd, mi_row, mi_col);
194 mbmi = &xd->mi[0]->mbmi;
196 // Set up destination pointers.
197 vp9_setup_dst_planes(xd->plane, get_frame_new_buffer(cm), mi_row, mi_col);
199 // Set up limit values for MV components.
200 // Mv beyond the range do not produce new/different prediction block.
201 x->mv_row_min = -(((mi_row + mi_height) * MI_SIZE) + VP9_INTERP_EXTEND);
202 x->mv_col_min = -(((mi_col + mi_width) * MI_SIZE) + VP9_INTERP_EXTEND);
203 x->mv_row_max = (cm->mi_rows - mi_row) * MI_SIZE + VP9_INTERP_EXTEND;
204 x->mv_col_max = (cm->mi_cols - mi_col) * MI_SIZE + VP9_INTERP_EXTEND;
206 // Set up distance of MB to edge of frame in 1/8th pel units.
207 assert(!(mi_col & (mi_width - 1)) && !(mi_row & (mi_height - 1)));
208 set_mi_row_col(xd, tile, mi_row, mi_height, mi_col, mi_width,
209 cm->mi_rows, cm->mi_cols);
211 // Set up source buffers.
212 vp9_setup_src_planes(x, cpi->Source, mi_row, mi_col);
215 x->rddiv = cpi->rd.RDDIV;
216 x->rdmult = cpi->rd.RDMULT;
220 if (cpi->oxcf.aq_mode != VARIANCE_AQ) {
221 const uint8_t *const map = seg->update_map ? cpi->segmentation_map
222 : cm->last_frame_seg_map;
223 mbmi->segment_id = vp9_get_segment_id(cm, map, bsize, mi_row, mi_col);
225 vp9_init_plane_quantizers(cpi, x);
227 x->encode_breakout = cpi->segment_encode_breakout[mbmi->segment_id];
229 mbmi->segment_id = 0;
230 x->encode_breakout = cpi->encode_breakout;
234 static void duplicate_mode_info_in_sb(VP9_COMMON *cm, MACROBLOCKD *xd,
235 int mi_row, int mi_col,
237 const int block_width = num_8x8_blocks_wide_lookup[bsize];
238 const int block_height = num_8x8_blocks_high_lookup[bsize];
240 for (j = 0; j < block_height; ++j)
241 for (i = 0; i < block_width; ++i) {
242 if (mi_row + j < cm->mi_rows && mi_col + i < cm->mi_cols)
243 xd->mi[j * xd->mi_stride + i] = xd->mi[0];
247 static void set_block_size(VP9_COMP * const cpi,
248 MACROBLOCKD *const xd,
249 int mi_row, int mi_col,
251 if (cpi->common.mi_cols > mi_col && cpi->common.mi_rows > mi_row) {
252 set_mode_info_offsets(&cpi->common, xd, mi_row, mi_col);
253 xd->mi[0]->mbmi.sb_type = bsize;
258 int64_t sum_square_error;
268 } partition_variance;
271 partition_variance part_variances;
276 partition_variance part_variances;
281 partition_variance part_variances;
286 partition_variance part_variances;
291 partition_variance part_variances;
296 partition_variance *part_variances;
306 static void tree_to_node(void *data, BLOCK_SIZE bsize, variance_node *node) {
308 node->part_variances = NULL;
311 v64x64 *vt = (v64x64 *) data;
312 node->part_variances = &vt->part_variances;
313 for (i = 0; i < 4; i++)
314 node->split[i] = &vt->split[i].part_variances.none;
318 v32x32 *vt = (v32x32 *) data;
319 node->part_variances = &vt->part_variances;
320 for (i = 0; i < 4; i++)
321 node->split[i] = &vt->split[i].part_variances.none;
325 v16x16 *vt = (v16x16 *) data;
326 node->part_variances = &vt->part_variances;
327 for (i = 0; i < 4; i++)
328 node->split[i] = &vt->split[i].part_variances.none;
332 v8x8 *vt = (v8x8 *) data;
333 node->part_variances = &vt->part_variances;
334 for (i = 0; i < 4; i++)
335 node->split[i] = &vt->split[i].part_variances.none;
339 v4x4 *vt = (v4x4 *) data;
340 node->part_variances = &vt->part_variances;
341 for (i = 0; i < 4; i++)
342 node->split[i] = &vt->split[i];
352 // Set variance values given sum square error, sum error, count.
353 static void fill_variance(int64_t s2, int64_t s, int c, var *v) {
354 v->sum_square_error = s2;
359 static void get_variance(var *v) {
360 v->variance = (int)(256 * (v->sum_square_error -
361 ((v->sum_error * v->sum_error) >> v->log2_count)) >> v->log2_count);
364 static void sum_2_variances(const var *a, const var *b, var *r) {
365 assert(a->log2_count == b->log2_count);
366 fill_variance(a->sum_square_error + b->sum_square_error,
367 a->sum_error + b->sum_error, a->log2_count + 1, r);
370 static void fill_variance_tree(void *data, BLOCK_SIZE bsize) {
372 tree_to_node(data, bsize, &node);
373 sum_2_variances(node.split[0], node.split[1], &node.part_variances->horz[0]);
374 sum_2_variances(node.split[2], node.split[3], &node.part_variances->horz[1]);
375 sum_2_variances(node.split[0], node.split[2], &node.part_variances->vert[0]);
376 sum_2_variances(node.split[1], node.split[3], &node.part_variances->vert[1]);
377 sum_2_variances(&node.part_variances->vert[0], &node.part_variances->vert[1],
378 &node.part_variances->none);
381 static int set_vt_partitioning(VP9_COMP *cpi,
382 MACROBLOCKD *const xd,
388 BLOCK_SIZE bsize_min,
390 VP9_COMMON * const cm = &cpi->common;
392 const int block_width = num_8x8_blocks_wide_lookup[bsize];
393 const int block_height = num_8x8_blocks_high_lookup[bsize];
394 const int low_res = (cm->width <= 352 && cm->height <= 288);
396 assert(block_height == block_width);
397 tree_to_node(data, bsize, &vt);
399 if (force_split == 1)
402 // For bsize=bsize_min (16x16/8x8 for 8x8/4x4 downsampling), select if
403 // variance is below threshold, otherwise split will be selected.
404 // No check for vert/horiz split as too few samples for variance.
405 if (bsize == bsize_min) {
406 // Variance already computed to set the force_split.
407 if (low_res || cm->frame_type == KEY_FRAME)
408 get_variance(&vt.part_variances->none);
409 if (mi_col + block_width / 2 < cm->mi_cols &&
410 mi_row + block_height / 2 < cm->mi_rows &&
411 vt.part_variances->none.variance < threshold) {
412 set_block_size(cpi, xd, mi_row, mi_col, bsize);
416 } else if (bsize > bsize_min) {
417 // Variance already computed to set the force_split.
418 if (low_res || cm->frame_type == KEY_FRAME)
419 get_variance(&vt.part_variances->none);
420 // For key frame: take split for bsize above 32X32 or very high variance.
421 if (cm->frame_type == KEY_FRAME &&
422 (bsize > BLOCK_32X32 ||
423 vt.part_variances->none.variance > (threshold << 4))) {
426 // If variance is low, take the bsize (no split).
427 if (mi_col + block_width / 2 < cm->mi_cols &&
428 mi_row + block_height / 2 < cm->mi_rows &&
429 vt.part_variances->none.variance < threshold) {
430 set_block_size(cpi, xd, mi_row, mi_col, bsize);
434 // Check vertical split.
435 if (mi_row + block_height / 2 < cm->mi_rows) {
436 BLOCK_SIZE subsize = get_subsize(bsize, PARTITION_VERT);
437 get_variance(&vt.part_variances->vert[0]);
438 get_variance(&vt.part_variances->vert[1]);
439 if (vt.part_variances->vert[0].variance < threshold &&
440 vt.part_variances->vert[1].variance < threshold &&
441 get_plane_block_size(subsize, &xd->plane[1]) < BLOCK_INVALID) {
442 set_block_size(cpi, xd, mi_row, mi_col, subsize);
443 set_block_size(cpi, xd, mi_row, mi_col + block_width / 2, subsize);
447 // Check horizontal split.
448 if (mi_col + block_width / 2 < cm->mi_cols) {
449 BLOCK_SIZE subsize = get_subsize(bsize, PARTITION_HORZ);
450 get_variance(&vt.part_variances->horz[0]);
451 get_variance(&vt.part_variances->horz[1]);
452 if (vt.part_variances->horz[0].variance < threshold &&
453 vt.part_variances->horz[1].variance < threshold &&
454 get_plane_block_size(subsize, &xd->plane[1]) < BLOCK_INVALID) {
455 set_block_size(cpi, xd, mi_row, mi_col, subsize);
456 set_block_size(cpi, xd, mi_row + block_height / 2, mi_col, subsize);
466 void vp9_set_vbp_thresholds(VP9_COMP *cpi, int q) {
467 SPEED_FEATURES *const sf = &cpi->sf;
468 if (sf->partition_search_type != VAR_BASED_PARTITION &&
469 sf->partition_search_type != REFERENCE_PARTITION) {
472 VP9_COMMON *const cm = &cpi->common;
473 const int is_key_frame = (cm->frame_type == KEY_FRAME);
474 const int threshold_multiplier = is_key_frame ? 20 : 1;
475 const int64_t threshold_base = (int64_t)(threshold_multiplier *
476 cpi->y_dequant[q][1]);
478 // TODO(marpan): Allow 4x4 partitions for inter-frames.
479 // use_4x4_partition = (variance4x4downsample[i2 + j] == 1);
480 // If 4x4 partition is not used, then 8x8 partition will be selected
481 // if variance of 16x16 block is very high, so use larger threshold
482 // for 16x16 (threshold_bsize_min) in that case.
484 // Array index: 0 - threshold_64x64; 1 - threshold_32x32;
485 // 2 - threshold_16x16; 3 - vbp_threshold_8x8;
487 cpi->vbp_thresholds[0] = threshold_base;
488 cpi->vbp_thresholds[1] = threshold_base >> 2;
489 cpi->vbp_thresholds[2] = threshold_base >> 2;
490 cpi->vbp_thresholds[3] = threshold_base << 2;
491 cpi->vbp_threshold_sad = 0;
492 cpi->vbp_bsize_min = BLOCK_8X8;
494 cpi->vbp_thresholds[1] = threshold_base;
495 if (cm->width <= 352 && cm->height <= 288) {
496 cpi->vbp_thresholds[0] = threshold_base >> 2;
497 cpi->vbp_thresholds[2] = threshold_base << 3;
498 cpi->vbp_threshold_sad = 100;
500 cpi->vbp_thresholds[0] = threshold_base;
501 cpi->vbp_thresholds[1] = (5 * threshold_base) >> 2;
502 cpi->vbp_thresholds[2] = threshold_base << cpi->oxcf.speed;
503 cpi->vbp_threshold_sad = (cpi->y_dequant[q][1] << 1) > 1000 ?
504 (cpi->y_dequant[q][1] << 1) : 1000;
506 cpi->vbp_bsize_min = BLOCK_16X16;
508 cpi->vbp_threshold_minmax = 15 + (q >> 3);
512 // Compute the minmax over the 8x8 subblocks.
513 static int compute_minmax_8x8(const uint8_t *s, int sp, const uint8_t *d,
514 int dp, int x16_idx, int y16_idx,
515 #if CONFIG_VP9_HIGHBITDEPTH
522 int minmax_min = 255;
523 // Loop over the 4 8x8 subblocks.
524 for (k = 0; k < 4; k++) {
525 int x8_idx = x16_idx + ((k & 1) << 3);
526 int y8_idx = y16_idx + ((k >> 1) << 3);
529 if (x8_idx < pixels_wide && y8_idx < pixels_high) {
530 #if CONFIG_VP9_HIGHBITDEPTH
531 if (highbd_flag & YV12_FLAG_HIGHBITDEPTH) {
532 vp9_highbd_minmax_8x8(s + y8_idx * sp + x8_idx, sp,
533 d + y8_idx * dp + x8_idx, dp,
536 vp9_minmax_8x8(s + y8_idx * sp + x8_idx, sp,
537 d + y8_idx * dp + x8_idx, dp,
541 vp9_minmax_8x8(s + y8_idx * sp + x8_idx, sp,
542 d + y8_idx * dp + x8_idx, dp,
545 if ((max - min) > minmax_max)
546 minmax_max = (max - min);
547 if ((max - min) < minmax_min)
548 minmax_min = (max - min);
551 return (minmax_max - minmax_min);
554 static void modify_vbp_thresholds(VP9_COMP *cpi, int64_t thresholds[], int q) {
555 VP9_COMMON *const cm = &cpi->common;
556 const int64_t threshold_base = (int64_t)(cpi->y_dequant[q][1]);
558 // Array index: 0 - threshold_64x64; 1 - threshold_32x32;
559 // 2 - threshold_16x16; 3 - vbp_threshold_8x8;
560 thresholds[1] = threshold_base;
561 if (cm->width <= 352 && cm->height <= 288) {
562 thresholds[0] = threshold_base >> 2;
563 thresholds[2] = threshold_base << 3;
565 thresholds[0] = threshold_base;
566 thresholds[1] = (5 * threshold_base) >> 2;
567 thresholds[2] = threshold_base << cpi->oxcf.speed;
571 static void fill_variance_4x4avg(const uint8_t *s, int sp, const uint8_t *d,
572 int dp, int x8_idx, int y8_idx, v8x8 *vst,
573 #if CONFIG_VP9_HIGHBITDEPTH
580 for (k = 0; k < 4; k++) {
581 int x4_idx = x8_idx + ((k & 1) << 2);
582 int y4_idx = y8_idx + ((k >> 1) << 2);
583 unsigned int sse = 0;
585 if (x4_idx < pixels_wide && y4_idx < pixels_high) {
588 #if CONFIG_VP9_HIGHBITDEPTH
589 if (highbd_flag & YV12_FLAG_HIGHBITDEPTH) {
590 s_avg = vp9_highbd_avg_4x4(s + y4_idx * sp + x4_idx, sp);
592 d_avg = vp9_highbd_avg_4x4(d + y4_idx * dp + x4_idx, dp);
594 s_avg = vp9_avg_4x4(s + y4_idx * sp + x4_idx, sp);
596 d_avg = vp9_avg_4x4(d + y4_idx * dp + x4_idx, dp);
599 s_avg = vp9_avg_4x4(s + y4_idx * sp + x4_idx, sp);
601 d_avg = vp9_avg_4x4(d + y4_idx * dp + x4_idx, dp);
606 fill_variance(sse, sum, 0, &vst->split[k].part_variances.none);
610 static void fill_variance_8x8avg(const uint8_t *s, int sp, const uint8_t *d,
611 int dp, int x16_idx, int y16_idx, v16x16 *vst,
612 #if CONFIG_VP9_HIGHBITDEPTH
619 for (k = 0; k < 4; k++) {
620 int x8_idx = x16_idx + ((k & 1) << 3);
621 int y8_idx = y16_idx + ((k >> 1) << 3);
622 unsigned int sse = 0;
624 if (x8_idx < pixels_wide && y8_idx < pixels_high) {
627 #if CONFIG_VP9_HIGHBITDEPTH
628 if (highbd_flag & YV12_FLAG_HIGHBITDEPTH) {
629 s_avg = vp9_highbd_avg_8x8(s + y8_idx * sp + x8_idx, sp);
631 d_avg = vp9_highbd_avg_8x8(d + y8_idx * dp + x8_idx, dp);
633 s_avg = vp9_avg_8x8(s + y8_idx * sp + x8_idx, sp);
635 d_avg = vp9_avg_8x8(d + y8_idx * dp + x8_idx, dp);
638 s_avg = vp9_avg_8x8(s + y8_idx * sp + x8_idx, sp);
640 d_avg = vp9_avg_8x8(d + y8_idx * dp + x8_idx, dp);
645 fill_variance(sse, sum, 0, &vst->split[k].part_variances.none);
649 // This function chooses partitioning based on the variance between source and
650 // reconstructed last, where variance is computed for down-sampled inputs.
651 static int choose_partitioning(VP9_COMP *cpi,
652 const TileInfo *const tile,
654 int mi_row, int mi_col) {
655 VP9_COMMON * const cm = &cpi->common;
656 MACROBLOCKD *xd = &x->e_mbd;
665 int pixels_wide = 64, pixels_high = 64;
666 int64_t thresholds[4] = {cpi->vbp_thresholds[0], cpi->vbp_thresholds[1],
667 cpi->vbp_thresholds[2], cpi->vbp_thresholds[3]};
669 // Always use 4x4 partition for key frame.
670 const int is_key_frame = (cm->frame_type == KEY_FRAME);
671 const int use_4x4_partition = is_key_frame;
672 const int low_res = (cm->width <= 352 && cm->height <= 288);
673 int variance4x4downsample[16];
675 int segment_id = CR_SEGMENT_ID_BASE;
676 if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ && cm->seg.enabled) {
677 const uint8_t *const map = cm->seg.update_map ? cpi->segmentation_map :
678 cm->last_frame_seg_map;
679 segment_id = vp9_get_segment_id(cm, map, BLOCK_64X64, mi_row, mi_col);
681 if (cyclic_refresh_segment_id_boosted(segment_id)) {
682 int q = vp9_get_qindex(&cm->seg, segment_id, cm->base_qindex);
683 modify_vbp_thresholds(cpi, thresholds, q);
687 set_offsets(cpi, tile, x, mi_row, mi_col, BLOCK_64X64);
689 if (xd->mb_to_right_edge < 0)
690 pixels_wide += (xd->mb_to_right_edge >> 3);
691 if (xd->mb_to_bottom_edge < 0)
692 pixels_high += (xd->mb_to_bottom_edge >> 3);
694 s = x->plane[0].src.buf;
695 sp = x->plane[0].src.stride;
698 MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
700 const YV12_BUFFER_CONFIG *yv12 = get_ref_frame_buffer(cpi, LAST_FRAME);
702 const YV12_BUFFER_CONFIG *yv12_g = get_ref_frame_buffer(cpi, GOLDEN_FRAME);
703 unsigned int y_sad, y_sad_g;
704 const BLOCK_SIZE bsize = BLOCK_32X32
705 + (mi_col + 4 < cm->mi_cols) * 2 + (mi_row + 4 < cm->mi_rows);
707 assert(yv12 != NULL);
708 if (yv12_g && yv12_g != yv12) {
709 vp9_setup_pre_planes(xd, 0, yv12_g, mi_row, mi_col,
710 &cm->frame_refs[GOLDEN_FRAME - 1].sf);
711 y_sad_g = cpi->fn_ptr[bsize].sdf(x->plane[0].src.buf,
712 x->plane[0].src.stride,
713 xd->plane[0].pre[0].buf,
714 xd->plane[0].pre[0].stride);
719 vp9_setup_pre_planes(xd, 0, yv12, mi_row, mi_col,
720 &cm->frame_refs[LAST_FRAME - 1].sf);
721 mbmi->ref_frame[0] = LAST_FRAME;
722 mbmi->ref_frame[1] = NONE;
723 mbmi->sb_type = BLOCK_64X64;
724 mbmi->mv[0].as_int = 0;
725 mbmi->interp_filter = BILINEAR;
727 y_sad = vp9_int_pro_motion_estimation(cpi, x, bsize, mi_row, mi_col);
728 if (y_sad_g < y_sad) {
729 vp9_setup_pre_planes(xd, 0, yv12_g, mi_row, mi_col,
730 &cm->frame_refs[GOLDEN_FRAME - 1].sf);
731 mbmi->ref_frame[0] = GOLDEN_FRAME;
732 mbmi->mv[0].as_int = 0;
735 x->pred_mv[LAST_FRAME] = mbmi->mv[0].as_mv;
738 vp9_build_inter_predictors_sb(xd, mi_row, mi_col, BLOCK_64X64);
740 for (i = 1; i <= 2; ++i) {
741 struct macroblock_plane *p = &x->plane[i];
742 struct macroblockd_plane *pd = &xd->plane[i];
743 const BLOCK_SIZE bs = get_plane_block_size(bsize, pd);
745 if (bs == BLOCK_INVALID)
748 uv_sad = cpi->fn_ptr[bs].sdf(p->src.buf, p->src.stride,
749 pd->dst.buf, pd->dst.stride);
751 x->color_sensitivity[i - 1] = uv_sad > (y_sad >> 2);
754 d = xd->plane[0].dst.buf;
755 dp = xd->plane[0].dst.stride;
757 // If the y_sad is very small, take 64x64 as partition and exit.
758 // Don't check on boosted segment for now, as 64x64 is suppressed there.
759 if (segment_id == CR_SEGMENT_ID_BASE &&
760 y_sad < cpi->vbp_threshold_sad) {
761 const int block_width = num_8x8_blocks_wide_lookup[BLOCK_64X64];
762 const int block_height = num_8x8_blocks_high_lookup[BLOCK_64X64];
763 if (mi_col + block_width / 2 < cm->mi_cols &&
764 mi_row + block_height / 2 < cm->mi_rows) {
765 set_block_size(cpi, xd, mi_row, mi_col, BLOCK_64X64);
772 #if CONFIG_VP9_HIGHBITDEPTH
773 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
776 d = CONVERT_TO_BYTEPTR(VP9_HIGH_VAR_OFFS_10);
779 d = CONVERT_TO_BYTEPTR(VP9_HIGH_VAR_OFFS_12);
783 d = CONVERT_TO_BYTEPTR(VP9_HIGH_VAR_OFFS_8);
787 #endif // CONFIG_VP9_HIGHBITDEPTH
790 // Index for force_split: 0 for 64x64, 1-4 for 32x32 blocks,
791 // 5-20 for the 16x16 blocks.
793 // Fill in the entire tree of 8x8 (or 4x4 under some conditions) variances
795 for (i = 0; i < 4; i++) {
796 const int x32_idx = ((i & 1) << 5);
797 const int y32_idx = ((i >> 1) << 5);
798 const int i2 = i << 2;
799 force_split[i + 1] = 0;
800 for (j = 0; j < 4; j++) {
801 const int x16_idx = x32_idx + ((j & 1) << 4);
802 const int y16_idx = y32_idx + ((j >> 1) << 4);
803 const int split_index = 5 + i2 + j;
804 v16x16 *vst = &vt.split[i].split[j];
805 force_split[split_index] = 0;
806 variance4x4downsample[i2 + j] = 0;
808 fill_variance_8x8avg(s, sp, d, dp, x16_idx, y16_idx, vst,
809 #if CONFIG_VP9_HIGHBITDEPTH
815 fill_variance_tree(&vt.split[i].split[j], BLOCK_16X16);
816 get_variance(&vt.split[i].split[j].part_variances.none);
817 if (vt.split[i].split[j].part_variances.none.variance >
819 // 16X16 variance is above threshold for split, so force split to 8x8
820 // for this 16x16 block (this also forces splits for upper levels).
821 force_split[split_index] = 1;
822 force_split[i + 1] = 1;
824 } else if (vt.split[i].split[j].part_variances.none.variance >
826 !cyclic_refresh_segment_id_boosted(segment_id)) {
827 // We have some nominal amount of 16x16 variance (based on average),
828 // compute the minmax over the 8x8 sub-blocks, and if above threshold,
829 // force split to 8x8 block for this 16x16 block.
830 int minmax = compute_minmax_8x8(s, sp, d, dp, x16_idx, y16_idx,
831 #if CONFIG_VP9_HIGHBITDEPTH
834 pixels_wide, pixels_high);
835 if (minmax > cpi->vbp_threshold_minmax) {
836 force_split[split_index] = 1;
837 force_split[i + 1] = 1;
842 if (is_key_frame || (low_res &&
843 vt.split[i].split[j].part_variances.none.variance >
844 (thresholds[1] << 1))) {
845 force_split[split_index] = 0;
846 // Go down to 4x4 down-sampling for variance.
847 variance4x4downsample[i2 + j] = 1;
848 for (k = 0; k < 4; k++) {
849 int x8_idx = x16_idx + ((k & 1) << 3);
850 int y8_idx = y16_idx + ((k >> 1) << 3);
851 v8x8 *vst2 = is_key_frame ? &vst->split[k] :
852 &vt2[i2 + j].split[k];
853 fill_variance_4x4avg(s, sp, d, dp, x8_idx, y8_idx, vst2,
854 #if CONFIG_VP9_HIGHBITDEPTH
865 // Fill the rest of the variance tree by summing split partition values.
866 for (i = 0; i < 4; i++) {
867 const int i2 = i << 2;
868 for (j = 0; j < 4; j++) {
869 if (variance4x4downsample[i2 + j] == 1) {
870 v16x16 *vtemp = (!is_key_frame) ? &vt2[i2 + j] :
871 &vt.split[i].split[j];
872 for (m = 0; m < 4; m++)
873 fill_variance_tree(&vtemp->split[m], BLOCK_8X8);
874 fill_variance_tree(vtemp, BLOCK_16X16);
877 fill_variance_tree(&vt.split[i], BLOCK_32X32);
878 // If variance of this 32x32 block is above the threshold, force the block
879 // to split. This also forces a split on the upper (64x64) level.
880 if (!force_split[i + 1]) {
881 get_variance(&vt.split[i].part_variances.none);
882 if (vt.split[i].part_variances.none.variance > thresholds[1]) {
883 force_split[i + 1] = 1;
888 if (!force_split[0]) {
889 fill_variance_tree(&vt, BLOCK_64X64);
890 get_variance(&vt.part_variances.none);
893 // Now go through the entire structure, splitting every block size until
894 // we get to one that's got a variance lower than our threshold.
895 if ( mi_col + 8 > cm->mi_cols || mi_row + 8 > cm->mi_rows ||
896 !set_vt_partitioning(cpi, xd, &vt, BLOCK_64X64, mi_row, mi_col,
897 thresholds[0], BLOCK_16X16, force_split[0])) {
898 for (i = 0; i < 4; ++i) {
899 const int x32_idx = ((i & 1) << 2);
900 const int y32_idx = ((i >> 1) << 2);
901 const int i2 = i << 2;
902 if (!set_vt_partitioning(cpi, xd, &vt.split[i], BLOCK_32X32,
903 (mi_row + y32_idx), (mi_col + x32_idx),
904 thresholds[1], BLOCK_16X16,
905 force_split[i + 1])) {
906 for (j = 0; j < 4; ++j) {
907 const int x16_idx = ((j & 1) << 1);
908 const int y16_idx = ((j >> 1) << 1);
909 // For inter frames: if variance4x4downsample[] == 1 for this 16x16
910 // block, then the variance is based on 4x4 down-sampling, so use vt2
911 // in set_vt_partioning(), otherwise use vt.
912 v16x16 *vtemp = (!is_key_frame &&
913 variance4x4downsample[i2 + j] == 1) ?
914 &vt2[i2 + j] : &vt.split[i].split[j];
915 if (!set_vt_partitioning(cpi, xd, vtemp, BLOCK_16X16,
916 mi_row + y32_idx + y16_idx,
917 mi_col + x32_idx + x16_idx,
920 force_split[5 + i2 + j])) {
921 for (k = 0; k < 4; ++k) {
922 const int x8_idx = (k & 1);
923 const int y8_idx = (k >> 1);
924 if (use_4x4_partition) {
925 if (!set_vt_partitioning(cpi, xd, &vtemp->split[k],
927 mi_row + y32_idx + y16_idx + y8_idx,
928 mi_col + x32_idx + x16_idx + x8_idx,
929 thresholds[3], BLOCK_8X8, 0)) {
930 set_block_size(cpi, xd,
931 (mi_row + y32_idx + y16_idx + y8_idx),
932 (mi_col + x32_idx + x16_idx + x8_idx),
936 set_block_size(cpi, xd,
937 (mi_row + y32_idx + y16_idx + y8_idx),
938 (mi_col + x32_idx + x16_idx + x8_idx),
950 static void update_state(VP9_COMP *cpi, ThreadData *td,
951 PICK_MODE_CONTEXT *ctx,
952 int mi_row, int mi_col, BLOCK_SIZE bsize,
953 int output_enabled) {
955 VP9_COMMON *const cm = &cpi->common;
956 RD_COUNTS *const rdc = &td->rd_counts;
957 MACROBLOCK *const x = &td->mb;
958 MACROBLOCKD *const xd = &x->e_mbd;
959 struct macroblock_plane *const p = x->plane;
960 struct macroblockd_plane *const pd = xd->plane;
961 MODE_INFO *mi = &ctx->mic;
962 MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
963 MODE_INFO *mi_addr = xd->mi[0];
964 const struct segmentation *const seg = &cm->seg;
965 const int bw = num_8x8_blocks_wide_lookup[mi->mbmi.sb_type];
966 const int bh = num_8x8_blocks_high_lookup[mi->mbmi.sb_type];
967 const int x_mis = MIN(bw, cm->mi_cols - mi_col);
968 const int y_mis = MIN(bh, cm->mi_rows - mi_row);
969 MV_REF *const frame_mvs =
970 cm->cur_frame->mvs + mi_row * cm->mi_cols + mi_col;
973 const int mis = cm->mi_stride;
974 const int mi_width = num_8x8_blocks_wide_lookup[bsize];
975 const int mi_height = num_8x8_blocks_high_lookup[bsize];
978 assert(mi->mbmi.sb_type == bsize);
982 // If segmentation in use
984 // For in frame complexity AQ copy the segment id from the segment map.
985 if (cpi->oxcf.aq_mode == COMPLEXITY_AQ) {
986 const uint8_t *const map = seg->update_map ? cpi->segmentation_map
987 : cm->last_frame_seg_map;
988 mi_addr->mbmi.segment_id =
989 vp9_get_segment_id(cm, map, bsize, mi_row, mi_col);
991 // Else for cyclic refresh mode update the segment map, set the segment id
992 // and then update the quantizer.
993 if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ) {
994 vp9_cyclic_refresh_update_segment(cpi, &xd->mi[0]->mbmi, mi_row,
995 mi_col, bsize, ctx->rate, ctx->dist,
1000 max_plane = is_inter_block(mbmi) ? MAX_MB_PLANE : 1;
1001 for (i = 0; i < max_plane; ++i) {
1002 p[i].coeff = ctx->coeff_pbuf[i][1];
1003 p[i].qcoeff = ctx->qcoeff_pbuf[i][1];
1004 pd[i].dqcoeff = ctx->dqcoeff_pbuf[i][1];
1005 p[i].eobs = ctx->eobs_pbuf[i][1];
1008 for (i = max_plane; i < MAX_MB_PLANE; ++i) {
1009 p[i].coeff = ctx->coeff_pbuf[i][2];
1010 p[i].qcoeff = ctx->qcoeff_pbuf[i][2];
1011 pd[i].dqcoeff = ctx->dqcoeff_pbuf[i][2];
1012 p[i].eobs = ctx->eobs_pbuf[i][2];
1015 // Restore the coding context of the MB to that that was in place
1016 // when the mode was picked for it
1017 for (y = 0; y < mi_height; y++)
1018 for (x_idx = 0; x_idx < mi_width; x_idx++)
1019 if ((xd->mb_to_right_edge >> (3 + MI_SIZE_LOG2)) + mi_width > x_idx
1020 && (xd->mb_to_bottom_edge >> (3 + MI_SIZE_LOG2)) + mi_height > y) {
1021 xd->mi[x_idx + y * mis] = mi_addr;
1024 if (cpi->oxcf.aq_mode)
1025 vp9_init_plane_quantizers(cpi, x);
1027 // FIXME(rbultje) I'm pretty sure this should go to the end of this block
1028 // (i.e. after the output_enabled)
1029 if (bsize < BLOCK_32X32) {
1030 if (bsize < BLOCK_16X16)
1031 ctx->tx_rd_diff[ALLOW_16X16] = ctx->tx_rd_diff[ALLOW_8X8];
1032 ctx->tx_rd_diff[ALLOW_32X32] = ctx->tx_rd_diff[ALLOW_16X16];
1035 if (is_inter_block(mbmi) && mbmi->sb_type < BLOCK_8X8) {
1036 mbmi->mv[0].as_int = mi->bmi[3].as_mv[0].as_int;
1037 mbmi->mv[1].as_int = mi->bmi[3].as_mv[1].as_int;
1040 x->skip = ctx->skip;
1041 memcpy(x->zcoeff_blk[mbmi->tx_size], ctx->zcoeff_blk,
1042 sizeof(uint8_t) * ctx->num_4x4_blk);
1044 if (!output_enabled)
1047 if (!vp9_segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_SKIP)) {
1048 for (i = 0; i < TX_MODES; i++)
1049 rdc->tx_select_diff[i] += ctx->tx_rd_diff[i];
1052 #if CONFIG_INTERNAL_STATS
1053 if (frame_is_intra_only(cm)) {
1054 static const int kf_mode_index[] = {
1056 THR_V_PRED /*V_PRED*/,
1057 THR_H_PRED /*H_PRED*/,
1058 THR_D45_PRED /*D45_PRED*/,
1059 THR_D135_PRED /*D135_PRED*/,
1060 THR_D117_PRED /*D117_PRED*/,
1061 THR_D153_PRED /*D153_PRED*/,
1062 THR_D207_PRED /*D207_PRED*/,
1063 THR_D63_PRED /*D63_PRED*/,
1066 ++cpi->mode_chosen_counts[kf_mode_index[mbmi->mode]];
1068 // Note how often each mode chosen as best
1069 ++cpi->mode_chosen_counts[ctx->best_mode_index];
1072 if (!frame_is_intra_only(cm)) {
1073 if (is_inter_block(mbmi)) {
1074 vp9_update_mv_count(td);
1076 if (cm->interp_filter == SWITCHABLE) {
1077 const int ctx = vp9_get_pred_context_switchable_interp(xd);
1078 ++td->counts->switchable_interp[ctx][mbmi->interp_filter];
1082 rdc->comp_pred_diff[SINGLE_REFERENCE] += ctx->single_pred_diff;
1083 rdc->comp_pred_diff[COMPOUND_REFERENCE] += ctx->comp_pred_diff;
1084 rdc->comp_pred_diff[REFERENCE_MODE_SELECT] += ctx->hybrid_pred_diff;
1086 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i)
1087 rdc->filter_diff[i] += ctx->best_filter_diff[i];
1090 for (h = 0; h < y_mis; ++h) {
1091 MV_REF *const frame_mv = frame_mvs + h * cm->mi_cols;
1092 for (w = 0; w < x_mis; ++w) {
1093 MV_REF *const mv = frame_mv + w;
1094 mv->ref_frame[0] = mi->mbmi.ref_frame[0];
1095 mv->ref_frame[1] = mi->mbmi.ref_frame[1];
1096 mv->mv[0].as_int = mi->mbmi.mv[0].as_int;
1097 mv->mv[1].as_int = mi->mbmi.mv[1].as_int;
1102 void vp9_setup_src_planes(MACROBLOCK *x, const YV12_BUFFER_CONFIG *src,
1103 int mi_row, int mi_col) {
1104 uint8_t *const buffers[3] = {src->y_buffer, src->u_buffer, src->v_buffer };
1105 const int strides[3] = {src->y_stride, src->uv_stride, src->uv_stride };
1108 // Set current frame pointer.
1109 x->e_mbd.cur_buf = src;
1111 for (i = 0; i < MAX_MB_PLANE; i++)
1112 setup_pred_plane(&x->plane[i].src, buffers[i], strides[i], mi_row, mi_col,
1113 NULL, x->e_mbd.plane[i].subsampling_x,
1114 x->e_mbd.plane[i].subsampling_y);
1117 static void set_mode_info_seg_skip(MACROBLOCK *x, TX_MODE tx_mode,
1118 RD_COST *rd_cost, BLOCK_SIZE bsize) {
1119 MACROBLOCKD *const xd = &x->e_mbd;
1120 MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
1121 INTERP_FILTER filter_ref;
1123 if (xd->up_available)
1124 filter_ref = xd->mi[-xd->mi_stride]->mbmi.interp_filter;
1125 else if (xd->left_available)
1126 filter_ref = xd->mi[-1]->mbmi.interp_filter;
1128 filter_ref = EIGHTTAP;
1130 mbmi->sb_type = bsize;
1131 mbmi->mode = ZEROMV;
1132 mbmi->tx_size = MIN(max_txsize_lookup[bsize],
1133 tx_mode_to_biggest_tx_size[tx_mode]);
1135 mbmi->uv_mode = DC_PRED;
1136 mbmi->ref_frame[0] = LAST_FRAME;
1137 mbmi->ref_frame[1] = NONE;
1138 mbmi->mv[0].as_int = 0;
1139 mbmi->interp_filter = filter_ref;
1141 xd->mi[0]->bmi[0].as_mv[0].as_int = 0;
1144 vp9_rd_cost_init(rd_cost);
1147 static int set_segment_rdmult(VP9_COMP *const cpi,
1148 MACROBLOCK *const x,
1149 int8_t segment_id) {
1151 VP9_COMMON *const cm = &cpi->common;
1152 vp9_init_plane_quantizers(cpi, x);
1153 vp9_clear_system_state();
1154 segment_qindex = vp9_get_qindex(&cm->seg, segment_id,
1156 return vp9_compute_rd_mult(cpi, segment_qindex + cm->y_dc_delta_q);
1159 static void rd_pick_sb_modes(VP9_COMP *cpi,
1160 TileDataEnc *tile_data,
1161 MACROBLOCK *const x,
1162 int mi_row, int mi_col, RD_COST *rd_cost,
1163 BLOCK_SIZE bsize, PICK_MODE_CONTEXT *ctx,
1165 VP9_COMMON *const cm = &cpi->common;
1166 TileInfo *const tile_info = &tile_data->tile_info;
1167 MACROBLOCKD *const xd = &x->e_mbd;
1169 struct macroblock_plane *const p = x->plane;
1170 struct macroblockd_plane *const pd = xd->plane;
1171 const AQ_MODE aq_mode = cpi->oxcf.aq_mode;
1174 vp9_clear_system_state();
1176 // Use the lower precision, but faster, 32x32 fdct for mode selection.
1177 x->use_lp32x32fdct = 1;
1179 set_offsets(cpi, tile_info, x, mi_row, mi_col, bsize);
1180 mbmi = &xd->mi[0]->mbmi;
1181 mbmi->sb_type = bsize;
1183 for (i = 0; i < MAX_MB_PLANE; ++i) {
1184 p[i].coeff = ctx->coeff_pbuf[i][0];
1185 p[i].qcoeff = ctx->qcoeff_pbuf[i][0];
1186 pd[i].dqcoeff = ctx->dqcoeff_pbuf[i][0];
1187 p[i].eobs = ctx->eobs_pbuf[i][0];
1191 ctx->pred_pixel_ready = 0;
1194 // Set to zero to make sure we do not use the previous encoded frame stats
1197 #if CONFIG_VP9_HIGHBITDEPTH
1198 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
1199 x->source_variance =
1200 vp9_high_get_sby_perpixel_variance(cpi, &x->plane[0].src,
1203 x->source_variance =
1204 vp9_get_sby_perpixel_variance(cpi, &x->plane[0].src, bsize);
1207 x->source_variance =
1208 vp9_get_sby_perpixel_variance(cpi, &x->plane[0].src, bsize);
1209 #endif // CONFIG_VP9_HIGHBITDEPTH
1211 // Save rdmult before it might be changed, so it can be restored later.
1212 orig_rdmult = x->rdmult;
1214 if (aq_mode == VARIANCE_AQ) {
1215 const int energy = bsize <= BLOCK_16X16 ? x->mb_energy
1216 : vp9_block_energy(cpi, x, bsize);
1217 if (cm->frame_type == KEY_FRAME ||
1218 cpi->refresh_alt_ref_frame ||
1219 (cpi->refresh_golden_frame && !cpi->rc.is_src_frame_alt_ref)) {
1220 mbmi->segment_id = vp9_vaq_segment_id(energy);
1222 const uint8_t *const map = cm->seg.update_map ? cpi->segmentation_map
1223 : cm->last_frame_seg_map;
1224 mbmi->segment_id = vp9_get_segment_id(cm, map, bsize, mi_row, mi_col);
1226 x->rdmult = set_segment_rdmult(cpi, x, mbmi->segment_id);
1227 } else if (aq_mode == COMPLEXITY_AQ) {
1228 x->rdmult = set_segment_rdmult(cpi, x, mbmi->segment_id);
1229 } else if (aq_mode == CYCLIC_REFRESH_AQ) {
1230 const uint8_t *const map = cm->seg.update_map ? cpi->segmentation_map
1231 : cm->last_frame_seg_map;
1232 // If segment is boosted, use rdmult for that segment.
1233 if (cyclic_refresh_segment_id_boosted(
1234 vp9_get_segment_id(cm, map, bsize, mi_row, mi_col)))
1235 x->rdmult = vp9_cyclic_refresh_get_rdmult(cpi->cyclic_refresh);
1238 // Find best coding mode & reconstruct the MB so it is available
1239 // as a predictor for MBs that follow in the SB
1240 if (frame_is_intra_only(cm)) {
1241 vp9_rd_pick_intra_mode_sb(cpi, x, rd_cost, bsize, ctx, best_rd);
1243 if (bsize >= BLOCK_8X8) {
1244 if (vp9_segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_SKIP))
1245 vp9_rd_pick_inter_mode_sb_seg_skip(cpi, tile_data, x, rd_cost, bsize,
1248 vp9_rd_pick_inter_mode_sb(cpi, tile_data, x, mi_row, mi_col,
1249 rd_cost, bsize, ctx, best_rd);
1251 vp9_rd_pick_inter_mode_sub8x8(cpi, tile_data, x, mi_row, mi_col,
1252 rd_cost, bsize, ctx, best_rd);
1257 // Examine the resulting rate and for AQ mode 2 make a segment choice.
1258 if ((rd_cost->rate != INT_MAX) &&
1259 (aq_mode == COMPLEXITY_AQ) && (bsize >= BLOCK_16X16) &&
1260 (cm->frame_type == KEY_FRAME ||
1261 cpi->refresh_alt_ref_frame ||
1262 (cpi->refresh_golden_frame && !cpi->rc.is_src_frame_alt_ref))) {
1263 vp9_caq_select_segment(cpi, x, bsize, mi_row, mi_col, rd_cost->rate);
1266 x->rdmult = orig_rdmult;
1268 // TODO(jingning) The rate-distortion optimization flow needs to be
1269 // refactored to provide proper exit/return handle.
1270 if (rd_cost->rate == INT_MAX)
1271 rd_cost->rdcost = INT64_MAX;
1273 ctx->rate = rd_cost->rate;
1274 ctx->dist = rd_cost->dist;
1277 static void update_stats(VP9_COMMON *cm, ThreadData *td) {
1278 const MACROBLOCK *x = &td->mb;
1279 const MACROBLOCKD *const xd = &x->e_mbd;
1280 const MODE_INFO *const mi = xd->mi[0];
1281 const MB_MODE_INFO *const mbmi = &mi->mbmi;
1282 const BLOCK_SIZE bsize = mbmi->sb_type;
1284 if (!frame_is_intra_only(cm)) {
1285 FRAME_COUNTS *const counts = td->counts;
1286 const int inter_block = is_inter_block(mbmi);
1287 const int seg_ref_active = vp9_segfeature_active(&cm->seg, mbmi->segment_id,
1289 if (!seg_ref_active) {
1290 counts->intra_inter[vp9_get_intra_inter_context(xd)][inter_block]++;
1291 // If the segment reference feature is enabled we have only a single
1292 // reference frame allowed for the segment so exclude it from
1293 // the reference frame counts used to work out probabilities.
1295 const MV_REFERENCE_FRAME ref0 = mbmi->ref_frame[0];
1296 if (cm->reference_mode == REFERENCE_MODE_SELECT)
1297 counts->comp_inter[vp9_get_reference_mode_context(cm, xd)]
1298 [has_second_ref(mbmi)]++;
1300 if (has_second_ref(mbmi)) {
1301 counts->comp_ref[vp9_get_pred_context_comp_ref_p(cm, xd)]
1302 [ref0 == GOLDEN_FRAME]++;
1304 counts->single_ref[vp9_get_pred_context_single_ref_p1(xd)][0]
1305 [ref0 != LAST_FRAME]++;
1306 if (ref0 != LAST_FRAME)
1307 counts->single_ref[vp9_get_pred_context_single_ref_p2(xd)][1]
1308 [ref0 != GOLDEN_FRAME]++;
1313 !vp9_segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_SKIP)) {
1314 const int mode_ctx = mbmi->mode_context[mbmi->ref_frame[0]];
1315 if (bsize >= BLOCK_8X8) {
1316 const PREDICTION_MODE mode = mbmi->mode;
1317 ++counts->inter_mode[mode_ctx][INTER_OFFSET(mode)];
1319 const int num_4x4_w = num_4x4_blocks_wide_lookup[bsize];
1320 const int num_4x4_h = num_4x4_blocks_high_lookup[bsize];
1322 for (idy = 0; idy < 2; idy += num_4x4_h) {
1323 for (idx = 0; idx < 2; idx += num_4x4_w) {
1324 const int j = idy * 2 + idx;
1325 const PREDICTION_MODE b_mode = mi->bmi[j].as_mode;
1326 ++counts->inter_mode[mode_ctx][INTER_OFFSET(b_mode)];
1334 static void restore_context(MACROBLOCK *const x, int mi_row, int mi_col,
1335 ENTROPY_CONTEXT a[16 * MAX_MB_PLANE],
1336 ENTROPY_CONTEXT l[16 * MAX_MB_PLANE],
1337 PARTITION_CONTEXT sa[8], PARTITION_CONTEXT sl[8],
1339 MACROBLOCKD *const xd = &x->e_mbd;
1341 const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
1342 const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
1343 int mi_width = num_8x8_blocks_wide_lookup[bsize];
1344 int mi_height = num_8x8_blocks_high_lookup[bsize];
1345 for (p = 0; p < MAX_MB_PLANE; p++) {
1347 xd->above_context[p] + ((mi_col * 2) >> xd->plane[p].subsampling_x),
1348 a + num_4x4_blocks_wide * p,
1349 (sizeof(ENTROPY_CONTEXT) * num_4x4_blocks_wide) >>
1350 xd->plane[p].subsampling_x);
1353 + ((mi_row & MI_MASK) * 2 >> xd->plane[p].subsampling_y),
1354 l + num_4x4_blocks_high * p,
1355 (sizeof(ENTROPY_CONTEXT) * num_4x4_blocks_high) >>
1356 xd->plane[p].subsampling_y);
1358 memcpy(xd->above_seg_context + mi_col, sa,
1359 sizeof(*xd->above_seg_context) * mi_width);
1360 memcpy(xd->left_seg_context + (mi_row & MI_MASK), sl,
1361 sizeof(xd->left_seg_context[0]) * mi_height);
1364 static void save_context(MACROBLOCK *const x, int mi_row, int mi_col,
1365 ENTROPY_CONTEXT a[16 * MAX_MB_PLANE],
1366 ENTROPY_CONTEXT l[16 * MAX_MB_PLANE],
1367 PARTITION_CONTEXT sa[8], PARTITION_CONTEXT sl[8],
1369 const MACROBLOCKD *const xd = &x->e_mbd;
1371 const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
1372 const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
1373 int mi_width = num_8x8_blocks_wide_lookup[bsize];
1374 int mi_height = num_8x8_blocks_high_lookup[bsize];
1376 // buffer the above/left context information of the block in search.
1377 for (p = 0; p < MAX_MB_PLANE; ++p) {
1379 a + num_4x4_blocks_wide * p,
1380 xd->above_context[p] + (mi_col * 2 >> xd->plane[p].subsampling_x),
1381 (sizeof(ENTROPY_CONTEXT) * num_4x4_blocks_wide) >>
1382 xd->plane[p].subsampling_x);
1384 l + num_4x4_blocks_high * p,
1386 + ((mi_row & MI_MASK) * 2 >> xd->plane[p].subsampling_y),
1387 (sizeof(ENTROPY_CONTEXT) * num_4x4_blocks_high) >>
1388 xd->plane[p].subsampling_y);
1390 memcpy(sa, xd->above_seg_context + mi_col,
1391 sizeof(*xd->above_seg_context) * mi_width);
1392 memcpy(sl, xd->left_seg_context + (mi_row & MI_MASK),
1393 sizeof(xd->left_seg_context[0]) * mi_height);
1396 static void encode_b(VP9_COMP *cpi, const TileInfo *const tile,
1398 TOKENEXTRA **tp, int mi_row, int mi_col,
1399 int output_enabled, BLOCK_SIZE bsize,
1400 PICK_MODE_CONTEXT *ctx) {
1401 MACROBLOCK *const x = &td->mb;
1402 set_offsets(cpi, tile, x, mi_row, mi_col, bsize);
1403 update_state(cpi, td, ctx, mi_row, mi_col, bsize, output_enabled);
1404 encode_superblock(cpi, td, tp, output_enabled, mi_row, mi_col, bsize, ctx);
1406 if (output_enabled) {
1407 update_stats(&cpi->common, td);
1409 (*tp)->token = EOSB_TOKEN;
1414 static void encode_sb(VP9_COMP *cpi, ThreadData *td,
1415 const TileInfo *const tile,
1416 TOKENEXTRA **tp, int mi_row, int mi_col,
1417 int output_enabled, BLOCK_SIZE bsize,
1419 VP9_COMMON *const cm = &cpi->common;
1420 MACROBLOCK *const x = &td->mb;
1421 MACROBLOCKD *const xd = &x->e_mbd;
1423 const int bsl = b_width_log2_lookup[bsize], hbs = (1 << bsl) / 4;
1425 PARTITION_TYPE partition;
1426 BLOCK_SIZE subsize = bsize;
1428 if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols)
1431 if (bsize >= BLOCK_8X8) {
1432 ctx = partition_plane_context(xd, mi_row, mi_col, bsize);
1433 subsize = get_subsize(bsize, pc_tree->partitioning);
1436 subsize = BLOCK_4X4;
1439 partition = partition_lookup[bsl][subsize];
1440 if (output_enabled && bsize != BLOCK_4X4)
1441 td->counts->partition[ctx][partition]++;
1443 switch (partition) {
1444 case PARTITION_NONE:
1445 encode_b(cpi, tile, td, tp, mi_row, mi_col, output_enabled, subsize,
1448 case PARTITION_VERT:
1449 encode_b(cpi, tile, td, tp, mi_row, mi_col, output_enabled, subsize,
1450 &pc_tree->vertical[0]);
1451 if (mi_col + hbs < cm->mi_cols && bsize > BLOCK_8X8) {
1452 encode_b(cpi, tile, td, tp, mi_row, mi_col + hbs, output_enabled,
1453 subsize, &pc_tree->vertical[1]);
1456 case PARTITION_HORZ:
1457 encode_b(cpi, tile, td, tp, mi_row, mi_col, output_enabled, subsize,
1458 &pc_tree->horizontal[0]);
1459 if (mi_row + hbs < cm->mi_rows && bsize > BLOCK_8X8) {
1460 encode_b(cpi, tile, td, tp, mi_row + hbs, mi_col, output_enabled,
1461 subsize, &pc_tree->horizontal[1]);
1464 case PARTITION_SPLIT:
1465 if (bsize == BLOCK_8X8) {
1466 encode_b(cpi, tile, td, tp, mi_row, mi_col, output_enabled, subsize,
1467 pc_tree->leaf_split[0]);
1469 encode_sb(cpi, td, tile, tp, mi_row, mi_col, output_enabled, subsize,
1471 encode_sb(cpi, td, tile, tp, mi_row, mi_col + hbs, output_enabled,
1472 subsize, pc_tree->split[1]);
1473 encode_sb(cpi, td, tile, tp, mi_row + hbs, mi_col, output_enabled,
1474 subsize, pc_tree->split[2]);
1475 encode_sb(cpi, td, tile, tp, mi_row + hbs, mi_col + hbs, output_enabled,
1476 subsize, pc_tree->split[3]);
1480 assert(0 && "Invalid partition type.");
1484 if (partition != PARTITION_SPLIT || bsize == BLOCK_8X8)
1485 update_partition_context(xd, mi_row, mi_col, subsize, bsize);
1488 // Check to see if the given partition size is allowed for a specified number
1489 // of 8x8 block rows and columns remaining in the image.
1490 // If not then return the largest allowed partition size
1491 static BLOCK_SIZE find_partition_size(BLOCK_SIZE bsize,
1492 int rows_left, int cols_left,
1494 if (rows_left <= 0 || cols_left <= 0) {
1495 return MIN(bsize, BLOCK_8X8);
1497 for (; bsize > 0; bsize -= 3) {
1498 *bh = num_8x8_blocks_high_lookup[bsize];
1499 *bw = num_8x8_blocks_wide_lookup[bsize];
1500 if ((*bh <= rows_left) && (*bw <= cols_left)) {
1508 static void set_partial_b64x64_partition(MODE_INFO *mi, int mis,
1509 int bh_in, int bw_in, int row8x8_remaining, int col8x8_remaining,
1510 BLOCK_SIZE bsize, MODE_INFO **mi_8x8) {
1513 for (r = 0; r < MI_BLOCK_SIZE; r += bh) {
1515 for (c = 0; c < MI_BLOCK_SIZE; c += bw) {
1516 const int index = r * mis + c;
1517 mi_8x8[index] = mi + index;
1518 mi_8x8[index]->mbmi.sb_type = find_partition_size(bsize,
1519 row8x8_remaining - r, col8x8_remaining - c, &bh, &bw);
1524 // This function attempts to set all mode info entries in a given SB64
1525 // to the same block partition size.
1526 // However, at the bottom and right borders of the image the requested size
1527 // may not be allowed in which case this code attempts to choose the largest
1528 // allowable partition.
1529 static void set_fixed_partitioning(VP9_COMP *cpi, const TileInfo *const tile,
1530 MODE_INFO **mi_8x8, int mi_row, int mi_col,
1532 VP9_COMMON *const cm = &cpi->common;
1533 const int mis = cm->mi_stride;
1534 const int row8x8_remaining = tile->mi_row_end - mi_row;
1535 const int col8x8_remaining = tile->mi_col_end - mi_col;
1536 int block_row, block_col;
1537 MODE_INFO *mi_upper_left = cm->mi + mi_row * mis + mi_col;
1538 int bh = num_8x8_blocks_high_lookup[bsize];
1539 int bw = num_8x8_blocks_wide_lookup[bsize];
1541 assert((row8x8_remaining > 0) && (col8x8_remaining > 0));
1543 // Apply the requested partition size to the SB64 if it is all "in image"
1544 if ((col8x8_remaining >= MI_BLOCK_SIZE) &&
1545 (row8x8_remaining >= MI_BLOCK_SIZE)) {
1546 for (block_row = 0; block_row < MI_BLOCK_SIZE; block_row += bh) {
1547 for (block_col = 0; block_col < MI_BLOCK_SIZE; block_col += bw) {
1548 int index = block_row * mis + block_col;
1549 mi_8x8[index] = mi_upper_left + index;
1550 mi_8x8[index]->mbmi.sb_type = bsize;
1554 // Else this is a partial SB64.
1555 set_partial_b64x64_partition(mi_upper_left, mis, bh, bw, row8x8_remaining,
1556 col8x8_remaining, bsize, mi_8x8);
1563 } coord_lookup[16] = {
1565 {0, 0}, {0, 2}, {2, 0}, {2, 2},
1567 {0, 4}, {0, 6}, {2, 4}, {2, 6},
1569 {4, 0}, {4, 2}, {6, 0}, {6, 2},
1571 {4, 4}, {4, 6}, {6, 4}, {6, 6},
1574 static void set_source_var_based_partition(VP9_COMP *cpi,
1575 const TileInfo *const tile,
1576 MACROBLOCK *const x,
1578 int mi_row, int mi_col) {
1579 VP9_COMMON *const cm = &cpi->common;
1580 const int mis = cm->mi_stride;
1581 const int row8x8_remaining = tile->mi_row_end - mi_row;
1582 const int col8x8_remaining = tile->mi_col_end - mi_col;
1583 MODE_INFO *mi_upper_left = cm->mi + mi_row * mis + mi_col;
1585 vp9_setup_src_planes(x, cpi->Source, mi_row, mi_col);
1587 assert((row8x8_remaining > 0) && (col8x8_remaining > 0));
1590 if ((col8x8_remaining >= MI_BLOCK_SIZE) &&
1591 (row8x8_remaining >= MI_BLOCK_SIZE)) {
1595 const int offset = (mi_row >> 1) * cm->mb_cols + (mi_col >> 1);
1596 int is_larger_better = 0;
1598 unsigned int thr = cpi->source_var_thresh;
1600 memset(d32, 0, 4 * sizeof(diff));
1602 for (i = 0; i < 4; i++) {
1605 for (j = 0; j < 4; j++) {
1606 int b_mi_row = coord_lookup[i * 4 + j].row;
1607 int b_mi_col = coord_lookup[i * 4 + j].col;
1608 int boffset = b_mi_row / 2 * cm->mb_cols +
1611 d16[j] = cpi->source_diff_var + offset + boffset;
1613 index = b_mi_row * mis + b_mi_col;
1614 mi_8x8[index] = mi_upper_left + index;
1615 mi_8x8[index]->mbmi.sb_type = BLOCK_16X16;
1617 // TODO(yunqingwang): If d16[j].var is very large, use 8x8 partition
1618 // size to further improve quality.
1621 is_larger_better = (d16[0]->var < thr) && (d16[1]->var < thr) &&
1622 (d16[2]->var < thr) && (d16[3]->var < thr);
1624 // Use 32x32 partition
1625 if (is_larger_better) {
1628 for (j = 0; j < 4; j++) {
1629 d32[i].sse += d16[j]->sse;
1630 d32[i].sum += d16[j]->sum;
1633 d32[i].var = d32[i].sse - (((int64_t)d32[i].sum * d32[i].sum) >> 10);
1635 index = coord_lookup[i*4].row * mis + coord_lookup[i*4].col;
1636 mi_8x8[index] = mi_upper_left + index;
1637 mi_8x8[index]->mbmi.sb_type = BLOCK_32X32;
1641 if (use32x32 == 4) {
1643 is_larger_better = (d32[0].var < thr) && (d32[1].var < thr) &&
1644 (d32[2].var < thr) && (d32[3].var < thr);
1646 // Use 64x64 partition
1647 if (is_larger_better) {
1648 mi_8x8[0] = mi_upper_left;
1649 mi_8x8[0]->mbmi.sb_type = BLOCK_64X64;
1652 } else { // partial in-image SB64
1653 int bh = num_8x8_blocks_high_lookup[BLOCK_16X16];
1654 int bw = num_8x8_blocks_wide_lookup[BLOCK_16X16];
1655 set_partial_b64x64_partition(mi_upper_left, mis, bh, bw,
1656 row8x8_remaining, col8x8_remaining, BLOCK_16X16, mi_8x8);
1660 static void update_state_rt(VP9_COMP *cpi, ThreadData *td,
1661 PICK_MODE_CONTEXT *ctx,
1662 int mi_row, int mi_col, int bsize) {
1663 VP9_COMMON *const cm = &cpi->common;
1664 MACROBLOCK *const x = &td->mb;
1665 MACROBLOCKD *const xd = &x->e_mbd;
1666 MODE_INFO *const mi = xd->mi[0];
1667 MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
1668 const struct segmentation *const seg = &cm->seg;
1669 const int bw = num_8x8_blocks_wide_lookup[mi->mbmi.sb_type];
1670 const int bh = num_8x8_blocks_high_lookup[mi->mbmi.sb_type];
1671 const int x_mis = MIN(bw, cm->mi_cols - mi_col);
1672 const int y_mis = MIN(bh, cm->mi_rows - mi_row);
1674 *(xd->mi[0]) = ctx->mic;
1676 if (seg->enabled && cpi->oxcf.aq_mode) {
1677 // For in frame complexity AQ or variance AQ, copy segment_id from
1678 // segmentation_map.
1679 if (cpi->oxcf.aq_mode == COMPLEXITY_AQ ||
1680 cpi->oxcf.aq_mode == VARIANCE_AQ ) {
1681 const uint8_t *const map = seg->update_map ? cpi->segmentation_map
1682 : cm->last_frame_seg_map;
1683 mbmi->segment_id = vp9_get_segment_id(cm, map, bsize, mi_row, mi_col);
1685 // Setting segmentation map for cyclic_refresh.
1686 vp9_cyclic_refresh_update_segment(cpi, mbmi, mi_row, mi_col, bsize,
1687 ctx->rate, ctx->dist, x->skip);
1689 vp9_init_plane_quantizers(cpi, x);
1692 if (is_inter_block(mbmi)) {
1693 vp9_update_mv_count(td);
1694 if (cm->interp_filter == SWITCHABLE) {
1695 const int pred_ctx = vp9_get_pred_context_switchable_interp(xd);
1696 ++td->counts->switchable_interp[pred_ctx][mbmi->interp_filter];
1699 if (mbmi->sb_type < BLOCK_8X8) {
1700 mbmi->mv[0].as_int = mi->bmi[3].as_mv[0].as_int;
1701 mbmi->mv[1].as_int = mi->bmi[3].as_mv[1].as_int;
1705 if (cm->use_prev_frame_mvs) {
1706 MV_REF *const frame_mvs =
1707 cm->cur_frame->mvs + mi_row * cm->mi_cols + mi_col;
1710 for (h = 0; h < y_mis; ++h) {
1711 MV_REF *const frame_mv = frame_mvs + h * cm->mi_cols;
1712 for (w = 0; w < x_mis; ++w) {
1713 MV_REF *const mv = frame_mv + w;
1714 mv->ref_frame[0] = mi->mbmi.ref_frame[0];
1715 mv->ref_frame[1] = mi->mbmi.ref_frame[1];
1716 mv->mv[0].as_int = mi->mbmi.mv[0].as_int;
1717 mv->mv[1].as_int = mi->mbmi.mv[1].as_int;
1722 x->skip = ctx->skip;
1723 x->skip_txfm[0] = mbmi->segment_id ? 0 : ctx->skip_txfm[0];
1726 static void encode_b_rt(VP9_COMP *cpi, ThreadData *td,
1727 const TileInfo *const tile,
1728 TOKENEXTRA **tp, int mi_row, int mi_col,
1729 int output_enabled, BLOCK_SIZE bsize,
1730 PICK_MODE_CONTEXT *ctx) {
1731 MACROBLOCK *const x = &td->mb;
1732 set_offsets(cpi, tile, x, mi_row, mi_col, bsize);
1733 update_state_rt(cpi, td, ctx, mi_row, mi_col, bsize);
1735 #if CONFIG_VP9_TEMPORAL_DENOISING
1736 if (cpi->oxcf.noise_sensitivity > 0 && output_enabled &&
1737 cpi->common.frame_type != KEY_FRAME) {
1738 vp9_denoiser_denoise(&cpi->denoiser, x, mi_row, mi_col,
1739 MAX(BLOCK_8X8, bsize), ctx);
1743 encode_superblock(cpi, td, tp, output_enabled, mi_row, mi_col, bsize, ctx);
1744 update_stats(&cpi->common, td);
1746 (*tp)->token = EOSB_TOKEN;
1750 static void encode_sb_rt(VP9_COMP *cpi, ThreadData *td,
1751 const TileInfo *const tile,
1752 TOKENEXTRA **tp, int mi_row, int mi_col,
1753 int output_enabled, BLOCK_SIZE bsize,
1755 VP9_COMMON *const cm = &cpi->common;
1756 MACROBLOCK *const x = &td->mb;
1757 MACROBLOCKD *const xd = &x->e_mbd;
1759 const int bsl = b_width_log2_lookup[bsize], hbs = (1 << bsl) / 4;
1761 PARTITION_TYPE partition;
1764 if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols)
1767 if (bsize >= BLOCK_8X8) {
1768 const int idx_str = xd->mi_stride * mi_row + mi_col;
1769 MODE_INFO ** mi_8x8 = cm->mi_grid_visible + idx_str;
1770 ctx = partition_plane_context(xd, mi_row, mi_col, bsize);
1771 subsize = mi_8x8[0]->mbmi.sb_type;
1774 subsize = BLOCK_4X4;
1777 partition = partition_lookup[bsl][subsize];
1778 if (output_enabled && bsize != BLOCK_4X4)
1779 td->counts->partition[ctx][partition]++;
1781 switch (partition) {
1782 case PARTITION_NONE:
1783 encode_b_rt(cpi, td, tile, tp, mi_row, mi_col, output_enabled, subsize,
1786 case PARTITION_VERT:
1787 encode_b_rt(cpi, td, tile, tp, mi_row, mi_col, output_enabled, subsize,
1788 &pc_tree->vertical[0]);
1789 if (mi_col + hbs < cm->mi_cols && bsize > BLOCK_8X8) {
1790 encode_b_rt(cpi, td, tile, tp, mi_row, mi_col + hbs, output_enabled,
1791 subsize, &pc_tree->vertical[1]);
1794 case PARTITION_HORZ:
1795 encode_b_rt(cpi, td, tile, tp, mi_row, mi_col, output_enabled, subsize,
1796 &pc_tree->horizontal[0]);
1797 if (mi_row + hbs < cm->mi_rows && bsize > BLOCK_8X8) {
1798 encode_b_rt(cpi, td, tile, tp, mi_row + hbs, mi_col, output_enabled,
1799 subsize, &pc_tree->horizontal[1]);
1802 case PARTITION_SPLIT:
1803 subsize = get_subsize(bsize, PARTITION_SPLIT);
1804 encode_sb_rt(cpi, td, tile, tp, mi_row, mi_col, output_enabled, subsize,
1806 encode_sb_rt(cpi, td, tile, tp, mi_row, mi_col + hbs, output_enabled,
1807 subsize, pc_tree->split[1]);
1808 encode_sb_rt(cpi, td, tile, tp, mi_row + hbs, mi_col, output_enabled,
1809 subsize, pc_tree->split[2]);
1810 encode_sb_rt(cpi, td, tile, tp, mi_row + hbs, mi_col + hbs,
1811 output_enabled, subsize, pc_tree->split[3]);
1814 assert(0 && "Invalid partition type.");
1818 if (partition != PARTITION_SPLIT || bsize == BLOCK_8X8)
1819 update_partition_context(xd, mi_row, mi_col, subsize, bsize);
1822 static void rd_use_partition(VP9_COMP *cpi,
1824 TileDataEnc *tile_data,
1825 MODE_INFO **mi_8x8, TOKENEXTRA **tp,
1826 int mi_row, int mi_col,
1828 int *rate, int64_t *dist,
1829 int do_recon, PC_TREE *pc_tree) {
1830 VP9_COMMON *const cm = &cpi->common;
1831 TileInfo *const tile_info = &tile_data->tile_info;
1832 MACROBLOCK *const x = &td->mb;
1833 MACROBLOCKD *const xd = &x->e_mbd;
1834 const int mis = cm->mi_stride;
1835 const int bsl = b_width_log2_lookup[bsize];
1836 const int mi_step = num_4x4_blocks_wide_lookup[bsize] / 2;
1837 const int bss = (1 << bsl) / 4;
1839 PARTITION_TYPE partition = PARTITION_NONE;
1841 ENTROPY_CONTEXT l[16 * MAX_MB_PLANE], a[16 * MAX_MB_PLANE];
1842 PARTITION_CONTEXT sl[8], sa[8];
1843 RD_COST last_part_rdc, none_rdc, chosen_rdc;
1844 BLOCK_SIZE sub_subsize = BLOCK_4X4;
1845 int splits_below = 0;
1846 BLOCK_SIZE bs_type = mi_8x8[0]->mbmi.sb_type;
1847 int do_partition_search = 1;
1848 PICK_MODE_CONTEXT *ctx = &pc_tree->none;
1850 if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols)
1853 assert(num_4x4_blocks_wide_lookup[bsize] ==
1854 num_4x4_blocks_high_lookup[bsize]);
1856 vp9_rd_cost_reset(&last_part_rdc);
1857 vp9_rd_cost_reset(&none_rdc);
1858 vp9_rd_cost_reset(&chosen_rdc);
1860 partition = partition_lookup[bsl][bs_type];
1861 subsize = get_subsize(bsize, partition);
1863 pc_tree->partitioning = partition;
1864 save_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
1866 if (bsize == BLOCK_16X16 && cpi->oxcf.aq_mode) {
1867 set_offsets(cpi, tile_info, x, mi_row, mi_col, bsize);
1868 x->mb_energy = vp9_block_energy(cpi, x, bsize);
1871 if (do_partition_search &&
1872 cpi->sf.partition_search_type == SEARCH_PARTITION &&
1873 cpi->sf.adjust_partitioning_from_last_frame) {
1874 // Check if any of the sub blocks are further split.
1875 if (partition == PARTITION_SPLIT && subsize > BLOCK_8X8) {
1876 sub_subsize = get_subsize(subsize, PARTITION_SPLIT);
1878 for (i = 0; i < 4; i++) {
1879 int jj = i >> 1, ii = i & 0x01;
1880 MODE_INFO *this_mi = mi_8x8[jj * bss * mis + ii * bss];
1881 if (this_mi && this_mi->mbmi.sb_type >= sub_subsize) {
1887 // If partition is not none try none unless each of the 4 splits are split
1889 if (partition != PARTITION_NONE && !splits_below &&
1890 mi_row + (mi_step >> 1) < cm->mi_rows &&
1891 mi_col + (mi_step >> 1) < cm->mi_cols) {
1892 pc_tree->partitioning = PARTITION_NONE;
1893 rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &none_rdc, bsize,
1896 pl = partition_plane_context(xd, mi_row, mi_col, bsize);
1898 if (none_rdc.rate < INT_MAX) {
1899 none_rdc.rate += cpi->partition_cost[pl][PARTITION_NONE];
1900 none_rdc.rdcost = RDCOST(x->rdmult, x->rddiv, none_rdc.rate,
1904 restore_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
1905 mi_8x8[0]->mbmi.sb_type = bs_type;
1906 pc_tree->partitioning = partition;
1910 switch (partition) {
1911 case PARTITION_NONE:
1912 rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &last_part_rdc,
1913 bsize, ctx, INT64_MAX);
1915 case PARTITION_HORZ:
1916 rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &last_part_rdc,
1917 subsize, &pc_tree->horizontal[0],
1919 if (last_part_rdc.rate != INT_MAX &&
1920 bsize >= BLOCK_8X8 && mi_row + (mi_step >> 1) < cm->mi_rows) {
1922 PICK_MODE_CONTEXT *ctx = &pc_tree->horizontal[0];
1923 vp9_rd_cost_init(&tmp_rdc);
1924 update_state(cpi, td, ctx, mi_row, mi_col, subsize, 0);
1925 encode_superblock(cpi, td, tp, 0, mi_row, mi_col, subsize, ctx);
1926 rd_pick_sb_modes(cpi, tile_data, x,
1927 mi_row + (mi_step >> 1), mi_col, &tmp_rdc,
1928 subsize, &pc_tree->horizontal[1], INT64_MAX);
1929 if (tmp_rdc.rate == INT_MAX || tmp_rdc.dist == INT64_MAX) {
1930 vp9_rd_cost_reset(&last_part_rdc);
1933 last_part_rdc.rate += tmp_rdc.rate;
1934 last_part_rdc.dist += tmp_rdc.dist;
1935 last_part_rdc.rdcost += tmp_rdc.rdcost;
1938 case PARTITION_VERT:
1939 rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &last_part_rdc,
1940 subsize, &pc_tree->vertical[0], INT64_MAX);
1941 if (last_part_rdc.rate != INT_MAX &&
1942 bsize >= BLOCK_8X8 && mi_col + (mi_step >> 1) < cm->mi_cols) {
1944 PICK_MODE_CONTEXT *ctx = &pc_tree->vertical[0];
1945 vp9_rd_cost_init(&tmp_rdc);
1946 update_state(cpi, td, ctx, mi_row, mi_col, subsize, 0);
1947 encode_superblock(cpi, td, tp, 0, mi_row, mi_col, subsize, ctx);
1948 rd_pick_sb_modes(cpi, tile_data, x,
1949 mi_row, mi_col + (mi_step >> 1), &tmp_rdc,
1950 subsize, &pc_tree->vertical[bsize > BLOCK_8X8],
1952 if (tmp_rdc.rate == INT_MAX || tmp_rdc.dist == INT64_MAX) {
1953 vp9_rd_cost_reset(&last_part_rdc);
1956 last_part_rdc.rate += tmp_rdc.rate;
1957 last_part_rdc.dist += tmp_rdc.dist;
1958 last_part_rdc.rdcost += tmp_rdc.rdcost;
1961 case PARTITION_SPLIT:
1962 if (bsize == BLOCK_8X8) {
1963 rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &last_part_rdc,
1964 subsize, pc_tree->leaf_split[0], INT64_MAX);
1967 last_part_rdc.rate = 0;
1968 last_part_rdc.dist = 0;
1969 last_part_rdc.rdcost = 0;
1970 for (i = 0; i < 4; i++) {
1971 int x_idx = (i & 1) * (mi_step >> 1);
1972 int y_idx = (i >> 1) * (mi_step >> 1);
1973 int jj = i >> 1, ii = i & 0x01;
1975 if ((mi_row + y_idx >= cm->mi_rows) || (mi_col + x_idx >= cm->mi_cols))
1978 vp9_rd_cost_init(&tmp_rdc);
1979 rd_use_partition(cpi, td, tile_data,
1980 mi_8x8 + jj * bss * mis + ii * bss, tp,
1981 mi_row + y_idx, mi_col + x_idx, subsize,
1982 &tmp_rdc.rate, &tmp_rdc.dist,
1983 i != 3, pc_tree->split[i]);
1984 if (tmp_rdc.rate == INT_MAX || tmp_rdc.dist == INT64_MAX) {
1985 vp9_rd_cost_reset(&last_part_rdc);
1988 last_part_rdc.rate += tmp_rdc.rate;
1989 last_part_rdc.dist += tmp_rdc.dist;
1997 pl = partition_plane_context(xd, mi_row, mi_col, bsize);
1998 if (last_part_rdc.rate < INT_MAX) {
1999 last_part_rdc.rate += cpi->partition_cost[pl][partition];
2000 last_part_rdc.rdcost = RDCOST(x->rdmult, x->rddiv,
2001 last_part_rdc.rate, last_part_rdc.dist);
2004 if (do_partition_search
2005 && cpi->sf.adjust_partitioning_from_last_frame
2006 && cpi->sf.partition_search_type == SEARCH_PARTITION
2007 && partition != PARTITION_SPLIT && bsize > BLOCK_8X8
2008 && (mi_row + mi_step < cm->mi_rows ||
2009 mi_row + (mi_step >> 1) == cm->mi_rows)
2010 && (mi_col + mi_step < cm->mi_cols ||
2011 mi_col + (mi_step >> 1) == cm->mi_cols)) {
2012 BLOCK_SIZE split_subsize = get_subsize(bsize, PARTITION_SPLIT);
2013 chosen_rdc.rate = 0;
2014 chosen_rdc.dist = 0;
2015 restore_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
2016 pc_tree->partitioning = PARTITION_SPLIT;
2019 for (i = 0; i < 4; i++) {
2020 int x_idx = (i & 1) * (mi_step >> 1);
2021 int y_idx = (i >> 1) * (mi_step >> 1);
2023 ENTROPY_CONTEXT l[16 * MAX_MB_PLANE], a[16 * MAX_MB_PLANE];
2024 PARTITION_CONTEXT sl[8], sa[8];
2026 if ((mi_row + y_idx >= cm->mi_rows) || (mi_col + x_idx >= cm->mi_cols))
2029 save_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
2030 pc_tree->split[i]->partitioning = PARTITION_NONE;
2031 rd_pick_sb_modes(cpi, tile_data, x,
2032 mi_row + y_idx, mi_col + x_idx, &tmp_rdc,
2033 split_subsize, &pc_tree->split[i]->none, INT64_MAX);
2035 restore_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
2037 if (tmp_rdc.rate == INT_MAX || tmp_rdc.dist == INT64_MAX) {
2038 vp9_rd_cost_reset(&chosen_rdc);
2042 chosen_rdc.rate += tmp_rdc.rate;
2043 chosen_rdc.dist += tmp_rdc.dist;
2046 encode_sb(cpi, td, tile_info, tp, mi_row + y_idx, mi_col + x_idx, 0,
2047 split_subsize, pc_tree->split[i]);
2049 pl = partition_plane_context(xd, mi_row + y_idx, mi_col + x_idx,
2051 chosen_rdc.rate += cpi->partition_cost[pl][PARTITION_NONE];
2053 pl = partition_plane_context(xd, mi_row, mi_col, bsize);
2054 if (chosen_rdc.rate < INT_MAX) {
2055 chosen_rdc.rate += cpi->partition_cost[pl][PARTITION_SPLIT];
2056 chosen_rdc.rdcost = RDCOST(x->rdmult, x->rddiv,
2057 chosen_rdc.rate, chosen_rdc.dist);
2061 // If last_part is better set the partitioning to that.
2062 if (last_part_rdc.rdcost < chosen_rdc.rdcost) {
2063 mi_8x8[0]->mbmi.sb_type = bsize;
2064 if (bsize >= BLOCK_8X8)
2065 pc_tree->partitioning = partition;
2066 chosen_rdc = last_part_rdc;
2068 // If none was better set the partitioning to that.
2069 if (none_rdc.rdcost < chosen_rdc.rdcost) {
2070 if (bsize >= BLOCK_8X8)
2071 pc_tree->partitioning = PARTITION_NONE;
2072 chosen_rdc = none_rdc;
2075 restore_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
2077 // We must have chosen a partitioning and encoding or we'll fail later on.
2078 // No other opportunities for success.
2079 if (bsize == BLOCK_64X64)
2080 assert(chosen_rdc.rate < INT_MAX && chosen_rdc.dist < INT64_MAX);
2083 int output_enabled = (bsize == BLOCK_64X64);
2084 encode_sb(cpi, td, tile_info, tp, mi_row, mi_col, output_enabled, bsize,
2088 *rate = chosen_rdc.rate;
2089 *dist = chosen_rdc.dist;
2092 static const BLOCK_SIZE min_partition_size[BLOCK_SIZES] = {
2093 BLOCK_4X4, BLOCK_4X4, BLOCK_4X4,
2094 BLOCK_4X4, BLOCK_4X4, BLOCK_4X4,
2095 BLOCK_8X8, BLOCK_8X8, BLOCK_8X8,
2096 BLOCK_16X16, BLOCK_16X16, BLOCK_16X16,
2100 static const BLOCK_SIZE max_partition_size[BLOCK_SIZES] = {
2101 BLOCK_8X8, BLOCK_16X16, BLOCK_16X16,
2102 BLOCK_16X16, BLOCK_32X32, BLOCK_32X32,
2103 BLOCK_32X32, BLOCK_64X64, BLOCK_64X64,
2104 BLOCK_64X64, BLOCK_64X64, BLOCK_64X64,
2108 // Look at all the mode_info entries for blocks that are part of this
2109 // partition and find the min and max values for sb_type.
2110 // At the moment this is designed to work on a 64x64 SB but could be
2111 // adjusted to use a size parameter.
2113 // The min and max are assumed to have been initialized prior to calling this
2114 // function so repeat calls can accumulate a min and max of more than one sb64.
2115 static void get_sb_partition_size_range(MACROBLOCKD *xd, MODE_INFO **mi_8x8,
2116 BLOCK_SIZE *min_block_size,
2117 BLOCK_SIZE *max_block_size,
2118 int bs_hist[BLOCK_SIZES]) {
2119 int sb_width_in_blocks = MI_BLOCK_SIZE;
2120 int sb_height_in_blocks = MI_BLOCK_SIZE;
2124 // Check the sb_type for each block that belongs to this region.
2125 for (i = 0; i < sb_height_in_blocks; ++i) {
2126 for (j = 0; j < sb_width_in_blocks; ++j) {
2127 MODE_INFO *mi = mi_8x8[index+j];
2128 BLOCK_SIZE sb_type = mi ? mi->mbmi.sb_type : 0;
2130 *min_block_size = MIN(*min_block_size, sb_type);
2131 *max_block_size = MAX(*max_block_size, sb_type);
2133 index += xd->mi_stride;
2137 // Next square block size less or equal than current block size.
2138 static const BLOCK_SIZE next_square_size[BLOCK_SIZES] = {
2139 BLOCK_4X4, BLOCK_4X4, BLOCK_4X4,
2140 BLOCK_8X8, BLOCK_8X8, BLOCK_8X8,
2141 BLOCK_16X16, BLOCK_16X16, BLOCK_16X16,
2142 BLOCK_32X32, BLOCK_32X32, BLOCK_32X32,
2146 // Look at neighboring blocks and set a min and max partition size based on
2148 static void rd_auto_partition_range(VP9_COMP *cpi, const TileInfo *const tile,
2149 MACROBLOCKD *const xd,
2150 int mi_row, int mi_col,
2151 BLOCK_SIZE *min_block_size,
2152 BLOCK_SIZE *max_block_size) {
2153 VP9_COMMON *const cm = &cpi->common;
2154 MODE_INFO **mi = xd->mi;
2155 const int left_in_image = xd->left_available && mi[-1];
2156 const int above_in_image = xd->up_available && mi[-xd->mi_stride];
2157 const int row8x8_remaining = tile->mi_row_end - mi_row;
2158 const int col8x8_remaining = tile->mi_col_end - mi_col;
2160 BLOCK_SIZE min_size = BLOCK_4X4;
2161 BLOCK_SIZE max_size = BLOCK_64X64;
2162 int bs_hist[BLOCK_SIZES] = {0};
2164 // Trap case where we do not have a prediction.
2165 if (left_in_image || above_in_image || cm->frame_type != KEY_FRAME) {
2166 // Default "min to max" and "max to min"
2167 min_size = BLOCK_64X64;
2168 max_size = BLOCK_4X4;
2170 // NOTE: each call to get_sb_partition_size_range() uses the previous
2171 // passed in values for min and max as a starting point.
2172 // Find the min and max partition used in previous frame at this location
2173 if (cm->frame_type != KEY_FRAME) {
2174 MODE_INFO **prev_mi =
2175 &cm->prev_mi_grid_visible[mi_row * xd->mi_stride + mi_col];
2176 get_sb_partition_size_range(xd, prev_mi, &min_size, &max_size, bs_hist);
2178 // Find the min and max partition sizes used in the left SB64
2179 if (left_in_image) {
2180 MODE_INFO **left_sb64_mi = &mi[-MI_BLOCK_SIZE];
2181 get_sb_partition_size_range(xd, left_sb64_mi, &min_size, &max_size,
2184 // Find the min and max partition sizes used in the above SB64.
2185 if (above_in_image) {
2186 MODE_INFO **above_sb64_mi = &mi[-xd->mi_stride * MI_BLOCK_SIZE];
2187 get_sb_partition_size_range(xd, above_sb64_mi, &min_size, &max_size,
2191 // Adjust observed min and max for "relaxed" auto partition case.
2192 if (cpi->sf.auto_min_max_partition_size == RELAXED_NEIGHBORING_MIN_MAX) {
2193 min_size = min_partition_size[min_size];
2194 max_size = max_partition_size[max_size];
2198 // Check border cases where max and min from neighbors may not be legal.
2199 max_size = find_partition_size(max_size,
2200 row8x8_remaining, col8x8_remaining,
2202 min_size = MIN(cpi->sf.rd_auto_partition_min_limit, MIN(min_size, max_size));
2204 // When use_square_partition_only is true, make sure at least one square
2205 // partition is allowed by selecting the next smaller square size as
2207 if (cpi->sf.use_square_partition_only &&
2208 next_square_size[max_size] < min_size) {
2209 min_size = next_square_size[max_size];
2212 *min_block_size = min_size;
2213 *max_block_size = max_size;
2216 static void auto_partition_range(VP9_COMP *cpi, const TileInfo *const tile,
2217 MACROBLOCKD *const xd,
2218 int mi_row, int mi_col,
2219 BLOCK_SIZE *min_block_size,
2220 BLOCK_SIZE *max_block_size) {
2221 VP9_COMMON *const cm = &cpi->common;
2222 MODE_INFO **mi_8x8 = xd->mi;
2223 const int left_in_image = xd->left_available && mi_8x8[-1];
2224 const int above_in_image = xd->up_available && mi_8x8[-xd->mi_stride];
2225 int row8x8_remaining = tile->mi_row_end - mi_row;
2226 int col8x8_remaining = tile->mi_col_end - mi_col;
2228 BLOCK_SIZE min_size = BLOCK_32X32;
2229 BLOCK_SIZE max_size = BLOCK_8X8;
2230 int bsl = mi_width_log2_lookup[BLOCK_64X64];
2231 const int search_range_ctrl = (((mi_row + mi_col) >> bsl) +
2232 get_chessboard_index(cm->current_video_frame)) & 0x1;
2233 // Trap case where we do not have a prediction.
2234 if (search_range_ctrl &&
2235 (left_in_image || above_in_image || cm->frame_type != KEY_FRAME)) {
2240 // Find the min and max partition sizes used in the left SB64.
2241 if (left_in_image) {
2244 for (block = 0; block < MI_BLOCK_SIZE; ++block) {
2245 cur_mi = mi[block * xd->mi_stride];
2246 sb_type = cur_mi ? cur_mi->mbmi.sb_type : 0;
2247 min_size = MIN(min_size, sb_type);
2248 max_size = MAX(max_size, sb_type);
2251 // Find the min and max partition sizes used in the above SB64.
2252 if (above_in_image) {
2253 mi = &mi_8x8[-xd->mi_stride * MI_BLOCK_SIZE];
2254 for (block = 0; block < MI_BLOCK_SIZE; ++block) {
2255 sb_type = mi[block] ? mi[block]->mbmi.sb_type : 0;
2256 min_size = MIN(min_size, sb_type);
2257 max_size = MAX(max_size, sb_type);
2261 min_size = min_partition_size[min_size];
2262 max_size = find_partition_size(max_size, row8x8_remaining, col8x8_remaining,
2264 min_size = MIN(min_size, max_size);
2265 min_size = MAX(min_size, BLOCK_8X8);
2266 max_size = MIN(max_size, BLOCK_32X32);
2268 min_size = BLOCK_8X8;
2269 max_size = BLOCK_32X32;
2272 *min_block_size = min_size;
2273 *max_block_size = max_size;
2276 // TODO(jingning) refactor functions setting partition search range
2277 static void set_partition_range(VP9_COMMON *cm, MACROBLOCKD *xd,
2278 int mi_row, int mi_col, BLOCK_SIZE bsize,
2279 BLOCK_SIZE *min_bs, BLOCK_SIZE *max_bs) {
2280 int mi_width = num_8x8_blocks_wide_lookup[bsize];
2281 int mi_height = num_8x8_blocks_high_lookup[bsize];
2285 const int idx_str = cm->mi_stride * mi_row + mi_col;
2286 MODE_INFO **prev_mi = &cm->prev_mi_grid_visible[idx_str];
2287 BLOCK_SIZE bs, min_size, max_size;
2289 min_size = BLOCK_64X64;
2290 max_size = BLOCK_4X4;
2293 for (idy = 0; idy < mi_height; ++idy) {
2294 for (idx = 0; idx < mi_width; ++idx) {
2295 mi = prev_mi[idy * cm->mi_stride + idx];
2296 bs = mi ? mi->mbmi.sb_type : bsize;
2297 min_size = MIN(min_size, bs);
2298 max_size = MAX(max_size, bs);
2303 if (xd->left_available) {
2304 for (idy = 0; idy < mi_height; ++idy) {
2305 mi = xd->mi[idy * cm->mi_stride - 1];
2306 bs = mi ? mi->mbmi.sb_type : bsize;
2307 min_size = MIN(min_size, bs);
2308 max_size = MAX(max_size, bs);
2312 if (xd->up_available) {
2313 for (idx = 0; idx < mi_width; ++idx) {
2314 mi = xd->mi[idx - cm->mi_stride];
2315 bs = mi ? mi->mbmi.sb_type : bsize;
2316 min_size = MIN(min_size, bs);
2317 max_size = MAX(max_size, bs);
2321 if (min_size == max_size) {
2322 min_size = min_partition_size[min_size];
2323 max_size = max_partition_size[max_size];
2330 static INLINE void store_pred_mv(MACROBLOCK *x, PICK_MODE_CONTEXT *ctx) {
2331 memcpy(ctx->pred_mv, x->pred_mv, sizeof(x->pred_mv));
2334 static INLINE void load_pred_mv(MACROBLOCK *x, PICK_MODE_CONTEXT *ctx) {
2335 memcpy(x->pred_mv, ctx->pred_mv, sizeof(x->pred_mv));
2338 #if CONFIG_FP_MB_STATS
2339 const int num_16x16_blocks_wide_lookup[BLOCK_SIZES] =
2340 {1, 1, 1, 1, 1, 1, 1, 1, 2, 2, 2, 4, 4};
2341 const int num_16x16_blocks_high_lookup[BLOCK_SIZES] =
2342 {1, 1, 1, 1, 1, 1, 1, 2, 1, 2, 4, 2, 4};
2343 const int qindex_skip_threshold_lookup[BLOCK_SIZES] =
2344 {0, 10, 10, 30, 40, 40, 60, 80, 80, 90, 100, 100, 120};
2345 const int qindex_split_threshold_lookup[BLOCK_SIZES] =
2346 {0, 3, 3, 7, 15, 15, 30, 40, 40, 60, 80, 80, 120};
2347 const int complexity_16x16_blocks_threshold[BLOCK_SIZES] =
2348 {1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 4, 4, 6};
2359 static INLINE MOTION_DIRECTION get_motion_direction_fp(uint8_t fp_byte) {
2360 if (fp_byte & FPMB_MOTION_ZERO_MASK) {
2362 } else if (fp_byte & FPMB_MOTION_LEFT_MASK) {
2364 } else if (fp_byte & FPMB_MOTION_RIGHT_MASK) {
2366 } else if (fp_byte & FPMB_MOTION_UP_MASK) {
2373 static INLINE int get_motion_inconsistency(MOTION_DIRECTION this_mv,
2374 MOTION_DIRECTION that_mv) {
2375 if (this_mv == that_mv) {
2378 return abs(this_mv - that_mv) == 2 ? 2 : 1;
2383 // TODO(jingning,jimbankoski,rbultje): properly skip partition types that are
2384 // unlikely to be selected depending on previous rate-distortion optimization
2385 // results, for encoding speed-up.
2386 static void rd_pick_partition(VP9_COMP *cpi, ThreadData *td,
2387 TileDataEnc *tile_data,
2388 TOKENEXTRA **tp, int mi_row, int mi_col,
2389 BLOCK_SIZE bsize, RD_COST *rd_cost,
2390 int64_t best_rd, PC_TREE *pc_tree) {
2391 VP9_COMMON *const cm = &cpi->common;
2392 TileInfo *const tile_info = &tile_data->tile_info;
2393 MACROBLOCK *const x = &td->mb;
2394 MACROBLOCKD *const xd = &x->e_mbd;
2395 const int mi_step = num_8x8_blocks_wide_lookup[bsize] / 2;
2396 ENTROPY_CONTEXT l[16 * MAX_MB_PLANE], a[16 * MAX_MB_PLANE];
2397 PARTITION_CONTEXT sl[8], sa[8];
2398 TOKENEXTRA *tp_orig = *tp;
2399 PICK_MODE_CONTEXT *ctx = &pc_tree->none;
2402 RD_COST this_rdc, sum_rdc, best_rdc;
2403 int do_split = bsize >= BLOCK_8X8;
2406 // Override skipping rectangular partition operations for edge blocks
2407 const int force_horz_split = (mi_row + mi_step >= cm->mi_rows);
2408 const int force_vert_split = (mi_col + mi_step >= cm->mi_cols);
2409 const int xss = x->e_mbd.plane[1].subsampling_x;
2410 const int yss = x->e_mbd.plane[1].subsampling_y;
2412 BLOCK_SIZE min_size = x->min_partition_size;
2413 BLOCK_SIZE max_size = x->max_partition_size;
2415 #if CONFIG_FP_MB_STATS
2416 unsigned int src_diff_var = UINT_MAX;
2417 int none_complexity = 0;
2420 int partition_none_allowed = !force_horz_split && !force_vert_split;
2421 int partition_horz_allowed = !force_vert_split && yss <= xss &&
2423 int partition_vert_allowed = !force_horz_split && xss <= yss &&
2427 assert(num_8x8_blocks_wide_lookup[bsize] ==
2428 num_8x8_blocks_high_lookup[bsize]);
2430 vp9_rd_cost_init(&this_rdc);
2431 vp9_rd_cost_init(&sum_rdc);
2432 vp9_rd_cost_reset(&best_rdc);
2433 best_rdc.rdcost = best_rd;
2435 set_offsets(cpi, tile_info, x, mi_row, mi_col, bsize);
2437 if (bsize == BLOCK_16X16 && cpi->oxcf.aq_mode)
2438 x->mb_energy = vp9_block_energy(cpi, x, bsize);
2440 if (cpi->sf.cb_partition_search && bsize == BLOCK_16X16) {
2441 int cb_partition_search_ctrl = ((pc_tree->index == 0 || pc_tree->index == 3)
2442 + get_chessboard_index(cm->current_video_frame)) & 0x1;
2444 if (cb_partition_search_ctrl && bsize > min_size && bsize < max_size)
2445 set_partition_range(cm, xd, mi_row, mi_col, bsize, &min_size, &max_size);
2448 // Determine partition types in search according to the speed features.
2449 // The threshold set here has to be of square block size.
2450 if (cpi->sf.auto_min_max_partition_size) {
2451 partition_none_allowed &= (bsize <= max_size && bsize >= min_size);
2452 partition_horz_allowed &= ((bsize <= max_size && bsize > min_size) ||
2454 partition_vert_allowed &= ((bsize <= max_size && bsize > min_size) ||
2456 do_split &= bsize > min_size;
2458 if (cpi->sf.use_square_partition_only) {
2459 partition_horz_allowed &= force_horz_split;
2460 partition_vert_allowed &= force_vert_split;
2463 save_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
2465 #if CONFIG_FP_MB_STATS
2466 if (cpi->use_fp_mb_stats) {
2467 set_offsets(cpi, tile_info, x, mi_row, mi_col, bsize);
2468 src_diff_var = get_sby_perpixel_diff_variance(cpi, &x->plane[0].src,
2469 mi_row, mi_col, bsize);
2473 #if CONFIG_FP_MB_STATS
2474 // Decide whether we shall split directly and skip searching NONE by using
2475 // the first pass block statistics
2476 if (cpi->use_fp_mb_stats && bsize >= BLOCK_32X32 && do_split &&
2477 partition_none_allowed && src_diff_var > 4 &&
2478 cm->base_qindex < qindex_split_threshold_lookup[bsize]) {
2479 int mb_row = mi_row >> 1;
2480 int mb_col = mi_col >> 1;
2482 MIN(mb_row + num_16x16_blocks_high_lookup[bsize], cm->mb_rows);
2484 MIN(mb_col + num_16x16_blocks_wide_lookup[bsize], cm->mb_cols);
2487 // compute a complexity measure, basically measure inconsistency of motion
2488 // vectors obtained from the first pass in the current block
2489 for (r = mb_row; r < mb_row_end ; r++) {
2490 for (c = mb_col; c < mb_col_end; c++) {
2491 const int mb_index = r * cm->mb_cols + c;
2493 MOTION_DIRECTION this_mv;
2494 MOTION_DIRECTION right_mv;
2495 MOTION_DIRECTION bottom_mv;
2498 get_motion_direction_fp(cpi->twopass.this_frame_mb_stats[mb_index]);
2501 if (c != mb_col_end - 1) {
2502 right_mv = get_motion_direction_fp(
2503 cpi->twopass.this_frame_mb_stats[mb_index + 1]);
2504 none_complexity += get_motion_inconsistency(this_mv, right_mv);
2508 if (r != mb_row_end - 1) {
2509 bottom_mv = get_motion_direction_fp(
2510 cpi->twopass.this_frame_mb_stats[mb_index + cm->mb_cols]);
2511 none_complexity += get_motion_inconsistency(this_mv, bottom_mv);
2514 // do not count its left and top neighbors to avoid double counting
2518 if (none_complexity > complexity_16x16_blocks_threshold[bsize]) {
2519 partition_none_allowed = 0;
2525 if (partition_none_allowed) {
2526 rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col,
2527 &this_rdc, bsize, ctx, best_rdc.rdcost);
2528 if (this_rdc.rate != INT_MAX) {
2529 if (bsize >= BLOCK_8X8) {
2530 pl = partition_plane_context(xd, mi_row, mi_col, bsize);
2531 this_rdc.rate += cpi->partition_cost[pl][PARTITION_NONE];
2532 this_rdc.rdcost = RDCOST(x->rdmult, x->rddiv,
2533 this_rdc.rate, this_rdc.dist);
2536 if (this_rdc.rdcost < best_rdc.rdcost) {
2537 int64_t dist_breakout_thr = cpi->sf.partition_search_breakout_dist_thr;
2538 int rate_breakout_thr = cpi->sf.partition_search_breakout_rate_thr;
2540 best_rdc = this_rdc;
2541 if (bsize >= BLOCK_8X8)
2542 pc_tree->partitioning = PARTITION_NONE;
2544 // Adjust dist breakout threshold according to the partition size.
2545 dist_breakout_thr >>= 8 - (b_width_log2_lookup[bsize] +
2546 b_height_log2_lookup[bsize]);
2548 rate_breakout_thr *= num_pels_log2_lookup[bsize];
2550 // If all y, u, v transform blocks in this partition are skippable, and
2551 // the dist & rate are within the thresholds, the partition search is
2552 // terminated for current branch of the partition search tree.
2553 // The dist & rate thresholds are set to 0 at speed 0 to disable the
2554 // early termination at that speed.
2555 if (!x->e_mbd.lossless &&
2556 (ctx->skippable && best_rdc.dist < dist_breakout_thr &&
2557 best_rdc.rate < rate_breakout_thr)) {
2562 #if CONFIG_FP_MB_STATS
2563 // Check if every 16x16 first pass block statistics has zero
2564 // motion and the corresponding first pass residue is small enough.
2565 // If that is the case, check the difference variance between the
2566 // current frame and the last frame. If the variance is small enough,
2567 // stop further splitting in RD optimization
2568 if (cpi->use_fp_mb_stats && do_split != 0 &&
2569 cm->base_qindex > qindex_skip_threshold_lookup[bsize]) {
2570 int mb_row = mi_row >> 1;
2571 int mb_col = mi_col >> 1;
2573 MIN(mb_row + num_16x16_blocks_high_lookup[bsize], cm->mb_rows);
2575 MIN(mb_col + num_16x16_blocks_wide_lookup[bsize], cm->mb_cols);
2579 for (r = mb_row; r < mb_row_end; r++) {
2580 for (c = mb_col; c < mb_col_end; c++) {
2581 const int mb_index = r * cm->mb_cols + c;
2582 if (!(cpi->twopass.this_frame_mb_stats[mb_index] &
2583 FPMB_MOTION_ZERO_MASK) ||
2584 !(cpi->twopass.this_frame_mb_stats[mb_index] &
2585 FPMB_ERROR_SMALL_MASK)) {
2595 if (src_diff_var == UINT_MAX) {
2596 set_offsets(cpi, tile_info, x, mi_row, mi_col, bsize);
2597 src_diff_var = get_sby_perpixel_diff_variance(
2598 cpi, &x->plane[0].src, mi_row, mi_col, bsize);
2600 if (src_diff_var < 8) {
2609 restore_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
2612 // store estimated motion vector
2613 if (cpi->sf.adaptive_motion_search)
2614 store_pred_mv(x, ctx);
2617 // TODO(jingning): use the motion vectors given by the above search as
2618 // the starting point of motion search in the following partition type check.
2620 subsize = get_subsize(bsize, PARTITION_SPLIT);
2621 if (bsize == BLOCK_8X8) {
2623 if (cpi->sf.adaptive_pred_interp_filter && partition_none_allowed)
2624 pc_tree->leaf_split[0]->pred_interp_filter =
2625 ctx->mic.mbmi.interp_filter;
2626 rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &sum_rdc, subsize,
2627 pc_tree->leaf_split[0], best_rdc.rdcost);
2628 if (sum_rdc.rate == INT_MAX)
2629 sum_rdc.rdcost = INT64_MAX;
2631 for (i = 0; i < 4 && sum_rdc.rdcost < best_rdc.rdcost; ++i) {
2632 const int x_idx = (i & 1) * mi_step;
2633 const int y_idx = (i >> 1) * mi_step;
2635 if (mi_row + y_idx >= cm->mi_rows || mi_col + x_idx >= cm->mi_cols)
2638 if (cpi->sf.adaptive_motion_search)
2639 load_pred_mv(x, ctx);
2641 pc_tree->split[i]->index = i;
2642 rd_pick_partition(cpi, td, tile_data, tp,
2643 mi_row + y_idx, mi_col + x_idx,
2645 best_rdc.rdcost - sum_rdc.rdcost, pc_tree->split[i]);
2647 if (this_rdc.rate == INT_MAX) {
2648 sum_rdc.rdcost = INT64_MAX;
2651 sum_rdc.rate += this_rdc.rate;
2652 sum_rdc.dist += this_rdc.dist;
2653 sum_rdc.rdcost += this_rdc.rdcost;
2658 if (sum_rdc.rdcost < best_rdc.rdcost && i == 4) {
2659 pl = partition_plane_context(xd, mi_row, mi_col, bsize);
2660 sum_rdc.rate += cpi->partition_cost[pl][PARTITION_SPLIT];
2661 sum_rdc.rdcost = RDCOST(x->rdmult, x->rddiv,
2662 sum_rdc.rate, sum_rdc.dist);
2664 if (sum_rdc.rdcost < best_rdc.rdcost) {
2666 pc_tree->partitioning = PARTITION_SPLIT;
2669 // skip rectangular partition test when larger block size
2670 // gives better rd cost
2671 if (cpi->sf.less_rectangular_check)
2672 do_rect &= !partition_none_allowed;
2674 restore_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
2678 if (partition_horz_allowed && do_rect) {
2679 subsize = get_subsize(bsize, PARTITION_HORZ);
2680 if (cpi->sf.adaptive_motion_search)
2681 load_pred_mv(x, ctx);
2682 if (cpi->sf.adaptive_pred_interp_filter && bsize == BLOCK_8X8 &&
2683 partition_none_allowed)
2684 pc_tree->horizontal[0].pred_interp_filter =
2685 ctx->mic.mbmi.interp_filter;
2686 rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &sum_rdc, subsize,
2687 &pc_tree->horizontal[0], best_rdc.rdcost);
2689 if (sum_rdc.rdcost < best_rdc.rdcost && mi_row + mi_step < cm->mi_rows &&
2690 bsize > BLOCK_8X8) {
2691 PICK_MODE_CONTEXT *ctx = &pc_tree->horizontal[0];
2692 update_state(cpi, td, ctx, mi_row, mi_col, subsize, 0);
2693 encode_superblock(cpi, td, tp, 0, mi_row, mi_col, subsize, ctx);
2695 if (cpi->sf.adaptive_motion_search)
2696 load_pred_mv(x, ctx);
2697 if (cpi->sf.adaptive_pred_interp_filter && bsize == BLOCK_8X8 &&
2698 partition_none_allowed)
2699 pc_tree->horizontal[1].pred_interp_filter =
2700 ctx->mic.mbmi.interp_filter;
2701 rd_pick_sb_modes(cpi, tile_data, x, mi_row + mi_step, mi_col,
2702 &this_rdc, subsize, &pc_tree->horizontal[1],
2703 best_rdc.rdcost - sum_rdc.rdcost);
2704 if (this_rdc.rate == INT_MAX) {
2705 sum_rdc.rdcost = INT64_MAX;
2707 sum_rdc.rate += this_rdc.rate;
2708 sum_rdc.dist += this_rdc.dist;
2709 sum_rdc.rdcost += this_rdc.rdcost;
2713 if (sum_rdc.rdcost < best_rdc.rdcost) {
2714 pl = partition_plane_context(xd, mi_row, mi_col, bsize);
2715 sum_rdc.rate += cpi->partition_cost[pl][PARTITION_HORZ];
2716 sum_rdc.rdcost = RDCOST(x->rdmult, x->rddiv, sum_rdc.rate, sum_rdc.dist);
2717 if (sum_rdc.rdcost < best_rdc.rdcost) {
2719 pc_tree->partitioning = PARTITION_HORZ;
2722 restore_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
2725 if (partition_vert_allowed && do_rect) {
2726 subsize = get_subsize(bsize, PARTITION_VERT);
2728 if (cpi->sf.adaptive_motion_search)
2729 load_pred_mv(x, ctx);
2730 if (cpi->sf.adaptive_pred_interp_filter && bsize == BLOCK_8X8 &&
2731 partition_none_allowed)
2732 pc_tree->vertical[0].pred_interp_filter =
2733 ctx->mic.mbmi.interp_filter;
2734 rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &sum_rdc, subsize,
2735 &pc_tree->vertical[0], best_rdc.rdcost);
2736 if (sum_rdc.rdcost < best_rdc.rdcost && mi_col + mi_step < cm->mi_cols &&
2737 bsize > BLOCK_8X8) {
2738 update_state(cpi, td, &pc_tree->vertical[0], mi_row, mi_col, subsize, 0);
2739 encode_superblock(cpi, td, tp, 0, mi_row, mi_col, subsize,
2740 &pc_tree->vertical[0]);
2742 if (cpi->sf.adaptive_motion_search)
2743 load_pred_mv(x, ctx);
2744 if (cpi->sf.adaptive_pred_interp_filter && bsize == BLOCK_8X8 &&
2745 partition_none_allowed)
2746 pc_tree->vertical[1].pred_interp_filter =
2747 ctx->mic.mbmi.interp_filter;
2748 rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col + mi_step,
2750 &pc_tree->vertical[1], best_rdc.rdcost - sum_rdc.rdcost);
2751 if (this_rdc.rate == INT_MAX) {
2752 sum_rdc.rdcost = INT64_MAX;
2754 sum_rdc.rate += this_rdc.rate;
2755 sum_rdc.dist += this_rdc.dist;
2756 sum_rdc.rdcost += this_rdc.rdcost;
2760 if (sum_rdc.rdcost < best_rdc.rdcost) {
2761 pl = partition_plane_context(xd, mi_row, mi_col, bsize);
2762 sum_rdc.rate += cpi->partition_cost[pl][PARTITION_VERT];
2763 sum_rdc.rdcost = RDCOST(x->rdmult, x->rddiv,
2764 sum_rdc.rate, sum_rdc.dist);
2765 if (sum_rdc.rdcost < best_rdc.rdcost) {
2767 pc_tree->partitioning = PARTITION_VERT;
2770 restore_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
2773 // TODO(jbb): This code added so that we avoid static analysis
2774 // warning related to the fact that best_rd isn't used after this
2775 // point. This code should be refactored so that the duplicate
2776 // checks occur in some sub function and thus are used...
2778 *rd_cost = best_rdc;
2781 if (best_rdc.rate < INT_MAX && best_rdc.dist < INT64_MAX &&
2782 pc_tree->index != 3) {
2783 int output_enabled = (bsize == BLOCK_64X64);
2784 encode_sb(cpi, td, tile_info, tp, mi_row, mi_col, output_enabled,
2788 if (bsize == BLOCK_64X64) {
2789 assert(tp_orig < *tp);
2790 assert(best_rdc.rate < INT_MAX);
2791 assert(best_rdc.dist < INT64_MAX);
2793 assert(tp_orig == *tp);
2797 static void encode_rd_sb_row(VP9_COMP *cpi,
2799 TileDataEnc *tile_data,
2802 VP9_COMMON *const cm = &cpi->common;
2803 TileInfo *const tile_info = &tile_data->tile_info;
2804 MACROBLOCK *const x = &td->mb;
2805 MACROBLOCKD *const xd = &x->e_mbd;
2806 SPEED_FEATURES *const sf = &cpi->sf;
2809 // Initialize the left context for the new SB row
2810 memset(&xd->left_context, 0, sizeof(xd->left_context));
2811 memset(xd->left_seg_context, 0, sizeof(xd->left_seg_context));
2813 // Code each SB in the row
2814 for (mi_col = tile_info->mi_col_start; mi_col < tile_info->mi_col_end;
2815 mi_col += MI_BLOCK_SIZE) {
2816 const struct segmentation *const seg = &cm->seg;
2823 const int idx_str = cm->mi_stride * mi_row + mi_col;
2824 MODE_INFO **mi = cm->mi_grid_visible + idx_str;
2826 if (sf->adaptive_pred_interp_filter) {
2827 for (i = 0; i < 64; ++i)
2828 td->leaf_tree[i].pred_interp_filter = SWITCHABLE;
2830 for (i = 0; i < 64; ++i) {
2831 td->pc_tree[i].vertical[0].pred_interp_filter = SWITCHABLE;
2832 td->pc_tree[i].vertical[1].pred_interp_filter = SWITCHABLE;
2833 td->pc_tree[i].horizontal[0].pred_interp_filter = SWITCHABLE;
2834 td->pc_tree[i].horizontal[1].pred_interp_filter = SWITCHABLE;
2838 vp9_zero(x->pred_mv);
2839 td->pc_root->index = 0;
2842 const uint8_t *const map = seg->update_map ? cpi->segmentation_map
2843 : cm->last_frame_seg_map;
2844 int segment_id = vp9_get_segment_id(cm, map, BLOCK_64X64, mi_row, mi_col);
2845 seg_skip = vp9_segfeature_active(seg, segment_id, SEG_LVL_SKIP);
2848 x->source_variance = UINT_MAX;
2849 if (sf->partition_search_type == FIXED_PARTITION || seg_skip) {
2850 const BLOCK_SIZE bsize =
2851 seg_skip ? BLOCK_64X64 : sf->always_this_block_size;
2852 set_offsets(cpi, tile_info, x, mi_row, mi_col, BLOCK_64X64);
2853 set_fixed_partitioning(cpi, tile_info, mi, mi_row, mi_col, bsize);
2854 rd_use_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col,
2855 BLOCK_64X64, &dummy_rate, &dummy_dist, 1, td->pc_root);
2856 } else if (cpi->partition_search_skippable_frame) {
2858 set_offsets(cpi, tile_info, x, mi_row, mi_col, BLOCK_64X64);
2859 bsize = get_rd_var_based_fixed_partition(cpi, x, mi_row, mi_col);
2860 set_fixed_partitioning(cpi, tile_info, mi, mi_row, mi_col, bsize);
2861 rd_use_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col,
2862 BLOCK_64X64, &dummy_rate, &dummy_dist, 1, td->pc_root);
2863 } else if (sf->partition_search_type == VAR_BASED_PARTITION &&
2864 cm->frame_type != KEY_FRAME) {
2865 choose_partitioning(cpi, tile_info, x, mi_row, mi_col);
2866 rd_use_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col,
2867 BLOCK_64X64, &dummy_rate, &dummy_dist, 1, td->pc_root);
2869 // If required set upper and lower partition size limits
2870 if (sf->auto_min_max_partition_size) {
2871 set_offsets(cpi, tile_info, x, mi_row, mi_col, BLOCK_64X64);
2872 rd_auto_partition_range(cpi, tile_info, xd, mi_row, mi_col,
2873 &x->min_partition_size,
2874 &x->max_partition_size);
2876 rd_pick_partition(cpi, td, tile_data, tp, mi_row, mi_col, BLOCK_64X64,
2877 &dummy_rdc, INT64_MAX, td->pc_root);
2882 static void init_encode_frame_mb_context(VP9_COMP *cpi) {
2883 MACROBLOCK *const x = &cpi->td.mb;
2884 VP9_COMMON *const cm = &cpi->common;
2885 MACROBLOCKD *const xd = &x->e_mbd;
2886 const int aligned_mi_cols = mi_cols_aligned_to_sb(cm->mi_cols);
2888 // Copy data over into macro block data structures.
2889 vp9_setup_src_planes(x, cpi->Source, 0, 0);
2891 vp9_setup_block_planes(&x->e_mbd, cm->subsampling_x, cm->subsampling_y);
2893 // Note: this memset assumes above_context[0], [1] and [2]
2894 // are allocated as part of the same buffer.
2895 memset(xd->above_context[0], 0,
2896 sizeof(*xd->above_context[0]) *
2897 2 * aligned_mi_cols * MAX_MB_PLANE);
2898 memset(xd->above_seg_context, 0,
2899 sizeof(*xd->above_seg_context) * aligned_mi_cols);
2902 static int check_dual_ref_flags(VP9_COMP *cpi) {
2903 const int ref_flags = cpi->ref_frame_flags;
2905 if (vp9_segfeature_active(&cpi->common.seg, 1, SEG_LVL_REF_FRAME)) {
2908 return (!!(ref_flags & VP9_GOLD_FLAG) + !!(ref_flags & VP9_LAST_FLAG)
2909 + !!(ref_flags & VP9_ALT_FLAG)) >= 2;
2913 static void reset_skip_tx_size(VP9_COMMON *cm, TX_SIZE max_tx_size) {
2915 const int mis = cm->mi_stride;
2916 MODE_INFO **mi_ptr = cm->mi_grid_visible;
2918 for (mi_row = 0; mi_row < cm->mi_rows; ++mi_row, mi_ptr += mis) {
2919 for (mi_col = 0; mi_col < cm->mi_cols; ++mi_col) {
2920 if (mi_ptr[mi_col]->mbmi.tx_size > max_tx_size)
2921 mi_ptr[mi_col]->mbmi.tx_size = max_tx_size;
2926 static MV_REFERENCE_FRAME get_frame_type(const VP9_COMP *cpi) {
2927 if (frame_is_intra_only(&cpi->common))
2929 else if (cpi->rc.is_src_frame_alt_ref && cpi->refresh_golden_frame)
2930 return ALTREF_FRAME;
2931 else if (cpi->refresh_golden_frame || cpi->refresh_alt_ref_frame)
2932 return GOLDEN_FRAME;
2937 static TX_MODE select_tx_mode(const VP9_COMP *cpi, MACROBLOCKD *const xd) {
2940 if (cpi->common.frame_type == KEY_FRAME &&
2941 cpi->sf.use_nonrd_pick_mode &&
2942 cpi->sf.partition_search_type == VAR_BASED_PARTITION)
2944 if (cpi->sf.tx_size_search_method == USE_LARGESTALL)
2946 else if (cpi->sf.tx_size_search_method == USE_FULL_RD||
2947 cpi->sf.tx_size_search_method == USE_TX_8X8)
2948 return TX_MODE_SELECT;
2950 return cpi->common.tx_mode;
2953 static void hybrid_intra_mode_search(VP9_COMP *cpi, MACROBLOCK *const x,
2954 RD_COST *rd_cost, BLOCK_SIZE bsize,
2955 PICK_MODE_CONTEXT *ctx) {
2956 if (bsize < BLOCK_16X16)
2957 vp9_rd_pick_intra_mode_sb(cpi, x, rd_cost, bsize, ctx, INT64_MAX);
2959 vp9_pick_intra_mode(cpi, x, rd_cost, bsize, ctx);
2962 static void nonrd_pick_sb_modes(VP9_COMP *cpi,
2963 TileDataEnc *tile_data, MACROBLOCK *const x,
2964 int mi_row, int mi_col, RD_COST *rd_cost,
2965 BLOCK_SIZE bsize, PICK_MODE_CONTEXT *ctx) {
2966 VP9_COMMON *const cm = &cpi->common;
2967 TileInfo *const tile_info = &tile_data->tile_info;
2968 MACROBLOCKD *const xd = &x->e_mbd;
2970 set_offsets(cpi, tile_info, x, mi_row, mi_col, bsize);
2971 mbmi = &xd->mi[0]->mbmi;
2972 mbmi->sb_type = bsize;
2974 if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ && cm->seg.enabled)
2975 if (cyclic_refresh_segment_id_boosted(mbmi->segment_id))
2976 x->rdmult = vp9_cyclic_refresh_get_rdmult(cpi->cyclic_refresh);
2978 if (cm->frame_type == KEY_FRAME)
2979 hybrid_intra_mode_search(cpi, x, rd_cost, bsize, ctx);
2980 else if (vp9_segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_SKIP))
2981 set_mode_info_seg_skip(x, cm->tx_mode, rd_cost, bsize);
2982 else if (bsize >= BLOCK_8X8)
2983 vp9_pick_inter_mode(cpi, x, tile_data, mi_row, mi_col,
2984 rd_cost, bsize, ctx);
2986 vp9_pick_inter_mode_sub8x8(cpi, x, tile_data, mi_row, mi_col,
2987 rd_cost, bsize, ctx);
2989 duplicate_mode_info_in_sb(cm, xd, mi_row, mi_col, bsize);
2991 if (rd_cost->rate == INT_MAX)
2992 vp9_rd_cost_reset(rd_cost);
2994 ctx->rate = rd_cost->rate;
2995 ctx->dist = rd_cost->dist;
2998 static void fill_mode_info_sb(VP9_COMMON *cm, MACROBLOCK *x,
2999 int mi_row, int mi_col,
3002 MACROBLOCKD *xd = &x->e_mbd;
3003 int bsl = b_width_log2_lookup[bsize], hbs = (1 << bsl) / 4;
3004 PARTITION_TYPE partition = pc_tree->partitioning;
3005 BLOCK_SIZE subsize = get_subsize(bsize, partition);
3007 assert(bsize >= BLOCK_8X8);
3009 if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols)
3012 switch (partition) {
3013 case PARTITION_NONE:
3014 set_mode_info_offsets(cm, xd, mi_row, mi_col);
3015 *(xd->mi[0]) = pc_tree->none.mic;
3016 duplicate_mode_info_in_sb(cm, xd, mi_row, mi_col, bsize);
3018 case PARTITION_VERT:
3019 set_mode_info_offsets(cm, xd, mi_row, mi_col);
3020 *(xd->mi[0]) = pc_tree->vertical[0].mic;
3021 duplicate_mode_info_in_sb(cm, xd, mi_row, mi_col, subsize);
3023 if (mi_col + hbs < cm->mi_cols) {
3024 set_mode_info_offsets(cm, xd, mi_row, mi_col + hbs);
3025 *(xd->mi[0]) = pc_tree->vertical[1].mic;
3026 duplicate_mode_info_in_sb(cm, xd, mi_row, mi_col + hbs, subsize);
3029 case PARTITION_HORZ:
3030 set_mode_info_offsets(cm, xd, mi_row, mi_col);
3031 *(xd->mi[0]) = pc_tree->horizontal[0].mic;
3032 duplicate_mode_info_in_sb(cm, xd, mi_row, mi_col, subsize);
3033 if (mi_row + hbs < cm->mi_rows) {
3034 set_mode_info_offsets(cm, xd, mi_row + hbs, mi_col);
3035 *(xd->mi[0]) = pc_tree->horizontal[1].mic;
3036 duplicate_mode_info_in_sb(cm, xd, mi_row + hbs, mi_col, subsize);
3039 case PARTITION_SPLIT: {
3040 fill_mode_info_sb(cm, x, mi_row, mi_col, subsize, pc_tree->split[0]);
3041 fill_mode_info_sb(cm, x, mi_row, mi_col + hbs, subsize,
3043 fill_mode_info_sb(cm, x, mi_row + hbs, mi_col, subsize,
3045 fill_mode_info_sb(cm, x, mi_row + hbs, mi_col + hbs, subsize,
3054 // Reset the prediction pixel ready flag recursively.
3055 static void pred_pixel_ready_reset(PC_TREE *pc_tree, BLOCK_SIZE bsize) {
3056 pc_tree->none.pred_pixel_ready = 0;
3057 pc_tree->horizontal[0].pred_pixel_ready = 0;
3058 pc_tree->horizontal[1].pred_pixel_ready = 0;
3059 pc_tree->vertical[0].pred_pixel_ready = 0;
3060 pc_tree->vertical[1].pred_pixel_ready = 0;
3062 if (bsize > BLOCK_8X8) {
3063 BLOCK_SIZE subsize = get_subsize(bsize, PARTITION_SPLIT);
3065 for (i = 0; i < 4; ++i)
3066 pred_pixel_ready_reset(pc_tree->split[i], subsize);
3070 static void nonrd_pick_partition(VP9_COMP *cpi, ThreadData *td,
3071 TileDataEnc *tile_data,
3072 TOKENEXTRA **tp, int mi_row,
3073 int mi_col, BLOCK_SIZE bsize, RD_COST *rd_cost,
3074 int do_recon, int64_t best_rd,
3076 const SPEED_FEATURES *const sf = &cpi->sf;
3077 VP9_COMMON *const cm = &cpi->common;
3078 TileInfo *const tile_info = &tile_data->tile_info;
3079 MACROBLOCK *const x = &td->mb;
3080 MACROBLOCKD *const xd = &x->e_mbd;
3081 const int ms = num_8x8_blocks_wide_lookup[bsize] / 2;
3082 TOKENEXTRA *tp_orig = *tp;
3083 PICK_MODE_CONTEXT *ctx = &pc_tree->none;
3085 BLOCK_SIZE subsize = bsize;
3086 RD_COST this_rdc, sum_rdc, best_rdc;
3087 int do_split = bsize >= BLOCK_8X8;
3089 // Override skipping rectangular partition operations for edge blocks
3090 const int force_horz_split = (mi_row + ms >= cm->mi_rows);
3091 const int force_vert_split = (mi_col + ms >= cm->mi_cols);
3092 const int xss = x->e_mbd.plane[1].subsampling_x;
3093 const int yss = x->e_mbd.plane[1].subsampling_y;
3095 int partition_none_allowed = !force_horz_split && !force_vert_split;
3096 int partition_horz_allowed = !force_vert_split && yss <= xss &&
3098 int partition_vert_allowed = !force_horz_split && xss <= yss &&
3102 assert(num_8x8_blocks_wide_lookup[bsize] ==
3103 num_8x8_blocks_high_lookup[bsize]);
3105 vp9_rd_cost_init(&sum_rdc);
3106 vp9_rd_cost_reset(&best_rdc);
3107 best_rdc.rdcost = best_rd;
3109 // Determine partition types in search according to the speed features.
3110 // The threshold set here has to be of square block size.
3111 if (sf->auto_min_max_partition_size) {
3112 partition_none_allowed &= (bsize <= x->max_partition_size &&
3113 bsize >= x->min_partition_size);
3114 partition_horz_allowed &= ((bsize <= x->max_partition_size &&
3115 bsize > x->min_partition_size) ||
3117 partition_vert_allowed &= ((bsize <= x->max_partition_size &&
3118 bsize > x->min_partition_size) ||
3120 do_split &= bsize > x->min_partition_size;
3122 if (sf->use_square_partition_only) {
3123 partition_horz_allowed &= force_horz_split;
3124 partition_vert_allowed &= force_vert_split;
3127 ctx->pred_pixel_ready = !(partition_vert_allowed ||
3128 partition_horz_allowed ||
3132 if (partition_none_allowed) {
3133 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col,
3134 &this_rdc, bsize, ctx);
3135 ctx->mic.mbmi = xd->mi[0]->mbmi;
3136 ctx->skip_txfm[0] = x->skip_txfm[0];
3137 ctx->skip = x->skip;
3139 if (this_rdc.rate != INT_MAX) {
3140 int pl = partition_plane_context(xd, mi_row, mi_col, bsize);
3141 this_rdc.rate += cpi->partition_cost[pl][PARTITION_NONE];
3142 this_rdc.rdcost = RDCOST(x->rdmult, x->rddiv,
3143 this_rdc.rate, this_rdc.dist);
3144 if (this_rdc.rdcost < best_rdc.rdcost) {
3145 int64_t dist_breakout_thr = sf->partition_search_breakout_dist_thr;
3146 int64_t rate_breakout_thr = sf->partition_search_breakout_rate_thr;
3148 dist_breakout_thr >>= 8 - (b_width_log2_lookup[bsize] +
3149 b_height_log2_lookup[bsize]);
3151 rate_breakout_thr *= num_pels_log2_lookup[bsize];
3153 best_rdc = this_rdc;
3154 if (bsize >= BLOCK_8X8)
3155 pc_tree->partitioning = PARTITION_NONE;
3157 if (!x->e_mbd.lossless &&
3158 this_rdc.rate < rate_breakout_thr &&
3159 this_rdc.dist < dist_breakout_thr) {
3167 // store estimated motion vector
3168 store_pred_mv(x, ctx);
3172 int pl = partition_plane_context(xd, mi_row, mi_col, bsize);
3173 sum_rdc.rate += cpi->partition_cost[pl][PARTITION_SPLIT];
3174 sum_rdc.rdcost = RDCOST(x->rdmult, x->rddiv, sum_rdc.rate, sum_rdc.dist);
3175 subsize = get_subsize(bsize, PARTITION_SPLIT);
3176 for (i = 0; i < 4 && sum_rdc.rdcost < best_rdc.rdcost; ++i) {
3177 const int x_idx = (i & 1) * ms;
3178 const int y_idx = (i >> 1) * ms;
3180 if (mi_row + y_idx >= cm->mi_rows || mi_col + x_idx >= cm->mi_cols)
3182 load_pred_mv(x, ctx);
3183 nonrd_pick_partition(cpi, td, tile_data, tp,
3184 mi_row + y_idx, mi_col + x_idx,
3185 subsize, &this_rdc, 0,
3186 best_rdc.rdcost - sum_rdc.rdcost, pc_tree->split[i]);
3188 if (this_rdc.rate == INT_MAX) {
3189 vp9_rd_cost_reset(&sum_rdc);
3191 sum_rdc.rate += this_rdc.rate;
3192 sum_rdc.dist += this_rdc.dist;
3193 sum_rdc.rdcost += this_rdc.rdcost;
3197 if (sum_rdc.rdcost < best_rdc.rdcost) {
3199 pc_tree->partitioning = PARTITION_SPLIT;
3201 // skip rectangular partition test when larger block size
3202 // gives better rd cost
3203 if (sf->less_rectangular_check)
3204 do_rect &= !partition_none_allowed;
3209 if (partition_horz_allowed && do_rect) {
3210 subsize = get_subsize(bsize, PARTITION_HORZ);
3211 if (sf->adaptive_motion_search)
3212 load_pred_mv(x, ctx);
3213 pc_tree->horizontal[0].pred_pixel_ready = 1;
3214 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &sum_rdc, subsize,
3215 &pc_tree->horizontal[0]);
3217 pc_tree->horizontal[0].mic.mbmi = xd->mi[0]->mbmi;
3218 pc_tree->horizontal[0].skip_txfm[0] = x->skip_txfm[0];
3219 pc_tree->horizontal[0].skip = x->skip;
3221 if (sum_rdc.rdcost < best_rdc.rdcost && mi_row + ms < cm->mi_rows) {
3222 load_pred_mv(x, ctx);
3223 pc_tree->horizontal[1].pred_pixel_ready = 1;
3224 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row + ms, mi_col,
3226 &pc_tree->horizontal[1]);
3228 pc_tree->horizontal[1].mic.mbmi = xd->mi[0]->mbmi;
3229 pc_tree->horizontal[1].skip_txfm[0] = x->skip_txfm[0];
3230 pc_tree->horizontal[1].skip = x->skip;
3232 if (this_rdc.rate == INT_MAX) {
3233 vp9_rd_cost_reset(&sum_rdc);
3235 int pl = partition_plane_context(xd, mi_row, mi_col, bsize);
3236 this_rdc.rate += cpi->partition_cost[pl][PARTITION_HORZ];
3237 sum_rdc.rate += this_rdc.rate;
3238 sum_rdc.dist += this_rdc.dist;
3239 sum_rdc.rdcost = RDCOST(x->rdmult, x->rddiv,
3240 sum_rdc.rate, sum_rdc.dist);
3244 if (sum_rdc.rdcost < best_rdc.rdcost) {
3246 pc_tree->partitioning = PARTITION_HORZ;
3248 pred_pixel_ready_reset(pc_tree, bsize);
3253 if (partition_vert_allowed && do_rect) {
3254 subsize = get_subsize(bsize, PARTITION_VERT);
3255 if (sf->adaptive_motion_search)
3256 load_pred_mv(x, ctx);
3257 pc_tree->vertical[0].pred_pixel_ready = 1;
3258 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &sum_rdc, subsize,
3259 &pc_tree->vertical[0]);
3260 pc_tree->vertical[0].mic.mbmi = xd->mi[0]->mbmi;
3261 pc_tree->vertical[0].skip_txfm[0] = x->skip_txfm[0];
3262 pc_tree->vertical[0].skip = x->skip;
3264 if (sum_rdc.rdcost < best_rdc.rdcost && mi_col + ms < cm->mi_cols) {
3265 load_pred_mv(x, ctx);
3266 pc_tree->vertical[1].pred_pixel_ready = 1;
3267 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col + ms,
3269 &pc_tree->vertical[1]);
3270 pc_tree->vertical[1].mic.mbmi = xd->mi[0]->mbmi;
3271 pc_tree->vertical[1].skip_txfm[0] = x->skip_txfm[0];
3272 pc_tree->vertical[1].skip = x->skip;
3274 if (this_rdc.rate == INT_MAX) {
3275 vp9_rd_cost_reset(&sum_rdc);
3277 int pl = partition_plane_context(xd, mi_row, mi_col, bsize);
3278 sum_rdc.rate += cpi->partition_cost[pl][PARTITION_VERT];
3279 sum_rdc.rate += this_rdc.rate;
3280 sum_rdc.dist += this_rdc.dist;
3281 sum_rdc.rdcost = RDCOST(x->rdmult, x->rddiv,
3282 sum_rdc.rate, sum_rdc.dist);
3286 if (sum_rdc.rdcost < best_rdc.rdcost) {
3288 pc_tree->partitioning = PARTITION_VERT;
3290 pred_pixel_ready_reset(pc_tree, bsize);
3294 *rd_cost = best_rdc;
3296 if (best_rdc.rate == INT_MAX) {
3297 vp9_rd_cost_reset(rd_cost);
3301 // update mode info array
3302 fill_mode_info_sb(cm, x, mi_row, mi_col, bsize, pc_tree);
3304 if (best_rdc.rate < INT_MAX && best_rdc.dist < INT64_MAX && do_recon) {
3305 int output_enabled = (bsize == BLOCK_64X64);
3306 encode_sb_rt(cpi, td, tile_info, tp, mi_row, mi_col, output_enabled,
3310 if (bsize == BLOCK_64X64 && do_recon) {
3311 assert(tp_orig < *tp);
3312 assert(best_rdc.rate < INT_MAX);
3313 assert(best_rdc.dist < INT64_MAX);
3315 assert(tp_orig == *tp);
3319 static void nonrd_select_partition(VP9_COMP *cpi,
3321 TileDataEnc *tile_data,
3324 int mi_row, int mi_col,
3325 BLOCK_SIZE bsize, int output_enabled,
3326 RD_COST *rd_cost, PC_TREE *pc_tree) {
3327 VP9_COMMON *const cm = &cpi->common;
3328 TileInfo *const tile_info = &tile_data->tile_info;
3329 MACROBLOCK *const x = &td->mb;
3330 MACROBLOCKD *const xd = &x->e_mbd;
3331 const int bsl = b_width_log2_lookup[bsize], hbs = (1 << bsl) / 4;
3332 const int mis = cm->mi_stride;
3333 PARTITION_TYPE partition;
3337 vp9_rd_cost_reset(&this_rdc);
3338 if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols)
3341 subsize = (bsize >= BLOCK_8X8) ? mi[0]->mbmi.sb_type : BLOCK_4X4;
3342 partition = partition_lookup[bsl][subsize];
3344 if (bsize == BLOCK_32X32 && partition != PARTITION_NONE &&
3345 subsize >= BLOCK_16X16) {
3346 x->max_partition_size = BLOCK_32X32;
3347 x->min_partition_size = BLOCK_8X8;
3348 nonrd_pick_partition(cpi, td, tile_data, tp, mi_row, mi_col, bsize,
3349 rd_cost, 0, INT64_MAX, pc_tree);
3350 } else if (bsize == BLOCK_16X16 && partition != PARTITION_NONE) {
3351 x->max_partition_size = BLOCK_16X16;
3352 x->min_partition_size = BLOCK_8X8;
3353 nonrd_pick_partition(cpi, td, tile_data, tp, mi_row, mi_col, bsize,
3354 rd_cost, 0, INT64_MAX, pc_tree);
3356 switch (partition) {
3357 case PARTITION_NONE:
3358 pc_tree->none.pred_pixel_ready = 1;
3359 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, rd_cost,
3360 subsize, &pc_tree->none);
3361 pc_tree->none.mic.mbmi = xd->mi[0]->mbmi;
3362 pc_tree->none.skip_txfm[0] = x->skip_txfm[0];
3363 pc_tree->none.skip = x->skip;
3365 case PARTITION_VERT:
3366 pc_tree->vertical[0].pred_pixel_ready = 1;
3367 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, rd_cost,
3368 subsize, &pc_tree->vertical[0]);
3369 pc_tree->vertical[0].mic.mbmi = xd->mi[0]->mbmi;
3370 pc_tree->vertical[0].skip_txfm[0] = x->skip_txfm[0];
3371 pc_tree->vertical[0].skip = x->skip;
3372 if (mi_col + hbs < cm->mi_cols) {
3373 pc_tree->vertical[1].pred_pixel_ready = 1;
3374 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col + hbs,
3375 &this_rdc, subsize, &pc_tree->vertical[1]);
3376 pc_tree->vertical[1].mic.mbmi = xd->mi[0]->mbmi;
3377 pc_tree->vertical[1].skip_txfm[0] = x->skip_txfm[0];
3378 pc_tree->vertical[1].skip = x->skip;
3379 if (this_rdc.rate != INT_MAX && this_rdc.dist != INT64_MAX &&
3380 rd_cost->rate != INT_MAX && rd_cost->dist != INT64_MAX) {
3381 rd_cost->rate += this_rdc.rate;
3382 rd_cost->dist += this_rdc.dist;
3386 case PARTITION_HORZ:
3387 pc_tree->horizontal[0].pred_pixel_ready = 1;
3388 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, rd_cost,
3389 subsize, &pc_tree->horizontal[0]);
3390 pc_tree->horizontal[0].mic.mbmi = xd->mi[0]->mbmi;
3391 pc_tree->horizontal[0].skip_txfm[0] = x->skip_txfm[0];
3392 pc_tree->horizontal[0].skip = x->skip;
3393 if (mi_row + hbs < cm->mi_rows) {
3394 pc_tree->horizontal[1].pred_pixel_ready = 1;
3395 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row + hbs, mi_col,
3396 &this_rdc, subsize, &pc_tree->horizontal[1]);
3397 pc_tree->horizontal[1].mic.mbmi = xd->mi[0]->mbmi;
3398 pc_tree->horizontal[1].skip_txfm[0] = x->skip_txfm[0];
3399 pc_tree->horizontal[1].skip = x->skip;
3400 if (this_rdc.rate != INT_MAX && this_rdc.dist != INT64_MAX &&
3401 rd_cost->rate != INT_MAX && rd_cost->dist != INT64_MAX) {
3402 rd_cost->rate += this_rdc.rate;
3403 rd_cost->dist += this_rdc.dist;
3407 case PARTITION_SPLIT:
3408 subsize = get_subsize(bsize, PARTITION_SPLIT);
3409 nonrd_select_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col,
3410 subsize, output_enabled, rd_cost,
3412 nonrd_select_partition(cpi, td, tile_data, mi + hbs, tp,
3413 mi_row, mi_col + hbs, subsize, output_enabled,
3414 &this_rdc, pc_tree->split[1]);
3415 if (this_rdc.rate != INT_MAX && this_rdc.dist != INT64_MAX &&
3416 rd_cost->rate != INT_MAX && rd_cost->dist != INT64_MAX) {
3417 rd_cost->rate += this_rdc.rate;
3418 rd_cost->dist += this_rdc.dist;
3420 nonrd_select_partition(cpi, td, tile_data, mi + hbs * mis, tp,
3421 mi_row + hbs, mi_col, subsize, output_enabled,
3422 &this_rdc, pc_tree->split[2]);
3423 if (this_rdc.rate != INT_MAX && this_rdc.dist != INT64_MAX &&
3424 rd_cost->rate != INT_MAX && rd_cost->dist != INT64_MAX) {
3425 rd_cost->rate += this_rdc.rate;
3426 rd_cost->dist += this_rdc.dist;
3428 nonrd_select_partition(cpi, td, tile_data, mi + hbs * mis + hbs, tp,
3429 mi_row + hbs, mi_col + hbs, subsize,
3430 output_enabled, &this_rdc, pc_tree->split[3]);
3431 if (this_rdc.rate != INT_MAX && this_rdc.dist != INT64_MAX &&
3432 rd_cost->rate != INT_MAX && rd_cost->dist != INT64_MAX) {
3433 rd_cost->rate += this_rdc.rate;
3434 rd_cost->dist += this_rdc.dist;
3438 assert(0 && "Invalid partition type.");
3443 if (bsize == BLOCK_64X64 && output_enabled)
3444 encode_sb_rt(cpi, td, tile_info, tp, mi_row, mi_col, 1, bsize, pc_tree);
3448 static void nonrd_use_partition(VP9_COMP *cpi,
3450 TileDataEnc *tile_data,
3453 int mi_row, int mi_col,
3454 BLOCK_SIZE bsize, int output_enabled,
3455 RD_COST *dummy_cost, PC_TREE *pc_tree) {
3456 VP9_COMMON *const cm = &cpi->common;
3457 TileInfo *tile_info = &tile_data->tile_info;
3458 MACROBLOCK *const x = &td->mb;
3459 MACROBLOCKD *const xd = &x->e_mbd;
3460 const int bsl = b_width_log2_lookup[bsize], hbs = (1 << bsl) / 4;
3461 const int mis = cm->mi_stride;
3462 PARTITION_TYPE partition;
3465 if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols)
3468 subsize = (bsize >= BLOCK_8X8) ? mi[0]->mbmi.sb_type : BLOCK_4X4;
3469 partition = partition_lookup[bsl][subsize];
3471 if (output_enabled && bsize != BLOCK_4X4) {
3472 int ctx = partition_plane_context(xd, mi_row, mi_col, bsize);
3473 td->counts->partition[ctx][partition]++;
3476 switch (partition) {
3477 case PARTITION_NONE:
3478 pc_tree->none.pred_pixel_ready = 1;
3479 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, dummy_cost,
3480 subsize, &pc_tree->none);
3481 pc_tree->none.mic.mbmi = xd->mi[0]->mbmi;
3482 pc_tree->none.skip_txfm[0] = x->skip_txfm[0];
3483 pc_tree->none.skip = x->skip;
3484 encode_b_rt(cpi, td, tile_info, tp, mi_row, mi_col, output_enabled,
3485 subsize, &pc_tree->none);
3487 case PARTITION_VERT:
3488 pc_tree->vertical[0].pred_pixel_ready = 1;
3489 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, dummy_cost,
3490 subsize, &pc_tree->vertical[0]);
3491 pc_tree->vertical[0].mic.mbmi = xd->mi[0]->mbmi;
3492 pc_tree->vertical[0].skip_txfm[0] = x->skip_txfm[0];
3493 pc_tree->vertical[0].skip = x->skip;
3494 encode_b_rt(cpi, td, tile_info, tp, mi_row, mi_col, output_enabled,
3495 subsize, &pc_tree->vertical[0]);
3496 if (mi_col + hbs < cm->mi_cols && bsize > BLOCK_8X8) {
3497 pc_tree->vertical[1].pred_pixel_ready = 1;
3498 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col + hbs,
3499 dummy_cost, subsize, &pc_tree->vertical[1]);
3500 pc_tree->vertical[1].mic.mbmi = xd->mi[0]->mbmi;
3501 pc_tree->vertical[1].skip_txfm[0] = x->skip_txfm[0];
3502 pc_tree->vertical[1].skip = x->skip;
3503 encode_b_rt(cpi, td, tile_info, tp, mi_row, mi_col + hbs,
3504 output_enabled, subsize, &pc_tree->vertical[1]);
3507 case PARTITION_HORZ:
3508 pc_tree->horizontal[0].pred_pixel_ready = 1;
3509 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, dummy_cost,
3510 subsize, &pc_tree->horizontal[0]);
3511 pc_tree->horizontal[0].mic.mbmi = xd->mi[0]->mbmi;
3512 pc_tree->horizontal[0].skip_txfm[0] = x->skip_txfm[0];
3513 pc_tree->horizontal[0].skip = x->skip;
3514 encode_b_rt(cpi, td, tile_info, tp, mi_row, mi_col, output_enabled,
3515 subsize, &pc_tree->horizontal[0]);
3517 if (mi_row + hbs < cm->mi_rows && bsize > BLOCK_8X8) {
3518 pc_tree->horizontal[1].pred_pixel_ready = 1;
3519 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row + hbs, mi_col,
3520 dummy_cost, subsize, &pc_tree->horizontal[1]);
3521 pc_tree->horizontal[1].mic.mbmi = xd->mi[0]->mbmi;
3522 pc_tree->horizontal[1].skip_txfm[0] = x->skip_txfm[0];
3523 pc_tree->horizontal[1].skip = x->skip;
3524 encode_b_rt(cpi, td, tile_info, tp, mi_row + hbs, mi_col,
3525 output_enabled, subsize, &pc_tree->horizontal[1]);
3528 case PARTITION_SPLIT:
3529 subsize = get_subsize(bsize, PARTITION_SPLIT);
3530 if (bsize == BLOCK_8X8) {
3531 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, dummy_cost,
3532 subsize, pc_tree->leaf_split[0]);
3533 encode_b_rt(cpi, td, tile_info, tp, mi_row, mi_col,
3534 output_enabled, subsize, pc_tree->leaf_split[0]);
3536 nonrd_use_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col,
3537 subsize, output_enabled, dummy_cost,
3539 nonrd_use_partition(cpi, td, tile_data, mi + hbs, tp,
3540 mi_row, mi_col + hbs, subsize, output_enabled,
3541 dummy_cost, pc_tree->split[1]);
3542 nonrd_use_partition(cpi, td, tile_data, mi + hbs * mis, tp,
3543 mi_row + hbs, mi_col, subsize, output_enabled,
3544 dummy_cost, pc_tree->split[2]);
3545 nonrd_use_partition(cpi, td, tile_data, mi + hbs * mis + hbs, tp,
3546 mi_row + hbs, mi_col + hbs, subsize, output_enabled,
3547 dummy_cost, pc_tree->split[3]);
3551 assert(0 && "Invalid partition type.");
3555 if (partition != PARTITION_SPLIT || bsize == BLOCK_8X8)
3556 update_partition_context(xd, mi_row, mi_col, subsize, bsize);
3559 static void encode_nonrd_sb_row(VP9_COMP *cpi,
3561 TileDataEnc *tile_data,
3564 SPEED_FEATURES *const sf = &cpi->sf;
3565 VP9_COMMON *const cm = &cpi->common;
3566 TileInfo *const tile_info = &tile_data->tile_info;
3567 MACROBLOCK *const x = &td->mb;
3568 MACROBLOCKD *const xd = &x->e_mbd;
3571 // Initialize the left context for the new SB row
3572 memset(&xd->left_context, 0, sizeof(xd->left_context));
3573 memset(xd->left_seg_context, 0, sizeof(xd->left_seg_context));
3575 // Code each SB in the row
3576 for (mi_col = tile_info->mi_col_start; mi_col < tile_info->mi_col_end;
3577 mi_col += MI_BLOCK_SIZE) {
3578 const struct segmentation *const seg = &cm->seg;
3580 const int idx_str = cm->mi_stride * mi_row + mi_col;
3581 MODE_INFO **mi = cm->mi_grid_visible + idx_str;
3582 PARTITION_SEARCH_TYPE partition_search_type = sf->partition_search_type;
3583 BLOCK_SIZE bsize = BLOCK_64X64;
3585 x->source_variance = UINT_MAX;
3586 vp9_zero(x->pred_mv);
3587 vp9_rd_cost_init(&dummy_rdc);
3588 x->color_sensitivity[0] = 0;
3589 x->color_sensitivity[1] = 0;
3592 const uint8_t *const map = seg->update_map ? cpi->segmentation_map
3593 : cm->last_frame_seg_map;
3594 int segment_id = vp9_get_segment_id(cm, map, BLOCK_64X64, mi_row, mi_col);
3595 seg_skip = vp9_segfeature_active(seg, segment_id, SEG_LVL_SKIP);
3597 partition_search_type = FIXED_PARTITION;
3601 // Set the partition type of the 64X64 block
3602 switch (partition_search_type) {
3603 case VAR_BASED_PARTITION:
3604 // TODO(jingning, marpan): The mode decision and encoding process
3605 // support both intra and inter sub8x8 block coding for RTC mode.
3606 // Tune the thresholds accordingly to use sub8x8 block coding for
3607 // coding performance improvement.
3608 choose_partitioning(cpi, tile_info, x, mi_row, mi_col);
3609 nonrd_use_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col,
3610 BLOCK_64X64, 1, &dummy_rdc, td->pc_root);
3612 case SOURCE_VAR_BASED_PARTITION:
3613 set_source_var_based_partition(cpi, tile_info, x, mi, mi_row, mi_col);
3614 nonrd_use_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col,
3615 BLOCK_64X64, 1, &dummy_rdc, td->pc_root);
3617 case FIXED_PARTITION:
3619 bsize = sf->always_this_block_size;
3620 set_fixed_partitioning(cpi, tile_info, mi, mi_row, mi_col, bsize);
3621 nonrd_use_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col,
3622 BLOCK_64X64, 1, &dummy_rdc, td->pc_root);
3624 case REFERENCE_PARTITION:
3625 set_offsets(cpi, tile_info, x, mi_row, mi_col, BLOCK_64X64);
3626 if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ && cm->seg.enabled &&
3627 xd->mi[0]->mbmi.segment_id) {
3628 x->max_partition_size = BLOCK_64X64;
3629 x->min_partition_size = BLOCK_8X8;
3630 nonrd_pick_partition(cpi, td, tile_data, tp, mi_row, mi_col,
3631 BLOCK_64X64, &dummy_rdc, 1,
3632 INT64_MAX, td->pc_root);
3634 choose_partitioning(cpi, tile_info, x, mi_row, mi_col);
3635 nonrd_select_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col,
3636 BLOCK_64X64, 1, &dummy_rdc, td->pc_root);
3646 // end RTC play code
3648 static int set_var_thresh_from_histogram(VP9_COMP *cpi) {
3649 const SPEED_FEATURES *const sf = &cpi->sf;
3650 const VP9_COMMON *const cm = &cpi->common;
3652 const uint8_t *src = cpi->Source->y_buffer;
3653 const uint8_t *last_src = cpi->Last_Source->y_buffer;
3654 const int src_stride = cpi->Source->y_stride;
3655 const int last_stride = cpi->Last_Source->y_stride;
3657 // Pick cutoff threshold
3658 const int cutoff = (MIN(cm->width, cm->height) >= 720) ?
3659 (cm->MBs * VAR_HIST_LARGE_CUT_OFF / 100) :
3660 (cm->MBs * VAR_HIST_SMALL_CUT_OFF / 100);
3661 DECLARE_ALIGNED(16, int, hist[VAR_HIST_BINS]);
3662 diff *var16 = cpi->source_diff_var;
3667 memset(hist, 0, VAR_HIST_BINS * sizeof(hist[0]));
3669 for (i = 0; i < cm->mb_rows; i++) {
3670 for (j = 0; j < cm->mb_cols; j++) {
3671 #if CONFIG_VP9_HIGHBITDEPTH
3672 if (cm->use_highbitdepth) {
3673 switch (cm->bit_depth) {
3675 vp9_highbd_get16x16var(src, src_stride, last_src, last_stride,
3676 &var16->sse, &var16->sum);
3679 vp9_highbd_10_get16x16var(src, src_stride, last_src, last_stride,
3680 &var16->sse, &var16->sum);
3683 vp9_highbd_12_get16x16var(src, src_stride, last_src, last_stride,
3684 &var16->sse, &var16->sum);
3687 assert(0 && "cm->bit_depth should be VPX_BITS_8, VPX_BITS_10"
3692 vp9_get16x16var(src, src_stride, last_src, last_stride,
3693 &var16->sse, &var16->sum);
3696 vp9_get16x16var(src, src_stride, last_src, last_stride,
3697 &var16->sse, &var16->sum);
3698 #endif // CONFIG_VP9_HIGHBITDEPTH
3699 var16->var = var16->sse -
3700 (((uint32_t)var16->sum * var16->sum) >> 8);
3702 if (var16->var >= VAR_HIST_MAX_BG_VAR)
3703 hist[VAR_HIST_BINS - 1]++;
3705 hist[var16->var / VAR_HIST_FACTOR]++;
3712 src = src - cm->mb_cols * 16 + 16 * src_stride;
3713 last_src = last_src - cm->mb_cols * 16 + 16 * last_stride;
3716 cpi->source_var_thresh = 0;
3718 if (hist[VAR_HIST_BINS - 1] < cutoff) {
3719 for (i = 0; i < VAR_HIST_BINS - 1; i++) {
3723 cpi->source_var_thresh = (i + 1) * VAR_HIST_FACTOR;
3729 return sf->search_type_check_frequency;
3732 static void source_var_based_partition_search_method(VP9_COMP *cpi) {
3733 VP9_COMMON *const cm = &cpi->common;
3734 SPEED_FEATURES *const sf = &cpi->sf;
3736 if (cm->frame_type == KEY_FRAME) {
3737 // For key frame, use SEARCH_PARTITION.
3738 sf->partition_search_type = SEARCH_PARTITION;
3739 } else if (cm->intra_only) {
3740 sf->partition_search_type = FIXED_PARTITION;
3742 if (cm->last_width != cm->width || cm->last_height != cm->height) {
3743 if (cpi->source_diff_var)
3744 vpx_free(cpi->source_diff_var);
3746 CHECK_MEM_ERROR(cm, cpi->source_diff_var,
3747 vpx_calloc(cm->MBs, sizeof(diff)));
3750 if (!cpi->frames_till_next_var_check)
3751 cpi->frames_till_next_var_check = set_var_thresh_from_histogram(cpi);
3753 if (cpi->frames_till_next_var_check > 0) {
3754 sf->partition_search_type = FIXED_PARTITION;
3755 cpi->frames_till_next_var_check--;
3760 static int get_skip_encode_frame(const VP9_COMMON *cm, ThreadData *const td) {
3761 unsigned int intra_count = 0, inter_count = 0;
3764 for (j = 0; j < INTRA_INTER_CONTEXTS; ++j) {
3765 intra_count += td->counts->intra_inter[j][0];
3766 inter_count += td->counts->intra_inter[j][1];
3769 return (intra_count << 2) < inter_count &&
3770 cm->frame_type != KEY_FRAME &&
3774 void vp9_init_tile_data(VP9_COMP *cpi) {
3775 VP9_COMMON *const cm = &cpi->common;
3776 const int tile_cols = 1 << cm->log2_tile_cols;
3777 const int tile_rows = 1 << cm->log2_tile_rows;
3778 int tile_col, tile_row;
3779 TOKENEXTRA *pre_tok = cpi->tile_tok[0][0];
3782 if (cpi->tile_data == NULL) {
3783 CHECK_MEM_ERROR(cm, cpi->tile_data,
3784 vpx_malloc(tile_cols * tile_rows * sizeof(*cpi->tile_data)));
3785 for (tile_row = 0; tile_row < tile_rows; ++tile_row)
3786 for (tile_col = 0; tile_col < tile_cols; ++tile_col) {
3787 TileDataEnc *tile_data =
3788 &cpi->tile_data[tile_row * tile_cols + tile_col];
3790 for (i = 0; i < BLOCK_SIZES; ++i) {
3791 for (j = 0; j < MAX_MODES; ++j) {
3792 tile_data->thresh_freq_fact[i][j] = 32;
3793 tile_data->mode_map[i][j] = j;
3799 for (tile_row = 0; tile_row < tile_rows; ++tile_row) {
3800 for (tile_col = 0; tile_col < tile_cols; ++tile_col) {
3801 TileInfo *tile_info =
3802 &cpi->tile_data[tile_row * tile_cols + tile_col].tile_info;
3803 vp9_tile_init(tile_info, cm, tile_row, tile_col);
3805 cpi->tile_tok[tile_row][tile_col] = pre_tok + tile_tok;
3806 pre_tok = cpi->tile_tok[tile_row][tile_col];
3807 tile_tok = allocated_tokens(*tile_info);
3812 void vp9_encode_tile(VP9_COMP *cpi, ThreadData *td,
3813 int tile_row, int tile_col) {
3814 VP9_COMMON *const cm = &cpi->common;
3815 const int tile_cols = 1 << cm->log2_tile_cols;
3816 TileDataEnc *this_tile =
3817 &cpi->tile_data[tile_row * tile_cols + tile_col];
3818 const TileInfo * const tile_info = &this_tile->tile_info;
3819 TOKENEXTRA *tok = cpi->tile_tok[tile_row][tile_col];
3822 for (mi_row = tile_info->mi_row_start; mi_row < tile_info->mi_row_end;
3823 mi_row += MI_BLOCK_SIZE) {
3824 if (cpi->sf.use_nonrd_pick_mode)
3825 encode_nonrd_sb_row(cpi, td, this_tile, mi_row, &tok);
3827 encode_rd_sb_row(cpi, td, this_tile, mi_row, &tok);
3829 cpi->tok_count[tile_row][tile_col] =
3830 (unsigned int)(tok - cpi->tile_tok[tile_row][tile_col]);
3831 assert(tok - cpi->tile_tok[tile_row][tile_col] <=
3832 allocated_tokens(*tile_info));
3835 static void encode_tiles(VP9_COMP *cpi) {
3836 VP9_COMMON *const cm = &cpi->common;
3837 const int tile_cols = 1 << cm->log2_tile_cols;
3838 const int tile_rows = 1 << cm->log2_tile_rows;
3839 int tile_col, tile_row;
3841 vp9_init_tile_data(cpi);
3843 for (tile_row = 0; tile_row < tile_rows; ++tile_row)
3844 for (tile_col = 0; tile_col < tile_cols; ++tile_col)
3845 vp9_encode_tile(cpi, &cpi->td, tile_row, tile_col);
3848 #if CONFIG_FP_MB_STATS
3849 static int input_fpmb_stats(FIRSTPASS_MB_STATS *firstpass_mb_stats,
3850 VP9_COMMON *cm, uint8_t **this_frame_mb_stats) {
3851 uint8_t *mb_stats_in = firstpass_mb_stats->mb_stats_start +
3852 cm->current_video_frame * cm->MBs * sizeof(uint8_t);
3854 if (mb_stats_in > firstpass_mb_stats->mb_stats_end)
3857 *this_frame_mb_stats = mb_stats_in;
3863 static void encode_frame_internal(VP9_COMP *cpi) {
3864 SPEED_FEATURES *const sf = &cpi->sf;
3865 RD_OPT *const rd_opt = &cpi->rd;
3866 ThreadData *const td = &cpi->td;
3867 MACROBLOCK *const x = &td->mb;
3868 VP9_COMMON *const cm = &cpi->common;
3869 MACROBLOCKD *const xd = &x->e_mbd;
3870 RD_COUNTS *const rdc = &cpi->td.rd_counts;
3872 xd->mi = cm->mi_grid_visible;
3875 vp9_zero(*td->counts);
3876 vp9_zero(rdc->coef_counts);
3877 vp9_zero(rdc->comp_pred_diff);
3878 vp9_zero(rdc->filter_diff);
3879 vp9_zero(rdc->tx_select_diff);
3880 vp9_zero(rd_opt->tx_select_threshes);
3882 xd->lossless = cm->base_qindex == 0 &&
3883 cm->y_dc_delta_q == 0 &&
3884 cm->uv_dc_delta_q == 0 &&
3885 cm->uv_ac_delta_q == 0;
3887 #if CONFIG_VP9_HIGHBITDEPTH
3888 if (cm->use_highbitdepth)
3889 x->fwd_txm4x4 = xd->lossless ? vp9_highbd_fwht4x4 : vp9_highbd_fdct4x4;
3891 x->fwd_txm4x4 = xd->lossless ? vp9_fwht4x4 : vp9_fdct4x4;
3892 x->highbd_itxm_add = xd->lossless ? vp9_highbd_iwht4x4_add :
3893 vp9_highbd_idct4x4_add;
3895 x->fwd_txm4x4 = xd->lossless ? vp9_fwht4x4 : vp9_fdct4x4;
3896 #endif // CONFIG_VP9_HIGHBITDEPTH
3897 x->itxm_add = xd->lossless ? vp9_iwht4x4_add : vp9_idct4x4_add;
3902 cm->tx_mode = select_tx_mode(cpi, xd);
3904 vp9_frame_init_quantizer(cpi);
3906 vp9_initialize_rd_consts(cpi);
3907 vp9_initialize_me_consts(cpi, x, cm->base_qindex);
3908 init_encode_frame_mb_context(cpi);
3909 cm->use_prev_frame_mvs = !cm->error_resilient_mode &&
3910 cm->width == cm->last_width &&
3911 cm->height == cm->last_height &&
3913 cm->last_show_frame;
3914 // Special case: set prev_mi to NULL when the previous mode info
3915 // context cannot be used.
3916 cm->prev_mi = cm->use_prev_frame_mvs ?
3917 cm->prev_mip + cm->mi_stride + 1 : NULL;
3919 x->quant_fp = cpi->sf.use_quant_fp;
3920 vp9_zero(x->skip_txfm);
3921 if (sf->use_nonrd_pick_mode) {
3922 // Initialize internal buffer pointers for rtc coding, where non-RD
3923 // mode decision is used and hence no buffer pointer swap needed.
3925 struct macroblock_plane *const p = x->plane;
3926 struct macroblockd_plane *const pd = xd->plane;
3927 PICK_MODE_CONTEXT *ctx = &cpi->td.pc_root->none;
3929 for (i = 0; i < MAX_MB_PLANE; ++i) {
3930 p[i].coeff = ctx->coeff_pbuf[i][0];
3931 p[i].qcoeff = ctx->qcoeff_pbuf[i][0];
3932 pd[i].dqcoeff = ctx->dqcoeff_pbuf[i][0];
3933 p[i].eobs = ctx->eobs_pbuf[i][0];
3935 vp9_zero(x->zcoeff_blk);
3937 if (cm->frame_type != KEY_FRAME && cpi->rc.frames_since_golden == 0)
3938 cpi->ref_frame_flags &= (~VP9_GOLD_FLAG);
3940 if (sf->partition_search_type == SOURCE_VAR_BASED_PARTITION)
3941 source_var_based_partition_search_method(cpi);
3945 struct vpx_usec_timer emr_timer;
3946 vpx_usec_timer_start(&emr_timer);
3948 #if CONFIG_FP_MB_STATS
3949 if (cpi->use_fp_mb_stats) {
3950 input_fpmb_stats(&cpi->twopass.firstpass_mb_stats, cm,
3951 &cpi->twopass.this_frame_mb_stats);
3955 // If allowed, encoding tiles in parallel with one thread handling one tile.
3956 if (MIN(cpi->oxcf.max_threads, 1 << cm->log2_tile_cols) > 1)
3957 vp9_encode_tiles_mt(cpi);
3961 vpx_usec_timer_mark(&emr_timer);
3962 cpi->time_encode_sb_row += vpx_usec_timer_elapsed(&emr_timer);
3965 sf->skip_encode_frame = sf->skip_encode_sb ?
3966 get_skip_encode_frame(cm, td) : 0;
3969 // Keep record of the total distortion this time around for future use
3970 cpi->last_frame_distortion = cpi->frame_distortion;
3974 static INTERP_FILTER get_interp_filter(
3975 const int64_t threshes[SWITCHABLE_FILTER_CONTEXTS], int is_alt_ref) {
3977 threshes[EIGHTTAP_SMOOTH] > threshes[EIGHTTAP] &&
3978 threshes[EIGHTTAP_SMOOTH] > threshes[EIGHTTAP_SHARP] &&
3979 threshes[EIGHTTAP_SMOOTH] > threshes[SWITCHABLE - 1]) {
3980 return EIGHTTAP_SMOOTH;
3981 } else if (threshes[EIGHTTAP_SHARP] > threshes[EIGHTTAP] &&
3982 threshes[EIGHTTAP_SHARP] > threshes[SWITCHABLE - 1]) {
3983 return EIGHTTAP_SHARP;
3984 } else if (threshes[EIGHTTAP] > threshes[SWITCHABLE - 1]) {
3991 void vp9_encode_frame(VP9_COMP *cpi) {
3992 VP9_COMMON *const cm = &cpi->common;
3994 // In the longer term the encoder should be generalized to match the
3995 // decoder such that we allow compound where one of the 3 buffers has a
3996 // different sign bias and that buffer is then the fixed ref. However, this
3997 // requires further work in the rd loop. For now the only supported encoder
3998 // side behavior is where the ALT ref buffer has opposite sign bias to
4000 if (!frame_is_intra_only(cm)) {
4001 if ((cm->ref_frame_sign_bias[ALTREF_FRAME] ==
4002 cm->ref_frame_sign_bias[GOLDEN_FRAME]) ||
4003 (cm->ref_frame_sign_bias[ALTREF_FRAME] ==
4004 cm->ref_frame_sign_bias[LAST_FRAME])) {
4005 cpi->allow_comp_inter_inter = 0;
4007 cpi->allow_comp_inter_inter = 1;
4008 cm->comp_fixed_ref = ALTREF_FRAME;
4009 cm->comp_var_ref[0] = LAST_FRAME;
4010 cm->comp_var_ref[1] = GOLDEN_FRAME;
4014 if (cpi->sf.frame_parameter_update) {
4016 RD_OPT *const rd_opt = &cpi->rd;
4017 FRAME_COUNTS *counts = cpi->td.counts;
4018 RD_COUNTS *const rdc = &cpi->td.rd_counts;
4020 // This code does a single RD pass over the whole frame assuming
4021 // either compound, single or hybrid prediction as per whatever has
4022 // worked best for that type of frame in the past.
4023 // It also predicts whether another coding mode would have worked
4024 // better that this coding mode. If that is the case, it remembers
4025 // that for subsequent frames.
4026 // It does the same analysis for transform size selection also.
4027 const MV_REFERENCE_FRAME frame_type = get_frame_type(cpi);
4028 int64_t *const mode_thrs = rd_opt->prediction_type_threshes[frame_type];
4029 int64_t *const filter_thrs = rd_opt->filter_threshes[frame_type];
4030 int *const tx_thrs = rd_opt->tx_select_threshes[frame_type];
4031 const int is_alt_ref = frame_type == ALTREF_FRAME;
4033 /* prediction (compound, single or hybrid) mode selection */
4034 if (is_alt_ref || !cpi->allow_comp_inter_inter)
4035 cm->reference_mode = SINGLE_REFERENCE;
4036 else if (mode_thrs[COMPOUND_REFERENCE] > mode_thrs[SINGLE_REFERENCE] &&
4037 mode_thrs[COMPOUND_REFERENCE] >
4038 mode_thrs[REFERENCE_MODE_SELECT] &&
4039 check_dual_ref_flags(cpi) &&
4040 cpi->static_mb_pct == 100)
4041 cm->reference_mode = COMPOUND_REFERENCE;
4042 else if (mode_thrs[SINGLE_REFERENCE] > mode_thrs[REFERENCE_MODE_SELECT])
4043 cm->reference_mode = SINGLE_REFERENCE;
4045 cm->reference_mode = REFERENCE_MODE_SELECT;
4047 if (cm->interp_filter == SWITCHABLE)
4048 cm->interp_filter = get_interp_filter(filter_thrs, is_alt_ref);
4050 encode_frame_internal(cpi);
4052 for (i = 0; i < REFERENCE_MODES; ++i)
4053 mode_thrs[i] = (mode_thrs[i] + rdc->comp_pred_diff[i] / cm->MBs) / 2;
4055 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i)
4056 filter_thrs[i] = (filter_thrs[i] + rdc->filter_diff[i] / cm->MBs) / 2;
4058 for (i = 0; i < TX_MODES; ++i) {
4059 int64_t pd = rdc->tx_select_diff[i];
4060 if (i == TX_MODE_SELECT)
4061 pd -= RDCOST(cpi->td.mb.rdmult, cpi->td.mb.rddiv, 2048 * (TX_SIZES - 1),
4063 tx_thrs[i] = (tx_thrs[i] + (int)(pd / cm->MBs)) / 2;
4066 if (cm->reference_mode == REFERENCE_MODE_SELECT) {
4067 int single_count_zero = 0;
4068 int comp_count_zero = 0;
4070 for (i = 0; i < COMP_INTER_CONTEXTS; i++) {
4071 single_count_zero += counts->comp_inter[i][0];
4072 comp_count_zero += counts->comp_inter[i][1];
4075 if (comp_count_zero == 0) {
4076 cm->reference_mode = SINGLE_REFERENCE;
4077 vp9_zero(counts->comp_inter);
4078 } else if (single_count_zero == 0) {
4079 cm->reference_mode = COMPOUND_REFERENCE;
4080 vp9_zero(counts->comp_inter);
4084 if (cm->tx_mode == TX_MODE_SELECT) {
4086 int count8x8_lp = 0, count8x8_8x8p = 0;
4087 int count16x16_16x16p = 0, count16x16_lp = 0;
4090 for (i = 0; i < TX_SIZE_CONTEXTS; ++i) {
4091 count4x4 += counts->tx.p32x32[i][TX_4X4];
4092 count4x4 += counts->tx.p16x16[i][TX_4X4];
4093 count4x4 += counts->tx.p8x8[i][TX_4X4];
4095 count8x8_lp += counts->tx.p32x32[i][TX_8X8];
4096 count8x8_lp += counts->tx.p16x16[i][TX_8X8];
4097 count8x8_8x8p += counts->tx.p8x8[i][TX_8X8];
4099 count16x16_16x16p += counts->tx.p16x16[i][TX_16X16];
4100 count16x16_lp += counts->tx.p32x32[i][TX_16X16];
4101 count32x32 += counts->tx.p32x32[i][TX_32X32];
4103 if (count4x4 == 0 && count16x16_lp == 0 && count16x16_16x16p == 0 &&
4105 cm->tx_mode = ALLOW_8X8;
4106 reset_skip_tx_size(cm, TX_8X8);
4107 } else if (count8x8_8x8p == 0 && count16x16_16x16p == 0 &&
4108 count8x8_lp == 0 && count16x16_lp == 0 && count32x32 == 0) {
4109 cm->tx_mode = ONLY_4X4;
4110 reset_skip_tx_size(cm, TX_4X4);
4111 } else if (count8x8_lp == 0 && count16x16_lp == 0 && count4x4 == 0) {
4112 cm->tx_mode = ALLOW_32X32;
4113 } else if (count32x32 == 0 && count8x8_lp == 0 && count4x4 == 0) {
4114 cm->tx_mode = ALLOW_16X16;
4115 reset_skip_tx_size(cm, TX_16X16);
4119 cm->reference_mode = SINGLE_REFERENCE;
4120 encode_frame_internal(cpi);
4124 static void sum_intra_stats(FRAME_COUNTS *counts, const MODE_INFO *mi) {
4125 const PREDICTION_MODE y_mode = mi->mbmi.mode;
4126 const PREDICTION_MODE uv_mode = mi->mbmi.uv_mode;
4127 const BLOCK_SIZE bsize = mi->mbmi.sb_type;
4129 if (bsize < BLOCK_8X8) {
4131 const int num_4x4_w = num_4x4_blocks_wide_lookup[bsize];
4132 const int num_4x4_h = num_4x4_blocks_high_lookup[bsize];
4133 for (idy = 0; idy < 2; idy += num_4x4_h)
4134 for (idx = 0; idx < 2; idx += num_4x4_w)
4135 ++counts->y_mode[0][mi->bmi[idy * 2 + idx].as_mode];
4137 ++counts->y_mode[size_group_lookup[bsize]][y_mode];
4140 ++counts->uv_mode[y_mode][uv_mode];
4143 static void encode_superblock(VP9_COMP *cpi, ThreadData *td,
4144 TOKENEXTRA **t, int output_enabled,
4145 int mi_row, int mi_col, BLOCK_SIZE bsize,
4146 PICK_MODE_CONTEXT *ctx) {
4147 VP9_COMMON *const cm = &cpi->common;
4148 MACROBLOCK *const x = &td->mb;
4149 MACROBLOCKD *const xd = &x->e_mbd;
4150 MODE_INFO **mi_8x8 = xd->mi;
4151 MODE_INFO *mi = mi_8x8[0];
4152 MB_MODE_INFO *mbmi = &mi->mbmi;
4153 const int seg_skip = vp9_segfeature_active(&cm->seg, mbmi->segment_id,
4155 const int mis = cm->mi_stride;
4156 const int mi_width = num_8x8_blocks_wide_lookup[bsize];
4157 const int mi_height = num_8x8_blocks_high_lookup[bsize];
4159 x->skip_recode = !x->select_tx_size && mbmi->sb_type >= BLOCK_8X8 &&
4160 cpi->oxcf.aq_mode != COMPLEXITY_AQ &&
4161 cpi->oxcf.aq_mode != CYCLIC_REFRESH_AQ &&
4162 cpi->sf.allow_skip_recode;
4164 if (!x->skip_recode && !cpi->sf.use_nonrd_pick_mode)
4165 memset(x->skip_txfm, 0, sizeof(x->skip_txfm));
4167 x->skip_optimize = ctx->is_coded;
4169 x->use_lp32x32fdct = cpi->sf.use_lp32x32fdct;
4170 x->skip_encode = (!output_enabled && cpi->sf.skip_encode_frame &&
4171 x->q_index < QIDX_SKIP_THRESH);
4176 set_ref_ptrs(cm, xd, mbmi->ref_frame[0], mbmi->ref_frame[1]);
4178 if (!is_inter_block(mbmi)) {
4181 for (plane = 0; plane < MAX_MB_PLANE; ++plane)
4182 vp9_encode_intra_block_plane(x, MAX(bsize, BLOCK_8X8), plane);
4184 sum_intra_stats(td->counts, mi);
4185 vp9_tokenize_sb(cpi, td, t, !output_enabled, MAX(bsize, BLOCK_8X8));
4188 const int is_compound = has_second_ref(mbmi);
4189 for (ref = 0; ref < 1 + is_compound; ++ref) {
4190 YV12_BUFFER_CONFIG *cfg = get_ref_frame_buffer(cpi,
4191 mbmi->ref_frame[ref]);
4192 assert(cfg != NULL);
4193 vp9_setup_pre_planes(xd, ref, cfg, mi_row, mi_col,
4194 &xd->block_refs[ref]->sf);
4196 if (!(cpi->sf.reuse_inter_pred_sby && ctx->pred_pixel_ready) || seg_skip)
4197 vp9_build_inter_predictors_sby(xd, mi_row, mi_col, MAX(bsize, BLOCK_8X8));
4199 vp9_build_inter_predictors_sbuv(xd, mi_row, mi_col, MAX(bsize, BLOCK_8X8));
4201 vp9_encode_sb(x, MAX(bsize, BLOCK_8X8));
4202 vp9_tokenize_sb(cpi, td, t, !output_enabled, MAX(bsize, BLOCK_8X8));
4205 if (output_enabled) {
4206 if (cm->tx_mode == TX_MODE_SELECT &&
4207 mbmi->sb_type >= BLOCK_8X8 &&
4208 !(is_inter_block(mbmi) && (mbmi->skip || seg_skip))) {
4209 ++get_tx_counts(max_txsize_lookup[bsize], vp9_get_tx_size_context(xd),
4210 &td->counts->tx)[mbmi->tx_size];
4214 // The new intra coding scheme requires no change of transform size
4215 if (is_inter_block(&mi->mbmi)) {
4216 tx_size = MIN(tx_mode_to_biggest_tx_size[cm->tx_mode],
4217 max_txsize_lookup[bsize]);
4219 tx_size = (bsize >= BLOCK_8X8) ? mbmi->tx_size : TX_4X4;
4222 for (y = 0; y < mi_height; y++)
4223 for (x = 0; x < mi_width; x++)
4224 if (mi_col + x < cm->mi_cols && mi_row + y < cm->mi_rows)
4225 mi_8x8[mis * y + x]->mbmi.tx_size = tx_size;
4227 ++td->counts->tx.tx_totals[mbmi->tx_size];
4228 ++td->counts->tx.tx_totals[get_uv_tx_size(mbmi, &xd->plane[1])];