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/vpx_timer.h"
20 #include "vp9/common/vp9_common.h"
21 #include "vp9/common/vp9_entropy.h"
22 #include "vp9/common/vp9_entropymode.h"
23 #include "vp9/common/vp9_idct.h"
24 #include "vp9/common/vp9_mvref_common.h"
25 #include "vp9/common/vp9_pred_common.h"
26 #include "vp9/common/vp9_quant_common.h"
27 #include "vp9/common/vp9_reconintra.h"
28 #include "vp9/common/vp9_reconinter.h"
29 #include "vp9/common/vp9_seg_common.h"
30 #include "vp9/common/vp9_systemdependent.h"
31 #include "vp9/common/vp9_tile_common.h"
33 #include "vp9/encoder/vp9_aq_complexity.h"
34 #include "vp9/encoder/vp9_aq_cyclicrefresh.h"
35 #include "vp9/encoder/vp9_aq_variance.h"
36 #include "vp9/encoder/vp9_encodeframe.h"
37 #include "vp9/encoder/vp9_encodemb.h"
38 #include "vp9/encoder/vp9_encodemv.h"
39 #include "vp9/encoder/vp9_ethread.h"
40 #include "vp9/encoder/vp9_extend.h"
41 #include "vp9/encoder/vp9_pickmode.h"
42 #include "vp9/encoder/vp9_rd.h"
43 #include "vp9/encoder/vp9_rdopt.h"
44 #include "vp9/encoder/vp9_segmentation.h"
45 #include "vp9/encoder/vp9_tokenize.h"
47 static void encode_superblock(VP9_COMP *cpi, ThreadData * td,
48 TOKENEXTRA **t, int output_enabled,
49 int mi_row, int mi_col, BLOCK_SIZE bsize,
50 PICK_MODE_CONTEXT *ctx);
52 // This is used as a reference when computing the source variance for the
53 // purposes of activity masking.
54 // Eventually this should be replaced by custom no-reference routines,
55 // which will be faster.
56 static const uint8_t VP9_VAR_OFFS[64] = {
57 128, 128, 128, 128, 128, 128, 128, 128,
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
67 #if CONFIG_VP9_HIGHBITDEPTH
68 static const uint16_t VP9_HIGH_VAR_OFFS_8[64] = {
69 128, 128, 128, 128, 128, 128, 128, 128,
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
79 static const uint16_t VP9_HIGH_VAR_OFFS_10[64] = {
80 128*4, 128*4, 128*4, 128*4, 128*4, 128*4, 128*4, 128*4,
81 128*4, 128*4, 128*4, 128*4, 128*4, 128*4, 128*4, 128*4,
82 128*4, 128*4, 128*4, 128*4, 128*4, 128*4, 128*4, 128*4,
83 128*4, 128*4, 128*4, 128*4, 128*4, 128*4, 128*4, 128*4,
84 128*4, 128*4, 128*4, 128*4, 128*4, 128*4, 128*4, 128*4,
85 128*4, 128*4, 128*4, 128*4, 128*4, 128*4, 128*4, 128*4,
86 128*4, 128*4, 128*4, 128*4, 128*4, 128*4, 128*4, 128*4,
87 128*4, 128*4, 128*4, 128*4, 128*4, 128*4, 128*4, 128*4
90 static const uint16_t VP9_HIGH_VAR_OFFS_12[64] = {
91 128*16, 128*16, 128*16, 128*16, 128*16, 128*16, 128*16, 128*16,
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
100 #endif // CONFIG_VP9_HIGHBITDEPTH
102 unsigned int vp9_get_sby_perpixel_variance(VP9_COMP *cpi,
103 const struct buf_2d *ref,
106 const unsigned int var = cpi->fn_ptr[bs].vf(ref->buf, ref->stride,
107 VP9_VAR_OFFS, 0, &sse);
108 return ROUND_POWER_OF_TWO(var, num_pels_log2_lookup[bs]);
111 #if CONFIG_VP9_HIGHBITDEPTH
112 unsigned int vp9_high_get_sby_perpixel_variance(
113 VP9_COMP *cpi, const struct buf_2d *ref, BLOCK_SIZE bs, int bd) {
114 unsigned int var, sse;
117 var = cpi->fn_ptr[bs].vf(ref->buf, ref->stride,
118 CONVERT_TO_BYTEPTR(VP9_HIGH_VAR_OFFS_10),
122 var = cpi->fn_ptr[bs].vf(ref->buf, ref->stride,
123 CONVERT_TO_BYTEPTR(VP9_HIGH_VAR_OFFS_12),
128 var = cpi->fn_ptr[bs].vf(ref->buf, ref->stride,
129 CONVERT_TO_BYTEPTR(VP9_HIGH_VAR_OFFS_8),
133 return ROUND_POWER_OF_TWO(var, num_pels_log2_lookup[bs]);
135 #endif // CONFIG_VP9_HIGHBITDEPTH
137 static unsigned int get_sby_perpixel_diff_variance(VP9_COMP *cpi,
138 const struct buf_2d *ref,
139 int mi_row, int mi_col,
141 unsigned int sse, var;
143 const YV12_BUFFER_CONFIG *last = get_ref_frame_buffer(cpi, LAST_FRAME);
145 assert(last != NULL);
147 &last->y_buffer[mi_row * MI_SIZE * last->y_stride + mi_col * MI_SIZE];
148 var = cpi->fn_ptr[bs].vf(ref->buf, ref->stride, last_y, last->y_stride, &sse);
149 return ROUND_POWER_OF_TWO(var, num_pels_log2_lookup[bs]);
152 static BLOCK_SIZE get_rd_var_based_fixed_partition(VP9_COMP *cpi, MACROBLOCK *x,
155 unsigned int var = get_sby_perpixel_diff_variance(cpi, &x->plane[0].src,
168 // Lighter version of set_offsets that only sets the mode info
170 static INLINE void set_mode_info_offsets(VP9_COMMON *const cm,
171 MACROBLOCKD *const xd,
174 const int idx_str = xd->mi_stride * mi_row + mi_col;
175 xd->mi = cm->mi + idx_str;
176 xd->mi[0].src_mi = &xd->mi[0];
179 static void set_offsets(VP9_COMP *cpi, const TileInfo *const tile,
180 MACROBLOCK *const x, int mi_row, int mi_col,
182 VP9_COMMON *const cm = &cpi->common;
183 MACROBLOCKD *const xd = &x->e_mbd;
185 const int mi_width = num_8x8_blocks_wide_lookup[bsize];
186 const int mi_height = num_8x8_blocks_high_lookup[bsize];
187 const struct segmentation *const seg = &cm->seg;
189 set_skip_context(xd, mi_row, mi_col);
191 set_mode_info_offsets(cm, xd, mi_row, mi_col);
193 mbmi = &xd->mi[0].src_mi->mbmi;
195 // Set up destination pointers.
196 vp9_setup_dst_planes(xd->plane, get_frame_new_buffer(cm), mi_row, mi_col);
198 // Set up limit values for MV components.
199 // Mv beyond the range do not produce new/different prediction block.
200 x->mv_row_min = -(((mi_row + mi_height) * MI_SIZE) + VP9_INTERP_EXTEND);
201 x->mv_col_min = -(((mi_col + mi_width) * MI_SIZE) + VP9_INTERP_EXTEND);
202 x->mv_row_max = (cm->mi_rows - mi_row) * MI_SIZE + VP9_INTERP_EXTEND;
203 x->mv_col_max = (cm->mi_cols - mi_col) * MI_SIZE + VP9_INTERP_EXTEND;
205 // Set up distance of MB to edge of frame in 1/8th pel units.
206 assert(!(mi_col & (mi_width - 1)) && !(mi_row & (mi_height - 1)));
207 set_mi_row_col(xd, tile, mi_row, mi_height, mi_col, mi_width,
208 cm->mi_rows, cm->mi_cols);
210 // Set up source buffers.
211 vp9_setup_src_planes(x, cpi->Source, mi_row, mi_col);
214 x->rddiv = cpi->rd.RDDIV;
215 x->rdmult = cpi->rd.RDMULT;
219 if (cpi->oxcf.aq_mode != VARIANCE_AQ) {
220 const uint8_t *const map = seg->update_map ? cpi->segmentation_map
221 : cm->last_frame_seg_map;
222 mbmi->segment_id = vp9_get_segment_id(cm, map, bsize, mi_row, mi_col);
224 vp9_init_plane_quantizers(cpi, x);
226 x->encode_breakout = cpi->segment_encode_breakout[mbmi->segment_id];
228 mbmi->segment_id = 0;
229 x->encode_breakout = cpi->encode_breakout;
233 static void duplicate_mode_info_in_sb(VP9_COMMON *cm, MACROBLOCKD *xd,
234 int mi_row, int mi_col,
236 const int block_width = num_8x8_blocks_wide_lookup[bsize];
237 const int block_height = num_8x8_blocks_high_lookup[bsize];
239 for (j = 0; j < block_height; ++j)
240 for (i = 0; i < block_width; ++i) {
241 if (mi_row + j < cm->mi_rows && mi_col + i < cm->mi_cols)
242 xd->mi[j * xd->mi_stride + i].src_mi = &xd->mi[0];
246 static void set_block_size(VP9_COMP * const cpi,
247 MACROBLOCKD *const xd,
248 int mi_row, int mi_col,
250 if (cpi->common.mi_cols > mi_col && cpi->common.mi_rows > mi_row) {
251 set_mode_info_offsets(&cpi->common, xd, mi_row, mi_col);
252 xd->mi[0].src_mi->mbmi.sb_type = bsize;
257 int64_t sum_square_error;
267 } partition_variance;
270 partition_variance part_variances;
275 partition_variance part_variances;
280 partition_variance part_variances;
285 partition_variance part_variances;
290 partition_variance part_variances;
295 partition_variance *part_variances;
305 static void tree_to_node(void *data, BLOCK_SIZE bsize, variance_node *node) {
307 node->part_variances = NULL;
310 v64x64 *vt = (v64x64 *) data;
311 node->part_variances = &vt->part_variances;
312 for (i = 0; i < 4; i++)
313 node->split[i] = &vt->split[i].part_variances.none;
317 v32x32 *vt = (v32x32 *) data;
318 node->part_variances = &vt->part_variances;
319 for (i = 0; i < 4; i++)
320 node->split[i] = &vt->split[i].part_variances.none;
324 v16x16 *vt = (v16x16 *) data;
325 node->part_variances = &vt->part_variances;
326 for (i = 0; i < 4; i++)
327 node->split[i] = &vt->split[i].part_variances.none;
331 v8x8 *vt = (v8x8 *) data;
332 node->part_variances = &vt->part_variances;
333 for (i = 0; i < 4; i++)
334 node->split[i] = &vt->split[i].part_variances.none;
338 v4x4 *vt = (v4x4 *) data;
339 node->part_variances = &vt->part_variances;
340 for (i = 0; i < 4; i++)
341 node->split[i] = &vt->split[i];
351 // Set variance values given sum square error, sum error, count.
352 static void fill_variance(int64_t s2, int64_t s, int c, var *v) {
353 v->sum_square_error = s2;
358 static void get_variance(var *v) {
359 v->variance = (int)(256 * (v->sum_square_error -
360 ((v->sum_error * v->sum_error) >> v->log2_count)) >> v->log2_count);
363 void sum_2_variances(const var *a, const var *b, var *r) {
364 assert(a->log2_count == b->log2_count);
365 fill_variance(a->sum_square_error + b->sum_square_error,
366 a->sum_error + b->sum_error, a->log2_count + 1, r);
369 static void fill_variance_tree(void *data, BLOCK_SIZE bsize) {
371 tree_to_node(data, bsize, &node);
372 sum_2_variances(node.split[0], node.split[1], &node.part_variances->horz[0]);
373 sum_2_variances(node.split[2], node.split[3], &node.part_variances->horz[1]);
374 sum_2_variances(node.split[0], node.split[2], &node.part_variances->vert[0]);
375 sum_2_variances(node.split[1], node.split[3], &node.part_variances->vert[1]);
376 sum_2_variances(&node.part_variances->vert[0], &node.part_variances->vert[1],
377 &node.part_variances->none);
380 static int set_vt_partitioning(VP9_COMP *cpi,
381 MACROBLOCKD *const xd,
387 BLOCK_SIZE bsize_min,
389 VP9_COMMON * const cm = &cpi->common;
391 const int block_width = num_8x8_blocks_wide_lookup[bsize];
392 const int block_height = num_8x8_blocks_high_lookup[bsize];
393 const int low_res = (cm->width <= 352 && cm->height <= 288);
395 assert(block_height == block_width);
396 tree_to_node(data, bsize, &vt);
398 if (force_split == 1)
401 // For bsize=bsize_min (16x16/8x8 for 8x8/4x4 downsampling), select if
402 // variance is below threshold, otherwise split will be selected.
403 // No check for vert/horiz split as too few samples for variance.
404 if (bsize == bsize_min) {
405 // Variance already computed to set the force_split.
406 if (low_res || cm->frame_type == KEY_FRAME)
407 get_variance(&vt.part_variances->none);
408 if (mi_col + block_width / 2 < cm->mi_cols &&
409 mi_row + block_height / 2 < cm->mi_rows &&
410 vt.part_variances->none.variance < threshold) {
411 set_block_size(cpi, xd, mi_row, mi_col, bsize);
415 } else if (bsize > bsize_min) {
416 // Variance already computed to set the force_split.
417 if (low_res || cm->frame_type == KEY_FRAME)
418 get_variance(&vt.part_variances->none);
419 // For key frame: take split for bsize above 32X32 or very high variance.
420 if (cm->frame_type == KEY_FRAME &&
421 (bsize > BLOCK_32X32 ||
422 vt.part_variances->none.variance > (threshold << 4))) {
425 // If variance is low, take the bsize (no split).
426 if (mi_col + block_width / 2 < cm->mi_cols &&
427 mi_row + block_height / 2 < cm->mi_rows &&
428 vt.part_variances->none.variance < threshold) {
429 set_block_size(cpi, xd, mi_row, mi_col, bsize);
433 // Check vertical split.
434 if (mi_row + block_height / 2 < cm->mi_rows) {
435 get_variance(&vt.part_variances->vert[0]);
436 get_variance(&vt.part_variances->vert[1]);
437 if (vt.part_variances->vert[0].variance < threshold &&
438 vt.part_variances->vert[1].variance < threshold) {
439 BLOCK_SIZE subsize = get_subsize(bsize, PARTITION_VERT);
440 set_block_size(cpi, xd, mi_row, mi_col, subsize);
441 set_block_size(cpi, xd, mi_row, mi_col + block_width / 2, subsize);
445 // Check horizontal split.
446 if (mi_col + block_width / 2 < cm->mi_cols) {
447 get_variance(&vt.part_variances->horz[0]);
448 get_variance(&vt.part_variances->horz[1]);
449 if (vt.part_variances->horz[0].variance < threshold &&
450 vt.part_variances->horz[1].variance < threshold) {
451 BLOCK_SIZE subsize = get_subsize(bsize, PARTITION_HORZ);
452 set_block_size(cpi, xd, mi_row, mi_col, subsize);
453 set_block_size(cpi, xd, mi_row + block_height / 2, mi_col, subsize);
463 void vp9_set_vbp_thresholds(VP9_COMP *cpi, int q) {
464 SPEED_FEATURES *const sf = &cpi->sf;
465 if (sf->partition_search_type != VAR_BASED_PARTITION &&
466 sf->partition_search_type != REFERENCE_PARTITION) {
469 VP9_COMMON *const cm = &cpi->common;
470 const int is_key_frame = (cm->frame_type == KEY_FRAME);
471 const int threshold_multiplier = is_key_frame ? 20 : 1;
472 const int64_t threshold_base = (int64_t)(threshold_multiplier *
473 cpi->y_dequant[q][1]);
475 // TODO(marpan): Allow 4x4 partitions for inter-frames.
476 // use_4x4_partition = (variance4x4downsample[i2 + j] == 1);
477 // If 4x4 partition is not used, then 8x8 partition will be selected
478 // if variance of 16x16 block is very high, so use larger threshold
479 // for 16x16 (threshold_bsize_min) in that case.
481 // Array index: 0 - threshold_64x64; 1 - threshold_32x32;
482 // 2 - threshold_16x16; 3 - vbp_threshold_8x8;
484 cpi->vbp_thresholds[0] = threshold_base;
485 cpi->vbp_thresholds[1] = threshold_base >> 2;
486 cpi->vbp_thresholds[2] = threshold_base >> 2;
487 cpi->vbp_thresholds[3] = threshold_base << 2;
488 cpi->vbp_threshold_sad = 0;
489 cpi->vbp_bsize_min = BLOCK_8X8;
491 cpi->vbp_thresholds[1] = threshold_base;
492 if (cm->width <= 352 && cm->height <= 288) {
493 cpi->vbp_thresholds[0] = threshold_base >> 2;
494 cpi->vbp_thresholds[2] = threshold_base << 3;
495 cpi->vbp_threshold_sad = 100;
497 cpi->vbp_thresholds[0] = threshold_base;
498 cpi->vbp_thresholds[1] = (5 * threshold_base) >> 2;
499 cpi->vbp_thresholds[2] = threshold_base << cpi->oxcf.speed;
500 cpi->vbp_threshold_sad = 1000;
502 cpi->vbp_bsize_min = BLOCK_16X16;
504 cpi->vbp_threshold_minmax = 15 + (q >> 3);
508 // Compute the minmax over the 8x8 subblocks.
509 static int compute_minmax_8x8(const uint8_t *s, int sp, const uint8_t *d,
510 int dp, int x16_idx, int y16_idx,
511 #if CONFIG_VP9_HIGHBITDEPTH
518 int minmax_min = 255;
519 // Loop over the 4 8x8 subblocks.
520 for (k = 0; k < 4; k++) {
521 int x8_idx = x16_idx + ((k & 1) << 3);
522 int y8_idx = y16_idx + ((k >> 1) << 3);
525 if (x8_idx < pixels_wide && y8_idx < pixels_high) {
526 #if CONFIG_VP9_HIGHBITDEPTH
527 if (highbd_flag & YV12_FLAG_HIGHBITDEPTH) {
528 vp9_highbd_minmax_8x8(s + y8_idx * sp + x8_idx, sp,
529 d + y8_idx * dp + x8_idx, dp,
532 vp9_minmax_8x8(s + y8_idx * sp + x8_idx, sp,
533 d + y8_idx * dp + x8_idx, dp,
537 vp9_minmax_8x8(s + y8_idx * sp + x8_idx, sp,
538 d + y8_idx * dp + x8_idx, dp,
541 if ((max - min) > minmax_max)
542 minmax_max = (max - min);
543 if ((max - min) < minmax_min)
544 minmax_min = (max - min);
547 return (minmax_max - minmax_min);
550 static void modify_vbp_thresholds(VP9_COMP *cpi, int64_t thresholds[], int q) {
551 VP9_COMMON *const cm = &cpi->common;
552 const int64_t threshold_base = (int64_t)(cpi->y_dequant[q][1]);
554 // Array index: 0 - threshold_64x64; 1 - threshold_32x32;
555 // 2 - threshold_16x16; 3 - vbp_threshold_8x8;
556 thresholds[1] = threshold_base;
557 if (cm->width <= 352 && cm->height <= 288) {
558 thresholds[0] = threshold_base >> 2;
559 thresholds[2] = threshold_base << 3;
561 thresholds[0] = threshold_base;
562 thresholds[1] = (5 * threshold_base) >> 2;
563 thresholds[2] = threshold_base << cpi->oxcf.speed;
567 static void fill_variance_4x4avg(const uint8_t *s, int sp, const uint8_t *d,
568 int dp, int x8_idx, int y8_idx, v8x8 *vst,
569 #if CONFIG_VP9_HIGHBITDEPTH
576 for (k = 0; k < 4; k++) {
577 int x4_idx = x8_idx + ((k & 1) << 2);
578 int y4_idx = y8_idx + ((k >> 1) << 2);
579 unsigned int sse = 0;
581 if (x4_idx < pixels_wide && y4_idx < pixels_high) {
584 #if CONFIG_VP9_HIGHBITDEPTH
585 if (highbd_flag & YV12_FLAG_HIGHBITDEPTH) {
586 s_avg = vp9_highbd_avg_4x4(s + y4_idx * sp + x4_idx, sp);
588 d_avg = vp9_highbd_avg_4x4(d + y4_idx * dp + x4_idx, dp);
590 s_avg = vp9_avg_4x4(s + y4_idx * sp + x4_idx, sp);
592 d_avg = vp9_avg_4x4(d + y4_idx * dp + x4_idx, dp);
595 s_avg = vp9_avg_4x4(s + y4_idx * sp + x4_idx, sp);
597 d_avg = vp9_avg_4x4(d + y4_idx * dp + x4_idx, dp);
602 fill_variance(sse, sum, 0, &vst->split[k].part_variances.none);
606 static void fill_variance_8x8avg(const uint8_t *s, int sp, const uint8_t *d,
607 int dp, int x16_idx, int y16_idx, v16x16 *vst,
608 #if CONFIG_VP9_HIGHBITDEPTH
615 for (k = 0; k < 4; k++) {
616 int x8_idx = x16_idx + ((k & 1) << 3);
617 int y8_idx = y16_idx + ((k >> 1) << 3);
618 unsigned int sse = 0;
620 if (x8_idx < pixels_wide && y8_idx < pixels_high) {
623 #if CONFIG_VP9_HIGHBITDEPTH
624 if (highbd_flag & YV12_FLAG_HIGHBITDEPTH) {
625 s_avg = vp9_highbd_avg_8x8(s + y8_idx * sp + x8_idx, sp);
627 d_avg = vp9_highbd_avg_8x8(d + y8_idx * dp + x8_idx, dp);
629 s_avg = vp9_avg_8x8(s + y8_idx * sp + x8_idx, sp);
631 d_avg = vp9_avg_8x8(d + y8_idx * dp + x8_idx, dp);
634 s_avg = vp9_avg_8x8(s + y8_idx * sp + x8_idx, sp);
636 d_avg = vp9_avg_8x8(d + y8_idx * dp + x8_idx, dp);
641 fill_variance(sse, sum, 0, &vst->split[k].part_variances.none);
645 // This function chooses partitioning based on the variance between source and
646 // reconstructed last, where variance is computed for down-sampled inputs.
647 static int choose_partitioning(VP9_COMP *cpi,
648 const TileInfo *const tile,
650 int mi_row, int mi_col) {
651 VP9_COMMON * const cm = &cpi->common;
652 MACROBLOCKD *xd = &x->e_mbd;
661 int pixels_wide = 64, pixels_high = 64;
662 int64_t thresholds[4] = {cpi->vbp_thresholds[0], cpi->vbp_thresholds[1],
663 cpi->vbp_thresholds[2], cpi->vbp_thresholds[3]};
665 // Always use 4x4 partition for key frame.
666 const int is_key_frame = (cm->frame_type == KEY_FRAME);
667 const int use_4x4_partition = is_key_frame;
668 const int low_res = (cm->width <= 352 && cm->height <= 288);
669 int variance4x4downsample[16];
671 int segment_id = CR_SEGMENT_ID_BASE;
672 if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ && cm->seg.enabled) {
673 const uint8_t *const map = cm->seg.update_map ? cpi->segmentation_map :
674 cm->last_frame_seg_map;
675 segment_id = vp9_get_segment_id(cm, map, BLOCK_64X64, mi_row, mi_col);
677 if (cyclic_refresh_segment_id_boosted(segment_id)) {
678 int q = vp9_get_qindex(&cm->seg, segment_id, cm->base_qindex);
679 modify_vbp_thresholds(cpi, thresholds, q);
683 set_offsets(cpi, tile, x, mi_row, mi_col, BLOCK_64X64);
685 if (xd->mb_to_right_edge < 0)
686 pixels_wide += (xd->mb_to_right_edge >> 3);
687 if (xd->mb_to_bottom_edge < 0)
688 pixels_high += (xd->mb_to_bottom_edge >> 3);
690 s = x->plane[0].src.buf;
691 sp = x->plane[0].src.stride;
694 MB_MODE_INFO *mbmi = &xd->mi[0].src_mi->mbmi;
696 const YV12_BUFFER_CONFIG *yv12 = get_ref_frame_buffer(cpi, LAST_FRAME);
698 const YV12_BUFFER_CONFIG *yv12_g = get_ref_frame_buffer(cpi, GOLDEN_FRAME);
699 unsigned int y_sad, y_sad_g;
700 const BLOCK_SIZE bsize = BLOCK_32X32
701 + (mi_col + 4 < cm->mi_cols) * 2 + (mi_row + 4 < cm->mi_rows);
703 assert(yv12 != NULL);
704 if (yv12_g && yv12_g != yv12) {
705 vp9_setup_pre_planes(xd, 0, yv12_g, mi_row, mi_col,
706 &cm->frame_refs[GOLDEN_FRAME - 1].sf);
707 y_sad_g = cpi->fn_ptr[bsize].sdf(x->plane[0].src.buf,
708 x->plane[0].src.stride,
709 xd->plane[0].pre[0].buf,
710 xd->plane[0].pre[0].stride);
715 vp9_setup_pre_planes(xd, 0, yv12, mi_row, mi_col,
716 &cm->frame_refs[LAST_FRAME - 1].sf);
717 mbmi->ref_frame[0] = LAST_FRAME;
718 mbmi->ref_frame[1] = NONE;
719 mbmi->sb_type = BLOCK_64X64;
720 mbmi->mv[0].as_int = 0;
721 mbmi->interp_filter = BILINEAR;
723 y_sad = vp9_int_pro_motion_estimation(cpi, x, bsize);
724 if (y_sad_g < y_sad) {
725 vp9_setup_pre_planes(xd, 0, yv12_g, mi_row, mi_col,
726 &cm->frame_refs[GOLDEN_FRAME - 1].sf);
727 mbmi->ref_frame[0] = GOLDEN_FRAME;
728 mbmi->mv[0].as_int = 0;
731 x->pred_mv[LAST_FRAME] = mbmi->mv[0].as_mv;
734 vp9_build_inter_predictors_sb(xd, mi_row, mi_col, BLOCK_64X64);
736 for (i = 1; i <= 2; ++i) {
737 struct macroblock_plane *p = &x->plane[i];
738 struct macroblockd_plane *pd = &xd->plane[i];
739 const BLOCK_SIZE bs = get_plane_block_size(bsize, pd);
741 if (bs == BLOCK_INVALID)
744 uv_sad = cpi->fn_ptr[bs].sdf(p->src.buf, p->src.stride,
745 pd->dst.buf, pd->dst.stride);
747 x->color_sensitivity[i - 1] = uv_sad > (y_sad >> 2);
750 d = xd->plane[0].dst.buf;
751 dp = xd->plane[0].dst.stride;
753 // If the y_sad is very small, take 64x64 as partition and exit.
754 // Don't check on boosted segment for now, as 64x64 is suppressed there.
755 if (segment_id == CR_SEGMENT_ID_BASE &&
756 y_sad < cpi->vbp_threshold_sad) {
757 const int block_width = num_8x8_blocks_wide_lookup[BLOCK_64X64];
758 const int block_height = num_8x8_blocks_high_lookup[BLOCK_64X64];
759 if (mi_col + block_width / 2 < cm->mi_cols &&
760 mi_row + block_height / 2 < cm->mi_rows) {
761 set_block_size(cpi, xd, mi_row, mi_col, BLOCK_64X64);
768 #if CONFIG_VP9_HIGHBITDEPTH
769 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
772 d = CONVERT_TO_BYTEPTR(VP9_HIGH_VAR_OFFS_10);
775 d = CONVERT_TO_BYTEPTR(VP9_HIGH_VAR_OFFS_12);
779 d = CONVERT_TO_BYTEPTR(VP9_HIGH_VAR_OFFS_8);
783 #endif // CONFIG_VP9_HIGHBITDEPTH
786 // Index for force_split: 0 for 64x64, 1-4 for 32x32 blocks,
787 // 5-20 for the 16x16 blocks.
789 // Fill in the entire tree of 8x8 (or 4x4 under some conditions) variances
791 for (i = 0; i < 4; i++) {
792 const int x32_idx = ((i & 1) << 5);
793 const int y32_idx = ((i >> 1) << 5);
794 const int i2 = i << 2;
795 force_split[i + 1] = 0;
796 for (j = 0; j < 4; j++) {
797 const int x16_idx = x32_idx + ((j & 1) << 4);
798 const int y16_idx = y32_idx + ((j >> 1) << 4);
799 const int split_index = 5 + i2 + j;
800 v16x16 *vst = &vt.split[i].split[j];
801 force_split[split_index] = 0;
802 variance4x4downsample[i2 + j] = 0;
804 fill_variance_8x8avg(s, sp, d, dp, x16_idx, y16_idx, vst,
805 #if CONFIG_VP9_HIGHBITDEPTH
811 fill_variance_tree(&vt.split[i].split[j], BLOCK_16X16);
812 get_variance(&vt.split[i].split[j].part_variances.none);
813 if (vt.split[i].split[j].part_variances.none.variance >
815 // 16X16 variance is above threshold for split, so force split to 8x8
816 // for this 16x16 block (this also forces splits for upper levels).
817 force_split[split_index] = 1;
818 force_split[i + 1] = 1;
820 } else if (vt.split[i].split[j].part_variances.none.variance >
822 !cyclic_refresh_segment_id_boosted(segment_id)) {
823 // We have some nominal amount of 16x16 variance (based on average),
824 // compute the minmax over the 8x8 sub-blocks, and if above threshold,
825 // force split to 8x8 block for this 16x16 block.
826 int minmax = compute_minmax_8x8(s, sp, d, dp, x16_idx, y16_idx,
827 #if CONFIG_VP9_HIGHBITDEPTH
830 pixels_wide, pixels_high);
831 if (minmax > cpi->vbp_threshold_minmax) {
832 force_split[split_index] = 1;
833 force_split[i + 1] = 1;
838 if (is_key_frame || (low_res &&
839 vt.split[i].split[j].part_variances.none.variance >
840 (thresholds[1] << 1))) {
841 force_split[split_index] = 0;
842 // Go down to 4x4 down-sampling for variance.
843 variance4x4downsample[i2 + j] = 1;
844 for (k = 0; k < 4; k++) {
845 int x8_idx = x16_idx + ((k & 1) << 3);
846 int y8_idx = y16_idx + ((k >> 1) << 3);
847 v8x8 *vst2 = is_key_frame ? &vst->split[k] :
848 &vt2[i2 + j].split[k];
849 fill_variance_4x4avg(s, sp, d, dp, x8_idx, y8_idx, vst2,
850 #if CONFIG_VP9_HIGHBITDEPTH
861 // Fill the rest of the variance tree by summing split partition values.
862 for (i = 0; i < 4; i++) {
863 const int i2 = i << 2;
864 for (j = 0; j < 4; j++) {
865 if (variance4x4downsample[i2 + j] == 1) {
866 v16x16 *vtemp = (!is_key_frame) ? &vt2[i2 + j] :
867 &vt.split[i].split[j];
868 for (m = 0; m < 4; m++)
869 fill_variance_tree(&vtemp->split[m], BLOCK_8X8);
870 fill_variance_tree(vtemp, BLOCK_16X16);
873 fill_variance_tree(&vt.split[i], BLOCK_32X32);
874 // If variance of this 32x32 block is above the threshold, force the block
875 // to split. This also forces a split on the upper (64x64) level.
876 if (!force_split[i + 1]) {
877 get_variance(&vt.split[i].part_variances.none);
878 if (vt.split[i].part_variances.none.variance > thresholds[1]) {
879 force_split[i + 1] = 1;
884 if (!force_split[0]) {
885 fill_variance_tree(&vt, BLOCK_64X64);
886 get_variance(&vt.part_variances.none);
889 // Now go through the entire structure, splitting every block size until
890 // we get to one that's got a variance lower than our threshold.
891 if ( mi_col + 8 > cm->mi_cols || mi_row + 8 > cm->mi_rows ||
892 !set_vt_partitioning(cpi, xd, &vt, BLOCK_64X64, mi_row, mi_col,
893 thresholds[0], BLOCK_16X16, force_split[0])) {
894 for (i = 0; i < 4; ++i) {
895 const int x32_idx = ((i & 1) << 2);
896 const int y32_idx = ((i >> 1) << 2);
897 const int i2 = i << 2;
898 if (!set_vt_partitioning(cpi, xd, &vt.split[i], BLOCK_32X32,
899 (mi_row + y32_idx), (mi_col + x32_idx),
900 thresholds[1], BLOCK_16X16,
901 force_split[i + 1])) {
902 for (j = 0; j < 4; ++j) {
903 const int x16_idx = ((j & 1) << 1);
904 const int y16_idx = ((j >> 1) << 1);
905 // For inter frames: if variance4x4downsample[] == 1 for this 16x16
906 // block, then the variance is based on 4x4 down-sampling, so use vt2
907 // in set_vt_partioning(), otherwise use vt.
908 v16x16 *vtemp = (!is_key_frame &&
909 variance4x4downsample[i2 + j] == 1) ?
910 &vt2[i2 + j] : &vt.split[i].split[j];
911 if (!set_vt_partitioning(cpi, xd, vtemp, BLOCK_16X16,
912 mi_row + y32_idx + y16_idx,
913 mi_col + x32_idx + x16_idx,
916 force_split[5 + i2 + j])) {
917 for (k = 0; k < 4; ++k) {
918 const int x8_idx = (k & 1);
919 const int y8_idx = (k >> 1);
920 if (use_4x4_partition) {
921 if (!set_vt_partitioning(cpi, xd, &vtemp->split[k],
923 mi_row + y32_idx + y16_idx + y8_idx,
924 mi_col + x32_idx + x16_idx + x8_idx,
925 thresholds[3], BLOCK_8X8, 0)) {
926 set_block_size(cpi, xd,
927 (mi_row + y32_idx + y16_idx + y8_idx),
928 (mi_col + x32_idx + x16_idx + x8_idx),
932 set_block_size(cpi, xd,
933 (mi_row + y32_idx + y16_idx + y8_idx),
934 (mi_col + x32_idx + x16_idx + x8_idx),
946 static void update_state(VP9_COMP *cpi, ThreadData *td,
947 PICK_MODE_CONTEXT *ctx,
948 int mi_row, int mi_col, BLOCK_SIZE bsize,
949 int output_enabled) {
951 VP9_COMMON *const cm = &cpi->common;
952 RD_COUNTS *const rdc = &td->rd_counts;
953 MACROBLOCK *const x = &td->mb;
954 MACROBLOCKD *const xd = &x->e_mbd;
955 struct macroblock_plane *const p = x->plane;
956 struct macroblockd_plane *const pd = xd->plane;
957 MODE_INFO *mi = &ctx->mic;
958 MB_MODE_INFO *const mbmi = &xd->mi[0].src_mi->mbmi;
959 MODE_INFO *mi_addr = &xd->mi[0];
960 const struct segmentation *const seg = &cm->seg;
961 const int bw = num_8x8_blocks_wide_lookup[mi->mbmi.sb_type];
962 const int bh = num_8x8_blocks_high_lookup[mi->mbmi.sb_type];
963 const int x_mis = MIN(bw, cm->mi_cols - mi_col);
964 const int y_mis = MIN(bh, cm->mi_rows - mi_row);
965 MV_REF *const frame_mvs =
966 cm->cur_frame->mvs + mi_row * cm->mi_cols + mi_col;
969 const int mis = cm->mi_stride;
970 const int mi_width = num_8x8_blocks_wide_lookup[bsize];
971 const int mi_height = num_8x8_blocks_high_lookup[bsize];
974 assert(mi->mbmi.sb_type == bsize);
977 mi_addr->src_mi = mi_addr;
979 // If segmentation in use
981 // For in frame complexity AQ copy the segment id from the segment map.
982 if (cpi->oxcf.aq_mode == COMPLEXITY_AQ) {
983 const uint8_t *const map = seg->update_map ? cpi->segmentation_map
984 : cm->last_frame_seg_map;
985 mi_addr->mbmi.segment_id =
986 vp9_get_segment_id(cm, map, bsize, mi_row, mi_col);
988 // Else for cyclic refresh mode update the segment map, set the segment id
989 // and then update the quantizer.
990 if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ) {
991 vp9_cyclic_refresh_update_segment(cpi, &xd->mi[0].src_mi->mbmi, mi_row,
992 mi_col, bsize, ctx->rate, ctx->dist,
997 max_plane = is_inter_block(mbmi) ? MAX_MB_PLANE : 1;
998 for (i = 0; i < max_plane; ++i) {
999 p[i].coeff = ctx->coeff_pbuf[i][1];
1000 p[i].qcoeff = ctx->qcoeff_pbuf[i][1];
1001 pd[i].dqcoeff = ctx->dqcoeff_pbuf[i][1];
1002 p[i].eobs = ctx->eobs_pbuf[i][1];
1005 for (i = max_plane; i < MAX_MB_PLANE; ++i) {
1006 p[i].coeff = ctx->coeff_pbuf[i][2];
1007 p[i].qcoeff = ctx->qcoeff_pbuf[i][2];
1008 pd[i].dqcoeff = ctx->dqcoeff_pbuf[i][2];
1009 p[i].eobs = ctx->eobs_pbuf[i][2];
1012 // Restore the coding context of the MB to that that was in place
1013 // when the mode was picked for it
1014 for (y = 0; y < mi_height; y++)
1015 for (x_idx = 0; x_idx < mi_width; x_idx++)
1016 if ((xd->mb_to_right_edge >> (3 + MI_SIZE_LOG2)) + mi_width > x_idx
1017 && (xd->mb_to_bottom_edge >> (3 + MI_SIZE_LOG2)) + mi_height > y) {
1018 xd->mi[x_idx + y * mis].src_mi = mi_addr;
1021 if (cpi->oxcf.aq_mode)
1022 vp9_init_plane_quantizers(cpi, x);
1024 // FIXME(rbultje) I'm pretty sure this should go to the end of this block
1025 // (i.e. after the output_enabled)
1026 if (bsize < BLOCK_32X32) {
1027 if (bsize < BLOCK_16X16)
1028 ctx->tx_rd_diff[ALLOW_16X16] = ctx->tx_rd_diff[ALLOW_8X8];
1029 ctx->tx_rd_diff[ALLOW_32X32] = ctx->tx_rd_diff[ALLOW_16X16];
1032 if (is_inter_block(mbmi) && mbmi->sb_type < BLOCK_8X8) {
1033 mbmi->mv[0].as_int = mi->bmi[3].as_mv[0].as_int;
1034 mbmi->mv[1].as_int = mi->bmi[3].as_mv[1].as_int;
1037 x->skip = ctx->skip;
1038 vpx_memcpy(x->zcoeff_blk[mbmi->tx_size], ctx->zcoeff_blk,
1039 sizeof(uint8_t) * ctx->num_4x4_blk);
1041 if (!output_enabled)
1044 if (!vp9_segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_SKIP)) {
1045 for (i = 0; i < TX_MODES; i++)
1046 rdc->tx_select_diff[i] += ctx->tx_rd_diff[i];
1049 #if CONFIG_INTERNAL_STATS
1050 if (frame_is_intra_only(cm)) {
1051 static const int kf_mode_index[] = {
1053 THR_V_PRED /*V_PRED*/,
1054 THR_H_PRED /*H_PRED*/,
1055 THR_D45_PRED /*D45_PRED*/,
1056 THR_D135_PRED /*D135_PRED*/,
1057 THR_D117_PRED /*D117_PRED*/,
1058 THR_D153_PRED /*D153_PRED*/,
1059 THR_D207_PRED /*D207_PRED*/,
1060 THR_D63_PRED /*D63_PRED*/,
1063 ++cpi->mode_chosen_counts[kf_mode_index[mbmi->mode]];
1065 // Note how often each mode chosen as best
1066 ++cpi->mode_chosen_counts[ctx->best_mode_index];
1069 if (!frame_is_intra_only(cm)) {
1070 if (is_inter_block(mbmi)) {
1071 vp9_update_mv_count(td);
1073 if (cm->interp_filter == SWITCHABLE) {
1074 const int ctx = vp9_get_pred_context_switchable_interp(xd);
1075 ++td->counts->switchable_interp[ctx][mbmi->interp_filter];
1079 rdc->comp_pred_diff[SINGLE_REFERENCE] += ctx->single_pred_diff;
1080 rdc->comp_pred_diff[COMPOUND_REFERENCE] += ctx->comp_pred_diff;
1081 rdc->comp_pred_diff[REFERENCE_MODE_SELECT] += ctx->hybrid_pred_diff;
1083 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i)
1084 rdc->filter_diff[i] += ctx->best_filter_diff[i];
1087 for (h = 0; h < y_mis; ++h) {
1088 MV_REF *const frame_mv = frame_mvs + h * cm->mi_cols;
1089 for (w = 0; w < x_mis; ++w) {
1090 MV_REF *const mv = frame_mv + w;
1091 mv->ref_frame[0] = mi->src_mi->mbmi.ref_frame[0];
1092 mv->ref_frame[1] = mi->src_mi->mbmi.ref_frame[1];
1093 mv->mv[0].as_int = mi->src_mi->mbmi.mv[0].as_int;
1094 mv->mv[1].as_int = mi->src_mi->mbmi.mv[1].as_int;
1099 void vp9_setup_src_planes(MACROBLOCK *x, const YV12_BUFFER_CONFIG *src,
1100 int mi_row, int mi_col) {
1101 uint8_t *const buffers[3] = {src->y_buffer, src->u_buffer, src->v_buffer };
1102 const int strides[3] = {src->y_stride, src->uv_stride, src->uv_stride };
1105 // Set current frame pointer.
1106 x->e_mbd.cur_buf = src;
1108 for (i = 0; i < MAX_MB_PLANE; i++)
1109 setup_pred_plane(&x->plane[i].src, buffers[i], strides[i], mi_row, mi_col,
1110 NULL, x->e_mbd.plane[i].subsampling_x,
1111 x->e_mbd.plane[i].subsampling_y);
1114 static void set_mode_info_seg_skip(MACROBLOCK *x, TX_MODE tx_mode,
1115 RD_COST *rd_cost, BLOCK_SIZE bsize) {
1116 MACROBLOCKD *const xd = &x->e_mbd;
1117 MB_MODE_INFO *const mbmi = &xd->mi[0].src_mi->mbmi;
1118 INTERP_FILTER filter_ref;
1120 if (xd->up_available)
1121 filter_ref = xd->mi[-xd->mi_stride].src_mi->mbmi.interp_filter;
1122 else if (xd->left_available)
1123 filter_ref = xd->mi[-1].src_mi->mbmi.interp_filter;
1125 filter_ref = EIGHTTAP;
1127 mbmi->sb_type = bsize;
1128 mbmi->mode = ZEROMV;
1129 mbmi->tx_size = MIN(max_txsize_lookup[bsize],
1130 tx_mode_to_biggest_tx_size[tx_mode]);
1132 mbmi->uv_mode = DC_PRED;
1133 mbmi->ref_frame[0] = LAST_FRAME;
1134 mbmi->ref_frame[1] = NONE;
1135 mbmi->mv[0].as_int = 0;
1136 mbmi->interp_filter = filter_ref;
1138 xd->mi[0].src_mi->bmi[0].as_mv[0].as_int = 0;
1141 vp9_rd_cost_init(rd_cost);
1144 static int set_segment_rdmult(VP9_COMP *const cpi,
1145 MACROBLOCK *const x,
1146 int8_t segment_id) {
1148 VP9_COMMON *const cm = &cpi->common;
1149 vp9_init_plane_quantizers(cpi, x);
1150 vp9_clear_system_state();
1151 segment_qindex = vp9_get_qindex(&cm->seg, segment_id,
1153 return vp9_compute_rd_mult(cpi, segment_qindex + cm->y_dc_delta_q);
1156 static void rd_pick_sb_modes(VP9_COMP *cpi,
1157 TileDataEnc *tile_data,
1158 MACROBLOCK *const x,
1159 int mi_row, int mi_col, RD_COST *rd_cost,
1160 BLOCK_SIZE bsize, PICK_MODE_CONTEXT *ctx,
1162 VP9_COMMON *const cm = &cpi->common;
1163 TileInfo *const tile_info = &tile_data->tile_info;
1164 MACROBLOCKD *const xd = &x->e_mbd;
1166 struct macroblock_plane *const p = x->plane;
1167 struct macroblockd_plane *const pd = xd->plane;
1168 const AQ_MODE aq_mode = cpi->oxcf.aq_mode;
1171 vp9_clear_system_state();
1173 // Use the lower precision, but faster, 32x32 fdct for mode selection.
1174 x->use_lp32x32fdct = 1;
1176 set_offsets(cpi, tile_info, x, mi_row, mi_col, bsize);
1177 mbmi = &xd->mi[0].src_mi->mbmi;
1178 mbmi->sb_type = bsize;
1180 for (i = 0; i < MAX_MB_PLANE; ++i) {
1181 p[i].coeff = ctx->coeff_pbuf[i][0];
1182 p[i].qcoeff = ctx->qcoeff_pbuf[i][0];
1183 pd[i].dqcoeff = ctx->dqcoeff_pbuf[i][0];
1184 p[i].eobs = ctx->eobs_pbuf[i][0];
1188 ctx->pred_pixel_ready = 0;
1191 // Set to zero to make sure we do not use the previous encoded frame stats
1194 #if CONFIG_VP9_HIGHBITDEPTH
1195 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
1196 x->source_variance =
1197 vp9_high_get_sby_perpixel_variance(cpi, &x->plane[0].src,
1200 x->source_variance =
1201 vp9_get_sby_perpixel_variance(cpi, &x->plane[0].src, bsize);
1204 x->source_variance =
1205 vp9_get_sby_perpixel_variance(cpi, &x->plane[0].src, bsize);
1206 #endif // CONFIG_VP9_HIGHBITDEPTH
1208 // Save rdmult before it might be changed, so it can be restored later.
1209 orig_rdmult = x->rdmult;
1211 if (aq_mode == VARIANCE_AQ) {
1212 const int energy = bsize <= BLOCK_16X16 ? x->mb_energy
1213 : vp9_block_energy(cpi, x, bsize);
1214 if (cm->frame_type == KEY_FRAME ||
1215 cpi->refresh_alt_ref_frame ||
1216 (cpi->refresh_golden_frame && !cpi->rc.is_src_frame_alt_ref)) {
1217 mbmi->segment_id = vp9_vaq_segment_id(energy);
1219 const uint8_t *const map = cm->seg.update_map ? cpi->segmentation_map
1220 : cm->last_frame_seg_map;
1221 mbmi->segment_id = vp9_get_segment_id(cm, map, bsize, mi_row, mi_col);
1223 x->rdmult = set_segment_rdmult(cpi, x, mbmi->segment_id);
1224 } else if (aq_mode == COMPLEXITY_AQ) {
1225 x->rdmult = set_segment_rdmult(cpi, x, mbmi->segment_id);
1226 } else if (aq_mode == CYCLIC_REFRESH_AQ) {
1227 const uint8_t *const map = cm->seg.update_map ? cpi->segmentation_map
1228 : cm->last_frame_seg_map;
1229 // If segment is boosted, use rdmult for that segment.
1230 if (cyclic_refresh_segment_id_boosted(
1231 vp9_get_segment_id(cm, map, bsize, mi_row, mi_col)))
1232 x->rdmult = vp9_cyclic_refresh_get_rdmult(cpi->cyclic_refresh);
1235 // Find best coding mode & reconstruct the MB so it is available
1236 // as a predictor for MBs that follow in the SB
1237 if (frame_is_intra_only(cm)) {
1238 vp9_rd_pick_intra_mode_sb(cpi, x, rd_cost, bsize, ctx, best_rd);
1240 if (bsize >= BLOCK_8X8) {
1241 if (vp9_segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_SKIP))
1242 vp9_rd_pick_inter_mode_sb_seg_skip(cpi, tile_data, x, rd_cost, bsize,
1245 vp9_rd_pick_inter_mode_sb(cpi, tile_data, x, mi_row, mi_col,
1246 rd_cost, bsize, ctx, best_rd);
1248 vp9_rd_pick_inter_mode_sub8x8(cpi, tile_data, x, mi_row, mi_col,
1249 rd_cost, bsize, ctx, best_rd);
1254 // Examine the resulting rate and for AQ mode 2 make a segment choice.
1255 if ((rd_cost->rate != INT_MAX) &&
1256 (aq_mode == COMPLEXITY_AQ) && (bsize >= BLOCK_16X16) &&
1257 (cm->frame_type == KEY_FRAME ||
1258 cpi->refresh_alt_ref_frame ||
1259 (cpi->refresh_golden_frame && !cpi->rc.is_src_frame_alt_ref))) {
1260 vp9_caq_select_segment(cpi, x, bsize, mi_row, mi_col, rd_cost->rate);
1263 x->rdmult = orig_rdmult;
1265 // TODO(jingning) The rate-distortion optimization flow needs to be
1266 // refactored to provide proper exit/return handle.
1267 if (rd_cost->rate == INT_MAX)
1268 rd_cost->rdcost = INT64_MAX;
1270 ctx->rate = rd_cost->rate;
1271 ctx->dist = rd_cost->dist;
1274 static void update_stats(VP9_COMMON *cm, ThreadData *td) {
1275 const MACROBLOCK *x = &td->mb;
1276 const MACROBLOCKD *const xd = &x->e_mbd;
1277 const MODE_INFO *const mi = xd->mi[0].src_mi;
1278 const MB_MODE_INFO *const mbmi = &mi->mbmi;
1279 const BLOCK_SIZE bsize = mbmi->sb_type;
1281 if (!frame_is_intra_only(cm)) {
1282 FRAME_COUNTS *const counts = td->counts;
1283 const int inter_block = is_inter_block(mbmi);
1284 const int seg_ref_active = vp9_segfeature_active(&cm->seg, mbmi->segment_id,
1286 if (!seg_ref_active) {
1287 counts->intra_inter[vp9_get_intra_inter_context(xd)][inter_block]++;
1288 // If the segment reference feature is enabled we have only a single
1289 // reference frame allowed for the segment so exclude it from
1290 // the reference frame counts used to work out probabilities.
1292 const MV_REFERENCE_FRAME ref0 = mbmi->ref_frame[0];
1293 if (cm->reference_mode == REFERENCE_MODE_SELECT)
1294 counts->comp_inter[vp9_get_reference_mode_context(cm, xd)]
1295 [has_second_ref(mbmi)]++;
1297 if (has_second_ref(mbmi)) {
1298 counts->comp_ref[vp9_get_pred_context_comp_ref_p(cm, xd)]
1299 [ref0 == GOLDEN_FRAME]++;
1301 counts->single_ref[vp9_get_pred_context_single_ref_p1(xd)][0]
1302 [ref0 != LAST_FRAME]++;
1303 if (ref0 != LAST_FRAME)
1304 counts->single_ref[vp9_get_pred_context_single_ref_p2(xd)][1]
1305 [ref0 != GOLDEN_FRAME]++;
1310 !vp9_segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_SKIP)) {
1311 const int mode_ctx = mbmi->mode_context[mbmi->ref_frame[0]];
1312 if (bsize >= BLOCK_8X8) {
1313 const PREDICTION_MODE mode = mbmi->mode;
1314 ++counts->inter_mode[mode_ctx][INTER_OFFSET(mode)];
1316 const int num_4x4_w = num_4x4_blocks_wide_lookup[bsize];
1317 const int num_4x4_h = num_4x4_blocks_high_lookup[bsize];
1319 for (idy = 0; idy < 2; idy += num_4x4_h) {
1320 for (idx = 0; idx < 2; idx += num_4x4_w) {
1321 const int j = idy * 2 + idx;
1322 const PREDICTION_MODE b_mode = mi->bmi[j].as_mode;
1323 ++counts->inter_mode[mode_ctx][INTER_OFFSET(b_mode)];
1331 static void restore_context(MACROBLOCK *const x, int mi_row, int mi_col,
1332 ENTROPY_CONTEXT a[16 * MAX_MB_PLANE],
1333 ENTROPY_CONTEXT l[16 * MAX_MB_PLANE],
1334 PARTITION_CONTEXT sa[8], PARTITION_CONTEXT sl[8],
1336 MACROBLOCKD *const xd = &x->e_mbd;
1338 const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
1339 const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
1340 int mi_width = num_8x8_blocks_wide_lookup[bsize];
1341 int mi_height = num_8x8_blocks_high_lookup[bsize];
1342 for (p = 0; p < MAX_MB_PLANE; p++) {
1344 xd->above_context[p] + ((mi_col * 2) >> xd->plane[p].subsampling_x),
1345 a + num_4x4_blocks_wide * p,
1346 (sizeof(ENTROPY_CONTEXT) * num_4x4_blocks_wide) >>
1347 xd->plane[p].subsampling_x);
1350 + ((mi_row & MI_MASK) * 2 >> xd->plane[p].subsampling_y),
1351 l + num_4x4_blocks_high * p,
1352 (sizeof(ENTROPY_CONTEXT) * num_4x4_blocks_high) >>
1353 xd->plane[p].subsampling_y);
1355 vpx_memcpy(xd->above_seg_context + mi_col, sa,
1356 sizeof(*xd->above_seg_context) * mi_width);
1357 vpx_memcpy(xd->left_seg_context + (mi_row & MI_MASK), sl,
1358 sizeof(xd->left_seg_context[0]) * mi_height);
1361 static void save_context(MACROBLOCK *const x, int mi_row, int mi_col,
1362 ENTROPY_CONTEXT a[16 * MAX_MB_PLANE],
1363 ENTROPY_CONTEXT l[16 * MAX_MB_PLANE],
1364 PARTITION_CONTEXT sa[8], PARTITION_CONTEXT sl[8],
1366 const MACROBLOCKD *const xd = &x->e_mbd;
1368 const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
1369 const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
1370 int mi_width = num_8x8_blocks_wide_lookup[bsize];
1371 int mi_height = num_8x8_blocks_high_lookup[bsize];
1373 // buffer the above/left context information of the block in search.
1374 for (p = 0; p < MAX_MB_PLANE; ++p) {
1376 a + num_4x4_blocks_wide * p,
1377 xd->above_context[p] + (mi_col * 2 >> xd->plane[p].subsampling_x),
1378 (sizeof(ENTROPY_CONTEXT) * num_4x4_blocks_wide) >>
1379 xd->plane[p].subsampling_x);
1381 l + num_4x4_blocks_high * p,
1383 + ((mi_row & MI_MASK) * 2 >> xd->plane[p].subsampling_y),
1384 (sizeof(ENTROPY_CONTEXT) * num_4x4_blocks_high) >>
1385 xd->plane[p].subsampling_y);
1387 vpx_memcpy(sa, xd->above_seg_context + mi_col,
1388 sizeof(*xd->above_seg_context) * mi_width);
1389 vpx_memcpy(sl, xd->left_seg_context + (mi_row & MI_MASK),
1390 sizeof(xd->left_seg_context[0]) * mi_height);
1393 static void encode_b(VP9_COMP *cpi, const TileInfo *const tile,
1395 TOKENEXTRA **tp, int mi_row, int mi_col,
1396 int output_enabled, BLOCK_SIZE bsize,
1397 PICK_MODE_CONTEXT *ctx) {
1398 MACROBLOCK *const x = &td->mb;
1399 set_offsets(cpi, tile, x, mi_row, mi_col, bsize);
1400 update_state(cpi, td, ctx, mi_row, mi_col, bsize, output_enabled);
1401 encode_superblock(cpi, td, tp, output_enabled, mi_row, mi_col, bsize, ctx);
1403 if (output_enabled) {
1404 update_stats(&cpi->common, td);
1406 (*tp)->token = EOSB_TOKEN;
1411 static void encode_sb(VP9_COMP *cpi, ThreadData *td,
1412 const TileInfo *const tile,
1413 TOKENEXTRA **tp, int mi_row, int mi_col,
1414 int output_enabled, BLOCK_SIZE bsize,
1416 VP9_COMMON *const cm = &cpi->common;
1417 MACROBLOCK *const x = &td->mb;
1418 MACROBLOCKD *const xd = &x->e_mbd;
1420 const int bsl = b_width_log2_lookup[bsize], hbs = (1 << bsl) / 4;
1422 PARTITION_TYPE partition;
1423 BLOCK_SIZE subsize = bsize;
1425 if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols)
1428 if (bsize >= BLOCK_8X8) {
1429 ctx = partition_plane_context(xd, mi_row, mi_col, bsize);
1430 subsize = get_subsize(bsize, pc_tree->partitioning);
1433 subsize = BLOCK_4X4;
1436 partition = partition_lookup[bsl][subsize];
1437 if (output_enabled && bsize != BLOCK_4X4)
1438 td->counts->partition[ctx][partition]++;
1440 switch (partition) {
1441 case PARTITION_NONE:
1442 encode_b(cpi, tile, td, tp, mi_row, mi_col, output_enabled, subsize,
1445 case PARTITION_VERT:
1446 encode_b(cpi, tile, td, tp, mi_row, mi_col, output_enabled, subsize,
1447 &pc_tree->vertical[0]);
1448 if (mi_col + hbs < cm->mi_cols && bsize > BLOCK_8X8) {
1449 encode_b(cpi, tile, td, tp, mi_row, mi_col + hbs, output_enabled,
1450 subsize, &pc_tree->vertical[1]);
1453 case PARTITION_HORZ:
1454 encode_b(cpi, tile, td, tp, mi_row, mi_col, output_enabled, subsize,
1455 &pc_tree->horizontal[0]);
1456 if (mi_row + hbs < cm->mi_rows && bsize > BLOCK_8X8) {
1457 encode_b(cpi, tile, td, tp, mi_row + hbs, mi_col, output_enabled,
1458 subsize, &pc_tree->horizontal[1]);
1461 case PARTITION_SPLIT:
1462 if (bsize == BLOCK_8X8) {
1463 encode_b(cpi, tile, td, tp, mi_row, mi_col, output_enabled, subsize,
1464 pc_tree->leaf_split[0]);
1466 encode_sb(cpi, td, tile, tp, mi_row, mi_col, output_enabled, subsize,
1468 encode_sb(cpi, td, tile, tp, mi_row, mi_col + hbs, output_enabled,
1469 subsize, pc_tree->split[1]);
1470 encode_sb(cpi, td, tile, tp, mi_row + hbs, mi_col, output_enabled,
1471 subsize, pc_tree->split[2]);
1472 encode_sb(cpi, td, tile, tp, mi_row + hbs, mi_col + hbs, output_enabled,
1473 subsize, pc_tree->split[3]);
1477 assert(0 && "Invalid partition type.");
1481 if (partition != PARTITION_SPLIT || bsize == BLOCK_8X8)
1482 update_partition_context(xd, mi_row, mi_col, subsize, bsize);
1485 // Check to see if the given partition size is allowed for a specified number
1486 // of 8x8 block rows and columns remaining in the image.
1487 // If not then return the largest allowed partition size
1488 static BLOCK_SIZE find_partition_size(BLOCK_SIZE bsize,
1489 int rows_left, int cols_left,
1491 if (rows_left <= 0 || cols_left <= 0) {
1492 return MIN(bsize, BLOCK_8X8);
1494 for (; bsize > 0; bsize -= 3) {
1495 *bh = num_8x8_blocks_high_lookup[bsize];
1496 *bw = num_8x8_blocks_wide_lookup[bsize];
1497 if ((*bh <= rows_left) && (*bw <= cols_left)) {
1505 static void set_partial_b64x64_partition(MODE_INFO *mi, int mis,
1506 int bh_in, int bw_in, int row8x8_remaining, int col8x8_remaining,
1507 BLOCK_SIZE bsize, MODE_INFO *mi_8x8) {
1510 for (r = 0; r < MI_BLOCK_SIZE; r += bh) {
1512 for (c = 0; c < MI_BLOCK_SIZE; c += bw) {
1513 const int index = r * mis + c;
1514 mi_8x8[index].src_mi = mi + index;
1515 mi_8x8[index].src_mi->mbmi.sb_type = find_partition_size(bsize,
1516 row8x8_remaining - r, col8x8_remaining - c, &bh, &bw);
1521 // This function attempts to set all mode info entries in a given SB64
1522 // to the same block partition size.
1523 // However, at the bottom and right borders of the image the requested size
1524 // may not be allowed in which case this code attempts to choose the largest
1525 // allowable partition.
1526 static void set_fixed_partitioning(VP9_COMP *cpi, const TileInfo *const tile,
1527 MODE_INFO *mi_8x8, int mi_row, int mi_col,
1529 VP9_COMMON *const cm = &cpi->common;
1530 const int mis = cm->mi_stride;
1531 const int row8x8_remaining = tile->mi_row_end - mi_row;
1532 const int col8x8_remaining = tile->mi_col_end - mi_col;
1533 int block_row, block_col;
1534 MODE_INFO *mi_upper_left = cm->mi + mi_row * mis + mi_col;
1535 int bh = num_8x8_blocks_high_lookup[bsize];
1536 int bw = num_8x8_blocks_wide_lookup[bsize];
1538 assert((row8x8_remaining > 0) && (col8x8_remaining > 0));
1540 // Apply the requested partition size to the SB64 if it is all "in image"
1541 if ((col8x8_remaining >= MI_BLOCK_SIZE) &&
1542 (row8x8_remaining >= MI_BLOCK_SIZE)) {
1543 for (block_row = 0; block_row < MI_BLOCK_SIZE; block_row += bh) {
1544 for (block_col = 0; block_col < MI_BLOCK_SIZE; block_col += bw) {
1545 int index = block_row * mis + block_col;
1546 mi_8x8[index].src_mi = mi_upper_left + index;
1547 mi_8x8[index].src_mi->mbmi.sb_type = bsize;
1551 // Else this is a partial SB64.
1552 set_partial_b64x64_partition(mi_upper_left, mis, bh, bw, row8x8_remaining,
1553 col8x8_remaining, bsize, mi_8x8);
1560 } coord_lookup[16] = {
1562 {0, 0}, {0, 2}, {2, 0}, {2, 2},
1564 {0, 4}, {0, 6}, {2, 4}, {2, 6},
1566 {4, 0}, {4, 2}, {6, 0}, {6, 2},
1568 {4, 4}, {4, 6}, {6, 4}, {6, 6},
1571 static void set_source_var_based_partition(VP9_COMP *cpi,
1572 const TileInfo *const tile,
1573 MACROBLOCK *const x,
1575 int mi_row, int mi_col) {
1576 VP9_COMMON *const cm = &cpi->common;
1577 const int mis = cm->mi_stride;
1578 const int row8x8_remaining = tile->mi_row_end - mi_row;
1579 const int col8x8_remaining = tile->mi_col_end - mi_col;
1580 MODE_INFO *mi_upper_left = cm->mi + mi_row * mis + mi_col;
1582 vp9_setup_src_planes(x, cpi->Source, mi_row, mi_col);
1584 assert((row8x8_remaining > 0) && (col8x8_remaining > 0));
1587 if ((col8x8_remaining >= MI_BLOCK_SIZE) &&
1588 (row8x8_remaining >= MI_BLOCK_SIZE)) {
1592 const int offset = (mi_row >> 1) * cm->mb_cols + (mi_col >> 1);
1593 int is_larger_better = 0;
1595 unsigned int thr = cpi->source_var_thresh;
1597 vpx_memset(d32, 0, 4 * sizeof(diff));
1599 for (i = 0; i < 4; i++) {
1602 for (j = 0; j < 4; j++) {
1603 int b_mi_row = coord_lookup[i * 4 + j].row;
1604 int b_mi_col = coord_lookup[i * 4 + j].col;
1605 int boffset = b_mi_row / 2 * cm->mb_cols +
1608 d16[j] = cpi->source_diff_var + offset + boffset;
1610 index = b_mi_row * mis + b_mi_col;
1611 mi_8x8[index].src_mi = mi_upper_left + index;
1612 mi_8x8[index].src_mi->mbmi.sb_type = BLOCK_16X16;
1614 // TODO(yunqingwang): If d16[j].var is very large, use 8x8 partition
1615 // size to further improve quality.
1618 is_larger_better = (d16[0]->var < thr) && (d16[1]->var < thr) &&
1619 (d16[2]->var < thr) && (d16[3]->var < thr);
1621 // Use 32x32 partition
1622 if (is_larger_better) {
1625 for (j = 0; j < 4; j++) {
1626 d32[i].sse += d16[j]->sse;
1627 d32[i].sum += d16[j]->sum;
1630 d32[i].var = d32[i].sse - (((int64_t)d32[i].sum * d32[i].sum) >> 10);
1632 index = coord_lookup[i*4].row * mis + coord_lookup[i*4].col;
1633 mi_8x8[index].src_mi = mi_upper_left + index;
1634 mi_8x8[index].src_mi->mbmi.sb_type = BLOCK_32X32;
1638 if (use32x32 == 4) {
1640 is_larger_better = (d32[0].var < thr) && (d32[1].var < thr) &&
1641 (d32[2].var < thr) && (d32[3].var < thr);
1643 // Use 64x64 partition
1644 if (is_larger_better) {
1645 mi_8x8[0].src_mi = mi_upper_left;
1646 mi_8x8[0].src_mi->mbmi.sb_type = BLOCK_64X64;
1649 } else { // partial in-image SB64
1650 int bh = num_8x8_blocks_high_lookup[BLOCK_16X16];
1651 int bw = num_8x8_blocks_wide_lookup[BLOCK_16X16];
1652 set_partial_b64x64_partition(mi_upper_left, mis, bh, bw,
1653 row8x8_remaining, col8x8_remaining, BLOCK_16X16, mi_8x8);
1657 static void update_state_rt(VP9_COMP *cpi, ThreadData *td,
1658 PICK_MODE_CONTEXT *ctx,
1659 int mi_row, int mi_col, int bsize) {
1660 VP9_COMMON *const cm = &cpi->common;
1661 MACROBLOCK *const x = &td->mb;
1662 MACROBLOCKD *const xd = &x->e_mbd;
1663 MODE_INFO *const mi = xd->mi[0].src_mi;
1664 MB_MODE_INFO *const mbmi = &xd->mi[0].src_mi->mbmi;
1665 const struct segmentation *const seg = &cm->seg;
1666 const int bw = num_8x8_blocks_wide_lookup[mi->mbmi.sb_type];
1667 const int bh = num_8x8_blocks_high_lookup[mi->mbmi.sb_type];
1668 const int x_mis = MIN(bw, cm->mi_cols - mi_col);
1669 const int y_mis = MIN(bh, cm->mi_rows - mi_row);
1671 xd->mi[0] = ctx->mic;
1672 xd->mi[0].src_mi = &xd->mi[0];
1674 if (seg->enabled && cpi->oxcf.aq_mode) {
1675 // For in frame complexity AQ or variance AQ, copy segment_id from
1676 // segmentation_map.
1677 if (cpi->oxcf.aq_mode == COMPLEXITY_AQ ||
1678 cpi->oxcf.aq_mode == VARIANCE_AQ ) {
1679 const uint8_t *const map = seg->update_map ? cpi->segmentation_map
1680 : cm->last_frame_seg_map;
1681 mbmi->segment_id = vp9_get_segment_id(cm, map, bsize, mi_row, mi_col);
1683 // Setting segmentation map for cyclic_refresh.
1684 vp9_cyclic_refresh_update_segment(cpi, mbmi, mi_row, mi_col, bsize,
1685 ctx->rate, ctx->dist, x->skip);
1687 vp9_init_plane_quantizers(cpi, x);
1690 if (is_inter_block(mbmi)) {
1691 vp9_update_mv_count(td);
1692 if (cm->interp_filter == SWITCHABLE) {
1693 const int pred_ctx = vp9_get_pred_context_switchable_interp(xd);
1694 ++td->counts->switchable_interp[pred_ctx][mbmi->interp_filter];
1697 if (mbmi->sb_type < BLOCK_8X8) {
1698 mbmi->mv[0].as_int = mi->bmi[3].as_mv[0].as_int;
1699 mbmi->mv[1].as_int = mi->bmi[3].as_mv[1].as_int;
1703 if (cm->use_prev_frame_mvs) {
1704 MV_REF *const frame_mvs =
1705 cm->cur_frame->mvs + mi_row * cm->mi_cols + mi_col;
1708 for (h = 0; h < y_mis; ++h) {
1709 MV_REF *const frame_mv = frame_mvs + h * cm->mi_cols;
1710 for (w = 0; w < x_mis; ++w) {
1711 MV_REF *const mv = frame_mv + w;
1712 mv->ref_frame[0] = mi->src_mi->mbmi.ref_frame[0];
1713 mv->ref_frame[1] = mi->src_mi->mbmi.ref_frame[1];
1714 mv->mv[0].as_int = mi->src_mi->mbmi.mv[0].as_int;
1715 mv->mv[1].as_int = mi->src_mi->mbmi.mv[1].as_int;
1720 x->skip = ctx->skip;
1721 x->skip_txfm[0] = mbmi->segment_id ? 0 : ctx->skip_txfm[0];
1724 static void encode_b_rt(VP9_COMP *cpi, ThreadData *td,
1725 const TileInfo *const tile,
1726 TOKENEXTRA **tp, int mi_row, int mi_col,
1727 int output_enabled, BLOCK_SIZE bsize,
1728 PICK_MODE_CONTEXT *ctx) {
1729 MACROBLOCK *const x = &td->mb;
1730 set_offsets(cpi, tile, x, mi_row, mi_col, bsize);
1731 update_state_rt(cpi, td, ctx, mi_row, mi_col, bsize);
1733 #if CONFIG_VP9_TEMPORAL_DENOISING
1734 if (cpi->oxcf.noise_sensitivity > 0 && output_enabled &&
1735 cpi->common.frame_type != KEY_FRAME) {
1736 vp9_denoiser_denoise(&cpi->denoiser, x, mi_row, mi_col,
1737 MAX(BLOCK_8X8, bsize), ctx);
1741 encode_superblock(cpi, td, tp, output_enabled, mi_row, mi_col, bsize, ctx);
1742 update_stats(&cpi->common, td);
1744 (*tp)->token = EOSB_TOKEN;
1748 static void encode_sb_rt(VP9_COMP *cpi, ThreadData *td,
1749 const TileInfo *const tile,
1750 TOKENEXTRA **tp, int mi_row, int mi_col,
1751 int output_enabled, BLOCK_SIZE bsize,
1753 VP9_COMMON *const cm = &cpi->common;
1754 MACROBLOCK *const x = &td->mb;
1755 MACROBLOCKD *const xd = &x->e_mbd;
1757 const int bsl = b_width_log2_lookup[bsize], hbs = (1 << bsl) / 4;
1759 PARTITION_TYPE partition;
1762 if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols)
1765 if (bsize >= BLOCK_8X8) {
1766 const int idx_str = xd->mi_stride * mi_row + mi_col;
1767 MODE_INFO *mi_8x8 = cm->mi[idx_str].src_mi;
1768 ctx = partition_plane_context(xd, mi_row, mi_col, bsize);
1769 subsize = mi_8x8[0].src_mi->mbmi.sb_type;
1772 subsize = BLOCK_4X4;
1775 partition = partition_lookup[bsl][subsize];
1776 if (output_enabled && bsize != BLOCK_4X4)
1777 td->counts->partition[ctx][partition]++;
1779 switch (partition) {
1780 case PARTITION_NONE:
1781 encode_b_rt(cpi, td, tile, tp, mi_row, mi_col, output_enabled, subsize,
1784 case PARTITION_VERT:
1785 encode_b_rt(cpi, td, tile, tp, mi_row, mi_col, output_enabled, subsize,
1786 &pc_tree->vertical[0]);
1787 if (mi_col + hbs < cm->mi_cols && bsize > BLOCK_8X8) {
1788 encode_b_rt(cpi, td, tile, tp, mi_row, mi_col + hbs, output_enabled,
1789 subsize, &pc_tree->vertical[1]);
1792 case PARTITION_HORZ:
1793 encode_b_rt(cpi, td, tile, tp, mi_row, mi_col, output_enabled, subsize,
1794 &pc_tree->horizontal[0]);
1795 if (mi_row + hbs < cm->mi_rows && bsize > BLOCK_8X8) {
1796 encode_b_rt(cpi, td, tile, tp, mi_row + hbs, mi_col, output_enabled,
1797 subsize, &pc_tree->horizontal[1]);
1800 case PARTITION_SPLIT:
1801 subsize = get_subsize(bsize, PARTITION_SPLIT);
1802 encode_sb_rt(cpi, td, tile, tp, mi_row, mi_col, output_enabled, subsize,
1804 encode_sb_rt(cpi, td, tile, tp, mi_row, mi_col + hbs, output_enabled,
1805 subsize, pc_tree->split[1]);
1806 encode_sb_rt(cpi, td, tile, tp, mi_row + hbs, mi_col, output_enabled,
1807 subsize, pc_tree->split[2]);
1808 encode_sb_rt(cpi, td, tile, tp, mi_row + hbs, mi_col + hbs,
1809 output_enabled, subsize, pc_tree->split[3]);
1812 assert(0 && "Invalid partition type.");
1816 if (partition != PARTITION_SPLIT || bsize == BLOCK_8X8)
1817 update_partition_context(xd, mi_row, mi_col, subsize, bsize);
1820 static void rd_use_partition(VP9_COMP *cpi,
1822 TileDataEnc *tile_data,
1823 MODE_INFO *mi_8x8, TOKENEXTRA **tp,
1824 int mi_row, int mi_col,
1826 int *rate, int64_t *dist,
1827 int do_recon, PC_TREE *pc_tree) {
1828 VP9_COMMON *const cm = &cpi->common;
1829 TileInfo *const tile_info = &tile_data->tile_info;
1830 MACROBLOCK *const x = &td->mb;
1831 MACROBLOCKD *const xd = &x->e_mbd;
1832 const int mis = cm->mi_stride;
1833 const int bsl = b_width_log2_lookup[bsize];
1834 const int mi_step = num_4x4_blocks_wide_lookup[bsize] / 2;
1835 const int bss = (1 << bsl) / 4;
1837 PARTITION_TYPE partition = PARTITION_NONE;
1839 ENTROPY_CONTEXT l[16 * MAX_MB_PLANE], a[16 * MAX_MB_PLANE];
1840 PARTITION_CONTEXT sl[8], sa[8];
1841 RD_COST last_part_rdc, none_rdc, chosen_rdc;
1842 BLOCK_SIZE sub_subsize = BLOCK_4X4;
1843 int splits_below = 0;
1844 BLOCK_SIZE bs_type = mi_8x8[0].src_mi->mbmi.sb_type;
1845 int do_partition_search = 1;
1846 PICK_MODE_CONTEXT *ctx = &pc_tree->none;
1848 if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols)
1851 assert(num_4x4_blocks_wide_lookup[bsize] ==
1852 num_4x4_blocks_high_lookup[bsize]);
1854 vp9_rd_cost_reset(&last_part_rdc);
1855 vp9_rd_cost_reset(&none_rdc);
1856 vp9_rd_cost_reset(&chosen_rdc);
1858 partition = partition_lookup[bsl][bs_type];
1859 subsize = get_subsize(bsize, partition);
1861 pc_tree->partitioning = partition;
1862 save_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
1864 if (bsize == BLOCK_16X16 && cpi->oxcf.aq_mode) {
1865 set_offsets(cpi, tile_info, x, mi_row, mi_col, bsize);
1866 x->mb_energy = vp9_block_energy(cpi, x, bsize);
1869 if (do_partition_search &&
1870 cpi->sf.partition_search_type == SEARCH_PARTITION &&
1871 cpi->sf.adjust_partitioning_from_last_frame) {
1872 // Check if any of the sub blocks are further split.
1873 if (partition == PARTITION_SPLIT && subsize > BLOCK_8X8) {
1874 sub_subsize = get_subsize(subsize, PARTITION_SPLIT);
1876 for (i = 0; i < 4; i++) {
1877 int jj = i >> 1, ii = i & 0x01;
1878 MODE_INFO *this_mi = mi_8x8[jj * bss * mis + ii * bss].src_mi;
1879 if (this_mi && this_mi->mbmi.sb_type >= sub_subsize) {
1885 // If partition is not none try none unless each of the 4 splits are split
1887 if (partition != PARTITION_NONE && !splits_below &&
1888 mi_row + (mi_step >> 1) < cm->mi_rows &&
1889 mi_col + (mi_step >> 1) < cm->mi_cols) {
1890 pc_tree->partitioning = PARTITION_NONE;
1891 rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &none_rdc, bsize,
1894 pl = partition_plane_context(xd, mi_row, mi_col, bsize);
1896 if (none_rdc.rate < INT_MAX) {
1897 none_rdc.rate += cpi->partition_cost[pl][PARTITION_NONE];
1898 none_rdc.rdcost = RDCOST(x->rdmult, x->rddiv, none_rdc.rate,
1902 restore_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
1903 mi_8x8[0].src_mi->mbmi.sb_type = bs_type;
1904 pc_tree->partitioning = partition;
1908 switch (partition) {
1909 case PARTITION_NONE:
1910 rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &last_part_rdc,
1911 bsize, ctx, INT64_MAX);
1913 case PARTITION_HORZ:
1914 rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &last_part_rdc,
1915 subsize, &pc_tree->horizontal[0],
1917 if (last_part_rdc.rate != INT_MAX &&
1918 bsize >= BLOCK_8X8 && mi_row + (mi_step >> 1) < cm->mi_rows) {
1920 PICK_MODE_CONTEXT *ctx = &pc_tree->horizontal[0];
1921 vp9_rd_cost_init(&tmp_rdc);
1922 update_state(cpi, td, ctx, mi_row, mi_col, subsize, 0);
1923 encode_superblock(cpi, td, tp, 0, mi_row, mi_col, subsize, ctx);
1924 rd_pick_sb_modes(cpi, tile_data, x,
1925 mi_row + (mi_step >> 1), mi_col, &tmp_rdc,
1926 subsize, &pc_tree->horizontal[1], INT64_MAX);
1927 if (tmp_rdc.rate == INT_MAX || tmp_rdc.dist == INT64_MAX) {
1928 vp9_rd_cost_reset(&last_part_rdc);
1931 last_part_rdc.rate += tmp_rdc.rate;
1932 last_part_rdc.dist += tmp_rdc.dist;
1933 last_part_rdc.rdcost += tmp_rdc.rdcost;
1936 case PARTITION_VERT:
1937 rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &last_part_rdc,
1938 subsize, &pc_tree->vertical[0], INT64_MAX);
1939 if (last_part_rdc.rate != INT_MAX &&
1940 bsize >= BLOCK_8X8 && mi_col + (mi_step >> 1) < cm->mi_cols) {
1942 PICK_MODE_CONTEXT *ctx = &pc_tree->vertical[0];
1943 vp9_rd_cost_init(&tmp_rdc);
1944 update_state(cpi, td, ctx, mi_row, mi_col, subsize, 0);
1945 encode_superblock(cpi, td, tp, 0, mi_row, mi_col, subsize, ctx);
1946 rd_pick_sb_modes(cpi, tile_data, x,
1947 mi_row, mi_col + (mi_step >> 1), &tmp_rdc,
1948 subsize, &pc_tree->vertical[bsize > BLOCK_8X8],
1950 if (tmp_rdc.rate == INT_MAX || tmp_rdc.dist == INT64_MAX) {
1951 vp9_rd_cost_reset(&last_part_rdc);
1954 last_part_rdc.rate += tmp_rdc.rate;
1955 last_part_rdc.dist += tmp_rdc.dist;
1956 last_part_rdc.rdcost += tmp_rdc.rdcost;
1959 case PARTITION_SPLIT:
1960 if (bsize == BLOCK_8X8) {
1961 rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &last_part_rdc,
1962 subsize, pc_tree->leaf_split[0], INT64_MAX);
1965 last_part_rdc.rate = 0;
1966 last_part_rdc.dist = 0;
1967 last_part_rdc.rdcost = 0;
1968 for (i = 0; i < 4; i++) {
1969 int x_idx = (i & 1) * (mi_step >> 1);
1970 int y_idx = (i >> 1) * (mi_step >> 1);
1971 int jj = i >> 1, ii = i & 0x01;
1973 if ((mi_row + y_idx >= cm->mi_rows) || (mi_col + x_idx >= cm->mi_cols))
1976 vp9_rd_cost_init(&tmp_rdc);
1977 rd_use_partition(cpi, td, tile_data,
1978 mi_8x8 + jj * bss * mis + ii * bss, tp,
1979 mi_row + y_idx, mi_col + x_idx, subsize,
1980 &tmp_rdc.rate, &tmp_rdc.dist,
1981 i != 3, pc_tree->split[i]);
1982 if (tmp_rdc.rate == INT_MAX || tmp_rdc.dist == INT64_MAX) {
1983 vp9_rd_cost_reset(&last_part_rdc);
1986 last_part_rdc.rate += tmp_rdc.rate;
1987 last_part_rdc.dist += tmp_rdc.dist;
1995 pl = partition_plane_context(xd, mi_row, mi_col, bsize);
1996 if (last_part_rdc.rate < INT_MAX) {
1997 last_part_rdc.rate += cpi->partition_cost[pl][partition];
1998 last_part_rdc.rdcost = RDCOST(x->rdmult, x->rddiv,
1999 last_part_rdc.rate, last_part_rdc.dist);
2002 if (do_partition_search
2003 && cpi->sf.adjust_partitioning_from_last_frame
2004 && cpi->sf.partition_search_type == SEARCH_PARTITION
2005 && partition != PARTITION_SPLIT && bsize > BLOCK_8X8
2006 && (mi_row + mi_step < cm->mi_rows ||
2007 mi_row + (mi_step >> 1) == cm->mi_rows)
2008 && (mi_col + mi_step < cm->mi_cols ||
2009 mi_col + (mi_step >> 1) == cm->mi_cols)) {
2010 BLOCK_SIZE split_subsize = get_subsize(bsize, PARTITION_SPLIT);
2011 chosen_rdc.rate = 0;
2012 chosen_rdc.dist = 0;
2013 restore_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
2014 pc_tree->partitioning = PARTITION_SPLIT;
2017 for (i = 0; i < 4; i++) {
2018 int x_idx = (i & 1) * (mi_step >> 1);
2019 int y_idx = (i >> 1) * (mi_step >> 1);
2021 ENTROPY_CONTEXT l[16 * MAX_MB_PLANE], a[16 * MAX_MB_PLANE];
2022 PARTITION_CONTEXT sl[8], sa[8];
2024 if ((mi_row + y_idx >= cm->mi_rows) || (mi_col + x_idx >= cm->mi_cols))
2027 save_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
2028 pc_tree->split[i]->partitioning = PARTITION_NONE;
2029 rd_pick_sb_modes(cpi, tile_data, x,
2030 mi_row + y_idx, mi_col + x_idx, &tmp_rdc,
2031 split_subsize, &pc_tree->split[i]->none, INT64_MAX);
2033 restore_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
2035 if (tmp_rdc.rate == INT_MAX || tmp_rdc.dist == INT64_MAX) {
2036 vp9_rd_cost_reset(&chosen_rdc);
2040 chosen_rdc.rate += tmp_rdc.rate;
2041 chosen_rdc.dist += tmp_rdc.dist;
2044 encode_sb(cpi, td, tile_info, tp, mi_row + y_idx, mi_col + x_idx, 0,
2045 split_subsize, pc_tree->split[i]);
2047 pl = partition_plane_context(xd, mi_row + y_idx, mi_col + x_idx,
2049 chosen_rdc.rate += cpi->partition_cost[pl][PARTITION_NONE];
2051 pl = partition_plane_context(xd, mi_row, mi_col, bsize);
2052 if (chosen_rdc.rate < INT_MAX) {
2053 chosen_rdc.rate += cpi->partition_cost[pl][PARTITION_SPLIT];
2054 chosen_rdc.rdcost = RDCOST(x->rdmult, x->rddiv,
2055 chosen_rdc.rate, chosen_rdc.dist);
2059 // If last_part is better set the partitioning to that.
2060 if (last_part_rdc.rdcost < chosen_rdc.rdcost) {
2061 mi_8x8[0].src_mi->mbmi.sb_type = bsize;
2062 if (bsize >= BLOCK_8X8)
2063 pc_tree->partitioning = partition;
2064 chosen_rdc = last_part_rdc;
2066 // If none was better set the partitioning to that.
2067 if (none_rdc.rdcost < chosen_rdc.rdcost) {
2068 if (bsize >= BLOCK_8X8)
2069 pc_tree->partitioning = PARTITION_NONE;
2070 chosen_rdc = none_rdc;
2073 restore_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
2075 // We must have chosen a partitioning and encoding or we'll fail later on.
2076 // No other opportunities for success.
2077 if (bsize == BLOCK_64X64)
2078 assert(chosen_rdc.rate < INT_MAX && chosen_rdc.dist < INT64_MAX);
2081 int output_enabled = (bsize == BLOCK_64X64);
2082 encode_sb(cpi, td, tile_info, tp, mi_row, mi_col, output_enabled, bsize,
2086 *rate = chosen_rdc.rate;
2087 *dist = chosen_rdc.dist;
2090 static const BLOCK_SIZE min_partition_size[BLOCK_SIZES] = {
2091 BLOCK_4X4, BLOCK_4X4, BLOCK_4X4,
2092 BLOCK_4X4, BLOCK_4X4, BLOCK_4X4,
2093 BLOCK_8X8, BLOCK_8X8, BLOCK_8X8,
2094 BLOCK_16X16, BLOCK_16X16, BLOCK_16X16,
2098 static const BLOCK_SIZE max_partition_size[BLOCK_SIZES] = {
2099 BLOCK_8X8, BLOCK_16X16, BLOCK_16X16,
2100 BLOCK_16X16, BLOCK_32X32, BLOCK_32X32,
2101 BLOCK_32X32, BLOCK_64X64, BLOCK_64X64,
2102 BLOCK_64X64, BLOCK_64X64, BLOCK_64X64,
2106 // Look at all the mode_info entries for blocks that are part of this
2107 // partition and find the min and max values for sb_type.
2108 // At the moment this is designed to work on a 64x64 SB but could be
2109 // adjusted to use a size parameter.
2111 // The min and max are assumed to have been initialized prior to calling this
2112 // function so repeat calls can accumulate a min and max of more than one sb64.
2113 static void get_sb_partition_size_range(MACROBLOCKD *xd, MODE_INFO *mi_8x8,
2114 BLOCK_SIZE *min_block_size,
2115 BLOCK_SIZE *max_block_size,
2116 int bs_hist[BLOCK_SIZES]) {
2117 int sb_width_in_blocks = MI_BLOCK_SIZE;
2118 int sb_height_in_blocks = MI_BLOCK_SIZE;
2122 // Check the sb_type for each block that belongs to this region.
2123 for (i = 0; i < sb_height_in_blocks; ++i) {
2124 for (j = 0; j < sb_width_in_blocks; ++j) {
2125 MODE_INFO *mi = mi_8x8[index+j].src_mi;
2126 BLOCK_SIZE sb_type = mi ? mi->mbmi.sb_type : 0;
2128 *min_block_size = MIN(*min_block_size, sb_type);
2129 *max_block_size = MAX(*max_block_size, sb_type);
2131 index += xd->mi_stride;
2135 // Next square block size less or equal than current block size.
2136 static const BLOCK_SIZE next_square_size[BLOCK_SIZES] = {
2137 BLOCK_4X4, BLOCK_4X4, BLOCK_4X4,
2138 BLOCK_8X8, BLOCK_8X8, BLOCK_8X8,
2139 BLOCK_16X16, BLOCK_16X16, BLOCK_16X16,
2140 BLOCK_32X32, BLOCK_32X32, BLOCK_32X32,
2144 // Look at neighboring blocks and set a min and max partition size based on
2146 static void rd_auto_partition_range(VP9_COMP *cpi, const TileInfo *const tile,
2147 MACROBLOCKD *const xd,
2148 int mi_row, int mi_col,
2149 BLOCK_SIZE *min_block_size,
2150 BLOCK_SIZE *max_block_size) {
2151 VP9_COMMON *const cm = &cpi->common;
2152 MODE_INFO *mi = xd->mi[0].src_mi;
2153 const int left_in_image = xd->left_available && mi[-1].src_mi;
2154 const int above_in_image = xd->up_available && mi[-xd->mi_stride].src_mi;
2155 const int row8x8_remaining = tile->mi_row_end - mi_row;
2156 const int col8x8_remaining = tile->mi_col_end - mi_col;
2158 BLOCK_SIZE min_size = BLOCK_4X4;
2159 BLOCK_SIZE max_size = BLOCK_64X64;
2161 int bs_hist[BLOCK_SIZES] = {0};
2163 // Trap case where we do not have a prediction.
2164 if (left_in_image || above_in_image || cm->frame_type != KEY_FRAME) {
2165 // Default "min to max" and "max to min"
2166 min_size = BLOCK_64X64;
2167 max_size = BLOCK_4X4;
2169 // NOTE: each call to get_sb_partition_size_range() uses the previous
2170 // passed in values for min and max as a starting point.
2171 // Find the min and max partition used in previous frame at this location
2172 if (cm->frame_type != KEY_FRAME) {
2173 MODE_INFO *prev_mi =
2174 cm->prev_mip + cm->mi_stride + 1 + 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].src_mi;
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].src_mi;
2187 get_sb_partition_size_range(xd, above_sb64_mi, &min_size, &max_size,
2191 // adjust observed min and max
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];
2195 } else if (cpi->sf.auto_min_max_partition_size ==
2196 CONSTRAIN_NEIGHBORING_MIN_MAX) {
2197 // adjust the search range based on the histogram of the observed
2198 // partition sizes from left, above the previous co-located blocks
2200 int first_moment = 0;
2201 int second_moment = 0;
2202 int var_unnormalized = 0;
2204 for (i = 0; i < BLOCK_SIZES; i++) {
2206 first_moment += bs_hist[i] * i;
2207 second_moment += bs_hist[i] * i * i;
2210 // if variance is small enough,
2211 // adjust the range around its mean size, which gives a tighter range
2212 var_unnormalized = second_moment - first_moment * first_moment / sum;
2213 if (var_unnormalized <= 4 * sum) {
2214 int mean = first_moment / sum;
2215 min_size = min_partition_size[mean];
2216 max_size = max_partition_size[mean];
2218 min_size = min_partition_size[min_size];
2219 max_size = max_partition_size[max_size];
2224 // Check border cases where max and min from neighbors may not be legal.
2225 max_size = find_partition_size(max_size,
2226 row8x8_remaining, col8x8_remaining,
2228 min_size = MIN(min_size, max_size);
2230 // When use_square_partition_only is true, make sure at least one square
2231 // partition is allowed by selecting the next smaller square size as
2233 if (cpi->sf.use_square_partition_only &&
2234 next_square_size[max_size] < min_size) {
2235 min_size = next_square_size[max_size];
2238 *min_block_size = min_size;
2239 *max_block_size = max_size;
2242 static void auto_partition_range(VP9_COMP *cpi, const TileInfo *const tile,
2243 MACROBLOCKD *const xd,
2244 int mi_row, int mi_col,
2245 BLOCK_SIZE *min_block_size,
2246 BLOCK_SIZE *max_block_size) {
2247 VP9_COMMON *const cm = &cpi->common;
2248 MODE_INFO *mi_8x8 = xd->mi;
2249 const int left_in_image = xd->left_available && mi_8x8[-1].src_mi;
2250 const int above_in_image = xd->up_available &&
2251 mi_8x8[-xd->mi_stride].src_mi;
2252 int row8x8_remaining = tile->mi_row_end - mi_row;
2253 int col8x8_remaining = tile->mi_col_end - mi_col;
2255 BLOCK_SIZE min_size = BLOCK_32X32;
2256 BLOCK_SIZE max_size = BLOCK_8X8;
2257 int bsl = mi_width_log2_lookup[BLOCK_64X64];
2258 const int search_range_ctrl = (((mi_row + mi_col) >> bsl) +
2259 get_chessboard_index(cm->current_video_frame)) & 0x1;
2260 // Trap case where we do not have a prediction.
2261 if (search_range_ctrl &&
2262 (left_in_image || above_in_image || cm->frame_type != KEY_FRAME)) {
2267 // Find the min and max partition sizes used in the left SB64.
2268 if (left_in_image) {
2270 mi = mi_8x8[-1].src_mi;
2271 for (block = 0; block < MI_BLOCK_SIZE; ++block) {
2272 cur_mi = mi[block * xd->mi_stride].src_mi;
2273 sb_type = cur_mi ? cur_mi->mbmi.sb_type : 0;
2274 min_size = MIN(min_size, sb_type);
2275 max_size = MAX(max_size, sb_type);
2278 // Find the min and max partition sizes used in the above SB64.
2279 if (above_in_image) {
2280 mi = mi_8x8[-xd->mi_stride * MI_BLOCK_SIZE].src_mi;
2281 for (block = 0; block < MI_BLOCK_SIZE; ++block) {
2282 sb_type = mi[block].src_mi ? mi[block].src_mi->mbmi.sb_type : 0;
2283 min_size = MIN(min_size, sb_type);
2284 max_size = MAX(max_size, sb_type);
2288 min_size = min_partition_size[min_size];
2289 max_size = find_partition_size(max_size, row8x8_remaining, col8x8_remaining,
2291 min_size = MIN(min_size, max_size);
2292 min_size = MAX(min_size, BLOCK_8X8);
2293 max_size = MIN(max_size, BLOCK_32X32);
2295 min_size = BLOCK_8X8;
2296 max_size = BLOCK_32X32;
2299 *min_block_size = min_size;
2300 *max_block_size = max_size;
2303 // TODO(jingning) refactor functions setting partition search range
2304 static void set_partition_range(VP9_COMMON *cm, MACROBLOCKD *xd,
2305 int mi_row, int mi_col, BLOCK_SIZE bsize,
2306 BLOCK_SIZE *min_bs, BLOCK_SIZE *max_bs) {
2307 int mi_width = num_8x8_blocks_wide_lookup[bsize];
2308 int mi_height = num_8x8_blocks_high_lookup[bsize];
2312 const int idx_str = cm->mi_stride * mi_row + mi_col;
2313 MODE_INFO *prev_mi = (cm->prev_mip + cm->mi_stride + 1 + idx_str)->src_mi;
2316 BLOCK_SIZE bs, min_size, max_size;
2318 min_size = BLOCK_64X64;
2319 max_size = BLOCK_4X4;
2322 for (idy = 0; idy < mi_height; ++idy) {
2323 for (idx = 0; idx < mi_width; ++idx) {
2324 mi = prev_mi[idy * cm->mi_stride + idx].src_mi;
2325 bs = mi ? mi->mbmi.sb_type : bsize;
2326 min_size = MIN(min_size, bs);
2327 max_size = MAX(max_size, bs);
2332 if (xd->left_available) {
2333 for (idy = 0; idy < mi_height; ++idy) {
2334 mi = xd->mi[idy * cm->mi_stride - 1].src_mi;
2335 bs = mi ? mi->mbmi.sb_type : bsize;
2336 min_size = MIN(min_size, bs);
2337 max_size = MAX(max_size, bs);
2341 if (xd->up_available) {
2342 for (idx = 0; idx < mi_width; ++idx) {
2343 mi = xd->mi[idx - cm->mi_stride].src_mi;
2344 bs = mi ? mi->mbmi.sb_type : bsize;
2345 min_size = MIN(min_size, bs);
2346 max_size = MAX(max_size, bs);
2350 if (min_size == max_size) {
2351 min_size = min_partition_size[min_size];
2352 max_size = max_partition_size[max_size];
2359 static INLINE void store_pred_mv(MACROBLOCK *x, PICK_MODE_CONTEXT *ctx) {
2360 vpx_memcpy(ctx->pred_mv, x->pred_mv, sizeof(x->pred_mv));
2363 static INLINE void load_pred_mv(MACROBLOCK *x, PICK_MODE_CONTEXT *ctx) {
2364 vpx_memcpy(x->pred_mv, ctx->pred_mv, sizeof(x->pred_mv));
2367 #if CONFIG_FP_MB_STATS
2368 const int num_16x16_blocks_wide_lookup[BLOCK_SIZES] =
2369 {1, 1, 1, 1, 1, 1, 1, 1, 2, 2, 2, 4, 4};
2370 const int num_16x16_blocks_high_lookup[BLOCK_SIZES] =
2371 {1, 1, 1, 1, 1, 1, 1, 2, 1, 2, 4, 2, 4};
2372 const int qindex_skip_threshold_lookup[BLOCK_SIZES] =
2373 {0, 10, 10, 30, 40, 40, 60, 80, 80, 90, 100, 100, 120};
2374 const int qindex_split_threshold_lookup[BLOCK_SIZES] =
2375 {0, 3, 3, 7, 15, 15, 30, 40, 40, 60, 80, 80, 120};
2376 const int complexity_16x16_blocks_threshold[BLOCK_SIZES] =
2377 {1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 4, 4, 6};
2388 static INLINE MOTION_DIRECTION get_motion_direction_fp(uint8_t fp_byte) {
2389 if (fp_byte & FPMB_MOTION_ZERO_MASK) {
2391 } else if (fp_byte & FPMB_MOTION_LEFT_MASK) {
2393 } else if (fp_byte & FPMB_MOTION_RIGHT_MASK) {
2395 } else if (fp_byte & FPMB_MOTION_UP_MASK) {
2402 static INLINE int get_motion_inconsistency(MOTION_DIRECTION this_mv,
2403 MOTION_DIRECTION that_mv) {
2404 if (this_mv == that_mv) {
2407 return abs(this_mv - that_mv) == 2 ? 2 : 1;
2412 // TODO(jingning,jimbankoski,rbultje): properly skip partition types that are
2413 // unlikely to be selected depending on previous rate-distortion optimization
2414 // results, for encoding speed-up.
2415 static void rd_pick_partition(VP9_COMP *cpi, ThreadData *td,
2416 TileDataEnc *tile_data,
2417 TOKENEXTRA **tp, int mi_row, int mi_col,
2418 BLOCK_SIZE bsize, RD_COST *rd_cost,
2419 int64_t best_rd, PC_TREE *pc_tree) {
2420 VP9_COMMON *const cm = &cpi->common;
2421 TileInfo *const tile_info = &tile_data->tile_info;
2422 MACROBLOCK *const x = &td->mb;
2423 MACROBLOCKD *const xd = &x->e_mbd;
2424 const int mi_step = num_8x8_blocks_wide_lookup[bsize] / 2;
2425 ENTROPY_CONTEXT l[16 * MAX_MB_PLANE], a[16 * MAX_MB_PLANE];
2426 PARTITION_CONTEXT sl[8], sa[8];
2427 TOKENEXTRA *tp_orig = *tp;
2428 PICK_MODE_CONTEXT *ctx = &pc_tree->none;
2431 RD_COST this_rdc, sum_rdc, best_rdc;
2432 int do_split = bsize >= BLOCK_8X8;
2435 // Override skipping rectangular partition operations for edge blocks
2436 const int force_horz_split = (mi_row + mi_step >= cm->mi_rows);
2437 const int force_vert_split = (mi_col + mi_step >= cm->mi_cols);
2438 const int xss = x->e_mbd.plane[1].subsampling_x;
2439 const int yss = x->e_mbd.plane[1].subsampling_y;
2441 BLOCK_SIZE min_size = x->min_partition_size;
2442 BLOCK_SIZE max_size = x->max_partition_size;
2444 #if CONFIG_FP_MB_STATS
2445 unsigned int src_diff_var = UINT_MAX;
2446 int none_complexity = 0;
2449 int partition_none_allowed = !force_horz_split && !force_vert_split;
2450 int partition_horz_allowed = !force_vert_split && yss <= xss &&
2452 int partition_vert_allowed = !force_horz_split && xss <= yss &&
2456 assert(num_8x8_blocks_wide_lookup[bsize] ==
2457 num_8x8_blocks_high_lookup[bsize]);
2459 vp9_rd_cost_init(&this_rdc);
2460 vp9_rd_cost_init(&sum_rdc);
2461 vp9_rd_cost_reset(&best_rdc);
2462 best_rdc.rdcost = best_rd;
2464 set_offsets(cpi, tile_info, x, mi_row, mi_col, bsize);
2466 if (bsize == BLOCK_16X16 && cpi->oxcf.aq_mode)
2467 x->mb_energy = vp9_block_energy(cpi, x, bsize);
2469 if (cpi->sf.cb_partition_search && bsize == BLOCK_16X16) {
2470 int cb_partition_search_ctrl = ((pc_tree->index == 0 || pc_tree->index == 3)
2471 + get_chessboard_index(cm->current_video_frame)) & 0x1;
2473 if (cb_partition_search_ctrl && bsize > min_size && bsize < max_size)
2474 set_partition_range(cm, xd, mi_row, mi_col, bsize, &min_size, &max_size);
2477 // Determine partition types in search according to the speed features.
2478 // The threshold set here has to be of square block size.
2479 if (cpi->sf.auto_min_max_partition_size) {
2480 partition_none_allowed &= (bsize <= max_size && bsize >= min_size);
2481 partition_horz_allowed &= ((bsize <= max_size && bsize > min_size) ||
2483 partition_vert_allowed &= ((bsize <= max_size && bsize > min_size) ||
2485 do_split &= bsize > min_size;
2487 if (cpi->sf.use_square_partition_only) {
2488 partition_horz_allowed &= force_horz_split;
2489 partition_vert_allowed &= force_vert_split;
2492 save_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
2494 #if CONFIG_FP_MB_STATS
2495 if (cpi->use_fp_mb_stats) {
2496 set_offsets(cpi, tile_info, x, mi_row, mi_col, bsize);
2497 src_diff_var = get_sby_perpixel_diff_variance(cpi, &x->plane[0].src,
2498 mi_row, mi_col, bsize);
2502 #if CONFIG_FP_MB_STATS
2503 // Decide whether we shall split directly and skip searching NONE by using
2504 // the first pass block statistics
2505 if (cpi->use_fp_mb_stats && bsize >= BLOCK_32X32 && do_split &&
2506 partition_none_allowed && src_diff_var > 4 &&
2507 cm->base_qindex < qindex_split_threshold_lookup[bsize]) {
2508 int mb_row = mi_row >> 1;
2509 int mb_col = mi_col >> 1;
2511 MIN(mb_row + num_16x16_blocks_high_lookup[bsize], cm->mb_rows);
2513 MIN(mb_col + num_16x16_blocks_wide_lookup[bsize], cm->mb_cols);
2516 // compute a complexity measure, basically measure inconsistency of motion
2517 // vectors obtained from the first pass in the current block
2518 for (r = mb_row; r < mb_row_end ; r++) {
2519 for (c = mb_col; c < mb_col_end; c++) {
2520 const int mb_index = r * cm->mb_cols + c;
2522 MOTION_DIRECTION this_mv;
2523 MOTION_DIRECTION right_mv;
2524 MOTION_DIRECTION bottom_mv;
2527 get_motion_direction_fp(cpi->twopass.this_frame_mb_stats[mb_index]);
2530 if (c != mb_col_end - 1) {
2531 right_mv = get_motion_direction_fp(
2532 cpi->twopass.this_frame_mb_stats[mb_index + 1]);
2533 none_complexity += get_motion_inconsistency(this_mv, right_mv);
2537 if (r != mb_row_end - 1) {
2538 bottom_mv = get_motion_direction_fp(
2539 cpi->twopass.this_frame_mb_stats[mb_index + cm->mb_cols]);
2540 none_complexity += get_motion_inconsistency(this_mv, bottom_mv);
2543 // do not count its left and top neighbors to avoid double counting
2547 if (none_complexity > complexity_16x16_blocks_threshold[bsize]) {
2548 partition_none_allowed = 0;
2554 if (partition_none_allowed) {
2555 rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col,
2556 &this_rdc, bsize, ctx, best_rdc.rdcost);
2557 if (this_rdc.rate != INT_MAX) {
2558 if (bsize >= BLOCK_8X8) {
2559 pl = partition_plane_context(xd, mi_row, mi_col, bsize);
2560 this_rdc.rate += cpi->partition_cost[pl][PARTITION_NONE];
2561 this_rdc.rdcost = RDCOST(x->rdmult, x->rddiv,
2562 this_rdc.rate, this_rdc.dist);
2565 if (this_rdc.rdcost < best_rdc.rdcost) {
2566 int64_t dist_breakout_thr = cpi->sf.partition_search_breakout_dist_thr;
2567 int rate_breakout_thr = cpi->sf.partition_search_breakout_rate_thr;
2569 best_rdc = this_rdc;
2570 if (bsize >= BLOCK_8X8)
2571 pc_tree->partitioning = PARTITION_NONE;
2573 // Adjust dist breakout threshold according to the partition size.
2574 dist_breakout_thr >>= 8 - (b_width_log2_lookup[bsize] +
2575 b_height_log2_lookup[bsize]);
2577 rate_breakout_thr *= num_pels_log2_lookup[bsize];
2579 // If all y, u, v transform blocks in this partition are skippable, and
2580 // the dist & rate are within the thresholds, the partition search is
2581 // terminated for current branch of the partition search tree.
2582 // The dist & rate thresholds are set to 0 at speed 0 to disable the
2583 // early termination at that speed.
2584 if (!x->e_mbd.lossless &&
2585 (ctx->skippable && best_rdc.dist < dist_breakout_thr &&
2586 best_rdc.rate < rate_breakout_thr)) {
2591 #if CONFIG_FP_MB_STATS
2592 // Check if every 16x16 first pass block statistics has zero
2593 // motion and the corresponding first pass residue is small enough.
2594 // If that is the case, check the difference variance between the
2595 // current frame and the last frame. If the variance is small enough,
2596 // stop further splitting in RD optimization
2597 if (cpi->use_fp_mb_stats && do_split != 0 &&
2598 cm->base_qindex > qindex_skip_threshold_lookup[bsize]) {
2599 int mb_row = mi_row >> 1;
2600 int mb_col = mi_col >> 1;
2602 MIN(mb_row + num_16x16_blocks_high_lookup[bsize], cm->mb_rows);
2604 MIN(mb_col + num_16x16_blocks_wide_lookup[bsize], cm->mb_cols);
2608 for (r = mb_row; r < mb_row_end; r++) {
2609 for (c = mb_col; c < mb_col_end; c++) {
2610 const int mb_index = r * cm->mb_cols + c;
2611 if (!(cpi->twopass.this_frame_mb_stats[mb_index] &
2612 FPMB_MOTION_ZERO_MASK) ||
2613 !(cpi->twopass.this_frame_mb_stats[mb_index] &
2614 FPMB_ERROR_SMALL_MASK)) {
2624 if (src_diff_var == UINT_MAX) {
2625 set_offsets(cpi, tile_info, x, mi_row, mi_col, bsize);
2626 src_diff_var = get_sby_perpixel_diff_variance(
2627 cpi, &x->plane[0].src, mi_row, mi_col, bsize);
2629 if (src_diff_var < 8) {
2638 restore_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
2641 // store estimated motion vector
2642 if (cpi->sf.adaptive_motion_search)
2643 store_pred_mv(x, ctx);
2646 // TODO(jingning): use the motion vectors given by the above search as
2647 // the starting point of motion search in the following partition type check.
2649 subsize = get_subsize(bsize, PARTITION_SPLIT);
2650 if (bsize == BLOCK_8X8) {
2652 if (cpi->sf.adaptive_pred_interp_filter && partition_none_allowed)
2653 pc_tree->leaf_split[0]->pred_interp_filter =
2654 ctx->mic.mbmi.interp_filter;
2655 rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &sum_rdc, subsize,
2656 pc_tree->leaf_split[0], best_rdc.rdcost);
2657 if (sum_rdc.rate == INT_MAX)
2658 sum_rdc.rdcost = INT64_MAX;
2660 for (i = 0; i < 4 && sum_rdc.rdcost < best_rdc.rdcost; ++i) {
2661 const int x_idx = (i & 1) * mi_step;
2662 const int y_idx = (i >> 1) * mi_step;
2664 if (mi_row + y_idx >= cm->mi_rows || mi_col + x_idx >= cm->mi_cols)
2667 if (cpi->sf.adaptive_motion_search)
2668 load_pred_mv(x, ctx);
2670 pc_tree->split[i]->index = i;
2671 rd_pick_partition(cpi, td, tile_data, tp,
2672 mi_row + y_idx, mi_col + x_idx,
2674 best_rdc.rdcost - sum_rdc.rdcost, pc_tree->split[i]);
2676 if (this_rdc.rate == INT_MAX) {
2677 sum_rdc.rdcost = INT64_MAX;
2680 sum_rdc.rate += this_rdc.rate;
2681 sum_rdc.dist += this_rdc.dist;
2682 sum_rdc.rdcost += this_rdc.rdcost;
2687 if (sum_rdc.rdcost < best_rdc.rdcost && i == 4) {
2688 pl = partition_plane_context(xd, mi_row, mi_col, bsize);
2689 sum_rdc.rate += cpi->partition_cost[pl][PARTITION_SPLIT];
2690 sum_rdc.rdcost = RDCOST(x->rdmult, x->rddiv,
2691 sum_rdc.rate, sum_rdc.dist);
2693 if (sum_rdc.rdcost < best_rdc.rdcost) {
2695 pc_tree->partitioning = PARTITION_SPLIT;
2698 // skip rectangular partition test when larger block size
2699 // gives better rd cost
2700 if (cpi->sf.less_rectangular_check)
2701 do_rect &= !partition_none_allowed;
2703 restore_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
2707 if (partition_horz_allowed && do_rect) {
2708 subsize = get_subsize(bsize, PARTITION_HORZ);
2709 if (cpi->sf.adaptive_motion_search)
2710 load_pred_mv(x, ctx);
2711 if (cpi->sf.adaptive_pred_interp_filter && bsize == BLOCK_8X8 &&
2712 partition_none_allowed)
2713 pc_tree->horizontal[0].pred_interp_filter =
2714 ctx->mic.mbmi.interp_filter;
2715 rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &sum_rdc, subsize,
2716 &pc_tree->horizontal[0], best_rdc.rdcost);
2718 if (sum_rdc.rdcost < best_rdc.rdcost && mi_row + mi_step < cm->mi_rows &&
2719 bsize > BLOCK_8X8) {
2720 PICK_MODE_CONTEXT *ctx = &pc_tree->horizontal[0];
2721 update_state(cpi, td, ctx, mi_row, mi_col, subsize, 0);
2722 encode_superblock(cpi, td, tp, 0, mi_row, mi_col, subsize, ctx);
2724 if (cpi->sf.adaptive_motion_search)
2725 load_pred_mv(x, ctx);
2726 if (cpi->sf.adaptive_pred_interp_filter && bsize == BLOCK_8X8 &&
2727 partition_none_allowed)
2728 pc_tree->horizontal[1].pred_interp_filter =
2729 ctx->mic.mbmi.interp_filter;
2730 rd_pick_sb_modes(cpi, tile_data, x, mi_row + mi_step, mi_col,
2731 &this_rdc, subsize, &pc_tree->horizontal[1],
2732 best_rdc.rdcost - sum_rdc.rdcost);
2733 if (this_rdc.rate == INT_MAX) {
2734 sum_rdc.rdcost = INT64_MAX;
2736 sum_rdc.rate += this_rdc.rate;
2737 sum_rdc.dist += this_rdc.dist;
2738 sum_rdc.rdcost += this_rdc.rdcost;
2742 if (sum_rdc.rdcost < best_rdc.rdcost) {
2743 pl = partition_plane_context(xd, mi_row, mi_col, bsize);
2744 sum_rdc.rate += cpi->partition_cost[pl][PARTITION_HORZ];
2745 sum_rdc.rdcost = RDCOST(x->rdmult, x->rddiv, sum_rdc.rate, sum_rdc.dist);
2746 if (sum_rdc.rdcost < best_rdc.rdcost) {
2748 pc_tree->partitioning = PARTITION_HORZ;
2751 restore_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
2754 if (partition_vert_allowed && do_rect) {
2755 subsize = get_subsize(bsize, PARTITION_VERT);
2757 if (cpi->sf.adaptive_motion_search)
2758 load_pred_mv(x, ctx);
2759 if (cpi->sf.adaptive_pred_interp_filter && bsize == BLOCK_8X8 &&
2760 partition_none_allowed)
2761 pc_tree->vertical[0].pred_interp_filter =
2762 ctx->mic.mbmi.interp_filter;
2763 rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &sum_rdc, subsize,
2764 &pc_tree->vertical[0], best_rdc.rdcost);
2765 if (sum_rdc.rdcost < best_rdc.rdcost && mi_col + mi_step < cm->mi_cols &&
2766 bsize > BLOCK_8X8) {
2767 update_state(cpi, td, &pc_tree->vertical[0], mi_row, mi_col, subsize, 0);
2768 encode_superblock(cpi, td, tp, 0, mi_row, mi_col, subsize,
2769 &pc_tree->vertical[0]);
2771 if (cpi->sf.adaptive_motion_search)
2772 load_pred_mv(x, ctx);
2773 if (cpi->sf.adaptive_pred_interp_filter && bsize == BLOCK_8X8 &&
2774 partition_none_allowed)
2775 pc_tree->vertical[1].pred_interp_filter =
2776 ctx->mic.mbmi.interp_filter;
2777 rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col + mi_step,
2779 &pc_tree->vertical[1], best_rdc.rdcost - sum_rdc.rdcost);
2780 if (this_rdc.rate == INT_MAX) {
2781 sum_rdc.rdcost = INT64_MAX;
2783 sum_rdc.rate += this_rdc.rate;
2784 sum_rdc.dist += this_rdc.dist;
2785 sum_rdc.rdcost += this_rdc.rdcost;
2789 if (sum_rdc.rdcost < best_rdc.rdcost) {
2790 pl = partition_plane_context(xd, mi_row, mi_col, bsize);
2791 sum_rdc.rate += cpi->partition_cost[pl][PARTITION_VERT];
2792 sum_rdc.rdcost = RDCOST(x->rdmult, x->rddiv,
2793 sum_rdc.rate, sum_rdc.dist);
2794 if (sum_rdc.rdcost < best_rdc.rdcost) {
2796 pc_tree->partitioning = PARTITION_VERT;
2799 restore_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
2802 // TODO(jbb): This code added so that we avoid static analysis
2803 // warning related to the fact that best_rd isn't used after this
2804 // point. This code should be refactored so that the duplicate
2805 // checks occur in some sub function and thus are used...
2807 *rd_cost = best_rdc;
2810 if (best_rdc.rate < INT_MAX && best_rdc.dist < INT64_MAX &&
2811 pc_tree->index != 3) {
2812 int output_enabled = (bsize == BLOCK_64X64);
2813 encode_sb(cpi, td, tile_info, tp, mi_row, mi_col, output_enabled,
2817 if (bsize == BLOCK_64X64) {
2818 assert(tp_orig < *tp);
2819 assert(best_rdc.rate < INT_MAX);
2820 assert(best_rdc.dist < INT64_MAX);
2822 assert(tp_orig == *tp);
2826 static void encode_rd_sb_row(VP9_COMP *cpi,
2828 TileDataEnc *tile_data,
2831 VP9_COMMON *const cm = &cpi->common;
2832 TileInfo *const tile_info = &tile_data->tile_info;
2833 MACROBLOCK *const x = &td->mb;
2834 MACROBLOCKD *const xd = &x->e_mbd;
2835 SPEED_FEATURES *const sf = &cpi->sf;
2838 // Initialize the left context for the new SB row
2839 vpx_memset(&xd->left_context, 0, sizeof(xd->left_context));
2840 vpx_memset(xd->left_seg_context, 0, sizeof(xd->left_seg_context));
2842 // Code each SB in the row
2843 for (mi_col = tile_info->mi_col_start; mi_col < tile_info->mi_col_end;
2844 mi_col += MI_BLOCK_SIZE) {
2845 const struct segmentation *const seg = &cm->seg;
2852 const int idx_str = cm->mi_stride * mi_row + mi_col;
2853 MODE_INFO *mi = cm->mi + idx_str;
2855 if (sf->adaptive_pred_interp_filter) {
2856 for (i = 0; i < 64; ++i)
2857 td->leaf_tree[i].pred_interp_filter = SWITCHABLE;
2859 for (i = 0; i < 64; ++i) {
2860 td->pc_tree[i].vertical[0].pred_interp_filter = SWITCHABLE;
2861 td->pc_tree[i].vertical[1].pred_interp_filter = SWITCHABLE;
2862 td->pc_tree[i].horizontal[0].pred_interp_filter = SWITCHABLE;
2863 td->pc_tree[i].horizontal[1].pred_interp_filter = SWITCHABLE;
2867 vp9_zero(x->pred_mv);
2868 td->pc_root->index = 0;
2871 const uint8_t *const map = seg->update_map ? cpi->segmentation_map
2872 : cm->last_frame_seg_map;
2873 int segment_id = vp9_get_segment_id(cm, map, BLOCK_64X64, mi_row, mi_col);
2874 seg_skip = vp9_segfeature_active(seg, segment_id, SEG_LVL_SKIP);
2877 x->source_variance = UINT_MAX;
2878 if (sf->partition_search_type == FIXED_PARTITION || seg_skip) {
2879 const BLOCK_SIZE bsize =
2880 seg_skip ? BLOCK_64X64 : sf->always_this_block_size;
2881 set_offsets(cpi, tile_info, x, mi_row, mi_col, BLOCK_64X64);
2882 set_fixed_partitioning(cpi, tile_info, mi, mi_row, mi_col, bsize);
2883 rd_use_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col,
2884 BLOCK_64X64, &dummy_rate, &dummy_dist, 1, td->pc_root);
2885 } else if (cpi->partition_search_skippable_frame) {
2887 set_offsets(cpi, tile_info, x, mi_row, mi_col, BLOCK_64X64);
2888 bsize = get_rd_var_based_fixed_partition(cpi, x, mi_row, mi_col);
2889 set_fixed_partitioning(cpi, tile_info, mi, mi_row, mi_col, bsize);
2890 rd_use_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col,
2891 BLOCK_64X64, &dummy_rate, &dummy_dist, 1, td->pc_root);
2892 } else if (sf->partition_search_type == VAR_BASED_PARTITION &&
2893 cm->frame_type != KEY_FRAME) {
2894 choose_partitioning(cpi, tile_info, x, mi_row, mi_col);
2895 rd_use_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col,
2896 BLOCK_64X64, &dummy_rate, &dummy_dist, 1, td->pc_root);
2898 // If required set upper and lower partition size limits
2899 if (sf->auto_min_max_partition_size) {
2900 set_offsets(cpi, tile_info, x, mi_row, mi_col, BLOCK_64X64);
2901 rd_auto_partition_range(cpi, tile_info, xd, mi_row, mi_col,
2902 &x->min_partition_size,
2903 &x->max_partition_size);
2905 rd_pick_partition(cpi, td, tile_data, tp, mi_row, mi_col, BLOCK_64X64,
2906 &dummy_rdc, INT64_MAX, td->pc_root);
2911 static void init_encode_frame_mb_context(VP9_COMP *cpi) {
2912 MACROBLOCK *const x = &cpi->td.mb;
2913 VP9_COMMON *const cm = &cpi->common;
2914 MACROBLOCKD *const xd = &x->e_mbd;
2915 const int aligned_mi_cols = mi_cols_aligned_to_sb(cm->mi_cols);
2917 // Copy data over into macro block data structures.
2918 vp9_setup_src_planes(x, cpi->Source, 0, 0);
2920 vp9_setup_block_planes(&x->e_mbd, cm->subsampling_x, cm->subsampling_y);
2922 // Note: this memset assumes above_context[0], [1] and [2]
2923 // are allocated as part of the same buffer.
2924 vpx_memset(xd->above_context[0], 0,
2925 sizeof(*xd->above_context[0]) *
2926 2 * aligned_mi_cols * MAX_MB_PLANE);
2927 vpx_memset(xd->above_seg_context, 0,
2928 sizeof(*xd->above_seg_context) * aligned_mi_cols);
2931 static int check_dual_ref_flags(VP9_COMP *cpi) {
2932 const int ref_flags = cpi->ref_frame_flags;
2934 if (vp9_segfeature_active(&cpi->common.seg, 1, SEG_LVL_REF_FRAME)) {
2937 return (!!(ref_flags & VP9_GOLD_FLAG) + !!(ref_flags & VP9_LAST_FLAG)
2938 + !!(ref_flags & VP9_ALT_FLAG)) >= 2;
2942 static void reset_skip_tx_size(VP9_COMMON *cm, TX_SIZE max_tx_size) {
2944 const int mis = cm->mi_stride;
2945 MODE_INFO *mi_ptr = cm->mi;
2947 for (mi_row = 0; mi_row < cm->mi_rows; ++mi_row, mi_ptr += mis) {
2948 for (mi_col = 0; mi_col < cm->mi_cols; ++mi_col) {
2949 if (mi_ptr[mi_col].src_mi->mbmi.tx_size > max_tx_size)
2950 mi_ptr[mi_col].src_mi->mbmi.tx_size = max_tx_size;
2955 static MV_REFERENCE_FRAME get_frame_type(const VP9_COMP *cpi) {
2956 if (frame_is_intra_only(&cpi->common))
2958 else if (cpi->rc.is_src_frame_alt_ref && cpi->refresh_golden_frame)
2959 return ALTREF_FRAME;
2960 else if (cpi->refresh_golden_frame || cpi->refresh_alt_ref_frame)
2961 return GOLDEN_FRAME;
2966 static TX_MODE select_tx_mode(const VP9_COMP *cpi, MACROBLOCKD *const xd) {
2969 if (cpi->common.frame_type == KEY_FRAME &&
2970 cpi->sf.use_nonrd_pick_mode &&
2971 cpi->sf.partition_search_type == VAR_BASED_PARTITION)
2973 if (cpi->sf.tx_size_search_method == USE_LARGESTALL)
2975 else if (cpi->sf.tx_size_search_method == USE_FULL_RD||
2976 cpi->sf.tx_size_search_method == USE_TX_8X8)
2977 return TX_MODE_SELECT;
2979 return cpi->common.tx_mode;
2982 static void hybrid_intra_mode_search(VP9_COMP *cpi, MACROBLOCK *const x,
2983 RD_COST *rd_cost, BLOCK_SIZE bsize,
2984 PICK_MODE_CONTEXT *ctx) {
2985 if (bsize < BLOCK_16X16)
2986 vp9_rd_pick_intra_mode_sb(cpi, x, rd_cost, bsize, ctx, INT64_MAX);
2988 vp9_pick_intra_mode(cpi, x, rd_cost, bsize, ctx);
2991 static void nonrd_pick_sb_modes(VP9_COMP *cpi,
2992 TileDataEnc *tile_data, MACROBLOCK *const x,
2993 int mi_row, int mi_col, RD_COST *rd_cost,
2994 BLOCK_SIZE bsize, PICK_MODE_CONTEXT *ctx) {
2995 VP9_COMMON *const cm = &cpi->common;
2996 TileInfo *const tile_info = &tile_data->tile_info;
2997 MACROBLOCKD *const xd = &x->e_mbd;
2999 set_offsets(cpi, tile_info, x, mi_row, mi_col, bsize);
3000 mbmi = &xd->mi[0].src_mi->mbmi;
3001 mbmi->sb_type = bsize;
3003 if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ && cm->seg.enabled)
3004 if (cyclic_refresh_segment_id_boosted(mbmi->segment_id))
3005 x->rdmult = vp9_cyclic_refresh_get_rdmult(cpi->cyclic_refresh);
3007 if (cm->frame_type == KEY_FRAME)
3008 hybrid_intra_mode_search(cpi, x, rd_cost, bsize, ctx);
3009 else if (vp9_segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_SKIP))
3010 set_mode_info_seg_skip(x, cm->tx_mode, rd_cost, bsize);
3011 else if (bsize >= BLOCK_8X8)
3012 vp9_pick_inter_mode(cpi, x, tile_data, mi_row, mi_col,
3013 rd_cost, bsize, ctx);
3015 vp9_pick_inter_mode_sub8x8(cpi, x, tile_data, mi_row, mi_col,
3016 rd_cost, bsize, ctx);
3018 duplicate_mode_info_in_sb(cm, xd, mi_row, mi_col, bsize);
3020 if (rd_cost->rate == INT_MAX)
3021 vp9_rd_cost_reset(rd_cost);
3023 ctx->rate = rd_cost->rate;
3024 ctx->dist = rd_cost->dist;
3027 static void fill_mode_info_sb(VP9_COMMON *cm, MACROBLOCK *x,
3028 int mi_row, int mi_col,
3031 MACROBLOCKD *xd = &x->e_mbd;
3032 int bsl = b_width_log2_lookup[bsize], hbs = (1 << bsl) / 4;
3033 PARTITION_TYPE partition = pc_tree->partitioning;
3034 BLOCK_SIZE subsize = get_subsize(bsize, partition);
3036 assert(bsize >= BLOCK_8X8);
3038 if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols)
3041 switch (partition) {
3042 case PARTITION_NONE:
3043 set_mode_info_offsets(cm, xd, mi_row, mi_col);
3044 *(xd->mi[0].src_mi) = pc_tree->none.mic;
3045 duplicate_mode_info_in_sb(cm, xd, mi_row, mi_col, bsize);
3047 case PARTITION_VERT:
3048 set_mode_info_offsets(cm, xd, mi_row, mi_col);
3049 *(xd->mi[0].src_mi) = pc_tree->vertical[0].mic;
3050 duplicate_mode_info_in_sb(cm, xd, mi_row, mi_col, subsize);
3052 if (mi_col + hbs < cm->mi_cols) {
3053 set_mode_info_offsets(cm, xd, mi_row, mi_col + hbs);
3054 *(xd->mi[0].src_mi) = pc_tree->vertical[1].mic;
3055 duplicate_mode_info_in_sb(cm, xd, mi_row, mi_col + hbs, subsize);
3058 case PARTITION_HORZ:
3059 set_mode_info_offsets(cm, xd, mi_row, mi_col);
3060 *(xd->mi[0].src_mi) = pc_tree->horizontal[0].mic;
3061 duplicate_mode_info_in_sb(cm, xd, mi_row, mi_col, subsize);
3062 if (mi_row + hbs < cm->mi_rows) {
3063 set_mode_info_offsets(cm, xd, mi_row + hbs, mi_col);
3064 *(xd->mi[0].src_mi) = pc_tree->horizontal[1].mic;
3065 duplicate_mode_info_in_sb(cm, xd, mi_row + hbs, mi_col, subsize);
3068 case PARTITION_SPLIT: {
3069 fill_mode_info_sb(cm, x, mi_row, mi_col, subsize, pc_tree->split[0]);
3070 fill_mode_info_sb(cm, x, mi_row, mi_col + hbs, subsize,
3072 fill_mode_info_sb(cm, x, mi_row + hbs, mi_col, subsize,
3074 fill_mode_info_sb(cm, x, mi_row + hbs, mi_col + hbs, subsize,
3083 // Reset the prediction pixel ready flag recursively.
3084 static void pred_pixel_ready_reset(PC_TREE *pc_tree, BLOCK_SIZE bsize) {
3085 pc_tree->none.pred_pixel_ready = 0;
3086 pc_tree->horizontal[0].pred_pixel_ready = 0;
3087 pc_tree->horizontal[1].pred_pixel_ready = 0;
3088 pc_tree->vertical[0].pred_pixel_ready = 0;
3089 pc_tree->vertical[1].pred_pixel_ready = 0;
3091 if (bsize > BLOCK_8X8) {
3092 BLOCK_SIZE subsize = get_subsize(bsize, PARTITION_SPLIT);
3094 for (i = 0; i < 4; ++i)
3095 pred_pixel_ready_reset(pc_tree->split[i], subsize);
3099 static void nonrd_pick_partition(VP9_COMP *cpi, ThreadData *td,
3100 TileDataEnc *tile_data,
3101 TOKENEXTRA **tp, int mi_row,
3102 int mi_col, BLOCK_SIZE bsize, RD_COST *rd_cost,
3103 int do_recon, int64_t best_rd,
3105 const SPEED_FEATURES *const sf = &cpi->sf;
3106 VP9_COMMON *const cm = &cpi->common;
3107 TileInfo *const tile_info = &tile_data->tile_info;
3108 MACROBLOCK *const x = &td->mb;
3109 MACROBLOCKD *const xd = &x->e_mbd;
3110 const int ms = num_8x8_blocks_wide_lookup[bsize] / 2;
3111 TOKENEXTRA *tp_orig = *tp;
3112 PICK_MODE_CONTEXT *ctx = &pc_tree->none;
3114 BLOCK_SIZE subsize = bsize;
3115 RD_COST this_rdc, sum_rdc, best_rdc;
3116 int do_split = bsize >= BLOCK_8X8;
3118 // Override skipping rectangular partition operations for edge blocks
3119 const int force_horz_split = (mi_row + ms >= cm->mi_rows);
3120 const int force_vert_split = (mi_col + ms >= cm->mi_cols);
3121 const int xss = x->e_mbd.plane[1].subsampling_x;
3122 const int yss = x->e_mbd.plane[1].subsampling_y;
3124 int partition_none_allowed = !force_horz_split && !force_vert_split;
3125 int partition_horz_allowed = !force_vert_split && yss <= xss &&
3127 int partition_vert_allowed = !force_horz_split && xss <= yss &&
3131 assert(num_8x8_blocks_wide_lookup[bsize] ==
3132 num_8x8_blocks_high_lookup[bsize]);
3134 vp9_rd_cost_init(&sum_rdc);
3135 vp9_rd_cost_reset(&best_rdc);
3136 best_rdc.rdcost = best_rd;
3138 // Determine partition types in search according to the speed features.
3139 // The threshold set here has to be of square block size.
3140 if (sf->auto_min_max_partition_size) {
3141 partition_none_allowed &= (bsize <= x->max_partition_size &&
3142 bsize >= x->min_partition_size);
3143 partition_horz_allowed &= ((bsize <= x->max_partition_size &&
3144 bsize > x->min_partition_size) ||
3146 partition_vert_allowed &= ((bsize <= x->max_partition_size &&
3147 bsize > x->min_partition_size) ||
3149 do_split &= bsize > x->min_partition_size;
3151 if (sf->use_square_partition_only) {
3152 partition_horz_allowed &= force_horz_split;
3153 partition_vert_allowed &= force_vert_split;
3156 ctx->pred_pixel_ready = !(partition_vert_allowed ||
3157 partition_horz_allowed ||
3161 if (partition_none_allowed) {
3162 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col,
3163 &this_rdc, bsize, ctx);
3164 ctx->mic.mbmi = xd->mi[0].src_mi->mbmi;
3165 ctx->skip_txfm[0] = x->skip_txfm[0];
3166 ctx->skip = x->skip;
3168 if (this_rdc.rate != INT_MAX) {
3169 int pl = partition_plane_context(xd, mi_row, mi_col, bsize);
3170 this_rdc.rate += cpi->partition_cost[pl][PARTITION_NONE];
3171 this_rdc.rdcost = RDCOST(x->rdmult, x->rddiv,
3172 this_rdc.rate, this_rdc.dist);
3173 if (this_rdc.rdcost < best_rdc.rdcost) {
3174 int64_t dist_breakout_thr = sf->partition_search_breakout_dist_thr;
3175 int64_t rate_breakout_thr = sf->partition_search_breakout_rate_thr;
3177 dist_breakout_thr >>= 8 - (b_width_log2_lookup[bsize] +
3178 b_height_log2_lookup[bsize]);
3180 rate_breakout_thr *= num_pels_log2_lookup[bsize];
3182 best_rdc = this_rdc;
3183 if (bsize >= BLOCK_8X8)
3184 pc_tree->partitioning = PARTITION_NONE;
3186 if (!x->e_mbd.lossless &&
3187 this_rdc.rate < rate_breakout_thr &&
3188 this_rdc.dist < dist_breakout_thr) {
3196 // store estimated motion vector
3197 store_pred_mv(x, ctx);
3201 int pl = partition_plane_context(xd, mi_row, mi_col, bsize);
3202 sum_rdc.rate += cpi->partition_cost[pl][PARTITION_SPLIT];
3203 sum_rdc.rdcost = RDCOST(x->rdmult, x->rddiv, sum_rdc.rate, sum_rdc.dist);
3204 subsize = get_subsize(bsize, PARTITION_SPLIT);
3205 for (i = 0; i < 4 && sum_rdc.rdcost < best_rdc.rdcost; ++i) {
3206 const int x_idx = (i & 1) * ms;
3207 const int y_idx = (i >> 1) * ms;
3209 if (mi_row + y_idx >= cm->mi_rows || mi_col + x_idx >= cm->mi_cols)
3211 load_pred_mv(x, ctx);
3212 nonrd_pick_partition(cpi, td, tile_data, tp,
3213 mi_row + y_idx, mi_col + x_idx,
3214 subsize, &this_rdc, 0,
3215 best_rdc.rdcost - sum_rdc.rdcost, pc_tree->split[i]);
3217 if (this_rdc.rate == INT_MAX) {
3218 vp9_rd_cost_reset(&sum_rdc);
3220 sum_rdc.rate += this_rdc.rate;
3221 sum_rdc.dist += this_rdc.dist;
3222 sum_rdc.rdcost += this_rdc.rdcost;
3226 if (sum_rdc.rdcost < best_rdc.rdcost) {
3228 pc_tree->partitioning = PARTITION_SPLIT;
3230 // skip rectangular partition test when larger block size
3231 // gives better rd cost
3232 if (sf->less_rectangular_check)
3233 do_rect &= !partition_none_allowed;
3238 if (partition_horz_allowed && do_rect) {
3239 subsize = get_subsize(bsize, PARTITION_HORZ);
3240 if (sf->adaptive_motion_search)
3241 load_pred_mv(x, ctx);
3242 pc_tree->horizontal[0].pred_pixel_ready = 1;
3243 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &sum_rdc, subsize,
3244 &pc_tree->horizontal[0]);
3246 pc_tree->horizontal[0].mic.mbmi = xd->mi[0].src_mi->mbmi;
3247 pc_tree->horizontal[0].skip_txfm[0] = x->skip_txfm[0];
3248 pc_tree->horizontal[0].skip = x->skip;
3250 if (sum_rdc.rdcost < best_rdc.rdcost && mi_row + ms < cm->mi_rows) {
3251 load_pred_mv(x, ctx);
3252 pc_tree->horizontal[1].pred_pixel_ready = 1;
3253 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row + ms, mi_col,
3255 &pc_tree->horizontal[1]);
3257 pc_tree->horizontal[1].mic.mbmi = xd->mi[0].src_mi->mbmi;
3258 pc_tree->horizontal[1].skip_txfm[0] = x->skip_txfm[0];
3259 pc_tree->horizontal[1].skip = x->skip;
3261 if (this_rdc.rate == INT_MAX) {
3262 vp9_rd_cost_reset(&sum_rdc);
3264 int pl = partition_plane_context(xd, mi_row, mi_col, bsize);
3265 this_rdc.rate += cpi->partition_cost[pl][PARTITION_HORZ];
3266 sum_rdc.rate += this_rdc.rate;
3267 sum_rdc.dist += this_rdc.dist;
3268 sum_rdc.rdcost = RDCOST(x->rdmult, x->rddiv,
3269 sum_rdc.rate, sum_rdc.dist);
3273 if (sum_rdc.rdcost < best_rdc.rdcost) {
3275 pc_tree->partitioning = PARTITION_HORZ;
3277 pred_pixel_ready_reset(pc_tree, bsize);
3282 if (partition_vert_allowed && do_rect) {
3283 subsize = get_subsize(bsize, PARTITION_VERT);
3284 if (sf->adaptive_motion_search)
3285 load_pred_mv(x, ctx);
3286 pc_tree->vertical[0].pred_pixel_ready = 1;
3287 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &sum_rdc, subsize,
3288 &pc_tree->vertical[0]);
3289 pc_tree->vertical[0].mic.mbmi = xd->mi[0].src_mi->mbmi;
3290 pc_tree->vertical[0].skip_txfm[0] = x->skip_txfm[0];
3291 pc_tree->vertical[0].skip = x->skip;
3293 if (sum_rdc.rdcost < best_rdc.rdcost && mi_col + ms < cm->mi_cols) {
3294 load_pred_mv(x, ctx);
3295 pc_tree->vertical[1].pred_pixel_ready = 1;
3296 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col + ms,
3298 &pc_tree->vertical[1]);
3299 pc_tree->vertical[1].mic.mbmi = xd->mi[0].src_mi->mbmi;
3300 pc_tree->vertical[1].skip_txfm[0] = x->skip_txfm[0];
3301 pc_tree->vertical[1].skip = x->skip;
3303 if (this_rdc.rate == INT_MAX) {
3304 vp9_rd_cost_reset(&sum_rdc);
3306 int pl = partition_plane_context(xd, mi_row, mi_col, bsize);
3307 sum_rdc.rate += cpi->partition_cost[pl][PARTITION_VERT];
3308 sum_rdc.rate += this_rdc.rate;
3309 sum_rdc.dist += this_rdc.dist;
3310 sum_rdc.rdcost = RDCOST(x->rdmult, x->rddiv,
3311 sum_rdc.rate, sum_rdc.dist);
3315 if (sum_rdc.rdcost < best_rdc.rdcost) {
3317 pc_tree->partitioning = PARTITION_VERT;
3319 pred_pixel_ready_reset(pc_tree, bsize);
3323 *rd_cost = best_rdc;
3325 if (best_rdc.rate == INT_MAX) {
3326 vp9_rd_cost_reset(rd_cost);
3330 // update mode info array
3331 fill_mode_info_sb(cm, x, mi_row, mi_col, bsize, pc_tree);
3333 if (best_rdc.rate < INT_MAX && best_rdc.dist < INT64_MAX && do_recon) {
3334 int output_enabled = (bsize == BLOCK_64X64);
3335 encode_sb_rt(cpi, td, tile_info, tp, mi_row, mi_col, output_enabled,
3339 if (bsize == BLOCK_64X64 && do_recon) {
3340 assert(tp_orig < *tp);
3341 assert(best_rdc.rate < INT_MAX);
3342 assert(best_rdc.dist < INT64_MAX);
3344 assert(tp_orig == *tp);
3348 static void nonrd_select_partition(VP9_COMP *cpi,
3350 TileDataEnc *tile_data,
3353 int mi_row, int mi_col,
3354 BLOCK_SIZE bsize, int output_enabled,
3355 RD_COST *rd_cost, PC_TREE *pc_tree) {
3356 VP9_COMMON *const cm = &cpi->common;
3357 TileInfo *const tile_info = &tile_data->tile_info;
3358 MACROBLOCK *const x = &td->mb;
3359 MACROBLOCKD *const xd = &x->e_mbd;
3360 const int bsl = b_width_log2_lookup[bsize], hbs = (1 << bsl) / 4;
3361 const int mis = cm->mi_stride;
3362 PARTITION_TYPE partition;
3366 vp9_rd_cost_reset(&this_rdc);
3367 if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols)
3370 subsize = (bsize >= BLOCK_8X8) ? mi[0].src_mi->mbmi.sb_type : BLOCK_4X4;
3371 partition = partition_lookup[bsl][subsize];
3373 if (bsize == BLOCK_32X32 && partition != PARTITION_NONE &&
3374 subsize >= BLOCK_16X16) {
3375 x->max_partition_size = BLOCK_32X32;
3376 x->min_partition_size = BLOCK_8X8;
3377 nonrd_pick_partition(cpi, td, tile_data, tp, mi_row, mi_col, bsize,
3378 rd_cost, 0, INT64_MAX, pc_tree);
3379 } else if (bsize == BLOCK_16X16 && partition != PARTITION_NONE) {
3380 x->max_partition_size = BLOCK_16X16;
3381 x->min_partition_size = BLOCK_8X8;
3382 nonrd_pick_partition(cpi, td, tile_data, tp, mi_row, mi_col, bsize,
3383 rd_cost, 0, INT64_MAX, pc_tree);
3385 switch (partition) {
3386 case PARTITION_NONE:
3387 pc_tree->none.pred_pixel_ready = 1;
3388 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, rd_cost,
3389 subsize, &pc_tree->none);
3390 pc_tree->none.mic.mbmi = xd->mi[0].src_mi->mbmi;
3391 pc_tree->none.skip_txfm[0] = x->skip_txfm[0];
3392 pc_tree->none.skip = x->skip;
3394 case PARTITION_VERT:
3395 pc_tree->vertical[0].pred_pixel_ready = 1;
3396 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, rd_cost,
3397 subsize, &pc_tree->vertical[0]);
3398 pc_tree->vertical[0].mic.mbmi = xd->mi[0].src_mi->mbmi;
3399 pc_tree->vertical[0].skip_txfm[0] = x->skip_txfm[0];
3400 pc_tree->vertical[0].skip = x->skip;
3401 if (mi_col + hbs < cm->mi_cols) {
3402 pc_tree->vertical[1].pred_pixel_ready = 1;
3403 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col + hbs,
3404 &this_rdc, subsize, &pc_tree->vertical[1]);
3405 pc_tree->vertical[1].mic.mbmi = xd->mi[0].src_mi->mbmi;
3406 pc_tree->vertical[1].skip_txfm[0] = x->skip_txfm[0];
3407 pc_tree->vertical[1].skip = x->skip;
3408 if (this_rdc.rate != INT_MAX && this_rdc.dist != INT64_MAX &&
3409 rd_cost->rate != INT_MAX && rd_cost->dist != INT64_MAX) {
3410 rd_cost->rate += this_rdc.rate;
3411 rd_cost->dist += this_rdc.dist;
3415 case PARTITION_HORZ:
3416 pc_tree->horizontal[0].pred_pixel_ready = 1;
3417 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, rd_cost,
3418 subsize, &pc_tree->horizontal[0]);
3419 pc_tree->horizontal[0].mic.mbmi = xd->mi[0].src_mi->mbmi;
3420 pc_tree->horizontal[0].skip_txfm[0] = x->skip_txfm[0];
3421 pc_tree->horizontal[0].skip = x->skip;
3422 if (mi_row + hbs < cm->mi_rows) {
3423 pc_tree->horizontal[1].pred_pixel_ready = 1;
3424 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row + hbs, mi_col,
3425 &this_rdc, subsize, &pc_tree->horizontal[1]);
3426 pc_tree->horizontal[1].mic.mbmi = xd->mi[0].src_mi->mbmi;
3427 pc_tree->horizontal[1].skip_txfm[0] = x->skip_txfm[0];
3428 pc_tree->horizontal[1].skip = x->skip;
3429 if (this_rdc.rate != INT_MAX && this_rdc.dist != INT64_MAX &&
3430 rd_cost->rate != INT_MAX && rd_cost->dist != INT64_MAX) {
3431 rd_cost->rate += this_rdc.rate;
3432 rd_cost->dist += this_rdc.dist;
3436 case PARTITION_SPLIT:
3437 subsize = get_subsize(bsize, PARTITION_SPLIT);
3438 nonrd_select_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col,
3439 subsize, output_enabled, rd_cost,
3441 nonrd_select_partition(cpi, td, tile_data, mi + hbs, tp,
3442 mi_row, mi_col + hbs, subsize, output_enabled,
3443 &this_rdc, pc_tree->split[1]);
3444 if (this_rdc.rate != INT_MAX && this_rdc.dist != INT64_MAX &&
3445 rd_cost->rate != INT_MAX && rd_cost->dist != INT64_MAX) {
3446 rd_cost->rate += this_rdc.rate;
3447 rd_cost->dist += this_rdc.dist;
3449 nonrd_select_partition(cpi, td, tile_data, mi + hbs * mis, tp,
3450 mi_row + hbs, mi_col, subsize, output_enabled,
3451 &this_rdc, pc_tree->split[2]);
3452 if (this_rdc.rate != INT_MAX && this_rdc.dist != INT64_MAX &&
3453 rd_cost->rate != INT_MAX && rd_cost->dist != INT64_MAX) {
3454 rd_cost->rate += this_rdc.rate;
3455 rd_cost->dist += this_rdc.dist;
3457 nonrd_select_partition(cpi, td, tile_data, mi + hbs * mis + hbs, tp,
3458 mi_row + hbs, mi_col + hbs, subsize,
3459 output_enabled, &this_rdc, pc_tree->split[3]);
3460 if (this_rdc.rate != INT_MAX && this_rdc.dist != INT64_MAX &&
3461 rd_cost->rate != INT_MAX && rd_cost->dist != INT64_MAX) {
3462 rd_cost->rate += this_rdc.rate;
3463 rd_cost->dist += this_rdc.dist;
3467 assert(0 && "Invalid partition type.");
3472 if (bsize == BLOCK_64X64 && output_enabled)
3473 encode_sb_rt(cpi, td, tile_info, tp, mi_row, mi_col, 1, bsize, pc_tree);
3477 static void nonrd_use_partition(VP9_COMP *cpi,
3479 TileDataEnc *tile_data,
3482 int mi_row, int mi_col,
3483 BLOCK_SIZE bsize, int output_enabled,
3484 RD_COST *dummy_cost, PC_TREE *pc_tree) {
3485 VP9_COMMON *const cm = &cpi->common;
3486 TileInfo *tile_info = &tile_data->tile_info;
3487 MACROBLOCK *const x = &td->mb;
3488 MACROBLOCKD *const xd = &x->e_mbd;
3489 const int bsl = b_width_log2_lookup[bsize], hbs = (1 << bsl) / 4;
3490 const int mis = cm->mi_stride;
3491 PARTITION_TYPE partition;
3494 if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols)
3497 subsize = (bsize >= BLOCK_8X8) ? mi[0].src_mi->mbmi.sb_type : BLOCK_4X4;
3498 partition = partition_lookup[bsl][subsize];
3500 if (output_enabled && bsize != BLOCK_4X4) {
3501 int ctx = partition_plane_context(xd, mi_row, mi_col, bsize);
3502 td->counts->partition[ctx][partition]++;
3505 switch (partition) {
3506 case PARTITION_NONE:
3507 pc_tree->none.pred_pixel_ready = 1;
3508 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, dummy_cost,
3509 subsize, &pc_tree->none);
3510 pc_tree->none.mic.mbmi = xd->mi[0].src_mi->mbmi;
3511 pc_tree->none.skip_txfm[0] = x->skip_txfm[0];
3512 pc_tree->none.skip = x->skip;
3513 encode_b_rt(cpi, td, tile_info, tp, mi_row, mi_col, output_enabled,
3514 subsize, &pc_tree->none);
3516 case PARTITION_VERT:
3517 pc_tree->vertical[0].pred_pixel_ready = 1;
3518 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, dummy_cost,
3519 subsize, &pc_tree->vertical[0]);
3520 pc_tree->vertical[0].mic.mbmi = xd->mi[0].src_mi->mbmi;
3521 pc_tree->vertical[0].skip_txfm[0] = x->skip_txfm[0];
3522 pc_tree->vertical[0].skip = x->skip;
3523 encode_b_rt(cpi, td, tile_info, tp, mi_row, mi_col, output_enabled,
3524 subsize, &pc_tree->vertical[0]);
3525 if (mi_col + hbs < cm->mi_cols && bsize > BLOCK_8X8) {
3526 pc_tree->vertical[1].pred_pixel_ready = 1;
3527 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col + hbs,
3528 dummy_cost, subsize, &pc_tree->vertical[1]);
3529 pc_tree->vertical[1].mic.mbmi = xd->mi[0].src_mi->mbmi;
3530 pc_tree->vertical[1].skip_txfm[0] = x->skip_txfm[0];
3531 pc_tree->vertical[1].skip = x->skip;
3532 encode_b_rt(cpi, td, tile_info, tp, mi_row, mi_col + hbs,
3533 output_enabled, subsize, &pc_tree->vertical[1]);
3536 case PARTITION_HORZ:
3537 pc_tree->horizontal[0].pred_pixel_ready = 1;
3538 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, dummy_cost,
3539 subsize, &pc_tree->horizontal[0]);
3540 pc_tree->horizontal[0].mic.mbmi = xd->mi[0].src_mi->mbmi;
3541 pc_tree->horizontal[0].skip_txfm[0] = x->skip_txfm[0];
3542 pc_tree->horizontal[0].skip = x->skip;
3543 encode_b_rt(cpi, td, tile_info, tp, mi_row, mi_col, output_enabled,
3544 subsize, &pc_tree->horizontal[0]);
3546 if (mi_row + hbs < cm->mi_rows && bsize > BLOCK_8X8) {
3547 pc_tree->horizontal[1].pred_pixel_ready = 1;
3548 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row + hbs, mi_col,
3549 dummy_cost, subsize, &pc_tree->horizontal[1]);
3550 pc_tree->horizontal[1].mic.mbmi = xd->mi[0].src_mi->mbmi;
3551 pc_tree->horizontal[1].skip_txfm[0] = x->skip_txfm[0];
3552 pc_tree->horizontal[1].skip = x->skip;
3553 encode_b_rt(cpi, td, tile_info, tp, mi_row + hbs, mi_col,
3554 output_enabled, subsize, &pc_tree->horizontal[1]);
3557 case PARTITION_SPLIT:
3558 subsize = get_subsize(bsize, PARTITION_SPLIT);
3559 if (bsize == BLOCK_8X8) {
3560 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, dummy_cost,
3561 subsize, pc_tree->leaf_split[0]);
3562 encode_b_rt(cpi, td, tile_info, tp, mi_row, mi_col,
3563 output_enabled, subsize, pc_tree->leaf_split[0]);
3565 nonrd_use_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col,
3566 subsize, output_enabled, dummy_cost,
3568 nonrd_use_partition(cpi, td, tile_data, mi + hbs, tp,
3569 mi_row, mi_col + hbs, subsize, output_enabled,
3570 dummy_cost, pc_tree->split[1]);
3571 nonrd_use_partition(cpi, td, tile_data, mi + hbs * mis, tp,
3572 mi_row + hbs, mi_col, subsize, output_enabled,
3573 dummy_cost, pc_tree->split[2]);
3574 nonrd_use_partition(cpi, td, tile_data, mi + hbs * mis + hbs, tp,
3575 mi_row + hbs, mi_col + hbs, subsize, output_enabled,
3576 dummy_cost, pc_tree->split[3]);
3580 assert(0 && "Invalid partition type.");
3584 if (partition != PARTITION_SPLIT || bsize == BLOCK_8X8)
3585 update_partition_context(xd, mi_row, mi_col, subsize, bsize);
3588 static void encode_nonrd_sb_row(VP9_COMP *cpi,
3590 TileDataEnc *tile_data,
3593 SPEED_FEATURES *const sf = &cpi->sf;
3594 VP9_COMMON *const cm = &cpi->common;
3595 TileInfo *const tile_info = &tile_data->tile_info;
3596 MACROBLOCK *const x = &td->mb;
3597 MACROBLOCKD *const xd = &x->e_mbd;
3600 // Initialize the left context for the new SB row
3601 vpx_memset(&xd->left_context, 0, sizeof(xd->left_context));
3602 vpx_memset(xd->left_seg_context, 0, sizeof(xd->left_seg_context));
3604 // Code each SB in the row
3605 for (mi_col = tile_info->mi_col_start; mi_col < tile_info->mi_col_end;
3606 mi_col += MI_BLOCK_SIZE) {
3607 const struct segmentation *const seg = &cm->seg;
3609 const int idx_str = cm->mi_stride * mi_row + mi_col;
3610 MODE_INFO *mi = cm->mi + idx_str;
3611 PARTITION_SEARCH_TYPE partition_search_type = sf->partition_search_type;
3612 BLOCK_SIZE bsize = BLOCK_64X64;
3614 x->source_variance = UINT_MAX;
3615 vp9_zero(x->pred_mv);
3616 vp9_rd_cost_init(&dummy_rdc);
3617 x->color_sensitivity[0] = 0;
3618 x->color_sensitivity[1] = 0;
3621 const uint8_t *const map = seg->update_map ? cpi->segmentation_map
3622 : cm->last_frame_seg_map;
3623 int segment_id = vp9_get_segment_id(cm, map, BLOCK_64X64, mi_row, mi_col);
3624 seg_skip = vp9_segfeature_active(seg, segment_id, SEG_LVL_SKIP);
3626 partition_search_type = FIXED_PARTITION;
3630 // Set the partition type of the 64X64 block
3631 switch (partition_search_type) {
3632 case VAR_BASED_PARTITION:
3633 // TODO(jingning, marpan): The mode decision and encoding process
3634 // support both intra and inter sub8x8 block coding for RTC mode.
3635 // Tune the thresholds accordingly to use sub8x8 block coding for
3636 // coding performance improvement.
3637 choose_partitioning(cpi, tile_info, x, mi_row, mi_col);
3638 nonrd_use_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col,
3639 BLOCK_64X64, 1, &dummy_rdc, td->pc_root);
3641 case SOURCE_VAR_BASED_PARTITION:
3642 set_source_var_based_partition(cpi, tile_info, x, mi, mi_row, mi_col);
3643 nonrd_use_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col,
3644 BLOCK_64X64, 1, &dummy_rdc, td->pc_root);
3646 case FIXED_PARTITION:
3648 bsize = sf->always_this_block_size;
3649 set_fixed_partitioning(cpi, tile_info, mi, mi_row, mi_col, bsize);
3650 nonrd_use_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col,
3651 BLOCK_64X64, 1, &dummy_rdc, td->pc_root);
3653 case REFERENCE_PARTITION:
3654 set_offsets(cpi, tile_info, x, mi_row, mi_col, BLOCK_64X64);
3655 if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ && cm->seg.enabled &&
3656 xd->mi[0].src_mi->mbmi.segment_id) {
3657 x->max_partition_size = BLOCK_64X64;
3658 x->min_partition_size = BLOCK_8X8;
3659 nonrd_pick_partition(cpi, td, tile_data, tp, mi_row, mi_col,
3660 BLOCK_64X64, &dummy_rdc, 1,
3661 INT64_MAX, td->pc_root);
3663 choose_partitioning(cpi, tile_info, x, mi_row, mi_col);
3664 nonrd_select_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col,
3665 BLOCK_64X64, 1, &dummy_rdc, td->pc_root);
3675 // end RTC play code
3677 static int set_var_thresh_from_histogram(VP9_COMP *cpi) {
3678 const SPEED_FEATURES *const sf = &cpi->sf;
3679 const VP9_COMMON *const cm = &cpi->common;
3681 const uint8_t *src = cpi->Source->y_buffer;
3682 const uint8_t *last_src = cpi->Last_Source->y_buffer;
3683 const int src_stride = cpi->Source->y_stride;
3684 const int last_stride = cpi->Last_Source->y_stride;
3686 // Pick cutoff threshold
3687 const int cutoff = (MIN(cm->width, cm->height) >= 720) ?
3688 (cm->MBs * VAR_HIST_LARGE_CUT_OFF / 100) :
3689 (cm->MBs * VAR_HIST_SMALL_CUT_OFF / 100);
3690 DECLARE_ALIGNED_ARRAY(16, int, hist, VAR_HIST_BINS);
3691 diff *var16 = cpi->source_diff_var;
3696 vpx_memset(hist, 0, VAR_HIST_BINS * sizeof(hist[0]));
3698 for (i = 0; i < cm->mb_rows; i++) {
3699 for (j = 0; j < cm->mb_cols; j++) {
3700 #if CONFIG_VP9_HIGHBITDEPTH
3701 if (cm->use_highbitdepth) {
3702 switch (cm->bit_depth) {
3704 vp9_highbd_get16x16var(src, src_stride, last_src, last_stride,
3705 &var16->sse, &var16->sum);
3708 vp9_highbd_10_get16x16var(src, src_stride, last_src, last_stride,
3709 &var16->sse, &var16->sum);
3712 vp9_highbd_12_get16x16var(src, src_stride, last_src, last_stride,
3713 &var16->sse, &var16->sum);
3716 assert(0 && "cm->bit_depth should be VPX_BITS_8, VPX_BITS_10"
3721 vp9_get16x16var(src, src_stride, last_src, last_stride,
3722 &var16->sse, &var16->sum);
3725 vp9_get16x16var(src, src_stride, last_src, last_stride,
3726 &var16->sse, &var16->sum);
3727 #endif // CONFIG_VP9_HIGHBITDEPTH
3728 var16->var = var16->sse -
3729 (((uint32_t)var16->sum * var16->sum) >> 8);
3731 if (var16->var >= VAR_HIST_MAX_BG_VAR)
3732 hist[VAR_HIST_BINS - 1]++;
3734 hist[var16->var / VAR_HIST_FACTOR]++;
3741 src = src - cm->mb_cols * 16 + 16 * src_stride;
3742 last_src = last_src - cm->mb_cols * 16 + 16 * last_stride;
3745 cpi->source_var_thresh = 0;
3747 if (hist[VAR_HIST_BINS - 1] < cutoff) {
3748 for (i = 0; i < VAR_HIST_BINS - 1; i++) {
3752 cpi->source_var_thresh = (i + 1) * VAR_HIST_FACTOR;
3758 return sf->search_type_check_frequency;
3761 static void source_var_based_partition_search_method(VP9_COMP *cpi) {
3762 VP9_COMMON *const cm = &cpi->common;
3763 SPEED_FEATURES *const sf = &cpi->sf;
3765 if (cm->frame_type == KEY_FRAME) {
3766 // For key frame, use SEARCH_PARTITION.
3767 sf->partition_search_type = SEARCH_PARTITION;
3768 } else if (cm->intra_only) {
3769 sf->partition_search_type = FIXED_PARTITION;
3771 if (cm->last_width != cm->width || cm->last_height != cm->height) {
3772 if (cpi->source_diff_var)
3773 vpx_free(cpi->source_diff_var);
3775 CHECK_MEM_ERROR(cm, cpi->source_diff_var,
3776 vpx_calloc(cm->MBs, sizeof(diff)));
3779 if (!cpi->frames_till_next_var_check)
3780 cpi->frames_till_next_var_check = set_var_thresh_from_histogram(cpi);
3782 if (cpi->frames_till_next_var_check > 0) {
3783 sf->partition_search_type = FIXED_PARTITION;
3784 cpi->frames_till_next_var_check--;
3789 static int get_skip_encode_frame(const VP9_COMMON *cm, ThreadData *const td) {
3790 unsigned int intra_count = 0, inter_count = 0;
3793 for (j = 0; j < INTRA_INTER_CONTEXTS; ++j) {
3794 intra_count += td->counts->intra_inter[j][0];
3795 inter_count += td->counts->intra_inter[j][1];
3798 return (intra_count << 2) < inter_count &&
3799 cm->frame_type != KEY_FRAME &&
3803 void vp9_init_tile_data(VP9_COMP *cpi) {
3804 VP9_COMMON *const cm = &cpi->common;
3805 const int tile_cols = 1 << cm->log2_tile_cols;
3806 const int tile_rows = 1 << cm->log2_tile_rows;
3807 int tile_col, tile_row;
3808 TOKENEXTRA *pre_tok = cpi->tile_tok[0][0];
3811 if (cpi->tile_data == NULL) {
3812 CHECK_MEM_ERROR(cm, cpi->tile_data,
3813 vpx_malloc(tile_cols * tile_rows * sizeof(*cpi->tile_data)));
3814 for (tile_row = 0; tile_row < tile_rows; ++tile_row)
3815 for (tile_col = 0; tile_col < tile_cols; ++tile_col) {
3816 TileDataEnc *tile_data =
3817 &cpi->tile_data[tile_row * tile_cols + tile_col];
3819 for (i = 0; i < BLOCK_SIZES; ++i) {
3820 for (j = 0; j < MAX_MODES; ++j) {
3821 tile_data->thresh_freq_fact[i][j] = 32;
3822 tile_data->mode_map[i][j] = j;
3828 for (tile_row = 0; tile_row < tile_rows; ++tile_row) {
3829 for (tile_col = 0; tile_col < tile_cols; ++tile_col) {
3830 TileInfo *tile_info =
3831 &cpi->tile_data[tile_row * tile_cols + tile_col].tile_info;
3832 vp9_tile_init(tile_info, cm, tile_row, tile_col);
3834 cpi->tile_tok[tile_row][tile_col] = pre_tok + tile_tok;
3835 pre_tok = cpi->tile_tok[tile_row][tile_col];
3836 tile_tok = allocated_tokens(*tile_info);
3841 void vp9_encode_tile(VP9_COMP *cpi, ThreadData *td,
3842 int tile_row, int tile_col) {
3843 VP9_COMMON *const cm = &cpi->common;
3844 const int tile_cols = 1 << cm->log2_tile_cols;
3845 TileDataEnc *this_tile =
3846 &cpi->tile_data[tile_row * tile_cols + tile_col];
3847 const TileInfo * const tile_info = &this_tile->tile_info;
3848 TOKENEXTRA *tok = cpi->tile_tok[tile_row][tile_col];
3851 for (mi_row = tile_info->mi_row_start; mi_row < tile_info->mi_row_end;
3852 mi_row += MI_BLOCK_SIZE) {
3853 if (cpi->sf.use_nonrd_pick_mode)
3854 encode_nonrd_sb_row(cpi, td, this_tile, mi_row, &tok);
3856 encode_rd_sb_row(cpi, td, this_tile, mi_row, &tok);
3858 cpi->tok_count[tile_row][tile_col] =
3859 (unsigned int)(tok - cpi->tile_tok[tile_row][tile_col]);
3860 assert(tok - cpi->tile_tok[tile_row][tile_col] <=
3861 allocated_tokens(*tile_info));
3864 static void encode_tiles(VP9_COMP *cpi) {
3865 VP9_COMMON *const cm = &cpi->common;
3866 const int tile_cols = 1 << cm->log2_tile_cols;
3867 const int tile_rows = 1 << cm->log2_tile_rows;
3868 int tile_col, tile_row;
3870 vp9_init_tile_data(cpi);
3872 for (tile_row = 0; tile_row < tile_rows; ++tile_row)
3873 for (tile_col = 0; tile_col < tile_cols; ++tile_col)
3874 vp9_encode_tile(cpi, &cpi->td, tile_row, tile_col);
3877 #if CONFIG_FP_MB_STATS
3878 static int input_fpmb_stats(FIRSTPASS_MB_STATS *firstpass_mb_stats,
3879 VP9_COMMON *cm, uint8_t **this_frame_mb_stats) {
3880 uint8_t *mb_stats_in = firstpass_mb_stats->mb_stats_start +
3881 cm->current_video_frame * cm->MBs * sizeof(uint8_t);
3883 if (mb_stats_in > firstpass_mb_stats->mb_stats_end)
3886 *this_frame_mb_stats = mb_stats_in;
3892 static void encode_frame_internal(VP9_COMP *cpi) {
3893 SPEED_FEATURES *const sf = &cpi->sf;
3894 RD_OPT *const rd_opt = &cpi->rd;
3895 ThreadData *const td = &cpi->td;
3896 MACROBLOCK *const x = &td->mb;
3897 VP9_COMMON *const cm = &cpi->common;
3898 MACROBLOCKD *const xd = &x->e_mbd;
3899 RD_COUNTS *const rdc = &cpi->td.rd_counts;
3902 xd->mi[0].src_mi = &xd->mi[0];
3904 vp9_zero(*td->counts);
3905 vp9_zero(rdc->coef_counts);
3906 vp9_zero(rdc->comp_pred_diff);
3907 vp9_zero(rdc->filter_diff);
3908 vp9_zero(rdc->tx_select_diff);
3909 vp9_zero(rd_opt->tx_select_threshes);
3911 xd->lossless = cm->base_qindex == 0 &&
3912 cm->y_dc_delta_q == 0 &&
3913 cm->uv_dc_delta_q == 0 &&
3914 cm->uv_ac_delta_q == 0;
3916 #if CONFIG_VP9_HIGHBITDEPTH
3917 if (cm->use_highbitdepth)
3918 x->fwd_txm4x4 = xd->lossless ? vp9_highbd_fwht4x4 : vp9_highbd_fdct4x4;
3920 x->fwd_txm4x4 = xd->lossless ? vp9_fwht4x4 : vp9_fdct4x4;
3921 x->highbd_itxm_add = xd->lossless ? vp9_highbd_iwht4x4_add :
3922 vp9_highbd_idct4x4_add;
3924 x->fwd_txm4x4 = xd->lossless ? vp9_fwht4x4 : vp9_fdct4x4;
3925 #endif // CONFIG_VP9_HIGHBITDEPTH
3926 x->itxm_add = xd->lossless ? vp9_iwht4x4_add : vp9_idct4x4_add;
3931 cm->tx_mode = select_tx_mode(cpi, xd);
3933 vp9_frame_init_quantizer(cpi);
3935 vp9_initialize_rd_consts(cpi);
3936 vp9_initialize_me_consts(cpi, x, cm->base_qindex);
3937 init_encode_frame_mb_context(cpi);
3938 cm->use_prev_frame_mvs = !cm->error_resilient_mode &&
3939 cm->width == cm->last_width &&
3940 cm->height == cm->last_height &&
3942 cm->last_show_frame;
3943 // Special case: set prev_mi to NULL when the previous mode info
3944 // context cannot be used.
3945 cm->prev_mi = cm->use_prev_frame_mvs ?
3946 cm->prev_mip + cm->mi_stride + 1 : NULL;
3948 x->quant_fp = cpi->sf.use_quant_fp;
3949 vp9_zero(x->skip_txfm);
3950 if (sf->use_nonrd_pick_mode) {
3951 // Initialize internal buffer pointers for rtc coding, where non-RD
3952 // mode decision is used and hence no buffer pointer swap needed.
3954 struct macroblock_plane *const p = x->plane;
3955 struct macroblockd_plane *const pd = xd->plane;
3956 PICK_MODE_CONTEXT *ctx = &cpi->td.pc_root->none;
3958 for (i = 0; i < MAX_MB_PLANE; ++i) {
3959 p[i].coeff = ctx->coeff_pbuf[i][0];
3960 p[i].qcoeff = ctx->qcoeff_pbuf[i][0];
3961 pd[i].dqcoeff = ctx->dqcoeff_pbuf[i][0];
3962 p[i].eobs = ctx->eobs_pbuf[i][0];
3964 vp9_zero(x->zcoeff_blk);
3966 if (cm->frame_type != KEY_FRAME && cpi->rc.frames_since_golden == 0)
3967 cpi->ref_frame_flags &= (~VP9_GOLD_FLAG);
3969 if (sf->partition_search_type == SOURCE_VAR_BASED_PARTITION)
3970 source_var_based_partition_search_method(cpi);
3974 struct vpx_usec_timer emr_timer;
3975 vpx_usec_timer_start(&emr_timer);
3977 #if CONFIG_FP_MB_STATS
3978 if (cpi->use_fp_mb_stats) {
3979 input_fpmb_stats(&cpi->twopass.firstpass_mb_stats, cm,
3980 &cpi->twopass.this_frame_mb_stats);
3984 // If allowed, encoding tiles in parallel with one thread handling one tile.
3985 if (MIN(cpi->oxcf.max_threads, 1 << cm->log2_tile_cols) > 1)
3986 vp9_encode_tiles_mt(cpi);
3990 vpx_usec_timer_mark(&emr_timer);
3991 cpi->time_encode_sb_row += vpx_usec_timer_elapsed(&emr_timer);
3994 sf->skip_encode_frame = sf->skip_encode_sb ?
3995 get_skip_encode_frame(cm, td) : 0;
3998 // Keep record of the total distortion this time around for future use
3999 cpi->last_frame_distortion = cpi->frame_distortion;
4003 static INTERP_FILTER get_interp_filter(
4004 const int64_t threshes[SWITCHABLE_FILTER_CONTEXTS], int is_alt_ref) {
4006 threshes[EIGHTTAP_SMOOTH] > threshes[EIGHTTAP] &&
4007 threshes[EIGHTTAP_SMOOTH] > threshes[EIGHTTAP_SHARP] &&
4008 threshes[EIGHTTAP_SMOOTH] > threshes[SWITCHABLE - 1]) {
4009 return EIGHTTAP_SMOOTH;
4010 } else if (threshes[EIGHTTAP_SHARP] > threshes[EIGHTTAP] &&
4011 threshes[EIGHTTAP_SHARP] > threshes[SWITCHABLE - 1]) {
4012 return EIGHTTAP_SHARP;
4013 } else if (threshes[EIGHTTAP] > threshes[SWITCHABLE - 1]) {
4020 void vp9_encode_frame(VP9_COMP *cpi) {
4021 VP9_COMMON *const cm = &cpi->common;
4023 // In the longer term the encoder should be generalized to match the
4024 // decoder such that we allow compound where one of the 3 buffers has a
4025 // different sign bias and that buffer is then the fixed ref. However, this
4026 // requires further work in the rd loop. For now the only supported encoder
4027 // side behavior is where the ALT ref buffer has opposite sign bias to
4029 if (!frame_is_intra_only(cm)) {
4030 if ((cm->ref_frame_sign_bias[ALTREF_FRAME] ==
4031 cm->ref_frame_sign_bias[GOLDEN_FRAME]) ||
4032 (cm->ref_frame_sign_bias[ALTREF_FRAME] ==
4033 cm->ref_frame_sign_bias[LAST_FRAME])) {
4034 cpi->allow_comp_inter_inter = 0;
4036 cpi->allow_comp_inter_inter = 1;
4037 cm->comp_fixed_ref = ALTREF_FRAME;
4038 cm->comp_var_ref[0] = LAST_FRAME;
4039 cm->comp_var_ref[1] = GOLDEN_FRAME;
4043 if (cpi->sf.frame_parameter_update) {
4045 RD_OPT *const rd_opt = &cpi->rd;
4046 FRAME_COUNTS *counts = cpi->td.counts;
4047 RD_COUNTS *const rdc = &cpi->td.rd_counts;
4049 // This code does a single RD pass over the whole frame assuming
4050 // either compound, single or hybrid prediction as per whatever has
4051 // worked best for that type of frame in the past.
4052 // It also predicts whether another coding mode would have worked
4053 // better that this coding mode. If that is the case, it remembers
4054 // that for subsequent frames.
4055 // It does the same analysis for transform size selection also.
4056 const MV_REFERENCE_FRAME frame_type = get_frame_type(cpi);
4057 int64_t *const mode_thrs = rd_opt->prediction_type_threshes[frame_type];
4058 int64_t *const filter_thrs = rd_opt->filter_threshes[frame_type];
4059 int *const tx_thrs = rd_opt->tx_select_threshes[frame_type];
4060 const int is_alt_ref = frame_type == ALTREF_FRAME;
4062 /* prediction (compound, single or hybrid) mode selection */
4063 if (is_alt_ref || !cpi->allow_comp_inter_inter)
4064 cm->reference_mode = SINGLE_REFERENCE;
4065 else if (mode_thrs[COMPOUND_REFERENCE] > mode_thrs[SINGLE_REFERENCE] &&
4066 mode_thrs[COMPOUND_REFERENCE] >
4067 mode_thrs[REFERENCE_MODE_SELECT] &&
4068 check_dual_ref_flags(cpi) &&
4069 cpi->static_mb_pct == 100)
4070 cm->reference_mode = COMPOUND_REFERENCE;
4071 else if (mode_thrs[SINGLE_REFERENCE] > mode_thrs[REFERENCE_MODE_SELECT])
4072 cm->reference_mode = SINGLE_REFERENCE;
4074 cm->reference_mode = REFERENCE_MODE_SELECT;
4076 if (cm->interp_filter == SWITCHABLE)
4077 cm->interp_filter = get_interp_filter(filter_thrs, is_alt_ref);
4079 encode_frame_internal(cpi);
4081 for (i = 0; i < REFERENCE_MODES; ++i)
4082 mode_thrs[i] = (mode_thrs[i] + rdc->comp_pred_diff[i] / cm->MBs) / 2;
4084 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i)
4085 filter_thrs[i] = (filter_thrs[i] + rdc->filter_diff[i] / cm->MBs) / 2;
4087 for (i = 0; i < TX_MODES; ++i) {
4088 int64_t pd = rdc->tx_select_diff[i];
4089 if (i == TX_MODE_SELECT)
4090 pd -= RDCOST(cpi->td.mb.rdmult, cpi->td.mb.rddiv, 2048 * (TX_SIZES - 1),
4092 tx_thrs[i] = (tx_thrs[i] + (int)(pd / cm->MBs)) / 2;
4095 if (cm->reference_mode == REFERENCE_MODE_SELECT) {
4096 int single_count_zero = 0;
4097 int comp_count_zero = 0;
4099 for (i = 0; i < COMP_INTER_CONTEXTS; i++) {
4100 single_count_zero += counts->comp_inter[i][0];
4101 comp_count_zero += counts->comp_inter[i][1];
4104 if (comp_count_zero == 0) {
4105 cm->reference_mode = SINGLE_REFERENCE;
4106 vp9_zero(counts->comp_inter);
4107 } else if (single_count_zero == 0) {
4108 cm->reference_mode = COMPOUND_REFERENCE;
4109 vp9_zero(counts->comp_inter);
4113 if (cm->tx_mode == TX_MODE_SELECT) {
4115 int count8x8_lp = 0, count8x8_8x8p = 0;
4116 int count16x16_16x16p = 0, count16x16_lp = 0;
4119 for (i = 0; i < TX_SIZE_CONTEXTS; ++i) {
4120 count4x4 += counts->tx.p32x32[i][TX_4X4];
4121 count4x4 += counts->tx.p16x16[i][TX_4X4];
4122 count4x4 += counts->tx.p8x8[i][TX_4X4];
4124 count8x8_lp += counts->tx.p32x32[i][TX_8X8];
4125 count8x8_lp += counts->tx.p16x16[i][TX_8X8];
4126 count8x8_8x8p += counts->tx.p8x8[i][TX_8X8];
4128 count16x16_16x16p += counts->tx.p16x16[i][TX_16X16];
4129 count16x16_lp += counts->tx.p32x32[i][TX_16X16];
4130 count32x32 += counts->tx.p32x32[i][TX_32X32];
4132 if (count4x4 == 0 && count16x16_lp == 0 && count16x16_16x16p == 0 &&
4134 cm->tx_mode = ALLOW_8X8;
4135 reset_skip_tx_size(cm, TX_8X8);
4136 } else if (count8x8_8x8p == 0 && count16x16_16x16p == 0 &&
4137 count8x8_lp == 0 && count16x16_lp == 0 && count32x32 == 0) {
4138 cm->tx_mode = ONLY_4X4;
4139 reset_skip_tx_size(cm, TX_4X4);
4140 } else if (count8x8_lp == 0 && count16x16_lp == 0 && count4x4 == 0) {
4141 cm->tx_mode = ALLOW_32X32;
4142 } else if (count32x32 == 0 && count8x8_lp == 0 && count4x4 == 0) {
4143 cm->tx_mode = ALLOW_16X16;
4144 reset_skip_tx_size(cm, TX_16X16);
4148 cm->reference_mode = SINGLE_REFERENCE;
4149 encode_frame_internal(cpi);
4153 static void sum_intra_stats(FRAME_COUNTS *counts, const MODE_INFO *mi) {
4154 const PREDICTION_MODE y_mode = mi->mbmi.mode;
4155 const PREDICTION_MODE uv_mode = mi->mbmi.uv_mode;
4156 const BLOCK_SIZE bsize = mi->mbmi.sb_type;
4158 if (bsize < BLOCK_8X8) {
4160 const int num_4x4_w = num_4x4_blocks_wide_lookup[bsize];
4161 const int num_4x4_h = num_4x4_blocks_high_lookup[bsize];
4162 for (idy = 0; idy < 2; idy += num_4x4_h)
4163 for (idx = 0; idx < 2; idx += num_4x4_w)
4164 ++counts->y_mode[0][mi->bmi[idy * 2 + idx].as_mode];
4166 ++counts->y_mode[size_group_lookup[bsize]][y_mode];
4169 ++counts->uv_mode[y_mode][uv_mode];
4172 static void encode_superblock(VP9_COMP *cpi, ThreadData *td,
4173 TOKENEXTRA **t, int output_enabled,
4174 int mi_row, int mi_col, BLOCK_SIZE bsize,
4175 PICK_MODE_CONTEXT *ctx) {
4176 VP9_COMMON *const cm = &cpi->common;
4177 MACROBLOCK *const x = &td->mb;
4178 MACROBLOCKD *const xd = &x->e_mbd;
4179 MODE_INFO *mi_8x8 = xd->mi;
4180 MODE_INFO *mi = mi_8x8;
4181 MB_MODE_INFO *mbmi = &mi->mbmi;
4182 const int seg_skip = vp9_segfeature_active(&cm->seg, mbmi->segment_id,
4184 const int mis = cm->mi_stride;
4185 const int mi_width = num_8x8_blocks_wide_lookup[bsize];
4186 const int mi_height = num_8x8_blocks_high_lookup[bsize];
4188 x->skip_recode = !x->select_tx_size && mbmi->sb_type >= BLOCK_8X8 &&
4189 cpi->oxcf.aq_mode != COMPLEXITY_AQ &&
4190 cpi->oxcf.aq_mode != CYCLIC_REFRESH_AQ &&
4191 cpi->sf.allow_skip_recode;
4193 if (!x->skip_recode && !cpi->sf.use_nonrd_pick_mode)
4194 vpx_memset(x->skip_txfm, 0, sizeof(x->skip_txfm));
4196 x->skip_optimize = ctx->is_coded;
4198 x->use_lp32x32fdct = cpi->sf.use_lp32x32fdct;
4199 x->skip_encode = (!output_enabled && cpi->sf.skip_encode_frame &&
4200 x->q_index < QIDX_SKIP_THRESH);
4205 set_ref_ptrs(cm, xd, mbmi->ref_frame[0], mbmi->ref_frame[1]);
4207 if (!is_inter_block(mbmi)) {
4210 for (plane = 0; plane < MAX_MB_PLANE; ++plane)
4211 vp9_encode_intra_block_plane(x, MAX(bsize, BLOCK_8X8), plane);
4213 sum_intra_stats(td->counts, mi);
4214 vp9_tokenize_sb(cpi, td, t, !output_enabled, MAX(bsize, BLOCK_8X8));
4217 const int is_compound = has_second_ref(mbmi);
4218 for (ref = 0; ref < 1 + is_compound; ++ref) {
4219 YV12_BUFFER_CONFIG *cfg = get_ref_frame_buffer(cpi,
4220 mbmi->ref_frame[ref]);
4221 assert(cfg != NULL);
4222 vp9_setup_pre_planes(xd, ref, cfg, mi_row, mi_col,
4223 &xd->block_refs[ref]->sf);
4225 if (!(cpi->sf.reuse_inter_pred_sby && ctx->pred_pixel_ready) || seg_skip)
4226 vp9_build_inter_predictors_sby(xd, mi_row, mi_col, MAX(bsize, BLOCK_8X8));
4228 vp9_build_inter_predictors_sbuv(xd, mi_row, mi_col, MAX(bsize, BLOCK_8X8));
4230 vp9_encode_sb(x, MAX(bsize, BLOCK_8X8));
4231 vp9_tokenize_sb(cpi, td, t, !output_enabled, MAX(bsize, BLOCK_8X8));
4234 if (output_enabled) {
4235 if (cm->tx_mode == TX_MODE_SELECT &&
4236 mbmi->sb_type >= BLOCK_8X8 &&
4237 !(is_inter_block(mbmi) && (mbmi->skip || seg_skip))) {
4238 ++get_tx_counts(max_txsize_lookup[bsize], vp9_get_tx_size_context(xd),
4239 &td->counts->tx)[mbmi->tx_size];
4243 // The new intra coding scheme requires no change of transform size
4244 if (is_inter_block(&mi->mbmi)) {
4245 tx_size = MIN(tx_mode_to_biggest_tx_size[cm->tx_mode],
4246 max_txsize_lookup[bsize]);
4248 tx_size = (bsize >= BLOCK_8X8) ? mbmi->tx_size : TX_4X4;
4251 for (y = 0; y < mi_height; y++)
4252 for (x = 0; x < mi_width; x++)
4253 if (mi_col + x < cm->mi_cols && mi_row + y < cm->mi_rows)
4254 mi_8x8[mis * y + x].src_mi->mbmi.tx_size = tx_size;
4256 ++td->counts->tx.tx_totals[mbmi->tx_size];
4257 ++td->counts->tx.tx_totals[get_uv_tx_size(mbmi, &xd->plane[1])];