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_grid_visible + idx_str;
176 xd->mi[0] = cm->mi + idx_str;
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]->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] = 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]->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 BLOCK_SIZE subsize = get_subsize(bsize, PARTITION_VERT);
436 get_variance(&vt.part_variances->vert[0]);
437 get_variance(&vt.part_variances->vert[1]);
438 if (vt.part_variances->vert[0].variance < threshold &&
439 vt.part_variances->vert[1].variance < threshold &&
440 get_plane_block_size(subsize, &xd->plane[1]) < BLOCK_INVALID) {
441 set_block_size(cpi, xd, mi_row, mi_col, subsize);
442 set_block_size(cpi, xd, mi_row, mi_col + block_width / 2, subsize);
446 // Check horizontal split.
447 if (mi_col + block_width / 2 < cm->mi_cols) {
448 BLOCK_SIZE subsize = get_subsize(bsize, PARTITION_HORZ);
449 get_variance(&vt.part_variances->horz[0]);
450 get_variance(&vt.part_variances->horz[1]);
451 if (vt.part_variances->horz[0].variance < threshold &&
452 vt.part_variances->horz[1].variance < threshold &&
453 get_plane_block_size(subsize, &xd->plane[1]) < BLOCK_INVALID) {
454 set_block_size(cpi, xd, mi_row, mi_col, subsize);
455 set_block_size(cpi, xd, mi_row + block_height / 2, mi_col, subsize);
465 void vp9_set_vbp_thresholds(VP9_COMP *cpi, int q) {
466 SPEED_FEATURES *const sf = &cpi->sf;
467 if (sf->partition_search_type != VAR_BASED_PARTITION &&
468 sf->partition_search_type != REFERENCE_PARTITION) {
471 VP9_COMMON *const cm = &cpi->common;
472 const int is_key_frame = (cm->frame_type == KEY_FRAME);
473 const int threshold_multiplier = is_key_frame ? 20 : 1;
474 const int64_t threshold_base = (int64_t)(threshold_multiplier *
475 cpi->y_dequant[q][1]);
477 // TODO(marpan): Allow 4x4 partitions for inter-frames.
478 // use_4x4_partition = (variance4x4downsample[i2 + j] == 1);
479 // If 4x4 partition is not used, then 8x8 partition will be selected
480 // if variance of 16x16 block is very high, so use larger threshold
481 // for 16x16 (threshold_bsize_min) in that case.
483 // Array index: 0 - threshold_64x64; 1 - threshold_32x32;
484 // 2 - threshold_16x16; 3 - vbp_threshold_8x8;
486 cpi->vbp_thresholds[0] = threshold_base;
487 cpi->vbp_thresholds[1] = threshold_base >> 2;
488 cpi->vbp_thresholds[2] = threshold_base >> 2;
489 cpi->vbp_thresholds[3] = threshold_base << 2;
490 cpi->vbp_threshold_sad = 0;
491 cpi->vbp_bsize_min = BLOCK_8X8;
493 cpi->vbp_thresholds[1] = threshold_base;
494 if (cm->width <= 352 && cm->height <= 288) {
495 cpi->vbp_thresholds[0] = threshold_base >> 2;
496 cpi->vbp_thresholds[2] = threshold_base << 3;
497 cpi->vbp_threshold_sad = 100;
499 cpi->vbp_thresholds[0] = threshold_base;
500 cpi->vbp_thresholds[1] = (5 * threshold_base) >> 2;
501 cpi->vbp_thresholds[2] = threshold_base << cpi->oxcf.speed;
502 cpi->vbp_threshold_sad = 1000;
504 cpi->vbp_bsize_min = BLOCK_16X16;
506 cpi->vbp_threshold_minmax = 15 + (q >> 3);
510 // Compute the minmax over the 8x8 subblocks.
511 static int compute_minmax_8x8(const uint8_t *s, int sp, const uint8_t *d,
512 int dp, int x16_idx, int y16_idx,
513 #if CONFIG_VP9_HIGHBITDEPTH
520 int minmax_min = 255;
521 // Loop over the 4 8x8 subblocks.
522 for (k = 0; k < 4; k++) {
523 int x8_idx = x16_idx + ((k & 1) << 3);
524 int y8_idx = y16_idx + ((k >> 1) << 3);
527 if (x8_idx < pixels_wide && y8_idx < pixels_high) {
528 #if CONFIG_VP9_HIGHBITDEPTH
529 if (highbd_flag & YV12_FLAG_HIGHBITDEPTH) {
530 vp9_highbd_minmax_8x8(s + y8_idx * sp + x8_idx, sp,
531 d + y8_idx * dp + x8_idx, dp,
534 vp9_minmax_8x8(s + y8_idx * sp + x8_idx, sp,
535 d + y8_idx * dp + x8_idx, dp,
539 vp9_minmax_8x8(s + y8_idx * sp + x8_idx, sp,
540 d + y8_idx * dp + x8_idx, dp,
543 if ((max - min) > minmax_max)
544 minmax_max = (max - min);
545 if ((max - min) < minmax_min)
546 minmax_min = (max - min);
549 return (minmax_max - minmax_min);
552 static void modify_vbp_thresholds(VP9_COMP *cpi, int64_t thresholds[], int q) {
553 VP9_COMMON *const cm = &cpi->common;
554 const int64_t threshold_base = (int64_t)(cpi->y_dequant[q][1]);
556 // Array index: 0 - threshold_64x64; 1 - threshold_32x32;
557 // 2 - threshold_16x16; 3 - vbp_threshold_8x8;
558 thresholds[1] = threshold_base;
559 if (cm->width <= 352 && cm->height <= 288) {
560 thresholds[0] = threshold_base >> 2;
561 thresholds[2] = threshold_base << 3;
563 thresholds[0] = threshold_base;
564 thresholds[1] = (5 * threshold_base) >> 2;
565 thresholds[2] = threshold_base << cpi->oxcf.speed;
569 static void fill_variance_4x4avg(const uint8_t *s, int sp, const uint8_t *d,
570 int dp, int x8_idx, int y8_idx, v8x8 *vst,
571 #if CONFIG_VP9_HIGHBITDEPTH
578 for (k = 0; k < 4; k++) {
579 int x4_idx = x8_idx + ((k & 1) << 2);
580 int y4_idx = y8_idx + ((k >> 1) << 2);
581 unsigned int sse = 0;
583 if (x4_idx < pixels_wide && y4_idx < pixels_high) {
586 #if CONFIG_VP9_HIGHBITDEPTH
587 if (highbd_flag & YV12_FLAG_HIGHBITDEPTH) {
588 s_avg = vp9_highbd_avg_4x4(s + y4_idx * sp + x4_idx, sp);
590 d_avg = vp9_highbd_avg_4x4(d + y4_idx * dp + x4_idx, dp);
592 s_avg = vp9_avg_4x4(s + y4_idx * sp + x4_idx, sp);
594 d_avg = vp9_avg_4x4(d + y4_idx * dp + x4_idx, dp);
597 s_avg = vp9_avg_4x4(s + y4_idx * sp + x4_idx, sp);
599 d_avg = vp9_avg_4x4(d + y4_idx * dp + x4_idx, dp);
604 fill_variance(sse, sum, 0, &vst->split[k].part_variances.none);
608 static void fill_variance_8x8avg(const uint8_t *s, int sp, const uint8_t *d,
609 int dp, int x16_idx, int y16_idx, v16x16 *vst,
610 #if CONFIG_VP9_HIGHBITDEPTH
617 for (k = 0; k < 4; k++) {
618 int x8_idx = x16_idx + ((k & 1) << 3);
619 int y8_idx = y16_idx + ((k >> 1) << 3);
620 unsigned int sse = 0;
622 if (x8_idx < pixels_wide && y8_idx < pixels_high) {
625 #if CONFIG_VP9_HIGHBITDEPTH
626 if (highbd_flag & YV12_FLAG_HIGHBITDEPTH) {
627 s_avg = vp9_highbd_avg_8x8(s + y8_idx * sp + x8_idx, sp);
629 d_avg = vp9_highbd_avg_8x8(d + y8_idx * dp + x8_idx, dp);
631 s_avg = vp9_avg_8x8(s + y8_idx * sp + x8_idx, sp);
633 d_avg = vp9_avg_8x8(d + y8_idx * dp + x8_idx, dp);
636 s_avg = vp9_avg_8x8(s + y8_idx * sp + x8_idx, sp);
638 d_avg = vp9_avg_8x8(d + y8_idx * dp + x8_idx, dp);
643 fill_variance(sse, sum, 0, &vst->split[k].part_variances.none);
647 // This function chooses partitioning based on the variance between source and
648 // reconstructed last, where variance is computed for down-sampled inputs.
649 static int choose_partitioning(VP9_COMP *cpi,
650 const TileInfo *const tile,
652 int mi_row, int mi_col) {
653 VP9_COMMON * const cm = &cpi->common;
654 MACROBLOCKD *xd = &x->e_mbd;
663 int pixels_wide = 64, pixels_high = 64;
664 int64_t thresholds[4] = {cpi->vbp_thresholds[0], cpi->vbp_thresholds[1],
665 cpi->vbp_thresholds[2], cpi->vbp_thresholds[3]};
667 // Always use 4x4 partition for key frame.
668 const int is_key_frame = (cm->frame_type == KEY_FRAME);
669 const int use_4x4_partition = is_key_frame;
670 const int low_res = (cm->width <= 352 && cm->height <= 288);
671 int variance4x4downsample[16];
673 int segment_id = CR_SEGMENT_ID_BASE;
674 if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ && cm->seg.enabled) {
675 const uint8_t *const map = cm->seg.update_map ? cpi->segmentation_map :
676 cm->last_frame_seg_map;
677 segment_id = vp9_get_segment_id(cm, map, BLOCK_64X64, mi_row, mi_col);
679 if (cyclic_refresh_segment_id_boosted(segment_id)) {
680 int q = vp9_get_qindex(&cm->seg, segment_id, cm->base_qindex);
681 modify_vbp_thresholds(cpi, thresholds, q);
685 set_offsets(cpi, tile, x, mi_row, mi_col, BLOCK_64X64);
687 if (xd->mb_to_right_edge < 0)
688 pixels_wide += (xd->mb_to_right_edge >> 3);
689 if (xd->mb_to_bottom_edge < 0)
690 pixels_high += (xd->mb_to_bottom_edge >> 3);
692 s = x->plane[0].src.buf;
693 sp = x->plane[0].src.stride;
696 MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
698 const YV12_BUFFER_CONFIG *yv12 = get_ref_frame_buffer(cpi, LAST_FRAME);
700 const YV12_BUFFER_CONFIG *yv12_g = get_ref_frame_buffer(cpi, GOLDEN_FRAME);
701 unsigned int y_sad, y_sad_g;
702 const BLOCK_SIZE bsize = BLOCK_32X32
703 + (mi_col + 4 < cm->mi_cols) * 2 + (mi_row + 4 < cm->mi_rows);
705 assert(yv12 != NULL);
706 if (yv12_g && yv12_g != yv12) {
707 vp9_setup_pre_planes(xd, 0, yv12_g, mi_row, mi_col,
708 &cm->frame_refs[GOLDEN_FRAME - 1].sf);
709 y_sad_g = cpi->fn_ptr[bsize].sdf(x->plane[0].src.buf,
710 x->plane[0].src.stride,
711 xd->plane[0].pre[0].buf,
712 xd->plane[0].pre[0].stride);
717 vp9_setup_pre_planes(xd, 0, yv12, mi_row, mi_col,
718 &cm->frame_refs[LAST_FRAME - 1].sf);
719 mbmi->ref_frame[0] = LAST_FRAME;
720 mbmi->ref_frame[1] = NONE;
721 mbmi->sb_type = BLOCK_64X64;
722 mbmi->mv[0].as_int = 0;
723 mbmi->interp_filter = BILINEAR;
725 y_sad = vp9_int_pro_motion_estimation(cpi, x, bsize);
726 if (y_sad_g < y_sad) {
727 vp9_setup_pre_planes(xd, 0, yv12_g, mi_row, mi_col,
728 &cm->frame_refs[GOLDEN_FRAME - 1].sf);
729 mbmi->ref_frame[0] = GOLDEN_FRAME;
730 mbmi->mv[0].as_int = 0;
733 x->pred_mv[LAST_FRAME] = mbmi->mv[0].as_mv;
736 vp9_build_inter_predictors_sb(xd, mi_row, mi_col, BLOCK_64X64);
738 for (i = 1; i <= 2; ++i) {
739 struct macroblock_plane *p = &x->plane[i];
740 struct macroblockd_plane *pd = &xd->plane[i];
741 const BLOCK_SIZE bs = get_plane_block_size(bsize, pd);
743 if (bs == BLOCK_INVALID)
746 uv_sad = cpi->fn_ptr[bs].sdf(p->src.buf, p->src.stride,
747 pd->dst.buf, pd->dst.stride);
749 x->color_sensitivity[i - 1] = uv_sad > (y_sad >> 2);
752 d = xd->plane[0].dst.buf;
753 dp = xd->plane[0].dst.stride;
755 // If the y_sad is very small, take 64x64 as partition and exit.
756 // Don't check on boosted segment for now, as 64x64 is suppressed there.
757 if (segment_id == CR_SEGMENT_ID_BASE &&
758 y_sad < cpi->vbp_threshold_sad) {
759 const int block_width = num_8x8_blocks_wide_lookup[BLOCK_64X64];
760 const int block_height = num_8x8_blocks_high_lookup[BLOCK_64X64];
761 if (mi_col + block_width / 2 < cm->mi_cols &&
762 mi_row + block_height / 2 < cm->mi_rows) {
763 set_block_size(cpi, xd, mi_row, mi_col, BLOCK_64X64);
770 #if CONFIG_VP9_HIGHBITDEPTH
771 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
774 d = CONVERT_TO_BYTEPTR(VP9_HIGH_VAR_OFFS_10);
777 d = CONVERT_TO_BYTEPTR(VP9_HIGH_VAR_OFFS_12);
781 d = CONVERT_TO_BYTEPTR(VP9_HIGH_VAR_OFFS_8);
785 #endif // CONFIG_VP9_HIGHBITDEPTH
788 // Index for force_split: 0 for 64x64, 1-4 for 32x32 blocks,
789 // 5-20 for the 16x16 blocks.
791 // Fill in the entire tree of 8x8 (or 4x4 under some conditions) variances
793 for (i = 0; i < 4; i++) {
794 const int x32_idx = ((i & 1) << 5);
795 const int y32_idx = ((i >> 1) << 5);
796 const int i2 = i << 2;
797 force_split[i + 1] = 0;
798 for (j = 0; j < 4; j++) {
799 const int x16_idx = x32_idx + ((j & 1) << 4);
800 const int y16_idx = y32_idx + ((j >> 1) << 4);
801 const int split_index = 5 + i2 + j;
802 v16x16 *vst = &vt.split[i].split[j];
803 force_split[split_index] = 0;
804 variance4x4downsample[i2 + j] = 0;
806 fill_variance_8x8avg(s, sp, d, dp, x16_idx, y16_idx, vst,
807 #if CONFIG_VP9_HIGHBITDEPTH
813 fill_variance_tree(&vt.split[i].split[j], BLOCK_16X16);
814 get_variance(&vt.split[i].split[j].part_variances.none);
815 if (vt.split[i].split[j].part_variances.none.variance >
817 // 16X16 variance is above threshold for split, so force split to 8x8
818 // for this 16x16 block (this also forces splits for upper levels).
819 force_split[split_index] = 1;
820 force_split[i + 1] = 1;
822 } else if (vt.split[i].split[j].part_variances.none.variance >
824 !cyclic_refresh_segment_id_boosted(segment_id)) {
825 // We have some nominal amount of 16x16 variance (based on average),
826 // compute the minmax over the 8x8 sub-blocks, and if above threshold,
827 // force split to 8x8 block for this 16x16 block.
828 int minmax = compute_minmax_8x8(s, sp, d, dp, x16_idx, y16_idx,
829 #if CONFIG_VP9_HIGHBITDEPTH
832 pixels_wide, pixels_high);
833 if (minmax > cpi->vbp_threshold_minmax) {
834 force_split[split_index] = 1;
835 force_split[i + 1] = 1;
840 if (is_key_frame || (low_res &&
841 vt.split[i].split[j].part_variances.none.variance >
842 (thresholds[1] << 1))) {
843 force_split[split_index] = 0;
844 // Go down to 4x4 down-sampling for variance.
845 variance4x4downsample[i2 + j] = 1;
846 for (k = 0; k < 4; k++) {
847 int x8_idx = x16_idx + ((k & 1) << 3);
848 int y8_idx = y16_idx + ((k >> 1) << 3);
849 v8x8 *vst2 = is_key_frame ? &vst->split[k] :
850 &vt2[i2 + j].split[k];
851 fill_variance_4x4avg(s, sp, d, dp, x8_idx, y8_idx, vst2,
852 #if CONFIG_VP9_HIGHBITDEPTH
863 // Fill the rest of the variance tree by summing split partition values.
864 for (i = 0; i < 4; i++) {
865 const int i2 = i << 2;
866 for (j = 0; j < 4; j++) {
867 if (variance4x4downsample[i2 + j] == 1) {
868 v16x16 *vtemp = (!is_key_frame) ? &vt2[i2 + j] :
869 &vt.split[i].split[j];
870 for (m = 0; m < 4; m++)
871 fill_variance_tree(&vtemp->split[m], BLOCK_8X8);
872 fill_variance_tree(vtemp, BLOCK_16X16);
875 fill_variance_tree(&vt.split[i], BLOCK_32X32);
876 // If variance of this 32x32 block is above the threshold, force the block
877 // to split. This also forces a split on the upper (64x64) level.
878 if (!force_split[i + 1]) {
879 get_variance(&vt.split[i].part_variances.none);
880 if (vt.split[i].part_variances.none.variance > thresholds[1]) {
881 force_split[i + 1] = 1;
886 if (!force_split[0]) {
887 fill_variance_tree(&vt, BLOCK_64X64);
888 get_variance(&vt.part_variances.none);
891 // Now go through the entire structure, splitting every block size until
892 // we get to one that's got a variance lower than our threshold.
893 if ( mi_col + 8 > cm->mi_cols || mi_row + 8 > cm->mi_rows ||
894 !set_vt_partitioning(cpi, xd, &vt, BLOCK_64X64, mi_row, mi_col,
895 thresholds[0], BLOCK_16X16, force_split[0])) {
896 for (i = 0; i < 4; ++i) {
897 const int x32_idx = ((i & 1) << 2);
898 const int y32_idx = ((i >> 1) << 2);
899 const int i2 = i << 2;
900 if (!set_vt_partitioning(cpi, xd, &vt.split[i], BLOCK_32X32,
901 (mi_row + y32_idx), (mi_col + x32_idx),
902 thresholds[1], BLOCK_16X16,
903 force_split[i + 1])) {
904 for (j = 0; j < 4; ++j) {
905 const int x16_idx = ((j & 1) << 1);
906 const int y16_idx = ((j >> 1) << 1);
907 // For inter frames: if variance4x4downsample[] == 1 for this 16x16
908 // block, then the variance is based on 4x4 down-sampling, so use vt2
909 // in set_vt_partioning(), otherwise use vt.
910 v16x16 *vtemp = (!is_key_frame &&
911 variance4x4downsample[i2 + j] == 1) ?
912 &vt2[i2 + j] : &vt.split[i].split[j];
913 if (!set_vt_partitioning(cpi, xd, vtemp, BLOCK_16X16,
914 mi_row + y32_idx + y16_idx,
915 mi_col + x32_idx + x16_idx,
918 force_split[5 + i2 + j])) {
919 for (k = 0; k < 4; ++k) {
920 const int x8_idx = (k & 1);
921 const int y8_idx = (k >> 1);
922 if (use_4x4_partition) {
923 if (!set_vt_partitioning(cpi, xd, &vtemp->split[k],
925 mi_row + y32_idx + y16_idx + y8_idx,
926 mi_col + x32_idx + x16_idx + x8_idx,
927 thresholds[3], BLOCK_8X8, 0)) {
928 set_block_size(cpi, xd,
929 (mi_row + y32_idx + y16_idx + y8_idx),
930 (mi_col + x32_idx + x16_idx + x8_idx),
934 set_block_size(cpi, xd,
935 (mi_row + y32_idx + y16_idx + y8_idx),
936 (mi_col + x32_idx + x16_idx + x8_idx),
948 static void update_state(VP9_COMP *cpi, ThreadData *td,
949 PICK_MODE_CONTEXT *ctx,
950 int mi_row, int mi_col, BLOCK_SIZE bsize,
951 int output_enabled) {
953 VP9_COMMON *const cm = &cpi->common;
954 RD_COUNTS *const rdc = &td->rd_counts;
955 MACROBLOCK *const x = &td->mb;
956 MACROBLOCKD *const xd = &x->e_mbd;
957 struct macroblock_plane *const p = x->plane;
958 struct macroblockd_plane *const pd = xd->plane;
959 MODE_INFO *mi = &ctx->mic;
960 MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
961 MODE_INFO *mi_addr = xd->mi[0];
962 const struct segmentation *const seg = &cm->seg;
963 const int bw = num_8x8_blocks_wide_lookup[mi->mbmi.sb_type];
964 const int bh = num_8x8_blocks_high_lookup[mi->mbmi.sb_type];
965 const int x_mis = MIN(bw, cm->mi_cols - mi_col);
966 const int y_mis = MIN(bh, cm->mi_rows - mi_row);
967 MV_REF *const frame_mvs =
968 cm->cur_frame->mvs + mi_row * cm->mi_cols + mi_col;
971 const int mis = cm->mi_stride;
972 const int mi_width = num_8x8_blocks_wide_lookup[bsize];
973 const int mi_height = num_8x8_blocks_high_lookup[bsize];
976 assert(mi->mbmi.sb_type == bsize);
980 // If segmentation in use
982 // For in frame complexity AQ copy the segment id from the segment map.
983 if (cpi->oxcf.aq_mode == COMPLEXITY_AQ) {
984 const uint8_t *const map = seg->update_map ? cpi->segmentation_map
985 : cm->last_frame_seg_map;
986 mi_addr->mbmi.segment_id =
987 vp9_get_segment_id(cm, map, bsize, mi_row, mi_col);
989 // Else for cyclic refresh mode update the segment map, set the segment id
990 // and then update the quantizer.
991 if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ) {
992 vp9_cyclic_refresh_update_segment(cpi, &xd->mi[0]->mbmi, mi_row,
993 mi_col, bsize, ctx->rate, ctx->dist,
998 max_plane = is_inter_block(mbmi) ? MAX_MB_PLANE : 1;
999 for (i = 0; i < max_plane; ++i) {
1000 p[i].coeff = ctx->coeff_pbuf[i][1];
1001 p[i].qcoeff = ctx->qcoeff_pbuf[i][1];
1002 pd[i].dqcoeff = ctx->dqcoeff_pbuf[i][1];
1003 p[i].eobs = ctx->eobs_pbuf[i][1];
1006 for (i = max_plane; i < MAX_MB_PLANE; ++i) {
1007 p[i].coeff = ctx->coeff_pbuf[i][2];
1008 p[i].qcoeff = ctx->qcoeff_pbuf[i][2];
1009 pd[i].dqcoeff = ctx->dqcoeff_pbuf[i][2];
1010 p[i].eobs = ctx->eobs_pbuf[i][2];
1013 // Restore the coding context of the MB to that that was in place
1014 // when the mode was picked for it
1015 for (y = 0; y < mi_height; y++)
1016 for (x_idx = 0; x_idx < mi_width; x_idx++)
1017 if ((xd->mb_to_right_edge >> (3 + MI_SIZE_LOG2)) + mi_width > x_idx
1018 && (xd->mb_to_bottom_edge >> (3 + MI_SIZE_LOG2)) + mi_height > y) {
1019 xd->mi[x_idx + y * mis] = mi_addr;
1022 if (cpi->oxcf.aq_mode)
1023 vp9_init_plane_quantizers(cpi, x);
1025 // FIXME(rbultje) I'm pretty sure this should go to the end of this block
1026 // (i.e. after the output_enabled)
1027 if (bsize < BLOCK_32X32) {
1028 if (bsize < BLOCK_16X16)
1029 ctx->tx_rd_diff[ALLOW_16X16] = ctx->tx_rd_diff[ALLOW_8X8];
1030 ctx->tx_rd_diff[ALLOW_32X32] = ctx->tx_rd_diff[ALLOW_16X16];
1033 if (is_inter_block(mbmi) && mbmi->sb_type < BLOCK_8X8) {
1034 mbmi->mv[0].as_int = mi->bmi[3].as_mv[0].as_int;
1035 mbmi->mv[1].as_int = mi->bmi[3].as_mv[1].as_int;
1038 x->skip = ctx->skip;
1039 memcpy(x->zcoeff_blk[mbmi->tx_size], ctx->zcoeff_blk,
1040 sizeof(uint8_t) * ctx->num_4x4_blk);
1042 if (!output_enabled)
1045 if (!vp9_segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_SKIP)) {
1046 for (i = 0; i < TX_MODES; i++)
1047 rdc->tx_select_diff[i] += ctx->tx_rd_diff[i];
1050 #if CONFIG_INTERNAL_STATS
1051 if (frame_is_intra_only(cm)) {
1052 static const int kf_mode_index[] = {
1054 THR_V_PRED /*V_PRED*/,
1055 THR_H_PRED /*H_PRED*/,
1056 THR_D45_PRED /*D45_PRED*/,
1057 THR_D135_PRED /*D135_PRED*/,
1058 THR_D117_PRED /*D117_PRED*/,
1059 THR_D153_PRED /*D153_PRED*/,
1060 THR_D207_PRED /*D207_PRED*/,
1061 THR_D63_PRED /*D63_PRED*/,
1064 ++cpi->mode_chosen_counts[kf_mode_index[mbmi->mode]];
1066 // Note how often each mode chosen as best
1067 ++cpi->mode_chosen_counts[ctx->best_mode_index];
1070 if (!frame_is_intra_only(cm)) {
1071 if (is_inter_block(mbmi)) {
1072 vp9_update_mv_count(td);
1074 if (cm->interp_filter == SWITCHABLE) {
1075 const int ctx = vp9_get_pred_context_switchable_interp(xd);
1076 ++td->counts->switchable_interp[ctx][mbmi->interp_filter];
1080 rdc->comp_pred_diff[SINGLE_REFERENCE] += ctx->single_pred_diff;
1081 rdc->comp_pred_diff[COMPOUND_REFERENCE] += ctx->comp_pred_diff;
1082 rdc->comp_pred_diff[REFERENCE_MODE_SELECT] += ctx->hybrid_pred_diff;
1084 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i)
1085 rdc->filter_diff[i] += ctx->best_filter_diff[i];
1088 for (h = 0; h < y_mis; ++h) {
1089 MV_REF *const frame_mv = frame_mvs + h * cm->mi_cols;
1090 for (w = 0; w < x_mis; ++w) {
1091 MV_REF *const mv = frame_mv + w;
1092 mv->ref_frame[0] = mi->mbmi.ref_frame[0];
1093 mv->ref_frame[1] = mi->mbmi.ref_frame[1];
1094 mv->mv[0].as_int = mi->mbmi.mv[0].as_int;
1095 mv->mv[1].as_int = mi->mbmi.mv[1].as_int;
1100 void vp9_setup_src_planes(MACROBLOCK *x, const YV12_BUFFER_CONFIG *src,
1101 int mi_row, int mi_col) {
1102 uint8_t *const buffers[3] = {src->y_buffer, src->u_buffer, src->v_buffer };
1103 const int strides[3] = {src->y_stride, src->uv_stride, src->uv_stride };
1106 // Set current frame pointer.
1107 x->e_mbd.cur_buf = src;
1109 for (i = 0; i < MAX_MB_PLANE; i++)
1110 setup_pred_plane(&x->plane[i].src, buffers[i], strides[i], mi_row, mi_col,
1111 NULL, x->e_mbd.plane[i].subsampling_x,
1112 x->e_mbd.plane[i].subsampling_y);
1115 static void set_mode_info_seg_skip(MACROBLOCK *x, TX_MODE tx_mode,
1116 RD_COST *rd_cost, BLOCK_SIZE bsize) {
1117 MACROBLOCKD *const xd = &x->e_mbd;
1118 MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
1119 INTERP_FILTER filter_ref;
1121 if (xd->up_available)
1122 filter_ref = xd->mi[-xd->mi_stride]->mbmi.interp_filter;
1123 else if (xd->left_available)
1124 filter_ref = xd->mi[-1]->mbmi.interp_filter;
1126 filter_ref = EIGHTTAP;
1128 mbmi->sb_type = bsize;
1129 mbmi->mode = ZEROMV;
1130 mbmi->tx_size = MIN(max_txsize_lookup[bsize],
1131 tx_mode_to_biggest_tx_size[tx_mode]);
1133 mbmi->uv_mode = DC_PRED;
1134 mbmi->ref_frame[0] = LAST_FRAME;
1135 mbmi->ref_frame[1] = NONE;
1136 mbmi->mv[0].as_int = 0;
1137 mbmi->interp_filter = filter_ref;
1139 xd->mi[0]->bmi[0].as_mv[0].as_int = 0;
1142 vp9_rd_cost_init(rd_cost);
1145 static int set_segment_rdmult(VP9_COMP *const cpi,
1146 MACROBLOCK *const x,
1147 int8_t segment_id) {
1149 VP9_COMMON *const cm = &cpi->common;
1150 vp9_init_plane_quantizers(cpi, x);
1151 vp9_clear_system_state();
1152 segment_qindex = vp9_get_qindex(&cm->seg, segment_id,
1154 return vp9_compute_rd_mult(cpi, segment_qindex + cm->y_dc_delta_q);
1157 static void rd_pick_sb_modes(VP9_COMP *cpi,
1158 TileDataEnc *tile_data,
1159 MACROBLOCK *const x,
1160 int mi_row, int mi_col, RD_COST *rd_cost,
1161 BLOCK_SIZE bsize, PICK_MODE_CONTEXT *ctx,
1163 VP9_COMMON *const cm = &cpi->common;
1164 TileInfo *const tile_info = &tile_data->tile_info;
1165 MACROBLOCKD *const xd = &x->e_mbd;
1167 struct macroblock_plane *const p = x->plane;
1168 struct macroblockd_plane *const pd = xd->plane;
1169 const AQ_MODE aq_mode = cpi->oxcf.aq_mode;
1172 vp9_clear_system_state();
1174 // Use the lower precision, but faster, 32x32 fdct for mode selection.
1175 x->use_lp32x32fdct = 1;
1177 set_offsets(cpi, tile_info, x, mi_row, mi_col, bsize);
1178 mbmi = &xd->mi[0]->mbmi;
1179 mbmi->sb_type = bsize;
1181 for (i = 0; i < MAX_MB_PLANE; ++i) {
1182 p[i].coeff = ctx->coeff_pbuf[i][0];
1183 p[i].qcoeff = ctx->qcoeff_pbuf[i][0];
1184 pd[i].dqcoeff = ctx->dqcoeff_pbuf[i][0];
1185 p[i].eobs = ctx->eobs_pbuf[i][0];
1189 ctx->pred_pixel_ready = 0;
1192 // Set to zero to make sure we do not use the previous encoded frame stats
1195 #if CONFIG_VP9_HIGHBITDEPTH
1196 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
1197 x->source_variance =
1198 vp9_high_get_sby_perpixel_variance(cpi, &x->plane[0].src,
1201 x->source_variance =
1202 vp9_get_sby_perpixel_variance(cpi, &x->plane[0].src, bsize);
1205 x->source_variance =
1206 vp9_get_sby_perpixel_variance(cpi, &x->plane[0].src, bsize);
1207 #endif // CONFIG_VP9_HIGHBITDEPTH
1209 // Save rdmult before it might be changed, so it can be restored later.
1210 orig_rdmult = x->rdmult;
1212 if (aq_mode == VARIANCE_AQ) {
1213 const int energy = bsize <= BLOCK_16X16 ? x->mb_energy
1214 : vp9_block_energy(cpi, x, bsize);
1215 if (cm->frame_type == KEY_FRAME ||
1216 cpi->refresh_alt_ref_frame ||
1217 (cpi->refresh_golden_frame && !cpi->rc.is_src_frame_alt_ref)) {
1218 mbmi->segment_id = vp9_vaq_segment_id(energy);
1220 const uint8_t *const map = cm->seg.update_map ? cpi->segmentation_map
1221 : cm->last_frame_seg_map;
1222 mbmi->segment_id = vp9_get_segment_id(cm, map, bsize, mi_row, mi_col);
1224 x->rdmult = set_segment_rdmult(cpi, x, mbmi->segment_id);
1225 } else if (aq_mode == COMPLEXITY_AQ) {
1226 x->rdmult = set_segment_rdmult(cpi, x, mbmi->segment_id);
1227 } else if (aq_mode == CYCLIC_REFRESH_AQ) {
1228 const uint8_t *const map = cm->seg.update_map ? cpi->segmentation_map
1229 : cm->last_frame_seg_map;
1230 // If segment is boosted, use rdmult for that segment.
1231 if (cyclic_refresh_segment_id_boosted(
1232 vp9_get_segment_id(cm, map, bsize, mi_row, mi_col)))
1233 x->rdmult = vp9_cyclic_refresh_get_rdmult(cpi->cyclic_refresh);
1236 // Find best coding mode & reconstruct the MB so it is available
1237 // as a predictor for MBs that follow in the SB
1238 if (frame_is_intra_only(cm)) {
1239 vp9_rd_pick_intra_mode_sb(cpi, x, rd_cost, bsize, ctx, best_rd);
1241 if (bsize >= BLOCK_8X8) {
1242 if (vp9_segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_SKIP))
1243 vp9_rd_pick_inter_mode_sb_seg_skip(cpi, tile_data, x, rd_cost, bsize,
1246 vp9_rd_pick_inter_mode_sb(cpi, tile_data, x, mi_row, mi_col,
1247 rd_cost, bsize, ctx, best_rd);
1249 vp9_rd_pick_inter_mode_sub8x8(cpi, tile_data, x, mi_row, mi_col,
1250 rd_cost, bsize, ctx, best_rd);
1255 // Examine the resulting rate and for AQ mode 2 make a segment choice.
1256 if ((rd_cost->rate != INT_MAX) &&
1257 (aq_mode == COMPLEXITY_AQ) && (bsize >= BLOCK_16X16) &&
1258 (cm->frame_type == KEY_FRAME ||
1259 cpi->refresh_alt_ref_frame ||
1260 (cpi->refresh_golden_frame && !cpi->rc.is_src_frame_alt_ref))) {
1261 vp9_caq_select_segment(cpi, x, bsize, mi_row, mi_col, rd_cost->rate);
1264 x->rdmult = orig_rdmult;
1266 // TODO(jingning) The rate-distortion optimization flow needs to be
1267 // refactored to provide proper exit/return handle.
1268 if (rd_cost->rate == INT_MAX)
1269 rd_cost->rdcost = INT64_MAX;
1271 ctx->rate = rd_cost->rate;
1272 ctx->dist = rd_cost->dist;
1275 static void update_stats(VP9_COMMON *cm, ThreadData *td) {
1276 const MACROBLOCK *x = &td->mb;
1277 const MACROBLOCKD *const xd = &x->e_mbd;
1278 const MODE_INFO *const mi = xd->mi[0];
1279 const MB_MODE_INFO *const mbmi = &mi->mbmi;
1280 const BLOCK_SIZE bsize = mbmi->sb_type;
1282 if (!frame_is_intra_only(cm)) {
1283 FRAME_COUNTS *const counts = td->counts;
1284 const int inter_block = is_inter_block(mbmi);
1285 const int seg_ref_active = vp9_segfeature_active(&cm->seg, mbmi->segment_id,
1287 if (!seg_ref_active) {
1288 counts->intra_inter[vp9_get_intra_inter_context(xd)][inter_block]++;
1289 // If the segment reference feature is enabled we have only a single
1290 // reference frame allowed for the segment so exclude it from
1291 // the reference frame counts used to work out probabilities.
1293 const MV_REFERENCE_FRAME ref0 = mbmi->ref_frame[0];
1294 if (cm->reference_mode == REFERENCE_MODE_SELECT)
1295 counts->comp_inter[vp9_get_reference_mode_context(cm, xd)]
1296 [has_second_ref(mbmi)]++;
1298 if (has_second_ref(mbmi)) {
1299 counts->comp_ref[vp9_get_pred_context_comp_ref_p(cm, xd)]
1300 [ref0 == GOLDEN_FRAME]++;
1302 counts->single_ref[vp9_get_pred_context_single_ref_p1(xd)][0]
1303 [ref0 != LAST_FRAME]++;
1304 if (ref0 != LAST_FRAME)
1305 counts->single_ref[vp9_get_pred_context_single_ref_p2(xd)][1]
1306 [ref0 != GOLDEN_FRAME]++;
1311 !vp9_segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_SKIP)) {
1312 const int mode_ctx = mbmi->mode_context[mbmi->ref_frame[0]];
1313 if (bsize >= BLOCK_8X8) {
1314 const PREDICTION_MODE mode = mbmi->mode;
1315 ++counts->inter_mode[mode_ctx][INTER_OFFSET(mode)];
1317 const int num_4x4_w = num_4x4_blocks_wide_lookup[bsize];
1318 const int num_4x4_h = num_4x4_blocks_high_lookup[bsize];
1320 for (idy = 0; idy < 2; idy += num_4x4_h) {
1321 for (idx = 0; idx < 2; idx += num_4x4_w) {
1322 const int j = idy * 2 + idx;
1323 const PREDICTION_MODE b_mode = mi->bmi[j].as_mode;
1324 ++counts->inter_mode[mode_ctx][INTER_OFFSET(b_mode)];
1332 static void restore_context(MACROBLOCK *const x, int mi_row, int mi_col,
1333 ENTROPY_CONTEXT a[16 * MAX_MB_PLANE],
1334 ENTROPY_CONTEXT l[16 * MAX_MB_PLANE],
1335 PARTITION_CONTEXT sa[8], PARTITION_CONTEXT sl[8],
1337 MACROBLOCKD *const xd = &x->e_mbd;
1339 const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
1340 const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
1341 int mi_width = num_8x8_blocks_wide_lookup[bsize];
1342 int mi_height = num_8x8_blocks_high_lookup[bsize];
1343 for (p = 0; p < MAX_MB_PLANE; p++) {
1345 xd->above_context[p] + ((mi_col * 2) >> xd->plane[p].subsampling_x),
1346 a + num_4x4_blocks_wide * p,
1347 (sizeof(ENTROPY_CONTEXT) * num_4x4_blocks_wide) >>
1348 xd->plane[p].subsampling_x);
1351 + ((mi_row & MI_MASK) * 2 >> xd->plane[p].subsampling_y),
1352 l + num_4x4_blocks_high * p,
1353 (sizeof(ENTROPY_CONTEXT) * num_4x4_blocks_high) >>
1354 xd->plane[p].subsampling_y);
1356 memcpy(xd->above_seg_context + mi_col, sa,
1357 sizeof(*xd->above_seg_context) * mi_width);
1358 memcpy(xd->left_seg_context + (mi_row & MI_MASK), sl,
1359 sizeof(xd->left_seg_context[0]) * mi_height);
1362 static void save_context(MACROBLOCK *const x, int mi_row, int mi_col,
1363 ENTROPY_CONTEXT a[16 * MAX_MB_PLANE],
1364 ENTROPY_CONTEXT l[16 * MAX_MB_PLANE],
1365 PARTITION_CONTEXT sa[8], PARTITION_CONTEXT sl[8],
1367 const MACROBLOCKD *const xd = &x->e_mbd;
1369 const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
1370 const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
1371 int mi_width = num_8x8_blocks_wide_lookup[bsize];
1372 int mi_height = num_8x8_blocks_high_lookup[bsize];
1374 // buffer the above/left context information of the block in search.
1375 for (p = 0; p < MAX_MB_PLANE; ++p) {
1377 a + num_4x4_blocks_wide * p,
1378 xd->above_context[p] + (mi_col * 2 >> xd->plane[p].subsampling_x),
1379 (sizeof(ENTROPY_CONTEXT) * num_4x4_blocks_wide) >>
1380 xd->plane[p].subsampling_x);
1382 l + num_4x4_blocks_high * p,
1384 + ((mi_row & MI_MASK) * 2 >> xd->plane[p].subsampling_y),
1385 (sizeof(ENTROPY_CONTEXT) * num_4x4_blocks_high) >>
1386 xd->plane[p].subsampling_y);
1388 memcpy(sa, xd->above_seg_context + mi_col,
1389 sizeof(*xd->above_seg_context) * mi_width);
1390 memcpy(sl, xd->left_seg_context + (mi_row & MI_MASK),
1391 sizeof(xd->left_seg_context[0]) * mi_height);
1394 static void encode_b(VP9_COMP *cpi, const TileInfo *const tile,
1396 TOKENEXTRA **tp, int mi_row, int mi_col,
1397 int output_enabled, BLOCK_SIZE bsize,
1398 PICK_MODE_CONTEXT *ctx) {
1399 MACROBLOCK *const x = &td->mb;
1400 set_offsets(cpi, tile, x, mi_row, mi_col, bsize);
1401 update_state(cpi, td, ctx, mi_row, mi_col, bsize, output_enabled);
1402 encode_superblock(cpi, td, tp, output_enabled, mi_row, mi_col, bsize, ctx);
1404 if (output_enabled) {
1405 update_stats(&cpi->common, td);
1407 (*tp)->token = EOSB_TOKEN;
1412 static void encode_sb(VP9_COMP *cpi, ThreadData *td,
1413 const TileInfo *const tile,
1414 TOKENEXTRA **tp, int mi_row, int mi_col,
1415 int output_enabled, BLOCK_SIZE bsize,
1417 VP9_COMMON *const cm = &cpi->common;
1418 MACROBLOCK *const x = &td->mb;
1419 MACROBLOCKD *const xd = &x->e_mbd;
1421 const int bsl = b_width_log2_lookup[bsize], hbs = (1 << bsl) / 4;
1423 PARTITION_TYPE partition;
1424 BLOCK_SIZE subsize = bsize;
1426 if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols)
1429 if (bsize >= BLOCK_8X8) {
1430 ctx = partition_plane_context(xd, mi_row, mi_col, bsize);
1431 subsize = get_subsize(bsize, pc_tree->partitioning);
1434 subsize = BLOCK_4X4;
1437 partition = partition_lookup[bsl][subsize];
1438 if (output_enabled && bsize != BLOCK_4X4)
1439 td->counts->partition[ctx][partition]++;
1441 switch (partition) {
1442 case PARTITION_NONE:
1443 encode_b(cpi, tile, td, tp, mi_row, mi_col, output_enabled, subsize,
1446 case PARTITION_VERT:
1447 encode_b(cpi, tile, td, tp, mi_row, mi_col, output_enabled, subsize,
1448 &pc_tree->vertical[0]);
1449 if (mi_col + hbs < cm->mi_cols && bsize > BLOCK_8X8) {
1450 encode_b(cpi, tile, td, tp, mi_row, mi_col + hbs, output_enabled,
1451 subsize, &pc_tree->vertical[1]);
1454 case PARTITION_HORZ:
1455 encode_b(cpi, tile, td, tp, mi_row, mi_col, output_enabled, subsize,
1456 &pc_tree->horizontal[0]);
1457 if (mi_row + hbs < cm->mi_rows && bsize > BLOCK_8X8) {
1458 encode_b(cpi, tile, td, tp, mi_row + hbs, mi_col, output_enabled,
1459 subsize, &pc_tree->horizontal[1]);
1462 case PARTITION_SPLIT:
1463 if (bsize == BLOCK_8X8) {
1464 encode_b(cpi, tile, td, tp, mi_row, mi_col, output_enabled, subsize,
1465 pc_tree->leaf_split[0]);
1467 encode_sb(cpi, td, tile, tp, mi_row, mi_col, output_enabled, subsize,
1469 encode_sb(cpi, td, tile, tp, mi_row, mi_col + hbs, output_enabled,
1470 subsize, pc_tree->split[1]);
1471 encode_sb(cpi, td, tile, tp, mi_row + hbs, mi_col, output_enabled,
1472 subsize, pc_tree->split[2]);
1473 encode_sb(cpi, td, tile, tp, mi_row + hbs, mi_col + hbs, output_enabled,
1474 subsize, pc_tree->split[3]);
1478 assert(0 && "Invalid partition type.");
1482 if (partition != PARTITION_SPLIT || bsize == BLOCK_8X8)
1483 update_partition_context(xd, mi_row, mi_col, subsize, bsize);
1486 // Check to see if the given partition size is allowed for a specified number
1487 // of 8x8 block rows and columns remaining in the image.
1488 // If not then return the largest allowed partition size
1489 static BLOCK_SIZE find_partition_size(BLOCK_SIZE bsize,
1490 int rows_left, int cols_left,
1492 if (rows_left <= 0 || cols_left <= 0) {
1493 return MIN(bsize, BLOCK_8X8);
1495 for (; bsize > 0; bsize -= 3) {
1496 *bh = num_8x8_blocks_high_lookup[bsize];
1497 *bw = num_8x8_blocks_wide_lookup[bsize];
1498 if ((*bh <= rows_left) && (*bw <= cols_left)) {
1506 static void set_partial_b64x64_partition(MODE_INFO *mi, int mis,
1507 int bh_in, int bw_in, int row8x8_remaining, int col8x8_remaining,
1508 BLOCK_SIZE bsize, MODE_INFO **mi_8x8) {
1511 for (r = 0; r < MI_BLOCK_SIZE; r += bh) {
1513 for (c = 0; c < MI_BLOCK_SIZE; c += bw) {
1514 const int index = r * mis + c;
1515 mi_8x8[index] = mi + index;
1516 mi_8x8[index]->mbmi.sb_type = find_partition_size(bsize,
1517 row8x8_remaining - r, col8x8_remaining - c, &bh, &bw);
1522 // This function attempts to set all mode info entries in a given SB64
1523 // to the same block partition size.
1524 // However, at the bottom and right borders of the image the requested size
1525 // may not be allowed in which case this code attempts to choose the largest
1526 // allowable partition.
1527 static void set_fixed_partitioning(VP9_COMP *cpi, const TileInfo *const tile,
1528 MODE_INFO **mi_8x8, int mi_row, int mi_col,
1530 VP9_COMMON *const cm = &cpi->common;
1531 const int mis = cm->mi_stride;
1532 const int row8x8_remaining = tile->mi_row_end - mi_row;
1533 const int col8x8_remaining = tile->mi_col_end - mi_col;
1534 int block_row, block_col;
1535 MODE_INFO *mi_upper_left = cm->mi + mi_row * mis + mi_col;
1536 int bh = num_8x8_blocks_high_lookup[bsize];
1537 int bw = num_8x8_blocks_wide_lookup[bsize];
1539 assert((row8x8_remaining > 0) && (col8x8_remaining > 0));
1541 // Apply the requested partition size to the SB64 if it is all "in image"
1542 if ((col8x8_remaining >= MI_BLOCK_SIZE) &&
1543 (row8x8_remaining >= MI_BLOCK_SIZE)) {
1544 for (block_row = 0; block_row < MI_BLOCK_SIZE; block_row += bh) {
1545 for (block_col = 0; block_col < MI_BLOCK_SIZE; block_col += bw) {
1546 int index = block_row * mis + block_col;
1547 mi_8x8[index] = mi_upper_left + index;
1548 mi_8x8[index]->mbmi.sb_type = bsize;
1552 // Else this is a partial SB64.
1553 set_partial_b64x64_partition(mi_upper_left, mis, bh, bw, row8x8_remaining,
1554 col8x8_remaining, bsize, mi_8x8);
1561 } coord_lookup[16] = {
1563 {0, 0}, {0, 2}, {2, 0}, {2, 2},
1565 {0, 4}, {0, 6}, {2, 4}, {2, 6},
1567 {4, 0}, {4, 2}, {6, 0}, {6, 2},
1569 {4, 4}, {4, 6}, {6, 4}, {6, 6},
1572 static void set_source_var_based_partition(VP9_COMP *cpi,
1573 const TileInfo *const tile,
1574 MACROBLOCK *const x,
1576 int mi_row, int mi_col) {
1577 VP9_COMMON *const cm = &cpi->common;
1578 const int mis = cm->mi_stride;
1579 const int row8x8_remaining = tile->mi_row_end - mi_row;
1580 const int col8x8_remaining = tile->mi_col_end - mi_col;
1581 MODE_INFO *mi_upper_left = cm->mi + mi_row * mis + mi_col;
1583 vp9_setup_src_planes(x, cpi->Source, mi_row, mi_col);
1585 assert((row8x8_remaining > 0) && (col8x8_remaining > 0));
1588 if ((col8x8_remaining >= MI_BLOCK_SIZE) &&
1589 (row8x8_remaining >= MI_BLOCK_SIZE)) {
1593 const int offset = (mi_row >> 1) * cm->mb_cols + (mi_col >> 1);
1594 int is_larger_better = 0;
1596 unsigned int thr = cpi->source_var_thresh;
1598 memset(d32, 0, 4 * sizeof(diff));
1600 for (i = 0; i < 4; i++) {
1603 for (j = 0; j < 4; j++) {
1604 int b_mi_row = coord_lookup[i * 4 + j].row;
1605 int b_mi_col = coord_lookup[i * 4 + j].col;
1606 int boffset = b_mi_row / 2 * cm->mb_cols +
1609 d16[j] = cpi->source_diff_var + offset + boffset;
1611 index = b_mi_row * mis + b_mi_col;
1612 mi_8x8[index] = mi_upper_left + index;
1613 mi_8x8[index]->mbmi.sb_type = BLOCK_16X16;
1615 // TODO(yunqingwang): If d16[j].var is very large, use 8x8 partition
1616 // size to further improve quality.
1619 is_larger_better = (d16[0]->var < thr) && (d16[1]->var < thr) &&
1620 (d16[2]->var < thr) && (d16[3]->var < thr);
1622 // Use 32x32 partition
1623 if (is_larger_better) {
1626 for (j = 0; j < 4; j++) {
1627 d32[i].sse += d16[j]->sse;
1628 d32[i].sum += d16[j]->sum;
1631 d32[i].var = d32[i].sse - (((int64_t)d32[i].sum * d32[i].sum) >> 10);
1633 index = coord_lookup[i*4].row * mis + coord_lookup[i*4].col;
1634 mi_8x8[index] = mi_upper_left + index;
1635 mi_8x8[index]->mbmi.sb_type = BLOCK_32X32;
1639 if (use32x32 == 4) {
1641 is_larger_better = (d32[0].var < thr) && (d32[1].var < thr) &&
1642 (d32[2].var < thr) && (d32[3].var < thr);
1644 // Use 64x64 partition
1645 if (is_larger_better) {
1646 mi_8x8[0] = mi_upper_left;
1647 mi_8x8[0]->mbmi.sb_type = BLOCK_64X64;
1650 } else { // partial in-image SB64
1651 int bh = num_8x8_blocks_high_lookup[BLOCK_16X16];
1652 int bw = num_8x8_blocks_wide_lookup[BLOCK_16X16];
1653 set_partial_b64x64_partition(mi_upper_left, mis, bh, bw,
1654 row8x8_remaining, col8x8_remaining, BLOCK_16X16, mi_8x8);
1658 static void update_state_rt(VP9_COMP *cpi, ThreadData *td,
1659 PICK_MODE_CONTEXT *ctx,
1660 int mi_row, int mi_col, int bsize) {
1661 VP9_COMMON *const cm = &cpi->common;
1662 MACROBLOCK *const x = &td->mb;
1663 MACROBLOCKD *const xd = &x->e_mbd;
1664 MODE_INFO *const mi = xd->mi[0];
1665 MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
1666 const struct segmentation *const seg = &cm->seg;
1667 const int bw = num_8x8_blocks_wide_lookup[mi->mbmi.sb_type];
1668 const int bh = num_8x8_blocks_high_lookup[mi->mbmi.sb_type];
1669 const int x_mis = MIN(bw, cm->mi_cols - mi_col);
1670 const int y_mis = MIN(bh, cm->mi_rows - mi_row);
1672 *(xd->mi[0]) = ctx->mic;
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->mbmi.ref_frame[0];
1713 mv->ref_frame[1] = mi->mbmi.ref_frame[1];
1714 mv->mv[0].as_int = mi->mbmi.mv[0].as_int;
1715 mv->mv[1].as_int = 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_grid_visible + idx_str;
1768 ctx = partition_plane_context(xd, mi_row, mi_col, bsize);
1769 subsize = mi_8x8[0]->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]->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];
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]->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]->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];
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;
2153 const int left_in_image = xd->left_available && mi[-1];
2154 const int above_in_image = xd->up_available && mi[-xd->mi_stride];
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_mi_grid_visible[mi_row * xd->mi_stride + mi_col];
2175 get_sb_partition_size_range(xd, prev_mi, &min_size, &max_size, bs_hist);
2177 // Find the min and max partition sizes used in the left SB64
2178 if (left_in_image) {
2179 MODE_INFO **left_sb64_mi = &mi[-MI_BLOCK_SIZE];
2180 get_sb_partition_size_range(xd, left_sb64_mi, &min_size, &max_size,
2183 // Find the min and max partition sizes used in the above SB64.
2184 if (above_in_image) {
2185 MODE_INFO **above_sb64_mi = &mi[-xd->mi_stride * MI_BLOCK_SIZE];
2186 get_sb_partition_size_range(xd, above_sb64_mi, &min_size, &max_size,
2190 // adjust observed min and max
2191 if (cpi->sf.auto_min_max_partition_size == RELAXED_NEIGHBORING_MIN_MAX) {
2192 min_size = min_partition_size[min_size];
2193 max_size = max_partition_size[max_size];
2194 } else if (cpi->sf.auto_min_max_partition_size ==
2195 CONSTRAIN_NEIGHBORING_MIN_MAX) {
2196 // adjust the search range based on the histogram of the observed
2197 // partition sizes from left, above the previous co-located blocks
2199 int first_moment = 0;
2200 int second_moment = 0;
2201 int var_unnormalized = 0;
2203 for (i = 0; i < BLOCK_SIZES; i++) {
2205 first_moment += bs_hist[i] * i;
2206 second_moment += bs_hist[i] * i * i;
2209 // if variance is small enough,
2210 // adjust the range around its mean size, which gives a tighter range
2211 var_unnormalized = second_moment - first_moment * first_moment / sum;
2212 if (var_unnormalized <= 4 * sum) {
2213 int mean = first_moment / sum;
2214 min_size = min_partition_size[mean];
2215 max_size = max_partition_size[mean];
2217 min_size = min_partition_size[min_size];
2218 max_size = max_partition_size[max_size];
2223 // Check border cases where max and min from neighbors may not be legal.
2224 max_size = find_partition_size(max_size,
2225 row8x8_remaining, col8x8_remaining,
2227 min_size = MIN(min_size, max_size);
2229 // When use_square_partition_only is true, make sure at least one square
2230 // partition is allowed by selecting the next smaller square size as
2232 if (cpi->sf.use_square_partition_only &&
2233 next_square_size[max_size] < min_size) {
2234 min_size = next_square_size[max_size];
2237 *min_block_size = min_size;
2238 *max_block_size = max_size;
2241 static void auto_partition_range(VP9_COMP *cpi, const TileInfo *const tile,
2242 MACROBLOCKD *const xd,
2243 int mi_row, int mi_col,
2244 BLOCK_SIZE *min_block_size,
2245 BLOCK_SIZE *max_block_size) {
2246 VP9_COMMON *const cm = &cpi->common;
2247 MODE_INFO **mi_8x8 = xd->mi;
2248 const int left_in_image = xd->left_available && mi_8x8[-1];
2249 const int above_in_image = xd->up_available && mi_8x8[-xd->mi_stride];
2250 int row8x8_remaining = tile->mi_row_end - mi_row;
2251 int col8x8_remaining = tile->mi_col_end - mi_col;
2253 BLOCK_SIZE min_size = BLOCK_32X32;
2254 BLOCK_SIZE max_size = BLOCK_8X8;
2255 int bsl = mi_width_log2_lookup[BLOCK_64X64];
2256 const int search_range_ctrl = (((mi_row + mi_col) >> bsl) +
2257 get_chessboard_index(cm->current_video_frame)) & 0x1;
2258 // Trap case where we do not have a prediction.
2259 if (search_range_ctrl &&
2260 (left_in_image || above_in_image || cm->frame_type != KEY_FRAME)) {
2265 // Find the min and max partition sizes used in the left SB64.
2266 if (left_in_image) {
2269 for (block = 0; block < MI_BLOCK_SIZE; ++block) {
2270 cur_mi = mi[block * xd->mi_stride];
2271 sb_type = cur_mi ? cur_mi->mbmi.sb_type : 0;
2272 min_size = MIN(min_size, sb_type);
2273 max_size = MAX(max_size, sb_type);
2276 // Find the min and max partition sizes used in the above SB64.
2277 if (above_in_image) {
2278 mi = &mi_8x8[-xd->mi_stride * MI_BLOCK_SIZE];
2279 for (block = 0; block < MI_BLOCK_SIZE; ++block) {
2280 sb_type = mi[block] ? mi[block]->mbmi.sb_type : 0;
2281 min_size = MIN(min_size, sb_type);
2282 max_size = MAX(max_size, sb_type);
2286 min_size = min_partition_size[min_size];
2287 max_size = find_partition_size(max_size, row8x8_remaining, col8x8_remaining,
2289 min_size = MIN(min_size, max_size);
2290 min_size = MAX(min_size, BLOCK_8X8);
2291 max_size = MIN(max_size, BLOCK_32X32);
2293 min_size = BLOCK_8X8;
2294 max_size = BLOCK_32X32;
2297 *min_block_size = min_size;
2298 *max_block_size = max_size;
2301 // TODO(jingning) refactor functions setting partition search range
2302 static void set_partition_range(VP9_COMMON *cm, MACROBLOCKD *xd,
2303 int mi_row, int mi_col, BLOCK_SIZE bsize,
2304 BLOCK_SIZE *min_bs, BLOCK_SIZE *max_bs) {
2305 int mi_width = num_8x8_blocks_wide_lookup[bsize];
2306 int mi_height = num_8x8_blocks_high_lookup[bsize];
2310 const int idx_str = cm->mi_stride * mi_row + mi_col;
2311 MODE_INFO **prev_mi = &cm->prev_mi_grid_visible[idx_str];
2312 BLOCK_SIZE bs, min_size, max_size;
2314 min_size = BLOCK_64X64;
2315 max_size = BLOCK_4X4;
2318 for (idy = 0; idy < mi_height; ++idy) {
2319 for (idx = 0; idx < mi_width; ++idx) {
2320 mi = prev_mi[idy * cm->mi_stride + idx];
2321 bs = mi ? mi->mbmi.sb_type : bsize;
2322 min_size = MIN(min_size, bs);
2323 max_size = MAX(max_size, bs);
2328 if (xd->left_available) {
2329 for (idy = 0; idy < mi_height; ++idy) {
2330 mi = xd->mi[idy * cm->mi_stride - 1];
2331 bs = mi ? mi->mbmi.sb_type : bsize;
2332 min_size = MIN(min_size, bs);
2333 max_size = MAX(max_size, bs);
2337 if (xd->up_available) {
2338 for (idx = 0; idx < mi_width; ++idx) {
2339 mi = xd->mi[idx - cm->mi_stride];
2340 bs = mi ? mi->mbmi.sb_type : bsize;
2341 min_size = MIN(min_size, bs);
2342 max_size = MAX(max_size, bs);
2346 if (min_size == max_size) {
2347 min_size = min_partition_size[min_size];
2348 max_size = max_partition_size[max_size];
2355 static INLINE void store_pred_mv(MACROBLOCK *x, PICK_MODE_CONTEXT *ctx) {
2356 memcpy(ctx->pred_mv, x->pred_mv, sizeof(x->pred_mv));
2359 static INLINE void load_pred_mv(MACROBLOCK *x, PICK_MODE_CONTEXT *ctx) {
2360 memcpy(x->pred_mv, ctx->pred_mv, sizeof(x->pred_mv));
2363 #if CONFIG_FP_MB_STATS
2364 const int num_16x16_blocks_wide_lookup[BLOCK_SIZES] =
2365 {1, 1, 1, 1, 1, 1, 1, 1, 2, 2, 2, 4, 4};
2366 const int num_16x16_blocks_high_lookup[BLOCK_SIZES] =
2367 {1, 1, 1, 1, 1, 1, 1, 2, 1, 2, 4, 2, 4};
2368 const int qindex_skip_threshold_lookup[BLOCK_SIZES] =
2369 {0, 10, 10, 30, 40, 40, 60, 80, 80, 90, 100, 100, 120};
2370 const int qindex_split_threshold_lookup[BLOCK_SIZES] =
2371 {0, 3, 3, 7, 15, 15, 30, 40, 40, 60, 80, 80, 120};
2372 const int complexity_16x16_blocks_threshold[BLOCK_SIZES] =
2373 {1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 4, 4, 6};
2384 static INLINE MOTION_DIRECTION get_motion_direction_fp(uint8_t fp_byte) {
2385 if (fp_byte & FPMB_MOTION_ZERO_MASK) {
2387 } else if (fp_byte & FPMB_MOTION_LEFT_MASK) {
2389 } else if (fp_byte & FPMB_MOTION_RIGHT_MASK) {
2391 } else if (fp_byte & FPMB_MOTION_UP_MASK) {
2398 static INLINE int get_motion_inconsistency(MOTION_DIRECTION this_mv,
2399 MOTION_DIRECTION that_mv) {
2400 if (this_mv == that_mv) {
2403 return abs(this_mv - that_mv) == 2 ? 2 : 1;
2408 // TODO(jingning,jimbankoski,rbultje): properly skip partition types that are
2409 // unlikely to be selected depending on previous rate-distortion optimization
2410 // results, for encoding speed-up.
2411 static void rd_pick_partition(VP9_COMP *cpi, ThreadData *td,
2412 TileDataEnc *tile_data,
2413 TOKENEXTRA **tp, int mi_row, int mi_col,
2414 BLOCK_SIZE bsize, RD_COST *rd_cost,
2415 int64_t best_rd, PC_TREE *pc_tree) {
2416 VP9_COMMON *const cm = &cpi->common;
2417 TileInfo *const tile_info = &tile_data->tile_info;
2418 MACROBLOCK *const x = &td->mb;
2419 MACROBLOCKD *const xd = &x->e_mbd;
2420 const int mi_step = num_8x8_blocks_wide_lookup[bsize] / 2;
2421 ENTROPY_CONTEXT l[16 * MAX_MB_PLANE], a[16 * MAX_MB_PLANE];
2422 PARTITION_CONTEXT sl[8], sa[8];
2423 TOKENEXTRA *tp_orig = *tp;
2424 PICK_MODE_CONTEXT *ctx = &pc_tree->none;
2427 RD_COST this_rdc, sum_rdc, best_rdc;
2428 int do_split = bsize >= BLOCK_8X8;
2431 // Override skipping rectangular partition operations for edge blocks
2432 const int force_horz_split = (mi_row + mi_step >= cm->mi_rows);
2433 const int force_vert_split = (mi_col + mi_step >= cm->mi_cols);
2434 const int xss = x->e_mbd.plane[1].subsampling_x;
2435 const int yss = x->e_mbd.plane[1].subsampling_y;
2437 BLOCK_SIZE min_size = x->min_partition_size;
2438 BLOCK_SIZE max_size = x->max_partition_size;
2440 #if CONFIG_FP_MB_STATS
2441 unsigned int src_diff_var = UINT_MAX;
2442 int none_complexity = 0;
2445 int partition_none_allowed = !force_horz_split && !force_vert_split;
2446 int partition_horz_allowed = !force_vert_split && yss <= xss &&
2448 int partition_vert_allowed = !force_horz_split && xss <= yss &&
2452 assert(num_8x8_blocks_wide_lookup[bsize] ==
2453 num_8x8_blocks_high_lookup[bsize]);
2455 vp9_rd_cost_init(&this_rdc);
2456 vp9_rd_cost_init(&sum_rdc);
2457 vp9_rd_cost_reset(&best_rdc);
2458 best_rdc.rdcost = best_rd;
2460 set_offsets(cpi, tile_info, x, mi_row, mi_col, bsize);
2462 if (bsize == BLOCK_16X16 && cpi->oxcf.aq_mode)
2463 x->mb_energy = vp9_block_energy(cpi, x, bsize);
2465 if (cpi->sf.cb_partition_search && bsize == BLOCK_16X16) {
2466 int cb_partition_search_ctrl = ((pc_tree->index == 0 || pc_tree->index == 3)
2467 + get_chessboard_index(cm->current_video_frame)) & 0x1;
2469 if (cb_partition_search_ctrl && bsize > min_size && bsize < max_size)
2470 set_partition_range(cm, xd, mi_row, mi_col, bsize, &min_size, &max_size);
2473 // Determine partition types in search according to the speed features.
2474 // The threshold set here has to be of square block size.
2475 if (cpi->sf.auto_min_max_partition_size) {
2476 partition_none_allowed &= (bsize <= max_size && bsize >= min_size);
2477 partition_horz_allowed &= ((bsize <= max_size && bsize > min_size) ||
2479 partition_vert_allowed &= ((bsize <= max_size && bsize > min_size) ||
2481 do_split &= bsize > min_size;
2483 if (cpi->sf.use_square_partition_only) {
2484 partition_horz_allowed &= force_horz_split;
2485 partition_vert_allowed &= force_vert_split;
2488 save_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
2490 #if CONFIG_FP_MB_STATS
2491 if (cpi->use_fp_mb_stats) {
2492 set_offsets(cpi, tile_info, x, mi_row, mi_col, bsize);
2493 src_diff_var = get_sby_perpixel_diff_variance(cpi, &x->plane[0].src,
2494 mi_row, mi_col, bsize);
2498 #if CONFIG_FP_MB_STATS
2499 // Decide whether we shall split directly and skip searching NONE by using
2500 // the first pass block statistics
2501 if (cpi->use_fp_mb_stats && bsize >= BLOCK_32X32 && do_split &&
2502 partition_none_allowed && src_diff_var > 4 &&
2503 cm->base_qindex < qindex_split_threshold_lookup[bsize]) {
2504 int mb_row = mi_row >> 1;
2505 int mb_col = mi_col >> 1;
2507 MIN(mb_row + num_16x16_blocks_high_lookup[bsize], cm->mb_rows);
2509 MIN(mb_col + num_16x16_blocks_wide_lookup[bsize], cm->mb_cols);
2512 // compute a complexity measure, basically measure inconsistency of motion
2513 // vectors obtained from the first pass in the current block
2514 for (r = mb_row; r < mb_row_end ; r++) {
2515 for (c = mb_col; c < mb_col_end; c++) {
2516 const int mb_index = r * cm->mb_cols + c;
2518 MOTION_DIRECTION this_mv;
2519 MOTION_DIRECTION right_mv;
2520 MOTION_DIRECTION bottom_mv;
2523 get_motion_direction_fp(cpi->twopass.this_frame_mb_stats[mb_index]);
2526 if (c != mb_col_end - 1) {
2527 right_mv = get_motion_direction_fp(
2528 cpi->twopass.this_frame_mb_stats[mb_index + 1]);
2529 none_complexity += get_motion_inconsistency(this_mv, right_mv);
2533 if (r != mb_row_end - 1) {
2534 bottom_mv = get_motion_direction_fp(
2535 cpi->twopass.this_frame_mb_stats[mb_index + cm->mb_cols]);
2536 none_complexity += get_motion_inconsistency(this_mv, bottom_mv);
2539 // do not count its left and top neighbors to avoid double counting
2543 if (none_complexity > complexity_16x16_blocks_threshold[bsize]) {
2544 partition_none_allowed = 0;
2550 if (partition_none_allowed) {
2551 rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col,
2552 &this_rdc, bsize, ctx, best_rdc.rdcost);
2553 if (this_rdc.rate != INT_MAX) {
2554 if (bsize >= BLOCK_8X8) {
2555 pl = partition_plane_context(xd, mi_row, mi_col, bsize);
2556 this_rdc.rate += cpi->partition_cost[pl][PARTITION_NONE];
2557 this_rdc.rdcost = RDCOST(x->rdmult, x->rddiv,
2558 this_rdc.rate, this_rdc.dist);
2561 if (this_rdc.rdcost < best_rdc.rdcost) {
2562 int64_t dist_breakout_thr = cpi->sf.partition_search_breakout_dist_thr;
2563 int rate_breakout_thr = cpi->sf.partition_search_breakout_rate_thr;
2565 best_rdc = this_rdc;
2566 if (bsize >= BLOCK_8X8)
2567 pc_tree->partitioning = PARTITION_NONE;
2569 // Adjust dist breakout threshold according to the partition size.
2570 dist_breakout_thr >>= 8 - (b_width_log2_lookup[bsize] +
2571 b_height_log2_lookup[bsize]);
2573 rate_breakout_thr *= num_pels_log2_lookup[bsize];
2575 // If all y, u, v transform blocks in this partition are skippable, and
2576 // the dist & rate are within the thresholds, the partition search is
2577 // terminated for current branch of the partition search tree.
2578 // The dist & rate thresholds are set to 0 at speed 0 to disable the
2579 // early termination at that speed.
2580 if (!x->e_mbd.lossless &&
2581 (ctx->skippable && best_rdc.dist < dist_breakout_thr &&
2582 best_rdc.rate < rate_breakout_thr)) {
2587 #if CONFIG_FP_MB_STATS
2588 // Check if every 16x16 first pass block statistics has zero
2589 // motion and the corresponding first pass residue is small enough.
2590 // If that is the case, check the difference variance between the
2591 // current frame and the last frame. If the variance is small enough,
2592 // stop further splitting in RD optimization
2593 if (cpi->use_fp_mb_stats && do_split != 0 &&
2594 cm->base_qindex > qindex_skip_threshold_lookup[bsize]) {
2595 int mb_row = mi_row >> 1;
2596 int mb_col = mi_col >> 1;
2598 MIN(mb_row + num_16x16_blocks_high_lookup[bsize], cm->mb_rows);
2600 MIN(mb_col + num_16x16_blocks_wide_lookup[bsize], cm->mb_cols);
2604 for (r = mb_row; r < mb_row_end; r++) {
2605 for (c = mb_col; c < mb_col_end; c++) {
2606 const int mb_index = r * cm->mb_cols + c;
2607 if (!(cpi->twopass.this_frame_mb_stats[mb_index] &
2608 FPMB_MOTION_ZERO_MASK) ||
2609 !(cpi->twopass.this_frame_mb_stats[mb_index] &
2610 FPMB_ERROR_SMALL_MASK)) {
2620 if (src_diff_var == UINT_MAX) {
2621 set_offsets(cpi, tile_info, x, mi_row, mi_col, bsize);
2622 src_diff_var = get_sby_perpixel_diff_variance(
2623 cpi, &x->plane[0].src, mi_row, mi_col, bsize);
2625 if (src_diff_var < 8) {
2634 restore_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
2637 // store estimated motion vector
2638 if (cpi->sf.adaptive_motion_search)
2639 store_pred_mv(x, ctx);
2642 // TODO(jingning): use the motion vectors given by the above search as
2643 // the starting point of motion search in the following partition type check.
2645 subsize = get_subsize(bsize, PARTITION_SPLIT);
2646 if (bsize == BLOCK_8X8) {
2648 if (cpi->sf.adaptive_pred_interp_filter && partition_none_allowed)
2649 pc_tree->leaf_split[0]->pred_interp_filter =
2650 ctx->mic.mbmi.interp_filter;
2651 rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &sum_rdc, subsize,
2652 pc_tree->leaf_split[0], best_rdc.rdcost);
2653 if (sum_rdc.rate == INT_MAX)
2654 sum_rdc.rdcost = INT64_MAX;
2656 for (i = 0; i < 4 && sum_rdc.rdcost < best_rdc.rdcost; ++i) {
2657 const int x_idx = (i & 1) * mi_step;
2658 const int y_idx = (i >> 1) * mi_step;
2660 if (mi_row + y_idx >= cm->mi_rows || mi_col + x_idx >= cm->mi_cols)
2663 if (cpi->sf.adaptive_motion_search)
2664 load_pred_mv(x, ctx);
2666 pc_tree->split[i]->index = i;
2667 rd_pick_partition(cpi, td, tile_data, tp,
2668 mi_row + y_idx, mi_col + x_idx,
2670 best_rdc.rdcost - sum_rdc.rdcost, pc_tree->split[i]);
2672 if (this_rdc.rate == INT_MAX) {
2673 sum_rdc.rdcost = INT64_MAX;
2676 sum_rdc.rate += this_rdc.rate;
2677 sum_rdc.dist += this_rdc.dist;
2678 sum_rdc.rdcost += this_rdc.rdcost;
2683 if (sum_rdc.rdcost < best_rdc.rdcost && i == 4) {
2684 pl = partition_plane_context(xd, mi_row, mi_col, bsize);
2685 sum_rdc.rate += cpi->partition_cost[pl][PARTITION_SPLIT];
2686 sum_rdc.rdcost = RDCOST(x->rdmult, x->rddiv,
2687 sum_rdc.rate, sum_rdc.dist);
2689 if (sum_rdc.rdcost < best_rdc.rdcost) {
2691 pc_tree->partitioning = PARTITION_SPLIT;
2694 // skip rectangular partition test when larger block size
2695 // gives better rd cost
2696 if (cpi->sf.less_rectangular_check)
2697 do_rect &= !partition_none_allowed;
2699 restore_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
2703 if (partition_horz_allowed && do_rect) {
2704 subsize = get_subsize(bsize, PARTITION_HORZ);
2705 if (cpi->sf.adaptive_motion_search)
2706 load_pred_mv(x, ctx);
2707 if (cpi->sf.adaptive_pred_interp_filter && bsize == BLOCK_8X8 &&
2708 partition_none_allowed)
2709 pc_tree->horizontal[0].pred_interp_filter =
2710 ctx->mic.mbmi.interp_filter;
2711 rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &sum_rdc, subsize,
2712 &pc_tree->horizontal[0], best_rdc.rdcost);
2714 if (sum_rdc.rdcost < best_rdc.rdcost && mi_row + mi_step < cm->mi_rows &&
2715 bsize > BLOCK_8X8) {
2716 PICK_MODE_CONTEXT *ctx = &pc_tree->horizontal[0];
2717 update_state(cpi, td, ctx, mi_row, mi_col, subsize, 0);
2718 encode_superblock(cpi, td, tp, 0, mi_row, mi_col, subsize, ctx);
2720 if (cpi->sf.adaptive_motion_search)
2721 load_pred_mv(x, ctx);
2722 if (cpi->sf.adaptive_pred_interp_filter && bsize == BLOCK_8X8 &&
2723 partition_none_allowed)
2724 pc_tree->horizontal[1].pred_interp_filter =
2725 ctx->mic.mbmi.interp_filter;
2726 rd_pick_sb_modes(cpi, tile_data, x, mi_row + mi_step, mi_col,
2727 &this_rdc, subsize, &pc_tree->horizontal[1],
2728 best_rdc.rdcost - sum_rdc.rdcost);
2729 if (this_rdc.rate == INT_MAX) {
2730 sum_rdc.rdcost = INT64_MAX;
2732 sum_rdc.rate += this_rdc.rate;
2733 sum_rdc.dist += this_rdc.dist;
2734 sum_rdc.rdcost += this_rdc.rdcost;
2738 if (sum_rdc.rdcost < best_rdc.rdcost) {
2739 pl = partition_plane_context(xd, mi_row, mi_col, bsize);
2740 sum_rdc.rate += cpi->partition_cost[pl][PARTITION_HORZ];
2741 sum_rdc.rdcost = RDCOST(x->rdmult, x->rddiv, sum_rdc.rate, sum_rdc.dist);
2742 if (sum_rdc.rdcost < best_rdc.rdcost) {
2744 pc_tree->partitioning = PARTITION_HORZ;
2747 restore_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
2750 if (partition_vert_allowed && do_rect) {
2751 subsize = get_subsize(bsize, PARTITION_VERT);
2753 if (cpi->sf.adaptive_motion_search)
2754 load_pred_mv(x, ctx);
2755 if (cpi->sf.adaptive_pred_interp_filter && bsize == BLOCK_8X8 &&
2756 partition_none_allowed)
2757 pc_tree->vertical[0].pred_interp_filter =
2758 ctx->mic.mbmi.interp_filter;
2759 rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &sum_rdc, subsize,
2760 &pc_tree->vertical[0], best_rdc.rdcost);
2761 if (sum_rdc.rdcost < best_rdc.rdcost && mi_col + mi_step < cm->mi_cols &&
2762 bsize > BLOCK_8X8) {
2763 update_state(cpi, td, &pc_tree->vertical[0], mi_row, mi_col, subsize, 0);
2764 encode_superblock(cpi, td, tp, 0, mi_row, mi_col, subsize,
2765 &pc_tree->vertical[0]);
2767 if (cpi->sf.adaptive_motion_search)
2768 load_pred_mv(x, ctx);
2769 if (cpi->sf.adaptive_pred_interp_filter && bsize == BLOCK_8X8 &&
2770 partition_none_allowed)
2771 pc_tree->vertical[1].pred_interp_filter =
2772 ctx->mic.mbmi.interp_filter;
2773 rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col + mi_step,
2775 &pc_tree->vertical[1], best_rdc.rdcost - sum_rdc.rdcost);
2776 if (this_rdc.rate == INT_MAX) {
2777 sum_rdc.rdcost = INT64_MAX;
2779 sum_rdc.rate += this_rdc.rate;
2780 sum_rdc.dist += this_rdc.dist;
2781 sum_rdc.rdcost += this_rdc.rdcost;
2785 if (sum_rdc.rdcost < best_rdc.rdcost) {
2786 pl = partition_plane_context(xd, mi_row, mi_col, bsize);
2787 sum_rdc.rate += cpi->partition_cost[pl][PARTITION_VERT];
2788 sum_rdc.rdcost = RDCOST(x->rdmult, x->rddiv,
2789 sum_rdc.rate, sum_rdc.dist);
2790 if (sum_rdc.rdcost < best_rdc.rdcost) {
2792 pc_tree->partitioning = PARTITION_VERT;
2795 restore_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
2798 // TODO(jbb): This code added so that we avoid static analysis
2799 // warning related to the fact that best_rd isn't used after this
2800 // point. This code should be refactored so that the duplicate
2801 // checks occur in some sub function and thus are used...
2803 *rd_cost = best_rdc;
2806 if (best_rdc.rate < INT_MAX && best_rdc.dist < INT64_MAX &&
2807 pc_tree->index != 3) {
2808 int output_enabled = (bsize == BLOCK_64X64);
2809 encode_sb(cpi, td, tile_info, tp, mi_row, mi_col, output_enabled,
2813 if (bsize == BLOCK_64X64) {
2814 assert(tp_orig < *tp);
2815 assert(best_rdc.rate < INT_MAX);
2816 assert(best_rdc.dist < INT64_MAX);
2818 assert(tp_orig == *tp);
2822 static void encode_rd_sb_row(VP9_COMP *cpi,
2824 TileDataEnc *tile_data,
2827 VP9_COMMON *const cm = &cpi->common;
2828 TileInfo *const tile_info = &tile_data->tile_info;
2829 MACROBLOCK *const x = &td->mb;
2830 MACROBLOCKD *const xd = &x->e_mbd;
2831 SPEED_FEATURES *const sf = &cpi->sf;
2834 // Initialize the left context for the new SB row
2835 memset(&xd->left_context, 0, sizeof(xd->left_context));
2836 memset(xd->left_seg_context, 0, sizeof(xd->left_seg_context));
2838 // Code each SB in the row
2839 for (mi_col = tile_info->mi_col_start; mi_col < tile_info->mi_col_end;
2840 mi_col += MI_BLOCK_SIZE) {
2841 const struct segmentation *const seg = &cm->seg;
2848 const int idx_str = cm->mi_stride * mi_row + mi_col;
2849 MODE_INFO **mi = cm->mi_grid_visible + idx_str;
2851 if (sf->adaptive_pred_interp_filter) {
2852 for (i = 0; i < 64; ++i)
2853 td->leaf_tree[i].pred_interp_filter = SWITCHABLE;
2855 for (i = 0; i < 64; ++i) {
2856 td->pc_tree[i].vertical[0].pred_interp_filter = SWITCHABLE;
2857 td->pc_tree[i].vertical[1].pred_interp_filter = SWITCHABLE;
2858 td->pc_tree[i].horizontal[0].pred_interp_filter = SWITCHABLE;
2859 td->pc_tree[i].horizontal[1].pred_interp_filter = SWITCHABLE;
2863 vp9_zero(x->pred_mv);
2864 td->pc_root->index = 0;
2867 const uint8_t *const map = seg->update_map ? cpi->segmentation_map
2868 : cm->last_frame_seg_map;
2869 int segment_id = vp9_get_segment_id(cm, map, BLOCK_64X64, mi_row, mi_col);
2870 seg_skip = vp9_segfeature_active(seg, segment_id, SEG_LVL_SKIP);
2873 x->source_variance = UINT_MAX;
2874 if (sf->partition_search_type == FIXED_PARTITION || seg_skip) {
2875 const BLOCK_SIZE bsize =
2876 seg_skip ? BLOCK_64X64 : sf->always_this_block_size;
2877 set_offsets(cpi, tile_info, x, mi_row, mi_col, BLOCK_64X64);
2878 set_fixed_partitioning(cpi, tile_info, mi, mi_row, mi_col, bsize);
2879 rd_use_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col,
2880 BLOCK_64X64, &dummy_rate, &dummy_dist, 1, td->pc_root);
2881 } else if (cpi->partition_search_skippable_frame) {
2883 set_offsets(cpi, tile_info, x, mi_row, mi_col, BLOCK_64X64);
2884 bsize = get_rd_var_based_fixed_partition(cpi, x, mi_row, mi_col);
2885 set_fixed_partitioning(cpi, tile_info, mi, mi_row, mi_col, bsize);
2886 rd_use_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col,
2887 BLOCK_64X64, &dummy_rate, &dummy_dist, 1, td->pc_root);
2888 } else if (sf->partition_search_type == VAR_BASED_PARTITION &&
2889 cm->frame_type != KEY_FRAME) {
2890 choose_partitioning(cpi, tile_info, x, mi_row, mi_col);
2891 rd_use_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col,
2892 BLOCK_64X64, &dummy_rate, &dummy_dist, 1, td->pc_root);
2894 // If required set upper and lower partition size limits
2895 if (sf->auto_min_max_partition_size) {
2896 set_offsets(cpi, tile_info, x, mi_row, mi_col, BLOCK_64X64);
2897 rd_auto_partition_range(cpi, tile_info, xd, mi_row, mi_col,
2898 &x->min_partition_size,
2899 &x->max_partition_size);
2901 rd_pick_partition(cpi, td, tile_data, tp, mi_row, mi_col, BLOCK_64X64,
2902 &dummy_rdc, INT64_MAX, td->pc_root);
2907 static void init_encode_frame_mb_context(VP9_COMP *cpi) {
2908 MACROBLOCK *const x = &cpi->td.mb;
2909 VP9_COMMON *const cm = &cpi->common;
2910 MACROBLOCKD *const xd = &x->e_mbd;
2911 const int aligned_mi_cols = mi_cols_aligned_to_sb(cm->mi_cols);
2913 // Copy data over into macro block data structures.
2914 vp9_setup_src_planes(x, cpi->Source, 0, 0);
2916 vp9_setup_block_planes(&x->e_mbd, cm->subsampling_x, cm->subsampling_y);
2918 // Note: this memset assumes above_context[0], [1] and [2]
2919 // are allocated as part of the same buffer.
2920 memset(xd->above_context[0], 0,
2921 sizeof(*xd->above_context[0]) *
2922 2 * aligned_mi_cols * MAX_MB_PLANE);
2923 memset(xd->above_seg_context, 0,
2924 sizeof(*xd->above_seg_context) * aligned_mi_cols);
2927 static int check_dual_ref_flags(VP9_COMP *cpi) {
2928 const int ref_flags = cpi->ref_frame_flags;
2930 if (vp9_segfeature_active(&cpi->common.seg, 1, SEG_LVL_REF_FRAME)) {
2933 return (!!(ref_flags & VP9_GOLD_FLAG) + !!(ref_flags & VP9_LAST_FLAG)
2934 + !!(ref_flags & VP9_ALT_FLAG)) >= 2;
2938 static void reset_skip_tx_size(VP9_COMMON *cm, TX_SIZE max_tx_size) {
2940 const int mis = cm->mi_stride;
2941 MODE_INFO **mi_ptr = cm->mi_grid_visible;
2943 for (mi_row = 0; mi_row < cm->mi_rows; ++mi_row, mi_ptr += mis) {
2944 for (mi_col = 0; mi_col < cm->mi_cols; ++mi_col) {
2945 if (mi_ptr[mi_col]->mbmi.tx_size > max_tx_size)
2946 mi_ptr[mi_col]->mbmi.tx_size = max_tx_size;
2951 static MV_REFERENCE_FRAME get_frame_type(const VP9_COMP *cpi) {
2952 if (frame_is_intra_only(&cpi->common))
2954 else if (cpi->rc.is_src_frame_alt_ref && cpi->refresh_golden_frame)
2955 return ALTREF_FRAME;
2956 else if (cpi->refresh_golden_frame || cpi->refresh_alt_ref_frame)
2957 return GOLDEN_FRAME;
2962 static TX_MODE select_tx_mode(const VP9_COMP *cpi, MACROBLOCKD *const xd) {
2965 if (cpi->common.frame_type == KEY_FRAME &&
2966 cpi->sf.use_nonrd_pick_mode &&
2967 cpi->sf.partition_search_type == VAR_BASED_PARTITION)
2969 if (cpi->sf.tx_size_search_method == USE_LARGESTALL)
2971 else if (cpi->sf.tx_size_search_method == USE_FULL_RD||
2972 cpi->sf.tx_size_search_method == USE_TX_8X8)
2973 return TX_MODE_SELECT;
2975 return cpi->common.tx_mode;
2978 static void hybrid_intra_mode_search(VP9_COMP *cpi, MACROBLOCK *const x,
2979 RD_COST *rd_cost, BLOCK_SIZE bsize,
2980 PICK_MODE_CONTEXT *ctx) {
2981 if (bsize < BLOCK_16X16)
2982 vp9_rd_pick_intra_mode_sb(cpi, x, rd_cost, bsize, ctx, INT64_MAX);
2984 vp9_pick_intra_mode(cpi, x, rd_cost, bsize, ctx);
2987 static void nonrd_pick_sb_modes(VP9_COMP *cpi,
2988 TileDataEnc *tile_data, MACROBLOCK *const x,
2989 int mi_row, int mi_col, RD_COST *rd_cost,
2990 BLOCK_SIZE bsize, PICK_MODE_CONTEXT *ctx) {
2991 VP9_COMMON *const cm = &cpi->common;
2992 TileInfo *const tile_info = &tile_data->tile_info;
2993 MACROBLOCKD *const xd = &x->e_mbd;
2995 set_offsets(cpi, tile_info, x, mi_row, mi_col, bsize);
2996 mbmi = &xd->mi[0]->mbmi;
2997 mbmi->sb_type = bsize;
2999 if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ && cm->seg.enabled)
3000 if (cyclic_refresh_segment_id_boosted(mbmi->segment_id))
3001 x->rdmult = vp9_cyclic_refresh_get_rdmult(cpi->cyclic_refresh);
3003 if (cm->frame_type == KEY_FRAME)
3004 hybrid_intra_mode_search(cpi, x, rd_cost, bsize, ctx);
3005 else if (vp9_segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_SKIP))
3006 set_mode_info_seg_skip(x, cm->tx_mode, rd_cost, bsize);
3007 else if (bsize >= BLOCK_8X8)
3008 vp9_pick_inter_mode(cpi, x, tile_data, mi_row, mi_col,
3009 rd_cost, bsize, ctx);
3011 vp9_pick_inter_mode_sub8x8(cpi, x, tile_data, mi_row, mi_col,
3012 rd_cost, bsize, ctx);
3014 duplicate_mode_info_in_sb(cm, xd, mi_row, mi_col, bsize);
3016 if (rd_cost->rate == INT_MAX)
3017 vp9_rd_cost_reset(rd_cost);
3019 ctx->rate = rd_cost->rate;
3020 ctx->dist = rd_cost->dist;
3023 static void fill_mode_info_sb(VP9_COMMON *cm, MACROBLOCK *x,
3024 int mi_row, int mi_col,
3027 MACROBLOCKD *xd = &x->e_mbd;
3028 int bsl = b_width_log2_lookup[bsize], hbs = (1 << bsl) / 4;
3029 PARTITION_TYPE partition = pc_tree->partitioning;
3030 BLOCK_SIZE subsize = get_subsize(bsize, partition);
3032 assert(bsize >= BLOCK_8X8);
3034 if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols)
3037 switch (partition) {
3038 case PARTITION_NONE:
3039 set_mode_info_offsets(cm, xd, mi_row, mi_col);
3040 *(xd->mi[0]) = pc_tree->none.mic;
3041 duplicate_mode_info_in_sb(cm, xd, mi_row, mi_col, bsize);
3043 case PARTITION_VERT:
3044 set_mode_info_offsets(cm, xd, mi_row, mi_col);
3045 *(xd->mi[0]) = pc_tree->vertical[0].mic;
3046 duplicate_mode_info_in_sb(cm, xd, mi_row, mi_col, subsize);
3048 if (mi_col + hbs < cm->mi_cols) {
3049 set_mode_info_offsets(cm, xd, mi_row, mi_col + hbs);
3050 *(xd->mi[0]) = pc_tree->vertical[1].mic;
3051 duplicate_mode_info_in_sb(cm, xd, mi_row, mi_col + hbs, subsize);
3054 case PARTITION_HORZ:
3055 set_mode_info_offsets(cm, xd, mi_row, mi_col);
3056 *(xd->mi[0]) = pc_tree->horizontal[0].mic;
3057 duplicate_mode_info_in_sb(cm, xd, mi_row, mi_col, subsize);
3058 if (mi_row + hbs < cm->mi_rows) {
3059 set_mode_info_offsets(cm, xd, mi_row + hbs, mi_col);
3060 *(xd->mi[0]) = pc_tree->horizontal[1].mic;
3061 duplicate_mode_info_in_sb(cm, xd, mi_row + hbs, mi_col, subsize);
3064 case PARTITION_SPLIT: {
3065 fill_mode_info_sb(cm, x, mi_row, mi_col, subsize, pc_tree->split[0]);
3066 fill_mode_info_sb(cm, x, mi_row, mi_col + hbs, subsize,
3068 fill_mode_info_sb(cm, x, mi_row + hbs, mi_col, subsize,
3070 fill_mode_info_sb(cm, x, mi_row + hbs, mi_col + hbs, subsize,
3079 // Reset the prediction pixel ready flag recursively.
3080 static void pred_pixel_ready_reset(PC_TREE *pc_tree, BLOCK_SIZE bsize) {
3081 pc_tree->none.pred_pixel_ready = 0;
3082 pc_tree->horizontal[0].pred_pixel_ready = 0;
3083 pc_tree->horizontal[1].pred_pixel_ready = 0;
3084 pc_tree->vertical[0].pred_pixel_ready = 0;
3085 pc_tree->vertical[1].pred_pixel_ready = 0;
3087 if (bsize > BLOCK_8X8) {
3088 BLOCK_SIZE subsize = get_subsize(bsize, PARTITION_SPLIT);
3090 for (i = 0; i < 4; ++i)
3091 pred_pixel_ready_reset(pc_tree->split[i], subsize);
3095 static void nonrd_pick_partition(VP9_COMP *cpi, ThreadData *td,
3096 TileDataEnc *tile_data,
3097 TOKENEXTRA **tp, int mi_row,
3098 int mi_col, BLOCK_SIZE bsize, RD_COST *rd_cost,
3099 int do_recon, int64_t best_rd,
3101 const SPEED_FEATURES *const sf = &cpi->sf;
3102 VP9_COMMON *const cm = &cpi->common;
3103 TileInfo *const tile_info = &tile_data->tile_info;
3104 MACROBLOCK *const x = &td->mb;
3105 MACROBLOCKD *const xd = &x->e_mbd;
3106 const int ms = num_8x8_blocks_wide_lookup[bsize] / 2;
3107 TOKENEXTRA *tp_orig = *tp;
3108 PICK_MODE_CONTEXT *ctx = &pc_tree->none;
3110 BLOCK_SIZE subsize = bsize;
3111 RD_COST this_rdc, sum_rdc, best_rdc;
3112 int do_split = bsize >= BLOCK_8X8;
3114 // Override skipping rectangular partition operations for edge blocks
3115 const int force_horz_split = (mi_row + ms >= cm->mi_rows);
3116 const int force_vert_split = (mi_col + ms >= cm->mi_cols);
3117 const int xss = x->e_mbd.plane[1].subsampling_x;
3118 const int yss = x->e_mbd.plane[1].subsampling_y;
3120 int partition_none_allowed = !force_horz_split && !force_vert_split;
3121 int partition_horz_allowed = !force_vert_split && yss <= xss &&
3123 int partition_vert_allowed = !force_horz_split && xss <= yss &&
3127 assert(num_8x8_blocks_wide_lookup[bsize] ==
3128 num_8x8_blocks_high_lookup[bsize]);
3130 vp9_rd_cost_init(&sum_rdc);
3131 vp9_rd_cost_reset(&best_rdc);
3132 best_rdc.rdcost = best_rd;
3134 // Determine partition types in search according to the speed features.
3135 // The threshold set here has to be of square block size.
3136 if (sf->auto_min_max_partition_size) {
3137 partition_none_allowed &= (bsize <= x->max_partition_size &&
3138 bsize >= x->min_partition_size);
3139 partition_horz_allowed &= ((bsize <= x->max_partition_size &&
3140 bsize > x->min_partition_size) ||
3142 partition_vert_allowed &= ((bsize <= x->max_partition_size &&
3143 bsize > x->min_partition_size) ||
3145 do_split &= bsize > x->min_partition_size;
3147 if (sf->use_square_partition_only) {
3148 partition_horz_allowed &= force_horz_split;
3149 partition_vert_allowed &= force_vert_split;
3152 ctx->pred_pixel_ready = !(partition_vert_allowed ||
3153 partition_horz_allowed ||
3157 if (partition_none_allowed) {
3158 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col,
3159 &this_rdc, bsize, ctx);
3160 ctx->mic.mbmi = xd->mi[0]->mbmi;
3161 ctx->skip_txfm[0] = x->skip_txfm[0];
3162 ctx->skip = x->skip;
3164 if (this_rdc.rate != INT_MAX) {
3165 int pl = partition_plane_context(xd, mi_row, mi_col, bsize);
3166 this_rdc.rate += cpi->partition_cost[pl][PARTITION_NONE];
3167 this_rdc.rdcost = RDCOST(x->rdmult, x->rddiv,
3168 this_rdc.rate, this_rdc.dist);
3169 if (this_rdc.rdcost < best_rdc.rdcost) {
3170 int64_t dist_breakout_thr = sf->partition_search_breakout_dist_thr;
3171 int64_t rate_breakout_thr = sf->partition_search_breakout_rate_thr;
3173 dist_breakout_thr >>= 8 - (b_width_log2_lookup[bsize] +
3174 b_height_log2_lookup[bsize]);
3176 rate_breakout_thr *= num_pels_log2_lookup[bsize];
3178 best_rdc = this_rdc;
3179 if (bsize >= BLOCK_8X8)
3180 pc_tree->partitioning = PARTITION_NONE;
3182 if (!x->e_mbd.lossless &&
3183 this_rdc.rate < rate_breakout_thr &&
3184 this_rdc.dist < dist_breakout_thr) {
3192 // store estimated motion vector
3193 store_pred_mv(x, ctx);
3197 int pl = partition_plane_context(xd, mi_row, mi_col, bsize);
3198 sum_rdc.rate += cpi->partition_cost[pl][PARTITION_SPLIT];
3199 sum_rdc.rdcost = RDCOST(x->rdmult, x->rddiv, sum_rdc.rate, sum_rdc.dist);
3200 subsize = get_subsize(bsize, PARTITION_SPLIT);
3201 for (i = 0; i < 4 && sum_rdc.rdcost < best_rdc.rdcost; ++i) {
3202 const int x_idx = (i & 1) * ms;
3203 const int y_idx = (i >> 1) * ms;
3205 if (mi_row + y_idx >= cm->mi_rows || mi_col + x_idx >= cm->mi_cols)
3207 load_pred_mv(x, ctx);
3208 nonrd_pick_partition(cpi, td, tile_data, tp,
3209 mi_row + y_idx, mi_col + x_idx,
3210 subsize, &this_rdc, 0,
3211 best_rdc.rdcost - sum_rdc.rdcost, pc_tree->split[i]);
3213 if (this_rdc.rate == INT_MAX) {
3214 vp9_rd_cost_reset(&sum_rdc);
3216 sum_rdc.rate += this_rdc.rate;
3217 sum_rdc.dist += this_rdc.dist;
3218 sum_rdc.rdcost += this_rdc.rdcost;
3222 if (sum_rdc.rdcost < best_rdc.rdcost) {
3224 pc_tree->partitioning = PARTITION_SPLIT;
3226 // skip rectangular partition test when larger block size
3227 // gives better rd cost
3228 if (sf->less_rectangular_check)
3229 do_rect &= !partition_none_allowed;
3234 if (partition_horz_allowed && do_rect) {
3235 subsize = get_subsize(bsize, PARTITION_HORZ);
3236 if (sf->adaptive_motion_search)
3237 load_pred_mv(x, ctx);
3238 pc_tree->horizontal[0].pred_pixel_ready = 1;
3239 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &sum_rdc, subsize,
3240 &pc_tree->horizontal[0]);
3242 pc_tree->horizontal[0].mic.mbmi = xd->mi[0]->mbmi;
3243 pc_tree->horizontal[0].skip_txfm[0] = x->skip_txfm[0];
3244 pc_tree->horizontal[0].skip = x->skip;
3246 if (sum_rdc.rdcost < best_rdc.rdcost && mi_row + ms < cm->mi_rows) {
3247 load_pred_mv(x, ctx);
3248 pc_tree->horizontal[1].pred_pixel_ready = 1;
3249 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row + ms, mi_col,
3251 &pc_tree->horizontal[1]);
3253 pc_tree->horizontal[1].mic.mbmi = xd->mi[0]->mbmi;
3254 pc_tree->horizontal[1].skip_txfm[0] = x->skip_txfm[0];
3255 pc_tree->horizontal[1].skip = x->skip;
3257 if (this_rdc.rate == INT_MAX) {
3258 vp9_rd_cost_reset(&sum_rdc);
3260 int pl = partition_plane_context(xd, mi_row, mi_col, bsize);
3261 this_rdc.rate += cpi->partition_cost[pl][PARTITION_HORZ];
3262 sum_rdc.rate += this_rdc.rate;
3263 sum_rdc.dist += this_rdc.dist;
3264 sum_rdc.rdcost = RDCOST(x->rdmult, x->rddiv,
3265 sum_rdc.rate, sum_rdc.dist);
3269 if (sum_rdc.rdcost < best_rdc.rdcost) {
3271 pc_tree->partitioning = PARTITION_HORZ;
3273 pred_pixel_ready_reset(pc_tree, bsize);
3278 if (partition_vert_allowed && do_rect) {
3279 subsize = get_subsize(bsize, PARTITION_VERT);
3280 if (sf->adaptive_motion_search)
3281 load_pred_mv(x, ctx);
3282 pc_tree->vertical[0].pred_pixel_ready = 1;
3283 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &sum_rdc, subsize,
3284 &pc_tree->vertical[0]);
3285 pc_tree->vertical[0].mic.mbmi = xd->mi[0]->mbmi;
3286 pc_tree->vertical[0].skip_txfm[0] = x->skip_txfm[0];
3287 pc_tree->vertical[0].skip = x->skip;
3289 if (sum_rdc.rdcost < best_rdc.rdcost && mi_col + ms < cm->mi_cols) {
3290 load_pred_mv(x, ctx);
3291 pc_tree->vertical[1].pred_pixel_ready = 1;
3292 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col + ms,
3294 &pc_tree->vertical[1]);
3295 pc_tree->vertical[1].mic.mbmi = xd->mi[0]->mbmi;
3296 pc_tree->vertical[1].skip_txfm[0] = x->skip_txfm[0];
3297 pc_tree->vertical[1].skip = x->skip;
3299 if (this_rdc.rate == INT_MAX) {
3300 vp9_rd_cost_reset(&sum_rdc);
3302 int pl = partition_plane_context(xd, mi_row, mi_col, bsize);
3303 sum_rdc.rate += cpi->partition_cost[pl][PARTITION_VERT];
3304 sum_rdc.rate += this_rdc.rate;
3305 sum_rdc.dist += this_rdc.dist;
3306 sum_rdc.rdcost = RDCOST(x->rdmult, x->rddiv,
3307 sum_rdc.rate, sum_rdc.dist);
3311 if (sum_rdc.rdcost < best_rdc.rdcost) {
3313 pc_tree->partitioning = PARTITION_VERT;
3315 pred_pixel_ready_reset(pc_tree, bsize);
3319 *rd_cost = best_rdc;
3321 if (best_rdc.rate == INT_MAX) {
3322 vp9_rd_cost_reset(rd_cost);
3326 // update mode info array
3327 fill_mode_info_sb(cm, x, mi_row, mi_col, bsize, pc_tree);
3329 if (best_rdc.rate < INT_MAX && best_rdc.dist < INT64_MAX && do_recon) {
3330 int output_enabled = (bsize == BLOCK_64X64);
3331 encode_sb_rt(cpi, td, tile_info, tp, mi_row, mi_col, output_enabled,
3335 if (bsize == BLOCK_64X64 && do_recon) {
3336 assert(tp_orig < *tp);
3337 assert(best_rdc.rate < INT_MAX);
3338 assert(best_rdc.dist < INT64_MAX);
3340 assert(tp_orig == *tp);
3344 static void nonrd_select_partition(VP9_COMP *cpi,
3346 TileDataEnc *tile_data,
3349 int mi_row, int mi_col,
3350 BLOCK_SIZE bsize, int output_enabled,
3351 RD_COST *rd_cost, PC_TREE *pc_tree) {
3352 VP9_COMMON *const cm = &cpi->common;
3353 TileInfo *const tile_info = &tile_data->tile_info;
3354 MACROBLOCK *const x = &td->mb;
3355 MACROBLOCKD *const xd = &x->e_mbd;
3356 const int bsl = b_width_log2_lookup[bsize], hbs = (1 << bsl) / 4;
3357 const int mis = cm->mi_stride;
3358 PARTITION_TYPE partition;
3362 vp9_rd_cost_reset(&this_rdc);
3363 if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols)
3366 subsize = (bsize >= BLOCK_8X8) ? mi[0]->mbmi.sb_type : BLOCK_4X4;
3367 partition = partition_lookup[bsl][subsize];
3369 if (bsize == BLOCK_32X32 && partition != PARTITION_NONE &&
3370 subsize >= BLOCK_16X16) {
3371 x->max_partition_size = BLOCK_32X32;
3372 x->min_partition_size = BLOCK_8X8;
3373 nonrd_pick_partition(cpi, td, tile_data, tp, mi_row, mi_col, bsize,
3374 rd_cost, 0, INT64_MAX, pc_tree);
3375 } else if (bsize == BLOCK_16X16 && partition != PARTITION_NONE) {
3376 x->max_partition_size = BLOCK_16X16;
3377 x->min_partition_size = BLOCK_8X8;
3378 nonrd_pick_partition(cpi, td, tile_data, tp, mi_row, mi_col, bsize,
3379 rd_cost, 0, INT64_MAX, pc_tree);
3381 switch (partition) {
3382 case PARTITION_NONE:
3383 pc_tree->none.pred_pixel_ready = 1;
3384 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, rd_cost,
3385 subsize, &pc_tree->none);
3386 pc_tree->none.mic.mbmi = xd->mi[0]->mbmi;
3387 pc_tree->none.skip_txfm[0] = x->skip_txfm[0];
3388 pc_tree->none.skip = x->skip;
3390 case PARTITION_VERT:
3391 pc_tree->vertical[0].pred_pixel_ready = 1;
3392 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, rd_cost,
3393 subsize, &pc_tree->vertical[0]);
3394 pc_tree->vertical[0].mic.mbmi = xd->mi[0]->mbmi;
3395 pc_tree->vertical[0].skip_txfm[0] = x->skip_txfm[0];
3396 pc_tree->vertical[0].skip = x->skip;
3397 if (mi_col + hbs < cm->mi_cols) {
3398 pc_tree->vertical[1].pred_pixel_ready = 1;
3399 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col + hbs,
3400 &this_rdc, subsize, &pc_tree->vertical[1]);
3401 pc_tree->vertical[1].mic.mbmi = xd->mi[0]->mbmi;
3402 pc_tree->vertical[1].skip_txfm[0] = x->skip_txfm[0];
3403 pc_tree->vertical[1].skip = x->skip;
3404 if (this_rdc.rate != INT_MAX && this_rdc.dist != INT64_MAX &&
3405 rd_cost->rate != INT_MAX && rd_cost->dist != INT64_MAX) {
3406 rd_cost->rate += this_rdc.rate;
3407 rd_cost->dist += this_rdc.dist;
3411 case PARTITION_HORZ:
3412 pc_tree->horizontal[0].pred_pixel_ready = 1;
3413 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, rd_cost,
3414 subsize, &pc_tree->horizontal[0]);
3415 pc_tree->horizontal[0].mic.mbmi = xd->mi[0]->mbmi;
3416 pc_tree->horizontal[0].skip_txfm[0] = x->skip_txfm[0];
3417 pc_tree->horizontal[0].skip = x->skip;
3418 if (mi_row + hbs < cm->mi_rows) {
3419 pc_tree->horizontal[1].pred_pixel_ready = 1;
3420 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row + hbs, mi_col,
3421 &this_rdc, subsize, &pc_tree->horizontal[1]);
3422 pc_tree->horizontal[1].mic.mbmi = xd->mi[0]->mbmi;
3423 pc_tree->horizontal[1].skip_txfm[0] = x->skip_txfm[0];
3424 pc_tree->horizontal[1].skip = x->skip;
3425 if (this_rdc.rate != INT_MAX && this_rdc.dist != INT64_MAX &&
3426 rd_cost->rate != INT_MAX && rd_cost->dist != INT64_MAX) {
3427 rd_cost->rate += this_rdc.rate;
3428 rd_cost->dist += this_rdc.dist;
3432 case PARTITION_SPLIT:
3433 subsize = get_subsize(bsize, PARTITION_SPLIT);
3434 nonrd_select_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col,
3435 subsize, output_enabled, rd_cost,
3437 nonrd_select_partition(cpi, td, tile_data, mi + hbs, tp,
3438 mi_row, mi_col + hbs, subsize, output_enabled,
3439 &this_rdc, pc_tree->split[1]);
3440 if (this_rdc.rate != INT_MAX && this_rdc.dist != INT64_MAX &&
3441 rd_cost->rate != INT_MAX && rd_cost->dist != INT64_MAX) {
3442 rd_cost->rate += this_rdc.rate;
3443 rd_cost->dist += this_rdc.dist;
3445 nonrd_select_partition(cpi, td, tile_data, mi + hbs * mis, tp,
3446 mi_row + hbs, mi_col, subsize, output_enabled,
3447 &this_rdc, pc_tree->split[2]);
3448 if (this_rdc.rate != INT_MAX && this_rdc.dist != INT64_MAX &&
3449 rd_cost->rate != INT_MAX && rd_cost->dist != INT64_MAX) {
3450 rd_cost->rate += this_rdc.rate;
3451 rd_cost->dist += this_rdc.dist;
3453 nonrd_select_partition(cpi, td, tile_data, mi + hbs * mis + hbs, tp,
3454 mi_row + hbs, mi_col + hbs, subsize,
3455 output_enabled, &this_rdc, pc_tree->split[3]);
3456 if (this_rdc.rate != INT_MAX && this_rdc.dist != INT64_MAX &&
3457 rd_cost->rate != INT_MAX && rd_cost->dist != INT64_MAX) {
3458 rd_cost->rate += this_rdc.rate;
3459 rd_cost->dist += this_rdc.dist;
3463 assert(0 && "Invalid partition type.");
3468 if (bsize == BLOCK_64X64 && output_enabled)
3469 encode_sb_rt(cpi, td, tile_info, tp, mi_row, mi_col, 1, bsize, pc_tree);
3473 static void nonrd_use_partition(VP9_COMP *cpi,
3475 TileDataEnc *tile_data,
3478 int mi_row, int mi_col,
3479 BLOCK_SIZE bsize, int output_enabled,
3480 RD_COST *dummy_cost, PC_TREE *pc_tree) {
3481 VP9_COMMON *const cm = &cpi->common;
3482 TileInfo *tile_info = &tile_data->tile_info;
3483 MACROBLOCK *const x = &td->mb;
3484 MACROBLOCKD *const xd = &x->e_mbd;
3485 const int bsl = b_width_log2_lookup[bsize], hbs = (1 << bsl) / 4;
3486 const int mis = cm->mi_stride;
3487 PARTITION_TYPE partition;
3490 if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols)
3493 subsize = (bsize >= BLOCK_8X8) ? mi[0]->mbmi.sb_type : BLOCK_4X4;
3494 partition = partition_lookup[bsl][subsize];
3496 if (output_enabled && bsize != BLOCK_4X4) {
3497 int ctx = partition_plane_context(xd, mi_row, mi_col, bsize);
3498 td->counts->partition[ctx][partition]++;
3501 switch (partition) {
3502 case PARTITION_NONE:
3503 pc_tree->none.pred_pixel_ready = 1;
3504 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, dummy_cost,
3505 subsize, &pc_tree->none);
3506 pc_tree->none.mic.mbmi = xd->mi[0]->mbmi;
3507 pc_tree->none.skip_txfm[0] = x->skip_txfm[0];
3508 pc_tree->none.skip = x->skip;
3509 encode_b_rt(cpi, td, tile_info, tp, mi_row, mi_col, output_enabled,
3510 subsize, &pc_tree->none);
3512 case PARTITION_VERT:
3513 pc_tree->vertical[0].pred_pixel_ready = 1;
3514 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, dummy_cost,
3515 subsize, &pc_tree->vertical[0]);
3516 pc_tree->vertical[0].mic.mbmi = xd->mi[0]->mbmi;
3517 pc_tree->vertical[0].skip_txfm[0] = x->skip_txfm[0];
3518 pc_tree->vertical[0].skip = x->skip;
3519 encode_b_rt(cpi, td, tile_info, tp, mi_row, mi_col, output_enabled,
3520 subsize, &pc_tree->vertical[0]);
3521 if (mi_col + hbs < cm->mi_cols && bsize > BLOCK_8X8) {
3522 pc_tree->vertical[1].pred_pixel_ready = 1;
3523 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col + hbs,
3524 dummy_cost, subsize, &pc_tree->vertical[1]);
3525 pc_tree->vertical[1].mic.mbmi = xd->mi[0]->mbmi;
3526 pc_tree->vertical[1].skip_txfm[0] = x->skip_txfm[0];
3527 pc_tree->vertical[1].skip = x->skip;
3528 encode_b_rt(cpi, td, tile_info, tp, mi_row, mi_col + hbs,
3529 output_enabled, subsize, &pc_tree->vertical[1]);
3532 case PARTITION_HORZ:
3533 pc_tree->horizontal[0].pred_pixel_ready = 1;
3534 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, dummy_cost,
3535 subsize, &pc_tree->horizontal[0]);
3536 pc_tree->horizontal[0].mic.mbmi = xd->mi[0]->mbmi;
3537 pc_tree->horizontal[0].skip_txfm[0] = x->skip_txfm[0];
3538 pc_tree->horizontal[0].skip = x->skip;
3539 encode_b_rt(cpi, td, tile_info, tp, mi_row, mi_col, output_enabled,
3540 subsize, &pc_tree->horizontal[0]);
3542 if (mi_row + hbs < cm->mi_rows && bsize > BLOCK_8X8) {
3543 pc_tree->horizontal[1].pred_pixel_ready = 1;
3544 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row + hbs, mi_col,
3545 dummy_cost, subsize, &pc_tree->horizontal[1]);
3546 pc_tree->horizontal[1].mic.mbmi = xd->mi[0]->mbmi;
3547 pc_tree->horizontal[1].skip_txfm[0] = x->skip_txfm[0];
3548 pc_tree->horizontal[1].skip = x->skip;
3549 encode_b_rt(cpi, td, tile_info, tp, mi_row + hbs, mi_col,
3550 output_enabled, subsize, &pc_tree->horizontal[1]);
3553 case PARTITION_SPLIT:
3554 subsize = get_subsize(bsize, PARTITION_SPLIT);
3555 if (bsize == BLOCK_8X8) {
3556 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, dummy_cost,
3557 subsize, pc_tree->leaf_split[0]);
3558 encode_b_rt(cpi, td, tile_info, tp, mi_row, mi_col,
3559 output_enabled, subsize, pc_tree->leaf_split[0]);
3561 nonrd_use_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col,
3562 subsize, output_enabled, dummy_cost,
3564 nonrd_use_partition(cpi, td, tile_data, mi + hbs, tp,
3565 mi_row, mi_col + hbs, subsize, output_enabled,
3566 dummy_cost, pc_tree->split[1]);
3567 nonrd_use_partition(cpi, td, tile_data, mi + hbs * mis, tp,
3568 mi_row + hbs, mi_col, subsize, output_enabled,
3569 dummy_cost, pc_tree->split[2]);
3570 nonrd_use_partition(cpi, td, tile_data, mi + hbs * mis + hbs, tp,
3571 mi_row + hbs, mi_col + hbs, subsize, output_enabled,
3572 dummy_cost, pc_tree->split[3]);
3576 assert(0 && "Invalid partition type.");
3580 if (partition != PARTITION_SPLIT || bsize == BLOCK_8X8)
3581 update_partition_context(xd, mi_row, mi_col, subsize, bsize);
3584 static void encode_nonrd_sb_row(VP9_COMP *cpi,
3586 TileDataEnc *tile_data,
3589 SPEED_FEATURES *const sf = &cpi->sf;
3590 VP9_COMMON *const cm = &cpi->common;
3591 TileInfo *const tile_info = &tile_data->tile_info;
3592 MACROBLOCK *const x = &td->mb;
3593 MACROBLOCKD *const xd = &x->e_mbd;
3596 // Initialize the left context for the new SB row
3597 memset(&xd->left_context, 0, sizeof(xd->left_context));
3598 memset(xd->left_seg_context, 0, sizeof(xd->left_seg_context));
3600 // Code each SB in the row
3601 for (mi_col = tile_info->mi_col_start; mi_col < tile_info->mi_col_end;
3602 mi_col += MI_BLOCK_SIZE) {
3603 const struct segmentation *const seg = &cm->seg;
3605 const int idx_str = cm->mi_stride * mi_row + mi_col;
3606 MODE_INFO **mi = cm->mi_grid_visible + idx_str;
3607 PARTITION_SEARCH_TYPE partition_search_type = sf->partition_search_type;
3608 BLOCK_SIZE bsize = BLOCK_64X64;
3610 x->source_variance = UINT_MAX;
3611 vp9_zero(x->pred_mv);
3612 vp9_rd_cost_init(&dummy_rdc);
3613 x->color_sensitivity[0] = 0;
3614 x->color_sensitivity[1] = 0;
3617 const uint8_t *const map = seg->update_map ? cpi->segmentation_map
3618 : cm->last_frame_seg_map;
3619 int segment_id = vp9_get_segment_id(cm, map, BLOCK_64X64, mi_row, mi_col);
3620 seg_skip = vp9_segfeature_active(seg, segment_id, SEG_LVL_SKIP);
3622 partition_search_type = FIXED_PARTITION;
3626 // Set the partition type of the 64X64 block
3627 switch (partition_search_type) {
3628 case VAR_BASED_PARTITION:
3629 // TODO(jingning, marpan): The mode decision and encoding process
3630 // support both intra and inter sub8x8 block coding for RTC mode.
3631 // Tune the thresholds accordingly to use sub8x8 block coding for
3632 // coding performance improvement.
3633 choose_partitioning(cpi, tile_info, x, mi_row, mi_col);
3634 nonrd_use_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col,
3635 BLOCK_64X64, 1, &dummy_rdc, td->pc_root);
3637 case SOURCE_VAR_BASED_PARTITION:
3638 set_source_var_based_partition(cpi, tile_info, x, mi, mi_row, mi_col);
3639 nonrd_use_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col,
3640 BLOCK_64X64, 1, &dummy_rdc, td->pc_root);
3642 case FIXED_PARTITION:
3644 bsize = sf->always_this_block_size;
3645 set_fixed_partitioning(cpi, tile_info, mi, mi_row, mi_col, bsize);
3646 nonrd_use_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col,
3647 BLOCK_64X64, 1, &dummy_rdc, td->pc_root);
3649 case REFERENCE_PARTITION:
3650 set_offsets(cpi, tile_info, x, mi_row, mi_col, BLOCK_64X64);
3651 if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ && cm->seg.enabled &&
3652 xd->mi[0]->mbmi.segment_id) {
3653 x->max_partition_size = BLOCK_64X64;
3654 x->min_partition_size = BLOCK_8X8;
3655 nonrd_pick_partition(cpi, td, tile_data, tp, mi_row, mi_col,
3656 BLOCK_64X64, &dummy_rdc, 1,
3657 INT64_MAX, td->pc_root);
3659 choose_partitioning(cpi, tile_info, x, mi_row, mi_col);
3660 nonrd_select_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col,
3661 BLOCK_64X64, 1, &dummy_rdc, td->pc_root);
3671 // end RTC play code
3673 static int set_var_thresh_from_histogram(VP9_COMP *cpi) {
3674 const SPEED_FEATURES *const sf = &cpi->sf;
3675 const VP9_COMMON *const cm = &cpi->common;
3677 const uint8_t *src = cpi->Source->y_buffer;
3678 const uint8_t *last_src = cpi->Last_Source->y_buffer;
3679 const int src_stride = cpi->Source->y_stride;
3680 const int last_stride = cpi->Last_Source->y_stride;
3682 // Pick cutoff threshold
3683 const int cutoff = (MIN(cm->width, cm->height) >= 720) ?
3684 (cm->MBs * VAR_HIST_LARGE_CUT_OFF / 100) :
3685 (cm->MBs * VAR_HIST_SMALL_CUT_OFF / 100);
3686 DECLARE_ALIGNED_ARRAY(16, int, hist, VAR_HIST_BINS);
3687 diff *var16 = cpi->source_diff_var;
3692 memset(hist, 0, VAR_HIST_BINS * sizeof(hist[0]));
3694 for (i = 0; i < cm->mb_rows; i++) {
3695 for (j = 0; j < cm->mb_cols; j++) {
3696 #if CONFIG_VP9_HIGHBITDEPTH
3697 if (cm->use_highbitdepth) {
3698 switch (cm->bit_depth) {
3700 vp9_highbd_get16x16var(src, src_stride, last_src, last_stride,
3701 &var16->sse, &var16->sum);
3704 vp9_highbd_10_get16x16var(src, src_stride, last_src, last_stride,
3705 &var16->sse, &var16->sum);
3708 vp9_highbd_12_get16x16var(src, src_stride, last_src, last_stride,
3709 &var16->sse, &var16->sum);
3712 assert(0 && "cm->bit_depth should be VPX_BITS_8, VPX_BITS_10"
3717 vp9_get16x16var(src, src_stride, last_src, last_stride,
3718 &var16->sse, &var16->sum);
3721 vp9_get16x16var(src, src_stride, last_src, last_stride,
3722 &var16->sse, &var16->sum);
3723 #endif // CONFIG_VP9_HIGHBITDEPTH
3724 var16->var = var16->sse -
3725 (((uint32_t)var16->sum * var16->sum) >> 8);
3727 if (var16->var >= VAR_HIST_MAX_BG_VAR)
3728 hist[VAR_HIST_BINS - 1]++;
3730 hist[var16->var / VAR_HIST_FACTOR]++;
3737 src = src - cm->mb_cols * 16 + 16 * src_stride;
3738 last_src = last_src - cm->mb_cols * 16 + 16 * last_stride;
3741 cpi->source_var_thresh = 0;
3743 if (hist[VAR_HIST_BINS - 1] < cutoff) {
3744 for (i = 0; i < VAR_HIST_BINS - 1; i++) {
3748 cpi->source_var_thresh = (i + 1) * VAR_HIST_FACTOR;
3754 return sf->search_type_check_frequency;
3757 static void source_var_based_partition_search_method(VP9_COMP *cpi) {
3758 VP9_COMMON *const cm = &cpi->common;
3759 SPEED_FEATURES *const sf = &cpi->sf;
3761 if (cm->frame_type == KEY_FRAME) {
3762 // For key frame, use SEARCH_PARTITION.
3763 sf->partition_search_type = SEARCH_PARTITION;
3764 } else if (cm->intra_only) {
3765 sf->partition_search_type = FIXED_PARTITION;
3767 if (cm->last_width != cm->width || cm->last_height != cm->height) {
3768 if (cpi->source_diff_var)
3769 vpx_free(cpi->source_diff_var);
3771 CHECK_MEM_ERROR(cm, cpi->source_diff_var,
3772 vpx_calloc(cm->MBs, sizeof(diff)));
3775 if (!cpi->frames_till_next_var_check)
3776 cpi->frames_till_next_var_check = set_var_thresh_from_histogram(cpi);
3778 if (cpi->frames_till_next_var_check > 0) {
3779 sf->partition_search_type = FIXED_PARTITION;
3780 cpi->frames_till_next_var_check--;
3785 static int get_skip_encode_frame(const VP9_COMMON *cm, ThreadData *const td) {
3786 unsigned int intra_count = 0, inter_count = 0;
3789 for (j = 0; j < INTRA_INTER_CONTEXTS; ++j) {
3790 intra_count += td->counts->intra_inter[j][0];
3791 inter_count += td->counts->intra_inter[j][1];
3794 return (intra_count << 2) < inter_count &&
3795 cm->frame_type != KEY_FRAME &&
3799 void vp9_init_tile_data(VP9_COMP *cpi) {
3800 VP9_COMMON *const cm = &cpi->common;
3801 const int tile_cols = 1 << cm->log2_tile_cols;
3802 const int tile_rows = 1 << cm->log2_tile_rows;
3803 int tile_col, tile_row;
3804 TOKENEXTRA *pre_tok = cpi->tile_tok[0][0];
3807 if (cpi->tile_data == NULL) {
3808 CHECK_MEM_ERROR(cm, cpi->tile_data,
3809 vpx_malloc(tile_cols * tile_rows * sizeof(*cpi->tile_data)));
3810 for (tile_row = 0; tile_row < tile_rows; ++tile_row)
3811 for (tile_col = 0; tile_col < tile_cols; ++tile_col) {
3812 TileDataEnc *tile_data =
3813 &cpi->tile_data[tile_row * tile_cols + tile_col];
3815 for (i = 0; i < BLOCK_SIZES; ++i) {
3816 for (j = 0; j < MAX_MODES; ++j) {
3817 tile_data->thresh_freq_fact[i][j] = 32;
3818 tile_data->mode_map[i][j] = j;
3824 for (tile_row = 0; tile_row < tile_rows; ++tile_row) {
3825 for (tile_col = 0; tile_col < tile_cols; ++tile_col) {
3826 TileInfo *tile_info =
3827 &cpi->tile_data[tile_row * tile_cols + tile_col].tile_info;
3828 vp9_tile_init(tile_info, cm, tile_row, tile_col);
3830 cpi->tile_tok[tile_row][tile_col] = pre_tok + tile_tok;
3831 pre_tok = cpi->tile_tok[tile_row][tile_col];
3832 tile_tok = allocated_tokens(*tile_info);
3837 void vp9_encode_tile(VP9_COMP *cpi, ThreadData *td,
3838 int tile_row, int tile_col) {
3839 VP9_COMMON *const cm = &cpi->common;
3840 const int tile_cols = 1 << cm->log2_tile_cols;
3841 TileDataEnc *this_tile =
3842 &cpi->tile_data[tile_row * tile_cols + tile_col];
3843 const TileInfo * const tile_info = &this_tile->tile_info;
3844 TOKENEXTRA *tok = cpi->tile_tok[tile_row][tile_col];
3847 for (mi_row = tile_info->mi_row_start; mi_row < tile_info->mi_row_end;
3848 mi_row += MI_BLOCK_SIZE) {
3849 if (cpi->sf.use_nonrd_pick_mode)
3850 encode_nonrd_sb_row(cpi, td, this_tile, mi_row, &tok);
3852 encode_rd_sb_row(cpi, td, this_tile, mi_row, &tok);
3854 cpi->tok_count[tile_row][tile_col] =
3855 (unsigned int)(tok - cpi->tile_tok[tile_row][tile_col]);
3856 assert(tok - cpi->tile_tok[tile_row][tile_col] <=
3857 allocated_tokens(*tile_info));
3860 static void encode_tiles(VP9_COMP *cpi) {
3861 VP9_COMMON *const cm = &cpi->common;
3862 const int tile_cols = 1 << cm->log2_tile_cols;
3863 const int tile_rows = 1 << cm->log2_tile_rows;
3864 int tile_col, tile_row;
3866 vp9_init_tile_data(cpi);
3868 for (tile_row = 0; tile_row < tile_rows; ++tile_row)
3869 for (tile_col = 0; tile_col < tile_cols; ++tile_col)
3870 vp9_encode_tile(cpi, &cpi->td, tile_row, tile_col);
3873 #if CONFIG_FP_MB_STATS
3874 static int input_fpmb_stats(FIRSTPASS_MB_STATS *firstpass_mb_stats,
3875 VP9_COMMON *cm, uint8_t **this_frame_mb_stats) {
3876 uint8_t *mb_stats_in = firstpass_mb_stats->mb_stats_start +
3877 cm->current_video_frame * cm->MBs * sizeof(uint8_t);
3879 if (mb_stats_in > firstpass_mb_stats->mb_stats_end)
3882 *this_frame_mb_stats = mb_stats_in;
3888 static void encode_frame_internal(VP9_COMP *cpi) {
3889 SPEED_FEATURES *const sf = &cpi->sf;
3890 RD_OPT *const rd_opt = &cpi->rd;
3891 ThreadData *const td = &cpi->td;
3892 MACROBLOCK *const x = &td->mb;
3893 VP9_COMMON *const cm = &cpi->common;
3894 MACROBLOCKD *const xd = &x->e_mbd;
3895 RD_COUNTS *const rdc = &cpi->td.rd_counts;
3897 xd->mi = cm->mi_grid_visible;
3900 vp9_zero(*td->counts);
3901 vp9_zero(rdc->coef_counts);
3902 vp9_zero(rdc->comp_pred_diff);
3903 vp9_zero(rdc->filter_diff);
3904 vp9_zero(rdc->tx_select_diff);
3905 vp9_zero(rd_opt->tx_select_threshes);
3907 xd->lossless = cm->base_qindex == 0 &&
3908 cm->y_dc_delta_q == 0 &&
3909 cm->uv_dc_delta_q == 0 &&
3910 cm->uv_ac_delta_q == 0;
3912 #if CONFIG_VP9_HIGHBITDEPTH
3913 if (cm->use_highbitdepth)
3914 x->fwd_txm4x4 = xd->lossless ? vp9_highbd_fwht4x4 : vp9_highbd_fdct4x4;
3916 x->fwd_txm4x4 = xd->lossless ? vp9_fwht4x4 : vp9_fdct4x4;
3917 x->highbd_itxm_add = xd->lossless ? vp9_highbd_iwht4x4_add :
3918 vp9_highbd_idct4x4_add;
3920 x->fwd_txm4x4 = xd->lossless ? vp9_fwht4x4 : vp9_fdct4x4;
3921 #endif // CONFIG_VP9_HIGHBITDEPTH
3922 x->itxm_add = xd->lossless ? vp9_iwht4x4_add : vp9_idct4x4_add;
3927 cm->tx_mode = select_tx_mode(cpi, xd);
3929 vp9_frame_init_quantizer(cpi);
3931 vp9_initialize_rd_consts(cpi);
3932 vp9_initialize_me_consts(cpi, x, cm->base_qindex);
3933 init_encode_frame_mb_context(cpi);
3934 cm->use_prev_frame_mvs = !cm->error_resilient_mode &&
3935 cm->width == cm->last_width &&
3936 cm->height == cm->last_height &&
3938 cm->last_show_frame;
3939 // Special case: set prev_mi to NULL when the previous mode info
3940 // context cannot be used.
3941 cm->prev_mi = cm->use_prev_frame_mvs ?
3942 cm->prev_mip + cm->mi_stride + 1 : NULL;
3944 x->quant_fp = cpi->sf.use_quant_fp;
3945 vp9_zero(x->skip_txfm);
3946 if (sf->use_nonrd_pick_mode) {
3947 // Initialize internal buffer pointers for rtc coding, where non-RD
3948 // mode decision is used and hence no buffer pointer swap needed.
3950 struct macroblock_plane *const p = x->plane;
3951 struct macroblockd_plane *const pd = xd->plane;
3952 PICK_MODE_CONTEXT *ctx = &cpi->td.pc_root->none;
3954 for (i = 0; i < MAX_MB_PLANE; ++i) {
3955 p[i].coeff = ctx->coeff_pbuf[i][0];
3956 p[i].qcoeff = ctx->qcoeff_pbuf[i][0];
3957 pd[i].dqcoeff = ctx->dqcoeff_pbuf[i][0];
3958 p[i].eobs = ctx->eobs_pbuf[i][0];
3960 vp9_zero(x->zcoeff_blk);
3962 if (cm->frame_type != KEY_FRAME && cpi->rc.frames_since_golden == 0)
3963 cpi->ref_frame_flags &= (~VP9_GOLD_FLAG);
3965 if (sf->partition_search_type == SOURCE_VAR_BASED_PARTITION)
3966 source_var_based_partition_search_method(cpi);
3970 struct vpx_usec_timer emr_timer;
3971 vpx_usec_timer_start(&emr_timer);
3973 #if CONFIG_FP_MB_STATS
3974 if (cpi->use_fp_mb_stats) {
3975 input_fpmb_stats(&cpi->twopass.firstpass_mb_stats, cm,
3976 &cpi->twopass.this_frame_mb_stats);
3980 // If allowed, encoding tiles in parallel with one thread handling one tile.
3981 if (MIN(cpi->oxcf.max_threads, 1 << cm->log2_tile_cols) > 1)
3982 vp9_encode_tiles_mt(cpi);
3986 vpx_usec_timer_mark(&emr_timer);
3987 cpi->time_encode_sb_row += vpx_usec_timer_elapsed(&emr_timer);
3990 sf->skip_encode_frame = sf->skip_encode_sb ?
3991 get_skip_encode_frame(cm, td) : 0;
3994 // Keep record of the total distortion this time around for future use
3995 cpi->last_frame_distortion = cpi->frame_distortion;
3999 static INTERP_FILTER get_interp_filter(
4000 const int64_t threshes[SWITCHABLE_FILTER_CONTEXTS], int is_alt_ref) {
4002 threshes[EIGHTTAP_SMOOTH] > threshes[EIGHTTAP] &&
4003 threshes[EIGHTTAP_SMOOTH] > threshes[EIGHTTAP_SHARP] &&
4004 threshes[EIGHTTAP_SMOOTH] > threshes[SWITCHABLE - 1]) {
4005 return EIGHTTAP_SMOOTH;
4006 } else if (threshes[EIGHTTAP_SHARP] > threshes[EIGHTTAP] &&
4007 threshes[EIGHTTAP_SHARP] > threshes[SWITCHABLE - 1]) {
4008 return EIGHTTAP_SHARP;
4009 } else if (threshes[EIGHTTAP] > threshes[SWITCHABLE - 1]) {
4016 void vp9_encode_frame(VP9_COMP *cpi) {
4017 VP9_COMMON *const cm = &cpi->common;
4019 // In the longer term the encoder should be generalized to match the
4020 // decoder such that we allow compound where one of the 3 buffers has a
4021 // different sign bias and that buffer is then the fixed ref. However, this
4022 // requires further work in the rd loop. For now the only supported encoder
4023 // side behavior is where the ALT ref buffer has opposite sign bias to
4025 if (!frame_is_intra_only(cm)) {
4026 if ((cm->ref_frame_sign_bias[ALTREF_FRAME] ==
4027 cm->ref_frame_sign_bias[GOLDEN_FRAME]) ||
4028 (cm->ref_frame_sign_bias[ALTREF_FRAME] ==
4029 cm->ref_frame_sign_bias[LAST_FRAME])) {
4030 cpi->allow_comp_inter_inter = 0;
4032 cpi->allow_comp_inter_inter = 1;
4033 cm->comp_fixed_ref = ALTREF_FRAME;
4034 cm->comp_var_ref[0] = LAST_FRAME;
4035 cm->comp_var_ref[1] = GOLDEN_FRAME;
4039 if (cpi->sf.frame_parameter_update) {
4041 RD_OPT *const rd_opt = &cpi->rd;
4042 FRAME_COUNTS *counts = cpi->td.counts;
4043 RD_COUNTS *const rdc = &cpi->td.rd_counts;
4045 // This code does a single RD pass over the whole frame assuming
4046 // either compound, single or hybrid prediction as per whatever has
4047 // worked best for that type of frame in the past.
4048 // It also predicts whether another coding mode would have worked
4049 // better that this coding mode. If that is the case, it remembers
4050 // that for subsequent frames.
4051 // It does the same analysis for transform size selection also.
4052 const MV_REFERENCE_FRAME frame_type = get_frame_type(cpi);
4053 int64_t *const mode_thrs = rd_opt->prediction_type_threshes[frame_type];
4054 int64_t *const filter_thrs = rd_opt->filter_threshes[frame_type];
4055 int *const tx_thrs = rd_opt->tx_select_threshes[frame_type];
4056 const int is_alt_ref = frame_type == ALTREF_FRAME;
4058 /* prediction (compound, single or hybrid) mode selection */
4059 if (is_alt_ref || !cpi->allow_comp_inter_inter)
4060 cm->reference_mode = SINGLE_REFERENCE;
4061 else if (mode_thrs[COMPOUND_REFERENCE] > mode_thrs[SINGLE_REFERENCE] &&
4062 mode_thrs[COMPOUND_REFERENCE] >
4063 mode_thrs[REFERENCE_MODE_SELECT] &&
4064 check_dual_ref_flags(cpi) &&
4065 cpi->static_mb_pct == 100)
4066 cm->reference_mode = COMPOUND_REFERENCE;
4067 else if (mode_thrs[SINGLE_REFERENCE] > mode_thrs[REFERENCE_MODE_SELECT])
4068 cm->reference_mode = SINGLE_REFERENCE;
4070 cm->reference_mode = REFERENCE_MODE_SELECT;
4072 if (cm->interp_filter == SWITCHABLE)
4073 cm->interp_filter = get_interp_filter(filter_thrs, is_alt_ref);
4075 encode_frame_internal(cpi);
4077 for (i = 0; i < REFERENCE_MODES; ++i)
4078 mode_thrs[i] = (mode_thrs[i] + rdc->comp_pred_diff[i] / cm->MBs) / 2;
4080 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i)
4081 filter_thrs[i] = (filter_thrs[i] + rdc->filter_diff[i] / cm->MBs) / 2;
4083 for (i = 0; i < TX_MODES; ++i) {
4084 int64_t pd = rdc->tx_select_diff[i];
4085 if (i == TX_MODE_SELECT)
4086 pd -= RDCOST(cpi->td.mb.rdmult, cpi->td.mb.rddiv, 2048 * (TX_SIZES - 1),
4088 tx_thrs[i] = (tx_thrs[i] + (int)(pd / cm->MBs)) / 2;
4091 if (cm->reference_mode == REFERENCE_MODE_SELECT) {
4092 int single_count_zero = 0;
4093 int comp_count_zero = 0;
4095 for (i = 0; i < COMP_INTER_CONTEXTS; i++) {
4096 single_count_zero += counts->comp_inter[i][0];
4097 comp_count_zero += counts->comp_inter[i][1];
4100 if (comp_count_zero == 0) {
4101 cm->reference_mode = SINGLE_REFERENCE;
4102 vp9_zero(counts->comp_inter);
4103 } else if (single_count_zero == 0) {
4104 cm->reference_mode = COMPOUND_REFERENCE;
4105 vp9_zero(counts->comp_inter);
4109 if (cm->tx_mode == TX_MODE_SELECT) {
4111 int count8x8_lp = 0, count8x8_8x8p = 0;
4112 int count16x16_16x16p = 0, count16x16_lp = 0;
4115 for (i = 0; i < TX_SIZE_CONTEXTS; ++i) {
4116 count4x4 += counts->tx.p32x32[i][TX_4X4];
4117 count4x4 += counts->tx.p16x16[i][TX_4X4];
4118 count4x4 += counts->tx.p8x8[i][TX_4X4];
4120 count8x8_lp += counts->tx.p32x32[i][TX_8X8];
4121 count8x8_lp += counts->tx.p16x16[i][TX_8X8];
4122 count8x8_8x8p += counts->tx.p8x8[i][TX_8X8];
4124 count16x16_16x16p += counts->tx.p16x16[i][TX_16X16];
4125 count16x16_lp += counts->tx.p32x32[i][TX_16X16];
4126 count32x32 += counts->tx.p32x32[i][TX_32X32];
4128 if (count4x4 == 0 && count16x16_lp == 0 && count16x16_16x16p == 0 &&
4130 cm->tx_mode = ALLOW_8X8;
4131 reset_skip_tx_size(cm, TX_8X8);
4132 } else if (count8x8_8x8p == 0 && count16x16_16x16p == 0 &&
4133 count8x8_lp == 0 && count16x16_lp == 0 && count32x32 == 0) {
4134 cm->tx_mode = ONLY_4X4;
4135 reset_skip_tx_size(cm, TX_4X4);
4136 } else if (count8x8_lp == 0 && count16x16_lp == 0 && count4x4 == 0) {
4137 cm->tx_mode = ALLOW_32X32;
4138 } else if (count32x32 == 0 && count8x8_lp == 0 && count4x4 == 0) {
4139 cm->tx_mode = ALLOW_16X16;
4140 reset_skip_tx_size(cm, TX_16X16);
4144 cm->reference_mode = SINGLE_REFERENCE;
4145 encode_frame_internal(cpi);
4149 static void sum_intra_stats(FRAME_COUNTS *counts, const MODE_INFO *mi) {
4150 const PREDICTION_MODE y_mode = mi->mbmi.mode;
4151 const PREDICTION_MODE uv_mode = mi->mbmi.uv_mode;
4152 const BLOCK_SIZE bsize = mi->mbmi.sb_type;
4154 if (bsize < BLOCK_8X8) {
4156 const int num_4x4_w = num_4x4_blocks_wide_lookup[bsize];
4157 const int num_4x4_h = num_4x4_blocks_high_lookup[bsize];
4158 for (idy = 0; idy < 2; idy += num_4x4_h)
4159 for (idx = 0; idx < 2; idx += num_4x4_w)
4160 ++counts->y_mode[0][mi->bmi[idy * 2 + idx].as_mode];
4162 ++counts->y_mode[size_group_lookup[bsize]][y_mode];
4165 ++counts->uv_mode[y_mode][uv_mode];
4168 static void encode_superblock(VP9_COMP *cpi, ThreadData *td,
4169 TOKENEXTRA **t, int output_enabled,
4170 int mi_row, int mi_col, BLOCK_SIZE bsize,
4171 PICK_MODE_CONTEXT *ctx) {
4172 VP9_COMMON *const cm = &cpi->common;
4173 MACROBLOCK *const x = &td->mb;
4174 MACROBLOCKD *const xd = &x->e_mbd;
4175 MODE_INFO **mi_8x8 = xd->mi;
4176 MODE_INFO *mi = mi_8x8[0];
4177 MB_MODE_INFO *mbmi = &mi->mbmi;
4178 const int seg_skip = vp9_segfeature_active(&cm->seg, mbmi->segment_id,
4180 const int mis = cm->mi_stride;
4181 const int mi_width = num_8x8_blocks_wide_lookup[bsize];
4182 const int mi_height = num_8x8_blocks_high_lookup[bsize];
4184 x->skip_recode = !x->select_tx_size && mbmi->sb_type >= BLOCK_8X8 &&
4185 cpi->oxcf.aq_mode != COMPLEXITY_AQ &&
4186 cpi->oxcf.aq_mode != CYCLIC_REFRESH_AQ &&
4187 cpi->sf.allow_skip_recode;
4189 if (!x->skip_recode && !cpi->sf.use_nonrd_pick_mode)
4190 memset(x->skip_txfm, 0, sizeof(x->skip_txfm));
4192 x->skip_optimize = ctx->is_coded;
4194 x->use_lp32x32fdct = cpi->sf.use_lp32x32fdct;
4195 x->skip_encode = (!output_enabled && cpi->sf.skip_encode_frame &&
4196 x->q_index < QIDX_SKIP_THRESH);
4201 set_ref_ptrs(cm, xd, mbmi->ref_frame[0], mbmi->ref_frame[1]);
4203 if (!is_inter_block(mbmi)) {
4206 for (plane = 0; plane < MAX_MB_PLANE; ++plane)
4207 vp9_encode_intra_block_plane(x, MAX(bsize, BLOCK_8X8), plane);
4209 sum_intra_stats(td->counts, mi);
4210 vp9_tokenize_sb(cpi, td, t, !output_enabled, MAX(bsize, BLOCK_8X8));
4213 const int is_compound = has_second_ref(mbmi);
4214 for (ref = 0; ref < 1 + is_compound; ++ref) {
4215 YV12_BUFFER_CONFIG *cfg = get_ref_frame_buffer(cpi,
4216 mbmi->ref_frame[ref]);
4217 assert(cfg != NULL);
4218 vp9_setup_pre_planes(xd, ref, cfg, mi_row, mi_col,
4219 &xd->block_refs[ref]->sf);
4221 if (!(cpi->sf.reuse_inter_pred_sby && ctx->pred_pixel_ready) || seg_skip)
4222 vp9_build_inter_predictors_sby(xd, mi_row, mi_col, MAX(bsize, BLOCK_8X8));
4224 vp9_build_inter_predictors_sbuv(xd, mi_row, mi_col, MAX(bsize, BLOCK_8X8));
4226 vp9_encode_sb(x, MAX(bsize, BLOCK_8X8));
4227 vp9_tokenize_sb(cpi, td, t, !output_enabled, MAX(bsize, BLOCK_8X8));
4230 if (output_enabled) {
4231 if (cm->tx_mode == TX_MODE_SELECT &&
4232 mbmi->sb_type >= BLOCK_8X8 &&
4233 !(is_inter_block(mbmi) && (mbmi->skip || seg_skip))) {
4234 ++get_tx_counts(max_txsize_lookup[bsize], vp9_get_tx_size_context(xd),
4235 &td->counts->tx)[mbmi->tx_size];
4239 // The new intra coding scheme requires no change of transform size
4240 if (is_inter_block(&mi->mbmi)) {
4241 tx_size = MIN(tx_mode_to_biggest_tx_size[cm->tx_mode],
4242 max_txsize_lookup[bsize]);
4244 tx_size = (bsize >= BLOCK_8X8) ? mbmi->tx_size : TX_4X4;
4247 for (y = 0; y < mi_height; y++)
4248 for (x = 0; x < mi_width; x++)
4249 if (mi_col + x < cm->mi_cols && mi_row + y < cm->mi_rows)
4250 mi_8x8[mis * y + x]->mbmi.tx_size = tx_size;
4252 ++td->counts->tx.tx_totals[mbmi->tx_size];
4253 ++td->counts->tx.tx_totals[get_uv_tx_size(mbmi, &xd->plane[1])];