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 static BLOCK_SIZE get_nonrd_var_based_fixed_partition(VP9_COMP *cpi,
172 unsigned int var = get_sby_perpixel_diff_variance(cpi, &x->plane[0].src,
183 // Lighter version of set_offsets that only sets the mode info
185 static INLINE void set_mode_info_offsets(VP9_COMMON *const cm,
186 MACROBLOCKD *const xd,
189 const int idx_str = xd->mi_stride * mi_row + mi_col;
190 xd->mi = cm->mi + idx_str;
191 xd->mi[0].src_mi = &xd->mi[0];
194 static void set_offsets(VP9_COMP *cpi, const TileInfo *const tile,
195 MACROBLOCK *const x, int mi_row, int mi_col,
197 VP9_COMMON *const cm = &cpi->common;
198 MACROBLOCKD *const xd = &x->e_mbd;
200 const int mi_width = num_8x8_blocks_wide_lookup[bsize];
201 const int mi_height = num_8x8_blocks_high_lookup[bsize];
202 const struct segmentation *const seg = &cm->seg;
204 set_skip_context(xd, mi_row, mi_col);
206 set_mode_info_offsets(cm, xd, mi_row, mi_col);
208 mbmi = &xd->mi[0].src_mi->mbmi;
210 // Set up destination pointers.
211 vp9_setup_dst_planes(xd->plane, get_frame_new_buffer(cm), mi_row, mi_col);
213 // Set up limit values for MV components.
214 // Mv beyond the range do not produce new/different prediction block.
215 x->mv_row_min = -(((mi_row + mi_height) * MI_SIZE) + VP9_INTERP_EXTEND);
216 x->mv_col_min = -(((mi_col + mi_width) * MI_SIZE) + VP9_INTERP_EXTEND);
217 x->mv_row_max = (cm->mi_rows - mi_row) * MI_SIZE + VP9_INTERP_EXTEND;
218 x->mv_col_max = (cm->mi_cols - mi_col) * MI_SIZE + VP9_INTERP_EXTEND;
220 // Set up distance of MB to edge of frame in 1/8th pel units.
221 assert(!(mi_col & (mi_width - 1)) && !(mi_row & (mi_height - 1)));
222 set_mi_row_col(xd, tile, mi_row, mi_height, mi_col, mi_width,
223 cm->mi_rows, cm->mi_cols);
225 // Set up source buffers.
226 vp9_setup_src_planes(x, cpi->Source, mi_row, mi_col);
229 x->rddiv = cpi->rd.RDDIV;
230 x->rdmult = cpi->rd.RDMULT;
234 if (cpi->oxcf.aq_mode != VARIANCE_AQ) {
235 const uint8_t *const map = seg->update_map ? cpi->segmentation_map
236 : cm->last_frame_seg_map;
237 mbmi->segment_id = vp9_get_segment_id(cm, map, bsize, mi_row, mi_col);
239 vp9_init_plane_quantizers(cpi, x);
241 x->encode_breakout = cpi->segment_encode_breakout[mbmi->segment_id];
243 mbmi->segment_id = 0;
244 x->encode_breakout = cpi->encode_breakout;
248 static void duplicate_mode_info_in_sb(VP9_COMMON *cm, MACROBLOCKD *xd,
249 int mi_row, int mi_col,
251 const int block_width = num_8x8_blocks_wide_lookup[bsize];
252 const int block_height = num_8x8_blocks_high_lookup[bsize];
254 for (j = 0; j < block_height; ++j)
255 for (i = 0; i < block_width; ++i) {
256 if (mi_row + j < cm->mi_rows && mi_col + i < cm->mi_cols)
257 xd->mi[j * xd->mi_stride + i].src_mi = &xd->mi[0];
261 static void set_block_size(VP9_COMP * const cpi,
262 MACROBLOCKD *const xd,
263 int mi_row, int mi_col,
265 if (cpi->common.mi_cols > mi_col && cpi->common.mi_rows > mi_row) {
266 set_mode_info_offsets(&cpi->common, xd, mi_row, mi_col);
267 xd->mi[0].src_mi->mbmi.sb_type = bsize;
272 int64_t sum_square_error;
282 } partition_variance;
285 partition_variance part_variances;
290 partition_variance part_variances;
295 partition_variance part_variances;
300 partition_variance part_variances;
305 partition_variance part_variances;
310 partition_variance *part_variances;
320 static void tree_to_node(void *data, BLOCK_SIZE bsize, variance_node *node) {
322 node->part_variances = NULL;
325 v64x64 *vt = (v64x64 *) data;
326 node->part_variances = &vt->part_variances;
327 for (i = 0; i < 4; i++)
328 node->split[i] = &vt->split[i].part_variances.none;
332 v32x32 *vt = (v32x32 *) data;
333 node->part_variances = &vt->part_variances;
334 for (i = 0; i < 4; i++)
335 node->split[i] = &vt->split[i].part_variances.none;
339 v16x16 *vt = (v16x16 *) data;
340 node->part_variances = &vt->part_variances;
341 for (i = 0; i < 4; i++)
342 node->split[i] = &vt->split[i].part_variances.none;
346 v8x8 *vt = (v8x8 *) data;
347 node->part_variances = &vt->part_variances;
348 for (i = 0; i < 4; i++)
349 node->split[i] = &vt->split[i].part_variances.none;
353 v4x4 *vt = (v4x4 *) data;
354 node->part_variances = &vt->part_variances;
355 for (i = 0; i < 4; i++)
356 node->split[i] = &vt->split[i];
366 // Set variance values given sum square error, sum error, count.
367 static void fill_variance(int64_t s2, int64_t s, int c, var *v) {
368 v->sum_square_error = s2;
373 static void get_variance(var *v) {
374 v->variance = (int)(256 * (v->sum_square_error -
375 ((v->sum_error * v->sum_error) >> v->log2_count)) >> v->log2_count);
378 void sum_2_variances(const var *a, const var *b, var *r) {
379 assert(a->log2_count == b->log2_count);
380 fill_variance(a->sum_square_error + b->sum_square_error,
381 a->sum_error + b->sum_error, a->log2_count + 1, r);
384 static void fill_variance_tree(void *data, BLOCK_SIZE bsize) {
386 tree_to_node(data, bsize, &node);
387 sum_2_variances(node.split[0], node.split[1], &node.part_variances->horz[0]);
388 sum_2_variances(node.split[2], node.split[3], &node.part_variances->horz[1]);
389 sum_2_variances(node.split[0], node.split[2], &node.part_variances->vert[0]);
390 sum_2_variances(node.split[1], node.split[3], &node.part_variances->vert[1]);
391 sum_2_variances(&node.part_variances->vert[0], &node.part_variances->vert[1],
392 &node.part_variances->none);
395 static int set_vt_partitioning(VP9_COMP *cpi,
396 MACROBLOCKD *const xd,
402 BLOCK_SIZE bsize_min,
404 VP9_COMMON * const cm = &cpi->common;
406 const int block_width = num_8x8_blocks_wide_lookup[bsize];
407 const int block_height = num_8x8_blocks_high_lookup[bsize];
409 assert(block_height == block_width);
410 tree_to_node(data, bsize, &vt);
415 // For bsize=bsize_min (16x16/8x8 for 8x8/4x4 downsampling), select if
416 // variance is below threshold, otherwise split will be selected.
417 // No check for vert/horiz split as too few samples for variance.
418 if (bsize == bsize_min) {
419 get_variance(&vt.part_variances->none);
420 if (mi_col + block_width / 2 < cm->mi_cols &&
421 mi_row + block_height / 2 < cm->mi_rows &&
422 vt.part_variances->none.variance < threshold) {
423 set_block_size(cpi, xd, mi_row, mi_col, bsize);
427 } else if (bsize > bsize_min) {
428 // Variance is already computed for 32x32 blocks to set the force_split.
429 if (bsize != BLOCK_32X32)
430 get_variance(&vt.part_variances->none);
431 // For key frame or low_res: for bsize above 32X32 or very high variance,
433 if (cm->frame_type == KEY_FRAME &&
434 (bsize > BLOCK_32X32 ||
435 vt.part_variances->none.variance > (threshold << 4))) {
438 // If variance is low, take the bsize (no split).
439 if (mi_col + block_width / 2 < cm->mi_cols &&
440 mi_row + block_height / 2 < cm->mi_rows &&
441 vt.part_variances->none.variance < threshold) {
442 set_block_size(cpi, xd, mi_row, mi_col, bsize);
446 // Check vertical split.
447 if (mi_row + block_height / 2 < cm->mi_rows) {
448 get_variance(&vt.part_variances->vert[0]);
449 get_variance(&vt.part_variances->vert[1]);
450 if (vt.part_variances->vert[0].variance < threshold &&
451 vt.part_variances->vert[1].variance < threshold) {
452 BLOCK_SIZE subsize = get_subsize(bsize, PARTITION_VERT);
453 set_block_size(cpi, xd, mi_row, mi_col, subsize);
454 set_block_size(cpi, xd, mi_row, mi_col + block_width / 2, subsize);
458 // Check horizontal split.
459 if (mi_col + block_width / 2 < cm->mi_cols) {
460 get_variance(&vt.part_variances->horz[0]);
461 get_variance(&vt.part_variances->horz[1]);
462 if (vt.part_variances->horz[0].variance < threshold &&
463 vt.part_variances->horz[1].variance < threshold) {
464 BLOCK_SIZE subsize = get_subsize(bsize, PARTITION_HORZ);
465 set_block_size(cpi, xd, mi_row, mi_col, subsize);
466 set_block_size(cpi, xd, mi_row + block_height / 2, mi_col, subsize);
477 void vp9_set_vbp_thresholds(VP9_COMP *cpi, int q) {
478 SPEED_FEATURES *const sf = &cpi->sf;
479 if (sf->partition_search_type != VAR_BASED_PARTITION &&
480 sf->partition_search_type != REFERENCE_PARTITION) {
483 VP9_COMMON *const cm = &cpi->common;
484 const int is_key_frame = (cm->frame_type == KEY_FRAME);
485 const int threshold_multiplier = is_key_frame ? 20 : 1;
486 const int64_t threshold_base = (int64_t)(threshold_multiplier *
487 cpi->y_dequant[q][1]);
489 // TODO(marpan): Allow 4x4 partitions for inter-frames.
490 // use_4x4_partition = (variance4x4downsample[i2 + j] == 1);
491 // If 4x4 partition is not used, then 8x8 partition will be selected
492 // if variance of 16x16 block is very high, so use larger threshold
493 // for 16x16 (threshold_bsize_min) in that case.
495 cpi->vbp_threshold_64x64 = threshold_base;
496 cpi->vbp_threshold_32x32 = threshold_base >> 2;
497 cpi->vbp_threshold_16x16 = threshold_base >> 2;
498 cpi->vbp_threshold_8x8 = threshold_base << 2;
499 cpi->vbp_bsize_min = BLOCK_8X8;
501 cpi->vbp_threshold_32x32 = threshold_base;
502 if (cm->width <= 352 && cm->height <= 288) {
503 cpi->vbp_threshold_64x64 = threshold_base >> 2;
504 cpi->vbp_threshold_16x16 = threshold_base << 3;
506 cpi->vbp_threshold_64x64 = threshold_base;
507 cpi->vbp_threshold_16x16 = threshold_base << cpi->oxcf.speed;
509 cpi->vbp_bsize_min = BLOCK_16X16;
514 static void fill_variance_4x4avg(const uint8_t *s, int sp, const uint8_t *d,
515 int dp, int x8_idx, int y8_idx, v8x8 *vst,
516 #if CONFIG_VP9_HIGHBITDEPTH
523 for (k = 0; k < 4; k++) {
524 int x4_idx = x8_idx + ((k & 1) << 2);
525 int y4_idx = y8_idx + ((k >> 1) << 2);
526 unsigned int sse = 0;
528 if (x4_idx < pixels_wide && y4_idx < pixels_high) {
531 #if CONFIG_VP9_HIGHBITDEPTH
532 if (highbd_flag & YV12_FLAG_HIGHBITDEPTH) {
533 s_avg = vp9_highbd_avg_4x4(s + y4_idx * sp + x4_idx, sp);
535 d_avg = vp9_highbd_avg_4x4(d + y4_idx * dp + x4_idx, dp);
537 s_avg = vp9_avg_4x4(s + y4_idx * sp + x4_idx, sp);
539 d_avg = vp9_avg_4x4(d + y4_idx * dp + x4_idx, dp);
542 s_avg = vp9_avg_4x4(s + y4_idx * sp + x4_idx, sp);
544 d_avg = vp9_avg_4x4(d + y4_idx * dp + x4_idx, dp);
549 fill_variance(sse, sum, 0, &vst->split[k].part_variances.none);
553 static void fill_variance_8x8avg(const uint8_t *s, int sp, const uint8_t *d,
554 int dp, int x16_idx, int y16_idx, v16x16 *vst,
555 #if CONFIG_VP9_HIGHBITDEPTH
562 for (k = 0; k < 4; k++) {
563 int x8_idx = x16_idx + ((k & 1) << 3);
564 int y8_idx = y16_idx + ((k >> 1) << 3);
565 unsigned int sse = 0;
567 if (x8_idx < pixels_wide && y8_idx < pixels_high) {
570 #if CONFIG_VP9_HIGHBITDEPTH
571 if (highbd_flag & YV12_FLAG_HIGHBITDEPTH) {
572 s_avg = vp9_highbd_avg_8x8(s + y8_idx * sp + x8_idx, sp);
574 d_avg = vp9_highbd_avg_8x8(d + y8_idx * dp + x8_idx, dp);
576 s_avg = vp9_avg_8x8(s + y8_idx * sp + x8_idx, sp);
578 d_avg = vp9_avg_8x8(d + y8_idx * dp + x8_idx, dp);
581 s_avg = vp9_avg_8x8(s + y8_idx * sp + x8_idx, sp);
583 d_avg = vp9_avg_8x8(d + y8_idx * dp + x8_idx, dp);
588 fill_variance(sse, sum, 0, &vst->split[k].part_variances.none);
592 // This function chooses partitioning based on the variance between source and
593 // reconstructed last, where variance is computed for down-sampled inputs.
594 static void choose_partitioning(VP9_COMP *cpi,
595 const TileInfo *const tile,
597 int mi_row, int mi_col) {
598 VP9_COMMON * const cm = &cpi->common;
599 MACROBLOCKD *xd = &x->e_mbd;
608 int pixels_wide = 64, pixels_high = 64;
610 // Always use 4x4 partition for key frame.
611 const int is_key_frame = (cm->frame_type == KEY_FRAME);
612 const int use_4x4_partition = is_key_frame;
613 const int low_res = (cm->width <= 352 && cm->height <= 288);
614 int variance4x4downsample[16];
616 int segment_id = CR_SEGMENT_ID_BASE;
617 if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ && cm->seg.enabled) {
618 const uint8_t *const map = cm->seg.update_map ? cpi->segmentation_map :
619 cm->last_frame_seg_map;
620 segment_id = vp9_get_segment_id(cm, map, BLOCK_64X64, mi_row, mi_col);
623 set_offsets(cpi, tile, x, mi_row, mi_col, BLOCK_64X64);
625 if (xd->mb_to_right_edge < 0)
626 pixels_wide += (xd->mb_to_right_edge >> 3);
627 if (xd->mb_to_bottom_edge < 0)
628 pixels_high += (xd->mb_to_bottom_edge >> 3);
630 s = x->plane[0].src.buf;
631 sp = x->plane[0].src.stride;
634 MB_MODE_INFO *mbmi = &xd->mi[0].src_mi->mbmi;
636 const YV12_BUFFER_CONFIG *yv12 = get_ref_frame_buffer(cpi, LAST_FRAME);
638 const YV12_BUFFER_CONFIG *yv12_g = get_ref_frame_buffer(cpi, GOLDEN_FRAME);
639 unsigned int y_sad, y_sad_g;
640 const BLOCK_SIZE bsize = BLOCK_32X32
641 + (mi_col + 4 < cm->mi_cols) * 2 + (mi_row + 4 < cm->mi_rows);
643 assert(yv12 != NULL);
644 if (yv12_g && yv12_g != yv12) {
645 vp9_setup_pre_planes(xd, 0, yv12_g, mi_row, mi_col,
646 &cm->frame_refs[GOLDEN_FRAME - 1].sf);
647 y_sad_g = cpi->fn_ptr[bsize].sdf(x->plane[0].src.buf,
648 x->plane[0].src.stride,
649 xd->plane[0].pre[0].buf,
650 xd->plane[0].pre[0].stride);
655 vp9_setup_pre_planes(xd, 0, yv12, mi_row, mi_col,
656 &cm->frame_refs[LAST_FRAME - 1].sf);
657 mbmi->ref_frame[0] = LAST_FRAME;
658 mbmi->ref_frame[1] = NONE;
659 mbmi->sb_type = BLOCK_64X64;
660 mbmi->mv[0].as_int = 0;
661 mbmi->interp_filter = BILINEAR;
663 y_sad = vp9_int_pro_motion_estimation(cpi, x, bsize);
664 if (y_sad_g < y_sad) {
665 vp9_setup_pre_planes(xd, 0, yv12_g, mi_row, mi_col,
666 &cm->frame_refs[GOLDEN_FRAME - 1].sf);
667 mbmi->ref_frame[0] = GOLDEN_FRAME;
668 mbmi->mv[0].as_int = 0;
671 x->pred_mv[LAST_FRAME] = mbmi->mv[0].as_mv;
674 vp9_build_inter_predictors_sb(xd, mi_row, mi_col, BLOCK_64X64);
676 for (i = 1; i <= 2; ++i) {
677 struct macroblock_plane *p = &x->plane[i];
678 struct macroblockd_plane *pd = &xd->plane[i];
679 const BLOCK_SIZE bs = get_plane_block_size(bsize, pd);
681 if (bs == BLOCK_INVALID)
684 uv_sad = cpi->fn_ptr[bs].sdf(p->src.buf, p->src.stride,
685 pd->dst.buf, pd->dst.stride);
687 x->color_sensitivity[i - 1] = uv_sad > (y_sad >> 2);
690 d = xd->plane[0].dst.buf;
691 dp = xd->plane[0].dst.stride;
695 #if CONFIG_VP9_HIGHBITDEPTH
696 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
699 d = CONVERT_TO_BYTEPTR(VP9_HIGH_VAR_OFFS_10);
702 d = CONVERT_TO_BYTEPTR(VP9_HIGH_VAR_OFFS_12);
706 d = CONVERT_TO_BYTEPTR(VP9_HIGH_VAR_OFFS_8);
710 #endif // CONFIG_VP9_HIGHBITDEPTH
713 // Index for force_split: 0 for 64x64, 1-4 for 32x32 blocks,
715 // Fill in the entire tree of 8x8 (or 4x4 under some conditions) variances
717 for (i = 0; i < 4; i++) {
718 const int x32_idx = ((i & 1) << 5);
719 const int y32_idx = ((i >> 1) << 5);
720 const int i2 = i << 2;
721 force_split[i + 1] = 0;
722 for (j = 0; j < 4; j++) {
723 const int x16_idx = x32_idx + ((j & 1) << 4);
724 const int y16_idx = y32_idx + ((j >> 1) << 4);
725 v16x16 *vst = &vt.split[i].split[j];
726 variance4x4downsample[i2 + j] = 0;
728 fill_variance_8x8avg(s, sp, d, dp, x16_idx, y16_idx, vst,
729 #if CONFIG_VP9_HIGHBITDEPTH
735 fill_variance_tree(&vt.split[i].split[j], BLOCK_16X16);
736 // For low-resolution, compute the variance based on 8x8 down-sampling,
737 // and if it is large (above the threshold) we go down for 4x4.
738 // For key frame we always go down to 4x4.
740 get_variance(&vt.split[i].split[j].part_variances.none);
742 if (is_key_frame || (low_res &&
743 vt.split[i].split[j].part_variances.none.variance >
744 (cpi->vbp_threshold_32x32 << 1))) {
745 // Go down to 4x4 down-sampling for variance.
746 variance4x4downsample[i2 + j] = 1;
747 for (k = 0; k < 4; k++) {
748 int x8_idx = x16_idx + ((k & 1) << 3);
749 int y8_idx = y16_idx + ((k >> 1) << 3);
750 v8x8 *vst2 = is_key_frame ? &vst->split[k] :
751 &vt2[i2 + j].split[k];
752 fill_variance_4x4avg(s, sp, d, dp, x8_idx, y8_idx, vst2,
753 #if CONFIG_VP9_HIGHBITDEPTH
764 // No 64x64 blocks on segments other than base (un-boosted) segment,
766 if (cyclic_refresh_segment_id_boosted(segment_id))
769 // Fill the rest of the variance tree by summing split partition values.
770 for (i = 0; i < 4; i++) {
771 const int i2 = i << 2;
772 for (j = 0; j < 4; j++) {
773 if (variance4x4downsample[i2 + j] == 1) {
774 v16x16 *vtemp = (!is_key_frame) ? &vt2[i2 + j] :
775 &vt.split[i].split[j];
776 for (m = 0; m < 4; m++)
777 fill_variance_tree(&vtemp->split[m], BLOCK_8X8);
778 fill_variance_tree(vtemp, BLOCK_16X16);
781 fill_variance_tree(&vt.split[i], BLOCK_32X32);
782 // If variance of this 32x32 block is above the threshold, force the block
783 // to split. This also forces a split on the upper (64x64) level.
784 get_variance(&vt.split[i].part_variances.none);
785 if (vt.split[i].part_variances.none.variance > cpi->vbp_threshold_32x32) {
786 force_split[i + 1] = 1;
791 fill_variance_tree(&vt, BLOCK_64X64);
793 // Now go through the entire structure, splitting every block size until
794 // we get to one that's got a variance lower than our threshold.
795 if ( mi_col + 8 > cm->mi_cols || mi_row + 8 > cm->mi_rows ||
796 !set_vt_partitioning(cpi, xd, &vt, BLOCK_64X64, mi_row, mi_col,
797 cpi->vbp_threshold_64x64, BLOCK_16X16,
799 for (i = 0; i < 4; ++i) {
800 const int x32_idx = ((i & 1) << 2);
801 const int y32_idx = ((i >> 1) << 2);
802 const int i2 = i << 2;
803 if (!set_vt_partitioning(cpi, xd, &vt.split[i], BLOCK_32X32,
804 (mi_row + y32_idx), (mi_col + x32_idx),
805 cpi->vbp_threshold_32x32,
806 BLOCK_16X16, force_split[i + 1])) {
807 for (j = 0; j < 4; ++j) {
808 const int x16_idx = ((j & 1) << 1);
809 const int y16_idx = ((j >> 1) << 1);
810 // For inter frames: if variance4x4downsample[] == 1 for this 16x16
811 // block, then the variance is based on 4x4 down-sampling, so use vt2
812 // in set_vt_partioning(), otherwise use vt.
813 v16x16 *vtemp = (!is_key_frame &&
814 variance4x4downsample[i2 + j] == 1) ?
815 &vt2[i2 + j] : &vt.split[i].split[j];
816 if (!set_vt_partitioning(cpi, xd, vtemp, BLOCK_16X16,
817 mi_row + y32_idx + y16_idx,
818 mi_col + x32_idx + x16_idx,
819 cpi->vbp_threshold_16x16,
820 cpi->vbp_bsize_min, 0)) {
821 for (k = 0; k < 4; ++k) {
822 const int x8_idx = (k & 1);
823 const int y8_idx = (k >> 1);
824 if (use_4x4_partition) {
825 if (!set_vt_partitioning(cpi, xd, &vtemp->split[k],
827 mi_row + y32_idx + y16_idx + y8_idx,
828 mi_col + x32_idx + x16_idx + x8_idx,
829 cpi->vbp_threshold_8x8,
831 set_block_size(cpi, xd,
832 (mi_row + y32_idx + y16_idx + y8_idx),
833 (mi_col + x32_idx + x16_idx + x8_idx),
837 set_block_size(cpi, xd,
838 (mi_row + y32_idx + y16_idx + y8_idx),
839 (mi_col + x32_idx + x16_idx + x8_idx),
850 static void update_state(VP9_COMP *cpi, ThreadData *td,
851 PICK_MODE_CONTEXT *ctx,
852 int mi_row, int mi_col, BLOCK_SIZE bsize,
853 int output_enabled) {
855 VP9_COMMON *const cm = &cpi->common;
856 RD_COUNTS *const rdc = &td->rd_counts;
857 MACROBLOCK *const x = &td->mb;
858 MACROBLOCKD *const xd = &x->e_mbd;
859 struct macroblock_plane *const p = x->plane;
860 struct macroblockd_plane *const pd = xd->plane;
861 MODE_INFO *mi = &ctx->mic;
862 MB_MODE_INFO *const mbmi = &xd->mi[0].src_mi->mbmi;
863 MODE_INFO *mi_addr = &xd->mi[0];
864 const struct segmentation *const seg = &cm->seg;
865 const int bw = num_8x8_blocks_wide_lookup[mi->mbmi.sb_type];
866 const int bh = num_8x8_blocks_high_lookup[mi->mbmi.sb_type];
867 const int x_mis = MIN(bw, cm->mi_cols - mi_col);
868 const int y_mis = MIN(bh, cm->mi_rows - mi_row);
869 MV_REF *const frame_mvs =
870 cm->cur_frame->mvs + mi_row * cm->mi_cols + mi_col;
873 const int mis = cm->mi_stride;
874 const int mi_width = num_8x8_blocks_wide_lookup[bsize];
875 const int mi_height = num_8x8_blocks_high_lookup[bsize];
878 assert(mi->mbmi.sb_type == bsize);
881 mi_addr->src_mi = mi_addr;
883 // If segmentation in use
885 // For in frame complexity AQ copy the segment id from the segment map.
886 if (cpi->oxcf.aq_mode == COMPLEXITY_AQ) {
887 const uint8_t *const map = seg->update_map ? cpi->segmentation_map
888 : cm->last_frame_seg_map;
889 mi_addr->mbmi.segment_id =
890 vp9_get_segment_id(cm, map, bsize, mi_row, mi_col);
892 // Else for cyclic refresh mode update the segment map, set the segment id
893 // and then update the quantizer.
894 if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ) {
895 vp9_cyclic_refresh_update_segment(cpi, &xd->mi[0].src_mi->mbmi, mi_row,
896 mi_col, bsize, ctx->rate, ctx->dist,
901 max_plane = is_inter_block(mbmi) ? MAX_MB_PLANE : 1;
902 for (i = 0; i < max_plane; ++i) {
903 p[i].coeff = ctx->coeff_pbuf[i][1];
904 p[i].qcoeff = ctx->qcoeff_pbuf[i][1];
905 pd[i].dqcoeff = ctx->dqcoeff_pbuf[i][1];
906 p[i].eobs = ctx->eobs_pbuf[i][1];
909 for (i = max_plane; i < MAX_MB_PLANE; ++i) {
910 p[i].coeff = ctx->coeff_pbuf[i][2];
911 p[i].qcoeff = ctx->qcoeff_pbuf[i][2];
912 pd[i].dqcoeff = ctx->dqcoeff_pbuf[i][2];
913 p[i].eobs = ctx->eobs_pbuf[i][2];
916 // Restore the coding context of the MB to that that was in place
917 // when the mode was picked for it
918 for (y = 0; y < mi_height; y++)
919 for (x_idx = 0; x_idx < mi_width; x_idx++)
920 if ((xd->mb_to_right_edge >> (3 + MI_SIZE_LOG2)) + mi_width > x_idx
921 && (xd->mb_to_bottom_edge >> (3 + MI_SIZE_LOG2)) + mi_height > y) {
922 xd->mi[x_idx + y * mis].src_mi = mi_addr;
925 if (cpi->oxcf.aq_mode)
926 vp9_init_plane_quantizers(cpi, x);
928 // FIXME(rbultje) I'm pretty sure this should go to the end of this block
929 // (i.e. after the output_enabled)
930 if (bsize < BLOCK_32X32) {
931 if (bsize < BLOCK_16X16)
932 ctx->tx_rd_diff[ALLOW_16X16] = ctx->tx_rd_diff[ALLOW_8X8];
933 ctx->tx_rd_diff[ALLOW_32X32] = ctx->tx_rd_diff[ALLOW_16X16];
936 if (is_inter_block(mbmi) && mbmi->sb_type < BLOCK_8X8) {
937 mbmi->mv[0].as_int = mi->bmi[3].as_mv[0].as_int;
938 mbmi->mv[1].as_int = mi->bmi[3].as_mv[1].as_int;
942 vpx_memcpy(x->zcoeff_blk[mbmi->tx_size], ctx->zcoeff_blk,
943 sizeof(uint8_t) * ctx->num_4x4_blk);
948 if (!vp9_segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_SKIP)) {
949 for (i = 0; i < TX_MODES; i++)
950 rdc->tx_select_diff[i] += ctx->tx_rd_diff[i];
953 #if CONFIG_INTERNAL_STATS
954 if (frame_is_intra_only(cm)) {
955 static const int kf_mode_index[] = {
957 THR_V_PRED /*V_PRED*/,
958 THR_H_PRED /*H_PRED*/,
959 THR_D45_PRED /*D45_PRED*/,
960 THR_D135_PRED /*D135_PRED*/,
961 THR_D117_PRED /*D117_PRED*/,
962 THR_D153_PRED /*D153_PRED*/,
963 THR_D207_PRED /*D207_PRED*/,
964 THR_D63_PRED /*D63_PRED*/,
967 ++cpi->mode_chosen_counts[kf_mode_index[mbmi->mode]];
969 // Note how often each mode chosen as best
970 ++cpi->mode_chosen_counts[ctx->best_mode_index];
973 if (!frame_is_intra_only(cm)) {
974 if (is_inter_block(mbmi)) {
975 vp9_update_mv_count(td);
977 if (cm->interp_filter == SWITCHABLE) {
978 const int ctx = vp9_get_pred_context_switchable_interp(xd);
979 ++td->counts->switchable_interp[ctx][mbmi->interp_filter];
983 rdc->comp_pred_diff[SINGLE_REFERENCE] += ctx->single_pred_diff;
984 rdc->comp_pred_diff[COMPOUND_REFERENCE] += ctx->comp_pred_diff;
985 rdc->comp_pred_diff[REFERENCE_MODE_SELECT] += ctx->hybrid_pred_diff;
987 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i)
988 rdc->filter_diff[i] += ctx->best_filter_diff[i];
991 for (h = 0; h < y_mis; ++h) {
992 MV_REF *const frame_mv = frame_mvs + h * cm->mi_cols;
993 for (w = 0; w < x_mis; ++w) {
994 MV_REF *const mv = frame_mv + w;
995 mv->ref_frame[0] = mi->src_mi->mbmi.ref_frame[0];
996 mv->ref_frame[1] = mi->src_mi->mbmi.ref_frame[1];
997 mv->mv[0].as_int = mi->src_mi->mbmi.mv[0].as_int;
998 mv->mv[1].as_int = mi->src_mi->mbmi.mv[1].as_int;
1003 void vp9_setup_src_planes(MACROBLOCK *x, const YV12_BUFFER_CONFIG *src,
1004 int mi_row, int mi_col) {
1005 uint8_t *const buffers[3] = {src->y_buffer, src->u_buffer, src->v_buffer };
1006 const int strides[3] = {src->y_stride, src->uv_stride, src->uv_stride };
1009 // Set current frame pointer.
1010 x->e_mbd.cur_buf = src;
1012 for (i = 0; i < MAX_MB_PLANE; i++)
1013 setup_pred_plane(&x->plane[i].src, buffers[i], strides[i], mi_row, mi_col,
1014 NULL, x->e_mbd.plane[i].subsampling_x,
1015 x->e_mbd.plane[i].subsampling_y);
1018 static void set_mode_info_seg_skip(MACROBLOCK *x, TX_MODE tx_mode,
1019 RD_COST *rd_cost, BLOCK_SIZE bsize) {
1020 MACROBLOCKD *const xd = &x->e_mbd;
1021 MB_MODE_INFO *const mbmi = &xd->mi[0].src_mi->mbmi;
1022 INTERP_FILTER filter_ref;
1024 if (xd->up_available)
1025 filter_ref = xd->mi[-xd->mi_stride].src_mi->mbmi.interp_filter;
1026 else if (xd->left_available)
1027 filter_ref = xd->mi[-1].src_mi->mbmi.interp_filter;
1029 filter_ref = EIGHTTAP;
1031 mbmi->sb_type = bsize;
1032 mbmi->mode = ZEROMV;
1033 mbmi->tx_size = MIN(max_txsize_lookup[bsize],
1034 tx_mode_to_biggest_tx_size[tx_mode]);
1036 mbmi->uv_mode = DC_PRED;
1037 mbmi->ref_frame[0] = LAST_FRAME;
1038 mbmi->ref_frame[1] = NONE;
1039 mbmi->mv[0].as_int = 0;
1040 mbmi->interp_filter = filter_ref;
1042 xd->mi[0].src_mi->bmi[0].as_mv[0].as_int = 0;
1045 vp9_rd_cost_init(rd_cost);
1048 static int set_segment_rdmult(VP9_COMP *const cpi,
1049 MACROBLOCK *const x,
1050 int8_t segment_id) {
1052 VP9_COMMON *const cm = &cpi->common;
1053 vp9_init_plane_quantizers(cpi, x);
1054 vp9_clear_system_state();
1055 segment_qindex = vp9_get_qindex(&cm->seg, segment_id,
1057 return vp9_compute_rd_mult(cpi, segment_qindex + cm->y_dc_delta_q);
1060 static void rd_pick_sb_modes(VP9_COMP *cpi,
1061 TileDataEnc *tile_data,
1062 MACROBLOCK *const x,
1063 int mi_row, int mi_col, RD_COST *rd_cost,
1064 BLOCK_SIZE bsize, PICK_MODE_CONTEXT *ctx,
1066 VP9_COMMON *const cm = &cpi->common;
1067 TileInfo *const tile_info = &tile_data->tile_info;
1068 MACROBLOCKD *const xd = &x->e_mbd;
1070 struct macroblock_plane *const p = x->plane;
1071 struct macroblockd_plane *const pd = xd->plane;
1072 const AQ_MODE aq_mode = cpi->oxcf.aq_mode;
1075 vp9_clear_system_state();
1077 // Use the lower precision, but faster, 32x32 fdct for mode selection.
1078 x->use_lp32x32fdct = 1;
1080 set_offsets(cpi, tile_info, x, mi_row, mi_col, bsize);
1081 mbmi = &xd->mi[0].src_mi->mbmi;
1082 mbmi->sb_type = bsize;
1084 for (i = 0; i < MAX_MB_PLANE; ++i) {
1085 p[i].coeff = ctx->coeff_pbuf[i][0];
1086 p[i].qcoeff = ctx->qcoeff_pbuf[i][0];
1087 pd[i].dqcoeff = ctx->dqcoeff_pbuf[i][0];
1088 p[i].eobs = ctx->eobs_pbuf[i][0];
1092 ctx->pred_pixel_ready = 0;
1095 // Set to zero to make sure we do not use the previous encoded frame stats
1098 #if CONFIG_VP9_HIGHBITDEPTH
1099 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
1100 x->source_variance =
1101 vp9_high_get_sby_perpixel_variance(cpi, &x->plane[0].src,
1104 x->source_variance =
1105 vp9_get_sby_perpixel_variance(cpi, &x->plane[0].src, bsize);
1108 x->source_variance =
1109 vp9_get_sby_perpixel_variance(cpi, &x->plane[0].src, bsize);
1110 #endif // CONFIG_VP9_HIGHBITDEPTH
1112 // Save rdmult before it might be changed, so it can be restored later.
1113 orig_rdmult = x->rdmult;
1115 if (aq_mode == VARIANCE_AQ) {
1116 const int energy = bsize <= BLOCK_16X16 ? x->mb_energy
1117 : vp9_block_energy(cpi, x, bsize);
1118 if (cm->frame_type == KEY_FRAME ||
1119 cpi->refresh_alt_ref_frame ||
1120 (cpi->refresh_golden_frame && !cpi->rc.is_src_frame_alt_ref)) {
1121 mbmi->segment_id = vp9_vaq_segment_id(energy);
1123 const uint8_t *const map = cm->seg.update_map ? cpi->segmentation_map
1124 : cm->last_frame_seg_map;
1125 mbmi->segment_id = vp9_get_segment_id(cm, map, bsize, mi_row, mi_col);
1127 x->rdmult = set_segment_rdmult(cpi, x, mbmi->segment_id);
1128 } else if (aq_mode == COMPLEXITY_AQ) {
1129 x->rdmult = set_segment_rdmult(cpi, x, mbmi->segment_id);
1130 } else if (aq_mode == CYCLIC_REFRESH_AQ) {
1131 const uint8_t *const map = cm->seg.update_map ? cpi->segmentation_map
1132 : cm->last_frame_seg_map;
1133 // If segment is boosted, use rdmult for that segment.
1134 if (cyclic_refresh_segment_id_boosted(
1135 vp9_get_segment_id(cm, map, bsize, mi_row, mi_col)))
1136 x->rdmult = vp9_cyclic_refresh_get_rdmult(cpi->cyclic_refresh);
1139 // Find best coding mode & reconstruct the MB so it is available
1140 // as a predictor for MBs that follow in the SB
1141 if (frame_is_intra_only(cm)) {
1142 vp9_rd_pick_intra_mode_sb(cpi, x, rd_cost, bsize, ctx, best_rd);
1144 if (bsize >= BLOCK_8X8) {
1145 if (vp9_segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_SKIP))
1146 vp9_rd_pick_inter_mode_sb_seg_skip(cpi, tile_data, x, rd_cost, bsize,
1149 vp9_rd_pick_inter_mode_sb(cpi, tile_data, x, mi_row, mi_col,
1150 rd_cost, bsize, ctx, best_rd);
1152 vp9_rd_pick_inter_mode_sub8x8(cpi, tile_data, x, mi_row, mi_col,
1153 rd_cost, bsize, ctx, best_rd);
1158 // Examine the resulting rate and for AQ mode 2 make a segment choice.
1159 if ((rd_cost->rate != INT_MAX) &&
1160 (aq_mode == COMPLEXITY_AQ) && (bsize >= BLOCK_16X16) &&
1161 (cm->frame_type == KEY_FRAME ||
1162 cpi->refresh_alt_ref_frame ||
1163 (cpi->refresh_golden_frame && !cpi->rc.is_src_frame_alt_ref))) {
1164 vp9_caq_select_segment(cpi, x, bsize, mi_row, mi_col, rd_cost->rate);
1167 x->rdmult = orig_rdmult;
1169 // TODO(jingning) The rate-distortion optimization flow needs to be
1170 // refactored to provide proper exit/return handle.
1171 if (rd_cost->rate == INT_MAX)
1172 rd_cost->rdcost = INT64_MAX;
1174 ctx->rate = rd_cost->rate;
1175 ctx->dist = rd_cost->dist;
1178 static void update_stats(VP9_COMMON *cm, ThreadData *td) {
1179 const MACROBLOCK *x = &td->mb;
1180 const MACROBLOCKD *const xd = &x->e_mbd;
1181 const MODE_INFO *const mi = xd->mi[0].src_mi;
1182 const MB_MODE_INFO *const mbmi = &mi->mbmi;
1183 const BLOCK_SIZE bsize = mbmi->sb_type;
1185 if (!frame_is_intra_only(cm)) {
1186 FRAME_COUNTS *const counts = td->counts;
1187 const int inter_block = is_inter_block(mbmi);
1188 const int seg_ref_active = vp9_segfeature_active(&cm->seg, mbmi->segment_id,
1190 if (!seg_ref_active) {
1191 counts->intra_inter[vp9_get_intra_inter_context(xd)][inter_block]++;
1192 // If the segment reference feature is enabled we have only a single
1193 // reference frame allowed for the segment so exclude it from
1194 // the reference frame counts used to work out probabilities.
1196 const MV_REFERENCE_FRAME ref0 = mbmi->ref_frame[0];
1197 if (cm->reference_mode == REFERENCE_MODE_SELECT)
1198 counts->comp_inter[vp9_get_reference_mode_context(cm, xd)]
1199 [has_second_ref(mbmi)]++;
1201 if (has_second_ref(mbmi)) {
1202 counts->comp_ref[vp9_get_pred_context_comp_ref_p(cm, xd)]
1203 [ref0 == GOLDEN_FRAME]++;
1205 counts->single_ref[vp9_get_pred_context_single_ref_p1(xd)][0]
1206 [ref0 != LAST_FRAME]++;
1207 if (ref0 != LAST_FRAME)
1208 counts->single_ref[vp9_get_pred_context_single_ref_p2(xd)][1]
1209 [ref0 != GOLDEN_FRAME]++;
1214 !vp9_segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_SKIP)) {
1215 const int mode_ctx = mbmi->mode_context[mbmi->ref_frame[0]];
1216 if (bsize >= BLOCK_8X8) {
1217 const PREDICTION_MODE mode = mbmi->mode;
1218 ++counts->inter_mode[mode_ctx][INTER_OFFSET(mode)];
1220 const int num_4x4_w = num_4x4_blocks_wide_lookup[bsize];
1221 const int num_4x4_h = num_4x4_blocks_high_lookup[bsize];
1223 for (idy = 0; idy < 2; idy += num_4x4_h) {
1224 for (idx = 0; idx < 2; idx += num_4x4_w) {
1225 const int j = idy * 2 + idx;
1226 const PREDICTION_MODE b_mode = mi->bmi[j].as_mode;
1227 ++counts->inter_mode[mode_ctx][INTER_OFFSET(b_mode)];
1235 static void restore_context(MACROBLOCK *const x, int mi_row, int mi_col,
1236 ENTROPY_CONTEXT a[16 * MAX_MB_PLANE],
1237 ENTROPY_CONTEXT l[16 * MAX_MB_PLANE],
1238 PARTITION_CONTEXT sa[8], PARTITION_CONTEXT sl[8],
1240 MACROBLOCKD *const xd = &x->e_mbd;
1242 const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
1243 const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
1244 int mi_width = num_8x8_blocks_wide_lookup[bsize];
1245 int mi_height = num_8x8_blocks_high_lookup[bsize];
1246 for (p = 0; p < MAX_MB_PLANE; p++) {
1248 xd->above_context[p] + ((mi_col * 2) >> xd->plane[p].subsampling_x),
1249 a + num_4x4_blocks_wide * p,
1250 (sizeof(ENTROPY_CONTEXT) * num_4x4_blocks_wide) >>
1251 xd->plane[p].subsampling_x);
1254 + ((mi_row & MI_MASK) * 2 >> xd->plane[p].subsampling_y),
1255 l + num_4x4_blocks_high * p,
1256 (sizeof(ENTROPY_CONTEXT) * num_4x4_blocks_high) >>
1257 xd->plane[p].subsampling_y);
1259 vpx_memcpy(xd->above_seg_context + mi_col, sa,
1260 sizeof(*xd->above_seg_context) * mi_width);
1261 vpx_memcpy(xd->left_seg_context + (mi_row & MI_MASK), sl,
1262 sizeof(xd->left_seg_context[0]) * mi_height);
1265 static void save_context(MACROBLOCK *const x, int mi_row, int mi_col,
1266 ENTROPY_CONTEXT a[16 * MAX_MB_PLANE],
1267 ENTROPY_CONTEXT l[16 * MAX_MB_PLANE],
1268 PARTITION_CONTEXT sa[8], PARTITION_CONTEXT sl[8],
1270 const MACROBLOCKD *const xd = &x->e_mbd;
1272 const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
1273 const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
1274 int mi_width = num_8x8_blocks_wide_lookup[bsize];
1275 int mi_height = num_8x8_blocks_high_lookup[bsize];
1277 // buffer the above/left context information of the block in search.
1278 for (p = 0; p < MAX_MB_PLANE; ++p) {
1280 a + num_4x4_blocks_wide * p,
1281 xd->above_context[p] + (mi_col * 2 >> xd->plane[p].subsampling_x),
1282 (sizeof(ENTROPY_CONTEXT) * num_4x4_blocks_wide) >>
1283 xd->plane[p].subsampling_x);
1285 l + num_4x4_blocks_high * p,
1287 + ((mi_row & MI_MASK) * 2 >> xd->plane[p].subsampling_y),
1288 (sizeof(ENTROPY_CONTEXT) * num_4x4_blocks_high) >>
1289 xd->plane[p].subsampling_y);
1291 vpx_memcpy(sa, xd->above_seg_context + mi_col,
1292 sizeof(*xd->above_seg_context) * mi_width);
1293 vpx_memcpy(sl, xd->left_seg_context + (mi_row & MI_MASK),
1294 sizeof(xd->left_seg_context[0]) * mi_height);
1297 static void encode_b(VP9_COMP *cpi, const TileInfo *const tile,
1299 TOKENEXTRA **tp, int mi_row, int mi_col,
1300 int output_enabled, BLOCK_SIZE bsize,
1301 PICK_MODE_CONTEXT *ctx) {
1302 MACROBLOCK *const x = &td->mb;
1303 set_offsets(cpi, tile, x, mi_row, mi_col, bsize);
1304 update_state(cpi, td, ctx, mi_row, mi_col, bsize, output_enabled);
1305 encode_superblock(cpi, td, tp, output_enabled, mi_row, mi_col, bsize, ctx);
1307 if (output_enabled) {
1308 update_stats(&cpi->common, td);
1310 (*tp)->token = EOSB_TOKEN;
1315 static void encode_sb(VP9_COMP *cpi, ThreadData *td,
1316 const TileInfo *const tile,
1317 TOKENEXTRA **tp, int mi_row, int mi_col,
1318 int output_enabled, BLOCK_SIZE bsize,
1320 VP9_COMMON *const cm = &cpi->common;
1321 MACROBLOCK *const x = &td->mb;
1322 MACROBLOCKD *const xd = &x->e_mbd;
1324 const int bsl = b_width_log2_lookup[bsize], hbs = (1 << bsl) / 4;
1326 PARTITION_TYPE partition;
1327 BLOCK_SIZE subsize = bsize;
1329 if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols)
1332 if (bsize >= BLOCK_8X8) {
1333 ctx = partition_plane_context(xd, mi_row, mi_col, bsize);
1334 subsize = get_subsize(bsize, pc_tree->partitioning);
1337 subsize = BLOCK_4X4;
1340 partition = partition_lookup[bsl][subsize];
1341 if (output_enabled && bsize != BLOCK_4X4)
1342 td->counts->partition[ctx][partition]++;
1344 switch (partition) {
1345 case PARTITION_NONE:
1346 encode_b(cpi, tile, td, tp, mi_row, mi_col, output_enabled, subsize,
1349 case PARTITION_VERT:
1350 encode_b(cpi, tile, td, tp, mi_row, mi_col, output_enabled, subsize,
1351 &pc_tree->vertical[0]);
1352 if (mi_col + hbs < cm->mi_cols && bsize > BLOCK_8X8) {
1353 encode_b(cpi, tile, td, tp, mi_row, mi_col + hbs, output_enabled,
1354 subsize, &pc_tree->vertical[1]);
1357 case PARTITION_HORZ:
1358 encode_b(cpi, tile, td, tp, mi_row, mi_col, output_enabled, subsize,
1359 &pc_tree->horizontal[0]);
1360 if (mi_row + hbs < cm->mi_rows && bsize > BLOCK_8X8) {
1361 encode_b(cpi, tile, td, tp, mi_row + hbs, mi_col, output_enabled,
1362 subsize, &pc_tree->horizontal[1]);
1365 case PARTITION_SPLIT:
1366 if (bsize == BLOCK_8X8) {
1367 encode_b(cpi, tile, td, tp, mi_row, mi_col, output_enabled, subsize,
1368 pc_tree->leaf_split[0]);
1370 encode_sb(cpi, td, tile, tp, mi_row, mi_col, output_enabled, subsize,
1372 encode_sb(cpi, td, tile, tp, mi_row, mi_col + hbs, output_enabled,
1373 subsize, pc_tree->split[1]);
1374 encode_sb(cpi, td, tile, tp, mi_row + hbs, mi_col, output_enabled,
1375 subsize, pc_tree->split[2]);
1376 encode_sb(cpi, td, tile, tp, mi_row + hbs, mi_col + hbs, output_enabled,
1377 subsize, pc_tree->split[3]);
1381 assert(0 && "Invalid partition type.");
1385 if (partition != PARTITION_SPLIT || bsize == BLOCK_8X8)
1386 update_partition_context(xd, mi_row, mi_col, subsize, bsize);
1389 // Check to see if the given partition size is allowed for a specified number
1390 // of 8x8 block rows and columns remaining in the image.
1391 // If not then return the largest allowed partition size
1392 static BLOCK_SIZE find_partition_size(BLOCK_SIZE bsize,
1393 int rows_left, int cols_left,
1395 if (rows_left <= 0 || cols_left <= 0) {
1396 return MIN(bsize, BLOCK_8X8);
1398 for (; bsize > 0; bsize -= 3) {
1399 *bh = num_8x8_blocks_high_lookup[bsize];
1400 *bw = num_8x8_blocks_wide_lookup[bsize];
1401 if ((*bh <= rows_left) && (*bw <= cols_left)) {
1409 static void set_partial_b64x64_partition(MODE_INFO *mi, int mis,
1410 int bh_in, int bw_in, int row8x8_remaining, int col8x8_remaining,
1411 BLOCK_SIZE bsize, MODE_INFO *mi_8x8) {
1414 for (r = 0; r < MI_BLOCK_SIZE; r += bh) {
1416 for (c = 0; c < MI_BLOCK_SIZE; c += bw) {
1417 const int index = r * mis + c;
1418 mi_8x8[index].src_mi = mi + index;
1419 mi_8x8[index].src_mi->mbmi.sb_type = find_partition_size(bsize,
1420 row8x8_remaining - r, col8x8_remaining - c, &bh, &bw);
1425 // This function attempts to set all mode info entries in a given SB64
1426 // to the same block partition size.
1427 // However, at the bottom and right borders of the image the requested size
1428 // may not be allowed in which case this code attempts to choose the largest
1429 // allowable partition.
1430 static void set_fixed_partitioning(VP9_COMP *cpi, const TileInfo *const tile,
1431 MODE_INFO *mi_8x8, int mi_row, int mi_col,
1433 VP9_COMMON *const cm = &cpi->common;
1434 const int mis = cm->mi_stride;
1435 const int row8x8_remaining = tile->mi_row_end - mi_row;
1436 const int col8x8_remaining = tile->mi_col_end - mi_col;
1437 int block_row, block_col;
1438 MODE_INFO *mi_upper_left = cm->mi + mi_row * mis + mi_col;
1439 int bh = num_8x8_blocks_high_lookup[bsize];
1440 int bw = num_8x8_blocks_wide_lookup[bsize];
1442 assert((row8x8_remaining > 0) && (col8x8_remaining > 0));
1444 // Apply the requested partition size to the SB64 if it is all "in image"
1445 if ((col8x8_remaining >= MI_BLOCK_SIZE) &&
1446 (row8x8_remaining >= MI_BLOCK_SIZE)) {
1447 for (block_row = 0; block_row < MI_BLOCK_SIZE; block_row += bh) {
1448 for (block_col = 0; block_col < MI_BLOCK_SIZE; block_col += bw) {
1449 int index = block_row * mis + block_col;
1450 mi_8x8[index].src_mi = mi_upper_left + index;
1451 mi_8x8[index].src_mi->mbmi.sb_type = bsize;
1455 // Else this is a partial SB64.
1456 set_partial_b64x64_partition(mi_upper_left, mis, bh, bw, row8x8_remaining,
1457 col8x8_remaining, bsize, mi_8x8);
1464 } coord_lookup[16] = {
1466 {0, 0}, {0, 2}, {2, 0}, {2, 2},
1468 {0, 4}, {0, 6}, {2, 4}, {2, 6},
1470 {4, 0}, {4, 2}, {6, 0}, {6, 2},
1472 {4, 4}, {4, 6}, {6, 4}, {6, 6},
1475 static void set_source_var_based_partition(VP9_COMP *cpi,
1476 const TileInfo *const tile,
1477 MACROBLOCK *const x,
1479 int mi_row, int mi_col) {
1480 VP9_COMMON *const cm = &cpi->common;
1481 const int mis = cm->mi_stride;
1482 const int row8x8_remaining = tile->mi_row_end - mi_row;
1483 const int col8x8_remaining = tile->mi_col_end - mi_col;
1484 MODE_INFO *mi_upper_left = cm->mi + mi_row * mis + mi_col;
1486 vp9_setup_src_planes(x, cpi->Source, mi_row, mi_col);
1488 assert((row8x8_remaining > 0) && (col8x8_remaining > 0));
1491 if ((col8x8_remaining >= MI_BLOCK_SIZE) &&
1492 (row8x8_remaining >= MI_BLOCK_SIZE)) {
1496 const int offset = (mi_row >> 1) * cm->mb_cols + (mi_col >> 1);
1497 int is_larger_better = 0;
1499 unsigned int thr = cpi->source_var_thresh;
1501 vpx_memset(d32, 0, 4 * sizeof(diff));
1503 for (i = 0; i < 4; i++) {
1506 for (j = 0; j < 4; j++) {
1507 int b_mi_row = coord_lookup[i * 4 + j].row;
1508 int b_mi_col = coord_lookup[i * 4 + j].col;
1509 int boffset = b_mi_row / 2 * cm->mb_cols +
1512 d16[j] = cpi->source_diff_var + offset + boffset;
1514 index = b_mi_row * mis + b_mi_col;
1515 mi_8x8[index].src_mi = mi_upper_left + index;
1516 mi_8x8[index].src_mi->mbmi.sb_type = BLOCK_16X16;
1518 // TODO(yunqingwang): If d16[j].var is very large, use 8x8 partition
1519 // size to further improve quality.
1522 is_larger_better = (d16[0]->var < thr) && (d16[1]->var < thr) &&
1523 (d16[2]->var < thr) && (d16[3]->var < thr);
1525 // Use 32x32 partition
1526 if (is_larger_better) {
1529 for (j = 0; j < 4; j++) {
1530 d32[i].sse += d16[j]->sse;
1531 d32[i].sum += d16[j]->sum;
1534 d32[i].var = d32[i].sse - (((int64_t)d32[i].sum * d32[i].sum) >> 10);
1536 index = coord_lookup[i*4].row * mis + coord_lookup[i*4].col;
1537 mi_8x8[index].src_mi = mi_upper_left + index;
1538 mi_8x8[index].src_mi->mbmi.sb_type = BLOCK_32X32;
1542 if (use32x32 == 4) {
1544 is_larger_better = (d32[0].var < thr) && (d32[1].var < thr) &&
1545 (d32[2].var < thr) && (d32[3].var < thr);
1547 // Use 64x64 partition
1548 if (is_larger_better) {
1549 mi_8x8[0].src_mi = mi_upper_left;
1550 mi_8x8[0].src_mi->mbmi.sb_type = BLOCK_64X64;
1553 } else { // partial in-image SB64
1554 int bh = num_8x8_blocks_high_lookup[BLOCK_16X16];
1555 int bw = num_8x8_blocks_wide_lookup[BLOCK_16X16];
1556 set_partial_b64x64_partition(mi_upper_left, mis, bh, bw,
1557 row8x8_remaining, col8x8_remaining, BLOCK_16X16, mi_8x8);
1561 static void update_state_rt(VP9_COMP *cpi, ThreadData *td,
1562 PICK_MODE_CONTEXT *ctx,
1563 int mi_row, int mi_col, int bsize) {
1564 VP9_COMMON *const cm = &cpi->common;
1565 MACROBLOCK *const x = &td->mb;
1566 MACROBLOCKD *const xd = &x->e_mbd;
1567 MODE_INFO *const mi = xd->mi[0].src_mi;
1568 MB_MODE_INFO *const mbmi = &xd->mi[0].src_mi->mbmi;
1569 const struct segmentation *const seg = &cm->seg;
1570 const int bw = num_8x8_blocks_wide_lookup[mi->mbmi.sb_type];
1571 const int bh = num_8x8_blocks_high_lookup[mi->mbmi.sb_type];
1572 const int x_mis = MIN(bw, cm->mi_cols - mi_col);
1573 const int y_mis = MIN(bh, cm->mi_rows - mi_row);
1575 xd->mi[0] = ctx->mic;
1576 xd->mi[0].src_mi = &xd->mi[0];
1578 if (seg->enabled && cpi->oxcf.aq_mode) {
1579 // For in frame complexity AQ or variance AQ, copy segment_id from
1580 // segmentation_map.
1581 if (cpi->oxcf.aq_mode == COMPLEXITY_AQ ||
1582 cpi->oxcf.aq_mode == VARIANCE_AQ ) {
1583 const uint8_t *const map = seg->update_map ? cpi->segmentation_map
1584 : cm->last_frame_seg_map;
1585 mbmi->segment_id = vp9_get_segment_id(cm, map, bsize, mi_row, mi_col);
1587 // Setting segmentation map for cyclic_refresh.
1588 vp9_cyclic_refresh_update_segment(cpi, mbmi, mi_row, mi_col, bsize,
1589 ctx->rate, ctx->dist, x->skip);
1591 vp9_init_plane_quantizers(cpi, x);
1594 if (is_inter_block(mbmi)) {
1595 vp9_update_mv_count(td);
1596 if (cm->interp_filter == SWITCHABLE) {
1597 const int pred_ctx = vp9_get_pred_context_switchable_interp(xd);
1598 ++td->counts->switchable_interp[pred_ctx][mbmi->interp_filter];
1601 if (mbmi->sb_type < BLOCK_8X8) {
1602 mbmi->mv[0].as_int = mi->bmi[3].as_mv[0].as_int;
1603 mbmi->mv[1].as_int = mi->bmi[3].as_mv[1].as_int;
1607 if (cm->use_prev_frame_mvs) {
1608 MV_REF *const frame_mvs =
1609 cm->cur_frame->mvs + mi_row * cm->mi_cols + mi_col;
1612 for (h = 0; h < y_mis; ++h) {
1613 MV_REF *const frame_mv = frame_mvs + h * cm->mi_cols;
1614 for (w = 0; w < x_mis; ++w) {
1615 MV_REF *const mv = frame_mv + w;
1616 mv->ref_frame[0] = mi->src_mi->mbmi.ref_frame[0];
1617 mv->ref_frame[1] = mi->src_mi->mbmi.ref_frame[1];
1618 mv->mv[0].as_int = mi->src_mi->mbmi.mv[0].as_int;
1619 mv->mv[1].as_int = mi->src_mi->mbmi.mv[1].as_int;
1624 x->skip = ctx->skip;
1625 x->skip_txfm[0] = mbmi->segment_id ? 0 : ctx->skip_txfm[0];
1628 static void encode_b_rt(VP9_COMP *cpi, ThreadData *td,
1629 const TileInfo *const tile,
1630 TOKENEXTRA **tp, int mi_row, int mi_col,
1631 int output_enabled, BLOCK_SIZE bsize,
1632 PICK_MODE_CONTEXT *ctx) {
1633 MACROBLOCK *const x = &td->mb;
1634 set_offsets(cpi, tile, x, mi_row, mi_col, bsize);
1635 update_state_rt(cpi, td, ctx, mi_row, mi_col, bsize);
1637 #if CONFIG_VP9_TEMPORAL_DENOISING
1638 if (cpi->oxcf.noise_sensitivity > 0 && output_enabled &&
1639 cpi->common.frame_type != KEY_FRAME) {
1640 vp9_denoiser_denoise(&cpi->denoiser, x, mi_row, mi_col,
1641 MAX(BLOCK_8X8, bsize), ctx);
1645 encode_superblock(cpi, td, tp, output_enabled, mi_row, mi_col, bsize, ctx);
1646 update_stats(&cpi->common, td);
1648 (*tp)->token = EOSB_TOKEN;
1652 static void encode_sb_rt(VP9_COMP *cpi, ThreadData *td,
1653 const TileInfo *const tile,
1654 TOKENEXTRA **tp, int mi_row, int mi_col,
1655 int output_enabled, BLOCK_SIZE bsize,
1657 VP9_COMMON *const cm = &cpi->common;
1658 MACROBLOCK *const x = &td->mb;
1659 MACROBLOCKD *const xd = &x->e_mbd;
1661 const int bsl = b_width_log2_lookup[bsize], hbs = (1 << bsl) / 4;
1663 PARTITION_TYPE partition;
1666 if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols)
1669 if (bsize >= BLOCK_8X8) {
1670 const int idx_str = xd->mi_stride * mi_row + mi_col;
1671 MODE_INFO *mi_8x8 = cm->mi[idx_str].src_mi;
1672 ctx = partition_plane_context(xd, mi_row, mi_col, bsize);
1673 subsize = mi_8x8[0].src_mi->mbmi.sb_type;
1676 subsize = BLOCK_4X4;
1679 partition = partition_lookup[bsl][subsize];
1680 if (output_enabled && bsize != BLOCK_4X4)
1681 td->counts->partition[ctx][partition]++;
1683 switch (partition) {
1684 case PARTITION_NONE:
1685 encode_b_rt(cpi, td, tile, tp, mi_row, mi_col, output_enabled, subsize,
1688 case PARTITION_VERT:
1689 encode_b_rt(cpi, td, tile, tp, mi_row, mi_col, output_enabled, subsize,
1690 &pc_tree->vertical[0]);
1691 if (mi_col + hbs < cm->mi_cols && bsize > BLOCK_8X8) {
1692 encode_b_rt(cpi, td, tile, tp, mi_row, mi_col + hbs, output_enabled,
1693 subsize, &pc_tree->vertical[1]);
1696 case PARTITION_HORZ:
1697 encode_b_rt(cpi, td, tile, tp, mi_row, mi_col, output_enabled, subsize,
1698 &pc_tree->horizontal[0]);
1699 if (mi_row + hbs < cm->mi_rows && bsize > BLOCK_8X8) {
1700 encode_b_rt(cpi, td, tile, tp, mi_row + hbs, mi_col, output_enabled,
1701 subsize, &pc_tree->horizontal[1]);
1704 case PARTITION_SPLIT:
1705 subsize = get_subsize(bsize, PARTITION_SPLIT);
1706 encode_sb_rt(cpi, td, tile, tp, mi_row, mi_col, output_enabled, subsize,
1708 encode_sb_rt(cpi, td, tile, tp, mi_row, mi_col + hbs, output_enabled,
1709 subsize, pc_tree->split[1]);
1710 encode_sb_rt(cpi, td, tile, tp, mi_row + hbs, mi_col, output_enabled,
1711 subsize, pc_tree->split[2]);
1712 encode_sb_rt(cpi, td, tile, tp, mi_row + hbs, mi_col + hbs,
1713 output_enabled, subsize, pc_tree->split[3]);
1716 assert(0 && "Invalid partition type.");
1720 if (partition != PARTITION_SPLIT || bsize == BLOCK_8X8)
1721 update_partition_context(xd, mi_row, mi_col, subsize, bsize);
1724 static void rd_use_partition(VP9_COMP *cpi,
1726 TileDataEnc *tile_data,
1727 MODE_INFO *mi_8x8, TOKENEXTRA **tp,
1728 int mi_row, int mi_col,
1730 int *rate, int64_t *dist,
1731 int do_recon, PC_TREE *pc_tree) {
1732 VP9_COMMON *const cm = &cpi->common;
1733 TileInfo *const tile_info = &tile_data->tile_info;
1734 MACROBLOCK *const x = &td->mb;
1735 MACROBLOCKD *const xd = &x->e_mbd;
1736 const int mis = cm->mi_stride;
1737 const int bsl = b_width_log2_lookup[bsize];
1738 const int mi_step = num_4x4_blocks_wide_lookup[bsize] / 2;
1739 const int bss = (1 << bsl) / 4;
1741 PARTITION_TYPE partition = PARTITION_NONE;
1743 ENTROPY_CONTEXT l[16 * MAX_MB_PLANE], a[16 * MAX_MB_PLANE];
1744 PARTITION_CONTEXT sl[8], sa[8];
1745 RD_COST last_part_rdc, none_rdc, chosen_rdc;
1746 BLOCK_SIZE sub_subsize = BLOCK_4X4;
1747 int splits_below = 0;
1748 BLOCK_SIZE bs_type = mi_8x8[0].src_mi->mbmi.sb_type;
1749 int do_partition_search = 1;
1750 PICK_MODE_CONTEXT *ctx = &pc_tree->none;
1752 if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols)
1755 assert(num_4x4_blocks_wide_lookup[bsize] ==
1756 num_4x4_blocks_high_lookup[bsize]);
1758 vp9_rd_cost_reset(&last_part_rdc);
1759 vp9_rd_cost_reset(&none_rdc);
1760 vp9_rd_cost_reset(&chosen_rdc);
1762 partition = partition_lookup[bsl][bs_type];
1763 subsize = get_subsize(bsize, partition);
1765 pc_tree->partitioning = partition;
1766 save_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
1768 if (bsize == BLOCK_16X16 && cpi->oxcf.aq_mode) {
1769 set_offsets(cpi, tile_info, x, mi_row, mi_col, bsize);
1770 x->mb_energy = vp9_block_energy(cpi, x, bsize);
1773 if (do_partition_search &&
1774 cpi->sf.partition_search_type == SEARCH_PARTITION &&
1775 cpi->sf.adjust_partitioning_from_last_frame) {
1776 // Check if any of the sub blocks are further split.
1777 if (partition == PARTITION_SPLIT && subsize > BLOCK_8X8) {
1778 sub_subsize = get_subsize(subsize, PARTITION_SPLIT);
1780 for (i = 0; i < 4; i++) {
1781 int jj = i >> 1, ii = i & 0x01;
1782 MODE_INFO *this_mi = mi_8x8[jj * bss * mis + ii * bss].src_mi;
1783 if (this_mi && this_mi->mbmi.sb_type >= sub_subsize) {
1789 // If partition is not none try none unless each of the 4 splits are split
1791 if (partition != PARTITION_NONE && !splits_below &&
1792 mi_row + (mi_step >> 1) < cm->mi_rows &&
1793 mi_col + (mi_step >> 1) < cm->mi_cols) {
1794 pc_tree->partitioning = PARTITION_NONE;
1795 rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &none_rdc, bsize,
1798 pl = partition_plane_context(xd, mi_row, mi_col, bsize);
1800 if (none_rdc.rate < INT_MAX) {
1801 none_rdc.rate += cpi->partition_cost[pl][PARTITION_NONE];
1802 none_rdc.rdcost = RDCOST(x->rdmult, x->rddiv, none_rdc.rate,
1806 restore_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
1807 mi_8x8[0].src_mi->mbmi.sb_type = bs_type;
1808 pc_tree->partitioning = partition;
1812 switch (partition) {
1813 case PARTITION_NONE:
1814 rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &last_part_rdc,
1815 bsize, ctx, INT64_MAX);
1817 case PARTITION_HORZ:
1818 rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &last_part_rdc,
1819 subsize, &pc_tree->horizontal[0],
1821 if (last_part_rdc.rate != INT_MAX &&
1822 bsize >= BLOCK_8X8 && mi_row + (mi_step >> 1) < cm->mi_rows) {
1824 PICK_MODE_CONTEXT *ctx = &pc_tree->horizontal[0];
1825 vp9_rd_cost_init(&tmp_rdc);
1826 update_state(cpi, td, ctx, mi_row, mi_col, subsize, 0);
1827 encode_superblock(cpi, td, tp, 0, mi_row, mi_col, subsize, ctx);
1828 rd_pick_sb_modes(cpi, tile_data, x,
1829 mi_row + (mi_step >> 1), mi_col, &tmp_rdc,
1830 subsize, &pc_tree->horizontal[1], INT64_MAX);
1831 if (tmp_rdc.rate == INT_MAX || tmp_rdc.dist == INT64_MAX) {
1832 vp9_rd_cost_reset(&last_part_rdc);
1835 last_part_rdc.rate += tmp_rdc.rate;
1836 last_part_rdc.dist += tmp_rdc.dist;
1837 last_part_rdc.rdcost += tmp_rdc.rdcost;
1840 case PARTITION_VERT:
1841 rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &last_part_rdc,
1842 subsize, &pc_tree->vertical[0], INT64_MAX);
1843 if (last_part_rdc.rate != INT_MAX &&
1844 bsize >= BLOCK_8X8 && mi_col + (mi_step >> 1) < cm->mi_cols) {
1846 PICK_MODE_CONTEXT *ctx = &pc_tree->vertical[0];
1847 vp9_rd_cost_init(&tmp_rdc);
1848 update_state(cpi, td, ctx, mi_row, mi_col, subsize, 0);
1849 encode_superblock(cpi, td, tp, 0, mi_row, mi_col, subsize, ctx);
1850 rd_pick_sb_modes(cpi, tile_data, x,
1851 mi_row, mi_col + (mi_step >> 1), &tmp_rdc,
1852 subsize, &pc_tree->vertical[bsize > BLOCK_8X8],
1854 if (tmp_rdc.rate == INT_MAX || tmp_rdc.dist == INT64_MAX) {
1855 vp9_rd_cost_reset(&last_part_rdc);
1858 last_part_rdc.rate += tmp_rdc.rate;
1859 last_part_rdc.dist += tmp_rdc.dist;
1860 last_part_rdc.rdcost += tmp_rdc.rdcost;
1863 case PARTITION_SPLIT:
1864 if (bsize == BLOCK_8X8) {
1865 rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &last_part_rdc,
1866 subsize, pc_tree->leaf_split[0], INT64_MAX);
1869 last_part_rdc.rate = 0;
1870 last_part_rdc.dist = 0;
1871 last_part_rdc.rdcost = 0;
1872 for (i = 0; i < 4; i++) {
1873 int x_idx = (i & 1) * (mi_step >> 1);
1874 int y_idx = (i >> 1) * (mi_step >> 1);
1875 int jj = i >> 1, ii = i & 0x01;
1877 if ((mi_row + y_idx >= cm->mi_rows) || (mi_col + x_idx >= cm->mi_cols))
1880 vp9_rd_cost_init(&tmp_rdc);
1881 rd_use_partition(cpi, td, tile_data,
1882 mi_8x8 + jj * bss * mis + ii * bss, tp,
1883 mi_row + y_idx, mi_col + x_idx, subsize,
1884 &tmp_rdc.rate, &tmp_rdc.dist,
1885 i != 3, pc_tree->split[i]);
1886 if (tmp_rdc.rate == INT_MAX || tmp_rdc.dist == INT64_MAX) {
1887 vp9_rd_cost_reset(&last_part_rdc);
1890 last_part_rdc.rate += tmp_rdc.rate;
1891 last_part_rdc.dist += tmp_rdc.dist;
1899 pl = partition_plane_context(xd, mi_row, mi_col, bsize);
1900 if (last_part_rdc.rate < INT_MAX) {
1901 last_part_rdc.rate += cpi->partition_cost[pl][partition];
1902 last_part_rdc.rdcost = RDCOST(x->rdmult, x->rddiv,
1903 last_part_rdc.rate, last_part_rdc.dist);
1906 if (do_partition_search
1907 && cpi->sf.adjust_partitioning_from_last_frame
1908 && cpi->sf.partition_search_type == SEARCH_PARTITION
1909 && partition != PARTITION_SPLIT && bsize > BLOCK_8X8
1910 && (mi_row + mi_step < cm->mi_rows ||
1911 mi_row + (mi_step >> 1) == cm->mi_rows)
1912 && (mi_col + mi_step < cm->mi_cols ||
1913 mi_col + (mi_step >> 1) == cm->mi_cols)) {
1914 BLOCK_SIZE split_subsize = get_subsize(bsize, PARTITION_SPLIT);
1915 chosen_rdc.rate = 0;
1916 chosen_rdc.dist = 0;
1917 restore_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
1918 pc_tree->partitioning = PARTITION_SPLIT;
1921 for (i = 0; i < 4; i++) {
1922 int x_idx = (i & 1) * (mi_step >> 1);
1923 int y_idx = (i >> 1) * (mi_step >> 1);
1925 ENTROPY_CONTEXT l[16 * MAX_MB_PLANE], a[16 * MAX_MB_PLANE];
1926 PARTITION_CONTEXT sl[8], sa[8];
1928 if ((mi_row + y_idx >= cm->mi_rows) || (mi_col + x_idx >= cm->mi_cols))
1931 save_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
1932 pc_tree->split[i]->partitioning = PARTITION_NONE;
1933 rd_pick_sb_modes(cpi, tile_data, x,
1934 mi_row + y_idx, mi_col + x_idx, &tmp_rdc,
1935 split_subsize, &pc_tree->split[i]->none, INT64_MAX);
1937 restore_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
1939 if (tmp_rdc.rate == INT_MAX || tmp_rdc.dist == INT64_MAX) {
1940 vp9_rd_cost_reset(&chosen_rdc);
1944 chosen_rdc.rate += tmp_rdc.rate;
1945 chosen_rdc.dist += tmp_rdc.dist;
1948 encode_sb(cpi, td, tile_info, tp, mi_row + y_idx, mi_col + x_idx, 0,
1949 split_subsize, pc_tree->split[i]);
1951 pl = partition_plane_context(xd, mi_row + y_idx, mi_col + x_idx,
1953 chosen_rdc.rate += cpi->partition_cost[pl][PARTITION_NONE];
1955 pl = partition_plane_context(xd, mi_row, mi_col, bsize);
1956 if (chosen_rdc.rate < INT_MAX) {
1957 chosen_rdc.rate += cpi->partition_cost[pl][PARTITION_SPLIT];
1958 chosen_rdc.rdcost = RDCOST(x->rdmult, x->rddiv,
1959 chosen_rdc.rate, chosen_rdc.dist);
1963 // If last_part is better set the partitioning to that.
1964 if (last_part_rdc.rdcost < chosen_rdc.rdcost) {
1965 mi_8x8[0].src_mi->mbmi.sb_type = bsize;
1966 if (bsize >= BLOCK_8X8)
1967 pc_tree->partitioning = partition;
1968 chosen_rdc = last_part_rdc;
1970 // If none was better set the partitioning to that.
1971 if (none_rdc.rdcost < chosen_rdc.rdcost) {
1972 if (bsize >= BLOCK_8X8)
1973 pc_tree->partitioning = PARTITION_NONE;
1974 chosen_rdc = none_rdc;
1977 restore_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
1979 // We must have chosen a partitioning and encoding or we'll fail later on.
1980 // No other opportunities for success.
1981 if (bsize == BLOCK_64X64)
1982 assert(chosen_rdc.rate < INT_MAX && chosen_rdc.dist < INT64_MAX);
1985 int output_enabled = (bsize == BLOCK_64X64);
1986 encode_sb(cpi, td, tile_info, tp, mi_row, mi_col, output_enabled, bsize,
1990 *rate = chosen_rdc.rate;
1991 *dist = chosen_rdc.dist;
1994 static const BLOCK_SIZE min_partition_size[BLOCK_SIZES] = {
1995 BLOCK_4X4, BLOCK_4X4, BLOCK_4X4,
1996 BLOCK_4X4, BLOCK_4X4, BLOCK_4X4,
1997 BLOCK_8X8, BLOCK_8X8, BLOCK_8X8,
1998 BLOCK_16X16, BLOCK_16X16, BLOCK_16X16,
2002 static const BLOCK_SIZE max_partition_size[BLOCK_SIZES] = {
2003 BLOCK_8X8, BLOCK_16X16, BLOCK_16X16,
2004 BLOCK_16X16, BLOCK_32X32, BLOCK_32X32,
2005 BLOCK_32X32, BLOCK_64X64, BLOCK_64X64,
2006 BLOCK_64X64, BLOCK_64X64, BLOCK_64X64,
2010 // Look at all the mode_info entries for blocks that are part of this
2011 // partition and find the min and max values for sb_type.
2012 // At the moment this is designed to work on a 64x64 SB but could be
2013 // adjusted to use a size parameter.
2015 // The min and max are assumed to have been initialized prior to calling this
2016 // function so repeat calls can accumulate a min and max of more than one sb64.
2017 static void get_sb_partition_size_range(MACROBLOCKD *xd, MODE_INFO *mi_8x8,
2018 BLOCK_SIZE *min_block_size,
2019 BLOCK_SIZE *max_block_size,
2020 int bs_hist[BLOCK_SIZES]) {
2021 int sb_width_in_blocks = MI_BLOCK_SIZE;
2022 int sb_height_in_blocks = MI_BLOCK_SIZE;
2026 // Check the sb_type for each block that belongs to this region.
2027 for (i = 0; i < sb_height_in_blocks; ++i) {
2028 for (j = 0; j < sb_width_in_blocks; ++j) {
2029 MODE_INFO *mi = mi_8x8[index+j].src_mi;
2030 BLOCK_SIZE sb_type = mi ? mi->mbmi.sb_type : 0;
2032 *min_block_size = MIN(*min_block_size, sb_type);
2033 *max_block_size = MAX(*max_block_size, sb_type);
2035 index += xd->mi_stride;
2039 // Next square block size less or equal than current block size.
2040 static const BLOCK_SIZE next_square_size[BLOCK_SIZES] = {
2041 BLOCK_4X4, BLOCK_4X4, BLOCK_4X4,
2042 BLOCK_8X8, BLOCK_8X8, BLOCK_8X8,
2043 BLOCK_16X16, BLOCK_16X16, BLOCK_16X16,
2044 BLOCK_32X32, BLOCK_32X32, BLOCK_32X32,
2048 // Look at neighboring blocks and set a min and max partition size based on
2050 static void rd_auto_partition_range(VP9_COMP *cpi, const TileInfo *const tile,
2051 MACROBLOCKD *const xd,
2052 int mi_row, int mi_col,
2053 BLOCK_SIZE *min_block_size,
2054 BLOCK_SIZE *max_block_size) {
2055 VP9_COMMON *const cm = &cpi->common;
2056 MODE_INFO *mi = xd->mi[0].src_mi;
2057 const int left_in_image = xd->left_available && mi[-1].src_mi;
2058 const int above_in_image = xd->up_available && mi[-xd->mi_stride].src_mi;
2059 const int row8x8_remaining = tile->mi_row_end - mi_row;
2060 const int col8x8_remaining = tile->mi_col_end - mi_col;
2062 BLOCK_SIZE min_size = BLOCK_4X4;
2063 BLOCK_SIZE max_size = BLOCK_64X64;
2065 int bs_hist[BLOCK_SIZES] = {0};
2067 // Trap case where we do not have a prediction.
2068 if (left_in_image || above_in_image || cm->frame_type != KEY_FRAME) {
2069 // Default "min to max" and "max to min"
2070 min_size = BLOCK_64X64;
2071 max_size = BLOCK_4X4;
2073 // NOTE: each call to get_sb_partition_size_range() uses the previous
2074 // passed in values for min and max as a starting point.
2075 // Find the min and max partition used in previous frame at this location
2076 if (cm->frame_type != KEY_FRAME) {
2077 MODE_INFO *prev_mi =
2078 cm->prev_mip + cm->mi_stride + 1 + mi_row * xd->mi_stride + mi_col;
2080 get_sb_partition_size_range(xd, prev_mi, &min_size, &max_size, bs_hist);
2082 // Find the min and max partition sizes used in the left SB64
2083 if (left_in_image) {
2084 MODE_INFO *left_sb64_mi = mi[-MI_BLOCK_SIZE].src_mi;
2085 get_sb_partition_size_range(xd, left_sb64_mi, &min_size, &max_size,
2088 // Find the min and max partition sizes used in the above SB64.
2089 if (above_in_image) {
2090 MODE_INFO *above_sb64_mi = mi[-xd->mi_stride * MI_BLOCK_SIZE].src_mi;
2091 get_sb_partition_size_range(xd, above_sb64_mi, &min_size, &max_size,
2095 // adjust observed min and max
2096 if (cpi->sf.auto_min_max_partition_size == RELAXED_NEIGHBORING_MIN_MAX) {
2097 min_size = min_partition_size[min_size];
2098 max_size = max_partition_size[max_size];
2099 } else if (cpi->sf.auto_min_max_partition_size ==
2100 CONSTRAIN_NEIGHBORING_MIN_MAX) {
2101 // adjust the search range based on the histogram of the observed
2102 // partition sizes from left, above the previous co-located blocks
2104 int first_moment = 0;
2105 int second_moment = 0;
2106 int var_unnormalized = 0;
2108 for (i = 0; i < BLOCK_SIZES; i++) {
2110 first_moment += bs_hist[i] * i;
2111 second_moment += bs_hist[i] * i * i;
2114 // if variance is small enough,
2115 // adjust the range around its mean size, which gives a tighter range
2116 var_unnormalized = second_moment - first_moment * first_moment / sum;
2117 if (var_unnormalized <= 4 * sum) {
2118 int mean = first_moment / sum;
2119 min_size = min_partition_size[mean];
2120 max_size = max_partition_size[mean];
2122 min_size = min_partition_size[min_size];
2123 max_size = max_partition_size[max_size];
2128 // Check border cases where max and min from neighbors may not be legal.
2129 max_size = find_partition_size(max_size,
2130 row8x8_remaining, col8x8_remaining,
2132 min_size = MIN(min_size, max_size);
2134 // When use_square_partition_only is true, make sure at least one square
2135 // partition is allowed by selecting the next smaller square size as
2137 if (cpi->sf.use_square_partition_only &&
2138 next_square_size[max_size] < min_size) {
2139 min_size = next_square_size[max_size];
2142 *min_block_size = min_size;
2143 *max_block_size = max_size;
2146 static void 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_8x8 = xd->mi;
2153 const int left_in_image = xd->left_available && mi_8x8[-1].src_mi;
2154 const int above_in_image = xd->up_available &&
2155 mi_8x8[-xd->mi_stride].src_mi;
2156 int row8x8_remaining = tile->mi_row_end - mi_row;
2157 int col8x8_remaining = tile->mi_col_end - mi_col;
2159 BLOCK_SIZE min_size = BLOCK_32X32;
2160 BLOCK_SIZE max_size = BLOCK_8X8;
2161 int bsl = mi_width_log2_lookup[BLOCK_64X64];
2162 const int search_range_ctrl = (((mi_row + mi_col) >> bsl) +
2163 get_chessboard_index(cm->current_video_frame)) & 0x1;
2164 // Trap case where we do not have a prediction.
2165 if (search_range_ctrl &&
2166 (left_in_image || above_in_image || cm->frame_type != KEY_FRAME)) {
2171 // Find the min and max partition sizes used in the left SB64.
2172 if (left_in_image) {
2174 mi = mi_8x8[-1].src_mi;
2175 for (block = 0; block < MI_BLOCK_SIZE; ++block) {
2176 cur_mi = mi[block * xd->mi_stride].src_mi;
2177 sb_type = cur_mi ? cur_mi->mbmi.sb_type : 0;
2178 min_size = MIN(min_size, sb_type);
2179 max_size = MAX(max_size, sb_type);
2182 // Find the min and max partition sizes used in the above SB64.
2183 if (above_in_image) {
2184 mi = mi_8x8[-xd->mi_stride * MI_BLOCK_SIZE].src_mi;
2185 for (block = 0; block < MI_BLOCK_SIZE; ++block) {
2186 sb_type = mi[block].src_mi ? mi[block].src_mi->mbmi.sb_type : 0;
2187 min_size = MIN(min_size, sb_type);
2188 max_size = MAX(max_size, sb_type);
2192 min_size = min_partition_size[min_size];
2193 max_size = find_partition_size(max_size, row8x8_remaining, col8x8_remaining,
2195 min_size = MIN(min_size, max_size);
2196 min_size = MAX(min_size, BLOCK_8X8);
2197 max_size = MIN(max_size, BLOCK_32X32);
2199 min_size = BLOCK_8X8;
2200 max_size = BLOCK_32X32;
2203 *min_block_size = min_size;
2204 *max_block_size = max_size;
2207 // TODO(jingning) refactor functions setting partition search range
2208 static void set_partition_range(VP9_COMMON *cm, MACROBLOCKD *xd,
2209 int mi_row, int mi_col, BLOCK_SIZE bsize,
2210 BLOCK_SIZE *min_bs, BLOCK_SIZE *max_bs) {
2211 int mi_width = num_8x8_blocks_wide_lookup[bsize];
2212 int mi_height = num_8x8_blocks_high_lookup[bsize];
2216 const int idx_str = cm->mi_stride * mi_row + mi_col;
2217 MODE_INFO *prev_mi = (cm->prev_mip + cm->mi_stride + 1 + idx_str)->src_mi;
2220 BLOCK_SIZE bs, min_size, max_size;
2222 min_size = BLOCK_64X64;
2223 max_size = BLOCK_4X4;
2226 for (idy = 0; idy < mi_height; ++idy) {
2227 for (idx = 0; idx < mi_width; ++idx) {
2228 mi = prev_mi[idy * cm->mi_stride + idx].src_mi;
2229 bs = mi ? mi->mbmi.sb_type : bsize;
2230 min_size = MIN(min_size, bs);
2231 max_size = MAX(max_size, bs);
2236 if (xd->left_available) {
2237 for (idy = 0; idy < mi_height; ++idy) {
2238 mi = xd->mi[idy * cm->mi_stride - 1].src_mi;
2239 bs = mi ? mi->mbmi.sb_type : bsize;
2240 min_size = MIN(min_size, bs);
2241 max_size = MAX(max_size, bs);
2245 if (xd->up_available) {
2246 for (idx = 0; idx < mi_width; ++idx) {
2247 mi = xd->mi[idx - cm->mi_stride].src_mi;
2248 bs = mi ? mi->mbmi.sb_type : bsize;
2249 min_size = MIN(min_size, bs);
2250 max_size = MAX(max_size, bs);
2254 if (min_size == max_size) {
2255 min_size = min_partition_size[min_size];
2256 max_size = max_partition_size[max_size];
2263 static INLINE void store_pred_mv(MACROBLOCK *x, PICK_MODE_CONTEXT *ctx) {
2264 vpx_memcpy(ctx->pred_mv, x->pred_mv, sizeof(x->pred_mv));
2267 static INLINE void load_pred_mv(MACROBLOCK *x, PICK_MODE_CONTEXT *ctx) {
2268 vpx_memcpy(x->pred_mv, ctx->pred_mv, sizeof(x->pred_mv));
2271 #if CONFIG_FP_MB_STATS
2272 const int num_16x16_blocks_wide_lookup[BLOCK_SIZES] =
2273 {1, 1, 1, 1, 1, 1, 1, 1, 2, 2, 2, 4, 4};
2274 const int num_16x16_blocks_high_lookup[BLOCK_SIZES] =
2275 {1, 1, 1, 1, 1, 1, 1, 2, 1, 2, 4, 2, 4};
2276 const int qindex_skip_threshold_lookup[BLOCK_SIZES] =
2277 {0, 10, 10, 30, 40, 40, 60, 80, 80, 90, 100, 100, 120};
2278 const int qindex_split_threshold_lookup[BLOCK_SIZES] =
2279 {0, 3, 3, 7, 15, 15, 30, 40, 40, 60, 80, 80, 120};
2280 const int complexity_16x16_blocks_threshold[BLOCK_SIZES] =
2281 {1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 4, 4, 6};
2292 static INLINE MOTION_DIRECTION get_motion_direction_fp(uint8_t fp_byte) {
2293 if (fp_byte & FPMB_MOTION_ZERO_MASK) {
2295 } else if (fp_byte & FPMB_MOTION_LEFT_MASK) {
2297 } else if (fp_byte & FPMB_MOTION_RIGHT_MASK) {
2299 } else if (fp_byte & FPMB_MOTION_UP_MASK) {
2306 static INLINE int get_motion_inconsistency(MOTION_DIRECTION this_mv,
2307 MOTION_DIRECTION that_mv) {
2308 if (this_mv == that_mv) {
2311 return abs(this_mv - that_mv) == 2 ? 2 : 1;
2316 // TODO(jingning,jimbankoski,rbultje): properly skip partition types that are
2317 // unlikely to be selected depending on previous rate-distortion optimization
2318 // results, for encoding speed-up.
2319 static void rd_pick_partition(VP9_COMP *cpi, ThreadData *td,
2320 TileDataEnc *tile_data,
2321 TOKENEXTRA **tp, int mi_row, int mi_col,
2322 BLOCK_SIZE bsize, RD_COST *rd_cost,
2323 int64_t best_rd, PC_TREE *pc_tree) {
2324 VP9_COMMON *const cm = &cpi->common;
2325 TileInfo *const tile_info = &tile_data->tile_info;
2326 MACROBLOCK *const x = &td->mb;
2327 MACROBLOCKD *const xd = &x->e_mbd;
2328 const int mi_step = num_8x8_blocks_wide_lookup[bsize] / 2;
2329 ENTROPY_CONTEXT l[16 * MAX_MB_PLANE], a[16 * MAX_MB_PLANE];
2330 PARTITION_CONTEXT sl[8], sa[8];
2331 TOKENEXTRA *tp_orig = *tp;
2332 PICK_MODE_CONTEXT *ctx = &pc_tree->none;
2335 RD_COST this_rdc, sum_rdc, best_rdc;
2336 int do_split = bsize >= BLOCK_8X8;
2339 // Override skipping rectangular partition operations for edge blocks
2340 const int force_horz_split = (mi_row + mi_step >= cm->mi_rows);
2341 const int force_vert_split = (mi_col + mi_step >= cm->mi_cols);
2342 const int xss = x->e_mbd.plane[1].subsampling_x;
2343 const int yss = x->e_mbd.plane[1].subsampling_y;
2345 BLOCK_SIZE min_size = x->min_partition_size;
2346 BLOCK_SIZE max_size = x->max_partition_size;
2348 #if CONFIG_FP_MB_STATS
2349 unsigned int src_diff_var = UINT_MAX;
2350 int none_complexity = 0;
2353 int partition_none_allowed = !force_horz_split && !force_vert_split;
2354 int partition_horz_allowed = !force_vert_split && yss <= xss &&
2356 int partition_vert_allowed = !force_horz_split && xss <= yss &&
2360 assert(num_8x8_blocks_wide_lookup[bsize] ==
2361 num_8x8_blocks_high_lookup[bsize]);
2363 vp9_rd_cost_init(&this_rdc);
2364 vp9_rd_cost_init(&sum_rdc);
2365 vp9_rd_cost_reset(&best_rdc);
2366 best_rdc.rdcost = best_rd;
2368 set_offsets(cpi, tile_info, x, mi_row, mi_col, bsize);
2370 if (bsize == BLOCK_16X16 && cpi->oxcf.aq_mode)
2371 x->mb_energy = vp9_block_energy(cpi, x, bsize);
2373 if (cpi->sf.cb_partition_search && bsize == BLOCK_16X16) {
2374 int cb_partition_search_ctrl = ((pc_tree->index == 0 || pc_tree->index == 3)
2375 + get_chessboard_index(cm->current_video_frame)) & 0x1;
2377 if (cb_partition_search_ctrl && bsize > min_size && bsize < max_size)
2378 set_partition_range(cm, xd, mi_row, mi_col, bsize, &min_size, &max_size);
2381 // Determine partition types in search according to the speed features.
2382 // The threshold set here has to be of square block size.
2383 if (cpi->sf.auto_min_max_partition_size) {
2384 partition_none_allowed &= (bsize <= max_size && bsize >= min_size);
2385 partition_horz_allowed &= ((bsize <= max_size && bsize > min_size) ||
2387 partition_vert_allowed &= ((bsize <= max_size && bsize > min_size) ||
2389 do_split &= bsize > min_size;
2391 if (cpi->sf.use_square_partition_only) {
2392 partition_horz_allowed &= force_horz_split;
2393 partition_vert_allowed &= force_vert_split;
2396 save_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
2398 #if CONFIG_FP_MB_STATS
2399 if (cpi->use_fp_mb_stats) {
2400 set_offsets(cpi, tile_info, x, mi_row, mi_col, bsize);
2401 src_diff_var = get_sby_perpixel_diff_variance(cpi, &x->plane[0].src,
2402 mi_row, mi_col, bsize);
2406 #if CONFIG_FP_MB_STATS
2407 // Decide whether we shall split directly and skip searching NONE by using
2408 // the first pass block statistics
2409 if (cpi->use_fp_mb_stats && bsize >= BLOCK_32X32 && do_split &&
2410 partition_none_allowed && src_diff_var > 4 &&
2411 cm->base_qindex < qindex_split_threshold_lookup[bsize]) {
2412 int mb_row = mi_row >> 1;
2413 int mb_col = mi_col >> 1;
2415 MIN(mb_row + num_16x16_blocks_high_lookup[bsize], cm->mb_rows);
2417 MIN(mb_col + num_16x16_blocks_wide_lookup[bsize], cm->mb_cols);
2420 // compute a complexity measure, basically measure inconsistency of motion
2421 // vectors obtained from the first pass in the current block
2422 for (r = mb_row; r < mb_row_end ; r++) {
2423 for (c = mb_col; c < mb_col_end; c++) {
2424 const int mb_index = r * cm->mb_cols + c;
2426 MOTION_DIRECTION this_mv;
2427 MOTION_DIRECTION right_mv;
2428 MOTION_DIRECTION bottom_mv;
2431 get_motion_direction_fp(cpi->twopass.this_frame_mb_stats[mb_index]);
2434 if (c != mb_col_end - 1) {
2435 right_mv = get_motion_direction_fp(
2436 cpi->twopass.this_frame_mb_stats[mb_index + 1]);
2437 none_complexity += get_motion_inconsistency(this_mv, right_mv);
2441 if (r != mb_row_end - 1) {
2442 bottom_mv = get_motion_direction_fp(
2443 cpi->twopass.this_frame_mb_stats[mb_index + cm->mb_cols]);
2444 none_complexity += get_motion_inconsistency(this_mv, bottom_mv);
2447 // do not count its left and top neighbors to avoid double counting
2451 if (none_complexity > complexity_16x16_blocks_threshold[bsize]) {
2452 partition_none_allowed = 0;
2458 if (partition_none_allowed) {
2459 rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col,
2460 &this_rdc, bsize, ctx, best_rdc.rdcost);
2461 if (this_rdc.rate != INT_MAX) {
2462 if (bsize >= BLOCK_8X8) {
2463 pl = partition_plane_context(xd, mi_row, mi_col, bsize);
2464 this_rdc.rate += cpi->partition_cost[pl][PARTITION_NONE];
2465 this_rdc.rdcost = RDCOST(x->rdmult, x->rddiv,
2466 this_rdc.rate, this_rdc.dist);
2469 if (this_rdc.rdcost < best_rdc.rdcost) {
2470 int64_t dist_breakout_thr = cpi->sf.partition_search_breakout_dist_thr;
2471 int rate_breakout_thr = cpi->sf.partition_search_breakout_rate_thr;
2473 best_rdc = this_rdc;
2474 if (bsize >= BLOCK_8X8)
2475 pc_tree->partitioning = PARTITION_NONE;
2477 // Adjust dist breakout threshold according to the partition size.
2478 dist_breakout_thr >>= 8 - (b_width_log2_lookup[bsize] +
2479 b_height_log2_lookup[bsize]);
2481 rate_breakout_thr *= num_pels_log2_lookup[bsize];
2483 // If all y, u, v transform blocks in this partition are skippable, and
2484 // the dist & rate are within the thresholds, the partition search is
2485 // terminated for current branch of the partition search tree.
2486 // The dist & rate thresholds are set to 0 at speed 0 to disable the
2487 // early termination at that speed.
2488 if (!x->e_mbd.lossless &&
2489 (ctx->skippable && best_rdc.dist < dist_breakout_thr &&
2490 best_rdc.rate < rate_breakout_thr)) {
2495 #if CONFIG_FP_MB_STATS
2496 // Check if every 16x16 first pass block statistics has zero
2497 // motion and the corresponding first pass residue is small enough.
2498 // If that is the case, check the difference variance between the
2499 // current frame and the last frame. If the variance is small enough,
2500 // stop further splitting in RD optimization
2501 if (cpi->use_fp_mb_stats && do_split != 0 &&
2502 cm->base_qindex > qindex_skip_threshold_lookup[bsize]) {
2503 int mb_row = mi_row >> 1;
2504 int mb_col = mi_col >> 1;
2506 MIN(mb_row + num_16x16_blocks_high_lookup[bsize], cm->mb_rows);
2508 MIN(mb_col + num_16x16_blocks_wide_lookup[bsize], cm->mb_cols);
2512 for (r = mb_row; r < mb_row_end; r++) {
2513 for (c = mb_col; c < mb_col_end; c++) {
2514 const int mb_index = r * cm->mb_cols + c;
2515 if (!(cpi->twopass.this_frame_mb_stats[mb_index] &
2516 FPMB_MOTION_ZERO_MASK) ||
2517 !(cpi->twopass.this_frame_mb_stats[mb_index] &
2518 FPMB_ERROR_SMALL_MASK)) {
2528 if (src_diff_var == UINT_MAX) {
2529 set_offsets(cpi, tile_info, x, mi_row, mi_col, bsize);
2530 src_diff_var = get_sby_perpixel_diff_variance(
2531 cpi, &x->plane[0].src, mi_row, mi_col, bsize);
2533 if (src_diff_var < 8) {
2542 restore_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
2545 // store estimated motion vector
2546 if (cpi->sf.adaptive_motion_search)
2547 store_pred_mv(x, ctx);
2550 // TODO(jingning): use the motion vectors given by the above search as
2551 // the starting point of motion search in the following partition type check.
2553 subsize = get_subsize(bsize, PARTITION_SPLIT);
2554 if (bsize == BLOCK_8X8) {
2556 if (cpi->sf.adaptive_pred_interp_filter && partition_none_allowed)
2557 pc_tree->leaf_split[0]->pred_interp_filter =
2558 ctx->mic.mbmi.interp_filter;
2559 rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &sum_rdc, subsize,
2560 pc_tree->leaf_split[0], best_rdc.rdcost);
2561 if (sum_rdc.rate == INT_MAX)
2562 sum_rdc.rdcost = INT64_MAX;
2564 for (i = 0; i < 4 && sum_rdc.rdcost < best_rdc.rdcost; ++i) {
2565 const int x_idx = (i & 1) * mi_step;
2566 const int y_idx = (i >> 1) * mi_step;
2568 if (mi_row + y_idx >= cm->mi_rows || mi_col + x_idx >= cm->mi_cols)
2571 if (cpi->sf.adaptive_motion_search)
2572 load_pred_mv(x, ctx);
2574 pc_tree->split[i]->index = i;
2575 rd_pick_partition(cpi, td, tile_data, tp,
2576 mi_row + y_idx, mi_col + x_idx,
2578 best_rdc.rdcost - sum_rdc.rdcost, pc_tree->split[i]);
2580 if (this_rdc.rate == INT_MAX) {
2581 sum_rdc.rdcost = INT64_MAX;
2584 sum_rdc.rate += this_rdc.rate;
2585 sum_rdc.dist += this_rdc.dist;
2586 sum_rdc.rdcost += this_rdc.rdcost;
2591 if (sum_rdc.rdcost < best_rdc.rdcost && i == 4) {
2592 pl = partition_plane_context(xd, mi_row, mi_col, bsize);
2593 sum_rdc.rate += cpi->partition_cost[pl][PARTITION_SPLIT];
2594 sum_rdc.rdcost = RDCOST(x->rdmult, x->rddiv,
2595 sum_rdc.rate, sum_rdc.dist);
2597 if (sum_rdc.rdcost < best_rdc.rdcost) {
2599 pc_tree->partitioning = PARTITION_SPLIT;
2602 // skip rectangular partition test when larger block size
2603 // gives better rd cost
2604 if (cpi->sf.less_rectangular_check)
2605 do_rect &= !partition_none_allowed;
2607 restore_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
2611 if (partition_horz_allowed && do_rect) {
2612 subsize = get_subsize(bsize, PARTITION_HORZ);
2613 if (cpi->sf.adaptive_motion_search)
2614 load_pred_mv(x, ctx);
2615 if (cpi->sf.adaptive_pred_interp_filter && bsize == BLOCK_8X8 &&
2616 partition_none_allowed)
2617 pc_tree->horizontal[0].pred_interp_filter =
2618 ctx->mic.mbmi.interp_filter;
2619 rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &sum_rdc, subsize,
2620 &pc_tree->horizontal[0], best_rdc.rdcost);
2622 if (sum_rdc.rdcost < best_rdc.rdcost && mi_row + mi_step < cm->mi_rows &&
2623 bsize > BLOCK_8X8) {
2624 PICK_MODE_CONTEXT *ctx = &pc_tree->horizontal[0];
2625 update_state(cpi, td, ctx, mi_row, mi_col, subsize, 0);
2626 encode_superblock(cpi, td, tp, 0, mi_row, mi_col, subsize, ctx);
2628 if (cpi->sf.adaptive_motion_search)
2629 load_pred_mv(x, ctx);
2630 if (cpi->sf.adaptive_pred_interp_filter && bsize == BLOCK_8X8 &&
2631 partition_none_allowed)
2632 pc_tree->horizontal[1].pred_interp_filter =
2633 ctx->mic.mbmi.interp_filter;
2634 rd_pick_sb_modes(cpi, tile_data, x, mi_row + mi_step, mi_col,
2635 &this_rdc, subsize, &pc_tree->horizontal[1],
2636 best_rdc.rdcost - sum_rdc.rdcost);
2637 if (this_rdc.rate == INT_MAX) {
2638 sum_rdc.rdcost = INT64_MAX;
2640 sum_rdc.rate += this_rdc.rate;
2641 sum_rdc.dist += this_rdc.dist;
2642 sum_rdc.rdcost += this_rdc.rdcost;
2646 if (sum_rdc.rdcost < best_rdc.rdcost) {
2647 pl = partition_plane_context(xd, mi_row, mi_col, bsize);
2648 sum_rdc.rate += cpi->partition_cost[pl][PARTITION_HORZ];
2649 sum_rdc.rdcost = RDCOST(x->rdmult, x->rddiv, sum_rdc.rate, sum_rdc.dist);
2650 if (sum_rdc.rdcost < best_rdc.rdcost) {
2652 pc_tree->partitioning = PARTITION_HORZ;
2655 restore_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
2658 if (partition_vert_allowed && do_rect) {
2659 subsize = get_subsize(bsize, PARTITION_VERT);
2661 if (cpi->sf.adaptive_motion_search)
2662 load_pred_mv(x, ctx);
2663 if (cpi->sf.adaptive_pred_interp_filter && bsize == BLOCK_8X8 &&
2664 partition_none_allowed)
2665 pc_tree->vertical[0].pred_interp_filter =
2666 ctx->mic.mbmi.interp_filter;
2667 rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &sum_rdc, subsize,
2668 &pc_tree->vertical[0], best_rdc.rdcost);
2669 if (sum_rdc.rdcost < best_rdc.rdcost && mi_col + mi_step < cm->mi_cols &&
2670 bsize > BLOCK_8X8) {
2671 update_state(cpi, td, &pc_tree->vertical[0], mi_row, mi_col, subsize, 0);
2672 encode_superblock(cpi, td, tp, 0, mi_row, mi_col, subsize,
2673 &pc_tree->vertical[0]);
2675 if (cpi->sf.adaptive_motion_search)
2676 load_pred_mv(x, ctx);
2677 if (cpi->sf.adaptive_pred_interp_filter && bsize == BLOCK_8X8 &&
2678 partition_none_allowed)
2679 pc_tree->vertical[1].pred_interp_filter =
2680 ctx->mic.mbmi.interp_filter;
2681 rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col + mi_step,
2683 &pc_tree->vertical[1], best_rdc.rdcost - sum_rdc.rdcost);
2684 if (this_rdc.rate == INT_MAX) {
2685 sum_rdc.rdcost = INT64_MAX;
2687 sum_rdc.rate += this_rdc.rate;
2688 sum_rdc.dist += this_rdc.dist;
2689 sum_rdc.rdcost += this_rdc.rdcost;
2693 if (sum_rdc.rdcost < best_rdc.rdcost) {
2694 pl = partition_plane_context(xd, mi_row, mi_col, bsize);
2695 sum_rdc.rate += cpi->partition_cost[pl][PARTITION_VERT];
2696 sum_rdc.rdcost = RDCOST(x->rdmult, x->rddiv,
2697 sum_rdc.rate, sum_rdc.dist);
2698 if (sum_rdc.rdcost < best_rdc.rdcost) {
2700 pc_tree->partitioning = PARTITION_VERT;
2703 restore_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
2706 // TODO(jbb): This code added so that we avoid static analysis
2707 // warning related to the fact that best_rd isn't used after this
2708 // point. This code should be refactored so that the duplicate
2709 // checks occur in some sub function and thus are used...
2711 *rd_cost = best_rdc;
2714 if (best_rdc.rate < INT_MAX && best_rdc.dist < INT64_MAX &&
2715 pc_tree->index != 3) {
2716 int output_enabled = (bsize == BLOCK_64X64);
2717 encode_sb(cpi, td, tile_info, tp, mi_row, mi_col, output_enabled,
2721 if (bsize == BLOCK_64X64) {
2722 assert(tp_orig < *tp);
2723 assert(best_rdc.rate < INT_MAX);
2724 assert(best_rdc.dist < INT64_MAX);
2726 assert(tp_orig == *tp);
2730 static void encode_rd_sb_row(VP9_COMP *cpi,
2732 TileDataEnc *tile_data,
2735 VP9_COMMON *const cm = &cpi->common;
2736 TileInfo *const tile_info = &tile_data->tile_info;
2737 MACROBLOCK *const x = &td->mb;
2738 MACROBLOCKD *const xd = &x->e_mbd;
2739 SPEED_FEATURES *const sf = &cpi->sf;
2742 // Initialize the left context for the new SB row
2743 vpx_memset(&xd->left_context, 0, sizeof(xd->left_context));
2744 vpx_memset(xd->left_seg_context, 0, sizeof(xd->left_seg_context));
2746 // Code each SB in the row
2747 for (mi_col = tile_info->mi_col_start; mi_col < tile_info->mi_col_end;
2748 mi_col += MI_BLOCK_SIZE) {
2749 const struct segmentation *const seg = &cm->seg;
2756 const int idx_str = cm->mi_stride * mi_row + mi_col;
2757 MODE_INFO *mi = cm->mi + idx_str;
2759 if (sf->adaptive_pred_interp_filter) {
2760 for (i = 0; i < 64; ++i)
2761 td->leaf_tree[i].pred_interp_filter = SWITCHABLE;
2763 for (i = 0; i < 64; ++i) {
2764 td->pc_tree[i].vertical[0].pred_interp_filter = SWITCHABLE;
2765 td->pc_tree[i].vertical[1].pred_interp_filter = SWITCHABLE;
2766 td->pc_tree[i].horizontal[0].pred_interp_filter = SWITCHABLE;
2767 td->pc_tree[i].horizontal[1].pred_interp_filter = SWITCHABLE;
2771 vp9_zero(x->pred_mv);
2772 td->pc_root->index = 0;
2775 const uint8_t *const map = seg->update_map ? cpi->segmentation_map
2776 : cm->last_frame_seg_map;
2777 int segment_id = vp9_get_segment_id(cm, map, BLOCK_64X64, mi_row, mi_col);
2778 seg_skip = vp9_segfeature_active(seg, segment_id, SEG_LVL_SKIP);
2781 x->source_variance = UINT_MAX;
2782 if (sf->partition_search_type == FIXED_PARTITION || seg_skip) {
2783 const BLOCK_SIZE bsize =
2784 seg_skip ? BLOCK_64X64 : sf->always_this_block_size;
2785 set_offsets(cpi, tile_info, x, mi_row, mi_col, BLOCK_64X64);
2786 set_fixed_partitioning(cpi, tile_info, mi, mi_row, mi_col, bsize);
2787 rd_use_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col,
2788 BLOCK_64X64, &dummy_rate, &dummy_dist, 1, td->pc_root);
2789 } else if (cpi->partition_search_skippable_frame) {
2791 set_offsets(cpi, tile_info, x, mi_row, mi_col, BLOCK_64X64);
2792 bsize = get_rd_var_based_fixed_partition(cpi, x, mi_row, mi_col);
2793 set_fixed_partitioning(cpi, tile_info, mi, mi_row, mi_col, bsize);
2794 rd_use_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col,
2795 BLOCK_64X64, &dummy_rate, &dummy_dist, 1, td->pc_root);
2796 } else if (sf->partition_search_type == VAR_BASED_PARTITION &&
2797 cm->frame_type != KEY_FRAME) {
2798 choose_partitioning(cpi, tile_info, x, mi_row, mi_col);
2799 rd_use_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col,
2800 BLOCK_64X64, &dummy_rate, &dummy_dist, 1, td->pc_root);
2802 // If required set upper and lower partition size limits
2803 if (sf->auto_min_max_partition_size) {
2804 set_offsets(cpi, tile_info, x, mi_row, mi_col, BLOCK_64X64);
2805 rd_auto_partition_range(cpi, tile_info, xd, mi_row, mi_col,
2806 &x->min_partition_size,
2807 &x->max_partition_size);
2809 rd_pick_partition(cpi, td, tile_data, tp, mi_row, mi_col, BLOCK_64X64,
2810 &dummy_rdc, INT64_MAX, td->pc_root);
2815 static void init_encode_frame_mb_context(VP9_COMP *cpi) {
2816 MACROBLOCK *const x = &cpi->td.mb;
2817 VP9_COMMON *const cm = &cpi->common;
2818 MACROBLOCKD *const xd = &x->e_mbd;
2819 const int aligned_mi_cols = mi_cols_aligned_to_sb(cm->mi_cols);
2821 // Copy data over into macro block data structures.
2822 vp9_setup_src_planes(x, cpi->Source, 0, 0);
2824 vp9_setup_block_planes(&x->e_mbd, cm->subsampling_x, cm->subsampling_y);
2826 // Note: this memset assumes above_context[0], [1] and [2]
2827 // are allocated as part of the same buffer.
2828 vpx_memset(xd->above_context[0], 0,
2829 sizeof(*xd->above_context[0]) *
2830 2 * aligned_mi_cols * MAX_MB_PLANE);
2831 vpx_memset(xd->above_seg_context, 0,
2832 sizeof(*xd->above_seg_context) * aligned_mi_cols);
2835 static int check_dual_ref_flags(VP9_COMP *cpi) {
2836 const int ref_flags = cpi->ref_frame_flags;
2838 if (vp9_segfeature_active(&cpi->common.seg, 1, SEG_LVL_REF_FRAME)) {
2841 return (!!(ref_flags & VP9_GOLD_FLAG) + !!(ref_flags & VP9_LAST_FLAG)
2842 + !!(ref_flags & VP9_ALT_FLAG)) >= 2;
2846 static void reset_skip_tx_size(VP9_COMMON *cm, TX_SIZE max_tx_size) {
2848 const int mis = cm->mi_stride;
2849 MODE_INFO *mi_ptr = cm->mi;
2851 for (mi_row = 0; mi_row < cm->mi_rows; ++mi_row, mi_ptr += mis) {
2852 for (mi_col = 0; mi_col < cm->mi_cols; ++mi_col) {
2853 if (mi_ptr[mi_col].src_mi->mbmi.tx_size > max_tx_size)
2854 mi_ptr[mi_col].src_mi->mbmi.tx_size = max_tx_size;
2859 static MV_REFERENCE_FRAME get_frame_type(const VP9_COMP *cpi) {
2860 if (frame_is_intra_only(&cpi->common))
2862 else if (cpi->rc.is_src_frame_alt_ref && cpi->refresh_golden_frame)
2863 return ALTREF_FRAME;
2864 else if (cpi->refresh_golden_frame || cpi->refresh_alt_ref_frame)
2865 return GOLDEN_FRAME;
2870 static TX_MODE select_tx_mode(const VP9_COMP *cpi, MACROBLOCKD *const xd) {
2873 if (cpi->common.frame_type == KEY_FRAME &&
2874 cpi->sf.use_nonrd_pick_mode &&
2875 cpi->sf.partition_search_type == VAR_BASED_PARTITION)
2877 if (cpi->sf.tx_size_search_method == USE_LARGESTALL)
2879 else if (cpi->sf.tx_size_search_method == USE_FULL_RD||
2880 cpi->sf.tx_size_search_method == USE_TX_8X8)
2881 return TX_MODE_SELECT;
2883 return cpi->common.tx_mode;
2886 static void hybrid_intra_mode_search(VP9_COMP *cpi, MACROBLOCK *const x,
2887 RD_COST *rd_cost, BLOCK_SIZE bsize,
2888 PICK_MODE_CONTEXT *ctx) {
2889 if (bsize < BLOCK_16X16)
2890 vp9_rd_pick_intra_mode_sb(cpi, x, rd_cost, bsize, ctx, INT64_MAX);
2892 vp9_pick_intra_mode(cpi, x, rd_cost, bsize, ctx);
2895 static void nonrd_pick_sb_modes(VP9_COMP *cpi,
2896 TileDataEnc *tile_data, MACROBLOCK *const x,
2897 int mi_row, int mi_col, RD_COST *rd_cost,
2898 BLOCK_SIZE bsize, PICK_MODE_CONTEXT *ctx) {
2899 VP9_COMMON *const cm = &cpi->common;
2900 TileInfo *const tile_info = &tile_data->tile_info;
2901 MACROBLOCKD *const xd = &x->e_mbd;
2903 set_offsets(cpi, tile_info, x, mi_row, mi_col, bsize);
2904 mbmi = &xd->mi[0].src_mi->mbmi;
2905 mbmi->sb_type = bsize;
2907 if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ && cm->seg.enabled)
2908 if (cyclic_refresh_segment_id_boosted(mbmi->segment_id))
2909 x->rdmult = vp9_cyclic_refresh_get_rdmult(cpi->cyclic_refresh);
2911 if (cm->frame_type == KEY_FRAME)
2912 hybrid_intra_mode_search(cpi, x, rd_cost, bsize, ctx);
2913 else if (vp9_segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_SKIP))
2914 set_mode_info_seg_skip(x, cm->tx_mode, rd_cost, bsize);
2915 else if (bsize >= BLOCK_8X8)
2916 vp9_pick_inter_mode(cpi, x, tile_data, mi_row, mi_col,
2917 rd_cost, bsize, ctx);
2919 vp9_pick_inter_mode_sub8x8(cpi, x, tile_data, mi_row, mi_col,
2920 rd_cost, bsize, ctx);
2922 duplicate_mode_info_in_sb(cm, xd, mi_row, mi_col, bsize);
2924 if (rd_cost->rate == INT_MAX)
2925 vp9_rd_cost_reset(rd_cost);
2927 ctx->rate = rd_cost->rate;
2928 ctx->dist = rd_cost->dist;
2931 static void fill_mode_info_sb(VP9_COMMON *cm, MACROBLOCK *x,
2932 int mi_row, int mi_col,
2935 MACROBLOCKD *xd = &x->e_mbd;
2936 int bsl = b_width_log2_lookup[bsize], hbs = (1 << bsl) / 4;
2937 PARTITION_TYPE partition = pc_tree->partitioning;
2938 BLOCK_SIZE subsize = get_subsize(bsize, partition);
2940 assert(bsize >= BLOCK_8X8);
2942 if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols)
2945 switch (partition) {
2946 case PARTITION_NONE:
2947 set_mode_info_offsets(cm, xd, mi_row, mi_col);
2948 *(xd->mi[0].src_mi) = pc_tree->none.mic;
2949 duplicate_mode_info_in_sb(cm, xd, mi_row, mi_col, bsize);
2951 case PARTITION_VERT:
2952 set_mode_info_offsets(cm, xd, mi_row, mi_col);
2953 *(xd->mi[0].src_mi) = pc_tree->vertical[0].mic;
2954 duplicate_mode_info_in_sb(cm, xd, mi_row, mi_col, subsize);
2956 if (mi_col + hbs < cm->mi_cols) {
2957 set_mode_info_offsets(cm, xd, mi_row, mi_col + hbs);
2958 *(xd->mi[0].src_mi) = pc_tree->vertical[1].mic;
2959 duplicate_mode_info_in_sb(cm, xd, mi_row, mi_col + hbs, subsize);
2962 case PARTITION_HORZ:
2963 set_mode_info_offsets(cm, xd, mi_row, mi_col);
2964 *(xd->mi[0].src_mi) = pc_tree->horizontal[0].mic;
2965 duplicate_mode_info_in_sb(cm, xd, mi_row, mi_col, subsize);
2966 if (mi_row + hbs < cm->mi_rows) {
2967 set_mode_info_offsets(cm, xd, mi_row + hbs, mi_col);
2968 *(xd->mi[0].src_mi) = pc_tree->horizontal[1].mic;
2969 duplicate_mode_info_in_sb(cm, xd, mi_row + hbs, mi_col, subsize);
2972 case PARTITION_SPLIT: {
2973 fill_mode_info_sb(cm, x, mi_row, mi_col, subsize, pc_tree->split[0]);
2974 fill_mode_info_sb(cm, x, mi_row, mi_col + hbs, subsize,
2976 fill_mode_info_sb(cm, x, mi_row + hbs, mi_col, subsize,
2978 fill_mode_info_sb(cm, x, mi_row + hbs, mi_col + hbs, subsize,
2987 // Reset the prediction pixel ready flag recursively.
2988 static void pred_pixel_ready_reset(PC_TREE *pc_tree, BLOCK_SIZE bsize) {
2989 pc_tree->none.pred_pixel_ready = 0;
2990 pc_tree->horizontal[0].pred_pixel_ready = 0;
2991 pc_tree->horizontal[1].pred_pixel_ready = 0;
2992 pc_tree->vertical[0].pred_pixel_ready = 0;
2993 pc_tree->vertical[1].pred_pixel_ready = 0;
2995 if (bsize > BLOCK_8X8) {
2996 BLOCK_SIZE subsize = get_subsize(bsize, PARTITION_SPLIT);
2998 for (i = 0; i < 4; ++i)
2999 pred_pixel_ready_reset(pc_tree->split[i], subsize);
3003 static void nonrd_pick_partition(VP9_COMP *cpi, ThreadData *td,
3004 TileDataEnc *tile_data,
3005 TOKENEXTRA **tp, int mi_row,
3006 int mi_col, BLOCK_SIZE bsize, RD_COST *rd_cost,
3007 int do_recon, int64_t best_rd,
3009 const SPEED_FEATURES *const sf = &cpi->sf;
3010 VP9_COMMON *const cm = &cpi->common;
3011 TileInfo *const tile_info = &tile_data->tile_info;
3012 MACROBLOCK *const x = &td->mb;
3013 MACROBLOCKD *const xd = &x->e_mbd;
3014 const int ms = num_8x8_blocks_wide_lookup[bsize] / 2;
3015 TOKENEXTRA *tp_orig = *tp;
3016 PICK_MODE_CONTEXT *ctx = &pc_tree->none;
3018 BLOCK_SIZE subsize = bsize;
3019 RD_COST this_rdc, sum_rdc, best_rdc;
3020 int do_split = bsize >= BLOCK_8X8;
3022 // Override skipping rectangular partition operations for edge blocks
3023 const int force_horz_split = (mi_row + ms >= cm->mi_rows);
3024 const int force_vert_split = (mi_col + ms >= cm->mi_cols);
3025 const int xss = x->e_mbd.plane[1].subsampling_x;
3026 const int yss = x->e_mbd.plane[1].subsampling_y;
3028 int partition_none_allowed = !force_horz_split && !force_vert_split;
3029 int partition_horz_allowed = !force_vert_split && yss <= xss &&
3031 int partition_vert_allowed = !force_horz_split && xss <= yss &&
3035 assert(num_8x8_blocks_wide_lookup[bsize] ==
3036 num_8x8_blocks_high_lookup[bsize]);
3038 vp9_rd_cost_init(&sum_rdc);
3039 vp9_rd_cost_reset(&best_rdc);
3040 best_rdc.rdcost = best_rd;
3042 // Determine partition types in search according to the speed features.
3043 // The threshold set here has to be of square block size.
3044 if (sf->auto_min_max_partition_size) {
3045 partition_none_allowed &= (bsize <= x->max_partition_size &&
3046 bsize >= x->min_partition_size);
3047 partition_horz_allowed &= ((bsize <= x->max_partition_size &&
3048 bsize > x->min_partition_size) ||
3050 partition_vert_allowed &= ((bsize <= x->max_partition_size &&
3051 bsize > x->min_partition_size) ||
3053 do_split &= bsize > x->min_partition_size;
3055 if (sf->use_square_partition_only) {
3056 partition_horz_allowed &= force_horz_split;
3057 partition_vert_allowed &= force_vert_split;
3060 ctx->pred_pixel_ready = !(partition_vert_allowed ||
3061 partition_horz_allowed ||
3065 if (partition_none_allowed) {
3066 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col,
3067 &this_rdc, bsize, ctx);
3068 ctx->mic.mbmi = xd->mi[0].src_mi->mbmi;
3069 ctx->skip_txfm[0] = x->skip_txfm[0];
3070 ctx->skip = x->skip;
3072 if (this_rdc.rate != INT_MAX) {
3073 int pl = partition_plane_context(xd, mi_row, mi_col, bsize);
3074 this_rdc.rate += cpi->partition_cost[pl][PARTITION_NONE];
3075 this_rdc.rdcost = RDCOST(x->rdmult, x->rddiv,
3076 this_rdc.rate, this_rdc.dist);
3077 if (this_rdc.rdcost < best_rdc.rdcost) {
3078 int64_t dist_breakout_thr = sf->partition_search_breakout_dist_thr;
3079 int64_t rate_breakout_thr = sf->partition_search_breakout_rate_thr;
3081 dist_breakout_thr >>= 8 - (b_width_log2_lookup[bsize] +
3082 b_height_log2_lookup[bsize]);
3084 rate_breakout_thr *= num_pels_log2_lookup[bsize];
3086 best_rdc = this_rdc;
3087 if (bsize >= BLOCK_8X8)
3088 pc_tree->partitioning = PARTITION_NONE;
3090 if (!x->e_mbd.lossless &&
3091 this_rdc.rate < rate_breakout_thr &&
3092 this_rdc.dist < dist_breakout_thr) {
3100 // store estimated motion vector
3101 store_pred_mv(x, ctx);
3105 int pl = partition_plane_context(xd, mi_row, mi_col, bsize);
3106 sum_rdc.rate += cpi->partition_cost[pl][PARTITION_SPLIT];
3107 sum_rdc.rdcost = RDCOST(x->rdmult, x->rddiv, sum_rdc.rate, sum_rdc.dist);
3108 subsize = get_subsize(bsize, PARTITION_SPLIT);
3109 for (i = 0; i < 4 && sum_rdc.rdcost < best_rdc.rdcost; ++i) {
3110 const int x_idx = (i & 1) * ms;
3111 const int y_idx = (i >> 1) * ms;
3113 if (mi_row + y_idx >= cm->mi_rows || mi_col + x_idx >= cm->mi_cols)
3115 load_pred_mv(x, ctx);
3116 nonrd_pick_partition(cpi, td, tile_data, tp,
3117 mi_row + y_idx, mi_col + x_idx,
3118 subsize, &this_rdc, 0,
3119 best_rdc.rdcost - sum_rdc.rdcost, pc_tree->split[i]);
3121 if (this_rdc.rate == INT_MAX) {
3122 vp9_rd_cost_reset(&sum_rdc);
3124 sum_rdc.rate += this_rdc.rate;
3125 sum_rdc.dist += this_rdc.dist;
3126 sum_rdc.rdcost += this_rdc.rdcost;
3130 if (sum_rdc.rdcost < best_rdc.rdcost) {
3132 pc_tree->partitioning = PARTITION_SPLIT;
3134 // skip rectangular partition test when larger block size
3135 // gives better rd cost
3136 if (sf->less_rectangular_check)
3137 do_rect &= !partition_none_allowed;
3142 if (partition_horz_allowed && do_rect) {
3143 subsize = get_subsize(bsize, PARTITION_HORZ);
3144 if (sf->adaptive_motion_search)
3145 load_pred_mv(x, ctx);
3146 pc_tree->horizontal[0].pred_pixel_ready = 1;
3147 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &sum_rdc, subsize,
3148 &pc_tree->horizontal[0]);
3150 pc_tree->horizontal[0].mic.mbmi = xd->mi[0].src_mi->mbmi;
3151 pc_tree->horizontal[0].skip_txfm[0] = x->skip_txfm[0];
3152 pc_tree->horizontal[0].skip = x->skip;
3154 if (sum_rdc.rdcost < best_rdc.rdcost && mi_row + ms < cm->mi_rows) {
3155 load_pred_mv(x, ctx);
3156 pc_tree->horizontal[1].pred_pixel_ready = 1;
3157 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row + ms, mi_col,
3159 &pc_tree->horizontal[1]);
3161 pc_tree->horizontal[1].mic.mbmi = xd->mi[0].src_mi->mbmi;
3162 pc_tree->horizontal[1].skip_txfm[0] = x->skip_txfm[0];
3163 pc_tree->horizontal[1].skip = x->skip;
3165 if (this_rdc.rate == INT_MAX) {
3166 vp9_rd_cost_reset(&sum_rdc);
3168 int pl = partition_plane_context(xd, mi_row, mi_col, bsize);
3169 this_rdc.rate += cpi->partition_cost[pl][PARTITION_HORZ];
3170 sum_rdc.rate += this_rdc.rate;
3171 sum_rdc.dist += this_rdc.dist;
3172 sum_rdc.rdcost = RDCOST(x->rdmult, x->rddiv,
3173 sum_rdc.rate, sum_rdc.dist);
3177 if (sum_rdc.rdcost < best_rdc.rdcost) {
3179 pc_tree->partitioning = PARTITION_HORZ;
3181 pred_pixel_ready_reset(pc_tree, bsize);
3186 if (partition_vert_allowed && do_rect) {
3187 subsize = get_subsize(bsize, PARTITION_VERT);
3188 if (sf->adaptive_motion_search)
3189 load_pred_mv(x, ctx);
3190 pc_tree->vertical[0].pred_pixel_ready = 1;
3191 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &sum_rdc, subsize,
3192 &pc_tree->vertical[0]);
3193 pc_tree->vertical[0].mic.mbmi = xd->mi[0].src_mi->mbmi;
3194 pc_tree->vertical[0].skip_txfm[0] = x->skip_txfm[0];
3195 pc_tree->vertical[0].skip = x->skip;
3197 if (sum_rdc.rdcost < best_rdc.rdcost && mi_col + ms < cm->mi_cols) {
3198 load_pred_mv(x, ctx);
3199 pc_tree->vertical[1].pred_pixel_ready = 1;
3200 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col + ms,
3202 &pc_tree->vertical[1]);
3203 pc_tree->vertical[1].mic.mbmi = xd->mi[0].src_mi->mbmi;
3204 pc_tree->vertical[1].skip_txfm[0] = x->skip_txfm[0];
3205 pc_tree->vertical[1].skip = x->skip;
3207 if (this_rdc.rate == INT_MAX) {
3208 vp9_rd_cost_reset(&sum_rdc);
3210 int pl = partition_plane_context(xd, mi_row, mi_col, bsize);
3211 sum_rdc.rate += cpi->partition_cost[pl][PARTITION_VERT];
3212 sum_rdc.rate += this_rdc.rate;
3213 sum_rdc.dist += this_rdc.dist;
3214 sum_rdc.rdcost = RDCOST(x->rdmult, x->rddiv,
3215 sum_rdc.rate, sum_rdc.dist);
3219 if (sum_rdc.rdcost < best_rdc.rdcost) {
3221 pc_tree->partitioning = PARTITION_VERT;
3223 pred_pixel_ready_reset(pc_tree, bsize);
3227 *rd_cost = best_rdc;
3229 if (best_rdc.rate == INT_MAX) {
3230 vp9_rd_cost_reset(rd_cost);
3234 // update mode info array
3235 fill_mode_info_sb(cm, x, mi_row, mi_col, bsize, pc_tree);
3237 if (best_rdc.rate < INT_MAX && best_rdc.dist < INT64_MAX && do_recon) {
3238 int output_enabled = (bsize == BLOCK_64X64);
3239 encode_sb_rt(cpi, td, tile_info, tp, mi_row, mi_col, output_enabled,
3243 if (bsize == BLOCK_64X64 && do_recon) {
3244 assert(tp_orig < *tp);
3245 assert(best_rdc.rate < INT_MAX);
3246 assert(best_rdc.dist < INT64_MAX);
3248 assert(tp_orig == *tp);
3252 static void nonrd_select_partition(VP9_COMP *cpi,
3254 TileDataEnc *tile_data,
3257 int mi_row, int mi_col,
3258 BLOCK_SIZE bsize, int output_enabled,
3259 RD_COST *rd_cost, PC_TREE *pc_tree) {
3260 VP9_COMMON *const cm = &cpi->common;
3261 TileInfo *const tile_info = &tile_data->tile_info;
3262 MACROBLOCK *const x = &td->mb;
3263 MACROBLOCKD *const xd = &x->e_mbd;
3264 const int bsl = b_width_log2_lookup[bsize], hbs = (1 << bsl) / 4;
3265 const int mis = cm->mi_stride;
3266 PARTITION_TYPE partition;
3270 vp9_rd_cost_reset(&this_rdc);
3271 if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols)
3274 subsize = (bsize >= BLOCK_8X8) ? mi[0].src_mi->mbmi.sb_type : BLOCK_4X4;
3275 partition = partition_lookup[bsl][subsize];
3277 if (bsize == BLOCK_32X32 && partition != PARTITION_NONE &&
3278 subsize >= BLOCK_16X16) {
3279 x->max_partition_size = BLOCK_32X32;
3280 x->min_partition_size = BLOCK_8X8;
3281 nonrd_pick_partition(cpi, td, tile_data, tp, mi_row, mi_col, bsize,
3282 rd_cost, 0, INT64_MAX, pc_tree);
3283 } else if (bsize == BLOCK_16X16 && partition != PARTITION_NONE) {
3284 x->max_partition_size = BLOCK_16X16;
3285 x->min_partition_size = BLOCK_8X8;
3286 nonrd_pick_partition(cpi, td, tile_data, tp, mi_row, mi_col, bsize,
3287 rd_cost, 0, INT64_MAX, pc_tree);
3289 switch (partition) {
3290 case PARTITION_NONE:
3291 pc_tree->none.pred_pixel_ready = 1;
3292 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, rd_cost,
3293 subsize, &pc_tree->none);
3294 pc_tree->none.mic.mbmi = xd->mi[0].src_mi->mbmi;
3295 pc_tree->none.skip_txfm[0] = x->skip_txfm[0];
3296 pc_tree->none.skip = x->skip;
3298 case PARTITION_VERT:
3299 pc_tree->vertical[0].pred_pixel_ready = 1;
3300 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, rd_cost,
3301 subsize, &pc_tree->vertical[0]);
3302 pc_tree->vertical[0].mic.mbmi = xd->mi[0].src_mi->mbmi;
3303 pc_tree->vertical[0].skip_txfm[0] = x->skip_txfm[0];
3304 pc_tree->vertical[0].skip = x->skip;
3305 if (mi_col + hbs < cm->mi_cols) {
3306 pc_tree->vertical[1].pred_pixel_ready = 1;
3307 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col + hbs,
3308 &this_rdc, subsize, &pc_tree->vertical[1]);
3309 pc_tree->vertical[1].mic.mbmi = xd->mi[0].src_mi->mbmi;
3310 pc_tree->vertical[1].skip_txfm[0] = x->skip_txfm[0];
3311 pc_tree->vertical[1].skip = x->skip;
3312 if (this_rdc.rate != INT_MAX && this_rdc.dist != INT64_MAX &&
3313 rd_cost->rate != INT_MAX && rd_cost->dist != INT64_MAX) {
3314 rd_cost->rate += this_rdc.rate;
3315 rd_cost->dist += this_rdc.dist;
3319 case PARTITION_HORZ:
3320 pc_tree->horizontal[0].pred_pixel_ready = 1;
3321 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, rd_cost,
3322 subsize, &pc_tree->horizontal[0]);
3323 pc_tree->horizontal[0].mic.mbmi = xd->mi[0].src_mi->mbmi;
3324 pc_tree->horizontal[0].skip_txfm[0] = x->skip_txfm[0];
3325 pc_tree->horizontal[0].skip = x->skip;
3326 if (mi_row + hbs < cm->mi_rows) {
3327 pc_tree->horizontal[1].pred_pixel_ready = 1;
3328 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row + hbs, mi_col,
3329 &this_rdc, subsize, &pc_tree->horizontal[1]);
3330 pc_tree->horizontal[1].mic.mbmi = xd->mi[0].src_mi->mbmi;
3331 pc_tree->horizontal[1].skip_txfm[0] = x->skip_txfm[0];
3332 pc_tree->horizontal[1].skip = x->skip;
3333 if (this_rdc.rate != INT_MAX && this_rdc.dist != INT64_MAX &&
3334 rd_cost->rate != INT_MAX && rd_cost->dist != INT64_MAX) {
3335 rd_cost->rate += this_rdc.rate;
3336 rd_cost->dist += this_rdc.dist;
3340 case PARTITION_SPLIT:
3341 subsize = get_subsize(bsize, PARTITION_SPLIT);
3342 nonrd_select_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col,
3343 subsize, output_enabled, rd_cost,
3345 nonrd_select_partition(cpi, td, tile_data, mi + hbs, tp,
3346 mi_row, mi_col + hbs, subsize, output_enabled,
3347 &this_rdc, pc_tree->split[1]);
3348 if (this_rdc.rate != INT_MAX && this_rdc.dist != INT64_MAX &&
3349 rd_cost->rate != INT_MAX && rd_cost->dist != INT64_MAX) {
3350 rd_cost->rate += this_rdc.rate;
3351 rd_cost->dist += this_rdc.dist;
3353 nonrd_select_partition(cpi, td, tile_data, mi + hbs * mis, tp,
3354 mi_row + hbs, mi_col, subsize, output_enabled,
3355 &this_rdc, pc_tree->split[2]);
3356 if (this_rdc.rate != INT_MAX && this_rdc.dist != INT64_MAX &&
3357 rd_cost->rate != INT_MAX && rd_cost->dist != INT64_MAX) {
3358 rd_cost->rate += this_rdc.rate;
3359 rd_cost->dist += this_rdc.dist;
3361 nonrd_select_partition(cpi, td, tile_data, mi + hbs * mis + hbs, tp,
3362 mi_row + hbs, mi_col + hbs, subsize,
3363 output_enabled, &this_rdc, pc_tree->split[3]);
3364 if (this_rdc.rate != INT_MAX && this_rdc.dist != INT64_MAX &&
3365 rd_cost->rate != INT_MAX && rd_cost->dist != INT64_MAX) {
3366 rd_cost->rate += this_rdc.rate;
3367 rd_cost->dist += this_rdc.dist;
3371 assert(0 && "Invalid partition type.");
3376 if (bsize == BLOCK_64X64 && output_enabled)
3377 encode_sb_rt(cpi, td, tile_info, tp, mi_row, mi_col, 1, bsize, pc_tree);
3381 static void nonrd_use_partition(VP9_COMP *cpi,
3383 TileDataEnc *tile_data,
3386 int mi_row, int mi_col,
3387 BLOCK_SIZE bsize, int output_enabled,
3388 RD_COST *dummy_cost, PC_TREE *pc_tree) {
3389 VP9_COMMON *const cm = &cpi->common;
3390 TileInfo *tile_info = &tile_data->tile_info;
3391 MACROBLOCK *const x = &td->mb;
3392 MACROBLOCKD *const xd = &x->e_mbd;
3393 const int bsl = b_width_log2_lookup[bsize], hbs = (1 << bsl) / 4;
3394 const int mis = cm->mi_stride;
3395 PARTITION_TYPE partition;
3398 if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols)
3401 subsize = (bsize >= BLOCK_8X8) ? mi[0].src_mi->mbmi.sb_type : BLOCK_4X4;
3402 partition = partition_lookup[bsl][subsize];
3404 if (output_enabled && bsize != BLOCK_4X4) {
3405 int ctx = partition_plane_context(xd, mi_row, mi_col, bsize);
3406 td->counts->partition[ctx][partition]++;
3409 switch (partition) {
3410 case PARTITION_NONE:
3411 pc_tree->none.pred_pixel_ready = 1;
3412 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, dummy_cost,
3413 subsize, &pc_tree->none);
3414 pc_tree->none.mic.mbmi = xd->mi[0].src_mi->mbmi;
3415 pc_tree->none.skip_txfm[0] = x->skip_txfm[0];
3416 pc_tree->none.skip = x->skip;
3417 encode_b_rt(cpi, td, tile_info, tp, mi_row, mi_col, output_enabled,
3418 subsize, &pc_tree->none);
3420 case PARTITION_VERT:
3421 pc_tree->vertical[0].pred_pixel_ready = 1;
3422 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, dummy_cost,
3423 subsize, &pc_tree->vertical[0]);
3424 pc_tree->vertical[0].mic.mbmi = xd->mi[0].src_mi->mbmi;
3425 pc_tree->vertical[0].skip_txfm[0] = x->skip_txfm[0];
3426 pc_tree->vertical[0].skip = x->skip;
3427 encode_b_rt(cpi, td, tile_info, tp, mi_row, mi_col, output_enabled,
3428 subsize, &pc_tree->vertical[0]);
3429 if (mi_col + hbs < cm->mi_cols && bsize > BLOCK_8X8) {
3430 pc_tree->vertical[1].pred_pixel_ready = 1;
3431 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col + hbs,
3432 dummy_cost, subsize, &pc_tree->vertical[1]);
3433 pc_tree->vertical[1].mic.mbmi = xd->mi[0].src_mi->mbmi;
3434 pc_tree->vertical[1].skip_txfm[0] = x->skip_txfm[0];
3435 pc_tree->vertical[1].skip = x->skip;
3436 encode_b_rt(cpi, td, tile_info, tp, mi_row, mi_col + hbs,
3437 output_enabled, subsize, &pc_tree->vertical[1]);
3440 case PARTITION_HORZ:
3441 pc_tree->horizontal[0].pred_pixel_ready = 1;
3442 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, dummy_cost,
3443 subsize, &pc_tree->horizontal[0]);
3444 pc_tree->horizontal[0].mic.mbmi = xd->mi[0].src_mi->mbmi;
3445 pc_tree->horizontal[0].skip_txfm[0] = x->skip_txfm[0];
3446 pc_tree->horizontal[0].skip = x->skip;
3447 encode_b_rt(cpi, td, tile_info, tp, mi_row, mi_col, output_enabled,
3448 subsize, &pc_tree->horizontal[0]);
3450 if (mi_row + hbs < cm->mi_rows && bsize > BLOCK_8X8) {
3451 pc_tree->horizontal[1].pred_pixel_ready = 1;
3452 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row + hbs, mi_col,
3453 dummy_cost, subsize, &pc_tree->horizontal[1]);
3454 pc_tree->horizontal[1].mic.mbmi = xd->mi[0].src_mi->mbmi;
3455 pc_tree->horizontal[1].skip_txfm[0] = x->skip_txfm[0];
3456 pc_tree->horizontal[1].skip = x->skip;
3457 encode_b_rt(cpi, td, tile_info, tp, mi_row + hbs, mi_col,
3458 output_enabled, subsize, &pc_tree->horizontal[1]);
3461 case PARTITION_SPLIT:
3462 subsize = get_subsize(bsize, PARTITION_SPLIT);
3463 if (bsize == BLOCK_8X8) {
3464 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, dummy_cost,
3465 subsize, pc_tree->leaf_split[0]);
3466 encode_b_rt(cpi, td, tile_info, tp, mi_row, mi_col,
3467 output_enabled, subsize, pc_tree->leaf_split[0]);
3469 nonrd_use_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col,
3470 subsize, output_enabled, dummy_cost,
3472 nonrd_use_partition(cpi, td, tile_data, mi + hbs, tp,
3473 mi_row, mi_col + hbs, subsize, output_enabled,
3474 dummy_cost, pc_tree->split[1]);
3475 nonrd_use_partition(cpi, td, tile_data, mi + hbs * mis, tp,
3476 mi_row + hbs, mi_col, subsize, output_enabled,
3477 dummy_cost, pc_tree->split[2]);
3478 nonrd_use_partition(cpi, td, tile_data, mi + hbs * mis + hbs, tp,
3479 mi_row + hbs, mi_col + hbs, subsize, output_enabled,
3480 dummy_cost, pc_tree->split[3]);
3484 assert(0 && "Invalid partition type.");
3488 if (partition != PARTITION_SPLIT || bsize == BLOCK_8X8)
3489 update_partition_context(xd, mi_row, mi_col, subsize, bsize);
3492 static void encode_nonrd_sb_row(VP9_COMP *cpi,
3494 TileDataEnc *tile_data,
3497 SPEED_FEATURES *const sf = &cpi->sf;
3498 VP9_COMMON *const cm = &cpi->common;
3499 TileInfo *const tile_info = &tile_data->tile_info;
3500 MACROBLOCK *const x = &td->mb;
3501 MACROBLOCKD *const xd = &x->e_mbd;
3504 // Initialize the left context for the new SB row
3505 vpx_memset(&xd->left_context, 0, sizeof(xd->left_context));
3506 vpx_memset(xd->left_seg_context, 0, sizeof(xd->left_seg_context));
3508 // Code each SB in the row
3509 for (mi_col = tile_info->mi_col_start; mi_col < tile_info->mi_col_end;
3510 mi_col += MI_BLOCK_SIZE) {
3511 const struct segmentation *const seg = &cm->seg;
3513 const int idx_str = cm->mi_stride * mi_row + mi_col;
3514 MODE_INFO *mi = cm->mi + idx_str;
3515 PARTITION_SEARCH_TYPE partition_search_type = sf->partition_search_type;
3516 BLOCK_SIZE bsize = BLOCK_64X64;
3518 x->source_variance = UINT_MAX;
3519 vp9_zero(x->pred_mv);
3520 vp9_rd_cost_init(&dummy_rdc);
3521 x->color_sensitivity[0] = 0;
3522 x->color_sensitivity[1] = 0;
3525 const uint8_t *const map = seg->update_map ? cpi->segmentation_map
3526 : cm->last_frame_seg_map;
3527 int segment_id = vp9_get_segment_id(cm, map, BLOCK_64X64, mi_row, mi_col);
3528 seg_skip = vp9_segfeature_active(seg, segment_id, SEG_LVL_SKIP);
3530 partition_search_type = FIXED_PARTITION;
3534 // Set the partition type of the 64X64 block
3535 switch (partition_search_type) {
3536 case VAR_BASED_PARTITION:
3537 // TODO(jingning, marpan): The mode decision and encoding process
3538 // support both intra and inter sub8x8 block coding for RTC mode.
3539 // Tune the thresholds accordingly to use sub8x8 block coding for
3540 // coding performance improvement.
3541 choose_partitioning(cpi, tile_info, x, mi_row, mi_col);
3542 nonrd_use_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col,
3543 BLOCK_64X64, 1, &dummy_rdc, td->pc_root);
3545 case SOURCE_VAR_BASED_PARTITION:
3546 set_source_var_based_partition(cpi, tile_info, x, mi, mi_row, mi_col);
3547 nonrd_use_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col,
3548 BLOCK_64X64, 1, &dummy_rdc, td->pc_root);
3550 case FIXED_PARTITION:
3552 bsize = sf->always_this_block_size;
3553 set_fixed_partitioning(cpi, tile_info, mi, mi_row, mi_col, bsize);
3554 nonrd_use_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col,
3555 BLOCK_64X64, 1, &dummy_rdc, td->pc_root);
3557 case REFERENCE_PARTITION:
3558 set_offsets(cpi, tile_info, x, mi_row, mi_col, BLOCK_64X64);
3559 if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ && cm->seg.enabled &&
3560 xd->mi[0].src_mi->mbmi.segment_id) {
3561 x->max_partition_size = BLOCK_64X64;
3562 x->min_partition_size = BLOCK_8X8;
3563 nonrd_pick_partition(cpi, td, tile_data, tp, mi_row, mi_col,
3564 BLOCK_64X64, &dummy_rdc, 1,
3565 INT64_MAX, td->pc_root);
3567 choose_partitioning(cpi, tile_info, x, mi_row, mi_col);
3568 nonrd_select_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col,
3569 BLOCK_64X64, 1, &dummy_rdc, td->pc_root);
3579 // end RTC play code
3581 static int set_var_thresh_from_histogram(VP9_COMP *cpi) {
3582 const SPEED_FEATURES *const sf = &cpi->sf;
3583 const VP9_COMMON *const cm = &cpi->common;
3585 const uint8_t *src = cpi->Source->y_buffer;
3586 const uint8_t *last_src = cpi->Last_Source->y_buffer;
3587 const int src_stride = cpi->Source->y_stride;
3588 const int last_stride = cpi->Last_Source->y_stride;
3590 // Pick cutoff threshold
3591 const int cutoff = (MIN(cm->width, cm->height) >= 720) ?
3592 (cm->MBs * VAR_HIST_LARGE_CUT_OFF / 100) :
3593 (cm->MBs * VAR_HIST_SMALL_CUT_OFF / 100);
3594 DECLARE_ALIGNED_ARRAY(16, int, hist, VAR_HIST_BINS);
3595 diff *var16 = cpi->source_diff_var;
3600 vpx_memset(hist, 0, VAR_HIST_BINS * sizeof(hist[0]));
3602 for (i = 0; i < cm->mb_rows; i++) {
3603 for (j = 0; j < cm->mb_cols; j++) {
3604 #if CONFIG_VP9_HIGHBITDEPTH
3605 if (cm->use_highbitdepth) {
3606 switch (cm->bit_depth) {
3608 vp9_highbd_get16x16var(src, src_stride, last_src, last_stride,
3609 &var16->sse, &var16->sum);
3612 vp9_highbd_10_get16x16var(src, src_stride, last_src, last_stride,
3613 &var16->sse, &var16->sum);
3616 vp9_highbd_12_get16x16var(src, src_stride, last_src, last_stride,
3617 &var16->sse, &var16->sum);
3620 assert(0 && "cm->bit_depth should be VPX_BITS_8, VPX_BITS_10"
3625 vp9_get16x16var(src, src_stride, last_src, last_stride,
3626 &var16->sse, &var16->sum);
3629 vp9_get16x16var(src, src_stride, last_src, last_stride,
3630 &var16->sse, &var16->sum);
3631 #endif // CONFIG_VP9_HIGHBITDEPTH
3632 var16->var = var16->sse -
3633 (((uint32_t)var16->sum * var16->sum) >> 8);
3635 if (var16->var >= VAR_HIST_MAX_BG_VAR)
3636 hist[VAR_HIST_BINS - 1]++;
3638 hist[var16->var / VAR_HIST_FACTOR]++;
3645 src = src - cm->mb_cols * 16 + 16 * src_stride;
3646 last_src = last_src - cm->mb_cols * 16 + 16 * last_stride;
3649 cpi->source_var_thresh = 0;
3651 if (hist[VAR_HIST_BINS - 1] < cutoff) {
3652 for (i = 0; i < VAR_HIST_BINS - 1; i++) {
3656 cpi->source_var_thresh = (i + 1) * VAR_HIST_FACTOR;
3662 return sf->search_type_check_frequency;
3665 static void source_var_based_partition_search_method(VP9_COMP *cpi) {
3666 VP9_COMMON *const cm = &cpi->common;
3667 SPEED_FEATURES *const sf = &cpi->sf;
3669 if (cm->frame_type == KEY_FRAME) {
3670 // For key frame, use SEARCH_PARTITION.
3671 sf->partition_search_type = SEARCH_PARTITION;
3672 } else if (cm->intra_only) {
3673 sf->partition_search_type = FIXED_PARTITION;
3675 if (cm->last_width != cm->width || cm->last_height != cm->height) {
3676 if (cpi->source_diff_var)
3677 vpx_free(cpi->source_diff_var);
3679 CHECK_MEM_ERROR(cm, cpi->source_diff_var,
3680 vpx_calloc(cm->MBs, sizeof(diff)));
3683 if (!cpi->frames_till_next_var_check)
3684 cpi->frames_till_next_var_check = set_var_thresh_from_histogram(cpi);
3686 if (cpi->frames_till_next_var_check > 0) {
3687 sf->partition_search_type = FIXED_PARTITION;
3688 cpi->frames_till_next_var_check--;
3693 static int get_skip_encode_frame(const VP9_COMMON *cm, ThreadData *const td) {
3694 unsigned int intra_count = 0, inter_count = 0;
3697 for (j = 0; j < INTRA_INTER_CONTEXTS; ++j) {
3698 intra_count += td->counts->intra_inter[j][0];
3699 inter_count += td->counts->intra_inter[j][1];
3702 return (intra_count << 2) < inter_count &&
3703 cm->frame_type != KEY_FRAME &&
3707 void vp9_init_tile_data(VP9_COMP *cpi) {
3708 VP9_COMMON *const cm = &cpi->common;
3709 const int tile_cols = 1 << cm->log2_tile_cols;
3710 const int tile_rows = 1 << cm->log2_tile_rows;
3711 int tile_col, tile_row;
3712 TOKENEXTRA *pre_tok = cpi->tile_tok[0][0];
3715 if (cpi->tile_data == NULL) {
3716 CHECK_MEM_ERROR(cm, cpi->tile_data,
3717 vpx_malloc(tile_cols * tile_rows * sizeof(*cpi->tile_data)));
3718 for (tile_row = 0; tile_row < tile_rows; ++tile_row)
3719 for (tile_col = 0; tile_col < tile_cols; ++tile_col) {
3720 TileDataEnc *tile_data =
3721 &cpi->tile_data[tile_row * tile_cols + tile_col];
3723 for (i = 0; i < BLOCK_SIZES; ++i) {
3724 for (j = 0; j < MAX_MODES; ++j) {
3725 tile_data->thresh_freq_fact[i][j] = 32;
3726 tile_data->mode_map[i][j] = j;
3732 for (tile_row = 0; tile_row < tile_rows; ++tile_row) {
3733 for (tile_col = 0; tile_col < tile_cols; ++tile_col) {
3734 TileInfo *tile_info =
3735 &cpi->tile_data[tile_row * tile_cols + tile_col].tile_info;
3736 vp9_tile_init(tile_info, cm, tile_row, tile_col);
3738 cpi->tile_tok[tile_row][tile_col] = pre_tok + tile_tok;
3739 pre_tok = cpi->tile_tok[tile_row][tile_col];
3740 tile_tok = allocated_tokens(*tile_info);
3745 void vp9_encode_tile(VP9_COMP *cpi, ThreadData *td,
3746 int tile_row, int tile_col) {
3747 VP9_COMMON *const cm = &cpi->common;
3748 const int tile_cols = 1 << cm->log2_tile_cols;
3749 TileDataEnc *this_tile =
3750 &cpi->tile_data[tile_row * tile_cols + tile_col];
3751 const TileInfo * const tile_info = &this_tile->tile_info;
3752 TOKENEXTRA *tok = cpi->tile_tok[tile_row][tile_col];
3755 for (mi_row = tile_info->mi_row_start; mi_row < tile_info->mi_row_end;
3756 mi_row += MI_BLOCK_SIZE) {
3757 if (cpi->sf.use_nonrd_pick_mode)
3758 encode_nonrd_sb_row(cpi, td, this_tile, mi_row, &tok);
3760 encode_rd_sb_row(cpi, td, this_tile, mi_row, &tok);
3762 cpi->tok_count[tile_row][tile_col] =
3763 (unsigned int)(tok - cpi->tile_tok[tile_row][tile_col]);
3764 assert(tok - cpi->tile_tok[tile_row][tile_col] <=
3765 allocated_tokens(*tile_info));
3768 static void encode_tiles(VP9_COMP *cpi) {
3769 VP9_COMMON *const cm = &cpi->common;
3770 const int tile_cols = 1 << cm->log2_tile_cols;
3771 const int tile_rows = 1 << cm->log2_tile_rows;
3772 int tile_col, tile_row;
3774 vp9_init_tile_data(cpi);
3776 for (tile_row = 0; tile_row < tile_rows; ++tile_row)
3777 for (tile_col = 0; tile_col < tile_cols; ++tile_col)
3778 vp9_encode_tile(cpi, &cpi->td, tile_row, tile_col);
3781 #if CONFIG_FP_MB_STATS
3782 static int input_fpmb_stats(FIRSTPASS_MB_STATS *firstpass_mb_stats,
3783 VP9_COMMON *cm, uint8_t **this_frame_mb_stats) {
3784 uint8_t *mb_stats_in = firstpass_mb_stats->mb_stats_start +
3785 cm->current_video_frame * cm->MBs * sizeof(uint8_t);
3787 if (mb_stats_in > firstpass_mb_stats->mb_stats_end)
3790 *this_frame_mb_stats = mb_stats_in;
3796 static void encode_frame_internal(VP9_COMP *cpi) {
3797 SPEED_FEATURES *const sf = &cpi->sf;
3798 RD_OPT *const rd_opt = &cpi->rd;
3799 ThreadData *const td = &cpi->td;
3800 MACROBLOCK *const x = &td->mb;
3801 VP9_COMMON *const cm = &cpi->common;
3802 MACROBLOCKD *const xd = &x->e_mbd;
3803 RD_COUNTS *const rdc = &cpi->td.rd_counts;
3806 xd->mi[0].src_mi = &xd->mi[0];
3808 vp9_zero(*td->counts);
3809 vp9_zero(rdc->coef_counts);
3810 vp9_zero(rdc->comp_pred_diff);
3811 vp9_zero(rdc->filter_diff);
3812 vp9_zero(rdc->tx_select_diff);
3813 vp9_zero(rd_opt->tx_select_threshes);
3815 xd->lossless = cm->base_qindex == 0 &&
3816 cm->y_dc_delta_q == 0 &&
3817 cm->uv_dc_delta_q == 0 &&
3818 cm->uv_ac_delta_q == 0;
3820 #if CONFIG_VP9_HIGHBITDEPTH
3821 if (cm->use_highbitdepth)
3822 x->fwd_txm4x4 = xd->lossless ? vp9_highbd_fwht4x4 : vp9_highbd_fdct4x4;
3824 x->fwd_txm4x4 = xd->lossless ? vp9_fwht4x4 : vp9_fdct4x4;
3825 x->highbd_itxm_add = xd->lossless ? vp9_highbd_iwht4x4_add :
3826 vp9_highbd_idct4x4_add;
3828 x->fwd_txm4x4 = xd->lossless ? vp9_fwht4x4 : vp9_fdct4x4;
3829 #endif // CONFIG_VP9_HIGHBITDEPTH
3830 x->itxm_add = xd->lossless ? vp9_iwht4x4_add : vp9_idct4x4_add;
3835 cm->tx_mode = select_tx_mode(cpi, xd);
3837 vp9_frame_init_quantizer(cpi);
3839 vp9_initialize_rd_consts(cpi);
3840 vp9_initialize_me_consts(cpi, x, cm->base_qindex);
3841 init_encode_frame_mb_context(cpi);
3842 cm->use_prev_frame_mvs = !cm->error_resilient_mode &&
3843 cm->width == cm->last_width &&
3844 cm->height == cm->last_height &&
3846 cm->last_show_frame;
3847 // Special case: set prev_mi to NULL when the previous mode info
3848 // context cannot be used.
3849 cm->prev_mi = cm->use_prev_frame_mvs ?
3850 cm->prev_mip + cm->mi_stride + 1 : NULL;
3852 x->quant_fp = cpi->sf.use_quant_fp;
3853 vp9_zero(x->skip_txfm);
3854 if (sf->use_nonrd_pick_mode) {
3855 // Initialize internal buffer pointers for rtc coding, where non-RD
3856 // mode decision is used and hence no buffer pointer swap needed.
3858 struct macroblock_plane *const p = x->plane;
3859 struct macroblockd_plane *const pd = xd->plane;
3860 PICK_MODE_CONTEXT *ctx = &cpi->td.pc_root->none;
3862 for (i = 0; i < MAX_MB_PLANE; ++i) {
3863 p[i].coeff = ctx->coeff_pbuf[i][0];
3864 p[i].qcoeff = ctx->qcoeff_pbuf[i][0];
3865 pd[i].dqcoeff = ctx->dqcoeff_pbuf[i][0];
3866 p[i].eobs = ctx->eobs_pbuf[i][0];
3868 vp9_zero(x->zcoeff_blk);
3870 if (sf->partition_search_type == SOURCE_VAR_BASED_PARTITION)
3871 source_var_based_partition_search_method(cpi);
3875 struct vpx_usec_timer emr_timer;
3876 vpx_usec_timer_start(&emr_timer);
3878 #if CONFIG_FP_MB_STATS
3879 if (cpi->use_fp_mb_stats) {
3880 input_fpmb_stats(&cpi->twopass.firstpass_mb_stats, cm,
3881 &cpi->twopass.this_frame_mb_stats);
3885 // If allowed, encoding tiles in parallel with one thread handling one tile.
3886 if (MIN(cpi->oxcf.max_threads, 1 << cm->log2_tile_cols) > 1)
3887 vp9_encode_tiles_mt(cpi);
3891 vpx_usec_timer_mark(&emr_timer);
3892 cpi->time_encode_sb_row += vpx_usec_timer_elapsed(&emr_timer);
3895 sf->skip_encode_frame = sf->skip_encode_sb ?
3896 get_skip_encode_frame(cm, td) : 0;
3899 // Keep record of the total distortion this time around for future use
3900 cpi->last_frame_distortion = cpi->frame_distortion;
3904 static INTERP_FILTER get_interp_filter(
3905 const int64_t threshes[SWITCHABLE_FILTER_CONTEXTS], int is_alt_ref) {
3907 threshes[EIGHTTAP_SMOOTH] > threshes[EIGHTTAP] &&
3908 threshes[EIGHTTAP_SMOOTH] > threshes[EIGHTTAP_SHARP] &&
3909 threshes[EIGHTTAP_SMOOTH] > threshes[SWITCHABLE - 1]) {
3910 return EIGHTTAP_SMOOTH;
3911 } else if (threshes[EIGHTTAP_SHARP] > threshes[EIGHTTAP] &&
3912 threshes[EIGHTTAP_SHARP] > threshes[SWITCHABLE - 1]) {
3913 return EIGHTTAP_SHARP;
3914 } else if (threshes[EIGHTTAP] > threshes[SWITCHABLE - 1]) {
3921 void vp9_encode_frame(VP9_COMP *cpi) {
3922 VP9_COMMON *const cm = &cpi->common;
3924 // In the longer term the encoder should be generalized to match the
3925 // decoder such that we allow compound where one of the 3 buffers has a
3926 // different sign bias and that buffer is then the fixed ref. However, this
3927 // requires further work in the rd loop. For now the only supported encoder
3928 // side behavior is where the ALT ref buffer has opposite sign bias to
3930 if (!frame_is_intra_only(cm)) {
3931 if ((cm->ref_frame_sign_bias[ALTREF_FRAME] ==
3932 cm->ref_frame_sign_bias[GOLDEN_FRAME]) ||
3933 (cm->ref_frame_sign_bias[ALTREF_FRAME] ==
3934 cm->ref_frame_sign_bias[LAST_FRAME])) {
3935 cpi->allow_comp_inter_inter = 0;
3937 cpi->allow_comp_inter_inter = 1;
3938 cm->comp_fixed_ref = ALTREF_FRAME;
3939 cm->comp_var_ref[0] = LAST_FRAME;
3940 cm->comp_var_ref[1] = GOLDEN_FRAME;
3944 if (cpi->sf.frame_parameter_update) {
3946 RD_OPT *const rd_opt = &cpi->rd;
3947 FRAME_COUNTS *counts = cpi->td.counts;
3948 RD_COUNTS *const rdc = &cpi->td.rd_counts;
3950 // This code does a single RD pass over the whole frame assuming
3951 // either compound, single or hybrid prediction as per whatever has
3952 // worked best for that type of frame in the past.
3953 // It also predicts whether another coding mode would have worked
3954 // better that this coding mode. If that is the case, it remembers
3955 // that for subsequent frames.
3956 // It does the same analysis for transform size selection also.
3957 const MV_REFERENCE_FRAME frame_type = get_frame_type(cpi);
3958 int64_t *const mode_thrs = rd_opt->prediction_type_threshes[frame_type];
3959 int64_t *const filter_thrs = rd_opt->filter_threshes[frame_type];
3960 int *const tx_thrs = rd_opt->tx_select_threshes[frame_type];
3961 const int is_alt_ref = frame_type == ALTREF_FRAME;
3963 /* prediction (compound, single or hybrid) mode selection */
3964 if (is_alt_ref || !cpi->allow_comp_inter_inter)
3965 cm->reference_mode = SINGLE_REFERENCE;
3966 else if (mode_thrs[COMPOUND_REFERENCE] > mode_thrs[SINGLE_REFERENCE] &&
3967 mode_thrs[COMPOUND_REFERENCE] >
3968 mode_thrs[REFERENCE_MODE_SELECT] &&
3969 check_dual_ref_flags(cpi) &&
3970 cpi->static_mb_pct == 100)
3971 cm->reference_mode = COMPOUND_REFERENCE;
3972 else if (mode_thrs[SINGLE_REFERENCE] > mode_thrs[REFERENCE_MODE_SELECT])
3973 cm->reference_mode = SINGLE_REFERENCE;
3975 cm->reference_mode = REFERENCE_MODE_SELECT;
3977 if (cm->interp_filter == SWITCHABLE)
3978 cm->interp_filter = get_interp_filter(filter_thrs, is_alt_ref);
3980 encode_frame_internal(cpi);
3982 for (i = 0; i < REFERENCE_MODES; ++i)
3983 mode_thrs[i] = (mode_thrs[i] + rdc->comp_pred_diff[i] / cm->MBs) / 2;
3985 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i)
3986 filter_thrs[i] = (filter_thrs[i] + rdc->filter_diff[i] / cm->MBs) / 2;
3988 for (i = 0; i < TX_MODES; ++i) {
3989 int64_t pd = rdc->tx_select_diff[i];
3990 if (i == TX_MODE_SELECT)
3991 pd -= RDCOST(cpi->td.mb.rdmult, cpi->td.mb.rddiv, 2048 * (TX_SIZES - 1),
3993 tx_thrs[i] = (tx_thrs[i] + (int)(pd / cm->MBs)) / 2;
3996 if (cm->reference_mode == REFERENCE_MODE_SELECT) {
3997 int single_count_zero = 0;
3998 int comp_count_zero = 0;
4000 for (i = 0; i < COMP_INTER_CONTEXTS; i++) {
4001 single_count_zero += counts->comp_inter[i][0];
4002 comp_count_zero += counts->comp_inter[i][1];
4005 if (comp_count_zero == 0) {
4006 cm->reference_mode = SINGLE_REFERENCE;
4007 vp9_zero(counts->comp_inter);
4008 } else if (single_count_zero == 0) {
4009 cm->reference_mode = COMPOUND_REFERENCE;
4010 vp9_zero(counts->comp_inter);
4014 if (cm->tx_mode == TX_MODE_SELECT) {
4016 int count8x8_lp = 0, count8x8_8x8p = 0;
4017 int count16x16_16x16p = 0, count16x16_lp = 0;
4020 for (i = 0; i < TX_SIZE_CONTEXTS; ++i) {
4021 count4x4 += counts->tx.p32x32[i][TX_4X4];
4022 count4x4 += counts->tx.p16x16[i][TX_4X4];
4023 count4x4 += counts->tx.p8x8[i][TX_4X4];
4025 count8x8_lp += counts->tx.p32x32[i][TX_8X8];
4026 count8x8_lp += counts->tx.p16x16[i][TX_8X8];
4027 count8x8_8x8p += counts->tx.p8x8[i][TX_8X8];
4029 count16x16_16x16p += counts->tx.p16x16[i][TX_16X16];
4030 count16x16_lp += counts->tx.p32x32[i][TX_16X16];
4031 count32x32 += counts->tx.p32x32[i][TX_32X32];
4033 if (count4x4 == 0 && count16x16_lp == 0 && count16x16_16x16p == 0 &&
4035 cm->tx_mode = ALLOW_8X8;
4036 reset_skip_tx_size(cm, TX_8X8);
4037 } else if (count8x8_8x8p == 0 && count16x16_16x16p == 0 &&
4038 count8x8_lp == 0 && count16x16_lp == 0 && count32x32 == 0) {
4039 cm->tx_mode = ONLY_4X4;
4040 reset_skip_tx_size(cm, TX_4X4);
4041 } else if (count8x8_lp == 0 && count16x16_lp == 0 && count4x4 == 0) {
4042 cm->tx_mode = ALLOW_32X32;
4043 } else if (count32x32 == 0 && count8x8_lp == 0 && count4x4 == 0) {
4044 cm->tx_mode = ALLOW_16X16;
4045 reset_skip_tx_size(cm, TX_16X16);
4049 cm->reference_mode = SINGLE_REFERENCE;
4050 encode_frame_internal(cpi);
4054 static void sum_intra_stats(FRAME_COUNTS *counts, const MODE_INFO *mi) {
4055 const PREDICTION_MODE y_mode = mi->mbmi.mode;
4056 const PREDICTION_MODE uv_mode = mi->mbmi.uv_mode;
4057 const BLOCK_SIZE bsize = mi->mbmi.sb_type;
4059 if (bsize < BLOCK_8X8) {
4061 const int num_4x4_w = num_4x4_blocks_wide_lookup[bsize];
4062 const int num_4x4_h = num_4x4_blocks_high_lookup[bsize];
4063 for (idy = 0; idy < 2; idy += num_4x4_h)
4064 for (idx = 0; idx < 2; idx += num_4x4_w)
4065 ++counts->y_mode[0][mi->bmi[idy * 2 + idx].as_mode];
4067 ++counts->y_mode[size_group_lookup[bsize]][y_mode];
4070 ++counts->uv_mode[y_mode][uv_mode];
4073 static void encode_superblock(VP9_COMP *cpi, ThreadData *td,
4074 TOKENEXTRA **t, int output_enabled,
4075 int mi_row, int mi_col, BLOCK_SIZE bsize,
4076 PICK_MODE_CONTEXT *ctx) {
4077 VP9_COMMON *const cm = &cpi->common;
4078 MACROBLOCK *const x = &td->mb;
4079 MACROBLOCKD *const xd = &x->e_mbd;
4080 MODE_INFO *mi_8x8 = xd->mi;
4081 MODE_INFO *mi = mi_8x8;
4082 MB_MODE_INFO *mbmi = &mi->mbmi;
4083 const int seg_skip = vp9_segfeature_active(&cm->seg, mbmi->segment_id,
4085 const int mis = cm->mi_stride;
4086 const int mi_width = num_8x8_blocks_wide_lookup[bsize];
4087 const int mi_height = num_8x8_blocks_high_lookup[bsize];
4089 x->skip_recode = !x->select_tx_size && mbmi->sb_type >= BLOCK_8X8 &&
4090 cpi->oxcf.aq_mode != COMPLEXITY_AQ &&
4091 cpi->oxcf.aq_mode != CYCLIC_REFRESH_AQ &&
4092 cpi->sf.allow_skip_recode;
4094 if (!x->skip_recode && !cpi->sf.use_nonrd_pick_mode)
4095 vpx_memset(x->skip_txfm, 0, sizeof(x->skip_txfm));
4097 x->skip_optimize = ctx->is_coded;
4099 x->use_lp32x32fdct = cpi->sf.use_lp32x32fdct;
4100 x->skip_encode = (!output_enabled && cpi->sf.skip_encode_frame &&
4101 x->q_index < QIDX_SKIP_THRESH);
4106 set_ref_ptrs(cm, xd, mbmi->ref_frame[0], mbmi->ref_frame[1]);
4108 if (!is_inter_block(mbmi)) {
4111 for (plane = 0; plane < MAX_MB_PLANE; ++plane)
4112 vp9_encode_intra_block_plane(x, MAX(bsize, BLOCK_8X8), plane);
4114 sum_intra_stats(td->counts, mi);
4115 vp9_tokenize_sb(cpi, td, t, !output_enabled, MAX(bsize, BLOCK_8X8));
4118 const int is_compound = has_second_ref(mbmi);
4119 for (ref = 0; ref < 1 + is_compound; ++ref) {
4120 YV12_BUFFER_CONFIG *cfg = get_ref_frame_buffer(cpi,
4121 mbmi->ref_frame[ref]);
4122 assert(cfg != NULL);
4123 vp9_setup_pre_planes(xd, ref, cfg, mi_row, mi_col,
4124 &xd->block_refs[ref]->sf);
4126 if (!(cpi->sf.reuse_inter_pred_sby && ctx->pred_pixel_ready) || seg_skip)
4127 vp9_build_inter_predictors_sby(xd, mi_row, mi_col, MAX(bsize, BLOCK_8X8));
4129 vp9_build_inter_predictors_sbuv(xd, mi_row, mi_col, MAX(bsize, BLOCK_8X8));
4131 vp9_encode_sb(x, MAX(bsize, BLOCK_8X8));
4132 vp9_tokenize_sb(cpi, td, t, !output_enabled, MAX(bsize, BLOCK_8X8));
4135 if (output_enabled) {
4136 if (cm->tx_mode == TX_MODE_SELECT &&
4137 mbmi->sb_type >= BLOCK_8X8 &&
4138 !(is_inter_block(mbmi) && (mbmi->skip || seg_skip))) {
4139 ++get_tx_counts(max_txsize_lookup[bsize], vp9_get_tx_size_context(xd),
4140 &td->counts->tx)[mbmi->tx_size];
4144 // The new intra coding scheme requires no change of transform size
4145 if (is_inter_block(&mi->mbmi)) {
4146 tx_size = MIN(tx_mode_to_biggest_tx_size[cm->tx_mode],
4147 max_txsize_lookup[bsize]);
4149 tx_size = (bsize >= BLOCK_8X8) ? mbmi->tx_size : TX_4X4;
4152 for (y = 0; y < mi_height; y++)
4153 for (x = 0; x < mi_width; x++)
4154 if (mi_col + x < cm->mi_cols && mi_row + y < cm->mi_rows)
4155 mi_8x8[mis * y + x].src_mi->mbmi.tx_size = tx_size;
4157 ++td->counts->tx.tx_totals[mbmi->tx_size];
4158 ++td->counts->tx.tx_totals[get_uv_tx_size(mbmi, &xd->plane[1])];