1 /*
2 * Copyright (c) 2010 The WebM project authors. All Rights Reserved.
3 *
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.
9 */
10
11 #include <limits.h>
12 #include <math.h>
13 #include <stdio.h>
14
15 #include "./vp9_rtcd.h"
16 #include "./vpx_dsp_rtcd.h"
17 #include "./vpx_config.h"
18
19 #include "vpx_ports/mem.h"
20 #include "vpx_ports/vpx_timer.h"
21 #include "vpx_ports/system_state.h"
22
23 #include "vp9/common/vp9_common.h"
24 #include "vp9/common/vp9_entropy.h"
25 #include "vp9/common/vp9_entropymode.h"
26 #include "vp9/common/vp9_idct.h"
27 #include "vp9/common/vp9_mvref_common.h"
28 #include "vp9/common/vp9_pred_common.h"
29 #include "vp9/common/vp9_quant_common.h"
30 #include "vp9/common/vp9_reconintra.h"
31 #include "vp9/common/vp9_reconinter.h"
32 #include "vp9/common/vp9_seg_common.h"
33 #include "vp9/common/vp9_tile_common.h"
34
35 #include "vp9/encoder/vp9_aq_complexity.h"
36 #include "vp9/encoder/vp9_aq_cyclicrefresh.h"
37 #include "vp9/encoder/vp9_aq_variance.h"
38 #include "vp9/encoder/vp9_encodeframe.h"
39 #include "vp9/encoder/vp9_encodemb.h"
40 #include "vp9/encoder/vp9_encodemv.h"
41 #include "vp9/encoder/vp9_ethread.h"
42 #include "vp9/encoder/vp9_extend.h"
43 #include "vp9/encoder/vp9_pickmode.h"
44 #include "vp9/encoder/vp9_rd.h"
45 #include "vp9/encoder/vp9_rdopt.h"
46 #include "vp9/encoder/vp9_segmentation.h"
47 #include "vp9/encoder/vp9_tokenize.h"
48
49 static void encode_superblock(VP9_COMP *cpi, ThreadData * td,
50 TOKENEXTRA **t, int output_enabled,
51 int mi_row, int mi_col, BLOCK_SIZE bsize,
52 PICK_MODE_CONTEXT *ctx);
53
54 // This is used as a reference when computing the source variance for the
55 // purposes of activity masking.
56 // Eventually this should be replaced by custom no-reference routines,
57 // which will be faster.
58 static const uint8_t VP9_VAR_OFFS[64] = {
59 128, 128, 128, 128, 128, 128, 128, 128,
60 128, 128, 128, 128, 128, 128, 128, 128,
61 128, 128, 128, 128, 128, 128, 128, 128,
62 128, 128, 128, 128, 128, 128, 128, 128,
63 128, 128, 128, 128, 128, 128, 128, 128,
64 128, 128, 128, 128, 128, 128, 128, 128,
65 128, 128, 128, 128, 128, 128, 128, 128,
66 128, 128, 128, 128, 128, 128, 128, 128
67 };
68
69 #if CONFIG_VP9_HIGHBITDEPTH
70 static const uint16_t VP9_HIGH_VAR_OFFS_8[64] = {
71 128, 128, 128, 128, 128, 128, 128, 128,
72 128, 128, 128, 128, 128, 128, 128, 128,
73 128, 128, 128, 128, 128, 128, 128, 128,
74 128, 128, 128, 128, 128, 128, 128, 128,
75 128, 128, 128, 128, 128, 128, 128, 128,
76 128, 128, 128, 128, 128, 128, 128, 128,
77 128, 128, 128, 128, 128, 128, 128, 128,
78 128, 128, 128, 128, 128, 128, 128, 128
79 };
80
81 static const uint16_t VP9_HIGH_VAR_OFFS_10[64] = {
82 128*4, 128*4, 128*4, 128*4, 128*4, 128*4, 128*4, 128*4,
83 128*4, 128*4, 128*4, 128*4, 128*4, 128*4, 128*4, 128*4,
84 128*4, 128*4, 128*4, 128*4, 128*4, 128*4, 128*4, 128*4,
85 128*4, 128*4, 128*4, 128*4, 128*4, 128*4, 128*4, 128*4,
86 128*4, 128*4, 128*4, 128*4, 128*4, 128*4, 128*4, 128*4,
87 128*4, 128*4, 128*4, 128*4, 128*4, 128*4, 128*4, 128*4,
88 128*4, 128*4, 128*4, 128*4, 128*4, 128*4, 128*4, 128*4,
89 128*4, 128*4, 128*4, 128*4, 128*4, 128*4, 128*4, 128*4
90 };
91
92 static const uint16_t VP9_HIGH_VAR_OFFS_12[64] = {
93 128*16, 128*16, 128*16, 128*16, 128*16, 128*16, 128*16, 128*16,
94 128*16, 128*16, 128*16, 128*16, 128*16, 128*16, 128*16, 128*16,
95 128*16, 128*16, 128*16, 128*16, 128*16, 128*16, 128*16, 128*16,
96 128*16, 128*16, 128*16, 128*16, 128*16, 128*16, 128*16, 128*16,
97 128*16, 128*16, 128*16, 128*16, 128*16, 128*16, 128*16, 128*16,
98 128*16, 128*16, 128*16, 128*16, 128*16, 128*16, 128*16, 128*16,
99 128*16, 128*16, 128*16, 128*16, 128*16, 128*16, 128*16, 128*16,
100 128*16, 128*16, 128*16, 128*16, 128*16, 128*16, 128*16, 128*16
101 };
102 #endif // CONFIG_VP9_HIGHBITDEPTH
103
vp9_get_sby_perpixel_variance(VP9_COMP * cpi,const struct buf_2d * ref,BLOCK_SIZE bs)104 unsigned int vp9_get_sby_perpixel_variance(VP9_COMP *cpi,
105 const struct buf_2d *ref,
106 BLOCK_SIZE bs) {
107 unsigned int sse;
108 const unsigned int var = cpi->fn_ptr[bs].vf(ref->buf, ref->stride,
109 VP9_VAR_OFFS, 0, &sse);
110 return ROUND_POWER_OF_TWO(var, num_pels_log2_lookup[bs]);
111 }
112
113 #if CONFIG_VP9_HIGHBITDEPTH
vp9_high_get_sby_perpixel_variance(VP9_COMP * cpi,const struct buf_2d * ref,BLOCK_SIZE bs,int bd)114 unsigned int vp9_high_get_sby_perpixel_variance(
115 VP9_COMP *cpi, const struct buf_2d *ref, BLOCK_SIZE bs, int bd) {
116 unsigned int var, sse;
117 switch (bd) {
118 case 10:
119 var = cpi->fn_ptr[bs].vf(ref->buf, ref->stride,
120 CONVERT_TO_BYTEPTR(VP9_HIGH_VAR_OFFS_10),
121 0, &sse);
122 break;
123 case 12:
124 var = cpi->fn_ptr[bs].vf(ref->buf, ref->stride,
125 CONVERT_TO_BYTEPTR(VP9_HIGH_VAR_OFFS_12),
126 0, &sse);
127 break;
128 case 8:
129 default:
130 var = cpi->fn_ptr[bs].vf(ref->buf, ref->stride,
131 CONVERT_TO_BYTEPTR(VP9_HIGH_VAR_OFFS_8),
132 0, &sse);
133 break;
134 }
135 return ROUND_POWER_OF_TWO(var, num_pels_log2_lookup[bs]);
136 }
137 #endif // CONFIG_VP9_HIGHBITDEPTH
138
get_sby_perpixel_diff_variance(VP9_COMP * cpi,const struct buf_2d * ref,int mi_row,int mi_col,BLOCK_SIZE bs)139 static unsigned int get_sby_perpixel_diff_variance(VP9_COMP *cpi,
140 const struct buf_2d *ref,
141 int mi_row, int mi_col,
142 BLOCK_SIZE bs) {
143 unsigned int sse, var;
144 uint8_t *last_y;
145 const YV12_BUFFER_CONFIG *last = get_ref_frame_buffer(cpi, LAST_FRAME);
146
147 assert(last != NULL);
148 last_y =
149 &last->y_buffer[mi_row * MI_SIZE * last->y_stride + mi_col * MI_SIZE];
150 var = cpi->fn_ptr[bs].vf(ref->buf, ref->stride, last_y, last->y_stride, &sse);
151 return ROUND_POWER_OF_TWO(var, num_pels_log2_lookup[bs]);
152 }
153
get_rd_var_based_fixed_partition(VP9_COMP * cpi,MACROBLOCK * x,int mi_row,int mi_col)154 static BLOCK_SIZE get_rd_var_based_fixed_partition(VP9_COMP *cpi, MACROBLOCK *x,
155 int mi_row,
156 int mi_col) {
157 unsigned int var = get_sby_perpixel_diff_variance(cpi, &x->plane[0].src,
158 mi_row, mi_col,
159 BLOCK_64X64);
160 if (var < 8)
161 return BLOCK_64X64;
162 else if (var < 128)
163 return BLOCK_32X32;
164 else if (var < 2048)
165 return BLOCK_16X16;
166 else
167 return BLOCK_8X8;
168 }
169
170 // Lighter version of set_offsets that only sets the mode info
171 // pointers.
set_mode_info_offsets(VP9_COMMON * const cm,MACROBLOCK * const x,MACROBLOCKD * const xd,int mi_row,int mi_col)172 static INLINE void set_mode_info_offsets(VP9_COMMON *const cm,
173 MACROBLOCK *const x,
174 MACROBLOCKD *const xd,
175 int mi_row,
176 int mi_col) {
177 const int idx_str = xd->mi_stride * mi_row + mi_col;
178 xd->mi = cm->mi_grid_visible + idx_str;
179 xd->mi[0] = cm->mi + idx_str;
180 x->mbmi_ext = x->mbmi_ext_base + (mi_row * cm->mi_cols + mi_col);
181 }
182
set_offsets(VP9_COMP * cpi,const TileInfo * const tile,MACROBLOCK * const x,int mi_row,int mi_col,BLOCK_SIZE bsize)183 static void set_offsets(VP9_COMP *cpi, const TileInfo *const tile,
184 MACROBLOCK *const x, int mi_row, int mi_col,
185 BLOCK_SIZE bsize) {
186 VP9_COMMON *const cm = &cpi->common;
187 MACROBLOCKD *const xd = &x->e_mbd;
188 MB_MODE_INFO *mbmi;
189 const int mi_width = num_8x8_blocks_wide_lookup[bsize];
190 const int mi_height = num_8x8_blocks_high_lookup[bsize];
191 const struct segmentation *const seg = &cm->seg;
192
193 set_skip_context(xd, mi_row, mi_col);
194
195 set_mode_info_offsets(cm, x, xd, mi_row, mi_col);
196
197 mbmi = &xd->mi[0]->mbmi;
198
199 // Set up destination pointers.
200 vp9_setup_dst_planes(xd->plane, get_frame_new_buffer(cm), mi_row, mi_col);
201
202 // Set up limit values for MV components.
203 // Mv beyond the range do not produce new/different prediction block.
204 x->mv_row_min = -(((mi_row + mi_height) * MI_SIZE) + VP9_INTERP_EXTEND);
205 x->mv_col_min = -(((mi_col + mi_width) * MI_SIZE) + VP9_INTERP_EXTEND);
206 x->mv_row_max = (cm->mi_rows - mi_row) * MI_SIZE + VP9_INTERP_EXTEND;
207 x->mv_col_max = (cm->mi_cols - mi_col) * MI_SIZE + VP9_INTERP_EXTEND;
208
209 // Set up distance of MB to edge of frame in 1/8th pel units.
210 assert(!(mi_col & (mi_width - 1)) && !(mi_row & (mi_height - 1)));
211 set_mi_row_col(xd, tile, mi_row, mi_height, mi_col, mi_width,
212 cm->mi_rows, cm->mi_cols);
213
214 // Set up source buffers.
215 vp9_setup_src_planes(x, cpi->Source, mi_row, mi_col);
216
217 // R/D setup.
218 x->rddiv = cpi->rd.RDDIV;
219 x->rdmult = cpi->rd.RDMULT;
220
221 // Setup segment ID.
222 if (seg->enabled) {
223 if (cpi->oxcf.aq_mode != VARIANCE_AQ) {
224 const uint8_t *const map = seg->update_map ? cpi->segmentation_map
225 : cm->last_frame_seg_map;
226 mbmi->segment_id = get_segment_id(cm, map, bsize, mi_row, mi_col);
227 }
228 vp9_init_plane_quantizers(cpi, x);
229
230 x->encode_breakout = cpi->segment_encode_breakout[mbmi->segment_id];
231 } else {
232 mbmi->segment_id = 0;
233 x->encode_breakout = cpi->encode_breakout;
234 }
235
236 // required by vp9_append_sub8x8_mvs_for_idx() and vp9_find_best_ref_mvs()
237 xd->tile = *tile;
238 }
239
duplicate_mode_info_in_sb(VP9_COMMON * cm,MACROBLOCKD * xd,int mi_row,int mi_col,BLOCK_SIZE bsize)240 static void duplicate_mode_info_in_sb(VP9_COMMON *cm, MACROBLOCKD *xd,
241 int mi_row, int mi_col,
242 BLOCK_SIZE bsize) {
243 const int block_width = num_8x8_blocks_wide_lookup[bsize];
244 const int block_height = num_8x8_blocks_high_lookup[bsize];
245 int i, j;
246 for (j = 0; j < block_height; ++j)
247 for (i = 0; i < block_width; ++i) {
248 if (mi_row + j < cm->mi_rows && mi_col + i < cm->mi_cols)
249 xd->mi[j * xd->mi_stride + i] = xd->mi[0];
250 }
251 }
252
set_block_size(VP9_COMP * const cpi,MACROBLOCK * const x,MACROBLOCKD * const xd,int mi_row,int mi_col,BLOCK_SIZE bsize)253 static void set_block_size(VP9_COMP * const cpi,
254 MACROBLOCK *const x,
255 MACROBLOCKD *const xd,
256 int mi_row, int mi_col,
257 BLOCK_SIZE bsize) {
258 if (cpi->common.mi_cols > mi_col && cpi->common.mi_rows > mi_row) {
259 set_mode_info_offsets(&cpi->common, x, xd, mi_row, mi_col);
260 xd->mi[0]->mbmi.sb_type = bsize;
261 }
262 }
263
264 typedef struct {
265 int64_t sum_square_error;
266 int64_t sum_error;
267 int log2_count;
268 int variance;
269 } var;
270
271 typedef struct {
272 var none;
273 var horz[2];
274 var vert[2];
275 } partition_variance;
276
277 typedef struct {
278 partition_variance part_variances;
279 var split[4];
280 } v4x4;
281
282 typedef struct {
283 partition_variance part_variances;
284 v4x4 split[4];
285 } v8x8;
286
287 typedef struct {
288 partition_variance part_variances;
289 v8x8 split[4];
290 } v16x16;
291
292 typedef struct {
293 partition_variance part_variances;
294 v16x16 split[4];
295 } v32x32;
296
297 typedef struct {
298 partition_variance part_variances;
299 v32x32 split[4];
300 } v64x64;
301
302 typedef struct {
303 partition_variance *part_variances;
304 var *split[4];
305 } variance_node;
306
307 typedef enum {
308 V16X16,
309 V32X32,
310 V64X64,
311 } TREE_LEVEL;
312
tree_to_node(void * data,BLOCK_SIZE bsize,variance_node * node)313 static void tree_to_node(void *data, BLOCK_SIZE bsize, variance_node *node) {
314 int i;
315 node->part_variances = NULL;
316 switch (bsize) {
317 case BLOCK_64X64: {
318 v64x64 *vt = (v64x64 *) data;
319 node->part_variances = &vt->part_variances;
320 for (i = 0; i < 4; i++)
321 node->split[i] = &vt->split[i].part_variances.none;
322 break;
323 }
324 case BLOCK_32X32: {
325 v32x32 *vt = (v32x32 *) 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;
329 break;
330 }
331 case BLOCK_16X16: {
332 v16x16 *vt = (v16x16 *) 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;
336 break;
337 }
338 case BLOCK_8X8: {
339 v8x8 *vt = (v8x8 *) 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;
343 break;
344 }
345 case BLOCK_4X4: {
346 v4x4 *vt = (v4x4 *) data;
347 node->part_variances = &vt->part_variances;
348 for (i = 0; i < 4; i++)
349 node->split[i] = &vt->split[i];
350 break;
351 }
352 default: {
353 assert(0);
354 break;
355 }
356 }
357 }
358
359 // Set variance values given sum square error, sum error, count.
fill_variance(int64_t s2,int64_t s,int c,var * v)360 static void fill_variance(int64_t s2, int64_t s, int c, var *v) {
361 v->sum_square_error = s2;
362 v->sum_error = s;
363 v->log2_count = c;
364 }
365
get_variance(var * v)366 static void get_variance(var *v) {
367 v->variance = (int)(256 * (v->sum_square_error -
368 ((v->sum_error * v->sum_error) >> v->log2_count)) >> v->log2_count);
369 }
370
sum_2_variances(const var * a,const var * b,var * r)371 static void sum_2_variances(const var *a, const var *b, var *r) {
372 assert(a->log2_count == b->log2_count);
373 fill_variance(a->sum_square_error + b->sum_square_error,
374 a->sum_error + b->sum_error, a->log2_count + 1, r);
375 }
376
fill_variance_tree(void * data,BLOCK_SIZE bsize)377 static void fill_variance_tree(void *data, BLOCK_SIZE bsize) {
378 variance_node node;
379 memset(&node, 0, sizeof(node));
380 tree_to_node(data, bsize, &node);
381 sum_2_variances(node.split[0], node.split[1], &node.part_variances->horz[0]);
382 sum_2_variances(node.split[2], node.split[3], &node.part_variances->horz[1]);
383 sum_2_variances(node.split[0], node.split[2], &node.part_variances->vert[0]);
384 sum_2_variances(node.split[1], node.split[3], &node.part_variances->vert[1]);
385 sum_2_variances(&node.part_variances->vert[0], &node.part_variances->vert[1],
386 &node.part_variances->none);
387 }
388
set_vt_partitioning(VP9_COMP * cpi,MACROBLOCK * const x,MACROBLOCKD * const xd,void * data,BLOCK_SIZE bsize,int mi_row,int mi_col,int64_t threshold,BLOCK_SIZE bsize_min,int force_split)389 static int set_vt_partitioning(VP9_COMP *cpi,
390 MACROBLOCK *const x,
391 MACROBLOCKD *const xd,
392 void *data,
393 BLOCK_SIZE bsize,
394 int mi_row,
395 int mi_col,
396 int64_t threshold,
397 BLOCK_SIZE bsize_min,
398 int force_split) {
399 VP9_COMMON * const cm = &cpi->common;
400 variance_node vt;
401 const int block_width = num_8x8_blocks_wide_lookup[bsize];
402 const int block_height = num_8x8_blocks_high_lookup[bsize];
403 const int low_res = (cm->width <= 352 && cm->height <= 288);
404
405 assert(block_height == block_width);
406 tree_to_node(data, bsize, &vt);
407
408 if (force_split == 1)
409 return 0;
410
411 // For bsize=bsize_min (16x16/8x8 for 8x8/4x4 downsampling), select if
412 // variance is below threshold, otherwise split will be selected.
413 // No check for vert/horiz split as too few samples for variance.
414 if (bsize == bsize_min) {
415 // Variance already computed to set the force_split.
416 if (low_res || cm->frame_type == KEY_FRAME)
417 get_variance(&vt.part_variances->none);
418 if (mi_col + block_width / 2 < cm->mi_cols &&
419 mi_row + block_height / 2 < cm->mi_rows &&
420 vt.part_variances->none.variance < threshold) {
421 set_block_size(cpi, x, xd, mi_row, mi_col, bsize);
422 return 1;
423 }
424 return 0;
425 } else if (bsize > bsize_min) {
426 // Variance already computed to set the force_split.
427 if (low_res || cm->frame_type == KEY_FRAME)
428 get_variance(&vt.part_variances->none);
429 // For key frame: take split for bsize above 32X32 or very high variance.
430 if (cm->frame_type == KEY_FRAME &&
431 (bsize > BLOCK_32X32 ||
432 vt.part_variances->none.variance > (threshold << 4))) {
433 return 0;
434 }
435 // If variance is low, take the bsize (no split).
436 if (mi_col + block_width / 2 < cm->mi_cols &&
437 mi_row + block_height / 2 < cm->mi_rows &&
438 vt.part_variances->none.variance < threshold) {
439 set_block_size(cpi, x, xd, mi_row, mi_col, bsize);
440 return 1;
441 }
442
443 // Check vertical split.
444 if (mi_row + block_height / 2 < cm->mi_rows) {
445 BLOCK_SIZE subsize = get_subsize(bsize, PARTITION_VERT);
446 get_variance(&vt.part_variances->vert[0]);
447 get_variance(&vt.part_variances->vert[1]);
448 if (vt.part_variances->vert[0].variance < threshold &&
449 vt.part_variances->vert[1].variance < threshold &&
450 get_plane_block_size(subsize, &xd->plane[1]) < BLOCK_INVALID) {
451 set_block_size(cpi, x, xd, mi_row, mi_col, subsize);
452 set_block_size(cpi, x, xd, mi_row, mi_col + block_width / 2, subsize);
453 return 1;
454 }
455 }
456 // Check horizontal split.
457 if (mi_col + block_width / 2 < cm->mi_cols) {
458 BLOCK_SIZE subsize = get_subsize(bsize, PARTITION_HORZ);
459 get_variance(&vt.part_variances->horz[0]);
460 get_variance(&vt.part_variances->horz[1]);
461 if (vt.part_variances->horz[0].variance < threshold &&
462 vt.part_variances->horz[1].variance < threshold &&
463 get_plane_block_size(subsize, &xd->plane[1]) < BLOCK_INVALID) {
464 set_block_size(cpi, x, xd, mi_row, mi_col, subsize);
465 set_block_size(cpi, x, xd, mi_row + block_height / 2, mi_col, subsize);
466 return 1;
467 }
468 }
469
470 return 0;
471 }
472 return 0;
473 }
474
475 // Set the variance split thresholds for following the block sizes:
476 // 0 - threshold_64x64, 1 - threshold_32x32, 2 - threshold_16x16,
477 // 3 - vbp_threshold_8x8. vbp_threshold_8x8 (to split to 4x4 partition) is
478 // currently only used on key frame.
set_vbp_thresholds(VP9_COMP * cpi,int64_t thresholds[],int q)479 static void set_vbp_thresholds(VP9_COMP *cpi, int64_t thresholds[], int q) {
480 VP9_COMMON *const cm = &cpi->common;
481 const int is_key_frame = (cm->frame_type == KEY_FRAME);
482 const int threshold_multiplier = is_key_frame ? 20 : 1;
483 const int64_t threshold_base = (int64_t)(threshold_multiplier *
484 cpi->y_dequant[q][1]);
485 if (is_key_frame) {
486 thresholds[0] = threshold_base;
487 thresholds[1] = threshold_base >> 2;
488 thresholds[2] = threshold_base >> 2;
489 thresholds[3] = threshold_base << 2;
490 } else {
491 thresholds[1] = threshold_base;
492 if (cm->width <= 352 && cm->height <= 288) {
493 thresholds[0] = threshold_base >> 2;
494 thresholds[2] = threshold_base << 3;
495 } else {
496 thresholds[0] = threshold_base;
497 thresholds[1] = (5 * threshold_base) >> 2;
498 if (cm->width >= 1920 && cm->height >= 1080)
499 thresholds[1] = (7 * threshold_base) >> 2;
500 thresholds[2] = threshold_base << cpi->oxcf.speed;
501 }
502 }
503 }
504
vp9_set_variance_partition_thresholds(VP9_COMP * cpi,int q)505 void vp9_set_variance_partition_thresholds(VP9_COMP *cpi, int q) {
506 VP9_COMMON *const cm = &cpi->common;
507 SPEED_FEATURES *const sf = &cpi->sf;
508 const int is_key_frame = (cm->frame_type == KEY_FRAME);
509 if (sf->partition_search_type != VAR_BASED_PARTITION &&
510 sf->partition_search_type != REFERENCE_PARTITION) {
511 return;
512 } else {
513 set_vbp_thresholds(cpi, cpi->vbp_thresholds, q);
514 // The thresholds below are not changed locally.
515 if (is_key_frame) {
516 cpi->vbp_threshold_sad = 0;
517 cpi->vbp_bsize_min = BLOCK_8X8;
518 } else {
519 if (cm->width <= 352 && cm->height <= 288)
520 cpi->vbp_threshold_sad = 100;
521 else
522 cpi->vbp_threshold_sad = (cpi->y_dequant[q][1] << 1) > 1000 ?
523 (cpi->y_dequant[q][1] << 1) : 1000;
524 cpi->vbp_bsize_min = BLOCK_16X16;
525 }
526 cpi->vbp_threshold_minmax = 15 + (q >> 3);
527 }
528 }
529
530 // Compute the minmax over the 8x8 subblocks.
compute_minmax_8x8(const uint8_t * s,int sp,const uint8_t * d,int dp,int x16_idx,int y16_idx,int highbd_flag,int pixels_wide,int pixels_high)531 static int compute_minmax_8x8(const uint8_t *s, int sp, const uint8_t *d,
532 int dp, int x16_idx, int y16_idx,
533 #if CONFIG_VP9_HIGHBITDEPTH
534 int highbd_flag,
535 #endif
536 int pixels_wide,
537 int pixels_high) {
538 int k;
539 int minmax_max = 0;
540 int minmax_min = 255;
541 // Loop over the 4 8x8 subblocks.
542 for (k = 0; k < 4; k++) {
543 int x8_idx = x16_idx + ((k & 1) << 3);
544 int y8_idx = y16_idx + ((k >> 1) << 3);
545 int min = 0;
546 int max = 0;
547 if (x8_idx < pixels_wide && y8_idx < pixels_high) {
548 #if CONFIG_VP9_HIGHBITDEPTH
549 if (highbd_flag & YV12_FLAG_HIGHBITDEPTH) {
550 vp9_highbd_minmax_8x8(s + y8_idx * sp + x8_idx, sp,
551 d + y8_idx * dp + x8_idx, dp,
552 &min, &max);
553 } else {
554 vp9_minmax_8x8(s + y8_idx * sp + x8_idx, sp,
555 d + y8_idx * dp + x8_idx, dp,
556 &min, &max);
557 }
558 #else
559 vp9_minmax_8x8(s + y8_idx * sp + x8_idx, sp,
560 d + y8_idx * dp + x8_idx, dp,
561 &min, &max);
562 #endif
563 if ((max - min) > minmax_max)
564 minmax_max = (max - min);
565 if ((max - min) < minmax_min)
566 minmax_min = (max - min);
567 }
568 }
569 return (minmax_max - minmax_min);
570 }
571
fill_variance_4x4avg(const uint8_t * s,int sp,const uint8_t * d,int dp,int x8_idx,int y8_idx,v8x8 * vst,int highbd_flag,int pixels_wide,int pixels_high,int is_key_frame)572 static void fill_variance_4x4avg(const uint8_t *s, int sp, const uint8_t *d,
573 int dp, int x8_idx, int y8_idx, v8x8 *vst,
574 #if CONFIG_VP9_HIGHBITDEPTH
575 int highbd_flag,
576 #endif
577 int pixels_wide,
578 int pixels_high,
579 int is_key_frame) {
580 int k;
581 for (k = 0; k < 4; k++) {
582 int x4_idx = x8_idx + ((k & 1) << 2);
583 int y4_idx = y8_idx + ((k >> 1) << 2);
584 unsigned int sse = 0;
585 int sum = 0;
586 if (x4_idx < pixels_wide && y4_idx < pixels_high) {
587 int s_avg;
588 int d_avg = 128;
589 #if CONFIG_VP9_HIGHBITDEPTH
590 if (highbd_flag & YV12_FLAG_HIGHBITDEPTH) {
591 s_avg = vp9_highbd_avg_4x4(s + y4_idx * sp + x4_idx, sp);
592 if (!is_key_frame)
593 d_avg = vp9_highbd_avg_4x4(d + y4_idx * dp + x4_idx, dp);
594 } else {
595 s_avg = vp9_avg_4x4(s + y4_idx * sp + x4_idx, sp);
596 if (!is_key_frame)
597 d_avg = vp9_avg_4x4(d + y4_idx * dp + x4_idx, dp);
598 }
599 #else
600 s_avg = vp9_avg_4x4(s + y4_idx * sp + x4_idx, sp);
601 if (!is_key_frame)
602 d_avg = vp9_avg_4x4(d + y4_idx * dp + x4_idx, dp);
603 #endif
604 sum = s_avg - d_avg;
605 sse = sum * sum;
606 }
607 fill_variance(sse, sum, 0, &vst->split[k].part_variances.none);
608 }
609 }
610
fill_variance_8x8avg(const uint8_t * s,int sp,const uint8_t * d,int dp,int x16_idx,int y16_idx,v16x16 * vst,int highbd_flag,int pixels_wide,int pixels_high,int is_key_frame)611 static void fill_variance_8x8avg(const uint8_t *s, int sp, const uint8_t *d,
612 int dp, int x16_idx, int y16_idx, v16x16 *vst,
613 #if CONFIG_VP9_HIGHBITDEPTH
614 int highbd_flag,
615 #endif
616 int pixels_wide,
617 int pixels_high,
618 int is_key_frame) {
619 int k;
620 for (k = 0; k < 4; k++) {
621 int x8_idx = x16_idx + ((k & 1) << 3);
622 int y8_idx = y16_idx + ((k >> 1) << 3);
623 unsigned int sse = 0;
624 int sum = 0;
625 if (x8_idx < pixels_wide && y8_idx < pixels_high) {
626 int s_avg;
627 int d_avg = 128;
628 #if CONFIG_VP9_HIGHBITDEPTH
629 if (highbd_flag & YV12_FLAG_HIGHBITDEPTH) {
630 s_avg = vp9_highbd_avg_8x8(s + y8_idx * sp + x8_idx, sp);
631 if (!is_key_frame)
632 d_avg = vp9_highbd_avg_8x8(d + y8_idx * dp + x8_idx, dp);
633 } else {
634 s_avg = vp9_avg_8x8(s + y8_idx * sp + x8_idx, sp);
635 if (!is_key_frame)
636 d_avg = vp9_avg_8x8(d + y8_idx * dp + x8_idx, dp);
637 }
638 #else
639 s_avg = vp9_avg_8x8(s + y8_idx * sp + x8_idx, sp);
640 if (!is_key_frame)
641 d_avg = vp9_avg_8x8(d + y8_idx * dp + x8_idx, dp);
642 #endif
643 sum = s_avg - d_avg;
644 sse = sum * sum;
645 }
646 fill_variance(sse, sum, 0, &vst->split[k].part_variances.none);
647 }
648 }
649
650 // This function chooses partitioning based on the variance between source and
651 // reconstructed last, where variance is computed for down-sampled inputs.
choose_partitioning(VP9_COMP * cpi,const TileInfo * const tile,MACROBLOCK * x,int mi_row,int mi_col)652 static int choose_partitioning(VP9_COMP *cpi,
653 const TileInfo *const tile,
654 MACROBLOCK *x,
655 int mi_row, int mi_col) {
656 VP9_COMMON * const cm = &cpi->common;
657 MACROBLOCKD *xd = &x->e_mbd;
658 int i, j, k, m;
659 v64x64 vt;
660 v16x16 vt2[16];
661 int force_split[21];
662 uint8_t *s;
663 const uint8_t *d;
664 int sp;
665 int dp;
666 int pixels_wide = 64, pixels_high = 64;
667 int64_t thresholds[4] = {cpi->vbp_thresholds[0], cpi->vbp_thresholds[1],
668 cpi->vbp_thresholds[2], cpi->vbp_thresholds[3]};
669
670 // Always use 4x4 partition for key frame.
671 const int is_key_frame = (cm->frame_type == KEY_FRAME);
672 const int use_4x4_partition = is_key_frame;
673 const int low_res = (cm->width <= 352 && cm->height <= 288);
674 int variance4x4downsample[16];
675
676 int segment_id = CR_SEGMENT_ID_BASE;
677 if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ && cm->seg.enabled) {
678 const uint8_t *const map = cm->seg.update_map ? cpi->segmentation_map :
679 cm->last_frame_seg_map;
680 segment_id = get_segment_id(cm, map, BLOCK_64X64, mi_row, mi_col);
681
682 if (cyclic_refresh_segment_id_boosted(segment_id)) {
683 int q = vp9_get_qindex(&cm->seg, segment_id, cm->base_qindex);
684 set_vbp_thresholds(cpi, thresholds, q);
685 }
686 }
687
688 set_offsets(cpi, tile, x, mi_row, mi_col, BLOCK_64X64);
689
690 if (xd->mb_to_right_edge < 0)
691 pixels_wide += (xd->mb_to_right_edge >> 3);
692 if (xd->mb_to_bottom_edge < 0)
693 pixels_high += (xd->mb_to_bottom_edge >> 3);
694
695 s = x->plane[0].src.buf;
696 sp = x->plane[0].src.stride;
697
698 if (!is_key_frame && !(is_one_pass_cbr_svc(cpi) &&
699 cpi->svc.layer_context[cpi->svc.temporal_layer_id].is_key_frame)) {
700 // In the case of spatial/temporal scalable coding, the assumption here is
701 // that the temporal reference frame will always be of type LAST_FRAME.
702 // TODO(marpan): If that assumption is broken, we need to revisit this code.
703 MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
704 unsigned int uv_sad;
705 const YV12_BUFFER_CONFIG *yv12 = get_ref_frame_buffer(cpi, LAST_FRAME);
706
707 const YV12_BUFFER_CONFIG *yv12_g = NULL;
708 unsigned int y_sad, y_sad_g;
709 const BLOCK_SIZE bsize = BLOCK_32X32
710 + (mi_col + 4 < cm->mi_cols) * 2 + (mi_row + 4 < cm->mi_rows);
711
712 assert(yv12 != NULL);
713
714 if (!(is_one_pass_cbr_svc(cpi) && cpi->svc.spatial_layer_id)) {
715 // For now, GOLDEN will not be used for non-zero spatial layers, since
716 // it may not be a temporal reference.
717 yv12_g = get_ref_frame_buffer(cpi, GOLDEN_FRAME);
718 }
719
720 if (yv12_g && yv12_g != yv12 &&
721 (cpi->ref_frame_flags & VP9_GOLD_FLAG)) {
722 vp9_setup_pre_planes(xd, 0, yv12_g, mi_row, mi_col,
723 &cm->frame_refs[GOLDEN_FRAME - 1].sf);
724 y_sad_g = cpi->fn_ptr[bsize].sdf(x->plane[0].src.buf,
725 x->plane[0].src.stride,
726 xd->plane[0].pre[0].buf,
727 xd->plane[0].pre[0].stride);
728 } else {
729 y_sad_g = UINT_MAX;
730 }
731
732 vp9_setup_pre_planes(xd, 0, yv12, mi_row, mi_col,
733 &cm->frame_refs[LAST_FRAME - 1].sf);
734 mbmi->ref_frame[0] = LAST_FRAME;
735 mbmi->ref_frame[1] = NONE;
736 mbmi->sb_type = BLOCK_64X64;
737 mbmi->mv[0].as_int = 0;
738 mbmi->interp_filter = BILINEAR;
739
740 y_sad = vp9_int_pro_motion_estimation(cpi, x, bsize, mi_row, mi_col);
741 if (y_sad_g < y_sad) {
742 vp9_setup_pre_planes(xd, 0, yv12_g, mi_row, mi_col,
743 &cm->frame_refs[GOLDEN_FRAME - 1].sf);
744 mbmi->ref_frame[0] = GOLDEN_FRAME;
745 mbmi->mv[0].as_int = 0;
746 y_sad = y_sad_g;
747 } else {
748 x->pred_mv[LAST_FRAME] = mbmi->mv[0].as_mv;
749 }
750
751 vp9_build_inter_predictors_sb(xd, mi_row, mi_col, BLOCK_64X64);
752
753 for (i = 1; i <= 2; ++i) {
754 struct macroblock_plane *p = &x->plane[i];
755 struct macroblockd_plane *pd = &xd->plane[i];
756 const BLOCK_SIZE bs = get_plane_block_size(bsize, pd);
757
758 if (bs == BLOCK_INVALID)
759 uv_sad = UINT_MAX;
760 else
761 uv_sad = cpi->fn_ptr[bs].sdf(p->src.buf, p->src.stride,
762 pd->dst.buf, pd->dst.stride);
763
764 x->color_sensitivity[i - 1] = uv_sad > (y_sad >> 2);
765 }
766
767 d = xd->plane[0].dst.buf;
768 dp = xd->plane[0].dst.stride;
769
770 // If the y_sad is very small, take 64x64 as partition and exit.
771 // Don't check on boosted segment for now, as 64x64 is suppressed there.
772 if (segment_id == CR_SEGMENT_ID_BASE &&
773 y_sad < cpi->vbp_threshold_sad) {
774 const int block_width = num_8x8_blocks_wide_lookup[BLOCK_64X64];
775 const int block_height = num_8x8_blocks_high_lookup[BLOCK_64X64];
776 if (mi_col + block_width / 2 < cm->mi_cols &&
777 mi_row + block_height / 2 < cm->mi_rows) {
778 set_block_size(cpi, x, xd, mi_row, mi_col, BLOCK_64X64);
779 return 0;
780 }
781 }
782 } else {
783 d = VP9_VAR_OFFS;
784 dp = 0;
785 #if CONFIG_VP9_HIGHBITDEPTH
786 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
787 switch (xd->bd) {
788 case 10:
789 d = CONVERT_TO_BYTEPTR(VP9_HIGH_VAR_OFFS_10);
790 break;
791 case 12:
792 d = CONVERT_TO_BYTEPTR(VP9_HIGH_VAR_OFFS_12);
793 break;
794 case 8:
795 default:
796 d = CONVERT_TO_BYTEPTR(VP9_HIGH_VAR_OFFS_8);
797 break;
798 }
799 }
800 #endif // CONFIG_VP9_HIGHBITDEPTH
801 }
802
803 // Index for force_split: 0 for 64x64, 1-4 for 32x32 blocks,
804 // 5-20 for the 16x16 blocks.
805 force_split[0] = 0;
806 // Fill in the entire tree of 8x8 (or 4x4 under some conditions) variances
807 // for splits.
808 for (i = 0; i < 4; i++) {
809 const int x32_idx = ((i & 1) << 5);
810 const int y32_idx = ((i >> 1) << 5);
811 const int i2 = i << 2;
812 force_split[i + 1] = 0;
813 for (j = 0; j < 4; j++) {
814 const int x16_idx = x32_idx + ((j & 1) << 4);
815 const int y16_idx = y32_idx + ((j >> 1) << 4);
816 const int split_index = 5 + i2 + j;
817 v16x16 *vst = &vt.split[i].split[j];
818 force_split[split_index] = 0;
819 variance4x4downsample[i2 + j] = 0;
820 if (!is_key_frame) {
821 fill_variance_8x8avg(s, sp, d, dp, x16_idx, y16_idx, vst,
822 #if CONFIG_VP9_HIGHBITDEPTH
823 xd->cur_buf->flags,
824 #endif
825 pixels_wide,
826 pixels_high,
827 is_key_frame);
828 fill_variance_tree(&vt.split[i].split[j], BLOCK_16X16);
829 get_variance(&vt.split[i].split[j].part_variances.none);
830 if (vt.split[i].split[j].part_variances.none.variance >
831 thresholds[2]) {
832 // 16X16 variance is above threshold for split, so force split to 8x8
833 // for this 16x16 block (this also forces splits for upper levels).
834 force_split[split_index] = 1;
835 force_split[i + 1] = 1;
836 force_split[0] = 1;
837 } else if (vt.split[i].split[j].part_variances.none.variance >
838 thresholds[1] &&
839 !cyclic_refresh_segment_id_boosted(segment_id)) {
840 // We have some nominal amount of 16x16 variance (based on average),
841 // compute the minmax over the 8x8 sub-blocks, and if above threshold,
842 // force split to 8x8 block for this 16x16 block.
843 int minmax = compute_minmax_8x8(s, sp, d, dp, x16_idx, y16_idx,
844 #if CONFIG_VP9_HIGHBITDEPTH
845 xd->cur_buf->flags,
846 #endif
847 pixels_wide, pixels_high);
848 if (minmax > cpi->vbp_threshold_minmax) {
849 force_split[split_index] = 1;
850 force_split[i + 1] = 1;
851 force_split[0] = 1;
852 }
853 }
854 }
855 // TODO(marpan): There is an issue with variance based on 4x4 average in
856 // svc mode, don't allow it for now.
857 if (is_key_frame || (low_res && !cpi->use_svc &&
858 vt.split[i].split[j].part_variances.none.variance >
859 (thresholds[1] << 1))) {
860 force_split[split_index] = 0;
861 // Go down to 4x4 down-sampling for variance.
862 variance4x4downsample[i2 + j] = 1;
863 for (k = 0; k < 4; k++) {
864 int x8_idx = x16_idx + ((k & 1) << 3);
865 int y8_idx = y16_idx + ((k >> 1) << 3);
866 v8x8 *vst2 = is_key_frame ? &vst->split[k] :
867 &vt2[i2 + j].split[k];
868 fill_variance_4x4avg(s, sp, d, dp, x8_idx, y8_idx, vst2,
869 #if CONFIG_VP9_HIGHBITDEPTH
870 xd->cur_buf->flags,
871 #endif
872 pixels_wide,
873 pixels_high,
874 is_key_frame);
875 }
876 }
877 }
878 }
879
880 // Fill the rest of the variance tree by summing split partition values.
881 for (i = 0; i < 4; i++) {
882 const int i2 = i << 2;
883 for (j = 0; j < 4; j++) {
884 if (variance4x4downsample[i2 + j] == 1) {
885 v16x16 *vtemp = (!is_key_frame) ? &vt2[i2 + j] :
886 &vt.split[i].split[j];
887 for (m = 0; m < 4; m++)
888 fill_variance_tree(&vtemp->split[m], BLOCK_8X8);
889 fill_variance_tree(vtemp, BLOCK_16X16);
890 }
891 }
892 fill_variance_tree(&vt.split[i], BLOCK_32X32);
893 // If variance of this 32x32 block is above the threshold, force the block
894 // to split. This also forces a split on the upper (64x64) level.
895 if (!force_split[i + 1]) {
896 get_variance(&vt.split[i].part_variances.none);
897 if (vt.split[i].part_variances.none.variance > thresholds[1]) {
898 force_split[i + 1] = 1;
899 force_split[0] = 1;
900 }
901 }
902 }
903 if (!force_split[0]) {
904 fill_variance_tree(&vt, BLOCK_64X64);
905 get_variance(&vt.part_variances.none);
906 }
907
908 // Now go through the entire structure, splitting every block size until
909 // we get to one that's got a variance lower than our threshold.
910 if ( mi_col + 8 > cm->mi_cols || mi_row + 8 > cm->mi_rows ||
911 !set_vt_partitioning(cpi, x, xd, &vt, BLOCK_64X64, mi_row, mi_col,
912 thresholds[0], BLOCK_16X16, force_split[0])) {
913 for (i = 0; i < 4; ++i) {
914 const int x32_idx = ((i & 1) << 2);
915 const int y32_idx = ((i >> 1) << 2);
916 const int i2 = i << 2;
917 if (!set_vt_partitioning(cpi, x, xd, &vt.split[i], BLOCK_32X32,
918 (mi_row + y32_idx), (mi_col + x32_idx),
919 thresholds[1], BLOCK_16X16,
920 force_split[i + 1])) {
921 for (j = 0; j < 4; ++j) {
922 const int x16_idx = ((j & 1) << 1);
923 const int y16_idx = ((j >> 1) << 1);
924 // For inter frames: if variance4x4downsample[] == 1 for this 16x16
925 // block, then the variance is based on 4x4 down-sampling, so use vt2
926 // in set_vt_partioning(), otherwise use vt.
927 v16x16 *vtemp = (!is_key_frame &&
928 variance4x4downsample[i2 + j] == 1) ?
929 &vt2[i2 + j] : &vt.split[i].split[j];
930 if (!set_vt_partitioning(cpi, x, xd, vtemp, BLOCK_16X16,
931 mi_row + y32_idx + y16_idx,
932 mi_col + x32_idx + x16_idx,
933 thresholds[2],
934 cpi->vbp_bsize_min,
935 force_split[5 + i2 + j])) {
936 for (k = 0; k < 4; ++k) {
937 const int x8_idx = (k & 1);
938 const int y8_idx = (k >> 1);
939 if (use_4x4_partition) {
940 if (!set_vt_partitioning(cpi, x, xd, &vtemp->split[k],
941 BLOCK_8X8,
942 mi_row + y32_idx + y16_idx + y8_idx,
943 mi_col + x32_idx + x16_idx + x8_idx,
944 thresholds[3], BLOCK_8X8, 0)) {
945 set_block_size(cpi, x, xd,
946 (mi_row + y32_idx + y16_idx + y8_idx),
947 (mi_col + x32_idx + x16_idx + x8_idx),
948 BLOCK_4X4);
949 }
950 } else {
951 set_block_size(cpi, x, xd,
952 (mi_row + y32_idx + y16_idx + y8_idx),
953 (mi_col + x32_idx + x16_idx + x8_idx),
954 BLOCK_8X8);
955 }
956 }
957 }
958 }
959 }
960 }
961 }
962 return 0;
963 }
964
update_state(VP9_COMP * cpi,ThreadData * td,PICK_MODE_CONTEXT * ctx,int mi_row,int mi_col,BLOCK_SIZE bsize,int output_enabled)965 static void update_state(VP9_COMP *cpi, ThreadData *td,
966 PICK_MODE_CONTEXT *ctx,
967 int mi_row, int mi_col, BLOCK_SIZE bsize,
968 int output_enabled) {
969 int i, x_idx, y;
970 VP9_COMMON *const cm = &cpi->common;
971 RD_COUNTS *const rdc = &td->rd_counts;
972 MACROBLOCK *const x = &td->mb;
973 MACROBLOCKD *const xd = &x->e_mbd;
974 struct macroblock_plane *const p = x->plane;
975 struct macroblockd_plane *const pd = xd->plane;
976 MODE_INFO *mi = &ctx->mic;
977 MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
978 MODE_INFO *mi_addr = xd->mi[0];
979 const struct segmentation *const seg = &cm->seg;
980 const int bw = num_8x8_blocks_wide_lookup[mi->mbmi.sb_type];
981 const int bh = num_8x8_blocks_high_lookup[mi->mbmi.sb_type];
982 const int x_mis = MIN(bw, cm->mi_cols - mi_col);
983 const int y_mis = MIN(bh, cm->mi_rows - mi_row);
984 MV_REF *const frame_mvs =
985 cm->cur_frame->mvs + mi_row * cm->mi_cols + mi_col;
986 int w, h;
987
988 const int mis = cm->mi_stride;
989 const int mi_width = num_8x8_blocks_wide_lookup[bsize];
990 const int mi_height = num_8x8_blocks_high_lookup[bsize];
991 int max_plane;
992
993 assert(mi->mbmi.sb_type == bsize);
994
995 *mi_addr = *mi;
996 *x->mbmi_ext = ctx->mbmi_ext;
997
998 // If segmentation in use
999 if (seg->enabled) {
1000 // For in frame complexity AQ copy the segment id from the segment map.
1001 if (cpi->oxcf.aq_mode == COMPLEXITY_AQ) {
1002 const uint8_t *const map = seg->update_map ? cpi->segmentation_map
1003 : cm->last_frame_seg_map;
1004 mi_addr->mbmi.segment_id =
1005 get_segment_id(cm, map, bsize, mi_row, mi_col);
1006 }
1007 // Else for cyclic refresh mode update the segment map, set the segment id
1008 // and then update the quantizer.
1009 if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ) {
1010 vp9_cyclic_refresh_update_segment(cpi, &xd->mi[0]->mbmi, mi_row,
1011 mi_col, bsize, ctx->rate, ctx->dist,
1012 x->skip);
1013 }
1014 }
1015
1016 max_plane = is_inter_block(mbmi) ? MAX_MB_PLANE : 1;
1017 for (i = 0; i < max_plane; ++i) {
1018 p[i].coeff = ctx->coeff_pbuf[i][1];
1019 p[i].qcoeff = ctx->qcoeff_pbuf[i][1];
1020 pd[i].dqcoeff = ctx->dqcoeff_pbuf[i][1];
1021 p[i].eobs = ctx->eobs_pbuf[i][1];
1022 }
1023
1024 for (i = max_plane; i < MAX_MB_PLANE; ++i) {
1025 p[i].coeff = ctx->coeff_pbuf[i][2];
1026 p[i].qcoeff = ctx->qcoeff_pbuf[i][2];
1027 pd[i].dqcoeff = ctx->dqcoeff_pbuf[i][2];
1028 p[i].eobs = ctx->eobs_pbuf[i][2];
1029 }
1030
1031 // Restore the coding context of the MB to that that was in place
1032 // when the mode was picked for it
1033 for (y = 0; y < mi_height; y++)
1034 for (x_idx = 0; x_idx < mi_width; x_idx++)
1035 if ((xd->mb_to_right_edge >> (3 + MI_SIZE_LOG2)) + mi_width > x_idx
1036 && (xd->mb_to_bottom_edge >> (3 + MI_SIZE_LOG2)) + mi_height > y) {
1037 xd->mi[x_idx + y * mis] = mi_addr;
1038 }
1039
1040 if (cpi->oxcf.aq_mode)
1041 vp9_init_plane_quantizers(cpi, x);
1042
1043 if (is_inter_block(mbmi) && mbmi->sb_type < BLOCK_8X8) {
1044 mbmi->mv[0].as_int = mi->bmi[3].as_mv[0].as_int;
1045 mbmi->mv[1].as_int = mi->bmi[3].as_mv[1].as_int;
1046 }
1047
1048 x->skip = ctx->skip;
1049 memcpy(x->zcoeff_blk[mbmi->tx_size], ctx->zcoeff_blk,
1050 sizeof(ctx->zcoeff_blk[0]) * ctx->num_4x4_blk);
1051
1052 if (!output_enabled)
1053 return;
1054
1055 #if CONFIG_INTERNAL_STATS
1056 if (frame_is_intra_only(cm)) {
1057 static const int kf_mode_index[] = {
1058 THR_DC /*DC_PRED*/,
1059 THR_V_PRED /*V_PRED*/,
1060 THR_H_PRED /*H_PRED*/,
1061 THR_D45_PRED /*D45_PRED*/,
1062 THR_D135_PRED /*D135_PRED*/,
1063 THR_D117_PRED /*D117_PRED*/,
1064 THR_D153_PRED /*D153_PRED*/,
1065 THR_D207_PRED /*D207_PRED*/,
1066 THR_D63_PRED /*D63_PRED*/,
1067 THR_TM /*TM_PRED*/,
1068 };
1069 ++cpi->mode_chosen_counts[kf_mode_index[mbmi->mode]];
1070 } else {
1071 // Note how often each mode chosen as best
1072 ++cpi->mode_chosen_counts[ctx->best_mode_index];
1073 }
1074 #endif
1075 if (!frame_is_intra_only(cm)) {
1076 if (is_inter_block(mbmi)) {
1077 vp9_update_mv_count(td);
1078
1079 if (cm->interp_filter == SWITCHABLE) {
1080 const int ctx = vp9_get_pred_context_switchable_interp(xd);
1081 ++td->counts->switchable_interp[ctx][mbmi->interp_filter];
1082 }
1083 }
1084
1085 rdc->comp_pred_diff[SINGLE_REFERENCE] += ctx->single_pred_diff;
1086 rdc->comp_pred_diff[COMPOUND_REFERENCE] += ctx->comp_pred_diff;
1087 rdc->comp_pred_diff[REFERENCE_MODE_SELECT] += ctx->hybrid_pred_diff;
1088
1089 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i)
1090 rdc->filter_diff[i] += ctx->best_filter_diff[i];
1091 }
1092
1093 for (h = 0; h < y_mis; ++h) {
1094 MV_REF *const frame_mv = frame_mvs + h * cm->mi_cols;
1095 for (w = 0; w < x_mis; ++w) {
1096 MV_REF *const mv = frame_mv + w;
1097 mv->ref_frame[0] = mi->mbmi.ref_frame[0];
1098 mv->ref_frame[1] = mi->mbmi.ref_frame[1];
1099 mv->mv[0].as_int = mi->mbmi.mv[0].as_int;
1100 mv->mv[1].as_int = mi->mbmi.mv[1].as_int;
1101 }
1102 }
1103 }
1104
vp9_setup_src_planes(MACROBLOCK * x,const YV12_BUFFER_CONFIG * src,int mi_row,int mi_col)1105 void vp9_setup_src_planes(MACROBLOCK *x, const YV12_BUFFER_CONFIG *src,
1106 int mi_row, int mi_col) {
1107 uint8_t *const buffers[3] = {src->y_buffer, src->u_buffer, src->v_buffer };
1108 const int strides[3] = {src->y_stride, src->uv_stride, src->uv_stride };
1109 int i;
1110
1111 // Set current frame pointer.
1112 x->e_mbd.cur_buf = src;
1113
1114 for (i = 0; i < MAX_MB_PLANE; i++)
1115 setup_pred_plane(&x->plane[i].src, buffers[i], strides[i], mi_row, mi_col,
1116 NULL, x->e_mbd.plane[i].subsampling_x,
1117 x->e_mbd.plane[i].subsampling_y);
1118 }
1119
set_mode_info_seg_skip(MACROBLOCK * x,TX_MODE tx_mode,RD_COST * rd_cost,BLOCK_SIZE bsize)1120 static void set_mode_info_seg_skip(MACROBLOCK *x, TX_MODE tx_mode,
1121 RD_COST *rd_cost, BLOCK_SIZE bsize) {
1122 MACROBLOCKD *const xd = &x->e_mbd;
1123 MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
1124 INTERP_FILTER filter_ref;
1125
1126 if (xd->up_available)
1127 filter_ref = xd->mi[-xd->mi_stride]->mbmi.interp_filter;
1128 else if (xd->left_available)
1129 filter_ref = xd->mi[-1]->mbmi.interp_filter;
1130 else
1131 filter_ref = EIGHTTAP;
1132
1133 mbmi->sb_type = bsize;
1134 mbmi->mode = ZEROMV;
1135 mbmi->tx_size = MIN(max_txsize_lookup[bsize],
1136 tx_mode_to_biggest_tx_size[tx_mode]);
1137 mbmi->skip = 1;
1138 mbmi->uv_mode = DC_PRED;
1139 mbmi->ref_frame[0] = LAST_FRAME;
1140 mbmi->ref_frame[1] = NONE;
1141 mbmi->mv[0].as_int = 0;
1142 mbmi->interp_filter = filter_ref;
1143
1144 xd->mi[0]->bmi[0].as_mv[0].as_int = 0;
1145 x->skip = 1;
1146
1147 vp9_rd_cost_init(rd_cost);
1148 }
1149
set_segment_rdmult(VP9_COMP * const cpi,MACROBLOCK * const x,int8_t segment_id)1150 static int set_segment_rdmult(VP9_COMP *const cpi,
1151 MACROBLOCK *const x,
1152 int8_t segment_id) {
1153 int segment_qindex;
1154 VP9_COMMON *const cm = &cpi->common;
1155 vp9_init_plane_quantizers(cpi, x);
1156 vpx_clear_system_state();
1157 segment_qindex = vp9_get_qindex(&cm->seg, segment_id,
1158 cm->base_qindex);
1159 return vp9_compute_rd_mult(cpi, segment_qindex + cm->y_dc_delta_q);
1160 }
1161
rd_pick_sb_modes(VP9_COMP * cpi,TileDataEnc * tile_data,MACROBLOCK * const x,int mi_row,int mi_col,RD_COST * rd_cost,BLOCK_SIZE bsize,PICK_MODE_CONTEXT * ctx,int64_t best_rd)1162 static void rd_pick_sb_modes(VP9_COMP *cpi,
1163 TileDataEnc *tile_data,
1164 MACROBLOCK *const x,
1165 int mi_row, int mi_col, RD_COST *rd_cost,
1166 BLOCK_SIZE bsize, PICK_MODE_CONTEXT *ctx,
1167 int64_t best_rd) {
1168 VP9_COMMON *const cm = &cpi->common;
1169 TileInfo *const tile_info = &tile_data->tile_info;
1170 MACROBLOCKD *const xd = &x->e_mbd;
1171 MB_MODE_INFO *mbmi;
1172 struct macroblock_plane *const p = x->plane;
1173 struct macroblockd_plane *const pd = xd->plane;
1174 const AQ_MODE aq_mode = cpi->oxcf.aq_mode;
1175 int i, orig_rdmult;
1176
1177 vpx_clear_system_state();
1178
1179 // Use the lower precision, but faster, 32x32 fdct for mode selection.
1180 x->use_lp32x32fdct = 1;
1181
1182 set_offsets(cpi, tile_info, x, mi_row, mi_col, bsize);
1183 mbmi = &xd->mi[0]->mbmi;
1184 mbmi->sb_type = bsize;
1185
1186 for (i = 0; i < MAX_MB_PLANE; ++i) {
1187 p[i].coeff = ctx->coeff_pbuf[i][0];
1188 p[i].qcoeff = ctx->qcoeff_pbuf[i][0];
1189 pd[i].dqcoeff = ctx->dqcoeff_pbuf[i][0];
1190 p[i].eobs = ctx->eobs_pbuf[i][0];
1191 }
1192 ctx->is_coded = 0;
1193 ctx->skippable = 0;
1194 ctx->pred_pixel_ready = 0;
1195 x->skip_recode = 0;
1196
1197 // Set to zero to make sure we do not use the previous encoded frame stats
1198 mbmi->skip = 0;
1199
1200 #if CONFIG_VP9_HIGHBITDEPTH
1201 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
1202 x->source_variance =
1203 vp9_high_get_sby_perpixel_variance(cpi, &x->plane[0].src,
1204 bsize, xd->bd);
1205 } else {
1206 x->source_variance =
1207 vp9_get_sby_perpixel_variance(cpi, &x->plane[0].src, bsize);
1208 }
1209 #else
1210 x->source_variance =
1211 vp9_get_sby_perpixel_variance(cpi, &x->plane[0].src, bsize);
1212 #endif // CONFIG_VP9_HIGHBITDEPTH
1213
1214 // Save rdmult before it might be changed, so it can be restored later.
1215 orig_rdmult = x->rdmult;
1216
1217 if (aq_mode == VARIANCE_AQ) {
1218 const int energy = bsize <= BLOCK_16X16 ? x->mb_energy
1219 : vp9_block_energy(cpi, x, bsize);
1220 if (cm->frame_type == KEY_FRAME ||
1221 cpi->refresh_alt_ref_frame ||
1222 (cpi->refresh_golden_frame && !cpi->rc.is_src_frame_alt_ref)) {
1223 mbmi->segment_id = vp9_vaq_segment_id(energy);
1224 } else {
1225 const uint8_t *const map = cm->seg.update_map ? cpi->segmentation_map
1226 : cm->last_frame_seg_map;
1227 mbmi->segment_id = get_segment_id(cm, map, bsize, mi_row, mi_col);
1228 }
1229 x->rdmult = set_segment_rdmult(cpi, x, mbmi->segment_id);
1230 } else if (aq_mode == COMPLEXITY_AQ) {
1231 x->rdmult = set_segment_rdmult(cpi, x, mbmi->segment_id);
1232 } else if (aq_mode == CYCLIC_REFRESH_AQ) {
1233 const uint8_t *const map = cm->seg.update_map ? cpi->segmentation_map
1234 : cm->last_frame_seg_map;
1235 // If segment is boosted, use rdmult for that segment.
1236 if (cyclic_refresh_segment_id_boosted(
1237 get_segment_id(cm, map, bsize, mi_row, mi_col)))
1238 x->rdmult = vp9_cyclic_refresh_get_rdmult(cpi->cyclic_refresh);
1239 }
1240
1241 // Find best coding mode & reconstruct the MB so it is available
1242 // as a predictor for MBs that follow in the SB
1243 if (frame_is_intra_only(cm)) {
1244 vp9_rd_pick_intra_mode_sb(cpi, x, rd_cost, bsize, ctx, best_rd);
1245 } else {
1246 if (bsize >= BLOCK_8X8) {
1247 if (segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_SKIP))
1248 vp9_rd_pick_inter_mode_sb_seg_skip(cpi, tile_data, x, rd_cost, bsize,
1249 ctx, best_rd);
1250 else
1251 vp9_rd_pick_inter_mode_sb(cpi, tile_data, x, mi_row, mi_col,
1252 rd_cost, bsize, ctx, best_rd);
1253 } else {
1254 vp9_rd_pick_inter_mode_sub8x8(cpi, tile_data, x, mi_row, mi_col,
1255 rd_cost, bsize, ctx, best_rd);
1256 }
1257 }
1258
1259
1260 // Examine the resulting rate and for AQ mode 2 make a segment choice.
1261 if ((rd_cost->rate != INT_MAX) &&
1262 (aq_mode == COMPLEXITY_AQ) && (bsize >= BLOCK_16X16) &&
1263 (cm->frame_type == KEY_FRAME ||
1264 cpi->refresh_alt_ref_frame ||
1265 (cpi->refresh_golden_frame && !cpi->rc.is_src_frame_alt_ref))) {
1266 vp9_caq_select_segment(cpi, x, bsize, mi_row, mi_col, rd_cost->rate);
1267 }
1268
1269 x->rdmult = orig_rdmult;
1270
1271 // TODO(jingning) The rate-distortion optimization flow needs to be
1272 // refactored to provide proper exit/return handle.
1273 if (rd_cost->rate == INT_MAX)
1274 rd_cost->rdcost = INT64_MAX;
1275
1276 ctx->rate = rd_cost->rate;
1277 ctx->dist = rd_cost->dist;
1278 }
1279
update_stats(VP9_COMMON * cm,ThreadData * td)1280 static void update_stats(VP9_COMMON *cm, ThreadData *td) {
1281 const MACROBLOCK *x = &td->mb;
1282 const MACROBLOCKD *const xd = &x->e_mbd;
1283 const MODE_INFO *const mi = xd->mi[0];
1284 const MB_MODE_INFO *const mbmi = &mi->mbmi;
1285 const MB_MODE_INFO_EXT *const mbmi_ext = x->mbmi_ext;
1286 const BLOCK_SIZE bsize = mbmi->sb_type;
1287
1288 if (!frame_is_intra_only(cm)) {
1289 FRAME_COUNTS *const counts = td->counts;
1290 const int inter_block = is_inter_block(mbmi);
1291 const int seg_ref_active = segfeature_active(&cm->seg, mbmi->segment_id,
1292 SEG_LVL_REF_FRAME);
1293 if (!seg_ref_active) {
1294 counts->intra_inter[vp9_get_intra_inter_context(xd)][inter_block]++;
1295 // If the segment reference feature is enabled we have only a single
1296 // reference frame allowed for the segment so exclude it from
1297 // the reference frame counts used to work out probabilities.
1298 if (inter_block) {
1299 const MV_REFERENCE_FRAME ref0 = mbmi->ref_frame[0];
1300 if (cm->reference_mode == REFERENCE_MODE_SELECT)
1301 counts->comp_inter[vp9_get_reference_mode_context(cm, xd)]
1302 [has_second_ref(mbmi)]++;
1303
1304 if (has_second_ref(mbmi)) {
1305 counts->comp_ref[vp9_get_pred_context_comp_ref_p(cm, xd)]
1306 [ref0 == GOLDEN_FRAME]++;
1307 } else {
1308 counts->single_ref[vp9_get_pred_context_single_ref_p1(xd)][0]
1309 [ref0 != LAST_FRAME]++;
1310 if (ref0 != LAST_FRAME)
1311 counts->single_ref[vp9_get_pred_context_single_ref_p2(xd)][1]
1312 [ref0 != GOLDEN_FRAME]++;
1313 }
1314 }
1315 }
1316 if (inter_block &&
1317 !segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_SKIP)) {
1318 const int mode_ctx = mbmi_ext->mode_context[mbmi->ref_frame[0]];
1319 if (bsize >= BLOCK_8X8) {
1320 const PREDICTION_MODE mode = mbmi->mode;
1321 ++counts->inter_mode[mode_ctx][INTER_OFFSET(mode)];
1322 } else {
1323 const int num_4x4_w = num_4x4_blocks_wide_lookup[bsize];
1324 const int num_4x4_h = num_4x4_blocks_high_lookup[bsize];
1325 int idx, idy;
1326 for (idy = 0; idy < 2; idy += num_4x4_h) {
1327 for (idx = 0; idx < 2; idx += num_4x4_w) {
1328 const int j = idy * 2 + idx;
1329 const PREDICTION_MODE b_mode = mi->bmi[j].as_mode;
1330 ++counts->inter_mode[mode_ctx][INTER_OFFSET(b_mode)];
1331 }
1332 }
1333 }
1334 }
1335 }
1336 }
1337
restore_context(MACROBLOCK * const x,int mi_row,int mi_col,ENTROPY_CONTEXT a[16* MAX_MB_PLANE],ENTROPY_CONTEXT l[16* MAX_MB_PLANE],PARTITION_CONTEXT sa[8],PARTITION_CONTEXT sl[8],BLOCK_SIZE bsize)1338 static void restore_context(MACROBLOCK *const x, int mi_row, int mi_col,
1339 ENTROPY_CONTEXT a[16 * MAX_MB_PLANE],
1340 ENTROPY_CONTEXT l[16 * MAX_MB_PLANE],
1341 PARTITION_CONTEXT sa[8], PARTITION_CONTEXT sl[8],
1342 BLOCK_SIZE bsize) {
1343 MACROBLOCKD *const xd = &x->e_mbd;
1344 int p;
1345 const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
1346 const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
1347 int mi_width = num_8x8_blocks_wide_lookup[bsize];
1348 int mi_height = num_8x8_blocks_high_lookup[bsize];
1349 for (p = 0; p < MAX_MB_PLANE; p++) {
1350 memcpy(
1351 xd->above_context[p] + ((mi_col * 2) >> xd->plane[p].subsampling_x),
1352 a + num_4x4_blocks_wide * p,
1353 (sizeof(ENTROPY_CONTEXT) * num_4x4_blocks_wide) >>
1354 xd->plane[p].subsampling_x);
1355 memcpy(
1356 xd->left_context[p]
1357 + ((mi_row & MI_MASK) * 2 >> xd->plane[p].subsampling_y),
1358 l + num_4x4_blocks_high * p,
1359 (sizeof(ENTROPY_CONTEXT) * num_4x4_blocks_high) >>
1360 xd->plane[p].subsampling_y);
1361 }
1362 memcpy(xd->above_seg_context + mi_col, sa,
1363 sizeof(*xd->above_seg_context) * mi_width);
1364 memcpy(xd->left_seg_context + (mi_row & MI_MASK), sl,
1365 sizeof(xd->left_seg_context[0]) * mi_height);
1366 }
1367
save_context(MACROBLOCK * const x,int mi_row,int mi_col,ENTROPY_CONTEXT a[16* MAX_MB_PLANE],ENTROPY_CONTEXT l[16* MAX_MB_PLANE],PARTITION_CONTEXT sa[8],PARTITION_CONTEXT sl[8],BLOCK_SIZE bsize)1368 static void save_context(MACROBLOCK *const x, int mi_row, int mi_col,
1369 ENTROPY_CONTEXT a[16 * MAX_MB_PLANE],
1370 ENTROPY_CONTEXT l[16 * MAX_MB_PLANE],
1371 PARTITION_CONTEXT sa[8], PARTITION_CONTEXT sl[8],
1372 BLOCK_SIZE bsize) {
1373 const MACROBLOCKD *const xd = &x->e_mbd;
1374 int p;
1375 const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
1376 const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
1377 int mi_width = num_8x8_blocks_wide_lookup[bsize];
1378 int mi_height = num_8x8_blocks_high_lookup[bsize];
1379
1380 // buffer the above/left context information of the block in search.
1381 for (p = 0; p < MAX_MB_PLANE; ++p) {
1382 memcpy(
1383 a + num_4x4_blocks_wide * p,
1384 xd->above_context[p] + (mi_col * 2 >> xd->plane[p].subsampling_x),
1385 (sizeof(ENTROPY_CONTEXT) * num_4x4_blocks_wide) >>
1386 xd->plane[p].subsampling_x);
1387 memcpy(
1388 l + num_4x4_blocks_high * p,
1389 xd->left_context[p]
1390 + ((mi_row & MI_MASK) * 2 >> xd->plane[p].subsampling_y),
1391 (sizeof(ENTROPY_CONTEXT) * num_4x4_blocks_high) >>
1392 xd->plane[p].subsampling_y);
1393 }
1394 memcpy(sa, xd->above_seg_context + mi_col,
1395 sizeof(*xd->above_seg_context) * mi_width);
1396 memcpy(sl, xd->left_seg_context + (mi_row & MI_MASK),
1397 sizeof(xd->left_seg_context[0]) * mi_height);
1398 }
1399
encode_b(VP9_COMP * cpi,const TileInfo * const tile,ThreadData * td,TOKENEXTRA ** tp,int mi_row,int mi_col,int output_enabled,BLOCK_SIZE bsize,PICK_MODE_CONTEXT * ctx)1400 static void encode_b(VP9_COMP *cpi, const TileInfo *const tile,
1401 ThreadData *td,
1402 TOKENEXTRA **tp, int mi_row, int mi_col,
1403 int output_enabled, BLOCK_SIZE bsize,
1404 PICK_MODE_CONTEXT *ctx) {
1405 MACROBLOCK *const x = &td->mb;
1406 set_offsets(cpi, tile, x, mi_row, mi_col, bsize);
1407 update_state(cpi, td, ctx, mi_row, mi_col, bsize, output_enabled);
1408 encode_superblock(cpi, td, tp, output_enabled, mi_row, mi_col, bsize, ctx);
1409
1410 if (output_enabled) {
1411 update_stats(&cpi->common, td);
1412
1413 (*tp)->token = EOSB_TOKEN;
1414 (*tp)++;
1415 }
1416 }
1417
encode_sb(VP9_COMP * cpi,ThreadData * td,const TileInfo * const tile,TOKENEXTRA ** tp,int mi_row,int mi_col,int output_enabled,BLOCK_SIZE bsize,PC_TREE * pc_tree)1418 static void encode_sb(VP9_COMP *cpi, ThreadData *td,
1419 const TileInfo *const tile,
1420 TOKENEXTRA **tp, int mi_row, int mi_col,
1421 int output_enabled, BLOCK_SIZE bsize,
1422 PC_TREE *pc_tree) {
1423 VP9_COMMON *const cm = &cpi->common;
1424 MACROBLOCK *const x = &td->mb;
1425 MACROBLOCKD *const xd = &x->e_mbd;
1426
1427 const int bsl = b_width_log2_lookup[bsize], hbs = (1 << bsl) / 4;
1428 int ctx;
1429 PARTITION_TYPE partition;
1430 BLOCK_SIZE subsize = bsize;
1431
1432 if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols)
1433 return;
1434
1435 if (bsize >= BLOCK_8X8) {
1436 ctx = partition_plane_context(xd, mi_row, mi_col, bsize);
1437 subsize = get_subsize(bsize, pc_tree->partitioning);
1438 } else {
1439 ctx = 0;
1440 subsize = BLOCK_4X4;
1441 }
1442
1443 partition = partition_lookup[bsl][subsize];
1444 if (output_enabled && bsize != BLOCK_4X4)
1445 td->counts->partition[ctx][partition]++;
1446
1447 switch (partition) {
1448 case PARTITION_NONE:
1449 encode_b(cpi, tile, td, tp, mi_row, mi_col, output_enabled, subsize,
1450 &pc_tree->none);
1451 break;
1452 case PARTITION_VERT:
1453 encode_b(cpi, tile, td, tp, mi_row, mi_col, output_enabled, subsize,
1454 &pc_tree->vertical[0]);
1455 if (mi_col + hbs < cm->mi_cols && bsize > BLOCK_8X8) {
1456 encode_b(cpi, tile, td, tp, mi_row, mi_col + hbs, output_enabled,
1457 subsize, &pc_tree->vertical[1]);
1458 }
1459 break;
1460 case PARTITION_HORZ:
1461 encode_b(cpi, tile, td, tp, mi_row, mi_col, output_enabled, subsize,
1462 &pc_tree->horizontal[0]);
1463 if (mi_row + hbs < cm->mi_rows && bsize > BLOCK_8X8) {
1464 encode_b(cpi, tile, td, tp, mi_row + hbs, mi_col, output_enabled,
1465 subsize, &pc_tree->horizontal[1]);
1466 }
1467 break;
1468 case PARTITION_SPLIT:
1469 if (bsize == BLOCK_8X8) {
1470 encode_b(cpi, tile, td, tp, mi_row, mi_col, output_enabled, subsize,
1471 pc_tree->leaf_split[0]);
1472 } else {
1473 encode_sb(cpi, td, tile, tp, mi_row, mi_col, output_enabled, subsize,
1474 pc_tree->split[0]);
1475 encode_sb(cpi, td, tile, tp, mi_row, mi_col + hbs, output_enabled,
1476 subsize, pc_tree->split[1]);
1477 encode_sb(cpi, td, tile, tp, mi_row + hbs, mi_col, output_enabled,
1478 subsize, pc_tree->split[2]);
1479 encode_sb(cpi, td, tile, tp, mi_row + hbs, mi_col + hbs, output_enabled,
1480 subsize, pc_tree->split[3]);
1481 }
1482 break;
1483 default:
1484 assert(0 && "Invalid partition type.");
1485 break;
1486 }
1487
1488 if (partition != PARTITION_SPLIT || bsize == BLOCK_8X8)
1489 update_partition_context(xd, mi_row, mi_col, subsize, bsize);
1490 }
1491
1492 // Check to see if the given partition size is allowed for a specified number
1493 // of 8x8 block rows and columns remaining in the image.
1494 // If not then return the largest allowed partition size
find_partition_size(BLOCK_SIZE bsize,int rows_left,int cols_left,int * bh,int * bw)1495 static BLOCK_SIZE find_partition_size(BLOCK_SIZE bsize,
1496 int rows_left, int cols_left,
1497 int *bh, int *bw) {
1498 if (rows_left <= 0 || cols_left <= 0) {
1499 return MIN(bsize, BLOCK_8X8);
1500 } else {
1501 for (; bsize > 0; bsize -= 3) {
1502 *bh = num_8x8_blocks_high_lookup[bsize];
1503 *bw = num_8x8_blocks_wide_lookup[bsize];
1504 if ((*bh <= rows_left) && (*bw <= cols_left)) {
1505 break;
1506 }
1507 }
1508 }
1509 return bsize;
1510 }
1511
set_partial_b64x64_partition(MODE_INFO * mi,int mis,int bh_in,int bw_in,int row8x8_remaining,int col8x8_remaining,BLOCK_SIZE bsize,MODE_INFO ** mi_8x8)1512 static void set_partial_b64x64_partition(MODE_INFO *mi, int mis,
1513 int bh_in, int bw_in, int row8x8_remaining, int col8x8_remaining,
1514 BLOCK_SIZE bsize, MODE_INFO **mi_8x8) {
1515 int bh = bh_in;
1516 int r, c;
1517 for (r = 0; r < MI_BLOCK_SIZE; r += bh) {
1518 int bw = bw_in;
1519 for (c = 0; c < MI_BLOCK_SIZE; c += bw) {
1520 const int index = r * mis + c;
1521 mi_8x8[index] = mi + index;
1522 mi_8x8[index]->mbmi.sb_type = find_partition_size(bsize,
1523 row8x8_remaining - r, col8x8_remaining - c, &bh, &bw);
1524 }
1525 }
1526 }
1527
1528 // This function attempts to set all mode info entries in a given SB64
1529 // to the same block partition size.
1530 // However, at the bottom and right borders of the image the requested size
1531 // may not be allowed in which case this code attempts to choose the largest
1532 // allowable partition.
set_fixed_partitioning(VP9_COMP * cpi,const TileInfo * const tile,MODE_INFO ** mi_8x8,int mi_row,int mi_col,BLOCK_SIZE bsize)1533 static void set_fixed_partitioning(VP9_COMP *cpi, const TileInfo *const tile,
1534 MODE_INFO **mi_8x8, int mi_row, int mi_col,
1535 BLOCK_SIZE bsize) {
1536 VP9_COMMON *const cm = &cpi->common;
1537 const int mis = cm->mi_stride;
1538 const int row8x8_remaining = tile->mi_row_end - mi_row;
1539 const int col8x8_remaining = tile->mi_col_end - mi_col;
1540 int block_row, block_col;
1541 MODE_INFO *mi_upper_left = cm->mi + mi_row * mis + mi_col;
1542 int bh = num_8x8_blocks_high_lookup[bsize];
1543 int bw = num_8x8_blocks_wide_lookup[bsize];
1544
1545 assert((row8x8_remaining > 0) && (col8x8_remaining > 0));
1546
1547 // Apply the requested partition size to the SB64 if it is all "in image"
1548 if ((col8x8_remaining >= MI_BLOCK_SIZE) &&
1549 (row8x8_remaining >= MI_BLOCK_SIZE)) {
1550 for (block_row = 0; block_row < MI_BLOCK_SIZE; block_row += bh) {
1551 for (block_col = 0; block_col < MI_BLOCK_SIZE; block_col += bw) {
1552 int index = block_row * mis + block_col;
1553 mi_8x8[index] = mi_upper_left + index;
1554 mi_8x8[index]->mbmi.sb_type = bsize;
1555 }
1556 }
1557 } else {
1558 // Else this is a partial SB64.
1559 set_partial_b64x64_partition(mi_upper_left, mis, bh, bw, row8x8_remaining,
1560 col8x8_remaining, bsize, mi_8x8);
1561 }
1562 }
1563
1564 static const struct {
1565 int row;
1566 int col;
1567 } coord_lookup[16] = {
1568 // 32x32 index = 0
1569 {0, 0}, {0, 2}, {2, 0}, {2, 2},
1570 // 32x32 index = 1
1571 {0, 4}, {0, 6}, {2, 4}, {2, 6},
1572 // 32x32 index = 2
1573 {4, 0}, {4, 2}, {6, 0}, {6, 2},
1574 // 32x32 index = 3
1575 {4, 4}, {4, 6}, {6, 4}, {6, 6},
1576 };
1577
set_source_var_based_partition(VP9_COMP * cpi,const TileInfo * const tile,MACROBLOCK * const x,MODE_INFO ** mi_8x8,int mi_row,int mi_col)1578 static void set_source_var_based_partition(VP9_COMP *cpi,
1579 const TileInfo *const tile,
1580 MACROBLOCK *const x,
1581 MODE_INFO **mi_8x8,
1582 int mi_row, int mi_col) {
1583 VP9_COMMON *const cm = &cpi->common;
1584 const int mis = cm->mi_stride;
1585 const int row8x8_remaining = tile->mi_row_end - mi_row;
1586 const int col8x8_remaining = tile->mi_col_end - mi_col;
1587 MODE_INFO *mi_upper_left = cm->mi + mi_row * mis + mi_col;
1588
1589 vp9_setup_src_planes(x, cpi->Source, mi_row, mi_col);
1590
1591 assert((row8x8_remaining > 0) && (col8x8_remaining > 0));
1592
1593 // In-image SB64
1594 if ((col8x8_remaining >= MI_BLOCK_SIZE) &&
1595 (row8x8_remaining >= MI_BLOCK_SIZE)) {
1596 int i, j;
1597 int index;
1598 diff d32[4];
1599 const int offset = (mi_row >> 1) * cm->mb_cols + (mi_col >> 1);
1600 int is_larger_better = 0;
1601 int use32x32 = 0;
1602 unsigned int thr = cpi->source_var_thresh;
1603
1604 memset(d32, 0, 4 * sizeof(diff));
1605
1606 for (i = 0; i < 4; i++) {
1607 diff *d16[4];
1608
1609 for (j = 0; j < 4; j++) {
1610 int b_mi_row = coord_lookup[i * 4 + j].row;
1611 int b_mi_col = coord_lookup[i * 4 + j].col;
1612 int boffset = b_mi_row / 2 * cm->mb_cols +
1613 b_mi_col / 2;
1614
1615 d16[j] = cpi->source_diff_var + offset + boffset;
1616
1617 index = b_mi_row * mis + b_mi_col;
1618 mi_8x8[index] = mi_upper_left + index;
1619 mi_8x8[index]->mbmi.sb_type = BLOCK_16X16;
1620
1621 // TODO(yunqingwang): If d16[j].var is very large, use 8x8 partition
1622 // size to further improve quality.
1623 }
1624
1625 is_larger_better = (d16[0]->var < thr) && (d16[1]->var < thr) &&
1626 (d16[2]->var < thr) && (d16[3]->var < thr);
1627
1628 // Use 32x32 partition
1629 if (is_larger_better) {
1630 use32x32 += 1;
1631
1632 for (j = 0; j < 4; j++) {
1633 d32[i].sse += d16[j]->sse;
1634 d32[i].sum += d16[j]->sum;
1635 }
1636
1637 d32[i].var = d32[i].sse - (((int64_t)d32[i].sum * d32[i].sum) >> 10);
1638
1639 index = coord_lookup[i*4].row * mis + coord_lookup[i*4].col;
1640 mi_8x8[index] = mi_upper_left + index;
1641 mi_8x8[index]->mbmi.sb_type = BLOCK_32X32;
1642 }
1643 }
1644
1645 if (use32x32 == 4) {
1646 thr <<= 1;
1647 is_larger_better = (d32[0].var < thr) && (d32[1].var < thr) &&
1648 (d32[2].var < thr) && (d32[3].var < thr);
1649
1650 // Use 64x64 partition
1651 if (is_larger_better) {
1652 mi_8x8[0] = mi_upper_left;
1653 mi_8x8[0]->mbmi.sb_type = BLOCK_64X64;
1654 }
1655 }
1656 } else { // partial in-image SB64
1657 int bh = num_8x8_blocks_high_lookup[BLOCK_16X16];
1658 int bw = num_8x8_blocks_wide_lookup[BLOCK_16X16];
1659 set_partial_b64x64_partition(mi_upper_left, mis, bh, bw,
1660 row8x8_remaining, col8x8_remaining, BLOCK_16X16, mi_8x8);
1661 }
1662 }
1663
update_state_rt(VP9_COMP * cpi,ThreadData * td,PICK_MODE_CONTEXT * ctx,int mi_row,int mi_col,int bsize)1664 static void update_state_rt(VP9_COMP *cpi, ThreadData *td,
1665 PICK_MODE_CONTEXT *ctx,
1666 int mi_row, int mi_col, int bsize) {
1667 VP9_COMMON *const cm = &cpi->common;
1668 MACROBLOCK *const x = &td->mb;
1669 MACROBLOCKD *const xd = &x->e_mbd;
1670 MODE_INFO *const mi = xd->mi[0];
1671 MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
1672 const struct segmentation *const seg = &cm->seg;
1673 const int bw = num_8x8_blocks_wide_lookup[mi->mbmi.sb_type];
1674 const int bh = num_8x8_blocks_high_lookup[mi->mbmi.sb_type];
1675 const int x_mis = MIN(bw, cm->mi_cols - mi_col);
1676 const int y_mis = MIN(bh, cm->mi_rows - mi_row);
1677
1678 *(xd->mi[0]) = ctx->mic;
1679 *(x->mbmi_ext) = ctx->mbmi_ext;
1680
1681 if (seg->enabled && cpi->oxcf.aq_mode) {
1682 // For in frame complexity AQ or variance AQ, copy segment_id from
1683 // segmentation_map.
1684 if (cpi->oxcf.aq_mode == COMPLEXITY_AQ ||
1685 cpi->oxcf.aq_mode == VARIANCE_AQ ) {
1686 const uint8_t *const map = seg->update_map ? cpi->segmentation_map
1687 : cm->last_frame_seg_map;
1688 mbmi->segment_id = get_segment_id(cm, map, bsize, mi_row, mi_col);
1689 } else {
1690 // Setting segmentation map for cyclic_refresh.
1691 vp9_cyclic_refresh_update_segment(cpi, mbmi, mi_row, mi_col, bsize,
1692 ctx->rate, ctx->dist, x->skip);
1693 }
1694 vp9_init_plane_quantizers(cpi, x);
1695 }
1696
1697 if (is_inter_block(mbmi)) {
1698 vp9_update_mv_count(td);
1699 if (cm->interp_filter == SWITCHABLE) {
1700 const int pred_ctx = vp9_get_pred_context_switchable_interp(xd);
1701 ++td->counts->switchable_interp[pred_ctx][mbmi->interp_filter];
1702 }
1703
1704 if (mbmi->sb_type < BLOCK_8X8) {
1705 mbmi->mv[0].as_int = mi->bmi[3].as_mv[0].as_int;
1706 mbmi->mv[1].as_int = mi->bmi[3].as_mv[1].as_int;
1707 }
1708 }
1709
1710 if (cm->use_prev_frame_mvs) {
1711 MV_REF *const frame_mvs =
1712 cm->cur_frame->mvs + mi_row * cm->mi_cols + mi_col;
1713 int w, h;
1714
1715 for (h = 0; h < y_mis; ++h) {
1716 MV_REF *const frame_mv = frame_mvs + h * cm->mi_cols;
1717 for (w = 0; w < x_mis; ++w) {
1718 MV_REF *const mv = frame_mv + w;
1719 mv->ref_frame[0] = mi->mbmi.ref_frame[0];
1720 mv->ref_frame[1] = mi->mbmi.ref_frame[1];
1721 mv->mv[0].as_int = mi->mbmi.mv[0].as_int;
1722 mv->mv[1].as_int = mi->mbmi.mv[1].as_int;
1723 }
1724 }
1725 }
1726
1727 x->skip = ctx->skip;
1728 x->skip_txfm[0] = mbmi->segment_id ? 0 : ctx->skip_txfm[0];
1729 }
1730
encode_b_rt(VP9_COMP * cpi,ThreadData * td,const TileInfo * const tile,TOKENEXTRA ** tp,int mi_row,int mi_col,int output_enabled,BLOCK_SIZE bsize,PICK_MODE_CONTEXT * ctx)1731 static void encode_b_rt(VP9_COMP *cpi, ThreadData *td,
1732 const TileInfo *const tile,
1733 TOKENEXTRA **tp, int mi_row, int mi_col,
1734 int output_enabled, BLOCK_SIZE bsize,
1735 PICK_MODE_CONTEXT *ctx) {
1736 MACROBLOCK *const x = &td->mb;
1737 set_offsets(cpi, tile, x, mi_row, mi_col, bsize);
1738 update_state_rt(cpi, td, ctx, mi_row, mi_col, bsize);
1739
1740 #if CONFIG_VP9_TEMPORAL_DENOISING
1741 if (cpi->oxcf.noise_sensitivity > 0 && output_enabled &&
1742 cpi->common.frame_type != KEY_FRAME) {
1743 vp9_denoiser_denoise(&cpi->denoiser, x, mi_row, mi_col,
1744 MAX(BLOCK_8X8, bsize), ctx);
1745 }
1746 #endif
1747
1748 encode_superblock(cpi, td, tp, output_enabled, mi_row, mi_col, bsize, ctx);
1749 update_stats(&cpi->common, td);
1750
1751 (*tp)->token = EOSB_TOKEN;
1752 (*tp)++;
1753 }
1754
encode_sb_rt(VP9_COMP * cpi,ThreadData * td,const TileInfo * const tile,TOKENEXTRA ** tp,int mi_row,int mi_col,int output_enabled,BLOCK_SIZE bsize,PC_TREE * pc_tree)1755 static void encode_sb_rt(VP9_COMP *cpi, ThreadData *td,
1756 const TileInfo *const tile,
1757 TOKENEXTRA **tp, int mi_row, int mi_col,
1758 int output_enabled, BLOCK_SIZE bsize,
1759 PC_TREE *pc_tree) {
1760 VP9_COMMON *const cm = &cpi->common;
1761 MACROBLOCK *const x = &td->mb;
1762 MACROBLOCKD *const xd = &x->e_mbd;
1763
1764 const int bsl = b_width_log2_lookup[bsize], hbs = (1 << bsl) / 4;
1765 int ctx;
1766 PARTITION_TYPE partition;
1767 BLOCK_SIZE subsize;
1768
1769 if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols)
1770 return;
1771
1772 if (bsize >= BLOCK_8X8) {
1773 const int idx_str = xd->mi_stride * mi_row + mi_col;
1774 MODE_INFO ** mi_8x8 = cm->mi_grid_visible + idx_str;
1775 ctx = partition_plane_context(xd, mi_row, mi_col, bsize);
1776 subsize = mi_8x8[0]->mbmi.sb_type;
1777 } else {
1778 ctx = 0;
1779 subsize = BLOCK_4X4;
1780 }
1781
1782 partition = partition_lookup[bsl][subsize];
1783 if (output_enabled && bsize != BLOCK_4X4)
1784 td->counts->partition[ctx][partition]++;
1785
1786 switch (partition) {
1787 case PARTITION_NONE:
1788 encode_b_rt(cpi, td, tile, tp, mi_row, mi_col, output_enabled, subsize,
1789 &pc_tree->none);
1790 break;
1791 case PARTITION_VERT:
1792 encode_b_rt(cpi, td, tile, tp, mi_row, mi_col, output_enabled, subsize,
1793 &pc_tree->vertical[0]);
1794 if (mi_col + hbs < cm->mi_cols && bsize > BLOCK_8X8) {
1795 encode_b_rt(cpi, td, tile, tp, mi_row, mi_col + hbs, output_enabled,
1796 subsize, &pc_tree->vertical[1]);
1797 }
1798 break;
1799 case PARTITION_HORZ:
1800 encode_b_rt(cpi, td, tile, tp, mi_row, mi_col, output_enabled, subsize,
1801 &pc_tree->horizontal[0]);
1802 if (mi_row + hbs < cm->mi_rows && bsize > BLOCK_8X8) {
1803 encode_b_rt(cpi, td, tile, tp, mi_row + hbs, mi_col, output_enabled,
1804 subsize, &pc_tree->horizontal[1]);
1805 }
1806 break;
1807 case PARTITION_SPLIT:
1808 subsize = get_subsize(bsize, PARTITION_SPLIT);
1809 encode_sb_rt(cpi, td, tile, tp, mi_row, mi_col, output_enabled, subsize,
1810 pc_tree->split[0]);
1811 encode_sb_rt(cpi, td, tile, tp, mi_row, mi_col + hbs, output_enabled,
1812 subsize, pc_tree->split[1]);
1813 encode_sb_rt(cpi, td, tile, tp, mi_row + hbs, mi_col, output_enabled,
1814 subsize, pc_tree->split[2]);
1815 encode_sb_rt(cpi, td, tile, tp, mi_row + hbs, mi_col + hbs,
1816 output_enabled, subsize, pc_tree->split[3]);
1817 break;
1818 default:
1819 assert(0 && "Invalid partition type.");
1820 break;
1821 }
1822
1823 if (partition != PARTITION_SPLIT || bsize == BLOCK_8X8)
1824 update_partition_context(xd, mi_row, mi_col, subsize, bsize);
1825 }
1826
rd_use_partition(VP9_COMP * cpi,ThreadData * td,TileDataEnc * tile_data,MODE_INFO ** mi_8x8,TOKENEXTRA ** tp,int mi_row,int mi_col,BLOCK_SIZE bsize,int * rate,int64_t * dist,int do_recon,PC_TREE * pc_tree)1827 static void rd_use_partition(VP9_COMP *cpi,
1828 ThreadData *td,
1829 TileDataEnc *tile_data,
1830 MODE_INFO **mi_8x8, TOKENEXTRA **tp,
1831 int mi_row, int mi_col,
1832 BLOCK_SIZE bsize,
1833 int *rate, int64_t *dist,
1834 int do_recon, PC_TREE *pc_tree) {
1835 VP9_COMMON *const cm = &cpi->common;
1836 TileInfo *const tile_info = &tile_data->tile_info;
1837 MACROBLOCK *const x = &td->mb;
1838 MACROBLOCKD *const xd = &x->e_mbd;
1839 const int mis = cm->mi_stride;
1840 const int bsl = b_width_log2_lookup[bsize];
1841 const int mi_step = num_4x4_blocks_wide_lookup[bsize] / 2;
1842 const int bss = (1 << bsl) / 4;
1843 int i, pl;
1844 PARTITION_TYPE partition = PARTITION_NONE;
1845 BLOCK_SIZE subsize;
1846 ENTROPY_CONTEXT l[16 * MAX_MB_PLANE], a[16 * MAX_MB_PLANE];
1847 PARTITION_CONTEXT sl[8], sa[8];
1848 RD_COST last_part_rdc, none_rdc, chosen_rdc;
1849 BLOCK_SIZE sub_subsize = BLOCK_4X4;
1850 int splits_below = 0;
1851 BLOCK_SIZE bs_type = mi_8x8[0]->mbmi.sb_type;
1852 int do_partition_search = 1;
1853 PICK_MODE_CONTEXT *ctx = &pc_tree->none;
1854
1855 if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols)
1856 return;
1857
1858 assert(num_4x4_blocks_wide_lookup[bsize] ==
1859 num_4x4_blocks_high_lookup[bsize]);
1860
1861 vp9_rd_cost_reset(&last_part_rdc);
1862 vp9_rd_cost_reset(&none_rdc);
1863 vp9_rd_cost_reset(&chosen_rdc);
1864
1865 partition = partition_lookup[bsl][bs_type];
1866 subsize = get_subsize(bsize, partition);
1867
1868 pc_tree->partitioning = partition;
1869 save_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
1870
1871 if (bsize == BLOCK_16X16 && cpi->oxcf.aq_mode) {
1872 set_offsets(cpi, tile_info, x, mi_row, mi_col, bsize);
1873 x->mb_energy = vp9_block_energy(cpi, x, bsize);
1874 }
1875
1876 if (do_partition_search &&
1877 cpi->sf.partition_search_type == SEARCH_PARTITION &&
1878 cpi->sf.adjust_partitioning_from_last_frame) {
1879 // Check if any of the sub blocks are further split.
1880 if (partition == PARTITION_SPLIT && subsize > BLOCK_8X8) {
1881 sub_subsize = get_subsize(subsize, PARTITION_SPLIT);
1882 splits_below = 1;
1883 for (i = 0; i < 4; i++) {
1884 int jj = i >> 1, ii = i & 0x01;
1885 MODE_INFO *this_mi = mi_8x8[jj * bss * mis + ii * bss];
1886 if (this_mi && this_mi->mbmi.sb_type >= sub_subsize) {
1887 splits_below = 0;
1888 }
1889 }
1890 }
1891
1892 // If partition is not none try none unless each of the 4 splits are split
1893 // even further..
1894 if (partition != PARTITION_NONE && !splits_below &&
1895 mi_row + (mi_step >> 1) < cm->mi_rows &&
1896 mi_col + (mi_step >> 1) < cm->mi_cols) {
1897 pc_tree->partitioning = PARTITION_NONE;
1898 rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &none_rdc, bsize,
1899 ctx, INT64_MAX);
1900
1901 pl = partition_plane_context(xd, mi_row, mi_col, bsize);
1902
1903 if (none_rdc.rate < INT_MAX) {
1904 none_rdc.rate += cpi->partition_cost[pl][PARTITION_NONE];
1905 none_rdc.rdcost = RDCOST(x->rdmult, x->rddiv, none_rdc.rate,
1906 none_rdc.dist);
1907 }
1908
1909 restore_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
1910 mi_8x8[0]->mbmi.sb_type = bs_type;
1911 pc_tree->partitioning = partition;
1912 }
1913 }
1914
1915 switch (partition) {
1916 case PARTITION_NONE:
1917 rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &last_part_rdc,
1918 bsize, ctx, INT64_MAX);
1919 break;
1920 case PARTITION_HORZ:
1921 rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &last_part_rdc,
1922 subsize, &pc_tree->horizontal[0],
1923 INT64_MAX);
1924 if (last_part_rdc.rate != INT_MAX &&
1925 bsize >= BLOCK_8X8 && mi_row + (mi_step >> 1) < cm->mi_rows) {
1926 RD_COST tmp_rdc;
1927 PICK_MODE_CONTEXT *ctx = &pc_tree->horizontal[0];
1928 vp9_rd_cost_init(&tmp_rdc);
1929 update_state(cpi, td, ctx, mi_row, mi_col, subsize, 0);
1930 encode_superblock(cpi, td, tp, 0, mi_row, mi_col, subsize, ctx);
1931 rd_pick_sb_modes(cpi, tile_data, x,
1932 mi_row + (mi_step >> 1), mi_col, &tmp_rdc,
1933 subsize, &pc_tree->horizontal[1], INT64_MAX);
1934 if (tmp_rdc.rate == INT_MAX || tmp_rdc.dist == INT64_MAX) {
1935 vp9_rd_cost_reset(&last_part_rdc);
1936 break;
1937 }
1938 last_part_rdc.rate += tmp_rdc.rate;
1939 last_part_rdc.dist += tmp_rdc.dist;
1940 last_part_rdc.rdcost += tmp_rdc.rdcost;
1941 }
1942 break;
1943 case PARTITION_VERT:
1944 rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &last_part_rdc,
1945 subsize, &pc_tree->vertical[0], INT64_MAX);
1946 if (last_part_rdc.rate != INT_MAX &&
1947 bsize >= BLOCK_8X8 && mi_col + (mi_step >> 1) < cm->mi_cols) {
1948 RD_COST tmp_rdc;
1949 PICK_MODE_CONTEXT *ctx = &pc_tree->vertical[0];
1950 vp9_rd_cost_init(&tmp_rdc);
1951 update_state(cpi, td, ctx, mi_row, mi_col, subsize, 0);
1952 encode_superblock(cpi, td, tp, 0, mi_row, mi_col, subsize, ctx);
1953 rd_pick_sb_modes(cpi, tile_data, x,
1954 mi_row, mi_col + (mi_step >> 1), &tmp_rdc,
1955 subsize, &pc_tree->vertical[bsize > BLOCK_8X8],
1956 INT64_MAX);
1957 if (tmp_rdc.rate == INT_MAX || tmp_rdc.dist == INT64_MAX) {
1958 vp9_rd_cost_reset(&last_part_rdc);
1959 break;
1960 }
1961 last_part_rdc.rate += tmp_rdc.rate;
1962 last_part_rdc.dist += tmp_rdc.dist;
1963 last_part_rdc.rdcost += tmp_rdc.rdcost;
1964 }
1965 break;
1966 case PARTITION_SPLIT:
1967 if (bsize == BLOCK_8X8) {
1968 rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &last_part_rdc,
1969 subsize, pc_tree->leaf_split[0], INT64_MAX);
1970 break;
1971 }
1972 last_part_rdc.rate = 0;
1973 last_part_rdc.dist = 0;
1974 last_part_rdc.rdcost = 0;
1975 for (i = 0; i < 4; i++) {
1976 int x_idx = (i & 1) * (mi_step >> 1);
1977 int y_idx = (i >> 1) * (mi_step >> 1);
1978 int jj = i >> 1, ii = i & 0x01;
1979 RD_COST tmp_rdc;
1980 if ((mi_row + y_idx >= cm->mi_rows) || (mi_col + x_idx >= cm->mi_cols))
1981 continue;
1982
1983 vp9_rd_cost_init(&tmp_rdc);
1984 rd_use_partition(cpi, td, tile_data,
1985 mi_8x8 + jj * bss * mis + ii * bss, tp,
1986 mi_row + y_idx, mi_col + x_idx, subsize,
1987 &tmp_rdc.rate, &tmp_rdc.dist,
1988 i != 3, pc_tree->split[i]);
1989 if (tmp_rdc.rate == INT_MAX || tmp_rdc.dist == INT64_MAX) {
1990 vp9_rd_cost_reset(&last_part_rdc);
1991 break;
1992 }
1993 last_part_rdc.rate += tmp_rdc.rate;
1994 last_part_rdc.dist += tmp_rdc.dist;
1995 }
1996 break;
1997 default:
1998 assert(0);
1999 break;
2000 }
2001
2002 pl = partition_plane_context(xd, mi_row, mi_col, bsize);
2003 if (last_part_rdc.rate < INT_MAX) {
2004 last_part_rdc.rate += cpi->partition_cost[pl][partition];
2005 last_part_rdc.rdcost = RDCOST(x->rdmult, x->rddiv,
2006 last_part_rdc.rate, last_part_rdc.dist);
2007 }
2008
2009 if (do_partition_search
2010 && cpi->sf.adjust_partitioning_from_last_frame
2011 && cpi->sf.partition_search_type == SEARCH_PARTITION
2012 && partition != PARTITION_SPLIT && bsize > BLOCK_8X8
2013 && (mi_row + mi_step < cm->mi_rows ||
2014 mi_row + (mi_step >> 1) == cm->mi_rows)
2015 && (mi_col + mi_step < cm->mi_cols ||
2016 mi_col + (mi_step >> 1) == cm->mi_cols)) {
2017 BLOCK_SIZE split_subsize = get_subsize(bsize, PARTITION_SPLIT);
2018 chosen_rdc.rate = 0;
2019 chosen_rdc.dist = 0;
2020 restore_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
2021 pc_tree->partitioning = PARTITION_SPLIT;
2022
2023 // Split partition.
2024 for (i = 0; i < 4; i++) {
2025 int x_idx = (i & 1) * (mi_step >> 1);
2026 int y_idx = (i >> 1) * (mi_step >> 1);
2027 RD_COST tmp_rdc;
2028 ENTROPY_CONTEXT l[16 * MAX_MB_PLANE], a[16 * MAX_MB_PLANE];
2029 PARTITION_CONTEXT sl[8], sa[8];
2030
2031 if ((mi_row + y_idx >= cm->mi_rows) || (mi_col + x_idx >= cm->mi_cols))
2032 continue;
2033
2034 save_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
2035 pc_tree->split[i]->partitioning = PARTITION_NONE;
2036 rd_pick_sb_modes(cpi, tile_data, x,
2037 mi_row + y_idx, mi_col + x_idx, &tmp_rdc,
2038 split_subsize, &pc_tree->split[i]->none, INT64_MAX);
2039
2040 restore_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
2041
2042 if (tmp_rdc.rate == INT_MAX || tmp_rdc.dist == INT64_MAX) {
2043 vp9_rd_cost_reset(&chosen_rdc);
2044 break;
2045 }
2046
2047 chosen_rdc.rate += tmp_rdc.rate;
2048 chosen_rdc.dist += tmp_rdc.dist;
2049
2050 if (i != 3)
2051 encode_sb(cpi, td, tile_info, tp, mi_row + y_idx, mi_col + x_idx, 0,
2052 split_subsize, pc_tree->split[i]);
2053
2054 pl = partition_plane_context(xd, mi_row + y_idx, mi_col + x_idx,
2055 split_subsize);
2056 chosen_rdc.rate += cpi->partition_cost[pl][PARTITION_NONE];
2057 }
2058 pl = partition_plane_context(xd, mi_row, mi_col, bsize);
2059 if (chosen_rdc.rate < INT_MAX) {
2060 chosen_rdc.rate += cpi->partition_cost[pl][PARTITION_SPLIT];
2061 chosen_rdc.rdcost = RDCOST(x->rdmult, x->rddiv,
2062 chosen_rdc.rate, chosen_rdc.dist);
2063 }
2064 }
2065
2066 // If last_part is better set the partitioning to that.
2067 if (last_part_rdc.rdcost < chosen_rdc.rdcost) {
2068 mi_8x8[0]->mbmi.sb_type = bsize;
2069 if (bsize >= BLOCK_8X8)
2070 pc_tree->partitioning = partition;
2071 chosen_rdc = last_part_rdc;
2072 }
2073 // If none was better set the partitioning to that.
2074 if (none_rdc.rdcost < chosen_rdc.rdcost) {
2075 if (bsize >= BLOCK_8X8)
2076 pc_tree->partitioning = PARTITION_NONE;
2077 chosen_rdc = none_rdc;
2078 }
2079
2080 restore_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
2081
2082 // We must have chosen a partitioning and encoding or we'll fail later on.
2083 // No other opportunities for success.
2084 if (bsize == BLOCK_64X64)
2085 assert(chosen_rdc.rate < INT_MAX && chosen_rdc.dist < INT64_MAX);
2086
2087 if (do_recon) {
2088 int output_enabled = (bsize == BLOCK_64X64);
2089 encode_sb(cpi, td, tile_info, tp, mi_row, mi_col, output_enabled, bsize,
2090 pc_tree);
2091 }
2092
2093 *rate = chosen_rdc.rate;
2094 *dist = chosen_rdc.dist;
2095 }
2096
2097 static const BLOCK_SIZE min_partition_size[BLOCK_SIZES] = {
2098 BLOCK_4X4, BLOCK_4X4, BLOCK_4X4,
2099 BLOCK_4X4, BLOCK_4X4, BLOCK_4X4,
2100 BLOCK_8X8, BLOCK_8X8, BLOCK_8X8,
2101 BLOCK_16X16, BLOCK_16X16, BLOCK_16X16,
2102 BLOCK_16X16
2103 };
2104
2105 static const BLOCK_SIZE max_partition_size[BLOCK_SIZES] = {
2106 BLOCK_8X8, BLOCK_16X16, BLOCK_16X16,
2107 BLOCK_16X16, BLOCK_32X32, BLOCK_32X32,
2108 BLOCK_32X32, BLOCK_64X64, BLOCK_64X64,
2109 BLOCK_64X64, BLOCK_64X64, BLOCK_64X64,
2110 BLOCK_64X64
2111 };
2112
2113
2114 // Look at all the mode_info entries for blocks that are part of this
2115 // partition and find the min and max values for sb_type.
2116 // At the moment this is designed to work on a 64x64 SB but could be
2117 // adjusted to use a size parameter.
2118 //
2119 // The min and max are assumed to have been initialized prior to calling this
2120 // function so repeat calls can accumulate a min and max of more than one sb64.
get_sb_partition_size_range(MACROBLOCKD * xd,MODE_INFO ** mi_8x8,BLOCK_SIZE * min_block_size,BLOCK_SIZE * max_block_size,int bs_hist[BLOCK_SIZES])2121 static void get_sb_partition_size_range(MACROBLOCKD *xd, MODE_INFO **mi_8x8,
2122 BLOCK_SIZE *min_block_size,
2123 BLOCK_SIZE *max_block_size,
2124 int bs_hist[BLOCK_SIZES]) {
2125 int sb_width_in_blocks = MI_BLOCK_SIZE;
2126 int sb_height_in_blocks = MI_BLOCK_SIZE;
2127 int i, j;
2128 int index = 0;
2129
2130 // Check the sb_type for each block that belongs to this region.
2131 for (i = 0; i < sb_height_in_blocks; ++i) {
2132 for (j = 0; j < sb_width_in_blocks; ++j) {
2133 MODE_INFO *mi = mi_8x8[index+j];
2134 BLOCK_SIZE sb_type = mi ? mi->mbmi.sb_type : 0;
2135 bs_hist[sb_type]++;
2136 *min_block_size = MIN(*min_block_size, sb_type);
2137 *max_block_size = MAX(*max_block_size, sb_type);
2138 }
2139 index += xd->mi_stride;
2140 }
2141 }
2142
2143 // Next square block size less or equal than current block size.
2144 static const BLOCK_SIZE next_square_size[BLOCK_SIZES] = {
2145 BLOCK_4X4, BLOCK_4X4, BLOCK_4X4,
2146 BLOCK_8X8, BLOCK_8X8, BLOCK_8X8,
2147 BLOCK_16X16, BLOCK_16X16, BLOCK_16X16,
2148 BLOCK_32X32, BLOCK_32X32, BLOCK_32X32,
2149 BLOCK_64X64
2150 };
2151
2152 // Look at neighboring blocks and set a min and max partition size based on
2153 // what they chose.
rd_auto_partition_range(VP9_COMP * cpi,const TileInfo * const tile,MACROBLOCKD * const xd,int mi_row,int mi_col,BLOCK_SIZE * min_block_size,BLOCK_SIZE * max_block_size)2154 static void rd_auto_partition_range(VP9_COMP *cpi, const TileInfo *const tile,
2155 MACROBLOCKD *const xd,
2156 int mi_row, int mi_col,
2157 BLOCK_SIZE *min_block_size,
2158 BLOCK_SIZE *max_block_size) {
2159 VP9_COMMON *const cm = &cpi->common;
2160 MODE_INFO **mi = xd->mi;
2161 const int left_in_image = xd->left_available && mi[-1];
2162 const int above_in_image = xd->up_available && mi[-xd->mi_stride];
2163 const int row8x8_remaining = tile->mi_row_end - mi_row;
2164 const int col8x8_remaining = tile->mi_col_end - mi_col;
2165 int bh, bw;
2166 BLOCK_SIZE min_size = BLOCK_4X4;
2167 BLOCK_SIZE max_size = BLOCK_64X64;
2168 int bs_hist[BLOCK_SIZES] = {0};
2169
2170 // Trap case where we do not have a prediction.
2171 if (left_in_image || above_in_image || cm->frame_type != KEY_FRAME) {
2172 // Default "min to max" and "max to min"
2173 min_size = BLOCK_64X64;
2174 max_size = BLOCK_4X4;
2175
2176 // NOTE: each call to get_sb_partition_size_range() uses the previous
2177 // passed in values for min and max as a starting point.
2178 // Find the min and max partition used in previous frame at this location
2179 if (cm->frame_type != KEY_FRAME) {
2180 MODE_INFO **prev_mi =
2181 &cm->prev_mi_grid_visible[mi_row * xd->mi_stride + mi_col];
2182 get_sb_partition_size_range(xd, prev_mi, &min_size, &max_size, bs_hist);
2183 }
2184 // Find the min and max partition sizes used in the left SB64
2185 if (left_in_image) {
2186 MODE_INFO **left_sb64_mi = &mi[-MI_BLOCK_SIZE];
2187 get_sb_partition_size_range(xd, left_sb64_mi, &min_size, &max_size,
2188 bs_hist);
2189 }
2190 // Find the min and max partition sizes used in the above SB64.
2191 if (above_in_image) {
2192 MODE_INFO **above_sb64_mi = &mi[-xd->mi_stride * MI_BLOCK_SIZE];
2193 get_sb_partition_size_range(xd, above_sb64_mi, &min_size, &max_size,
2194 bs_hist);
2195 }
2196
2197 // Adjust observed min and max for "relaxed" auto partition case.
2198 if (cpi->sf.auto_min_max_partition_size == RELAXED_NEIGHBORING_MIN_MAX) {
2199 min_size = min_partition_size[min_size];
2200 max_size = max_partition_size[max_size];
2201 }
2202 }
2203
2204 // Check border cases where max and min from neighbors may not be legal.
2205 max_size = find_partition_size(max_size,
2206 row8x8_remaining, col8x8_remaining,
2207 &bh, &bw);
2208 // Test for blocks at the edge of the active image.
2209 // This may be the actual edge of the image or where there are formatting
2210 // bars.
2211 if (vp9_active_edge_sb(cpi, mi_row, mi_col)) {
2212 min_size = BLOCK_4X4;
2213 } else {
2214 min_size = MIN(cpi->sf.rd_auto_partition_min_limit,
2215 MIN(min_size, max_size));
2216 }
2217
2218 // When use_square_partition_only is true, make sure at least one square
2219 // partition is allowed by selecting the next smaller square size as
2220 // *min_block_size.
2221 if (cpi->sf.use_square_partition_only &&
2222 next_square_size[max_size] < min_size) {
2223 min_size = next_square_size[max_size];
2224 }
2225
2226 *min_block_size = min_size;
2227 *max_block_size = max_size;
2228 }
2229
2230 // TODO(jingning) refactor functions setting partition search range
set_partition_range(VP9_COMMON * cm,MACROBLOCKD * xd,int mi_row,int mi_col,BLOCK_SIZE bsize,BLOCK_SIZE * min_bs,BLOCK_SIZE * max_bs)2231 static void set_partition_range(VP9_COMMON *cm, MACROBLOCKD *xd,
2232 int mi_row, int mi_col, BLOCK_SIZE bsize,
2233 BLOCK_SIZE *min_bs, BLOCK_SIZE *max_bs) {
2234 int mi_width = num_8x8_blocks_wide_lookup[bsize];
2235 int mi_height = num_8x8_blocks_high_lookup[bsize];
2236 int idx, idy;
2237
2238 MODE_INFO *mi;
2239 const int idx_str = cm->mi_stride * mi_row + mi_col;
2240 MODE_INFO **prev_mi = &cm->prev_mi_grid_visible[idx_str];
2241 BLOCK_SIZE bs, min_size, max_size;
2242
2243 min_size = BLOCK_64X64;
2244 max_size = BLOCK_4X4;
2245
2246 if (prev_mi) {
2247 for (idy = 0; idy < mi_height; ++idy) {
2248 for (idx = 0; idx < mi_width; ++idx) {
2249 mi = prev_mi[idy * cm->mi_stride + idx];
2250 bs = mi ? mi->mbmi.sb_type : bsize;
2251 min_size = MIN(min_size, bs);
2252 max_size = MAX(max_size, bs);
2253 }
2254 }
2255 }
2256
2257 if (xd->left_available) {
2258 for (idy = 0; idy < mi_height; ++idy) {
2259 mi = xd->mi[idy * cm->mi_stride - 1];
2260 bs = mi ? mi->mbmi.sb_type : bsize;
2261 min_size = MIN(min_size, bs);
2262 max_size = MAX(max_size, bs);
2263 }
2264 }
2265
2266 if (xd->up_available) {
2267 for (idx = 0; idx < mi_width; ++idx) {
2268 mi = xd->mi[idx - cm->mi_stride];
2269 bs = mi ? mi->mbmi.sb_type : bsize;
2270 min_size = MIN(min_size, bs);
2271 max_size = MAX(max_size, bs);
2272 }
2273 }
2274
2275 if (min_size == max_size) {
2276 min_size = min_partition_size[min_size];
2277 max_size = max_partition_size[max_size];
2278 }
2279
2280 *min_bs = min_size;
2281 *max_bs = max_size;
2282 }
2283
store_pred_mv(MACROBLOCK * x,PICK_MODE_CONTEXT * ctx)2284 static INLINE void store_pred_mv(MACROBLOCK *x, PICK_MODE_CONTEXT *ctx) {
2285 memcpy(ctx->pred_mv, x->pred_mv, sizeof(x->pred_mv));
2286 }
2287
load_pred_mv(MACROBLOCK * x,PICK_MODE_CONTEXT * ctx)2288 static INLINE void load_pred_mv(MACROBLOCK *x, PICK_MODE_CONTEXT *ctx) {
2289 memcpy(x->pred_mv, ctx->pred_mv, sizeof(x->pred_mv));
2290 }
2291
2292 #if CONFIG_FP_MB_STATS
2293 const int num_16x16_blocks_wide_lookup[BLOCK_SIZES] =
2294 {1, 1, 1, 1, 1, 1, 1, 1, 2, 2, 2, 4, 4};
2295 const int num_16x16_blocks_high_lookup[BLOCK_SIZES] =
2296 {1, 1, 1, 1, 1, 1, 1, 2, 1, 2, 4, 2, 4};
2297 const int qindex_skip_threshold_lookup[BLOCK_SIZES] =
2298 {0, 10, 10, 30, 40, 40, 60, 80, 80, 90, 100, 100, 120};
2299 const int qindex_split_threshold_lookup[BLOCK_SIZES] =
2300 {0, 3, 3, 7, 15, 15, 30, 40, 40, 60, 80, 80, 120};
2301 const int complexity_16x16_blocks_threshold[BLOCK_SIZES] =
2302 {1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 4, 4, 6};
2303
2304 typedef enum {
2305 MV_ZERO = 0,
2306 MV_LEFT = 1,
2307 MV_UP = 2,
2308 MV_RIGHT = 3,
2309 MV_DOWN = 4,
2310 MV_INVALID
2311 } MOTION_DIRECTION;
2312
get_motion_direction_fp(uint8_t fp_byte)2313 static INLINE MOTION_DIRECTION get_motion_direction_fp(uint8_t fp_byte) {
2314 if (fp_byte & FPMB_MOTION_ZERO_MASK) {
2315 return MV_ZERO;
2316 } else if (fp_byte & FPMB_MOTION_LEFT_MASK) {
2317 return MV_LEFT;
2318 } else if (fp_byte & FPMB_MOTION_RIGHT_MASK) {
2319 return MV_RIGHT;
2320 } else if (fp_byte & FPMB_MOTION_UP_MASK) {
2321 return MV_UP;
2322 } else {
2323 return MV_DOWN;
2324 }
2325 }
2326
get_motion_inconsistency(MOTION_DIRECTION this_mv,MOTION_DIRECTION that_mv)2327 static INLINE int get_motion_inconsistency(MOTION_DIRECTION this_mv,
2328 MOTION_DIRECTION that_mv) {
2329 if (this_mv == that_mv) {
2330 return 0;
2331 } else {
2332 return abs(this_mv - that_mv) == 2 ? 2 : 1;
2333 }
2334 }
2335 #endif
2336
2337 // TODO(jingning,jimbankoski,rbultje): properly skip partition types that are
2338 // unlikely to be selected depending on previous rate-distortion optimization
2339 // results, for encoding speed-up.
rd_pick_partition(VP9_COMP * cpi,ThreadData * td,TileDataEnc * tile_data,TOKENEXTRA ** tp,int mi_row,int mi_col,BLOCK_SIZE bsize,RD_COST * rd_cost,int64_t best_rd,PC_TREE * pc_tree)2340 static void rd_pick_partition(VP9_COMP *cpi, ThreadData *td,
2341 TileDataEnc *tile_data,
2342 TOKENEXTRA **tp, int mi_row, int mi_col,
2343 BLOCK_SIZE bsize, RD_COST *rd_cost,
2344 int64_t best_rd, PC_TREE *pc_tree) {
2345 VP9_COMMON *const cm = &cpi->common;
2346 TileInfo *const tile_info = &tile_data->tile_info;
2347 MACROBLOCK *const x = &td->mb;
2348 MACROBLOCKD *const xd = &x->e_mbd;
2349 const int mi_step = num_8x8_blocks_wide_lookup[bsize] / 2;
2350 ENTROPY_CONTEXT l[16 * MAX_MB_PLANE], a[16 * MAX_MB_PLANE];
2351 PARTITION_CONTEXT sl[8], sa[8];
2352 TOKENEXTRA *tp_orig = *tp;
2353 PICK_MODE_CONTEXT *ctx = &pc_tree->none;
2354 int i, pl;
2355 BLOCK_SIZE subsize;
2356 RD_COST this_rdc, sum_rdc, best_rdc;
2357 int do_split = bsize >= BLOCK_8X8;
2358 int do_rect = 1;
2359
2360 // Override skipping rectangular partition operations for edge blocks
2361 const int force_horz_split = (mi_row + mi_step >= cm->mi_rows);
2362 const int force_vert_split = (mi_col + mi_step >= cm->mi_cols);
2363 const int xss = x->e_mbd.plane[1].subsampling_x;
2364 const int yss = x->e_mbd.plane[1].subsampling_y;
2365
2366 BLOCK_SIZE min_size = x->min_partition_size;
2367 BLOCK_SIZE max_size = x->max_partition_size;
2368
2369 #if CONFIG_FP_MB_STATS
2370 unsigned int src_diff_var = UINT_MAX;
2371 int none_complexity = 0;
2372 #endif
2373
2374 int partition_none_allowed = !force_horz_split && !force_vert_split;
2375 int partition_horz_allowed = !force_vert_split && yss <= xss &&
2376 bsize >= BLOCK_8X8;
2377 int partition_vert_allowed = !force_horz_split && xss <= yss &&
2378 bsize >= BLOCK_8X8;
2379 (void) *tp_orig;
2380
2381 assert(num_8x8_blocks_wide_lookup[bsize] ==
2382 num_8x8_blocks_high_lookup[bsize]);
2383
2384 vp9_rd_cost_init(&this_rdc);
2385 vp9_rd_cost_init(&sum_rdc);
2386 vp9_rd_cost_reset(&best_rdc);
2387 best_rdc.rdcost = best_rd;
2388
2389 set_offsets(cpi, tile_info, x, mi_row, mi_col, bsize);
2390
2391 if (bsize == BLOCK_16X16 && cpi->oxcf.aq_mode)
2392 x->mb_energy = vp9_block_energy(cpi, x, bsize);
2393
2394 if (cpi->sf.cb_partition_search && bsize == BLOCK_16X16) {
2395 int cb_partition_search_ctrl = ((pc_tree->index == 0 || pc_tree->index == 3)
2396 + get_chessboard_index(cm->current_video_frame)) & 0x1;
2397
2398 if (cb_partition_search_ctrl && bsize > min_size && bsize < max_size)
2399 set_partition_range(cm, xd, mi_row, mi_col, bsize, &min_size, &max_size);
2400 }
2401
2402 // Determine partition types in search according to the speed features.
2403 // The threshold set here has to be of square block size.
2404 if (cpi->sf.auto_min_max_partition_size) {
2405 partition_none_allowed &= (bsize <= max_size && bsize >= min_size);
2406 partition_horz_allowed &= ((bsize <= max_size && bsize > min_size) ||
2407 force_horz_split);
2408 partition_vert_allowed &= ((bsize <= max_size && bsize > min_size) ||
2409 force_vert_split);
2410 do_split &= bsize > min_size;
2411 }
2412 if (cpi->sf.use_square_partition_only) {
2413 partition_horz_allowed &= force_horz_split;
2414 partition_vert_allowed &= force_vert_split;
2415 }
2416
2417 save_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
2418
2419 #if CONFIG_FP_MB_STATS
2420 if (cpi->use_fp_mb_stats) {
2421 set_offsets(cpi, tile_info, x, mi_row, mi_col, bsize);
2422 src_diff_var = get_sby_perpixel_diff_variance(cpi, &x->plane[0].src,
2423 mi_row, mi_col, bsize);
2424 }
2425 #endif
2426
2427 #if CONFIG_FP_MB_STATS
2428 // Decide whether we shall split directly and skip searching NONE by using
2429 // the first pass block statistics
2430 if (cpi->use_fp_mb_stats && bsize >= BLOCK_32X32 && do_split &&
2431 partition_none_allowed && src_diff_var > 4 &&
2432 cm->base_qindex < qindex_split_threshold_lookup[bsize]) {
2433 int mb_row = mi_row >> 1;
2434 int mb_col = mi_col >> 1;
2435 int mb_row_end =
2436 MIN(mb_row + num_16x16_blocks_high_lookup[bsize], cm->mb_rows);
2437 int mb_col_end =
2438 MIN(mb_col + num_16x16_blocks_wide_lookup[bsize], cm->mb_cols);
2439 int r, c;
2440
2441 // compute a complexity measure, basically measure inconsistency of motion
2442 // vectors obtained from the first pass in the current block
2443 for (r = mb_row; r < mb_row_end ; r++) {
2444 for (c = mb_col; c < mb_col_end; c++) {
2445 const int mb_index = r * cm->mb_cols + c;
2446
2447 MOTION_DIRECTION this_mv;
2448 MOTION_DIRECTION right_mv;
2449 MOTION_DIRECTION bottom_mv;
2450
2451 this_mv =
2452 get_motion_direction_fp(cpi->twopass.this_frame_mb_stats[mb_index]);
2453
2454 // to its right
2455 if (c != mb_col_end - 1) {
2456 right_mv = get_motion_direction_fp(
2457 cpi->twopass.this_frame_mb_stats[mb_index + 1]);
2458 none_complexity += get_motion_inconsistency(this_mv, right_mv);
2459 }
2460
2461 // to its bottom
2462 if (r != mb_row_end - 1) {
2463 bottom_mv = get_motion_direction_fp(
2464 cpi->twopass.this_frame_mb_stats[mb_index + cm->mb_cols]);
2465 none_complexity += get_motion_inconsistency(this_mv, bottom_mv);
2466 }
2467
2468 // do not count its left and top neighbors to avoid double counting
2469 }
2470 }
2471
2472 if (none_complexity > complexity_16x16_blocks_threshold[bsize]) {
2473 partition_none_allowed = 0;
2474 }
2475 }
2476 #endif
2477
2478 // PARTITION_NONE
2479 if (partition_none_allowed) {
2480 rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col,
2481 &this_rdc, bsize, ctx, best_rdc.rdcost);
2482 if (this_rdc.rate != INT_MAX) {
2483 if (bsize >= BLOCK_8X8) {
2484 pl = partition_plane_context(xd, mi_row, mi_col, bsize);
2485 this_rdc.rate += cpi->partition_cost[pl][PARTITION_NONE];
2486 this_rdc.rdcost = RDCOST(x->rdmult, x->rddiv,
2487 this_rdc.rate, this_rdc.dist);
2488 }
2489
2490 if (this_rdc.rdcost < best_rdc.rdcost) {
2491 int64_t dist_breakout_thr = cpi->sf.partition_search_breakout_dist_thr;
2492 int rate_breakout_thr = cpi->sf.partition_search_breakout_rate_thr;
2493
2494 best_rdc = this_rdc;
2495 if (bsize >= BLOCK_8X8)
2496 pc_tree->partitioning = PARTITION_NONE;
2497
2498 // Adjust dist breakout threshold according to the partition size.
2499 dist_breakout_thr >>= 8 - (b_width_log2_lookup[bsize] +
2500 b_height_log2_lookup[bsize]);
2501
2502 rate_breakout_thr *= num_pels_log2_lookup[bsize];
2503
2504 // If all y, u, v transform blocks in this partition are skippable, and
2505 // the dist & rate are within the thresholds, the partition search is
2506 // terminated for current branch of the partition search tree.
2507 // The dist & rate thresholds are set to 0 at speed 0 to disable the
2508 // early termination at that speed.
2509 if (!x->e_mbd.lossless &&
2510 (ctx->skippable && best_rdc.dist < dist_breakout_thr &&
2511 best_rdc.rate < rate_breakout_thr)) {
2512 do_split = 0;
2513 do_rect = 0;
2514 }
2515
2516 #if CONFIG_FP_MB_STATS
2517 // Check if every 16x16 first pass block statistics has zero
2518 // motion and the corresponding first pass residue is small enough.
2519 // If that is the case, check the difference variance between the
2520 // current frame and the last frame. If the variance is small enough,
2521 // stop further splitting in RD optimization
2522 if (cpi->use_fp_mb_stats && do_split != 0 &&
2523 cm->base_qindex > qindex_skip_threshold_lookup[bsize]) {
2524 int mb_row = mi_row >> 1;
2525 int mb_col = mi_col >> 1;
2526 int mb_row_end =
2527 MIN(mb_row + num_16x16_blocks_high_lookup[bsize], cm->mb_rows);
2528 int mb_col_end =
2529 MIN(mb_col + num_16x16_blocks_wide_lookup[bsize], cm->mb_cols);
2530 int r, c;
2531
2532 int skip = 1;
2533 for (r = mb_row; r < mb_row_end; r++) {
2534 for (c = mb_col; c < mb_col_end; c++) {
2535 const int mb_index = r * cm->mb_cols + c;
2536 if (!(cpi->twopass.this_frame_mb_stats[mb_index] &
2537 FPMB_MOTION_ZERO_MASK) ||
2538 !(cpi->twopass.this_frame_mb_stats[mb_index] &
2539 FPMB_ERROR_SMALL_MASK)) {
2540 skip = 0;
2541 break;
2542 }
2543 }
2544 if (skip == 0) {
2545 break;
2546 }
2547 }
2548 if (skip) {
2549 if (src_diff_var == UINT_MAX) {
2550 set_offsets(cpi, tile_info, x, mi_row, mi_col, bsize);
2551 src_diff_var = get_sby_perpixel_diff_variance(
2552 cpi, &x->plane[0].src, mi_row, mi_col, bsize);
2553 }
2554 if (src_diff_var < 8) {
2555 do_split = 0;
2556 do_rect = 0;
2557 }
2558 }
2559 }
2560 #endif
2561 }
2562 }
2563 restore_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
2564 }
2565
2566 // store estimated motion vector
2567 if (cpi->sf.adaptive_motion_search)
2568 store_pred_mv(x, ctx);
2569
2570 // PARTITION_SPLIT
2571 // TODO(jingning): use the motion vectors given by the above search as
2572 // the starting point of motion search in the following partition type check.
2573 if (do_split) {
2574 subsize = get_subsize(bsize, PARTITION_SPLIT);
2575 if (bsize == BLOCK_8X8) {
2576 i = 4;
2577 if (cpi->sf.adaptive_pred_interp_filter && partition_none_allowed)
2578 pc_tree->leaf_split[0]->pred_interp_filter =
2579 ctx->mic.mbmi.interp_filter;
2580 rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &sum_rdc, subsize,
2581 pc_tree->leaf_split[0], best_rdc.rdcost);
2582 if (sum_rdc.rate == INT_MAX)
2583 sum_rdc.rdcost = INT64_MAX;
2584 } else {
2585 for (i = 0; i < 4 && sum_rdc.rdcost < best_rdc.rdcost; ++i) {
2586 const int x_idx = (i & 1) * mi_step;
2587 const int y_idx = (i >> 1) * mi_step;
2588
2589 if (mi_row + y_idx >= cm->mi_rows || mi_col + x_idx >= cm->mi_cols)
2590 continue;
2591
2592 if (cpi->sf.adaptive_motion_search)
2593 load_pred_mv(x, ctx);
2594
2595 pc_tree->split[i]->index = i;
2596 rd_pick_partition(cpi, td, tile_data, tp,
2597 mi_row + y_idx, mi_col + x_idx,
2598 subsize, &this_rdc,
2599 best_rdc.rdcost - sum_rdc.rdcost, pc_tree->split[i]);
2600
2601 if (this_rdc.rate == INT_MAX) {
2602 sum_rdc.rdcost = INT64_MAX;
2603 break;
2604 } else {
2605 sum_rdc.rate += this_rdc.rate;
2606 sum_rdc.dist += this_rdc.dist;
2607 sum_rdc.rdcost += this_rdc.rdcost;
2608 }
2609 }
2610 }
2611
2612 if (sum_rdc.rdcost < best_rdc.rdcost && i == 4) {
2613 pl = partition_plane_context(xd, mi_row, mi_col, bsize);
2614 sum_rdc.rate += cpi->partition_cost[pl][PARTITION_SPLIT];
2615 sum_rdc.rdcost = RDCOST(x->rdmult, x->rddiv,
2616 sum_rdc.rate, sum_rdc.dist);
2617
2618 if (sum_rdc.rdcost < best_rdc.rdcost) {
2619 best_rdc = sum_rdc;
2620 pc_tree->partitioning = PARTITION_SPLIT;
2621 }
2622 } else {
2623 // skip rectangular partition test when larger block size
2624 // gives better rd cost
2625 if (cpi->sf.less_rectangular_check)
2626 do_rect &= !partition_none_allowed;
2627 }
2628 restore_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
2629 }
2630
2631 // PARTITION_HORZ
2632 if (partition_horz_allowed &&
2633 (do_rect || vp9_active_h_edge(cpi, mi_row, mi_step))) {
2634 subsize = get_subsize(bsize, PARTITION_HORZ);
2635 if (cpi->sf.adaptive_motion_search)
2636 load_pred_mv(x, ctx);
2637 if (cpi->sf.adaptive_pred_interp_filter && bsize == BLOCK_8X8 &&
2638 partition_none_allowed)
2639 pc_tree->horizontal[0].pred_interp_filter =
2640 ctx->mic.mbmi.interp_filter;
2641 rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &sum_rdc, subsize,
2642 &pc_tree->horizontal[0], best_rdc.rdcost);
2643
2644 if (sum_rdc.rdcost < best_rdc.rdcost && mi_row + mi_step < cm->mi_rows &&
2645 bsize > BLOCK_8X8) {
2646 PICK_MODE_CONTEXT *ctx = &pc_tree->horizontal[0];
2647 update_state(cpi, td, ctx, mi_row, mi_col, subsize, 0);
2648 encode_superblock(cpi, td, tp, 0, mi_row, mi_col, subsize, ctx);
2649
2650 if (cpi->sf.adaptive_motion_search)
2651 load_pred_mv(x, ctx);
2652 if (cpi->sf.adaptive_pred_interp_filter && bsize == BLOCK_8X8 &&
2653 partition_none_allowed)
2654 pc_tree->horizontal[1].pred_interp_filter =
2655 ctx->mic.mbmi.interp_filter;
2656 rd_pick_sb_modes(cpi, tile_data, x, mi_row + mi_step, mi_col,
2657 &this_rdc, subsize, &pc_tree->horizontal[1],
2658 best_rdc.rdcost - sum_rdc.rdcost);
2659 if (this_rdc.rate == INT_MAX) {
2660 sum_rdc.rdcost = INT64_MAX;
2661 } else {
2662 sum_rdc.rate += this_rdc.rate;
2663 sum_rdc.dist += this_rdc.dist;
2664 sum_rdc.rdcost += this_rdc.rdcost;
2665 }
2666 }
2667
2668 if (sum_rdc.rdcost < best_rdc.rdcost) {
2669 pl = partition_plane_context(xd, mi_row, mi_col, bsize);
2670 sum_rdc.rate += cpi->partition_cost[pl][PARTITION_HORZ];
2671 sum_rdc.rdcost = RDCOST(x->rdmult, x->rddiv, sum_rdc.rate, sum_rdc.dist);
2672 if (sum_rdc.rdcost < best_rdc.rdcost) {
2673 best_rdc = sum_rdc;
2674 pc_tree->partitioning = PARTITION_HORZ;
2675 }
2676 }
2677 restore_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
2678 }
2679 // PARTITION_VERT
2680 if (partition_vert_allowed &&
2681 (do_rect || vp9_active_v_edge(cpi, mi_col, mi_step))) {
2682 subsize = get_subsize(bsize, PARTITION_VERT);
2683
2684 if (cpi->sf.adaptive_motion_search)
2685 load_pred_mv(x, ctx);
2686 if (cpi->sf.adaptive_pred_interp_filter && bsize == BLOCK_8X8 &&
2687 partition_none_allowed)
2688 pc_tree->vertical[0].pred_interp_filter =
2689 ctx->mic.mbmi.interp_filter;
2690 rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &sum_rdc, subsize,
2691 &pc_tree->vertical[0], best_rdc.rdcost);
2692 if (sum_rdc.rdcost < best_rdc.rdcost && mi_col + mi_step < cm->mi_cols &&
2693 bsize > BLOCK_8X8) {
2694 update_state(cpi, td, &pc_tree->vertical[0], mi_row, mi_col, subsize, 0);
2695 encode_superblock(cpi, td, tp, 0, mi_row, mi_col, subsize,
2696 &pc_tree->vertical[0]);
2697
2698 if (cpi->sf.adaptive_motion_search)
2699 load_pred_mv(x, ctx);
2700 if (cpi->sf.adaptive_pred_interp_filter && bsize == BLOCK_8X8 &&
2701 partition_none_allowed)
2702 pc_tree->vertical[1].pred_interp_filter =
2703 ctx->mic.mbmi.interp_filter;
2704 rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col + mi_step,
2705 &this_rdc, subsize,
2706 &pc_tree->vertical[1], best_rdc.rdcost - sum_rdc.rdcost);
2707 if (this_rdc.rate == INT_MAX) {
2708 sum_rdc.rdcost = INT64_MAX;
2709 } else {
2710 sum_rdc.rate += this_rdc.rate;
2711 sum_rdc.dist += this_rdc.dist;
2712 sum_rdc.rdcost += this_rdc.rdcost;
2713 }
2714 }
2715
2716 if (sum_rdc.rdcost < best_rdc.rdcost) {
2717 pl = partition_plane_context(xd, mi_row, mi_col, bsize);
2718 sum_rdc.rate += cpi->partition_cost[pl][PARTITION_VERT];
2719 sum_rdc.rdcost = RDCOST(x->rdmult, x->rddiv,
2720 sum_rdc.rate, sum_rdc.dist);
2721 if (sum_rdc.rdcost < best_rdc.rdcost) {
2722 best_rdc = sum_rdc;
2723 pc_tree->partitioning = PARTITION_VERT;
2724 }
2725 }
2726 restore_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
2727 }
2728
2729 // TODO(jbb): This code added so that we avoid static analysis
2730 // warning related to the fact that best_rd isn't used after this
2731 // point. This code should be refactored so that the duplicate
2732 // checks occur in some sub function and thus are used...
2733 (void) best_rd;
2734 *rd_cost = best_rdc;
2735
2736
2737 if (best_rdc.rate < INT_MAX && best_rdc.dist < INT64_MAX &&
2738 pc_tree->index != 3) {
2739 int output_enabled = (bsize == BLOCK_64X64);
2740 encode_sb(cpi, td, tile_info, tp, mi_row, mi_col, output_enabled,
2741 bsize, pc_tree);
2742 }
2743
2744 if (bsize == BLOCK_64X64) {
2745 assert(tp_orig < *tp);
2746 assert(best_rdc.rate < INT_MAX);
2747 assert(best_rdc.dist < INT64_MAX);
2748 } else {
2749 assert(tp_orig == *tp);
2750 }
2751 }
2752
encode_rd_sb_row(VP9_COMP * cpi,ThreadData * td,TileDataEnc * tile_data,int mi_row,TOKENEXTRA ** tp)2753 static void encode_rd_sb_row(VP9_COMP *cpi,
2754 ThreadData *td,
2755 TileDataEnc *tile_data,
2756 int mi_row,
2757 TOKENEXTRA **tp) {
2758 VP9_COMMON *const cm = &cpi->common;
2759 TileInfo *const tile_info = &tile_data->tile_info;
2760 MACROBLOCK *const x = &td->mb;
2761 MACROBLOCKD *const xd = &x->e_mbd;
2762 SPEED_FEATURES *const sf = &cpi->sf;
2763 int mi_col;
2764
2765 // Initialize the left context for the new SB row
2766 memset(&xd->left_context, 0, sizeof(xd->left_context));
2767 memset(xd->left_seg_context, 0, sizeof(xd->left_seg_context));
2768
2769 // Code each SB in the row
2770 for (mi_col = tile_info->mi_col_start; mi_col < tile_info->mi_col_end;
2771 mi_col += MI_BLOCK_SIZE) {
2772 const struct segmentation *const seg = &cm->seg;
2773 int dummy_rate;
2774 int64_t dummy_dist;
2775 RD_COST dummy_rdc;
2776 int i;
2777 int seg_skip = 0;
2778
2779 const int idx_str = cm->mi_stride * mi_row + mi_col;
2780 MODE_INFO **mi = cm->mi_grid_visible + idx_str;
2781
2782 if (sf->adaptive_pred_interp_filter) {
2783 for (i = 0; i < 64; ++i)
2784 td->leaf_tree[i].pred_interp_filter = SWITCHABLE;
2785
2786 for (i = 0; i < 64; ++i) {
2787 td->pc_tree[i].vertical[0].pred_interp_filter = SWITCHABLE;
2788 td->pc_tree[i].vertical[1].pred_interp_filter = SWITCHABLE;
2789 td->pc_tree[i].horizontal[0].pred_interp_filter = SWITCHABLE;
2790 td->pc_tree[i].horizontal[1].pred_interp_filter = SWITCHABLE;
2791 }
2792 }
2793
2794 vp9_zero(x->pred_mv);
2795 td->pc_root->index = 0;
2796
2797 if (seg->enabled) {
2798 const uint8_t *const map = seg->update_map ? cpi->segmentation_map
2799 : cm->last_frame_seg_map;
2800 int segment_id = get_segment_id(cm, map, BLOCK_64X64, mi_row, mi_col);
2801 seg_skip = segfeature_active(seg, segment_id, SEG_LVL_SKIP);
2802 }
2803
2804 x->source_variance = UINT_MAX;
2805 if (sf->partition_search_type == FIXED_PARTITION || seg_skip) {
2806 const BLOCK_SIZE bsize =
2807 seg_skip ? BLOCK_64X64 : sf->always_this_block_size;
2808 set_offsets(cpi, tile_info, x, mi_row, mi_col, BLOCK_64X64);
2809 set_fixed_partitioning(cpi, tile_info, mi, mi_row, mi_col, bsize);
2810 rd_use_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col,
2811 BLOCK_64X64, &dummy_rate, &dummy_dist, 1, td->pc_root);
2812 } else if (cpi->partition_search_skippable_frame) {
2813 BLOCK_SIZE bsize;
2814 set_offsets(cpi, tile_info, x, mi_row, mi_col, BLOCK_64X64);
2815 bsize = get_rd_var_based_fixed_partition(cpi, x, mi_row, mi_col);
2816 set_fixed_partitioning(cpi, tile_info, mi, mi_row, mi_col, bsize);
2817 rd_use_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col,
2818 BLOCK_64X64, &dummy_rate, &dummy_dist, 1, td->pc_root);
2819 } else if (sf->partition_search_type == VAR_BASED_PARTITION &&
2820 cm->frame_type != KEY_FRAME) {
2821 choose_partitioning(cpi, tile_info, x, mi_row, mi_col);
2822 rd_use_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col,
2823 BLOCK_64X64, &dummy_rate, &dummy_dist, 1, td->pc_root);
2824 } else {
2825 // If required set upper and lower partition size limits
2826 if (sf->auto_min_max_partition_size) {
2827 set_offsets(cpi, tile_info, x, mi_row, mi_col, BLOCK_64X64);
2828 rd_auto_partition_range(cpi, tile_info, xd, mi_row, mi_col,
2829 &x->min_partition_size,
2830 &x->max_partition_size);
2831 }
2832 rd_pick_partition(cpi, td, tile_data, tp, mi_row, mi_col, BLOCK_64X64,
2833 &dummy_rdc, INT64_MAX, td->pc_root);
2834 }
2835 }
2836 }
2837
init_encode_frame_mb_context(VP9_COMP * cpi)2838 static void init_encode_frame_mb_context(VP9_COMP *cpi) {
2839 MACROBLOCK *const x = &cpi->td.mb;
2840 VP9_COMMON *const cm = &cpi->common;
2841 MACROBLOCKD *const xd = &x->e_mbd;
2842 const int aligned_mi_cols = mi_cols_aligned_to_sb(cm->mi_cols);
2843
2844 // Copy data over into macro block data structures.
2845 vp9_setup_src_planes(x, cpi->Source, 0, 0);
2846
2847 vp9_setup_block_planes(&x->e_mbd, cm->subsampling_x, cm->subsampling_y);
2848
2849 // Note: this memset assumes above_context[0], [1] and [2]
2850 // are allocated as part of the same buffer.
2851 memset(xd->above_context[0], 0,
2852 sizeof(*xd->above_context[0]) *
2853 2 * aligned_mi_cols * MAX_MB_PLANE);
2854 memset(xd->above_seg_context, 0,
2855 sizeof(*xd->above_seg_context) * aligned_mi_cols);
2856 }
2857
check_dual_ref_flags(VP9_COMP * cpi)2858 static int check_dual_ref_flags(VP9_COMP *cpi) {
2859 const int ref_flags = cpi->ref_frame_flags;
2860
2861 if (segfeature_active(&cpi->common.seg, 1, SEG_LVL_REF_FRAME)) {
2862 return 0;
2863 } else {
2864 return (!!(ref_flags & VP9_GOLD_FLAG) + !!(ref_flags & VP9_LAST_FLAG)
2865 + !!(ref_flags & VP9_ALT_FLAG)) >= 2;
2866 }
2867 }
2868
reset_skip_tx_size(VP9_COMMON * cm,TX_SIZE max_tx_size)2869 static void reset_skip_tx_size(VP9_COMMON *cm, TX_SIZE max_tx_size) {
2870 int mi_row, mi_col;
2871 const int mis = cm->mi_stride;
2872 MODE_INFO **mi_ptr = cm->mi_grid_visible;
2873
2874 for (mi_row = 0; mi_row < cm->mi_rows; ++mi_row, mi_ptr += mis) {
2875 for (mi_col = 0; mi_col < cm->mi_cols; ++mi_col) {
2876 if (mi_ptr[mi_col]->mbmi.tx_size > max_tx_size)
2877 mi_ptr[mi_col]->mbmi.tx_size = max_tx_size;
2878 }
2879 }
2880 }
2881
get_frame_type(const VP9_COMP * cpi)2882 static MV_REFERENCE_FRAME get_frame_type(const VP9_COMP *cpi) {
2883 if (frame_is_intra_only(&cpi->common))
2884 return INTRA_FRAME;
2885 else if (cpi->rc.is_src_frame_alt_ref && cpi->refresh_golden_frame)
2886 return ALTREF_FRAME;
2887 else if (cpi->refresh_golden_frame || cpi->refresh_alt_ref_frame)
2888 return GOLDEN_FRAME;
2889 else
2890 return LAST_FRAME;
2891 }
2892
select_tx_mode(const VP9_COMP * cpi,MACROBLOCKD * const xd)2893 static TX_MODE select_tx_mode(const VP9_COMP *cpi, MACROBLOCKD *const xd) {
2894 if (xd->lossless)
2895 return ONLY_4X4;
2896 if (cpi->common.frame_type == KEY_FRAME &&
2897 cpi->sf.use_nonrd_pick_mode)
2898 return ALLOW_16X16;
2899 if (cpi->sf.tx_size_search_method == USE_LARGESTALL)
2900 return ALLOW_32X32;
2901 else if (cpi->sf.tx_size_search_method == USE_FULL_RD||
2902 cpi->sf.tx_size_search_method == USE_TX_8X8)
2903 return TX_MODE_SELECT;
2904 else
2905 return cpi->common.tx_mode;
2906 }
2907
hybrid_intra_mode_search(VP9_COMP * cpi,MACROBLOCK * const x,RD_COST * rd_cost,BLOCK_SIZE bsize,PICK_MODE_CONTEXT * ctx)2908 static void hybrid_intra_mode_search(VP9_COMP *cpi, MACROBLOCK *const x,
2909 RD_COST *rd_cost, BLOCK_SIZE bsize,
2910 PICK_MODE_CONTEXT *ctx) {
2911 if (bsize < BLOCK_16X16)
2912 vp9_rd_pick_intra_mode_sb(cpi, x, rd_cost, bsize, ctx, INT64_MAX);
2913 else
2914 vp9_pick_intra_mode(cpi, x, rd_cost, bsize, ctx);
2915 }
2916
nonrd_pick_sb_modes(VP9_COMP * cpi,TileDataEnc * tile_data,MACROBLOCK * const x,int mi_row,int mi_col,RD_COST * rd_cost,BLOCK_SIZE bsize,PICK_MODE_CONTEXT * ctx)2917 static void nonrd_pick_sb_modes(VP9_COMP *cpi,
2918 TileDataEnc *tile_data, MACROBLOCK *const x,
2919 int mi_row, int mi_col, RD_COST *rd_cost,
2920 BLOCK_SIZE bsize, PICK_MODE_CONTEXT *ctx) {
2921 VP9_COMMON *const cm = &cpi->common;
2922 TileInfo *const tile_info = &tile_data->tile_info;
2923 MACROBLOCKD *const xd = &x->e_mbd;
2924 MB_MODE_INFO *mbmi;
2925 set_offsets(cpi, tile_info, x, mi_row, mi_col, bsize);
2926 mbmi = &xd->mi[0]->mbmi;
2927 mbmi->sb_type = bsize;
2928
2929 if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ && cm->seg.enabled)
2930 if (cyclic_refresh_segment_id_boosted(mbmi->segment_id))
2931 x->rdmult = vp9_cyclic_refresh_get_rdmult(cpi->cyclic_refresh);
2932
2933 if (cm->frame_type == KEY_FRAME)
2934 hybrid_intra_mode_search(cpi, x, rd_cost, bsize, ctx);
2935 else if (segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_SKIP))
2936 set_mode_info_seg_skip(x, cm->tx_mode, rd_cost, bsize);
2937 else if (bsize >= BLOCK_8X8)
2938 vp9_pick_inter_mode(cpi, x, tile_data, mi_row, mi_col,
2939 rd_cost, bsize, ctx);
2940 else
2941 vp9_pick_inter_mode_sub8x8(cpi, x, mi_row, mi_col,
2942 rd_cost, bsize, ctx);
2943
2944 duplicate_mode_info_in_sb(cm, xd, mi_row, mi_col, bsize);
2945
2946 if (rd_cost->rate == INT_MAX)
2947 vp9_rd_cost_reset(rd_cost);
2948
2949 ctx->rate = rd_cost->rate;
2950 ctx->dist = rd_cost->dist;
2951 }
2952
fill_mode_info_sb(VP9_COMMON * cm,MACROBLOCK * x,int mi_row,int mi_col,BLOCK_SIZE bsize,PC_TREE * pc_tree)2953 static void fill_mode_info_sb(VP9_COMMON *cm, MACROBLOCK *x,
2954 int mi_row, int mi_col,
2955 BLOCK_SIZE bsize,
2956 PC_TREE *pc_tree) {
2957 MACROBLOCKD *xd = &x->e_mbd;
2958 int bsl = b_width_log2_lookup[bsize], hbs = (1 << bsl) / 4;
2959 PARTITION_TYPE partition = pc_tree->partitioning;
2960 BLOCK_SIZE subsize = get_subsize(bsize, partition);
2961
2962 assert(bsize >= BLOCK_8X8);
2963
2964 if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols)
2965 return;
2966
2967 switch (partition) {
2968 case PARTITION_NONE:
2969 set_mode_info_offsets(cm, x, xd, mi_row, mi_col);
2970 *(xd->mi[0]) = pc_tree->none.mic;
2971 *(x->mbmi_ext) = pc_tree->none.mbmi_ext;
2972 duplicate_mode_info_in_sb(cm, xd, mi_row, mi_col, bsize);
2973 break;
2974 case PARTITION_VERT:
2975 set_mode_info_offsets(cm, x, xd, mi_row, mi_col);
2976 *(xd->mi[0]) = pc_tree->vertical[0].mic;
2977 *(x->mbmi_ext) = pc_tree->vertical[0].mbmi_ext;
2978 duplicate_mode_info_in_sb(cm, xd, mi_row, mi_col, subsize);
2979
2980 if (mi_col + hbs < cm->mi_cols) {
2981 set_mode_info_offsets(cm, x, xd, mi_row, mi_col + hbs);
2982 *(xd->mi[0]) = pc_tree->vertical[1].mic;
2983 *(x->mbmi_ext) = pc_tree->vertical[1].mbmi_ext;
2984 duplicate_mode_info_in_sb(cm, xd, mi_row, mi_col + hbs, subsize);
2985 }
2986 break;
2987 case PARTITION_HORZ:
2988 set_mode_info_offsets(cm, x, xd, mi_row, mi_col);
2989 *(xd->mi[0]) = pc_tree->horizontal[0].mic;
2990 *(x->mbmi_ext) = pc_tree->horizontal[0].mbmi_ext;
2991 duplicate_mode_info_in_sb(cm, xd, mi_row, mi_col, subsize);
2992 if (mi_row + hbs < cm->mi_rows) {
2993 set_mode_info_offsets(cm, x, xd, mi_row + hbs, mi_col);
2994 *(xd->mi[0]) = pc_tree->horizontal[1].mic;
2995 *(x->mbmi_ext) = pc_tree->horizontal[1].mbmi_ext;
2996 duplicate_mode_info_in_sb(cm, xd, mi_row + hbs, mi_col, subsize);
2997 }
2998 break;
2999 case PARTITION_SPLIT: {
3000 fill_mode_info_sb(cm, x, mi_row, mi_col, subsize, pc_tree->split[0]);
3001 fill_mode_info_sb(cm, x, mi_row, mi_col + hbs, subsize,
3002 pc_tree->split[1]);
3003 fill_mode_info_sb(cm, x, mi_row + hbs, mi_col, subsize,
3004 pc_tree->split[2]);
3005 fill_mode_info_sb(cm, x, mi_row + hbs, mi_col + hbs, subsize,
3006 pc_tree->split[3]);
3007 break;
3008 }
3009 default:
3010 break;
3011 }
3012 }
3013
3014 // Reset the prediction pixel ready flag recursively.
pred_pixel_ready_reset(PC_TREE * pc_tree,BLOCK_SIZE bsize)3015 static void pred_pixel_ready_reset(PC_TREE *pc_tree, BLOCK_SIZE bsize) {
3016 pc_tree->none.pred_pixel_ready = 0;
3017 pc_tree->horizontal[0].pred_pixel_ready = 0;
3018 pc_tree->horizontal[1].pred_pixel_ready = 0;
3019 pc_tree->vertical[0].pred_pixel_ready = 0;
3020 pc_tree->vertical[1].pred_pixel_ready = 0;
3021
3022 if (bsize > BLOCK_8X8) {
3023 BLOCK_SIZE subsize = get_subsize(bsize, PARTITION_SPLIT);
3024 int i;
3025 for (i = 0; i < 4; ++i)
3026 pred_pixel_ready_reset(pc_tree->split[i], subsize);
3027 }
3028 }
3029
nonrd_pick_partition(VP9_COMP * cpi,ThreadData * td,TileDataEnc * tile_data,TOKENEXTRA ** tp,int mi_row,int mi_col,BLOCK_SIZE bsize,RD_COST * rd_cost,int do_recon,int64_t best_rd,PC_TREE * pc_tree)3030 static void nonrd_pick_partition(VP9_COMP *cpi, ThreadData *td,
3031 TileDataEnc *tile_data,
3032 TOKENEXTRA **tp, int mi_row,
3033 int mi_col, BLOCK_SIZE bsize, RD_COST *rd_cost,
3034 int do_recon, int64_t best_rd,
3035 PC_TREE *pc_tree) {
3036 const SPEED_FEATURES *const sf = &cpi->sf;
3037 VP9_COMMON *const cm = &cpi->common;
3038 TileInfo *const tile_info = &tile_data->tile_info;
3039 MACROBLOCK *const x = &td->mb;
3040 MACROBLOCKD *const xd = &x->e_mbd;
3041 const int ms = num_8x8_blocks_wide_lookup[bsize] / 2;
3042 TOKENEXTRA *tp_orig = *tp;
3043 PICK_MODE_CONTEXT *ctx = &pc_tree->none;
3044 int i;
3045 BLOCK_SIZE subsize = bsize;
3046 RD_COST this_rdc, sum_rdc, best_rdc;
3047 int do_split = bsize >= BLOCK_8X8;
3048 int do_rect = 1;
3049 // Override skipping rectangular partition operations for edge blocks
3050 const int force_horz_split = (mi_row + ms >= cm->mi_rows);
3051 const int force_vert_split = (mi_col + ms >= cm->mi_cols);
3052 const int xss = x->e_mbd.plane[1].subsampling_x;
3053 const int yss = x->e_mbd.plane[1].subsampling_y;
3054
3055 int partition_none_allowed = !force_horz_split && !force_vert_split;
3056 int partition_horz_allowed = !force_vert_split && yss <= xss &&
3057 bsize >= BLOCK_8X8;
3058 int partition_vert_allowed = !force_horz_split && xss <= yss &&
3059 bsize >= BLOCK_8X8;
3060 (void) *tp_orig;
3061
3062 assert(num_8x8_blocks_wide_lookup[bsize] ==
3063 num_8x8_blocks_high_lookup[bsize]);
3064
3065 vp9_rd_cost_init(&sum_rdc);
3066 vp9_rd_cost_reset(&best_rdc);
3067 best_rdc.rdcost = best_rd;
3068
3069 // Determine partition types in search according to the speed features.
3070 // The threshold set here has to be of square block size.
3071 if (sf->auto_min_max_partition_size) {
3072 partition_none_allowed &= (bsize <= x->max_partition_size &&
3073 bsize >= x->min_partition_size);
3074 partition_horz_allowed &= ((bsize <= x->max_partition_size &&
3075 bsize > x->min_partition_size) ||
3076 force_horz_split);
3077 partition_vert_allowed &= ((bsize <= x->max_partition_size &&
3078 bsize > x->min_partition_size) ||
3079 force_vert_split);
3080 do_split &= bsize > x->min_partition_size;
3081 }
3082 if (sf->use_square_partition_only) {
3083 partition_horz_allowed &= force_horz_split;
3084 partition_vert_allowed &= force_vert_split;
3085 }
3086
3087 ctx->pred_pixel_ready = !(partition_vert_allowed ||
3088 partition_horz_allowed ||
3089 do_split);
3090
3091 // PARTITION_NONE
3092 if (partition_none_allowed) {
3093 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col,
3094 &this_rdc, bsize, ctx);
3095 ctx->mic.mbmi = xd->mi[0]->mbmi;
3096 ctx->mbmi_ext = *x->mbmi_ext;
3097 ctx->skip_txfm[0] = x->skip_txfm[0];
3098 ctx->skip = x->skip;
3099
3100 if (this_rdc.rate != INT_MAX) {
3101 int pl = partition_plane_context(xd, mi_row, mi_col, bsize);
3102 this_rdc.rate += cpi->partition_cost[pl][PARTITION_NONE];
3103 this_rdc.rdcost = RDCOST(x->rdmult, x->rddiv,
3104 this_rdc.rate, this_rdc.dist);
3105 if (this_rdc.rdcost < best_rdc.rdcost) {
3106 int64_t dist_breakout_thr = sf->partition_search_breakout_dist_thr;
3107 int64_t rate_breakout_thr = sf->partition_search_breakout_rate_thr;
3108
3109 dist_breakout_thr >>= 8 - (b_width_log2_lookup[bsize] +
3110 b_height_log2_lookup[bsize]);
3111
3112 rate_breakout_thr *= num_pels_log2_lookup[bsize];
3113
3114 best_rdc = this_rdc;
3115 if (bsize >= BLOCK_8X8)
3116 pc_tree->partitioning = PARTITION_NONE;
3117
3118 if (!x->e_mbd.lossless &&
3119 this_rdc.rate < rate_breakout_thr &&
3120 this_rdc.dist < dist_breakout_thr) {
3121 do_split = 0;
3122 do_rect = 0;
3123 }
3124 }
3125 }
3126 }
3127
3128 // store estimated motion vector
3129 store_pred_mv(x, ctx);
3130
3131 // PARTITION_SPLIT
3132 if (do_split) {
3133 int pl = partition_plane_context(xd, mi_row, mi_col, bsize);
3134 sum_rdc.rate += cpi->partition_cost[pl][PARTITION_SPLIT];
3135 sum_rdc.rdcost = RDCOST(x->rdmult, x->rddiv, sum_rdc.rate, sum_rdc.dist);
3136 subsize = get_subsize(bsize, PARTITION_SPLIT);
3137 for (i = 0; i < 4 && sum_rdc.rdcost < best_rdc.rdcost; ++i) {
3138 const int x_idx = (i & 1) * ms;
3139 const int y_idx = (i >> 1) * ms;
3140
3141 if (mi_row + y_idx >= cm->mi_rows || mi_col + x_idx >= cm->mi_cols)
3142 continue;
3143 load_pred_mv(x, ctx);
3144 nonrd_pick_partition(cpi, td, tile_data, tp,
3145 mi_row + y_idx, mi_col + x_idx,
3146 subsize, &this_rdc, 0,
3147 best_rdc.rdcost - sum_rdc.rdcost, pc_tree->split[i]);
3148
3149 if (this_rdc.rate == INT_MAX) {
3150 vp9_rd_cost_reset(&sum_rdc);
3151 } else {
3152 sum_rdc.rate += this_rdc.rate;
3153 sum_rdc.dist += this_rdc.dist;
3154 sum_rdc.rdcost += this_rdc.rdcost;
3155 }
3156 }
3157
3158 if (sum_rdc.rdcost < best_rdc.rdcost) {
3159 best_rdc = sum_rdc;
3160 pc_tree->partitioning = PARTITION_SPLIT;
3161 } else {
3162 // skip rectangular partition test when larger block size
3163 // gives better rd cost
3164 if (sf->less_rectangular_check)
3165 do_rect &= !partition_none_allowed;
3166 }
3167 }
3168
3169 // PARTITION_HORZ
3170 if (partition_horz_allowed && do_rect) {
3171 subsize = get_subsize(bsize, PARTITION_HORZ);
3172 if (sf->adaptive_motion_search)
3173 load_pred_mv(x, ctx);
3174 pc_tree->horizontal[0].pred_pixel_ready = 1;
3175 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &sum_rdc, subsize,
3176 &pc_tree->horizontal[0]);
3177
3178 pc_tree->horizontal[0].mic.mbmi = xd->mi[0]->mbmi;
3179 pc_tree->horizontal[0].mbmi_ext = *x->mbmi_ext;
3180 pc_tree->horizontal[0].skip_txfm[0] = x->skip_txfm[0];
3181 pc_tree->horizontal[0].skip = x->skip;
3182
3183 if (sum_rdc.rdcost < best_rdc.rdcost && mi_row + ms < cm->mi_rows) {
3184 load_pred_mv(x, ctx);
3185 pc_tree->horizontal[1].pred_pixel_ready = 1;
3186 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row + ms, mi_col,
3187 &this_rdc, subsize,
3188 &pc_tree->horizontal[1]);
3189
3190 pc_tree->horizontal[1].mic.mbmi = xd->mi[0]->mbmi;
3191 pc_tree->horizontal[1].mbmi_ext = *x->mbmi_ext;
3192 pc_tree->horizontal[1].skip_txfm[0] = x->skip_txfm[0];
3193 pc_tree->horizontal[1].skip = x->skip;
3194
3195 if (this_rdc.rate == INT_MAX) {
3196 vp9_rd_cost_reset(&sum_rdc);
3197 } else {
3198 int pl = partition_plane_context(xd, mi_row, mi_col, bsize);
3199 this_rdc.rate += cpi->partition_cost[pl][PARTITION_HORZ];
3200 sum_rdc.rate += this_rdc.rate;
3201 sum_rdc.dist += this_rdc.dist;
3202 sum_rdc.rdcost = RDCOST(x->rdmult, x->rddiv,
3203 sum_rdc.rate, sum_rdc.dist);
3204 }
3205 }
3206
3207 if (sum_rdc.rdcost < best_rdc.rdcost) {
3208 best_rdc = sum_rdc;
3209 pc_tree->partitioning = PARTITION_HORZ;
3210 } else {
3211 pred_pixel_ready_reset(pc_tree, bsize);
3212 }
3213 }
3214
3215 // PARTITION_VERT
3216 if (partition_vert_allowed && do_rect) {
3217 subsize = get_subsize(bsize, PARTITION_VERT);
3218 if (sf->adaptive_motion_search)
3219 load_pred_mv(x, ctx);
3220 pc_tree->vertical[0].pred_pixel_ready = 1;
3221 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &sum_rdc, subsize,
3222 &pc_tree->vertical[0]);
3223 pc_tree->vertical[0].mic.mbmi = xd->mi[0]->mbmi;
3224 pc_tree->vertical[0].mbmi_ext = *x->mbmi_ext;
3225 pc_tree->vertical[0].skip_txfm[0] = x->skip_txfm[0];
3226 pc_tree->vertical[0].skip = x->skip;
3227
3228 if (sum_rdc.rdcost < best_rdc.rdcost && mi_col + ms < cm->mi_cols) {
3229 load_pred_mv(x, ctx);
3230 pc_tree->vertical[1].pred_pixel_ready = 1;
3231 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col + ms,
3232 &this_rdc, subsize,
3233 &pc_tree->vertical[1]);
3234 pc_tree->vertical[1].mic.mbmi = xd->mi[0]->mbmi;
3235 pc_tree->vertical[1].mbmi_ext = *x->mbmi_ext;
3236 pc_tree->vertical[1].skip_txfm[0] = x->skip_txfm[0];
3237 pc_tree->vertical[1].skip = x->skip;
3238
3239 if (this_rdc.rate == INT_MAX) {
3240 vp9_rd_cost_reset(&sum_rdc);
3241 } else {
3242 int pl = partition_plane_context(xd, mi_row, mi_col, bsize);
3243 sum_rdc.rate += cpi->partition_cost[pl][PARTITION_VERT];
3244 sum_rdc.rate += this_rdc.rate;
3245 sum_rdc.dist += this_rdc.dist;
3246 sum_rdc.rdcost = RDCOST(x->rdmult, x->rddiv,
3247 sum_rdc.rate, sum_rdc.dist);
3248 }
3249 }
3250
3251 if (sum_rdc.rdcost < best_rdc.rdcost) {
3252 best_rdc = sum_rdc;
3253 pc_tree->partitioning = PARTITION_VERT;
3254 } else {
3255 pred_pixel_ready_reset(pc_tree, bsize);
3256 }
3257 }
3258
3259 *rd_cost = best_rdc;
3260
3261 if (best_rdc.rate == INT_MAX) {
3262 vp9_rd_cost_reset(rd_cost);
3263 return;
3264 }
3265
3266 // update mode info array
3267 fill_mode_info_sb(cm, x, mi_row, mi_col, bsize, pc_tree);
3268
3269 if (best_rdc.rate < INT_MAX && best_rdc.dist < INT64_MAX && do_recon) {
3270 int output_enabled = (bsize == BLOCK_64X64);
3271 encode_sb_rt(cpi, td, tile_info, tp, mi_row, mi_col, output_enabled,
3272 bsize, pc_tree);
3273 }
3274
3275 if (bsize == BLOCK_64X64 && do_recon) {
3276 assert(tp_orig < *tp);
3277 assert(best_rdc.rate < INT_MAX);
3278 assert(best_rdc.dist < INT64_MAX);
3279 } else {
3280 assert(tp_orig == *tp);
3281 }
3282 }
3283
nonrd_select_partition(VP9_COMP * cpi,ThreadData * td,TileDataEnc * tile_data,MODE_INFO ** mi,TOKENEXTRA ** tp,int mi_row,int mi_col,BLOCK_SIZE bsize,int output_enabled,RD_COST * rd_cost,PC_TREE * pc_tree)3284 static void nonrd_select_partition(VP9_COMP *cpi,
3285 ThreadData *td,
3286 TileDataEnc *tile_data,
3287 MODE_INFO **mi,
3288 TOKENEXTRA **tp,
3289 int mi_row, int mi_col,
3290 BLOCK_SIZE bsize, int output_enabled,
3291 RD_COST *rd_cost, PC_TREE *pc_tree) {
3292 VP9_COMMON *const cm = &cpi->common;
3293 TileInfo *const tile_info = &tile_data->tile_info;
3294 MACROBLOCK *const x = &td->mb;
3295 MACROBLOCKD *const xd = &x->e_mbd;
3296 const int bsl = b_width_log2_lookup[bsize], hbs = (1 << bsl) / 4;
3297 const int mis = cm->mi_stride;
3298 PARTITION_TYPE partition;
3299 BLOCK_SIZE subsize;
3300 RD_COST this_rdc;
3301
3302 vp9_rd_cost_reset(&this_rdc);
3303 if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols)
3304 return;
3305
3306 subsize = (bsize >= BLOCK_8X8) ? mi[0]->mbmi.sb_type : BLOCK_4X4;
3307 partition = partition_lookup[bsl][subsize];
3308
3309 if (bsize == BLOCK_32X32 && subsize == BLOCK_32X32) {
3310 x->max_partition_size = BLOCK_32X32;
3311 x->min_partition_size = BLOCK_16X16;
3312 nonrd_pick_partition(cpi, td, tile_data, tp, mi_row, mi_col, bsize,
3313 rd_cost, 0, INT64_MAX, pc_tree);
3314 } else if (bsize == BLOCK_32X32 && partition != PARTITION_NONE &&
3315 subsize >= BLOCK_16X16) {
3316 x->max_partition_size = BLOCK_32X32;
3317 x->min_partition_size = BLOCK_8X8;
3318 nonrd_pick_partition(cpi, td, tile_data, tp, mi_row, mi_col, bsize,
3319 rd_cost, 0, INT64_MAX, pc_tree);
3320 } else if (bsize == BLOCK_16X16 && partition != PARTITION_NONE) {
3321 x->max_partition_size = BLOCK_16X16;
3322 x->min_partition_size = BLOCK_8X8;
3323 nonrd_pick_partition(cpi, td, tile_data, tp, mi_row, mi_col, bsize,
3324 rd_cost, 0, INT64_MAX, pc_tree);
3325 } else {
3326 switch (partition) {
3327 case PARTITION_NONE:
3328 pc_tree->none.pred_pixel_ready = 1;
3329 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, rd_cost,
3330 subsize, &pc_tree->none);
3331 pc_tree->none.mic.mbmi = xd->mi[0]->mbmi;
3332 pc_tree->none.mbmi_ext = *x->mbmi_ext;
3333 pc_tree->none.skip_txfm[0] = x->skip_txfm[0];
3334 pc_tree->none.skip = x->skip;
3335 break;
3336 case PARTITION_VERT:
3337 pc_tree->vertical[0].pred_pixel_ready = 1;
3338 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, rd_cost,
3339 subsize, &pc_tree->vertical[0]);
3340 pc_tree->vertical[0].mic.mbmi = xd->mi[0]->mbmi;
3341 pc_tree->vertical[0].mbmi_ext = *x->mbmi_ext;
3342 pc_tree->vertical[0].skip_txfm[0] = x->skip_txfm[0];
3343 pc_tree->vertical[0].skip = x->skip;
3344 if (mi_col + hbs < cm->mi_cols) {
3345 pc_tree->vertical[1].pred_pixel_ready = 1;
3346 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col + hbs,
3347 &this_rdc, subsize, &pc_tree->vertical[1]);
3348 pc_tree->vertical[1].mic.mbmi = xd->mi[0]->mbmi;
3349 pc_tree->vertical[1].mbmi_ext = *x->mbmi_ext;
3350 pc_tree->vertical[1].skip_txfm[0] = x->skip_txfm[0];
3351 pc_tree->vertical[1].skip = x->skip;
3352 if (this_rdc.rate != INT_MAX && this_rdc.dist != INT64_MAX &&
3353 rd_cost->rate != INT_MAX && rd_cost->dist != INT64_MAX) {
3354 rd_cost->rate += this_rdc.rate;
3355 rd_cost->dist += this_rdc.dist;
3356 }
3357 }
3358 break;
3359 case PARTITION_HORZ:
3360 pc_tree->horizontal[0].pred_pixel_ready = 1;
3361 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, rd_cost,
3362 subsize, &pc_tree->horizontal[0]);
3363 pc_tree->horizontal[0].mic.mbmi = xd->mi[0]->mbmi;
3364 pc_tree->horizontal[0].mbmi_ext = *x->mbmi_ext;
3365 pc_tree->horizontal[0].skip_txfm[0] = x->skip_txfm[0];
3366 pc_tree->horizontal[0].skip = x->skip;
3367 if (mi_row + hbs < cm->mi_rows) {
3368 pc_tree->horizontal[1].pred_pixel_ready = 1;
3369 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row + hbs, mi_col,
3370 &this_rdc, subsize, &pc_tree->horizontal[1]);
3371 pc_tree->horizontal[1].mic.mbmi = xd->mi[0]->mbmi;
3372 pc_tree->horizontal[1].mbmi_ext = *x->mbmi_ext;
3373 pc_tree->horizontal[1].skip_txfm[0] = x->skip_txfm[0];
3374 pc_tree->horizontal[1].skip = x->skip;
3375 if (this_rdc.rate != INT_MAX && this_rdc.dist != INT64_MAX &&
3376 rd_cost->rate != INT_MAX && rd_cost->dist != INT64_MAX) {
3377 rd_cost->rate += this_rdc.rate;
3378 rd_cost->dist += this_rdc.dist;
3379 }
3380 }
3381 break;
3382 case PARTITION_SPLIT:
3383 subsize = get_subsize(bsize, PARTITION_SPLIT);
3384 nonrd_select_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col,
3385 subsize, output_enabled, rd_cost,
3386 pc_tree->split[0]);
3387 nonrd_select_partition(cpi, td, tile_data, mi + hbs, tp,
3388 mi_row, mi_col + hbs, subsize, output_enabled,
3389 &this_rdc, pc_tree->split[1]);
3390 if (this_rdc.rate != INT_MAX && this_rdc.dist != INT64_MAX &&
3391 rd_cost->rate != INT_MAX && rd_cost->dist != INT64_MAX) {
3392 rd_cost->rate += this_rdc.rate;
3393 rd_cost->dist += this_rdc.dist;
3394 }
3395 nonrd_select_partition(cpi, td, tile_data, mi + hbs * mis, tp,
3396 mi_row + hbs, mi_col, subsize, output_enabled,
3397 &this_rdc, pc_tree->split[2]);
3398 if (this_rdc.rate != INT_MAX && this_rdc.dist != INT64_MAX &&
3399 rd_cost->rate != INT_MAX && rd_cost->dist != INT64_MAX) {
3400 rd_cost->rate += this_rdc.rate;
3401 rd_cost->dist += this_rdc.dist;
3402 }
3403 nonrd_select_partition(cpi, td, tile_data, mi + hbs * mis + hbs, tp,
3404 mi_row + hbs, mi_col + hbs, subsize,
3405 output_enabled, &this_rdc, pc_tree->split[3]);
3406 if (this_rdc.rate != INT_MAX && this_rdc.dist != INT64_MAX &&
3407 rd_cost->rate != INT_MAX && rd_cost->dist != INT64_MAX) {
3408 rd_cost->rate += this_rdc.rate;
3409 rd_cost->dist += this_rdc.dist;
3410 }
3411 break;
3412 default:
3413 assert(0 && "Invalid partition type.");
3414 break;
3415 }
3416 }
3417
3418 if (bsize == BLOCK_64X64 && output_enabled)
3419 encode_sb_rt(cpi, td, tile_info, tp, mi_row, mi_col, 1, bsize, pc_tree);
3420 }
3421
3422
nonrd_use_partition(VP9_COMP * cpi,ThreadData * td,TileDataEnc * tile_data,MODE_INFO ** mi,TOKENEXTRA ** tp,int mi_row,int mi_col,BLOCK_SIZE bsize,int output_enabled,RD_COST * dummy_cost,PC_TREE * pc_tree)3423 static void nonrd_use_partition(VP9_COMP *cpi,
3424 ThreadData *td,
3425 TileDataEnc *tile_data,
3426 MODE_INFO **mi,
3427 TOKENEXTRA **tp,
3428 int mi_row, int mi_col,
3429 BLOCK_SIZE bsize, int output_enabled,
3430 RD_COST *dummy_cost, PC_TREE *pc_tree) {
3431 VP9_COMMON *const cm = &cpi->common;
3432 TileInfo *tile_info = &tile_data->tile_info;
3433 MACROBLOCK *const x = &td->mb;
3434 MACROBLOCKD *const xd = &x->e_mbd;
3435 const int bsl = b_width_log2_lookup[bsize], hbs = (1 << bsl) / 4;
3436 const int mis = cm->mi_stride;
3437 PARTITION_TYPE partition;
3438 BLOCK_SIZE subsize;
3439
3440 if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols)
3441 return;
3442
3443 subsize = (bsize >= BLOCK_8X8) ? mi[0]->mbmi.sb_type : BLOCK_4X4;
3444 partition = partition_lookup[bsl][subsize];
3445
3446 if (output_enabled && bsize != BLOCK_4X4) {
3447 int ctx = partition_plane_context(xd, mi_row, mi_col, bsize);
3448 td->counts->partition[ctx][partition]++;
3449 }
3450
3451 switch (partition) {
3452 case PARTITION_NONE:
3453 pc_tree->none.pred_pixel_ready = 1;
3454 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, dummy_cost,
3455 subsize, &pc_tree->none);
3456 pc_tree->none.mic.mbmi = xd->mi[0]->mbmi;
3457 pc_tree->none.mbmi_ext = *x->mbmi_ext;
3458 pc_tree->none.skip_txfm[0] = x->skip_txfm[0];
3459 pc_tree->none.skip = x->skip;
3460 encode_b_rt(cpi, td, tile_info, tp, mi_row, mi_col, output_enabled,
3461 subsize, &pc_tree->none);
3462 break;
3463 case PARTITION_VERT:
3464 pc_tree->vertical[0].pred_pixel_ready = 1;
3465 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, dummy_cost,
3466 subsize, &pc_tree->vertical[0]);
3467 pc_tree->vertical[0].mic.mbmi = xd->mi[0]->mbmi;
3468 pc_tree->vertical[0].mbmi_ext = *x->mbmi_ext;
3469 pc_tree->vertical[0].skip_txfm[0] = x->skip_txfm[0];
3470 pc_tree->vertical[0].skip = x->skip;
3471 encode_b_rt(cpi, td, tile_info, tp, mi_row, mi_col, output_enabled,
3472 subsize, &pc_tree->vertical[0]);
3473 if (mi_col + hbs < cm->mi_cols && bsize > BLOCK_8X8) {
3474 pc_tree->vertical[1].pred_pixel_ready = 1;
3475 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col + hbs,
3476 dummy_cost, subsize, &pc_tree->vertical[1]);
3477 pc_tree->vertical[1].mic.mbmi = xd->mi[0]->mbmi;
3478 pc_tree->vertical[1].mbmi_ext = *x->mbmi_ext;
3479 pc_tree->vertical[1].skip_txfm[0] = x->skip_txfm[0];
3480 pc_tree->vertical[1].skip = x->skip;
3481 encode_b_rt(cpi, td, tile_info, tp, mi_row, mi_col + hbs,
3482 output_enabled, subsize, &pc_tree->vertical[1]);
3483 }
3484 break;
3485 case PARTITION_HORZ:
3486 pc_tree->horizontal[0].pred_pixel_ready = 1;
3487 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, dummy_cost,
3488 subsize, &pc_tree->horizontal[0]);
3489 pc_tree->horizontal[0].mic.mbmi = xd->mi[0]->mbmi;
3490 pc_tree->horizontal[0].mbmi_ext = *x->mbmi_ext;
3491 pc_tree->horizontal[0].skip_txfm[0] = x->skip_txfm[0];
3492 pc_tree->horizontal[0].skip = x->skip;
3493 encode_b_rt(cpi, td, tile_info, tp, mi_row, mi_col, output_enabled,
3494 subsize, &pc_tree->horizontal[0]);
3495
3496 if (mi_row + hbs < cm->mi_rows && bsize > BLOCK_8X8) {
3497 pc_tree->horizontal[1].pred_pixel_ready = 1;
3498 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row + hbs, mi_col,
3499 dummy_cost, subsize, &pc_tree->horizontal[1]);
3500 pc_tree->horizontal[1].mic.mbmi = xd->mi[0]->mbmi;
3501 pc_tree->horizontal[1].mbmi_ext = *x->mbmi_ext;
3502 pc_tree->horizontal[1].skip_txfm[0] = x->skip_txfm[0];
3503 pc_tree->horizontal[1].skip = x->skip;
3504 encode_b_rt(cpi, td, tile_info, tp, mi_row + hbs, mi_col,
3505 output_enabled, subsize, &pc_tree->horizontal[1]);
3506 }
3507 break;
3508 case PARTITION_SPLIT:
3509 subsize = get_subsize(bsize, PARTITION_SPLIT);
3510 if (bsize == BLOCK_8X8) {
3511 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, dummy_cost,
3512 subsize, pc_tree->leaf_split[0]);
3513 encode_b_rt(cpi, td, tile_info, tp, mi_row, mi_col,
3514 output_enabled, subsize, pc_tree->leaf_split[0]);
3515 } else {
3516 nonrd_use_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col,
3517 subsize, output_enabled, dummy_cost,
3518 pc_tree->split[0]);
3519 nonrd_use_partition(cpi, td, tile_data, mi + hbs, tp,
3520 mi_row, mi_col + hbs, subsize, output_enabled,
3521 dummy_cost, pc_tree->split[1]);
3522 nonrd_use_partition(cpi, td, tile_data, mi + hbs * mis, tp,
3523 mi_row + hbs, mi_col, subsize, output_enabled,
3524 dummy_cost, pc_tree->split[2]);
3525 nonrd_use_partition(cpi, td, tile_data, mi + hbs * mis + hbs, tp,
3526 mi_row + hbs, mi_col + hbs, subsize, output_enabled,
3527 dummy_cost, pc_tree->split[3]);
3528 }
3529 break;
3530 default:
3531 assert(0 && "Invalid partition type.");
3532 break;
3533 }
3534
3535 if (partition != PARTITION_SPLIT || bsize == BLOCK_8X8)
3536 update_partition_context(xd, mi_row, mi_col, subsize, bsize);
3537 }
3538
encode_nonrd_sb_row(VP9_COMP * cpi,ThreadData * td,TileDataEnc * tile_data,int mi_row,TOKENEXTRA ** tp)3539 static void encode_nonrd_sb_row(VP9_COMP *cpi,
3540 ThreadData *td,
3541 TileDataEnc *tile_data,
3542 int mi_row,
3543 TOKENEXTRA **tp) {
3544 SPEED_FEATURES *const sf = &cpi->sf;
3545 VP9_COMMON *const cm = &cpi->common;
3546 TileInfo *const tile_info = &tile_data->tile_info;
3547 MACROBLOCK *const x = &td->mb;
3548 MACROBLOCKD *const xd = &x->e_mbd;
3549 int mi_col;
3550
3551 // Initialize the left context for the new SB row
3552 memset(&xd->left_context, 0, sizeof(xd->left_context));
3553 memset(xd->left_seg_context, 0, sizeof(xd->left_seg_context));
3554
3555 // Code each SB in the row
3556 for (mi_col = tile_info->mi_col_start; mi_col < tile_info->mi_col_end;
3557 mi_col += MI_BLOCK_SIZE) {
3558 const struct segmentation *const seg = &cm->seg;
3559 RD_COST dummy_rdc;
3560 const int idx_str = cm->mi_stride * mi_row + mi_col;
3561 MODE_INFO **mi = cm->mi_grid_visible + idx_str;
3562 PARTITION_SEARCH_TYPE partition_search_type = sf->partition_search_type;
3563 BLOCK_SIZE bsize = BLOCK_64X64;
3564 int seg_skip = 0;
3565 x->source_variance = UINT_MAX;
3566 vp9_zero(x->pred_mv);
3567 vp9_rd_cost_init(&dummy_rdc);
3568 x->color_sensitivity[0] = 0;
3569 x->color_sensitivity[1] = 0;
3570
3571 if (seg->enabled) {
3572 const uint8_t *const map = seg->update_map ? cpi->segmentation_map
3573 : cm->last_frame_seg_map;
3574 int segment_id = get_segment_id(cm, map, BLOCK_64X64, mi_row, mi_col);
3575 seg_skip = segfeature_active(seg, segment_id, SEG_LVL_SKIP);
3576 if (seg_skip) {
3577 partition_search_type = FIXED_PARTITION;
3578 }
3579 }
3580
3581 // Set the partition type of the 64X64 block
3582 switch (partition_search_type) {
3583 case VAR_BASED_PARTITION:
3584 // TODO(jingning, marpan): The mode decision and encoding process
3585 // support both intra and inter sub8x8 block coding for RTC mode.
3586 // Tune the thresholds accordingly to use sub8x8 block coding for
3587 // coding performance improvement.
3588 choose_partitioning(cpi, tile_info, x, mi_row, mi_col);
3589 nonrd_use_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col,
3590 BLOCK_64X64, 1, &dummy_rdc, td->pc_root);
3591 break;
3592 case SOURCE_VAR_BASED_PARTITION:
3593 set_source_var_based_partition(cpi, tile_info, x, mi, mi_row, mi_col);
3594 nonrd_use_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col,
3595 BLOCK_64X64, 1, &dummy_rdc, td->pc_root);
3596 break;
3597 case FIXED_PARTITION:
3598 if (!seg_skip)
3599 bsize = sf->always_this_block_size;
3600 set_fixed_partitioning(cpi, tile_info, mi, mi_row, mi_col, bsize);
3601 nonrd_use_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col,
3602 BLOCK_64X64, 1, &dummy_rdc, td->pc_root);
3603 break;
3604 case REFERENCE_PARTITION:
3605 set_offsets(cpi, tile_info, x, mi_row, mi_col, BLOCK_64X64);
3606 if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ && cm->seg.enabled &&
3607 xd->mi[0]->mbmi.segment_id) {
3608 // Use lower max_partition_size for low resoultions.
3609 if (cm->width <= 352 && cm->height <= 288)
3610 x->max_partition_size = BLOCK_32X32;
3611 else
3612 x->max_partition_size = BLOCK_64X64;
3613 x->min_partition_size = BLOCK_8X8;
3614 nonrd_pick_partition(cpi, td, tile_data, tp, mi_row, mi_col,
3615 BLOCK_64X64, &dummy_rdc, 1,
3616 INT64_MAX, td->pc_root);
3617 } else {
3618 choose_partitioning(cpi, tile_info, x, mi_row, mi_col);
3619 // TODO(marpan): Seems like nonrd_select_partition does not support
3620 // 4x4 partition. Since 4x4 is used on key frame, use this switch
3621 // for now.
3622 if (cm->frame_type == KEY_FRAME)
3623 nonrd_use_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col,
3624 BLOCK_64X64, 1, &dummy_rdc, td->pc_root);
3625 else
3626 nonrd_select_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col,
3627 BLOCK_64X64, 1, &dummy_rdc, td->pc_root);
3628 }
3629
3630 break;
3631 default:
3632 assert(0);
3633 break;
3634 }
3635 }
3636 }
3637 // end RTC play code
3638
set_var_thresh_from_histogram(VP9_COMP * cpi)3639 static int set_var_thresh_from_histogram(VP9_COMP *cpi) {
3640 const SPEED_FEATURES *const sf = &cpi->sf;
3641 const VP9_COMMON *const cm = &cpi->common;
3642
3643 const uint8_t *src = cpi->Source->y_buffer;
3644 const uint8_t *last_src = cpi->Last_Source->y_buffer;
3645 const int src_stride = cpi->Source->y_stride;
3646 const int last_stride = cpi->Last_Source->y_stride;
3647
3648 // Pick cutoff threshold
3649 const int cutoff = (MIN(cm->width, cm->height) >= 720) ?
3650 (cm->MBs * VAR_HIST_LARGE_CUT_OFF / 100) :
3651 (cm->MBs * VAR_HIST_SMALL_CUT_OFF / 100);
3652 DECLARE_ALIGNED(16, int, hist[VAR_HIST_BINS]);
3653 diff *var16 = cpi->source_diff_var;
3654
3655 int sum = 0;
3656 int i, j;
3657
3658 memset(hist, 0, VAR_HIST_BINS * sizeof(hist[0]));
3659
3660 for (i = 0; i < cm->mb_rows; i++) {
3661 for (j = 0; j < cm->mb_cols; j++) {
3662 #if CONFIG_VP9_HIGHBITDEPTH
3663 if (cm->use_highbitdepth) {
3664 switch (cm->bit_depth) {
3665 case VPX_BITS_8:
3666 vpx_highbd_8_get16x16var(src, src_stride, last_src, last_stride,
3667 &var16->sse, &var16->sum);
3668 break;
3669 case VPX_BITS_10:
3670 vpx_highbd_10_get16x16var(src, src_stride, last_src, last_stride,
3671 &var16->sse, &var16->sum);
3672 break;
3673 case VPX_BITS_12:
3674 vpx_highbd_12_get16x16var(src, src_stride, last_src, last_stride,
3675 &var16->sse, &var16->sum);
3676 break;
3677 default:
3678 assert(0 && "cm->bit_depth should be VPX_BITS_8, VPX_BITS_10"
3679 " or VPX_BITS_12");
3680 return -1;
3681 }
3682 } else {
3683 vpx_get16x16var(src, src_stride, last_src, last_stride,
3684 &var16->sse, &var16->sum);
3685 }
3686 #else
3687 vpx_get16x16var(src, src_stride, last_src, last_stride,
3688 &var16->sse, &var16->sum);
3689 #endif // CONFIG_VP9_HIGHBITDEPTH
3690 var16->var = var16->sse -
3691 (((uint32_t)var16->sum * var16->sum) >> 8);
3692
3693 if (var16->var >= VAR_HIST_MAX_BG_VAR)
3694 hist[VAR_HIST_BINS - 1]++;
3695 else
3696 hist[var16->var / VAR_HIST_FACTOR]++;
3697
3698 src += 16;
3699 last_src += 16;
3700 var16++;
3701 }
3702
3703 src = src - cm->mb_cols * 16 + 16 * src_stride;
3704 last_src = last_src - cm->mb_cols * 16 + 16 * last_stride;
3705 }
3706
3707 cpi->source_var_thresh = 0;
3708
3709 if (hist[VAR_HIST_BINS - 1] < cutoff) {
3710 for (i = 0; i < VAR_HIST_BINS - 1; i++) {
3711 sum += hist[i];
3712
3713 if (sum > cutoff) {
3714 cpi->source_var_thresh = (i + 1) * VAR_HIST_FACTOR;
3715 return 0;
3716 }
3717 }
3718 }
3719
3720 return sf->search_type_check_frequency;
3721 }
3722
source_var_based_partition_search_method(VP9_COMP * cpi)3723 static void source_var_based_partition_search_method(VP9_COMP *cpi) {
3724 VP9_COMMON *const cm = &cpi->common;
3725 SPEED_FEATURES *const sf = &cpi->sf;
3726
3727 if (cm->frame_type == KEY_FRAME) {
3728 // For key frame, use SEARCH_PARTITION.
3729 sf->partition_search_type = SEARCH_PARTITION;
3730 } else if (cm->intra_only) {
3731 sf->partition_search_type = FIXED_PARTITION;
3732 } else {
3733 if (cm->last_width != cm->width || cm->last_height != cm->height) {
3734 if (cpi->source_diff_var)
3735 vpx_free(cpi->source_diff_var);
3736
3737 CHECK_MEM_ERROR(cm, cpi->source_diff_var,
3738 vpx_calloc(cm->MBs, sizeof(diff)));
3739 }
3740
3741 if (!cpi->frames_till_next_var_check)
3742 cpi->frames_till_next_var_check = set_var_thresh_from_histogram(cpi);
3743
3744 if (cpi->frames_till_next_var_check > 0) {
3745 sf->partition_search_type = FIXED_PARTITION;
3746 cpi->frames_till_next_var_check--;
3747 }
3748 }
3749 }
3750
get_skip_encode_frame(const VP9_COMMON * cm,ThreadData * const td)3751 static int get_skip_encode_frame(const VP9_COMMON *cm, ThreadData *const td) {
3752 unsigned int intra_count = 0, inter_count = 0;
3753 int j;
3754
3755 for (j = 0; j < INTRA_INTER_CONTEXTS; ++j) {
3756 intra_count += td->counts->intra_inter[j][0];
3757 inter_count += td->counts->intra_inter[j][1];
3758 }
3759
3760 return (intra_count << 2) < inter_count &&
3761 cm->frame_type != KEY_FRAME &&
3762 cm->show_frame;
3763 }
3764
vp9_init_tile_data(VP9_COMP * cpi)3765 void vp9_init_tile_data(VP9_COMP *cpi) {
3766 VP9_COMMON *const cm = &cpi->common;
3767 const int tile_cols = 1 << cm->log2_tile_cols;
3768 const int tile_rows = 1 << cm->log2_tile_rows;
3769 int tile_col, tile_row;
3770 TOKENEXTRA *pre_tok = cpi->tile_tok[0][0];
3771 int tile_tok = 0;
3772
3773 if (cpi->tile_data == NULL || cpi->allocated_tiles < tile_cols * tile_rows) {
3774 if (cpi->tile_data != NULL)
3775 vpx_free(cpi->tile_data);
3776 CHECK_MEM_ERROR(cm, cpi->tile_data,
3777 vpx_malloc(tile_cols * tile_rows * sizeof(*cpi->tile_data)));
3778 cpi->allocated_tiles = tile_cols * tile_rows;
3779
3780 for (tile_row = 0; tile_row < tile_rows; ++tile_row)
3781 for (tile_col = 0; tile_col < tile_cols; ++tile_col) {
3782 TileDataEnc *tile_data =
3783 &cpi->tile_data[tile_row * tile_cols + tile_col];
3784 int i, j;
3785 for (i = 0; i < BLOCK_SIZES; ++i) {
3786 for (j = 0; j < MAX_MODES; ++j) {
3787 tile_data->thresh_freq_fact[i][j] = 32;
3788 tile_data->mode_map[i][j] = j;
3789 }
3790 }
3791 }
3792 }
3793
3794 for (tile_row = 0; tile_row < tile_rows; ++tile_row) {
3795 for (tile_col = 0; tile_col < tile_cols; ++tile_col) {
3796 TileInfo *tile_info =
3797 &cpi->tile_data[tile_row * tile_cols + tile_col].tile_info;
3798 vp9_tile_init(tile_info, cm, tile_row, tile_col);
3799
3800 cpi->tile_tok[tile_row][tile_col] = pre_tok + tile_tok;
3801 pre_tok = cpi->tile_tok[tile_row][tile_col];
3802 tile_tok = allocated_tokens(*tile_info);
3803 }
3804 }
3805 }
3806
vp9_encode_tile(VP9_COMP * cpi,ThreadData * td,int tile_row,int tile_col)3807 void vp9_encode_tile(VP9_COMP *cpi, ThreadData *td,
3808 int tile_row, int tile_col) {
3809 VP9_COMMON *const cm = &cpi->common;
3810 const int tile_cols = 1 << cm->log2_tile_cols;
3811 TileDataEnc *this_tile =
3812 &cpi->tile_data[tile_row * tile_cols + tile_col];
3813 const TileInfo * const tile_info = &this_tile->tile_info;
3814 TOKENEXTRA *tok = cpi->tile_tok[tile_row][tile_col];
3815 int mi_row;
3816
3817 for (mi_row = tile_info->mi_row_start; mi_row < tile_info->mi_row_end;
3818 mi_row += MI_BLOCK_SIZE) {
3819 if (cpi->sf.use_nonrd_pick_mode)
3820 encode_nonrd_sb_row(cpi, td, this_tile, mi_row, &tok);
3821 else
3822 encode_rd_sb_row(cpi, td, this_tile, mi_row, &tok);
3823 }
3824 cpi->tok_count[tile_row][tile_col] =
3825 (unsigned int)(tok - cpi->tile_tok[tile_row][tile_col]);
3826 assert(tok - cpi->tile_tok[tile_row][tile_col] <=
3827 allocated_tokens(*tile_info));
3828 }
3829
encode_tiles(VP9_COMP * cpi)3830 static void encode_tiles(VP9_COMP *cpi) {
3831 VP9_COMMON *const cm = &cpi->common;
3832 const int tile_cols = 1 << cm->log2_tile_cols;
3833 const int tile_rows = 1 << cm->log2_tile_rows;
3834 int tile_col, tile_row;
3835
3836 vp9_init_tile_data(cpi);
3837
3838 for (tile_row = 0; tile_row < tile_rows; ++tile_row)
3839 for (tile_col = 0; tile_col < tile_cols; ++tile_col)
3840 vp9_encode_tile(cpi, &cpi->td, tile_row, tile_col);
3841 }
3842
3843 #if CONFIG_FP_MB_STATS
input_fpmb_stats(FIRSTPASS_MB_STATS * firstpass_mb_stats,VP9_COMMON * cm,uint8_t ** this_frame_mb_stats)3844 static int input_fpmb_stats(FIRSTPASS_MB_STATS *firstpass_mb_stats,
3845 VP9_COMMON *cm, uint8_t **this_frame_mb_stats) {
3846 uint8_t *mb_stats_in = firstpass_mb_stats->mb_stats_start +
3847 cm->current_video_frame * cm->MBs * sizeof(uint8_t);
3848
3849 if (mb_stats_in > firstpass_mb_stats->mb_stats_end)
3850 return EOF;
3851
3852 *this_frame_mb_stats = mb_stats_in;
3853
3854 return 1;
3855 }
3856 #endif
3857
encode_frame_internal(VP9_COMP * cpi)3858 static void encode_frame_internal(VP9_COMP *cpi) {
3859 SPEED_FEATURES *const sf = &cpi->sf;
3860 ThreadData *const td = &cpi->td;
3861 MACROBLOCK *const x = &td->mb;
3862 VP9_COMMON *const cm = &cpi->common;
3863 MACROBLOCKD *const xd = &x->e_mbd;
3864 RD_COUNTS *const rdc = &cpi->td.rd_counts;
3865
3866 xd->mi = cm->mi_grid_visible;
3867 xd->mi[0] = cm->mi;
3868
3869 vp9_zero(*td->counts);
3870 vp9_zero(rdc->coef_counts);
3871 vp9_zero(rdc->comp_pred_diff);
3872 vp9_zero(rdc->filter_diff);
3873
3874 xd->lossless = cm->base_qindex == 0 &&
3875 cm->y_dc_delta_q == 0 &&
3876 cm->uv_dc_delta_q == 0 &&
3877 cm->uv_ac_delta_q == 0;
3878
3879 #if CONFIG_VP9_HIGHBITDEPTH
3880 if (cm->use_highbitdepth)
3881 x->fwd_txm4x4 = xd->lossless ? vp9_highbd_fwht4x4 : vpx_highbd_fdct4x4;
3882 else
3883 x->fwd_txm4x4 = xd->lossless ? vp9_fwht4x4 : vpx_fdct4x4;
3884 x->highbd_itxm_add = xd->lossless ? vp9_highbd_iwht4x4_add :
3885 vp9_highbd_idct4x4_add;
3886 #else
3887 x->fwd_txm4x4 = xd->lossless ? vp9_fwht4x4 : vpx_fdct4x4;
3888 #endif // CONFIG_VP9_HIGHBITDEPTH
3889 x->itxm_add = xd->lossless ? vp9_iwht4x4_add : vp9_idct4x4_add;
3890
3891 if (xd->lossless)
3892 x->optimize = 0;
3893
3894 cm->tx_mode = select_tx_mode(cpi, xd);
3895
3896 vp9_frame_init_quantizer(cpi);
3897
3898 vp9_initialize_rd_consts(cpi);
3899 vp9_initialize_me_consts(cpi, x, cm->base_qindex);
3900 init_encode_frame_mb_context(cpi);
3901 cm->use_prev_frame_mvs = !cm->error_resilient_mode &&
3902 cm->width == cm->last_width &&
3903 cm->height == cm->last_height &&
3904 !cm->intra_only &&
3905 cm->last_show_frame;
3906 // Special case: set prev_mi to NULL when the previous mode info
3907 // context cannot be used.
3908 cm->prev_mi = cm->use_prev_frame_mvs ?
3909 cm->prev_mip + cm->mi_stride + 1 : NULL;
3910
3911 x->quant_fp = cpi->sf.use_quant_fp;
3912 vp9_zero(x->skip_txfm);
3913 if (sf->use_nonrd_pick_mode) {
3914 // Initialize internal buffer pointers for rtc coding, where non-RD
3915 // mode decision is used and hence no buffer pointer swap needed.
3916 int i;
3917 struct macroblock_plane *const p = x->plane;
3918 struct macroblockd_plane *const pd = xd->plane;
3919 PICK_MODE_CONTEXT *ctx = &cpi->td.pc_root->none;
3920
3921 for (i = 0; i < MAX_MB_PLANE; ++i) {
3922 p[i].coeff = ctx->coeff_pbuf[i][0];
3923 p[i].qcoeff = ctx->qcoeff_pbuf[i][0];
3924 pd[i].dqcoeff = ctx->dqcoeff_pbuf[i][0];
3925 p[i].eobs = ctx->eobs_pbuf[i][0];
3926 }
3927 vp9_zero(x->zcoeff_blk);
3928
3929 if (cm->frame_type != KEY_FRAME &&
3930 cpi->rc.frames_since_golden == 0 &&
3931 !cpi->use_svc)
3932 cpi->ref_frame_flags &= (~VP9_GOLD_FLAG);
3933
3934 if (sf->partition_search_type == SOURCE_VAR_BASED_PARTITION)
3935 source_var_based_partition_search_method(cpi);
3936 }
3937
3938 {
3939 struct vpx_usec_timer emr_timer;
3940 vpx_usec_timer_start(&emr_timer);
3941
3942 #if CONFIG_FP_MB_STATS
3943 if (cpi->use_fp_mb_stats) {
3944 input_fpmb_stats(&cpi->twopass.firstpass_mb_stats, cm,
3945 &cpi->twopass.this_frame_mb_stats);
3946 }
3947 #endif
3948
3949 // If allowed, encoding tiles in parallel with one thread handling one tile.
3950 if (MIN(cpi->oxcf.max_threads, 1 << cm->log2_tile_cols) > 1)
3951 vp9_encode_tiles_mt(cpi);
3952 else
3953 encode_tiles(cpi);
3954
3955 vpx_usec_timer_mark(&emr_timer);
3956 cpi->time_encode_sb_row += vpx_usec_timer_elapsed(&emr_timer);
3957 }
3958
3959 sf->skip_encode_frame = sf->skip_encode_sb ?
3960 get_skip_encode_frame(cm, td) : 0;
3961
3962 #if 0
3963 // Keep record of the total distortion this time around for future use
3964 cpi->last_frame_distortion = cpi->frame_distortion;
3965 #endif
3966 }
3967
get_interp_filter(const int64_t threshes[SWITCHABLE_FILTER_CONTEXTS],int is_alt_ref)3968 static INTERP_FILTER get_interp_filter(
3969 const int64_t threshes[SWITCHABLE_FILTER_CONTEXTS], int is_alt_ref) {
3970 if (!is_alt_ref &&
3971 threshes[EIGHTTAP_SMOOTH] > threshes[EIGHTTAP] &&
3972 threshes[EIGHTTAP_SMOOTH] > threshes[EIGHTTAP_SHARP] &&
3973 threshes[EIGHTTAP_SMOOTH] > threshes[SWITCHABLE - 1]) {
3974 return EIGHTTAP_SMOOTH;
3975 } else if (threshes[EIGHTTAP_SHARP] > threshes[EIGHTTAP] &&
3976 threshes[EIGHTTAP_SHARP] > threshes[SWITCHABLE - 1]) {
3977 return EIGHTTAP_SHARP;
3978 } else if (threshes[EIGHTTAP] > threshes[SWITCHABLE - 1]) {
3979 return EIGHTTAP;
3980 } else {
3981 return SWITCHABLE;
3982 }
3983 }
3984
vp9_encode_frame(VP9_COMP * cpi)3985 void vp9_encode_frame(VP9_COMP *cpi) {
3986 VP9_COMMON *const cm = &cpi->common;
3987
3988 // In the longer term the encoder should be generalized to match the
3989 // decoder such that we allow compound where one of the 3 buffers has a
3990 // different sign bias and that buffer is then the fixed ref. However, this
3991 // requires further work in the rd loop. For now the only supported encoder
3992 // side behavior is where the ALT ref buffer has opposite sign bias to
3993 // the other two.
3994 if (!frame_is_intra_only(cm)) {
3995 if ((cm->ref_frame_sign_bias[ALTREF_FRAME] ==
3996 cm->ref_frame_sign_bias[GOLDEN_FRAME]) ||
3997 (cm->ref_frame_sign_bias[ALTREF_FRAME] ==
3998 cm->ref_frame_sign_bias[LAST_FRAME])) {
3999 cpi->allow_comp_inter_inter = 0;
4000 } else {
4001 cpi->allow_comp_inter_inter = 1;
4002 cm->comp_fixed_ref = ALTREF_FRAME;
4003 cm->comp_var_ref[0] = LAST_FRAME;
4004 cm->comp_var_ref[1] = GOLDEN_FRAME;
4005 }
4006 }
4007
4008 if (cpi->sf.frame_parameter_update) {
4009 int i;
4010 RD_OPT *const rd_opt = &cpi->rd;
4011 FRAME_COUNTS *counts = cpi->td.counts;
4012 RD_COUNTS *const rdc = &cpi->td.rd_counts;
4013
4014 // This code does a single RD pass over the whole frame assuming
4015 // either compound, single or hybrid prediction as per whatever has
4016 // worked best for that type of frame in the past.
4017 // It also predicts whether another coding mode would have worked
4018 // better that this coding mode. If that is the case, it remembers
4019 // that for subsequent frames.
4020 // It does the same analysis for transform size selection also.
4021 const MV_REFERENCE_FRAME frame_type = get_frame_type(cpi);
4022 int64_t *const mode_thrs = rd_opt->prediction_type_threshes[frame_type];
4023 int64_t *const filter_thrs = rd_opt->filter_threshes[frame_type];
4024 const int is_alt_ref = frame_type == ALTREF_FRAME;
4025
4026 /* prediction (compound, single or hybrid) mode selection */
4027 if (is_alt_ref || !cpi->allow_comp_inter_inter)
4028 cm->reference_mode = SINGLE_REFERENCE;
4029 else if (mode_thrs[COMPOUND_REFERENCE] > mode_thrs[SINGLE_REFERENCE] &&
4030 mode_thrs[COMPOUND_REFERENCE] >
4031 mode_thrs[REFERENCE_MODE_SELECT] &&
4032 check_dual_ref_flags(cpi) &&
4033 cpi->static_mb_pct == 100)
4034 cm->reference_mode = COMPOUND_REFERENCE;
4035 else if (mode_thrs[SINGLE_REFERENCE] > mode_thrs[REFERENCE_MODE_SELECT])
4036 cm->reference_mode = SINGLE_REFERENCE;
4037 else
4038 cm->reference_mode = REFERENCE_MODE_SELECT;
4039
4040 if (cm->interp_filter == SWITCHABLE)
4041 cm->interp_filter = get_interp_filter(filter_thrs, is_alt_ref);
4042
4043 encode_frame_internal(cpi);
4044
4045 for (i = 0; i < REFERENCE_MODES; ++i)
4046 mode_thrs[i] = (mode_thrs[i] + rdc->comp_pred_diff[i] / cm->MBs) / 2;
4047
4048 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i)
4049 filter_thrs[i] = (filter_thrs[i] + rdc->filter_diff[i] / cm->MBs) / 2;
4050
4051 if (cm->reference_mode == REFERENCE_MODE_SELECT) {
4052 int single_count_zero = 0;
4053 int comp_count_zero = 0;
4054
4055 for (i = 0; i < COMP_INTER_CONTEXTS; i++) {
4056 single_count_zero += counts->comp_inter[i][0];
4057 comp_count_zero += counts->comp_inter[i][1];
4058 }
4059
4060 if (comp_count_zero == 0) {
4061 cm->reference_mode = SINGLE_REFERENCE;
4062 vp9_zero(counts->comp_inter);
4063 } else if (single_count_zero == 0) {
4064 cm->reference_mode = COMPOUND_REFERENCE;
4065 vp9_zero(counts->comp_inter);
4066 }
4067 }
4068
4069 if (cm->tx_mode == TX_MODE_SELECT) {
4070 int count4x4 = 0;
4071 int count8x8_lp = 0, count8x8_8x8p = 0;
4072 int count16x16_16x16p = 0, count16x16_lp = 0;
4073 int count32x32 = 0;
4074
4075 for (i = 0; i < TX_SIZE_CONTEXTS; ++i) {
4076 count4x4 += counts->tx.p32x32[i][TX_4X4];
4077 count4x4 += counts->tx.p16x16[i][TX_4X4];
4078 count4x4 += counts->tx.p8x8[i][TX_4X4];
4079
4080 count8x8_lp += counts->tx.p32x32[i][TX_8X8];
4081 count8x8_lp += counts->tx.p16x16[i][TX_8X8];
4082 count8x8_8x8p += counts->tx.p8x8[i][TX_8X8];
4083
4084 count16x16_16x16p += counts->tx.p16x16[i][TX_16X16];
4085 count16x16_lp += counts->tx.p32x32[i][TX_16X16];
4086 count32x32 += counts->tx.p32x32[i][TX_32X32];
4087 }
4088 if (count4x4 == 0 && count16x16_lp == 0 && count16x16_16x16p == 0 &&
4089 count32x32 == 0) {
4090 cm->tx_mode = ALLOW_8X8;
4091 reset_skip_tx_size(cm, TX_8X8);
4092 } else if (count8x8_8x8p == 0 && count16x16_16x16p == 0 &&
4093 count8x8_lp == 0 && count16x16_lp == 0 && count32x32 == 0) {
4094 cm->tx_mode = ONLY_4X4;
4095 reset_skip_tx_size(cm, TX_4X4);
4096 } else if (count8x8_lp == 0 && count16x16_lp == 0 && count4x4 == 0) {
4097 cm->tx_mode = ALLOW_32X32;
4098 } else if (count32x32 == 0 && count8x8_lp == 0 && count4x4 == 0) {
4099 cm->tx_mode = ALLOW_16X16;
4100 reset_skip_tx_size(cm, TX_16X16);
4101 }
4102 }
4103 } else {
4104 cm->reference_mode = SINGLE_REFERENCE;
4105 encode_frame_internal(cpi);
4106 }
4107 }
4108
sum_intra_stats(FRAME_COUNTS * counts,const MODE_INFO * mi)4109 static void sum_intra_stats(FRAME_COUNTS *counts, const MODE_INFO *mi) {
4110 const PREDICTION_MODE y_mode = mi->mbmi.mode;
4111 const PREDICTION_MODE uv_mode = mi->mbmi.uv_mode;
4112 const BLOCK_SIZE bsize = mi->mbmi.sb_type;
4113
4114 if (bsize < BLOCK_8X8) {
4115 int idx, idy;
4116 const int num_4x4_w = num_4x4_blocks_wide_lookup[bsize];
4117 const int num_4x4_h = num_4x4_blocks_high_lookup[bsize];
4118 for (idy = 0; idy < 2; idy += num_4x4_h)
4119 for (idx = 0; idx < 2; idx += num_4x4_w)
4120 ++counts->y_mode[0][mi->bmi[idy * 2 + idx].as_mode];
4121 } else {
4122 ++counts->y_mode[size_group_lookup[bsize]][y_mode];
4123 }
4124
4125 ++counts->uv_mode[y_mode][uv_mode];
4126 }
4127
encode_superblock(VP9_COMP * cpi,ThreadData * td,TOKENEXTRA ** t,int output_enabled,int mi_row,int mi_col,BLOCK_SIZE bsize,PICK_MODE_CONTEXT * ctx)4128 static void encode_superblock(VP9_COMP *cpi, ThreadData *td,
4129 TOKENEXTRA **t, int output_enabled,
4130 int mi_row, int mi_col, BLOCK_SIZE bsize,
4131 PICK_MODE_CONTEXT *ctx) {
4132 VP9_COMMON *const cm = &cpi->common;
4133 MACROBLOCK *const x = &td->mb;
4134 MACROBLOCKD *const xd = &x->e_mbd;
4135 MODE_INFO **mi_8x8 = xd->mi;
4136 MODE_INFO *mi = mi_8x8[0];
4137 MB_MODE_INFO *mbmi = &mi->mbmi;
4138 const int seg_skip = segfeature_active(&cm->seg, mbmi->segment_id,
4139 SEG_LVL_SKIP);
4140 const int mis = cm->mi_stride;
4141 const int mi_width = num_8x8_blocks_wide_lookup[bsize];
4142 const int mi_height = num_8x8_blocks_high_lookup[bsize];
4143
4144 x->skip_recode = !x->select_tx_size && mbmi->sb_type >= BLOCK_8X8 &&
4145 cpi->oxcf.aq_mode != COMPLEXITY_AQ &&
4146 cpi->oxcf.aq_mode != CYCLIC_REFRESH_AQ &&
4147 cpi->sf.allow_skip_recode;
4148
4149 if (!x->skip_recode && !cpi->sf.use_nonrd_pick_mode)
4150 memset(x->skip_txfm, 0, sizeof(x->skip_txfm));
4151
4152 x->skip_optimize = ctx->is_coded;
4153 ctx->is_coded = 1;
4154 x->use_lp32x32fdct = cpi->sf.use_lp32x32fdct;
4155 x->skip_encode = (!output_enabled && cpi->sf.skip_encode_frame &&
4156 x->q_index < QIDX_SKIP_THRESH);
4157
4158 if (x->skip_encode)
4159 return;
4160
4161 if (!is_inter_block(mbmi)) {
4162 int plane;
4163 mbmi->skip = 1;
4164 for (plane = 0; plane < MAX_MB_PLANE; ++plane)
4165 vp9_encode_intra_block_plane(x, MAX(bsize, BLOCK_8X8), plane);
4166 if (output_enabled)
4167 sum_intra_stats(td->counts, mi);
4168 vp9_tokenize_sb(cpi, td, t, !output_enabled, MAX(bsize, BLOCK_8X8));
4169 } else {
4170 int ref;
4171 const int is_compound = has_second_ref(mbmi);
4172 set_ref_ptrs(cm, xd, mbmi->ref_frame[0], mbmi->ref_frame[1]);
4173 for (ref = 0; ref < 1 + is_compound; ++ref) {
4174 YV12_BUFFER_CONFIG *cfg = get_ref_frame_buffer(cpi,
4175 mbmi->ref_frame[ref]);
4176 assert(cfg != NULL);
4177 vp9_setup_pre_planes(xd, ref, cfg, mi_row, mi_col,
4178 &xd->block_refs[ref]->sf);
4179 }
4180 if (!(cpi->sf.reuse_inter_pred_sby && ctx->pred_pixel_ready) || seg_skip)
4181 vp9_build_inter_predictors_sby(xd, mi_row, mi_col, MAX(bsize, BLOCK_8X8));
4182
4183 vp9_build_inter_predictors_sbuv(xd, mi_row, mi_col, MAX(bsize, BLOCK_8X8));
4184
4185 vp9_encode_sb(x, MAX(bsize, BLOCK_8X8));
4186 vp9_tokenize_sb(cpi, td, t, !output_enabled, MAX(bsize, BLOCK_8X8));
4187 }
4188
4189 if (output_enabled) {
4190 if (cm->tx_mode == TX_MODE_SELECT &&
4191 mbmi->sb_type >= BLOCK_8X8 &&
4192 !(is_inter_block(mbmi) && (mbmi->skip || seg_skip))) {
4193 ++get_tx_counts(max_txsize_lookup[bsize], get_tx_size_context(xd),
4194 &td->counts->tx)[mbmi->tx_size];
4195 } else {
4196 int x, y;
4197 TX_SIZE tx_size;
4198 // The new intra coding scheme requires no change of transform size
4199 if (is_inter_block(&mi->mbmi)) {
4200 tx_size = MIN(tx_mode_to_biggest_tx_size[cm->tx_mode],
4201 max_txsize_lookup[bsize]);
4202 } else {
4203 tx_size = (bsize >= BLOCK_8X8) ? mbmi->tx_size : TX_4X4;
4204 }
4205
4206 for (y = 0; y < mi_height; y++)
4207 for (x = 0; x < mi_width; x++)
4208 if (mi_col + x < cm->mi_cols && mi_row + y < cm->mi_rows)
4209 mi_8x8[mis * y + x]->mbmi.tx_size = tx_size;
4210 }
4211 ++td->counts->tx.tx_totals[mbmi->tx_size];
4212 ++td->counts->tx.tx_totals[get_uv_tx_size(mbmi, &xd->plane[1])];
4213 }
4214 }
4215