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_dsp/vpx_dsp_common.h"
20 #include "vpx_ports/mem.h"
21 #include "vpx_ports/vpx_timer.h"
22 #include "vpx_ports/system_state.h"
23
24 #include "vp9/common/vp9_common.h"
25 #include "vp9/common/vp9_entropy.h"
26 #include "vp9/common/vp9_entropymode.h"
27 #include "vp9/common/vp9_idct.h"
28 #include "vp9/common/vp9_mvref_common.h"
29 #include "vp9/common/vp9_pred_common.h"
30 #include "vp9/common/vp9_quant_common.h"
31 #include "vp9/common/vp9_reconintra.h"
32 #include "vp9/common/vp9_reconinter.h"
33 #include "vp9/common/vp9_seg_common.h"
34 #include "vp9/common/vp9_tile_common.h"
35
36 #include "vp9/encoder/vp9_aq_360.h"
37 #include "vp9/encoder/vp9_aq_complexity.h"
38 #include "vp9/encoder/vp9_aq_cyclicrefresh.h"
39 #include "vp9/encoder/vp9_aq_variance.h"
40 #include "vp9/encoder/vp9_encodeframe.h"
41 #include "vp9/encoder/vp9_encodemb.h"
42 #include "vp9/encoder/vp9_encodemv.h"
43 #include "vp9/encoder/vp9_ethread.h"
44 #include "vp9/encoder/vp9_extend.h"
45 #include "vp9/encoder/vp9_pickmode.h"
46 #include "vp9/encoder/vp9_rd.h"
47 #include "vp9/encoder/vp9_rdopt.h"
48 #include "vp9/encoder/vp9_segmentation.h"
49 #include "vp9/encoder/vp9_tokenize.h"
50
51 static void encode_superblock(VP9_COMP *cpi, ThreadData *td, TOKENEXTRA **t,
52 int output_enabled, int mi_row, int mi_col,
53 BLOCK_SIZE bsize, PICK_MODE_CONTEXT *ctx);
54
55 // Machine learning-based early termination parameters.
56 static const double train_mean[24] = {
57 303501.697372, 3042630.372158, 24.694696, 1.392182,
58 689.413511, 162.027012, 1.478213, 0.0,
59 135382.260230, 912738.513263, 28.845217, 1.515230,
60 544.158492, 131.807995, 1.436863, 0.0,
61 43682.377587, 208131.711766, 28.084737, 1.356677,
62 138.254122, 119.522553, 1.252322, 0.0
63 };
64
65 static const double train_stdm[24] = {
66 673689.212982, 5996652.516628, 0.024449, 1.989792,
67 985.880847, 0.014638, 2.001898, 0.0,
68 208798.775332, 1812548.443284, 0.018693, 1.838009,
69 396.986910, 0.015657, 1.332541, 0.0,
70 55888.847031, 448587.962714, 0.017900, 1.904776,
71 98.652832, 0.016598, 1.320992, 0.0
72 };
73
74 // Error tolerance: 0.01%-0.0.05%-0.1%
75 static const double classifiers[24] = {
76 0.111736, 0.289977, 0.042219, 0.204765, 0.120410, -0.143863,
77 0.282376, 0.847811, 0.637161, 0.131570, 0.018636, 0.202134,
78 0.112797, 0.028162, 0.182450, 1.124367, 0.386133, 0.083700,
79 0.050028, 0.150873, 0.061119, 0.109318, 0.127255, 0.625211
80 };
81
82 // This is used as a reference when computing the source variance for the
83 // purpose of activity masking.
84 // Eventually this should be replaced by custom no-reference routines,
85 // which will be faster.
86 static const uint8_t VP9_VAR_OFFS[64] = {
87 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128,
88 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128,
89 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128,
90 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128,
91 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128
92 };
93
94 #if CONFIG_VP9_HIGHBITDEPTH
95 static const uint16_t VP9_HIGH_VAR_OFFS_8[64] = {
96 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128,
97 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128,
98 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128,
99 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128,
100 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128
101 };
102
103 static const uint16_t VP9_HIGH_VAR_OFFS_10[64] = {
104 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4,
105 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4,
106 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4,
107 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4,
108 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4,
109 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4,
110 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4,
111 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4
112 };
113
114 static const uint16_t VP9_HIGH_VAR_OFFS_12[64] = {
115 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16,
116 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16,
117 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16,
118 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16,
119 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16,
120 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16,
121 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16,
122 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16,
123 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16,
124 128 * 16
125 };
126 #endif // CONFIG_VP9_HIGHBITDEPTH
127
vp9_get_sby_perpixel_variance(VP9_COMP * cpi,const struct buf_2d * ref,BLOCK_SIZE bs)128 unsigned int vp9_get_sby_perpixel_variance(VP9_COMP *cpi,
129 const struct buf_2d *ref,
130 BLOCK_SIZE bs) {
131 unsigned int sse;
132 const unsigned int var =
133 cpi->fn_ptr[bs].vf(ref->buf, ref->stride, VP9_VAR_OFFS, 0, &sse);
134 return ROUND_POWER_OF_TWO(var, num_pels_log2_lookup[bs]);
135 }
136
137 #if CONFIG_VP9_HIGHBITDEPTH
vp9_high_get_sby_perpixel_variance(VP9_COMP * cpi,const struct buf_2d * ref,BLOCK_SIZE bs,int bd)138 unsigned int vp9_high_get_sby_perpixel_variance(VP9_COMP *cpi,
139 const struct buf_2d *ref,
140 BLOCK_SIZE bs, int bd) {
141 unsigned int var, sse;
142 switch (bd) {
143 case 10:
144 var =
145 cpi->fn_ptr[bs].vf(ref->buf, ref->stride,
146 CONVERT_TO_BYTEPTR(VP9_HIGH_VAR_OFFS_10), 0, &sse);
147 break;
148 case 12:
149 var =
150 cpi->fn_ptr[bs].vf(ref->buf, ref->stride,
151 CONVERT_TO_BYTEPTR(VP9_HIGH_VAR_OFFS_12), 0, &sse);
152 break;
153 case 8:
154 default:
155 var =
156 cpi->fn_ptr[bs].vf(ref->buf, ref->stride,
157 CONVERT_TO_BYTEPTR(VP9_HIGH_VAR_OFFS_8), 0, &sse);
158 break;
159 }
160 return (unsigned int)ROUND64_POWER_OF_TWO((int64_t)var,
161 num_pels_log2_lookup[bs]);
162 }
163 #endif // CONFIG_VP9_HIGHBITDEPTH
164
get_sby_perpixel_diff_variance(VP9_COMP * cpi,const struct buf_2d * ref,int mi_row,int mi_col,BLOCK_SIZE bs)165 static unsigned int get_sby_perpixel_diff_variance(VP9_COMP *cpi,
166 const struct buf_2d *ref,
167 int mi_row, int mi_col,
168 BLOCK_SIZE bs) {
169 unsigned int sse, var;
170 uint8_t *last_y;
171 const YV12_BUFFER_CONFIG *last = get_ref_frame_buffer(cpi, LAST_FRAME);
172
173 assert(last != NULL);
174 last_y =
175 &last->y_buffer[mi_row * MI_SIZE * last->y_stride + mi_col * MI_SIZE];
176 var = cpi->fn_ptr[bs].vf(ref->buf, ref->stride, last_y, last->y_stride, &sse);
177 return ROUND_POWER_OF_TWO(var, num_pels_log2_lookup[bs]);
178 }
179
get_rd_var_based_fixed_partition(VP9_COMP * cpi,MACROBLOCK * x,int mi_row,int mi_col)180 static BLOCK_SIZE get_rd_var_based_fixed_partition(VP9_COMP *cpi, MACROBLOCK *x,
181 int mi_row, int mi_col) {
182 unsigned int var = get_sby_perpixel_diff_variance(
183 cpi, &x->plane[0].src, mi_row, mi_col, BLOCK_64X64);
184 if (var < 8)
185 return BLOCK_64X64;
186 else if (var < 128)
187 return BLOCK_32X32;
188 else if (var < 2048)
189 return BLOCK_16X16;
190 else
191 return BLOCK_8X8;
192 }
193
194 // Lighter version of set_offsets that only sets the mode info
195 // pointers.
set_mode_info_offsets(VP9_COMMON * const cm,MACROBLOCK * const x,MACROBLOCKD * const xd,int mi_row,int mi_col)196 static INLINE void set_mode_info_offsets(VP9_COMMON *const cm,
197 MACROBLOCK *const x,
198 MACROBLOCKD *const xd, int mi_row,
199 int mi_col) {
200 const int idx_str = xd->mi_stride * mi_row + mi_col;
201 xd->mi = cm->mi_grid_visible + idx_str;
202 xd->mi[0] = cm->mi + idx_str;
203 x->mbmi_ext = x->mbmi_ext_base + (mi_row * cm->mi_cols + mi_col);
204 }
205
set_offsets(VP9_COMP * cpi,const TileInfo * const tile,MACROBLOCK * const x,int mi_row,int mi_col,BLOCK_SIZE bsize)206 static void set_offsets(VP9_COMP *cpi, const TileInfo *const tile,
207 MACROBLOCK *const x, int mi_row, int mi_col,
208 BLOCK_SIZE bsize) {
209 VP9_COMMON *const cm = &cpi->common;
210 MACROBLOCKD *const xd = &x->e_mbd;
211 MODE_INFO *mi;
212 const int mi_width = num_8x8_blocks_wide_lookup[bsize];
213 const int mi_height = num_8x8_blocks_high_lookup[bsize];
214 const struct segmentation *const seg = &cm->seg;
215 MvLimits *const mv_limits = &x->mv_limits;
216
217 set_skip_context(xd, mi_row, mi_col);
218
219 set_mode_info_offsets(cm, x, xd, mi_row, mi_col);
220
221 mi = xd->mi[0];
222
223 // Set up destination pointers.
224 vp9_setup_dst_planes(xd->plane, get_frame_new_buffer(cm), mi_row, mi_col);
225
226 // Set up limit values for MV components.
227 // Mv beyond the range do not produce new/different prediction block.
228 mv_limits->row_min = -(((mi_row + mi_height) * MI_SIZE) + VP9_INTERP_EXTEND);
229 mv_limits->col_min = -(((mi_col + mi_width) * MI_SIZE) + VP9_INTERP_EXTEND);
230 mv_limits->row_max = (cm->mi_rows - mi_row) * MI_SIZE + VP9_INTERP_EXTEND;
231 mv_limits->col_max = (cm->mi_cols - mi_col) * MI_SIZE + VP9_INTERP_EXTEND;
232
233 // Set up distance of MB to edge of frame in 1/8th pel units.
234 assert(!(mi_col & (mi_width - 1)) && !(mi_row & (mi_height - 1)));
235 set_mi_row_col(xd, tile, mi_row, mi_height, mi_col, mi_width, cm->mi_rows,
236 cm->mi_cols);
237
238 // Set up source buffers.
239 vp9_setup_src_planes(x, cpi->Source, mi_row, mi_col);
240
241 // R/D setup.
242 x->rddiv = cpi->rd.RDDIV;
243 x->rdmult = cpi->rd.RDMULT;
244
245 // Setup segment ID.
246 if (seg->enabled) {
247 if (cpi->oxcf.aq_mode != VARIANCE_AQ && cpi->oxcf.aq_mode != LOOKAHEAD_AQ &&
248 cpi->oxcf.aq_mode != EQUATOR360_AQ) {
249 const uint8_t *const map =
250 seg->update_map ? cpi->segmentation_map : cm->last_frame_seg_map;
251 mi->segment_id = get_segment_id(cm, map, bsize, mi_row, mi_col);
252 }
253 vp9_init_plane_quantizers(cpi, x);
254
255 x->encode_breakout = cpi->segment_encode_breakout[mi->segment_id];
256 } else {
257 mi->segment_id = 0;
258 x->encode_breakout = cpi->encode_breakout;
259 }
260
261 // required by vp9_append_sub8x8_mvs_for_idx() and vp9_find_best_ref_mvs()
262 xd->tile = *tile;
263 }
264
duplicate_mode_info_in_sb(VP9_COMMON * cm,MACROBLOCKD * xd,int mi_row,int mi_col,BLOCK_SIZE bsize)265 static void duplicate_mode_info_in_sb(VP9_COMMON *cm, MACROBLOCKD *xd,
266 int mi_row, int mi_col,
267 BLOCK_SIZE bsize) {
268 const int block_width =
269 VPXMIN(num_8x8_blocks_wide_lookup[bsize], cm->mi_cols - mi_col);
270 const int block_height =
271 VPXMIN(num_8x8_blocks_high_lookup[bsize], cm->mi_rows - mi_row);
272 const int mi_stride = xd->mi_stride;
273 MODE_INFO *const src_mi = xd->mi[0];
274 int i, j;
275
276 for (j = 0; j < block_height; ++j)
277 for (i = 0; i < block_width; ++i) xd->mi[j * mi_stride + i] = src_mi;
278 }
279
set_block_size(VP9_COMP * const cpi,MACROBLOCK * const x,MACROBLOCKD * const xd,int mi_row,int mi_col,BLOCK_SIZE bsize)280 static void set_block_size(VP9_COMP *const cpi, MACROBLOCK *const x,
281 MACROBLOCKD *const xd, int mi_row, int mi_col,
282 BLOCK_SIZE bsize) {
283 if (cpi->common.mi_cols > mi_col && cpi->common.mi_rows > mi_row) {
284 set_mode_info_offsets(&cpi->common, x, xd, mi_row, mi_col);
285 xd->mi[0]->sb_type = bsize;
286 }
287 }
288
289 typedef struct {
290 int64_t sum_square_error;
291 int64_t sum_error;
292 int log2_count;
293 int variance;
294 } var;
295
296 typedef struct {
297 var none;
298 var horz[2];
299 var vert[2];
300 } partition_variance;
301
302 typedef struct {
303 partition_variance part_variances;
304 var split[4];
305 } v4x4;
306
307 typedef struct {
308 partition_variance part_variances;
309 v4x4 split[4];
310 } v8x8;
311
312 typedef struct {
313 partition_variance part_variances;
314 v8x8 split[4];
315 } v16x16;
316
317 typedef struct {
318 partition_variance part_variances;
319 v16x16 split[4];
320 } v32x32;
321
322 typedef struct {
323 partition_variance part_variances;
324 v32x32 split[4];
325 } v64x64;
326
327 typedef struct {
328 partition_variance *part_variances;
329 var *split[4];
330 } variance_node;
331
332 typedef enum {
333 V16X16,
334 V32X32,
335 V64X64,
336 } TREE_LEVEL;
337
tree_to_node(void * data,BLOCK_SIZE bsize,variance_node * node)338 static void tree_to_node(void *data, BLOCK_SIZE bsize, variance_node *node) {
339 int i;
340 node->part_variances = NULL;
341 switch (bsize) {
342 case BLOCK_64X64: {
343 v64x64 *vt = (v64x64 *)data;
344 node->part_variances = &vt->part_variances;
345 for (i = 0; i < 4; i++)
346 node->split[i] = &vt->split[i].part_variances.none;
347 break;
348 }
349 case BLOCK_32X32: {
350 v32x32 *vt = (v32x32 *)data;
351 node->part_variances = &vt->part_variances;
352 for (i = 0; i < 4; i++)
353 node->split[i] = &vt->split[i].part_variances.none;
354 break;
355 }
356 case BLOCK_16X16: {
357 v16x16 *vt = (v16x16 *)data;
358 node->part_variances = &vt->part_variances;
359 for (i = 0; i < 4; i++)
360 node->split[i] = &vt->split[i].part_variances.none;
361 break;
362 }
363 case BLOCK_8X8: {
364 v8x8 *vt = (v8x8 *)data;
365 node->part_variances = &vt->part_variances;
366 for (i = 0; i < 4; i++)
367 node->split[i] = &vt->split[i].part_variances.none;
368 break;
369 }
370 case BLOCK_4X4: {
371 v4x4 *vt = (v4x4 *)data;
372 node->part_variances = &vt->part_variances;
373 for (i = 0; i < 4; i++) node->split[i] = &vt->split[i];
374 break;
375 }
376 default: {
377 assert(0);
378 break;
379 }
380 }
381 }
382
383 // Set variance values given sum square error, sum error, count.
fill_variance(int64_t s2,int64_t s,int c,var * v)384 static void fill_variance(int64_t s2, int64_t s, int c, var *v) {
385 v->sum_square_error = s2;
386 v->sum_error = s;
387 v->log2_count = c;
388 }
389
get_variance(var * v)390 static void get_variance(var *v) {
391 v->variance =
392 (int)(256 * (v->sum_square_error -
393 ((v->sum_error * v->sum_error) >> v->log2_count)) >>
394 v->log2_count);
395 }
396
sum_2_variances(const var * a,const var * b,var * r)397 static void sum_2_variances(const var *a, const var *b, var *r) {
398 assert(a->log2_count == b->log2_count);
399 fill_variance(a->sum_square_error + b->sum_square_error,
400 a->sum_error + b->sum_error, a->log2_count + 1, r);
401 }
402
fill_variance_tree(void * data,BLOCK_SIZE bsize)403 static void fill_variance_tree(void *data, BLOCK_SIZE bsize) {
404 variance_node node;
405 memset(&node, 0, sizeof(node));
406 tree_to_node(data, bsize, &node);
407 sum_2_variances(node.split[0], node.split[1], &node.part_variances->horz[0]);
408 sum_2_variances(node.split[2], node.split[3], &node.part_variances->horz[1]);
409 sum_2_variances(node.split[0], node.split[2], &node.part_variances->vert[0]);
410 sum_2_variances(node.split[1], node.split[3], &node.part_variances->vert[1]);
411 sum_2_variances(&node.part_variances->vert[0], &node.part_variances->vert[1],
412 &node.part_variances->none);
413 }
414
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)415 static int set_vt_partitioning(VP9_COMP *cpi, MACROBLOCK *const x,
416 MACROBLOCKD *const xd, void *data,
417 BLOCK_SIZE bsize, int mi_row, int mi_col,
418 int64_t threshold, BLOCK_SIZE bsize_min,
419 int force_split) {
420 VP9_COMMON *const cm = &cpi->common;
421 variance_node vt;
422 const int block_width = num_8x8_blocks_wide_lookup[bsize];
423 const int block_height = num_8x8_blocks_high_lookup[bsize];
424
425 assert(block_height == block_width);
426 tree_to_node(data, bsize, &vt);
427
428 if (force_split == 1) return 0;
429
430 // For bsize=bsize_min (16x16/8x8 for 8x8/4x4 downsampling), select if
431 // variance is below threshold, otherwise split will be selected.
432 // No check for vert/horiz split as too few samples for variance.
433 if (bsize == bsize_min) {
434 // Variance already computed to set the force_split.
435 if (cm->frame_type == KEY_FRAME) get_variance(&vt.part_variances->none);
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 return 0;
443 } else if (bsize > bsize_min) {
444 // Variance already computed to set the force_split.
445 if (cm->frame_type == KEY_FRAME) get_variance(&vt.part_variances->none);
446 // For key frame: take split for bsize above 32X32 or very high variance.
447 if (cm->frame_type == KEY_FRAME &&
448 (bsize > BLOCK_32X32 ||
449 vt.part_variances->none.variance > (threshold << 4))) {
450 return 0;
451 }
452 // If variance is low, take the bsize (no split).
453 if (mi_col + block_width / 2 < cm->mi_cols &&
454 mi_row + block_height / 2 < cm->mi_rows &&
455 vt.part_variances->none.variance < threshold) {
456 set_block_size(cpi, x, xd, mi_row, mi_col, bsize);
457 return 1;
458 }
459
460 // Check vertical split.
461 if (mi_row + block_height / 2 < cm->mi_rows) {
462 BLOCK_SIZE subsize = get_subsize(bsize, PARTITION_VERT);
463 get_variance(&vt.part_variances->vert[0]);
464 get_variance(&vt.part_variances->vert[1]);
465 if (vt.part_variances->vert[0].variance < threshold &&
466 vt.part_variances->vert[1].variance < threshold &&
467 get_plane_block_size(subsize, &xd->plane[1]) < BLOCK_INVALID) {
468 set_block_size(cpi, x, xd, mi_row, mi_col, subsize);
469 set_block_size(cpi, x, xd, mi_row, mi_col + block_width / 2, subsize);
470 return 1;
471 }
472 }
473 // Check horizontal split.
474 if (mi_col + block_width / 2 < cm->mi_cols) {
475 BLOCK_SIZE subsize = get_subsize(bsize, PARTITION_HORZ);
476 get_variance(&vt.part_variances->horz[0]);
477 get_variance(&vt.part_variances->horz[1]);
478 if (vt.part_variances->horz[0].variance < threshold &&
479 vt.part_variances->horz[1].variance < threshold &&
480 get_plane_block_size(subsize, &xd->plane[1]) < BLOCK_INVALID) {
481 set_block_size(cpi, x, xd, mi_row, mi_col, subsize);
482 set_block_size(cpi, x, xd, mi_row + block_height / 2, mi_col, subsize);
483 return 1;
484 }
485 }
486
487 return 0;
488 }
489 return 0;
490 }
491
scale_part_thresh_sumdiff(int64_t threshold_base,int speed,int width,int height,int content_state)492 int64_t scale_part_thresh_sumdiff(int64_t threshold_base, int speed, int width,
493 int height, int content_state) {
494 if (speed >= 8) {
495 if (width <= 640 && height <= 480)
496 return (5 * threshold_base) >> 2;
497 else if ((content_state == kLowSadLowSumdiff) ||
498 (content_state == kHighSadLowSumdiff) ||
499 (content_state == kLowVarHighSumdiff))
500 return (5 * threshold_base) >> 2;
501 } else if (speed == 7) {
502 if ((content_state == kLowSadLowSumdiff) ||
503 (content_state == kHighSadLowSumdiff) ||
504 (content_state == kLowVarHighSumdiff)) {
505 return (5 * threshold_base) >> 2;
506 }
507 }
508 return threshold_base;
509 }
510
511 // Set the variance split thresholds for following the block sizes:
512 // 0 - threshold_64x64, 1 - threshold_32x32, 2 - threshold_16x16,
513 // 3 - vbp_threshold_8x8. vbp_threshold_8x8 (to split to 4x4 partition) is
514 // currently only used on key frame.
set_vbp_thresholds(VP9_COMP * cpi,int64_t thresholds[],int q,int content_state)515 static void set_vbp_thresholds(VP9_COMP *cpi, int64_t thresholds[], int q,
516 int content_state) {
517 VP9_COMMON *const cm = &cpi->common;
518 const int is_key_frame = (cm->frame_type == KEY_FRAME);
519 const int threshold_multiplier = is_key_frame ? 20 : 1;
520 int64_t threshold_base =
521 (int64_t)(threshold_multiplier * cpi->y_dequant[q][1]);
522
523 if (is_key_frame) {
524 thresholds[0] = threshold_base;
525 thresholds[1] = threshold_base >> 2;
526 thresholds[2] = threshold_base >> 2;
527 thresholds[3] = threshold_base << 2;
528 } else {
529 // Increase base variance threshold based on estimated noise level.
530 if (cpi->noise_estimate.enabled && cm->width >= 640 && cm->height >= 480) {
531 NOISE_LEVEL noise_level =
532 vp9_noise_estimate_extract_level(&cpi->noise_estimate);
533 if (noise_level == kHigh)
534 threshold_base = 3 * threshold_base;
535 else if (noise_level == kMedium)
536 threshold_base = threshold_base << 1;
537 else if (noise_level < kLow)
538 threshold_base = (7 * threshold_base) >> 3;
539 }
540 #if CONFIG_VP9_TEMPORAL_DENOISING
541 if (cpi->oxcf.noise_sensitivity > 0 && denoise_svc(cpi) &&
542 cpi->oxcf.speed > 5 && cpi->denoiser.denoising_level >= kDenLow)
543 threshold_base =
544 vp9_scale_part_thresh(threshold_base, cpi->denoiser.denoising_level,
545 content_state, cpi->svc.temporal_layer_id);
546 else
547 threshold_base =
548 scale_part_thresh_sumdiff(threshold_base, cpi->oxcf.speed, cm->width,
549 cm->height, content_state);
550 #else
551 // Increase base variance threshold based on content_state/sum_diff level.
552 threshold_base = scale_part_thresh_sumdiff(
553 threshold_base, cpi->oxcf.speed, cm->width, cm->height, content_state);
554 #endif
555 thresholds[0] = threshold_base;
556 thresholds[2] = threshold_base << cpi->oxcf.speed;
557 if (cm->width <= 352 && cm->height <= 288) {
558 thresholds[0] = threshold_base >> 3;
559 thresholds[1] = threshold_base >> 1;
560 thresholds[2] = threshold_base << 3;
561 } else if (cm->width < 1280 && cm->height < 720) {
562 thresholds[1] = (5 * threshold_base) >> 2;
563 } else if (cm->width < 1920 && cm->height < 1080) {
564 thresholds[1] = threshold_base << 1;
565 } else {
566 thresholds[1] = (5 * threshold_base) >> 1;
567 }
568 }
569 }
570
vp9_set_variance_partition_thresholds(VP9_COMP * cpi,int q,int content_state)571 void vp9_set_variance_partition_thresholds(VP9_COMP *cpi, int q,
572 int content_state) {
573 VP9_COMMON *const cm = &cpi->common;
574 SPEED_FEATURES *const sf = &cpi->sf;
575 const int is_key_frame = (cm->frame_type == KEY_FRAME);
576 if (sf->partition_search_type != VAR_BASED_PARTITION &&
577 sf->partition_search_type != REFERENCE_PARTITION) {
578 return;
579 } else {
580 set_vbp_thresholds(cpi, cpi->vbp_thresholds, q, content_state);
581 // The thresholds below are not changed locally.
582 if (is_key_frame) {
583 cpi->vbp_threshold_sad = 0;
584 cpi->vbp_threshold_copy = 0;
585 cpi->vbp_bsize_min = BLOCK_8X8;
586 } else {
587 if (cm->width <= 352 && cm->height <= 288)
588 cpi->vbp_threshold_sad = 10;
589 else
590 cpi->vbp_threshold_sad = (cpi->y_dequant[q][1] << 1) > 1000
591 ? (cpi->y_dequant[q][1] << 1)
592 : 1000;
593 cpi->vbp_bsize_min = BLOCK_16X16;
594 if (cm->width <= 352 && cm->height <= 288)
595 cpi->vbp_threshold_copy = 4000;
596 else if (cm->width <= 640 && cm->height <= 360)
597 cpi->vbp_threshold_copy = 8000;
598 else
599 cpi->vbp_threshold_copy = (cpi->y_dequant[q][1] << 3) > 8000
600 ? (cpi->y_dequant[q][1] << 3)
601 : 8000;
602 }
603 cpi->vbp_threshold_minmax = 15 + (q >> 3);
604 }
605 }
606
607 // 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)608 static int compute_minmax_8x8(const uint8_t *s, int sp, const uint8_t *d,
609 int dp, int x16_idx, int y16_idx,
610 #if CONFIG_VP9_HIGHBITDEPTH
611 int highbd_flag,
612 #endif
613 int pixels_wide, int pixels_high) {
614 int k;
615 int minmax_max = 0;
616 int minmax_min = 255;
617 // Loop over the 4 8x8 subblocks.
618 for (k = 0; k < 4; k++) {
619 int x8_idx = x16_idx + ((k & 1) << 3);
620 int y8_idx = y16_idx + ((k >> 1) << 3);
621 int min = 0;
622 int max = 0;
623 if (x8_idx < pixels_wide && y8_idx < pixels_high) {
624 #if CONFIG_VP9_HIGHBITDEPTH
625 if (highbd_flag & YV12_FLAG_HIGHBITDEPTH) {
626 vpx_highbd_minmax_8x8(s + y8_idx * sp + x8_idx, sp,
627 d + y8_idx * dp + x8_idx, dp, &min, &max);
628 } else {
629 vpx_minmax_8x8(s + y8_idx * sp + x8_idx, sp, d + y8_idx * dp + x8_idx,
630 dp, &min, &max);
631 }
632 #else
633 vpx_minmax_8x8(s + y8_idx * sp + x8_idx, sp, d + y8_idx * dp + x8_idx, dp,
634 &min, &max);
635 #endif
636 if ((max - min) > minmax_max) minmax_max = (max - min);
637 if ((max - min) < minmax_min) minmax_min = (max - min);
638 }
639 }
640 return (minmax_max - minmax_min);
641 }
642
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)643 static void fill_variance_4x4avg(const uint8_t *s, int sp, const uint8_t *d,
644 int dp, int x8_idx, int y8_idx, v8x8 *vst,
645 #if CONFIG_VP9_HIGHBITDEPTH
646 int highbd_flag,
647 #endif
648 int pixels_wide, int pixels_high,
649 int is_key_frame) {
650 int k;
651 for (k = 0; k < 4; k++) {
652 int x4_idx = x8_idx + ((k & 1) << 2);
653 int y4_idx = y8_idx + ((k >> 1) << 2);
654 unsigned int sse = 0;
655 int sum = 0;
656 if (x4_idx < pixels_wide && y4_idx < pixels_high) {
657 int s_avg;
658 int d_avg = 128;
659 #if CONFIG_VP9_HIGHBITDEPTH
660 if (highbd_flag & YV12_FLAG_HIGHBITDEPTH) {
661 s_avg = vpx_highbd_avg_4x4(s + y4_idx * sp + x4_idx, sp);
662 if (!is_key_frame)
663 d_avg = vpx_highbd_avg_4x4(d + y4_idx * dp + x4_idx, dp);
664 } else {
665 s_avg = vpx_avg_4x4(s + y4_idx * sp + x4_idx, sp);
666 if (!is_key_frame) d_avg = vpx_avg_4x4(d + y4_idx * dp + x4_idx, dp);
667 }
668 #else
669 s_avg = vpx_avg_4x4(s + y4_idx * sp + x4_idx, sp);
670 if (!is_key_frame) d_avg = vpx_avg_4x4(d + y4_idx * dp + x4_idx, dp);
671 #endif
672 sum = s_avg - d_avg;
673 sse = sum * sum;
674 }
675 fill_variance(sse, sum, 0, &vst->split[k].part_variances.none);
676 }
677 }
678
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)679 static void fill_variance_8x8avg(const uint8_t *s, int sp, const uint8_t *d,
680 int dp, int x16_idx, int y16_idx, v16x16 *vst,
681 #if CONFIG_VP9_HIGHBITDEPTH
682 int highbd_flag,
683 #endif
684 int pixels_wide, int pixels_high,
685 int is_key_frame) {
686 int k;
687 for (k = 0; k < 4; k++) {
688 int x8_idx = x16_idx + ((k & 1) << 3);
689 int y8_idx = y16_idx + ((k >> 1) << 3);
690 unsigned int sse = 0;
691 int sum = 0;
692 if (x8_idx < pixels_wide && y8_idx < pixels_high) {
693 int s_avg;
694 int d_avg = 128;
695 #if CONFIG_VP9_HIGHBITDEPTH
696 if (highbd_flag & YV12_FLAG_HIGHBITDEPTH) {
697 s_avg = vpx_highbd_avg_8x8(s + y8_idx * sp + x8_idx, sp);
698 if (!is_key_frame)
699 d_avg = vpx_highbd_avg_8x8(d + y8_idx * dp + x8_idx, dp);
700 } else {
701 s_avg = vpx_avg_8x8(s + y8_idx * sp + x8_idx, sp);
702 if (!is_key_frame) d_avg = vpx_avg_8x8(d + y8_idx * dp + x8_idx, dp);
703 }
704 #else
705 s_avg = vpx_avg_8x8(s + y8_idx * sp + x8_idx, sp);
706 if (!is_key_frame) d_avg = vpx_avg_8x8(d + y8_idx * dp + x8_idx, dp);
707 #endif
708 sum = s_avg - d_avg;
709 sse = sum * sum;
710 }
711 fill_variance(sse, sum, 0, &vst->split[k].part_variances.none);
712 }
713 }
714
715 // Check if most of the superblock is skin content, and if so, force split to
716 // 32x32, and set x->sb_is_skin for use in mode selection.
skin_sb_split(VP9_COMP * cpi,MACROBLOCK * x,const int low_res,int mi_row,int mi_col,int * force_split)717 static int skin_sb_split(VP9_COMP *cpi, MACROBLOCK *x, const int low_res,
718 int mi_row, int mi_col, int *force_split) {
719 VP9_COMMON *const cm = &cpi->common;
720 #if CONFIG_VP9_HIGHBITDEPTH
721 if (cm->use_highbitdepth) return 0;
722 #endif
723 // Avoid checking superblocks on/near boundary and avoid low resolutions.
724 // Note superblock may still pick 64X64 if y_sad is very small
725 // (i.e., y_sad < cpi->vbp_threshold_sad) below. For now leave this as is.
726 if (!low_res && (mi_col >= 8 && mi_col + 8 < cm->mi_cols && mi_row >= 8 &&
727 mi_row + 8 < cm->mi_rows)) {
728 int num_16x16_skin = 0;
729 int num_16x16_nonskin = 0;
730 uint8_t *ysignal = x->plane[0].src.buf;
731 uint8_t *usignal = x->plane[1].src.buf;
732 uint8_t *vsignal = x->plane[2].src.buf;
733 int sp = x->plane[0].src.stride;
734 int spuv = x->plane[1].src.stride;
735 const int block_index = mi_row * cm->mi_cols + mi_col;
736 const int bw = num_8x8_blocks_wide_lookup[BLOCK_64X64];
737 const int bh = num_8x8_blocks_high_lookup[BLOCK_64X64];
738 const int xmis = VPXMIN(cm->mi_cols - mi_col, bw);
739 const int ymis = VPXMIN(cm->mi_rows - mi_row, bh);
740 // Loop through the 16x16 sub-blocks.
741 int i, j;
742 for (i = 0; i < ymis; i += 2) {
743 for (j = 0; j < xmis; j += 2) {
744 int bl_index = block_index + i * cm->mi_cols + j;
745 int bl_index1 = bl_index + 1;
746 int bl_index2 = bl_index + cm->mi_cols;
747 int bl_index3 = bl_index2 + 1;
748 int consec_zeromv =
749 VPXMIN(cpi->consec_zero_mv[bl_index],
750 VPXMIN(cpi->consec_zero_mv[bl_index1],
751 VPXMIN(cpi->consec_zero_mv[bl_index2],
752 cpi->consec_zero_mv[bl_index3])));
753 int is_skin = vp9_compute_skin_block(
754 ysignal, usignal, vsignal, sp, spuv, BLOCK_16X16, consec_zeromv, 0);
755 num_16x16_skin += is_skin;
756 num_16x16_nonskin += (1 - is_skin);
757 if (num_16x16_nonskin > 3) {
758 // Exit loop if at least 4 of the 16x16 blocks are not skin.
759 i = ymis;
760 break;
761 }
762 ysignal += 16;
763 usignal += 8;
764 vsignal += 8;
765 }
766 ysignal += (sp << 4) - 64;
767 usignal += (spuv << 3) - 32;
768 vsignal += (spuv << 3) - 32;
769 }
770 if (num_16x16_skin > 12) {
771 *force_split = 1;
772 return 1;
773 }
774 }
775 return 0;
776 }
777
set_low_temp_var_flag(VP9_COMP * cpi,MACROBLOCK * x,MACROBLOCKD * xd,v64x64 * vt,int64_t thresholds[],MV_REFERENCE_FRAME ref_frame_partition,int mi_col,int mi_row)778 static void set_low_temp_var_flag(VP9_COMP *cpi, MACROBLOCK *x, MACROBLOCKD *xd,
779 v64x64 *vt, int64_t thresholds[],
780 MV_REFERENCE_FRAME ref_frame_partition,
781 int mi_col, int mi_row) {
782 int i, j;
783 VP9_COMMON *const cm = &cpi->common;
784 const int mv_thr = cm->width > 640 ? 8 : 4;
785 // Check temporal variance for bsize >= 16x16, if LAST_FRAME was selected and
786 // int_pro mv is small. If the temporal variance is small set the flag
787 // variance_low for the block. The variance threshold can be adjusted, the
788 // higher the more aggressive.
789 if (ref_frame_partition == LAST_FRAME &&
790 (cpi->sf.short_circuit_low_temp_var == 1 ||
791 (xd->mi[0]->mv[0].as_mv.col < mv_thr &&
792 xd->mi[0]->mv[0].as_mv.col > -mv_thr &&
793 xd->mi[0]->mv[0].as_mv.row < mv_thr &&
794 xd->mi[0]->mv[0].as_mv.row > -mv_thr))) {
795 if (xd->mi[0]->sb_type == BLOCK_64X64) {
796 if ((vt->part_variances).none.variance < (thresholds[0] >> 1))
797 x->variance_low[0] = 1;
798 } else if (xd->mi[0]->sb_type == BLOCK_64X32) {
799 for (i = 0; i < 2; i++) {
800 if (vt->part_variances.horz[i].variance < (thresholds[0] >> 2))
801 x->variance_low[i + 1] = 1;
802 }
803 } else if (xd->mi[0]->sb_type == BLOCK_32X64) {
804 for (i = 0; i < 2; i++) {
805 if (vt->part_variances.vert[i].variance < (thresholds[0] >> 2))
806 x->variance_low[i + 3] = 1;
807 }
808 } else {
809 for (i = 0; i < 4; i++) {
810 const int idx[4][2] = { { 0, 0 }, { 0, 4 }, { 4, 0 }, { 4, 4 } };
811 const int idx_str =
812 cm->mi_stride * (mi_row + idx[i][0]) + mi_col + idx[i][1];
813 MODE_INFO **this_mi = cm->mi_grid_visible + idx_str;
814
815 if (cm->mi_cols <= mi_col + idx[i][1] ||
816 cm->mi_rows <= mi_row + idx[i][0])
817 continue;
818
819 if ((*this_mi)->sb_type == BLOCK_32X32) {
820 int64_t threshold_32x32 = (cpi->sf.short_circuit_low_temp_var == 1 ||
821 cpi->sf.short_circuit_low_temp_var == 3)
822 ? ((5 * thresholds[1]) >> 3)
823 : (thresholds[1] >> 1);
824 if (vt->split[i].part_variances.none.variance < threshold_32x32)
825 x->variance_low[i + 5] = 1;
826 } else if (cpi->sf.short_circuit_low_temp_var >= 2) {
827 // For 32x16 and 16x32 blocks, the flag is set on each 16x16 block
828 // inside.
829 if ((*this_mi)->sb_type == BLOCK_16X16 ||
830 (*this_mi)->sb_type == BLOCK_32X16 ||
831 (*this_mi)->sb_type == BLOCK_16X32) {
832 for (j = 0; j < 4; j++) {
833 if (vt->split[i].split[j].part_variances.none.variance <
834 (thresholds[2] >> 8))
835 x->variance_low[(i << 2) + j + 9] = 1;
836 }
837 }
838 }
839 }
840 }
841 }
842 }
843
copy_partitioning_helper(VP9_COMP * cpi,MACROBLOCK * x,MACROBLOCKD * xd,BLOCK_SIZE bsize,int mi_row,int mi_col)844 static void copy_partitioning_helper(VP9_COMP *cpi, MACROBLOCK *x,
845 MACROBLOCKD *xd, BLOCK_SIZE bsize,
846 int mi_row, int mi_col) {
847 VP9_COMMON *const cm = &cpi->common;
848 BLOCK_SIZE *prev_part = cpi->prev_partition;
849 int start_pos = mi_row * cm->mi_stride + mi_col;
850
851 const int bsl = b_width_log2_lookup[bsize];
852 const int bs = (1 << bsl) / 4;
853 BLOCK_SIZE subsize;
854 PARTITION_TYPE partition;
855
856 if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols) return;
857
858 partition = partition_lookup[bsl][prev_part[start_pos]];
859 subsize = get_subsize(bsize, partition);
860
861 if (subsize < BLOCK_8X8) {
862 set_block_size(cpi, x, xd, mi_row, mi_col, bsize);
863 } else {
864 switch (partition) {
865 case PARTITION_NONE:
866 set_block_size(cpi, x, xd, mi_row, mi_col, bsize);
867 break;
868 case PARTITION_HORZ:
869 set_block_size(cpi, x, xd, mi_row, mi_col, subsize);
870 set_block_size(cpi, x, xd, mi_row + bs, mi_col, subsize);
871 break;
872 case PARTITION_VERT:
873 set_block_size(cpi, x, xd, mi_row, mi_col, subsize);
874 set_block_size(cpi, x, xd, mi_row, mi_col + bs, subsize);
875 break;
876 case PARTITION_SPLIT:
877 copy_partitioning_helper(cpi, x, xd, subsize, mi_row, mi_col);
878 copy_partitioning_helper(cpi, x, xd, subsize, mi_row + bs, mi_col);
879 copy_partitioning_helper(cpi, x, xd, subsize, mi_row, mi_col + bs);
880 copy_partitioning_helper(cpi, x, xd, subsize, mi_row + bs, mi_col + bs);
881 break;
882 default: assert(0);
883 }
884 }
885 }
886
copy_partitioning(VP9_COMP * cpi,MACROBLOCK * x,MACROBLOCKD * xd,int mi_row,int mi_col,int segment_id,int sb_offset)887 static int copy_partitioning(VP9_COMP *cpi, MACROBLOCK *x, MACROBLOCKD *xd,
888 int mi_row, int mi_col, int segment_id,
889 int sb_offset) {
890 int svc_copy_allowed = 1;
891 int frames_since_key_thresh = 1;
892 if (cpi->use_svc) {
893 // For SVC, don't allow copy if base spatial layer is key frame, or if
894 // frame is not a temporal enhancement layer frame.
895 int layer = LAYER_IDS_TO_IDX(0, cpi->svc.temporal_layer_id,
896 cpi->svc.number_temporal_layers);
897 const LAYER_CONTEXT *lc = &cpi->svc.layer_context[layer];
898 if (lc->is_key_frame ||
899 (cpi->svc.temporal_layer_id != cpi->svc.number_temporal_layers - 1 &&
900 cpi->svc.number_temporal_layers > 1))
901 svc_copy_allowed = 0;
902 frames_since_key_thresh = cpi->svc.number_spatial_layers << 1;
903 }
904 if (cpi->rc.frames_since_key > frames_since_key_thresh && svc_copy_allowed &&
905 !cpi->resize_pending && segment_id == CR_SEGMENT_ID_BASE &&
906 cpi->prev_segment_id[sb_offset] == CR_SEGMENT_ID_BASE &&
907 cpi->copied_frame_cnt[sb_offset] < cpi->max_copied_frame) {
908 if (cpi->prev_partition != NULL) {
909 copy_partitioning_helper(cpi, x, xd, BLOCK_64X64, mi_row, mi_col);
910 cpi->copied_frame_cnt[sb_offset] += 1;
911 memcpy(x->variance_low, &(cpi->prev_variance_low[sb_offset * 25]),
912 sizeof(x->variance_low));
913 return 1;
914 }
915 }
916
917 return 0;
918 }
919
update_prev_partition(VP9_COMP * cpi,BLOCK_SIZE bsize,int mi_row,int mi_col)920 static void update_prev_partition(VP9_COMP *cpi, BLOCK_SIZE bsize, int mi_row,
921 int mi_col) {
922 VP9_COMMON *const cm = &cpi->common;
923 BLOCK_SIZE *prev_part = cpi->prev_partition;
924 int start_pos = mi_row * cm->mi_stride + mi_col;
925 const int bsl = b_width_log2_lookup[bsize];
926 const int bs = (1 << bsl) / 4;
927 BLOCK_SIZE subsize;
928 PARTITION_TYPE partition;
929 const MODE_INFO *mi = NULL;
930
931 if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols) return;
932
933 mi = cm->mi_grid_visible[start_pos];
934 partition = partition_lookup[bsl][mi->sb_type];
935 subsize = get_subsize(bsize, partition);
936 if (subsize < BLOCK_8X8) {
937 prev_part[start_pos] = bsize;
938 } else {
939 switch (partition) {
940 case PARTITION_NONE: prev_part[start_pos] = bsize; break;
941 case PARTITION_HORZ:
942 prev_part[start_pos] = subsize;
943 if (mi_row + bs < cm->mi_rows)
944 prev_part[start_pos + bs * cm->mi_stride] = subsize;
945 break;
946 case PARTITION_VERT:
947 prev_part[start_pos] = subsize;
948 if (mi_col + bs < cm->mi_cols) prev_part[start_pos + bs] = subsize;
949 break;
950 case PARTITION_SPLIT:
951 update_prev_partition(cpi, subsize, mi_row, mi_col);
952 update_prev_partition(cpi, subsize, mi_row + bs, mi_col);
953 update_prev_partition(cpi, subsize, mi_row, mi_col + bs);
954 update_prev_partition(cpi, subsize, mi_row + bs, mi_col + bs);
955 break;
956 default: assert(0);
957 }
958 }
959 }
960
chroma_check(VP9_COMP * cpi,MACROBLOCK * x,int bsize,unsigned int y_sad,int is_key_frame)961 static void chroma_check(VP9_COMP *cpi, MACROBLOCK *x, int bsize,
962 unsigned int y_sad, int is_key_frame) {
963 int i;
964 MACROBLOCKD *xd = &x->e_mbd;
965
966 if (is_key_frame) return;
967
968 // For speed >= 8, avoid the chroma check if y_sad is above threshold.
969 if (cpi->oxcf.speed >= 8) {
970 if (y_sad > cpi->vbp_thresholds[1] &&
971 (!cpi->noise_estimate.enabled ||
972 vp9_noise_estimate_extract_level(&cpi->noise_estimate) < kMedium))
973 return;
974 }
975
976 for (i = 1; i <= 2; ++i) {
977 unsigned int uv_sad = UINT_MAX;
978 struct macroblock_plane *p = &x->plane[i];
979 struct macroblockd_plane *pd = &xd->plane[i];
980 const BLOCK_SIZE bs = get_plane_block_size(bsize, pd);
981
982 if (bs != BLOCK_INVALID)
983 uv_sad = cpi->fn_ptr[bs].sdf(p->src.buf, p->src.stride, pd->dst.buf,
984 pd->dst.stride);
985
986 // TODO(marpan): Investigate if we should lower this threshold if
987 // superblock is detected as skin.
988 x->color_sensitivity[i - 1] = uv_sad > (y_sad >> 2);
989 }
990 }
991
avg_source_sad(VP9_COMP * cpi,MACROBLOCK * x,int shift,int sb_offset)992 static void avg_source_sad(VP9_COMP *cpi, MACROBLOCK *x, int shift,
993 int sb_offset) {
994 unsigned int tmp_sse;
995 uint64_t tmp_sad;
996 unsigned int tmp_variance;
997 const BLOCK_SIZE bsize = BLOCK_64X64;
998 uint8_t *src_y = cpi->Source->y_buffer;
999 int src_ystride = cpi->Source->y_stride;
1000 uint8_t *last_src_y = cpi->Last_Source->y_buffer;
1001 int last_src_ystride = cpi->Last_Source->y_stride;
1002 uint64_t avg_source_sad_threshold = 10000;
1003 uint64_t avg_source_sad_threshold2 = 12000;
1004 #if CONFIG_VP9_HIGHBITDEPTH
1005 if (cpi->common.use_highbitdepth) return;
1006 #endif
1007 src_y += shift;
1008 last_src_y += shift;
1009 tmp_sad =
1010 cpi->fn_ptr[bsize].sdf(src_y, src_ystride, last_src_y, last_src_ystride);
1011 tmp_variance = vpx_variance64x64(src_y, src_ystride, last_src_y,
1012 last_src_ystride, &tmp_sse);
1013 // Note: tmp_sse - tmp_variance = ((sum * sum) >> 12)
1014 if (tmp_sad < avg_source_sad_threshold)
1015 x->content_state_sb = ((tmp_sse - tmp_variance) < 25) ? kLowSadLowSumdiff
1016 : kLowSadHighSumdiff;
1017 else
1018 x->content_state_sb = ((tmp_sse - tmp_variance) < 25) ? kHighSadLowSumdiff
1019 : kHighSadHighSumdiff;
1020
1021 // Detect large lighting change.
1022 if (tmp_variance < (tmp_sse >> 3) && (tmp_sse - tmp_variance) > 10000)
1023 x->content_state_sb = kLowVarHighSumdiff;
1024
1025 if (cpi->content_state_sb_fd != NULL) {
1026 if (tmp_sad < avg_source_sad_threshold2) {
1027 // Cap the increment to 255.
1028 if (cpi->content_state_sb_fd[sb_offset] < 255)
1029 cpi->content_state_sb_fd[sb_offset]++;
1030 } else {
1031 cpi->content_state_sb_fd[sb_offset] = 0;
1032 }
1033 }
1034 return;
1035 }
1036
1037 // This function chooses partitioning based on the variance between source and
1038 // 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)1039 static int choose_partitioning(VP9_COMP *cpi, const TileInfo *const tile,
1040 MACROBLOCK *x, int mi_row, int mi_col) {
1041 VP9_COMMON *const cm = &cpi->common;
1042 MACROBLOCKD *xd = &x->e_mbd;
1043 int i, j, k, m;
1044 v64x64 vt;
1045 v16x16 vt2[16];
1046 int force_split[21];
1047 int avg_32x32;
1048 int max_var_32x32 = 0;
1049 int min_var_32x32 = INT_MAX;
1050 int var_32x32;
1051 int avg_16x16[4];
1052 int maxvar_16x16[4];
1053 int minvar_16x16[4];
1054 int64_t threshold_4x4avg;
1055 NOISE_LEVEL noise_level = kLow;
1056 int content_state = 0;
1057 uint8_t *s;
1058 const uint8_t *d;
1059 int sp;
1060 int dp;
1061 unsigned int y_sad = UINT_MAX;
1062 BLOCK_SIZE bsize = BLOCK_64X64;
1063 // Ref frame used in partitioning.
1064 MV_REFERENCE_FRAME ref_frame_partition = LAST_FRAME;
1065 int pixels_wide = 64, pixels_high = 64;
1066 int64_t thresholds[4] = { cpi->vbp_thresholds[0], cpi->vbp_thresholds[1],
1067 cpi->vbp_thresholds[2], cpi->vbp_thresholds[3] };
1068
1069 // For the variance computation under SVC mode, we treat the frame as key if
1070 // the reference (base layer frame) is key frame (i.e., is_key_frame == 1).
1071 const int is_key_frame =
1072 (cm->frame_type == KEY_FRAME ||
1073 (is_one_pass_cbr_svc(cpi) &&
1074 cpi->svc.layer_context[cpi->svc.temporal_layer_id].is_key_frame));
1075 // Always use 4x4 partition for key frame.
1076 const int use_4x4_partition = cm->frame_type == KEY_FRAME;
1077 const int low_res = (cm->width <= 352 && cm->height <= 288);
1078 int variance4x4downsample[16];
1079 int segment_id;
1080 int sb_offset = (cm->mi_stride >> 3) * (mi_row >> 3) + (mi_col >> 3);
1081
1082 set_offsets(cpi, tile, x, mi_row, mi_col, BLOCK_64X64);
1083 segment_id = xd->mi[0]->segment_id;
1084
1085 if (cpi->sf.use_source_sad && !is_key_frame) {
1086 int sb_offset2 = ((cm->mi_cols + 7) >> 3) * (mi_row >> 3) + (mi_col >> 3);
1087 content_state = x->content_state_sb;
1088 x->skip_low_source_sad = (content_state == kLowSadLowSumdiff ||
1089 content_state == kLowSadHighSumdiff)
1090 ? 1
1091 : 0;
1092 x->lowvar_highsumdiff = (content_state == kLowVarHighSumdiff) ? 1 : 0;
1093 if (cpi->content_state_sb_fd != NULL)
1094 x->last_sb_high_content = cpi->content_state_sb_fd[sb_offset2];
1095 // If source_sad is low copy the partition without computing the y_sad.
1096 if (x->skip_low_source_sad && cpi->sf.copy_partition_flag &&
1097 copy_partitioning(cpi, x, xd, mi_row, mi_col, segment_id, sb_offset)) {
1098 return 0;
1099 }
1100 }
1101
1102 if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ && cm->seg.enabled &&
1103 cyclic_refresh_segment_id_boosted(segment_id)) {
1104 int q = vp9_get_qindex(&cm->seg, segment_id, cm->base_qindex);
1105 set_vbp_thresholds(cpi, thresholds, q, content_state);
1106 } else {
1107 set_vbp_thresholds(cpi, thresholds, cm->base_qindex, content_state);
1108 }
1109
1110 // For non keyframes, disable 4x4 average for low resolution when speed = 8
1111 threshold_4x4avg = (cpi->oxcf.speed < 8) ? thresholds[1] << 1 : INT64_MAX;
1112
1113 memset(x->variance_low, 0, sizeof(x->variance_low));
1114
1115 if (xd->mb_to_right_edge < 0) pixels_wide += (xd->mb_to_right_edge >> 3);
1116 if (xd->mb_to_bottom_edge < 0) pixels_high += (xd->mb_to_bottom_edge >> 3);
1117
1118 s = x->plane[0].src.buf;
1119 sp = x->plane[0].src.stride;
1120
1121 // Index for force_split: 0 for 64x64, 1-4 for 32x32 blocks,
1122 // 5-20 for the 16x16 blocks.
1123 force_split[0] = 0;
1124
1125 if (!is_key_frame) {
1126 // In the case of spatial/temporal scalable coding, the assumption here is
1127 // that the temporal reference frame will always be of type LAST_FRAME.
1128 // TODO(marpan): If that assumption is broken, we need to revisit this code.
1129 MODE_INFO *mi = xd->mi[0];
1130 YV12_BUFFER_CONFIG *yv12 = get_ref_frame_buffer(cpi, LAST_FRAME);
1131
1132 const YV12_BUFFER_CONFIG *yv12_g = NULL;
1133 unsigned int y_sad_g, y_sad_thr, y_sad_last;
1134 bsize = BLOCK_32X32 + (mi_col + 4 < cm->mi_cols) * 2 +
1135 (mi_row + 4 < cm->mi_rows);
1136
1137 assert(yv12 != NULL);
1138
1139 if (!(is_one_pass_cbr_svc(cpi) && cpi->svc.spatial_layer_id)) {
1140 // For now, GOLDEN will not be used for non-zero spatial layers, since
1141 // it may not be a temporal reference.
1142 yv12_g = get_ref_frame_buffer(cpi, GOLDEN_FRAME);
1143 }
1144
1145 // Only compute y_sad_g (sad for golden reference) for speed < 8.
1146 if (cpi->oxcf.speed < 8 && yv12_g && yv12_g != yv12 &&
1147 (cpi->ref_frame_flags & VP9_GOLD_FLAG)) {
1148 vp9_setup_pre_planes(xd, 0, yv12_g, mi_row, mi_col,
1149 &cm->frame_refs[GOLDEN_FRAME - 1].sf);
1150 y_sad_g = cpi->fn_ptr[bsize].sdf(
1151 x->plane[0].src.buf, x->plane[0].src.stride, xd->plane[0].pre[0].buf,
1152 xd->plane[0].pre[0].stride);
1153 } else {
1154 y_sad_g = UINT_MAX;
1155 }
1156
1157 if (cpi->oxcf.lag_in_frames > 0 && cpi->oxcf.rc_mode == VPX_VBR &&
1158 cpi->rc.is_src_frame_alt_ref) {
1159 yv12 = get_ref_frame_buffer(cpi, ALTREF_FRAME);
1160 vp9_setup_pre_planes(xd, 0, yv12, mi_row, mi_col,
1161 &cm->frame_refs[ALTREF_FRAME - 1].sf);
1162 mi->ref_frame[0] = ALTREF_FRAME;
1163 y_sad_g = UINT_MAX;
1164 } else {
1165 vp9_setup_pre_planes(xd, 0, yv12, mi_row, mi_col,
1166 &cm->frame_refs[LAST_FRAME - 1].sf);
1167 mi->ref_frame[0] = LAST_FRAME;
1168 }
1169 mi->ref_frame[1] = NONE;
1170 mi->sb_type = BLOCK_64X64;
1171 mi->mv[0].as_int = 0;
1172 mi->interp_filter = BILINEAR;
1173
1174 if (cpi->oxcf.speed >= 8 && !low_res)
1175 y_sad = cpi->fn_ptr[bsize].sdf(
1176 x->plane[0].src.buf, x->plane[0].src.stride, xd->plane[0].pre[0].buf,
1177 xd->plane[0].pre[0].stride);
1178 else
1179 y_sad = vp9_int_pro_motion_estimation(cpi, x, bsize, mi_row, mi_col);
1180
1181 y_sad_last = y_sad;
1182 // Pick ref frame for partitioning, bias last frame when y_sad_g and y_sad
1183 // are close if short_circuit_low_temp_var is on.
1184 y_sad_thr = cpi->sf.short_circuit_low_temp_var ? (y_sad * 7) >> 3 : y_sad;
1185 if (y_sad_g < y_sad_thr) {
1186 vp9_setup_pre_planes(xd, 0, yv12_g, mi_row, mi_col,
1187 &cm->frame_refs[GOLDEN_FRAME - 1].sf);
1188 mi->ref_frame[0] = GOLDEN_FRAME;
1189 mi->mv[0].as_int = 0;
1190 y_sad = y_sad_g;
1191 ref_frame_partition = GOLDEN_FRAME;
1192 } else {
1193 x->pred_mv[LAST_FRAME] = mi->mv[0].as_mv;
1194 ref_frame_partition = LAST_FRAME;
1195 }
1196
1197 set_ref_ptrs(cm, xd, mi->ref_frame[0], mi->ref_frame[1]);
1198 vp9_build_inter_predictors_sb(xd, mi_row, mi_col, BLOCK_64X64);
1199
1200 x->sb_is_skin = skin_sb_split(cpi, x, low_res, mi_row, mi_col, force_split);
1201
1202 d = xd->plane[0].dst.buf;
1203 dp = xd->plane[0].dst.stride;
1204
1205 // If the y_sad is very small, take 64x64 as partition and exit.
1206 // Don't check on boosted segment for now, as 64x64 is suppressed there.
1207 if (segment_id == CR_SEGMENT_ID_BASE && y_sad < cpi->vbp_threshold_sad) {
1208 const int block_width = num_8x8_blocks_wide_lookup[BLOCK_64X64];
1209 const int block_height = num_8x8_blocks_high_lookup[BLOCK_64X64];
1210 if (mi_col + block_width / 2 < cm->mi_cols &&
1211 mi_row + block_height / 2 < cm->mi_rows) {
1212 set_block_size(cpi, x, xd, mi_row, mi_col, BLOCK_64X64);
1213 x->variance_low[0] = 1;
1214 chroma_check(cpi, x, bsize, y_sad, is_key_frame);
1215 return 0;
1216 }
1217 }
1218
1219 // If the y_sad is small enough, copy the partition of the superblock in the
1220 // last frame to current frame only if the last frame is not a keyframe.
1221 // Stop the copy every cpi->max_copied_frame to refresh the partition.
1222 // TODO(jianj) : tune the threshold.
1223 if (cpi->sf.copy_partition_flag && y_sad_last < cpi->vbp_threshold_copy &&
1224 copy_partitioning(cpi, x, xd, mi_row, mi_col, segment_id, sb_offset)) {
1225 chroma_check(cpi, x, bsize, y_sad, is_key_frame);
1226 return 0;
1227 }
1228 } else {
1229 d = VP9_VAR_OFFS;
1230 dp = 0;
1231 #if CONFIG_VP9_HIGHBITDEPTH
1232 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
1233 switch (xd->bd) {
1234 case 10: d = CONVERT_TO_BYTEPTR(VP9_HIGH_VAR_OFFS_10); break;
1235 case 12: d = CONVERT_TO_BYTEPTR(VP9_HIGH_VAR_OFFS_12); break;
1236 case 8:
1237 default: d = CONVERT_TO_BYTEPTR(VP9_HIGH_VAR_OFFS_8); break;
1238 }
1239 }
1240 #endif // CONFIG_VP9_HIGHBITDEPTH
1241 }
1242
1243 // Fill in the entire tree of 8x8 (or 4x4 under some conditions) variances
1244 // for splits.
1245 for (i = 0; i < 4; i++) {
1246 const int x32_idx = ((i & 1) << 5);
1247 const int y32_idx = ((i >> 1) << 5);
1248 const int i2 = i << 2;
1249 force_split[i + 1] = 0;
1250 avg_16x16[i] = 0;
1251 maxvar_16x16[i] = 0;
1252 minvar_16x16[i] = INT_MAX;
1253 for (j = 0; j < 4; j++) {
1254 const int x16_idx = x32_idx + ((j & 1) << 4);
1255 const int y16_idx = y32_idx + ((j >> 1) << 4);
1256 const int split_index = 5 + i2 + j;
1257 v16x16 *vst = &vt.split[i].split[j];
1258 force_split[split_index] = 0;
1259 variance4x4downsample[i2 + j] = 0;
1260 if (!is_key_frame) {
1261 fill_variance_8x8avg(s, sp, d, dp, x16_idx, y16_idx, vst,
1262 #if CONFIG_VP9_HIGHBITDEPTH
1263 xd->cur_buf->flags,
1264 #endif
1265 pixels_wide, pixels_high, is_key_frame);
1266 fill_variance_tree(&vt.split[i].split[j], BLOCK_16X16);
1267 get_variance(&vt.split[i].split[j].part_variances.none);
1268 avg_16x16[i] += vt.split[i].split[j].part_variances.none.variance;
1269 if (vt.split[i].split[j].part_variances.none.variance < minvar_16x16[i])
1270 minvar_16x16[i] = vt.split[i].split[j].part_variances.none.variance;
1271 if (vt.split[i].split[j].part_variances.none.variance > maxvar_16x16[i])
1272 maxvar_16x16[i] = vt.split[i].split[j].part_variances.none.variance;
1273 if (vt.split[i].split[j].part_variances.none.variance > thresholds[2]) {
1274 // 16X16 variance is above threshold for split, so force split to 8x8
1275 // for this 16x16 block (this also forces splits for upper levels).
1276 force_split[split_index] = 1;
1277 force_split[i + 1] = 1;
1278 force_split[0] = 1;
1279 } else if (cpi->oxcf.speed < 8 &&
1280 vt.split[i].split[j].part_variances.none.variance >
1281 thresholds[1] &&
1282 !cyclic_refresh_segment_id_boosted(segment_id)) {
1283 // We have some nominal amount of 16x16 variance (based on average),
1284 // compute the minmax over the 8x8 sub-blocks, and if above threshold,
1285 // force split to 8x8 block for this 16x16 block.
1286 int minmax = compute_minmax_8x8(s, sp, d, dp, x16_idx, y16_idx,
1287 #if CONFIG_VP9_HIGHBITDEPTH
1288 xd->cur_buf->flags,
1289 #endif
1290 pixels_wide, pixels_high);
1291 if (minmax > cpi->vbp_threshold_minmax) {
1292 force_split[split_index] = 1;
1293 force_split[i + 1] = 1;
1294 force_split[0] = 1;
1295 }
1296 }
1297 }
1298 if (is_key_frame || (low_res &&
1299 vt.split[i].split[j].part_variances.none.variance >
1300 threshold_4x4avg)) {
1301 force_split[split_index] = 0;
1302 // Go down to 4x4 down-sampling for variance.
1303 variance4x4downsample[i2 + j] = 1;
1304 for (k = 0; k < 4; k++) {
1305 int x8_idx = x16_idx + ((k & 1) << 3);
1306 int y8_idx = y16_idx + ((k >> 1) << 3);
1307 v8x8 *vst2 = is_key_frame ? &vst->split[k] : &vt2[i2 + j].split[k];
1308 fill_variance_4x4avg(s, sp, d, dp, x8_idx, y8_idx, vst2,
1309 #if CONFIG_VP9_HIGHBITDEPTH
1310 xd->cur_buf->flags,
1311 #endif
1312 pixels_wide, pixels_high, is_key_frame);
1313 }
1314 }
1315 }
1316 }
1317 if (cpi->noise_estimate.enabled)
1318 noise_level = vp9_noise_estimate_extract_level(&cpi->noise_estimate);
1319 // Fill the rest of the variance tree by summing split partition values.
1320 avg_32x32 = 0;
1321 for (i = 0; i < 4; i++) {
1322 const int i2 = i << 2;
1323 for (j = 0; j < 4; j++) {
1324 if (variance4x4downsample[i2 + j] == 1) {
1325 v16x16 *vtemp = (!is_key_frame) ? &vt2[i2 + j] : &vt.split[i].split[j];
1326 for (m = 0; m < 4; m++) fill_variance_tree(&vtemp->split[m], BLOCK_8X8);
1327 fill_variance_tree(vtemp, BLOCK_16X16);
1328 // If variance of this 16x16 block is above the threshold, force block
1329 // to split. This also forces a split on the upper levels.
1330 get_variance(&vtemp->part_variances.none);
1331 if (vtemp->part_variances.none.variance > thresholds[2]) {
1332 force_split[5 + i2 + j] = 1;
1333 force_split[i + 1] = 1;
1334 force_split[0] = 1;
1335 }
1336 }
1337 }
1338 fill_variance_tree(&vt.split[i], BLOCK_32X32);
1339 // If variance of this 32x32 block is above the threshold, or if its above
1340 // (some threshold of) the average variance over the sub-16x16 blocks, then
1341 // force this block to split. This also forces a split on the upper
1342 // (64x64) level.
1343 if (!force_split[i + 1]) {
1344 get_variance(&vt.split[i].part_variances.none);
1345 var_32x32 = vt.split[i].part_variances.none.variance;
1346 max_var_32x32 = VPXMAX(var_32x32, max_var_32x32);
1347 min_var_32x32 = VPXMIN(var_32x32, min_var_32x32);
1348 if (vt.split[i].part_variances.none.variance > thresholds[1] ||
1349 (!is_key_frame &&
1350 vt.split[i].part_variances.none.variance > (thresholds[1] >> 1) &&
1351 vt.split[i].part_variances.none.variance > (avg_16x16[i] >> 1))) {
1352 force_split[i + 1] = 1;
1353 force_split[0] = 1;
1354 } else if (!is_key_frame && noise_level < kLow && cm->height <= 360 &&
1355 (maxvar_16x16[i] - minvar_16x16[i]) > (thresholds[1] >> 1) &&
1356 maxvar_16x16[i] > thresholds[1]) {
1357 force_split[i + 1] = 1;
1358 force_split[0] = 1;
1359 }
1360 avg_32x32 += var_32x32;
1361 }
1362 }
1363 if (!force_split[0]) {
1364 fill_variance_tree(&vt, BLOCK_64X64);
1365 get_variance(&vt.part_variances.none);
1366 // If variance of this 64x64 block is above (some threshold of) the average
1367 // variance over the sub-32x32 blocks, then force this block to split.
1368 // Only checking this for noise level >= medium for now.
1369 if (!is_key_frame && noise_level >= kMedium &&
1370 vt.part_variances.none.variance > (9 * avg_32x32) >> 5)
1371 force_split[0] = 1;
1372 // Else if the maximum 32x32 variance minus the miniumum 32x32 variance in
1373 // a 64x64 block is greater than threshold and the maximum 32x32 variance is
1374 // above a miniumum threshold, then force the split of a 64x64 block
1375 // Only check this for low noise.
1376 else if (!is_key_frame && noise_level < kMedium &&
1377 (max_var_32x32 - min_var_32x32) > 3 * (thresholds[0] >> 3) &&
1378 max_var_32x32 > thresholds[0] >> 1)
1379 force_split[0] = 1;
1380 }
1381
1382 // Now go through the entire structure, splitting every block size until
1383 // we get to one that's got a variance lower than our threshold.
1384 if (mi_col + 8 > cm->mi_cols || mi_row + 8 > cm->mi_rows ||
1385 !set_vt_partitioning(cpi, x, xd, &vt, BLOCK_64X64, mi_row, mi_col,
1386 thresholds[0], BLOCK_16X16, force_split[0])) {
1387 for (i = 0; i < 4; ++i) {
1388 const int x32_idx = ((i & 1) << 2);
1389 const int y32_idx = ((i >> 1) << 2);
1390 const int i2 = i << 2;
1391 if (!set_vt_partitioning(cpi, x, xd, &vt.split[i], BLOCK_32X32,
1392 (mi_row + y32_idx), (mi_col + x32_idx),
1393 thresholds[1], BLOCK_16X16,
1394 force_split[i + 1])) {
1395 for (j = 0; j < 4; ++j) {
1396 const int x16_idx = ((j & 1) << 1);
1397 const int y16_idx = ((j >> 1) << 1);
1398 // For inter frames: if variance4x4downsample[] == 1 for this 16x16
1399 // block, then the variance is based on 4x4 down-sampling, so use vt2
1400 // in set_vt_partioning(), otherwise use vt.
1401 v16x16 *vtemp = (!is_key_frame && variance4x4downsample[i2 + j] == 1)
1402 ? &vt2[i2 + j]
1403 : &vt.split[i].split[j];
1404 if (!set_vt_partitioning(
1405 cpi, x, xd, vtemp, BLOCK_16X16, mi_row + y32_idx + y16_idx,
1406 mi_col + x32_idx + x16_idx, thresholds[2], cpi->vbp_bsize_min,
1407 force_split[5 + i2 + j])) {
1408 for (k = 0; k < 4; ++k) {
1409 const int x8_idx = (k & 1);
1410 const int y8_idx = (k >> 1);
1411 if (use_4x4_partition) {
1412 if (!set_vt_partitioning(cpi, x, xd, &vtemp->split[k],
1413 BLOCK_8X8,
1414 mi_row + y32_idx + y16_idx + y8_idx,
1415 mi_col + x32_idx + x16_idx + x8_idx,
1416 thresholds[3], BLOCK_8X8, 0)) {
1417 set_block_size(
1418 cpi, x, xd, (mi_row + y32_idx + y16_idx + y8_idx),
1419 (mi_col + x32_idx + x16_idx + x8_idx), BLOCK_4X4);
1420 }
1421 } else {
1422 set_block_size(
1423 cpi, x, xd, (mi_row + y32_idx + y16_idx + y8_idx),
1424 (mi_col + x32_idx + x16_idx + x8_idx), BLOCK_8X8);
1425 }
1426 }
1427 }
1428 }
1429 }
1430 }
1431 }
1432
1433 if (cm->frame_type != KEY_FRAME && cpi->sf.copy_partition_flag) {
1434 update_prev_partition(cpi, BLOCK_64X64, mi_row, mi_col);
1435 cpi->prev_segment_id[sb_offset] = segment_id;
1436 memcpy(&(cpi->prev_variance_low[sb_offset * 25]), x->variance_low,
1437 sizeof(x->variance_low));
1438 // Reset the counter for copy partitioning
1439 if (cpi->copied_frame_cnt[sb_offset] == cpi->max_copied_frame)
1440 cpi->copied_frame_cnt[sb_offset] = 0;
1441 }
1442
1443 if (cpi->sf.short_circuit_low_temp_var) {
1444 set_low_temp_var_flag(cpi, x, xd, &vt, thresholds, ref_frame_partition,
1445 mi_col, mi_row);
1446 }
1447
1448 chroma_check(cpi, x, bsize, y_sad, is_key_frame);
1449 return 0;
1450 }
1451
update_state(VP9_COMP * cpi,ThreadData * td,PICK_MODE_CONTEXT * ctx,int mi_row,int mi_col,BLOCK_SIZE bsize,int output_enabled)1452 static void update_state(VP9_COMP *cpi, ThreadData *td, PICK_MODE_CONTEXT *ctx,
1453 int mi_row, int mi_col, BLOCK_SIZE bsize,
1454 int output_enabled) {
1455 int i, x_idx, y;
1456 VP9_COMMON *const cm = &cpi->common;
1457 RD_COUNTS *const rdc = &td->rd_counts;
1458 MACROBLOCK *const x = &td->mb;
1459 MACROBLOCKD *const xd = &x->e_mbd;
1460 struct macroblock_plane *const p = x->plane;
1461 struct macroblockd_plane *const pd = xd->plane;
1462 MODE_INFO *mi = &ctx->mic;
1463 MODE_INFO *const xdmi = xd->mi[0];
1464 MODE_INFO *mi_addr = xd->mi[0];
1465 const struct segmentation *const seg = &cm->seg;
1466 const int bw = num_8x8_blocks_wide_lookup[mi->sb_type];
1467 const int bh = num_8x8_blocks_high_lookup[mi->sb_type];
1468 const int x_mis = VPXMIN(bw, cm->mi_cols - mi_col);
1469 const int y_mis = VPXMIN(bh, cm->mi_rows - mi_row);
1470 MV_REF *const frame_mvs = cm->cur_frame->mvs + mi_row * cm->mi_cols + mi_col;
1471 int w, h;
1472
1473 const int mis = cm->mi_stride;
1474 const int mi_width = num_8x8_blocks_wide_lookup[bsize];
1475 const int mi_height = num_8x8_blocks_high_lookup[bsize];
1476 int max_plane;
1477
1478 assert(mi->sb_type == bsize);
1479
1480 *mi_addr = *mi;
1481 *x->mbmi_ext = ctx->mbmi_ext;
1482
1483 // If segmentation in use
1484 if (seg->enabled) {
1485 // For in frame complexity AQ copy the segment id from the segment map.
1486 if (cpi->oxcf.aq_mode == COMPLEXITY_AQ) {
1487 const uint8_t *const map =
1488 seg->update_map ? cpi->segmentation_map : cm->last_frame_seg_map;
1489 mi_addr->segment_id = get_segment_id(cm, map, bsize, mi_row, mi_col);
1490 }
1491 // Else for cyclic refresh mode update the segment map, set the segment id
1492 // and then update the quantizer.
1493 if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ) {
1494 vp9_cyclic_refresh_update_segment(cpi, xd->mi[0], mi_row, mi_col, bsize,
1495 ctx->rate, ctx->dist, x->skip, p);
1496 }
1497 }
1498
1499 max_plane = is_inter_block(xdmi) ? MAX_MB_PLANE : 1;
1500 for (i = 0; i < max_plane; ++i) {
1501 p[i].coeff = ctx->coeff_pbuf[i][1];
1502 p[i].qcoeff = ctx->qcoeff_pbuf[i][1];
1503 pd[i].dqcoeff = ctx->dqcoeff_pbuf[i][1];
1504 p[i].eobs = ctx->eobs_pbuf[i][1];
1505 }
1506
1507 for (i = max_plane; i < MAX_MB_PLANE; ++i) {
1508 p[i].coeff = ctx->coeff_pbuf[i][2];
1509 p[i].qcoeff = ctx->qcoeff_pbuf[i][2];
1510 pd[i].dqcoeff = ctx->dqcoeff_pbuf[i][2];
1511 p[i].eobs = ctx->eobs_pbuf[i][2];
1512 }
1513
1514 // Restore the coding context of the MB to that that was in place
1515 // when the mode was picked for it
1516 for (y = 0; y < mi_height; y++)
1517 for (x_idx = 0; x_idx < mi_width; x_idx++)
1518 if ((xd->mb_to_right_edge >> (3 + MI_SIZE_LOG2)) + mi_width > x_idx &&
1519 (xd->mb_to_bottom_edge >> (3 + MI_SIZE_LOG2)) + mi_height > y) {
1520 xd->mi[x_idx + y * mis] = mi_addr;
1521 }
1522
1523 if (cpi->oxcf.aq_mode != NO_AQ) vp9_init_plane_quantizers(cpi, x);
1524
1525 if (is_inter_block(xdmi) && xdmi->sb_type < BLOCK_8X8) {
1526 xdmi->mv[0].as_int = mi->bmi[3].as_mv[0].as_int;
1527 xdmi->mv[1].as_int = mi->bmi[3].as_mv[1].as_int;
1528 }
1529
1530 x->skip = ctx->skip;
1531 memcpy(x->zcoeff_blk[xdmi->tx_size], ctx->zcoeff_blk,
1532 sizeof(ctx->zcoeff_blk[0]) * ctx->num_4x4_blk);
1533
1534 if (!output_enabled) return;
1535
1536 #if CONFIG_INTERNAL_STATS
1537 if (frame_is_intra_only(cm)) {
1538 static const int kf_mode_index[] = {
1539 THR_DC /*DC_PRED*/, THR_V_PRED /*V_PRED*/,
1540 THR_H_PRED /*H_PRED*/, THR_D45_PRED /*D45_PRED*/,
1541 THR_D135_PRED /*D135_PRED*/, THR_D117_PRED /*D117_PRED*/,
1542 THR_D153_PRED /*D153_PRED*/, THR_D207_PRED /*D207_PRED*/,
1543 THR_D63_PRED /*D63_PRED*/, THR_TM /*TM_PRED*/,
1544 };
1545 ++cpi->mode_chosen_counts[kf_mode_index[xdmi->mode]];
1546 } else {
1547 // Note how often each mode chosen as best
1548 ++cpi->mode_chosen_counts[ctx->best_mode_index];
1549 }
1550 #endif
1551 if (!frame_is_intra_only(cm)) {
1552 if (is_inter_block(xdmi)) {
1553 vp9_update_mv_count(td);
1554
1555 if (cm->interp_filter == SWITCHABLE) {
1556 const int ctx = get_pred_context_switchable_interp(xd);
1557 ++td->counts->switchable_interp[ctx][xdmi->interp_filter];
1558 }
1559 }
1560
1561 rdc->comp_pred_diff[SINGLE_REFERENCE] += ctx->single_pred_diff;
1562 rdc->comp_pred_diff[COMPOUND_REFERENCE] += ctx->comp_pred_diff;
1563 rdc->comp_pred_diff[REFERENCE_MODE_SELECT] += ctx->hybrid_pred_diff;
1564
1565 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i)
1566 rdc->filter_diff[i] += ctx->best_filter_diff[i];
1567 }
1568
1569 for (h = 0; h < y_mis; ++h) {
1570 MV_REF *const frame_mv = frame_mvs + h * cm->mi_cols;
1571 for (w = 0; w < x_mis; ++w) {
1572 MV_REF *const mv = frame_mv + w;
1573 mv->ref_frame[0] = mi->ref_frame[0];
1574 mv->ref_frame[1] = mi->ref_frame[1];
1575 mv->mv[0].as_int = mi->mv[0].as_int;
1576 mv->mv[1].as_int = mi->mv[1].as_int;
1577 }
1578 }
1579 }
1580
vp9_setup_src_planes(MACROBLOCK * x,const YV12_BUFFER_CONFIG * src,int mi_row,int mi_col)1581 void vp9_setup_src_planes(MACROBLOCK *x, const YV12_BUFFER_CONFIG *src,
1582 int mi_row, int mi_col) {
1583 uint8_t *const buffers[3] = { src->y_buffer, src->u_buffer, src->v_buffer };
1584 const int strides[3] = { src->y_stride, src->uv_stride, src->uv_stride };
1585 int i;
1586
1587 // Set current frame pointer.
1588 x->e_mbd.cur_buf = src;
1589
1590 for (i = 0; i < MAX_MB_PLANE; i++)
1591 setup_pred_plane(&x->plane[i].src, buffers[i], strides[i], mi_row, mi_col,
1592 NULL, x->e_mbd.plane[i].subsampling_x,
1593 x->e_mbd.plane[i].subsampling_y);
1594 }
1595
set_mode_info_seg_skip(MACROBLOCK * x,TX_MODE tx_mode,RD_COST * rd_cost,BLOCK_SIZE bsize)1596 static void set_mode_info_seg_skip(MACROBLOCK *x, TX_MODE tx_mode,
1597 RD_COST *rd_cost, BLOCK_SIZE bsize) {
1598 MACROBLOCKD *const xd = &x->e_mbd;
1599 MODE_INFO *const mi = xd->mi[0];
1600 INTERP_FILTER filter_ref;
1601
1602 filter_ref = get_pred_context_switchable_interp(xd);
1603 if (filter_ref == SWITCHABLE_FILTERS) filter_ref = EIGHTTAP;
1604
1605 mi->sb_type = bsize;
1606 mi->mode = ZEROMV;
1607 mi->tx_size =
1608 VPXMIN(max_txsize_lookup[bsize], tx_mode_to_biggest_tx_size[tx_mode]);
1609 mi->skip = 1;
1610 mi->uv_mode = DC_PRED;
1611 mi->ref_frame[0] = LAST_FRAME;
1612 mi->ref_frame[1] = NONE;
1613 mi->mv[0].as_int = 0;
1614 mi->interp_filter = filter_ref;
1615
1616 xd->mi[0]->bmi[0].as_mv[0].as_int = 0;
1617 x->skip = 1;
1618
1619 vp9_rd_cost_init(rd_cost);
1620 }
1621
set_segment_rdmult(VP9_COMP * const cpi,MACROBLOCK * const x,int8_t segment_id)1622 static int set_segment_rdmult(VP9_COMP *const cpi, MACROBLOCK *const x,
1623 int8_t segment_id) {
1624 int segment_qindex;
1625 VP9_COMMON *const cm = &cpi->common;
1626 vp9_init_plane_quantizers(cpi, x);
1627 vpx_clear_system_state();
1628 segment_qindex = vp9_get_qindex(&cm->seg, segment_id, cm->base_qindex);
1629 return vp9_compute_rd_mult(cpi, segment_qindex + cm->y_dc_delta_q);
1630 }
1631
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)1632 static void rd_pick_sb_modes(VP9_COMP *cpi, TileDataEnc *tile_data,
1633 MACROBLOCK *const x, int mi_row, int mi_col,
1634 RD_COST *rd_cost, BLOCK_SIZE bsize,
1635 PICK_MODE_CONTEXT *ctx, int64_t best_rd) {
1636 VP9_COMMON *const cm = &cpi->common;
1637 TileInfo *const tile_info = &tile_data->tile_info;
1638 MACROBLOCKD *const xd = &x->e_mbd;
1639 MODE_INFO *mi;
1640 struct macroblock_plane *const p = x->plane;
1641 struct macroblockd_plane *const pd = xd->plane;
1642 const AQ_MODE aq_mode = cpi->oxcf.aq_mode;
1643 int i, orig_rdmult;
1644
1645 vpx_clear_system_state();
1646
1647 // Use the lower precision, but faster, 32x32 fdct for mode selection.
1648 x->use_lp32x32fdct = 1;
1649
1650 set_offsets(cpi, tile_info, x, mi_row, mi_col, bsize);
1651 mi = xd->mi[0];
1652 mi->sb_type = bsize;
1653
1654 for (i = 0; i < MAX_MB_PLANE; ++i) {
1655 p[i].coeff = ctx->coeff_pbuf[i][0];
1656 p[i].qcoeff = ctx->qcoeff_pbuf[i][0];
1657 pd[i].dqcoeff = ctx->dqcoeff_pbuf[i][0];
1658 p[i].eobs = ctx->eobs_pbuf[i][0];
1659 }
1660 ctx->is_coded = 0;
1661 ctx->skippable = 0;
1662 ctx->pred_pixel_ready = 0;
1663 x->skip_recode = 0;
1664
1665 // Set to zero to make sure we do not use the previous encoded frame stats
1666 mi->skip = 0;
1667
1668 #if CONFIG_VP9_HIGHBITDEPTH
1669 if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
1670 x->source_variance = vp9_high_get_sby_perpixel_variance(
1671 cpi, &x->plane[0].src, bsize, xd->bd);
1672 } else {
1673 x->source_variance =
1674 vp9_get_sby_perpixel_variance(cpi, &x->plane[0].src, bsize);
1675 }
1676 #else
1677 x->source_variance =
1678 vp9_get_sby_perpixel_variance(cpi, &x->plane[0].src, bsize);
1679 #endif // CONFIG_VP9_HIGHBITDEPTH
1680
1681 // Save rdmult before it might be changed, so it can be restored later.
1682 orig_rdmult = x->rdmult;
1683
1684 if ((cpi->sf.tx_domain_thresh > 0.0) || (cpi->sf.quant_opt_thresh > 0.0)) {
1685 double logvar = vp9_log_block_var(cpi, x, bsize);
1686 // Check block complexity as part of descision on using pixel or transform
1687 // domain distortion in rd tests.
1688 x->block_tx_domain = cpi->sf.allow_txfm_domain_distortion &&
1689 (logvar >= cpi->sf.tx_domain_thresh);
1690
1691 // Check block complexity as part of descision on using quantized
1692 // coefficient optimisation inside the rd loop.
1693 x->block_qcoeff_opt =
1694 cpi->sf.allow_quant_coeff_opt && (logvar <= cpi->sf.quant_opt_thresh);
1695 } else {
1696 x->block_tx_domain = cpi->sf.allow_txfm_domain_distortion;
1697 x->block_qcoeff_opt = cpi->sf.allow_quant_coeff_opt;
1698 }
1699
1700 if (aq_mode == VARIANCE_AQ) {
1701 const int energy =
1702 bsize <= BLOCK_16X16 ? x->mb_energy : vp9_block_energy(cpi, x, bsize);
1703
1704 if (cm->frame_type == KEY_FRAME || cpi->refresh_alt_ref_frame ||
1705 cpi->force_update_segmentation ||
1706 (cpi->refresh_golden_frame && !cpi->rc.is_src_frame_alt_ref)) {
1707 mi->segment_id = vp9_vaq_segment_id(energy);
1708 } else {
1709 const uint8_t *const map =
1710 cm->seg.update_map ? cpi->segmentation_map : cm->last_frame_seg_map;
1711 mi->segment_id = get_segment_id(cm, map, bsize, mi_row, mi_col);
1712 }
1713 x->rdmult = set_segment_rdmult(cpi, x, mi->segment_id);
1714 } else if (aq_mode == LOOKAHEAD_AQ) {
1715 const uint8_t *const map = cpi->segmentation_map;
1716
1717 // I do not change rdmult here consciously.
1718 mi->segment_id = get_segment_id(cm, map, bsize, mi_row, mi_col);
1719 } else if (aq_mode == EQUATOR360_AQ) {
1720 if (cm->frame_type == KEY_FRAME || cpi->force_update_segmentation) {
1721 mi->segment_id = vp9_360aq_segment_id(mi_row, cm->mi_rows);
1722 } else {
1723 const uint8_t *const map =
1724 cm->seg.update_map ? cpi->segmentation_map : cm->last_frame_seg_map;
1725 mi->segment_id = get_segment_id(cm, map, bsize, mi_row, mi_col);
1726 }
1727 x->rdmult = set_segment_rdmult(cpi, x, mi->segment_id);
1728 } else if (aq_mode == COMPLEXITY_AQ) {
1729 x->rdmult = set_segment_rdmult(cpi, x, mi->segment_id);
1730 } else if (aq_mode == CYCLIC_REFRESH_AQ) {
1731 const uint8_t *const map =
1732 cm->seg.update_map ? cpi->segmentation_map : cm->last_frame_seg_map;
1733 // If segment is boosted, use rdmult for that segment.
1734 if (cyclic_refresh_segment_id_boosted(
1735 get_segment_id(cm, map, bsize, mi_row, mi_col)))
1736 x->rdmult = vp9_cyclic_refresh_get_rdmult(cpi->cyclic_refresh);
1737 }
1738
1739 // Find best coding mode & reconstruct the MB so it is available
1740 // as a predictor for MBs that follow in the SB
1741 if (frame_is_intra_only(cm)) {
1742 vp9_rd_pick_intra_mode_sb(cpi, x, rd_cost, bsize, ctx, best_rd);
1743 } else {
1744 if (bsize >= BLOCK_8X8) {
1745 if (segfeature_active(&cm->seg, mi->segment_id, SEG_LVL_SKIP))
1746 vp9_rd_pick_inter_mode_sb_seg_skip(cpi, tile_data, x, rd_cost, bsize,
1747 ctx, best_rd);
1748 else
1749 vp9_rd_pick_inter_mode_sb(cpi, tile_data, x, mi_row, mi_col, rd_cost,
1750 bsize, ctx, best_rd);
1751 } else {
1752 vp9_rd_pick_inter_mode_sub8x8(cpi, tile_data, x, mi_row, mi_col, rd_cost,
1753 bsize, ctx, best_rd);
1754 }
1755 }
1756
1757 // Examine the resulting rate and for AQ mode 2 make a segment choice.
1758 if ((rd_cost->rate != INT_MAX) && (aq_mode == COMPLEXITY_AQ) &&
1759 (bsize >= BLOCK_16X16) &&
1760 (cm->frame_type == KEY_FRAME || cpi->refresh_alt_ref_frame ||
1761 (cpi->refresh_golden_frame && !cpi->rc.is_src_frame_alt_ref))) {
1762 vp9_caq_select_segment(cpi, x, bsize, mi_row, mi_col, rd_cost->rate);
1763 }
1764
1765 x->rdmult = orig_rdmult;
1766
1767 // TODO(jingning) The rate-distortion optimization flow needs to be
1768 // refactored to provide proper exit/return handle.
1769 if (rd_cost->rate == INT_MAX) rd_cost->rdcost = INT64_MAX;
1770
1771 ctx->rate = rd_cost->rate;
1772 ctx->dist = rd_cost->dist;
1773 }
1774
update_stats(VP9_COMMON * cm,ThreadData * td)1775 static void update_stats(VP9_COMMON *cm, ThreadData *td) {
1776 const MACROBLOCK *x = &td->mb;
1777 const MACROBLOCKD *const xd = &x->e_mbd;
1778 const MODE_INFO *const mi = xd->mi[0];
1779 const MB_MODE_INFO_EXT *const mbmi_ext = x->mbmi_ext;
1780 const BLOCK_SIZE bsize = mi->sb_type;
1781
1782 if (!frame_is_intra_only(cm)) {
1783 FRAME_COUNTS *const counts = td->counts;
1784 const int inter_block = is_inter_block(mi);
1785 const int seg_ref_active =
1786 segfeature_active(&cm->seg, mi->segment_id, SEG_LVL_REF_FRAME);
1787 if (!seg_ref_active) {
1788 counts->intra_inter[get_intra_inter_context(xd)][inter_block]++;
1789 // If the segment reference feature is enabled we have only a single
1790 // reference frame allowed for the segment so exclude it from
1791 // the reference frame counts used to work out probabilities.
1792 if (inter_block) {
1793 const MV_REFERENCE_FRAME ref0 = mi->ref_frame[0];
1794 if (cm->reference_mode == REFERENCE_MODE_SELECT)
1795 counts->comp_inter[vp9_get_reference_mode_context(cm, xd)]
1796 [has_second_ref(mi)]++;
1797
1798 if (has_second_ref(mi)) {
1799 counts->comp_ref[vp9_get_pred_context_comp_ref_p(cm, xd)]
1800 [ref0 == GOLDEN_FRAME]++;
1801 } else {
1802 counts->single_ref[vp9_get_pred_context_single_ref_p1(xd)][0]
1803 [ref0 != LAST_FRAME]++;
1804 if (ref0 != LAST_FRAME)
1805 counts->single_ref[vp9_get_pred_context_single_ref_p2(xd)][1]
1806 [ref0 != GOLDEN_FRAME]++;
1807 }
1808 }
1809 }
1810 if (inter_block &&
1811 !segfeature_active(&cm->seg, mi->segment_id, SEG_LVL_SKIP)) {
1812 const int mode_ctx = mbmi_ext->mode_context[mi->ref_frame[0]];
1813 if (bsize >= BLOCK_8X8) {
1814 const PREDICTION_MODE mode = mi->mode;
1815 ++counts->inter_mode[mode_ctx][INTER_OFFSET(mode)];
1816 } else {
1817 const int num_4x4_w = num_4x4_blocks_wide_lookup[bsize];
1818 const int num_4x4_h = num_4x4_blocks_high_lookup[bsize];
1819 int idx, idy;
1820 for (idy = 0; idy < 2; idy += num_4x4_h) {
1821 for (idx = 0; idx < 2; idx += num_4x4_w) {
1822 const int j = idy * 2 + idx;
1823 const PREDICTION_MODE b_mode = mi->bmi[j].as_mode;
1824 ++counts->inter_mode[mode_ctx][INTER_OFFSET(b_mode)];
1825 }
1826 }
1827 }
1828 }
1829 }
1830 }
1831
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)1832 static void restore_context(MACROBLOCK *const x, int mi_row, int mi_col,
1833 ENTROPY_CONTEXT a[16 * MAX_MB_PLANE],
1834 ENTROPY_CONTEXT l[16 * MAX_MB_PLANE],
1835 PARTITION_CONTEXT sa[8], PARTITION_CONTEXT sl[8],
1836 BLOCK_SIZE bsize) {
1837 MACROBLOCKD *const xd = &x->e_mbd;
1838 int p;
1839 const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
1840 const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
1841 int mi_width = num_8x8_blocks_wide_lookup[bsize];
1842 int mi_height = num_8x8_blocks_high_lookup[bsize];
1843 for (p = 0; p < MAX_MB_PLANE; p++) {
1844 memcpy(xd->above_context[p] + ((mi_col * 2) >> xd->plane[p].subsampling_x),
1845 a + num_4x4_blocks_wide * p,
1846 (sizeof(ENTROPY_CONTEXT) * num_4x4_blocks_wide) >>
1847 xd->plane[p].subsampling_x);
1848 memcpy(xd->left_context[p] +
1849 ((mi_row & MI_MASK) * 2 >> xd->plane[p].subsampling_y),
1850 l + num_4x4_blocks_high * p,
1851 (sizeof(ENTROPY_CONTEXT) * num_4x4_blocks_high) >>
1852 xd->plane[p].subsampling_y);
1853 }
1854 memcpy(xd->above_seg_context + mi_col, sa,
1855 sizeof(*xd->above_seg_context) * mi_width);
1856 memcpy(xd->left_seg_context + (mi_row & MI_MASK), sl,
1857 sizeof(xd->left_seg_context[0]) * mi_height);
1858 }
1859
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)1860 static void save_context(MACROBLOCK *const x, int mi_row, int mi_col,
1861 ENTROPY_CONTEXT a[16 * MAX_MB_PLANE],
1862 ENTROPY_CONTEXT l[16 * MAX_MB_PLANE],
1863 PARTITION_CONTEXT sa[8], PARTITION_CONTEXT sl[8],
1864 BLOCK_SIZE bsize) {
1865 const MACROBLOCKD *const xd = &x->e_mbd;
1866 int p;
1867 const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
1868 const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
1869 int mi_width = num_8x8_blocks_wide_lookup[bsize];
1870 int mi_height = num_8x8_blocks_high_lookup[bsize];
1871
1872 // buffer the above/left context information of the block in search.
1873 for (p = 0; p < MAX_MB_PLANE; ++p) {
1874 memcpy(a + num_4x4_blocks_wide * p,
1875 xd->above_context[p] + (mi_col * 2 >> xd->plane[p].subsampling_x),
1876 (sizeof(ENTROPY_CONTEXT) * num_4x4_blocks_wide) >>
1877 xd->plane[p].subsampling_x);
1878 memcpy(l + num_4x4_blocks_high * p,
1879 xd->left_context[p] +
1880 ((mi_row & MI_MASK) * 2 >> xd->plane[p].subsampling_y),
1881 (sizeof(ENTROPY_CONTEXT) * num_4x4_blocks_high) >>
1882 xd->plane[p].subsampling_y);
1883 }
1884 memcpy(sa, xd->above_seg_context + mi_col,
1885 sizeof(*xd->above_seg_context) * mi_width);
1886 memcpy(sl, xd->left_seg_context + (mi_row & MI_MASK),
1887 sizeof(xd->left_seg_context[0]) * mi_height);
1888 }
1889
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)1890 static void encode_b(VP9_COMP *cpi, const TileInfo *const tile, ThreadData *td,
1891 TOKENEXTRA **tp, int mi_row, int mi_col,
1892 int output_enabled, BLOCK_SIZE bsize,
1893 PICK_MODE_CONTEXT *ctx) {
1894 MACROBLOCK *const x = &td->mb;
1895 set_offsets(cpi, tile, x, mi_row, mi_col, bsize);
1896 update_state(cpi, td, ctx, mi_row, mi_col, bsize, output_enabled);
1897 encode_superblock(cpi, td, tp, output_enabled, mi_row, mi_col, bsize, ctx);
1898
1899 if (output_enabled) {
1900 update_stats(&cpi->common, td);
1901
1902 (*tp)->token = EOSB_TOKEN;
1903 (*tp)++;
1904 }
1905 }
1906
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)1907 static void encode_sb(VP9_COMP *cpi, ThreadData *td, const TileInfo *const tile,
1908 TOKENEXTRA **tp, int mi_row, int mi_col,
1909 int output_enabled, BLOCK_SIZE bsize, PC_TREE *pc_tree) {
1910 VP9_COMMON *const cm = &cpi->common;
1911 MACROBLOCK *const x = &td->mb;
1912 MACROBLOCKD *const xd = &x->e_mbd;
1913
1914 const int bsl = b_width_log2_lookup[bsize], hbs = (1 << bsl) / 4;
1915 int ctx;
1916 PARTITION_TYPE partition;
1917 BLOCK_SIZE subsize = bsize;
1918
1919 if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols) return;
1920
1921 if (bsize >= BLOCK_8X8) {
1922 ctx = partition_plane_context(xd, mi_row, mi_col, bsize);
1923 subsize = get_subsize(bsize, pc_tree->partitioning);
1924 } else {
1925 ctx = 0;
1926 subsize = BLOCK_4X4;
1927 }
1928
1929 partition = partition_lookup[bsl][subsize];
1930 if (output_enabled && bsize != BLOCK_4X4)
1931 td->counts->partition[ctx][partition]++;
1932
1933 switch (partition) {
1934 case PARTITION_NONE:
1935 encode_b(cpi, tile, td, tp, mi_row, mi_col, output_enabled, subsize,
1936 &pc_tree->none);
1937 break;
1938 case PARTITION_VERT:
1939 encode_b(cpi, tile, td, tp, mi_row, mi_col, output_enabled, subsize,
1940 &pc_tree->vertical[0]);
1941 if (mi_col + hbs < cm->mi_cols && bsize > BLOCK_8X8) {
1942 encode_b(cpi, tile, td, tp, mi_row, mi_col + hbs, output_enabled,
1943 subsize, &pc_tree->vertical[1]);
1944 }
1945 break;
1946 case PARTITION_HORZ:
1947 encode_b(cpi, tile, td, tp, mi_row, mi_col, output_enabled, subsize,
1948 &pc_tree->horizontal[0]);
1949 if (mi_row + hbs < cm->mi_rows && bsize > BLOCK_8X8) {
1950 encode_b(cpi, tile, td, tp, mi_row + hbs, mi_col, output_enabled,
1951 subsize, &pc_tree->horizontal[1]);
1952 }
1953 break;
1954 case PARTITION_SPLIT:
1955 if (bsize == BLOCK_8X8) {
1956 encode_b(cpi, tile, td, tp, mi_row, mi_col, output_enabled, subsize,
1957 pc_tree->leaf_split[0]);
1958 } else {
1959 encode_sb(cpi, td, tile, tp, mi_row, mi_col, output_enabled, subsize,
1960 pc_tree->split[0]);
1961 encode_sb(cpi, td, tile, tp, mi_row, mi_col + hbs, output_enabled,
1962 subsize, pc_tree->split[1]);
1963 encode_sb(cpi, td, tile, tp, mi_row + hbs, mi_col, output_enabled,
1964 subsize, pc_tree->split[2]);
1965 encode_sb(cpi, td, tile, tp, mi_row + hbs, mi_col + hbs, output_enabled,
1966 subsize, pc_tree->split[3]);
1967 }
1968 break;
1969 default: assert(0 && "Invalid partition type."); break;
1970 }
1971
1972 if (partition != PARTITION_SPLIT || bsize == BLOCK_8X8)
1973 update_partition_context(xd, mi_row, mi_col, subsize, bsize);
1974 }
1975
1976 // Check to see if the given partition size is allowed for a specified number
1977 // of 8x8 block rows and columns remaining in the image.
1978 // 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)1979 static BLOCK_SIZE find_partition_size(BLOCK_SIZE bsize, int rows_left,
1980 int cols_left, int *bh, int *bw) {
1981 if (rows_left <= 0 || cols_left <= 0) {
1982 return VPXMIN(bsize, BLOCK_8X8);
1983 } else {
1984 for (; bsize > 0; bsize -= 3) {
1985 *bh = num_8x8_blocks_high_lookup[bsize];
1986 *bw = num_8x8_blocks_wide_lookup[bsize];
1987 if ((*bh <= rows_left) && (*bw <= cols_left)) {
1988 break;
1989 }
1990 }
1991 }
1992 return bsize;
1993 }
1994
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)1995 static void set_partial_b64x64_partition(MODE_INFO *mi, int mis, int bh_in,
1996 int bw_in, int row8x8_remaining,
1997 int col8x8_remaining, BLOCK_SIZE bsize,
1998 MODE_INFO **mi_8x8) {
1999 int bh = bh_in;
2000 int r, c;
2001 for (r = 0; r < MI_BLOCK_SIZE; r += bh) {
2002 int bw = bw_in;
2003 for (c = 0; c < MI_BLOCK_SIZE; c += bw) {
2004 const int index = r * mis + c;
2005 mi_8x8[index] = mi + index;
2006 mi_8x8[index]->sb_type = find_partition_size(
2007 bsize, row8x8_remaining - r, col8x8_remaining - c, &bh, &bw);
2008 }
2009 }
2010 }
2011
2012 // This function attempts to set all mode info entries in a given SB64
2013 // to the same block partition size.
2014 // However, at the bottom and right borders of the image the requested size
2015 // may not be allowed in which case this code attempts to choose the largest
2016 // 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)2017 static void set_fixed_partitioning(VP9_COMP *cpi, const TileInfo *const tile,
2018 MODE_INFO **mi_8x8, int mi_row, int mi_col,
2019 BLOCK_SIZE bsize) {
2020 VP9_COMMON *const cm = &cpi->common;
2021 const int mis = cm->mi_stride;
2022 const int row8x8_remaining = tile->mi_row_end - mi_row;
2023 const int col8x8_remaining = tile->mi_col_end - mi_col;
2024 int block_row, block_col;
2025 MODE_INFO *mi_upper_left = cm->mi + mi_row * mis + mi_col;
2026 int bh = num_8x8_blocks_high_lookup[bsize];
2027 int bw = num_8x8_blocks_wide_lookup[bsize];
2028
2029 assert((row8x8_remaining > 0) && (col8x8_remaining > 0));
2030
2031 // Apply the requested partition size to the SB64 if it is all "in image"
2032 if ((col8x8_remaining >= MI_BLOCK_SIZE) &&
2033 (row8x8_remaining >= MI_BLOCK_SIZE)) {
2034 for (block_row = 0; block_row < MI_BLOCK_SIZE; block_row += bh) {
2035 for (block_col = 0; block_col < MI_BLOCK_SIZE; block_col += bw) {
2036 int index = block_row * mis + block_col;
2037 mi_8x8[index] = mi_upper_left + index;
2038 mi_8x8[index]->sb_type = bsize;
2039 }
2040 }
2041 } else {
2042 // Else this is a partial SB64.
2043 set_partial_b64x64_partition(mi_upper_left, mis, bh, bw, row8x8_remaining,
2044 col8x8_remaining, bsize, mi_8x8);
2045 }
2046 }
2047
2048 static const struct {
2049 int row;
2050 int col;
2051 } coord_lookup[16] = {
2052 // 32x32 index = 0
2053 { 0, 0 },
2054 { 0, 2 },
2055 { 2, 0 },
2056 { 2, 2 },
2057 // 32x32 index = 1
2058 { 0, 4 },
2059 { 0, 6 },
2060 { 2, 4 },
2061 { 2, 6 },
2062 // 32x32 index = 2
2063 { 4, 0 },
2064 { 4, 2 },
2065 { 6, 0 },
2066 { 6, 2 },
2067 // 32x32 index = 3
2068 { 4, 4 },
2069 { 4, 6 },
2070 { 6, 4 },
2071 { 6, 6 },
2072 };
2073
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)2074 static void set_source_var_based_partition(VP9_COMP *cpi,
2075 const TileInfo *const tile,
2076 MACROBLOCK *const x,
2077 MODE_INFO **mi_8x8, int mi_row,
2078 int mi_col) {
2079 VP9_COMMON *const cm = &cpi->common;
2080 const int mis = cm->mi_stride;
2081 const int row8x8_remaining = tile->mi_row_end - mi_row;
2082 const int col8x8_remaining = tile->mi_col_end - mi_col;
2083 MODE_INFO *mi_upper_left = cm->mi + mi_row * mis + mi_col;
2084
2085 vp9_setup_src_planes(x, cpi->Source, mi_row, mi_col);
2086
2087 assert((row8x8_remaining > 0) && (col8x8_remaining > 0));
2088
2089 // In-image SB64
2090 if ((col8x8_remaining >= MI_BLOCK_SIZE) &&
2091 (row8x8_remaining >= MI_BLOCK_SIZE)) {
2092 int i, j;
2093 int index;
2094 diff d32[4];
2095 const int offset = (mi_row >> 1) * cm->mb_cols + (mi_col >> 1);
2096 int is_larger_better = 0;
2097 int use32x32 = 0;
2098 unsigned int thr = cpi->source_var_thresh;
2099
2100 memset(d32, 0, 4 * sizeof(diff));
2101
2102 for (i = 0; i < 4; i++) {
2103 diff *d16[4];
2104
2105 for (j = 0; j < 4; j++) {
2106 int b_mi_row = coord_lookup[i * 4 + j].row;
2107 int b_mi_col = coord_lookup[i * 4 + j].col;
2108 int boffset = b_mi_row / 2 * cm->mb_cols + b_mi_col / 2;
2109
2110 d16[j] = cpi->source_diff_var + offset + boffset;
2111
2112 index = b_mi_row * mis + b_mi_col;
2113 mi_8x8[index] = mi_upper_left + index;
2114 mi_8x8[index]->sb_type = BLOCK_16X16;
2115
2116 // TODO(yunqingwang): If d16[j].var is very large, use 8x8 partition
2117 // size to further improve quality.
2118 }
2119
2120 is_larger_better = (d16[0]->var < thr) && (d16[1]->var < thr) &&
2121 (d16[2]->var < thr) && (d16[3]->var < thr);
2122
2123 // Use 32x32 partition
2124 if (is_larger_better) {
2125 use32x32 += 1;
2126
2127 for (j = 0; j < 4; j++) {
2128 d32[i].sse += d16[j]->sse;
2129 d32[i].sum += d16[j]->sum;
2130 }
2131
2132 d32[i].var =
2133 (unsigned int)(d32[i].sse -
2134 (unsigned int)(((int64_t)d32[i].sum * d32[i].sum) >>
2135 10));
2136
2137 index = coord_lookup[i * 4].row * mis + coord_lookup[i * 4].col;
2138 mi_8x8[index] = mi_upper_left + index;
2139 mi_8x8[index]->sb_type = BLOCK_32X32;
2140 }
2141 }
2142
2143 if (use32x32 == 4) {
2144 thr <<= 1;
2145 is_larger_better = (d32[0].var < thr) && (d32[1].var < thr) &&
2146 (d32[2].var < thr) && (d32[3].var < thr);
2147
2148 // Use 64x64 partition
2149 if (is_larger_better) {
2150 mi_8x8[0] = mi_upper_left;
2151 mi_8x8[0]->sb_type = BLOCK_64X64;
2152 }
2153 }
2154 } else { // partial in-image SB64
2155 int bh = num_8x8_blocks_high_lookup[BLOCK_16X16];
2156 int bw = num_8x8_blocks_wide_lookup[BLOCK_16X16];
2157 set_partial_b64x64_partition(mi_upper_left, mis, bh, bw, row8x8_remaining,
2158 col8x8_remaining, BLOCK_16X16, mi_8x8);
2159 }
2160 }
2161
update_state_rt(VP9_COMP * cpi,ThreadData * td,PICK_MODE_CONTEXT * ctx,int mi_row,int mi_col,int bsize)2162 static void update_state_rt(VP9_COMP *cpi, ThreadData *td,
2163 PICK_MODE_CONTEXT *ctx, int mi_row, int mi_col,
2164 int bsize) {
2165 VP9_COMMON *const cm = &cpi->common;
2166 MACROBLOCK *const x = &td->mb;
2167 MACROBLOCKD *const xd = &x->e_mbd;
2168 MODE_INFO *const mi = xd->mi[0];
2169 struct macroblock_plane *const p = x->plane;
2170 const struct segmentation *const seg = &cm->seg;
2171 const int bw = num_8x8_blocks_wide_lookup[mi->sb_type];
2172 const int bh = num_8x8_blocks_high_lookup[mi->sb_type];
2173 const int x_mis = VPXMIN(bw, cm->mi_cols - mi_col);
2174 const int y_mis = VPXMIN(bh, cm->mi_rows - mi_row);
2175
2176 *(xd->mi[0]) = ctx->mic;
2177 *(x->mbmi_ext) = ctx->mbmi_ext;
2178
2179 if (seg->enabled && cpi->oxcf.aq_mode != NO_AQ) {
2180 // For in frame complexity AQ or variance AQ, copy segment_id from
2181 // segmentation_map.
2182 if (cpi->oxcf.aq_mode != CYCLIC_REFRESH_AQ) {
2183 const uint8_t *const map =
2184 seg->update_map ? cpi->segmentation_map : cm->last_frame_seg_map;
2185 mi->segment_id = get_segment_id(cm, map, bsize, mi_row, mi_col);
2186 } else {
2187 // Setting segmentation map for cyclic_refresh.
2188 vp9_cyclic_refresh_update_segment(cpi, mi, mi_row, mi_col, bsize,
2189 ctx->rate, ctx->dist, x->skip, p);
2190 }
2191 vp9_init_plane_quantizers(cpi, x);
2192 }
2193
2194 if (is_inter_block(mi)) {
2195 vp9_update_mv_count(td);
2196 if (cm->interp_filter == SWITCHABLE) {
2197 const int pred_ctx = get_pred_context_switchable_interp(xd);
2198 ++td->counts->switchable_interp[pred_ctx][mi->interp_filter];
2199 }
2200
2201 if (mi->sb_type < BLOCK_8X8) {
2202 mi->mv[0].as_int = mi->bmi[3].as_mv[0].as_int;
2203 mi->mv[1].as_int = mi->bmi[3].as_mv[1].as_int;
2204 }
2205 }
2206
2207 if (cm->use_prev_frame_mvs || !cm->error_resilient_mode ||
2208 (cpi->svc.use_base_mv && cpi->svc.number_spatial_layers > 1 &&
2209 cpi->svc.spatial_layer_id != cpi->svc.number_spatial_layers - 1)) {
2210 MV_REF *const frame_mvs =
2211 cm->cur_frame->mvs + mi_row * cm->mi_cols + mi_col;
2212 int w, h;
2213
2214 for (h = 0; h < y_mis; ++h) {
2215 MV_REF *const frame_mv = frame_mvs + h * cm->mi_cols;
2216 for (w = 0; w < x_mis; ++w) {
2217 MV_REF *const mv = frame_mv + w;
2218 mv->ref_frame[0] = mi->ref_frame[0];
2219 mv->ref_frame[1] = mi->ref_frame[1];
2220 mv->mv[0].as_int = mi->mv[0].as_int;
2221 mv->mv[1].as_int = mi->mv[1].as_int;
2222 }
2223 }
2224 }
2225
2226 x->skip = ctx->skip;
2227 x->skip_txfm[0] = mi->segment_id ? 0 : ctx->skip_txfm[0];
2228 }
2229
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)2230 static void encode_b_rt(VP9_COMP *cpi, ThreadData *td,
2231 const TileInfo *const tile, TOKENEXTRA **tp, int mi_row,
2232 int mi_col, int output_enabled, BLOCK_SIZE bsize,
2233 PICK_MODE_CONTEXT *ctx) {
2234 MACROBLOCK *const x = &td->mb;
2235 set_offsets(cpi, tile, x, mi_row, mi_col, bsize);
2236 update_state_rt(cpi, td, ctx, mi_row, mi_col, bsize);
2237
2238 encode_superblock(cpi, td, tp, output_enabled, mi_row, mi_col, bsize, ctx);
2239 update_stats(&cpi->common, td);
2240
2241 (*tp)->token = EOSB_TOKEN;
2242 (*tp)++;
2243 }
2244
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)2245 static void encode_sb_rt(VP9_COMP *cpi, ThreadData *td,
2246 const TileInfo *const tile, TOKENEXTRA **tp,
2247 int mi_row, int mi_col, int output_enabled,
2248 BLOCK_SIZE bsize, PC_TREE *pc_tree) {
2249 VP9_COMMON *const cm = &cpi->common;
2250 MACROBLOCK *const x = &td->mb;
2251 MACROBLOCKD *const xd = &x->e_mbd;
2252
2253 const int bsl = b_width_log2_lookup[bsize], hbs = (1 << bsl) / 4;
2254 int ctx;
2255 PARTITION_TYPE partition;
2256 BLOCK_SIZE subsize;
2257
2258 if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols) return;
2259
2260 if (bsize >= BLOCK_8X8) {
2261 const int idx_str = xd->mi_stride * mi_row + mi_col;
2262 MODE_INFO **mi_8x8 = cm->mi_grid_visible + idx_str;
2263 ctx = partition_plane_context(xd, mi_row, mi_col, bsize);
2264 subsize = mi_8x8[0]->sb_type;
2265 } else {
2266 ctx = 0;
2267 subsize = BLOCK_4X4;
2268 }
2269
2270 partition = partition_lookup[bsl][subsize];
2271 if (output_enabled && bsize != BLOCK_4X4)
2272 td->counts->partition[ctx][partition]++;
2273
2274 switch (partition) {
2275 case PARTITION_NONE:
2276 encode_b_rt(cpi, td, tile, tp, mi_row, mi_col, output_enabled, subsize,
2277 &pc_tree->none);
2278 break;
2279 case PARTITION_VERT:
2280 encode_b_rt(cpi, td, tile, tp, mi_row, mi_col, output_enabled, subsize,
2281 &pc_tree->vertical[0]);
2282 if (mi_col + hbs < cm->mi_cols && bsize > BLOCK_8X8) {
2283 encode_b_rt(cpi, td, tile, tp, mi_row, mi_col + hbs, output_enabled,
2284 subsize, &pc_tree->vertical[1]);
2285 }
2286 break;
2287 case PARTITION_HORZ:
2288 encode_b_rt(cpi, td, tile, tp, mi_row, mi_col, output_enabled, subsize,
2289 &pc_tree->horizontal[0]);
2290 if (mi_row + hbs < cm->mi_rows && bsize > BLOCK_8X8) {
2291 encode_b_rt(cpi, td, tile, tp, mi_row + hbs, mi_col, output_enabled,
2292 subsize, &pc_tree->horizontal[1]);
2293 }
2294 break;
2295 case PARTITION_SPLIT:
2296 subsize = get_subsize(bsize, PARTITION_SPLIT);
2297 encode_sb_rt(cpi, td, tile, tp, mi_row, mi_col, output_enabled, subsize,
2298 pc_tree->split[0]);
2299 encode_sb_rt(cpi, td, tile, tp, mi_row, mi_col + hbs, output_enabled,
2300 subsize, pc_tree->split[1]);
2301 encode_sb_rt(cpi, td, tile, tp, mi_row + hbs, mi_col, output_enabled,
2302 subsize, pc_tree->split[2]);
2303 encode_sb_rt(cpi, td, tile, tp, mi_row + hbs, mi_col + hbs,
2304 output_enabled, subsize, pc_tree->split[3]);
2305 break;
2306 default: assert(0 && "Invalid partition type."); break;
2307 }
2308
2309 if (partition != PARTITION_SPLIT || bsize == BLOCK_8X8)
2310 update_partition_context(xd, mi_row, mi_col, subsize, bsize);
2311 }
2312
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)2313 static void rd_use_partition(VP9_COMP *cpi, ThreadData *td,
2314 TileDataEnc *tile_data, MODE_INFO **mi_8x8,
2315 TOKENEXTRA **tp, int mi_row, int mi_col,
2316 BLOCK_SIZE bsize, int *rate, int64_t *dist,
2317 int do_recon, PC_TREE *pc_tree) {
2318 VP9_COMMON *const cm = &cpi->common;
2319 TileInfo *const tile_info = &tile_data->tile_info;
2320 MACROBLOCK *const x = &td->mb;
2321 MACROBLOCKD *const xd = &x->e_mbd;
2322 const int mis = cm->mi_stride;
2323 const int bsl = b_width_log2_lookup[bsize];
2324 const int mi_step = num_4x4_blocks_wide_lookup[bsize] / 2;
2325 const int bss = (1 << bsl) / 4;
2326 int i, pl;
2327 PARTITION_TYPE partition = PARTITION_NONE;
2328 BLOCK_SIZE subsize;
2329 ENTROPY_CONTEXT l[16 * MAX_MB_PLANE], a[16 * MAX_MB_PLANE];
2330 PARTITION_CONTEXT sl[8], sa[8];
2331 RD_COST last_part_rdc, none_rdc, chosen_rdc;
2332 BLOCK_SIZE sub_subsize = BLOCK_4X4;
2333 int splits_below = 0;
2334 BLOCK_SIZE bs_type = mi_8x8[0]->sb_type;
2335 int do_partition_search = 1;
2336 PICK_MODE_CONTEXT *ctx = &pc_tree->none;
2337
2338 if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols) return;
2339
2340 assert(num_4x4_blocks_wide_lookup[bsize] ==
2341 num_4x4_blocks_high_lookup[bsize]);
2342
2343 vp9_rd_cost_reset(&last_part_rdc);
2344 vp9_rd_cost_reset(&none_rdc);
2345 vp9_rd_cost_reset(&chosen_rdc);
2346
2347 partition = partition_lookup[bsl][bs_type];
2348 subsize = get_subsize(bsize, partition);
2349
2350 pc_tree->partitioning = partition;
2351 save_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
2352
2353 if (bsize == BLOCK_16X16 && cpi->oxcf.aq_mode != NO_AQ) {
2354 set_offsets(cpi, tile_info, x, mi_row, mi_col, bsize);
2355 x->mb_energy = vp9_block_energy(cpi, x, bsize);
2356 }
2357
2358 if (do_partition_search &&
2359 cpi->sf.partition_search_type == SEARCH_PARTITION &&
2360 cpi->sf.adjust_partitioning_from_last_frame) {
2361 // Check if any of the sub blocks are further split.
2362 if (partition == PARTITION_SPLIT && subsize > BLOCK_8X8) {
2363 sub_subsize = get_subsize(subsize, PARTITION_SPLIT);
2364 splits_below = 1;
2365 for (i = 0; i < 4; i++) {
2366 int jj = i >> 1, ii = i & 0x01;
2367 MODE_INFO *this_mi = mi_8x8[jj * bss * mis + ii * bss];
2368 if (this_mi && this_mi->sb_type >= sub_subsize) {
2369 splits_below = 0;
2370 }
2371 }
2372 }
2373
2374 // If partition is not none try none unless each of the 4 splits are split
2375 // even further..
2376 if (partition != PARTITION_NONE && !splits_below &&
2377 mi_row + (mi_step >> 1) < cm->mi_rows &&
2378 mi_col + (mi_step >> 1) < cm->mi_cols) {
2379 pc_tree->partitioning = PARTITION_NONE;
2380 rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &none_rdc, bsize, ctx,
2381 INT64_MAX);
2382
2383 pl = partition_plane_context(xd, mi_row, mi_col, bsize);
2384
2385 if (none_rdc.rate < INT_MAX) {
2386 none_rdc.rate += cpi->partition_cost[pl][PARTITION_NONE];
2387 none_rdc.rdcost =
2388 RDCOST(x->rdmult, x->rddiv, none_rdc.rate, none_rdc.dist);
2389 }
2390
2391 restore_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
2392 mi_8x8[0]->sb_type = bs_type;
2393 pc_tree->partitioning = partition;
2394 }
2395 }
2396
2397 switch (partition) {
2398 case PARTITION_NONE:
2399 rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &last_part_rdc, bsize,
2400 ctx, INT64_MAX);
2401 break;
2402 case PARTITION_HORZ:
2403 rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &last_part_rdc,
2404 subsize, &pc_tree->horizontal[0], INT64_MAX);
2405 if (last_part_rdc.rate != INT_MAX && bsize >= BLOCK_8X8 &&
2406 mi_row + (mi_step >> 1) < cm->mi_rows) {
2407 RD_COST tmp_rdc;
2408 PICK_MODE_CONTEXT *ctx = &pc_tree->horizontal[0];
2409 vp9_rd_cost_init(&tmp_rdc);
2410 update_state(cpi, td, ctx, mi_row, mi_col, subsize, 0);
2411 encode_superblock(cpi, td, tp, 0, mi_row, mi_col, subsize, ctx);
2412 rd_pick_sb_modes(cpi, tile_data, x, mi_row + (mi_step >> 1), mi_col,
2413 &tmp_rdc, subsize, &pc_tree->horizontal[1], INT64_MAX);
2414 if (tmp_rdc.rate == INT_MAX || tmp_rdc.dist == INT64_MAX) {
2415 vp9_rd_cost_reset(&last_part_rdc);
2416 break;
2417 }
2418 last_part_rdc.rate += tmp_rdc.rate;
2419 last_part_rdc.dist += tmp_rdc.dist;
2420 last_part_rdc.rdcost += tmp_rdc.rdcost;
2421 }
2422 break;
2423 case PARTITION_VERT:
2424 rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &last_part_rdc,
2425 subsize, &pc_tree->vertical[0], INT64_MAX);
2426 if (last_part_rdc.rate != INT_MAX && bsize >= BLOCK_8X8 &&
2427 mi_col + (mi_step >> 1) < cm->mi_cols) {
2428 RD_COST tmp_rdc;
2429 PICK_MODE_CONTEXT *ctx = &pc_tree->vertical[0];
2430 vp9_rd_cost_init(&tmp_rdc);
2431 update_state(cpi, td, ctx, mi_row, mi_col, subsize, 0);
2432 encode_superblock(cpi, td, tp, 0, mi_row, mi_col, subsize, ctx);
2433 rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col + (mi_step >> 1),
2434 &tmp_rdc, subsize,
2435 &pc_tree->vertical[bsize > BLOCK_8X8], INT64_MAX);
2436 if (tmp_rdc.rate == INT_MAX || tmp_rdc.dist == INT64_MAX) {
2437 vp9_rd_cost_reset(&last_part_rdc);
2438 break;
2439 }
2440 last_part_rdc.rate += tmp_rdc.rate;
2441 last_part_rdc.dist += tmp_rdc.dist;
2442 last_part_rdc.rdcost += tmp_rdc.rdcost;
2443 }
2444 break;
2445 case PARTITION_SPLIT:
2446 if (bsize == BLOCK_8X8) {
2447 rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &last_part_rdc,
2448 subsize, pc_tree->leaf_split[0], INT64_MAX);
2449 break;
2450 }
2451 last_part_rdc.rate = 0;
2452 last_part_rdc.dist = 0;
2453 last_part_rdc.rdcost = 0;
2454 for (i = 0; i < 4; i++) {
2455 int x_idx = (i & 1) * (mi_step >> 1);
2456 int y_idx = (i >> 1) * (mi_step >> 1);
2457 int jj = i >> 1, ii = i & 0x01;
2458 RD_COST tmp_rdc;
2459 if ((mi_row + y_idx >= cm->mi_rows) || (mi_col + x_idx >= cm->mi_cols))
2460 continue;
2461
2462 vp9_rd_cost_init(&tmp_rdc);
2463 rd_use_partition(cpi, td, tile_data, mi_8x8 + jj * bss * mis + ii * bss,
2464 tp, mi_row + y_idx, mi_col + x_idx, subsize,
2465 &tmp_rdc.rate, &tmp_rdc.dist, i != 3,
2466 pc_tree->split[i]);
2467 if (tmp_rdc.rate == INT_MAX || tmp_rdc.dist == INT64_MAX) {
2468 vp9_rd_cost_reset(&last_part_rdc);
2469 break;
2470 }
2471 last_part_rdc.rate += tmp_rdc.rate;
2472 last_part_rdc.dist += tmp_rdc.dist;
2473 }
2474 break;
2475 default: assert(0); break;
2476 }
2477
2478 pl = partition_plane_context(xd, mi_row, mi_col, bsize);
2479 if (last_part_rdc.rate < INT_MAX) {
2480 last_part_rdc.rate += cpi->partition_cost[pl][partition];
2481 last_part_rdc.rdcost =
2482 RDCOST(x->rdmult, x->rddiv, last_part_rdc.rate, last_part_rdc.dist);
2483 }
2484
2485 if (do_partition_search && cpi->sf.adjust_partitioning_from_last_frame &&
2486 cpi->sf.partition_search_type == SEARCH_PARTITION &&
2487 partition != PARTITION_SPLIT && bsize > BLOCK_8X8 &&
2488 (mi_row + mi_step < cm->mi_rows ||
2489 mi_row + (mi_step >> 1) == cm->mi_rows) &&
2490 (mi_col + mi_step < cm->mi_cols ||
2491 mi_col + (mi_step >> 1) == cm->mi_cols)) {
2492 BLOCK_SIZE split_subsize = get_subsize(bsize, PARTITION_SPLIT);
2493 chosen_rdc.rate = 0;
2494 chosen_rdc.dist = 0;
2495 restore_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
2496 pc_tree->partitioning = PARTITION_SPLIT;
2497
2498 // Split partition.
2499 for (i = 0; i < 4; i++) {
2500 int x_idx = (i & 1) * (mi_step >> 1);
2501 int y_idx = (i >> 1) * (mi_step >> 1);
2502 RD_COST tmp_rdc;
2503 ENTROPY_CONTEXT l[16 * MAX_MB_PLANE], a[16 * MAX_MB_PLANE];
2504 PARTITION_CONTEXT sl[8], sa[8];
2505
2506 if ((mi_row + y_idx >= cm->mi_rows) || (mi_col + x_idx >= cm->mi_cols))
2507 continue;
2508
2509 save_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
2510 pc_tree->split[i]->partitioning = PARTITION_NONE;
2511 rd_pick_sb_modes(cpi, tile_data, x, mi_row + y_idx, mi_col + x_idx,
2512 &tmp_rdc, split_subsize, &pc_tree->split[i]->none,
2513 INT64_MAX);
2514
2515 restore_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
2516
2517 if (tmp_rdc.rate == INT_MAX || tmp_rdc.dist == INT64_MAX) {
2518 vp9_rd_cost_reset(&chosen_rdc);
2519 break;
2520 }
2521
2522 chosen_rdc.rate += tmp_rdc.rate;
2523 chosen_rdc.dist += tmp_rdc.dist;
2524
2525 if (i != 3)
2526 encode_sb(cpi, td, tile_info, tp, mi_row + y_idx, mi_col + x_idx, 0,
2527 split_subsize, pc_tree->split[i]);
2528
2529 pl = partition_plane_context(xd, mi_row + y_idx, mi_col + x_idx,
2530 split_subsize);
2531 chosen_rdc.rate += cpi->partition_cost[pl][PARTITION_NONE];
2532 }
2533 pl = partition_plane_context(xd, mi_row, mi_col, bsize);
2534 if (chosen_rdc.rate < INT_MAX) {
2535 chosen_rdc.rate += cpi->partition_cost[pl][PARTITION_SPLIT];
2536 chosen_rdc.rdcost =
2537 RDCOST(x->rdmult, x->rddiv, chosen_rdc.rate, chosen_rdc.dist);
2538 }
2539 }
2540
2541 // If last_part is better set the partitioning to that.
2542 if (last_part_rdc.rdcost < chosen_rdc.rdcost) {
2543 mi_8x8[0]->sb_type = bsize;
2544 if (bsize >= BLOCK_8X8) pc_tree->partitioning = partition;
2545 chosen_rdc = last_part_rdc;
2546 }
2547 // If none was better set the partitioning to that.
2548 if (none_rdc.rdcost < chosen_rdc.rdcost) {
2549 if (bsize >= BLOCK_8X8) pc_tree->partitioning = PARTITION_NONE;
2550 chosen_rdc = none_rdc;
2551 }
2552
2553 restore_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
2554
2555 // We must have chosen a partitioning and encoding or we'll fail later on.
2556 // No other opportunities for success.
2557 if (bsize == BLOCK_64X64)
2558 assert(chosen_rdc.rate < INT_MAX && chosen_rdc.dist < INT64_MAX);
2559
2560 if (do_recon) {
2561 int output_enabled = (bsize == BLOCK_64X64);
2562 encode_sb(cpi, td, tile_info, tp, mi_row, mi_col, output_enabled, bsize,
2563 pc_tree);
2564 }
2565
2566 *rate = chosen_rdc.rate;
2567 *dist = chosen_rdc.dist;
2568 }
2569
2570 static const BLOCK_SIZE min_partition_size[BLOCK_SIZES] = {
2571 BLOCK_4X4, BLOCK_4X4, BLOCK_4X4, BLOCK_4X4, BLOCK_4X4,
2572 BLOCK_4X4, BLOCK_8X8, BLOCK_8X8, BLOCK_8X8, BLOCK_16X16,
2573 BLOCK_16X16, BLOCK_16X16, BLOCK_16X16
2574 };
2575
2576 static const BLOCK_SIZE max_partition_size[BLOCK_SIZES] = {
2577 BLOCK_8X8, BLOCK_16X16, BLOCK_16X16, BLOCK_16X16, BLOCK_32X32,
2578 BLOCK_32X32, BLOCK_32X32, BLOCK_64X64, BLOCK_64X64, BLOCK_64X64,
2579 BLOCK_64X64, BLOCK_64X64, BLOCK_64X64
2580 };
2581
2582 // Look at all the mode_info entries for blocks that are part of this
2583 // partition and find the min and max values for sb_type.
2584 // At the moment this is designed to work on a 64x64 SB but could be
2585 // adjusted to use a size parameter.
2586 //
2587 // The min and max are assumed to have been initialized prior to calling this
2588 // 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])2589 static void get_sb_partition_size_range(MACROBLOCKD *xd, MODE_INFO **mi_8x8,
2590 BLOCK_SIZE *min_block_size,
2591 BLOCK_SIZE *max_block_size,
2592 int bs_hist[BLOCK_SIZES]) {
2593 int sb_width_in_blocks = MI_BLOCK_SIZE;
2594 int sb_height_in_blocks = MI_BLOCK_SIZE;
2595 int i, j;
2596 int index = 0;
2597
2598 // Check the sb_type for each block that belongs to this region.
2599 for (i = 0; i < sb_height_in_blocks; ++i) {
2600 for (j = 0; j < sb_width_in_blocks; ++j) {
2601 MODE_INFO *mi = mi_8x8[index + j];
2602 BLOCK_SIZE sb_type = mi ? mi->sb_type : 0;
2603 bs_hist[sb_type]++;
2604 *min_block_size = VPXMIN(*min_block_size, sb_type);
2605 *max_block_size = VPXMAX(*max_block_size, sb_type);
2606 }
2607 index += xd->mi_stride;
2608 }
2609 }
2610
2611 // Next square block size less or equal than current block size.
2612 static const BLOCK_SIZE next_square_size[BLOCK_SIZES] = {
2613 BLOCK_4X4, BLOCK_4X4, BLOCK_4X4, BLOCK_8X8, BLOCK_8X8,
2614 BLOCK_8X8, BLOCK_16X16, BLOCK_16X16, BLOCK_16X16, BLOCK_32X32,
2615 BLOCK_32X32, BLOCK_32X32, BLOCK_64X64
2616 };
2617
2618 // Look at neighboring blocks and set a min and max partition size based on
2619 // 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)2620 static void rd_auto_partition_range(VP9_COMP *cpi, const TileInfo *const tile,
2621 MACROBLOCKD *const xd, int mi_row,
2622 int mi_col, BLOCK_SIZE *min_block_size,
2623 BLOCK_SIZE *max_block_size) {
2624 VP9_COMMON *const cm = &cpi->common;
2625 MODE_INFO **mi = xd->mi;
2626 const int left_in_image = !!xd->left_mi;
2627 const int above_in_image = !!xd->above_mi;
2628 const int row8x8_remaining = tile->mi_row_end - mi_row;
2629 const int col8x8_remaining = tile->mi_col_end - mi_col;
2630 int bh, bw;
2631 BLOCK_SIZE min_size = BLOCK_4X4;
2632 BLOCK_SIZE max_size = BLOCK_64X64;
2633 int bs_hist[BLOCK_SIZES] = { 0 };
2634
2635 // Trap case where we do not have a prediction.
2636 if (left_in_image || above_in_image || cm->frame_type != KEY_FRAME) {
2637 // Default "min to max" and "max to min"
2638 min_size = BLOCK_64X64;
2639 max_size = BLOCK_4X4;
2640
2641 // NOTE: each call to get_sb_partition_size_range() uses the previous
2642 // passed in values for min and max as a starting point.
2643 // Find the min and max partition used in previous frame at this location
2644 if (cm->frame_type != KEY_FRAME) {
2645 MODE_INFO **prev_mi =
2646 &cm->prev_mi_grid_visible[mi_row * xd->mi_stride + mi_col];
2647 get_sb_partition_size_range(xd, prev_mi, &min_size, &max_size, bs_hist);
2648 }
2649 // Find the min and max partition sizes used in the left SB64
2650 if (left_in_image) {
2651 MODE_INFO **left_sb64_mi = &mi[-MI_BLOCK_SIZE];
2652 get_sb_partition_size_range(xd, left_sb64_mi, &min_size, &max_size,
2653 bs_hist);
2654 }
2655 // Find the min and max partition sizes used in the above SB64.
2656 if (above_in_image) {
2657 MODE_INFO **above_sb64_mi = &mi[-xd->mi_stride * MI_BLOCK_SIZE];
2658 get_sb_partition_size_range(xd, above_sb64_mi, &min_size, &max_size,
2659 bs_hist);
2660 }
2661
2662 // Adjust observed min and max for "relaxed" auto partition case.
2663 if (cpi->sf.auto_min_max_partition_size == RELAXED_NEIGHBORING_MIN_MAX) {
2664 min_size = min_partition_size[min_size];
2665 max_size = max_partition_size[max_size];
2666 }
2667 }
2668
2669 // Check border cases where max and min from neighbors may not be legal.
2670 max_size = find_partition_size(max_size, row8x8_remaining, col8x8_remaining,
2671 &bh, &bw);
2672 // Test for blocks at the edge of the active image.
2673 // This may be the actual edge of the image or where there are formatting
2674 // bars.
2675 if (vp9_active_edge_sb(cpi, mi_row, mi_col)) {
2676 min_size = BLOCK_4X4;
2677 } else {
2678 min_size =
2679 VPXMIN(cpi->sf.rd_auto_partition_min_limit, VPXMIN(min_size, max_size));
2680 }
2681
2682 // When use_square_partition_only is true, make sure at least one square
2683 // partition is allowed by selecting the next smaller square size as
2684 // *min_block_size.
2685 if (cpi->sf.use_square_partition_only &&
2686 next_square_size[max_size] < min_size) {
2687 min_size = next_square_size[max_size];
2688 }
2689
2690 *min_block_size = min_size;
2691 *max_block_size = max_size;
2692 }
2693
2694 // 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)2695 static void set_partition_range(VP9_COMMON *cm, MACROBLOCKD *xd, int mi_row,
2696 int mi_col, BLOCK_SIZE bsize,
2697 BLOCK_SIZE *min_bs, BLOCK_SIZE *max_bs) {
2698 int mi_width = num_8x8_blocks_wide_lookup[bsize];
2699 int mi_height = num_8x8_blocks_high_lookup[bsize];
2700 int idx, idy;
2701
2702 MODE_INFO *mi;
2703 const int idx_str = cm->mi_stride * mi_row + mi_col;
2704 MODE_INFO **prev_mi = &cm->prev_mi_grid_visible[idx_str];
2705 BLOCK_SIZE bs, min_size, max_size;
2706
2707 min_size = BLOCK_64X64;
2708 max_size = BLOCK_4X4;
2709
2710 if (prev_mi) {
2711 for (idy = 0; idy < mi_height; ++idy) {
2712 for (idx = 0; idx < mi_width; ++idx) {
2713 mi = prev_mi[idy * cm->mi_stride + idx];
2714 bs = mi ? mi->sb_type : bsize;
2715 min_size = VPXMIN(min_size, bs);
2716 max_size = VPXMAX(max_size, bs);
2717 }
2718 }
2719 }
2720
2721 if (xd->left_mi) {
2722 for (idy = 0; idy < mi_height; ++idy) {
2723 mi = xd->mi[idy * cm->mi_stride - 1];
2724 bs = mi ? mi->sb_type : bsize;
2725 min_size = VPXMIN(min_size, bs);
2726 max_size = VPXMAX(max_size, bs);
2727 }
2728 }
2729
2730 if (xd->above_mi) {
2731 for (idx = 0; idx < mi_width; ++idx) {
2732 mi = xd->mi[idx - cm->mi_stride];
2733 bs = mi ? mi->sb_type : bsize;
2734 min_size = VPXMIN(min_size, bs);
2735 max_size = VPXMAX(max_size, bs);
2736 }
2737 }
2738
2739 if (min_size == max_size) {
2740 min_size = min_partition_size[min_size];
2741 max_size = max_partition_size[max_size];
2742 }
2743
2744 *min_bs = min_size;
2745 *max_bs = max_size;
2746 }
2747
store_pred_mv(MACROBLOCK * x,PICK_MODE_CONTEXT * ctx)2748 static INLINE void store_pred_mv(MACROBLOCK *x, PICK_MODE_CONTEXT *ctx) {
2749 memcpy(ctx->pred_mv, x->pred_mv, sizeof(x->pred_mv));
2750 }
2751
load_pred_mv(MACROBLOCK * x,PICK_MODE_CONTEXT * ctx)2752 static INLINE void load_pred_mv(MACROBLOCK *x, PICK_MODE_CONTEXT *ctx) {
2753 memcpy(x->pred_mv, ctx->pred_mv, sizeof(x->pred_mv));
2754 }
2755
2756 #if CONFIG_FP_MB_STATS
2757 const int num_16x16_blocks_wide_lookup[BLOCK_SIZES] = { 1, 1, 1, 1, 1, 1, 1,
2758 1, 2, 2, 2, 4, 4 };
2759 const int num_16x16_blocks_high_lookup[BLOCK_SIZES] = { 1, 1, 1, 1, 1, 1, 1,
2760 2, 1, 2, 4, 2, 4 };
2761 const int qindex_skip_threshold_lookup[BLOCK_SIZES] = {
2762 0, 10, 10, 30, 40, 40, 60, 80, 80, 90, 100, 100, 120
2763 };
2764 const int qindex_split_threshold_lookup[BLOCK_SIZES] = {
2765 0, 3, 3, 7, 15, 15, 30, 40, 40, 60, 80, 80, 120
2766 };
2767 const int complexity_16x16_blocks_threshold[BLOCK_SIZES] = {
2768 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 4, 4, 6
2769 };
2770
2771 typedef enum {
2772 MV_ZERO = 0,
2773 MV_LEFT = 1,
2774 MV_UP = 2,
2775 MV_RIGHT = 3,
2776 MV_DOWN = 4,
2777 MV_INVALID
2778 } MOTION_DIRECTION;
2779
get_motion_direction_fp(uint8_t fp_byte)2780 static INLINE MOTION_DIRECTION get_motion_direction_fp(uint8_t fp_byte) {
2781 if (fp_byte & FPMB_MOTION_ZERO_MASK) {
2782 return MV_ZERO;
2783 } else if (fp_byte & FPMB_MOTION_LEFT_MASK) {
2784 return MV_LEFT;
2785 } else if (fp_byte & FPMB_MOTION_RIGHT_MASK) {
2786 return MV_RIGHT;
2787 } else if (fp_byte & FPMB_MOTION_UP_MASK) {
2788 return MV_UP;
2789 } else {
2790 return MV_DOWN;
2791 }
2792 }
2793
get_motion_inconsistency(MOTION_DIRECTION this_mv,MOTION_DIRECTION that_mv)2794 static INLINE int get_motion_inconsistency(MOTION_DIRECTION this_mv,
2795 MOTION_DIRECTION that_mv) {
2796 if (this_mv == that_mv) {
2797 return 0;
2798 } else {
2799 return abs(this_mv - that_mv) == 2 ? 2 : 1;
2800 }
2801 }
2802 #endif
2803
2804 // Calculate the score used in machine-learning based partition search early
2805 // termination.
compute_score(VP9_COMMON * const cm,MACROBLOCKD * const xd,PICK_MODE_CONTEXT * ctx,int mi_row,int mi_col,BLOCK_SIZE bsize)2806 static double compute_score(VP9_COMMON *const cm, MACROBLOCKD *const xd,
2807 PICK_MODE_CONTEXT *ctx, int mi_row, int mi_col,
2808 BLOCK_SIZE bsize) {
2809 const double *clf;
2810 const double *mean;
2811 const double *sd;
2812 const int mag_mv =
2813 abs(ctx->mic.mv[0].as_mv.col) + abs(ctx->mic.mv[0].as_mv.row);
2814 const int left_in_image = !!xd->left_mi;
2815 const int above_in_image = !!xd->above_mi;
2816 MODE_INFO **prev_mi =
2817 &cm->prev_mi_grid_visible[mi_col + cm->mi_stride * mi_row];
2818 int above_par = 0; // above_partitioning
2819 int left_par = 0; // left_partitioning
2820 int last_par = 0; // last_partitioning
2821 BLOCK_SIZE context_size;
2822 double score;
2823 int offset = 0;
2824
2825 assert(b_width_log2_lookup[bsize] == b_height_log2_lookup[bsize]);
2826
2827 if (above_in_image) {
2828 context_size = xd->above_mi->sb_type;
2829 if (context_size < bsize)
2830 above_par = 2;
2831 else if (context_size == bsize)
2832 above_par = 1;
2833 }
2834
2835 if (left_in_image) {
2836 context_size = xd->left_mi->sb_type;
2837 if (context_size < bsize)
2838 left_par = 2;
2839 else if (context_size == bsize)
2840 left_par = 1;
2841 }
2842
2843 if (prev_mi) {
2844 context_size = prev_mi[0]->sb_type;
2845 if (context_size < bsize)
2846 last_par = 2;
2847 else if (context_size == bsize)
2848 last_par = 1;
2849 }
2850
2851 if (bsize == BLOCK_64X64)
2852 offset = 0;
2853 else if (bsize == BLOCK_32X32)
2854 offset = 8;
2855 else if (bsize == BLOCK_16X16)
2856 offset = 16;
2857
2858 // early termination score calculation
2859 clf = &classifiers[offset];
2860 mean = &train_mean[offset];
2861 sd = &train_stdm[offset];
2862 score = clf[0] * (((double)ctx->rate - mean[0]) / sd[0]) +
2863 clf[1] * (((double)ctx->dist - mean[1]) / sd[1]) +
2864 clf[2] * (((double)mag_mv / 2 - mean[2]) * sd[2]) +
2865 clf[3] * (((double)(left_par + above_par) / 2 - mean[3]) * sd[3]) +
2866 clf[4] * (((double)ctx->sum_y_eobs - mean[4]) / sd[4]) +
2867 clf[5] * (((double)cm->base_qindex - mean[5]) * sd[5]) +
2868 clf[6] * (((double)last_par - mean[6]) * sd[6]) + clf[7];
2869 return score;
2870 }
2871
2872 // TODO(jingning,jimbankoski,rbultje): properly skip partition types that are
2873 // unlikely to be selected depending on previous rate-distortion optimization
2874 // 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)2875 static void rd_pick_partition(VP9_COMP *cpi, ThreadData *td,
2876 TileDataEnc *tile_data, TOKENEXTRA **tp,
2877 int mi_row, int mi_col, BLOCK_SIZE bsize,
2878 RD_COST *rd_cost, int64_t best_rd,
2879 PC_TREE *pc_tree) {
2880 VP9_COMMON *const cm = &cpi->common;
2881 TileInfo *const tile_info = &tile_data->tile_info;
2882 MACROBLOCK *const x = &td->mb;
2883 MACROBLOCKD *const xd = &x->e_mbd;
2884 const int mi_step = num_8x8_blocks_wide_lookup[bsize] / 2;
2885 ENTROPY_CONTEXT l[16 * MAX_MB_PLANE], a[16 * MAX_MB_PLANE];
2886 PARTITION_CONTEXT sl[8], sa[8];
2887 TOKENEXTRA *tp_orig = *tp;
2888 PICK_MODE_CONTEXT *ctx = &pc_tree->none;
2889 int i;
2890 const int pl = partition_plane_context(xd, mi_row, mi_col, bsize);
2891 BLOCK_SIZE subsize;
2892 RD_COST this_rdc, sum_rdc, best_rdc;
2893 int do_split = bsize >= BLOCK_8X8;
2894 int do_rect = 1;
2895 INTERP_FILTER pred_interp_filter;
2896
2897 // Override skipping rectangular partition operations for edge blocks
2898 const int force_horz_split = (mi_row + mi_step >= cm->mi_rows);
2899 const int force_vert_split = (mi_col + mi_step >= cm->mi_cols);
2900 const int xss = x->e_mbd.plane[1].subsampling_x;
2901 const int yss = x->e_mbd.plane[1].subsampling_y;
2902
2903 BLOCK_SIZE min_size = x->min_partition_size;
2904 BLOCK_SIZE max_size = x->max_partition_size;
2905
2906 #if CONFIG_FP_MB_STATS
2907 unsigned int src_diff_var = UINT_MAX;
2908 int none_complexity = 0;
2909 #endif
2910
2911 int partition_none_allowed = !force_horz_split && !force_vert_split;
2912 int partition_horz_allowed =
2913 !force_vert_split && yss <= xss && bsize >= BLOCK_8X8;
2914 int partition_vert_allowed =
2915 !force_horz_split && xss <= yss && bsize >= BLOCK_8X8;
2916
2917 int64_t dist_breakout_thr = cpi->sf.partition_search_breakout_thr.dist;
2918 int rate_breakout_thr = cpi->sf.partition_search_breakout_thr.rate;
2919
2920 (void)*tp_orig;
2921
2922 assert(num_8x8_blocks_wide_lookup[bsize] ==
2923 num_8x8_blocks_high_lookup[bsize]);
2924
2925 // Adjust dist breakout threshold according to the partition size.
2926 dist_breakout_thr >>=
2927 8 - (b_width_log2_lookup[bsize] + b_height_log2_lookup[bsize]);
2928 rate_breakout_thr *= num_pels_log2_lookup[bsize];
2929
2930 vp9_rd_cost_init(&this_rdc);
2931 vp9_rd_cost_init(&sum_rdc);
2932 vp9_rd_cost_reset(&best_rdc);
2933 best_rdc.rdcost = best_rd;
2934
2935 set_offsets(cpi, tile_info, x, mi_row, mi_col, bsize);
2936
2937 if (bsize == BLOCK_16X16 && cpi->oxcf.aq_mode != NO_AQ &&
2938 cpi->oxcf.aq_mode != LOOKAHEAD_AQ)
2939 x->mb_energy = vp9_block_energy(cpi, x, bsize);
2940
2941 if (cpi->sf.cb_partition_search && bsize == BLOCK_16X16) {
2942 int cb_partition_search_ctrl =
2943 ((pc_tree->index == 0 || pc_tree->index == 3) +
2944 get_chessboard_index(cm->current_video_frame)) &
2945 0x1;
2946
2947 if (cb_partition_search_ctrl && bsize > min_size && bsize < max_size)
2948 set_partition_range(cm, xd, mi_row, mi_col, bsize, &min_size, &max_size);
2949 }
2950
2951 // Determine partition types in search according to the speed features.
2952 // The threshold set here has to be of square block size.
2953 if (cpi->sf.auto_min_max_partition_size) {
2954 partition_none_allowed &= (bsize <= max_size && bsize >= min_size);
2955 partition_horz_allowed &=
2956 ((bsize <= max_size && bsize > min_size) || force_horz_split);
2957 partition_vert_allowed &=
2958 ((bsize <= max_size && bsize > min_size) || force_vert_split);
2959 do_split &= bsize > min_size;
2960 }
2961
2962 if (cpi->sf.use_square_partition_only &&
2963 bsize > cpi->sf.use_square_only_threshold) {
2964 if (cpi->use_svc) {
2965 if (!vp9_active_h_edge(cpi, mi_row, mi_step) || x->e_mbd.lossless)
2966 partition_horz_allowed &= force_horz_split;
2967 if (!vp9_active_v_edge(cpi, mi_row, mi_step) || x->e_mbd.lossless)
2968 partition_vert_allowed &= force_vert_split;
2969 } else {
2970 partition_horz_allowed &= force_horz_split;
2971 partition_vert_allowed &= force_vert_split;
2972 }
2973 }
2974
2975 save_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
2976
2977 #if CONFIG_FP_MB_STATS
2978 if (cpi->use_fp_mb_stats) {
2979 set_offsets(cpi, tile_info, x, mi_row, mi_col, bsize);
2980 src_diff_var = get_sby_perpixel_diff_variance(cpi, &x->plane[0].src, mi_row,
2981 mi_col, bsize);
2982 }
2983 #endif
2984
2985 #if CONFIG_FP_MB_STATS
2986 // Decide whether we shall split directly and skip searching NONE by using
2987 // the first pass block statistics
2988 if (cpi->use_fp_mb_stats && bsize >= BLOCK_32X32 && do_split &&
2989 partition_none_allowed && src_diff_var > 4 &&
2990 cm->base_qindex < qindex_split_threshold_lookup[bsize]) {
2991 int mb_row = mi_row >> 1;
2992 int mb_col = mi_col >> 1;
2993 int mb_row_end =
2994 VPXMIN(mb_row + num_16x16_blocks_high_lookup[bsize], cm->mb_rows);
2995 int mb_col_end =
2996 VPXMIN(mb_col + num_16x16_blocks_wide_lookup[bsize], cm->mb_cols);
2997 int r, c;
2998
2999 // compute a complexity measure, basically measure inconsistency of motion
3000 // vectors obtained from the first pass in the current block
3001 for (r = mb_row; r < mb_row_end; r++) {
3002 for (c = mb_col; c < mb_col_end; c++) {
3003 const int mb_index = r * cm->mb_cols + c;
3004
3005 MOTION_DIRECTION this_mv;
3006 MOTION_DIRECTION right_mv;
3007 MOTION_DIRECTION bottom_mv;
3008
3009 this_mv =
3010 get_motion_direction_fp(cpi->twopass.this_frame_mb_stats[mb_index]);
3011
3012 // to its right
3013 if (c != mb_col_end - 1) {
3014 right_mv = get_motion_direction_fp(
3015 cpi->twopass.this_frame_mb_stats[mb_index + 1]);
3016 none_complexity += get_motion_inconsistency(this_mv, right_mv);
3017 }
3018
3019 // to its bottom
3020 if (r != mb_row_end - 1) {
3021 bottom_mv = get_motion_direction_fp(
3022 cpi->twopass.this_frame_mb_stats[mb_index + cm->mb_cols]);
3023 none_complexity += get_motion_inconsistency(this_mv, bottom_mv);
3024 }
3025
3026 // do not count its left and top neighbors to avoid double counting
3027 }
3028 }
3029
3030 if (none_complexity > complexity_16x16_blocks_threshold[bsize]) {
3031 partition_none_allowed = 0;
3032 }
3033 }
3034 #endif
3035
3036 // PARTITION_NONE
3037 if (partition_none_allowed) {
3038 rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &this_rdc, bsize, ctx,
3039 best_rdc.rdcost);
3040 if (this_rdc.rate != INT_MAX) {
3041 if (bsize >= BLOCK_8X8) {
3042 this_rdc.rate += cpi->partition_cost[pl][PARTITION_NONE];
3043 this_rdc.rdcost =
3044 RDCOST(x->rdmult, x->rddiv, this_rdc.rate, this_rdc.dist);
3045 }
3046
3047 if (this_rdc.rdcost < best_rdc.rdcost) {
3048 MODE_INFO *mi = xd->mi[0];
3049
3050 best_rdc = this_rdc;
3051 if (bsize >= BLOCK_8X8) pc_tree->partitioning = PARTITION_NONE;
3052
3053 if (!cpi->sf.ml_partition_search_early_termination) {
3054 // If all y, u, v transform blocks in this partition are skippable,
3055 // and the dist & rate are within the thresholds, the partition search
3056 // is terminated for current branch of the partition search tree.
3057 if (!x->e_mbd.lossless && ctx->skippable &&
3058 ((best_rdc.dist < (dist_breakout_thr >> 2)) ||
3059 (best_rdc.dist < dist_breakout_thr &&
3060 best_rdc.rate < rate_breakout_thr))) {
3061 do_split = 0;
3062 do_rect = 0;
3063 }
3064 } else {
3065 // Currently, the machine-learning based partition search early
3066 // termination is only used while bsize is 16x16, 32x32 or 64x64,
3067 // VPXMIN(cm->width, cm->height) >= 480, and speed = 0.
3068 if (!x->e_mbd.lossless &&
3069 !segfeature_active(&cm->seg, mi->segment_id, SEG_LVL_SKIP) &&
3070 ctx->mic.mode >= INTRA_MODES && bsize >= BLOCK_16X16) {
3071 if (compute_score(cm, xd, ctx, mi_row, mi_col, bsize) < 0.0) {
3072 do_split = 0;
3073 do_rect = 0;
3074 }
3075 }
3076 }
3077
3078 #if CONFIG_FP_MB_STATS
3079 // Check if every 16x16 first pass block statistics has zero
3080 // motion and the corresponding first pass residue is small enough.
3081 // If that is the case, check the difference variance between the
3082 // current frame and the last frame. If the variance is small enough,
3083 // stop further splitting in RD optimization
3084 if (cpi->use_fp_mb_stats && do_split != 0 &&
3085 cm->base_qindex > qindex_skip_threshold_lookup[bsize]) {
3086 int mb_row = mi_row >> 1;
3087 int mb_col = mi_col >> 1;
3088 int mb_row_end =
3089 VPXMIN(mb_row + num_16x16_blocks_high_lookup[bsize], cm->mb_rows);
3090 int mb_col_end =
3091 VPXMIN(mb_col + num_16x16_blocks_wide_lookup[bsize], cm->mb_cols);
3092 int r, c;
3093
3094 int skip = 1;
3095 for (r = mb_row; r < mb_row_end; r++) {
3096 for (c = mb_col; c < mb_col_end; c++) {
3097 const int mb_index = r * cm->mb_cols + c;
3098 if (!(cpi->twopass.this_frame_mb_stats[mb_index] &
3099 FPMB_MOTION_ZERO_MASK) ||
3100 !(cpi->twopass.this_frame_mb_stats[mb_index] &
3101 FPMB_ERROR_SMALL_MASK)) {
3102 skip = 0;
3103 break;
3104 }
3105 }
3106 if (skip == 0) {
3107 break;
3108 }
3109 }
3110
3111 if (skip) {
3112 if (src_diff_var == UINT_MAX) {
3113 set_offsets(cpi, tile_info, x, mi_row, mi_col, bsize);
3114 src_diff_var = get_sby_perpixel_diff_variance(
3115 cpi, &x->plane[0].src, mi_row, mi_col, bsize);
3116 }
3117 if (src_diff_var < 8) {
3118 do_split = 0;
3119 do_rect = 0;
3120 }
3121 }
3122 }
3123 #endif
3124 }
3125 }
3126 restore_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
3127 }
3128
3129 // store estimated motion vector
3130 if (cpi->sf.adaptive_motion_search) store_pred_mv(x, ctx);
3131
3132 // If the interp_filter is marked as SWITCHABLE_FILTERS, it was for an
3133 // intra block and used for context purposes.
3134 if (ctx->mic.interp_filter == SWITCHABLE_FILTERS) {
3135 pred_interp_filter = EIGHTTAP;
3136 } else {
3137 pred_interp_filter = ctx->mic.interp_filter;
3138 }
3139
3140 // PARTITION_SPLIT
3141 // TODO(jingning): use the motion vectors given by the above search as
3142 // the starting point of motion search in the following partition type check.
3143 if (do_split) {
3144 subsize = get_subsize(bsize, PARTITION_SPLIT);
3145 if (bsize == BLOCK_8X8) {
3146 i = 4;
3147 if (cpi->sf.adaptive_pred_interp_filter && partition_none_allowed)
3148 pc_tree->leaf_split[0]->pred_interp_filter = pred_interp_filter;
3149 rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &sum_rdc, subsize,
3150 pc_tree->leaf_split[0], best_rdc.rdcost);
3151
3152 if (sum_rdc.rate == INT_MAX) sum_rdc.rdcost = INT64_MAX;
3153 } else {
3154 for (i = 0; i < 4 && sum_rdc.rdcost < best_rdc.rdcost; ++i) {
3155 const int x_idx = (i & 1) * mi_step;
3156 const int y_idx = (i >> 1) * mi_step;
3157
3158 if (mi_row + y_idx >= cm->mi_rows || mi_col + x_idx >= cm->mi_cols)
3159 continue;
3160
3161 if (cpi->sf.adaptive_motion_search) load_pred_mv(x, ctx);
3162
3163 pc_tree->split[i]->index = i;
3164 rd_pick_partition(cpi, td, tile_data, tp, mi_row + y_idx,
3165 mi_col + x_idx, subsize, &this_rdc,
3166 best_rdc.rdcost - sum_rdc.rdcost, pc_tree->split[i]);
3167
3168 if (this_rdc.rate == INT_MAX) {
3169 sum_rdc.rdcost = INT64_MAX;
3170 break;
3171 } else {
3172 sum_rdc.rate += this_rdc.rate;
3173 sum_rdc.dist += this_rdc.dist;
3174 sum_rdc.rdcost += this_rdc.rdcost;
3175 }
3176 }
3177 }
3178
3179 if (sum_rdc.rdcost < best_rdc.rdcost && i == 4) {
3180 sum_rdc.rate += cpi->partition_cost[pl][PARTITION_SPLIT];
3181 sum_rdc.rdcost = RDCOST(x->rdmult, x->rddiv, sum_rdc.rate, sum_rdc.dist);
3182
3183 if (sum_rdc.rdcost < best_rdc.rdcost) {
3184 best_rdc = sum_rdc;
3185 pc_tree->partitioning = PARTITION_SPLIT;
3186
3187 // Rate and distortion based partition search termination clause.
3188 if (!cpi->sf.ml_partition_search_early_termination &&
3189 !x->e_mbd.lossless && ((best_rdc.dist < (dist_breakout_thr >> 2)) ||
3190 (best_rdc.dist < dist_breakout_thr &&
3191 best_rdc.rate < rate_breakout_thr))) {
3192 do_rect = 0;
3193 }
3194 }
3195 } else {
3196 // skip rectangular partition test when larger block size
3197 // gives better rd cost
3198 if ((cpi->sf.less_rectangular_check) &&
3199 ((bsize > cpi->sf.use_square_only_threshold) ||
3200 (best_rdc.dist < dist_breakout_thr)))
3201 do_rect &= !partition_none_allowed;
3202 }
3203 restore_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
3204 }
3205
3206 // PARTITION_HORZ
3207 if (partition_horz_allowed &&
3208 (do_rect || vp9_active_h_edge(cpi, mi_row, mi_step))) {
3209 subsize = get_subsize(bsize, PARTITION_HORZ);
3210 if (cpi->sf.adaptive_motion_search) load_pred_mv(x, ctx);
3211 if (cpi->sf.adaptive_pred_interp_filter && bsize == BLOCK_8X8 &&
3212 partition_none_allowed)
3213 pc_tree->horizontal[0].pred_interp_filter = pred_interp_filter;
3214 rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &sum_rdc, subsize,
3215 &pc_tree->horizontal[0], best_rdc.rdcost);
3216
3217 if (sum_rdc.rdcost < best_rdc.rdcost && mi_row + mi_step < cm->mi_rows &&
3218 bsize > BLOCK_8X8) {
3219 PICK_MODE_CONTEXT *ctx = &pc_tree->horizontal[0];
3220 update_state(cpi, td, ctx, mi_row, mi_col, subsize, 0);
3221 encode_superblock(cpi, td, tp, 0, mi_row, mi_col, subsize, ctx);
3222
3223 if (cpi->sf.adaptive_motion_search) load_pred_mv(x, ctx);
3224 if (cpi->sf.adaptive_pred_interp_filter && bsize == BLOCK_8X8 &&
3225 partition_none_allowed)
3226 pc_tree->horizontal[1].pred_interp_filter = pred_interp_filter;
3227 rd_pick_sb_modes(cpi, tile_data, x, mi_row + mi_step, mi_col, &this_rdc,
3228 subsize, &pc_tree->horizontal[1],
3229 best_rdc.rdcost - sum_rdc.rdcost);
3230 if (this_rdc.rate == INT_MAX) {
3231 sum_rdc.rdcost = INT64_MAX;
3232 } else {
3233 sum_rdc.rate += this_rdc.rate;
3234 sum_rdc.dist += this_rdc.dist;
3235 sum_rdc.rdcost += this_rdc.rdcost;
3236 }
3237 }
3238
3239 if (sum_rdc.rdcost < best_rdc.rdcost) {
3240 sum_rdc.rate += cpi->partition_cost[pl][PARTITION_HORZ];
3241 sum_rdc.rdcost = RDCOST(x->rdmult, x->rddiv, sum_rdc.rate, sum_rdc.dist);
3242 if (sum_rdc.rdcost < best_rdc.rdcost) {
3243 best_rdc = sum_rdc;
3244 pc_tree->partitioning = PARTITION_HORZ;
3245
3246 if ((cpi->sf.less_rectangular_check) &&
3247 (bsize > cpi->sf.use_square_only_threshold))
3248 do_rect = 0;
3249 }
3250 }
3251 restore_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
3252 }
3253
3254 // PARTITION_VERT
3255 if (partition_vert_allowed &&
3256 (do_rect || vp9_active_v_edge(cpi, mi_col, mi_step))) {
3257 subsize = get_subsize(bsize, PARTITION_VERT);
3258
3259 if (cpi->sf.adaptive_motion_search) load_pred_mv(x, ctx);
3260 if (cpi->sf.adaptive_pred_interp_filter && bsize == BLOCK_8X8 &&
3261 partition_none_allowed)
3262 pc_tree->vertical[0].pred_interp_filter = pred_interp_filter;
3263 rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &sum_rdc, subsize,
3264 &pc_tree->vertical[0], best_rdc.rdcost);
3265 if (sum_rdc.rdcost < best_rdc.rdcost && mi_col + mi_step < cm->mi_cols &&
3266 bsize > BLOCK_8X8) {
3267 update_state(cpi, td, &pc_tree->vertical[0], mi_row, mi_col, subsize, 0);
3268 encode_superblock(cpi, td, tp, 0, mi_row, mi_col, subsize,
3269 &pc_tree->vertical[0]);
3270
3271 if (cpi->sf.adaptive_motion_search) load_pred_mv(x, ctx);
3272 if (cpi->sf.adaptive_pred_interp_filter && bsize == BLOCK_8X8 &&
3273 partition_none_allowed)
3274 pc_tree->vertical[1].pred_interp_filter = pred_interp_filter;
3275 rd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col + mi_step, &this_rdc,
3276 subsize, &pc_tree->vertical[1],
3277 best_rdc.rdcost - sum_rdc.rdcost);
3278 if (this_rdc.rate == INT_MAX) {
3279 sum_rdc.rdcost = INT64_MAX;
3280 } else {
3281 sum_rdc.rate += this_rdc.rate;
3282 sum_rdc.dist += this_rdc.dist;
3283 sum_rdc.rdcost += this_rdc.rdcost;
3284 }
3285 }
3286
3287 if (sum_rdc.rdcost < best_rdc.rdcost) {
3288 sum_rdc.rate += cpi->partition_cost[pl][PARTITION_VERT];
3289 sum_rdc.rdcost = RDCOST(x->rdmult, x->rddiv, sum_rdc.rate, sum_rdc.dist);
3290 if (sum_rdc.rdcost < best_rdc.rdcost) {
3291 best_rdc = sum_rdc;
3292 pc_tree->partitioning = PARTITION_VERT;
3293 }
3294 }
3295 restore_context(x, mi_row, mi_col, a, l, sa, sl, bsize);
3296 }
3297
3298 // TODO(jbb): This code added so that we avoid static analysis
3299 // warning related to the fact that best_rd isn't used after this
3300 // point. This code should be refactored so that the duplicate
3301 // checks occur in some sub function and thus are used...
3302 (void)best_rd;
3303 *rd_cost = best_rdc;
3304
3305 if (best_rdc.rate < INT_MAX && best_rdc.dist < INT64_MAX &&
3306 pc_tree->index != 3) {
3307 int output_enabled = (bsize == BLOCK_64X64);
3308 encode_sb(cpi, td, tile_info, tp, mi_row, mi_col, output_enabled, bsize,
3309 pc_tree);
3310 }
3311
3312 if (bsize == BLOCK_64X64) {
3313 assert(tp_orig < *tp);
3314 assert(best_rdc.rate < INT_MAX);
3315 assert(best_rdc.dist < INT64_MAX);
3316 } else {
3317 assert(tp_orig == *tp);
3318 }
3319 }
3320
encode_rd_sb_row(VP9_COMP * cpi,ThreadData * td,TileDataEnc * tile_data,int mi_row,TOKENEXTRA ** tp)3321 static void encode_rd_sb_row(VP9_COMP *cpi, ThreadData *td,
3322 TileDataEnc *tile_data, int mi_row,
3323 TOKENEXTRA **tp) {
3324 VP9_COMMON *const cm = &cpi->common;
3325 TileInfo *const tile_info = &tile_data->tile_info;
3326 MACROBLOCK *const x = &td->mb;
3327 MACROBLOCKD *const xd = &x->e_mbd;
3328 SPEED_FEATURES *const sf = &cpi->sf;
3329 const int mi_col_start = tile_info->mi_col_start;
3330 const int mi_col_end = tile_info->mi_col_end;
3331 int mi_col;
3332 const int sb_row = mi_row >> MI_BLOCK_SIZE_LOG2;
3333 const int num_sb_cols =
3334 get_num_cols(tile_data->tile_info, MI_BLOCK_SIZE_LOG2);
3335 int sb_col_in_tile;
3336
3337 // Initialize the left context for the new SB row
3338 memset(&xd->left_context, 0, sizeof(xd->left_context));
3339 memset(xd->left_seg_context, 0, sizeof(xd->left_seg_context));
3340
3341 // Code each SB in the row
3342 for (mi_col = mi_col_start, sb_col_in_tile = 0; mi_col < mi_col_end;
3343 mi_col += MI_BLOCK_SIZE, sb_col_in_tile++) {
3344 const struct segmentation *const seg = &cm->seg;
3345 int dummy_rate;
3346 int64_t dummy_dist;
3347 RD_COST dummy_rdc;
3348 int i;
3349 int seg_skip = 0;
3350
3351 const int idx_str = cm->mi_stride * mi_row + mi_col;
3352 MODE_INFO **mi = cm->mi_grid_visible + idx_str;
3353
3354 (*(cpi->row_mt_sync_read_ptr))(&tile_data->row_mt_sync, sb_row,
3355 sb_col_in_tile);
3356
3357 if (sf->adaptive_pred_interp_filter) {
3358 for (i = 0; i < 64; ++i) td->leaf_tree[i].pred_interp_filter = SWITCHABLE;
3359
3360 for (i = 0; i < 64; ++i) {
3361 td->pc_tree[i].vertical[0].pred_interp_filter = SWITCHABLE;
3362 td->pc_tree[i].vertical[1].pred_interp_filter = SWITCHABLE;
3363 td->pc_tree[i].horizontal[0].pred_interp_filter = SWITCHABLE;
3364 td->pc_tree[i].horizontal[1].pred_interp_filter = SWITCHABLE;
3365 }
3366 }
3367
3368 vp9_zero(x->pred_mv);
3369 td->pc_root->index = 0;
3370
3371 if (seg->enabled) {
3372 const uint8_t *const map =
3373 seg->update_map ? cpi->segmentation_map : cm->last_frame_seg_map;
3374 int segment_id = get_segment_id(cm, map, BLOCK_64X64, mi_row, mi_col);
3375 seg_skip = segfeature_active(seg, segment_id, SEG_LVL_SKIP);
3376 }
3377
3378 x->source_variance = UINT_MAX;
3379 if (sf->partition_search_type == FIXED_PARTITION || seg_skip) {
3380 const BLOCK_SIZE bsize =
3381 seg_skip ? BLOCK_64X64 : sf->always_this_block_size;
3382 set_offsets(cpi, tile_info, x, mi_row, mi_col, BLOCK_64X64);
3383 set_fixed_partitioning(cpi, tile_info, mi, mi_row, mi_col, bsize);
3384 rd_use_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col, BLOCK_64X64,
3385 &dummy_rate, &dummy_dist, 1, td->pc_root);
3386 } else if (cpi->partition_search_skippable_frame) {
3387 BLOCK_SIZE bsize;
3388 set_offsets(cpi, tile_info, x, mi_row, mi_col, BLOCK_64X64);
3389 bsize = get_rd_var_based_fixed_partition(cpi, x, mi_row, mi_col);
3390 set_fixed_partitioning(cpi, tile_info, mi, mi_row, mi_col, bsize);
3391 rd_use_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col, BLOCK_64X64,
3392 &dummy_rate, &dummy_dist, 1, td->pc_root);
3393 } else if (sf->partition_search_type == VAR_BASED_PARTITION &&
3394 cm->frame_type != KEY_FRAME) {
3395 choose_partitioning(cpi, tile_info, x, mi_row, mi_col);
3396 rd_use_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col, BLOCK_64X64,
3397 &dummy_rate, &dummy_dist, 1, td->pc_root);
3398 } else {
3399 // If required set upper and lower partition size limits
3400 if (sf->auto_min_max_partition_size) {
3401 set_offsets(cpi, tile_info, x, mi_row, mi_col, BLOCK_64X64);
3402 rd_auto_partition_range(cpi, tile_info, xd, mi_row, mi_col,
3403 &x->min_partition_size, &x->max_partition_size);
3404 }
3405 rd_pick_partition(cpi, td, tile_data, tp, mi_row, mi_col, BLOCK_64X64,
3406 &dummy_rdc, INT64_MAX, td->pc_root);
3407 }
3408 (*(cpi->row_mt_sync_write_ptr))(&tile_data->row_mt_sync, sb_row,
3409 sb_col_in_tile, num_sb_cols);
3410 }
3411 }
3412
init_encode_frame_mb_context(VP9_COMP * cpi)3413 static void init_encode_frame_mb_context(VP9_COMP *cpi) {
3414 MACROBLOCK *const x = &cpi->td.mb;
3415 VP9_COMMON *const cm = &cpi->common;
3416 MACROBLOCKD *const xd = &x->e_mbd;
3417 const int aligned_mi_cols = mi_cols_aligned_to_sb(cm->mi_cols);
3418
3419 // Copy data over into macro block data structures.
3420 vp9_setup_src_planes(x, cpi->Source, 0, 0);
3421
3422 vp9_setup_block_planes(&x->e_mbd, cm->subsampling_x, cm->subsampling_y);
3423
3424 // Note: this memset assumes above_context[0], [1] and [2]
3425 // are allocated as part of the same buffer.
3426 memset(xd->above_context[0], 0,
3427 sizeof(*xd->above_context[0]) * 2 * aligned_mi_cols * MAX_MB_PLANE);
3428 memset(xd->above_seg_context, 0,
3429 sizeof(*xd->above_seg_context) * aligned_mi_cols);
3430 }
3431
check_dual_ref_flags(VP9_COMP * cpi)3432 static int check_dual_ref_flags(VP9_COMP *cpi) {
3433 const int ref_flags = cpi->ref_frame_flags;
3434
3435 if (segfeature_active(&cpi->common.seg, 1, SEG_LVL_REF_FRAME)) {
3436 return 0;
3437 } else {
3438 return (!!(ref_flags & VP9_GOLD_FLAG) + !!(ref_flags & VP9_LAST_FLAG) +
3439 !!(ref_flags & VP9_ALT_FLAG)) >= 2;
3440 }
3441 }
3442
reset_skip_tx_size(VP9_COMMON * cm,TX_SIZE max_tx_size)3443 static void reset_skip_tx_size(VP9_COMMON *cm, TX_SIZE max_tx_size) {
3444 int mi_row, mi_col;
3445 const int mis = cm->mi_stride;
3446 MODE_INFO **mi_ptr = cm->mi_grid_visible;
3447
3448 for (mi_row = 0; mi_row < cm->mi_rows; ++mi_row, mi_ptr += mis) {
3449 for (mi_col = 0; mi_col < cm->mi_cols; ++mi_col) {
3450 if (mi_ptr[mi_col]->tx_size > max_tx_size)
3451 mi_ptr[mi_col]->tx_size = max_tx_size;
3452 }
3453 }
3454 }
3455
get_frame_type(const VP9_COMP * cpi)3456 static MV_REFERENCE_FRAME get_frame_type(const VP9_COMP *cpi) {
3457 if (frame_is_intra_only(&cpi->common))
3458 return INTRA_FRAME;
3459 else if (cpi->rc.is_src_frame_alt_ref && cpi->refresh_golden_frame)
3460 return ALTREF_FRAME;
3461 else if (cpi->refresh_golden_frame || cpi->refresh_alt_ref_frame)
3462 return GOLDEN_FRAME;
3463 else
3464 return LAST_FRAME;
3465 }
3466
select_tx_mode(const VP9_COMP * cpi,MACROBLOCKD * const xd)3467 static TX_MODE select_tx_mode(const VP9_COMP *cpi, MACROBLOCKD *const xd) {
3468 if (xd->lossless) return ONLY_4X4;
3469 if (cpi->common.frame_type == KEY_FRAME && cpi->sf.use_nonrd_pick_mode)
3470 return ALLOW_16X16;
3471 if (cpi->sf.tx_size_search_method == USE_LARGESTALL)
3472 return ALLOW_32X32;
3473 else if (cpi->sf.tx_size_search_method == USE_FULL_RD ||
3474 cpi->sf.tx_size_search_method == USE_TX_8X8)
3475 return TX_MODE_SELECT;
3476 else
3477 return cpi->common.tx_mode;
3478 }
3479
hybrid_intra_mode_search(VP9_COMP * cpi,MACROBLOCK * const x,RD_COST * rd_cost,BLOCK_SIZE bsize,PICK_MODE_CONTEXT * ctx)3480 static void hybrid_intra_mode_search(VP9_COMP *cpi, MACROBLOCK *const x,
3481 RD_COST *rd_cost, BLOCK_SIZE bsize,
3482 PICK_MODE_CONTEXT *ctx) {
3483 if (bsize < BLOCK_16X16)
3484 vp9_rd_pick_intra_mode_sb(cpi, x, rd_cost, bsize, ctx, INT64_MAX);
3485 else
3486 vp9_pick_intra_mode(cpi, x, rd_cost, bsize, ctx);
3487 }
3488
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)3489 static void nonrd_pick_sb_modes(VP9_COMP *cpi, TileDataEnc *tile_data,
3490 MACROBLOCK *const x, int mi_row, int mi_col,
3491 RD_COST *rd_cost, BLOCK_SIZE bsize,
3492 PICK_MODE_CONTEXT *ctx) {
3493 VP9_COMMON *const cm = &cpi->common;
3494 TileInfo *const tile_info = &tile_data->tile_info;
3495 MACROBLOCKD *const xd = &x->e_mbd;
3496 MODE_INFO *mi;
3497 ENTROPY_CONTEXT l[16 * MAX_MB_PLANE], a[16 * MAX_MB_PLANE];
3498 BLOCK_SIZE bs = VPXMAX(bsize, BLOCK_8X8); // processing unit block size
3499 const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bs];
3500 const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bs];
3501 int plane;
3502
3503 set_offsets(cpi, tile_info, x, mi_row, mi_col, bsize);
3504 mi = xd->mi[0];
3505 mi->sb_type = bsize;
3506
3507 for (plane = 0; plane < MAX_MB_PLANE; ++plane) {
3508 struct macroblockd_plane *pd = &xd->plane[plane];
3509 memcpy(a + num_4x4_blocks_wide * plane, pd->above_context,
3510 (sizeof(a[0]) * num_4x4_blocks_wide) >> pd->subsampling_x);
3511 memcpy(l + num_4x4_blocks_high * plane, pd->left_context,
3512 (sizeof(l[0]) * num_4x4_blocks_high) >> pd->subsampling_y);
3513 }
3514
3515 if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ && cm->seg.enabled)
3516 if (cyclic_refresh_segment_id_boosted(mi->segment_id))
3517 x->rdmult = vp9_cyclic_refresh_get_rdmult(cpi->cyclic_refresh);
3518
3519 if (cm->frame_type == KEY_FRAME)
3520 hybrid_intra_mode_search(cpi, x, rd_cost, bsize, ctx);
3521 else if (segfeature_active(&cm->seg, mi->segment_id, SEG_LVL_SKIP))
3522 set_mode_info_seg_skip(x, cm->tx_mode, rd_cost, bsize);
3523 else if (bsize >= BLOCK_8X8)
3524 vp9_pick_inter_mode(cpi, x, tile_data, mi_row, mi_col, rd_cost, bsize, ctx);
3525 else
3526 vp9_pick_inter_mode_sub8x8(cpi, x, mi_row, mi_col, rd_cost, bsize, ctx);
3527
3528 duplicate_mode_info_in_sb(cm, xd, mi_row, mi_col, bsize);
3529
3530 for (plane = 0; plane < MAX_MB_PLANE; ++plane) {
3531 struct macroblockd_plane *pd = &xd->plane[plane];
3532 memcpy(pd->above_context, a + num_4x4_blocks_wide * plane,
3533 (sizeof(a[0]) * num_4x4_blocks_wide) >> pd->subsampling_x);
3534 memcpy(pd->left_context, l + num_4x4_blocks_high * plane,
3535 (sizeof(l[0]) * num_4x4_blocks_high) >> pd->subsampling_y);
3536 }
3537
3538 if (rd_cost->rate == INT_MAX) vp9_rd_cost_reset(rd_cost);
3539
3540 ctx->rate = rd_cost->rate;
3541 ctx->dist = rd_cost->dist;
3542 }
3543
fill_mode_info_sb(VP9_COMMON * cm,MACROBLOCK * x,int mi_row,int mi_col,BLOCK_SIZE bsize,PC_TREE * pc_tree)3544 static void fill_mode_info_sb(VP9_COMMON *cm, MACROBLOCK *x, int mi_row,
3545 int mi_col, BLOCK_SIZE bsize, PC_TREE *pc_tree) {
3546 MACROBLOCKD *xd = &x->e_mbd;
3547 int bsl = b_width_log2_lookup[bsize], hbs = (1 << bsl) / 4;
3548 PARTITION_TYPE partition = pc_tree->partitioning;
3549 BLOCK_SIZE subsize = get_subsize(bsize, partition);
3550
3551 assert(bsize >= BLOCK_8X8);
3552
3553 if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols) return;
3554
3555 switch (partition) {
3556 case PARTITION_NONE:
3557 set_mode_info_offsets(cm, x, xd, mi_row, mi_col);
3558 *(xd->mi[0]) = pc_tree->none.mic;
3559 *(x->mbmi_ext) = pc_tree->none.mbmi_ext;
3560 duplicate_mode_info_in_sb(cm, xd, mi_row, mi_col, bsize);
3561 break;
3562 case PARTITION_VERT:
3563 set_mode_info_offsets(cm, x, xd, mi_row, mi_col);
3564 *(xd->mi[0]) = pc_tree->vertical[0].mic;
3565 *(x->mbmi_ext) = pc_tree->vertical[0].mbmi_ext;
3566 duplicate_mode_info_in_sb(cm, xd, mi_row, mi_col, subsize);
3567
3568 if (mi_col + hbs < cm->mi_cols) {
3569 set_mode_info_offsets(cm, x, xd, mi_row, mi_col + hbs);
3570 *(xd->mi[0]) = pc_tree->vertical[1].mic;
3571 *(x->mbmi_ext) = pc_tree->vertical[1].mbmi_ext;
3572 duplicate_mode_info_in_sb(cm, xd, mi_row, mi_col + hbs, subsize);
3573 }
3574 break;
3575 case PARTITION_HORZ:
3576 set_mode_info_offsets(cm, x, xd, mi_row, mi_col);
3577 *(xd->mi[0]) = pc_tree->horizontal[0].mic;
3578 *(x->mbmi_ext) = pc_tree->horizontal[0].mbmi_ext;
3579 duplicate_mode_info_in_sb(cm, xd, mi_row, mi_col, subsize);
3580 if (mi_row + hbs < cm->mi_rows) {
3581 set_mode_info_offsets(cm, x, xd, mi_row + hbs, mi_col);
3582 *(xd->mi[0]) = pc_tree->horizontal[1].mic;
3583 *(x->mbmi_ext) = pc_tree->horizontal[1].mbmi_ext;
3584 duplicate_mode_info_in_sb(cm, xd, mi_row + hbs, mi_col, subsize);
3585 }
3586 break;
3587 case PARTITION_SPLIT: {
3588 fill_mode_info_sb(cm, x, mi_row, mi_col, subsize, pc_tree->split[0]);
3589 fill_mode_info_sb(cm, x, mi_row, mi_col + hbs, subsize,
3590 pc_tree->split[1]);
3591 fill_mode_info_sb(cm, x, mi_row + hbs, mi_col, subsize,
3592 pc_tree->split[2]);
3593 fill_mode_info_sb(cm, x, mi_row + hbs, mi_col + hbs, subsize,
3594 pc_tree->split[3]);
3595 break;
3596 }
3597 default: break;
3598 }
3599 }
3600
3601 // Reset the prediction pixel ready flag recursively.
pred_pixel_ready_reset(PC_TREE * pc_tree,BLOCK_SIZE bsize)3602 static void pred_pixel_ready_reset(PC_TREE *pc_tree, BLOCK_SIZE bsize) {
3603 pc_tree->none.pred_pixel_ready = 0;
3604 pc_tree->horizontal[0].pred_pixel_ready = 0;
3605 pc_tree->horizontal[1].pred_pixel_ready = 0;
3606 pc_tree->vertical[0].pred_pixel_ready = 0;
3607 pc_tree->vertical[1].pred_pixel_ready = 0;
3608
3609 if (bsize > BLOCK_8X8) {
3610 BLOCK_SIZE subsize = get_subsize(bsize, PARTITION_SPLIT);
3611 int i;
3612 for (i = 0; i < 4; ++i) pred_pixel_ready_reset(pc_tree->split[i], subsize);
3613 }
3614 }
3615
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)3616 static void nonrd_pick_partition(VP9_COMP *cpi, ThreadData *td,
3617 TileDataEnc *tile_data, TOKENEXTRA **tp,
3618 int mi_row, int mi_col, BLOCK_SIZE bsize,
3619 RD_COST *rd_cost, int do_recon,
3620 int64_t best_rd, PC_TREE *pc_tree) {
3621 const SPEED_FEATURES *const sf = &cpi->sf;
3622 VP9_COMMON *const cm = &cpi->common;
3623 TileInfo *const tile_info = &tile_data->tile_info;
3624 MACROBLOCK *const x = &td->mb;
3625 MACROBLOCKD *const xd = &x->e_mbd;
3626 const int ms = num_8x8_blocks_wide_lookup[bsize] / 2;
3627 TOKENEXTRA *tp_orig = *tp;
3628 PICK_MODE_CONTEXT *ctx = &pc_tree->none;
3629 int i;
3630 BLOCK_SIZE subsize = bsize;
3631 RD_COST this_rdc, sum_rdc, best_rdc;
3632 int do_split = bsize >= BLOCK_8X8;
3633 int do_rect = 1;
3634 // Override skipping rectangular partition operations for edge blocks
3635 const int force_horz_split = (mi_row + ms >= cm->mi_rows);
3636 const int force_vert_split = (mi_col + ms >= cm->mi_cols);
3637 const int xss = x->e_mbd.plane[1].subsampling_x;
3638 const int yss = x->e_mbd.plane[1].subsampling_y;
3639
3640 int partition_none_allowed = !force_horz_split && !force_vert_split;
3641 int partition_horz_allowed =
3642 !force_vert_split && yss <= xss && bsize >= BLOCK_8X8;
3643 int partition_vert_allowed =
3644 !force_horz_split && xss <= yss && bsize >= BLOCK_8X8;
3645 (void)*tp_orig;
3646
3647 assert(num_8x8_blocks_wide_lookup[bsize] ==
3648 num_8x8_blocks_high_lookup[bsize]);
3649
3650 vp9_rd_cost_init(&sum_rdc);
3651 vp9_rd_cost_reset(&best_rdc);
3652 best_rdc.rdcost = best_rd;
3653
3654 // Determine partition types in search according to the speed features.
3655 // The threshold set here has to be of square block size.
3656 if (sf->auto_min_max_partition_size) {
3657 partition_none_allowed &=
3658 (bsize <= x->max_partition_size && bsize >= x->min_partition_size);
3659 partition_horz_allowed &=
3660 ((bsize <= x->max_partition_size && bsize > x->min_partition_size) ||
3661 force_horz_split);
3662 partition_vert_allowed &=
3663 ((bsize <= x->max_partition_size && bsize > x->min_partition_size) ||
3664 force_vert_split);
3665 do_split &= bsize > x->min_partition_size;
3666 }
3667 if (sf->use_square_partition_only) {
3668 partition_horz_allowed &= force_horz_split;
3669 partition_vert_allowed &= force_vert_split;
3670 }
3671
3672 ctx->pred_pixel_ready =
3673 !(partition_vert_allowed || partition_horz_allowed || do_split);
3674
3675 // PARTITION_NONE
3676 if (partition_none_allowed) {
3677 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &this_rdc, bsize,
3678 ctx);
3679 ctx->mic = *xd->mi[0];
3680 ctx->mbmi_ext = *x->mbmi_ext;
3681 ctx->skip_txfm[0] = x->skip_txfm[0];
3682 ctx->skip = x->skip;
3683
3684 if (this_rdc.rate != INT_MAX) {
3685 int pl = partition_plane_context(xd, mi_row, mi_col, bsize);
3686 this_rdc.rate += cpi->partition_cost[pl][PARTITION_NONE];
3687 this_rdc.rdcost =
3688 RDCOST(x->rdmult, x->rddiv, this_rdc.rate, this_rdc.dist);
3689 if (this_rdc.rdcost < best_rdc.rdcost) {
3690 int64_t dist_breakout_thr = sf->partition_search_breakout_thr.dist;
3691 int64_t rate_breakout_thr = sf->partition_search_breakout_thr.rate;
3692
3693 dist_breakout_thr >>=
3694 8 - (b_width_log2_lookup[bsize] + b_height_log2_lookup[bsize]);
3695
3696 rate_breakout_thr *= num_pels_log2_lookup[bsize];
3697
3698 best_rdc = this_rdc;
3699 if (bsize >= BLOCK_8X8) pc_tree->partitioning = PARTITION_NONE;
3700
3701 if (!x->e_mbd.lossless && this_rdc.rate < rate_breakout_thr &&
3702 this_rdc.dist < dist_breakout_thr) {
3703 do_split = 0;
3704 do_rect = 0;
3705 }
3706 }
3707 }
3708 }
3709
3710 // store estimated motion vector
3711 store_pred_mv(x, ctx);
3712
3713 // PARTITION_SPLIT
3714 if (do_split) {
3715 int pl = partition_plane_context(xd, mi_row, mi_col, bsize);
3716 sum_rdc.rate += cpi->partition_cost[pl][PARTITION_SPLIT];
3717 sum_rdc.rdcost = RDCOST(x->rdmult, x->rddiv, sum_rdc.rate, sum_rdc.dist);
3718 subsize = get_subsize(bsize, PARTITION_SPLIT);
3719 for (i = 0; i < 4 && sum_rdc.rdcost < best_rdc.rdcost; ++i) {
3720 const int x_idx = (i & 1) * ms;
3721 const int y_idx = (i >> 1) * ms;
3722
3723 if (mi_row + y_idx >= cm->mi_rows || mi_col + x_idx >= cm->mi_cols)
3724 continue;
3725 load_pred_mv(x, ctx);
3726 nonrd_pick_partition(cpi, td, tile_data, tp, mi_row + y_idx,
3727 mi_col + x_idx, subsize, &this_rdc, 0,
3728 best_rdc.rdcost - sum_rdc.rdcost, pc_tree->split[i]);
3729
3730 if (this_rdc.rate == INT_MAX) {
3731 vp9_rd_cost_reset(&sum_rdc);
3732 } else {
3733 sum_rdc.rate += this_rdc.rate;
3734 sum_rdc.dist += this_rdc.dist;
3735 sum_rdc.rdcost += this_rdc.rdcost;
3736 }
3737 }
3738
3739 if (sum_rdc.rdcost < best_rdc.rdcost) {
3740 best_rdc = sum_rdc;
3741 pc_tree->partitioning = PARTITION_SPLIT;
3742 } else {
3743 // skip rectangular partition test when larger block size
3744 // gives better rd cost
3745 if (sf->less_rectangular_check) do_rect &= !partition_none_allowed;
3746 }
3747 }
3748
3749 // PARTITION_HORZ
3750 if (partition_horz_allowed && do_rect) {
3751 subsize = get_subsize(bsize, PARTITION_HORZ);
3752 if (sf->adaptive_motion_search) load_pred_mv(x, ctx);
3753 pc_tree->horizontal[0].pred_pixel_ready = 1;
3754 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &sum_rdc, subsize,
3755 &pc_tree->horizontal[0]);
3756
3757 pc_tree->horizontal[0].mic = *xd->mi[0];
3758 pc_tree->horizontal[0].mbmi_ext = *x->mbmi_ext;
3759 pc_tree->horizontal[0].skip_txfm[0] = x->skip_txfm[0];
3760 pc_tree->horizontal[0].skip = x->skip;
3761
3762 if (sum_rdc.rdcost < best_rdc.rdcost && mi_row + ms < cm->mi_rows) {
3763 load_pred_mv(x, ctx);
3764 pc_tree->horizontal[1].pred_pixel_ready = 1;
3765 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row + ms, mi_col, &this_rdc,
3766 subsize, &pc_tree->horizontal[1]);
3767
3768 pc_tree->horizontal[1].mic = *xd->mi[0];
3769 pc_tree->horizontal[1].mbmi_ext = *x->mbmi_ext;
3770 pc_tree->horizontal[1].skip_txfm[0] = x->skip_txfm[0];
3771 pc_tree->horizontal[1].skip = x->skip;
3772
3773 if (this_rdc.rate == INT_MAX) {
3774 vp9_rd_cost_reset(&sum_rdc);
3775 } else {
3776 int pl = partition_plane_context(xd, mi_row, mi_col, bsize);
3777 this_rdc.rate += cpi->partition_cost[pl][PARTITION_HORZ];
3778 sum_rdc.rate += this_rdc.rate;
3779 sum_rdc.dist += this_rdc.dist;
3780 sum_rdc.rdcost =
3781 RDCOST(x->rdmult, x->rddiv, sum_rdc.rate, sum_rdc.dist);
3782 }
3783 }
3784
3785 if (sum_rdc.rdcost < best_rdc.rdcost) {
3786 best_rdc = sum_rdc;
3787 pc_tree->partitioning = PARTITION_HORZ;
3788 } else {
3789 pred_pixel_ready_reset(pc_tree, bsize);
3790 }
3791 }
3792
3793 // PARTITION_VERT
3794 if (partition_vert_allowed && do_rect) {
3795 subsize = get_subsize(bsize, PARTITION_VERT);
3796 if (sf->adaptive_motion_search) load_pred_mv(x, ctx);
3797 pc_tree->vertical[0].pred_pixel_ready = 1;
3798 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &sum_rdc, subsize,
3799 &pc_tree->vertical[0]);
3800 pc_tree->vertical[0].mic = *xd->mi[0];
3801 pc_tree->vertical[0].mbmi_ext = *x->mbmi_ext;
3802 pc_tree->vertical[0].skip_txfm[0] = x->skip_txfm[0];
3803 pc_tree->vertical[0].skip = x->skip;
3804
3805 if (sum_rdc.rdcost < best_rdc.rdcost && mi_col + ms < cm->mi_cols) {
3806 load_pred_mv(x, ctx);
3807 pc_tree->vertical[1].pred_pixel_ready = 1;
3808 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col + ms, &this_rdc,
3809 subsize, &pc_tree->vertical[1]);
3810 pc_tree->vertical[1].mic = *xd->mi[0];
3811 pc_tree->vertical[1].mbmi_ext = *x->mbmi_ext;
3812 pc_tree->vertical[1].skip_txfm[0] = x->skip_txfm[0];
3813 pc_tree->vertical[1].skip = x->skip;
3814
3815 if (this_rdc.rate == INT_MAX) {
3816 vp9_rd_cost_reset(&sum_rdc);
3817 } else {
3818 int pl = partition_plane_context(xd, mi_row, mi_col, bsize);
3819 sum_rdc.rate += cpi->partition_cost[pl][PARTITION_VERT];
3820 sum_rdc.rate += this_rdc.rate;
3821 sum_rdc.dist += this_rdc.dist;
3822 sum_rdc.rdcost =
3823 RDCOST(x->rdmult, x->rddiv, sum_rdc.rate, sum_rdc.dist);
3824 }
3825 }
3826
3827 if (sum_rdc.rdcost < best_rdc.rdcost) {
3828 best_rdc = sum_rdc;
3829 pc_tree->partitioning = PARTITION_VERT;
3830 } else {
3831 pred_pixel_ready_reset(pc_tree, bsize);
3832 }
3833 }
3834
3835 *rd_cost = best_rdc;
3836
3837 if (best_rdc.rate == INT_MAX) {
3838 vp9_rd_cost_reset(rd_cost);
3839 return;
3840 }
3841
3842 // update mode info array
3843 fill_mode_info_sb(cm, x, mi_row, mi_col, bsize, pc_tree);
3844
3845 if (best_rdc.rate < INT_MAX && best_rdc.dist < INT64_MAX && do_recon) {
3846 int output_enabled = (bsize == BLOCK_64X64);
3847 encode_sb_rt(cpi, td, tile_info, tp, mi_row, mi_col, output_enabled, bsize,
3848 pc_tree);
3849 }
3850
3851 if (bsize == BLOCK_64X64 && do_recon) {
3852 assert(tp_orig < *tp);
3853 assert(best_rdc.rate < INT_MAX);
3854 assert(best_rdc.dist < INT64_MAX);
3855 } else {
3856 assert(tp_orig == *tp);
3857 }
3858 }
3859
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)3860 static void nonrd_select_partition(VP9_COMP *cpi, ThreadData *td,
3861 TileDataEnc *tile_data, MODE_INFO **mi,
3862 TOKENEXTRA **tp, int mi_row, int mi_col,
3863 BLOCK_SIZE bsize, int output_enabled,
3864 RD_COST *rd_cost, PC_TREE *pc_tree) {
3865 VP9_COMMON *const cm = &cpi->common;
3866 TileInfo *const tile_info = &tile_data->tile_info;
3867 MACROBLOCK *const x = &td->mb;
3868 MACROBLOCKD *const xd = &x->e_mbd;
3869 const int bsl = b_width_log2_lookup[bsize], hbs = (1 << bsl) / 4;
3870 const int mis = cm->mi_stride;
3871 PARTITION_TYPE partition;
3872 BLOCK_SIZE subsize;
3873 RD_COST this_rdc;
3874
3875 vp9_rd_cost_reset(&this_rdc);
3876 if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols) return;
3877
3878 subsize = (bsize >= BLOCK_8X8) ? mi[0]->sb_type : BLOCK_4X4;
3879 partition = partition_lookup[bsl][subsize];
3880
3881 if (bsize == BLOCK_32X32 && subsize == BLOCK_32X32) {
3882 x->max_partition_size = BLOCK_32X32;
3883 x->min_partition_size = BLOCK_16X16;
3884 nonrd_pick_partition(cpi, td, tile_data, tp, mi_row, mi_col, bsize, rd_cost,
3885 0, INT64_MAX, pc_tree);
3886 } else if (bsize == BLOCK_32X32 && partition != PARTITION_NONE &&
3887 subsize >= BLOCK_16X16) {
3888 x->max_partition_size = BLOCK_32X32;
3889 x->min_partition_size = BLOCK_8X8;
3890 nonrd_pick_partition(cpi, td, tile_data, tp, mi_row, mi_col, bsize, rd_cost,
3891 0, INT64_MAX, pc_tree);
3892 } else if (bsize == BLOCK_16X16 && partition != PARTITION_NONE) {
3893 x->max_partition_size = BLOCK_16X16;
3894 x->min_partition_size = BLOCK_8X8;
3895 nonrd_pick_partition(cpi, td, tile_data, tp, mi_row, mi_col, bsize, rd_cost,
3896 0, INT64_MAX, pc_tree);
3897 } else {
3898 switch (partition) {
3899 case PARTITION_NONE:
3900 pc_tree->none.pred_pixel_ready = 1;
3901 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, rd_cost, subsize,
3902 &pc_tree->none);
3903 pc_tree->none.mic = *xd->mi[0];
3904 pc_tree->none.mbmi_ext = *x->mbmi_ext;
3905 pc_tree->none.skip_txfm[0] = x->skip_txfm[0];
3906 pc_tree->none.skip = x->skip;
3907 break;
3908 case PARTITION_VERT:
3909 pc_tree->vertical[0].pred_pixel_ready = 1;
3910 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, rd_cost, subsize,
3911 &pc_tree->vertical[0]);
3912 pc_tree->vertical[0].mic = *xd->mi[0];
3913 pc_tree->vertical[0].mbmi_ext = *x->mbmi_ext;
3914 pc_tree->vertical[0].skip_txfm[0] = x->skip_txfm[0];
3915 pc_tree->vertical[0].skip = x->skip;
3916 if (mi_col + hbs < cm->mi_cols) {
3917 pc_tree->vertical[1].pred_pixel_ready = 1;
3918 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col + hbs,
3919 &this_rdc, subsize, &pc_tree->vertical[1]);
3920 pc_tree->vertical[1].mic = *xd->mi[0];
3921 pc_tree->vertical[1].mbmi_ext = *x->mbmi_ext;
3922 pc_tree->vertical[1].skip_txfm[0] = x->skip_txfm[0];
3923 pc_tree->vertical[1].skip = x->skip;
3924 if (this_rdc.rate != INT_MAX && this_rdc.dist != INT64_MAX &&
3925 rd_cost->rate != INT_MAX && rd_cost->dist != INT64_MAX) {
3926 rd_cost->rate += this_rdc.rate;
3927 rd_cost->dist += this_rdc.dist;
3928 }
3929 }
3930 break;
3931 case PARTITION_HORZ:
3932 pc_tree->horizontal[0].pred_pixel_ready = 1;
3933 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, rd_cost, subsize,
3934 &pc_tree->horizontal[0]);
3935 pc_tree->horizontal[0].mic = *xd->mi[0];
3936 pc_tree->horizontal[0].mbmi_ext = *x->mbmi_ext;
3937 pc_tree->horizontal[0].skip_txfm[0] = x->skip_txfm[0];
3938 pc_tree->horizontal[0].skip = x->skip;
3939 if (mi_row + hbs < cm->mi_rows) {
3940 pc_tree->horizontal[1].pred_pixel_ready = 1;
3941 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row + hbs, mi_col,
3942 &this_rdc, subsize, &pc_tree->horizontal[1]);
3943 pc_tree->horizontal[1].mic = *xd->mi[0];
3944 pc_tree->horizontal[1].mbmi_ext = *x->mbmi_ext;
3945 pc_tree->horizontal[1].skip_txfm[0] = x->skip_txfm[0];
3946 pc_tree->horizontal[1].skip = x->skip;
3947 if (this_rdc.rate != INT_MAX && this_rdc.dist != INT64_MAX &&
3948 rd_cost->rate != INT_MAX && rd_cost->dist != INT64_MAX) {
3949 rd_cost->rate += this_rdc.rate;
3950 rd_cost->dist += this_rdc.dist;
3951 }
3952 }
3953 break;
3954 case PARTITION_SPLIT:
3955 subsize = get_subsize(bsize, PARTITION_SPLIT);
3956 nonrd_select_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col,
3957 subsize, output_enabled, rd_cost,
3958 pc_tree->split[0]);
3959 nonrd_select_partition(cpi, td, tile_data, mi + hbs, tp, mi_row,
3960 mi_col + hbs, subsize, output_enabled, &this_rdc,
3961 pc_tree->split[1]);
3962 if (this_rdc.rate != INT_MAX && this_rdc.dist != INT64_MAX &&
3963 rd_cost->rate != INT_MAX && rd_cost->dist != INT64_MAX) {
3964 rd_cost->rate += this_rdc.rate;
3965 rd_cost->dist += this_rdc.dist;
3966 }
3967 nonrd_select_partition(cpi, td, tile_data, mi + hbs * mis, tp,
3968 mi_row + hbs, mi_col, subsize, output_enabled,
3969 &this_rdc, pc_tree->split[2]);
3970 if (this_rdc.rate != INT_MAX && this_rdc.dist != INT64_MAX &&
3971 rd_cost->rate != INT_MAX && rd_cost->dist != INT64_MAX) {
3972 rd_cost->rate += this_rdc.rate;
3973 rd_cost->dist += this_rdc.dist;
3974 }
3975 nonrd_select_partition(cpi, td, tile_data, mi + hbs * mis + hbs, tp,
3976 mi_row + hbs, mi_col + hbs, subsize,
3977 output_enabled, &this_rdc, pc_tree->split[3]);
3978 if (this_rdc.rate != INT_MAX && this_rdc.dist != INT64_MAX &&
3979 rd_cost->rate != INT_MAX && rd_cost->dist != INT64_MAX) {
3980 rd_cost->rate += this_rdc.rate;
3981 rd_cost->dist += this_rdc.dist;
3982 }
3983 break;
3984 default: assert(0 && "Invalid partition type."); break;
3985 }
3986 }
3987
3988 if (bsize == BLOCK_64X64 && output_enabled)
3989 encode_sb_rt(cpi, td, tile_info, tp, mi_row, mi_col, 1, bsize, pc_tree);
3990 }
3991
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)3992 static void nonrd_use_partition(VP9_COMP *cpi, ThreadData *td,
3993 TileDataEnc *tile_data, MODE_INFO **mi,
3994 TOKENEXTRA **tp, int mi_row, int mi_col,
3995 BLOCK_SIZE bsize, int output_enabled,
3996 RD_COST *dummy_cost, PC_TREE *pc_tree) {
3997 VP9_COMMON *const cm = &cpi->common;
3998 TileInfo *tile_info = &tile_data->tile_info;
3999 MACROBLOCK *const x = &td->mb;
4000 MACROBLOCKD *const xd = &x->e_mbd;
4001 const int bsl = b_width_log2_lookup[bsize], hbs = (1 << bsl) / 4;
4002 const int mis = cm->mi_stride;
4003 PARTITION_TYPE partition;
4004 BLOCK_SIZE subsize;
4005
4006 if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols) return;
4007
4008 subsize = (bsize >= BLOCK_8X8) ? mi[0]->sb_type : BLOCK_4X4;
4009 partition = partition_lookup[bsl][subsize];
4010
4011 if (output_enabled && bsize != BLOCK_4X4) {
4012 int ctx = partition_plane_context(xd, mi_row, mi_col, bsize);
4013 td->counts->partition[ctx][partition]++;
4014 }
4015
4016 switch (partition) {
4017 case PARTITION_NONE:
4018 pc_tree->none.pred_pixel_ready = 1;
4019 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, dummy_cost,
4020 subsize, &pc_tree->none);
4021 pc_tree->none.mic = *xd->mi[0];
4022 pc_tree->none.mbmi_ext = *x->mbmi_ext;
4023 pc_tree->none.skip_txfm[0] = x->skip_txfm[0];
4024 pc_tree->none.skip = x->skip;
4025 encode_b_rt(cpi, td, tile_info, tp, mi_row, mi_col, output_enabled,
4026 subsize, &pc_tree->none);
4027 break;
4028 case PARTITION_VERT:
4029 pc_tree->vertical[0].pred_pixel_ready = 1;
4030 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, dummy_cost,
4031 subsize, &pc_tree->vertical[0]);
4032 pc_tree->vertical[0].mic = *xd->mi[0];
4033 pc_tree->vertical[0].mbmi_ext = *x->mbmi_ext;
4034 pc_tree->vertical[0].skip_txfm[0] = x->skip_txfm[0];
4035 pc_tree->vertical[0].skip = x->skip;
4036 encode_b_rt(cpi, td, tile_info, tp, mi_row, mi_col, output_enabled,
4037 subsize, &pc_tree->vertical[0]);
4038 if (mi_col + hbs < cm->mi_cols && bsize > BLOCK_8X8) {
4039 pc_tree->vertical[1].pred_pixel_ready = 1;
4040 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col + hbs, dummy_cost,
4041 subsize, &pc_tree->vertical[1]);
4042 pc_tree->vertical[1].mic = *xd->mi[0];
4043 pc_tree->vertical[1].mbmi_ext = *x->mbmi_ext;
4044 pc_tree->vertical[1].skip_txfm[0] = x->skip_txfm[0];
4045 pc_tree->vertical[1].skip = x->skip;
4046 encode_b_rt(cpi, td, tile_info, tp, mi_row, mi_col + hbs,
4047 output_enabled, subsize, &pc_tree->vertical[1]);
4048 }
4049 break;
4050 case PARTITION_HORZ:
4051 pc_tree->horizontal[0].pred_pixel_ready = 1;
4052 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, dummy_cost,
4053 subsize, &pc_tree->horizontal[0]);
4054 pc_tree->horizontal[0].mic = *xd->mi[0];
4055 pc_tree->horizontal[0].mbmi_ext = *x->mbmi_ext;
4056 pc_tree->horizontal[0].skip_txfm[0] = x->skip_txfm[0];
4057 pc_tree->horizontal[0].skip = x->skip;
4058 encode_b_rt(cpi, td, tile_info, tp, mi_row, mi_col, output_enabled,
4059 subsize, &pc_tree->horizontal[0]);
4060
4061 if (mi_row + hbs < cm->mi_rows && bsize > BLOCK_8X8) {
4062 pc_tree->horizontal[1].pred_pixel_ready = 1;
4063 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row + hbs, mi_col, dummy_cost,
4064 subsize, &pc_tree->horizontal[1]);
4065 pc_tree->horizontal[1].mic = *xd->mi[0];
4066 pc_tree->horizontal[1].mbmi_ext = *x->mbmi_ext;
4067 pc_tree->horizontal[1].skip_txfm[0] = x->skip_txfm[0];
4068 pc_tree->horizontal[1].skip = x->skip;
4069 encode_b_rt(cpi, td, tile_info, tp, mi_row + hbs, mi_col,
4070 output_enabled, subsize, &pc_tree->horizontal[1]);
4071 }
4072 break;
4073 case PARTITION_SPLIT:
4074 subsize = get_subsize(bsize, PARTITION_SPLIT);
4075 if (bsize == BLOCK_8X8) {
4076 nonrd_pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, dummy_cost,
4077 subsize, pc_tree->leaf_split[0]);
4078 encode_b_rt(cpi, td, tile_info, tp, mi_row, mi_col, output_enabled,
4079 subsize, pc_tree->leaf_split[0]);
4080 } else {
4081 nonrd_use_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col, subsize,
4082 output_enabled, dummy_cost, pc_tree->split[0]);
4083 nonrd_use_partition(cpi, td, tile_data, mi + hbs, tp, mi_row,
4084 mi_col + hbs, subsize, output_enabled, dummy_cost,
4085 pc_tree->split[1]);
4086 nonrd_use_partition(cpi, td, tile_data, mi + hbs * mis, tp,
4087 mi_row + hbs, mi_col, subsize, output_enabled,
4088 dummy_cost, pc_tree->split[2]);
4089 nonrd_use_partition(cpi, td, tile_data, mi + hbs * mis + hbs, tp,
4090 mi_row + hbs, mi_col + hbs, subsize, output_enabled,
4091 dummy_cost, pc_tree->split[3]);
4092 }
4093 break;
4094 default: assert(0 && "Invalid partition type."); break;
4095 }
4096
4097 if (partition != PARTITION_SPLIT || bsize == BLOCK_8X8)
4098 update_partition_context(xd, mi_row, mi_col, subsize, bsize);
4099 }
4100
encode_nonrd_sb_row(VP9_COMP * cpi,ThreadData * td,TileDataEnc * tile_data,int mi_row,TOKENEXTRA ** tp)4101 static void encode_nonrd_sb_row(VP9_COMP *cpi, ThreadData *td,
4102 TileDataEnc *tile_data, int mi_row,
4103 TOKENEXTRA **tp) {
4104 SPEED_FEATURES *const sf = &cpi->sf;
4105 VP9_COMMON *const cm = &cpi->common;
4106 TileInfo *const tile_info = &tile_data->tile_info;
4107 MACROBLOCK *const x = &td->mb;
4108 MACROBLOCKD *const xd = &x->e_mbd;
4109 const int mi_col_start = tile_info->mi_col_start;
4110 const int mi_col_end = tile_info->mi_col_end;
4111 int mi_col;
4112 const int sb_row = mi_row >> MI_BLOCK_SIZE_LOG2;
4113 const int num_sb_cols =
4114 get_num_cols(tile_data->tile_info, MI_BLOCK_SIZE_LOG2);
4115 int sb_col_in_tile;
4116
4117 // Initialize the left context for the new SB row
4118 memset(&xd->left_context, 0, sizeof(xd->left_context));
4119 memset(xd->left_seg_context, 0, sizeof(xd->left_seg_context));
4120
4121 // Code each SB in the row
4122 for (mi_col = mi_col_start, sb_col_in_tile = 0; mi_col < mi_col_end;
4123 mi_col += MI_BLOCK_SIZE, ++sb_col_in_tile) {
4124 const struct segmentation *const seg = &cm->seg;
4125 RD_COST dummy_rdc;
4126 const int idx_str = cm->mi_stride * mi_row + mi_col;
4127 MODE_INFO **mi = cm->mi_grid_visible + idx_str;
4128 PARTITION_SEARCH_TYPE partition_search_type = sf->partition_search_type;
4129 BLOCK_SIZE bsize = BLOCK_64X64;
4130 int seg_skip = 0;
4131
4132 (*(cpi->row_mt_sync_read_ptr))(&tile_data->row_mt_sync, sb_row,
4133 sb_col_in_tile);
4134
4135 x->source_variance = UINT_MAX;
4136 vp9_zero(x->pred_mv);
4137 vp9_rd_cost_init(&dummy_rdc);
4138 x->color_sensitivity[0] = 0;
4139 x->color_sensitivity[1] = 0;
4140 x->sb_is_skin = 0;
4141 x->skip_low_source_sad = 0;
4142 x->lowvar_highsumdiff = 0;
4143 x->content_state_sb = 0;
4144
4145 if (seg->enabled) {
4146 const uint8_t *const map =
4147 seg->update_map ? cpi->segmentation_map : cm->last_frame_seg_map;
4148 int segment_id = get_segment_id(cm, map, BLOCK_64X64, mi_row, mi_col);
4149 seg_skip = segfeature_active(seg, segment_id, SEG_LVL_SKIP);
4150 if (seg_skip) {
4151 partition_search_type = FIXED_PARTITION;
4152 }
4153 }
4154
4155 if (cpi->compute_source_sad_onepass && cpi->sf.use_source_sad) {
4156 int shift = cpi->Source->y_stride * (mi_row << 3) + (mi_col << 3);
4157 int sb_offset2 = ((cm->mi_cols + 7) >> 3) * (mi_row >> 3) + (mi_col >> 3);
4158 avg_source_sad(cpi, x, shift, sb_offset2);
4159 }
4160
4161 // Set the partition type of the 64X64 block
4162 switch (partition_search_type) {
4163 case VAR_BASED_PARTITION:
4164 // TODO(jingning, marpan): The mode decision and encoding process
4165 // support both intra and inter sub8x8 block coding for RTC mode.
4166 // Tune the thresholds accordingly to use sub8x8 block coding for
4167 // coding performance improvement.
4168 choose_partitioning(cpi, tile_info, x, mi_row, mi_col);
4169 nonrd_use_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col,
4170 BLOCK_64X64, 1, &dummy_rdc, td->pc_root);
4171 break;
4172 case SOURCE_VAR_BASED_PARTITION:
4173 set_source_var_based_partition(cpi, tile_info, x, mi, mi_row, mi_col);
4174 nonrd_use_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col,
4175 BLOCK_64X64, 1, &dummy_rdc, td->pc_root);
4176 break;
4177 case FIXED_PARTITION:
4178 if (!seg_skip) bsize = sf->always_this_block_size;
4179 set_fixed_partitioning(cpi, tile_info, mi, mi_row, mi_col, bsize);
4180 nonrd_use_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col,
4181 BLOCK_64X64, 1, &dummy_rdc, td->pc_root);
4182 break;
4183 case REFERENCE_PARTITION:
4184 set_offsets(cpi, tile_info, x, mi_row, mi_col, BLOCK_64X64);
4185 // Use nonrd_pick_partition on scene-cut for VBR mode.
4186 // nonrd_pick_partition does not support 4x4 partition, so avoid it
4187 // on key frame for now.
4188 if ((cpi->oxcf.rc_mode == VPX_VBR && cpi->rc.high_source_sad &&
4189 cm->frame_type != KEY_FRAME)) {
4190 // Use lower max_partition_size for low resoultions.
4191 if (cm->width <= 352 && cm->height <= 288)
4192 x->max_partition_size = BLOCK_32X32;
4193 else
4194 x->max_partition_size = BLOCK_64X64;
4195 x->min_partition_size = BLOCK_8X8;
4196 nonrd_pick_partition(cpi, td, tile_data, tp, mi_row, mi_col,
4197 BLOCK_64X64, &dummy_rdc, 1, INT64_MAX,
4198 td->pc_root);
4199 } else {
4200 choose_partitioning(cpi, tile_info, x, mi_row, mi_col);
4201 // TODO(marpan): Seems like nonrd_select_partition does not support
4202 // 4x4 partition. Since 4x4 is used on key frame, use this switch
4203 // for now.
4204 if (cm->frame_type == KEY_FRAME)
4205 nonrd_use_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col,
4206 BLOCK_64X64, 1, &dummy_rdc, td->pc_root);
4207 else
4208 nonrd_select_partition(cpi, td, tile_data, mi, tp, mi_row, mi_col,
4209 BLOCK_64X64, 1, &dummy_rdc, td->pc_root);
4210 }
4211
4212 break;
4213 default: assert(0); break;
4214 }
4215
4216 (*(cpi->row_mt_sync_write_ptr))(&tile_data->row_mt_sync, sb_row,
4217 sb_col_in_tile, num_sb_cols);
4218 }
4219 }
4220 // end RTC play code
4221
set_var_thresh_from_histogram(VP9_COMP * cpi)4222 static int set_var_thresh_from_histogram(VP9_COMP *cpi) {
4223 const SPEED_FEATURES *const sf = &cpi->sf;
4224 const VP9_COMMON *const cm = &cpi->common;
4225
4226 const uint8_t *src = cpi->Source->y_buffer;
4227 const uint8_t *last_src = cpi->Last_Source->y_buffer;
4228 const int src_stride = cpi->Source->y_stride;
4229 const int last_stride = cpi->Last_Source->y_stride;
4230
4231 // Pick cutoff threshold
4232 const int cutoff = (VPXMIN(cm->width, cm->height) >= 720)
4233 ? (cm->MBs * VAR_HIST_LARGE_CUT_OFF / 100)
4234 : (cm->MBs * VAR_HIST_SMALL_CUT_OFF / 100);
4235 DECLARE_ALIGNED(16, int, hist[VAR_HIST_BINS]);
4236 diff *var16 = cpi->source_diff_var;
4237
4238 int sum = 0;
4239 int i, j;
4240
4241 memset(hist, 0, VAR_HIST_BINS * sizeof(hist[0]));
4242
4243 for (i = 0; i < cm->mb_rows; i++) {
4244 for (j = 0; j < cm->mb_cols; j++) {
4245 #if CONFIG_VP9_HIGHBITDEPTH
4246 if (cm->use_highbitdepth) {
4247 switch (cm->bit_depth) {
4248 case VPX_BITS_8:
4249 vpx_highbd_8_get16x16var(src, src_stride, last_src, last_stride,
4250 &var16->sse, &var16->sum);
4251 break;
4252 case VPX_BITS_10:
4253 vpx_highbd_10_get16x16var(src, src_stride, last_src, last_stride,
4254 &var16->sse, &var16->sum);
4255 break;
4256 case VPX_BITS_12:
4257 vpx_highbd_12_get16x16var(src, src_stride, last_src, last_stride,
4258 &var16->sse, &var16->sum);
4259 break;
4260 default:
4261 assert(0 &&
4262 "cm->bit_depth should be VPX_BITS_8, VPX_BITS_10"
4263 " or VPX_BITS_12");
4264 return -1;
4265 }
4266 } else {
4267 vpx_get16x16var(src, src_stride, last_src, last_stride, &var16->sse,
4268 &var16->sum);
4269 }
4270 #else
4271 vpx_get16x16var(src, src_stride, last_src, last_stride, &var16->sse,
4272 &var16->sum);
4273 #endif // CONFIG_VP9_HIGHBITDEPTH
4274 var16->var = var16->sse - (((uint32_t)var16->sum * var16->sum) >> 8);
4275
4276 if (var16->var >= VAR_HIST_MAX_BG_VAR)
4277 hist[VAR_HIST_BINS - 1]++;
4278 else
4279 hist[var16->var / VAR_HIST_FACTOR]++;
4280
4281 src += 16;
4282 last_src += 16;
4283 var16++;
4284 }
4285
4286 src = src - cm->mb_cols * 16 + 16 * src_stride;
4287 last_src = last_src - cm->mb_cols * 16 + 16 * last_stride;
4288 }
4289
4290 cpi->source_var_thresh = 0;
4291
4292 if (hist[VAR_HIST_BINS - 1] < cutoff) {
4293 for (i = 0; i < VAR_HIST_BINS - 1; i++) {
4294 sum += hist[i];
4295
4296 if (sum > cutoff) {
4297 cpi->source_var_thresh = (i + 1) * VAR_HIST_FACTOR;
4298 return 0;
4299 }
4300 }
4301 }
4302
4303 return sf->search_type_check_frequency;
4304 }
4305
source_var_based_partition_search_method(VP9_COMP * cpi)4306 static void source_var_based_partition_search_method(VP9_COMP *cpi) {
4307 VP9_COMMON *const cm = &cpi->common;
4308 SPEED_FEATURES *const sf = &cpi->sf;
4309
4310 if (cm->frame_type == KEY_FRAME) {
4311 // For key frame, use SEARCH_PARTITION.
4312 sf->partition_search_type = SEARCH_PARTITION;
4313 } else if (cm->intra_only) {
4314 sf->partition_search_type = FIXED_PARTITION;
4315 } else {
4316 if (cm->last_width != cm->width || cm->last_height != cm->height) {
4317 if (cpi->source_diff_var) vpx_free(cpi->source_diff_var);
4318
4319 CHECK_MEM_ERROR(cm, cpi->source_diff_var,
4320 vpx_calloc(cm->MBs, sizeof(diff)));
4321 }
4322
4323 if (!cpi->frames_till_next_var_check)
4324 cpi->frames_till_next_var_check = set_var_thresh_from_histogram(cpi);
4325
4326 if (cpi->frames_till_next_var_check > 0) {
4327 sf->partition_search_type = FIXED_PARTITION;
4328 cpi->frames_till_next_var_check--;
4329 }
4330 }
4331 }
4332
get_skip_encode_frame(const VP9_COMMON * cm,ThreadData * const td)4333 static int get_skip_encode_frame(const VP9_COMMON *cm, ThreadData *const td) {
4334 unsigned int intra_count = 0, inter_count = 0;
4335 int j;
4336
4337 for (j = 0; j < INTRA_INTER_CONTEXTS; ++j) {
4338 intra_count += td->counts->intra_inter[j][0];
4339 inter_count += td->counts->intra_inter[j][1];
4340 }
4341
4342 return (intra_count << 2) < inter_count && cm->frame_type != KEY_FRAME &&
4343 cm->show_frame;
4344 }
4345
vp9_init_tile_data(VP9_COMP * cpi)4346 void vp9_init_tile_data(VP9_COMP *cpi) {
4347 VP9_COMMON *const cm = &cpi->common;
4348 const int tile_cols = 1 << cm->log2_tile_cols;
4349 const int tile_rows = 1 << cm->log2_tile_rows;
4350 int tile_col, tile_row;
4351 TOKENEXTRA *pre_tok = cpi->tile_tok[0][0];
4352 TOKENLIST *tplist = cpi->tplist[0][0];
4353 int tile_tok = 0;
4354 int tplist_count = 0;
4355
4356 if (cpi->tile_data == NULL || cpi->allocated_tiles < tile_cols * tile_rows) {
4357 if (cpi->tile_data != NULL) vpx_free(cpi->tile_data);
4358 CHECK_MEM_ERROR(cm, cpi->tile_data, vpx_malloc(tile_cols * tile_rows *
4359 sizeof(*cpi->tile_data)));
4360 cpi->allocated_tiles = tile_cols * tile_rows;
4361
4362 for (tile_row = 0; tile_row < tile_rows; ++tile_row)
4363 for (tile_col = 0; tile_col < tile_cols; ++tile_col) {
4364 TileDataEnc *tile_data =
4365 &cpi->tile_data[tile_row * tile_cols + tile_col];
4366 int i, j;
4367 for (i = 0; i < BLOCK_SIZES; ++i) {
4368 for (j = 0; j < MAX_MODES; ++j) {
4369 tile_data->thresh_freq_fact[i][j] = RD_THRESH_INIT_FACT;
4370 tile_data->mode_map[i][j] = j;
4371 }
4372 }
4373 #if CONFIG_MULTITHREAD
4374 tile_data->row_base_thresh_freq_fact = NULL;
4375 #endif
4376 }
4377 }
4378
4379 for (tile_row = 0; tile_row < tile_rows; ++tile_row) {
4380 for (tile_col = 0; tile_col < tile_cols; ++tile_col) {
4381 TileDataEnc *this_tile = &cpi->tile_data[tile_row * tile_cols + tile_col];
4382 TileInfo *tile_info = &this_tile->tile_info;
4383 vp9_tile_init(tile_info, cm, tile_row, tile_col);
4384
4385 cpi->tile_tok[tile_row][tile_col] = pre_tok + tile_tok;
4386 pre_tok = cpi->tile_tok[tile_row][tile_col];
4387 tile_tok = allocated_tokens(*tile_info);
4388
4389 cpi->tplist[tile_row][tile_col] = tplist + tplist_count;
4390 tplist = cpi->tplist[tile_row][tile_col];
4391 tplist_count = get_num_vert_units(*tile_info, MI_BLOCK_SIZE_LOG2);
4392 }
4393 }
4394 }
4395
vp9_encode_sb_row(VP9_COMP * cpi,ThreadData * td,int tile_row,int tile_col,int mi_row)4396 void vp9_encode_sb_row(VP9_COMP *cpi, ThreadData *td, int tile_row,
4397 int tile_col, int mi_row) {
4398 VP9_COMMON *const cm = &cpi->common;
4399 const int tile_cols = 1 << cm->log2_tile_cols;
4400 TileDataEnc *this_tile = &cpi->tile_data[tile_row * tile_cols + tile_col];
4401 const TileInfo *const tile_info = &this_tile->tile_info;
4402 TOKENEXTRA *tok = NULL;
4403 int tile_sb_row;
4404 int tile_mb_cols = (tile_info->mi_col_end - tile_info->mi_col_start + 1) >> 1;
4405
4406 tile_sb_row = mi_cols_aligned_to_sb(mi_row - tile_info->mi_row_start) >>
4407 MI_BLOCK_SIZE_LOG2;
4408 get_start_tok(cpi, tile_row, tile_col, mi_row, &tok);
4409 cpi->tplist[tile_row][tile_col][tile_sb_row].start = tok;
4410
4411 if (cpi->sf.use_nonrd_pick_mode)
4412 encode_nonrd_sb_row(cpi, td, this_tile, mi_row, &tok);
4413 else
4414 encode_rd_sb_row(cpi, td, this_tile, mi_row, &tok);
4415
4416 cpi->tplist[tile_row][tile_col][tile_sb_row].stop = tok;
4417 cpi->tplist[tile_row][tile_col][tile_sb_row].count =
4418 (unsigned int)(cpi->tplist[tile_row][tile_col][tile_sb_row].stop -
4419 cpi->tplist[tile_row][tile_col][tile_sb_row].start);
4420 assert(tok - cpi->tplist[tile_row][tile_col][tile_sb_row].start <=
4421 get_token_alloc(MI_BLOCK_SIZE >> 1, tile_mb_cols));
4422
4423 (void)tile_mb_cols;
4424 }
4425
vp9_encode_tile(VP9_COMP * cpi,ThreadData * td,int tile_row,int tile_col)4426 void vp9_encode_tile(VP9_COMP *cpi, ThreadData *td, int tile_row,
4427 int tile_col) {
4428 VP9_COMMON *const cm = &cpi->common;
4429 const int tile_cols = 1 << cm->log2_tile_cols;
4430 TileDataEnc *this_tile = &cpi->tile_data[tile_row * tile_cols + tile_col];
4431 const TileInfo *const tile_info = &this_tile->tile_info;
4432 const int mi_row_start = tile_info->mi_row_start;
4433 const int mi_row_end = tile_info->mi_row_end;
4434 int mi_row;
4435
4436 for (mi_row = mi_row_start; mi_row < mi_row_end; mi_row += MI_BLOCK_SIZE)
4437 vp9_encode_sb_row(cpi, td, tile_row, tile_col, mi_row);
4438 }
4439
encode_tiles(VP9_COMP * cpi)4440 static void encode_tiles(VP9_COMP *cpi) {
4441 VP9_COMMON *const cm = &cpi->common;
4442 const int tile_cols = 1 << cm->log2_tile_cols;
4443 const int tile_rows = 1 << cm->log2_tile_rows;
4444 int tile_col, tile_row;
4445
4446 vp9_init_tile_data(cpi);
4447
4448 for (tile_row = 0; tile_row < tile_rows; ++tile_row)
4449 for (tile_col = 0; tile_col < tile_cols; ++tile_col)
4450 vp9_encode_tile(cpi, &cpi->td, tile_row, tile_col);
4451 }
4452
4453 #if CONFIG_FP_MB_STATS
input_fpmb_stats(FIRSTPASS_MB_STATS * firstpass_mb_stats,VP9_COMMON * cm,uint8_t ** this_frame_mb_stats)4454 static int input_fpmb_stats(FIRSTPASS_MB_STATS *firstpass_mb_stats,
4455 VP9_COMMON *cm, uint8_t **this_frame_mb_stats) {
4456 uint8_t *mb_stats_in = firstpass_mb_stats->mb_stats_start +
4457 cm->current_video_frame * cm->MBs * sizeof(uint8_t);
4458
4459 if (mb_stats_in > firstpass_mb_stats->mb_stats_end) return EOF;
4460
4461 *this_frame_mb_stats = mb_stats_in;
4462
4463 return 1;
4464 }
4465 #endif
4466
encode_frame_internal(VP9_COMP * cpi)4467 static void encode_frame_internal(VP9_COMP *cpi) {
4468 SPEED_FEATURES *const sf = &cpi->sf;
4469 ThreadData *const td = &cpi->td;
4470 MACROBLOCK *const x = &td->mb;
4471 VP9_COMMON *const cm = &cpi->common;
4472 MACROBLOCKD *const xd = &x->e_mbd;
4473
4474 xd->mi = cm->mi_grid_visible;
4475 xd->mi[0] = cm->mi;
4476
4477 vp9_zero(*td->counts);
4478 vp9_zero(cpi->td.rd_counts);
4479
4480 xd->lossless = cm->base_qindex == 0 && cm->y_dc_delta_q == 0 &&
4481 cm->uv_dc_delta_q == 0 && cm->uv_ac_delta_q == 0;
4482
4483 #if CONFIG_VP9_HIGHBITDEPTH
4484 if (cm->use_highbitdepth)
4485 x->fwd_txm4x4 = xd->lossless ? vp9_highbd_fwht4x4 : vpx_highbd_fdct4x4;
4486 else
4487 x->fwd_txm4x4 = xd->lossless ? vp9_fwht4x4 : vpx_fdct4x4;
4488 x->highbd_itxm_add =
4489 xd->lossless ? vp9_highbd_iwht4x4_add : vp9_highbd_idct4x4_add;
4490 #else
4491 x->fwd_txm4x4 = xd->lossless ? vp9_fwht4x4 : vpx_fdct4x4;
4492 #endif // CONFIG_VP9_HIGHBITDEPTH
4493 x->itxm_add = xd->lossless ? vp9_iwht4x4_add : vp9_idct4x4_add;
4494
4495 if (xd->lossless) x->optimize = 0;
4496
4497 cm->tx_mode = select_tx_mode(cpi, xd);
4498
4499 vp9_frame_init_quantizer(cpi);
4500
4501 vp9_initialize_rd_consts(cpi);
4502 vp9_initialize_me_consts(cpi, x, cm->base_qindex);
4503 init_encode_frame_mb_context(cpi);
4504 cm->use_prev_frame_mvs =
4505 !cm->error_resilient_mode && cm->width == cm->last_width &&
4506 cm->height == cm->last_height && !cm->intra_only && cm->last_show_frame;
4507 // Special case: set prev_mi to NULL when the previous mode info
4508 // context cannot be used.
4509 cm->prev_mi =
4510 cm->use_prev_frame_mvs ? cm->prev_mip + cm->mi_stride + 1 : NULL;
4511
4512 x->quant_fp = cpi->sf.use_quant_fp;
4513 vp9_zero(x->skip_txfm);
4514 if (sf->use_nonrd_pick_mode) {
4515 // Initialize internal buffer pointers for rtc coding, where non-RD
4516 // mode decision is used and hence no buffer pointer swap needed.
4517 int i;
4518 struct macroblock_plane *const p = x->plane;
4519 struct macroblockd_plane *const pd = xd->plane;
4520 PICK_MODE_CONTEXT *ctx = &cpi->td.pc_root->none;
4521
4522 for (i = 0; i < MAX_MB_PLANE; ++i) {
4523 p[i].coeff = ctx->coeff_pbuf[i][0];
4524 p[i].qcoeff = ctx->qcoeff_pbuf[i][0];
4525 pd[i].dqcoeff = ctx->dqcoeff_pbuf[i][0];
4526 p[i].eobs = ctx->eobs_pbuf[i][0];
4527 }
4528 vp9_zero(x->zcoeff_blk);
4529
4530 if (cm->frame_type != KEY_FRAME && cpi->rc.frames_since_golden == 0 &&
4531 !(cpi->oxcf.lag_in_frames > 0 && cpi->oxcf.rc_mode == VPX_VBR) &&
4532 !cpi->use_svc)
4533 cpi->ref_frame_flags &= (~VP9_GOLD_FLAG);
4534
4535 if (sf->partition_search_type == SOURCE_VAR_BASED_PARTITION)
4536 source_var_based_partition_search_method(cpi);
4537 }
4538
4539 {
4540 struct vpx_usec_timer emr_timer;
4541 vpx_usec_timer_start(&emr_timer);
4542
4543 #if CONFIG_FP_MB_STATS
4544 if (cpi->use_fp_mb_stats) {
4545 input_fpmb_stats(&cpi->twopass.firstpass_mb_stats, cm,
4546 &cpi->twopass.this_frame_mb_stats);
4547 }
4548 #endif
4549
4550 if (!cpi->row_mt) {
4551 cpi->row_mt_sync_read_ptr = vp9_row_mt_sync_read_dummy;
4552 cpi->row_mt_sync_write_ptr = vp9_row_mt_sync_write_dummy;
4553 // If allowed, encoding tiles in parallel with one thread handling one
4554 // tile when row based multi-threading is disabled.
4555 if (VPXMIN(cpi->oxcf.max_threads, 1 << cm->log2_tile_cols) > 1)
4556 vp9_encode_tiles_mt(cpi);
4557 else
4558 encode_tiles(cpi);
4559 } else {
4560 cpi->row_mt_sync_read_ptr = vp9_row_mt_sync_read;
4561 cpi->row_mt_sync_write_ptr = vp9_row_mt_sync_write;
4562 vp9_encode_tiles_row_mt(cpi);
4563 }
4564
4565 vpx_usec_timer_mark(&emr_timer);
4566 cpi->time_encode_sb_row += vpx_usec_timer_elapsed(&emr_timer);
4567 }
4568
4569 sf->skip_encode_frame =
4570 sf->skip_encode_sb ? get_skip_encode_frame(cm, td) : 0;
4571
4572 #if 0
4573 // Keep record of the total distortion this time around for future use
4574 cpi->last_frame_distortion = cpi->frame_distortion;
4575 #endif
4576 }
4577
get_interp_filter(const int64_t threshes[SWITCHABLE_FILTER_CONTEXTS],int is_alt_ref)4578 static INTERP_FILTER get_interp_filter(
4579 const int64_t threshes[SWITCHABLE_FILTER_CONTEXTS], int is_alt_ref) {
4580 if (!is_alt_ref && threshes[EIGHTTAP_SMOOTH] > threshes[EIGHTTAP] &&
4581 threshes[EIGHTTAP_SMOOTH] > threshes[EIGHTTAP_SHARP] &&
4582 threshes[EIGHTTAP_SMOOTH] > threshes[SWITCHABLE - 1]) {
4583 return EIGHTTAP_SMOOTH;
4584 } else if (threshes[EIGHTTAP_SHARP] > threshes[EIGHTTAP] &&
4585 threshes[EIGHTTAP_SHARP] > threshes[SWITCHABLE - 1]) {
4586 return EIGHTTAP_SHARP;
4587 } else if (threshes[EIGHTTAP] > threshes[SWITCHABLE - 1]) {
4588 return EIGHTTAP;
4589 } else {
4590 return SWITCHABLE;
4591 }
4592 }
4593
compute_frame_aq_offset(struct VP9_COMP * cpi)4594 static int compute_frame_aq_offset(struct VP9_COMP *cpi) {
4595 VP9_COMMON *const cm = &cpi->common;
4596 MODE_INFO **mi_8x8_ptr = cm->mi_grid_visible;
4597 struct segmentation *const seg = &cm->seg;
4598
4599 int mi_row, mi_col;
4600 int sum_delta = 0;
4601 int map_index = 0;
4602 int qdelta_index;
4603 int segment_id;
4604
4605 for (mi_row = 0; mi_row < cm->mi_rows; mi_row++) {
4606 MODE_INFO **mi_8x8 = mi_8x8_ptr;
4607 for (mi_col = 0; mi_col < cm->mi_cols; mi_col++, mi_8x8++) {
4608 segment_id = mi_8x8[0]->segment_id;
4609 qdelta_index = get_segdata(seg, segment_id, SEG_LVL_ALT_Q);
4610 sum_delta += qdelta_index;
4611 map_index++;
4612 }
4613 mi_8x8_ptr += cm->mi_stride;
4614 }
4615
4616 return sum_delta / (cm->mi_rows * cm->mi_cols);
4617 }
4618
vp9_encode_frame(VP9_COMP * cpi)4619 void vp9_encode_frame(VP9_COMP *cpi) {
4620 VP9_COMMON *const cm = &cpi->common;
4621
4622 // In the longer term the encoder should be generalized to match the
4623 // decoder such that we allow compound where one of the 3 buffers has a
4624 // different sign bias and that buffer is then the fixed ref. However, this
4625 // requires further work in the rd loop. For now the only supported encoder
4626 // side behavior is where the ALT ref buffer has opposite sign bias to
4627 // the other two.
4628 if (!frame_is_intra_only(cm)) {
4629 if ((cm->ref_frame_sign_bias[ALTREF_FRAME] ==
4630 cm->ref_frame_sign_bias[GOLDEN_FRAME]) ||
4631 (cm->ref_frame_sign_bias[ALTREF_FRAME] ==
4632 cm->ref_frame_sign_bias[LAST_FRAME])) {
4633 cpi->allow_comp_inter_inter = 0;
4634 } else {
4635 cpi->allow_comp_inter_inter = 1;
4636 cm->comp_fixed_ref = ALTREF_FRAME;
4637 cm->comp_var_ref[0] = LAST_FRAME;
4638 cm->comp_var_ref[1] = GOLDEN_FRAME;
4639 }
4640 }
4641
4642 if (cpi->sf.frame_parameter_update) {
4643 int i;
4644 RD_OPT *const rd_opt = &cpi->rd;
4645 FRAME_COUNTS *counts = cpi->td.counts;
4646 RD_COUNTS *const rdc = &cpi->td.rd_counts;
4647
4648 // This code does a single RD pass over the whole frame assuming
4649 // either compound, single or hybrid prediction as per whatever has
4650 // worked best for that type of frame in the past.
4651 // It also predicts whether another coding mode would have worked
4652 // better than this coding mode. If that is the case, it remembers
4653 // that for subsequent frames.
4654 // It also does the same analysis for transform size selection.
4655 const MV_REFERENCE_FRAME frame_type = get_frame_type(cpi);
4656 int64_t *const mode_thrs = rd_opt->prediction_type_threshes[frame_type];
4657 int64_t *const filter_thrs = rd_opt->filter_threshes[frame_type];
4658 const int is_alt_ref = frame_type == ALTREF_FRAME;
4659
4660 /* prediction (compound, single or hybrid) mode selection */
4661 if (is_alt_ref || !cpi->allow_comp_inter_inter)
4662 cm->reference_mode = SINGLE_REFERENCE;
4663 else if (mode_thrs[COMPOUND_REFERENCE] > mode_thrs[SINGLE_REFERENCE] &&
4664 mode_thrs[COMPOUND_REFERENCE] > mode_thrs[REFERENCE_MODE_SELECT] &&
4665 check_dual_ref_flags(cpi) && cpi->static_mb_pct == 100)
4666 cm->reference_mode = COMPOUND_REFERENCE;
4667 else if (mode_thrs[SINGLE_REFERENCE] > mode_thrs[REFERENCE_MODE_SELECT])
4668 cm->reference_mode = SINGLE_REFERENCE;
4669 else
4670 cm->reference_mode = REFERENCE_MODE_SELECT;
4671
4672 if (cm->interp_filter == SWITCHABLE)
4673 cm->interp_filter = get_interp_filter(filter_thrs, is_alt_ref);
4674
4675 encode_frame_internal(cpi);
4676
4677 for (i = 0; i < REFERENCE_MODES; ++i)
4678 mode_thrs[i] = (mode_thrs[i] + rdc->comp_pred_diff[i] / cm->MBs) / 2;
4679
4680 for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i)
4681 filter_thrs[i] = (filter_thrs[i] + rdc->filter_diff[i] / cm->MBs) / 2;
4682
4683 if (cm->reference_mode == REFERENCE_MODE_SELECT) {
4684 int single_count_zero = 0;
4685 int comp_count_zero = 0;
4686
4687 for (i = 0; i < COMP_INTER_CONTEXTS; i++) {
4688 single_count_zero += counts->comp_inter[i][0];
4689 comp_count_zero += counts->comp_inter[i][1];
4690 }
4691
4692 if (comp_count_zero == 0) {
4693 cm->reference_mode = SINGLE_REFERENCE;
4694 vp9_zero(counts->comp_inter);
4695 } else if (single_count_zero == 0) {
4696 cm->reference_mode = COMPOUND_REFERENCE;
4697 vp9_zero(counts->comp_inter);
4698 }
4699 }
4700
4701 if (cm->tx_mode == TX_MODE_SELECT) {
4702 int count4x4 = 0;
4703 int count8x8_lp = 0, count8x8_8x8p = 0;
4704 int count16x16_16x16p = 0, count16x16_lp = 0;
4705 int count32x32 = 0;
4706
4707 for (i = 0; i < TX_SIZE_CONTEXTS; ++i) {
4708 count4x4 += counts->tx.p32x32[i][TX_4X4];
4709 count4x4 += counts->tx.p16x16[i][TX_4X4];
4710 count4x4 += counts->tx.p8x8[i][TX_4X4];
4711
4712 count8x8_lp += counts->tx.p32x32[i][TX_8X8];
4713 count8x8_lp += counts->tx.p16x16[i][TX_8X8];
4714 count8x8_8x8p += counts->tx.p8x8[i][TX_8X8];
4715
4716 count16x16_16x16p += counts->tx.p16x16[i][TX_16X16];
4717 count16x16_lp += counts->tx.p32x32[i][TX_16X16];
4718 count32x32 += counts->tx.p32x32[i][TX_32X32];
4719 }
4720 if (count4x4 == 0 && count16x16_lp == 0 && count16x16_16x16p == 0 &&
4721 count32x32 == 0) {
4722 cm->tx_mode = ALLOW_8X8;
4723 reset_skip_tx_size(cm, TX_8X8);
4724 } else if (count8x8_8x8p == 0 && count16x16_16x16p == 0 &&
4725 count8x8_lp == 0 && count16x16_lp == 0 && count32x32 == 0) {
4726 cm->tx_mode = ONLY_4X4;
4727 reset_skip_tx_size(cm, TX_4X4);
4728 } else if (count8x8_lp == 0 && count16x16_lp == 0 && count4x4 == 0) {
4729 cm->tx_mode = ALLOW_32X32;
4730 } else if (count32x32 == 0 && count8x8_lp == 0 && count4x4 == 0) {
4731 cm->tx_mode = ALLOW_16X16;
4732 reset_skip_tx_size(cm, TX_16X16);
4733 }
4734 }
4735 } else {
4736 cm->reference_mode = SINGLE_REFERENCE;
4737 encode_frame_internal(cpi);
4738 }
4739
4740 // If segmented AQ is enabled compute the average AQ weighting.
4741 if (cm->seg.enabled && (cpi->oxcf.aq_mode != NO_AQ) &&
4742 (cm->seg.update_map || cm->seg.update_data)) {
4743 cm->seg.aq_av_offset = compute_frame_aq_offset(cpi);
4744 }
4745 }
4746
sum_intra_stats(FRAME_COUNTS * counts,const MODE_INFO * mi)4747 static void sum_intra_stats(FRAME_COUNTS *counts, const MODE_INFO *mi) {
4748 const PREDICTION_MODE y_mode = mi->mode;
4749 const PREDICTION_MODE uv_mode = mi->uv_mode;
4750 const BLOCK_SIZE bsize = mi->sb_type;
4751
4752 if (bsize < BLOCK_8X8) {
4753 int idx, idy;
4754 const int num_4x4_w = num_4x4_blocks_wide_lookup[bsize];
4755 const int num_4x4_h = num_4x4_blocks_high_lookup[bsize];
4756 for (idy = 0; idy < 2; idy += num_4x4_h)
4757 for (idx = 0; idx < 2; idx += num_4x4_w)
4758 ++counts->y_mode[0][mi->bmi[idy * 2 + idx].as_mode];
4759 } else {
4760 ++counts->y_mode[size_group_lookup[bsize]][y_mode];
4761 }
4762
4763 ++counts->uv_mode[y_mode][uv_mode];
4764 }
4765
update_zeromv_cnt(VP9_COMP * const cpi,const MODE_INFO * const mi,int mi_row,int mi_col,BLOCK_SIZE bsize)4766 static void update_zeromv_cnt(VP9_COMP *const cpi, const MODE_INFO *const mi,
4767 int mi_row, int mi_col, BLOCK_SIZE bsize) {
4768 const VP9_COMMON *const cm = &cpi->common;
4769 MV mv = mi->mv[0].as_mv;
4770 const int bw = num_8x8_blocks_wide_lookup[bsize];
4771 const int bh = num_8x8_blocks_high_lookup[bsize];
4772 const int xmis = VPXMIN(cm->mi_cols - mi_col, bw);
4773 const int ymis = VPXMIN(cm->mi_rows - mi_row, bh);
4774 const int block_index = mi_row * cm->mi_cols + mi_col;
4775 int x, y;
4776 for (y = 0; y < ymis; y++)
4777 for (x = 0; x < xmis; x++) {
4778 int map_offset = block_index + y * cm->mi_cols + x;
4779 if (is_inter_block(mi) && mi->segment_id <= CR_SEGMENT_ID_BOOST2) {
4780 if (abs(mv.row) < 8 && abs(mv.col) < 8) {
4781 if (cpi->consec_zero_mv[map_offset] < 255)
4782 cpi->consec_zero_mv[map_offset]++;
4783 } else {
4784 cpi->consec_zero_mv[map_offset] = 0;
4785 }
4786 }
4787 }
4788 }
4789
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)4790 static void encode_superblock(VP9_COMP *cpi, ThreadData *td, TOKENEXTRA **t,
4791 int output_enabled, int mi_row, int mi_col,
4792 BLOCK_SIZE bsize, PICK_MODE_CONTEXT *ctx) {
4793 VP9_COMMON *const cm = &cpi->common;
4794 MACROBLOCK *const x = &td->mb;
4795 MACROBLOCKD *const xd = &x->e_mbd;
4796 MODE_INFO *mi = xd->mi[0];
4797 const int seg_skip =
4798 segfeature_active(&cm->seg, mi->segment_id, SEG_LVL_SKIP);
4799 x->skip_recode = !x->select_tx_size && mi->sb_type >= BLOCK_8X8 &&
4800 cpi->oxcf.aq_mode != COMPLEXITY_AQ &&
4801 cpi->oxcf.aq_mode != CYCLIC_REFRESH_AQ &&
4802 cpi->sf.allow_skip_recode;
4803
4804 if (!x->skip_recode && !cpi->sf.use_nonrd_pick_mode)
4805 memset(x->skip_txfm, 0, sizeof(x->skip_txfm));
4806
4807 x->skip_optimize = ctx->is_coded;
4808 ctx->is_coded = 1;
4809 x->use_lp32x32fdct = cpi->sf.use_lp32x32fdct;
4810 x->skip_encode = (!output_enabled && cpi->sf.skip_encode_frame &&
4811 x->q_index < QIDX_SKIP_THRESH);
4812
4813 if (x->skip_encode) return;
4814
4815 if (!is_inter_block(mi)) {
4816 int plane;
4817 #if CONFIG_BETTER_HW_COMPATIBILITY && CONFIG_VP9_HIGHBITDEPTH
4818 if ((xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) &&
4819 (xd->above_mi == NULL || xd->left_mi == NULL) &&
4820 need_top_left[mi->uv_mode])
4821 assert(0);
4822 #endif // CONFIG_BETTER_HW_COMPATIBILITY && CONFIG_VP9_HIGHBITDEPTH
4823 mi->skip = 1;
4824 for (plane = 0; plane < MAX_MB_PLANE; ++plane)
4825 vp9_encode_intra_block_plane(x, VPXMAX(bsize, BLOCK_8X8), plane, 1);
4826 if (output_enabled) sum_intra_stats(td->counts, mi);
4827 vp9_tokenize_sb(cpi, td, t, !output_enabled, seg_skip,
4828 VPXMAX(bsize, BLOCK_8X8));
4829 } else {
4830 int ref;
4831 const int is_compound = has_second_ref(mi);
4832 set_ref_ptrs(cm, xd, mi->ref_frame[0], mi->ref_frame[1]);
4833 for (ref = 0; ref < 1 + is_compound; ++ref) {
4834 YV12_BUFFER_CONFIG *cfg = get_ref_frame_buffer(cpi, mi->ref_frame[ref]);
4835 assert(cfg != NULL);
4836 vp9_setup_pre_planes(xd, ref, cfg, mi_row, mi_col,
4837 &xd->block_refs[ref]->sf);
4838 }
4839 if (!(cpi->sf.reuse_inter_pred_sby && ctx->pred_pixel_ready) || seg_skip)
4840 vp9_build_inter_predictors_sby(xd, mi_row, mi_col,
4841 VPXMAX(bsize, BLOCK_8X8));
4842
4843 vp9_build_inter_predictors_sbuv(xd, mi_row, mi_col,
4844 VPXMAX(bsize, BLOCK_8X8));
4845
4846 vp9_encode_sb(x, VPXMAX(bsize, BLOCK_8X8));
4847 vp9_tokenize_sb(cpi, td, t, !output_enabled, seg_skip,
4848 VPXMAX(bsize, BLOCK_8X8));
4849 }
4850
4851 if (seg_skip) {
4852 assert(mi->skip);
4853 }
4854
4855 if (output_enabled) {
4856 if (cm->tx_mode == TX_MODE_SELECT && mi->sb_type >= BLOCK_8X8 &&
4857 !(is_inter_block(mi) && mi->skip)) {
4858 ++get_tx_counts(max_txsize_lookup[bsize], get_tx_size_context(xd),
4859 &td->counts->tx)[mi->tx_size];
4860 } else {
4861 // The new intra coding scheme requires no change of transform size
4862 if (is_inter_block(mi)) {
4863 mi->tx_size = VPXMIN(tx_mode_to_biggest_tx_size[cm->tx_mode],
4864 max_txsize_lookup[bsize]);
4865 } else {
4866 mi->tx_size = (bsize >= BLOCK_8X8) ? mi->tx_size : TX_4X4;
4867 }
4868 }
4869
4870 ++td->counts->tx.tx_totals[mi->tx_size];
4871 ++td->counts->tx.tx_totals[get_uv_tx_size(mi, &xd->plane[1])];
4872 if (cm->seg.enabled && cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ)
4873 vp9_cyclic_refresh_update_sb_postencode(cpi, mi, mi_row, mi_col, bsize);
4874 if (cpi->oxcf.pass == 0 && cpi->svc.temporal_layer_id == 0)
4875 update_zeromv_cnt(cpi, mi, mi_row, mi_col, bsize);
4876 }
4877 }
4878