• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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_config.h"
17 
18 #include "vpx_ports/vpx_timer.h"
19 
20 #include "vp9/common/vp9_common.h"
21 #include "vp9/common/vp9_entropy.h"
22 #include "vp9/common/vp9_entropymode.h"
23 #include "vp9/common/vp9_idct.h"
24 #include "vp9/common/vp9_mvref_common.h"
25 #include "vp9/common/vp9_pred_common.h"
26 #include "vp9/common/vp9_quant_common.h"
27 #include "vp9/common/vp9_reconintra.h"
28 #include "vp9/common/vp9_reconinter.h"
29 #include "vp9/common/vp9_seg_common.h"
30 #include "vp9/common/vp9_systemdependent.h"
31 #include "vp9/common/vp9_tile_common.h"
32 
33 #include "vp9/encoder/vp9_aq_complexity.h"
34 #include "vp9/encoder/vp9_aq_cyclicrefresh.h"
35 #include "vp9/encoder/vp9_aq_variance.h"
36 #include "vp9/encoder/vp9_encodeframe.h"
37 #include "vp9/encoder/vp9_encodemb.h"
38 #include "vp9/encoder/vp9_encodemv.h"
39 #include "vp9/encoder/vp9_extend.h"
40 #include "vp9/encoder/vp9_pickmode.h"
41 #include "vp9/encoder/vp9_rd.h"
42 #include "vp9/encoder/vp9_rdopt.h"
43 #include "vp9/encoder/vp9_segmentation.h"
44 #include "vp9/encoder/vp9_tokenize.h"
45 
46 #define GF_ZEROMV_ZBIN_BOOST 0
47 #define LF_ZEROMV_ZBIN_BOOST 0
48 #define MV_ZBIN_BOOST        0
49 #define SPLIT_MV_ZBIN_BOOST  0
50 #define INTRA_ZBIN_BOOST     0
51 
52 static void encode_superblock(VP9_COMP *cpi, TOKENEXTRA **t, int output_enabled,
53                               int mi_row, int mi_col, BLOCK_SIZE bsize,
54                               PICK_MODE_CONTEXT *ctx);
55 
56 // Motion vector component magnitude threshold for defining fast motion.
57 #define FAST_MOTION_MV_THRESH 24
58 
59 // This is used as a reference when computing the source variance for the
60 //  purposes of activity masking.
61 // Eventually this should be replaced by custom no-reference routines,
62 //  which will be faster.
63 static const uint8_t VP9_VAR_OFFS[64] = {
64   128, 128, 128, 128, 128, 128, 128, 128,
65   128, 128, 128, 128, 128, 128, 128, 128,
66   128, 128, 128, 128, 128, 128, 128, 128,
67   128, 128, 128, 128, 128, 128, 128, 128,
68   128, 128, 128, 128, 128, 128, 128, 128,
69   128, 128, 128, 128, 128, 128, 128, 128,
70   128, 128, 128, 128, 128, 128, 128, 128,
71   128, 128, 128, 128, 128, 128, 128, 128
72 };
73 
get_sby_perpixel_variance(VP9_COMP * cpi,const struct buf_2d * ref,BLOCK_SIZE bs)74 static unsigned int get_sby_perpixel_variance(VP9_COMP *cpi,
75                                               const struct buf_2d *ref,
76                                               BLOCK_SIZE bs) {
77   unsigned int sse;
78   const unsigned int var = cpi->fn_ptr[bs].vf(ref->buf, ref->stride,
79                                               VP9_VAR_OFFS, 0, &sse);
80   return ROUND_POWER_OF_TWO(var, num_pels_log2_lookup[bs]);
81 }
82 
get_sby_perpixel_diff_variance(VP9_COMP * cpi,const struct buf_2d * ref,int mi_row,int mi_col,BLOCK_SIZE bs)83 static unsigned int get_sby_perpixel_diff_variance(VP9_COMP *cpi,
84                                                    const struct buf_2d *ref,
85                                                    int mi_row, int mi_col,
86                                                    BLOCK_SIZE bs) {
87   const YV12_BUFFER_CONFIG *last = get_ref_frame_buffer(cpi, LAST_FRAME);
88   const uint8_t* last_y = &last->y_buffer[mi_row * MI_SIZE * last->y_stride +
89                                               mi_col * MI_SIZE];
90   unsigned int sse;
91   const unsigned int var = cpi->fn_ptr[bs].vf(ref->buf, ref->stride,
92                                               last_y, last->y_stride, &sse);
93   return ROUND_POWER_OF_TWO(var, num_pels_log2_lookup[bs]);
94 }
95 
get_rd_var_based_fixed_partition(VP9_COMP * cpi,int mi_row,int mi_col)96 static BLOCK_SIZE get_rd_var_based_fixed_partition(VP9_COMP *cpi,
97                                                    int mi_row,
98                                                    int mi_col) {
99   unsigned int var = get_sby_perpixel_diff_variance(cpi, &cpi->mb.plane[0].src,
100                                                     mi_row, mi_col,
101                                                     BLOCK_64X64);
102   if (var < 8)
103     return BLOCK_64X64;
104   else if (var < 128)
105     return BLOCK_32X32;
106   else if (var < 2048)
107     return BLOCK_16X16;
108   else
109     return BLOCK_8X8;
110 }
111 
get_nonrd_var_based_fixed_partition(VP9_COMP * cpi,int mi_row,int mi_col)112 static BLOCK_SIZE get_nonrd_var_based_fixed_partition(VP9_COMP *cpi,
113                                                       int mi_row,
114                                                       int mi_col) {
115   unsigned int var = get_sby_perpixel_diff_variance(cpi, &cpi->mb.plane[0].src,
116                                                     mi_row, mi_col,
117                                                     BLOCK_64X64);
118   if (var < 4)
119     return BLOCK_64X64;
120   else if (var < 10)
121     return BLOCK_32X32;
122   else
123     return BLOCK_16X16;
124 }
125 
126 // Lighter version of set_offsets that only sets the mode info
127 // pointers.
set_modeinfo_offsets(VP9_COMMON * const cm,MACROBLOCKD * const xd,int mi_row,int mi_col)128 static INLINE void set_modeinfo_offsets(VP9_COMMON *const cm,
129                                         MACROBLOCKD *const xd,
130                                         int mi_row,
131                                         int mi_col) {
132   const int idx_str = xd->mi_stride * mi_row + mi_col;
133   xd->mi = cm->mi + idx_str;
134   xd->mi[0].src_mi = &xd->mi[0];
135 }
136 
set_offsets(VP9_COMP * cpi,const TileInfo * const tile,int mi_row,int mi_col,BLOCK_SIZE bsize)137 static void set_offsets(VP9_COMP *cpi, const TileInfo *const tile,
138                         int mi_row, int mi_col, BLOCK_SIZE bsize) {
139   MACROBLOCK *const x = &cpi->mb;
140   VP9_COMMON *const cm = &cpi->common;
141   MACROBLOCKD *const xd = &x->e_mbd;
142   MB_MODE_INFO *mbmi;
143   const int mi_width = num_8x8_blocks_wide_lookup[bsize];
144   const int mi_height = num_8x8_blocks_high_lookup[bsize];
145   const struct segmentation *const seg = &cm->seg;
146 
147   set_skip_context(xd, mi_row, mi_col);
148 
149   set_modeinfo_offsets(cm, xd, mi_row, mi_col);
150 
151   mbmi = &xd->mi[0].src_mi->mbmi;
152 
153   // Set up destination pointers.
154   vp9_setup_dst_planes(xd->plane, get_frame_new_buffer(cm), mi_row, mi_col);
155 
156   // Set up limit values for MV components.
157   // Mv beyond the range do not produce new/different prediction block.
158   x->mv_row_min = -(((mi_row + mi_height) * MI_SIZE) + VP9_INTERP_EXTEND);
159   x->mv_col_min = -(((mi_col + mi_width) * MI_SIZE) + VP9_INTERP_EXTEND);
160   x->mv_row_max = (cm->mi_rows - mi_row) * MI_SIZE + VP9_INTERP_EXTEND;
161   x->mv_col_max = (cm->mi_cols - mi_col) * MI_SIZE + VP9_INTERP_EXTEND;
162 
163   // Set up distance of MB to edge of frame in 1/8th pel units.
164   assert(!(mi_col & (mi_width - 1)) && !(mi_row & (mi_height - 1)));
165   set_mi_row_col(xd, tile, mi_row, mi_height, mi_col, mi_width,
166                  cm->mi_rows, cm->mi_cols);
167 
168   // Set up source buffers.
169   vp9_setup_src_planes(x, cpi->Source, mi_row, mi_col);
170 
171   // R/D setup.
172   x->rddiv = cpi->rd.RDDIV;
173   x->rdmult = cpi->rd.RDMULT;
174 
175   // Setup segment ID.
176   if (seg->enabled) {
177     if (cpi->oxcf.aq_mode != VARIANCE_AQ) {
178       const uint8_t *const map = seg->update_map ? cpi->segmentation_map
179                                                  : cm->last_frame_seg_map;
180       mbmi->segment_id = vp9_get_segment_id(cm, map, bsize, mi_row, mi_col);
181     }
182     vp9_init_plane_quantizers(cpi, x);
183 
184     x->encode_breakout = cpi->segment_encode_breakout[mbmi->segment_id];
185   } else {
186     mbmi->segment_id = 0;
187     x->encode_breakout = cpi->encode_breakout;
188   }
189 }
190 
duplicate_mode_info_in_sb(VP9_COMMON * cm,MACROBLOCKD * xd,int mi_row,int mi_col,BLOCK_SIZE bsize)191 static void duplicate_mode_info_in_sb(VP9_COMMON *cm, MACROBLOCKD *xd,
192                                       int mi_row, int mi_col,
193                                       BLOCK_SIZE bsize) {
194   const int block_width = num_8x8_blocks_wide_lookup[bsize];
195   const int block_height = num_8x8_blocks_high_lookup[bsize];
196   int i, j;
197   for (j = 0; j < block_height; ++j)
198     for (i = 0; i < block_width; ++i) {
199       if (mi_row + j < cm->mi_rows && mi_col + i < cm->mi_cols)
200         xd->mi[j * xd->mi_stride + i].src_mi = &xd->mi[0];
201     }
202 }
203 
set_block_size(VP9_COMP * const cpi,int mi_row,int mi_col,BLOCK_SIZE bsize)204 static void set_block_size(VP9_COMP * const cpi,
205                            int mi_row, int mi_col,
206                            BLOCK_SIZE bsize) {
207   if (cpi->common.mi_cols > mi_col && cpi->common.mi_rows > mi_row) {
208     MACROBLOCKD *const xd = &cpi->mb.e_mbd;
209     set_modeinfo_offsets(&cpi->common, xd, mi_row, mi_col);
210     xd->mi[0].src_mi->mbmi.sb_type = bsize;
211     duplicate_mode_info_in_sb(&cpi->common, xd, mi_row, mi_col, bsize);
212   }
213 }
214 
215 typedef struct {
216   int64_t sum_square_error;
217   int64_t sum_error;
218   int count;
219   int variance;
220 } var;
221 
222 typedef struct {
223   var none;
224   var horz[2];
225   var vert[2];
226 } partition_variance;
227 
228 typedef struct {
229   partition_variance part_variances;
230   var split[4];
231 } v8x8;
232 
233 typedef struct {
234   partition_variance part_variances;
235   v8x8 split[4];
236 } v16x16;
237 
238 typedef struct {
239   partition_variance part_variances;
240   v16x16 split[4];
241 } v32x32;
242 
243 typedef struct {
244   partition_variance part_variances;
245   v32x32 split[4];
246 } v64x64;
247 
248 typedef struct {
249   partition_variance *part_variances;
250   var *split[4];
251 } variance_node;
252 
253 typedef enum {
254   V16X16,
255   V32X32,
256   V64X64,
257 } TREE_LEVEL;
258 
tree_to_node(void * data,BLOCK_SIZE bsize,variance_node * node)259 static void tree_to_node(void *data, BLOCK_SIZE bsize, variance_node *node) {
260   int i;
261   node->part_variances = NULL;
262   vpx_memset(node->split, 0, sizeof(node->split));
263   switch (bsize) {
264     case BLOCK_64X64: {
265       v64x64 *vt = (v64x64 *) data;
266       node->part_variances = &vt->part_variances;
267       for (i = 0; i < 4; i++)
268         node->split[i] = &vt->split[i].part_variances.none;
269       break;
270     }
271     case BLOCK_32X32: {
272       v32x32 *vt = (v32x32 *) data;
273       node->part_variances = &vt->part_variances;
274       for (i = 0; i < 4; i++)
275         node->split[i] = &vt->split[i].part_variances.none;
276       break;
277     }
278     case BLOCK_16X16: {
279       v16x16 *vt = (v16x16 *) data;
280       node->part_variances = &vt->part_variances;
281       for (i = 0; i < 4; i++)
282         node->split[i] = &vt->split[i].part_variances.none;
283       break;
284     }
285     case BLOCK_8X8: {
286       v8x8 *vt = (v8x8 *) data;
287       node->part_variances = &vt->part_variances;
288       for (i = 0; i < 4; i++)
289         node->split[i] = &vt->split[i];
290       break;
291     }
292     default: {
293       assert(0);
294       break;
295     }
296   }
297 }
298 
299 // Set variance values given sum square error, sum error, count.
fill_variance(int64_t s2,int64_t s,int c,var * v)300 static void fill_variance(int64_t s2, int64_t s, int c, var *v) {
301   v->sum_square_error = s2;
302   v->sum_error = s;
303   v->count = c;
304   if (c > 0)
305     v->variance = (int)(256 *
306                         (v->sum_square_error - v->sum_error * v->sum_error /
307                          v->count) / v->count);
308   else
309     v->variance = 0;
310 }
311 
sum_2_variances(const var * a,const var * b,var * r)312 void sum_2_variances(const var *a, const var *b, var *r) {
313   fill_variance(a->sum_square_error + b->sum_square_error,
314                 a->sum_error + b->sum_error, a->count + b->count, r);
315 }
316 
fill_variance_tree(void * data,BLOCK_SIZE bsize)317 static void fill_variance_tree(void *data, BLOCK_SIZE bsize) {
318   variance_node node;
319   tree_to_node(data, bsize, &node);
320   sum_2_variances(node.split[0], node.split[1], &node.part_variances->horz[0]);
321   sum_2_variances(node.split[2], node.split[3], &node.part_variances->horz[1]);
322   sum_2_variances(node.split[0], node.split[2], &node.part_variances->vert[0]);
323   sum_2_variances(node.split[1], node.split[3], &node.part_variances->vert[1]);
324   sum_2_variances(&node.part_variances->vert[0], &node.part_variances->vert[1],
325                   &node.part_variances->none);
326 }
327 
set_vt_partitioning(VP9_COMP * cpi,void * data,BLOCK_SIZE bsize,int mi_row,int mi_col)328 static int set_vt_partitioning(VP9_COMP *cpi,
329                                void *data,
330                                BLOCK_SIZE bsize,
331                                int mi_row,
332                                int mi_col) {
333   VP9_COMMON * const cm = &cpi->common;
334   variance_node vt;
335   const int block_width = num_8x8_blocks_wide_lookup[bsize];
336   const int block_height = num_8x8_blocks_high_lookup[bsize];
337   // TODO(debargha): Choose this more intelligently.
338   const int64_t threshold_multiplier = 25;
339   int64_t threshold = threshold_multiplier * cpi->common.base_qindex;
340   assert(block_height == block_width);
341 
342   tree_to_node(data, bsize, &vt);
343 
344   // Split none is available only if we have more than half a block size
345   // in width and height inside the visible image.
346   if (mi_col + block_width / 2 < cm->mi_cols &&
347       mi_row + block_height / 2 < cm->mi_rows &&
348       vt.part_variances->none.variance < threshold) {
349     set_block_size(cpi, mi_row, mi_col, bsize);
350     return 1;
351   }
352 
353   // Vertical split is available on all but the bottom border.
354   if (mi_row + block_height / 2 < cm->mi_rows &&
355       vt.part_variances->vert[0].variance < threshold &&
356       vt.part_variances->vert[1].variance < threshold) {
357     BLOCK_SIZE subsize = get_subsize(bsize, PARTITION_VERT);
358     set_block_size(cpi, mi_row, mi_col, subsize);
359     set_block_size(cpi, mi_row, mi_col + block_width / 2, subsize);
360     return 1;
361   }
362 
363   // Horizontal split is available on all but the right border.
364   if (mi_col + block_width / 2 < cm->mi_cols &&
365       vt.part_variances->horz[0].variance < threshold &&
366       vt.part_variances->horz[1].variance < threshold) {
367     BLOCK_SIZE subsize = get_subsize(bsize, PARTITION_HORZ);
368     set_block_size(cpi, mi_row, mi_col, subsize);
369     set_block_size(cpi, mi_row + block_height / 2, mi_col, subsize);
370     return 1;
371   }
372   return 0;
373 }
374 
375 // TODO(debargha): Fix this function and make it work as expected.
choose_partitioning(VP9_COMP * cpi,const TileInfo * const tile,int mi_row,int mi_col)376 static void choose_partitioning(VP9_COMP *cpi,
377                                 const TileInfo *const tile,
378                                 int mi_row, int mi_col) {
379   VP9_COMMON * const cm = &cpi->common;
380   MACROBLOCK *x = &cpi->mb;
381   MACROBLOCKD *xd = &cpi->mb.e_mbd;
382 
383   int i, j, k;
384   v64x64 vt;
385   uint8_t *s;
386   const uint8_t *d;
387   int sp;
388   int dp;
389   int pixels_wide = 64, pixels_high = 64;
390   int_mv nearest_mv, near_mv;
391   const YV12_BUFFER_CONFIG *yv12 = get_ref_frame_buffer(cpi, LAST_FRAME);
392   const struct scale_factors *const sf = &cm->frame_refs[LAST_FRAME - 1].sf;
393 
394   vp9_zero(vt);
395   set_offsets(cpi, tile, mi_row, mi_col, BLOCK_64X64);
396 
397   if (xd->mb_to_right_edge < 0)
398     pixels_wide += (xd->mb_to_right_edge >> 3);
399   if (xd->mb_to_bottom_edge < 0)
400     pixels_high += (xd->mb_to_bottom_edge >> 3);
401 
402   s = x->plane[0].src.buf;
403   sp = x->plane[0].src.stride;
404 
405   if (cm->frame_type != KEY_FRAME) {
406     vp9_setup_pre_planes(xd, 0, yv12, mi_row, mi_col, sf);
407 
408     xd->mi[0].src_mi->mbmi.ref_frame[0] = LAST_FRAME;
409     xd->mi[0].src_mi->mbmi.sb_type = BLOCK_64X64;
410     vp9_find_best_ref_mvs(xd, cm->allow_high_precision_mv,
411                           xd->mi[0].src_mi->mbmi.ref_mvs[LAST_FRAME],
412                           &nearest_mv, &near_mv);
413 
414     xd->mi[0].src_mi->mbmi.mv[0] = nearest_mv;
415     vp9_build_inter_predictors_sby(xd, mi_row, mi_col, BLOCK_64X64);
416 
417     d = xd->plane[0].dst.buf;
418     dp = xd->plane[0].dst.stride;
419   } else {
420     d = VP9_VAR_OFFS;
421     dp = 0;
422   }
423 
424   // Fill in the entire tree of 8x8 variances for splits.
425   for (i = 0; i < 4; i++) {
426     const int x32_idx = ((i & 1) << 5);
427     const int y32_idx = ((i >> 1) << 5);
428     for (j = 0; j < 4; j++) {
429       const int x16_idx = x32_idx + ((j & 1) << 4);
430       const int y16_idx = y32_idx + ((j >> 1) << 4);
431       v16x16 *vst = &vt.split[i].split[j];
432       for (k = 0; k < 4; k++) {
433         int x_idx = x16_idx + ((k & 1) << 3);
434         int y_idx = y16_idx + ((k >> 1) << 3);
435         unsigned int sse = 0;
436         int sum = 0;
437         if (x_idx < pixels_wide && y_idx < pixels_high)
438           vp9_get8x8var(s + y_idx * sp + x_idx, sp,
439                         d + y_idx * dp + x_idx, dp, &sse, &sum);
440         fill_variance(sse, sum, 64, &vst->split[k].part_variances.none);
441       }
442     }
443   }
444   // Fill the rest of the variance tree by summing split partition values.
445   for (i = 0; i < 4; i++) {
446     for (j = 0; j < 4; j++) {
447       fill_variance_tree(&vt.split[i].split[j], BLOCK_16X16);
448     }
449     fill_variance_tree(&vt.split[i], BLOCK_32X32);
450   }
451   fill_variance_tree(&vt, BLOCK_64X64);
452 
453   // Now go through the entire structure,  splitting every block size until
454   // we get to one that's got a variance lower than our threshold,  or we
455   // hit 8x8.
456   if (!set_vt_partitioning(cpi, &vt, BLOCK_64X64,
457                            mi_row, mi_col)) {
458     for (i = 0; i < 4; ++i) {
459       const int x32_idx = ((i & 1) << 2);
460       const int y32_idx = ((i >> 1) << 2);
461       if (!set_vt_partitioning(cpi, &vt.split[i], BLOCK_32X32,
462                                (mi_row + y32_idx), (mi_col + x32_idx))) {
463         for (j = 0; j < 4; ++j) {
464           const int x16_idx = ((j & 1) << 1);
465           const int y16_idx = ((j >> 1) << 1);
466           // NOTE: This is a temporary hack to disable 8x8 partitions,
467           // since it works really bad - possibly due to a bug
468 #define DISABLE_8X8_VAR_BASED_PARTITION
469 #ifdef DISABLE_8X8_VAR_BASED_PARTITION
470           if (mi_row + y32_idx + y16_idx + 1 < cm->mi_rows &&
471               mi_row + x32_idx + x16_idx + 1 < cm->mi_cols) {
472             set_block_size(cpi,
473                            (mi_row + y32_idx + y16_idx),
474                            (mi_col + x32_idx + x16_idx),
475                            BLOCK_16X16);
476           } else {
477             for (k = 0; k < 4; ++k) {
478               const int x8_idx = (k & 1);
479               const int y8_idx = (k >> 1);
480               set_block_size(cpi,
481                              (mi_row + y32_idx + y16_idx + y8_idx),
482                              (mi_col + x32_idx + x16_idx + x8_idx),
483                              BLOCK_8X8);
484             }
485           }
486 #else
487           if (!set_vt_partitioning(cpi, &vt.split[i].split[j], tile,
488                                    BLOCK_16X16,
489                                    (mi_row + y32_idx + y16_idx),
490                                    (mi_col + x32_idx + x16_idx), 2)) {
491             for (k = 0; k < 4; ++k) {
492               const int x8_idx = (k & 1);
493               const int y8_idx = (k >> 1);
494               set_block_size(cpi,
495                              (mi_row + y32_idx + y16_idx + y8_idx),
496                              (mi_col + x32_idx + x16_idx + x8_idx),
497                              BLOCK_8X8);
498             }
499           }
500 #endif
501         }
502       }
503     }
504   }
505 }
506 
update_state(VP9_COMP * cpi,PICK_MODE_CONTEXT * ctx,int mi_row,int mi_col,BLOCK_SIZE bsize,int output_enabled)507 static void update_state(VP9_COMP *cpi, PICK_MODE_CONTEXT *ctx,
508                          int mi_row, int mi_col, BLOCK_SIZE bsize,
509                          int output_enabled) {
510   int i, x_idx, y;
511   VP9_COMMON *const cm = &cpi->common;
512   RD_OPT *const rd_opt = &cpi->rd;
513   MACROBLOCK *const x = &cpi->mb;
514   MACROBLOCKD *const xd = &x->e_mbd;
515   struct macroblock_plane *const p = x->plane;
516   struct macroblockd_plane *const pd = xd->plane;
517   MODE_INFO *mi = &ctx->mic;
518   MB_MODE_INFO *const mbmi = &xd->mi[0].src_mi->mbmi;
519   MODE_INFO *mi_addr = &xd->mi[0];
520   const struct segmentation *const seg = &cm->seg;
521 
522   const int mis = cm->mi_stride;
523   const int mi_width = num_8x8_blocks_wide_lookup[bsize];
524   const int mi_height = num_8x8_blocks_high_lookup[bsize];
525   int max_plane;
526 
527   assert(mi->mbmi.sb_type == bsize);
528 
529   *mi_addr = *mi;
530   mi_addr->src_mi = mi_addr;
531 
532   // If segmentation in use
533   if (seg->enabled && output_enabled) {
534     // For in frame complexity AQ copy the segment id from the segment map.
535     if (cpi->oxcf.aq_mode == COMPLEXITY_AQ) {
536       const uint8_t *const map = seg->update_map ? cpi->segmentation_map
537                                                  : cm->last_frame_seg_map;
538       mi_addr->mbmi.segment_id =
539         vp9_get_segment_id(cm, map, bsize, mi_row, mi_col);
540     }
541     // Else for cyclic refresh mode update the segment map, set the segment id
542     // and then update the quantizer.
543     if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ) {
544       vp9_cyclic_refresh_update_segment(cpi, &xd->mi[0].src_mi->mbmi,
545                                         mi_row, mi_col, bsize, 1);
546     }
547   }
548 
549   max_plane = is_inter_block(mbmi) ? MAX_MB_PLANE : 1;
550   for (i = 0; i < max_plane; ++i) {
551     p[i].coeff = ctx->coeff_pbuf[i][1];
552     p[i].qcoeff = ctx->qcoeff_pbuf[i][1];
553     pd[i].dqcoeff = ctx->dqcoeff_pbuf[i][1];
554     p[i].eobs = ctx->eobs_pbuf[i][1];
555   }
556 
557   for (i = max_plane; i < MAX_MB_PLANE; ++i) {
558     p[i].coeff = ctx->coeff_pbuf[i][2];
559     p[i].qcoeff = ctx->qcoeff_pbuf[i][2];
560     pd[i].dqcoeff = ctx->dqcoeff_pbuf[i][2];
561     p[i].eobs = ctx->eobs_pbuf[i][2];
562   }
563 
564   // Restore the coding context of the MB to that that was in place
565   // when the mode was picked for it
566   for (y = 0; y < mi_height; y++)
567     for (x_idx = 0; x_idx < mi_width; x_idx++)
568       if ((xd->mb_to_right_edge >> (3 + MI_SIZE_LOG2)) + mi_width > x_idx
569         && (xd->mb_to_bottom_edge >> (3 + MI_SIZE_LOG2)) + mi_height > y) {
570         xd->mi[x_idx + y * mis].src_mi = mi_addr;
571       }
572 
573   if (cpi->oxcf.aq_mode)
574     vp9_init_plane_quantizers(cpi, x);
575 
576   // FIXME(rbultje) I'm pretty sure this should go to the end of this block
577   // (i.e. after the output_enabled)
578   if (bsize < BLOCK_32X32) {
579     if (bsize < BLOCK_16X16)
580       ctx->tx_rd_diff[ALLOW_16X16] = ctx->tx_rd_diff[ALLOW_8X8];
581     ctx->tx_rd_diff[ALLOW_32X32] = ctx->tx_rd_diff[ALLOW_16X16];
582   }
583 
584   if (is_inter_block(mbmi) && mbmi->sb_type < BLOCK_8X8) {
585     mbmi->mv[0].as_int = mi->bmi[3].as_mv[0].as_int;
586     mbmi->mv[1].as_int = mi->bmi[3].as_mv[1].as_int;
587   }
588 
589   x->skip = ctx->skip;
590   vpx_memcpy(x->zcoeff_blk[mbmi->tx_size], ctx->zcoeff_blk,
591              sizeof(uint8_t) * ctx->num_4x4_blk);
592 
593   if (!output_enabled)
594     return;
595 
596   if (!vp9_segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_SKIP)) {
597     for (i = 0; i < TX_MODES; i++)
598       rd_opt->tx_select_diff[i] += ctx->tx_rd_diff[i];
599   }
600 
601 #if CONFIG_INTERNAL_STATS
602   if (frame_is_intra_only(cm)) {
603     static const int kf_mode_index[] = {
604       THR_DC        /*DC_PRED*/,
605       THR_V_PRED    /*V_PRED*/,
606       THR_H_PRED    /*H_PRED*/,
607       THR_D45_PRED  /*D45_PRED*/,
608       THR_D135_PRED /*D135_PRED*/,
609       THR_D117_PRED /*D117_PRED*/,
610       THR_D153_PRED /*D153_PRED*/,
611       THR_D207_PRED /*D207_PRED*/,
612       THR_D63_PRED  /*D63_PRED*/,
613       THR_TM        /*TM_PRED*/,
614     };
615     ++cpi->mode_chosen_counts[kf_mode_index[mbmi->mode]];
616   } else {
617     // Note how often each mode chosen as best
618     ++cpi->mode_chosen_counts[ctx->best_mode_index];
619   }
620 #endif
621   if (!frame_is_intra_only(cm)) {
622     if (is_inter_block(mbmi)) {
623       vp9_update_mv_count(cm, xd);
624 
625       if (cm->interp_filter == SWITCHABLE) {
626         const int ctx = vp9_get_pred_context_switchable_interp(xd);
627         ++cm->counts.switchable_interp[ctx][mbmi->interp_filter];
628       }
629     }
630 
631     rd_opt->comp_pred_diff[SINGLE_REFERENCE] += ctx->single_pred_diff;
632     rd_opt->comp_pred_diff[COMPOUND_REFERENCE] += ctx->comp_pred_diff;
633     rd_opt->comp_pred_diff[REFERENCE_MODE_SELECT] += ctx->hybrid_pred_diff;
634 
635     for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i)
636       rd_opt->filter_diff[i] += ctx->best_filter_diff[i];
637   }
638 }
639 
vp9_setup_src_planes(MACROBLOCK * x,const YV12_BUFFER_CONFIG * src,int mi_row,int mi_col)640 void vp9_setup_src_planes(MACROBLOCK *x, const YV12_BUFFER_CONFIG *src,
641                           int mi_row, int mi_col) {
642   uint8_t *const buffers[3] = {src->y_buffer, src->u_buffer, src->v_buffer };
643   const int strides[3] = {src->y_stride, src->uv_stride, src->uv_stride };
644   int i;
645 
646   // Set current frame pointer.
647   x->e_mbd.cur_buf = src;
648 
649   for (i = 0; i < MAX_MB_PLANE; i++)
650     setup_pred_plane(&x->plane[i].src, buffers[i], strides[i], mi_row, mi_col,
651                      NULL, x->e_mbd.plane[i].subsampling_x,
652                      x->e_mbd.plane[i].subsampling_y);
653 }
654 
set_mode_info_seg_skip(MACROBLOCK * x,TX_MODE tx_mode,int * rate,int64_t * dist,BLOCK_SIZE bsize)655 static void set_mode_info_seg_skip(MACROBLOCK *x, TX_MODE tx_mode, int *rate,
656                                    int64_t *dist, BLOCK_SIZE bsize) {
657   MACROBLOCKD *const xd = &x->e_mbd;
658   MB_MODE_INFO *const mbmi = &xd->mi[0].src_mi->mbmi;
659   INTERP_FILTER filter_ref;
660 
661   if (xd->up_available)
662     filter_ref = xd->mi[-xd->mi_stride].src_mi->mbmi.interp_filter;
663   else if (xd->left_available)
664     filter_ref = xd->mi[-1].src_mi->mbmi.interp_filter;
665   else
666     filter_ref = EIGHTTAP;
667 
668   mbmi->sb_type = bsize;
669   mbmi->mode = ZEROMV;
670   mbmi->tx_size = MIN(max_txsize_lookup[bsize],
671                       tx_mode_to_biggest_tx_size[tx_mode]);
672   mbmi->skip = 1;
673   mbmi->uv_mode = DC_PRED;
674   mbmi->ref_frame[0] = LAST_FRAME;
675   mbmi->ref_frame[1] = NONE;
676   mbmi->mv[0].as_int = 0;
677   mbmi->interp_filter = filter_ref;
678 
679   xd->mi[0].src_mi->bmi[0].as_mv[0].as_int = 0;
680   x->skip = 1;
681 
682   *rate = 0;
683   *dist = 0;
684 }
685 
rd_pick_sb_modes(VP9_COMP * cpi,const TileInfo * const tile,int mi_row,int mi_col,int * totalrate,int64_t * totaldist,BLOCK_SIZE bsize,PICK_MODE_CONTEXT * ctx,int64_t best_rd,int block)686 static void rd_pick_sb_modes(VP9_COMP *cpi, const TileInfo *const tile,
687                              int mi_row, int mi_col,
688                              int *totalrate, int64_t *totaldist,
689                              BLOCK_SIZE bsize, PICK_MODE_CONTEXT *ctx,
690                              int64_t best_rd, int block) {
691   VP9_COMMON *const cm = &cpi->common;
692   MACROBLOCK *const x = &cpi->mb;
693   MACROBLOCKD *const xd = &x->e_mbd;
694   MB_MODE_INFO *mbmi;
695   struct macroblock_plane *const p = x->plane;
696   struct macroblockd_plane *const pd = xd->plane;
697   const AQ_MODE aq_mode = cpi->oxcf.aq_mode;
698   int i, orig_rdmult;
699   double rdmult_ratio;
700 
701   vp9_clear_system_state();
702   rdmult_ratio = 1.0;  // avoid uninitialized warnings
703 
704   // Use the lower precision, but faster, 32x32 fdct for mode selection.
705   x->use_lp32x32fdct = 1;
706 
707   // TODO(JBB): Most other places in the code instead of calling the function
708   // and then checking if its not the first 8x8 we put the check in the
709   // calling function.  Do that here.
710   if (bsize < BLOCK_8X8) {
711     // When ab_index = 0 all sub-blocks are handled, so for ab_index != 0
712     // there is nothing to be done.
713     if (block != 0) {
714       *totalrate = 0;
715       *totaldist = 0;
716       return;
717     }
718   }
719 
720   set_offsets(cpi, tile, mi_row, mi_col, bsize);
721   mbmi = &xd->mi[0].src_mi->mbmi;
722   mbmi->sb_type = bsize;
723 
724   for (i = 0; i < MAX_MB_PLANE; ++i) {
725     p[i].coeff = ctx->coeff_pbuf[i][0];
726     p[i].qcoeff = ctx->qcoeff_pbuf[i][0];
727     pd[i].dqcoeff = ctx->dqcoeff_pbuf[i][0];
728     p[i].eobs = ctx->eobs_pbuf[i][0];
729   }
730   ctx->is_coded = 0;
731   ctx->skippable = 0;
732   x->skip_recode = 0;
733 
734   // Set to zero to make sure we do not use the previous encoded frame stats
735   mbmi->skip = 0;
736 
737   x->source_variance = get_sby_perpixel_variance(cpi, &x->plane[0].src, bsize);
738 
739   // Save rdmult before it might be changed, so it can be restored later.
740   orig_rdmult = x->rdmult;
741 
742   if (aq_mode == VARIANCE_AQ) {
743     const int energy = bsize <= BLOCK_16X16 ? x->mb_energy
744                                             : vp9_block_energy(cpi, x, bsize);
745     if (cm->frame_type == KEY_FRAME ||
746         cpi->refresh_alt_ref_frame ||
747         (cpi->refresh_golden_frame && !cpi->rc.is_src_frame_alt_ref)) {
748       mbmi->segment_id = vp9_vaq_segment_id(energy);
749     } else {
750       const uint8_t *const map = cm->seg.update_map ? cpi->segmentation_map
751                                                     : cm->last_frame_seg_map;
752       mbmi->segment_id = vp9_get_segment_id(cm, map, bsize, mi_row, mi_col);
753     }
754 
755     rdmult_ratio = vp9_vaq_rdmult_ratio(energy);
756     vp9_init_plane_quantizers(cpi, x);
757     vp9_clear_system_state();
758     x->rdmult = (int)round(x->rdmult * rdmult_ratio);
759   } else if (aq_mode == COMPLEXITY_AQ) {
760     const int mi_offset = mi_row * cm->mi_cols + mi_col;
761     unsigned char complexity = cpi->complexity_map[mi_offset];
762     const int is_edge = (mi_row <= 1) || (mi_row >= (cm->mi_rows - 2)) ||
763                         (mi_col <= 1) || (mi_col >= (cm->mi_cols - 2));
764     if (!is_edge && (complexity > 128))
765       x->rdmult += ((x->rdmult * (complexity - 128)) / 256);
766   } else if (aq_mode == CYCLIC_REFRESH_AQ) {
767     const uint8_t *const map = cm->seg.update_map ? cpi->segmentation_map
768                                                   : cm->last_frame_seg_map;
769     // If segment 1, use rdmult for that segment.
770     if (vp9_get_segment_id(cm, map, bsize, mi_row, mi_col))
771       x->rdmult = vp9_cyclic_refresh_get_rdmult(cpi->cyclic_refresh);
772   }
773 
774   // Find best coding mode & reconstruct the MB so it is available
775   // as a predictor for MBs that follow in the SB
776   if (frame_is_intra_only(cm)) {
777     vp9_rd_pick_intra_mode_sb(cpi, x, totalrate, totaldist, bsize, ctx,
778                               best_rd);
779   } else {
780     if (bsize >= BLOCK_8X8) {
781       if (vp9_segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_SKIP))
782         vp9_rd_pick_inter_mode_sb_seg_skip(cpi, x, totalrate, totaldist, bsize,
783                                            ctx, best_rd);
784       else
785         vp9_rd_pick_inter_mode_sb(cpi, x, tile, mi_row, mi_col,
786                                   totalrate, totaldist, bsize, ctx, best_rd);
787     } else {
788       vp9_rd_pick_inter_mode_sub8x8(cpi, x, tile, mi_row, mi_col, totalrate,
789                                     totaldist, bsize, ctx, best_rd);
790     }
791   }
792 
793   x->rdmult = orig_rdmult;
794 
795   if (aq_mode == VARIANCE_AQ && *totalrate != INT_MAX) {
796     vp9_clear_system_state();
797     *totalrate = (int)round(*totalrate * rdmult_ratio);
798   }
799 }
800 
update_stats(VP9_COMMON * cm,const MACROBLOCK * x)801 static void update_stats(VP9_COMMON *cm, const MACROBLOCK *x) {
802   const MACROBLOCKD *const xd = &x->e_mbd;
803   const MODE_INFO *const mi = xd->mi[0].src_mi;
804   const MB_MODE_INFO *const mbmi = &mi->mbmi;
805 
806   if (!frame_is_intra_only(cm)) {
807     const int seg_ref_active = vp9_segfeature_active(&cm->seg, mbmi->segment_id,
808                                                      SEG_LVL_REF_FRAME);
809     if (!seg_ref_active) {
810       FRAME_COUNTS *const counts = &cm->counts;
811       const int inter_block = is_inter_block(mbmi);
812 
813       counts->intra_inter[vp9_get_intra_inter_context(xd)][inter_block]++;
814 
815       // If the segment reference feature is enabled we have only a single
816       // reference frame allowed for the segment so exclude it from
817       // the reference frame counts used to work out probabilities.
818       if (inter_block) {
819         const MV_REFERENCE_FRAME ref0 = mbmi->ref_frame[0];
820 
821         if (cm->reference_mode == REFERENCE_MODE_SELECT)
822           counts->comp_inter[vp9_get_reference_mode_context(cm, xd)]
823                             [has_second_ref(mbmi)]++;
824 
825         if (has_second_ref(mbmi)) {
826           counts->comp_ref[vp9_get_pred_context_comp_ref_p(cm, xd)]
827                           [ref0 == GOLDEN_FRAME]++;
828         } else {
829           counts->single_ref[vp9_get_pred_context_single_ref_p1(xd)][0]
830                             [ref0 != LAST_FRAME]++;
831           if (ref0 != LAST_FRAME)
832             counts->single_ref[vp9_get_pred_context_single_ref_p2(xd)][1]
833                               [ref0 != GOLDEN_FRAME]++;
834         }
835       }
836     }
837   }
838 }
839 
restore_context(VP9_COMP * cpi,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)840 static void restore_context(VP9_COMP *cpi, int mi_row, int mi_col,
841                             ENTROPY_CONTEXT a[16 * MAX_MB_PLANE],
842                             ENTROPY_CONTEXT l[16 * MAX_MB_PLANE],
843                             PARTITION_CONTEXT sa[8], PARTITION_CONTEXT sl[8],
844                             BLOCK_SIZE bsize) {
845   MACROBLOCK *const x = &cpi->mb;
846   MACROBLOCKD *const xd = &x->e_mbd;
847   int p;
848   const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
849   const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
850   int mi_width = num_8x8_blocks_wide_lookup[bsize];
851   int mi_height = num_8x8_blocks_high_lookup[bsize];
852   for (p = 0; p < MAX_MB_PLANE; p++) {
853     vpx_memcpy(
854         xd->above_context[p] + ((mi_col * 2) >> xd->plane[p].subsampling_x),
855         a + num_4x4_blocks_wide * p,
856         (sizeof(ENTROPY_CONTEXT) * num_4x4_blocks_wide) >>
857         xd->plane[p].subsampling_x);
858     vpx_memcpy(
859         xd->left_context[p]
860             + ((mi_row & MI_MASK) * 2 >> xd->plane[p].subsampling_y),
861         l + num_4x4_blocks_high * p,
862         (sizeof(ENTROPY_CONTEXT) * num_4x4_blocks_high) >>
863         xd->plane[p].subsampling_y);
864   }
865   vpx_memcpy(xd->above_seg_context + mi_col, sa,
866              sizeof(*xd->above_seg_context) * mi_width);
867   vpx_memcpy(xd->left_seg_context + (mi_row & MI_MASK), sl,
868              sizeof(xd->left_seg_context[0]) * mi_height);
869 }
870 
save_context(VP9_COMP * cpi,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)871 static void save_context(VP9_COMP *cpi, int mi_row, int mi_col,
872                          ENTROPY_CONTEXT a[16 * MAX_MB_PLANE],
873                          ENTROPY_CONTEXT l[16 * MAX_MB_PLANE],
874                          PARTITION_CONTEXT sa[8], PARTITION_CONTEXT sl[8],
875                          BLOCK_SIZE bsize) {
876   const MACROBLOCK *const x = &cpi->mb;
877   const MACROBLOCKD *const xd = &x->e_mbd;
878   int p;
879   const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
880   const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
881   int mi_width = num_8x8_blocks_wide_lookup[bsize];
882   int mi_height = num_8x8_blocks_high_lookup[bsize];
883 
884   // buffer the above/left context information of the block in search.
885   for (p = 0; p < MAX_MB_PLANE; ++p) {
886     vpx_memcpy(
887         a + num_4x4_blocks_wide * p,
888         xd->above_context[p] + (mi_col * 2 >> xd->plane[p].subsampling_x),
889         (sizeof(ENTROPY_CONTEXT) * num_4x4_blocks_wide) >>
890         xd->plane[p].subsampling_x);
891     vpx_memcpy(
892         l + num_4x4_blocks_high * p,
893         xd->left_context[p]
894             + ((mi_row & MI_MASK) * 2 >> xd->plane[p].subsampling_y),
895         (sizeof(ENTROPY_CONTEXT) * num_4x4_blocks_high) >>
896         xd->plane[p].subsampling_y);
897   }
898   vpx_memcpy(sa, xd->above_seg_context + mi_col,
899              sizeof(*xd->above_seg_context) * mi_width);
900   vpx_memcpy(sl, xd->left_seg_context + (mi_row & MI_MASK),
901              sizeof(xd->left_seg_context[0]) * mi_height);
902 }
903 
encode_b(VP9_COMP * cpi,const TileInfo * const tile,TOKENEXTRA ** tp,int mi_row,int mi_col,int output_enabled,BLOCK_SIZE bsize,PICK_MODE_CONTEXT * ctx)904 static void encode_b(VP9_COMP *cpi, const TileInfo *const tile,
905                      TOKENEXTRA **tp, int mi_row, int mi_col,
906                      int output_enabled, BLOCK_SIZE bsize,
907                      PICK_MODE_CONTEXT *ctx) {
908   set_offsets(cpi, tile, mi_row, mi_col, bsize);
909   update_state(cpi, ctx, mi_row, mi_col, bsize, output_enabled);
910   encode_superblock(cpi, tp, output_enabled, mi_row, mi_col, bsize, ctx);
911 
912   if (output_enabled) {
913     update_stats(&cpi->common, &cpi->mb);
914 
915     (*tp)->token = EOSB_TOKEN;
916     (*tp)++;
917   }
918 }
919 
encode_sb(VP9_COMP * cpi,const TileInfo * const tile,TOKENEXTRA ** tp,int mi_row,int mi_col,int output_enabled,BLOCK_SIZE bsize,PC_TREE * pc_tree)920 static void encode_sb(VP9_COMP *cpi, const TileInfo *const tile,
921                       TOKENEXTRA **tp, int mi_row, int mi_col,
922                       int output_enabled, BLOCK_SIZE bsize,
923                       PC_TREE *pc_tree) {
924   VP9_COMMON *const cm = &cpi->common;
925   MACROBLOCK *const x = &cpi->mb;
926   MACROBLOCKD *const xd = &x->e_mbd;
927 
928   const int bsl = b_width_log2(bsize), hbs = (1 << bsl) / 4;
929   int ctx;
930   PARTITION_TYPE partition;
931   BLOCK_SIZE subsize = bsize;
932 
933   if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols)
934     return;
935 
936   if (bsize >= BLOCK_8X8) {
937     ctx = partition_plane_context(xd, mi_row, mi_col, bsize);
938     subsize = get_subsize(bsize, pc_tree->partitioning);
939   } else {
940     ctx = 0;
941     subsize = BLOCK_4X4;
942   }
943 
944   partition = partition_lookup[bsl][subsize];
945   if (output_enabled && bsize != BLOCK_4X4)
946     cm->counts.partition[ctx][partition]++;
947 
948   switch (partition) {
949     case PARTITION_NONE:
950       encode_b(cpi, tile, tp, mi_row, mi_col, output_enabled, subsize,
951                &pc_tree->none);
952       break;
953     case PARTITION_VERT:
954       encode_b(cpi, tile, tp, mi_row, mi_col, output_enabled, subsize,
955                &pc_tree->vertical[0]);
956       if (mi_col + hbs < cm->mi_cols && bsize > BLOCK_8X8) {
957         encode_b(cpi, tile, tp, mi_row, mi_col + hbs, output_enabled, subsize,
958                  &pc_tree->vertical[1]);
959       }
960       break;
961     case PARTITION_HORZ:
962       encode_b(cpi, tile, tp, mi_row, mi_col, output_enabled, subsize,
963                &pc_tree->horizontal[0]);
964       if (mi_row + hbs < cm->mi_rows && bsize > BLOCK_8X8) {
965         encode_b(cpi, tile, tp, mi_row + hbs, mi_col, output_enabled, subsize,
966                  &pc_tree->horizontal[1]);
967       }
968       break;
969     case PARTITION_SPLIT:
970       if (bsize == BLOCK_8X8) {
971         encode_b(cpi, tile, tp, mi_row, mi_col, output_enabled, subsize,
972                  pc_tree->leaf_split[0]);
973       } else {
974         encode_sb(cpi, tile, tp, mi_row, mi_col, output_enabled, subsize,
975                   pc_tree->split[0]);
976         encode_sb(cpi, tile, tp, mi_row, mi_col + hbs, output_enabled, subsize,
977                   pc_tree->split[1]);
978         encode_sb(cpi, tile, tp, mi_row + hbs, mi_col, output_enabled, subsize,
979                   pc_tree->split[2]);
980         encode_sb(cpi, tile, tp, mi_row + hbs, mi_col + hbs, output_enabled,
981                   subsize, pc_tree->split[3]);
982       }
983       break;
984     default:
985       assert("Invalid partition type.");
986       break;
987   }
988 
989   if (partition != PARTITION_SPLIT || bsize == BLOCK_8X8)
990     update_partition_context(xd, mi_row, mi_col, subsize, bsize);
991 }
992 
993 // Check to see if the given partition size is allowed for a specified number
994 // of 8x8 block rows and columns remaining in the image.
995 // 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)996 static BLOCK_SIZE find_partition_size(BLOCK_SIZE bsize,
997                                       int rows_left, int cols_left,
998                                       int *bh, int *bw) {
999   if (rows_left <= 0 || cols_left <= 0) {
1000     return MIN(bsize, BLOCK_8X8);
1001   } else {
1002     for (; bsize > 0; bsize -= 3) {
1003       *bh = num_8x8_blocks_high_lookup[bsize];
1004       *bw = num_8x8_blocks_wide_lookup[bsize];
1005       if ((*bh <= rows_left) && (*bw <= cols_left)) {
1006         break;
1007       }
1008     }
1009   }
1010   return bsize;
1011 }
1012 
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)1013 static void set_partial_b64x64_partition(MODE_INFO *mi, int mis,
1014     int bh_in, int bw_in, int row8x8_remaining, int col8x8_remaining,
1015     BLOCK_SIZE bsize, MODE_INFO *mi_8x8) {
1016   int bh = bh_in;
1017   int r, c;
1018   for (r = 0; r < MI_BLOCK_SIZE; r += bh) {
1019     int bw = bw_in;
1020     for (c = 0; c < MI_BLOCK_SIZE; c += bw) {
1021       const int index = r * mis + c;
1022       mi_8x8[index].src_mi = mi + index;
1023       mi_8x8[index].src_mi->mbmi.sb_type = find_partition_size(bsize,
1024           row8x8_remaining - r, col8x8_remaining - c, &bh, &bw);
1025     }
1026   }
1027 }
1028 
1029 // This function attempts to set all mode info entries in a given SB64
1030 // to the same block partition size.
1031 // However, at the bottom and right borders of the image the requested size
1032 // may not be allowed in which case this code attempts to choose the largest
1033 // 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)1034 static void set_fixed_partitioning(VP9_COMP *cpi, const TileInfo *const tile,
1035                                    MODE_INFO *mi_8x8, int mi_row, int mi_col,
1036                                    BLOCK_SIZE bsize) {
1037   VP9_COMMON *const cm = &cpi->common;
1038   const int mis = cm->mi_stride;
1039   const int row8x8_remaining = tile->mi_row_end - mi_row;
1040   const int col8x8_remaining = tile->mi_col_end - mi_col;
1041   int block_row, block_col;
1042   MODE_INFO *mi_upper_left = cm->mi + mi_row * mis + mi_col;
1043   int bh = num_8x8_blocks_high_lookup[bsize];
1044   int bw = num_8x8_blocks_wide_lookup[bsize];
1045 
1046   assert((row8x8_remaining > 0) && (col8x8_remaining > 0));
1047 
1048   // Apply the requested partition size to the SB64 if it is all "in image"
1049   if ((col8x8_remaining >= MI_BLOCK_SIZE) &&
1050       (row8x8_remaining >= MI_BLOCK_SIZE)) {
1051     for (block_row = 0; block_row < MI_BLOCK_SIZE; block_row += bh) {
1052       for (block_col = 0; block_col < MI_BLOCK_SIZE; block_col += bw) {
1053         int index = block_row * mis + block_col;
1054         mi_8x8[index].src_mi = mi_upper_left + index;
1055         mi_8x8[index].src_mi->mbmi.sb_type = bsize;
1056       }
1057     }
1058   } else {
1059     // Else this is a partial SB64.
1060     set_partial_b64x64_partition(mi_upper_left, mis, bh, bw, row8x8_remaining,
1061         col8x8_remaining, bsize, mi_8x8);
1062   }
1063 }
1064 
copy_partitioning(VP9_COMMON * cm,MODE_INFO * mi_8x8,MODE_INFO * prev_mi_8x8)1065 static void copy_partitioning(VP9_COMMON *cm, MODE_INFO *mi_8x8,
1066   MODE_INFO *prev_mi_8x8) {
1067   const int mis = cm->mi_stride;
1068   int block_row, block_col;
1069 
1070   for (block_row = 0; block_row < 8; ++block_row) {
1071     for (block_col = 0; block_col < 8; ++block_col) {
1072       MODE_INFO *const prev_mi =
1073           prev_mi_8x8[block_row * mis + block_col].src_mi;
1074       const BLOCK_SIZE sb_type = prev_mi ? prev_mi->mbmi.sb_type : 0;
1075 
1076       if (prev_mi) {
1077         const ptrdiff_t offset = prev_mi - cm->prev_mi;
1078         mi_8x8[block_row * mis + block_col].src_mi = cm->mi + offset;
1079         mi_8x8[block_row * mis + block_col].src_mi->mbmi.sb_type = sb_type;
1080       }
1081     }
1082   }
1083 }
1084 
constrain_copy_partitioning(VP9_COMP * const cpi,const TileInfo * const tile,MODE_INFO * mi_8x8,MODE_INFO * prev_mi_8x8,int mi_row,int mi_col,BLOCK_SIZE bsize)1085 static void constrain_copy_partitioning(VP9_COMP *const cpi,
1086                                         const TileInfo *const tile,
1087                                         MODE_INFO *mi_8x8,
1088                                         MODE_INFO *prev_mi_8x8,
1089                                         int mi_row, int mi_col,
1090                                         BLOCK_SIZE bsize) {
1091   VP9_COMMON *const cm = &cpi->common;
1092   const int mis = cm->mi_stride;
1093   const int row8x8_remaining = tile->mi_row_end - mi_row;
1094   const int col8x8_remaining = tile->mi_col_end - mi_col;
1095   MODE_INFO *const mi_upper_left = cm->mi + mi_row * mis + mi_col;
1096   const int bh = num_8x8_blocks_high_lookup[bsize];
1097   const int bw = num_8x8_blocks_wide_lookup[bsize];
1098   int block_row, block_col;
1099 
1100   assert((row8x8_remaining > 0) && (col8x8_remaining > 0));
1101 
1102   // If the SB64 if it is all "in image".
1103   if ((col8x8_remaining >= MI_BLOCK_SIZE) &&
1104       (row8x8_remaining >= MI_BLOCK_SIZE)) {
1105     for (block_row = 0; block_row < MI_BLOCK_SIZE; block_row += bh) {
1106       for (block_col = 0; block_col < MI_BLOCK_SIZE; block_col += bw) {
1107         const int index = block_row * mis + block_col;
1108         MODE_INFO *prev_mi = prev_mi_8x8[index].src_mi;
1109         const BLOCK_SIZE sb_type = prev_mi ? prev_mi->mbmi.sb_type : 0;
1110         // Use previous partition if block size is not larger than bsize.
1111         if (prev_mi && sb_type <= bsize) {
1112           int block_row2, block_col2;
1113           for (block_row2 = 0; block_row2 < bh; ++block_row2) {
1114             for (block_col2 = 0; block_col2 < bw; ++block_col2) {
1115               const int index2 = (block_row + block_row2) * mis +
1116                   block_col + block_col2;
1117               prev_mi = prev_mi_8x8[index2].src_mi;
1118               if (prev_mi) {
1119                 const ptrdiff_t offset = prev_mi - cm->prev_mi;
1120                 mi_8x8[index2].src_mi = cm->mi + offset;
1121                 mi_8x8[index2].src_mi->mbmi.sb_type = prev_mi->mbmi.sb_type;
1122               }
1123             }
1124           }
1125         } else {
1126           // Otherwise, use fixed partition of size bsize.
1127           mi_8x8[index].src_mi = mi_upper_left + index;
1128           mi_8x8[index].src_mi->mbmi.sb_type = bsize;
1129         }
1130       }
1131     }
1132   } else {
1133     // Else this is a partial SB64, copy previous partition.
1134     copy_partitioning(cm, mi_8x8, prev_mi_8x8);
1135   }
1136 }
1137 
1138 const struct {
1139   int row;
1140   int col;
1141 } coord_lookup[16] = {
1142     // 32x32 index = 0
1143     {0, 0}, {0, 2}, {2, 0}, {2, 2},
1144     // 32x32 index = 1
1145     {0, 4}, {0, 6}, {2, 4}, {2, 6},
1146     // 32x32 index = 2
1147     {4, 0}, {4, 2}, {6, 0}, {6, 2},
1148     // 32x32 index = 3
1149     {4, 4}, {4, 6}, {6, 4}, {6, 6},
1150 };
1151 
set_source_var_based_partition(VP9_COMP * cpi,const TileInfo * const tile,MODE_INFO * mi_8x8,int mi_row,int mi_col)1152 static void set_source_var_based_partition(VP9_COMP *cpi,
1153                                            const TileInfo *const tile,
1154                                            MODE_INFO *mi_8x8,
1155                                            int mi_row, int mi_col) {
1156   VP9_COMMON *const cm = &cpi->common;
1157   MACROBLOCK *const x = &cpi->mb;
1158   const int mis = cm->mi_stride;
1159   const int row8x8_remaining = tile->mi_row_end - mi_row;
1160   const int col8x8_remaining = tile->mi_col_end - mi_col;
1161   MODE_INFO *mi_upper_left = cm->mi + mi_row * mis + mi_col;
1162 
1163   vp9_setup_src_planes(x, cpi->Source, mi_row, mi_col);
1164 
1165   assert((row8x8_remaining > 0) && (col8x8_remaining > 0));
1166 
1167   // In-image SB64
1168   if ((col8x8_remaining >= MI_BLOCK_SIZE) &&
1169       (row8x8_remaining >= MI_BLOCK_SIZE)) {
1170     int i, j;
1171     int index;
1172     diff d32[4];
1173     const int offset = (mi_row >> 1) * cm->mb_cols + (mi_col >> 1);
1174     int is_larger_better = 0;
1175     int use32x32 = 0;
1176     unsigned int thr = cpi->source_var_thresh;
1177 
1178     vpx_memset(d32, 0, 4 * sizeof(diff));
1179 
1180     for (i = 0; i < 4; i++) {
1181       diff *d16[4];
1182 
1183       for (j = 0; j < 4; j++) {
1184         int b_mi_row = coord_lookup[i * 4 + j].row;
1185         int b_mi_col = coord_lookup[i * 4 + j].col;
1186         int boffset = b_mi_row / 2 * cm->mb_cols +
1187                       b_mi_col / 2;
1188 
1189         d16[j] = cpi->source_diff_var + offset + boffset;
1190 
1191         index = b_mi_row * mis + b_mi_col;
1192         mi_8x8[index].src_mi = mi_upper_left + index;
1193         mi_8x8[index].src_mi->mbmi.sb_type = BLOCK_16X16;
1194 
1195         // TODO(yunqingwang): If d16[j].var is very large, use 8x8 partition
1196         // size to further improve quality.
1197       }
1198 
1199       is_larger_better = (d16[0]->var < thr) && (d16[1]->var < thr) &&
1200           (d16[2]->var < thr) && (d16[3]->var < thr);
1201 
1202       // Use 32x32 partition
1203       if (is_larger_better) {
1204         use32x32 += 1;
1205 
1206         for (j = 0; j < 4; j++) {
1207           d32[i].sse += d16[j]->sse;
1208           d32[i].sum += d16[j]->sum;
1209         }
1210 
1211         d32[i].var = d32[i].sse - (((int64_t)d32[i].sum * d32[i].sum) >> 10);
1212 
1213         index = coord_lookup[i*4].row * mis + coord_lookup[i*4].col;
1214         mi_8x8[index].src_mi = mi_upper_left + index;
1215         mi_8x8[index].src_mi->mbmi.sb_type = BLOCK_32X32;
1216       }
1217     }
1218 
1219     if (use32x32 == 4) {
1220       thr <<= 1;
1221       is_larger_better = (d32[0].var < thr) && (d32[1].var < thr) &&
1222           (d32[2].var < thr) && (d32[3].var < thr);
1223 
1224       // Use 64x64 partition
1225       if (is_larger_better) {
1226         mi_8x8[0].src_mi = mi_upper_left;
1227         mi_8x8[0].src_mi->mbmi.sb_type = BLOCK_64X64;
1228       }
1229     }
1230   } else {   // partial in-image SB64
1231     int bh = num_8x8_blocks_high_lookup[BLOCK_16X16];
1232     int bw = num_8x8_blocks_wide_lookup[BLOCK_16X16];
1233     set_partial_b64x64_partition(mi_upper_left, mis, bh, bw,
1234         row8x8_remaining, col8x8_remaining, BLOCK_16X16, mi_8x8);
1235   }
1236 }
1237 
is_background(const VP9_COMP * cpi,const TileInfo * const tile,int mi_row,int mi_col)1238 static int is_background(const VP9_COMP *cpi, const TileInfo *const tile,
1239                          int mi_row, int mi_col) {
1240   // This assumes the input source frames are of the same dimension.
1241   const int row8x8_remaining = tile->mi_row_end - mi_row;
1242   const int col8x8_remaining = tile->mi_col_end - mi_col;
1243   const int x = mi_col * MI_SIZE;
1244   const int y = mi_row * MI_SIZE;
1245   const int src_stride = cpi->Source->y_stride;
1246   const uint8_t *const src = &cpi->Source->y_buffer[y * src_stride + x];
1247   const int pre_stride = cpi->Last_Source->y_stride;
1248   const uint8_t *const pre = &cpi->Last_Source->y_buffer[y * pre_stride + x];
1249   int this_sad = 0;
1250   int threshold = 0;
1251 
1252   if (row8x8_remaining >= MI_BLOCK_SIZE &&
1253       col8x8_remaining >= MI_BLOCK_SIZE) {
1254     this_sad = cpi->fn_ptr[BLOCK_64X64].sdf(src, src_stride, pre, pre_stride);
1255     threshold = (1 << 12);
1256   } else {
1257     int r, c;
1258     for (r = 0; r < row8x8_remaining; r += 2)
1259       for (c = 0; c < col8x8_remaining; c += 2)
1260         this_sad += cpi->fn_ptr[BLOCK_16X16].sdf(src, src_stride,
1261                                                  pre, pre_stride);
1262     threshold = (row8x8_remaining * col8x8_remaining) << 6;
1263   }
1264 
1265   return this_sad < 2 * threshold;
1266 }
1267 
sb_has_motion(const VP9_COMMON * cm,MODE_INFO * prev_mi_8x8,const int motion_thresh)1268 static int sb_has_motion(const VP9_COMMON *cm, MODE_INFO *prev_mi_8x8,
1269                          const int motion_thresh) {
1270   const int mis = cm->mi_stride;
1271   int block_row, block_col;
1272 
1273   if (cm->prev_mi) {
1274     for (block_row = 0; block_row < 8; ++block_row) {
1275       for (block_col = 0; block_col < 8; ++block_col) {
1276         const MODE_INFO *prev_mi =
1277             prev_mi_8x8[block_row * mis + block_col].src_mi;
1278         if (prev_mi) {
1279           if (abs(prev_mi->mbmi.mv[0].as_mv.row) > motion_thresh ||
1280               abs(prev_mi->mbmi.mv[0].as_mv.col) > motion_thresh)
1281             return 1;
1282         }
1283       }
1284     }
1285   }
1286   return 0;
1287 }
1288 
update_state_rt(VP9_COMP * cpi,PICK_MODE_CONTEXT * ctx,int mi_row,int mi_col,int bsize)1289 static void update_state_rt(VP9_COMP *cpi, PICK_MODE_CONTEXT *ctx,
1290                             int mi_row, int mi_col, int bsize) {
1291   VP9_COMMON *const cm = &cpi->common;
1292   MACROBLOCK *const x = &cpi->mb;
1293   MACROBLOCKD *const xd = &x->e_mbd;
1294   MB_MODE_INFO *const mbmi = &xd->mi[0].src_mi->mbmi;
1295   const struct segmentation *const seg = &cm->seg;
1296 
1297   *(xd->mi[0].src_mi) = ctx->mic;
1298   xd->mi[0].src_mi = &xd->mi[0];
1299 
1300 
1301   // For in frame adaptive Q, check for reseting the segment_id and updating
1302   // the cyclic refresh map.
1303   if ((cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ) && seg->enabled) {
1304     vp9_cyclic_refresh_update_segment(cpi, &xd->mi[0].src_mi->mbmi,
1305                                       mi_row, mi_col, bsize, 1);
1306     vp9_init_plane_quantizers(cpi, x);
1307   }
1308 
1309   if (is_inter_block(mbmi)) {
1310     vp9_update_mv_count(cm, xd);
1311 
1312     if (cm->interp_filter == SWITCHABLE) {
1313       const int pred_ctx = vp9_get_pred_context_switchable_interp(xd);
1314       ++cm->counts.switchable_interp[pred_ctx][mbmi->interp_filter];
1315     }
1316   }
1317 
1318   x->skip = ctx->skip;
1319   x->skip_txfm[0] = mbmi->segment_id ? 0 : ctx->skip_txfm[0];
1320 }
1321 
encode_b_rt(VP9_COMP * cpi,const TileInfo * const tile,TOKENEXTRA ** tp,int mi_row,int mi_col,int output_enabled,BLOCK_SIZE bsize,PICK_MODE_CONTEXT * ctx)1322 static void encode_b_rt(VP9_COMP *cpi, const TileInfo *const tile,
1323                         TOKENEXTRA **tp, int mi_row, int mi_col,
1324                      int output_enabled, BLOCK_SIZE bsize,
1325                      PICK_MODE_CONTEXT *ctx) {
1326   set_offsets(cpi, tile, mi_row, mi_col, bsize);
1327   update_state_rt(cpi, ctx, mi_row, mi_col, bsize);
1328 
1329 #if CONFIG_VP9_TEMPORAL_DENOISING
1330   if (cpi->oxcf.noise_sensitivity > 0 && output_enabled) {
1331     vp9_denoiser_denoise(&cpi->denoiser, &cpi->mb, mi_row, mi_col,
1332                          MAX(BLOCK_8X8, bsize), ctx);
1333   }
1334 #endif
1335 
1336   encode_superblock(cpi, tp, output_enabled, mi_row, mi_col, bsize, ctx);
1337   update_stats(&cpi->common, &cpi->mb);
1338 
1339   (*tp)->token = EOSB_TOKEN;
1340   (*tp)++;
1341 }
1342 
encode_sb_rt(VP9_COMP * cpi,const TileInfo * const tile,TOKENEXTRA ** tp,int mi_row,int mi_col,int output_enabled,BLOCK_SIZE bsize,PC_TREE * pc_tree)1343 static void encode_sb_rt(VP9_COMP *cpi, const TileInfo *const tile,
1344                          TOKENEXTRA **tp, int mi_row, int mi_col,
1345                          int output_enabled, BLOCK_SIZE bsize,
1346                          PC_TREE *pc_tree) {
1347   VP9_COMMON *const cm = &cpi->common;
1348   MACROBLOCK *const x = &cpi->mb;
1349   MACROBLOCKD *const xd = &x->e_mbd;
1350 
1351   const int bsl = b_width_log2(bsize), hbs = (1 << bsl) / 4;
1352   int ctx;
1353   PARTITION_TYPE partition;
1354   BLOCK_SIZE subsize;
1355 
1356   if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols)
1357     return;
1358 
1359   if (bsize >= BLOCK_8X8) {
1360     const int idx_str = xd->mi_stride * mi_row + mi_col;
1361     MODE_INFO *mi_8x8 = cm->mi[idx_str].src_mi;
1362     ctx = partition_plane_context(xd, mi_row, mi_col, bsize);
1363     subsize = mi_8x8[0].src_mi->mbmi.sb_type;
1364   } else {
1365     ctx = 0;
1366     subsize = BLOCK_4X4;
1367   }
1368 
1369   partition = partition_lookup[bsl][subsize];
1370   if (output_enabled && bsize != BLOCK_4X4)
1371     cm->counts.partition[ctx][partition]++;
1372 
1373   switch (partition) {
1374     case PARTITION_NONE:
1375       encode_b_rt(cpi, tile, tp, mi_row, mi_col, output_enabled, subsize,
1376                   &pc_tree->none);
1377       break;
1378     case PARTITION_VERT:
1379       encode_b_rt(cpi, tile, tp, mi_row, mi_col, output_enabled, subsize,
1380                   &pc_tree->vertical[0]);
1381       if (mi_col + hbs < cm->mi_cols && bsize > BLOCK_8X8) {
1382         encode_b_rt(cpi, tile, tp, mi_row, mi_col + hbs, output_enabled,
1383                     subsize, &pc_tree->vertical[1]);
1384       }
1385       break;
1386     case PARTITION_HORZ:
1387       encode_b_rt(cpi, tile, tp, mi_row, mi_col, output_enabled, subsize,
1388                   &pc_tree->horizontal[0]);
1389       if (mi_row + hbs < cm->mi_rows && bsize > BLOCK_8X8) {
1390         encode_b_rt(cpi, tile, tp, mi_row + hbs, mi_col, output_enabled,
1391                     subsize, &pc_tree->horizontal[1]);
1392       }
1393       break;
1394     case PARTITION_SPLIT:
1395       subsize = get_subsize(bsize, PARTITION_SPLIT);
1396       encode_sb_rt(cpi, tile, tp, mi_row, mi_col, output_enabled, subsize,
1397                    pc_tree->split[0]);
1398       encode_sb_rt(cpi, tile, tp, mi_row, mi_col + hbs, output_enabled,
1399                    subsize, pc_tree->split[1]);
1400       encode_sb_rt(cpi, tile, tp, mi_row + hbs, mi_col, output_enabled,
1401                    subsize, pc_tree->split[2]);
1402       encode_sb_rt(cpi, tile, tp, mi_row + hbs, mi_col + hbs, output_enabled,
1403                    subsize, pc_tree->split[3]);
1404       break;
1405     default:
1406       assert("Invalid partition type.");
1407       break;
1408   }
1409 
1410   if (partition != PARTITION_SPLIT || bsize == BLOCK_8X8)
1411     update_partition_context(xd, mi_row, mi_col, subsize, bsize);
1412 }
1413 
rd_use_partition(VP9_COMP * cpi,const TileInfo * const tile,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)1414 static void rd_use_partition(VP9_COMP *cpi,
1415                              const TileInfo *const tile,
1416                              MODE_INFO *mi_8x8,
1417                              TOKENEXTRA **tp, int mi_row, int mi_col,
1418                              BLOCK_SIZE bsize, int *rate, int64_t *dist,
1419                              int do_recon, PC_TREE *pc_tree) {
1420   VP9_COMMON *const cm = &cpi->common;
1421   MACROBLOCK *const x = &cpi->mb;
1422   MACROBLOCKD *const xd = &x->e_mbd;
1423   const int mis = cm->mi_stride;
1424   const int bsl = b_width_log2(bsize);
1425   const int mi_step = num_4x4_blocks_wide_lookup[bsize] / 2;
1426   const int bss = (1 << bsl) / 4;
1427   int i, pl;
1428   PARTITION_TYPE partition = PARTITION_NONE;
1429   BLOCK_SIZE subsize;
1430   ENTROPY_CONTEXT l[16 * MAX_MB_PLANE], a[16 * MAX_MB_PLANE];
1431   PARTITION_CONTEXT sl[8], sa[8];
1432   int last_part_rate = INT_MAX;
1433   int64_t last_part_dist = INT64_MAX;
1434   int64_t last_part_rd = INT64_MAX;
1435   int none_rate = INT_MAX;
1436   int64_t none_dist = INT64_MAX;
1437   int64_t none_rd = INT64_MAX;
1438   int chosen_rate = INT_MAX;
1439   int64_t chosen_dist = INT64_MAX;
1440   int64_t chosen_rd = INT64_MAX;
1441   BLOCK_SIZE sub_subsize = BLOCK_4X4;
1442   int splits_below = 0;
1443   BLOCK_SIZE bs_type = mi_8x8[0].src_mi->mbmi.sb_type;
1444   int do_partition_search = 1;
1445   PICK_MODE_CONTEXT *ctx = &pc_tree->none;
1446 
1447   if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols)
1448     return;
1449 
1450   assert(num_4x4_blocks_wide_lookup[bsize] ==
1451          num_4x4_blocks_high_lookup[bsize]);
1452 
1453   partition = partition_lookup[bsl][bs_type];
1454   subsize = get_subsize(bsize, partition);
1455 
1456   pc_tree->partitioning = partition;
1457   save_context(cpi, mi_row, mi_col, a, l, sa, sl, bsize);
1458 
1459   if (bsize == BLOCK_16X16 && cpi->oxcf.aq_mode) {
1460     set_offsets(cpi, tile, mi_row, mi_col, bsize);
1461     x->mb_energy = vp9_block_energy(cpi, x, bsize);
1462   }
1463 
1464   if (do_partition_search &&
1465       cpi->sf.partition_search_type == SEARCH_PARTITION &&
1466       cpi->sf.adjust_partitioning_from_last_frame) {
1467     // Check if any of the sub blocks are further split.
1468     if (partition == PARTITION_SPLIT && subsize > BLOCK_8X8) {
1469       sub_subsize = get_subsize(subsize, PARTITION_SPLIT);
1470       splits_below = 1;
1471       for (i = 0; i < 4; i++) {
1472         int jj = i >> 1, ii = i & 0x01;
1473         MODE_INFO *this_mi = mi_8x8[jj * bss * mis + ii * bss].src_mi;
1474         if (this_mi && this_mi->mbmi.sb_type >= sub_subsize) {
1475           splits_below = 0;
1476         }
1477       }
1478     }
1479 
1480     // If partition is not none try none unless each of the 4 splits are split
1481     // even further..
1482     if (partition != PARTITION_NONE && !splits_below &&
1483         mi_row + (mi_step >> 1) < cm->mi_rows &&
1484         mi_col + (mi_step >> 1) < cm->mi_cols) {
1485       pc_tree->partitioning = PARTITION_NONE;
1486       rd_pick_sb_modes(cpi, tile, mi_row, mi_col, &none_rate, &none_dist, bsize,
1487                        ctx, INT64_MAX, 0);
1488 
1489       pl = partition_plane_context(xd, mi_row, mi_col, bsize);
1490 
1491       if (none_rate < INT_MAX) {
1492         none_rate += cpi->partition_cost[pl][PARTITION_NONE];
1493         none_rd = RDCOST(x->rdmult, x->rddiv, none_rate, none_dist);
1494       }
1495 
1496       restore_context(cpi, mi_row, mi_col, a, l, sa, sl, bsize);
1497       mi_8x8[0].src_mi->mbmi.sb_type = bs_type;
1498       pc_tree->partitioning = partition;
1499     }
1500   }
1501 
1502   switch (partition) {
1503     case PARTITION_NONE:
1504       rd_pick_sb_modes(cpi, tile, mi_row, mi_col, &last_part_rate,
1505                        &last_part_dist, bsize, ctx, INT64_MAX, 0);
1506       break;
1507     case PARTITION_HORZ:
1508       rd_pick_sb_modes(cpi, tile, mi_row, mi_col, &last_part_rate,
1509                        &last_part_dist, subsize, &pc_tree->horizontal[0],
1510                        INT64_MAX, 0);
1511       if (last_part_rate != INT_MAX &&
1512           bsize >= BLOCK_8X8 && mi_row + (mi_step >> 1) < cm->mi_rows) {
1513         int rt = 0;
1514         int64_t dt = 0;
1515         PICK_MODE_CONTEXT *ctx = &pc_tree->horizontal[0];
1516         update_state(cpi, ctx, mi_row, mi_col, subsize, 0);
1517         encode_superblock(cpi, tp, 0, mi_row, mi_col, subsize, ctx);
1518         rd_pick_sb_modes(cpi, tile, mi_row + (mi_step >> 1), mi_col, &rt, &dt,
1519                          subsize, &pc_tree->horizontal[1], INT64_MAX, 1);
1520         if (rt == INT_MAX || dt == INT64_MAX) {
1521           last_part_rate = INT_MAX;
1522           last_part_dist = INT64_MAX;
1523           break;
1524         }
1525 
1526         last_part_rate += rt;
1527         last_part_dist += dt;
1528       }
1529       break;
1530     case PARTITION_VERT:
1531       rd_pick_sb_modes(cpi, tile, mi_row, mi_col, &last_part_rate,
1532                        &last_part_dist, subsize, &pc_tree->vertical[0],
1533                        INT64_MAX, 0);
1534       if (last_part_rate != INT_MAX &&
1535           bsize >= BLOCK_8X8 && mi_col + (mi_step >> 1) < cm->mi_cols) {
1536         int rt = 0;
1537         int64_t dt = 0;
1538         PICK_MODE_CONTEXT *ctx = &pc_tree->vertical[0];
1539         update_state(cpi, ctx, mi_row, mi_col, subsize, 0);
1540         encode_superblock(cpi, tp, 0, mi_row, mi_col, subsize, ctx);
1541         rd_pick_sb_modes(cpi, tile, mi_row, mi_col + (mi_step >> 1), &rt, &dt,
1542                          subsize, &pc_tree->vertical[bsize > BLOCK_8X8],
1543                          INT64_MAX, 1);
1544         if (rt == INT_MAX || dt == INT64_MAX) {
1545           last_part_rate = INT_MAX;
1546           last_part_dist = INT64_MAX;
1547           break;
1548         }
1549         last_part_rate += rt;
1550         last_part_dist += dt;
1551       }
1552       break;
1553     case PARTITION_SPLIT:
1554       if (bsize == BLOCK_8X8) {
1555         rd_pick_sb_modes(cpi, tile, mi_row, mi_col, &last_part_rate,
1556                          &last_part_dist, subsize, pc_tree->leaf_split[0],
1557                          INT64_MAX, 0);
1558         break;
1559       }
1560       last_part_rate = 0;
1561       last_part_dist = 0;
1562       for (i = 0; i < 4; i++) {
1563         int x_idx = (i & 1) * (mi_step >> 1);
1564         int y_idx = (i >> 1) * (mi_step >> 1);
1565         int jj = i >> 1, ii = i & 0x01;
1566         int rt;
1567         int64_t dt;
1568 
1569         if ((mi_row + y_idx >= cm->mi_rows) || (mi_col + x_idx >= cm->mi_cols))
1570           continue;
1571 
1572         rd_use_partition(cpi, tile, mi_8x8 + jj * bss * mis + ii * bss, tp,
1573                          mi_row + y_idx, mi_col + x_idx, subsize, &rt, &dt,
1574                          i != 3, pc_tree->split[i]);
1575         if (rt == INT_MAX || dt == INT64_MAX) {
1576           last_part_rate = INT_MAX;
1577           last_part_dist = INT64_MAX;
1578           break;
1579         }
1580         last_part_rate += rt;
1581         last_part_dist += dt;
1582       }
1583       break;
1584     default:
1585       assert(0);
1586       break;
1587   }
1588 
1589   pl = partition_plane_context(xd, mi_row, mi_col, bsize);
1590   if (last_part_rate < INT_MAX) {
1591     last_part_rate += cpi->partition_cost[pl][partition];
1592     last_part_rd = RDCOST(x->rdmult, x->rddiv, last_part_rate, last_part_dist);
1593   }
1594 
1595   if (do_partition_search
1596       && cpi->sf.adjust_partitioning_from_last_frame
1597       && cpi->sf.partition_search_type == SEARCH_PARTITION
1598       && partition != PARTITION_SPLIT && bsize > BLOCK_8X8
1599       && (mi_row + mi_step < cm->mi_rows ||
1600           mi_row + (mi_step >> 1) == cm->mi_rows)
1601       && (mi_col + mi_step < cm->mi_cols ||
1602           mi_col + (mi_step >> 1) == cm->mi_cols)) {
1603     BLOCK_SIZE split_subsize = get_subsize(bsize, PARTITION_SPLIT);
1604     chosen_rate = 0;
1605     chosen_dist = 0;
1606     restore_context(cpi, mi_row, mi_col, a, l, sa, sl, bsize);
1607     pc_tree->partitioning = PARTITION_SPLIT;
1608 
1609     // Split partition.
1610     for (i = 0; i < 4; i++) {
1611       int x_idx = (i & 1) * (mi_step >> 1);
1612       int y_idx = (i >> 1) * (mi_step >> 1);
1613       int rt = 0;
1614       int64_t dt = 0;
1615       ENTROPY_CONTEXT l[16 * MAX_MB_PLANE], a[16 * MAX_MB_PLANE];
1616       PARTITION_CONTEXT sl[8], sa[8];
1617 
1618       if ((mi_row + y_idx >= cm->mi_rows) || (mi_col + x_idx >= cm->mi_cols))
1619         continue;
1620 
1621       save_context(cpi, mi_row, mi_col, a, l, sa, sl, bsize);
1622       pc_tree->split[i]->partitioning = PARTITION_NONE;
1623       rd_pick_sb_modes(cpi, tile, mi_row + y_idx, mi_col + x_idx, &rt, &dt,
1624                        split_subsize, &pc_tree->split[i]->none,
1625                        INT64_MAX, i);
1626 
1627       restore_context(cpi, mi_row, mi_col, a, l, sa, sl, bsize);
1628 
1629       if (rt == INT_MAX || dt == INT64_MAX) {
1630         chosen_rate = INT_MAX;
1631         chosen_dist = INT64_MAX;
1632         break;
1633       }
1634 
1635       chosen_rate += rt;
1636       chosen_dist += dt;
1637 
1638       if (i != 3)
1639         encode_sb(cpi, tile, tp,  mi_row + y_idx, mi_col + x_idx, 0,
1640                   split_subsize, pc_tree->split[i]);
1641 
1642       pl = partition_plane_context(xd, mi_row + y_idx, mi_col + x_idx,
1643                                    split_subsize);
1644       chosen_rate += cpi->partition_cost[pl][PARTITION_NONE];
1645     }
1646     pl = partition_plane_context(xd, mi_row, mi_col, bsize);
1647     if (chosen_rate < INT_MAX) {
1648       chosen_rate += cpi->partition_cost[pl][PARTITION_SPLIT];
1649       chosen_rd = RDCOST(x->rdmult, x->rddiv, chosen_rate, chosen_dist);
1650     }
1651   }
1652 
1653   // If last_part is better set the partitioning to that.
1654   if (last_part_rd < chosen_rd) {
1655     mi_8x8[0].src_mi->mbmi.sb_type = bsize;
1656     if (bsize >= BLOCK_8X8)
1657       pc_tree->partitioning = partition;
1658     chosen_rate = last_part_rate;
1659     chosen_dist = last_part_dist;
1660     chosen_rd = last_part_rd;
1661   }
1662   // If none was better set the partitioning to that.
1663   if (none_rd < chosen_rd) {
1664     if (bsize >= BLOCK_8X8)
1665       pc_tree->partitioning = PARTITION_NONE;
1666     chosen_rate = none_rate;
1667     chosen_dist = none_dist;
1668   }
1669 
1670   restore_context(cpi, mi_row, mi_col, a, l, sa, sl, bsize);
1671 
1672   // We must have chosen a partitioning and encoding or we'll fail later on.
1673   // No other opportunities for success.
1674   if ( bsize == BLOCK_64X64)
1675     assert(chosen_rate < INT_MAX && chosen_dist < INT64_MAX);
1676 
1677   if (do_recon) {
1678     int output_enabled = (bsize == BLOCK_64X64);
1679 
1680     // Check the projected output rate for this SB against it's target
1681     // and and if necessary apply a Q delta using segmentation to get
1682     // closer to the target.
1683     if ((cpi->oxcf.aq_mode == COMPLEXITY_AQ) && cm->seg.update_map) {
1684       vp9_select_in_frame_q_segment(cpi, mi_row, mi_col,
1685                                     output_enabled, chosen_rate);
1686     }
1687 
1688     if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ)
1689       vp9_cyclic_refresh_set_rate_and_dist_sb(cpi->cyclic_refresh,
1690                                               chosen_rate, chosen_dist);
1691     encode_sb(cpi, tile, tp, mi_row, mi_col, output_enabled, bsize,
1692               pc_tree);
1693   }
1694 
1695   *rate = chosen_rate;
1696   *dist = chosen_dist;
1697 }
1698 
1699 static const BLOCK_SIZE min_partition_size[BLOCK_SIZES] = {
1700   BLOCK_4X4,   BLOCK_4X4,   BLOCK_4X4,
1701   BLOCK_4X4,   BLOCK_4X4,   BLOCK_4X4,
1702   BLOCK_8X8,   BLOCK_8X8,   BLOCK_8X8,
1703   BLOCK_16X16, BLOCK_16X16, BLOCK_16X16,
1704   BLOCK_16X16
1705 };
1706 
1707 static const BLOCK_SIZE max_partition_size[BLOCK_SIZES] = {
1708   BLOCK_8X8,   BLOCK_16X16, BLOCK_16X16,
1709   BLOCK_16X16, BLOCK_32X32, BLOCK_32X32,
1710   BLOCK_32X32, BLOCK_64X64, BLOCK_64X64,
1711   BLOCK_64X64, BLOCK_64X64, BLOCK_64X64,
1712   BLOCK_64X64
1713 };
1714 
1715 // Look at all the mode_info entries for blocks that are part of this
1716 // partition and find the min and max values for sb_type.
1717 // At the moment this is designed to work on a 64x64 SB but could be
1718 // adjusted to use a size parameter.
1719 //
1720 // The min and max are assumed to have been initialized prior to calling this
1721 // 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])1722 static void get_sb_partition_size_range(MACROBLOCKD *xd, MODE_INFO *mi_8x8,
1723                                         BLOCK_SIZE *min_block_size,
1724                                         BLOCK_SIZE *max_block_size,
1725                                         int bs_hist[BLOCK_SIZES]) {
1726   int sb_width_in_blocks = MI_BLOCK_SIZE;
1727   int sb_height_in_blocks  = MI_BLOCK_SIZE;
1728   int i, j;
1729   int index = 0;
1730 
1731   // Check the sb_type for each block that belongs to this region.
1732   for (i = 0; i < sb_height_in_blocks; ++i) {
1733     for (j = 0; j < sb_width_in_blocks; ++j) {
1734       MODE_INFO *mi = mi_8x8[index+j].src_mi;
1735       BLOCK_SIZE sb_type = mi ? mi->mbmi.sb_type : 0;
1736       bs_hist[sb_type]++;
1737       *min_block_size = MIN(*min_block_size, sb_type);
1738       *max_block_size = MAX(*max_block_size, sb_type);
1739     }
1740     index += xd->mi_stride;
1741   }
1742 }
1743 
1744 // Next square block size less or equal than current block size.
1745 static const BLOCK_SIZE next_square_size[BLOCK_SIZES] = {
1746   BLOCK_4X4, BLOCK_4X4, BLOCK_4X4,
1747   BLOCK_8X8, BLOCK_8X8, BLOCK_8X8,
1748   BLOCK_16X16, BLOCK_16X16, BLOCK_16X16,
1749   BLOCK_32X32, BLOCK_32X32, BLOCK_32X32,
1750   BLOCK_64X64
1751 };
1752 
1753 // Look at neighboring blocks and set a min and max partition size based on
1754 // what they chose.
rd_auto_partition_range(VP9_COMP * cpi,const TileInfo * const tile,int mi_row,int mi_col,BLOCK_SIZE * min_block_size,BLOCK_SIZE * max_block_size)1755 static void rd_auto_partition_range(VP9_COMP *cpi, const TileInfo *const tile,
1756                                     int mi_row, int mi_col,
1757                                     BLOCK_SIZE *min_block_size,
1758                                     BLOCK_SIZE *max_block_size) {
1759   VP9_COMMON *const cm = &cpi->common;
1760   MACROBLOCKD *const xd = &cpi->mb.e_mbd;
1761   MODE_INFO *mi = xd->mi[0].src_mi;
1762   const int left_in_image = xd->left_available && mi[-1].src_mi;
1763   const int above_in_image = xd->up_available && mi[-xd->mi_stride].src_mi;
1764   const int row8x8_remaining = tile->mi_row_end - mi_row;
1765   const int col8x8_remaining = tile->mi_col_end - mi_col;
1766   int bh, bw;
1767   BLOCK_SIZE min_size = BLOCK_4X4;
1768   BLOCK_SIZE max_size = BLOCK_64X64;
1769   int i = 0;
1770   int bs_hist[BLOCK_SIZES] = {0};
1771 
1772   // Trap case where we do not have a prediction.
1773   if (left_in_image || above_in_image || cm->frame_type != KEY_FRAME) {
1774     // Default "min to max" and "max to min"
1775     min_size = BLOCK_64X64;
1776     max_size = BLOCK_4X4;
1777 
1778     // NOTE: each call to get_sb_partition_size_range() uses the previous
1779     // passed in values for min and max as a starting point.
1780     // Find the min and max partition used in previous frame at this location
1781     if (cm->frame_type != KEY_FRAME) {
1782       MODE_INFO *prev_mi =
1783           cm->prev_mip + cm->mi_stride + 1 + mi_row * xd->mi_stride + mi_col;
1784 
1785       get_sb_partition_size_range(xd, prev_mi, &min_size, &max_size, bs_hist);
1786     }
1787     // Find the min and max partition sizes used in the left SB64
1788     if (left_in_image) {
1789       MODE_INFO *left_sb64_mi = mi[-MI_BLOCK_SIZE].src_mi;
1790       get_sb_partition_size_range(xd, left_sb64_mi, &min_size, &max_size,
1791                                   bs_hist);
1792     }
1793     // Find the min and max partition sizes used in the above SB64.
1794     if (above_in_image) {
1795       MODE_INFO *above_sb64_mi = mi[-xd->mi_stride * MI_BLOCK_SIZE].src_mi;
1796       get_sb_partition_size_range(xd, above_sb64_mi, &min_size, &max_size,
1797                                   bs_hist);
1798     }
1799 
1800     // adjust observed min and max
1801     if (cpi->sf.auto_min_max_partition_size == RELAXED_NEIGHBORING_MIN_MAX) {
1802       min_size = min_partition_size[min_size];
1803       max_size = max_partition_size[max_size];
1804     } else if (cpi->sf.auto_min_max_partition_size ==
1805                CONSTRAIN_NEIGHBORING_MIN_MAX) {
1806       // adjust the search range based on the histogram of the observed
1807       // partition sizes from left, above the previous co-located blocks
1808       int sum = 0;
1809       int first_moment = 0;
1810       int second_moment = 0;
1811       int var_unnormalized = 0;
1812 
1813       for (i = 0; i < BLOCK_SIZES; i++) {
1814         sum += bs_hist[i];
1815         first_moment += bs_hist[i] * i;
1816         second_moment += bs_hist[i] * i * i;
1817       }
1818 
1819       // if variance is small enough,
1820       // adjust the range around its mean size, which gives a tighter range
1821       var_unnormalized = second_moment - first_moment * first_moment / sum;
1822       if (var_unnormalized <= 4 * sum) {
1823         int mean = first_moment / sum;
1824         min_size = min_partition_size[mean];
1825         max_size = max_partition_size[mean];
1826       } else {
1827         min_size = min_partition_size[min_size];
1828         max_size = max_partition_size[max_size];
1829       }
1830     }
1831   }
1832 
1833   // Check border cases where max and min from neighbors may not be legal.
1834   max_size = find_partition_size(max_size,
1835                                  row8x8_remaining, col8x8_remaining,
1836                                  &bh, &bw);
1837   min_size = MIN(min_size, max_size);
1838 
1839   // When use_square_partition_only is true, make sure at least one square
1840   // partition is allowed by selecting the next smaller square size as
1841   // *min_block_size.
1842   if (cpi->sf.use_square_partition_only &&
1843       next_square_size[max_size] < min_size) {
1844      min_size = next_square_size[max_size];
1845   }
1846 
1847   *min_block_size = min_size;
1848   *max_block_size = max_size;
1849 }
1850 
auto_partition_range(VP9_COMP * cpi,const TileInfo * const tile,int mi_row,int mi_col,BLOCK_SIZE * min_block_size,BLOCK_SIZE * max_block_size)1851 static void auto_partition_range(VP9_COMP *cpi, const TileInfo *const tile,
1852                                  int mi_row, int mi_col,
1853                                  BLOCK_SIZE *min_block_size,
1854                                  BLOCK_SIZE *max_block_size) {
1855   VP9_COMMON *const cm = &cpi->common;
1856   MACROBLOCKD *const xd = &cpi->mb.e_mbd;
1857   MODE_INFO *mi_8x8 = xd->mi;
1858   const int left_in_image = xd->left_available && mi_8x8[-1].src_mi;
1859   const int above_in_image = xd->up_available &&
1860                              mi_8x8[-xd->mi_stride].src_mi;
1861   int row8x8_remaining = tile->mi_row_end - mi_row;
1862   int col8x8_remaining = tile->mi_col_end - mi_col;
1863   int bh, bw;
1864   BLOCK_SIZE min_size = BLOCK_32X32;
1865   BLOCK_SIZE max_size = BLOCK_8X8;
1866   int bsl = mi_width_log2(BLOCK_64X64);
1867   const int search_range_ctrl = (((mi_row + mi_col) >> bsl) +
1868                        get_chessboard_index(cm->current_video_frame)) & 0x1;
1869   // Trap case where we do not have a prediction.
1870   if (search_range_ctrl &&
1871       (left_in_image || above_in_image || cm->frame_type != KEY_FRAME)) {
1872     int block;
1873     MODE_INFO *mi;
1874     BLOCK_SIZE sb_type;
1875 
1876     // Find the min and max partition sizes used in the left SB64.
1877     if (left_in_image) {
1878       MODE_INFO *cur_mi;
1879       mi = mi_8x8[-1].src_mi;
1880       for (block = 0; block < MI_BLOCK_SIZE; ++block) {
1881         cur_mi = mi[block * xd->mi_stride].src_mi;
1882         sb_type = cur_mi ? cur_mi->mbmi.sb_type : 0;
1883         min_size = MIN(min_size, sb_type);
1884         max_size = MAX(max_size, sb_type);
1885       }
1886     }
1887     // Find the min and max partition sizes used in the above SB64.
1888     if (above_in_image) {
1889       mi = mi_8x8[-xd->mi_stride * MI_BLOCK_SIZE].src_mi;
1890       for (block = 0; block < MI_BLOCK_SIZE; ++block) {
1891         sb_type = mi[block].src_mi ? mi[block].src_mi->mbmi.sb_type : 0;
1892         min_size = MIN(min_size, sb_type);
1893         max_size = MAX(max_size, sb_type);
1894       }
1895     }
1896 
1897     min_size = min_partition_size[min_size];
1898     max_size = find_partition_size(max_size, row8x8_remaining, col8x8_remaining,
1899                                    &bh, &bw);
1900     min_size = MIN(min_size, max_size);
1901     min_size = MAX(min_size, BLOCK_8X8);
1902     max_size = MIN(max_size, BLOCK_32X32);
1903   } else {
1904     min_size = BLOCK_8X8;
1905     max_size = BLOCK_32X32;
1906   }
1907 
1908   *min_block_size = min_size;
1909   *max_block_size = max_size;
1910 }
1911 
1912 // 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)1913 static void set_partition_range(VP9_COMMON *cm, MACROBLOCKD *xd,
1914                                 int mi_row, int mi_col, BLOCK_SIZE bsize,
1915                                 BLOCK_SIZE *min_bs, BLOCK_SIZE *max_bs) {
1916   int mi_width  = num_8x8_blocks_wide_lookup[bsize];
1917   int mi_height = num_8x8_blocks_high_lookup[bsize];
1918   int idx, idy;
1919 
1920   MODE_INFO *mi;
1921   const int idx_str = cm->mi_stride * mi_row + mi_col;
1922   MODE_INFO *prev_mi = (cm->prev_mip + cm->mi_stride + 1 + idx_str)->src_mi;
1923 
1924 
1925   BLOCK_SIZE bs, min_size, max_size;
1926 
1927   min_size = BLOCK_64X64;
1928   max_size = BLOCK_4X4;
1929 
1930   if (prev_mi) {
1931     for (idy = 0; idy < mi_height; ++idy) {
1932       for (idx = 0; idx < mi_width; ++idx) {
1933         mi = prev_mi[idy * cm->mi_stride + idx].src_mi;
1934         bs = mi ? mi->mbmi.sb_type : bsize;
1935         min_size = MIN(min_size, bs);
1936         max_size = MAX(max_size, bs);
1937       }
1938     }
1939   }
1940 
1941   if (xd->left_available) {
1942     for (idy = 0; idy < mi_height; ++idy) {
1943       mi = xd->mi[idy * cm->mi_stride - 1].src_mi;
1944       bs = mi ? mi->mbmi.sb_type : bsize;
1945       min_size = MIN(min_size, bs);
1946       max_size = MAX(max_size, bs);
1947     }
1948   }
1949 
1950   if (xd->up_available) {
1951     for (idx = 0; idx < mi_width; ++idx) {
1952       mi = xd->mi[idx - cm->mi_stride].src_mi;
1953       bs = mi ? mi->mbmi.sb_type : bsize;
1954       min_size = MIN(min_size, bs);
1955       max_size = MAX(max_size, bs);
1956     }
1957   }
1958 
1959   if (min_size == max_size) {
1960     min_size = min_partition_size[min_size];
1961     max_size = max_partition_size[max_size];
1962   }
1963 
1964   *min_bs = min_size;
1965   *max_bs = max_size;
1966 }
1967 
store_pred_mv(MACROBLOCK * x,PICK_MODE_CONTEXT * ctx)1968 static INLINE void store_pred_mv(MACROBLOCK *x, PICK_MODE_CONTEXT *ctx) {
1969   vpx_memcpy(ctx->pred_mv, x->pred_mv, sizeof(x->pred_mv));
1970 }
1971 
load_pred_mv(MACROBLOCK * x,PICK_MODE_CONTEXT * ctx)1972 static INLINE void load_pred_mv(MACROBLOCK *x, PICK_MODE_CONTEXT *ctx) {
1973   vpx_memcpy(x->pred_mv, ctx->pred_mv, sizeof(x->pred_mv));
1974 }
1975 
1976 #if CONFIG_FP_MB_STATS
1977 const int num_16x16_blocks_wide_lookup[BLOCK_SIZES] =
1978   {1, 1, 1, 1, 1, 1, 1, 1, 2, 2, 2, 4, 4};
1979 const int num_16x16_blocks_high_lookup[BLOCK_SIZES] =
1980   {1, 1, 1, 1, 1, 1, 1, 2, 1, 2, 4, 2, 4};
1981 const int qindex_skip_threshold_lookup[BLOCK_SIZES] =
1982   {0, 10, 10, 30, 40, 40, 60, 80, 80, 90, 100, 100, 120};
1983 const int qindex_split_threshold_lookup[BLOCK_SIZES] =
1984   {0, 3, 3, 7, 15, 15, 30, 40, 40, 60, 80, 80, 120};
1985 const int complexity_16x16_blocks_threshold[BLOCK_SIZES] =
1986   {1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 4, 4, 6};
1987 
1988 typedef enum {
1989   MV_ZERO = 0,
1990   MV_LEFT = 1,
1991   MV_UP = 2,
1992   MV_RIGHT = 3,
1993   MV_DOWN = 4,
1994   MV_INVALID
1995 } MOTION_DIRECTION;
1996 
get_motion_direction_fp(uint8_t fp_byte)1997 static INLINE MOTION_DIRECTION get_motion_direction_fp(uint8_t fp_byte) {
1998   if (fp_byte & FPMB_MOTION_ZERO_MASK) {
1999     return MV_ZERO;
2000   } else if (fp_byte & FPMB_MOTION_LEFT_MASK) {
2001     return MV_LEFT;
2002   } else if (fp_byte & FPMB_MOTION_RIGHT_MASK) {
2003     return MV_RIGHT;
2004   } else if (fp_byte & FPMB_MOTION_UP_MASK) {
2005     return MV_UP;
2006   } else {
2007     return MV_DOWN;
2008   }
2009 }
2010 
get_motion_inconsistency(MOTION_DIRECTION this_mv,MOTION_DIRECTION that_mv)2011 static INLINE int get_motion_inconsistency(MOTION_DIRECTION this_mv,
2012                                            MOTION_DIRECTION that_mv) {
2013   if (this_mv == that_mv) {
2014     return 0;
2015   } else {
2016     return abs(this_mv - that_mv) == 2 ? 2 : 1;
2017   }
2018 }
2019 #endif
2020 
2021 // TODO(jingning,jimbankoski,rbultje): properly skip partition types that are
2022 // unlikely to be selected depending on previous rate-distortion optimization
2023 // results, for encoding speed-up.
rd_pick_partition(VP9_COMP * cpi,const TileInfo * const tile,TOKENEXTRA ** tp,int mi_row,int mi_col,BLOCK_SIZE bsize,int * rate,int64_t * dist,int64_t best_rd,PC_TREE * pc_tree)2024 static void rd_pick_partition(VP9_COMP *cpi, const TileInfo *const tile,
2025                               TOKENEXTRA **tp, int mi_row,
2026                               int mi_col, BLOCK_SIZE bsize, int *rate,
2027                               int64_t *dist, int64_t best_rd,
2028                               PC_TREE *pc_tree) {
2029   VP9_COMMON *const cm = &cpi->common;
2030   MACROBLOCK *const x = &cpi->mb;
2031   MACROBLOCKD *const xd = &x->e_mbd;
2032   const int mi_step = num_8x8_blocks_wide_lookup[bsize] / 2;
2033   ENTROPY_CONTEXT l[16 * MAX_MB_PLANE], a[16 * MAX_MB_PLANE];
2034   PARTITION_CONTEXT sl[8], sa[8];
2035   TOKENEXTRA *tp_orig = *tp;
2036   PICK_MODE_CONTEXT *ctx = &pc_tree->none;
2037   int i, pl;
2038   BLOCK_SIZE subsize;
2039   int this_rate, sum_rate = 0, best_rate = INT_MAX;
2040   int64_t this_dist, sum_dist = 0, best_dist = INT64_MAX;
2041   int64_t sum_rd = 0;
2042   int do_split = bsize >= BLOCK_8X8;
2043   int do_rect = 1;
2044 
2045   // Override skipping rectangular partition operations for edge blocks
2046   const int force_horz_split = (mi_row + mi_step >= cm->mi_rows);
2047   const int force_vert_split = (mi_col + mi_step >= cm->mi_cols);
2048   const int xss = x->e_mbd.plane[1].subsampling_x;
2049   const int yss = x->e_mbd.plane[1].subsampling_y;
2050 
2051   BLOCK_SIZE min_size = cpi->sf.min_partition_size;
2052   BLOCK_SIZE max_size = cpi->sf.max_partition_size;
2053 
2054 #if CONFIG_FP_MB_STATS
2055   unsigned int src_diff_var = UINT_MAX;
2056   int none_complexity = 0;
2057 #endif
2058 
2059   int partition_none_allowed = !force_horz_split && !force_vert_split;
2060   int partition_horz_allowed = !force_vert_split && yss <= xss &&
2061                                bsize >= BLOCK_8X8;
2062   int partition_vert_allowed = !force_horz_split && xss <= yss &&
2063                                bsize >= BLOCK_8X8;
2064   (void) *tp_orig;
2065 
2066   assert(num_8x8_blocks_wide_lookup[bsize] ==
2067              num_8x8_blocks_high_lookup[bsize]);
2068 
2069   set_offsets(cpi, tile, mi_row, mi_col, bsize);
2070 
2071   if (bsize == BLOCK_16X16 && cpi->oxcf.aq_mode)
2072     x->mb_energy = vp9_block_energy(cpi, x, bsize);
2073 
2074   if (cpi->sf.cb_partition_search && bsize == BLOCK_16X16) {
2075     int cb_partition_search_ctrl = ((pc_tree->index == 0 || pc_tree->index == 3)
2076         + get_chessboard_index(cm->current_video_frame)) & 0x1;
2077 
2078     if (cb_partition_search_ctrl && bsize > min_size && bsize < max_size)
2079       set_partition_range(cm, xd, mi_row, mi_col, bsize, &min_size, &max_size);
2080   }
2081 
2082   // Determine partition types in search according to the speed features.
2083   // The threshold set here has to be of square block size.
2084   if (cpi->sf.auto_min_max_partition_size) {
2085     partition_none_allowed &= (bsize <= max_size && bsize >= min_size);
2086     partition_horz_allowed &= ((bsize <= max_size && bsize > min_size) ||
2087                                 force_horz_split);
2088     partition_vert_allowed &= ((bsize <= max_size && bsize > min_size) ||
2089                                 force_vert_split);
2090     do_split &= bsize > min_size;
2091   }
2092   if (cpi->sf.use_square_partition_only) {
2093     partition_horz_allowed &= force_horz_split;
2094     partition_vert_allowed &= force_vert_split;
2095   }
2096 
2097   save_context(cpi, mi_row, mi_col, a, l, sa, sl, bsize);
2098 
2099 #if CONFIG_FP_MB_STATS
2100   if (cpi->use_fp_mb_stats) {
2101     set_offsets(cpi, tile, mi_row, mi_col, bsize);
2102     src_diff_var = get_sby_perpixel_diff_variance(cpi, &cpi->mb.plane[0].src,
2103                                                   mi_row, mi_col, bsize);
2104   }
2105 #endif
2106 
2107 #if CONFIG_FP_MB_STATS
2108   // Decide whether we shall split directly and skip searching NONE by using
2109   // the first pass block statistics
2110   if (cpi->use_fp_mb_stats && bsize >= BLOCK_32X32 && do_split &&
2111       partition_none_allowed && src_diff_var > 4 &&
2112       cm->base_qindex < qindex_split_threshold_lookup[bsize]) {
2113     int mb_row = mi_row >> 1;
2114     int mb_col = mi_col >> 1;
2115     int mb_row_end =
2116         MIN(mb_row + num_16x16_blocks_high_lookup[bsize], cm->mb_rows);
2117     int mb_col_end =
2118         MIN(mb_col + num_16x16_blocks_wide_lookup[bsize], cm->mb_cols);
2119     int r, c;
2120 
2121     // compute a complexity measure, basically measure inconsistency of motion
2122     // vectors obtained from the first pass in the current block
2123     for (r = mb_row; r < mb_row_end ; r++) {
2124       for (c = mb_col; c < mb_col_end; c++) {
2125         const int mb_index = r * cm->mb_cols + c;
2126 
2127         MOTION_DIRECTION this_mv;
2128         MOTION_DIRECTION right_mv;
2129         MOTION_DIRECTION bottom_mv;
2130 
2131         this_mv =
2132             get_motion_direction_fp(cpi->twopass.this_frame_mb_stats[mb_index]);
2133 
2134         // to its right
2135         if (c != mb_col_end - 1) {
2136           right_mv = get_motion_direction_fp(
2137               cpi->twopass.this_frame_mb_stats[mb_index + 1]);
2138           none_complexity += get_motion_inconsistency(this_mv, right_mv);
2139         }
2140 
2141         // to its bottom
2142         if (r != mb_row_end - 1) {
2143           bottom_mv = get_motion_direction_fp(
2144               cpi->twopass.this_frame_mb_stats[mb_index + cm->mb_cols]);
2145           none_complexity += get_motion_inconsistency(this_mv, bottom_mv);
2146         }
2147 
2148         // do not count its left and top neighbors to avoid double counting
2149       }
2150     }
2151 
2152     if (none_complexity > complexity_16x16_blocks_threshold[bsize]) {
2153       partition_none_allowed = 0;
2154     }
2155   }
2156 #endif
2157 
2158   // PARTITION_NONE
2159   if (partition_none_allowed) {
2160     rd_pick_sb_modes(cpi, tile, mi_row, mi_col, &this_rate, &this_dist, bsize,
2161                      ctx, best_rd, 0);
2162     if (this_rate != INT_MAX) {
2163       if (bsize >= BLOCK_8X8) {
2164         pl = partition_plane_context(xd, mi_row, mi_col, bsize);
2165         this_rate += cpi->partition_cost[pl][PARTITION_NONE];
2166       }
2167       sum_rd = RDCOST(x->rdmult, x->rddiv, this_rate, this_dist);
2168 
2169       if (sum_rd < best_rd) {
2170         int64_t dist_breakout_thr = cpi->sf.partition_search_breakout_dist_thr;
2171         int rate_breakout_thr = cpi->sf.partition_search_breakout_rate_thr;
2172 
2173         best_rate = this_rate;
2174         best_dist = this_dist;
2175         best_rd = sum_rd;
2176         if (bsize >= BLOCK_8X8)
2177           pc_tree->partitioning = PARTITION_NONE;
2178 
2179         // Adjust dist breakout threshold according to the partition size.
2180         dist_breakout_thr >>= 8 - (b_width_log2(bsize) +
2181             b_height_log2(bsize));
2182 
2183         // If all y, u, v transform blocks in this partition are skippable, and
2184         // the dist & rate are within the thresholds, the partition search is
2185         // terminated for current branch of the partition search tree.
2186         // The dist & rate thresholds are set to 0 at speed 0 to disable the
2187         // early termination at that speed.
2188         if (!x->e_mbd.lossless &&
2189             (ctx->skippable && best_dist < dist_breakout_thr &&
2190             best_rate < rate_breakout_thr)) {
2191           do_split = 0;
2192           do_rect = 0;
2193         }
2194 
2195 #if CONFIG_FP_MB_STATS
2196         // Check if every 16x16 first pass block statistics has zero
2197         // motion and the corresponding first pass residue is small enough.
2198         // If that is the case, check the difference variance between the
2199         // current frame and the last frame. If the variance is small enough,
2200         // stop further splitting in RD optimization
2201         if (cpi->use_fp_mb_stats && do_split != 0 &&
2202             cm->base_qindex > qindex_skip_threshold_lookup[bsize]) {
2203           int mb_row = mi_row >> 1;
2204           int mb_col = mi_col >> 1;
2205           int mb_row_end =
2206               MIN(mb_row + num_16x16_blocks_high_lookup[bsize], cm->mb_rows);
2207           int mb_col_end =
2208               MIN(mb_col + num_16x16_blocks_wide_lookup[bsize], cm->mb_cols);
2209           int r, c;
2210 
2211           int skip = 1;
2212           for (r = mb_row; r < mb_row_end; r++) {
2213             for (c = mb_col; c < mb_col_end; c++) {
2214               const int mb_index = r * cm->mb_cols + c;
2215               if (!(cpi->twopass.this_frame_mb_stats[mb_index] &
2216                     FPMB_MOTION_ZERO_MASK) ||
2217                   !(cpi->twopass.this_frame_mb_stats[mb_index] &
2218                     FPMB_ERROR_SMALL_MASK)) {
2219                 skip = 0;
2220                 break;
2221               }
2222             }
2223             if (skip == 0) {
2224               break;
2225             }
2226           }
2227           if (skip) {
2228             if (src_diff_var == UINT_MAX) {
2229               set_offsets(cpi, tile, mi_row, mi_col, bsize);
2230               src_diff_var = get_sby_perpixel_diff_variance(
2231                   cpi, &cpi->mb.plane[0].src, mi_row, mi_col, bsize);
2232             }
2233             if (src_diff_var < 8) {
2234               do_split = 0;
2235               do_rect = 0;
2236             }
2237           }
2238         }
2239 #endif
2240       }
2241     }
2242     restore_context(cpi, mi_row, mi_col, a, l, sa, sl, bsize);
2243   }
2244 
2245   // store estimated motion vector
2246   if (cpi->sf.adaptive_motion_search)
2247     store_pred_mv(x, ctx);
2248 
2249   // PARTITION_SPLIT
2250   sum_rd = 0;
2251   // TODO(jingning): use the motion vectors given by the above search as
2252   // the starting point of motion search in the following partition type check.
2253   if (do_split) {
2254     subsize = get_subsize(bsize, PARTITION_SPLIT);
2255     if (bsize == BLOCK_8X8) {
2256       i = 4;
2257       if (cpi->sf.adaptive_pred_interp_filter && partition_none_allowed)
2258         pc_tree->leaf_split[0]->pred_interp_filter =
2259             ctx->mic.mbmi.interp_filter;
2260       rd_pick_sb_modes(cpi, tile, mi_row, mi_col, &sum_rate, &sum_dist, subsize,
2261                        pc_tree->leaf_split[0], best_rd, 0);
2262       if (sum_rate == INT_MAX)
2263         sum_rd = INT64_MAX;
2264       else
2265         sum_rd = RDCOST(x->rdmult, x->rddiv, sum_rate, sum_dist);
2266     } else {
2267       for (i = 0; i < 4 && sum_rd < best_rd; ++i) {
2268       const int x_idx = (i & 1) * mi_step;
2269       const int y_idx = (i >> 1) * mi_step;
2270 
2271         if (mi_row + y_idx >= cm->mi_rows || mi_col + x_idx >= cm->mi_cols)
2272           continue;
2273 
2274         if (cpi->sf.adaptive_motion_search)
2275           load_pred_mv(x, ctx);
2276 
2277         pc_tree->split[i]->index = i;
2278         rd_pick_partition(cpi, tile, tp, mi_row + y_idx, mi_col + x_idx,
2279                           subsize, &this_rate, &this_dist,
2280                           best_rd - sum_rd, pc_tree->split[i]);
2281 
2282         if (this_rate == INT_MAX) {
2283           sum_rd = INT64_MAX;
2284         } else {
2285           sum_rate += this_rate;
2286           sum_dist += this_dist;
2287           sum_rd = RDCOST(x->rdmult, x->rddiv, sum_rate, sum_dist);
2288         }
2289       }
2290     }
2291 
2292     if (sum_rd < best_rd && i == 4) {
2293       pl = partition_plane_context(xd, mi_row, mi_col, bsize);
2294       sum_rate += cpi->partition_cost[pl][PARTITION_SPLIT];
2295       sum_rd = RDCOST(x->rdmult, x->rddiv, sum_rate, sum_dist);
2296 
2297       if (sum_rd < best_rd) {
2298         best_rate = sum_rate;
2299         best_dist = sum_dist;
2300         best_rd = sum_rd;
2301         pc_tree->partitioning = PARTITION_SPLIT;
2302       }
2303     } else {
2304       // skip rectangular partition test when larger block size
2305       // gives better rd cost
2306       if (cpi->sf.less_rectangular_check)
2307         do_rect &= !partition_none_allowed;
2308     }
2309     restore_context(cpi, mi_row, mi_col, a, l, sa, sl, bsize);
2310   }
2311 
2312   // PARTITION_HORZ
2313   if (partition_horz_allowed && do_rect) {
2314     subsize = get_subsize(bsize, PARTITION_HORZ);
2315     if (cpi->sf.adaptive_motion_search)
2316       load_pred_mv(x, ctx);
2317     if (cpi->sf.adaptive_pred_interp_filter && bsize == BLOCK_8X8 &&
2318         partition_none_allowed)
2319       pc_tree->horizontal[0].pred_interp_filter =
2320           ctx->mic.mbmi.interp_filter;
2321     rd_pick_sb_modes(cpi, tile, mi_row, mi_col, &sum_rate, &sum_dist, subsize,
2322                      &pc_tree->horizontal[0], best_rd, 0);
2323     sum_rd = RDCOST(x->rdmult, x->rddiv, sum_rate, sum_dist);
2324 
2325     if (sum_rd < best_rd && mi_row + mi_step < cm->mi_rows) {
2326       PICK_MODE_CONTEXT *ctx = &pc_tree->horizontal[0];
2327       update_state(cpi, ctx, mi_row, mi_col, subsize, 0);
2328       encode_superblock(cpi, tp, 0, mi_row, mi_col, subsize, ctx);
2329 
2330       if (cpi->sf.adaptive_motion_search)
2331         load_pred_mv(x, ctx);
2332       if (cpi->sf.adaptive_pred_interp_filter && bsize == BLOCK_8X8 &&
2333           partition_none_allowed)
2334         pc_tree->horizontal[1].pred_interp_filter =
2335             ctx->mic.mbmi.interp_filter;
2336       rd_pick_sb_modes(cpi, tile, mi_row + mi_step, mi_col, &this_rate,
2337                        &this_dist, subsize, &pc_tree->horizontal[1],
2338                        best_rd - sum_rd, 1);
2339       if (this_rate == INT_MAX) {
2340         sum_rd = INT64_MAX;
2341       } else {
2342         sum_rate += this_rate;
2343         sum_dist += this_dist;
2344         sum_rd = RDCOST(x->rdmult, x->rddiv, sum_rate, sum_dist);
2345       }
2346     }
2347     if (sum_rd < best_rd) {
2348       pl = partition_plane_context(xd, mi_row, mi_col, bsize);
2349       sum_rate += cpi->partition_cost[pl][PARTITION_HORZ];
2350       sum_rd = RDCOST(x->rdmult, x->rddiv, sum_rate, sum_dist);
2351       if (sum_rd < best_rd) {
2352         best_rd = sum_rd;
2353         best_rate = sum_rate;
2354         best_dist = sum_dist;
2355         pc_tree->partitioning = PARTITION_HORZ;
2356       }
2357     }
2358     restore_context(cpi, mi_row, mi_col, a, l, sa, sl, bsize);
2359   }
2360   // PARTITION_VERT
2361   if (partition_vert_allowed && do_rect) {
2362     subsize = get_subsize(bsize, PARTITION_VERT);
2363 
2364     if (cpi->sf.adaptive_motion_search)
2365       load_pred_mv(x, ctx);
2366     if (cpi->sf.adaptive_pred_interp_filter && bsize == BLOCK_8X8 &&
2367         partition_none_allowed)
2368       pc_tree->vertical[0].pred_interp_filter =
2369           ctx->mic.mbmi.interp_filter;
2370     rd_pick_sb_modes(cpi, tile, mi_row, mi_col, &sum_rate, &sum_dist, subsize,
2371                      &pc_tree->vertical[0], best_rd, 0);
2372     sum_rd = RDCOST(x->rdmult, x->rddiv, sum_rate, sum_dist);
2373     if (sum_rd < best_rd && mi_col + mi_step < cm->mi_cols) {
2374       update_state(cpi, &pc_tree->vertical[0], mi_row, mi_col, subsize, 0);
2375       encode_superblock(cpi, tp, 0, mi_row, mi_col, subsize,
2376                         &pc_tree->vertical[0]);
2377 
2378       if (cpi->sf.adaptive_motion_search)
2379         load_pred_mv(x, ctx);
2380       if (cpi->sf.adaptive_pred_interp_filter && bsize == BLOCK_8X8 &&
2381           partition_none_allowed)
2382         pc_tree->vertical[1].pred_interp_filter =
2383             ctx->mic.mbmi.interp_filter;
2384       rd_pick_sb_modes(cpi, tile, mi_row, mi_col + mi_step, &this_rate,
2385                        &this_dist, subsize,
2386                        &pc_tree->vertical[1], best_rd - sum_rd,
2387                        1);
2388       if (this_rate == INT_MAX) {
2389         sum_rd = INT64_MAX;
2390       } else {
2391         sum_rate += this_rate;
2392         sum_dist += this_dist;
2393         sum_rd = RDCOST(x->rdmult, x->rddiv, sum_rate, sum_dist);
2394       }
2395     }
2396     if (sum_rd < best_rd) {
2397       pl = partition_plane_context(xd, mi_row, mi_col, bsize);
2398       sum_rate += cpi->partition_cost[pl][PARTITION_VERT];
2399       sum_rd = RDCOST(x->rdmult, x->rddiv, sum_rate, sum_dist);
2400       if (sum_rd < best_rd) {
2401         best_rate = sum_rate;
2402         best_dist = sum_dist;
2403         best_rd = sum_rd;
2404         pc_tree->partitioning = PARTITION_VERT;
2405       }
2406     }
2407     restore_context(cpi, mi_row, mi_col, a, l, sa, sl, bsize);
2408   }
2409 
2410   // TODO(jbb): This code added so that we avoid static analysis
2411   // warning related to the fact that best_rd isn't used after this
2412   // point.  This code should be refactored so that the duplicate
2413   // checks occur in some sub function and thus are used...
2414   (void) best_rd;
2415   *rate = best_rate;
2416   *dist = best_dist;
2417 
2418   if (best_rate < INT_MAX && best_dist < INT64_MAX && pc_tree->index != 3) {
2419     int output_enabled = (bsize == BLOCK_64X64);
2420 
2421     // Check the projected output rate for this SB against it's target
2422     // and and if necessary apply a Q delta using segmentation to get
2423     // closer to the target.
2424     if ((cpi->oxcf.aq_mode == COMPLEXITY_AQ) && cm->seg.update_map)
2425       vp9_select_in_frame_q_segment(cpi, mi_row, mi_col, output_enabled,
2426                                     best_rate);
2427     if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ)
2428       vp9_cyclic_refresh_set_rate_and_dist_sb(cpi->cyclic_refresh,
2429                                               best_rate, best_dist);
2430 
2431     encode_sb(cpi, tile, tp, mi_row, mi_col, output_enabled, bsize, pc_tree);
2432   }
2433 
2434   if (bsize == BLOCK_64X64) {
2435     assert(tp_orig < *tp);
2436     assert(best_rate < INT_MAX);
2437     assert(best_dist < INT64_MAX);
2438   } else {
2439     assert(tp_orig == *tp);
2440   }
2441 }
2442 
encode_rd_sb_row(VP9_COMP * cpi,const TileInfo * const tile,int mi_row,TOKENEXTRA ** tp)2443 static void encode_rd_sb_row(VP9_COMP *cpi, const TileInfo *const tile,
2444                              int mi_row, TOKENEXTRA **tp) {
2445   VP9_COMMON *const cm = &cpi->common;
2446   MACROBLOCKD *const xd = &cpi->mb.e_mbd;
2447   SPEED_FEATURES *const sf = &cpi->sf;
2448   int mi_col;
2449 
2450   // Initialize the left context for the new SB row
2451   vpx_memset(&xd->left_context, 0, sizeof(xd->left_context));
2452   vpx_memset(xd->left_seg_context, 0, sizeof(xd->left_seg_context));
2453 
2454   // Code each SB in the row
2455   for (mi_col = tile->mi_col_start; mi_col < tile->mi_col_end;
2456        mi_col += MI_BLOCK_SIZE) {
2457     int dummy_rate;
2458     int64_t dummy_dist;
2459 
2460     int i;
2461 
2462     if (sf->adaptive_pred_interp_filter) {
2463       for (i = 0; i < 64; ++i)
2464         cpi->leaf_tree[i].pred_interp_filter = SWITCHABLE;
2465 
2466       for (i = 0; i < 64; ++i) {
2467         cpi->pc_tree[i].vertical[0].pred_interp_filter = SWITCHABLE;
2468         cpi->pc_tree[i].vertical[1].pred_interp_filter = SWITCHABLE;
2469         cpi->pc_tree[i].horizontal[0].pred_interp_filter = SWITCHABLE;
2470         cpi->pc_tree[i].horizontal[1].pred_interp_filter = SWITCHABLE;
2471       }
2472     }
2473 
2474     vp9_zero(cpi->mb.pred_mv);
2475     cpi->pc_root->index = 0;
2476 
2477     // TODO(yunqingwang): use_lastframe_partitioning is no longer used in good-
2478     // quality encoding. Need to evaluate it in real-time encoding later to
2479     // decide if it can be removed too. And then, do the code cleanup.
2480     if ((sf->partition_search_type == SEARCH_PARTITION &&
2481          sf->use_lastframe_partitioning) ||
2482          sf->partition_search_type == FIXED_PARTITION ||
2483          sf->partition_search_type == VAR_BASED_PARTITION ||
2484          sf->partition_search_type == VAR_BASED_FIXED_PARTITION) {
2485       const int idx_str = cm->mi_stride * mi_row + mi_col;
2486       MODE_INFO *mi = cm->mi + idx_str;
2487       MODE_INFO *prev_mi = (cm->prev_mip + cm->mi_stride + 1 + idx_str)->src_mi;
2488       cpi->mb.source_variance = UINT_MAX;
2489       if (sf->partition_search_type == FIXED_PARTITION) {
2490         set_offsets(cpi, tile, mi_row, mi_col, BLOCK_64X64);
2491         set_fixed_partitioning(cpi, tile, mi, mi_row, mi_col,
2492                                sf->always_this_block_size);
2493         rd_use_partition(cpi, tile, mi, tp, mi_row, mi_col, BLOCK_64X64,
2494                          &dummy_rate, &dummy_dist, 1, cpi->pc_root);
2495       } else if (cpi->skippable_frame ||
2496                  sf->partition_search_type == VAR_BASED_FIXED_PARTITION) {
2497         BLOCK_SIZE bsize;
2498         set_offsets(cpi, tile, mi_row, mi_col, BLOCK_64X64);
2499         bsize = get_rd_var_based_fixed_partition(cpi, mi_row, mi_col);
2500         set_fixed_partitioning(cpi, tile, mi, mi_row, mi_col, bsize);
2501         rd_use_partition(cpi, tile, mi, tp, mi_row, mi_col, BLOCK_64X64,
2502                          &dummy_rate, &dummy_dist, 1, cpi->pc_root);
2503       } else if (sf->partition_search_type == VAR_BASED_PARTITION) {
2504         choose_partitioning(cpi, tile, mi_row, mi_col);
2505         rd_use_partition(cpi, tile, mi, tp, mi_row, mi_col, BLOCK_64X64,
2506                          &dummy_rate, &dummy_dist, 1, cpi->pc_root);
2507       } else {
2508         GF_GROUP * gf_grp = &cpi->twopass.gf_group;
2509         int last_was_mid_sequence_overlay = 0;
2510         if ((cpi->oxcf.pass == 2) && (gf_grp->index)) {
2511           if (gf_grp->update_type[gf_grp->index - 1] == OVERLAY_UPDATE)
2512             last_was_mid_sequence_overlay = 1;
2513         }
2514         if ((cpi->rc.frames_since_key
2515             % sf->last_partitioning_redo_frequency) == 0
2516             || last_was_mid_sequence_overlay
2517             || cm->prev_mi == 0
2518             || cm->show_frame == 0
2519             || cm->frame_type == KEY_FRAME
2520             || cpi->rc.is_src_frame_alt_ref
2521             || ((sf->use_lastframe_partitioning ==
2522                  LAST_FRAME_PARTITION_LOW_MOTION) &&
2523                  sb_has_motion(cm, prev_mi, sf->lf_motion_threshold))) {
2524           // If required set upper and lower partition size limits
2525           if (sf->auto_min_max_partition_size) {
2526             set_offsets(cpi, tile, mi_row, mi_col, BLOCK_64X64);
2527             rd_auto_partition_range(cpi, tile, mi_row, mi_col,
2528                                     &sf->min_partition_size,
2529                                     &sf->max_partition_size);
2530           }
2531           rd_pick_partition(cpi, tile, tp, mi_row, mi_col, BLOCK_64X64,
2532                             &dummy_rate, &dummy_dist, INT64_MAX,
2533                             cpi->pc_root);
2534         } else {
2535           if (sf->constrain_copy_partition &&
2536               sb_has_motion(cm, prev_mi, sf->lf_motion_threshold))
2537             constrain_copy_partitioning(cpi, tile, mi, prev_mi,
2538                                         mi_row, mi_col, BLOCK_16X16);
2539           else
2540             copy_partitioning(cm, mi, prev_mi);
2541           rd_use_partition(cpi, tile, mi, tp, mi_row, mi_col, BLOCK_64X64,
2542                            &dummy_rate, &dummy_dist, 1, cpi->pc_root);
2543         }
2544       }
2545     } else {
2546       // If required set upper and lower partition size limits
2547       if (sf->auto_min_max_partition_size) {
2548         set_offsets(cpi, tile, mi_row, mi_col, BLOCK_64X64);
2549         rd_auto_partition_range(cpi, tile, mi_row, mi_col,
2550                                 &sf->min_partition_size,
2551                                 &sf->max_partition_size);
2552       }
2553       rd_pick_partition(cpi, tile, tp, mi_row, mi_col, BLOCK_64X64,
2554                         &dummy_rate, &dummy_dist, INT64_MAX, cpi->pc_root);
2555     }
2556   }
2557 }
2558 
init_encode_frame_mb_context(VP9_COMP * cpi)2559 static void init_encode_frame_mb_context(VP9_COMP *cpi) {
2560   MACROBLOCK *const x = &cpi->mb;
2561   VP9_COMMON *const cm = &cpi->common;
2562   MACROBLOCKD *const xd = &x->e_mbd;
2563   const int aligned_mi_cols = mi_cols_aligned_to_sb(cm->mi_cols);
2564 
2565   // Copy data over into macro block data structures.
2566   vp9_setup_src_planes(x, cpi->Source, 0, 0);
2567 
2568   vp9_setup_block_planes(&x->e_mbd, cm->subsampling_x, cm->subsampling_y);
2569 
2570   // Note: this memset assumes above_context[0], [1] and [2]
2571   // are allocated as part of the same buffer.
2572   vpx_memset(xd->above_context[0], 0,
2573              sizeof(*xd->above_context[0]) *
2574              2 * aligned_mi_cols * MAX_MB_PLANE);
2575   vpx_memset(xd->above_seg_context, 0,
2576              sizeof(*xd->above_seg_context) * aligned_mi_cols);
2577 }
2578 
check_dual_ref_flags(VP9_COMP * cpi)2579 static int check_dual_ref_flags(VP9_COMP *cpi) {
2580   const int ref_flags = cpi->ref_frame_flags;
2581 
2582   if (vp9_segfeature_active(&cpi->common.seg, 1, SEG_LVL_REF_FRAME)) {
2583     return 0;
2584   } else {
2585     return (!!(ref_flags & VP9_GOLD_FLAG) + !!(ref_flags & VP9_LAST_FLAG)
2586         + !!(ref_flags & VP9_ALT_FLAG)) >= 2;
2587   }
2588 }
2589 
reset_skip_tx_size(VP9_COMMON * cm,TX_SIZE max_tx_size)2590 static void reset_skip_tx_size(VP9_COMMON *cm, TX_SIZE max_tx_size) {
2591   int mi_row, mi_col;
2592   const int mis = cm->mi_stride;
2593   MODE_INFO *mi_ptr = cm->mi;
2594 
2595   for (mi_row = 0; mi_row < cm->mi_rows; ++mi_row, mi_ptr += mis) {
2596     for (mi_col = 0; mi_col < cm->mi_cols; ++mi_col) {
2597       if (mi_ptr[mi_col].src_mi->mbmi.tx_size > max_tx_size)
2598         mi_ptr[mi_col].src_mi->mbmi.tx_size = max_tx_size;
2599     }
2600   }
2601 }
2602 
get_frame_type(const VP9_COMP * cpi)2603 static MV_REFERENCE_FRAME get_frame_type(const VP9_COMP *cpi) {
2604   if (frame_is_intra_only(&cpi->common))
2605     return INTRA_FRAME;
2606   else if (cpi->rc.is_src_frame_alt_ref && cpi->refresh_golden_frame)
2607     return ALTREF_FRAME;
2608   else if (cpi->refresh_golden_frame || cpi->refresh_alt_ref_frame)
2609     return GOLDEN_FRAME;
2610   else
2611     return LAST_FRAME;
2612 }
2613 
select_tx_mode(const VP9_COMP * cpi)2614 static TX_MODE select_tx_mode(const VP9_COMP *cpi) {
2615   if (cpi->mb.e_mbd.lossless)
2616     return ONLY_4X4;
2617   if (cpi->sf.tx_size_search_method == USE_LARGESTALL)
2618     return ALLOW_32X32;
2619   else if (cpi->sf.tx_size_search_method == USE_FULL_RD||
2620            cpi->sf.tx_size_search_method == USE_TX_8X8)
2621     return TX_MODE_SELECT;
2622   else
2623     return cpi->common.tx_mode;
2624 }
2625 
nonrd_pick_sb_modes(VP9_COMP * cpi,const TileInfo * const tile,int mi_row,int mi_col,int * rate,int64_t * dist,BLOCK_SIZE bsize,PICK_MODE_CONTEXT * ctx)2626 static void nonrd_pick_sb_modes(VP9_COMP *cpi, const TileInfo *const tile,
2627                                 int mi_row, int mi_col,
2628                                 int *rate, int64_t *dist,
2629                                 BLOCK_SIZE bsize, PICK_MODE_CONTEXT *ctx) {
2630   VP9_COMMON *const cm = &cpi->common;
2631   MACROBLOCK *const x = &cpi->mb;
2632   MACROBLOCKD *const xd = &x->e_mbd;
2633   MB_MODE_INFO *mbmi;
2634   set_offsets(cpi, tile, mi_row, mi_col, bsize);
2635   mbmi = &xd->mi[0].src_mi->mbmi;
2636   mbmi->sb_type = bsize;
2637 
2638   if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ && cm->seg.enabled)
2639     if (mbmi->segment_id && x->in_static_area)
2640       x->rdmult = vp9_cyclic_refresh_get_rdmult(cpi->cyclic_refresh);
2641 
2642   if (vp9_segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_SKIP))
2643     set_mode_info_seg_skip(x, cm->tx_mode, rate, dist, bsize);
2644   else
2645     vp9_pick_inter_mode(cpi, x, tile, mi_row, mi_col, rate, dist, bsize, ctx);
2646 
2647   duplicate_mode_info_in_sb(cm, xd, mi_row, mi_col, bsize);
2648 }
2649 
fill_mode_info_sb(VP9_COMMON * cm,MACROBLOCK * x,int mi_row,int mi_col,BLOCK_SIZE bsize,BLOCK_SIZE subsize,PC_TREE * pc_tree)2650 static void fill_mode_info_sb(VP9_COMMON *cm, MACROBLOCK *x,
2651                               int mi_row, int mi_col,
2652                               BLOCK_SIZE bsize, BLOCK_SIZE subsize,
2653                               PC_TREE *pc_tree) {
2654   MACROBLOCKD *xd = &x->e_mbd;
2655   int bsl = b_width_log2(bsize), hbs = (1 << bsl) / 4;
2656   PARTITION_TYPE partition = pc_tree->partitioning;
2657 
2658   assert(bsize >= BLOCK_8X8);
2659 
2660   if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols)
2661     return;
2662 
2663   switch (partition) {
2664     case PARTITION_NONE:
2665       set_modeinfo_offsets(cm, xd, mi_row, mi_col);
2666       *(xd->mi[0].src_mi) = pc_tree->none.mic;
2667       duplicate_mode_info_in_sb(cm, xd, mi_row, mi_col, bsize);
2668       break;
2669     case PARTITION_VERT:
2670       set_modeinfo_offsets(cm, xd, mi_row, mi_col);
2671       *(xd->mi[0].src_mi) = pc_tree->vertical[0].mic;
2672       duplicate_mode_info_in_sb(cm, xd, mi_row, mi_col, bsize);
2673 
2674       if (mi_col + hbs < cm->mi_cols) {
2675         set_modeinfo_offsets(cm, xd, mi_row, mi_col + hbs);
2676         *(xd->mi[0].src_mi) = pc_tree->vertical[1].mic;
2677         duplicate_mode_info_in_sb(cm, xd, mi_row, mi_col + hbs, bsize);
2678       }
2679       break;
2680     case PARTITION_HORZ:
2681       set_modeinfo_offsets(cm, xd, mi_row, mi_col);
2682       *(xd->mi[0].src_mi) = pc_tree->horizontal[0].mic;
2683       duplicate_mode_info_in_sb(cm, xd, mi_row, mi_col, bsize);
2684       if (mi_row + hbs < cm->mi_rows) {
2685         set_modeinfo_offsets(cm, xd, mi_row + hbs, mi_col);
2686         *(xd->mi[0].src_mi) = pc_tree->horizontal[1].mic;
2687         duplicate_mode_info_in_sb(cm, xd, mi_row + hbs, mi_col, bsize);
2688       }
2689       break;
2690     case PARTITION_SPLIT: {
2691       BLOCK_SIZE subsubsize = get_subsize(subsize, PARTITION_SPLIT);
2692       fill_mode_info_sb(cm, x, mi_row, mi_col, subsize,
2693                         subsubsize, pc_tree->split[0]);
2694       fill_mode_info_sb(cm, x, mi_row, mi_col + hbs, subsize,
2695                         subsubsize, pc_tree->split[1]);
2696       fill_mode_info_sb(cm, x, mi_row + hbs, mi_col, subsize,
2697                         subsubsize, pc_tree->split[2]);
2698       fill_mode_info_sb(cm, x, mi_row + hbs, mi_col + hbs, subsize,
2699                         subsubsize, pc_tree->split[3]);
2700       break;
2701     }
2702     default:
2703       break;
2704   }
2705 }
2706 
nonrd_pick_partition(VP9_COMP * cpi,const TileInfo * const tile,TOKENEXTRA ** tp,int mi_row,int mi_col,BLOCK_SIZE bsize,int * rate,int64_t * dist,int do_recon,int64_t best_rd,PC_TREE * pc_tree)2707 static void nonrd_pick_partition(VP9_COMP *cpi, const TileInfo *const tile,
2708                                  TOKENEXTRA **tp, int mi_row,
2709                                  int mi_col, BLOCK_SIZE bsize, int *rate,
2710                                  int64_t *dist, int do_recon, int64_t best_rd,
2711                                  PC_TREE *pc_tree) {
2712   const SPEED_FEATURES *const sf = &cpi->sf;
2713   const VP9EncoderConfig *const oxcf = &cpi->oxcf;
2714   VP9_COMMON *const cm = &cpi->common;
2715   MACROBLOCK *const x = &cpi->mb;
2716   MACROBLOCKD *const xd = &x->e_mbd;
2717   const int ms = num_8x8_blocks_wide_lookup[bsize] / 2;
2718   TOKENEXTRA *tp_orig = *tp;
2719   PICK_MODE_CONTEXT *ctx = &pc_tree->none;
2720   int i;
2721   BLOCK_SIZE subsize = bsize;
2722   int this_rate, sum_rate = 0, best_rate = INT_MAX;
2723   int64_t this_dist, sum_dist = 0, best_dist = INT64_MAX;
2724   int64_t sum_rd = 0;
2725   int do_split = bsize >= BLOCK_8X8;
2726   int do_rect = 1;
2727   // Override skipping rectangular partition operations for edge blocks
2728   const int force_horz_split = (mi_row + ms >= cm->mi_rows);
2729   const int force_vert_split = (mi_col + ms >= cm->mi_cols);
2730   const int xss = x->e_mbd.plane[1].subsampling_x;
2731   const int yss = x->e_mbd.plane[1].subsampling_y;
2732 
2733   int partition_none_allowed = !force_horz_split && !force_vert_split;
2734   int partition_horz_allowed = !force_vert_split && yss <= xss &&
2735                                bsize >= BLOCK_8X8;
2736   int partition_vert_allowed = !force_horz_split && xss <= yss &&
2737                                bsize >= BLOCK_8X8;
2738   (void) *tp_orig;
2739 
2740   assert(num_8x8_blocks_wide_lookup[bsize] ==
2741              num_8x8_blocks_high_lookup[bsize]);
2742 
2743   // Determine partition types in search according to the speed features.
2744   // The threshold set here has to be of square block size.
2745   if (sf->auto_min_max_partition_size) {
2746     partition_none_allowed &= (bsize <= sf->max_partition_size &&
2747                                bsize >= sf->min_partition_size);
2748     partition_horz_allowed &= ((bsize <= sf->max_partition_size &&
2749                                 bsize > sf->min_partition_size) ||
2750                                 force_horz_split);
2751     partition_vert_allowed &= ((bsize <= sf->max_partition_size &&
2752                                 bsize > sf->min_partition_size) ||
2753                                 force_vert_split);
2754     do_split &= bsize > sf->min_partition_size;
2755   }
2756   if (sf->use_square_partition_only) {
2757     partition_horz_allowed &= force_horz_split;
2758     partition_vert_allowed &= force_vert_split;
2759   }
2760 
2761   // PARTITION_NONE
2762   if (partition_none_allowed) {
2763     nonrd_pick_sb_modes(cpi, tile, mi_row, mi_col,
2764                         &this_rate, &this_dist, bsize, ctx);
2765     ctx->mic.mbmi = xd->mi[0].src_mi->mbmi;
2766     ctx->skip_txfm[0] = x->skip_txfm[0];
2767     ctx->skip = x->skip;
2768 
2769     if (this_rate != INT_MAX) {
2770       int pl = partition_plane_context(xd, mi_row, mi_col, bsize);
2771       this_rate += cpi->partition_cost[pl][PARTITION_NONE];
2772       sum_rd = RDCOST(x->rdmult, x->rddiv, this_rate, this_dist);
2773       if (sum_rd < best_rd) {
2774         int64_t stop_thresh = 4096;
2775         int64_t stop_thresh_rd;
2776 
2777         best_rate = this_rate;
2778         best_dist = this_dist;
2779         best_rd = sum_rd;
2780         if (bsize >= BLOCK_8X8)
2781           pc_tree->partitioning = PARTITION_NONE;
2782 
2783         // Adjust threshold according to partition size.
2784         stop_thresh >>= 8 - (b_width_log2(bsize) +
2785             b_height_log2(bsize));
2786 
2787         stop_thresh_rd = RDCOST(x->rdmult, x->rddiv, 0, stop_thresh);
2788         // If obtained distortion is very small, choose current partition
2789         // and stop splitting.
2790         if (!x->e_mbd.lossless && best_rd < stop_thresh_rd) {
2791           do_split = 0;
2792           do_rect = 0;
2793         }
2794       }
2795     }
2796   }
2797 
2798   // store estimated motion vector
2799   store_pred_mv(x, ctx);
2800 
2801   // PARTITION_SPLIT
2802   sum_rd = 0;
2803   if (do_split) {
2804     int pl = partition_plane_context(xd, mi_row, mi_col, bsize);
2805     sum_rate += cpi->partition_cost[pl][PARTITION_SPLIT];
2806     subsize = get_subsize(bsize, PARTITION_SPLIT);
2807     for (i = 0; i < 4 && sum_rd < best_rd; ++i) {
2808       const int x_idx = (i & 1) * ms;
2809       const int y_idx = (i >> 1) * ms;
2810 
2811       if (mi_row + y_idx >= cm->mi_rows || mi_col + x_idx >= cm->mi_cols)
2812         continue;
2813       load_pred_mv(x, ctx);
2814       nonrd_pick_partition(cpi, tile, tp, mi_row + y_idx, mi_col + x_idx,
2815                            subsize, &this_rate, &this_dist, 0,
2816                            best_rd - sum_rd, pc_tree->split[i]);
2817 
2818       if (this_rate == INT_MAX) {
2819         sum_rd = INT64_MAX;
2820       } else {
2821         sum_rate += this_rate;
2822         sum_dist += this_dist;
2823         sum_rd = RDCOST(x->rdmult, x->rddiv, sum_rate, sum_dist);
2824       }
2825     }
2826 
2827     if (sum_rd < best_rd) {
2828       best_rate = sum_rate;
2829       best_dist = sum_dist;
2830       best_rd = sum_rd;
2831       pc_tree->partitioning = PARTITION_SPLIT;
2832     } else {
2833       // skip rectangular partition test when larger block size
2834       // gives better rd cost
2835       if (sf->less_rectangular_check)
2836         do_rect &= !partition_none_allowed;
2837     }
2838   }
2839 
2840   // PARTITION_HORZ
2841   if (partition_horz_allowed && do_rect) {
2842     subsize = get_subsize(bsize, PARTITION_HORZ);
2843     if (sf->adaptive_motion_search)
2844       load_pred_mv(x, ctx);
2845 
2846     nonrd_pick_sb_modes(cpi, tile, mi_row, mi_col,
2847                         &this_rate, &this_dist, subsize,
2848                         &pc_tree->horizontal[0]);
2849 
2850     pc_tree->horizontal[0].mic.mbmi = xd->mi[0].src_mi->mbmi;
2851     pc_tree->horizontal[0].skip_txfm[0] = x->skip_txfm[0];
2852     pc_tree->horizontal[0].skip = x->skip;
2853 
2854     sum_rd = RDCOST(x->rdmult, x->rddiv, sum_rate, sum_dist);
2855 
2856     if (sum_rd < best_rd && mi_row + ms < cm->mi_rows) {
2857       load_pred_mv(x, ctx);
2858       nonrd_pick_sb_modes(cpi, tile, mi_row + ms, mi_col,
2859                           &this_rate, &this_dist, subsize,
2860                           &pc_tree->horizontal[1]);
2861 
2862       pc_tree->horizontal[1].mic.mbmi = xd->mi[0].src_mi->mbmi;
2863       pc_tree->horizontal[1].skip_txfm[0] = x->skip_txfm[0];
2864       pc_tree->horizontal[1].skip = x->skip;
2865 
2866       if (this_rate == INT_MAX) {
2867         sum_rd = INT64_MAX;
2868       } else {
2869         int pl = partition_plane_context(xd, mi_row, mi_col, bsize);
2870         this_rate += cpi->partition_cost[pl][PARTITION_HORZ];
2871         sum_rate += this_rate;
2872         sum_dist += this_dist;
2873         sum_rd = RDCOST(x->rdmult, x->rddiv, sum_rate, sum_dist);
2874       }
2875     }
2876     if (sum_rd < best_rd) {
2877       best_rd = sum_rd;
2878       best_rate = sum_rate;
2879       best_dist = sum_dist;
2880       pc_tree->partitioning = PARTITION_HORZ;
2881     }
2882   }
2883 
2884   // PARTITION_VERT
2885   if (partition_vert_allowed && do_rect) {
2886     subsize = get_subsize(bsize, PARTITION_VERT);
2887 
2888     if (sf->adaptive_motion_search)
2889       load_pred_mv(x, ctx);
2890 
2891     nonrd_pick_sb_modes(cpi, tile, mi_row, mi_col,
2892                         &this_rate, &this_dist, subsize,
2893                         &pc_tree->vertical[0]);
2894     pc_tree->vertical[0].mic.mbmi = xd->mi[0].src_mi->mbmi;
2895     pc_tree->vertical[0].skip_txfm[0] = x->skip_txfm[0];
2896     pc_tree->vertical[0].skip = x->skip;
2897     sum_rd = RDCOST(x->rdmult, x->rddiv, sum_rate, sum_dist);
2898     if (sum_rd < best_rd && mi_col + ms < cm->mi_cols) {
2899       load_pred_mv(x, ctx);
2900       nonrd_pick_sb_modes(cpi, tile, mi_row, mi_col + ms,
2901                           &this_rate, &this_dist, subsize,
2902                           &pc_tree->vertical[1]);
2903       pc_tree->vertical[1].mic.mbmi = xd->mi[0].src_mi->mbmi;
2904       pc_tree->vertical[1].skip_txfm[0] = x->skip_txfm[0];
2905       pc_tree->vertical[1].skip = x->skip;
2906       if (this_rate == INT_MAX) {
2907         sum_rd = INT64_MAX;
2908       } else {
2909         int pl = partition_plane_context(xd, mi_row, mi_col, bsize);
2910         this_rate += cpi->partition_cost[pl][PARTITION_VERT];
2911         sum_rate += this_rate;
2912         sum_dist += this_dist;
2913         sum_rd = RDCOST(x->rdmult, x->rddiv, sum_rate, sum_dist);
2914       }
2915     }
2916     if (sum_rd < best_rd) {
2917       best_rate = sum_rate;
2918       best_dist = sum_dist;
2919       best_rd = sum_rd;
2920       pc_tree->partitioning = PARTITION_VERT;
2921     }
2922   }
2923   // TODO(JBB): The following line is here just to avoid a static warning
2924   // that occurs because at this point we never again reuse best_rd
2925   // despite setting it here.  The code should be refactored to avoid this.
2926   (void) best_rd;
2927 
2928   *rate = best_rate;
2929   *dist = best_dist;
2930 
2931   if (best_rate == INT_MAX)
2932     return;
2933 
2934   // update mode info array
2935   subsize = get_subsize(bsize, pc_tree->partitioning);
2936   fill_mode_info_sb(cm, x, mi_row, mi_col, bsize, subsize,
2937                     pc_tree);
2938 
2939   if (best_rate < INT_MAX && best_dist < INT64_MAX && do_recon) {
2940     int output_enabled = (bsize == BLOCK_64X64);
2941 
2942     // Check the projected output rate for this SB against it's target
2943     // and and if necessary apply a Q delta using segmentation to get
2944     // closer to the target.
2945     if ((oxcf->aq_mode == COMPLEXITY_AQ) && cm->seg.update_map) {
2946       vp9_select_in_frame_q_segment(cpi, mi_row, mi_col, output_enabled,
2947                                     best_rate);
2948     }
2949 
2950     if (oxcf->aq_mode == CYCLIC_REFRESH_AQ)
2951       vp9_cyclic_refresh_set_rate_and_dist_sb(cpi->cyclic_refresh,
2952                                               best_rate, best_dist);
2953 
2954     encode_sb_rt(cpi, tile, tp, mi_row, mi_col, output_enabled, bsize, pc_tree);
2955   }
2956 
2957   if (bsize == BLOCK_64X64) {
2958     assert(tp_orig < *tp);
2959     assert(best_rate < INT_MAX);
2960     assert(best_dist < INT64_MAX);
2961   } else {
2962     assert(tp_orig == *tp);
2963   }
2964 }
2965 
nonrd_use_partition(VP9_COMP * cpi,const TileInfo * const tile,MODE_INFO * mi,TOKENEXTRA ** tp,int mi_row,int mi_col,BLOCK_SIZE bsize,int output_enabled,int * totrate,int64_t * totdist,PC_TREE * pc_tree)2966 static void nonrd_use_partition(VP9_COMP *cpi,
2967                                 const TileInfo *const tile,
2968                                 MODE_INFO *mi,
2969                                 TOKENEXTRA **tp,
2970                                 int mi_row, int mi_col,
2971                                 BLOCK_SIZE bsize, int output_enabled,
2972                                 int *totrate, int64_t *totdist,
2973                                 PC_TREE *pc_tree) {
2974   VP9_COMMON *const cm = &cpi->common;
2975   MACROBLOCK *const x = &cpi->mb;
2976   MACROBLOCKD *const xd = &x->e_mbd;
2977   const int bsl = b_width_log2(bsize), hbs = (1 << bsl) / 4;
2978   const int mis = cm->mi_stride;
2979   PARTITION_TYPE partition;
2980   BLOCK_SIZE subsize;
2981   int rate = INT_MAX;
2982   int64_t dist = INT64_MAX;
2983 
2984   if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols)
2985     return;
2986 
2987   subsize = (bsize >= BLOCK_8X8) ? mi[0].src_mi->mbmi.sb_type : BLOCK_4X4;
2988   partition = partition_lookup[bsl][subsize];
2989 
2990   switch (partition) {
2991     case PARTITION_NONE:
2992       nonrd_pick_sb_modes(cpi, tile, mi_row, mi_col, totrate, totdist,
2993                           subsize, &pc_tree->none);
2994       pc_tree->none.mic.mbmi = xd->mi[0].src_mi->mbmi;
2995       pc_tree->none.skip_txfm[0] = x->skip_txfm[0];
2996       pc_tree->none.skip = x->skip;
2997       break;
2998     case PARTITION_VERT:
2999       nonrd_pick_sb_modes(cpi, tile, mi_row, mi_col, totrate, totdist,
3000                           subsize, &pc_tree->vertical[0]);
3001       pc_tree->vertical[0].mic.mbmi = xd->mi[0].src_mi->mbmi;
3002       pc_tree->vertical[0].skip_txfm[0] = x->skip_txfm[0];
3003       pc_tree->vertical[0].skip = x->skip;
3004       if (mi_col + hbs < cm->mi_cols) {
3005         nonrd_pick_sb_modes(cpi, tile, mi_row, mi_col + hbs,
3006                             &rate, &dist, subsize, &pc_tree->vertical[1]);
3007         pc_tree->vertical[1].mic.mbmi = xd->mi[0].src_mi->mbmi;
3008         pc_tree->vertical[1].skip_txfm[0] = x->skip_txfm[0];
3009         pc_tree->vertical[1].skip = x->skip;
3010         if (rate != INT_MAX && dist != INT64_MAX &&
3011             *totrate != INT_MAX && *totdist != INT64_MAX) {
3012           *totrate += rate;
3013           *totdist += dist;
3014         }
3015       }
3016       break;
3017     case PARTITION_HORZ:
3018       nonrd_pick_sb_modes(cpi, tile, mi_row, mi_col, totrate, totdist,
3019                           subsize, &pc_tree->horizontal[0]);
3020       pc_tree->horizontal[0].mic.mbmi = xd->mi[0].src_mi->mbmi;
3021       pc_tree->horizontal[0].skip_txfm[0] = x->skip_txfm[0];
3022       pc_tree->horizontal[0].skip = x->skip;
3023       if (mi_row + hbs < cm->mi_rows) {
3024         nonrd_pick_sb_modes(cpi, tile, mi_row + hbs, mi_col,
3025                             &rate, &dist, subsize, &pc_tree->horizontal[0]);
3026         pc_tree->horizontal[1].mic.mbmi = xd->mi[0].src_mi->mbmi;
3027         pc_tree->horizontal[1].skip_txfm[0] = x->skip_txfm[0];
3028         pc_tree->horizontal[1].skip = x->skip;
3029         if (rate != INT_MAX && dist != INT64_MAX &&
3030             *totrate != INT_MAX && *totdist != INT64_MAX) {
3031           *totrate += rate;
3032           *totdist += dist;
3033         }
3034       }
3035       break;
3036     case PARTITION_SPLIT:
3037       subsize = get_subsize(bsize, PARTITION_SPLIT);
3038       nonrd_use_partition(cpi, tile, mi, tp, mi_row, mi_col,
3039                           subsize, output_enabled, totrate, totdist,
3040                           pc_tree->split[0]);
3041       nonrd_use_partition(cpi, tile, mi + hbs, tp,
3042                           mi_row, mi_col + hbs, subsize, output_enabled,
3043                           &rate, &dist, pc_tree->split[1]);
3044       if (rate != INT_MAX && dist != INT64_MAX &&
3045           *totrate != INT_MAX && *totdist != INT64_MAX) {
3046         *totrate += rate;
3047         *totdist += dist;
3048       }
3049       nonrd_use_partition(cpi, tile, mi + hbs * mis, tp,
3050                           mi_row + hbs, mi_col, subsize, output_enabled,
3051                           &rate, &dist, pc_tree->split[2]);
3052       if (rate != INT_MAX && dist != INT64_MAX &&
3053           *totrate != INT_MAX && *totdist != INT64_MAX) {
3054         *totrate += rate;
3055         *totdist += dist;
3056       }
3057       nonrd_use_partition(cpi, tile, mi + hbs * mis + hbs, tp,
3058                           mi_row + hbs, mi_col + hbs, subsize, output_enabled,
3059                           &rate, &dist, pc_tree->split[3]);
3060       if (rate != INT_MAX && dist != INT64_MAX &&
3061           *totrate != INT_MAX && *totdist != INT64_MAX) {
3062         *totrate += rate;
3063         *totdist += dist;
3064       }
3065       break;
3066     default:
3067       assert("Invalid partition type.");
3068       break;
3069   }
3070 
3071   if (bsize == BLOCK_64X64 && output_enabled) {
3072     if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ)
3073       vp9_cyclic_refresh_set_rate_and_dist_sb(cpi->cyclic_refresh,
3074                                               *totrate, *totdist);
3075     encode_sb_rt(cpi, tile, tp, mi_row, mi_col, 1, bsize, pc_tree);
3076   }
3077 }
3078 
encode_nonrd_sb_row(VP9_COMP * cpi,const TileInfo * const tile,int mi_row,TOKENEXTRA ** tp)3079 static void encode_nonrd_sb_row(VP9_COMP *cpi, const TileInfo *const tile,
3080                                 int mi_row, TOKENEXTRA **tp) {
3081   SPEED_FEATURES *const sf = &cpi->sf;
3082   VP9_COMMON *const cm = &cpi->common;
3083   MACROBLOCK *const x = &cpi->mb;
3084   MACROBLOCKD *const xd = &x->e_mbd;
3085   int mi_col;
3086 
3087   // Initialize the left context for the new SB row
3088   vpx_memset(&xd->left_context, 0, sizeof(xd->left_context));
3089   vpx_memset(xd->left_seg_context, 0, sizeof(xd->left_seg_context));
3090 
3091   // Code each SB in the row
3092   for (mi_col = tile->mi_col_start; mi_col < tile->mi_col_end;
3093        mi_col += MI_BLOCK_SIZE) {
3094     int dummy_rate = 0;
3095     int64_t dummy_dist = 0;
3096     const int idx_str = cm->mi_stride * mi_row + mi_col;
3097     MODE_INFO *mi = cm->mi + idx_str;
3098     MODE_INFO *prev_mi = (cm->prev_mip + cm->mi_stride + 1 + idx_str)->src_mi;
3099     BLOCK_SIZE bsize;
3100     x->in_static_area = 0;
3101     x->source_variance = UINT_MAX;
3102     vp9_zero(x->pred_mv);
3103 
3104     // Set the partition type of the 64X64 block
3105     switch (sf->partition_search_type) {
3106       case VAR_BASED_PARTITION:
3107         choose_partitioning(cpi, tile, mi_row, mi_col);
3108         nonrd_use_partition(cpi, tile, mi, tp, mi_row, mi_col, BLOCK_64X64,
3109                             1, &dummy_rate, &dummy_dist, cpi->pc_root);
3110         break;
3111       case SOURCE_VAR_BASED_PARTITION:
3112         set_source_var_based_partition(cpi, tile, mi, mi_row, mi_col);
3113         nonrd_use_partition(cpi, tile, mi, tp, mi_row, mi_col, BLOCK_64X64,
3114                             1, &dummy_rate, &dummy_dist, cpi->pc_root);
3115         break;
3116       case VAR_BASED_FIXED_PARTITION:
3117       case FIXED_PARTITION:
3118         bsize = sf->partition_search_type == FIXED_PARTITION ?
3119                 sf->always_this_block_size :
3120                 get_nonrd_var_based_fixed_partition(cpi, mi_row, mi_col);
3121         set_fixed_partitioning(cpi, tile, mi, mi_row, mi_col, bsize);
3122         nonrd_use_partition(cpi, tile, mi, tp, mi_row, mi_col, BLOCK_64X64,
3123                             1, &dummy_rate, &dummy_dist, cpi->pc_root);
3124         break;
3125       case REFERENCE_PARTITION:
3126         if (sf->partition_check ||
3127             !(x->in_static_area = is_background(cpi, tile, mi_row, mi_col))) {
3128           set_modeinfo_offsets(cm, xd, mi_row, mi_col);
3129           auto_partition_range(cpi, tile, mi_row, mi_col,
3130                                &sf->min_partition_size,
3131                                &sf->max_partition_size);
3132           nonrd_pick_partition(cpi, tile, tp, mi_row, mi_col, BLOCK_64X64,
3133                                &dummy_rate, &dummy_dist, 1, INT64_MAX,
3134                                cpi->pc_root);
3135         } else {
3136           copy_partitioning(cm, mi, prev_mi);
3137           nonrd_use_partition(cpi, tile, mi, tp, mi_row, mi_col,
3138                               BLOCK_64X64, 1, &dummy_rate, &dummy_dist,
3139                               cpi->pc_root);
3140         }
3141         break;
3142       default:
3143         assert(0);
3144         break;
3145     }
3146   }
3147 }
3148 // end RTC play code
3149 
set_var_thresh_from_histogram(VP9_COMP * cpi)3150 static int set_var_thresh_from_histogram(VP9_COMP *cpi) {
3151   const SPEED_FEATURES *const sf = &cpi->sf;
3152   const VP9_COMMON *const cm = &cpi->common;
3153 
3154   const uint8_t *src = cpi->Source->y_buffer;
3155   const uint8_t *last_src = cpi->Last_Source->y_buffer;
3156   const int src_stride = cpi->Source->y_stride;
3157   const int last_stride = cpi->Last_Source->y_stride;
3158 
3159   // Pick cutoff threshold
3160   const int cutoff = (MIN(cm->width, cm->height) >= 720) ?
3161       (cm->MBs * VAR_HIST_LARGE_CUT_OFF / 100) :
3162       (cm->MBs * VAR_HIST_SMALL_CUT_OFF / 100);
3163   DECLARE_ALIGNED_ARRAY(16, int, hist, VAR_HIST_BINS);
3164   diff *var16 = cpi->source_diff_var;
3165 
3166   int sum = 0;
3167   int i, j;
3168 
3169   vpx_memset(hist, 0, VAR_HIST_BINS * sizeof(hist[0]));
3170 
3171   for (i = 0; i < cm->mb_rows; i++) {
3172     for (j = 0; j < cm->mb_cols; j++) {
3173       vp9_get16x16var(src, src_stride, last_src, last_stride,
3174                       &var16->sse, &var16->sum);
3175 
3176       var16->var = var16->sse -
3177           (((uint32_t)var16->sum * var16->sum) >> 8);
3178 
3179       if (var16->var >= VAR_HIST_MAX_BG_VAR)
3180         hist[VAR_HIST_BINS - 1]++;
3181       else
3182         hist[var16->var / VAR_HIST_FACTOR]++;
3183 
3184       src += 16;
3185       last_src += 16;
3186       var16++;
3187     }
3188 
3189     src = src - cm->mb_cols * 16 + 16 * src_stride;
3190     last_src = last_src - cm->mb_cols * 16 + 16 * last_stride;
3191   }
3192 
3193   cpi->source_var_thresh = 0;
3194 
3195   if (hist[VAR_HIST_BINS - 1] < cutoff) {
3196     for (i = 0; i < VAR_HIST_BINS - 1; i++) {
3197       sum += hist[i];
3198 
3199       if (sum > cutoff) {
3200         cpi->source_var_thresh = (i + 1) * VAR_HIST_FACTOR;
3201         return 0;
3202       }
3203     }
3204   }
3205 
3206   return sf->search_type_check_frequency;
3207 }
3208 
source_var_based_partition_search_method(VP9_COMP * cpi)3209 static void source_var_based_partition_search_method(VP9_COMP *cpi) {
3210   VP9_COMMON *const cm = &cpi->common;
3211   SPEED_FEATURES *const sf = &cpi->sf;
3212 
3213   if (cm->frame_type == KEY_FRAME) {
3214     // For key frame, use SEARCH_PARTITION.
3215     sf->partition_search_type = SEARCH_PARTITION;
3216   } else if (cm->intra_only) {
3217     sf->partition_search_type = FIXED_PARTITION;
3218   } else {
3219     if (cm->last_width != cm->width || cm->last_height != cm->height) {
3220       if (cpi->source_diff_var)
3221         vpx_free(cpi->source_diff_var);
3222 
3223         CHECK_MEM_ERROR(cm, cpi->source_diff_var,
3224                         vpx_calloc(cm->MBs, sizeof(diff)));
3225       }
3226 
3227     if (!cpi->frames_till_next_var_check)
3228       cpi->frames_till_next_var_check = set_var_thresh_from_histogram(cpi);
3229 
3230     if (cpi->frames_till_next_var_check > 0) {
3231       sf->partition_search_type = FIXED_PARTITION;
3232       cpi->frames_till_next_var_check--;
3233     }
3234   }
3235 }
3236 
get_skip_encode_frame(const VP9_COMMON * cm)3237 static int get_skip_encode_frame(const VP9_COMMON *cm) {
3238   unsigned int intra_count = 0, inter_count = 0;
3239   int j;
3240 
3241   for (j = 0; j < INTRA_INTER_CONTEXTS; ++j) {
3242     intra_count += cm->counts.intra_inter[j][0];
3243     inter_count += cm->counts.intra_inter[j][1];
3244   }
3245 
3246   return (intra_count << 2) < inter_count &&
3247          cm->frame_type != KEY_FRAME &&
3248          cm->show_frame;
3249 }
3250 
encode_tiles(VP9_COMP * cpi)3251 static void encode_tiles(VP9_COMP *cpi) {
3252   const VP9_COMMON *const cm = &cpi->common;
3253   const int tile_cols = 1 << cm->log2_tile_cols;
3254   const int tile_rows = 1 << cm->log2_tile_rows;
3255   int tile_col, tile_row;
3256   TOKENEXTRA *tok = cpi->tok;
3257 
3258   for (tile_row = 0; tile_row < tile_rows; ++tile_row) {
3259     for (tile_col = 0; tile_col < tile_cols; ++tile_col) {
3260       TileInfo tile;
3261       TOKENEXTRA *old_tok = tok;
3262       int mi_row;
3263 
3264       vp9_tile_init(&tile, cm, tile_row, tile_col);
3265       for (mi_row = tile.mi_row_start; mi_row < tile.mi_row_end;
3266            mi_row += MI_BLOCK_SIZE) {
3267         if (cpi->sf.use_nonrd_pick_mode && !frame_is_intra_only(cm))
3268           encode_nonrd_sb_row(cpi, &tile, mi_row, &tok);
3269         else
3270           encode_rd_sb_row(cpi, &tile, mi_row, &tok);
3271       }
3272       cpi->tok_count[tile_row][tile_col] = (unsigned int)(tok - old_tok);
3273       assert(tok - cpi->tok <= get_token_alloc(cm->mb_rows, cm->mb_cols));
3274     }
3275   }
3276 }
3277 
3278 #if CONFIG_FP_MB_STATS
input_fpmb_stats(FIRSTPASS_MB_STATS * firstpass_mb_stats,VP9_COMMON * cm,uint8_t ** this_frame_mb_stats)3279 static int input_fpmb_stats(FIRSTPASS_MB_STATS *firstpass_mb_stats,
3280                             VP9_COMMON *cm, uint8_t **this_frame_mb_stats) {
3281   uint8_t *mb_stats_in = firstpass_mb_stats->mb_stats_start +
3282       cm->current_video_frame * cm->MBs * sizeof(uint8_t);
3283 
3284   if (mb_stats_in > firstpass_mb_stats->mb_stats_end)
3285     return EOF;
3286 
3287   *this_frame_mb_stats = mb_stats_in;
3288 
3289   return 1;
3290 }
3291 #endif
3292 
encode_frame_internal(VP9_COMP * cpi)3293 static void encode_frame_internal(VP9_COMP *cpi) {
3294   SPEED_FEATURES *const sf = &cpi->sf;
3295   RD_OPT *const rd_opt = &cpi->rd;
3296   MACROBLOCK *const x = &cpi->mb;
3297   VP9_COMMON *const cm = &cpi->common;
3298   MACROBLOCKD *const xd = &x->e_mbd;
3299 
3300   xd->mi = cm->mi;
3301   xd->mi[0].src_mi = &xd->mi[0];
3302 
3303   vp9_zero(cm->counts);
3304   vp9_zero(cpi->coef_counts);
3305   vp9_zero(rd_opt->comp_pred_diff);
3306   vp9_zero(rd_opt->filter_diff);
3307   vp9_zero(rd_opt->tx_select_diff);
3308   vp9_zero(rd_opt->tx_select_threshes);
3309 
3310   xd->lossless = cm->base_qindex == 0 &&
3311                  cm->y_dc_delta_q == 0 &&
3312                  cm->uv_dc_delta_q == 0 &&
3313                  cm->uv_ac_delta_q == 0;
3314 
3315   cm->tx_mode = select_tx_mode(cpi);
3316 
3317   x->fwd_txm4x4 = xd->lossless ? vp9_fwht4x4 : vp9_fdct4x4;
3318   x->itxm_add = xd->lossless ? vp9_iwht4x4_add : vp9_idct4x4_add;
3319 
3320   if (xd->lossless) {
3321     x->optimize = 0;
3322     cm->lf.filter_level = 0;
3323     cpi->zbin_mode_boost_enabled = 0;
3324   }
3325 
3326   vp9_frame_init_quantizer(cpi);
3327 
3328   vp9_initialize_rd_consts(cpi);
3329   vp9_initialize_me_consts(cpi, cm->base_qindex);
3330   init_encode_frame_mb_context(cpi);
3331   set_prev_mi(cm);
3332 
3333   x->quant_fp = cpi->sf.use_quant_fp;
3334   vp9_zero(x->skip_txfm);
3335   if (sf->use_nonrd_pick_mode) {
3336     // Initialize internal buffer pointers for rtc coding, where non-RD
3337     // mode decision is used and hence no buffer pointer swap needed.
3338     int i;
3339     struct macroblock_plane *const p = x->plane;
3340     struct macroblockd_plane *const pd = xd->plane;
3341     PICK_MODE_CONTEXT *ctx = &cpi->pc_root->none;
3342 
3343     for (i = 0; i < MAX_MB_PLANE; ++i) {
3344       p[i].coeff = ctx->coeff_pbuf[i][0];
3345       p[i].qcoeff = ctx->qcoeff_pbuf[i][0];
3346       pd[i].dqcoeff = ctx->dqcoeff_pbuf[i][0];
3347       p[i].eobs = ctx->eobs_pbuf[i][0];
3348     }
3349     vp9_zero(x->zcoeff_blk);
3350 
3351     if (sf->partition_search_type == SOURCE_VAR_BASED_PARTITION)
3352       source_var_based_partition_search_method(cpi);
3353   }
3354 
3355   {
3356     struct vpx_usec_timer emr_timer;
3357     vpx_usec_timer_start(&emr_timer);
3358 
3359 #if CONFIG_FP_MB_STATS
3360   if (cpi->use_fp_mb_stats) {
3361     input_fpmb_stats(&cpi->twopass.firstpass_mb_stats, cm,
3362                      &cpi->twopass.this_frame_mb_stats);
3363   }
3364 #endif
3365 
3366     encode_tiles(cpi);
3367 
3368     vpx_usec_timer_mark(&emr_timer);
3369     cpi->time_encode_sb_row += vpx_usec_timer_elapsed(&emr_timer);
3370   }
3371 
3372   sf->skip_encode_frame = sf->skip_encode_sb ? get_skip_encode_frame(cm) : 0;
3373 
3374 #if 0
3375   // Keep record of the total distortion this time around for future use
3376   cpi->last_frame_distortion = cpi->frame_distortion;
3377 #endif
3378 }
3379 
get_interp_filter(const int64_t threshes[SWITCHABLE_FILTER_CONTEXTS],int is_alt_ref)3380 static INTERP_FILTER get_interp_filter(
3381     const int64_t threshes[SWITCHABLE_FILTER_CONTEXTS], int is_alt_ref) {
3382   if (!is_alt_ref &&
3383       threshes[EIGHTTAP_SMOOTH] > threshes[EIGHTTAP] &&
3384       threshes[EIGHTTAP_SMOOTH] > threshes[EIGHTTAP_SHARP] &&
3385       threshes[EIGHTTAP_SMOOTH] > threshes[SWITCHABLE - 1]) {
3386     return EIGHTTAP_SMOOTH;
3387   } else if (threshes[EIGHTTAP_SHARP] > threshes[EIGHTTAP] &&
3388              threshes[EIGHTTAP_SHARP] > threshes[SWITCHABLE - 1]) {
3389     return EIGHTTAP_SHARP;
3390   } else if (threshes[EIGHTTAP] > threshes[SWITCHABLE - 1]) {
3391     return EIGHTTAP;
3392   } else {
3393     return SWITCHABLE;
3394   }
3395 }
3396 
vp9_encode_frame(VP9_COMP * cpi)3397 void vp9_encode_frame(VP9_COMP *cpi) {
3398   VP9_COMMON *const cm = &cpi->common;
3399   RD_OPT *const rd_opt = &cpi->rd;
3400 
3401   // In the longer term the encoder should be generalized to match the
3402   // decoder such that we allow compound where one of the 3 buffers has a
3403   // different sign bias and that buffer is then the fixed ref. However, this
3404   // requires further work in the rd loop. For now the only supported encoder
3405   // side behavior is where the ALT ref buffer has opposite sign bias to
3406   // the other two.
3407   if (!frame_is_intra_only(cm)) {
3408     if ((cm->ref_frame_sign_bias[ALTREF_FRAME] ==
3409              cm->ref_frame_sign_bias[GOLDEN_FRAME]) ||
3410         (cm->ref_frame_sign_bias[ALTREF_FRAME] ==
3411              cm->ref_frame_sign_bias[LAST_FRAME])) {
3412       cm->allow_comp_inter_inter = 0;
3413     } else {
3414       cm->allow_comp_inter_inter = 1;
3415       cm->comp_fixed_ref = ALTREF_FRAME;
3416       cm->comp_var_ref[0] = LAST_FRAME;
3417       cm->comp_var_ref[1] = GOLDEN_FRAME;
3418     }
3419   }
3420 
3421   if (cpi->sf.frame_parameter_update) {
3422     int i;
3423 
3424     // This code does a single RD pass over the whole frame assuming
3425     // either compound, single or hybrid prediction as per whatever has
3426     // worked best for that type of frame in the past.
3427     // It also predicts whether another coding mode would have worked
3428     // better that this coding mode. If that is the case, it remembers
3429     // that for subsequent frames.
3430     // It does the same analysis for transform size selection also.
3431     const MV_REFERENCE_FRAME frame_type = get_frame_type(cpi);
3432     int64_t *const mode_thrs = rd_opt->prediction_type_threshes[frame_type];
3433     int64_t *const filter_thrs = rd_opt->filter_threshes[frame_type];
3434     int *const tx_thrs = rd_opt->tx_select_threshes[frame_type];
3435     const int is_alt_ref = frame_type == ALTREF_FRAME;
3436 
3437     /* prediction (compound, single or hybrid) mode selection */
3438     if (is_alt_ref || !cm->allow_comp_inter_inter)
3439       cm->reference_mode = SINGLE_REFERENCE;
3440     else if (mode_thrs[COMPOUND_REFERENCE] > mode_thrs[SINGLE_REFERENCE] &&
3441              mode_thrs[COMPOUND_REFERENCE] >
3442                  mode_thrs[REFERENCE_MODE_SELECT] &&
3443              check_dual_ref_flags(cpi) &&
3444              cpi->static_mb_pct == 100)
3445       cm->reference_mode = COMPOUND_REFERENCE;
3446     else if (mode_thrs[SINGLE_REFERENCE] > mode_thrs[REFERENCE_MODE_SELECT])
3447       cm->reference_mode = SINGLE_REFERENCE;
3448     else
3449       cm->reference_mode = REFERENCE_MODE_SELECT;
3450 
3451     if (cm->interp_filter == SWITCHABLE)
3452       cm->interp_filter = get_interp_filter(filter_thrs, is_alt_ref);
3453 
3454     encode_frame_internal(cpi);
3455 
3456     for (i = 0; i < REFERENCE_MODES; ++i)
3457       mode_thrs[i] = (mode_thrs[i] + rd_opt->comp_pred_diff[i] / cm->MBs) / 2;
3458 
3459     for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i)
3460       filter_thrs[i] = (filter_thrs[i] + rd_opt->filter_diff[i] / cm->MBs) / 2;
3461 
3462     for (i = 0; i < TX_MODES; ++i) {
3463       int64_t pd = rd_opt->tx_select_diff[i];
3464       if (i == TX_MODE_SELECT)
3465         pd -= RDCOST(cpi->mb.rdmult, cpi->mb.rddiv, 2048 * (TX_SIZES - 1), 0);
3466       tx_thrs[i] = (tx_thrs[i] + (int)(pd / cm->MBs)) / 2;
3467     }
3468 
3469     if (cm->reference_mode == REFERENCE_MODE_SELECT) {
3470       int single_count_zero = 0;
3471       int comp_count_zero = 0;
3472 
3473       for (i = 0; i < COMP_INTER_CONTEXTS; i++) {
3474         single_count_zero += cm->counts.comp_inter[i][0];
3475         comp_count_zero += cm->counts.comp_inter[i][1];
3476       }
3477 
3478       if (comp_count_zero == 0) {
3479         cm->reference_mode = SINGLE_REFERENCE;
3480         vp9_zero(cm->counts.comp_inter);
3481       } else if (single_count_zero == 0) {
3482         cm->reference_mode = COMPOUND_REFERENCE;
3483         vp9_zero(cm->counts.comp_inter);
3484       }
3485     }
3486 
3487     if (cm->tx_mode == TX_MODE_SELECT) {
3488       int count4x4 = 0;
3489       int count8x8_lp = 0, count8x8_8x8p = 0;
3490       int count16x16_16x16p = 0, count16x16_lp = 0;
3491       int count32x32 = 0;
3492 
3493       for (i = 0; i < TX_SIZE_CONTEXTS; ++i) {
3494         count4x4 += cm->counts.tx.p32x32[i][TX_4X4];
3495         count4x4 += cm->counts.tx.p16x16[i][TX_4X4];
3496         count4x4 += cm->counts.tx.p8x8[i][TX_4X4];
3497 
3498         count8x8_lp += cm->counts.tx.p32x32[i][TX_8X8];
3499         count8x8_lp += cm->counts.tx.p16x16[i][TX_8X8];
3500         count8x8_8x8p += cm->counts.tx.p8x8[i][TX_8X8];
3501 
3502         count16x16_16x16p += cm->counts.tx.p16x16[i][TX_16X16];
3503         count16x16_lp += cm->counts.tx.p32x32[i][TX_16X16];
3504         count32x32 += cm->counts.tx.p32x32[i][TX_32X32];
3505       }
3506 
3507       if (count4x4 == 0 && count16x16_lp == 0 && count16x16_16x16p == 0 &&
3508           count32x32 == 0) {
3509         cm->tx_mode = ALLOW_8X8;
3510         reset_skip_tx_size(cm, TX_8X8);
3511       } else if (count8x8_8x8p == 0 && count16x16_16x16p == 0 &&
3512                  count8x8_lp == 0 && count16x16_lp == 0 && count32x32 == 0) {
3513         cm->tx_mode = ONLY_4X4;
3514         reset_skip_tx_size(cm, TX_4X4);
3515       } else if (count8x8_lp == 0 && count16x16_lp == 0 && count4x4 == 0) {
3516         cm->tx_mode = ALLOW_32X32;
3517       } else if (count32x32 == 0 && count8x8_lp == 0 && count4x4 == 0) {
3518         cm->tx_mode = ALLOW_16X16;
3519         reset_skip_tx_size(cm, TX_16X16);
3520       }
3521     }
3522   } else {
3523     cm->reference_mode = SINGLE_REFERENCE;
3524     encode_frame_internal(cpi);
3525   }
3526 }
3527 
sum_intra_stats(FRAME_COUNTS * counts,const MODE_INFO * mi)3528 static void sum_intra_stats(FRAME_COUNTS *counts, const MODE_INFO *mi) {
3529   const PREDICTION_MODE y_mode = mi->mbmi.mode;
3530   const PREDICTION_MODE uv_mode = mi->mbmi.uv_mode;
3531   const BLOCK_SIZE bsize = mi->mbmi.sb_type;
3532 
3533   if (bsize < BLOCK_8X8) {
3534     int idx, idy;
3535     const int num_4x4_w = num_4x4_blocks_wide_lookup[bsize];
3536     const int num_4x4_h = num_4x4_blocks_high_lookup[bsize];
3537     for (idy = 0; idy < 2; idy += num_4x4_h)
3538       for (idx = 0; idx < 2; idx += num_4x4_w)
3539         ++counts->y_mode[0][mi->bmi[idy * 2 + idx].as_mode];
3540   } else {
3541     ++counts->y_mode[size_group_lookup[bsize]][y_mode];
3542   }
3543 
3544   ++counts->uv_mode[y_mode][uv_mode];
3545 }
3546 
get_zbin_mode_boost(const MB_MODE_INFO * mbmi,int enabled)3547 static int get_zbin_mode_boost(const MB_MODE_INFO *mbmi, int enabled) {
3548   if (enabled) {
3549     if (is_inter_block(mbmi)) {
3550       if (mbmi->mode == ZEROMV) {
3551         return mbmi->ref_frame[0] != LAST_FRAME ? GF_ZEROMV_ZBIN_BOOST
3552                                                 : LF_ZEROMV_ZBIN_BOOST;
3553       } else {
3554         return mbmi->sb_type < BLOCK_8X8 ? SPLIT_MV_ZBIN_BOOST
3555                                          : MV_ZBIN_BOOST;
3556       }
3557     } else {
3558       return INTRA_ZBIN_BOOST;
3559     }
3560   } else {
3561     return 0;
3562   }
3563 }
3564 
encode_superblock(VP9_COMP * cpi,TOKENEXTRA ** t,int output_enabled,int mi_row,int mi_col,BLOCK_SIZE bsize,PICK_MODE_CONTEXT * ctx)3565 static void encode_superblock(VP9_COMP *cpi, TOKENEXTRA **t, int output_enabled,
3566                               int mi_row, int mi_col, BLOCK_SIZE bsize,
3567                               PICK_MODE_CONTEXT *ctx) {
3568   VP9_COMMON *const cm = &cpi->common;
3569   MACROBLOCK *const x = &cpi->mb;
3570   MACROBLOCKD *const xd = &x->e_mbd;
3571   MODE_INFO *mi_8x8 = xd->mi;
3572   MODE_INFO *mi = mi_8x8;
3573   MB_MODE_INFO *mbmi = &mi->mbmi;
3574   const int seg_skip = vp9_segfeature_active(&cm->seg, mbmi->segment_id,
3575                                              SEG_LVL_SKIP);
3576   const int mis = cm->mi_stride;
3577   const int mi_width = num_8x8_blocks_wide_lookup[bsize];
3578   const int mi_height = num_8x8_blocks_high_lookup[bsize];
3579 
3580   x->skip_recode = !x->select_tx_size && mbmi->sb_type >= BLOCK_8X8 &&
3581                    cpi->oxcf.aq_mode != COMPLEXITY_AQ &&
3582                    cpi->oxcf.aq_mode != CYCLIC_REFRESH_AQ &&
3583                    cpi->sf.allow_skip_recode;
3584 
3585   if (!x->skip_recode && !cpi->sf.use_nonrd_pick_mode)
3586     vpx_memset(x->skip_txfm, 0, sizeof(x->skip_txfm));
3587 
3588   x->skip_optimize = ctx->is_coded;
3589   ctx->is_coded = 1;
3590   x->use_lp32x32fdct = cpi->sf.use_lp32x32fdct;
3591   x->skip_encode = (!output_enabled && cpi->sf.skip_encode_frame &&
3592                     x->q_index < QIDX_SKIP_THRESH);
3593 
3594   if (x->skip_encode)
3595     return;
3596 
3597   set_ref_ptrs(cm, xd, mbmi->ref_frame[0], mbmi->ref_frame[1]);
3598 
3599   // Experimental code. Special case for gf and arf zeromv modes.
3600   // Increase zbin size to suppress noise
3601   cpi->zbin_mode_boost = get_zbin_mode_boost(mbmi,
3602                                              cpi->zbin_mode_boost_enabled);
3603   vp9_update_zbin_extra(cpi, x);
3604 
3605   if (!is_inter_block(mbmi)) {
3606     int plane;
3607     mbmi->skip = 1;
3608     for (plane = 0; plane < MAX_MB_PLANE; ++plane)
3609       vp9_encode_intra_block_plane(x, MAX(bsize, BLOCK_8X8), plane);
3610     if (output_enabled)
3611       sum_intra_stats(&cm->counts, mi);
3612     vp9_tokenize_sb(cpi, t, !output_enabled, MAX(bsize, BLOCK_8X8));
3613   } else {
3614     int ref;
3615     const int is_compound = has_second_ref(mbmi);
3616     for (ref = 0; ref < 1 + is_compound; ++ref) {
3617       YV12_BUFFER_CONFIG *cfg = get_ref_frame_buffer(cpi,
3618                                                      mbmi->ref_frame[ref]);
3619       vp9_setup_pre_planes(xd, ref, cfg, mi_row, mi_col,
3620                            &xd->block_refs[ref]->sf);
3621     }
3622     if (!cpi->sf.reuse_inter_pred_sby || seg_skip)
3623       vp9_build_inter_predictors_sby(xd, mi_row, mi_col, MAX(bsize, BLOCK_8X8));
3624 
3625     vp9_build_inter_predictors_sbuv(xd, mi_row, mi_col, MAX(bsize, BLOCK_8X8));
3626 
3627     vp9_encode_sb(x, MAX(bsize, BLOCK_8X8));
3628     vp9_tokenize_sb(cpi, t, !output_enabled, MAX(bsize, BLOCK_8X8));
3629   }
3630 
3631   if (output_enabled) {
3632     if (cm->tx_mode == TX_MODE_SELECT &&
3633         mbmi->sb_type >= BLOCK_8X8  &&
3634         !(is_inter_block(mbmi) && (mbmi->skip || seg_skip))) {
3635       ++get_tx_counts(max_txsize_lookup[bsize], vp9_get_tx_size_context(xd),
3636                       &cm->counts.tx)[mbmi->tx_size];
3637     } else {
3638       int x, y;
3639       TX_SIZE tx_size;
3640       // The new intra coding scheme requires no change of transform size
3641       if (is_inter_block(&mi->mbmi)) {
3642         tx_size = MIN(tx_mode_to_biggest_tx_size[cm->tx_mode],
3643                       max_txsize_lookup[bsize]);
3644       } else {
3645         tx_size = (bsize >= BLOCK_8X8) ? mbmi->tx_size : TX_4X4;
3646       }
3647 
3648       for (y = 0; y < mi_height; y++)
3649         for (x = 0; x < mi_width; x++)
3650           if (mi_col + x < cm->mi_cols && mi_row + y < cm->mi_rows)
3651             mi_8x8[mis * y + x].src_mi->mbmi.tx_size = tx_size;
3652     }
3653   }
3654 }
3655