• 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 <assert.h>
12 #include <math.h>
13 
14 #include "./vp9_rtcd.h"
15 #include "./vpx_dsp_rtcd.h"
16 
17 #include "vpx_dsp/vpx_dsp_common.h"
18 #include "vpx_mem/vpx_mem.h"
19 #include "vpx_ports/mem.h"
20 #include "vpx_ports/system_state.h"
21 
22 #include "vp9/common/vp9_common.h"
23 #include "vp9/common/vp9_entropy.h"
24 #include "vp9/common/vp9_entropymode.h"
25 #include "vp9/common/vp9_idct.h"
26 #include "vp9/common/vp9_mvref_common.h"
27 #include "vp9/common/vp9_pred_common.h"
28 #include "vp9/common/vp9_quant_common.h"
29 #include "vp9/common/vp9_reconinter.h"
30 #include "vp9/common/vp9_reconintra.h"
31 #include "vp9/common/vp9_scan.h"
32 #include "vp9/common/vp9_seg_common.h"
33 
34 #if !CONFIG_REALTIME_ONLY
35 #include "vp9/encoder/vp9_aq_variance.h"
36 #endif
37 #include "vp9/encoder/vp9_cost.h"
38 #include "vp9/encoder/vp9_encodemb.h"
39 #include "vp9/encoder/vp9_encodemv.h"
40 #include "vp9/encoder/vp9_encoder.h"
41 #include "vp9/encoder/vp9_mcomp.h"
42 #include "vp9/encoder/vp9_quantize.h"
43 #include "vp9/encoder/vp9_ratectrl.h"
44 #include "vp9/encoder/vp9_rd.h"
45 #include "vp9/encoder/vp9_rdopt.h"
46 
47 #define LAST_FRAME_MODE_MASK \
48   ((1 << GOLDEN_FRAME) | (1 << ALTREF_FRAME) | (1 << INTRA_FRAME))
49 #define GOLDEN_FRAME_MODE_MASK \
50   ((1 << LAST_FRAME) | (1 << ALTREF_FRAME) | (1 << INTRA_FRAME))
51 #define ALT_REF_MODE_MASK \
52   ((1 << LAST_FRAME) | (1 << GOLDEN_FRAME) | (1 << INTRA_FRAME))
53 
54 #define SECOND_REF_FRAME_MASK ((1 << ALTREF_FRAME) | 0x01)
55 
56 #define MIN_EARLY_TERM_INDEX 3
57 #define NEW_MV_DISCOUNT_FACTOR 8
58 
59 typedef struct {
60   PREDICTION_MODE mode;
61   MV_REFERENCE_FRAME ref_frame[2];
62 } MODE_DEFINITION;
63 
64 typedef struct {
65   MV_REFERENCE_FRAME ref_frame[2];
66 } REF_DEFINITION;
67 
68 struct rdcost_block_args {
69   const VP9_COMP *cpi;
70   MACROBLOCK *x;
71   ENTROPY_CONTEXT t_above[16];
72   ENTROPY_CONTEXT t_left[16];
73   int this_rate;
74   int64_t this_dist;
75   int64_t this_sse;
76   int64_t this_rd;
77   int64_t best_rd;
78   int exit_early;
79   int use_fast_coef_costing;
80   const scan_order *so;
81   uint8_t skippable;
82   struct buf_2d *this_recon;
83 };
84 
85 #define LAST_NEW_MV_INDEX 6
86 
87 #if !CONFIG_REALTIME_ONLY
88 static const MODE_DEFINITION vp9_mode_order[MAX_MODES] = {
89   { NEARESTMV, { LAST_FRAME, NONE } },
90   { NEARESTMV, { ALTREF_FRAME, NONE } },
91   { NEARESTMV, { GOLDEN_FRAME, NONE } },
92 
93   { DC_PRED, { INTRA_FRAME, NONE } },
94 
95   { NEWMV, { LAST_FRAME, NONE } },
96   { NEWMV, { ALTREF_FRAME, NONE } },
97   { NEWMV, { GOLDEN_FRAME, NONE } },
98 
99   { NEARMV, { LAST_FRAME, NONE } },
100   { NEARMV, { ALTREF_FRAME, NONE } },
101   { NEARMV, { GOLDEN_FRAME, NONE } },
102 
103   { ZEROMV, { LAST_FRAME, NONE } },
104   { ZEROMV, { GOLDEN_FRAME, NONE } },
105   { ZEROMV, { ALTREF_FRAME, NONE } },
106 
107   { NEARESTMV, { LAST_FRAME, ALTREF_FRAME } },
108   { NEARESTMV, { GOLDEN_FRAME, ALTREF_FRAME } },
109 
110   { TM_PRED, { INTRA_FRAME, NONE } },
111 
112   { NEARMV, { LAST_FRAME, ALTREF_FRAME } },
113   { NEWMV, { LAST_FRAME, ALTREF_FRAME } },
114   { NEARMV, { GOLDEN_FRAME, ALTREF_FRAME } },
115   { NEWMV, { GOLDEN_FRAME, ALTREF_FRAME } },
116 
117   { ZEROMV, { LAST_FRAME, ALTREF_FRAME } },
118   { ZEROMV, { GOLDEN_FRAME, ALTREF_FRAME } },
119 
120   { H_PRED, { INTRA_FRAME, NONE } },
121   { V_PRED, { INTRA_FRAME, NONE } },
122   { D135_PRED, { INTRA_FRAME, NONE } },
123   { D207_PRED, { INTRA_FRAME, NONE } },
124   { D153_PRED, { INTRA_FRAME, NONE } },
125   { D63_PRED, { INTRA_FRAME, NONE } },
126   { D117_PRED, { INTRA_FRAME, NONE } },
127   { D45_PRED, { INTRA_FRAME, NONE } },
128 };
129 
130 static const REF_DEFINITION vp9_ref_order[MAX_REFS] = {
131   { { LAST_FRAME, NONE } },           { { GOLDEN_FRAME, NONE } },
132   { { ALTREF_FRAME, NONE } },         { { LAST_FRAME, ALTREF_FRAME } },
133   { { GOLDEN_FRAME, ALTREF_FRAME } }, { { INTRA_FRAME, NONE } },
134 };
135 #endif  // !CONFIG_REALTIME_ONLY
136 
swap_block_ptr(MACROBLOCK * x,PICK_MODE_CONTEXT * ctx,int m,int n,int min_plane,int max_plane)137 static void swap_block_ptr(MACROBLOCK *x, PICK_MODE_CONTEXT *ctx, int m, int n,
138                            int min_plane, int max_plane) {
139   int i;
140 
141   for (i = min_plane; i < max_plane; ++i) {
142     struct macroblock_plane *const p = &x->plane[i];
143     struct macroblockd_plane *const pd = &x->e_mbd.plane[i];
144 
145     p->coeff = ctx->coeff_pbuf[i][m];
146     p->qcoeff = ctx->qcoeff_pbuf[i][m];
147     pd->dqcoeff = ctx->dqcoeff_pbuf[i][m];
148     p->eobs = ctx->eobs_pbuf[i][m];
149 
150     ctx->coeff_pbuf[i][m] = ctx->coeff_pbuf[i][n];
151     ctx->qcoeff_pbuf[i][m] = ctx->qcoeff_pbuf[i][n];
152     ctx->dqcoeff_pbuf[i][m] = ctx->dqcoeff_pbuf[i][n];
153     ctx->eobs_pbuf[i][m] = ctx->eobs_pbuf[i][n];
154 
155     ctx->coeff_pbuf[i][n] = p->coeff;
156     ctx->qcoeff_pbuf[i][n] = p->qcoeff;
157     ctx->dqcoeff_pbuf[i][n] = pd->dqcoeff;
158     ctx->eobs_pbuf[i][n] = p->eobs;
159   }
160 }
161 
162 #if !CONFIG_REALTIME_ONLY
model_rd_for_sb(VP9_COMP * cpi,BLOCK_SIZE bsize,MACROBLOCK * x,MACROBLOCKD * xd,int * out_rate_sum,int64_t * out_dist_sum,int * skip_txfm_sb,int64_t * skip_sse_sb)163 static void model_rd_for_sb(VP9_COMP *cpi, BLOCK_SIZE bsize, MACROBLOCK *x,
164                             MACROBLOCKD *xd, int *out_rate_sum,
165                             int64_t *out_dist_sum, int *skip_txfm_sb,
166                             int64_t *skip_sse_sb) {
167   // Note our transform coeffs are 8 times an orthogonal transform.
168   // Hence quantizer step is also 8 times. To get effective quantizer
169   // we need to divide by 8 before sending to modeling function.
170   int i;
171   int64_t rate_sum = 0;
172   int64_t dist_sum = 0;
173   const int ref = xd->mi[0]->ref_frame[0];
174   unsigned int sse;
175   unsigned int var = 0;
176   int64_t total_sse = 0;
177   int skip_flag = 1;
178   const int shift = 6;
179   int64_t dist;
180   const int dequant_shift =
181 #if CONFIG_VP9_HIGHBITDEPTH
182       (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) ? xd->bd - 5 :
183 #endif  // CONFIG_VP9_HIGHBITDEPTH
184                                                     3;
185   unsigned int qstep_vec[MAX_MB_PLANE];
186   unsigned int nlog2_vec[MAX_MB_PLANE];
187   unsigned int sum_sse_vec[MAX_MB_PLANE];
188   int any_zero_sum_sse = 0;
189 
190   x->pred_sse[ref] = 0;
191 
192   for (i = 0; i < MAX_MB_PLANE; ++i) {
193     struct macroblock_plane *const p = &x->plane[i];
194     struct macroblockd_plane *const pd = &xd->plane[i];
195     const BLOCK_SIZE bs = get_plane_block_size(bsize, pd);
196     const TX_SIZE max_tx_size = max_txsize_lookup[bs];
197     const BLOCK_SIZE unit_size = txsize_to_bsize[max_tx_size];
198     const int64_t dc_thr = p->quant_thred[0] >> shift;
199     const int64_t ac_thr = p->quant_thred[1] >> shift;
200     unsigned int sum_sse = 0;
201     // The low thresholds are used to measure if the prediction errors are
202     // low enough so that we can skip the mode search.
203     const int64_t low_dc_thr = VPXMIN(50, dc_thr >> 2);
204     const int64_t low_ac_thr = VPXMIN(80, ac_thr >> 2);
205     int bw = 1 << (b_width_log2_lookup[bs] - b_width_log2_lookup[unit_size]);
206     int bh = 1 << (b_height_log2_lookup[bs] - b_width_log2_lookup[unit_size]);
207     int idx, idy;
208     int lw = b_width_log2_lookup[unit_size] + 2;
209     int lh = b_height_log2_lookup[unit_size] + 2;
210 
211     for (idy = 0; idy < bh; ++idy) {
212       for (idx = 0; idx < bw; ++idx) {
213         uint8_t *src = p->src.buf + (idy * p->src.stride << lh) + (idx << lw);
214         uint8_t *dst = pd->dst.buf + (idy * pd->dst.stride << lh) + (idx << lh);
215         int block_idx = (idy << 1) + idx;
216         int low_err_skip = 0;
217 
218         var = cpi->fn_ptr[unit_size].vf(src, p->src.stride, dst, pd->dst.stride,
219                                         &sse);
220         x->bsse[(i << 2) + block_idx] = sse;
221         sum_sse += sse;
222 
223         x->skip_txfm[(i << 2) + block_idx] = SKIP_TXFM_NONE;
224         if (!x->select_tx_size) {
225           // Check if all ac coefficients can be quantized to zero.
226           if (var < ac_thr || var == 0) {
227             x->skip_txfm[(i << 2) + block_idx] = SKIP_TXFM_AC_ONLY;
228 
229             // Check if dc coefficient can be quantized to zero.
230             if (sse - var < dc_thr || sse == var) {
231               x->skip_txfm[(i << 2) + block_idx] = SKIP_TXFM_AC_DC;
232 
233               if (!sse || (var < low_ac_thr && sse - var < low_dc_thr))
234                 low_err_skip = 1;
235             }
236           }
237         }
238 
239         if (skip_flag && !low_err_skip) skip_flag = 0;
240 
241         if (i == 0) x->pred_sse[ref] += sse;
242       }
243     }
244 
245     total_sse += sum_sse;
246     sum_sse_vec[i] = sum_sse;
247     any_zero_sum_sse = any_zero_sum_sse || (sum_sse == 0);
248     qstep_vec[i] = pd->dequant[1] >> dequant_shift;
249     nlog2_vec[i] = num_pels_log2_lookup[bs];
250   }
251 
252   // Fast approximate the modelling function.
253   if (cpi->sf.simple_model_rd_from_var) {
254     for (i = 0; i < MAX_MB_PLANE; ++i) {
255       int64_t rate;
256       const int64_t square_error = sum_sse_vec[i];
257       int quantizer = qstep_vec[i];
258 
259       if (quantizer < 120)
260         rate = (square_error * (280 - quantizer)) >> (16 - VP9_PROB_COST_SHIFT);
261       else
262         rate = 0;
263       dist = (square_error * quantizer) >> 8;
264       rate_sum += rate;
265       dist_sum += dist;
266     }
267   } else {
268     if (any_zero_sum_sse) {
269       for (i = 0; i < MAX_MB_PLANE; ++i) {
270         int rate;
271         vp9_model_rd_from_var_lapndz(sum_sse_vec[i], nlog2_vec[i], qstep_vec[i],
272                                      &rate, &dist);
273         rate_sum += rate;
274         dist_sum += dist;
275       }
276     } else {
277       vp9_model_rd_from_var_lapndz_vec(sum_sse_vec, nlog2_vec, qstep_vec,
278                                        &rate_sum, &dist_sum);
279     }
280   }
281 
282   *skip_txfm_sb = skip_flag;
283   *skip_sse_sb = total_sse << VP9_DIST_SCALE_LOG2;
284   *out_rate_sum = (int)rate_sum;
285   *out_dist_sum = dist_sum << VP9_DIST_SCALE_LOG2;
286 }
287 #endif  // !CONFIG_REALTIME_ONLY
288 
289 #if CONFIG_VP9_HIGHBITDEPTH
vp9_highbd_block_error_c(const tran_low_t * coeff,const tran_low_t * dqcoeff,intptr_t block_size,int64_t * ssz,int bd)290 int64_t vp9_highbd_block_error_c(const tran_low_t *coeff,
291                                  const tran_low_t *dqcoeff, intptr_t block_size,
292                                  int64_t *ssz, int bd) {
293   int i;
294   int64_t error = 0, sqcoeff = 0;
295   int shift = 2 * (bd - 8);
296   int rounding = shift > 0 ? 1 << (shift - 1) : 0;
297 
298   for (i = 0; i < block_size; i++) {
299     const int64_t diff = coeff[i] - dqcoeff[i];
300     error += diff * diff;
301     sqcoeff += (int64_t)coeff[i] * (int64_t)coeff[i];
302   }
303   assert(error >= 0 && sqcoeff >= 0);
304   error = (error + rounding) >> shift;
305   sqcoeff = (sqcoeff + rounding) >> shift;
306 
307   *ssz = sqcoeff;
308   return error;
309 }
310 
vp9_highbd_block_error_dispatch(const tran_low_t * coeff,const tran_low_t * dqcoeff,intptr_t block_size,int64_t * ssz,int bd)311 static int64_t vp9_highbd_block_error_dispatch(const tran_low_t *coeff,
312                                                const tran_low_t *dqcoeff,
313                                                intptr_t block_size,
314                                                int64_t *ssz, int bd) {
315   if (bd == 8) {
316     return vp9_block_error(coeff, dqcoeff, block_size, ssz);
317   } else {
318     return vp9_highbd_block_error(coeff, dqcoeff, block_size, ssz, bd);
319   }
320 }
321 #endif  // CONFIG_VP9_HIGHBITDEPTH
322 
vp9_block_error_c(const tran_low_t * coeff,const tran_low_t * dqcoeff,intptr_t block_size,int64_t * ssz)323 int64_t vp9_block_error_c(const tran_low_t *coeff, const tran_low_t *dqcoeff,
324                           intptr_t block_size, int64_t *ssz) {
325   int i;
326   int64_t error = 0, sqcoeff = 0;
327 
328   for (i = 0; i < block_size; i++) {
329     const int diff = coeff[i] - dqcoeff[i];
330     error += diff * diff;
331     sqcoeff += coeff[i] * coeff[i];
332   }
333 
334   *ssz = sqcoeff;
335   return error;
336 }
337 
vp9_block_error_fp_c(const tran_low_t * coeff,const tran_low_t * dqcoeff,int block_size)338 int64_t vp9_block_error_fp_c(const tran_low_t *coeff, const tran_low_t *dqcoeff,
339                              int block_size) {
340   int i;
341   int64_t error = 0;
342 
343   for (i = 0; i < block_size; i++) {
344     const int diff = coeff[i] - dqcoeff[i];
345     error += diff * diff;
346   }
347 
348   return error;
349 }
350 
351 /* The trailing '0' is a terminator which is used inside cost_coeffs() to
352  * decide whether to include cost of a trailing EOB node or not (i.e. we
353  * can skip this if the last coefficient in this transform block, e.g. the
354  * 16th coefficient in a 4x4 block or the 64th coefficient in a 8x8 block,
355  * were non-zero). */
356 static const int16_t band_counts[TX_SIZES][8] = {
357   { 1, 2, 3, 4, 3, 16 - 13, 0 },
358   { 1, 2, 3, 4, 11, 64 - 21, 0 },
359   { 1, 2, 3, 4, 11, 256 - 21, 0 },
360   { 1, 2, 3, 4, 11, 1024 - 21, 0 },
361 };
cost_coeffs(MACROBLOCK * x,int plane,int block,TX_SIZE tx_size,int pt,const int16_t * scan,const int16_t * nb,int use_fast_coef_costing)362 static int cost_coeffs(MACROBLOCK *x, int plane, int block, TX_SIZE tx_size,
363                        int pt, const int16_t *scan, const int16_t *nb,
364                        int use_fast_coef_costing) {
365   MACROBLOCKD *const xd = &x->e_mbd;
366   MODE_INFO *mi = xd->mi[0];
367   const struct macroblock_plane *p = &x->plane[plane];
368   const PLANE_TYPE type = get_plane_type(plane);
369   const int16_t *band_count = &band_counts[tx_size][1];
370   const int eob = p->eobs[block];
371   const tran_low_t *const qcoeff = BLOCK_OFFSET(p->qcoeff, block);
372   unsigned int(*token_costs)[2][COEFF_CONTEXTS][ENTROPY_TOKENS] =
373       x->token_costs[tx_size][type][is_inter_block(mi)];
374   uint8_t token_cache[32 * 32];
375   int cost;
376 #if CONFIG_VP9_HIGHBITDEPTH
377   const uint16_t *cat6_high_cost = vp9_get_high_cost_table(xd->bd);
378 #else
379   const uint16_t *cat6_high_cost = vp9_get_high_cost_table(8);
380 #endif
381 
382   // Check for consistency of tx_size with mode info
383   assert(type == PLANE_TYPE_Y
384              ? mi->tx_size == tx_size
385              : get_uv_tx_size(mi, &xd->plane[plane]) == tx_size);
386 
387   if (eob == 0) {
388     // single eob token
389     cost = token_costs[0][0][pt][EOB_TOKEN];
390   } else {
391     if (use_fast_coef_costing) {
392       int band_left = *band_count++;
393       int c;
394 
395       // dc token
396       int v = qcoeff[0];
397       int16_t prev_t;
398       cost = vp9_get_token_cost(v, &prev_t, cat6_high_cost);
399       cost += (*token_costs)[0][pt][prev_t];
400 
401       token_cache[0] = vp9_pt_energy_class[prev_t];
402       ++token_costs;
403 
404       // ac tokens
405       for (c = 1; c < eob; c++) {
406         const int rc = scan[c];
407         int16_t t;
408 
409         v = qcoeff[rc];
410         cost += vp9_get_token_cost(v, &t, cat6_high_cost);
411         cost += (*token_costs)[!prev_t][!prev_t][t];
412         prev_t = t;
413         if (!--band_left) {
414           band_left = *band_count++;
415           ++token_costs;
416         }
417       }
418 
419       // eob token
420       if (band_left) cost += (*token_costs)[0][!prev_t][EOB_TOKEN];
421 
422     } else {  // !use_fast_coef_costing
423       int band_left = *band_count++;
424       int c;
425 
426       // dc token
427       int v = qcoeff[0];
428       int16_t tok;
429       unsigned int(*tok_cost_ptr)[COEFF_CONTEXTS][ENTROPY_TOKENS];
430       cost = vp9_get_token_cost(v, &tok, cat6_high_cost);
431       cost += (*token_costs)[0][pt][tok];
432 
433       token_cache[0] = vp9_pt_energy_class[tok];
434       ++token_costs;
435 
436       tok_cost_ptr = &((*token_costs)[!tok]);
437 
438       // ac tokens
439       for (c = 1; c < eob; c++) {
440         const int rc = scan[c];
441 
442         v = qcoeff[rc];
443         cost += vp9_get_token_cost(v, &tok, cat6_high_cost);
444         pt = get_coef_context(nb, token_cache, c);
445         cost += (*tok_cost_ptr)[pt][tok];
446         token_cache[rc] = vp9_pt_energy_class[tok];
447         if (!--band_left) {
448           band_left = *band_count++;
449           ++token_costs;
450         }
451         tok_cost_ptr = &((*token_costs)[!tok]);
452       }
453 
454       // eob token
455       if (band_left) {
456         pt = get_coef_context(nb, token_cache, c);
457         cost += (*token_costs)[0][pt][EOB_TOKEN];
458       }
459     }
460   }
461 
462   return cost;
463 }
464 
num_4x4_to_edge(int plane_4x4_dim,int mb_to_edge_dim,int subsampling_dim,int blk_dim)465 static INLINE int num_4x4_to_edge(int plane_4x4_dim, int mb_to_edge_dim,
466                                   int subsampling_dim, int blk_dim) {
467   return plane_4x4_dim + (mb_to_edge_dim >> (5 + subsampling_dim)) - blk_dim;
468 }
469 
470 // Copy all visible 4x4s in the transform block.
copy_block_visible(const MACROBLOCKD * xd,const struct macroblockd_plane * const pd,const uint8_t * src,const int src_stride,uint8_t * dst,const int dst_stride,int blk_row,int blk_col,const BLOCK_SIZE plane_bsize,const BLOCK_SIZE tx_bsize)471 static void copy_block_visible(const MACROBLOCKD *xd,
472                                const struct macroblockd_plane *const pd,
473                                const uint8_t *src, const int src_stride,
474                                uint8_t *dst, const int dst_stride, int blk_row,
475                                int blk_col, const BLOCK_SIZE plane_bsize,
476                                const BLOCK_SIZE tx_bsize) {
477   const int plane_4x4_w = num_4x4_blocks_wide_lookup[plane_bsize];
478   const int plane_4x4_h = num_4x4_blocks_high_lookup[plane_bsize];
479   const int tx_4x4_w = num_4x4_blocks_wide_lookup[tx_bsize];
480   const int tx_4x4_h = num_4x4_blocks_high_lookup[tx_bsize];
481   int b4x4s_to_right_edge = num_4x4_to_edge(plane_4x4_w, xd->mb_to_right_edge,
482                                             pd->subsampling_x, blk_col);
483   int b4x4s_to_bottom_edge = num_4x4_to_edge(plane_4x4_h, xd->mb_to_bottom_edge,
484                                              pd->subsampling_y, blk_row);
485   const int is_highbd = xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH;
486   if (tx_bsize == BLOCK_4X4 ||
487       (b4x4s_to_right_edge >= tx_4x4_w && b4x4s_to_bottom_edge >= tx_4x4_h)) {
488     const int w = tx_4x4_w << 2;
489     const int h = tx_4x4_h << 2;
490 #if CONFIG_VP9_HIGHBITDEPTH
491     if (is_highbd) {
492       vpx_highbd_convolve_copy(CONVERT_TO_SHORTPTR(src), src_stride,
493                                CONVERT_TO_SHORTPTR(dst), dst_stride, NULL, 0, 0,
494                                0, 0, w, h, xd->bd);
495     } else {
496 #endif
497       vpx_convolve_copy(src, src_stride, dst, dst_stride, NULL, 0, 0, 0, 0, w,
498                         h);
499 #if CONFIG_VP9_HIGHBITDEPTH
500     }
501 #endif
502   } else {
503     int r, c;
504     int max_r = VPXMIN(b4x4s_to_bottom_edge, tx_4x4_h);
505     int max_c = VPXMIN(b4x4s_to_right_edge, tx_4x4_w);
506     // if we are in the unrestricted motion border.
507     for (r = 0; r < max_r; ++r) {
508       // Skip visiting the sub blocks that are wholly within the UMV.
509       for (c = 0; c < max_c; ++c) {
510         const uint8_t *src_ptr = src + r * src_stride * 4 + c * 4;
511         uint8_t *dst_ptr = dst + r * dst_stride * 4 + c * 4;
512 #if CONFIG_VP9_HIGHBITDEPTH
513         if (is_highbd) {
514           vpx_highbd_convolve_copy(CONVERT_TO_SHORTPTR(src_ptr), src_stride,
515                                    CONVERT_TO_SHORTPTR(dst_ptr), dst_stride,
516                                    NULL, 0, 0, 0, 0, 4, 4, xd->bd);
517         } else {
518 #endif
519           vpx_convolve_copy(src_ptr, src_stride, dst_ptr, dst_stride, NULL, 0,
520                             0, 0, 0, 4, 4);
521 #if CONFIG_VP9_HIGHBITDEPTH
522         }
523 #endif
524       }
525     }
526   }
527   (void)is_highbd;
528 }
529 
530 // Compute the pixel domain sum square error on all visible 4x4s in the
531 // transform block.
pixel_sse(const VP9_COMP * const cpi,const MACROBLOCKD * xd,const struct macroblockd_plane * const pd,const uint8_t * src,const int src_stride,const uint8_t * dst,const int dst_stride,int blk_row,int blk_col,const BLOCK_SIZE plane_bsize,const BLOCK_SIZE tx_bsize)532 static unsigned pixel_sse(const VP9_COMP *const cpi, const MACROBLOCKD *xd,
533                           const struct macroblockd_plane *const pd,
534                           const uint8_t *src, const int src_stride,
535                           const uint8_t *dst, const int dst_stride, int blk_row,
536                           int blk_col, const BLOCK_SIZE plane_bsize,
537                           const BLOCK_SIZE tx_bsize) {
538   unsigned int sse = 0;
539   const int plane_4x4_w = num_4x4_blocks_wide_lookup[plane_bsize];
540   const int plane_4x4_h = num_4x4_blocks_high_lookup[plane_bsize];
541   const int tx_4x4_w = num_4x4_blocks_wide_lookup[tx_bsize];
542   const int tx_4x4_h = num_4x4_blocks_high_lookup[tx_bsize];
543   int b4x4s_to_right_edge = num_4x4_to_edge(plane_4x4_w, xd->mb_to_right_edge,
544                                             pd->subsampling_x, blk_col);
545   int b4x4s_to_bottom_edge = num_4x4_to_edge(plane_4x4_h, xd->mb_to_bottom_edge,
546                                              pd->subsampling_y, blk_row);
547   if (tx_bsize == BLOCK_4X4 ||
548       (b4x4s_to_right_edge >= tx_4x4_w && b4x4s_to_bottom_edge >= tx_4x4_h)) {
549     cpi->fn_ptr[tx_bsize].vf(src, src_stride, dst, dst_stride, &sse);
550   } else {
551     const vpx_variance_fn_t vf_4x4 = cpi->fn_ptr[BLOCK_4X4].vf;
552     int r, c;
553     unsigned this_sse = 0;
554     int max_r = VPXMIN(b4x4s_to_bottom_edge, tx_4x4_h);
555     int max_c = VPXMIN(b4x4s_to_right_edge, tx_4x4_w);
556     sse = 0;
557     // if we are in the unrestricted motion border.
558     for (r = 0; r < max_r; ++r) {
559       // Skip visiting the sub blocks that are wholly within the UMV.
560       for (c = 0; c < max_c; ++c) {
561         vf_4x4(src + r * src_stride * 4 + c * 4, src_stride,
562                dst + r * dst_stride * 4 + c * 4, dst_stride, &this_sse);
563         sse += this_sse;
564       }
565     }
566   }
567   return sse;
568 }
569 
570 // Compute the squares sum squares on all visible 4x4s in the transform block.
sum_squares_visible(const MACROBLOCKD * xd,const struct macroblockd_plane * const pd,const int16_t * diff,const int diff_stride,int blk_row,int blk_col,const BLOCK_SIZE plane_bsize,const BLOCK_SIZE tx_bsize)571 static int64_t sum_squares_visible(const MACROBLOCKD *xd,
572                                    const struct macroblockd_plane *const pd,
573                                    const int16_t *diff, const int diff_stride,
574                                    int blk_row, int blk_col,
575                                    const BLOCK_SIZE plane_bsize,
576                                    const BLOCK_SIZE tx_bsize) {
577   int64_t sse;
578   const int plane_4x4_w = num_4x4_blocks_wide_lookup[plane_bsize];
579   const int plane_4x4_h = num_4x4_blocks_high_lookup[plane_bsize];
580   const int tx_4x4_w = num_4x4_blocks_wide_lookup[tx_bsize];
581   const int tx_4x4_h = num_4x4_blocks_high_lookup[tx_bsize];
582   int b4x4s_to_right_edge = num_4x4_to_edge(plane_4x4_w, xd->mb_to_right_edge,
583                                             pd->subsampling_x, blk_col);
584   int b4x4s_to_bottom_edge = num_4x4_to_edge(plane_4x4_h, xd->mb_to_bottom_edge,
585                                              pd->subsampling_y, blk_row);
586   if (tx_bsize == BLOCK_4X4 ||
587       (b4x4s_to_right_edge >= tx_4x4_w && b4x4s_to_bottom_edge >= tx_4x4_h)) {
588     assert(tx_4x4_w == tx_4x4_h);
589     sse = (int64_t)vpx_sum_squares_2d_i16(diff, diff_stride, tx_4x4_w << 2);
590   } else {
591     int r, c;
592     int max_r = VPXMIN(b4x4s_to_bottom_edge, tx_4x4_h);
593     int max_c = VPXMIN(b4x4s_to_right_edge, tx_4x4_w);
594     sse = 0;
595     // if we are in the unrestricted motion border.
596     for (r = 0; r < max_r; ++r) {
597       // Skip visiting the sub blocks that are wholly within the UMV.
598       for (c = 0; c < max_c; ++c) {
599         sse += (int64_t)vpx_sum_squares_2d_i16(
600             diff + r * diff_stride * 4 + c * 4, diff_stride, 4);
601       }
602     }
603   }
604   return sse;
605 }
606 
dist_block(const VP9_COMP * cpi,MACROBLOCK * x,int plane,BLOCK_SIZE plane_bsize,int block,int blk_row,int blk_col,TX_SIZE tx_size,int64_t * out_dist,int64_t * out_sse,struct buf_2d * out_recon)607 static void dist_block(const VP9_COMP *cpi, MACROBLOCK *x, int plane,
608                        BLOCK_SIZE plane_bsize, int block, int blk_row,
609                        int blk_col, TX_SIZE tx_size, int64_t *out_dist,
610                        int64_t *out_sse, struct buf_2d *out_recon) {
611   MACROBLOCKD *const xd = &x->e_mbd;
612   const struct macroblock_plane *const p = &x->plane[plane];
613   const struct macroblockd_plane *const pd = &xd->plane[plane];
614   const int eob = p->eobs[block];
615 
616   if (!out_recon && x->block_tx_domain && eob) {
617     const int ss_txfrm_size = tx_size << 1;
618     int64_t this_sse;
619     const int shift = tx_size == TX_32X32 ? 0 : 2;
620     const tran_low_t *const coeff = BLOCK_OFFSET(p->coeff, block);
621     const tran_low_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
622 #if CONFIG_VP9_HIGHBITDEPTH
623     const int bd = (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) ? xd->bd : 8;
624     *out_dist = vp9_highbd_block_error_dispatch(
625                     coeff, dqcoeff, 16 << ss_txfrm_size, &this_sse, bd) >>
626                 shift;
627 #else
628     *out_dist =
629         vp9_block_error(coeff, dqcoeff, 16 << ss_txfrm_size, &this_sse) >>
630         shift;
631 #endif  // CONFIG_VP9_HIGHBITDEPTH
632     *out_sse = this_sse >> shift;
633 
634     if (x->skip_encode && !is_inter_block(xd->mi[0])) {
635       // TODO(jingning): tune the model to better capture the distortion.
636       const int64_t p =
637           (pd->dequant[1] * pd->dequant[1] * (1 << ss_txfrm_size)) >>
638 #if CONFIG_VP9_HIGHBITDEPTH
639           (shift + 2 + (bd - 8) * 2);
640 #else
641           (shift + 2);
642 #endif  // CONFIG_VP9_HIGHBITDEPTH
643       *out_dist += (p >> 4);
644       *out_sse += p;
645     }
646   } else {
647     const BLOCK_SIZE tx_bsize = txsize_to_bsize[tx_size];
648     const int bs = 4 * num_4x4_blocks_wide_lookup[tx_bsize];
649     const int src_stride = p->src.stride;
650     const int dst_stride = pd->dst.stride;
651     const int src_idx = 4 * (blk_row * src_stride + blk_col);
652     const int dst_idx = 4 * (blk_row * dst_stride + blk_col);
653     const uint8_t *src = &p->src.buf[src_idx];
654     const uint8_t *dst = &pd->dst.buf[dst_idx];
655     uint8_t *out_recon_ptr = 0;
656 
657     const tran_low_t *dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
658     unsigned int tmp;
659 
660     tmp = pixel_sse(cpi, xd, pd, src, src_stride, dst, dst_stride, blk_row,
661                     blk_col, plane_bsize, tx_bsize);
662     *out_sse = (int64_t)tmp * 16;
663     if (out_recon) {
664       const int out_recon_idx = 4 * (blk_row * out_recon->stride + blk_col);
665       out_recon_ptr = &out_recon->buf[out_recon_idx];
666       copy_block_visible(xd, pd, dst, dst_stride, out_recon_ptr,
667                          out_recon->stride, blk_row, blk_col, plane_bsize,
668                          tx_bsize);
669     }
670 
671     if (eob) {
672 #if CONFIG_VP9_HIGHBITDEPTH
673       DECLARE_ALIGNED(16, uint16_t, recon16[1024]);
674       uint8_t *recon = (uint8_t *)recon16;
675 #else
676       DECLARE_ALIGNED(16, uint8_t, recon[1024]);
677 #endif  // CONFIG_VP9_HIGHBITDEPTH
678 
679 #if CONFIG_VP9_HIGHBITDEPTH
680       if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
681         vpx_highbd_convolve_copy(CONVERT_TO_SHORTPTR(dst), dst_stride, recon16,
682                                  32, NULL, 0, 0, 0, 0, bs, bs, xd->bd);
683         if (xd->lossless) {
684           vp9_highbd_iwht4x4_add(dqcoeff, recon16, 32, eob, xd->bd);
685         } else {
686           switch (tx_size) {
687             case TX_4X4:
688               vp9_highbd_idct4x4_add(dqcoeff, recon16, 32, eob, xd->bd);
689               break;
690             case TX_8X8:
691               vp9_highbd_idct8x8_add(dqcoeff, recon16, 32, eob, xd->bd);
692               break;
693             case TX_16X16:
694               vp9_highbd_idct16x16_add(dqcoeff, recon16, 32, eob, xd->bd);
695               break;
696             default:
697               assert(tx_size == TX_32X32);
698               vp9_highbd_idct32x32_add(dqcoeff, recon16, 32, eob, xd->bd);
699               break;
700           }
701         }
702         recon = CONVERT_TO_BYTEPTR(recon16);
703       } else {
704 #endif  // CONFIG_VP9_HIGHBITDEPTH
705         vpx_convolve_copy(dst, dst_stride, recon, 32, NULL, 0, 0, 0, 0, bs, bs);
706         switch (tx_size) {
707           case TX_32X32: vp9_idct32x32_add(dqcoeff, recon, 32, eob); break;
708           case TX_16X16: vp9_idct16x16_add(dqcoeff, recon, 32, eob); break;
709           case TX_8X8: vp9_idct8x8_add(dqcoeff, recon, 32, eob); break;
710           default:
711             assert(tx_size == TX_4X4);
712             // this is like vp9_short_idct4x4 but has a special case around
713             // eob<=1, which is significant (not just an optimization) for
714             // the lossless case.
715             x->inv_txfm_add(dqcoeff, recon, 32, eob);
716             break;
717         }
718 #if CONFIG_VP9_HIGHBITDEPTH
719       }
720 #endif  // CONFIG_VP9_HIGHBITDEPTH
721 
722       tmp = pixel_sse(cpi, xd, pd, src, src_stride, recon, 32, blk_row, blk_col,
723                       plane_bsize, tx_bsize);
724       if (out_recon) {
725         copy_block_visible(xd, pd, recon, 32, out_recon_ptr, out_recon->stride,
726                            blk_row, blk_col, plane_bsize, tx_bsize);
727       }
728     }
729 
730     *out_dist = (int64_t)tmp * 16;
731   }
732 }
733 
rate_block(int plane,int block,TX_SIZE tx_size,int coeff_ctx,struct rdcost_block_args * args)734 static int rate_block(int plane, int block, TX_SIZE tx_size, int coeff_ctx,
735                       struct rdcost_block_args *args) {
736   return cost_coeffs(args->x, plane, block, tx_size, coeff_ctx, args->so->scan,
737                      args->so->neighbors, args->use_fast_coef_costing);
738 }
739 
block_rd_txfm(int plane,int block,int blk_row,int blk_col,BLOCK_SIZE plane_bsize,TX_SIZE tx_size,void * arg)740 static void block_rd_txfm(int plane, int block, int blk_row, int blk_col,
741                           BLOCK_SIZE plane_bsize, TX_SIZE tx_size, void *arg) {
742   struct rdcost_block_args *args = arg;
743   MACROBLOCK *const x = args->x;
744   MACROBLOCKD *const xd = &x->e_mbd;
745   MODE_INFO *const mi = xd->mi[0];
746   int64_t rd1, rd2, rd;
747   int rate;
748   int64_t dist = INT64_MAX;
749   int64_t sse = INT64_MAX;
750   const int coeff_ctx =
751       combine_entropy_contexts(args->t_left[blk_row], args->t_above[blk_col]);
752   struct buf_2d *recon = args->this_recon;
753   const BLOCK_SIZE tx_bsize = txsize_to_bsize[tx_size];
754   const struct macroblockd_plane *const pd = &xd->plane[plane];
755   const int dst_stride = pd->dst.stride;
756   const uint8_t *dst = &pd->dst.buf[4 * (blk_row * dst_stride + blk_col)];
757 
758   if (args->exit_early) return;
759 
760   if (!is_inter_block(mi)) {
761 #if CONFIG_MISMATCH_DEBUG
762     struct encode_b_args intra_arg = {
763       x, x->block_qcoeff_opt, args->t_above, args->t_left, &mi->skip, 0, 0, 0
764     };
765 #else
766     struct encode_b_args intra_arg = { x, x->block_qcoeff_opt, args->t_above,
767                                        args->t_left, &mi->skip };
768 #endif
769     vp9_encode_block_intra(plane, block, blk_row, blk_col, plane_bsize, tx_size,
770                            &intra_arg);
771     if (recon) {
772       uint8_t *rec_ptr = &recon->buf[4 * (blk_row * recon->stride + blk_col)];
773       copy_block_visible(xd, pd, dst, dst_stride, rec_ptr, recon->stride,
774                          blk_row, blk_col, plane_bsize, tx_bsize);
775     }
776     if (x->block_tx_domain) {
777       dist_block(args->cpi, x, plane, plane_bsize, block, blk_row, blk_col,
778                  tx_size, &dist, &sse, /*recon =*/0);
779     } else {
780       const struct macroblock_plane *const p = &x->plane[plane];
781       const int src_stride = p->src.stride;
782       const int diff_stride = 4 * num_4x4_blocks_wide_lookup[plane_bsize];
783       const uint8_t *src = &p->src.buf[4 * (blk_row * src_stride + blk_col)];
784       const int16_t *diff = &p->src_diff[4 * (blk_row * diff_stride + blk_col)];
785       unsigned int tmp;
786       sse = sum_squares_visible(xd, pd, diff, diff_stride, blk_row, blk_col,
787                                 plane_bsize, tx_bsize);
788 #if CONFIG_VP9_HIGHBITDEPTH
789       if ((xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) && (xd->bd > 8))
790         sse = ROUND64_POWER_OF_TWO(sse, (xd->bd - 8) * 2);
791 #endif  // CONFIG_VP9_HIGHBITDEPTH
792       sse = sse * 16;
793       tmp = pixel_sse(args->cpi, xd, pd, src, src_stride, dst, dst_stride,
794                       blk_row, blk_col, plane_bsize, tx_bsize);
795       dist = (int64_t)tmp * 16;
796     }
797   } else {
798     int skip_txfm_flag = SKIP_TXFM_NONE;
799     if (max_txsize_lookup[plane_bsize] == tx_size)
800       skip_txfm_flag = x->skip_txfm[(plane << 2) + (block >> (tx_size << 1))];
801 
802     // This reduces the risk of bad perceptual quality due to bad prediction.
803     // We always force the encoder to perform transform and quantization.
804     if (!args->cpi->sf.allow_skip_txfm_ac_dc &&
805         skip_txfm_flag == SKIP_TXFM_AC_DC) {
806       skip_txfm_flag = SKIP_TXFM_NONE;
807     }
808 
809     if (skip_txfm_flag == SKIP_TXFM_NONE ||
810         (recon && skip_txfm_flag == SKIP_TXFM_AC_ONLY)) {
811       // full forward transform and quantization
812       vp9_xform_quant(x, plane, block, blk_row, blk_col, plane_bsize, tx_size);
813       if (x->block_qcoeff_opt)
814         vp9_optimize_b(x, plane, block, tx_size, coeff_ctx);
815       dist_block(args->cpi, x, plane, plane_bsize, block, blk_row, blk_col,
816                  tx_size, &dist, &sse, recon);
817     } else if (skip_txfm_flag == SKIP_TXFM_AC_ONLY) {
818       // compute DC coefficient
819       tran_low_t *const coeff = BLOCK_OFFSET(x->plane[plane].coeff, block);
820       tran_low_t *const dqcoeff = BLOCK_OFFSET(xd->plane[plane].dqcoeff, block);
821       vp9_xform_quant_dc(x, plane, block, blk_row, blk_col, plane_bsize,
822                          tx_size);
823       sse = x->bsse[(plane << 2) + (block >> (tx_size << 1))] << 4;
824       dist = sse;
825       if (x->plane[plane].eobs[block]) {
826         const int64_t orig_sse = (int64_t)coeff[0] * coeff[0];
827         const int64_t resd_sse = coeff[0] - dqcoeff[0];
828         int64_t dc_correct = orig_sse - resd_sse * resd_sse;
829 #if CONFIG_VP9_HIGHBITDEPTH
830         dc_correct >>= ((xd->bd - 8) * 2);
831 #endif
832         if (tx_size != TX_32X32) dc_correct >>= 2;
833 
834         dist = VPXMAX(0, sse - dc_correct);
835       }
836     } else {
837       assert(0 && "allow_skip_txfm_ac_dc does not allow SKIP_TXFM_AC_DC.");
838     }
839   }
840 
841   rd = RDCOST(x->rdmult, x->rddiv, 0, dist);
842   if (args->this_rd + rd > args->best_rd) {
843     args->exit_early = 1;
844     return;
845   }
846 
847   rate = rate_block(plane, block, tx_size, coeff_ctx, args);
848   args->t_above[blk_col] = (x->plane[plane].eobs[block] > 0) ? 1 : 0;
849   args->t_left[blk_row] = (x->plane[plane].eobs[block] > 0) ? 1 : 0;
850   rd1 = RDCOST(x->rdmult, x->rddiv, rate, dist);
851   rd2 = RDCOST(x->rdmult, x->rddiv, 0, sse);
852 
853   // TODO(jingning): temporarily enabled only for luma component
854   rd = VPXMIN(rd1, rd2);
855   if (plane == 0) {
856     x->zcoeff_blk[tx_size][block] =
857         !x->plane[plane].eobs[block] ||
858         (x->sharpness == 0 && rd1 > rd2 && !xd->lossless);
859     x->sum_y_eobs[tx_size] += x->plane[plane].eobs[block];
860   }
861 
862   args->this_rate += rate;
863   args->this_dist += dist;
864   args->this_sse += sse;
865   args->this_rd += rd;
866 
867   if (args->this_rd > args->best_rd) {
868     args->exit_early = 1;
869     return;
870   }
871 
872   args->skippable &= !x->plane[plane].eobs[block];
873 }
874 
txfm_rd_in_plane(const VP9_COMP * cpi,MACROBLOCK * x,int * rate,int64_t * distortion,int * skippable,int64_t * sse,int64_t ref_best_rd,int plane,BLOCK_SIZE bsize,TX_SIZE tx_size,int use_fast_coef_costing,struct buf_2d * recon)875 static void txfm_rd_in_plane(const VP9_COMP *cpi, MACROBLOCK *x, int *rate,
876                              int64_t *distortion, int *skippable, int64_t *sse,
877                              int64_t ref_best_rd, int plane, BLOCK_SIZE bsize,
878                              TX_SIZE tx_size, int use_fast_coef_costing,
879                              struct buf_2d *recon) {
880   MACROBLOCKD *const xd = &x->e_mbd;
881   const struct macroblockd_plane *const pd = &xd->plane[plane];
882   struct rdcost_block_args args;
883   vp9_zero(args);
884   args.cpi = cpi;
885   args.x = x;
886   args.best_rd = ref_best_rd;
887   args.use_fast_coef_costing = use_fast_coef_costing;
888   args.skippable = 1;
889   args.this_recon = recon;
890 
891   if (plane == 0) xd->mi[0]->tx_size = tx_size;
892 
893   vp9_get_entropy_contexts(bsize, tx_size, pd, args.t_above, args.t_left);
894 
895   args.so = get_scan(xd, tx_size, get_plane_type(plane), 0);
896 
897   vp9_foreach_transformed_block_in_plane(xd, bsize, plane, block_rd_txfm,
898                                          &args);
899   if (args.exit_early) {
900     *rate = INT_MAX;
901     *distortion = INT64_MAX;
902     *sse = INT64_MAX;
903     *skippable = 0;
904   } else {
905     *distortion = args.this_dist;
906     *rate = args.this_rate;
907     *sse = args.this_sse;
908     *skippable = args.skippable;
909   }
910 }
911 
choose_largest_tx_size(VP9_COMP * cpi,MACROBLOCK * x,int * rate,int64_t * distortion,int * skip,int64_t * sse,int64_t ref_best_rd,BLOCK_SIZE bs,struct buf_2d * recon)912 static void choose_largest_tx_size(VP9_COMP *cpi, MACROBLOCK *x, int *rate,
913                                    int64_t *distortion, int *skip, int64_t *sse,
914                                    int64_t ref_best_rd, BLOCK_SIZE bs,
915                                    struct buf_2d *recon) {
916   const TX_SIZE max_tx_size = max_txsize_lookup[bs];
917   VP9_COMMON *const cm = &cpi->common;
918   const TX_SIZE largest_tx_size = tx_mode_to_biggest_tx_size[cm->tx_mode];
919   MACROBLOCKD *const xd = &x->e_mbd;
920   MODE_INFO *const mi = xd->mi[0];
921 
922   mi->tx_size = VPXMIN(max_tx_size, largest_tx_size);
923 
924   txfm_rd_in_plane(cpi, x, rate, distortion, skip, sse, ref_best_rd, 0, bs,
925                    mi->tx_size, cpi->sf.use_fast_coef_costing, recon);
926 }
927 
choose_tx_size_from_rd(VP9_COMP * cpi,MACROBLOCK * x,int * rate,int64_t * distortion,int * skip,int64_t * psse,int64_t ref_best_rd,BLOCK_SIZE bs,struct buf_2d * recon)928 static void choose_tx_size_from_rd(VP9_COMP *cpi, MACROBLOCK *x, int *rate,
929                                    int64_t *distortion, int *skip,
930                                    int64_t *psse, int64_t ref_best_rd,
931                                    BLOCK_SIZE bs, struct buf_2d *recon) {
932   const TX_SIZE max_tx_size = max_txsize_lookup[bs];
933   VP9_COMMON *const cm = &cpi->common;
934   MACROBLOCKD *const xd = &x->e_mbd;
935   MODE_INFO *const mi = xd->mi[0];
936   vpx_prob skip_prob = vp9_get_skip_prob(cm, xd);
937   int r[TX_SIZES][2], s[TX_SIZES];
938   int64_t d[TX_SIZES], sse[TX_SIZES];
939   int64_t rd[TX_SIZES][2] = { { INT64_MAX, INT64_MAX },
940                               { INT64_MAX, INT64_MAX },
941                               { INT64_MAX, INT64_MAX },
942                               { INT64_MAX, INT64_MAX } };
943   int n;
944   int s0, s1;
945   int64_t best_rd = ref_best_rd;
946   TX_SIZE best_tx = max_tx_size;
947   int start_tx, end_tx;
948   const int tx_size_ctx = get_tx_size_context(xd);
949 #if CONFIG_VP9_HIGHBITDEPTH
950   DECLARE_ALIGNED(16, uint16_t, recon_buf16[TX_SIZES][64 * 64]);
951   uint8_t *recon_buf[TX_SIZES];
952   for (n = 0; n < TX_SIZES; ++n) {
953     if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
954       recon_buf[n] = CONVERT_TO_BYTEPTR(recon_buf16[n]);
955     } else {
956       recon_buf[n] = (uint8_t *)recon_buf16[n];
957     }
958   }
959 #else
960   DECLARE_ALIGNED(16, uint8_t, recon_buf[TX_SIZES][64 * 64]);
961 #endif  // CONFIG_VP9_HIGHBITDEPTH
962 
963   assert(skip_prob > 0);
964   s0 = vp9_cost_bit(skip_prob, 0);
965   s1 = vp9_cost_bit(skip_prob, 1);
966 
967   if (cm->tx_mode == TX_MODE_SELECT) {
968     start_tx = max_tx_size;
969     end_tx = VPXMAX(start_tx - cpi->sf.tx_size_search_depth, 0);
970     if (bs > BLOCK_32X32) end_tx = VPXMIN(end_tx + 1, start_tx);
971   } else {
972     TX_SIZE chosen_tx_size =
973         VPXMIN(max_tx_size, tx_mode_to_biggest_tx_size[cm->tx_mode]);
974     start_tx = chosen_tx_size;
975     end_tx = chosen_tx_size;
976   }
977 
978   for (n = start_tx; n >= end_tx; n--) {
979     const int r_tx_size = cpi->tx_size_cost[max_tx_size - 1][tx_size_ctx][n];
980     if (recon) {
981       struct buf_2d this_recon;
982       this_recon.buf = recon_buf[n];
983       this_recon.stride = recon->stride;
984       txfm_rd_in_plane(cpi, x, &r[n][0], &d[n], &s[n], &sse[n], best_rd, 0, bs,
985                        n, cpi->sf.use_fast_coef_costing, &this_recon);
986     } else {
987       txfm_rd_in_plane(cpi, x, &r[n][0], &d[n], &s[n], &sse[n], best_rd, 0, bs,
988                        n, cpi->sf.use_fast_coef_costing, 0);
989     }
990     r[n][1] = r[n][0];
991     if (r[n][0] < INT_MAX) {
992       r[n][1] += r_tx_size;
993     }
994     if (d[n] == INT64_MAX || r[n][0] == INT_MAX) {
995       rd[n][0] = rd[n][1] = INT64_MAX;
996     } else if (s[n]) {
997       if (is_inter_block(mi)) {
998         rd[n][0] = rd[n][1] = RDCOST(x->rdmult, x->rddiv, s1, sse[n]);
999         r[n][1] -= r_tx_size;
1000       } else {
1001         rd[n][0] = RDCOST(x->rdmult, x->rddiv, s1, sse[n]);
1002         rd[n][1] = RDCOST(x->rdmult, x->rddiv, s1 + r_tx_size, sse[n]);
1003       }
1004     } else {
1005       rd[n][0] = RDCOST(x->rdmult, x->rddiv, r[n][0] + s0, d[n]);
1006       rd[n][1] = RDCOST(x->rdmult, x->rddiv, r[n][1] + s0, d[n]);
1007     }
1008 
1009     if (is_inter_block(mi) && !xd->lossless && !s[n] && sse[n] != INT64_MAX) {
1010       rd[n][0] = VPXMIN(rd[n][0], RDCOST(x->rdmult, x->rddiv, s1, sse[n]));
1011       rd[n][1] = VPXMIN(rd[n][1], RDCOST(x->rdmult, x->rddiv, s1, sse[n]));
1012     }
1013 
1014     // Early termination in transform size search.
1015     if (cpi->sf.tx_size_search_breakout &&
1016         (rd[n][1] == INT64_MAX ||
1017          (n < (int)max_tx_size && rd[n][1] > rd[n + 1][1]) || s[n] == 1))
1018       break;
1019 
1020     if (rd[n][1] < best_rd) {
1021       best_tx = n;
1022       best_rd = rd[n][1];
1023     }
1024   }
1025   mi->tx_size = best_tx;
1026 
1027   *distortion = d[mi->tx_size];
1028   *rate = r[mi->tx_size][cm->tx_mode == TX_MODE_SELECT];
1029   *skip = s[mi->tx_size];
1030   *psse = sse[mi->tx_size];
1031   if (recon) {
1032 #if CONFIG_VP9_HIGHBITDEPTH
1033     if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
1034       memcpy(CONVERT_TO_SHORTPTR(recon->buf),
1035              CONVERT_TO_SHORTPTR(recon_buf[mi->tx_size]),
1036              64 * 64 * sizeof(uint16_t));
1037     } else {
1038 #endif
1039       memcpy(recon->buf, recon_buf[mi->tx_size], 64 * 64);
1040 #if CONFIG_VP9_HIGHBITDEPTH
1041     }
1042 #endif
1043   }
1044 }
1045 
super_block_yrd(VP9_COMP * cpi,MACROBLOCK * x,int * rate,int64_t * distortion,int * skip,int64_t * psse,BLOCK_SIZE bs,int64_t ref_best_rd,struct buf_2d * recon)1046 static void super_block_yrd(VP9_COMP *cpi, MACROBLOCK *x, int *rate,
1047                             int64_t *distortion, int *skip, int64_t *psse,
1048                             BLOCK_SIZE bs, int64_t ref_best_rd,
1049                             struct buf_2d *recon) {
1050   MACROBLOCKD *xd = &x->e_mbd;
1051   int64_t sse;
1052   int64_t *ret_sse = psse ? psse : &sse;
1053 
1054   assert(bs == xd->mi[0]->sb_type);
1055 
1056   if (cpi->sf.tx_size_search_method == USE_LARGESTALL || xd->lossless) {
1057     choose_largest_tx_size(cpi, x, rate, distortion, skip, ret_sse, ref_best_rd,
1058                            bs, recon);
1059   } else {
1060     choose_tx_size_from_rd(cpi, x, rate, distortion, skip, ret_sse, ref_best_rd,
1061                            bs, recon);
1062   }
1063 }
1064 
conditional_skipintra(PREDICTION_MODE mode,PREDICTION_MODE best_intra_mode)1065 static int conditional_skipintra(PREDICTION_MODE mode,
1066                                  PREDICTION_MODE best_intra_mode) {
1067   if (mode == D117_PRED && best_intra_mode != V_PRED &&
1068       best_intra_mode != D135_PRED)
1069     return 1;
1070   if (mode == D63_PRED && best_intra_mode != V_PRED &&
1071       best_intra_mode != D45_PRED)
1072     return 1;
1073   if (mode == D207_PRED && best_intra_mode != H_PRED &&
1074       best_intra_mode != D45_PRED)
1075     return 1;
1076   if (mode == D153_PRED && best_intra_mode != H_PRED &&
1077       best_intra_mode != D135_PRED)
1078     return 1;
1079   return 0;
1080 }
1081 
rd_pick_intra4x4block(VP9_COMP * cpi,MACROBLOCK * x,int row,int col,PREDICTION_MODE * best_mode,const int * bmode_costs,ENTROPY_CONTEXT * a,ENTROPY_CONTEXT * l,int * bestrate,int * bestratey,int64_t * bestdistortion,BLOCK_SIZE bsize,int64_t rd_thresh)1082 static int64_t rd_pick_intra4x4block(VP9_COMP *cpi, MACROBLOCK *x, int row,
1083                                      int col, PREDICTION_MODE *best_mode,
1084                                      const int *bmode_costs, ENTROPY_CONTEXT *a,
1085                                      ENTROPY_CONTEXT *l, int *bestrate,
1086                                      int *bestratey, int64_t *bestdistortion,
1087                                      BLOCK_SIZE bsize, int64_t rd_thresh) {
1088   PREDICTION_MODE mode;
1089   MACROBLOCKD *const xd = &x->e_mbd;
1090   int64_t best_rd = rd_thresh;
1091   struct macroblock_plane *p = &x->plane[0];
1092   struct macroblockd_plane *pd = &xd->plane[0];
1093   const int src_stride = p->src.stride;
1094   const int dst_stride = pd->dst.stride;
1095   const uint8_t *src_init = &p->src.buf[row * 4 * src_stride + col * 4];
1096   uint8_t *dst_init = &pd->dst.buf[row * 4 * src_stride + col * 4];
1097   ENTROPY_CONTEXT ta[2], tempa[2];
1098   ENTROPY_CONTEXT tl[2], templ[2];
1099   const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
1100   const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
1101   int idx, idy;
1102   uint8_t best_dst[8 * 8];
1103 #if CONFIG_VP9_HIGHBITDEPTH
1104   uint16_t best_dst16[8 * 8];
1105 #endif
1106   memcpy(ta, a, num_4x4_blocks_wide * sizeof(a[0]));
1107   memcpy(tl, l, num_4x4_blocks_high * sizeof(l[0]));
1108 
1109   xd->mi[0]->tx_size = TX_4X4;
1110 
1111   assert(!x->skip_block);
1112 
1113 #if CONFIG_VP9_HIGHBITDEPTH
1114   if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
1115     for (mode = DC_PRED; mode <= TM_PRED; ++mode) {
1116       int64_t this_rd;
1117       int ratey = 0;
1118       int64_t distortion = 0;
1119       int rate = bmode_costs[mode];
1120 
1121       if (!(cpi->sf.intra_y_mode_mask[TX_4X4] & (1 << mode))) continue;
1122 
1123       // Only do the oblique modes if the best so far is
1124       // one of the neighboring directional modes
1125       if (cpi->sf.mode_search_skip_flags & FLAG_SKIP_INTRA_DIRMISMATCH) {
1126         if (conditional_skipintra(mode, *best_mode)) continue;
1127       }
1128 
1129       memcpy(tempa, ta, num_4x4_blocks_wide * sizeof(ta[0]));
1130       memcpy(templ, tl, num_4x4_blocks_high * sizeof(tl[0]));
1131 
1132       for (idy = 0; idy < num_4x4_blocks_high; ++idy) {
1133         for (idx = 0; idx < num_4x4_blocks_wide; ++idx) {
1134           const int block = (row + idy) * 2 + (col + idx);
1135           const uint8_t *const src = &src_init[idx * 4 + idy * 4 * src_stride];
1136           uint8_t *const dst = &dst_init[idx * 4 + idy * 4 * dst_stride];
1137           uint16_t *const dst16 = CONVERT_TO_SHORTPTR(dst);
1138           int16_t *const src_diff =
1139               vp9_raster_block_offset_int16(BLOCK_8X8, block, p->src_diff);
1140           tran_low_t *const coeff = BLOCK_OFFSET(p->coeff, block);
1141           tran_low_t *const qcoeff = BLOCK_OFFSET(p->qcoeff, block);
1142           tran_low_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
1143           uint16_t *const eob = &p->eobs[block];
1144           xd->mi[0]->bmi[block].as_mode = mode;
1145           vp9_predict_intra_block(xd, 1, TX_4X4, mode,
1146                                   x->skip_encode ? src : dst,
1147                                   x->skip_encode ? src_stride : dst_stride, dst,
1148                                   dst_stride, col + idx, row + idy, 0);
1149           vpx_highbd_subtract_block(4, 4, src_diff, 8, src, src_stride, dst,
1150                                     dst_stride, xd->bd);
1151           if (xd->lossless) {
1152             const scan_order *so = &vp9_default_scan_orders[TX_4X4];
1153             const int coeff_ctx =
1154                 combine_entropy_contexts(tempa[idx], templ[idy]);
1155             vp9_highbd_fwht4x4(src_diff, coeff, 8);
1156             vpx_highbd_quantize_b(coeff, 4 * 4, p->zbin, p->round, p->quant,
1157                                   p->quant_shift, qcoeff, dqcoeff, pd->dequant,
1158                                   eob, so->scan, so->iscan);
1159             ratey += cost_coeffs(x, 0, block, TX_4X4, coeff_ctx, so->scan,
1160                                  so->neighbors, cpi->sf.use_fast_coef_costing);
1161             tempa[idx] = templ[idy] = (x->plane[0].eobs[block] > 0 ? 1 : 0);
1162             if (RDCOST(x->rdmult, x->rddiv, ratey, distortion) >= best_rd)
1163               goto next_highbd;
1164             vp9_highbd_iwht4x4_add(BLOCK_OFFSET(pd->dqcoeff, block), dst16,
1165                                    dst_stride, p->eobs[block], xd->bd);
1166           } else {
1167             int64_t unused;
1168             const TX_TYPE tx_type = get_tx_type_4x4(PLANE_TYPE_Y, xd, block);
1169             const scan_order *so = &vp9_scan_orders[TX_4X4][tx_type];
1170             const int coeff_ctx =
1171                 combine_entropy_contexts(tempa[idx], templ[idy]);
1172             if (tx_type == DCT_DCT)
1173               vpx_highbd_fdct4x4(src_diff, coeff, 8);
1174             else
1175               vp9_highbd_fht4x4(src_diff, coeff, 8, tx_type);
1176             vpx_highbd_quantize_b(coeff, 4 * 4, p->zbin, p->round, p->quant,
1177                                   p->quant_shift, qcoeff, dqcoeff, pd->dequant,
1178                                   eob, so->scan, so->iscan);
1179             ratey += cost_coeffs(x, 0, block, TX_4X4, coeff_ctx, so->scan,
1180                                  so->neighbors, cpi->sf.use_fast_coef_costing);
1181             distortion += vp9_highbd_block_error_dispatch(
1182                               coeff, BLOCK_OFFSET(pd->dqcoeff, block), 16,
1183                               &unused, xd->bd) >>
1184                           2;
1185             tempa[idx] = templ[idy] = (x->plane[0].eobs[block] > 0 ? 1 : 0);
1186             if (RDCOST(x->rdmult, x->rddiv, ratey, distortion) >= best_rd)
1187               goto next_highbd;
1188             vp9_highbd_iht4x4_add(tx_type, BLOCK_OFFSET(pd->dqcoeff, block),
1189                                   dst16, dst_stride, p->eobs[block], xd->bd);
1190           }
1191         }
1192       }
1193 
1194       rate += ratey;
1195       this_rd = RDCOST(x->rdmult, x->rddiv, rate, distortion);
1196 
1197       if (this_rd < best_rd) {
1198         *bestrate = rate;
1199         *bestratey = ratey;
1200         *bestdistortion = distortion;
1201         best_rd = this_rd;
1202         *best_mode = mode;
1203         memcpy(a, tempa, num_4x4_blocks_wide * sizeof(tempa[0]));
1204         memcpy(l, templ, num_4x4_blocks_high * sizeof(templ[0]));
1205         for (idy = 0; idy < num_4x4_blocks_high * 4; ++idy) {
1206           memcpy(best_dst16 + idy * 8,
1207                  CONVERT_TO_SHORTPTR(dst_init + idy * dst_stride),
1208                  num_4x4_blocks_wide * 4 * sizeof(uint16_t));
1209         }
1210       }
1211     next_highbd : {}
1212     }
1213     if (best_rd >= rd_thresh || x->skip_encode) return best_rd;
1214 
1215     for (idy = 0; idy < num_4x4_blocks_high * 4; ++idy) {
1216       memcpy(CONVERT_TO_SHORTPTR(dst_init + idy * dst_stride),
1217              best_dst16 + idy * 8, num_4x4_blocks_wide * 4 * sizeof(uint16_t));
1218     }
1219 
1220     return best_rd;
1221   }
1222 #endif  // CONFIG_VP9_HIGHBITDEPTH
1223 
1224   for (mode = DC_PRED; mode <= TM_PRED; ++mode) {
1225     int64_t this_rd;
1226     int ratey = 0;
1227     int64_t distortion = 0;
1228     int rate = bmode_costs[mode];
1229 
1230     if (!(cpi->sf.intra_y_mode_mask[TX_4X4] & (1 << mode))) continue;
1231 
1232     // Only do the oblique modes if the best so far is
1233     // one of the neighboring directional modes
1234     if (cpi->sf.mode_search_skip_flags & FLAG_SKIP_INTRA_DIRMISMATCH) {
1235       if (conditional_skipintra(mode, *best_mode)) continue;
1236     }
1237 
1238     memcpy(tempa, ta, num_4x4_blocks_wide * sizeof(ta[0]));
1239     memcpy(templ, tl, num_4x4_blocks_high * sizeof(tl[0]));
1240 
1241     for (idy = 0; idy < num_4x4_blocks_high; ++idy) {
1242       for (idx = 0; idx < num_4x4_blocks_wide; ++idx) {
1243         const int block = (row + idy) * 2 + (col + idx);
1244         const uint8_t *const src = &src_init[idx * 4 + idy * 4 * src_stride];
1245         uint8_t *const dst = &dst_init[idx * 4 + idy * 4 * dst_stride];
1246         int16_t *const src_diff =
1247             vp9_raster_block_offset_int16(BLOCK_8X8, block, p->src_diff);
1248         tran_low_t *const coeff = BLOCK_OFFSET(p->coeff, block);
1249         tran_low_t *const qcoeff = BLOCK_OFFSET(p->qcoeff, block);
1250         tran_low_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
1251         uint16_t *const eob = &p->eobs[block];
1252         xd->mi[0]->bmi[block].as_mode = mode;
1253         vp9_predict_intra_block(xd, 1, TX_4X4, mode, x->skip_encode ? src : dst,
1254                                 x->skip_encode ? src_stride : dst_stride, dst,
1255                                 dst_stride, col + idx, row + idy, 0);
1256         vpx_subtract_block(4, 4, src_diff, 8, src, src_stride, dst, dst_stride);
1257 
1258         if (xd->lossless) {
1259           const scan_order *so = &vp9_default_scan_orders[TX_4X4];
1260           const int coeff_ctx =
1261               combine_entropy_contexts(tempa[idx], templ[idy]);
1262           vp9_fwht4x4(src_diff, coeff, 8);
1263           vpx_quantize_b(coeff, 4 * 4, p->zbin, p->round, p->quant,
1264                          p->quant_shift, qcoeff, dqcoeff, pd->dequant, eob,
1265                          so->scan, so->iscan);
1266           ratey += cost_coeffs(x, 0, block, TX_4X4, coeff_ctx, so->scan,
1267                                so->neighbors, cpi->sf.use_fast_coef_costing);
1268           tempa[idx] = templ[idy] = (x->plane[0].eobs[block] > 0) ? 1 : 0;
1269           if (RDCOST(x->rdmult, x->rddiv, ratey, distortion) >= best_rd)
1270             goto next;
1271           vp9_iwht4x4_add(BLOCK_OFFSET(pd->dqcoeff, block), dst, dst_stride,
1272                           p->eobs[block]);
1273         } else {
1274           int64_t unused;
1275           const TX_TYPE tx_type = get_tx_type_4x4(PLANE_TYPE_Y, xd, block);
1276           const scan_order *so = &vp9_scan_orders[TX_4X4][tx_type];
1277           const int coeff_ctx =
1278               combine_entropy_contexts(tempa[idx], templ[idy]);
1279           vp9_fht4x4(src_diff, coeff, 8, tx_type);
1280           vpx_quantize_b(coeff, 4 * 4, p->zbin, p->round, p->quant,
1281                          p->quant_shift, qcoeff, dqcoeff, pd->dequant, eob,
1282                          so->scan, so->iscan);
1283           ratey += cost_coeffs(x, 0, block, TX_4X4, coeff_ctx, so->scan,
1284                                so->neighbors, cpi->sf.use_fast_coef_costing);
1285           tempa[idx] = templ[idy] = (x->plane[0].eobs[block] > 0) ? 1 : 0;
1286           distortion += vp9_block_error(coeff, BLOCK_OFFSET(pd->dqcoeff, block),
1287                                         16, &unused) >>
1288                         2;
1289           if (RDCOST(x->rdmult, x->rddiv, ratey, distortion) >= best_rd)
1290             goto next;
1291           vp9_iht4x4_add(tx_type, BLOCK_OFFSET(pd->dqcoeff, block), dst,
1292                          dst_stride, p->eobs[block]);
1293         }
1294       }
1295     }
1296 
1297     rate += ratey;
1298     this_rd = RDCOST(x->rdmult, x->rddiv, rate, distortion);
1299 
1300     if (this_rd < best_rd) {
1301       *bestrate = rate;
1302       *bestratey = ratey;
1303       *bestdistortion = distortion;
1304       best_rd = this_rd;
1305       *best_mode = mode;
1306       memcpy(a, tempa, num_4x4_blocks_wide * sizeof(tempa[0]));
1307       memcpy(l, templ, num_4x4_blocks_high * sizeof(templ[0]));
1308       for (idy = 0; idy < num_4x4_blocks_high * 4; ++idy)
1309         memcpy(best_dst + idy * 8, dst_init + idy * dst_stride,
1310                num_4x4_blocks_wide * 4);
1311     }
1312   next : {}
1313   }
1314 
1315   if (best_rd >= rd_thresh || x->skip_encode) return best_rd;
1316 
1317   for (idy = 0; idy < num_4x4_blocks_high * 4; ++idy)
1318     memcpy(dst_init + idy * dst_stride, best_dst + idy * 8,
1319            num_4x4_blocks_wide * 4);
1320 
1321   return best_rd;
1322 }
1323 
rd_pick_intra_sub_8x8_y_mode(VP9_COMP * cpi,MACROBLOCK * mb,int * rate,int * rate_y,int64_t * distortion,int64_t best_rd)1324 static int64_t rd_pick_intra_sub_8x8_y_mode(VP9_COMP *cpi, MACROBLOCK *mb,
1325                                             int *rate, int *rate_y,
1326                                             int64_t *distortion,
1327                                             int64_t best_rd) {
1328   int i, j;
1329   const MACROBLOCKD *const xd = &mb->e_mbd;
1330   MODE_INFO *const mic = xd->mi[0];
1331   const MODE_INFO *above_mi = xd->above_mi;
1332   const MODE_INFO *left_mi = xd->left_mi;
1333   const BLOCK_SIZE bsize = xd->mi[0]->sb_type;
1334   const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
1335   const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
1336   int idx, idy;
1337   int cost = 0;
1338   int64_t total_distortion = 0;
1339   int tot_rate_y = 0;
1340   int64_t total_rd = 0;
1341   const int *bmode_costs = cpi->mbmode_cost;
1342 
1343   // Pick modes for each sub-block (of size 4x4, 4x8, or 8x4) in an 8x8 block.
1344   for (idy = 0; idy < 2; idy += num_4x4_blocks_high) {
1345     for (idx = 0; idx < 2; idx += num_4x4_blocks_wide) {
1346       PREDICTION_MODE best_mode = DC_PRED;
1347       int r = INT_MAX, ry = INT_MAX;
1348       int64_t d = INT64_MAX, this_rd = INT64_MAX;
1349       i = idy * 2 + idx;
1350       if (cpi->common.frame_type == KEY_FRAME) {
1351         const PREDICTION_MODE A = vp9_above_block_mode(mic, above_mi, i);
1352         const PREDICTION_MODE L = vp9_left_block_mode(mic, left_mi, i);
1353 
1354         bmode_costs = cpi->y_mode_costs[A][L];
1355       }
1356 
1357       this_rd = rd_pick_intra4x4block(
1358           cpi, mb, idy, idx, &best_mode, bmode_costs,
1359           xd->plane[0].above_context + idx, xd->plane[0].left_context + idy, &r,
1360           &ry, &d, bsize, best_rd - total_rd);
1361 
1362       if (this_rd >= best_rd - total_rd) return INT64_MAX;
1363 
1364       total_rd += this_rd;
1365       cost += r;
1366       total_distortion += d;
1367       tot_rate_y += ry;
1368 
1369       mic->bmi[i].as_mode = best_mode;
1370       for (j = 1; j < num_4x4_blocks_high; ++j)
1371         mic->bmi[i + j * 2].as_mode = best_mode;
1372       for (j = 1; j < num_4x4_blocks_wide; ++j)
1373         mic->bmi[i + j].as_mode = best_mode;
1374 
1375       if (total_rd >= best_rd) return INT64_MAX;
1376     }
1377   }
1378 
1379   *rate = cost;
1380   *rate_y = tot_rate_y;
1381   *distortion = total_distortion;
1382   mic->mode = mic->bmi[3].as_mode;
1383 
1384   return RDCOST(mb->rdmult, mb->rddiv, cost, total_distortion);
1385 }
1386 
1387 // This function is used only for intra_only frames
rd_pick_intra_sby_mode(VP9_COMP * cpi,MACROBLOCK * x,int * rate,int * rate_tokenonly,int64_t * distortion,int * skippable,BLOCK_SIZE bsize,int64_t best_rd)1388 static int64_t rd_pick_intra_sby_mode(VP9_COMP *cpi, MACROBLOCK *x, int *rate,
1389                                       int *rate_tokenonly, int64_t *distortion,
1390                                       int *skippable, BLOCK_SIZE bsize,
1391                                       int64_t best_rd) {
1392   PREDICTION_MODE mode;
1393   PREDICTION_MODE mode_selected = DC_PRED;
1394   MACROBLOCKD *const xd = &x->e_mbd;
1395   MODE_INFO *const mic = xd->mi[0];
1396   int this_rate, this_rate_tokenonly, s;
1397   int64_t this_distortion, this_rd;
1398   TX_SIZE best_tx = TX_4X4;
1399   int *bmode_costs;
1400   const MODE_INFO *above_mi = xd->above_mi;
1401   const MODE_INFO *left_mi = xd->left_mi;
1402   const PREDICTION_MODE A = vp9_above_block_mode(mic, above_mi, 0);
1403   const PREDICTION_MODE L = vp9_left_block_mode(mic, left_mi, 0);
1404   bmode_costs = cpi->y_mode_costs[A][L];
1405 
1406   memset(x->skip_txfm, SKIP_TXFM_NONE, sizeof(x->skip_txfm));
1407   /* Y Search for intra prediction mode */
1408   for (mode = DC_PRED; mode <= TM_PRED; mode++) {
1409     if (cpi->sf.use_nonrd_pick_mode) {
1410       // These speed features are turned on in hybrid non-RD and RD mode
1411       // for key frame coding in the context of real-time setting.
1412       if (conditional_skipintra(mode, mode_selected)) continue;
1413       if (*skippable) break;
1414     }
1415 
1416     mic->mode = mode;
1417 
1418     super_block_yrd(cpi, x, &this_rate_tokenonly, &this_distortion, &s, NULL,
1419                     bsize, best_rd, /*recon = */ 0);
1420 
1421     if (this_rate_tokenonly == INT_MAX) continue;
1422 
1423     this_rate = this_rate_tokenonly + bmode_costs[mode];
1424     this_rd = RDCOST(x->rdmult, x->rddiv, this_rate, this_distortion);
1425 
1426     if (this_rd < best_rd) {
1427       mode_selected = mode;
1428       best_rd = this_rd;
1429       best_tx = mic->tx_size;
1430       *rate = this_rate;
1431       *rate_tokenonly = this_rate_tokenonly;
1432       *distortion = this_distortion;
1433       *skippable = s;
1434     }
1435   }
1436 
1437   mic->mode = mode_selected;
1438   mic->tx_size = best_tx;
1439 
1440   return best_rd;
1441 }
1442 
1443 // Return value 0: early termination triggered, no valid rd cost available;
1444 //              1: rd cost values are valid.
super_block_uvrd(const VP9_COMP * cpi,MACROBLOCK * x,int * rate,int64_t * distortion,int * skippable,int64_t * sse,BLOCK_SIZE bsize,int64_t ref_best_rd)1445 static int super_block_uvrd(const VP9_COMP *cpi, MACROBLOCK *x, int *rate,
1446                             int64_t *distortion, int *skippable, int64_t *sse,
1447                             BLOCK_SIZE bsize, int64_t ref_best_rd) {
1448   MACROBLOCKD *const xd = &x->e_mbd;
1449   MODE_INFO *const mi = xd->mi[0];
1450   const TX_SIZE uv_tx_size = get_uv_tx_size(mi, &xd->plane[1]);
1451   int plane;
1452   int pnrate = 0, pnskip = 1;
1453   int64_t pndist = 0, pnsse = 0;
1454   int is_cost_valid = 1;
1455 
1456   if (ref_best_rd < 0) is_cost_valid = 0;
1457 
1458   if (is_inter_block(mi) && is_cost_valid) {
1459     int plane;
1460     for (plane = 1; plane < MAX_MB_PLANE; ++plane)
1461       vp9_subtract_plane(x, bsize, plane);
1462   }
1463 
1464   *rate = 0;
1465   *distortion = 0;
1466   *sse = 0;
1467   *skippable = 1;
1468 
1469   for (plane = 1; plane < MAX_MB_PLANE; ++plane) {
1470     txfm_rd_in_plane(cpi, x, &pnrate, &pndist, &pnskip, &pnsse, ref_best_rd,
1471                      plane, bsize, uv_tx_size, cpi->sf.use_fast_coef_costing,
1472                      /*recon = */ 0);
1473     if (pnrate == INT_MAX) {
1474       is_cost_valid = 0;
1475       break;
1476     }
1477     *rate += pnrate;
1478     *distortion += pndist;
1479     *sse += pnsse;
1480     *skippable &= pnskip;
1481   }
1482 
1483   if (!is_cost_valid) {
1484     // reset cost value
1485     *rate = INT_MAX;
1486     *distortion = INT64_MAX;
1487     *sse = INT64_MAX;
1488     *skippable = 0;
1489   }
1490 
1491   return is_cost_valid;
1492 }
1493 
rd_pick_intra_sbuv_mode(VP9_COMP * cpi,MACROBLOCK * x,PICK_MODE_CONTEXT * ctx,int * rate,int * rate_tokenonly,int64_t * distortion,int * skippable,BLOCK_SIZE bsize,TX_SIZE max_tx_size)1494 static int64_t rd_pick_intra_sbuv_mode(VP9_COMP *cpi, MACROBLOCK *x,
1495                                        PICK_MODE_CONTEXT *ctx, int *rate,
1496                                        int *rate_tokenonly, int64_t *distortion,
1497                                        int *skippable, BLOCK_SIZE bsize,
1498                                        TX_SIZE max_tx_size) {
1499   MACROBLOCKD *xd = &x->e_mbd;
1500   PREDICTION_MODE mode;
1501   PREDICTION_MODE mode_selected = DC_PRED;
1502   int64_t best_rd = INT64_MAX, this_rd;
1503   int this_rate_tokenonly, this_rate, s;
1504   int64_t this_distortion, this_sse;
1505 
1506   memset(x->skip_txfm, SKIP_TXFM_NONE, sizeof(x->skip_txfm));
1507   for (mode = DC_PRED; mode <= TM_PRED; ++mode) {
1508     if (!(cpi->sf.intra_uv_mode_mask[max_tx_size] & (1 << mode))) continue;
1509 #if CONFIG_BETTER_HW_COMPATIBILITY && CONFIG_VP9_HIGHBITDEPTH
1510     if ((xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) &&
1511         (xd->above_mi == NULL || xd->left_mi == NULL) && need_top_left[mode])
1512       continue;
1513 #endif  // CONFIG_BETTER_HW_COMPATIBILITY && CONFIG_VP9_HIGHBITDEPTH
1514 
1515     xd->mi[0]->uv_mode = mode;
1516 
1517     if (!super_block_uvrd(cpi, x, &this_rate_tokenonly, &this_distortion, &s,
1518                           &this_sse, bsize, best_rd))
1519       continue;
1520     this_rate =
1521         this_rate_tokenonly +
1522         cpi->intra_uv_mode_cost[cpi->common.frame_type][xd->mi[0]->mode][mode];
1523     this_rd = RDCOST(x->rdmult, x->rddiv, this_rate, this_distortion);
1524 
1525     if (this_rd < best_rd) {
1526       mode_selected = mode;
1527       best_rd = this_rd;
1528       *rate = this_rate;
1529       *rate_tokenonly = this_rate_tokenonly;
1530       *distortion = this_distortion;
1531       *skippable = s;
1532       if (!x->select_tx_size) swap_block_ptr(x, ctx, 2, 0, 1, MAX_MB_PLANE);
1533     }
1534   }
1535 
1536   xd->mi[0]->uv_mode = mode_selected;
1537   return best_rd;
1538 }
1539 
1540 #if !CONFIG_REALTIME_ONLY
rd_sbuv_dcpred(const VP9_COMP * cpi,MACROBLOCK * x,int * rate,int * rate_tokenonly,int64_t * distortion,int * skippable,BLOCK_SIZE bsize)1541 static int64_t rd_sbuv_dcpred(const VP9_COMP *cpi, MACROBLOCK *x, int *rate,
1542                               int *rate_tokenonly, int64_t *distortion,
1543                               int *skippable, BLOCK_SIZE bsize) {
1544   const VP9_COMMON *cm = &cpi->common;
1545   int64_t unused;
1546 
1547   x->e_mbd.mi[0]->uv_mode = DC_PRED;
1548   memset(x->skip_txfm, SKIP_TXFM_NONE, sizeof(x->skip_txfm));
1549   super_block_uvrd(cpi, x, rate_tokenonly, distortion, skippable, &unused,
1550                    bsize, INT64_MAX);
1551   *rate =
1552       *rate_tokenonly +
1553       cpi->intra_uv_mode_cost[cm->frame_type][x->e_mbd.mi[0]->mode][DC_PRED];
1554   return RDCOST(x->rdmult, x->rddiv, *rate, *distortion);
1555 }
1556 
choose_intra_uv_mode(VP9_COMP * cpi,MACROBLOCK * const x,PICK_MODE_CONTEXT * ctx,BLOCK_SIZE bsize,TX_SIZE max_tx_size,int * rate_uv,int * rate_uv_tokenonly,int64_t * dist_uv,int * skip_uv,PREDICTION_MODE * mode_uv)1557 static void choose_intra_uv_mode(VP9_COMP *cpi, MACROBLOCK *const x,
1558                                  PICK_MODE_CONTEXT *ctx, BLOCK_SIZE bsize,
1559                                  TX_SIZE max_tx_size, int *rate_uv,
1560                                  int *rate_uv_tokenonly, int64_t *dist_uv,
1561                                  int *skip_uv, PREDICTION_MODE *mode_uv) {
1562   // Use an estimated rd for uv_intra based on DC_PRED if the
1563   // appropriate speed flag is set.
1564   if (cpi->sf.use_uv_intra_rd_estimate) {
1565     rd_sbuv_dcpred(cpi, x, rate_uv, rate_uv_tokenonly, dist_uv, skip_uv,
1566                    bsize < BLOCK_8X8 ? BLOCK_8X8 : bsize);
1567     // Else do a proper rd search for each possible transform size that may
1568     // be considered in the main rd loop.
1569   } else {
1570     rd_pick_intra_sbuv_mode(cpi, x, ctx, rate_uv, rate_uv_tokenonly, dist_uv,
1571                             skip_uv, bsize < BLOCK_8X8 ? BLOCK_8X8 : bsize,
1572                             max_tx_size);
1573   }
1574   *mode_uv = x->e_mbd.mi[0]->uv_mode;
1575 }
1576 
cost_mv_ref(const VP9_COMP * cpi,PREDICTION_MODE mode,int mode_context)1577 static int cost_mv_ref(const VP9_COMP *cpi, PREDICTION_MODE mode,
1578                        int mode_context) {
1579   assert(is_inter_mode(mode));
1580   return cpi->inter_mode_cost[mode_context][INTER_OFFSET(mode)];
1581 }
1582 
set_and_cost_bmi_mvs(VP9_COMP * cpi,MACROBLOCK * x,MACROBLOCKD * xd,int i,PREDICTION_MODE mode,int_mv this_mv[2],int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES],int_mv seg_mvs[MAX_REF_FRAMES],int_mv * best_ref_mv[2],const int * mvjcost,int * mvcost[2])1583 static int set_and_cost_bmi_mvs(VP9_COMP *cpi, MACROBLOCK *x, MACROBLOCKD *xd,
1584                                 int i, PREDICTION_MODE mode, int_mv this_mv[2],
1585                                 int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES],
1586                                 int_mv seg_mvs[MAX_REF_FRAMES],
1587                                 int_mv *best_ref_mv[2], const int *mvjcost,
1588                                 int *mvcost[2]) {
1589   MODE_INFO *const mi = xd->mi[0];
1590   const MB_MODE_INFO_EXT *const mbmi_ext = x->mbmi_ext;
1591   int thismvcost = 0;
1592   int idx, idy;
1593   const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[mi->sb_type];
1594   const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[mi->sb_type];
1595   const int is_compound = has_second_ref(mi);
1596 
1597   switch (mode) {
1598     case NEWMV:
1599       this_mv[0].as_int = seg_mvs[mi->ref_frame[0]].as_int;
1600       thismvcost += vp9_mv_bit_cost(&this_mv[0].as_mv, &best_ref_mv[0]->as_mv,
1601                                     mvjcost, mvcost, MV_COST_WEIGHT_SUB);
1602       if (is_compound) {
1603         this_mv[1].as_int = seg_mvs[mi->ref_frame[1]].as_int;
1604         thismvcost += vp9_mv_bit_cost(&this_mv[1].as_mv, &best_ref_mv[1]->as_mv,
1605                                       mvjcost, mvcost, MV_COST_WEIGHT_SUB);
1606       }
1607       break;
1608     case NEARMV:
1609     case NEARESTMV:
1610       this_mv[0].as_int = frame_mv[mode][mi->ref_frame[0]].as_int;
1611       if (is_compound)
1612         this_mv[1].as_int = frame_mv[mode][mi->ref_frame[1]].as_int;
1613       break;
1614     default:
1615       assert(mode == ZEROMV);
1616       this_mv[0].as_int = 0;
1617       if (is_compound) this_mv[1].as_int = 0;
1618       break;
1619   }
1620 
1621   mi->bmi[i].as_mv[0].as_int = this_mv[0].as_int;
1622   if (is_compound) mi->bmi[i].as_mv[1].as_int = this_mv[1].as_int;
1623 
1624   mi->bmi[i].as_mode = mode;
1625 
1626   for (idy = 0; idy < num_4x4_blocks_high; ++idy)
1627     for (idx = 0; idx < num_4x4_blocks_wide; ++idx)
1628       memmove(&mi->bmi[i + idy * 2 + idx], &mi->bmi[i], sizeof(mi->bmi[i]));
1629 
1630   return cost_mv_ref(cpi, mode, mbmi_ext->mode_context[mi->ref_frame[0]]) +
1631          thismvcost;
1632 }
1633 
encode_inter_mb_segment(VP9_COMP * cpi,MACROBLOCK * x,int64_t best_yrd,int i,int * labelyrate,int64_t * distortion,int64_t * sse,ENTROPY_CONTEXT * ta,ENTROPY_CONTEXT * tl,int mi_row,int mi_col)1634 static int64_t encode_inter_mb_segment(VP9_COMP *cpi, MACROBLOCK *x,
1635                                        int64_t best_yrd, int i, int *labelyrate,
1636                                        int64_t *distortion, int64_t *sse,
1637                                        ENTROPY_CONTEXT *ta, ENTROPY_CONTEXT *tl,
1638                                        int mi_row, int mi_col) {
1639   int k;
1640   MACROBLOCKD *xd = &x->e_mbd;
1641   struct macroblockd_plane *const pd = &xd->plane[0];
1642   struct macroblock_plane *const p = &x->plane[0];
1643   MODE_INFO *const mi = xd->mi[0];
1644   const BLOCK_SIZE plane_bsize = get_plane_block_size(mi->sb_type, pd);
1645   const int width = 4 * num_4x4_blocks_wide_lookup[plane_bsize];
1646   const int height = 4 * num_4x4_blocks_high_lookup[plane_bsize];
1647   int idx, idy;
1648 
1649   const uint8_t *const src =
1650       &p->src.buf[vp9_raster_block_offset(BLOCK_8X8, i, p->src.stride)];
1651   uint8_t *const dst =
1652       &pd->dst.buf[vp9_raster_block_offset(BLOCK_8X8, i, pd->dst.stride)];
1653   int64_t thisdistortion = 0, thissse = 0;
1654   int thisrate = 0, ref;
1655   const scan_order *so = &vp9_default_scan_orders[TX_4X4];
1656   const int is_compound = has_second_ref(mi);
1657   const InterpKernel *kernel = vp9_filter_kernels[mi->interp_filter];
1658 
1659   assert(!x->skip_block);
1660 
1661   for (ref = 0; ref < 1 + is_compound; ++ref) {
1662     const int bw = b_width_log2_lookup[BLOCK_8X8];
1663     const int h = 4 * (i >> bw);
1664     const int w = 4 * (i & ((1 << bw) - 1));
1665     const struct scale_factors *sf = &xd->block_refs[ref]->sf;
1666     int y_stride = pd->pre[ref].stride;
1667     uint8_t *pre = pd->pre[ref].buf + (h * pd->pre[ref].stride + w);
1668 
1669     if (vp9_is_scaled(sf)) {
1670       const int x_start = (-xd->mb_to_left_edge >> (3 + pd->subsampling_x));
1671       const int y_start = (-xd->mb_to_top_edge >> (3 + pd->subsampling_y));
1672 
1673       y_stride = xd->block_refs[ref]->buf->y_stride;
1674       pre = xd->block_refs[ref]->buf->y_buffer;
1675       pre += scaled_buffer_offset(x_start + w, y_start + h, y_stride, sf);
1676     }
1677 #if CONFIG_VP9_HIGHBITDEPTH
1678     if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
1679       vp9_highbd_build_inter_predictor(
1680           CONVERT_TO_SHORTPTR(pre), y_stride, CONVERT_TO_SHORTPTR(dst),
1681           pd->dst.stride, &mi->bmi[i].as_mv[ref].as_mv,
1682           &xd->block_refs[ref]->sf, width, height, ref, kernel, MV_PRECISION_Q3,
1683           mi_col * MI_SIZE + 4 * (i % 2), mi_row * MI_SIZE + 4 * (i / 2),
1684           xd->bd);
1685     } else {
1686       vp9_build_inter_predictor(
1687           pre, y_stride, dst, pd->dst.stride, &mi->bmi[i].as_mv[ref].as_mv,
1688           &xd->block_refs[ref]->sf, width, height, ref, kernel, MV_PRECISION_Q3,
1689           mi_col * MI_SIZE + 4 * (i % 2), mi_row * MI_SIZE + 4 * (i / 2));
1690     }
1691 #else
1692     vp9_build_inter_predictor(
1693         pre, y_stride, dst, pd->dst.stride, &mi->bmi[i].as_mv[ref].as_mv,
1694         &xd->block_refs[ref]->sf, width, height, ref, kernel, MV_PRECISION_Q3,
1695         mi_col * MI_SIZE + 4 * (i % 2), mi_row * MI_SIZE + 4 * (i / 2));
1696 #endif  // CONFIG_VP9_HIGHBITDEPTH
1697   }
1698 
1699 #if CONFIG_VP9_HIGHBITDEPTH
1700   if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
1701     vpx_highbd_subtract_block(
1702         height, width, vp9_raster_block_offset_int16(BLOCK_8X8, i, p->src_diff),
1703         8, src, p->src.stride, dst, pd->dst.stride, xd->bd);
1704   } else {
1705     vpx_subtract_block(height, width,
1706                        vp9_raster_block_offset_int16(BLOCK_8X8, i, p->src_diff),
1707                        8, src, p->src.stride, dst, pd->dst.stride);
1708   }
1709 #else
1710   vpx_subtract_block(height, width,
1711                      vp9_raster_block_offset_int16(BLOCK_8X8, i, p->src_diff),
1712                      8, src, p->src.stride, dst, pd->dst.stride);
1713 #endif  // CONFIG_VP9_HIGHBITDEPTH
1714 
1715   k = i;
1716   for (idy = 0; idy < height / 4; ++idy) {
1717     for (idx = 0; idx < width / 4; ++idx) {
1718 #if CONFIG_VP9_HIGHBITDEPTH
1719       const int bd = (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) ? xd->bd : 8;
1720 #endif
1721       int64_t ssz, rd, rd1, rd2;
1722       tran_low_t *coeff, *qcoeff, *dqcoeff;
1723       uint16_t *eob;
1724       int coeff_ctx;
1725       k += (idy * 2 + idx);
1726       coeff_ctx = combine_entropy_contexts(ta[k & 1], tl[k >> 1]);
1727       coeff = BLOCK_OFFSET(p->coeff, k);
1728       qcoeff = BLOCK_OFFSET(p->qcoeff, k);
1729       dqcoeff = BLOCK_OFFSET(pd->dqcoeff, k);
1730       eob = &p->eobs[k];
1731 
1732       x->fwd_txfm4x4(vp9_raster_block_offset_int16(BLOCK_8X8, k, p->src_diff),
1733                      coeff, 8);
1734 #if CONFIG_VP9_HIGHBITDEPTH
1735       vpx_highbd_quantize_b(coeff, 4 * 4, p->zbin, p->round, p->quant,
1736                             p->quant_shift, qcoeff, dqcoeff, pd->dequant, eob,
1737                             so->scan, so->iscan);
1738       thisdistortion += vp9_highbd_block_error_dispatch(
1739           coeff, BLOCK_OFFSET(pd->dqcoeff, k), 16, &ssz, bd);
1740 #else
1741       vpx_quantize_b(coeff, 4 * 4, p->zbin, p->round, p->quant, p->quant_shift,
1742                      qcoeff, dqcoeff, pd->dequant, eob, so->scan, so->iscan);
1743       thisdistortion +=
1744           vp9_block_error(coeff, BLOCK_OFFSET(pd->dqcoeff, k), 16, &ssz);
1745 #endif  // CONFIG_VP9_HIGHBITDEPTH
1746       thissse += ssz;
1747       thisrate += cost_coeffs(x, 0, k, TX_4X4, coeff_ctx, so->scan,
1748                               so->neighbors, cpi->sf.use_fast_coef_costing);
1749       ta[k & 1] = tl[k >> 1] = (x->plane[0].eobs[k] > 0) ? 1 : 0;
1750       rd1 = RDCOST(x->rdmult, x->rddiv, thisrate, thisdistortion >> 2);
1751       rd2 = RDCOST(x->rdmult, x->rddiv, 0, thissse >> 2);
1752       rd = VPXMIN(rd1, rd2);
1753       if (rd >= best_yrd) return INT64_MAX;
1754     }
1755   }
1756 
1757   *distortion = thisdistortion >> 2;
1758   *labelyrate = thisrate;
1759   *sse = thissse >> 2;
1760 
1761   return RDCOST(x->rdmult, x->rddiv, *labelyrate, *distortion);
1762 }
1763 #endif  // !CONFIG_REALTIME_ONLY
1764 
1765 typedef struct {
1766   int eobs;
1767   int brate;
1768   int byrate;
1769   int64_t bdist;
1770   int64_t bsse;
1771   int64_t brdcost;
1772   int_mv mvs[2];
1773   ENTROPY_CONTEXT ta[2];
1774   ENTROPY_CONTEXT tl[2];
1775 } SEG_RDSTAT;
1776 
1777 typedef struct {
1778   int_mv *ref_mv[2];
1779   int_mv mvp;
1780 
1781   int64_t segment_rd;
1782   int r;
1783   int64_t d;
1784   int64_t sse;
1785   int segment_yrate;
1786   PREDICTION_MODE modes[4];
1787   SEG_RDSTAT rdstat[4][INTER_MODES];
1788   int mvthresh;
1789 } BEST_SEG_INFO;
1790 
1791 #if !CONFIG_REALTIME_ONLY
mv_check_bounds(const MvLimits * mv_limits,const MV * mv)1792 static INLINE int mv_check_bounds(const MvLimits *mv_limits, const MV *mv) {
1793   return (mv->row >> 3) < mv_limits->row_min ||
1794          (mv->row >> 3) > mv_limits->row_max ||
1795          (mv->col >> 3) < mv_limits->col_min ||
1796          (mv->col >> 3) > mv_limits->col_max;
1797 }
1798 
mi_buf_shift(MACROBLOCK * x,int i)1799 static INLINE void mi_buf_shift(MACROBLOCK *x, int i) {
1800   MODE_INFO *const mi = x->e_mbd.mi[0];
1801   struct macroblock_plane *const p = &x->plane[0];
1802   struct macroblockd_plane *const pd = &x->e_mbd.plane[0];
1803 
1804   p->src.buf =
1805       &p->src.buf[vp9_raster_block_offset(BLOCK_8X8, i, p->src.stride)];
1806   assert(((intptr_t)pd->pre[0].buf & 0x7) == 0);
1807   pd->pre[0].buf =
1808       &pd->pre[0].buf[vp9_raster_block_offset(BLOCK_8X8, i, pd->pre[0].stride)];
1809   if (has_second_ref(mi))
1810     pd->pre[1].buf =
1811         &pd->pre[1]
1812              .buf[vp9_raster_block_offset(BLOCK_8X8, i, pd->pre[1].stride)];
1813 }
1814 
mi_buf_restore(MACROBLOCK * x,struct buf_2d orig_src,struct buf_2d orig_pre[2])1815 static INLINE void mi_buf_restore(MACROBLOCK *x, struct buf_2d orig_src,
1816                                   struct buf_2d orig_pre[2]) {
1817   MODE_INFO *mi = x->e_mbd.mi[0];
1818   x->plane[0].src = orig_src;
1819   x->e_mbd.plane[0].pre[0] = orig_pre[0];
1820   if (has_second_ref(mi)) x->e_mbd.plane[0].pre[1] = orig_pre[1];
1821 }
1822 
mv_has_subpel(const MV * mv)1823 static INLINE int mv_has_subpel(const MV *mv) {
1824   return (mv->row & 0x0F) || (mv->col & 0x0F);
1825 }
1826 
1827 // Check if NEARESTMV/NEARMV/ZEROMV is the cheapest way encode zero motion.
1828 // TODO(aconverse): Find out if this is still productive then clean up or remove
check_best_zero_mv(const VP9_COMP * cpi,const uint8_t mode_context[MAX_REF_FRAMES],int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES],int this_mode,const MV_REFERENCE_FRAME ref_frames[2])1829 static int check_best_zero_mv(const VP9_COMP *cpi,
1830                               const uint8_t mode_context[MAX_REF_FRAMES],
1831                               int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES],
1832                               int this_mode,
1833                               const MV_REFERENCE_FRAME ref_frames[2]) {
1834   if ((this_mode == NEARMV || this_mode == NEARESTMV || this_mode == ZEROMV) &&
1835       frame_mv[this_mode][ref_frames[0]].as_int == 0 &&
1836       (ref_frames[1] == NONE ||
1837        frame_mv[this_mode][ref_frames[1]].as_int == 0)) {
1838     int rfc = mode_context[ref_frames[0]];
1839     int c1 = cost_mv_ref(cpi, NEARMV, rfc);
1840     int c2 = cost_mv_ref(cpi, NEARESTMV, rfc);
1841     int c3 = cost_mv_ref(cpi, ZEROMV, rfc);
1842 
1843     if (this_mode == NEARMV) {
1844       if (c1 > c3) return 0;
1845     } else if (this_mode == NEARESTMV) {
1846       if (c2 > c3) return 0;
1847     } else {
1848       assert(this_mode == ZEROMV);
1849       if (ref_frames[1] == NONE) {
1850         if ((c3 >= c2 && frame_mv[NEARESTMV][ref_frames[0]].as_int == 0) ||
1851             (c3 >= c1 && frame_mv[NEARMV][ref_frames[0]].as_int == 0))
1852           return 0;
1853       } else {
1854         if ((c3 >= c2 && frame_mv[NEARESTMV][ref_frames[0]].as_int == 0 &&
1855              frame_mv[NEARESTMV][ref_frames[1]].as_int == 0) ||
1856             (c3 >= c1 && frame_mv[NEARMV][ref_frames[0]].as_int == 0 &&
1857              frame_mv[NEARMV][ref_frames[1]].as_int == 0))
1858           return 0;
1859       }
1860     }
1861   }
1862   return 1;
1863 }
1864 
joint_motion_search(VP9_COMP * cpi,MACROBLOCK * x,BLOCK_SIZE bsize,int_mv * frame_mv,int mi_row,int mi_col,int_mv single_newmv[MAX_REF_FRAMES],int * rate_mv)1865 static void joint_motion_search(VP9_COMP *cpi, MACROBLOCK *x, BLOCK_SIZE bsize,
1866                                 int_mv *frame_mv, int mi_row, int mi_col,
1867                                 int_mv single_newmv[MAX_REF_FRAMES],
1868                                 int *rate_mv) {
1869   const VP9_COMMON *const cm = &cpi->common;
1870   const int pw = 4 * num_4x4_blocks_wide_lookup[bsize];
1871   const int ph = 4 * num_4x4_blocks_high_lookup[bsize];
1872   MACROBLOCKD *xd = &x->e_mbd;
1873   MODE_INFO *mi = xd->mi[0];
1874   const int refs[2] = { mi->ref_frame[0],
1875                         mi->ref_frame[1] < 0 ? 0 : mi->ref_frame[1] };
1876   int_mv ref_mv[2];
1877   int ite, ref;
1878   const InterpKernel *kernel = vp9_filter_kernels[mi->interp_filter];
1879   struct scale_factors sf;
1880 
1881   // Do joint motion search in compound mode to get more accurate mv.
1882   struct buf_2d backup_yv12[2][MAX_MB_PLANE];
1883   uint32_t last_besterr[2] = { UINT_MAX, UINT_MAX };
1884   const YV12_BUFFER_CONFIG *const scaled_ref_frame[2] = {
1885     vp9_get_scaled_ref_frame(cpi, mi->ref_frame[0]),
1886     vp9_get_scaled_ref_frame(cpi, mi->ref_frame[1])
1887   };
1888 
1889 // Prediction buffer from second frame.
1890 #if CONFIG_VP9_HIGHBITDEPTH
1891   DECLARE_ALIGNED(16, uint16_t, second_pred_alloc_16[64 * 64]);
1892   uint8_t *second_pred;
1893 #else
1894   DECLARE_ALIGNED(16, uint8_t, second_pred[64 * 64]);
1895 #endif  // CONFIG_VP9_HIGHBITDEPTH
1896 
1897   for (ref = 0; ref < 2; ++ref) {
1898     ref_mv[ref] = x->mbmi_ext->ref_mvs[refs[ref]][0];
1899 
1900     if (scaled_ref_frame[ref]) {
1901       int i;
1902       // Swap out the reference frame for a version that's been scaled to
1903       // match the resolution of the current frame, allowing the existing
1904       // motion search code to be used without additional modifications.
1905       for (i = 0; i < MAX_MB_PLANE; i++)
1906         backup_yv12[ref][i] = xd->plane[i].pre[ref];
1907       vp9_setup_pre_planes(xd, ref, scaled_ref_frame[ref], mi_row, mi_col,
1908                            NULL);
1909     }
1910 
1911     frame_mv[refs[ref]].as_int = single_newmv[refs[ref]].as_int;
1912   }
1913 
1914 // Since we have scaled the reference frames to match the size of the current
1915 // frame we must use a unit scaling factor during mode selection.
1916 #if CONFIG_VP9_HIGHBITDEPTH
1917   vp9_setup_scale_factors_for_frame(&sf, cm->width, cm->height, cm->width,
1918                                     cm->height, cm->use_highbitdepth);
1919 #else
1920   vp9_setup_scale_factors_for_frame(&sf, cm->width, cm->height, cm->width,
1921                                     cm->height);
1922 #endif  // CONFIG_VP9_HIGHBITDEPTH
1923 
1924   // Allow joint search multiple times iteratively for each reference frame
1925   // and break out of the search loop if it couldn't find a better mv.
1926   for (ite = 0; ite < 4; ite++) {
1927     struct buf_2d ref_yv12[2];
1928     uint32_t bestsme = UINT_MAX;
1929     int sadpb = x->sadperbit16;
1930     MV tmp_mv;
1931     int search_range = 3;
1932 
1933     const MvLimits tmp_mv_limits = x->mv_limits;
1934     int id = ite % 2;  // Even iterations search in the first reference frame,
1935                        // odd iterations search in the second. The predictor
1936                        // found for the 'other' reference frame is factored in.
1937 
1938     // Initialized here because of compiler problem in Visual Studio.
1939     ref_yv12[0] = xd->plane[0].pre[0];
1940     ref_yv12[1] = xd->plane[0].pre[1];
1941 
1942 // Get the prediction block from the 'other' reference frame.
1943 #if CONFIG_VP9_HIGHBITDEPTH
1944     if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
1945       second_pred = CONVERT_TO_BYTEPTR(second_pred_alloc_16);
1946       vp9_highbd_build_inter_predictor(
1947           CONVERT_TO_SHORTPTR(ref_yv12[!id].buf), ref_yv12[!id].stride,
1948           second_pred_alloc_16, pw, &frame_mv[refs[!id]].as_mv, &sf, pw, ph, 0,
1949           kernel, MV_PRECISION_Q3, mi_col * MI_SIZE, mi_row * MI_SIZE, xd->bd);
1950     } else {
1951       second_pred = (uint8_t *)second_pred_alloc_16;
1952       vp9_build_inter_predictor(ref_yv12[!id].buf, ref_yv12[!id].stride,
1953                                 second_pred, pw, &frame_mv[refs[!id]].as_mv,
1954                                 &sf, pw, ph, 0, kernel, MV_PRECISION_Q3,
1955                                 mi_col * MI_SIZE, mi_row * MI_SIZE);
1956     }
1957 #else
1958     vp9_build_inter_predictor(ref_yv12[!id].buf, ref_yv12[!id].stride,
1959                               second_pred, pw, &frame_mv[refs[!id]].as_mv, &sf,
1960                               pw, ph, 0, kernel, MV_PRECISION_Q3,
1961                               mi_col * MI_SIZE, mi_row * MI_SIZE);
1962 #endif  // CONFIG_VP9_HIGHBITDEPTH
1963 
1964     // Do compound motion search on the current reference frame.
1965     if (id) xd->plane[0].pre[0] = ref_yv12[id];
1966     vp9_set_mv_search_range(&x->mv_limits, &ref_mv[id].as_mv);
1967 
1968     // Use the mv result from the single mode as mv predictor.
1969     tmp_mv = frame_mv[refs[id]].as_mv;
1970 
1971     tmp_mv.col >>= 3;
1972     tmp_mv.row >>= 3;
1973 
1974     // Small-range full-pixel motion search.
1975     bestsme = vp9_refining_search_8p_c(x, &tmp_mv, sadpb, search_range,
1976                                        &cpi->fn_ptr[bsize], &ref_mv[id].as_mv,
1977                                        second_pred);
1978     if (bestsme < UINT_MAX)
1979       bestsme = vp9_get_mvpred_av_var(x, &tmp_mv, &ref_mv[id].as_mv,
1980                                       second_pred, &cpi->fn_ptr[bsize], 1);
1981 
1982     x->mv_limits = tmp_mv_limits;
1983 
1984     if (bestsme < UINT_MAX) {
1985       uint32_t dis; /* TODO: use dis in distortion calculation later. */
1986       uint32_t sse;
1987       bestsme = cpi->find_fractional_mv_step(
1988           x, &tmp_mv, &ref_mv[id].as_mv, cpi->common.allow_high_precision_mv,
1989           x->errorperbit, &cpi->fn_ptr[bsize], 0,
1990           cpi->sf.mv.subpel_search_level, NULL, x->nmvjointcost, x->mvcost,
1991           &dis, &sse, second_pred, pw, ph, cpi->sf.use_accurate_subpel_search);
1992     }
1993 
1994     // Restore the pointer to the first (possibly scaled) prediction buffer.
1995     if (id) xd->plane[0].pre[0] = ref_yv12[0];
1996 
1997     if (bestsme < last_besterr[id]) {
1998       frame_mv[refs[id]].as_mv = tmp_mv;
1999       last_besterr[id] = bestsme;
2000     } else {
2001       break;
2002     }
2003   }
2004 
2005   *rate_mv = 0;
2006 
2007   for (ref = 0; ref < 2; ++ref) {
2008     if (scaled_ref_frame[ref]) {
2009       // Restore the prediction frame pointers to their unscaled versions.
2010       int i;
2011       for (i = 0; i < MAX_MB_PLANE; i++)
2012         xd->plane[i].pre[ref] = backup_yv12[ref][i];
2013     }
2014 
2015     *rate_mv += vp9_mv_bit_cost(&frame_mv[refs[ref]].as_mv,
2016                                 &x->mbmi_ext->ref_mvs[refs[ref]][0].as_mv,
2017                                 x->nmvjointcost, x->mvcost, MV_COST_WEIGHT);
2018   }
2019 }
2020 
rd_pick_best_sub8x8_mode(VP9_COMP * cpi,MACROBLOCK * x,int_mv * best_ref_mv,int_mv * second_best_ref_mv,int64_t best_rd,int * returntotrate,int * returnyrate,int64_t * returndistortion,int * skippable,int64_t * psse,int mvthresh,int_mv seg_mvs[4][MAX_REF_FRAMES],BEST_SEG_INFO * bsi_buf,int filter_idx,int mi_row,int mi_col)2021 static int64_t rd_pick_best_sub8x8_mode(
2022     VP9_COMP *cpi, MACROBLOCK *x, int_mv *best_ref_mv,
2023     int_mv *second_best_ref_mv, int64_t best_rd, int *returntotrate,
2024     int *returnyrate, int64_t *returndistortion, int *skippable, int64_t *psse,
2025     int mvthresh, int_mv seg_mvs[4][MAX_REF_FRAMES], BEST_SEG_INFO *bsi_buf,
2026     int filter_idx, int mi_row, int mi_col) {
2027   int i;
2028   BEST_SEG_INFO *bsi = bsi_buf + filter_idx;
2029   MACROBLOCKD *xd = &x->e_mbd;
2030   MODE_INFO *mi = xd->mi[0];
2031   int mode_idx;
2032   int k, br = 0, idx, idy;
2033   int64_t bd = 0, block_sse = 0;
2034   PREDICTION_MODE this_mode;
2035   VP9_COMMON *cm = &cpi->common;
2036   struct macroblock_plane *const p = &x->plane[0];
2037   struct macroblockd_plane *const pd = &xd->plane[0];
2038   const int label_count = 4;
2039   int64_t this_segment_rd = 0;
2040   int label_mv_thresh;
2041   int segmentyrate = 0;
2042   const BLOCK_SIZE bsize = mi->sb_type;
2043   const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
2044   const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
2045   const int pw = num_4x4_blocks_wide << 2;
2046   const int ph = num_4x4_blocks_high << 2;
2047   ENTROPY_CONTEXT t_above[2], t_left[2];
2048   int subpelmv = 1, have_ref = 0;
2049   SPEED_FEATURES *const sf = &cpi->sf;
2050   const int has_second_rf = has_second_ref(mi);
2051   const int inter_mode_mask = sf->inter_mode_mask[bsize];
2052   MB_MODE_INFO_EXT *const mbmi_ext = x->mbmi_ext;
2053 
2054   vp9_zero(*bsi);
2055 
2056   bsi->segment_rd = best_rd;
2057   bsi->ref_mv[0] = best_ref_mv;
2058   bsi->ref_mv[1] = second_best_ref_mv;
2059   bsi->mvp.as_int = best_ref_mv->as_int;
2060   bsi->mvthresh = mvthresh;
2061 
2062   for (i = 0; i < 4; i++) bsi->modes[i] = ZEROMV;
2063 
2064   memcpy(t_above, pd->above_context, sizeof(t_above));
2065   memcpy(t_left, pd->left_context, sizeof(t_left));
2066 
2067   // 64 makes this threshold really big effectively
2068   // making it so that we very rarely check mvs on
2069   // segments.   setting this to 1 would make mv thresh
2070   // roughly equal to what it is for macroblocks
2071   label_mv_thresh = 1 * bsi->mvthresh / label_count;
2072 
2073   // Segmentation method overheads
2074   for (idy = 0; idy < 2; idy += num_4x4_blocks_high) {
2075     for (idx = 0; idx < 2; idx += num_4x4_blocks_wide) {
2076       // TODO(jingning,rbultje): rewrite the rate-distortion optimization
2077       // loop for 4x4/4x8/8x4 block coding. to be replaced with new rd loop
2078       int_mv mode_mv[MB_MODE_COUNT][2];
2079       int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES];
2080       PREDICTION_MODE mode_selected = ZEROMV;
2081       int64_t best_rd = INT64_MAX;
2082       const int i = idy * 2 + idx;
2083       int ref;
2084 
2085       for (ref = 0; ref < 1 + has_second_rf; ++ref) {
2086         const MV_REFERENCE_FRAME frame = mi->ref_frame[ref];
2087         frame_mv[ZEROMV][frame].as_int = 0;
2088         vp9_append_sub8x8_mvs_for_idx(
2089             cm, xd, i, ref, mi_row, mi_col, &frame_mv[NEARESTMV][frame],
2090             &frame_mv[NEARMV][frame], mbmi_ext->mode_context);
2091       }
2092 
2093       // search for the best motion vector on this segment
2094       for (this_mode = NEARESTMV; this_mode <= NEWMV; ++this_mode) {
2095         const struct buf_2d orig_src = x->plane[0].src;
2096         struct buf_2d orig_pre[2];
2097 
2098         mode_idx = INTER_OFFSET(this_mode);
2099         bsi->rdstat[i][mode_idx].brdcost = INT64_MAX;
2100         if (!(inter_mode_mask & (1 << this_mode))) continue;
2101 
2102         if (!check_best_zero_mv(cpi, mbmi_ext->mode_context, frame_mv,
2103                                 this_mode, mi->ref_frame))
2104           continue;
2105 
2106         memcpy(orig_pre, pd->pre, sizeof(orig_pre));
2107         memcpy(bsi->rdstat[i][mode_idx].ta, t_above,
2108                sizeof(bsi->rdstat[i][mode_idx].ta));
2109         memcpy(bsi->rdstat[i][mode_idx].tl, t_left,
2110                sizeof(bsi->rdstat[i][mode_idx].tl));
2111 
2112         // motion search for newmv (single predictor case only)
2113         if (!has_second_rf && this_mode == NEWMV &&
2114             seg_mvs[i][mi->ref_frame[0]].as_int == INVALID_MV) {
2115           MV *const new_mv = &mode_mv[NEWMV][0].as_mv;
2116           int step_param = 0;
2117           uint32_t bestsme = UINT_MAX;
2118           int sadpb = x->sadperbit4;
2119           MV mvp_full;
2120           int max_mv;
2121           int cost_list[5];
2122           const MvLimits tmp_mv_limits = x->mv_limits;
2123 
2124           /* Is the best so far sufficiently good that we cant justify doing
2125            * and new motion search. */
2126           if (best_rd < label_mv_thresh) break;
2127 
2128           if (cpi->oxcf.mode != BEST) {
2129             // use previous block's result as next block's MV predictor.
2130             if (i > 0) {
2131               bsi->mvp.as_int = mi->bmi[i - 1].as_mv[0].as_int;
2132               if (i == 2) bsi->mvp.as_int = mi->bmi[i - 2].as_mv[0].as_int;
2133             }
2134           }
2135           if (i == 0)
2136             max_mv = x->max_mv_context[mi->ref_frame[0]];
2137           else
2138             max_mv =
2139                 VPXMAX(abs(bsi->mvp.as_mv.row), abs(bsi->mvp.as_mv.col)) >> 3;
2140 
2141           if (sf->mv.auto_mv_step_size && cm->show_frame) {
2142             // Take wtd average of the step_params based on the last frame's
2143             // max mv magnitude and the best ref mvs of the current block for
2144             // the given reference.
2145             step_param =
2146                 (vp9_init_search_range(max_mv) + cpi->mv_step_param) / 2;
2147           } else {
2148             step_param = cpi->mv_step_param;
2149           }
2150 
2151           mvp_full.row = bsi->mvp.as_mv.row >> 3;
2152           mvp_full.col = bsi->mvp.as_mv.col >> 3;
2153 
2154           if (sf->adaptive_motion_search) {
2155             if (x->pred_mv[mi->ref_frame[0]].row != INT16_MAX &&
2156                 x->pred_mv[mi->ref_frame[0]].col != INT16_MAX) {
2157               mvp_full.row = x->pred_mv[mi->ref_frame[0]].row >> 3;
2158               mvp_full.col = x->pred_mv[mi->ref_frame[0]].col >> 3;
2159             }
2160             step_param = VPXMAX(step_param, 8);
2161           }
2162 
2163           // adjust src pointer for this block
2164           mi_buf_shift(x, i);
2165 
2166           vp9_set_mv_search_range(&x->mv_limits, &bsi->ref_mv[0]->as_mv);
2167 
2168           bestsme = vp9_full_pixel_search(
2169               cpi, x, bsize, &mvp_full, step_param, cpi->sf.mv.search_method,
2170               sadpb,
2171               sf->mv.subpel_search_method != SUBPEL_TREE ? cost_list : NULL,
2172               &bsi->ref_mv[0]->as_mv, new_mv, INT_MAX, 1);
2173 
2174           x->mv_limits = tmp_mv_limits;
2175 
2176           if (bestsme < UINT_MAX) {
2177             uint32_t distortion;
2178             cpi->find_fractional_mv_step(
2179                 x, new_mv, &bsi->ref_mv[0]->as_mv, cm->allow_high_precision_mv,
2180                 x->errorperbit, &cpi->fn_ptr[bsize], sf->mv.subpel_force_stop,
2181                 sf->mv.subpel_search_level, cond_cost_list(cpi, cost_list),
2182                 x->nmvjointcost, x->mvcost, &distortion,
2183                 &x->pred_sse[mi->ref_frame[0]], NULL, pw, ph,
2184                 cpi->sf.use_accurate_subpel_search);
2185 
2186             // save motion search result for use in compound prediction
2187             seg_mvs[i][mi->ref_frame[0]].as_mv = *new_mv;
2188           }
2189 
2190           x->pred_mv[mi->ref_frame[0]] = *new_mv;
2191 
2192           // restore src pointers
2193           mi_buf_restore(x, orig_src, orig_pre);
2194         }
2195 
2196         if (has_second_rf) {
2197           if (seg_mvs[i][mi->ref_frame[1]].as_int == INVALID_MV ||
2198               seg_mvs[i][mi->ref_frame[0]].as_int == INVALID_MV)
2199             continue;
2200         }
2201 
2202         if (has_second_rf && this_mode == NEWMV &&
2203             mi->interp_filter == EIGHTTAP) {
2204           // adjust src pointers
2205           mi_buf_shift(x, i);
2206           if (sf->comp_inter_joint_search_thresh <= bsize) {
2207             int rate_mv;
2208             joint_motion_search(cpi, x, bsize, frame_mv[this_mode], mi_row,
2209                                 mi_col, seg_mvs[i], &rate_mv);
2210             seg_mvs[i][mi->ref_frame[0]].as_int =
2211                 frame_mv[this_mode][mi->ref_frame[0]].as_int;
2212             seg_mvs[i][mi->ref_frame[1]].as_int =
2213                 frame_mv[this_mode][mi->ref_frame[1]].as_int;
2214           }
2215           // restore src pointers
2216           mi_buf_restore(x, orig_src, orig_pre);
2217         }
2218 
2219         bsi->rdstat[i][mode_idx].brate = set_and_cost_bmi_mvs(
2220             cpi, x, xd, i, this_mode, mode_mv[this_mode], frame_mv, seg_mvs[i],
2221             bsi->ref_mv, x->nmvjointcost, x->mvcost);
2222 
2223         for (ref = 0; ref < 1 + has_second_rf; ++ref) {
2224           bsi->rdstat[i][mode_idx].mvs[ref].as_int =
2225               mode_mv[this_mode][ref].as_int;
2226           if (num_4x4_blocks_wide > 1)
2227             bsi->rdstat[i + 1][mode_idx].mvs[ref].as_int =
2228                 mode_mv[this_mode][ref].as_int;
2229           if (num_4x4_blocks_high > 1)
2230             bsi->rdstat[i + 2][mode_idx].mvs[ref].as_int =
2231                 mode_mv[this_mode][ref].as_int;
2232         }
2233 
2234         // Trap vectors that reach beyond the UMV borders
2235         if (mv_check_bounds(&x->mv_limits, &mode_mv[this_mode][0].as_mv) ||
2236             (has_second_rf &&
2237              mv_check_bounds(&x->mv_limits, &mode_mv[this_mode][1].as_mv)))
2238           continue;
2239 
2240         if (filter_idx > 0) {
2241           BEST_SEG_INFO *ref_bsi = bsi_buf;
2242           subpelmv = 0;
2243           have_ref = 1;
2244 
2245           for (ref = 0; ref < 1 + has_second_rf; ++ref) {
2246             subpelmv |= mv_has_subpel(&mode_mv[this_mode][ref].as_mv);
2247             have_ref &= mode_mv[this_mode][ref].as_int ==
2248                         ref_bsi->rdstat[i][mode_idx].mvs[ref].as_int;
2249           }
2250 
2251           if (filter_idx > 1 && !subpelmv && !have_ref) {
2252             ref_bsi = bsi_buf + 1;
2253             have_ref = 1;
2254             for (ref = 0; ref < 1 + has_second_rf; ++ref)
2255               have_ref &= mode_mv[this_mode][ref].as_int ==
2256                           ref_bsi->rdstat[i][mode_idx].mvs[ref].as_int;
2257           }
2258 
2259           if (!subpelmv && have_ref &&
2260               ref_bsi->rdstat[i][mode_idx].brdcost < INT64_MAX) {
2261             memcpy(&bsi->rdstat[i][mode_idx], &ref_bsi->rdstat[i][mode_idx],
2262                    sizeof(SEG_RDSTAT));
2263             if (num_4x4_blocks_wide > 1)
2264               bsi->rdstat[i + 1][mode_idx].eobs =
2265                   ref_bsi->rdstat[i + 1][mode_idx].eobs;
2266             if (num_4x4_blocks_high > 1)
2267               bsi->rdstat[i + 2][mode_idx].eobs =
2268                   ref_bsi->rdstat[i + 2][mode_idx].eobs;
2269 
2270             if (bsi->rdstat[i][mode_idx].brdcost < best_rd) {
2271               mode_selected = this_mode;
2272               best_rd = bsi->rdstat[i][mode_idx].brdcost;
2273             }
2274             continue;
2275           }
2276         }
2277 
2278         bsi->rdstat[i][mode_idx].brdcost = encode_inter_mb_segment(
2279             cpi, x, bsi->segment_rd - this_segment_rd, i,
2280             &bsi->rdstat[i][mode_idx].byrate, &bsi->rdstat[i][mode_idx].bdist,
2281             &bsi->rdstat[i][mode_idx].bsse, bsi->rdstat[i][mode_idx].ta,
2282             bsi->rdstat[i][mode_idx].tl, mi_row, mi_col);
2283         if (bsi->rdstat[i][mode_idx].brdcost < INT64_MAX) {
2284           bsi->rdstat[i][mode_idx].brdcost +=
2285               RDCOST(x->rdmult, x->rddiv, bsi->rdstat[i][mode_idx].brate, 0);
2286           bsi->rdstat[i][mode_idx].brate += bsi->rdstat[i][mode_idx].byrate;
2287           bsi->rdstat[i][mode_idx].eobs = p->eobs[i];
2288           if (num_4x4_blocks_wide > 1)
2289             bsi->rdstat[i + 1][mode_idx].eobs = p->eobs[i + 1];
2290           if (num_4x4_blocks_high > 1)
2291             bsi->rdstat[i + 2][mode_idx].eobs = p->eobs[i + 2];
2292         }
2293 
2294         if (bsi->rdstat[i][mode_idx].brdcost < best_rd) {
2295           mode_selected = this_mode;
2296           best_rd = bsi->rdstat[i][mode_idx].brdcost;
2297         }
2298       } /*for each 4x4 mode*/
2299 
2300       if (best_rd == INT64_MAX) {
2301         int iy, midx;
2302         for (iy = i + 1; iy < 4; ++iy)
2303           for (midx = 0; midx < INTER_MODES; ++midx)
2304             bsi->rdstat[iy][midx].brdcost = INT64_MAX;
2305         bsi->segment_rd = INT64_MAX;
2306         return INT64_MAX;
2307       }
2308 
2309       mode_idx = INTER_OFFSET(mode_selected);
2310       memcpy(t_above, bsi->rdstat[i][mode_idx].ta, sizeof(t_above));
2311       memcpy(t_left, bsi->rdstat[i][mode_idx].tl, sizeof(t_left));
2312 
2313       set_and_cost_bmi_mvs(cpi, x, xd, i, mode_selected, mode_mv[mode_selected],
2314                            frame_mv, seg_mvs[i], bsi->ref_mv, x->nmvjointcost,
2315                            x->mvcost);
2316 
2317       br += bsi->rdstat[i][mode_idx].brate;
2318       bd += bsi->rdstat[i][mode_idx].bdist;
2319       block_sse += bsi->rdstat[i][mode_idx].bsse;
2320       segmentyrate += bsi->rdstat[i][mode_idx].byrate;
2321       this_segment_rd += bsi->rdstat[i][mode_idx].brdcost;
2322 
2323       if (this_segment_rd > bsi->segment_rd) {
2324         int iy, midx;
2325         for (iy = i + 1; iy < 4; ++iy)
2326           for (midx = 0; midx < INTER_MODES; ++midx)
2327             bsi->rdstat[iy][midx].brdcost = INT64_MAX;
2328         bsi->segment_rd = INT64_MAX;
2329         return INT64_MAX;
2330       }
2331     }
2332   } /* for each label */
2333 
2334   bsi->r = br;
2335   bsi->d = bd;
2336   bsi->segment_yrate = segmentyrate;
2337   bsi->segment_rd = this_segment_rd;
2338   bsi->sse = block_sse;
2339 
2340   // update the coding decisions
2341   for (k = 0; k < 4; ++k) bsi->modes[k] = mi->bmi[k].as_mode;
2342 
2343   if (bsi->segment_rd > best_rd) return INT64_MAX;
2344   /* set it to the best */
2345   for (i = 0; i < 4; i++) {
2346     mode_idx = INTER_OFFSET(bsi->modes[i]);
2347     mi->bmi[i].as_mv[0].as_int = bsi->rdstat[i][mode_idx].mvs[0].as_int;
2348     if (has_second_ref(mi))
2349       mi->bmi[i].as_mv[1].as_int = bsi->rdstat[i][mode_idx].mvs[1].as_int;
2350     x->plane[0].eobs[i] = bsi->rdstat[i][mode_idx].eobs;
2351     mi->bmi[i].as_mode = bsi->modes[i];
2352   }
2353 
2354   /*
2355    * used to set mbmi->mv.as_int
2356    */
2357   *returntotrate = bsi->r;
2358   *returndistortion = bsi->d;
2359   *returnyrate = bsi->segment_yrate;
2360   *skippable = vp9_is_skippable_in_plane(x, BLOCK_8X8, 0);
2361   *psse = bsi->sse;
2362   mi->mode = bsi->modes[3];
2363 
2364   return bsi->segment_rd;
2365 }
2366 
estimate_ref_frame_costs(const VP9_COMMON * cm,const MACROBLOCKD * xd,int segment_id,unsigned int * ref_costs_single,unsigned int * ref_costs_comp,vpx_prob * comp_mode_p)2367 static void estimate_ref_frame_costs(const VP9_COMMON *cm,
2368                                      const MACROBLOCKD *xd, int segment_id,
2369                                      unsigned int *ref_costs_single,
2370                                      unsigned int *ref_costs_comp,
2371                                      vpx_prob *comp_mode_p) {
2372   int seg_ref_active =
2373       segfeature_active(&cm->seg, segment_id, SEG_LVL_REF_FRAME);
2374   if (seg_ref_active) {
2375     memset(ref_costs_single, 0, MAX_REF_FRAMES * sizeof(*ref_costs_single));
2376     memset(ref_costs_comp, 0, MAX_REF_FRAMES * sizeof(*ref_costs_comp));
2377     *comp_mode_p = 128;
2378   } else {
2379     vpx_prob intra_inter_p = vp9_get_intra_inter_prob(cm, xd);
2380     vpx_prob comp_inter_p = 128;
2381 
2382     if (cm->reference_mode == REFERENCE_MODE_SELECT) {
2383       comp_inter_p = vp9_get_reference_mode_prob(cm, xd);
2384       *comp_mode_p = comp_inter_p;
2385     } else {
2386       *comp_mode_p = 128;
2387     }
2388 
2389     ref_costs_single[INTRA_FRAME] = vp9_cost_bit(intra_inter_p, 0);
2390 
2391     if (cm->reference_mode != COMPOUND_REFERENCE) {
2392       vpx_prob ref_single_p1 = vp9_get_pred_prob_single_ref_p1(cm, xd);
2393       vpx_prob ref_single_p2 = vp9_get_pred_prob_single_ref_p2(cm, xd);
2394       unsigned int base_cost = vp9_cost_bit(intra_inter_p, 1);
2395 
2396       if (cm->reference_mode == REFERENCE_MODE_SELECT)
2397         base_cost += vp9_cost_bit(comp_inter_p, 0);
2398 
2399       ref_costs_single[LAST_FRAME] = ref_costs_single[GOLDEN_FRAME] =
2400           ref_costs_single[ALTREF_FRAME] = base_cost;
2401       ref_costs_single[LAST_FRAME] += vp9_cost_bit(ref_single_p1, 0);
2402       ref_costs_single[GOLDEN_FRAME] += vp9_cost_bit(ref_single_p1, 1);
2403       ref_costs_single[ALTREF_FRAME] += vp9_cost_bit(ref_single_p1, 1);
2404       ref_costs_single[GOLDEN_FRAME] += vp9_cost_bit(ref_single_p2, 0);
2405       ref_costs_single[ALTREF_FRAME] += vp9_cost_bit(ref_single_p2, 1);
2406     } else {
2407       ref_costs_single[LAST_FRAME] = 512;
2408       ref_costs_single[GOLDEN_FRAME] = 512;
2409       ref_costs_single[ALTREF_FRAME] = 512;
2410     }
2411     if (cm->reference_mode != SINGLE_REFERENCE) {
2412       vpx_prob ref_comp_p = vp9_get_pred_prob_comp_ref_p(cm, xd);
2413       unsigned int base_cost = vp9_cost_bit(intra_inter_p, 1);
2414 
2415       if (cm->reference_mode == REFERENCE_MODE_SELECT)
2416         base_cost += vp9_cost_bit(comp_inter_p, 1);
2417 
2418       ref_costs_comp[LAST_FRAME] = base_cost + vp9_cost_bit(ref_comp_p, 0);
2419       ref_costs_comp[GOLDEN_FRAME] = base_cost + vp9_cost_bit(ref_comp_p, 1);
2420     } else {
2421       ref_costs_comp[LAST_FRAME] = 512;
2422       ref_costs_comp[GOLDEN_FRAME] = 512;
2423     }
2424   }
2425 }
2426 
store_coding_context(MACROBLOCK * x,PICK_MODE_CONTEXT * ctx,int mode_index,int64_t comp_pred_diff[REFERENCE_MODES],int64_t best_filter_diff[SWITCHABLE_FILTER_CONTEXTS],int skippable)2427 static void store_coding_context(
2428     MACROBLOCK *x, PICK_MODE_CONTEXT *ctx, int mode_index,
2429     int64_t comp_pred_diff[REFERENCE_MODES],
2430     int64_t best_filter_diff[SWITCHABLE_FILTER_CONTEXTS], int skippable) {
2431   MACROBLOCKD *const xd = &x->e_mbd;
2432 
2433   // Take a snapshot of the coding context so it can be
2434   // restored if we decide to encode this way
2435   ctx->skip = x->skip;
2436   ctx->skippable = skippable;
2437   ctx->best_mode_index = mode_index;
2438   ctx->mic = *xd->mi[0];
2439   ctx->mbmi_ext = *x->mbmi_ext;
2440   ctx->single_pred_diff = (int)comp_pred_diff[SINGLE_REFERENCE];
2441   ctx->comp_pred_diff = (int)comp_pred_diff[COMPOUND_REFERENCE];
2442   ctx->hybrid_pred_diff = (int)comp_pred_diff[REFERENCE_MODE_SELECT];
2443 
2444   memcpy(ctx->best_filter_diff, best_filter_diff,
2445          sizeof(*best_filter_diff) * SWITCHABLE_FILTER_CONTEXTS);
2446 }
2447 
setup_buffer_inter(VP9_COMP * cpi,MACROBLOCK * x,MV_REFERENCE_FRAME ref_frame,BLOCK_SIZE block_size,int mi_row,int mi_col,int_mv frame_nearest_mv[MAX_REF_FRAMES],int_mv frame_near_mv[MAX_REF_FRAMES],struct buf_2d yv12_mb[4][MAX_MB_PLANE])2448 static void setup_buffer_inter(VP9_COMP *cpi, MACROBLOCK *x,
2449                                MV_REFERENCE_FRAME ref_frame,
2450                                BLOCK_SIZE block_size, int mi_row, int mi_col,
2451                                int_mv frame_nearest_mv[MAX_REF_FRAMES],
2452                                int_mv frame_near_mv[MAX_REF_FRAMES],
2453                                struct buf_2d yv12_mb[4][MAX_MB_PLANE]) {
2454   const VP9_COMMON *cm = &cpi->common;
2455   const YV12_BUFFER_CONFIG *yv12 = get_ref_frame_buffer(cpi, ref_frame);
2456   MACROBLOCKD *const xd = &x->e_mbd;
2457   MODE_INFO *const mi = xd->mi[0];
2458   int_mv *const candidates = x->mbmi_ext->ref_mvs[ref_frame];
2459   const struct scale_factors *const sf = &cm->frame_refs[ref_frame - 1].sf;
2460   MB_MODE_INFO_EXT *const mbmi_ext = x->mbmi_ext;
2461 
2462   assert(yv12 != NULL);
2463 
2464   // TODO(jkoleszar): Is the UV buffer ever used here? If so, need to make this
2465   // use the UV scaling factors.
2466   vp9_setup_pred_block(xd, yv12_mb[ref_frame], yv12, mi_row, mi_col, sf, sf);
2467 
2468   // Gets an initial list of candidate vectors from neighbours and orders them
2469   vp9_find_mv_refs(cm, xd, mi, ref_frame, candidates, mi_row, mi_col,
2470                    mbmi_ext->mode_context);
2471 
2472   // Candidate refinement carried out at encoder and decoder
2473   vp9_find_best_ref_mvs(xd, cm->allow_high_precision_mv, candidates,
2474                         &frame_nearest_mv[ref_frame],
2475                         &frame_near_mv[ref_frame]);
2476 
2477   // Further refinement that is encode side only to test the top few candidates
2478   // in full and choose the best as the centre point for subsequent searches.
2479   // The current implementation doesn't support scaling.
2480   if (!vp9_is_scaled(sf) && block_size >= BLOCK_8X8)
2481     vp9_mv_pred(cpi, x, yv12_mb[ref_frame][0].buf, yv12->y_stride, ref_frame,
2482                 block_size);
2483 }
2484 
2485 #if CONFIG_NON_GREEDY_MV
ref_frame_to_gf_rf_idx(int ref_frame)2486 static int ref_frame_to_gf_rf_idx(int ref_frame) {
2487   if (ref_frame == GOLDEN_FRAME) {
2488     return 0;
2489   }
2490   if (ref_frame == LAST_FRAME) {
2491     return 1;
2492   }
2493   if (ref_frame == ALTREF_FRAME) {
2494     return 2;
2495   }
2496   assert(0);
2497   return -1;
2498 }
2499 #endif
2500 
single_motion_search(VP9_COMP * cpi,MACROBLOCK * x,BLOCK_SIZE bsize,int mi_row,int mi_col,int_mv * tmp_mv,int * rate_mv)2501 static void single_motion_search(VP9_COMP *cpi, MACROBLOCK *x, BLOCK_SIZE bsize,
2502                                  int mi_row, int mi_col, int_mv *tmp_mv,
2503                                  int *rate_mv) {
2504   MACROBLOCKD *xd = &x->e_mbd;
2505   const VP9_COMMON *cm = &cpi->common;
2506   MODE_INFO *mi = xd->mi[0];
2507   struct buf_2d backup_yv12[MAX_MB_PLANE] = { { 0, 0 } };
2508   int step_param;
2509   MV mvp_full;
2510   int ref = mi->ref_frame[0];
2511   MV ref_mv = x->mbmi_ext->ref_mvs[ref][0].as_mv;
2512   const MvLimits tmp_mv_limits = x->mv_limits;
2513   int cost_list[5];
2514   const int best_predmv_idx = x->mv_best_ref_index[ref];
2515   const YV12_BUFFER_CONFIG *scaled_ref_frame =
2516       vp9_get_scaled_ref_frame(cpi, ref);
2517   const int pw = num_4x4_blocks_wide_lookup[bsize] << 2;
2518   const int ph = num_4x4_blocks_high_lookup[bsize] << 2;
2519   MV pred_mv[3];
2520 
2521   int bestsme = INT_MAX;
2522 #if CONFIG_NON_GREEDY_MV
2523   int gf_group_idx = cpi->twopass.gf_group.index;
2524   int gf_rf_idx = ref_frame_to_gf_rf_idx(ref);
2525   BLOCK_SIZE square_bsize = get_square_block_size(bsize);
2526   int_mv nb_full_mvs[NB_MVS_NUM] = { 0 };
2527   MotionField *motion_field = vp9_motion_field_info_get_motion_field(
2528       &cpi->motion_field_info, gf_group_idx, gf_rf_idx, square_bsize);
2529   const int nb_full_mv_num =
2530       vp9_prepare_nb_full_mvs(motion_field, mi_row, mi_col, nb_full_mvs);
2531   const int lambda = (pw * ph) / 4;
2532   assert(pw * ph == lambda << 2);
2533 #else   // CONFIG_NON_GREEDY_MV
2534   int sadpb = x->sadperbit16;
2535 #endif  // CONFIG_NON_GREEDY_MV
2536 
2537   pred_mv[0] = x->mbmi_ext->ref_mvs[ref][0].as_mv;
2538   pred_mv[1] = x->mbmi_ext->ref_mvs[ref][1].as_mv;
2539   pred_mv[2] = x->pred_mv[ref];
2540 
2541   if (scaled_ref_frame) {
2542     int i;
2543     // Swap out the reference frame for a version that's been scaled to
2544     // match the resolution of the current frame, allowing the existing
2545     // motion search code to be used without additional modifications.
2546     for (i = 0; i < MAX_MB_PLANE; i++) backup_yv12[i] = xd->plane[i].pre[0];
2547 
2548     vp9_setup_pre_planes(xd, 0, scaled_ref_frame, mi_row, mi_col, NULL);
2549   }
2550 
2551   // Work out the size of the first step in the mv step search.
2552   // 0 here is maximum length first step. 1 is VPXMAX >> 1 etc.
2553   if (cpi->sf.mv.auto_mv_step_size && cm->show_frame) {
2554     // Take wtd average of the step_params based on the last frame's
2555     // max mv magnitude and that based on the best ref mvs of the current
2556     // block for the given reference.
2557     step_param =
2558         (vp9_init_search_range(x->max_mv_context[ref]) + cpi->mv_step_param) /
2559         2;
2560   } else {
2561     step_param = cpi->mv_step_param;
2562   }
2563 
2564   if (cpi->sf.adaptive_motion_search && bsize < BLOCK_64X64) {
2565     const int boffset =
2566         2 * (b_width_log2_lookup[BLOCK_64X64] -
2567              VPXMIN(b_height_log2_lookup[bsize], b_width_log2_lookup[bsize]));
2568     step_param = VPXMAX(step_param, boffset);
2569   }
2570 
2571   if (cpi->sf.adaptive_motion_search) {
2572     int bwl = b_width_log2_lookup[bsize];
2573     int bhl = b_height_log2_lookup[bsize];
2574     int tlevel = x->pred_mv_sad[ref] >> (bwl + bhl + 4);
2575 
2576     if (tlevel < 5) step_param += 2;
2577 
2578     // prev_mv_sad is not setup for dynamically scaled frames.
2579     if (cpi->oxcf.resize_mode != RESIZE_DYNAMIC) {
2580       int i;
2581       for (i = LAST_FRAME; i <= ALTREF_FRAME && cm->show_frame; ++i) {
2582         if ((x->pred_mv_sad[ref] >> 3) > x->pred_mv_sad[i]) {
2583           x->pred_mv[ref].row = INT16_MAX;
2584           x->pred_mv[ref].col = INT16_MAX;
2585           tmp_mv->as_int = INVALID_MV;
2586 
2587           if (scaled_ref_frame) {
2588             int i;
2589             for (i = 0; i < MAX_MB_PLANE; ++i)
2590               xd->plane[i].pre[0] = backup_yv12[i];
2591           }
2592           return;
2593         }
2594       }
2595     }
2596   }
2597 
2598   // Note: MV limits are modified here. Always restore the original values
2599   // after full-pixel motion search.
2600   vp9_set_mv_search_range(&x->mv_limits, &ref_mv);
2601 
2602   mvp_full = pred_mv[best_predmv_idx];
2603   mvp_full.col >>= 3;
2604   mvp_full.row >>= 3;
2605 
2606 #if CONFIG_NON_GREEDY_MV
2607   bestsme = vp9_full_pixel_diamond_new(cpi, x, bsize, &mvp_full, step_param,
2608                                        lambda, 1, nb_full_mvs, nb_full_mv_num,
2609                                        &tmp_mv->as_mv);
2610 #else   // CONFIG_NON_GREEDY_MV
2611   bestsme = vp9_full_pixel_search(
2612       cpi, x, bsize, &mvp_full, step_param, cpi->sf.mv.search_method, sadpb,
2613       cond_cost_list(cpi, cost_list), &ref_mv, &tmp_mv->as_mv, INT_MAX, 1);
2614 #endif  // CONFIG_NON_GREEDY_MV
2615 
2616   if (cpi->sf.enhanced_full_pixel_motion_search) {
2617     int i;
2618     for (i = 0; i < 3; ++i) {
2619       int this_me;
2620       MV this_mv;
2621       int diff_row;
2622       int diff_col;
2623       int step;
2624 
2625       if (pred_mv[i].row == INT16_MAX || pred_mv[i].col == INT16_MAX) continue;
2626       if (i == best_predmv_idx) continue;
2627 
2628       diff_row = ((int)pred_mv[i].row -
2629                   pred_mv[i > 0 ? (i - 1) : best_predmv_idx].row) >>
2630                  3;
2631       diff_col = ((int)pred_mv[i].col -
2632                   pred_mv[i > 0 ? (i - 1) : best_predmv_idx].col) >>
2633                  3;
2634       if (diff_row == 0 && diff_col == 0) continue;
2635       if (diff_row < 0) diff_row = -diff_row;
2636       if (diff_col < 0) diff_col = -diff_col;
2637       step = get_msb((diff_row + diff_col + 1) >> 1);
2638       if (step <= 0) continue;
2639 
2640       mvp_full = pred_mv[i];
2641       mvp_full.col >>= 3;
2642       mvp_full.row >>= 3;
2643 #if CONFIG_NON_GREEDY_MV
2644       this_me = vp9_full_pixel_diamond_new(
2645           cpi, x, bsize, &mvp_full,
2646           VPXMAX(step_param, MAX_MVSEARCH_STEPS - step), lambda, 1, nb_full_mvs,
2647           nb_full_mv_num, &this_mv);
2648 #else   // CONFIG_NON_GREEDY_MV
2649       this_me = vp9_full_pixel_search(
2650           cpi, x, bsize, &mvp_full,
2651           VPXMAX(step_param, MAX_MVSEARCH_STEPS - step),
2652           cpi->sf.mv.search_method, sadpb, cond_cost_list(cpi, cost_list),
2653           &ref_mv, &this_mv, INT_MAX, 1);
2654 #endif  // CONFIG_NON_GREEDY_MV
2655       if (this_me < bestsme) {
2656         tmp_mv->as_mv = this_mv;
2657         bestsme = this_me;
2658       }
2659     }
2660   }
2661 
2662   x->mv_limits = tmp_mv_limits;
2663 
2664   if (bestsme < INT_MAX) {
2665     uint32_t dis; /* TODO: use dis in distortion calculation later. */
2666     cpi->find_fractional_mv_step(
2667         x, &tmp_mv->as_mv, &ref_mv, cm->allow_high_precision_mv, x->errorperbit,
2668         &cpi->fn_ptr[bsize], cpi->sf.mv.subpel_force_stop,
2669         cpi->sf.mv.subpel_search_level, cond_cost_list(cpi, cost_list),
2670         x->nmvjointcost, x->mvcost, &dis, &x->pred_sse[ref], NULL, pw, ph,
2671         cpi->sf.use_accurate_subpel_search);
2672   }
2673   *rate_mv = vp9_mv_bit_cost(&tmp_mv->as_mv, &ref_mv, x->nmvjointcost,
2674                              x->mvcost, MV_COST_WEIGHT);
2675 
2676   x->pred_mv[ref] = tmp_mv->as_mv;
2677 
2678   if (scaled_ref_frame) {
2679     int i;
2680     for (i = 0; i < MAX_MB_PLANE; i++) xd->plane[i].pre[0] = backup_yv12[i];
2681   }
2682 }
2683 
restore_dst_buf(MACROBLOCKD * xd,uint8_t * orig_dst[MAX_MB_PLANE],int orig_dst_stride[MAX_MB_PLANE])2684 static INLINE void restore_dst_buf(MACROBLOCKD *xd,
2685                                    uint8_t *orig_dst[MAX_MB_PLANE],
2686                                    int orig_dst_stride[MAX_MB_PLANE]) {
2687   int i;
2688   for (i = 0; i < MAX_MB_PLANE; i++) {
2689     xd->plane[i].dst.buf = orig_dst[i];
2690     xd->plane[i].dst.stride = orig_dst_stride[i];
2691   }
2692 }
2693 
2694 // In some situations we want to discount tha pparent cost of a new motion
2695 // vector. Where there is a subtle motion field and especially where there is
2696 // low spatial complexity then it can be hard to cover the cost of a new motion
2697 // vector in a single block, even if that motion vector reduces distortion.
2698 // However, once established that vector may be usable through the nearest and
2699 // near mv modes to reduce distortion in subsequent blocks and also improve
2700 // visual quality.
discount_newmv_test(VP9_COMP * cpi,int this_mode,int_mv this_mv,int_mv (* mode_mv)[MAX_REF_FRAMES],int ref_frame,int mi_row,int mi_col,BLOCK_SIZE bsize)2701 static int discount_newmv_test(VP9_COMP *cpi, int this_mode, int_mv this_mv,
2702                                int_mv (*mode_mv)[MAX_REF_FRAMES], int ref_frame,
2703                                int mi_row, int mi_col, BLOCK_SIZE bsize) {
2704 #if CONFIG_NON_GREEDY_MV
2705   (void)mode_mv;
2706   (void)this_mv;
2707   if (this_mode == NEWMV && bsize >= BLOCK_8X8 && cpi->tpl_ready) {
2708     const int gf_group_idx = cpi->twopass.gf_group.index;
2709     const int gf_rf_idx = ref_frame_to_gf_rf_idx(ref_frame);
2710     const TplDepFrame tpl_frame = cpi->tpl_stats[gf_group_idx];
2711     const MotionField *motion_field = vp9_motion_field_info_get_motion_field(
2712         &cpi->motion_field_info, gf_group_idx, gf_rf_idx, cpi->tpl_bsize);
2713     const int tpl_block_mi_h = num_8x8_blocks_high_lookup[cpi->tpl_bsize];
2714     const int tpl_block_mi_w = num_8x8_blocks_wide_lookup[cpi->tpl_bsize];
2715     const int tpl_mi_row = mi_row - (mi_row % tpl_block_mi_h);
2716     const int tpl_mi_col = mi_col - (mi_col % tpl_block_mi_w);
2717     const int mv_mode =
2718         tpl_frame
2719             .mv_mode_arr[gf_rf_idx][tpl_mi_row * tpl_frame.stride + tpl_mi_col];
2720     if (mv_mode == NEW_MV_MODE) {
2721       int_mv tpl_new_mv =
2722           vp9_motion_field_mi_get_mv(motion_field, tpl_mi_row, tpl_mi_col);
2723       int row_diff = abs(tpl_new_mv.as_mv.row - this_mv.as_mv.row);
2724       int col_diff = abs(tpl_new_mv.as_mv.col - this_mv.as_mv.col);
2725       if (VPXMAX(row_diff, col_diff) <= 8) {
2726         return 1;
2727       } else {
2728         return 0;
2729       }
2730     } else {
2731       return 0;
2732     }
2733   } else {
2734     return 0;
2735   }
2736 #else
2737   (void)mi_row;
2738   (void)mi_col;
2739   (void)bsize;
2740   return (!cpi->rc.is_src_frame_alt_ref && (this_mode == NEWMV) &&
2741           (this_mv.as_int != 0) &&
2742           ((mode_mv[NEARESTMV][ref_frame].as_int == 0) ||
2743            (mode_mv[NEARESTMV][ref_frame].as_int == INVALID_MV)) &&
2744           ((mode_mv[NEARMV][ref_frame].as_int == 0) ||
2745            (mode_mv[NEARMV][ref_frame].as_int == INVALID_MV)));
2746 #endif
2747 }
2748 
handle_inter_mode(VP9_COMP * cpi,MACROBLOCK * x,BLOCK_SIZE bsize,int * rate2,int64_t * distortion,int * skippable,int * rate_y,int * rate_uv,struct buf_2d * recon,int * disable_skip,int_mv (* mode_mv)[MAX_REF_FRAMES],int mi_row,int mi_col,int_mv single_newmv[MAX_REF_FRAMES],INTERP_FILTER (* single_filter)[MAX_REF_FRAMES],int (* single_skippable)[MAX_REF_FRAMES],int64_t * psse,const int64_t ref_best_rd,int64_t * mask_filter,int64_t filter_cache[])2749 static int64_t handle_inter_mode(
2750     VP9_COMP *cpi, MACROBLOCK *x, BLOCK_SIZE bsize, int *rate2,
2751     int64_t *distortion, int *skippable, int *rate_y, int *rate_uv,
2752     struct buf_2d *recon, int *disable_skip, int_mv (*mode_mv)[MAX_REF_FRAMES],
2753     int mi_row, int mi_col, int_mv single_newmv[MAX_REF_FRAMES],
2754     INTERP_FILTER (*single_filter)[MAX_REF_FRAMES],
2755     int (*single_skippable)[MAX_REF_FRAMES], int64_t *psse,
2756     const int64_t ref_best_rd, int64_t *mask_filter, int64_t filter_cache[]) {
2757   VP9_COMMON *cm = &cpi->common;
2758   MACROBLOCKD *xd = &x->e_mbd;
2759   MODE_INFO *mi = xd->mi[0];
2760   MB_MODE_INFO_EXT *const mbmi_ext = x->mbmi_ext;
2761   const int is_comp_pred = has_second_ref(mi);
2762   const int this_mode = mi->mode;
2763   int_mv *frame_mv = mode_mv[this_mode];
2764   int i;
2765   int refs[2] = { mi->ref_frame[0],
2766                   (mi->ref_frame[1] < 0 ? 0 : mi->ref_frame[1]) };
2767   int_mv cur_mv[2];
2768 #if CONFIG_VP9_HIGHBITDEPTH
2769   DECLARE_ALIGNED(16, uint16_t, tmp_buf16[MAX_MB_PLANE * 64 * 64]);
2770   uint8_t *tmp_buf;
2771 #else
2772   DECLARE_ALIGNED(16, uint8_t, tmp_buf[MAX_MB_PLANE * 64 * 64]);
2773 #endif  // CONFIG_VP9_HIGHBITDEPTH
2774   int pred_exists = 0;
2775   int intpel_mv;
2776   int64_t rd, tmp_rd, best_rd = INT64_MAX;
2777   int best_needs_copy = 0;
2778   uint8_t *orig_dst[MAX_MB_PLANE];
2779   int orig_dst_stride[MAX_MB_PLANE];
2780   int rs = 0;
2781   INTERP_FILTER best_filter = SWITCHABLE;
2782   uint8_t skip_txfm[MAX_MB_PLANE << 2] = { 0 };
2783   int64_t bsse[MAX_MB_PLANE << 2] = { 0 };
2784 
2785   int bsl = mi_width_log2_lookup[bsize];
2786   int pred_filter_search =
2787       cpi->sf.cb_pred_filter_search
2788           ? (((mi_row + mi_col) >> bsl) +
2789              get_chessboard_index(cm->current_video_frame)) &
2790                 0x1
2791           : 0;
2792 
2793   int skip_txfm_sb = 0;
2794   int64_t skip_sse_sb = INT64_MAX;
2795   int64_t distortion_y = 0, distortion_uv = 0;
2796 
2797 #if CONFIG_VP9_HIGHBITDEPTH
2798   if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
2799     tmp_buf = CONVERT_TO_BYTEPTR(tmp_buf16);
2800   } else {
2801     tmp_buf = (uint8_t *)tmp_buf16;
2802   }
2803 #endif  // CONFIG_VP9_HIGHBITDEPTH
2804 
2805   if (pred_filter_search) {
2806     INTERP_FILTER af = SWITCHABLE, lf = SWITCHABLE;
2807     if (xd->above_mi && is_inter_block(xd->above_mi))
2808       af = xd->above_mi->interp_filter;
2809     if (xd->left_mi && is_inter_block(xd->left_mi))
2810       lf = xd->left_mi->interp_filter;
2811 
2812     if ((this_mode != NEWMV) || (af == lf)) best_filter = af;
2813   }
2814 
2815   if (is_comp_pred) {
2816     if (frame_mv[refs[0]].as_int == INVALID_MV ||
2817         frame_mv[refs[1]].as_int == INVALID_MV)
2818       return INT64_MAX;
2819 
2820     if (cpi->sf.adaptive_mode_search) {
2821       if (single_filter[this_mode][refs[0]] ==
2822           single_filter[this_mode][refs[1]])
2823         best_filter = single_filter[this_mode][refs[0]];
2824     }
2825   }
2826 
2827   if (this_mode == NEWMV) {
2828     int rate_mv;
2829     if (is_comp_pred) {
2830       // Initialize mv using single prediction mode result.
2831       frame_mv[refs[0]].as_int = single_newmv[refs[0]].as_int;
2832       frame_mv[refs[1]].as_int = single_newmv[refs[1]].as_int;
2833 
2834       if (cpi->sf.comp_inter_joint_search_thresh <= bsize) {
2835         joint_motion_search(cpi, x, bsize, frame_mv, mi_row, mi_col,
2836                             single_newmv, &rate_mv);
2837       } else {
2838         rate_mv = vp9_mv_bit_cost(&frame_mv[refs[0]].as_mv,
2839                                   &x->mbmi_ext->ref_mvs[refs[0]][0].as_mv,
2840                                   x->nmvjointcost, x->mvcost, MV_COST_WEIGHT);
2841         rate_mv += vp9_mv_bit_cost(&frame_mv[refs[1]].as_mv,
2842                                    &x->mbmi_ext->ref_mvs[refs[1]][0].as_mv,
2843                                    x->nmvjointcost, x->mvcost, MV_COST_WEIGHT);
2844       }
2845       *rate2 += rate_mv;
2846     } else {
2847       int_mv tmp_mv;
2848       single_motion_search(cpi, x, bsize, mi_row, mi_col, &tmp_mv, &rate_mv);
2849       if (tmp_mv.as_int == INVALID_MV) return INT64_MAX;
2850 
2851       frame_mv[refs[0]].as_int = xd->mi[0]->bmi[0].as_mv[0].as_int =
2852           tmp_mv.as_int;
2853       single_newmv[refs[0]].as_int = tmp_mv.as_int;
2854 
2855       // Estimate the rate implications of a new mv but discount this
2856       // under certain circumstances where we want to help initiate a weak
2857       // motion field, where the distortion gain for a single block may not
2858       // be enough to overcome the cost of a new mv.
2859       if (discount_newmv_test(cpi, this_mode, tmp_mv, mode_mv, refs[0], mi_row,
2860                               mi_col, bsize)) {
2861         *rate2 += VPXMAX((rate_mv / NEW_MV_DISCOUNT_FACTOR), 1);
2862       } else {
2863         *rate2 += rate_mv;
2864       }
2865     }
2866   }
2867 
2868   for (i = 0; i < is_comp_pred + 1; ++i) {
2869     cur_mv[i] = frame_mv[refs[i]];
2870     // Clip "next_nearest" so that it does not extend to far out of image
2871     if (this_mode != NEWMV) clamp_mv2(&cur_mv[i].as_mv, xd);
2872 
2873     if (mv_check_bounds(&x->mv_limits, &cur_mv[i].as_mv)) return INT64_MAX;
2874     mi->mv[i].as_int = cur_mv[i].as_int;
2875   }
2876 
2877   // do first prediction into the destination buffer. Do the next
2878   // prediction into a temporary buffer. Then keep track of which one
2879   // of these currently holds the best predictor, and use the other
2880   // one for future predictions. In the end, copy from tmp_buf to
2881   // dst if necessary.
2882   for (i = 0; i < MAX_MB_PLANE; i++) {
2883     orig_dst[i] = xd->plane[i].dst.buf;
2884     orig_dst_stride[i] = xd->plane[i].dst.stride;
2885   }
2886 
2887   // We don't include the cost of the second reference here, because there
2888   // are only two options: Last/ARF or Golden/ARF; The second one is always
2889   // known, which is ARF.
2890   //
2891   // Under some circumstances we discount the cost of new mv mode to encourage
2892   // initiation of a motion field.
2893   if (discount_newmv_test(cpi, this_mode, frame_mv[refs[0]], mode_mv, refs[0],
2894                           mi_row, mi_col, bsize)) {
2895     *rate2 +=
2896         VPXMIN(cost_mv_ref(cpi, this_mode, mbmi_ext->mode_context[refs[0]]),
2897                cost_mv_ref(cpi, NEARESTMV, mbmi_ext->mode_context[refs[0]]));
2898   } else {
2899     *rate2 += cost_mv_ref(cpi, this_mode, mbmi_ext->mode_context[refs[0]]);
2900   }
2901 
2902   if (RDCOST(x->rdmult, x->rddiv, *rate2, 0) > ref_best_rd &&
2903       mi->mode != NEARESTMV)
2904     return INT64_MAX;
2905 
2906   pred_exists = 0;
2907   // Are all MVs integer pel for Y and UV
2908   intpel_mv = !mv_has_subpel(&mi->mv[0].as_mv);
2909   if (is_comp_pred) intpel_mv &= !mv_has_subpel(&mi->mv[1].as_mv);
2910 
2911   // Search for best switchable filter by checking the variance of
2912   // pred error irrespective of whether the filter will be used
2913   for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i) filter_cache[i] = INT64_MAX;
2914 
2915   if (cm->interp_filter != BILINEAR) {
2916     if (x->source_variance < cpi->sf.disable_filter_search_var_thresh) {
2917       best_filter = EIGHTTAP;
2918     } else if (best_filter == SWITCHABLE) {
2919       int newbest;
2920       int tmp_rate_sum = 0;
2921       int64_t tmp_dist_sum = 0;
2922 
2923       for (i = 0; i < SWITCHABLE_FILTERS; ++i) {
2924         int j;
2925         int64_t rs_rd;
2926         int tmp_skip_sb = 0;
2927         int64_t tmp_skip_sse = INT64_MAX;
2928 
2929         mi->interp_filter = i;
2930         rs = vp9_get_switchable_rate(cpi, xd);
2931         rs_rd = RDCOST(x->rdmult, x->rddiv, rs, 0);
2932 
2933         if (i > 0 && intpel_mv) {
2934           rd = RDCOST(x->rdmult, x->rddiv, tmp_rate_sum, tmp_dist_sum);
2935           filter_cache[i] = rd;
2936           filter_cache[SWITCHABLE_FILTERS] =
2937               VPXMIN(filter_cache[SWITCHABLE_FILTERS], rd + rs_rd);
2938           if (cm->interp_filter == SWITCHABLE) rd += rs_rd;
2939           *mask_filter = VPXMAX(*mask_filter, rd);
2940         } else {
2941           int rate_sum = 0;
2942           int64_t dist_sum = 0;
2943           if (i > 0 && cpi->sf.adaptive_interp_filter_search &&
2944               (cpi->sf.interp_filter_search_mask & (1 << i))) {
2945             rate_sum = INT_MAX;
2946             dist_sum = INT64_MAX;
2947             continue;
2948           }
2949 
2950           if ((cm->interp_filter == SWITCHABLE && (!i || best_needs_copy)) ||
2951               (cm->interp_filter != SWITCHABLE &&
2952                (cm->interp_filter == mi->interp_filter ||
2953                 (i == 0 && intpel_mv)))) {
2954             restore_dst_buf(xd, orig_dst, orig_dst_stride);
2955           } else {
2956             for (j = 0; j < MAX_MB_PLANE; j++) {
2957               xd->plane[j].dst.buf = tmp_buf + j * 64 * 64;
2958               xd->plane[j].dst.stride = 64;
2959             }
2960           }
2961           vp9_build_inter_predictors_sb(xd, mi_row, mi_col, bsize);
2962           model_rd_for_sb(cpi, bsize, x, xd, &rate_sum, &dist_sum, &tmp_skip_sb,
2963                           &tmp_skip_sse);
2964 
2965           rd = RDCOST(x->rdmult, x->rddiv, rate_sum, dist_sum);
2966           filter_cache[i] = rd;
2967           filter_cache[SWITCHABLE_FILTERS] =
2968               VPXMIN(filter_cache[SWITCHABLE_FILTERS], rd + rs_rd);
2969           if (cm->interp_filter == SWITCHABLE) rd += rs_rd;
2970           *mask_filter = VPXMAX(*mask_filter, rd);
2971 
2972           if (i == 0 && intpel_mv) {
2973             tmp_rate_sum = rate_sum;
2974             tmp_dist_sum = dist_sum;
2975           }
2976         }
2977 
2978         if (i == 0 && cpi->sf.use_rd_breakout && ref_best_rd < INT64_MAX) {
2979           if (rd / 2 > ref_best_rd) {
2980             restore_dst_buf(xd, orig_dst, orig_dst_stride);
2981             return INT64_MAX;
2982           }
2983         }
2984         newbest = i == 0 || rd < best_rd;
2985 
2986         if (newbest) {
2987           best_rd = rd;
2988           best_filter = mi->interp_filter;
2989           if (cm->interp_filter == SWITCHABLE && i && !intpel_mv)
2990             best_needs_copy = !best_needs_copy;
2991         }
2992 
2993         if ((cm->interp_filter == SWITCHABLE && newbest) ||
2994             (cm->interp_filter != SWITCHABLE &&
2995              cm->interp_filter == mi->interp_filter)) {
2996           pred_exists = 1;
2997           tmp_rd = best_rd;
2998 
2999           skip_txfm_sb = tmp_skip_sb;
3000           skip_sse_sb = tmp_skip_sse;
3001           memcpy(skip_txfm, x->skip_txfm, sizeof(skip_txfm));
3002           memcpy(bsse, x->bsse, sizeof(bsse));
3003         }
3004       }
3005       restore_dst_buf(xd, orig_dst, orig_dst_stride);
3006     }
3007   }
3008   // Set the appropriate filter
3009   mi->interp_filter =
3010       cm->interp_filter != SWITCHABLE ? cm->interp_filter : best_filter;
3011   rs = cm->interp_filter == SWITCHABLE ? vp9_get_switchable_rate(cpi, xd) : 0;
3012 
3013   if (pred_exists) {
3014     if (best_needs_copy) {
3015       // again temporarily set the buffers to local memory to prevent a memcpy
3016       for (i = 0; i < MAX_MB_PLANE; i++) {
3017         xd->plane[i].dst.buf = tmp_buf + i * 64 * 64;
3018         xd->plane[i].dst.stride = 64;
3019       }
3020     }
3021     rd = tmp_rd + RDCOST(x->rdmult, x->rddiv, rs, 0);
3022   } else {
3023     int tmp_rate;
3024     int64_t tmp_dist;
3025     // Handles the special case when a filter that is not in the
3026     // switchable list (ex. bilinear) is indicated at the frame level, or
3027     // skip condition holds.
3028     vp9_build_inter_predictors_sb(xd, mi_row, mi_col, bsize);
3029     model_rd_for_sb(cpi, bsize, x, xd, &tmp_rate, &tmp_dist, &skip_txfm_sb,
3030                     &skip_sse_sb);
3031     rd = RDCOST(x->rdmult, x->rddiv, rs + tmp_rate, tmp_dist);
3032     memcpy(skip_txfm, x->skip_txfm, sizeof(skip_txfm));
3033     memcpy(bsse, x->bsse, sizeof(bsse));
3034   }
3035 
3036   if (!is_comp_pred) single_filter[this_mode][refs[0]] = mi->interp_filter;
3037 
3038   if (cpi->sf.adaptive_mode_search)
3039     if (is_comp_pred)
3040       if (single_skippable[this_mode][refs[0]] &&
3041           single_skippable[this_mode][refs[1]])
3042         memset(skip_txfm, SKIP_TXFM_AC_DC, sizeof(skip_txfm));
3043 
3044   if (cpi->sf.use_rd_breakout && ref_best_rd < INT64_MAX) {
3045     // if current pred_error modeled rd is substantially more than the best
3046     // so far, do not bother doing full rd
3047     if (rd / 2 > ref_best_rd) {
3048       restore_dst_buf(xd, orig_dst, orig_dst_stride);
3049       return INT64_MAX;
3050     }
3051   }
3052 
3053   if (cm->interp_filter == SWITCHABLE) *rate2 += rs;
3054 
3055   memcpy(x->skip_txfm, skip_txfm, sizeof(skip_txfm));
3056   memcpy(x->bsse, bsse, sizeof(bsse));
3057 
3058   if (!skip_txfm_sb || xd->lossless) {
3059     int skippable_y, skippable_uv;
3060     int64_t sseuv = INT64_MAX;
3061     int64_t rdcosty = INT64_MAX;
3062 
3063     // Y cost and distortion
3064     vp9_subtract_plane(x, bsize, 0);
3065     super_block_yrd(cpi, x, rate_y, &distortion_y, &skippable_y, psse, bsize,
3066                     ref_best_rd, recon);
3067 
3068     if (*rate_y == INT_MAX) {
3069       *rate2 = INT_MAX;
3070       *distortion = INT64_MAX;
3071       restore_dst_buf(xd, orig_dst, orig_dst_stride);
3072       return INT64_MAX;
3073     }
3074 
3075     *rate2 += *rate_y;
3076     *distortion += distortion_y;
3077 
3078     rdcosty = RDCOST(x->rdmult, x->rddiv, *rate2, *distortion);
3079     rdcosty = VPXMIN(rdcosty, RDCOST(x->rdmult, x->rddiv, 0, *psse));
3080 
3081     if (!super_block_uvrd(cpi, x, rate_uv, &distortion_uv, &skippable_uv,
3082                           &sseuv, bsize, ref_best_rd - rdcosty)) {
3083       *rate2 = INT_MAX;
3084       *distortion = INT64_MAX;
3085       restore_dst_buf(xd, orig_dst, orig_dst_stride);
3086       return INT64_MAX;
3087     }
3088 
3089     *psse += sseuv;
3090     *rate2 += *rate_uv;
3091     *distortion += distortion_uv;
3092     *skippable = skippable_y && skippable_uv;
3093   } else {
3094     x->skip = 1;
3095     *disable_skip = 1;
3096 
3097     // The cost of skip bit needs to be added.
3098     *rate2 += vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1);
3099 
3100     *distortion = skip_sse_sb;
3101   }
3102 
3103   if (!is_comp_pred) single_skippable[this_mode][refs[0]] = *skippable;
3104 
3105   restore_dst_buf(xd, orig_dst, orig_dst_stride);
3106   return 0;  // The rate-distortion cost will be re-calculated by caller.
3107 }
3108 #endif  // !CONFIG_REALTIME_ONLY
3109 
vp9_rd_pick_intra_mode_sb(VP9_COMP * cpi,MACROBLOCK * x,RD_COST * rd_cost,BLOCK_SIZE bsize,PICK_MODE_CONTEXT * ctx,int64_t best_rd)3110 void vp9_rd_pick_intra_mode_sb(VP9_COMP *cpi, MACROBLOCK *x, RD_COST *rd_cost,
3111                                BLOCK_SIZE bsize, PICK_MODE_CONTEXT *ctx,
3112                                int64_t best_rd) {
3113   VP9_COMMON *const cm = &cpi->common;
3114   MACROBLOCKD *const xd = &x->e_mbd;
3115   struct macroblockd_plane *const pd = xd->plane;
3116   int rate_y = 0, rate_uv = 0, rate_y_tokenonly = 0, rate_uv_tokenonly = 0;
3117   int y_skip = 0, uv_skip = 0;
3118   int64_t dist_y = 0, dist_uv = 0;
3119   TX_SIZE max_uv_tx_size;
3120   x->skip_encode = 0;
3121   ctx->skip = 0;
3122   xd->mi[0]->ref_frame[0] = INTRA_FRAME;
3123   xd->mi[0]->ref_frame[1] = NONE;
3124   // Initialize interp_filter here so we do not have to check for inter block
3125   // modes in get_pred_context_switchable_interp()
3126   xd->mi[0]->interp_filter = SWITCHABLE_FILTERS;
3127 
3128   if (bsize >= BLOCK_8X8) {
3129     if (rd_pick_intra_sby_mode(cpi, x, &rate_y, &rate_y_tokenonly, &dist_y,
3130                                &y_skip, bsize, best_rd) >= best_rd) {
3131       rd_cost->rate = INT_MAX;
3132       return;
3133     }
3134   } else {
3135     y_skip = 0;
3136     if (rd_pick_intra_sub_8x8_y_mode(cpi, x, &rate_y, &rate_y_tokenonly,
3137                                      &dist_y, best_rd) >= best_rd) {
3138       rd_cost->rate = INT_MAX;
3139       return;
3140     }
3141   }
3142   max_uv_tx_size = uv_txsize_lookup[bsize][xd->mi[0]->tx_size]
3143                                    [pd[1].subsampling_x][pd[1].subsampling_y];
3144   rd_pick_intra_sbuv_mode(cpi, x, ctx, &rate_uv, &rate_uv_tokenonly, &dist_uv,
3145                           &uv_skip, VPXMAX(BLOCK_8X8, bsize), max_uv_tx_size);
3146 
3147   if (y_skip && uv_skip) {
3148     rd_cost->rate = rate_y + rate_uv - rate_y_tokenonly - rate_uv_tokenonly +
3149                     vp9_cost_bit(vp9_get_skip_prob(cm, xd), 1);
3150     rd_cost->dist = dist_y + dist_uv;
3151   } else {
3152     rd_cost->rate =
3153         rate_y + rate_uv + vp9_cost_bit(vp9_get_skip_prob(cm, xd), 0);
3154     rd_cost->dist = dist_y + dist_uv;
3155   }
3156 
3157   ctx->mic = *xd->mi[0];
3158   ctx->mbmi_ext = *x->mbmi_ext;
3159   rd_cost->rdcost = RDCOST(x->rdmult, x->rddiv, rd_cost->rate, rd_cost->dist);
3160 }
3161 
3162 #if !CONFIG_REALTIME_ONLY
3163 // This function is designed to apply a bias or adjustment to an rd value based
3164 // on the relative variance of the source and reconstruction.
3165 #define LOW_VAR_THRESH 250
3166 #define VAR_MULT 250
3167 static unsigned int max_var_adjust[VP9E_CONTENT_INVALID] = { 16, 16, 250 };
3168 
rd_variance_adjustment(VP9_COMP * cpi,MACROBLOCK * x,BLOCK_SIZE bsize,int64_t * this_rd,struct buf_2d * recon,MV_REFERENCE_FRAME ref_frame,MV_REFERENCE_FRAME second_ref_frame,PREDICTION_MODE this_mode)3169 static void rd_variance_adjustment(VP9_COMP *cpi, MACROBLOCK *x,
3170                                    BLOCK_SIZE bsize, int64_t *this_rd,
3171                                    struct buf_2d *recon,
3172                                    MV_REFERENCE_FRAME ref_frame,
3173                                    MV_REFERENCE_FRAME second_ref_frame,
3174                                    PREDICTION_MODE this_mode) {
3175   MACROBLOCKD *const xd = &x->e_mbd;
3176   unsigned int rec_variance;
3177   unsigned int src_variance;
3178   unsigned int src_rec_min;
3179   unsigned int var_diff = 0;
3180   unsigned int var_factor = 0;
3181   unsigned int adj_max;
3182   unsigned int low_var_thresh = LOW_VAR_THRESH;
3183   const int bw = num_8x8_blocks_wide_lookup[bsize];
3184   const int bh = num_8x8_blocks_high_lookup[bsize];
3185   vp9e_tune_content content_type = cpi->oxcf.content;
3186 
3187   if (*this_rd == INT64_MAX) return;
3188 
3189 #if CONFIG_VP9_HIGHBITDEPTH
3190   if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
3191     rec_variance = vp9_high_get_sby_variance(cpi, recon, bsize, xd->bd);
3192     src_variance =
3193         vp9_high_get_sby_variance(cpi, &x->plane[0].src, bsize, xd->bd);
3194   } else {
3195     rec_variance = vp9_get_sby_variance(cpi, recon, bsize);
3196     src_variance = vp9_get_sby_variance(cpi, &x->plane[0].src, bsize);
3197   }
3198 #else
3199   rec_variance = vp9_get_sby_variance(cpi, recon, bsize);
3200   src_variance = vp9_get_sby_variance(cpi, &x->plane[0].src, bsize);
3201 #endif  // CONFIG_VP9_HIGHBITDEPTH
3202 
3203   // Scale based on area in 8x8 blocks
3204   rec_variance /= (bw * bh);
3205   src_variance /= (bw * bh);
3206 
3207   if (content_type == VP9E_CONTENT_FILM) {
3208     if (cpi->oxcf.pass == 2) {
3209       // Adjust low variance threshold based on estimated group noise enegry.
3210       double noise_factor =
3211           (double)cpi->twopass.gf_group.group_noise_energy / SECTION_NOISE_DEF;
3212       low_var_thresh = (unsigned int)(low_var_thresh * noise_factor);
3213 
3214       if (ref_frame == INTRA_FRAME) {
3215         low_var_thresh *= 2;
3216         if (this_mode == DC_PRED) low_var_thresh *= 5;
3217       } else if (second_ref_frame > INTRA_FRAME) {
3218         low_var_thresh *= 2;
3219       }
3220     }
3221   } else {
3222     low_var_thresh = LOW_VAR_THRESH / 2;
3223   }
3224 
3225   // Lower of source (raw per pixel value) and recon variance. Note that
3226   // if the source per pixel is 0 then the recon value here will not be per
3227   // pixel (see above) so will likely be much larger.
3228   src_rec_min = VPXMIN(src_variance, rec_variance);
3229 
3230   if (src_rec_min > low_var_thresh) return;
3231 
3232   // We care more when the reconstruction has lower variance so give this case
3233   // a stronger weighting.
3234   var_diff = (src_variance > rec_variance) ? (src_variance - rec_variance) * 2
3235                                            : (rec_variance - src_variance) / 2;
3236 
3237   adj_max = max_var_adjust[content_type];
3238 
3239   var_factor =
3240       (unsigned int)((int64_t)VAR_MULT * var_diff) / VPXMAX(1, src_variance);
3241   var_factor = VPXMIN(adj_max, var_factor);
3242 
3243   if ((content_type == VP9E_CONTENT_FILM) &&
3244       ((ref_frame == INTRA_FRAME) || (second_ref_frame > INTRA_FRAME))) {
3245     var_factor *= 2;
3246   }
3247 
3248   *this_rd += (*this_rd * var_factor) / 100;
3249 
3250   (void)xd;
3251 }
3252 #endif  // !CONFIG_REALTIME_ONLY
3253 
3254 // Do we have an internal image edge (e.g. formatting bars).
vp9_internal_image_edge(VP9_COMP * cpi)3255 int vp9_internal_image_edge(VP9_COMP *cpi) {
3256   return (cpi->oxcf.pass == 2) &&
3257          ((cpi->twopass.this_frame_stats.inactive_zone_rows > 0) ||
3258           (cpi->twopass.this_frame_stats.inactive_zone_cols > 0));
3259 }
3260 
3261 // Checks to see if a super block is on a horizontal image edge.
3262 // In most cases this is the "real" edge unless there are formatting
3263 // bars embedded in the stream.
vp9_active_h_edge(VP9_COMP * cpi,int mi_row,int mi_step)3264 int vp9_active_h_edge(VP9_COMP *cpi, int mi_row, int mi_step) {
3265   int top_edge = 0;
3266   int bottom_edge = cpi->common.mi_rows;
3267   int is_active_h_edge = 0;
3268 
3269   // For two pass account for any formatting bars detected.
3270   if (cpi->oxcf.pass == 2) {
3271     TWO_PASS *twopass = &cpi->twopass;
3272     vpx_clear_system_state();
3273 
3274     // The inactive region is specified in MBs not mi units.
3275     // The image edge is in the following MB row.
3276     top_edge += (int)(twopass->this_frame_stats.inactive_zone_rows * 2);
3277 
3278     bottom_edge -= (int)(twopass->this_frame_stats.inactive_zone_rows * 2);
3279     bottom_edge = VPXMAX(top_edge, bottom_edge);
3280   }
3281 
3282   if (((top_edge >= mi_row) && (top_edge < (mi_row + mi_step))) ||
3283       ((bottom_edge >= mi_row) && (bottom_edge < (mi_row + mi_step)))) {
3284     is_active_h_edge = 1;
3285   }
3286   return is_active_h_edge;
3287 }
3288 
3289 // Checks to see if a super block is on a vertical image edge.
3290 // In most cases this is the "real" edge unless there are formatting
3291 // bars embedded in the stream.
vp9_active_v_edge(VP9_COMP * cpi,int mi_col,int mi_step)3292 int vp9_active_v_edge(VP9_COMP *cpi, int mi_col, int mi_step) {
3293   int left_edge = 0;
3294   int right_edge = cpi->common.mi_cols;
3295   int is_active_v_edge = 0;
3296 
3297   // For two pass account for any formatting bars detected.
3298   if (cpi->oxcf.pass == 2) {
3299     TWO_PASS *twopass = &cpi->twopass;
3300     vpx_clear_system_state();
3301 
3302     // The inactive region is specified in MBs not mi units.
3303     // The image edge is in the following MB row.
3304     left_edge += (int)(twopass->this_frame_stats.inactive_zone_cols * 2);
3305 
3306     right_edge -= (int)(twopass->this_frame_stats.inactive_zone_cols * 2);
3307     right_edge = VPXMAX(left_edge, right_edge);
3308   }
3309 
3310   if (((left_edge >= mi_col) && (left_edge < (mi_col + mi_step))) ||
3311       ((right_edge >= mi_col) && (right_edge < (mi_col + mi_step)))) {
3312     is_active_v_edge = 1;
3313   }
3314   return is_active_v_edge;
3315 }
3316 
3317 // Checks to see if a super block is at the edge of the active image.
3318 // In most cases this is the "real" edge unless there are formatting
3319 // bars embedded in the stream.
vp9_active_edge_sb(VP9_COMP * cpi,int mi_row,int mi_col)3320 int vp9_active_edge_sb(VP9_COMP *cpi, int mi_row, int mi_col) {
3321   return vp9_active_h_edge(cpi, mi_row, MI_BLOCK_SIZE) ||
3322          vp9_active_v_edge(cpi, mi_col, MI_BLOCK_SIZE);
3323 }
3324 
3325 #if !CONFIG_REALTIME_ONLY
vp9_rd_pick_inter_mode_sb(VP9_COMP * cpi,TileDataEnc * tile_data,MACROBLOCK * x,int mi_row,int mi_col,RD_COST * rd_cost,BLOCK_SIZE bsize,PICK_MODE_CONTEXT * ctx,int64_t best_rd_so_far)3326 void vp9_rd_pick_inter_mode_sb(VP9_COMP *cpi, TileDataEnc *tile_data,
3327                                MACROBLOCK *x, int mi_row, int mi_col,
3328                                RD_COST *rd_cost, BLOCK_SIZE bsize,
3329                                PICK_MODE_CONTEXT *ctx, int64_t best_rd_so_far) {
3330   VP9_COMMON *const cm = &cpi->common;
3331   TileInfo *const tile_info = &tile_data->tile_info;
3332   RD_OPT *const rd_opt = &cpi->rd;
3333   SPEED_FEATURES *const sf = &cpi->sf;
3334   MACROBLOCKD *const xd = &x->e_mbd;
3335   MODE_INFO *const mi = xd->mi[0];
3336   MB_MODE_INFO_EXT *const mbmi_ext = x->mbmi_ext;
3337   const struct segmentation *const seg = &cm->seg;
3338   PREDICTION_MODE this_mode;
3339   MV_REFERENCE_FRAME ref_frame, second_ref_frame;
3340   unsigned char segment_id = mi->segment_id;
3341   int comp_pred, i, k;
3342   int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES];
3343   struct buf_2d yv12_mb[4][MAX_MB_PLANE];
3344   int_mv single_newmv[MAX_REF_FRAMES] = { { 0 } };
3345   INTERP_FILTER single_inter_filter[MB_MODE_COUNT][MAX_REF_FRAMES];
3346   int single_skippable[MB_MODE_COUNT][MAX_REF_FRAMES];
3347   int64_t best_rd = best_rd_so_far;
3348   int64_t best_pred_diff[REFERENCE_MODES];
3349   int64_t best_pred_rd[REFERENCE_MODES];
3350   int64_t best_filter_rd[SWITCHABLE_FILTER_CONTEXTS];
3351   int64_t best_filter_diff[SWITCHABLE_FILTER_CONTEXTS];
3352   MODE_INFO best_mbmode;
3353   int best_mode_skippable = 0;
3354   int midx, best_mode_index = -1;
3355   unsigned int ref_costs_single[MAX_REF_FRAMES], ref_costs_comp[MAX_REF_FRAMES];
3356   vpx_prob comp_mode_p;
3357   int64_t best_intra_rd = INT64_MAX;
3358   unsigned int best_pred_sse = UINT_MAX;
3359   PREDICTION_MODE best_intra_mode = DC_PRED;
3360   int rate_uv_intra[TX_SIZES], rate_uv_tokenonly[TX_SIZES];
3361   int64_t dist_uv[TX_SIZES];
3362   int skip_uv[TX_SIZES];
3363   PREDICTION_MODE mode_uv[TX_SIZES];
3364   const int intra_cost_penalty =
3365       vp9_get_intra_cost_penalty(cpi, bsize, cm->base_qindex, cm->y_dc_delta_q);
3366   int best_skip2 = 0;
3367   uint8_t ref_frame_skip_mask[2] = { 0, 1 };
3368   uint16_t mode_skip_mask[MAX_REF_FRAMES] = { 0 };
3369   int mode_skip_start = sf->mode_skip_start + 1;
3370   const int *const rd_threshes = rd_opt->threshes[segment_id][bsize];
3371   const int *const rd_thresh_freq_fact = tile_data->thresh_freq_fact[bsize];
3372   int64_t mode_threshold[MAX_MODES];
3373   int8_t *tile_mode_map = tile_data->mode_map[bsize];
3374   int8_t mode_map[MAX_MODES];  // Maintain mode_map information locally to avoid
3375                                // lock mechanism involved with reads from
3376                                // tile_mode_map
3377   const int mode_search_skip_flags = sf->mode_search_skip_flags;
3378   const int is_rect_partition =
3379       num_4x4_blocks_wide_lookup[bsize] != num_4x4_blocks_high_lookup[bsize];
3380   int64_t mask_filter = 0;
3381   int64_t filter_cache[SWITCHABLE_FILTER_CONTEXTS];
3382 
3383   struct buf_2d *recon;
3384   struct buf_2d recon_buf;
3385 #if CONFIG_VP9_HIGHBITDEPTH
3386   DECLARE_ALIGNED(16, uint16_t, recon16[64 * 64]);
3387   recon_buf.buf = xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH
3388                       ? CONVERT_TO_BYTEPTR(recon16)
3389                       : (uint8_t *)recon16;
3390 #else
3391   DECLARE_ALIGNED(16, uint8_t, recon8[64 * 64]);
3392   recon_buf.buf = recon8;
3393 #endif  // CONFIG_VP9_HIGHBITDEPTH
3394   recon_buf.stride = 64;
3395   recon = cpi->oxcf.content == VP9E_CONTENT_FILM ? &recon_buf : 0;
3396 
3397   vp9_zero(best_mbmode);
3398 
3399   x->skip_encode = sf->skip_encode_frame && x->q_index < QIDX_SKIP_THRESH;
3400 
3401   for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i) filter_cache[i] = INT64_MAX;
3402 
3403   estimate_ref_frame_costs(cm, xd, segment_id, ref_costs_single, ref_costs_comp,
3404                            &comp_mode_p);
3405 
3406   for (i = 0; i < REFERENCE_MODES; ++i) best_pred_rd[i] = INT64_MAX;
3407   for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++)
3408     best_filter_rd[i] = INT64_MAX;
3409   for (i = 0; i < TX_SIZES; i++) rate_uv_intra[i] = INT_MAX;
3410   for (i = 0; i < MAX_REF_FRAMES; ++i) x->pred_sse[i] = INT_MAX;
3411   for (i = 0; i < MB_MODE_COUNT; ++i) {
3412     for (k = 0; k < MAX_REF_FRAMES; ++k) {
3413       single_inter_filter[i][k] = SWITCHABLE;
3414       single_skippable[i][k] = 0;
3415     }
3416   }
3417 
3418   rd_cost->rate = INT_MAX;
3419 
3420   for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ++ref_frame) {
3421     x->pred_mv_sad[ref_frame] = INT_MAX;
3422     if ((cpi->ref_frame_flags & ref_frame_to_flag(ref_frame)) &&
3423         !(is_rect_partition && (ctx->skip_ref_frame_mask & (1 << ref_frame)))) {
3424       assert(get_ref_frame_buffer(cpi, ref_frame) != NULL);
3425       setup_buffer_inter(cpi, x, ref_frame, bsize, mi_row, mi_col,
3426                          frame_mv[NEARESTMV], frame_mv[NEARMV], yv12_mb);
3427     }
3428     frame_mv[NEWMV][ref_frame].as_int = INVALID_MV;
3429     frame_mv[ZEROMV][ref_frame].as_int = 0;
3430   }
3431 
3432   for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ++ref_frame) {
3433     if (!(cpi->ref_frame_flags & ref_frame_to_flag(ref_frame))) {
3434       // Skip checking missing references in both single and compound reference
3435       // modes. Note that a mode will be skipped if both reference frames
3436       // are masked out.
3437       ref_frame_skip_mask[0] |= (1 << ref_frame);
3438       ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
3439     } else if (sf->reference_masking) {
3440       for (i = LAST_FRAME; i <= ALTREF_FRAME; ++i) {
3441         // Skip fixed mv modes for poor references
3442         if ((x->pred_mv_sad[ref_frame] >> 2) > x->pred_mv_sad[i]) {
3443           mode_skip_mask[ref_frame] |= INTER_NEAREST_NEAR_ZERO;
3444           break;
3445         }
3446       }
3447     }
3448     // If the segment reference frame feature is enabled....
3449     // then do nothing if the current ref frame is not allowed..
3450     if (segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME) &&
3451         get_segdata(seg, segment_id, SEG_LVL_REF_FRAME) != (int)ref_frame) {
3452       ref_frame_skip_mask[0] |= (1 << ref_frame);
3453       ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
3454     }
3455   }
3456 
3457   // Disable this drop out case if the ref frame
3458   // segment level feature is enabled for this segment. This is to
3459   // prevent the possibility that we end up unable to pick any mode.
3460   if (!segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME)) {
3461     // Only consider ZEROMV/ALTREF_FRAME for alt ref frame,
3462     // unless ARNR filtering is enabled in which case we want
3463     // an unfiltered alternative. We allow near/nearest as well
3464     // because they may result in zero-zero MVs but be cheaper.
3465     if (cpi->rc.is_src_frame_alt_ref && (cpi->oxcf.arnr_max_frames == 0)) {
3466       ref_frame_skip_mask[0] = (1 << LAST_FRAME) | (1 << GOLDEN_FRAME);
3467       ref_frame_skip_mask[1] = SECOND_REF_FRAME_MASK;
3468       mode_skip_mask[ALTREF_FRAME] = ~INTER_NEAREST_NEAR_ZERO;
3469       if (frame_mv[NEARMV][ALTREF_FRAME].as_int != 0)
3470         mode_skip_mask[ALTREF_FRAME] |= (1 << NEARMV);
3471       if (frame_mv[NEARESTMV][ALTREF_FRAME].as_int != 0)
3472         mode_skip_mask[ALTREF_FRAME] |= (1 << NEARESTMV);
3473     }
3474   }
3475 
3476   if (cpi->rc.is_src_frame_alt_ref) {
3477     if (sf->alt_ref_search_fp) {
3478       mode_skip_mask[ALTREF_FRAME] = 0;
3479       ref_frame_skip_mask[0] = ~(1 << ALTREF_FRAME) & 0xff;
3480       ref_frame_skip_mask[1] = SECOND_REF_FRAME_MASK;
3481     }
3482   }
3483 
3484   if (sf->alt_ref_search_fp)
3485     if (!cm->show_frame && x->pred_mv_sad[GOLDEN_FRAME] < INT_MAX)
3486       if (x->pred_mv_sad[ALTREF_FRAME] > (x->pred_mv_sad[GOLDEN_FRAME] << 1))
3487         mode_skip_mask[ALTREF_FRAME] |= INTER_ALL;
3488 
3489   if (sf->adaptive_mode_search) {
3490     if (cm->show_frame && !cpi->rc.is_src_frame_alt_ref &&
3491         cpi->rc.frames_since_golden >= 3)
3492       if (x->pred_mv_sad[GOLDEN_FRAME] > (x->pred_mv_sad[LAST_FRAME] << 1))
3493         mode_skip_mask[GOLDEN_FRAME] |= INTER_ALL;
3494   }
3495 
3496   if (bsize > sf->max_intra_bsize) {
3497     ref_frame_skip_mask[0] |= (1 << INTRA_FRAME);
3498     ref_frame_skip_mask[1] |= (1 << INTRA_FRAME);
3499   }
3500 
3501   mode_skip_mask[INTRA_FRAME] |=
3502       (uint16_t) ~(sf->intra_y_mode_mask[max_txsize_lookup[bsize]]);
3503 
3504   for (i = 0; i <= LAST_NEW_MV_INDEX; ++i) mode_threshold[i] = 0;
3505 
3506   for (i = LAST_NEW_MV_INDEX + 1; i < MAX_MODES; ++i)
3507     mode_threshold[i] = ((int64_t)rd_threshes[i] * rd_thresh_freq_fact[i]) >> 5;
3508 
3509   midx = sf->schedule_mode_search ? mode_skip_start : 0;
3510 
3511   while (midx > 4) {
3512     uint8_t end_pos = 0;
3513     for (i = 5; i < midx; ++i) {
3514       if (mode_threshold[tile_mode_map[i - 1]] >
3515           mode_threshold[tile_mode_map[i]]) {
3516         uint8_t tmp = tile_mode_map[i];
3517         tile_mode_map[i] = tile_mode_map[i - 1];
3518         tile_mode_map[i - 1] = tmp;
3519         end_pos = i;
3520       }
3521     }
3522     midx = end_pos;
3523   }
3524 
3525   memcpy(mode_map, tile_mode_map, sizeof(mode_map));
3526 
3527   for (midx = 0; midx < MAX_MODES; ++midx) {
3528     int mode_index = mode_map[midx];
3529     int mode_excluded = 0;
3530     int64_t this_rd = INT64_MAX;
3531     int disable_skip = 0;
3532     int compmode_cost = 0;
3533     int rate2 = 0, rate_y = 0, rate_uv = 0;
3534     int64_t distortion2 = 0, distortion_y = 0, distortion_uv = 0;
3535     int skippable = 0;
3536     int this_skip2 = 0;
3537     int64_t total_sse = INT64_MAX;
3538     int early_term = 0;
3539 
3540     this_mode = vp9_mode_order[mode_index].mode;
3541     ref_frame = vp9_mode_order[mode_index].ref_frame[0];
3542     second_ref_frame = vp9_mode_order[mode_index].ref_frame[1];
3543 
3544     vp9_zero(x->sum_y_eobs);
3545 
3546     if (is_rect_partition) {
3547       if (ctx->skip_ref_frame_mask & (1 << ref_frame)) continue;
3548       if (second_ref_frame > 0 &&
3549           (ctx->skip_ref_frame_mask & (1 << second_ref_frame)))
3550         continue;
3551     }
3552 
3553     // Look at the reference frame of the best mode so far and set the
3554     // skip mask to look at a subset of the remaining modes.
3555     if (midx == mode_skip_start && best_mode_index >= 0) {
3556       switch (best_mbmode.ref_frame[0]) {
3557         case INTRA_FRAME: break;
3558         case LAST_FRAME: ref_frame_skip_mask[0] |= LAST_FRAME_MODE_MASK; break;
3559         case GOLDEN_FRAME:
3560           ref_frame_skip_mask[0] |= GOLDEN_FRAME_MODE_MASK;
3561           break;
3562         case ALTREF_FRAME: ref_frame_skip_mask[0] |= ALT_REF_MODE_MASK; break;
3563         case NONE:
3564         case MAX_REF_FRAMES: assert(0 && "Invalid Reference frame"); break;
3565       }
3566     }
3567 
3568     if ((ref_frame_skip_mask[0] & (1 << ref_frame)) &&
3569         (ref_frame_skip_mask[1] & (1 << VPXMAX(0, second_ref_frame))))
3570       continue;
3571 
3572     if (mode_skip_mask[ref_frame] & (1 << this_mode)) continue;
3573 
3574     // Test best rd so far against threshold for trying this mode.
3575     if (best_mode_skippable && sf->schedule_mode_search)
3576       mode_threshold[mode_index] <<= 1;
3577 
3578     if (best_rd < mode_threshold[mode_index]) continue;
3579 
3580     // This is only used in motion vector unit test.
3581     if (cpi->oxcf.motion_vector_unit_test && ref_frame == INTRA_FRAME) continue;
3582 
3583     if (sf->motion_field_mode_search) {
3584       const int mi_width = VPXMIN(num_8x8_blocks_wide_lookup[bsize],
3585                                   tile_info->mi_col_end - mi_col);
3586       const int mi_height = VPXMIN(num_8x8_blocks_high_lookup[bsize],
3587                                    tile_info->mi_row_end - mi_row);
3588       const int bsl = mi_width_log2_lookup[bsize];
3589       int cb_partition_search_ctrl =
3590           (((mi_row + mi_col) >> bsl) +
3591            get_chessboard_index(cm->current_video_frame)) &
3592           0x1;
3593       MODE_INFO *ref_mi;
3594       int const_motion = 1;
3595       int skip_ref_frame = !cb_partition_search_ctrl;
3596       MV_REFERENCE_FRAME rf = NONE;
3597       int_mv ref_mv;
3598       ref_mv.as_int = INVALID_MV;
3599 
3600       if ((mi_row - 1) >= tile_info->mi_row_start) {
3601         ref_mv = xd->mi[-xd->mi_stride]->mv[0];
3602         rf = xd->mi[-xd->mi_stride]->ref_frame[0];
3603         for (i = 0; i < mi_width; ++i) {
3604           ref_mi = xd->mi[-xd->mi_stride + i];
3605           const_motion &= (ref_mv.as_int == ref_mi->mv[0].as_int) &&
3606                           (ref_frame == ref_mi->ref_frame[0]);
3607           skip_ref_frame &= (rf == ref_mi->ref_frame[0]);
3608         }
3609       }
3610 
3611       if ((mi_col - 1) >= tile_info->mi_col_start) {
3612         if (ref_mv.as_int == INVALID_MV) ref_mv = xd->mi[-1]->mv[0];
3613         if (rf == NONE) rf = xd->mi[-1]->ref_frame[0];
3614         for (i = 0; i < mi_height; ++i) {
3615           ref_mi = xd->mi[i * xd->mi_stride - 1];
3616           const_motion &= (ref_mv.as_int == ref_mi->mv[0].as_int) &&
3617                           (ref_frame == ref_mi->ref_frame[0]);
3618           skip_ref_frame &= (rf == ref_mi->ref_frame[0]);
3619         }
3620       }
3621 
3622       if (skip_ref_frame && this_mode != NEARESTMV && this_mode != NEWMV)
3623         if (rf > INTRA_FRAME)
3624           if (ref_frame != rf) continue;
3625 
3626       if (const_motion)
3627         if (this_mode == NEARMV || this_mode == ZEROMV) continue;
3628     }
3629 
3630     comp_pred = second_ref_frame > INTRA_FRAME;
3631     if (comp_pred) {
3632       if (!cpi->allow_comp_inter_inter) continue;
3633 
3634       if (cm->ref_frame_sign_bias[ref_frame] ==
3635           cm->ref_frame_sign_bias[second_ref_frame])
3636         continue;
3637 
3638       // Skip compound inter modes if ARF is not available.
3639       if (!(cpi->ref_frame_flags & ref_frame_to_flag(second_ref_frame)))
3640         continue;
3641 
3642       // Do not allow compound prediction if the segment level reference frame
3643       // feature is in use as in this case there can only be one reference.
3644       if (segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME)) continue;
3645 
3646       if ((mode_search_skip_flags & FLAG_SKIP_COMP_BESTINTRA) &&
3647           best_mode_index >= 0 && best_mbmode.ref_frame[0] == INTRA_FRAME)
3648         continue;
3649 
3650       mode_excluded = cm->reference_mode == SINGLE_REFERENCE;
3651     } else {
3652       if (ref_frame != INTRA_FRAME)
3653         mode_excluded = cm->reference_mode == COMPOUND_REFERENCE;
3654     }
3655 
3656     if (ref_frame == INTRA_FRAME) {
3657       if (sf->adaptive_mode_search)
3658         if ((x->source_variance << num_pels_log2_lookup[bsize]) > best_pred_sse)
3659           continue;
3660 
3661       if (this_mode != DC_PRED) {
3662         // Disable intra modes other than DC_PRED for blocks with low variance
3663         // Threshold for intra skipping based on source variance
3664         // TODO(debargha): Specialize the threshold for super block sizes
3665         const unsigned int skip_intra_var_thresh =
3666             (cpi->oxcf.content == VP9E_CONTENT_FILM) ? 0 : 64;
3667         if ((mode_search_skip_flags & FLAG_SKIP_INTRA_LOWVAR) &&
3668             x->source_variance < skip_intra_var_thresh)
3669           continue;
3670         // Only search the oblique modes if the best so far is
3671         // one of the neighboring directional modes
3672         if ((mode_search_skip_flags & FLAG_SKIP_INTRA_BESTINTER) &&
3673             (this_mode >= D45_PRED && this_mode <= TM_PRED)) {
3674           if (best_mode_index >= 0 && best_mbmode.ref_frame[0] > INTRA_FRAME)
3675             continue;
3676         }
3677         if (mode_search_skip_flags & FLAG_SKIP_INTRA_DIRMISMATCH) {
3678           if (conditional_skipintra(this_mode, best_intra_mode)) continue;
3679         }
3680       }
3681     } else {
3682       const MV_REFERENCE_FRAME ref_frames[2] = { ref_frame, second_ref_frame };
3683       if (!check_best_zero_mv(cpi, mbmi_ext->mode_context, frame_mv, this_mode,
3684                               ref_frames))
3685         continue;
3686     }
3687 
3688     mi->mode = this_mode;
3689     mi->uv_mode = DC_PRED;
3690     mi->ref_frame[0] = ref_frame;
3691     mi->ref_frame[1] = second_ref_frame;
3692     // Evaluate all sub-pel filters irrespective of whether we can use
3693     // them for this frame.
3694     mi->interp_filter =
3695         cm->interp_filter == SWITCHABLE ? EIGHTTAP : cm->interp_filter;
3696     mi->mv[0].as_int = mi->mv[1].as_int = 0;
3697 
3698     x->skip = 0;
3699     set_ref_ptrs(cm, xd, ref_frame, second_ref_frame);
3700 
3701     // Select prediction reference frames.
3702     for (i = 0; i < MAX_MB_PLANE; i++) {
3703       xd->plane[i].pre[0] = yv12_mb[ref_frame][i];
3704       if (comp_pred) xd->plane[i].pre[1] = yv12_mb[second_ref_frame][i];
3705     }
3706 
3707     if (ref_frame == INTRA_FRAME) {
3708       TX_SIZE uv_tx;
3709       struct macroblockd_plane *const pd = &xd->plane[1];
3710       memset(x->skip_txfm, 0, sizeof(x->skip_txfm));
3711       super_block_yrd(cpi, x, &rate_y, &distortion_y, &skippable, NULL, bsize,
3712                       best_rd, recon);
3713       if (rate_y == INT_MAX) continue;
3714 
3715       uv_tx = uv_txsize_lookup[bsize][mi->tx_size][pd->subsampling_x]
3716                               [pd->subsampling_y];
3717       if (rate_uv_intra[uv_tx] == INT_MAX) {
3718         choose_intra_uv_mode(cpi, x, ctx, bsize, uv_tx, &rate_uv_intra[uv_tx],
3719                              &rate_uv_tokenonly[uv_tx], &dist_uv[uv_tx],
3720                              &skip_uv[uv_tx], &mode_uv[uv_tx]);
3721       }
3722 
3723       rate_uv = rate_uv_tokenonly[uv_tx];
3724       distortion_uv = dist_uv[uv_tx];
3725       skippable = skippable && skip_uv[uv_tx];
3726       mi->uv_mode = mode_uv[uv_tx];
3727 
3728       rate2 = rate_y + cpi->mbmode_cost[mi->mode] + rate_uv_intra[uv_tx];
3729       if (this_mode != DC_PRED && this_mode != TM_PRED)
3730         rate2 += intra_cost_penalty;
3731       distortion2 = distortion_y + distortion_uv;
3732     } else {
3733       this_rd = handle_inter_mode(
3734           cpi, x, bsize, &rate2, &distortion2, &skippable, &rate_y, &rate_uv,
3735           recon, &disable_skip, frame_mv, mi_row, mi_col, single_newmv,
3736           single_inter_filter, single_skippable, &total_sse, best_rd,
3737           &mask_filter, filter_cache);
3738       if (this_rd == INT64_MAX) continue;
3739 
3740       compmode_cost = vp9_cost_bit(comp_mode_p, comp_pred);
3741 
3742       if (cm->reference_mode == REFERENCE_MODE_SELECT) rate2 += compmode_cost;
3743     }
3744 
3745     // Estimate the reference frame signaling cost and add it
3746     // to the rolling cost variable.
3747     if (comp_pred) {
3748       rate2 += ref_costs_comp[ref_frame];
3749     } else {
3750       rate2 += ref_costs_single[ref_frame];
3751     }
3752 
3753     if (!disable_skip) {
3754       const vpx_prob skip_prob = vp9_get_skip_prob(cm, xd);
3755       const int skip_cost0 = vp9_cost_bit(skip_prob, 0);
3756       const int skip_cost1 = vp9_cost_bit(skip_prob, 1);
3757 
3758       if (skippable) {
3759         // Back out the coefficient coding costs
3760         rate2 -= (rate_y + rate_uv);
3761 
3762         // Cost the skip mb case
3763         rate2 += skip_cost1;
3764       } else if (ref_frame != INTRA_FRAME && !xd->lossless &&
3765                  !cpi->oxcf.sharpness) {
3766         if (RDCOST(x->rdmult, x->rddiv, rate_y + rate_uv + skip_cost0,
3767                    distortion2) <
3768             RDCOST(x->rdmult, x->rddiv, skip_cost1, total_sse)) {
3769           // Add in the cost of the no skip flag.
3770           rate2 += skip_cost0;
3771         } else {
3772           // FIXME(rbultje) make this work for splitmv also
3773           assert(total_sse >= 0);
3774 
3775           rate2 += skip_cost1;
3776           distortion2 = total_sse;
3777           rate2 -= (rate_y + rate_uv);
3778           this_skip2 = 1;
3779         }
3780       } else {
3781         // Add in the cost of the no skip flag.
3782         rate2 += skip_cost0;
3783       }
3784 
3785       // Calculate the final RD estimate for this mode.
3786       this_rd = RDCOST(x->rdmult, x->rddiv, rate2, distortion2);
3787     }
3788 
3789     if (recon) {
3790       // In film mode bias against DC pred and other intra if there is a
3791       // significant difference between the variance of the sub blocks in the
3792       // the source. Also apply some bias against compound modes which also
3793       // tend to blur fine texture such as film grain over time.
3794       //
3795       // The sub block test here acts in the case where one or more sub
3796       // blocks have high relatively variance but others relatively low
3797       // variance. Here the high variance sub blocks may push the
3798       // total variance for the current block size over the thresholds
3799       // used in rd_variance_adjustment() below.
3800       if (cpi->oxcf.content == VP9E_CONTENT_FILM) {
3801         if (bsize >= BLOCK_16X16) {
3802           int min_energy, max_energy;
3803           vp9_get_sub_block_energy(cpi, x, mi_row, mi_col, bsize, &min_energy,
3804                                    &max_energy);
3805           if (max_energy > min_energy) {
3806             if (ref_frame == INTRA_FRAME) {
3807               if (this_mode == DC_PRED)
3808                 this_rd += (this_rd * (max_energy - min_energy));
3809               else
3810                 this_rd += (this_rd * (max_energy - min_energy)) / 4;
3811             } else if (second_ref_frame > INTRA_FRAME) {
3812               this_rd += this_rd / 4;
3813             }
3814           }
3815         }
3816       }
3817       // Apply an adjustment to the rd value based on the similarity of the
3818       // source variance and reconstructed variance.
3819       rd_variance_adjustment(cpi, x, bsize, &this_rd, recon, ref_frame,
3820                              second_ref_frame, this_mode);
3821     }
3822 
3823     if (ref_frame == INTRA_FRAME) {
3824       // Keep record of best intra rd
3825       if (this_rd < best_intra_rd) {
3826         best_intra_rd = this_rd;
3827         best_intra_mode = mi->mode;
3828       }
3829     }
3830 
3831     if (!disable_skip && ref_frame == INTRA_FRAME) {
3832       for (i = 0; i < REFERENCE_MODES; ++i)
3833         best_pred_rd[i] = VPXMIN(best_pred_rd[i], this_rd);
3834       for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++)
3835         best_filter_rd[i] = VPXMIN(best_filter_rd[i], this_rd);
3836     }
3837 
3838     // Did this mode help.. i.e. is it the new best mode
3839     if (this_rd < best_rd || x->skip) {
3840       int max_plane = MAX_MB_PLANE;
3841       if (!mode_excluded) {
3842         // Note index of best mode so far
3843         best_mode_index = mode_index;
3844 
3845         if (ref_frame == INTRA_FRAME) {
3846           /* required for left and above block mv */
3847           mi->mv[0].as_int = 0;
3848           max_plane = 1;
3849           // Initialize interp_filter here so we do not have to check for
3850           // inter block modes in get_pred_context_switchable_interp()
3851           mi->interp_filter = SWITCHABLE_FILTERS;
3852         } else {
3853           best_pred_sse = x->pred_sse[ref_frame];
3854         }
3855 
3856         rd_cost->rate = rate2;
3857         rd_cost->dist = distortion2;
3858         rd_cost->rdcost = this_rd;
3859         best_rd = this_rd;
3860         best_mbmode = *mi;
3861         best_skip2 = this_skip2;
3862         best_mode_skippable = skippable;
3863 
3864         if (!x->select_tx_size) swap_block_ptr(x, ctx, 1, 0, 0, max_plane);
3865         memcpy(ctx->zcoeff_blk, x->zcoeff_blk[mi->tx_size],
3866                sizeof(ctx->zcoeff_blk[0]) * ctx->num_4x4_blk);
3867         ctx->sum_y_eobs = x->sum_y_eobs[mi->tx_size];
3868 
3869         // TODO(debargha): enhance this test with a better distortion prediction
3870         // based on qp, activity mask and history
3871         if ((mode_search_skip_flags & FLAG_EARLY_TERMINATE) &&
3872             (mode_index > MIN_EARLY_TERM_INDEX)) {
3873           int qstep = xd->plane[0].dequant[1];
3874           // TODO(debargha): Enhance this by specializing for each mode_index
3875           int scale = 4;
3876 #if CONFIG_VP9_HIGHBITDEPTH
3877           if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
3878             qstep >>= (xd->bd - 8);
3879           }
3880 #endif  // CONFIG_VP9_HIGHBITDEPTH
3881           if (x->source_variance < UINT_MAX) {
3882             const int var_adjust = (x->source_variance < 16);
3883             scale -= var_adjust;
3884           }
3885           if (ref_frame > INTRA_FRAME && distortion2 * scale < qstep * qstep) {
3886             early_term = 1;
3887           }
3888         }
3889       }
3890     }
3891 
3892     /* keep record of best compound/single-only prediction */
3893     if (!disable_skip && ref_frame != INTRA_FRAME) {
3894       int64_t single_rd, hybrid_rd, single_rate, hybrid_rate;
3895 
3896       if (cm->reference_mode == REFERENCE_MODE_SELECT) {
3897         single_rate = rate2 - compmode_cost;
3898         hybrid_rate = rate2;
3899       } else {
3900         single_rate = rate2;
3901         hybrid_rate = rate2 + compmode_cost;
3902       }
3903 
3904       single_rd = RDCOST(x->rdmult, x->rddiv, single_rate, distortion2);
3905       hybrid_rd = RDCOST(x->rdmult, x->rddiv, hybrid_rate, distortion2);
3906 
3907       if (!comp_pred) {
3908         if (single_rd < best_pred_rd[SINGLE_REFERENCE])
3909           best_pred_rd[SINGLE_REFERENCE] = single_rd;
3910       } else {
3911         if (single_rd < best_pred_rd[COMPOUND_REFERENCE])
3912           best_pred_rd[COMPOUND_REFERENCE] = single_rd;
3913       }
3914       if (hybrid_rd < best_pred_rd[REFERENCE_MODE_SELECT])
3915         best_pred_rd[REFERENCE_MODE_SELECT] = hybrid_rd;
3916 
3917       /* keep record of best filter type */
3918       if (!mode_excluded && cm->interp_filter != BILINEAR) {
3919         int64_t ref =
3920             filter_cache[cm->interp_filter == SWITCHABLE ? SWITCHABLE_FILTERS
3921                                                          : cm->interp_filter];
3922 
3923         for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++) {
3924           int64_t adj_rd;
3925           if (ref == INT64_MAX)
3926             adj_rd = 0;
3927           else if (filter_cache[i] == INT64_MAX)
3928             // when early termination is triggered, the encoder does not have
3929             // access to the rate-distortion cost. it only knows that the cost
3930             // should be above the maximum valid value. hence it takes the known
3931             // maximum plus an arbitrary constant as the rate-distortion cost.
3932             adj_rd = mask_filter - ref + 10;
3933           else
3934             adj_rd = filter_cache[i] - ref;
3935 
3936           adj_rd += this_rd;
3937           best_filter_rd[i] = VPXMIN(best_filter_rd[i], adj_rd);
3938         }
3939       }
3940     }
3941 
3942     if (early_term) break;
3943 
3944     if (x->skip && !comp_pred) break;
3945   }
3946 
3947   // The inter modes' rate costs are not calculated precisely in some cases.
3948   // Therefore, sometimes, NEWMV is chosen instead of NEARESTMV, NEARMV, and
3949   // ZEROMV. Here, checks are added for those cases, and the mode decisions
3950   // are corrected.
3951   if (best_mbmode.mode == NEWMV) {
3952     const MV_REFERENCE_FRAME refs[2] = { best_mbmode.ref_frame[0],
3953                                          best_mbmode.ref_frame[1] };
3954     int comp_pred_mode = refs[1] > INTRA_FRAME;
3955 
3956     if (frame_mv[NEARESTMV][refs[0]].as_int == best_mbmode.mv[0].as_int &&
3957         ((comp_pred_mode &&
3958           frame_mv[NEARESTMV][refs[1]].as_int == best_mbmode.mv[1].as_int) ||
3959          !comp_pred_mode))
3960       best_mbmode.mode = NEARESTMV;
3961     else if (frame_mv[NEARMV][refs[0]].as_int == best_mbmode.mv[0].as_int &&
3962              ((comp_pred_mode &&
3963                frame_mv[NEARMV][refs[1]].as_int == best_mbmode.mv[1].as_int) ||
3964               !comp_pred_mode))
3965       best_mbmode.mode = NEARMV;
3966     else if (best_mbmode.mv[0].as_int == 0 &&
3967              ((comp_pred_mode && best_mbmode.mv[1].as_int == 0) ||
3968               !comp_pred_mode))
3969       best_mbmode.mode = ZEROMV;
3970   }
3971 
3972   if (best_mode_index < 0 || best_rd >= best_rd_so_far) {
3973 // If adaptive interp filter is enabled, then the current leaf node of 8x8
3974 // data is needed for sub8x8. Hence preserve the context.
3975 #if CONFIG_CONSISTENT_RECODE
3976     if (bsize == BLOCK_8X8) ctx->mic = *xd->mi[0];
3977 #else
3978     if (cpi->row_mt && bsize == BLOCK_8X8) ctx->mic = *xd->mi[0];
3979 #endif
3980     rd_cost->rate = INT_MAX;
3981     rd_cost->rdcost = INT64_MAX;
3982     return;
3983   }
3984 
3985   // If we used an estimate for the uv intra rd in the loop above...
3986   if (sf->use_uv_intra_rd_estimate) {
3987     // Do Intra UV best rd mode selection if best mode choice above was intra.
3988     if (best_mbmode.ref_frame[0] == INTRA_FRAME) {
3989       TX_SIZE uv_tx_size;
3990       *mi = best_mbmode;
3991       uv_tx_size = get_uv_tx_size(mi, &xd->plane[1]);
3992       rd_pick_intra_sbuv_mode(cpi, x, ctx, &rate_uv_intra[uv_tx_size],
3993                               &rate_uv_tokenonly[uv_tx_size],
3994                               &dist_uv[uv_tx_size], &skip_uv[uv_tx_size],
3995                               bsize < BLOCK_8X8 ? BLOCK_8X8 : bsize,
3996                               uv_tx_size);
3997     }
3998   }
3999 
4000   assert((cm->interp_filter == SWITCHABLE) ||
4001          (cm->interp_filter == best_mbmode.interp_filter) ||
4002          !is_inter_block(&best_mbmode));
4003 
4004   if (!cpi->rc.is_src_frame_alt_ref)
4005     vp9_update_rd_thresh_fact(tile_data->thresh_freq_fact,
4006                               sf->adaptive_rd_thresh, bsize, best_mode_index);
4007 
4008   // macroblock modes
4009   *mi = best_mbmode;
4010   x->skip |= best_skip2;
4011 
4012   for (i = 0; i < REFERENCE_MODES; ++i) {
4013     if (best_pred_rd[i] == INT64_MAX)
4014       best_pred_diff[i] = INT_MIN;
4015     else
4016       best_pred_diff[i] = best_rd - best_pred_rd[i];
4017   }
4018 
4019   if (!x->skip) {
4020     for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++) {
4021       if (best_filter_rd[i] == INT64_MAX)
4022         best_filter_diff[i] = 0;
4023       else
4024         best_filter_diff[i] = best_rd - best_filter_rd[i];
4025     }
4026     if (cm->interp_filter == SWITCHABLE)
4027       assert(best_filter_diff[SWITCHABLE_FILTERS] == 0);
4028   } else {
4029     vp9_zero(best_filter_diff);
4030   }
4031 
4032   // TODO(yunqingwang): Moving this line in front of the above best_filter_diff
4033   // updating code causes PSNR loss. Need to figure out the confliction.
4034   x->skip |= best_mode_skippable;
4035 
4036   if (!x->skip && !x->select_tx_size) {
4037     int has_high_freq_coeff = 0;
4038     int plane;
4039     int max_plane = is_inter_block(xd->mi[0]) ? MAX_MB_PLANE : 1;
4040     for (plane = 0; plane < max_plane; ++plane) {
4041       x->plane[plane].eobs = ctx->eobs_pbuf[plane][1];
4042       has_high_freq_coeff |= vp9_has_high_freq_in_plane(x, bsize, plane);
4043     }
4044 
4045     for (plane = max_plane; plane < MAX_MB_PLANE; ++plane) {
4046       x->plane[plane].eobs = ctx->eobs_pbuf[plane][2];
4047       has_high_freq_coeff |= vp9_has_high_freq_in_plane(x, bsize, plane);
4048     }
4049 
4050     best_mode_skippable |= !has_high_freq_coeff;
4051   }
4052 
4053   assert(best_mode_index >= 0);
4054 
4055   store_coding_context(x, ctx, best_mode_index, best_pred_diff,
4056                        best_filter_diff, best_mode_skippable);
4057 }
4058 
vp9_rd_pick_inter_mode_sb_seg_skip(VP9_COMP * cpi,TileDataEnc * tile_data,MACROBLOCK * x,RD_COST * rd_cost,BLOCK_SIZE bsize,PICK_MODE_CONTEXT * ctx,int64_t best_rd_so_far)4059 void vp9_rd_pick_inter_mode_sb_seg_skip(VP9_COMP *cpi, TileDataEnc *tile_data,
4060                                         MACROBLOCK *x, RD_COST *rd_cost,
4061                                         BLOCK_SIZE bsize,
4062                                         PICK_MODE_CONTEXT *ctx,
4063                                         int64_t best_rd_so_far) {
4064   VP9_COMMON *const cm = &cpi->common;
4065   MACROBLOCKD *const xd = &x->e_mbd;
4066   MODE_INFO *const mi = xd->mi[0];
4067   unsigned char segment_id = mi->segment_id;
4068   const int comp_pred = 0;
4069   int i;
4070   int64_t best_pred_diff[REFERENCE_MODES];
4071   int64_t best_filter_diff[SWITCHABLE_FILTER_CONTEXTS];
4072   unsigned int ref_costs_single[MAX_REF_FRAMES], ref_costs_comp[MAX_REF_FRAMES];
4073   vpx_prob comp_mode_p;
4074   INTERP_FILTER best_filter = SWITCHABLE;
4075   int64_t this_rd = INT64_MAX;
4076   int rate2 = 0;
4077   const int64_t distortion2 = 0;
4078 
4079   x->skip_encode = cpi->sf.skip_encode_frame && x->q_index < QIDX_SKIP_THRESH;
4080 
4081   estimate_ref_frame_costs(cm, xd, segment_id, ref_costs_single, ref_costs_comp,
4082                            &comp_mode_p);
4083 
4084   for (i = 0; i < MAX_REF_FRAMES; ++i) x->pred_sse[i] = INT_MAX;
4085   for (i = LAST_FRAME; i < MAX_REF_FRAMES; ++i) x->pred_mv_sad[i] = INT_MAX;
4086 
4087   rd_cost->rate = INT_MAX;
4088 
4089   assert(segfeature_active(&cm->seg, segment_id, SEG_LVL_SKIP));
4090 
4091   mi->mode = ZEROMV;
4092   mi->uv_mode = DC_PRED;
4093   mi->ref_frame[0] = LAST_FRAME;
4094   mi->ref_frame[1] = NONE;
4095   mi->mv[0].as_int = 0;
4096   x->skip = 1;
4097 
4098   ctx->sum_y_eobs = 0;
4099 
4100   if (cm->interp_filter != BILINEAR) {
4101     best_filter = EIGHTTAP;
4102     if (cm->interp_filter == SWITCHABLE &&
4103         x->source_variance >= cpi->sf.disable_filter_search_var_thresh) {
4104       int rs;
4105       int best_rs = INT_MAX;
4106       for (i = 0; i < SWITCHABLE_FILTERS; ++i) {
4107         mi->interp_filter = i;
4108         rs = vp9_get_switchable_rate(cpi, xd);
4109         if (rs < best_rs) {
4110           best_rs = rs;
4111           best_filter = mi->interp_filter;
4112         }
4113       }
4114     }
4115   }
4116   // Set the appropriate filter
4117   if (cm->interp_filter == SWITCHABLE) {
4118     mi->interp_filter = best_filter;
4119     rate2 += vp9_get_switchable_rate(cpi, xd);
4120   } else {
4121     mi->interp_filter = cm->interp_filter;
4122   }
4123 
4124   if (cm->reference_mode == REFERENCE_MODE_SELECT)
4125     rate2 += vp9_cost_bit(comp_mode_p, comp_pred);
4126 
4127   // Estimate the reference frame signaling cost and add it
4128   // to the rolling cost variable.
4129   rate2 += ref_costs_single[LAST_FRAME];
4130   this_rd = RDCOST(x->rdmult, x->rddiv, rate2, distortion2);
4131 
4132   rd_cost->rate = rate2;
4133   rd_cost->dist = distortion2;
4134   rd_cost->rdcost = this_rd;
4135 
4136   if (this_rd >= best_rd_so_far) {
4137     rd_cost->rate = INT_MAX;
4138     rd_cost->rdcost = INT64_MAX;
4139     return;
4140   }
4141 
4142   assert((cm->interp_filter == SWITCHABLE) ||
4143          (cm->interp_filter == mi->interp_filter));
4144 
4145   vp9_update_rd_thresh_fact(tile_data->thresh_freq_fact,
4146                             cpi->sf.adaptive_rd_thresh, bsize, THR_ZEROMV);
4147 
4148   vp9_zero(best_pred_diff);
4149   vp9_zero(best_filter_diff);
4150 
4151   if (!x->select_tx_size) swap_block_ptr(x, ctx, 1, 0, 0, MAX_MB_PLANE);
4152   store_coding_context(x, ctx, THR_ZEROMV, best_pred_diff, best_filter_diff, 0);
4153 }
4154 
vp9_rd_pick_inter_mode_sub8x8(VP9_COMP * cpi,TileDataEnc * tile_data,MACROBLOCK * x,int mi_row,int mi_col,RD_COST * rd_cost,BLOCK_SIZE bsize,PICK_MODE_CONTEXT * ctx,int64_t best_rd_so_far)4155 void vp9_rd_pick_inter_mode_sub8x8(VP9_COMP *cpi, TileDataEnc *tile_data,
4156                                    MACROBLOCK *x, int mi_row, int mi_col,
4157                                    RD_COST *rd_cost, BLOCK_SIZE bsize,
4158                                    PICK_MODE_CONTEXT *ctx,
4159                                    int64_t best_rd_so_far) {
4160   VP9_COMMON *const cm = &cpi->common;
4161   RD_OPT *const rd_opt = &cpi->rd;
4162   SPEED_FEATURES *const sf = &cpi->sf;
4163   MACROBLOCKD *const xd = &x->e_mbd;
4164   MODE_INFO *const mi = xd->mi[0];
4165   const struct segmentation *const seg = &cm->seg;
4166   MV_REFERENCE_FRAME ref_frame, second_ref_frame;
4167   unsigned char segment_id = mi->segment_id;
4168   int comp_pred, i;
4169   int_mv frame_mv[MB_MODE_COUNT][MAX_REF_FRAMES];
4170   struct buf_2d yv12_mb[4][MAX_MB_PLANE];
4171   int64_t best_rd = best_rd_so_far;
4172   int64_t best_yrd = best_rd_so_far;  // FIXME(rbultje) more precise
4173   int64_t best_pred_diff[REFERENCE_MODES];
4174   int64_t best_pred_rd[REFERENCE_MODES];
4175   int64_t best_filter_rd[SWITCHABLE_FILTER_CONTEXTS];
4176   int64_t best_filter_diff[SWITCHABLE_FILTER_CONTEXTS];
4177   MODE_INFO best_mbmode;
4178   int ref_index, best_ref_index = 0;
4179   unsigned int ref_costs_single[MAX_REF_FRAMES], ref_costs_comp[MAX_REF_FRAMES];
4180   vpx_prob comp_mode_p;
4181   INTERP_FILTER tmp_best_filter = SWITCHABLE;
4182   int rate_uv_intra, rate_uv_tokenonly;
4183   int64_t dist_uv;
4184   int skip_uv;
4185   PREDICTION_MODE mode_uv = DC_PRED;
4186   const int intra_cost_penalty =
4187       vp9_get_intra_cost_penalty(cpi, bsize, cm->base_qindex, cm->y_dc_delta_q);
4188   int_mv seg_mvs[4][MAX_REF_FRAMES];
4189   b_mode_info best_bmodes[4];
4190   int best_skip2 = 0;
4191   int ref_frame_skip_mask[2] = { 0 };
4192   int64_t mask_filter = 0;
4193   int64_t filter_cache[SWITCHABLE_FILTER_CONTEXTS];
4194   int internal_active_edge =
4195       vp9_active_edge_sb(cpi, mi_row, mi_col) && vp9_internal_image_edge(cpi);
4196   const int *const rd_thresh_freq_fact = tile_data->thresh_freq_fact[bsize];
4197 
4198   x->skip_encode = sf->skip_encode_frame && x->q_index < QIDX_SKIP_THRESH;
4199   memset(x->zcoeff_blk[TX_4X4], 0, 4);
4200   vp9_zero(best_mbmode);
4201 
4202   for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i) filter_cache[i] = INT64_MAX;
4203 
4204   for (i = 0; i < 4; i++) {
4205     int j;
4206     for (j = 0; j < MAX_REF_FRAMES; j++) seg_mvs[i][j].as_int = INVALID_MV;
4207   }
4208 
4209   estimate_ref_frame_costs(cm, xd, segment_id, ref_costs_single, ref_costs_comp,
4210                            &comp_mode_p);
4211 
4212   for (i = 0; i < REFERENCE_MODES; ++i) best_pred_rd[i] = INT64_MAX;
4213   for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++)
4214     best_filter_rd[i] = INT64_MAX;
4215   rate_uv_intra = INT_MAX;
4216 
4217   rd_cost->rate = INT_MAX;
4218 
4219   for (ref_frame = LAST_FRAME; ref_frame <= ALTREF_FRAME; ref_frame++) {
4220     if (cpi->ref_frame_flags & ref_frame_to_flag(ref_frame)) {
4221       setup_buffer_inter(cpi, x, ref_frame, bsize, mi_row, mi_col,
4222                          frame_mv[NEARESTMV], frame_mv[NEARMV], yv12_mb);
4223     } else {
4224       ref_frame_skip_mask[0] |= (1 << ref_frame);
4225       ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
4226     }
4227     frame_mv[NEWMV][ref_frame].as_int = INVALID_MV;
4228     frame_mv[ZEROMV][ref_frame].as_int = 0;
4229   }
4230 
4231   for (ref_index = 0; ref_index < MAX_REFS; ++ref_index) {
4232     int mode_excluded = 0;
4233     int64_t this_rd = INT64_MAX;
4234     int disable_skip = 0;
4235     int compmode_cost = 0;
4236     int rate2 = 0, rate_y = 0, rate_uv = 0;
4237     int64_t distortion2 = 0, distortion_y = 0, distortion_uv = 0;
4238     int skippable = 0;
4239     int i;
4240     int this_skip2 = 0;
4241     int64_t total_sse = INT_MAX;
4242     int early_term = 0;
4243     struct buf_2d backup_yv12[2][MAX_MB_PLANE];
4244 
4245     ref_frame = vp9_ref_order[ref_index].ref_frame[0];
4246     second_ref_frame = vp9_ref_order[ref_index].ref_frame[1];
4247 
4248     vp9_zero(x->sum_y_eobs);
4249 
4250 #if CONFIG_BETTER_HW_COMPATIBILITY
4251     // forbid 8X4 and 4X8 partitions if any reference frame is scaled.
4252     if (bsize == BLOCK_8X4 || bsize == BLOCK_4X8) {
4253       int ref_scaled = ref_frame > INTRA_FRAME &&
4254                        vp9_is_scaled(&cm->frame_refs[ref_frame - 1].sf);
4255       if (second_ref_frame > INTRA_FRAME)
4256         ref_scaled += vp9_is_scaled(&cm->frame_refs[second_ref_frame - 1].sf);
4257       if (ref_scaled) continue;
4258     }
4259 #endif
4260     // Look at the reference frame of the best mode so far and set the
4261     // skip mask to look at a subset of the remaining modes.
4262     if (ref_index > 2 && sf->mode_skip_start < MAX_MODES) {
4263       if (ref_index == 3) {
4264         switch (best_mbmode.ref_frame[0]) {
4265           case INTRA_FRAME: break;
4266           case LAST_FRAME:
4267             ref_frame_skip_mask[0] |= (1 << GOLDEN_FRAME) | (1 << ALTREF_FRAME);
4268             ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
4269             break;
4270           case GOLDEN_FRAME:
4271             ref_frame_skip_mask[0] |= (1 << LAST_FRAME) | (1 << ALTREF_FRAME);
4272             ref_frame_skip_mask[1] |= SECOND_REF_FRAME_MASK;
4273             break;
4274           case ALTREF_FRAME:
4275             ref_frame_skip_mask[0] |= (1 << GOLDEN_FRAME) | (1 << LAST_FRAME);
4276             break;
4277           case NONE:
4278           case MAX_REF_FRAMES: assert(0 && "Invalid Reference frame"); break;
4279         }
4280       }
4281     }
4282 
4283     if ((ref_frame_skip_mask[0] & (1 << ref_frame)) &&
4284         (ref_frame_skip_mask[1] & (1 << VPXMAX(0, second_ref_frame))))
4285       continue;
4286 
4287     // Test best rd so far against threshold for trying this mode.
4288     if (!internal_active_edge &&
4289         rd_less_than_thresh(best_rd,
4290                             rd_opt->threshes[segment_id][bsize][ref_index],
4291                             &rd_thresh_freq_fact[ref_index]))
4292       continue;
4293 
4294     // This is only used in motion vector unit test.
4295     if (cpi->oxcf.motion_vector_unit_test && ref_frame == INTRA_FRAME) continue;
4296 
4297     comp_pred = second_ref_frame > INTRA_FRAME;
4298     if (comp_pred) {
4299       if (!cpi->allow_comp_inter_inter) continue;
4300 
4301       if (cm->ref_frame_sign_bias[ref_frame] ==
4302           cm->ref_frame_sign_bias[second_ref_frame])
4303         continue;
4304 
4305       if (!(cpi->ref_frame_flags & ref_frame_to_flag(second_ref_frame)))
4306         continue;
4307       // Do not allow compound prediction if the segment level reference frame
4308       // feature is in use as in this case there can only be one reference.
4309       if (segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME)) continue;
4310 
4311       if ((sf->mode_search_skip_flags & FLAG_SKIP_COMP_BESTINTRA) &&
4312           best_mbmode.ref_frame[0] == INTRA_FRAME)
4313         continue;
4314     }
4315 
4316     if (comp_pred)
4317       mode_excluded = cm->reference_mode == SINGLE_REFERENCE;
4318     else if (ref_frame != INTRA_FRAME)
4319       mode_excluded = cm->reference_mode == COMPOUND_REFERENCE;
4320 
4321     // If the segment reference frame feature is enabled....
4322     // then do nothing if the current ref frame is not allowed..
4323     if (segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME) &&
4324         get_segdata(seg, segment_id, SEG_LVL_REF_FRAME) != (int)ref_frame) {
4325       continue;
4326       // Disable this drop out case if the ref frame
4327       // segment level feature is enabled for this segment. This is to
4328       // prevent the possibility that we end up unable to pick any mode.
4329     } else if (!segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME)) {
4330       // Only consider ZEROMV/ALTREF_FRAME for alt ref frame,
4331       // unless ARNR filtering is enabled in which case we want
4332       // an unfiltered alternative. We allow near/nearest as well
4333       // because they may result in zero-zero MVs but be cheaper.
4334       if (cpi->rc.is_src_frame_alt_ref && (cpi->oxcf.arnr_max_frames == 0))
4335         continue;
4336     }
4337 
4338     mi->tx_size = TX_4X4;
4339     mi->uv_mode = DC_PRED;
4340     mi->ref_frame[0] = ref_frame;
4341     mi->ref_frame[1] = second_ref_frame;
4342     // Evaluate all sub-pel filters irrespective of whether we can use
4343     // them for this frame.
4344     mi->interp_filter =
4345         cm->interp_filter == SWITCHABLE ? EIGHTTAP : cm->interp_filter;
4346     x->skip = 0;
4347     set_ref_ptrs(cm, xd, ref_frame, second_ref_frame);
4348 
4349     // Select prediction reference frames.
4350     for (i = 0; i < MAX_MB_PLANE; i++) {
4351       xd->plane[i].pre[0] = yv12_mb[ref_frame][i];
4352       if (comp_pred) xd->plane[i].pre[1] = yv12_mb[second_ref_frame][i];
4353     }
4354 
4355     if (ref_frame == INTRA_FRAME) {
4356       int rate;
4357       if (rd_pick_intra_sub_8x8_y_mode(cpi, x, &rate, &rate_y, &distortion_y,
4358                                        best_rd) >= best_rd)
4359         continue;
4360       rate2 += rate;
4361       rate2 += intra_cost_penalty;
4362       distortion2 += distortion_y;
4363 
4364       if (rate_uv_intra == INT_MAX) {
4365         choose_intra_uv_mode(cpi, x, ctx, bsize, TX_4X4, &rate_uv_intra,
4366                              &rate_uv_tokenonly, &dist_uv, &skip_uv, &mode_uv);
4367       }
4368       rate2 += rate_uv_intra;
4369       rate_uv = rate_uv_tokenonly;
4370       distortion2 += dist_uv;
4371       distortion_uv = dist_uv;
4372       mi->uv_mode = mode_uv;
4373     } else {
4374       int rate;
4375       int64_t distortion;
4376       int64_t this_rd_thresh;
4377       int64_t tmp_rd, tmp_best_rd = INT64_MAX, tmp_best_rdu = INT64_MAX;
4378       int tmp_best_rate = INT_MAX, tmp_best_ratey = INT_MAX;
4379       int64_t tmp_best_distortion = INT_MAX, tmp_best_sse, uv_sse;
4380       int tmp_best_skippable = 0;
4381       int switchable_filter_index;
4382       int_mv *second_ref =
4383           comp_pred ? &x->mbmi_ext->ref_mvs[second_ref_frame][0] : NULL;
4384       b_mode_info tmp_best_bmodes[16];
4385       MODE_INFO tmp_best_mbmode;
4386       BEST_SEG_INFO bsi[SWITCHABLE_FILTERS];
4387       int pred_exists = 0;
4388       int uv_skippable;
4389 
4390       YV12_BUFFER_CONFIG *scaled_ref_frame[2] = { NULL, NULL };
4391       int ref;
4392 
4393       for (ref = 0; ref < 2; ++ref) {
4394         scaled_ref_frame[ref] =
4395             mi->ref_frame[ref] > INTRA_FRAME
4396                 ? vp9_get_scaled_ref_frame(cpi, mi->ref_frame[ref])
4397                 : NULL;
4398 
4399         if (scaled_ref_frame[ref]) {
4400           int i;
4401           // Swap out the reference frame for a version that's been scaled to
4402           // match the resolution of the current frame, allowing the existing
4403           // motion search code to be used without additional modifications.
4404           for (i = 0; i < MAX_MB_PLANE; i++)
4405             backup_yv12[ref][i] = xd->plane[i].pre[ref];
4406           vp9_setup_pre_planes(xd, ref, scaled_ref_frame[ref], mi_row, mi_col,
4407                                NULL);
4408         }
4409       }
4410 
4411       this_rd_thresh = (ref_frame == LAST_FRAME)
4412                            ? rd_opt->threshes[segment_id][bsize][THR_LAST]
4413                            : rd_opt->threshes[segment_id][bsize][THR_ALTR];
4414       this_rd_thresh = (ref_frame == GOLDEN_FRAME)
4415                            ? rd_opt->threshes[segment_id][bsize][THR_GOLD]
4416                            : this_rd_thresh;
4417       for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i)
4418         filter_cache[i] = INT64_MAX;
4419 
4420       if (cm->interp_filter != BILINEAR) {
4421         tmp_best_filter = EIGHTTAP;
4422         if (x->source_variance < sf->disable_filter_search_var_thresh) {
4423           tmp_best_filter = EIGHTTAP;
4424         } else if (sf->adaptive_pred_interp_filter == 1 &&
4425                    ctx->pred_interp_filter < SWITCHABLE) {
4426           tmp_best_filter = ctx->pred_interp_filter;
4427         } else if (sf->adaptive_pred_interp_filter == 2) {
4428           tmp_best_filter = ctx->pred_interp_filter < SWITCHABLE
4429                                 ? ctx->pred_interp_filter
4430                                 : 0;
4431         } else {
4432           for (switchable_filter_index = 0;
4433                switchable_filter_index < SWITCHABLE_FILTERS;
4434                ++switchable_filter_index) {
4435             int newbest, rs;
4436             int64_t rs_rd;
4437             MB_MODE_INFO_EXT *mbmi_ext = x->mbmi_ext;
4438             mi->interp_filter = switchable_filter_index;
4439             tmp_rd = rd_pick_best_sub8x8_mode(
4440                 cpi, x, &mbmi_ext->ref_mvs[ref_frame][0], second_ref, best_yrd,
4441                 &rate, &rate_y, &distortion, &skippable, &total_sse,
4442                 (int)this_rd_thresh, seg_mvs, bsi, switchable_filter_index,
4443                 mi_row, mi_col);
4444 
4445             if (tmp_rd == INT64_MAX) continue;
4446             rs = vp9_get_switchable_rate(cpi, xd);
4447             rs_rd = RDCOST(x->rdmult, x->rddiv, rs, 0);
4448             filter_cache[switchable_filter_index] = tmp_rd;
4449             filter_cache[SWITCHABLE_FILTERS] =
4450                 VPXMIN(filter_cache[SWITCHABLE_FILTERS], tmp_rd + rs_rd);
4451             if (cm->interp_filter == SWITCHABLE) tmp_rd += rs_rd;
4452 
4453             mask_filter = VPXMAX(mask_filter, tmp_rd);
4454 
4455             newbest = (tmp_rd < tmp_best_rd);
4456             if (newbest) {
4457               tmp_best_filter = mi->interp_filter;
4458               tmp_best_rd = tmp_rd;
4459             }
4460             if ((newbest && cm->interp_filter == SWITCHABLE) ||
4461                 (mi->interp_filter == cm->interp_filter &&
4462                  cm->interp_filter != SWITCHABLE)) {
4463               tmp_best_rdu = tmp_rd;
4464               tmp_best_rate = rate;
4465               tmp_best_ratey = rate_y;
4466               tmp_best_distortion = distortion;
4467               tmp_best_sse = total_sse;
4468               tmp_best_skippable = skippable;
4469               tmp_best_mbmode = *mi;
4470               x->sum_y_eobs[TX_4X4] = 0;
4471               for (i = 0; i < 4; i++) {
4472                 tmp_best_bmodes[i] = xd->mi[0]->bmi[i];
4473                 x->zcoeff_blk[TX_4X4][i] = !x->plane[0].eobs[i];
4474                 x->sum_y_eobs[TX_4X4] += x->plane[0].eobs[i];
4475               }
4476               pred_exists = 1;
4477               if (switchable_filter_index == 0 && sf->use_rd_breakout &&
4478                   best_rd < INT64_MAX) {
4479                 if (tmp_best_rdu / 2 > best_rd) {
4480                   // skip searching the other filters if the first is
4481                   // already substantially larger than the best so far
4482                   tmp_best_filter = mi->interp_filter;
4483                   tmp_best_rdu = INT64_MAX;
4484                   break;
4485                 }
4486               }
4487             }
4488           }  // switchable_filter_index loop
4489         }
4490       }
4491 
4492       if (tmp_best_rdu == INT64_MAX && pred_exists) continue;
4493 
4494       mi->interp_filter = (cm->interp_filter == SWITCHABLE ? tmp_best_filter
4495                                                            : cm->interp_filter);
4496       if (!pred_exists) {
4497         // Handles the special case when a filter that is not in the
4498         // switchable list (bilinear, 6-tap) is indicated at the frame level
4499         tmp_rd = rd_pick_best_sub8x8_mode(
4500             cpi, x, &x->mbmi_ext->ref_mvs[ref_frame][0], second_ref, best_yrd,
4501             &rate, &rate_y, &distortion, &skippable, &total_sse,
4502             (int)this_rd_thresh, seg_mvs, bsi, 0, mi_row, mi_col);
4503         if (tmp_rd == INT64_MAX) continue;
4504         x->sum_y_eobs[TX_4X4] = 0;
4505         for (i = 0; i < 4; i++) {
4506           x->zcoeff_blk[TX_4X4][i] = !x->plane[0].eobs[i];
4507           x->sum_y_eobs[TX_4X4] += x->plane[0].eobs[i];
4508         }
4509       } else {
4510         total_sse = tmp_best_sse;
4511         rate = tmp_best_rate;
4512         rate_y = tmp_best_ratey;
4513         distortion = tmp_best_distortion;
4514         skippable = tmp_best_skippable;
4515         *mi = tmp_best_mbmode;
4516         for (i = 0; i < 4; i++) xd->mi[0]->bmi[i] = tmp_best_bmodes[i];
4517       }
4518 
4519       rate2 += rate;
4520       distortion2 += distortion;
4521 
4522       if (cm->interp_filter == SWITCHABLE)
4523         rate2 += vp9_get_switchable_rate(cpi, xd);
4524 
4525       if (!mode_excluded)
4526         mode_excluded = comp_pred ? cm->reference_mode == SINGLE_REFERENCE
4527                                   : cm->reference_mode == COMPOUND_REFERENCE;
4528 
4529       compmode_cost = vp9_cost_bit(comp_mode_p, comp_pred);
4530 
4531       tmp_best_rdu =
4532           best_rd - VPXMIN(RDCOST(x->rdmult, x->rddiv, rate2, distortion2),
4533                            RDCOST(x->rdmult, x->rddiv, 0, total_sse));
4534 
4535       if (tmp_best_rdu > 0) {
4536         // If even the 'Y' rd value of split is higher than best so far
4537         // then dont bother looking at UV
4538         vp9_build_inter_predictors_sbuv(&x->e_mbd, mi_row, mi_col, BLOCK_8X8);
4539         memset(x->skip_txfm, SKIP_TXFM_NONE, sizeof(x->skip_txfm));
4540         if (!super_block_uvrd(cpi, x, &rate_uv, &distortion_uv, &uv_skippable,
4541                               &uv_sse, BLOCK_8X8, tmp_best_rdu)) {
4542           for (ref = 0; ref < 2; ++ref) {
4543             if (scaled_ref_frame[ref]) {
4544               int i;
4545               for (i = 0; i < MAX_MB_PLANE; ++i)
4546                 xd->plane[i].pre[ref] = backup_yv12[ref][i];
4547             }
4548           }
4549           continue;
4550         }
4551 
4552         rate2 += rate_uv;
4553         distortion2 += distortion_uv;
4554         skippable = skippable && uv_skippable;
4555         total_sse += uv_sse;
4556       }
4557 
4558       for (ref = 0; ref < 2; ++ref) {
4559         if (scaled_ref_frame[ref]) {
4560           // Restore the prediction frame pointers to their unscaled versions.
4561           int i;
4562           for (i = 0; i < MAX_MB_PLANE; ++i)
4563             xd->plane[i].pre[ref] = backup_yv12[ref][i];
4564         }
4565       }
4566     }
4567 
4568     if (cm->reference_mode == REFERENCE_MODE_SELECT) rate2 += compmode_cost;
4569 
4570     // Estimate the reference frame signaling cost and add it
4571     // to the rolling cost variable.
4572     if (second_ref_frame > INTRA_FRAME) {
4573       rate2 += ref_costs_comp[ref_frame];
4574     } else {
4575       rate2 += ref_costs_single[ref_frame];
4576     }
4577 
4578     if (!disable_skip) {
4579       const vpx_prob skip_prob = vp9_get_skip_prob(cm, xd);
4580       const int skip_cost0 = vp9_cost_bit(skip_prob, 0);
4581       const int skip_cost1 = vp9_cost_bit(skip_prob, 1);
4582 
4583       // Skip is never coded at the segment level for sub8x8 blocks and instead
4584       // always coded in the bitstream at the mode info level.
4585       if (ref_frame != INTRA_FRAME && !xd->lossless) {
4586         if (RDCOST(x->rdmult, x->rddiv, rate_y + rate_uv + skip_cost0,
4587                    distortion2) <
4588             RDCOST(x->rdmult, x->rddiv, skip_cost1, total_sse)) {
4589           // Add in the cost of the no skip flag.
4590           rate2 += skip_cost0;
4591         } else {
4592           // FIXME(rbultje) make this work for splitmv also
4593           rate2 += skip_cost1;
4594           distortion2 = total_sse;
4595           assert(total_sse >= 0);
4596           rate2 -= (rate_y + rate_uv);
4597           rate_y = 0;
4598           rate_uv = 0;
4599           this_skip2 = 1;
4600         }
4601       } else {
4602         // Add in the cost of the no skip flag.
4603         rate2 += skip_cost0;
4604       }
4605 
4606       // Calculate the final RD estimate for this mode.
4607       this_rd = RDCOST(x->rdmult, x->rddiv, rate2, distortion2);
4608     }
4609 
4610     if (!disable_skip && ref_frame == INTRA_FRAME) {
4611       for (i = 0; i < REFERENCE_MODES; ++i)
4612         best_pred_rd[i] = VPXMIN(best_pred_rd[i], this_rd);
4613       for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++)
4614         best_filter_rd[i] = VPXMIN(best_filter_rd[i], this_rd);
4615     }
4616 
4617     // Did this mode help.. i.e. is it the new best mode
4618     if (this_rd < best_rd || x->skip) {
4619       if (!mode_excluded) {
4620         int max_plane = MAX_MB_PLANE;
4621         // Note index of best mode so far
4622         best_ref_index = ref_index;
4623 
4624         if (ref_frame == INTRA_FRAME) {
4625           /* required for left and above block mv */
4626           mi->mv[0].as_int = 0;
4627           max_plane = 1;
4628           // Initialize interp_filter here so we do not have to check for
4629           // inter block modes in get_pred_context_switchable_interp()
4630           mi->interp_filter = SWITCHABLE_FILTERS;
4631         }
4632 
4633         rd_cost->rate = rate2;
4634         rd_cost->dist = distortion2;
4635         rd_cost->rdcost = this_rd;
4636         best_rd = this_rd;
4637         best_yrd =
4638             best_rd - RDCOST(x->rdmult, x->rddiv, rate_uv, distortion_uv);
4639         best_mbmode = *mi;
4640         best_skip2 = this_skip2;
4641         if (!x->select_tx_size) swap_block_ptr(x, ctx, 1, 0, 0, max_plane);
4642         memcpy(ctx->zcoeff_blk, x->zcoeff_blk[TX_4X4],
4643                sizeof(ctx->zcoeff_blk[0]) * ctx->num_4x4_blk);
4644         ctx->sum_y_eobs = x->sum_y_eobs[TX_4X4];
4645 
4646         for (i = 0; i < 4; i++) best_bmodes[i] = xd->mi[0]->bmi[i];
4647 
4648         // TODO(debargha): enhance this test with a better distortion prediction
4649         // based on qp, activity mask and history
4650         if ((sf->mode_search_skip_flags & FLAG_EARLY_TERMINATE) &&
4651             (ref_index > MIN_EARLY_TERM_INDEX)) {
4652           int qstep = xd->plane[0].dequant[1];
4653           // TODO(debargha): Enhance this by specializing for each mode_index
4654           int scale = 4;
4655 #if CONFIG_VP9_HIGHBITDEPTH
4656           if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
4657             qstep >>= (xd->bd - 8);
4658           }
4659 #endif  // CONFIG_VP9_HIGHBITDEPTH
4660           if (x->source_variance < UINT_MAX) {
4661             const int var_adjust = (x->source_variance < 16);
4662             scale -= var_adjust;
4663           }
4664           if (ref_frame > INTRA_FRAME && distortion2 * scale < qstep * qstep) {
4665             early_term = 1;
4666           }
4667         }
4668       }
4669     }
4670 
4671     /* keep record of best compound/single-only prediction */
4672     if (!disable_skip && ref_frame != INTRA_FRAME) {
4673       int64_t single_rd, hybrid_rd, single_rate, hybrid_rate;
4674 
4675       if (cm->reference_mode == REFERENCE_MODE_SELECT) {
4676         single_rate = rate2 - compmode_cost;
4677         hybrid_rate = rate2;
4678       } else {
4679         single_rate = rate2;
4680         hybrid_rate = rate2 + compmode_cost;
4681       }
4682 
4683       single_rd = RDCOST(x->rdmult, x->rddiv, single_rate, distortion2);
4684       hybrid_rd = RDCOST(x->rdmult, x->rddiv, hybrid_rate, distortion2);
4685 
4686       if (!comp_pred && single_rd < best_pred_rd[SINGLE_REFERENCE])
4687         best_pred_rd[SINGLE_REFERENCE] = single_rd;
4688       else if (comp_pred && single_rd < best_pred_rd[COMPOUND_REFERENCE])
4689         best_pred_rd[COMPOUND_REFERENCE] = single_rd;
4690 
4691       if (hybrid_rd < best_pred_rd[REFERENCE_MODE_SELECT])
4692         best_pred_rd[REFERENCE_MODE_SELECT] = hybrid_rd;
4693     }
4694 
4695     /* keep record of best filter type */
4696     if (!mode_excluded && !disable_skip && ref_frame != INTRA_FRAME &&
4697         cm->interp_filter != BILINEAR) {
4698       int64_t ref =
4699           filter_cache[cm->interp_filter == SWITCHABLE ? SWITCHABLE_FILTERS
4700                                                        : cm->interp_filter];
4701       int64_t adj_rd;
4702       for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++) {
4703         if (ref == INT64_MAX)
4704           adj_rd = 0;
4705         else if (filter_cache[i] == INT64_MAX)
4706           // when early termination is triggered, the encoder does not have
4707           // access to the rate-distortion cost. it only knows that the cost
4708           // should be above the maximum valid value. hence it takes the known
4709           // maximum plus an arbitrary constant as the rate-distortion cost.
4710           adj_rd = mask_filter - ref + 10;
4711         else
4712           adj_rd = filter_cache[i] - ref;
4713 
4714         adj_rd += this_rd;
4715         best_filter_rd[i] = VPXMIN(best_filter_rd[i], adj_rd);
4716       }
4717     }
4718 
4719     if (early_term) break;
4720 
4721     if (x->skip && !comp_pred) break;
4722   }
4723 
4724   if (best_rd >= best_rd_so_far) {
4725     rd_cost->rate = INT_MAX;
4726     rd_cost->rdcost = INT64_MAX;
4727     return;
4728   }
4729 
4730   // If we used an estimate for the uv intra rd in the loop above...
4731   if (sf->use_uv_intra_rd_estimate) {
4732     // Do Intra UV best rd mode selection if best mode choice above was intra.
4733     if (best_mbmode.ref_frame[0] == INTRA_FRAME) {
4734       *mi = best_mbmode;
4735       rd_pick_intra_sbuv_mode(cpi, x, ctx, &rate_uv_intra, &rate_uv_tokenonly,
4736                               &dist_uv, &skip_uv, BLOCK_8X8, TX_4X4);
4737     }
4738   }
4739 
4740   if (best_rd == INT64_MAX) {
4741     rd_cost->rate = INT_MAX;
4742     rd_cost->dist = INT64_MAX;
4743     rd_cost->rdcost = INT64_MAX;
4744     return;
4745   }
4746 
4747   assert((cm->interp_filter == SWITCHABLE) ||
4748          (cm->interp_filter == best_mbmode.interp_filter) ||
4749          !is_inter_block(&best_mbmode));
4750 
4751   vp9_update_rd_thresh_fact(tile_data->thresh_freq_fact, sf->adaptive_rd_thresh,
4752                             bsize, best_ref_index);
4753 
4754   // macroblock modes
4755   *mi = best_mbmode;
4756   x->skip |= best_skip2;
4757   if (!is_inter_block(&best_mbmode)) {
4758     for (i = 0; i < 4; i++) xd->mi[0]->bmi[i].as_mode = best_bmodes[i].as_mode;
4759   } else {
4760     for (i = 0; i < 4; ++i)
4761       memcpy(&xd->mi[0]->bmi[i], &best_bmodes[i], sizeof(b_mode_info));
4762 
4763     mi->mv[0].as_int = xd->mi[0]->bmi[3].as_mv[0].as_int;
4764     mi->mv[1].as_int = xd->mi[0]->bmi[3].as_mv[1].as_int;
4765   }
4766   // If the second reference does not exist, set the corresponding mv to zero.
4767   if (mi->ref_frame[1] == NONE) {
4768     mi->mv[1].as_int = 0;
4769     for (i = 0; i < 4; ++i) {
4770       mi->bmi[i].as_mv[1].as_int = 0;
4771     }
4772   }
4773 
4774   for (i = 0; i < REFERENCE_MODES; ++i) {
4775     if (best_pred_rd[i] == INT64_MAX)
4776       best_pred_diff[i] = INT_MIN;
4777     else
4778       best_pred_diff[i] = best_rd - best_pred_rd[i];
4779   }
4780 
4781   if (!x->skip) {
4782     for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; i++) {
4783       if (best_filter_rd[i] == INT64_MAX)
4784         best_filter_diff[i] = 0;
4785       else
4786         best_filter_diff[i] = best_rd - best_filter_rd[i];
4787     }
4788     if (cm->interp_filter == SWITCHABLE)
4789       assert(best_filter_diff[SWITCHABLE_FILTERS] == 0);
4790   } else {
4791     vp9_zero(best_filter_diff);
4792   }
4793 
4794   store_coding_context(x, ctx, best_ref_index, best_pred_diff, best_filter_diff,
4795                        0);
4796 }
4797 #endif  // !CONFIG_REALTIME_ONLY
4798